From 465132c5e343dbda4d13fdaaf8575d4896d738e9 Mon Sep 17 00:00:00 2001 From: eeb Date: Sat, 15 Oct 2005 16:46:21 +0000 Subject: [PATCH] * First cut iiblnd (compiles but untested) * Removed #if LNET_SINGLE_THREADED and replaced with #if !HAVE_LIBPTHREAD * Fixed LND module descriptions to say LND (not NAL) * viblnd cleanups (removed unused struct members, fixed some formatting etc) * minor cleanup in text buffer allocations (lnet/lnet/config.c) * format string fix in klnd/ptllnd.c * fixed lustre/utils/obd.c to work without libpthread/fork (disables --threads) --- lnet/autoconf/lustre-lnet.m4 | 1 - lnet/include/lnet/lib-lnet.h | 2 +- lnet/include/lnet/lib-types.h | 2 +- lnet/include/lnet/ptllnd.h | 4 +- lnet/klnds/gmlnd/gmlnd_module.c | 2 +- lnet/klnds/iiblnd/Makefile.in | 2 +- lnet/klnds/iiblnd/iiblnd.c | 1622 ++++++++------ lnet/klnds/iiblnd/iiblnd.h | 593 +++--- lnet/klnds/iiblnd/iiblnd_cb.c | 3886 +++++++++++++++++----------------- lnet/klnds/iiblnd/iiblnd_modparams.c | 160 ++ lnet/klnds/openiblnd/openiblnd.c | 2 +- lnet/klnds/ralnd/ralnd.c | 2 +- lnet/klnds/socklnd/socklnd.c | 4 +- lnet/klnds/viblnd/viblnd.c | 10 +- lnet/klnds/viblnd/viblnd.h | 11 +- lnet/klnds/viblnd/viblnd_cb.c | 16 +- lnet/lnet/api-ni.c | 8 +- lnet/lnet/config.c | 6 +- lnet/lnet/lib-eq.c | 4 +- lnet/lnet/lib-msg.c | 2 +- lnet/lnet/peer.c | 2 + lnet/ulnds/ptllnd/ptllnd.c | 4 +- lnet/ulnds/socklnd/procapi.c | 2 +- 23 files changed, 3514 insertions(+), 2833 deletions(-) create mode 100644 lnet/klnds/iiblnd/iiblnd_modparams.c diff --git a/lnet/autoconf/lustre-lnet.m4 b/lnet/autoconf/lustre-lnet.m4 index 8f0abf2..bd8d455 100644 --- a/lnet/autoconf/lustre-lnet.m4 +++ b/lnet/autoconf/lustre-lnet.m4 @@ -863,7 +863,6 @@ if test x$enable_liblustre = xyes ; then AC_DEFINE([HAVE_LIBPTHREAD], 1, [use libpthread]) else PTHREAD_LIBS="" - AC_DEFINE([LNET_SINGLE_THREADED], 1, [lnet single threaded]) fi AC_SUBST(PTHREAD_LIBS) fi diff --git a/lnet/include/lnet/lib-lnet.h b/lnet/include/lnet/lib-lnet.h index 8ff67d0..02c562e 100644 --- a/lnet/include/lnet/lib-lnet.h +++ b/lnet/include/lnet/lib-lnet.h @@ -43,7 +43,7 @@ static inline int lnet_md_exhausted (lnet_libmd_t *md) #define LNET_MUTEX_DOWN(m) mutex_down(m) #define LNET_MUTEX_UP(m) mutex_up(m) #else -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD #define LNET_SINGLE_THREADED_LOCK(l) \ do { \ LASSERT ((l) == 0); \ diff --git a/lnet/include/lnet/lib-types.h b/lnet/include/lnet/lib-types.h index dc063fb..dfbefdd 100644 --- a/lnet/include/lnet/lib-types.h +++ b/lnet/include/lnet/lib-types.h @@ -412,7 +412,7 @@ typedef struct struct semaphore ln_api_mutex; struct semaphore ln_lnd_mutex; #else -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD int ln_lock; int ln_api_mutex; int ln_lnd_mutex; diff --git a/lnet/include/lnet/ptllnd.h b/lnet/include/lnet/ptllnd.h index 61805d8..9fc47b8 100755 --- a/lnet/include/lnet/ptllnd.h +++ b/lnet/include/lnet/ptllnd.h @@ -31,6 +31,7 @@ /* NIDs are 64-bits on Lustre Portals */ #define FMT_NID LPX64 +#define FMT_PID "%d" /* When using Lustre Portals Lustre completion semantics are imlicit*/ #define PTL_MD_LUSTRE_COMPLETION_SEMANTICS 0 @@ -40,8 +41,9 @@ /* Explicit NULL function pointer for EQ handler */ #define PTL_EQ_HANDLER_NONE 0 -/* NIDs are integers on Lustre Portals */ +/* NIDs are integers on Cray Portals */ #define FMT_NID "%x" +#define FMT_PID "%d" /* When using Cray Portals this is defined in the Cray Portals Header*/ /*#define PTL_MD_LUSTRE_COMPLETION_SEMANTICS */ diff --git a/lnet/klnds/gmlnd/gmlnd_module.c b/lnet/klnds/gmlnd/gmlnd_module.c index b74e21d..91ecf04 100644 --- a/lnet/klnds/gmlnd/gmlnd_module.c +++ b/lnet/klnds/gmlnd/gmlnd_module.c @@ -119,5 +119,5 @@ module_init(gmnal_load); module_exit(gmnal_unload); MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel GM NAL v1.01"); +MODULE_DESCRIPTION("Kernel GM LND v1.01"); MODULE_LICENSE("GPL"); diff --git a/lnet/klnds/iiblnd/Makefile.in b/lnet/klnds/iiblnd/Makefile.in index 8733812..7ee9b64 100644 --- a/lnet/klnds/iiblnd/Makefile.in +++ b/lnet/klnds/iiblnd/Makefile.in @@ -1,5 +1,5 @@ MODULES := kiiblnd -kiiblnd-objs := iiblnd.o iiblnd_cb.o +kiiblnd-objs := iiblnd.o iiblnd_cb.o iiblnd_modparams.o EXTRA_POST_CFLAGS := @IIBCPPFLAGS@ diff --git a/lnet/klnds/iiblnd/iiblnd.c b/lnet/klnds/iiblnd/iiblnd.c index d8253b9..281b9dc 100644 --- a/lnet/klnds/iiblnd/iiblnd.c +++ b/lnet/klnds/iiblnd/iiblnd.c @@ -32,119 +32,435 @@ lnd_t the_kiblnd = { .lnd_recv = kibnal_recv, }; -kib_tunables_t kibnal_tunables; +kib_data_t kibnal_data; -kib_data_t kibnal_data = { - .kib_service_id = IBNAL_SERVICE_NUMBER, -}; - -#ifdef CONFIG_SYSCTL -#define IBNAL_SYSCTL 202 +__u32 +kibnal_cksum (void *ptr, int nob) +{ + char *c = ptr; + __u32 sum = 0; -#define IBNAL_SYSCTL_TIMEOUT 1 + while (nob-- > 0) + sum = ((sum << 1) | (sum >> 31)) + *c++; + + /* ensure I don't return 0 (== no checksum) */ + return (sum == 0) ? 1 : sum; +} -static ctl_table kibnal_ctl_table[] = { - {IBNAL_SYSCTL_TIMEOUT, "timeout", - &kibnal_tunables.kib_io_timeout, sizeof (int), - 0644, NULL, &proc_dointvec}, - { 0 } -}; +void +kibnal_init_msg(kib_msg_t *msg, int type, int body_nob) +{ + msg->ibm_type = type; + msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob; +} -static ctl_table kibnal_top_ctl_table[] = { - {IBNAL_SYSCTL, "iibnal", NULL, 0, 0555, kibnal_ctl_table}, - { 0 } -}; -#endif +void +kibnal_pack_msg(kib_msg_t *msg, int credits, lnet_nid_t dstnid, + __u64 dststamp, __u64 seq) +{ + /* CAVEAT EMPTOR! all message fields not set here should have been + * initialised previously. */ + msg->ibm_magic = IBNAL_MSG_MAGIC; + msg->ibm_version = IBNAL_MSG_VERSION; + /* ibm_type */ + msg->ibm_credits = credits; + /* ibm_nob */ + msg->ibm_cksum = 0; + msg->ibm_srcnid = lnet_ptlcompat_srcnid(kibnal_data.kib_ni->ni_nid, + dstnid); + msg->ibm_srcstamp = kibnal_data.kib_incarnation; + msg->ibm_dstnid = dstnid; + msg->ibm_dststamp = dststamp; + msg->ibm_seq = seq; + + if (*kibnal_tunables.kib_cksum) { + /* NB ibm_cksum zero while computing cksum */ + msg->ibm_cksum = kibnal_cksum(msg, msg->ibm_nob); + } +} -#ifdef unused void -print_service(IB_SERVICE_RECORD *service, char *tag, int rc) +kibnal_pack_connmsg(kib_msg_t *msg, int nob, int type, + lnet_nid_t dstnid, __u64 dststamp) { - char name[32]; - lnet_nid_t nid; + LASSERT (nob >= offsetof(kib_msg_t, ibm_u) + sizeof(kib_connparams_t)); - if (service == NULL) - { - CWARN("tag : %s\n" - "status : %d (NULL)\n", tag, rc); - return; + memset(msg, 0, nob); + kibnal_init_msg(msg, type, sizeof(kib_connparams_t)); + + msg->ibm_u.connparams.ibcp_queue_depth = IBNAL_MSG_QUEUE_SIZE; + msg->ibm_u.connparams.ibcp_max_msg_size = IBNAL_MSG_SIZE; + msg->ibm_u.connparams.ibcp_max_frags = IBNAL_MAX_RDMA_FRAGS; + + kibnal_pack_msg(msg, 0, dstnid, dststamp, 0); +} + +int +kibnal_unpack_msg(kib_msg_t *msg, int nob) +{ + const int hdr_size = offsetof(kib_msg_t, ibm_u); + __u32 msg_cksum; + int flip; + int msg_nob; +#if !IBNAL_USE_FMR + int i; + int n; +#endif + /* 6 bytes are enough to have received magic + version */ + if (nob < 6) { + CERROR("Short message: %d\n", nob); + return -EPROTO; + } + + if (msg->ibm_magic == IBNAL_MSG_MAGIC) { + flip = 0; + } else if (msg->ibm_magic == __swab32(IBNAL_MSG_MAGIC)) { + flip = 1; + } else { + CERROR("Bad magic: %08x\n", msg->ibm_magic); + return -EPROTO; + } + + if (msg->ibm_version != + (flip ? __swab16(IBNAL_MSG_VERSION) : IBNAL_MSG_VERSION)) { + CERROR("Bad version: %d\n", msg->ibm_version); + return -EPROTO; + } + + if (nob < hdr_size) { + CERROR("Short message: %d\n", nob); + return -EPROTO; + } + + msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob; + if (msg_nob > nob) { + CERROR("Short message: got %d, wanted %d\n", nob, msg_nob); + return -EPROTO; + } + + /* checksum must be computed with ibm_cksum zero and BEFORE anything + * gets flipped */ + msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum; + msg->ibm_cksum = 0; + if (msg_cksum != 0 && + msg_cksum != kibnal_cksum(msg, msg_nob)) { + CERROR("Bad checksum\n"); + return -EPROTO; } - strncpy (name, service->ServiceName, sizeof(name)-1); - name[sizeof(name)-1] = 0; - nid = *kibnal_service_nid_field(service); + msg->ibm_cksum = msg_cksum; - CWARN("tag : %s\n" - "status : %d\n" - "service id: "LPX64"\n" - "name : %s\n" - "NID : %s\n", tag, rc, - service->RID.ServiceID, name, - libcfs_nid2str(nid)); -} + if (flip) { + /* leave magic unflipped as a clue to peer endianness */ + __swab16s(&msg->ibm_version); + CLASSERT (sizeof(msg->ibm_type) == 1); + CLASSERT (sizeof(msg->ibm_credits) == 1); + msg->ibm_nob = msg_nob; + __swab64s(&msg->ibm_srcnid); + __swab64s(&msg->ibm_srcstamp); + __swab64s(&msg->ibm_dstnid); + __swab64s(&msg->ibm_dststamp); + __swab64s(&msg->ibm_seq); + } + + if (msg->ibm_srcnid == LNET_NID_ANY) { + CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid)); + return -EPROTO; + } + + switch (msg->ibm_type) { + default: + CERROR("Unknown message type %x\n", msg->ibm_type); + return -EPROTO; + + case IBNAL_MSG_NOOP: + break; + + case IBNAL_MSG_IMMEDIATE: + if (msg_nob < offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0])) { + CERROR("Short IMMEDIATE: %d(%d)\n", msg_nob, + (int)offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0])); + return -EPROTO; + } + break; + + case IBNAL_MSG_PUT_REQ: + if (msg_nob < hdr_size + sizeof(msg->ibm_u.putreq)) { + CERROR("Short PUT_REQ: %d(%d)\n", msg_nob, + (int)(hdr_size + sizeof(msg->ibm_u.putreq))); + return -EPROTO; + } + break; + + case IBNAL_MSG_PUT_ACK: +#if IBNAL_USE_FMR + if (msg_nob < hdr_size + sizeof(msg->ibm_u.putack)) { + CERROR("Short PUT_ACK: %d(%d)\n", msg_nob, + (int)(hdr_size + sizeof(msg->ibm_u.putack))); + return -EPROTO; + } + + if (flip) { + __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_addr); + __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nob); + __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key); + } +#else + if (flip) { + __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key); + __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nfrag); + } + + n = msg->ibm_u.putack.ibpam_rd.rd_nfrag; + if (n <= 0 || n > IBNAL_MAX_RDMA_FRAGS) { + CERROR("Bad PUT_ACK nfrags: %d, should be 0 < n <= %d\n", + n, IBNAL_MAX_RDMA_FRAGS); + return -EPROTO; + } + + if (msg_nob < offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n])) { + CERROR("Short PUT_ACK: %d(%d)\n", msg_nob, + (int)offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n])); + return -EPROTO; + } + + if (flip) { + for (i = 0; i < n; i++) { + __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_nob); + __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_addr); + } + } #endif + break; -static void -kibnal_service_setunset_done (void *arg, FABRIC_OPERATION_DATA *fod, - FSTATUS frc, uint32 madrc) + case IBNAL_MSG_GET_REQ: + if (msg_nob < hdr_size + sizeof(msg->ibm_u.get)) { + CERROR("Short GET_REQ: %d(%d)\n", msg_nob, + (int)(hdr_size + sizeof(msg->ibm_u.get))); + return -EPROTO; + } +#if IBNAL_USE_FMR + if (flip) { + __swab64s(&msg->ibm_u.get.ibgm_rd.rd_addr); + __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nob); + __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key); + } +#else + if (flip) { + __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key); + __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nfrag); + } + + n = msg->ibm_u.get.ibgm_rd.rd_nfrag; + if (n <= 0 || n > IBNAL_MAX_RDMA_FRAGS) { + CERROR("Bad GET_REQ nfrags: %d, should be 0 < n <= %d\n", + n, IBNAL_MAX_RDMA_FRAGS); + return -EPROTO; + } + + if (msg_nob < offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n])) { + CERROR("Short GET_REQ: %d(%d)\n", msg_nob, + (int)offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n])); + return -EPROTO; + } + + if (flip) + for (i = 0; i < msg->ibm_u.get.ibgm_rd.rd_nfrag; i++) { + __swab32s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_nob); + __swab64s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_addr); + } +#endif + break; + + case IBNAL_MSG_PUT_NAK: + case IBNAL_MSG_PUT_DONE: + case IBNAL_MSG_GET_DONE: + if (msg_nob < hdr_size + sizeof(msg->ibm_u.completion)) { + CERROR("Short RDMA completion: %d(%d)\n", msg_nob, + (int)(hdr_size + sizeof(msg->ibm_u.completion))); + return -EPROTO; + } + if (flip) + __swab32s(&msg->ibm_u.completion.ibcm_status); + break; + + case IBNAL_MSG_CONNREQ: + case IBNAL_MSG_CONNACK: + if (msg_nob < hdr_size + sizeof(msg->ibm_u.connparams)) { + CERROR("Short connreq/ack: %d(%d)\n", msg_nob, + (int)(hdr_size + sizeof(msg->ibm_u.connparams))); + return -EPROTO; + } + if (flip) { + __swab32s(&msg->ibm_u.connparams.ibcp_queue_depth); + __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size); + __swab32s(&msg->ibm_u.connparams.ibcp_max_frags); + } + break; + } + return 0; +} + +IB_HANDLE +kibnal_create_cep(lnet_nid_t nid) { - *(FSTATUS *)arg = frc; - up (&kibnal_data.kib_nid_signal); + FSTATUS frc; + __u32 u32val; + IB_HANDLE cep; + + cep = iibt_cm_create_cep(CM_RC_TYPE); + if (cep == NULL) { + CERROR ("Can't create CEP for %s\n", + (nid == LNET_NID_ANY) ? "listener" : + libcfs_nid2str(nid)); + return NULL; + } + + if (nid == LNET_NID_ANY) { + u32val = 1; + frc = iibt_cm_modify_cep(cep, CM_FLAG_ASYNC_ACCEPT, + (char *)&u32val, sizeof(u32val), 0); + if (frc != FSUCCESS) { + CERROR("Can't set async_accept: %d\n", frc); + goto failed; + } + + u32val = 0; /* sets system max */ + frc = iibt_cm_modify_cep(cep, CM_FLAG_LISTEN_BACKLOG, + (char *)&u32val, sizeof(u32val), 0); + if (frc != FSUCCESS) { + CERROR("Can't set listen backlog: %d\n", frc); + goto failed; + } + } + + u32val = 1; + frc = iibt_cm_modify_cep(cep, CM_FLAG_TIMEWAIT_CALLBACK, + (char *)&u32val, sizeof(u32val), 0); + if (frc != FSUCCESS) { + CERROR("Can't set timewait_callback for %s: %d\n", + (nid == LNET_NID_ANY) ? "listener" : + libcfs_nid2str(nid), frc); + goto failed; + } + + return cep; + + failed: + iibt_cm_destroy_cep(cep); + return NULL; } #if IBNAL_CHECK_ADVERT -static void +void kibnal_service_query_done (void *arg, QUERY *qry, QUERY_RESULT_VALUES *qry_result) { - FSTATUS frc = qry_result->Status; + int *rcp = arg; + FSTATUS frc = qry_result->Status; + SERVICE_RECORD_RESULTS *svc_rslt; + SERVICE_RECORD *svc; + lnet_nid_t nid; + + if (frc != FSUCCESS || qry_result->ResultDataSize == 0) { + CERROR("Error checking advert: status %d data size %d\n", + frc, qry_result->ResultDataSize); + *rcp = -EIO; + goto out; + } + + svc_rslt = (SERVICE_RECORD_RESULTS *)qry_result->QueryResult; - if (frc != FSUCCESS && - qry_result->ResultDataSize == 0) - frc = FERROR; + if (svc_rslt->NumServiceRecords < 1) { + CERROR("Check advert: %d records\n", + svc->NumServiceRecords); + *rcp = -ENOENT; + goto out; + } + + svc = &svc_rslt->ServiceRecords[0]; + nid = le64_to_cpu(*kibnal_service_nid_field(svc)); - *(FSTATUS *)arg = frc; - up (&kibnal_data.kib_nid_signal); + if (nid != kibnal_data.kib_ni->ni_nid) { + CERROR("Check advert: Bad NID %s (%s expected)\n", + nid, kibnal_data.kib_ni->ni_nid); + *rcp = -EINVAL; + goto out; + } + + if (svc->RID.ServiceID != *kibnal_tunables.kib_service_number) { + CERROR("Check advert: Bad ServiceID "LPX64" ("LPX64" expected)\n", + svc->RID.ServiceID, + *kibnal_tunables.kib_service_number); + *rcp = -EINVAL; + goto out; + } + + if (svc->RID.ServiceGID.Type.Global.InterfaceID != + kibnal_data.kib_port_guid) { + CERROR("Check advert: Bad GUID "LPX64" ("LPX64" expected)\n", + svc->RID.ServiceGID.Type.Global.InterfaceID, + kibnal_data.kib_port_guid); + *rcp = -EINVAL; + goto out; + } + + if (svc->RID.ServiceP_Key != kibnal_data.kib_port_pkey) { + CERROR("Check advert: Bad PKEY %04x (%04x expected)\n", + svc->RID.ServiceP_Key, kibnal_data.kib_port_pkey); + *rcp = -EINVAL; + goto out; + } + + CDEBUG(D_WARNING, "Check advert OK\n"); + *rcp = 0; + + out: + up (&kibnal_data.kib_listener_signal); } -static void +int kibnal_check_advert (void) { - QUERY *qry; - IB_SERVICE_RECORD *svc; - FSTATUS frc; - FSTATUS frc2; + /* single-threaded */ + static QUERY qry; - LIBCFS_ALLOC(qry, sizeof(*qry)); - if (qry == NULL) - return; + FSTATUS frc; + int rc; - memset (qry, 0, sizeof(*qry)); - qry->InputType = InputTypeServiceRecord; - qry->OutputType = OutputTypeServiceRecord; - qry->InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK; - svc = &qry->InputValue.ServiceRecordValue.ServiceRecord; - kibnal_set_service_keys(svc, kibnal_data.kib_ni->ni_nid); + memset (&qry, 0, sizeof(qry)); + qry.InputType = InputTypeServiceRecord; + qry.OutputType = OutputTypeServiceRecord; + kibnal_set_service_keys(&qry.InputValue.ServiceRecordValue.ServiceRecord, + kibnal_data.kib_ni->ni_nid); + qry.InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK; - frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, + frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, kibnal_data.kib_port_guid, - qry, + &qry, kibnal_service_query_done, - NULL, &frc2); - if (frc != FSUCCESS && frc != FPENDING) { + &kibnal_data.kib_sdretry, + &rc); + if (frc != FPENDING) { CERROR ("Immediate error %d checking SM service\n", frc); - } else { - down (&kibnal_data.kib_nid_signal); - frc = frc2; - - if (frc != 0) - CERROR ("Error %d checking SM service\n", rc); + return -EIO; } - - return (rc); + + down (&kibnal_data.kib_listener_signal); + + if (rc != 0) + CERROR ("Error %d checking SM service\n", rc); + return rc; +} +#else +int +kibnal_check_advert(void) +{ + return 0; } #endif -static void fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type) +void +kibnal_fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type) { IB_SERVICE_RECORD *svc; @@ -152,7 +468,7 @@ static void fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type) fod->Type = type; svc = &fod->Value.ServiceRecordValue.ServiceRecord; - svc->RID.ServiceID = kibnal_data.kib_service_id; + svc->RID.ServiceID = *kibnal_tunables.kib_service_number; svc->RID.ServiceGID.Type.Global.InterfaceID = kibnal_data.kib_port_guid; svc->RID.ServiceGID.Type.Global.SubnetPrefix = DEFAULT_SUBNET_PREFIX; svc->RID.ServiceP_Key = kibnal_data.kib_port_pkey; @@ -161,86 +477,96 @@ static void fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type) kibnal_set_service_keys(svc, kibnal_data.kib_ni->ni_nid); } -static int -kibnal_advertise (void) +void +kibnal_service_setunset_done (void *arg, FABRIC_OPERATION_DATA *fod, + FSTATUS frc, uint32 madrc) { - FABRIC_OPERATION_DATA *fod; - IB_SERVICE_RECORD *svc; - FSTATUS frc; - FSTATUS frc2; - - LASSERT (kibnal_data.kib_ni->ni_nid != LNET_NID_ANY); + *(FSTATUS *)arg = frc; + up (&kibnal_data.kib_listener_signal); +} - LIBCFS_ALLOC(fod, sizeof(*fod)); - if (fod == NULL) - return (-ENOMEM); +int +kibnal_advertise (void) +{ + /* Single threaded here */ + static FABRIC_OPERATION_DATA fod; + + IB_SERVICE_RECORD *svc = &fod.Value.ServiceRecordValue.ServiceRecord; + FSTATUS frc; + FSTATUS frc2; + + if (strlen(*kibnal_tunables.kib_service_name) >= + sizeof(svc->ServiceName)) { + CERROR("Service name '%s' too long (%d chars max)\n", + *kibnal_tunables.kib_service_name, + sizeof(svc->ServiceName) - 1); + return -EINVAL; + } - fill_fod(fod, FabOpSetServiceRecord); - svc = &fod->Value.ServiceRecordValue.ServiceRecord; + kibnal_fill_fod(&fod, FabOpSetServiceRecord); CDEBUG(D_NET, "Advertising service id "LPX64" %s:%s\n", - svc->RID.ServiceID, - svc->ServiceName, - libcfs_nid2str(*kibnal_service_nid_field(svc))); + svc->RID.ServiceID, svc->ServiceName, + libcfs_nid2str(le64_to_cpu(*kibnal_service_nid_field(svc)))); frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd, kibnal_data.kib_port_guid, - fod, kibnal_service_setunset_done, - NULL, &frc2); + &fod, + kibnal_service_setunset_done, + &kibnal_data.kib_sdretry, + &frc2); if (frc != FSUCCESS && frc != FPENDING) { CERROR ("Immediate error %d advertising NID %s\n", frc, libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); - goto out; + return -EIO; } - down (&kibnal_data.kib_nid_signal); + down (&kibnal_data.kib_listener_signal); frc = frc2; - if (frc != FSUCCESS) - CERROR ("Error %d advertising BUD %s\n", - frc, libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); -out: - LIBCFS_FREE(fod, sizeof(*fod)); - return (frc == FSUCCESS) ? 0 : -EINVAL; + if (frc == FSUCCESS) + return 0; + + CERROR ("Error %d advertising %s\n", + frc, libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); + return -EIO; } -static void +void kibnal_unadvertise (int expect_success) { - FABRIC_OPERATION_DATA *fod; - IB_SERVICE_RECORD *svc; - FSTATUS frc; - FSTATUS frc2; + /* single threaded */ + static FABRIC_OPERATION_DATA fod; - LASSERT (kibnal_data.kib_ni->ni_nid != LNET_NID_ANY); + IB_SERVICE_RECORD *svc = &fod.Value.ServiceRecordValue.ServiceRecord; + FSTATUS frc; + FSTATUS frc2; - LIBCFS_ALLOC(fod, sizeof(*fod)); - if (fod == NULL) - return; + LASSERT (kibnal_data.kib_ni->ni_nid != LNET_NID_ANY); - fill_fod(fod, FabOpDeleteServiceRecord); - svc = &fod->Value.ServiceRecordValue.ServiceRecord; + kibnal_fill_fod(&fod, FabOpDeleteServiceRecord); CDEBUG(D_NET, "Unadvertising service %s:%s\n", svc->ServiceName, - libcfs_nid2str(*kibnal_service_nid_field(svc))); + libcfs_nid2str(le64_to_cpu(*kibnal_service_nid_field(svc)))); frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd, kibnal_data.kib_port_guid, - fod, kibnal_service_setunset_done, - NULL, &frc2); - + &fod, + kibnal_service_setunset_done, + &kibnal_data.kib_sdretry, + &frc2); if (frc != FSUCCESS && frc != FPENDING) { CERROR ("Immediate error %d unadvertising NID %s\n", frc, libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); - goto out; + return; } - down (&kibnal_data.kib_nid_signal); + down (&kibnal_data.kib_listener_signal); - if ((frc2 == FSUCCESS) == !!expect_success) - goto out; + if ((frc2 -= FSUCCESS) == !!expect_success) + return; if (expect_success) CERROR("Error %d unadvertising NID %s\n", @@ -248,118 +574,104 @@ kibnal_unadvertise (int expect_success) else CWARN("Removed conflicting NID %s\n", libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); - out: - LIBCFS_FREE(fod, sizeof(*fod)); } -static int -kibnal_set_mynid(lnet_nid_t nid) +void +kibnal_stop_listener(int normal_shutdown) { - struct timeval tv; - int rc; + /* NB this also disables peer creation and destroys all existing + * peers */ + IB_HANDLE cep = kibnal_data.kib_listener_cep; + unsigned long flags; FSTATUS frc; - CDEBUG(D_IOCTL, "setting mynid to %s (old nid=%s)\n", - libcfs_nid2str(nid), - libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); + LASSERT (cep != NULL); - do_gettimeofday(&tv); + kibnal_unadvertise(normal_shutdown); - down (&kibnal_data.kib_nid_mutex); + frc = iibt_cm_cancel(cep); + if (frc != FSUCCESS && frc != FPENDING) + CERROR ("Error %d stopping listener\n", frc); - if (nid == kibnal_data.kib_ni->ni_nid) { - /* no change of NID */ - up (&kibnal_data.kib_nid_mutex); - return (0); - } + down(&kibnal_data.kib_listener_signal); - CDEBUG(D_NET, "NID %s(%s)\n", - libcfs_nid2str(kibnal_data.kib_ni->ni_nid), - libcfs_nid2str(nid)); - - if (kibnal_data.kib_ni->ni_nid != LNET_NID_ANY) { + frc = iibt_cm_destroy_cep(cep); + if (frc != FSUCCESS) + CERROR ("Error %d destroying listener CEP\n", frc); - kibnal_unadvertise (1); + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + /* This assignment disables peer creation */ + kibnal_data.kib_listener_cep = NULL; + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); - frc = iibt_cm_cancel(kibnal_data.kib_cep); - if (frc != FSUCCESS && frc != FPENDING) - CERROR ("Error %d stopping listener\n", frc); + /* Start to tear down any peers created while the listener was + * running */ + kibnal_del_peer(LNET_NID_ANY); +} - frc = iibt_cm_destroy_cep(kibnal_data.kib_cep); - if (frc != FSUCCESS) - CERROR ("Error %d destroying CEP\n", frc); +int +kibnal_start_listener(void) +{ + /* NB this also enables peer creation */ - kibnal_data.kib_cep = NULL; - } - - kibnal_data.kib_ni->ni_nid = nid; - kibnal_data.kib_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; - - /* Delete all existing peers and their connections after new - * NID/incarnation set to ensure no old connections in our brave - * new world. */ - kibnal_del_peer(LNET_NID_ANY); + IB_HANDLE cep; + CM_LISTEN_INFO info; + unsigned long flags; + int rc; + FSTATUS frc; + __u32 u32val; - if (kibnal_data.kib_ni->ni_nid == LNET_NID_ANY) { - /* No new NID to install */ - up (&kibnal_data.kib_nid_mutex); - return (0); - } + LASSERT (kibnal_data.kib_listener_cep == NULL); + init_MUTEX_LOCKED (&kibnal_data.kib_listener_signal); /* remove any previous advert (crashed node etc) */ kibnal_unadvertise(0); - kibnal_data.kib_cep = iibt_cm_create_cep(CM_RC_TYPE); - if (kibnal_data.kib_cep == NULL) { - CERROR ("Can't create CEP\n"); - rc = -ENOMEM; - } else { - CM_LISTEN_INFO info; - memset (&info, 0, sizeof(info)); - info.ListenAddr.EndPt.SID = kibnal_data.kib_service_id; - - frc = iibt_cm_listen(kibnal_data.kib_cep, &info, - kibnal_listen_callback, NULL); - if (frc != FSUCCESS && frc != FPENDING) { - CERROR ("iibt_cm_listen error: %d\n", frc); - rc = -EINVAL; - } else { - rc = 0; - } - } - - if (rc == 0) { - rc = kibnal_advertise(); - if (rc == 0) { -#if IBNAL_CHECK_ADVERT - kibnal_check_advert(); -#endif - up (&kibnal_data.kib_nid_mutex); - return (0); - } - - iibt_cm_cancel (kibnal_data.kib_cep); - iibt_cm_destroy_cep (kibnal_data.kib_cep); - /* remove any peers that sprung up while I failed to - * advertise myself */ - kibnal_del_peer(LNET_NID_ANY); + cep = kibnal_create_cep(LNET_NID_ANY); + if (cep == NULL) + return -ENOMEM; + + memset (&info, 0, sizeof(info)); + info.ListenAddr.EndPt.SID = *kibnal_tunables.kib_service_number; + + frc = iibt_cm_listen(cep, &info, kibnal_listen_callback, NULL); + if (frc != FSUCCESS && frc != FPENDING) { + CERROR ("iibt_cm_listen error: %d\n", frc); + + iibt_cm_destroy_cep(cep); + return -EIO; } - kibnal_data.kib_ni->ni_nid = LNET_NID_ANY; - up (&kibnal_data.kib_nid_mutex); - return (rc); + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + /* This assignment enables peer creation */ + kibnal_data.kib_listener_cep = cep; + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); + + rc = kibnal_advertise(); + if (rc == 0) + rc = kibnal_check_advert(); + + if (rc == 0) + return 0; + + kibnal_stop_listener(0); + return rc; } -kib_peer_t * -kibnal_create_peer (lnet_nid_t nid) +int +kibnal_create_peer (kib_peer_t **peerp, lnet_nid_t nid) { - kib_peer_t *peer; + kib_peer_t *peer; + unsigned long flags; + int rc; LASSERT (nid != LNET_NID_ANY); LIBCFS_ALLOC (peer, sizeof (*peer)); - if (peer == NULL) - return (NULL); + if (peer == NULL) { + CERROR("Cannot allocate peer\n"); + return -ENOMEM; + } memset(peer, 0, sizeof(*peer)); /* zero flags etc */ @@ -372,8 +684,31 @@ kibnal_create_peer (lnet_nid_t nid) peer->ibp_reconnect_interval = 0; /* OK to connect at any time */ - atomic_inc (&kibnal_data.kib_npeers); - return (peer); + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + + if (atomic_read(&kibnal_data.kib_npeers) >= + *kibnal_tunables.kib_concurrent_peers) { + rc = -EOVERFLOW; /* !! but at least it distinguishes */ + } else if (kibnal_data.kib_listener_cep == NULL) { + rc = -ESHUTDOWN; /* shutdown has started */ + } else { + rc = 0; + /* npeers only grows with the global lock held */ + atomic_inc(&kibnal_data.kib_npeers); + } + + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); + + if (rc != 0) { + CERROR("Can't create peer: %s\n", + (rc == -ESHUTDOWN) ? "shutting down" : + "too many peers"); + LIBCFS_FREE(peer, sizeof(*peer)); + } else { + *peerp = peer; + } + + return rc; } void @@ -424,21 +759,6 @@ kibnal_find_peer_locked (lnet_nid_t nid) return (NULL); } -kib_peer_t * -kibnal_get_peer (lnet_nid_t nid) -{ - kib_peer_t *peer; - unsigned long flags; - - read_lock_irqsave(&kibnal_data.kib_global_lock, flags); - peer = kibnal_find_peer_locked (nid); - if (peer != NULL) /* +1 ref for caller? */ - kib_peer_addref(peer); - read_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); - - return (peer); -} - void kibnal_unlink_peer_locked (kib_peer_t *peer) { @@ -448,10 +768,10 @@ kibnal_unlink_peer_locked (kib_peer_t *peer) LASSERT (kibnal_peer_active(peer)); list_del_init (&peer->ibp_list); /* lose peerlist's ref */ - kib_peer_decref(peer); + kibnal_peer_decref(peer); } -static int +int kibnal_get_peer_info (int index, lnet_nid_t *nidp, int *persistencep) { kib_peer_t *peer; @@ -486,25 +806,26 @@ kibnal_get_peer_info (int index, lnet_nid_t *nidp, int *persistencep) return (-ENOENT); } -static int +int kibnal_add_persistent_peer (lnet_nid_t nid) { unsigned long flags; kib_peer_t *peer; kib_peer_t *peer2; + int rc; if (nid == LNET_NID_ANY) return (-EINVAL); - peer = kibnal_create_peer (nid); - if (peer == NULL) - return (-ENOMEM); + rc = kibnal_create_peer(&peer, nid); + if (rc != 0) + return rc; write_lock_irqsave (&kibnal_data.kib_global_lock, flags); peer2 = kibnal_find_peer_locked (nid); if (peer2 != NULL) { - kib_peer_decref (peer); + kibnal_peer_decref (peer); peer = peer2; } else { /* peer table takes existing ref on peer */ @@ -518,7 +839,7 @@ kibnal_add_persistent_peer (lnet_nid_t nid) return (0); } -static void +void kibnal_del_peer_locked (kib_peer_t *peer) { struct list_head *ctmp; @@ -583,7 +904,7 @@ kibnal_del_peer (lnet_nid_t nid) return (rc); } -static kib_conn_t * +kib_conn_t * kibnal_get_conn_by_idx (int index) { kib_peer_t *peer; @@ -608,11 +929,7 @@ kibnal_get_conn_by_idx (int index) continue; conn = list_entry (ctmp, kib_conn_t, ibc_list); - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); + kibnal_conn_addref(conn); read_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); return (conn); @@ -624,13 +941,90 @@ kibnal_get_conn_by_idx (int index) return (NULL); } +int +kibnal_conn_rts(kib_conn_t *conn, + __u32 qpn, __u8 resp_res, __u8 init_depth, __u32 psn) +{ + IB_PATH_RECORD *path = &conn->ibc_cvars->cv_path; + IB_HANDLE qp = conn->ibc_qp; + IB_QP_ATTRIBUTES_MODIFY modify_attr; + FSTATUS frc; + int rc; + + if (resp_res > kibnal_data.kib_hca_attrs.MaxQPResponderResources) + resp_res = kibnal_data.kib_hca_attrs.MaxQPResponderResources; + + if (init_depth > kibnal_data.kib_hca_attrs.MaxQPInitiatorDepth) + init_depth = kibnal_data.kib_hca_attrs.MaxQPInitiatorDepth; + + modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { + .RequestState = QPStateReadyToRecv, + .RecvPSN = IBNAL_STARTING_PSN, + .DestQPNumber = qpn, + .ResponderResources = resp_res, + .MinRnrTimer = UsecToRnrNakTimer(2000), /* 20 ms */ + .Attrs = (IB_QP_ATTR_RECVPSN | + IB_QP_ATTR_DESTQPNUMBER | + IB_QP_ATTR_RESPONDERRESOURCES | + IB_QP_ATTR_DESTAV | + IB_QP_ATTR_PATHMTU | + IB_QP_ATTR_MINRNRTIMER), + }; + GetAVFromPath(0, path, &modify_attr.PathMTU, NULL, + &modify_attr.DestAV); + + frc = iibt_qp_modify(qp, &modify_attr, NULL); + if (frc != FSUCCESS) { + CERROR("Can't set QP %s ready to receive: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + return -EIO; + } + + rc = kibnal_post_receives(conn); + if (rc != 0) { + CERROR("Can't post receives for %s: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); + return rc; + } + + modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { + .RequestState = QPStateReadyToSend, + .FlowControl = TRUE, + .InitiatorDepth = init_depth, + .SendPSN = psn, + .LocalAckTimeout = path->PktLifeTime + 2, /* 2 or 1? */ + .RetryCount = IBNAL_RETRY, + .RnrRetryCount = IBNAL_RNR_RETRY, + .Attrs = (IB_QP_ATTR_FLOWCONTROL | + IB_QP_ATTR_INITIATORDEPTH | + IB_QP_ATTR_SENDPSN | + IB_QP_ATTR_LOCALACKTIMEOUT | + IB_QP_ATTR_RETRYCOUNT | + IB_QP_ATTR_RNRRETRYCOUNT), + }; + + frc = iibt_qp_modify(qp, &modify_attr, NULL); + if (frc != FSUCCESS) { + CERROR("Can't set QP %s ready to send: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + return -EIO; + } + + frc = iibt_qp_query(conn->ibc_qp, &conn->ibc_cvars->cv_qpattrs, NULL); + if (frc != FSUCCESS) { + CERROR ("Can't query QP %s attributes: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + return -EIO; + } + + return 0; +} + kib_conn_t * -kibnal_create_conn (void) +kibnal_create_conn (lnet_nid_t nid) { kib_conn_t *conn; int i; - __u64 vaddr = 0; - __u64 vaddr_base; int page_offset; int ipage; int rc; @@ -642,13 +1036,16 @@ kibnal_create_conn (void) LIBCFS_ALLOC (conn, sizeof (*conn)); if (conn == NULL) { - CERROR ("Can't allocate connection\n"); + CERROR ("Can't allocate connection for %s\n", + libcfs_nid2str(nid)); return (NULL); } /* zero flags, NULL pointers etc... */ memset (conn, 0, sizeof (*conn)); + conn->ibc_state = IBNAL_CONN_INIT_NOTHING; + INIT_LIST_HEAD (&conn->ibc_early_rxs); INIT_LIST_HEAD (&conn->ibc_tx_queue); INIT_LIST_HEAD (&conn->ibc_active_txs); spin_lock_init (&conn->ibc_lock); @@ -656,34 +1053,39 @@ kibnal_create_conn (void) atomic_inc (&kibnal_data.kib_nconns); /* well not really, but I call destroy() on failure, which decrements */ - LIBCFS_ALLOC (conn->ibc_rxs, IBNAL_RX_MSGS * sizeof (kib_rx_t)); - if (conn->ibc_rxs == NULL) + LIBCFS_ALLOC (conn->ibc_cvars, sizeof (*conn->ibc_cvars)); + if (conn->ibc_cvars == NULL) { + CERROR ("Can't allocate connvars for %s\n", + libcfs_nid2str(nid)); goto failed; - memset (conn->ibc_rxs, 0, IBNAL_RX_MSGS * sizeof(kib_rx_t)); + } + memset(conn->ibc_cvars, 0, sizeof (*conn->ibc_cvars)); - rc = kibnal_alloc_pages(&conn->ibc_rx_pages, IBNAL_RX_MSG_PAGES, 1); - if (rc != 0) + LIBCFS_ALLOC(conn->ibc_rxs, IBNAL_RX_MSGS * sizeof (kib_rx_t)); + if (conn->ibc_rxs == NULL) { + CERROR("Cannot allocate RX descriptors for %s\n", + libcfs_nid2str(nid)); goto failed; + } + memset (conn->ibc_rxs, 0, IBNAL_RX_MSGS * sizeof(kib_rx_t)); - vaddr_base = vaddr = conn->ibc_rx_pages->ibp_vaddr; - + rc = kibnal_alloc_pages(&conn->ibc_rx_pages, IBNAL_RX_MSG_PAGES); + if (rc != 0) { + CERROR("Can't allocate RX buffers for %s\n", + libcfs_nid2str(nid)); + goto failed; + } + for (i = ipage = page_offset = 0; i < IBNAL_RX_MSGS; i++) { struct page *page = conn->ibc_rx_pages->ibp_pages[ipage]; - kib_rx_t *rx = &conn->ibc_rxs[i]; + kib_rx_t *rx = &conn->ibc_rxs[i]; rx->rx_conn = conn; rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) + page_offset); - if (kibnal_whole_mem()) - rx->rx_vaddr = kibnal_page2phys(page) + - page_offset + - kibnal_data.kib_md.md_addr; - else - rx->rx_vaddr = vaddr; - - vaddr += IBNAL_MSG_SIZE; - LASSERT (vaddr <= vaddr_base + IBNAL_RX_MSG_BYTES); + rx->rx_hca_msg = kibnal_data.kib_whole_mem.md_addr + + kibnal_page2phys(page) + page_offset; page_offset += IBNAL_MSG_SIZE; LASSERT (page_offset <= PAGE_SIZE); @@ -697,7 +1099,7 @@ kibnal_create_conn (void) params.qp_create = (IB_QP_ATTRIBUTES_CREATE) { .Type = QPTypeReliableConnected, - .SendQDepth = IBNAL_TX_MAX_SG * + .SendQDepth = (IBNAL_MAX_RDMA_FRAGS + 1) * IBNAL_MSG_QUEUE_SIZE, .RecvQDepth = IBNAL_MSG_QUEUE_SIZE, .SendDSListDepth = 1, @@ -708,14 +1110,14 @@ kibnal_create_conn (void) .SendSignaledCompletions = TRUE, }; frc = iibt_qp_create(kibnal_data.kib_hca, ¶ms.qp_create, NULL, - &conn->ibc_qp, &conn->ibc_qp_attrs); - if (rc != 0) { - CERROR ("Failed to create queue pair: %d\n", rc); + &conn->ibc_qp, &conn->ibc_cvars->cv_qpattrs); + if (frc != 0) { + CERROR ("Can't create QP %s: %d\n", libcfs_nid2str(nid), frc); goto failed; } /* Mark QP created */ - conn->ibc_state = IBNAL_CONN_INIT_QP; + kibnal_set_conn_state(conn, IBNAL_CONN_INIT_QP); params.qp_attr = (IB_QP_ATTRIBUTES_MODIFY) { .RequestState = QPStateInit, @@ -724,16 +1126,24 @@ kibnal_create_conn (void) IB_QP_ATTR_ACCESSCONTROL), .PortGUID = kibnal_data.kib_port_guid, .PkeyIndex = 0, - .AccessControl = { + .AccessControl = { .s = { .RdmaWrite = 1, .RdmaRead = 1, }, }, }; - rc = iibt_qp_modify(conn->ibc_qp, ¶ms.qp_attr, NULL); - if (rc != 0) { - CERROR ("Failed to modify queue pair: %d\n", rc); + frc = iibt_qp_modify(conn->ibc_qp, ¶ms.qp_attr, NULL); + if (frc != 0) { + CERROR ("Can't set QP %s state to INIT: %d\n", + libcfs_nid2str(nid), frc); + goto failed; + } + + frc = iibt_qp_query(conn->ibc_qp, &conn->ibc_cvars->cv_qpattrs, NULL); + if (frc != FSUCCESS) { + CERROR ("Can't query QP %s attributes: %d\n", + libcfs_nid2str(nid), frc); goto failed; } @@ -749,42 +1159,46 @@ kibnal_create_conn (void) void kibnal_destroy_conn (kib_conn_t *conn) { - int rc; + int rc; FSTATUS frc; + + LASSERT (!in_interrupt()); CDEBUG (D_NET, "connection %p\n", conn); LASSERT (atomic_read (&conn->ibc_refcount) == 0); + LASSERT (list_empty(&conn->ibc_early_rxs)); LASSERT (list_empty(&conn->ibc_tx_queue)); LASSERT (list_empty(&conn->ibc_active_txs)); LASSERT (conn->ibc_nsends_posted == 0); - LASSERT (conn->ibc_connreq == NULL); switch (conn->ibc_state) { - case IBNAL_CONN_DISCONNECTED: - /* called after connection sequence initiated */ - /* fall through */ - - case IBNAL_CONN_INIT_QP: - /* _destroy includes an implicit Reset of the QP which - * discards posted work */ - rc = iibt_qp_destroy(conn->ibc_qp); - if (rc != 0) - CERROR("Can't destroy QP: %d\n", rc); - /* fall through */ - case IBNAL_CONN_INIT_NOTHING: + case IBNAL_CONN_INIT_QP: + case IBNAL_CONN_DISCONNECTED: break; default: - LASSERT (0); + /* conn must either have never engaged with the CM, or have + * completely disengaged from it */ + CERROR("Bad conn %s state %d\n", + (conn->ibc_peer) == NULL ? "" : + libcfs_nid2str(conn->ibc_peer->ibp_nid), conn->ibc_state); + LBUG(); } if (conn->ibc_cep != NULL) { frc = iibt_cm_destroy_cep(conn->ibc_cep); - if (frc != 0) - CERROR("Can't destroy CEP %p: %d\n", conn->ibc_cep, - frc); + if (frc != FSUCCESS) + CERROR("Error destroying CEP %p: %d\n", + conn->ibc_cep, frc); + } + + if (conn->ibc_qp != NULL) { + frc = iibt_qp_destroy(conn->ibc_qp); + if (frc != FSUCCESS) + CERROR("Error destroying QP %p: %d\n", + conn->ibc_qp, frc); } if (conn->ibc_rx_pages != NULL) @@ -794,48 +1208,18 @@ kibnal_destroy_conn (kib_conn_t *conn) LIBCFS_FREE(conn->ibc_rxs, IBNAL_RX_MSGS * sizeof(kib_rx_t)); + if (conn->ibc_cvars != NULL) + LIBCFS_FREE(conn->ibc_cvars, sizeof(*conn->ibc_cvars)); + if (conn->ibc_peer != NULL) - kib_peer_decref(conn->ibc_peer); + kibnal_peer_decref(conn->ibc_peer); LIBCFS_FREE(conn, sizeof (*conn)); atomic_dec(&kibnal_data.kib_nconns); - - if (atomic_read (&kibnal_data.kib_nconns) == 0 && - kibnal_data.kib_shutdown) { - /* I just nuked the last connection on shutdown; wake up - * everyone so they can exit. */ - wake_up_all(&kibnal_data.kib_sched_waitq); - wake_up_all(&kibnal_data.kib_connd_waitq); - } -} - -void -kibnal_put_conn (kib_conn_t *conn) -{ - unsigned long flags; - - CDEBUG (D_NET, "putting conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - - LASSERT (atomic_read (&conn->ibc_refcount) > 0); - if (!atomic_dec_and_test (&conn->ibc_refcount)) - return; - - /* must disconnect before dropping the final ref */ - LASSERT (conn->ibc_state == IBNAL_CONN_DISCONNECTED); - - spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); - - list_add (&conn->ibc_list, &kibnal_data.kib_connd_conns); - wake_up (&kibnal_data.kib_connd_waitq); - - spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); } -static int +int kibnal_close_peer_conns_locked (kib_peer_t *peer, int why) { kib_conn_t *conn; @@ -878,7 +1262,7 @@ kibnal_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation) return (count); } -static int +int kibnal_close_matching_conns (lnet_nid_t nid) { unsigned long flags; @@ -935,7 +1319,7 @@ kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) switch(cmd) { case IOC_LIBCFS_GET_PEER: { lnet_nid_t nid = 0; - int share_count = 0; + int share_count = 0; rc = kibnal_get_peer_info(data->ioc_count, &nid, &share_count); @@ -959,7 +1343,7 @@ kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) else { rc = 0; data->ioc_nid = conn->ibc_peer->ibp_nid; - kibnal_put_conn (conn); + kibnal_conn_decref(conn); } break; } @@ -968,10 +1352,14 @@ kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) break; } case IOC_LIBCFS_REGISTER_MYNID: { - if (data->ioc_nid == LNET_NID_ANY) + if (ni->ni_nid == data->ioc_nid) { + rc = 0; + } else { + CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n", + libcfs_nid2str(data->ioc_nid), + libcfs_nid2str(ni->ni_nid)); rc = -EINVAL; - else - rc = kibnal_set_mynid (data->ioc_nid); + } break; } } @@ -983,15 +1371,8 @@ void kibnal_free_pages (kib_pages_t *p) { int npages = p->ibp_npages; - int rc; int i; - if (p->ibp_mapped) { - rc = iibt_deregister_memory(p->ibp_handle); - if (rc != 0) - CERROR ("Deregister error: %d\n", rc); - } - for (i = 0; i < npages; i++) if (p->ibp_pages[i] != NULL) __free_page(p->ibp_pages[i]); @@ -1000,19 +1381,10 @@ kibnal_free_pages (kib_pages_t *p) } int -kibnal_alloc_pages (kib_pages_t **pp, int npages, int allow_write) +kibnal_alloc_pages (kib_pages_t **pp, int npages) { - kib_pages_t *p; - __u64 *phys_pages; - int i; - FSTATUS frc; - IB_ACCESS_CONTROL access; - - memset(&access, 0, sizeof(access)); - access.s.MWBindable = 1; - access.s.LocalWrite = 1; - access.s.RdmaRead = 1; - access.s.RdmaWrite = 1; + kib_pages_t *p; + int i; LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages])); if (p == NULL) { @@ -1032,123 +1404,215 @@ kibnal_alloc_pages (kib_pages_t **pp, int npages, int allow_write) } } - if (kibnal_whole_mem()) - goto out; - - LIBCFS_ALLOC(phys_pages, npages * sizeof(*phys_pages)); - if (phys_pages == NULL) { - CERROR ("Can't allocate physarray for %d pages\n", npages); - /* XXX free ibp_pages? */ - kibnal_free_pages(p); - return (-ENOMEM); - } + *pp = p; + return (0); +} - /* if we were using the _contig_ registration variant we would have - * an array of PhysAddr/Length pairs, but the discontiguous variant - * just takes the PhysAddr */ - for (i = 0; i < npages; i++) - phys_pages[i] = kibnal_page2phys(p->ibp_pages[i]); - - frc = iibt_register_physical_memory(kibnal_data.kib_hca, - 0, /* requested vaddr */ - phys_pages, npages, - 0, /* offset */ - kibnal_data.kib_pd, - access, - &p->ibp_handle, &p->ibp_vaddr, - &p->ibp_lkey, &p->ibp_rkey); +int +kibnal_alloc_tx_descs (void) +{ + int i; - LIBCFS_FREE(phys_pages, npages * sizeof(*phys_pages)); + LIBCFS_ALLOC (kibnal_data.kib_tx_descs, + IBNAL_TX_MSGS() * sizeof(kib_tx_t)); + if (kibnal_data.kib_tx_descs == NULL) + return -ENOMEM; - if (frc != FSUCCESS) { - CERROR ("Error %d mapping %d pages\n", frc, npages); - kibnal_free_pages(p); - return (-ENOMEM); + memset(kibnal_data.kib_tx_descs, 0, + IBNAL_TX_MSGS() * sizeof(kib_tx_t)); + + for (i = 0; i < IBNAL_TX_MSGS(); i++) { + kib_tx_t *tx = &kibnal_data.kib_tx_descs[i]; + +#if IBNAL_USE_FMR + LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV * + sizeof(*tx->tx_pages)); + if (tx->tx_pages == NULL) + return -ENOMEM; +#else + LIBCFS_ALLOC(tx->tx_wrq, + (1 + IBNAL_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_wrq)); + if (tx->tx_wrq == NULL) + return -ENOMEM; + + LIBCFS_ALLOC(tx->tx_gl, + (1 + IBNAL_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_gl)); + if (tx->tx_gl == NULL) + return -ENOMEM; + + LIBCFS_ALLOC(tx->tx_rd, + offsetof(kib_rdma_desc_t, + rd_frags[IBNAL_MAX_RDMA_FRAGS])); + if (tx->tx_rd == NULL) + return -ENOMEM; +#endif } - CDEBUG(D_NET, "registered %d pages; handle: %p vaddr "LPX64" " - "lkey %x rkey %x\n", npages, p->ibp_handle, - p->ibp_vaddr, p->ibp_lkey, p->ibp_rkey); - - p->ibp_mapped = 1; -out: - *pp = p; - return (0); + return 0; +} + +void +kibnal_free_tx_descs (void) +{ + int i; + + if (kibnal_data.kib_tx_descs == NULL) + return; + + for (i = 0; i < IBNAL_TX_MSGS(); i++) { + kib_tx_t *tx = &kibnal_data.kib_tx_descs[i]; + +#if IBNAL_USE_FMR + if (tx->tx_pages != NULL) + LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV * + sizeof(*tx->tx_pages)); +#else + if (tx->tx_wrq != NULL) + LIBCFS_FREE(tx->tx_wrq, + (1 + IBNAL_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_wrq)); + + if (tx->tx_gl != NULL) + LIBCFS_FREE(tx->tx_gl, + (1 + IBNAL_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_gl)); + + if (tx->tx_rd != NULL) + LIBCFS_FREE(tx->tx_rd, + offsetof(kib_rdma_desc_t, + rd_frags[IBNAL_MAX_RDMA_FRAGS])); +#endif + } + + LIBCFS_FREE(kibnal_data.kib_tx_descs, + IBNAL_TX_MSGS() * sizeof(kib_tx_t)); } -static int +int kibnal_setup_tx_descs (void) { int ipage = 0; int page_offset = 0; - __u64 vaddr; - __u64 vaddr_base; struct page *page; kib_tx_t *tx; int i; int rc; /* pre-mapped messages are not bigger than 1 page */ - LASSERT (IBNAL_MSG_SIZE <= PAGE_SIZE); + CLASSERT (IBNAL_MSG_SIZE <= PAGE_SIZE); /* No fancy arithmetic when we do the buffer calculations */ - LASSERT (PAGE_SIZE % IBNAL_MSG_SIZE == 0); + CLASSERT (PAGE_SIZE % IBNAL_MSG_SIZE == 0); - rc = kibnal_alloc_pages(&kibnal_data.kib_tx_pages, IBNAL_TX_MSG_PAGES, - 0); + rc = kibnal_alloc_pages(&kibnal_data.kib_tx_pages, + IBNAL_TX_MSG_PAGES()); if (rc != 0) return (rc); - /* ignored for the whole_mem case */ - vaddr = vaddr_base = kibnal_data.kib_tx_pages->ibp_vaddr; - - for (i = 0; i < IBNAL_TX_MSGS; i++) { + for (i = 0; i < IBNAL_TX_MSGS(); i++) { page = kibnal_data.kib_tx_pages->ibp_pages[ipage]; tx = &kibnal_data.kib_tx_descs[i]; - memset (tx, 0, sizeof(*tx)); /* zero flags etc */ - +#if IBNAL_USE_FMR + /* Allocate an FMR for this TX so it can map src/sink buffers + * for large transfers */ +#endif tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + page_offset); - if (kibnal_whole_mem()) - tx->tx_vaddr = kibnal_page2phys(page) + - page_offset + - kibnal_data.kib_md.md_addr; - else - tx->tx_vaddr = vaddr; - - tx->tx_mapped = KIB_TX_UNMAPPED; + tx->tx_hca_msg = kibnal_data.kib_whole_mem.md_addr + + kibnal_page2phys(page) + page_offset; CDEBUG(D_NET, "Tx[%d] %p->%p - "LPX64"\n", - i, tx, tx->tx_msg, tx->tx_vaddr); + i, tx, tx->tx_msg, tx->tx_hca_msg); list_add (&tx->tx_list, &kibnal_data.kib_idle_txs); - vaddr += IBNAL_MSG_SIZE; - LASSERT (vaddr <= vaddr_base + IBNAL_TX_MSG_BYTES); - page_offset += IBNAL_MSG_SIZE; LASSERT (page_offset <= PAGE_SIZE); if (page_offset == PAGE_SIZE) { page_offset = 0; ipage++; - LASSERT (ipage <= IBNAL_TX_MSG_PAGES); + LASSERT (ipage <= IBNAL_TX_MSG_PAGES()); } } return (0); } +int +kibnal_register_all_memory(void) +{ + /* CAVEAT EMPTOR: this assumes all physical memory is in 1 contiguous + * chunk starting at 0 */ + struct sysinfo si; + __u64 total; + __u64 roundup = (128<<20); /* round up in big chunks */ + IB_MR_PHYS_BUFFER phys; + IB_ACCESS_CONTROL access; + FSTATUS frc; + int rc; + + memset(&access, 0, sizeof(access)); + access.s.MWBindable = 1; + access.s.LocalWrite = 1; + access.s.RdmaRead = 1; + access.s.RdmaWrite = 1; + + /* XXX we don't bother with first-gen cards */ + if (kibnal_data.kib_hca_attrs.VendorId == 0xd0b7 && + kibnal_data.kib_hca_attrs.DeviceId == 0x3101) { + CERROR("Can't register all memory on first generation HCAs\n"); + return -EINVAL; + } + + si_meminfo(&si); + total = ((__u64)si.totalram) * si.mem_unit; + + if (total < ((__u64)max_mapnr) * PAGE_SIZE) + total = ((__u64)max_mapnr) * PAGE_SIZE; + + if (total == 0) { + CERROR("Can't determine memory size\n"); + return -ENOMEM; + } + + roundup = (128<<20); + total = (total + (roundup - 1)) & ~(roundup - 1); + + phys.PhysAddr = 0; + phys.Length = total; + + frc = iibt_register_contig_physical_memory( + kibnal_data.kib_hca, 0, &phys, 1, 0, + kibnal_data.kib_pd, access, + &kibnal_data.kib_whole_mem.md_handle, + &kibnal_data.kib_whole_mem.md_addr, + &kibnal_data.kib_whole_mem.md_lkey, + &kibnal_data.kib_whole_mem.md_rkey); + + if (frc != FSUCCESS) { + CERROR("registering physical memory failed: %d\n", frc); + return -EIO; + } + + CDEBUG(D_WARNING, "registered phys mem from "LPX64" for "LPU64"\n", + phys.PhysAddr, phys.Length); + + return 0; +} + void kibnal_shutdown (lnet_ni_t *ni) { int i; int rc; - LASSERT (ni->ni_data == &kibnal_data); LASSERT (ni == kibnal_data.kib_ni); + LASSERT (ni->ni_data == &kibnal_data); CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n", atomic_read (&libcfs_kmemory)); @@ -1159,16 +1623,16 @@ kibnal_shutdown (lnet_ni_t *ni) LBUG(); case IBNAL_INIT_ALL: - /* resetting my NID to unadvertises me, removes my - * listener and nukes all current peers */ - kibnal_set_mynid (LNET_NID_ANY); + /* stop accepting connections, prevent new peers and start to + * tear down all existing ones... */ + kibnal_stop_listener(1); - /* Wait for all peer state to clean up (crazy) */ + /* Wait for all peer state to clean up */ i = 2; while (atomic_read (&kibnal_data.kib_npeers) != 0) { i++; CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ - "waiting for %d peers to disconnect (can take a few seconds)\n", + "waiting for %d peers to disconnect\n", atomic_read (&kibnal_data.kib_npeers)); set_current_state (TASK_UNINTERRUPTIBLE); schedule_timeout (HZ); @@ -1185,21 +1649,12 @@ kibnal_shutdown (lnet_ni_t *ni) kibnal_free_pages (kibnal_data.kib_tx_pages); /* fall through */ - case IBNAL_INIT_MR: - if (kibnal_data.kib_md.md_handle != NULL) { - rc = iibt_deregister_memory(kibnal_data.kib_md.md_handle); - if (rc != FSUCCESS) - CERROR ("Deregister memory: %d\n", rc); - } + case IBNAL_INIT_MD: + rc = iibt_deregister_memory(kibnal_data.kib_whole_mem.md_handle); + if (rc != FSUCCESS) + CERROR ("Deregister memory: %d\n", rc); /* fall through */ -#if IBNAL_FMR - case IBNAL_INIT_FMR: - rc = ib_fmr_pool_destroy (kibnal_data.kib_fmr_pool); - if (rc != 0) - CERROR ("Destroy FMR pool error: %d\n", rc); - /* fall through */ -#endif case IBNAL_INIT_PD: rc = iibt_pd_free(kibnal_data.kib_pd); if (rc != 0) @@ -1212,10 +1667,6 @@ kibnal_shutdown (lnet_ni_t *ni) CERROR ("Deregister SD error: %d\n", rc); /* fall through */ - case IBNAL_INIT_PORT: - /* XXX ??? */ - /* fall through */ - case IBNAL_INIT_PORTATTRS: LIBCFS_FREE(kibnal_data.kib_hca_attrs.PortAttributesList, kibnal_data.kib_hca_attrs.PortAttributesListSize); @@ -1228,16 +1679,13 @@ kibnal_shutdown (lnet_ni_t *ni) /* fall through */ case IBNAL_INIT_DATA: - /* Module refcount only gets to zero when all peers - * have been closed so all lists must be empty */ LASSERT (atomic_read (&kibnal_data.kib_npeers) == 0); LASSERT (kibnal_data.kib_peers != NULL); for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) { LASSERT (list_empty (&kibnal_data.kib_peers[i])); } LASSERT (atomic_read (&kibnal_data.kib_nconns) == 0); - LASSERT (list_empty (&kibnal_data.kib_sched_rxq)); - LASSERT (list_empty (&kibnal_data.kib_sched_txq)); + LASSERT (list_empty (&kibnal_data.kib_connd_zombies)); LASSERT (list_empty (&kibnal_data.kib_connd_conns)); LASSERT (list_empty (&kibnal_data.kib_connd_peers)); @@ -1261,9 +1709,7 @@ kibnal_shutdown (lnet_ni_t *ni) break; } - if (kibnal_data.kib_tx_descs != NULL) - LIBCFS_FREE (kibnal_data.kib_tx_descs, - IBNAL_TX_MSGS * sizeof(kib_tx_t)); + kibnal_free_tx_descs(); if (kibnal_data.kib_peers != NULL) LIBCFS_FREE (kibnal_data.kib_peers, @@ -1277,32 +1723,19 @@ kibnal_shutdown (lnet_ni_t *ni) PORTAL_MODULE_UNUSE; } -#define roundup_power(val, power) \ - ( (val + (__u64)(power - 1)) & ~((__u64)(power - 1)) ) - -/* this isn't very portable or sturdy in the face of funny mem/bus configs */ -static __u64 max_phys_mem(IB_CA_ATTRIBUTES *ca_attr) -{ - struct sysinfo si; - __u64 ret; - - /* XXX we don't bother with first-gen cards */ - if (ca_attr->VendorId == 0xd0b7 && ca_attr->DeviceId == 0x3101) - return 0ULL; - - si_meminfo(&si); - ret = (__u64)max(si.totalram, max_mapnr) * si.mem_unit; - return roundup_power(ret, 128 * 1024 * 1024); -} -#undef roundup_power - int kibnal_startup (lnet_ni_t *ni) { + char ipif_name[32]; + __u32 ip; + __u32 netmask; + int up; + int nob; + struct timeval tv; IB_PORT_ATTRIBUTES *pattr; FSTATUS frc; int rc; - int n; + __u32 n; int i; LASSERT (ni->ni_lnd == &the_kiblnd); @@ -1313,37 +1746,81 @@ kibnal_startup (lnet_ni_t *ni) return -EPERM; } - if (ni->ni_interfaces[0] != NULL) { - CERROR("Explicit interface config not supported\n"); - return -EPERM; - } - if (IBNAL_CREDITS > IBNAL_NTX) { CERROR ("Can't set credits(%d) > ntx(%d)\n", IBNAL_CREDITS, IBNAL_NTX); return -EINVAL; } - ni->ni_maxtxcredits = IBNAL_CREDITS; - ni->ni_peertxcredits = IBNAL_PEERCREDITS; + if (*kibnal_tunables.kib_credits > *kibnal_tunables.kib_ntx) { + CERROR ("Can't set credits(%d) > ntx(%d)\n", + *kibnal_tunables.kib_credits, + *kibnal_tunables.kib_ntx); + return -EINVAL; + } + + ni->ni_maxtxcredits = *kibnal_tunables.kib_credits; + ni->ni_peertxcredits = *kibnal_tunables.kib_peercredits; + + CLASSERT (LNET_MAX_INTERFACES > 1); + + if (ni->ni_interfaces[0] == NULL) { + kibnal_data.kib_hca_idx = 0; + } else { + /* Use the HCA specified in 'networks=' */ + if (ni->ni_interfaces[1] != NULL) { + CERROR("Multiple interfaces not supported\n"); + return -EPERM; + } + + /* Parse into kib_hca_idx */ + nob = strlen(ni->ni_interfaces[0]); + if (sscanf(ni->ni_interfaces[0], "%d%n", + &kibnal_data.kib_hca_idx, &nob) < 1 || + nob != strlen(ni->ni_interfaces[0])) { + CERROR("Can't parse interface '%s'\n", + ni->ni_interfaces[0]); + return -EINVAL; + } + } + + /* Find IP address from */ + snprintf(ipif_name, sizeof(ipif_name), "%s%d", + *kibnal_tunables.kib_ipif_basename, kibnal_data.kib_hca_idx); + if (strlen(ipif_name) == sizeof(ipif_name - 1)) { + CERROR("IPoIB interface name %s truncated\n", ipif_name); + return -EINVAL; + } + + rc = libcfs_ipif_query(ipif_name, &up, &ip, &netmask); + if (rc != 0) { + CERROR("Can't query IPoIB interface %s: %d\n", ipif_name, rc); + return -ENETDOWN; + } + + if (!up) { + CERROR("Can't query IPoIB interface %s: it's down\n", ipif_name); + return -ENETDOWN; + } + + ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ip); ni->ni_data = &kibnal_data; kibnal_data.kib_ni = ni; + do_gettimeofday(&tv); + kibnal_data.kib_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; + frc = IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2, &kibnal_data.kib_interfaces); if (frc != FSUCCESS) { CERROR("IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2) = %d\n", - frc); + frc); return -ENOSYS; } PORTAL_MODULE_USE; - init_MUTEX (&kibnal_data.kib_nid_mutex); - init_MUTEX_LOCKED (&kibnal_data.kib_nid_signal); - kibnal_data.kib_ni->ni_nid = LNET_NID_ANY; - rwlock_init(&kibnal_data.kib_global_lock); kibnal_data.kib_peer_hash_size = IBNAL_PEER_HASH_SIZE; @@ -1358,20 +1835,18 @@ kibnal_startup (lnet_ni_t *ni) spin_lock_init (&kibnal_data.kib_connd_lock); INIT_LIST_HEAD (&kibnal_data.kib_connd_peers); INIT_LIST_HEAD (&kibnal_data.kib_connd_conns); + INIT_LIST_HEAD (&kibnal_data.kib_connd_zombies); init_waitqueue_head (&kibnal_data.kib_connd_waitq); spin_lock_init (&kibnal_data.kib_sched_lock); - INIT_LIST_HEAD (&kibnal_data.kib_sched_txq); - INIT_LIST_HEAD (&kibnal_data.kib_sched_rxq); init_waitqueue_head (&kibnal_data.kib_sched_waitq); spin_lock_init (&kibnal_data.kib_tx_lock); INIT_LIST_HEAD (&kibnal_data.kib_idle_txs); - LIBCFS_ALLOC (kibnal_data.kib_tx_descs, - IBNAL_TX_MSGS * sizeof(kib_tx_t)); - if (kibnal_data.kib_tx_descs == NULL) { - CERROR ("Can't allocate tx descs\n"); + rc = kibnal_alloc_tx_descs(); + if (rc != 0) { + CERROR("Can't allocate tx descs\n"); goto failed; } @@ -1379,6 +1854,10 @@ kibnal_startup (lnet_ni_t *ni) kibnal_data.kib_init = IBNAL_INIT_DATA; /*****************************************************/ + kibnal_data.kib_sdretry.RetryCount = *kibnal_tunables.kib_sd_retries; + kibnal_data.kib_sdretry.Timeout = (*kibnal_tunables.kib_timeout * 1000)/ + *kibnal_tunables.kib_sd_retries; + for (i = 0; i < IBNAL_N_SCHED; i++) { rc = kibnal_thread_start (kibnal_scheduler, (void *)i); if (rc != 0) { @@ -1398,22 +1877,30 @@ kibnal_startup (lnet_ni_t *ni) sizeof(kibnal_data.kib_hca_guids[0]); frc = iibt_get_hca_guids(&n, kibnal_data.kib_hca_guids); if (frc != FSUCCESS) { - CERROR ("Can't get channel adapter guids: %d\n", frc); + CERROR ("Can't get HCA guids: %d\n", frc); goto failed; } + if (n == 0) { - CERROR ("No channel adapters found\n"); + CERROR ("No HCAs found\n"); goto failed; } - /* Infinicon has per-HCA rather than per CQ completion handlers */ - frc = iibt_open_hca(kibnal_data.kib_hca_guids[0], - kibnal_ca_callback, - kibnal_ca_async_callback, - &kibnal_data.kib_hca, + if (n <= kibnal_data.kib_hca_idx) { + CERROR("Invalid HCA %d requested: (must be 0 - %d inclusive)\n", + kibnal_data.kib_hca_idx, n - 1); + goto failed; + } + + /* Infinicon has per-HCA notification callbacks */ + frc = iibt_open_hca(kibnal_data.kib_hca_guids[kibnal_data.kib_hca_idx], + kibnal_hca_callback, + kibnal_hca_async_callback, + NULL, &kibnal_data.kib_hca); if (frc != FSUCCESS) { - CERROR ("Can't open CA[0]: %d\n", frc); + CERROR ("Can't open HCA[%d]: %d\n", + kibnal_data.kib_hca_idx, frc); goto failed; } @@ -1442,7 +1929,8 @@ kibnal_startup (lnet_ni_t *ni) frc = iibt_query_hca(kibnal_data.kib_hca, &kibnal_data.kib_hca_attrs, NULL); if (frc != FSUCCESS) { - CERROR ("Can't get port attrs for CA 0: %d\n", frc); + CERROR ("Can't get port attrs for HCA %d: %d\n", + kibnal_data.kib_hca_idx, frc); goto failed; } @@ -1481,10 +1969,6 @@ kibnal_startup (lnet_ni_t *ni) CDEBUG(D_NET, "got guid "LPX64"\n", kibnal_data.kib_port_guid); - /* Active port found */ - kibnal_data.kib_init = IBNAL_INIT_PORT; - /*****************************************************/ - frc = iibt_sd_register(&kibnal_data.kib_sd, NULL); if (frc != FSUCCESS) { CERROR ("Can't register with SD: %d\n", frc); @@ -1505,73 +1989,14 @@ kibnal_startup (lnet_ni_t *ni) kibnal_data.kib_init = IBNAL_INIT_PD; /*****************************************************/ -#if IBNAL_FMR - { - const int pool_size = IBNAL_NTX; - struct ib_fmr_pool_param params = { - .max_pages_per_fmr = LNET_MTU/PAGE_SIZE, - .access = (IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE | - IB_ACCESS_REMOTE_READ), - .pool_size = pool_size, - .dirty_watermark = (pool_size * 3)/4, - .flush_function = NULL, - .flush_arg = NULL, - .cache = 1, - }; - rc = ib_fmr_pool_create(kibnal_data.kib_pd, ¶ms, - &kibnal_data.kib_fmr_pool); - if (rc != 0) { - CERROR ("Can't create FMR pool size %d: %d\n", - pool_size, rc); - goto failed; - } - } - - /* flag FMR pool initialised */ - kibnal_data.kib_init = IBNAL_INIT_FMR; -#endif - /*****************************************************/ - if (IBNAL_WHOLE_MEM) { - IB_MR_PHYS_BUFFER phys; - IB_ACCESS_CONTROL access; - kib_md_t *md = &kibnal_data.kib_md; - - memset(&access, 0, sizeof(access)); - access.s.MWBindable = 1; - access.s.LocalWrite = 1; - access.s.RdmaRead = 1; - access.s.RdmaWrite = 1; - - phys.PhysAddr = 0; - phys.Length = max_phys_mem(&kibnal_data.kib_hca_attrs); - if (phys.Length == 0) { - CERROR ("couldn't determine the end of phys mem\n"); - goto failed; - } - - rc = iibt_register_contig_physical_memory(kibnal_data.kib_hca, - 0, - &phys, 1, - 0, - kibnal_data.kib_pd, - access, - &md->md_handle, - &md->md_addr, - &md->md_lkey, - &md->md_rkey); - if (rc != FSUCCESS) { - CERROR("registering physical memory failed: %d\n", - rc); - CERROR("falling back to registration per-rdma\n"); - md->md_handle = NULL; - } else { - CDEBUG(D_NET, "registered "LPU64" bytes of mem\n", - phys.Length); - kibnal_data.kib_init = IBNAL_INIT_MR; - } + rc = kibnal_register_all_memory(); + if (rc != 0) { + CERROR ("Can't register all memory\n"); + goto failed; } - + + /* flag whole memory MD initialised */ + kibnal_data.kib_init = IBNAL_INIT_MD; /*****************************************************/ rc = kibnal_setup_tx_descs(); @@ -1584,33 +2009,36 @@ kibnal_startup (lnet_ni_t *ni) kibnal_data.kib_init = IBNAL_INIT_TXD; /*****************************************************/ - { - uint32 nentries; - - frc = iibt_cq_create(kibnal_data.kib_hca, IBNAL_CQ_ENTRIES, - &kibnal_data.kib_cq, &kibnal_data.kib_cq, - &nentries); - if (frc != FSUCCESS) { - CERROR ("Can't create RX CQ: %d\n", frc); - goto failed; - } - - /* flag CQ initialised */ - kibnal_data.kib_init = IBNAL_INIT_CQ; + frc = iibt_cq_create(kibnal_data.kib_hca, IBNAL_CQ_ENTRIES(), + &kibnal_data.kib_cq, &kibnal_data.kib_cq, + &n); + if (frc != FSUCCESS) { + CERROR ("Can't create RX CQ: %d\n", frc); + goto failed; + } - if (nentries < IBNAL_CQ_ENTRIES) { - CERROR ("CQ only has %d entries, need %d\n", - nentries, IBNAL_CQ_ENTRIES); - goto failed; - } + /* flag CQ initialised */ + kibnal_data.kib_init = IBNAL_INIT_CQ; + /*****************************************************/ + + if (n < IBNAL_CQ_ENTRIES()) { + CERROR ("CQ only has %d entries: %d needed\n", + n, IBNAL_CQ_ENTRIES()); + goto failed; + } - rc = iibt_cq_rearm(kibnal_data.kib_cq, CQEventSelNextWC); - if (rc != 0) { - CERROR ("Failed to re-arm completion queue: %d\n", rc); - goto failed; - } + rc = iibt_cq_rearm(kibnal_data.kib_cq, CQEventSelNextWC); + if (rc != 0) { + CERROR ("Failed to re-arm completion queue: %d\n", rc); + goto failed; } + rc = kibnal_start_listener(); + if (rc != 0) { + CERROR("Can't start listener: %d\n", rc); + goto failed; + } + /* flag everything initialised */ kibnal_data.kib_init = IBNAL_INIT_ALL; /*****************************************************/ @@ -1625,11 +2053,8 @@ kibnal_startup (lnet_ni_t *ni) void __exit kibnal_module_fini (void) { -#ifdef CONFIG_SYSCTL - if (kibnal_tunables.kib_sysctl != NULL) - unregister_sysctl_table (kibnal_tunables.kib_sysctl); -#endif lnet_unregister_lnd(&the_kiblnd); + kibnal_tunables_fini(); } int __init @@ -1642,32 +2067,17 @@ kibnal_module_init (void) return -ENODEV; } - if (sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN) { - CERROR("sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN\n"); - return -EINVAL; - } - - /* the following must be sizeof(int) for proc_dointvec() */ - if (sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)) { - CERROR("sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)\n"); - return -EINVAL; - } - - /* Initialise dynamic tunables to defaults once only */ - kibnal_tunables.kib_io_timeout = IBNAL_IO_TIMEOUT; + rc = kibnal_tunables_init(); + if (rc != 0) + return rc; lnet_register_lnd(&the_kiblnd); - -#ifdef CONFIG_SYSCTL - /* Press on regardless even if registering sysctl doesn't work */ - kibnal_tunables.kib_sysctl = - register_sysctl_table (kibnal_top_ctl_table, 0); -#endif - return (0); + + return 0; } MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel Infinicon IB NAL v0.01"); +MODULE_DESCRIPTION("Kernel Infinicon IB LND v1.00"); MODULE_LICENSE("GPL"); module_init(kibnal_module_init); diff --git a/lnet/klnds/iiblnd/iiblnd.h b/lnet/klnds/iiblnd/iiblnd.h index 6585f51..6d796b9 100644 --- a/lnet/klnds/iiblnd/iiblnd.h +++ b/lnet/klnds/iiblnd/iiblnd.h @@ -68,93 +68,96 @@ #error Invalid GCC version. Must use GCC >= 3.2.3 #endif -#define IBNAL_SERVICE_NAME "iibnal" -#define IBNAL_SERVICE_NUMBER 0x11b9a1 - #if CONFIG_SMP # define IBNAL_N_SCHED num_online_cpus() /* # schedulers */ #else # define IBNAL_N_SCHED 1 /* # schedulers */ #endif -#define IBNAL_MIN_RECONNECT_INTERVAL HZ /* first failed connection retry... */ -#define IBNAL_MAX_RECONNECT_INTERVAL (60*HZ) /* ...exponentially increasing to this */ - -#define IBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ - -#define IBNAL_MSG_QUEUE_SIZE 8 /* # messages/RDMAs in-flight */ -#define IBNAL_CREDIT_HIGHWATER 7 /* when to eagerly return credits */ +#define IBNAL_FMR 0 /* map on demand v. use whole mem mapping */ + +/* defaults for modparams/tunables */ +#define IBNAL_IPIF_BASENAME "ib" /* IPoIB interface basename */ +#define IBNAL_SERVICE_NAME "iibnal" /* global service name */ +#define IBNAL_SERVICE_NUMBER 0x11b9a2 /* global service number */ +#define IBNAL_MIN_RECONNECT_INTERVAL 1 /* first failed connection retry... */ +#define IBNAL_MAX_RECONNECT_INTERVAL 60 /* ...exponentially increasing to this */ +#define IBNAL_CONCURRENT_PEERS 1152 /* # nodes all talking at once to me */ +#define IBNAL_CKSUM 0 /* checksum kib_msg_t? */ +#define IBNAL_TIMEOUT 50 /* default comms timeout (seconds) */ +#define IBNAL_NTX 256 /* # tx descs */ +#define IBNAL_CREDITS 128 /* # concurrent sends */ +#define IBNAL_PEERCREDITS 8 /* # concurrent sends to 1 peer */ +#define IBNAL_SD_RETRIES 8 /* # times to retry SD reqs */ +#define IBNAL_LISTEN_BACKLOG 64 /* listener backlog */ + +/* tunables fixed at compile time */ +#define IBNAL_PEER_HASH_SIZE 101 /* # peer lists */ +#define IBNAL_RESCHED 100 /* # scheduler loops before reschedule */ +#define IBNAL_MSG_QUEUE_SIZE 8 /* # messages/RDMAs in-flight */ +#define IBNAL_CREDIT_HIGHWATER 7 /* when to eagerly return credits */ +#define IBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ +#define IBNAL_RDMA_BASE 0x0eeb0000 +#define IBNAL_STARTING_PSN 1 + +/* QP tunables */ /* 7 indicates infinite retry attempts, Infinicon recommended 5 */ -#define IBNAL_RETRY 5 /* # times to retry */ -#define IBNAL_RNR_RETRY 5 /* */ -#define IBNAL_CM_RETRY 5 /* # times to retry connection */ -#define IBNAL_FLOW_CONTROL 1 -#define IBNAL_ACK_TIMEOUT 20 /* supposedly 4 secs */ - -#define IBNAL_NTX 64 /* # tx descs */ -/* this had to be dropped down so that we only register < 255 pages per - * region. this will change if we register all memory. */ -#define IBNAL_CREDITS 32 -#define IBNAL_PEERCREDITS 8 - -#define IBNAL_PEER_HASH_SIZE 101 /* # peer lists */ - -#define IBNAL_RESCHED 100 /* # scheduler loops before reschedule */ - -#define IBNAL_CONCURRENT_PEERS 1000 /* # nodes all talking at once to me */ - -/* default vals for runtime tunables */ -#define IBNAL_IO_TIMEOUT 50 /* default comms timeout (seconds) */ +#define IBNAL_RETRY 5 /* # times to retry */ +#define IBNAL_RNR_RETRY 5 /* */ +#define IBNAL_CM_RETRY 5 /* # times to retry connection */ +#define IBNAL_FLOW_CONTROL 1 +#define IBNAL_ACK_TIMEOUT 20 /* supposedly 4 secs */ +#define IBNAL_EE_FLOW 1 +#define IBNAL_LOCAL_SUB 1 +#define IBNAL_FAILOVER_ACCEPTED 0 /************************/ /* derived constants... */ /* TX messages (shared by all connections) */ -#define IBNAL_TX_MSGS IBNAL_NTX -#define IBNAL_TX_MSG_BYTES (IBNAL_TX_MSGS * IBNAL_MSG_SIZE) -#define IBNAL_TX_MSG_PAGES ((IBNAL_TX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) +#define IBNAL_TX_MSGS() (*kibnal_tunables.kib_ntx) +#define IBNAL_TX_MSG_BYTES() (IBNAL_TX_MSGS() * IBNAL_MSG_SIZE) +#define IBNAL_TX_MSG_PAGES() ((IBNAL_TX_MSG_BYTES() + PAGE_SIZE - 1)/PAGE_SIZE) -#define IBNAL_TX_MAX_SG (LNET_MAX_IOV + 1) +#if IBNAL_USE_FMR +# define IBNAL_MAX_RDMA_FRAGS 1 +#else +# define IBNAL_MAX_RDMA_FRAGS LNET_MAX_IOV +#endif /* RX messages (per connection) */ -#define IBNAL_RX_MSGS IBNAL_MSG_QUEUE_SIZE -#define IBNAL_RX_MSG_BYTES (IBNAL_RX_MSGS * IBNAL_MSG_SIZE) -#define IBNAL_RX_MSG_PAGES ((IBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) +#define IBNAL_RX_MSGS IBNAL_MSG_QUEUE_SIZE +#define IBNAL_RX_MSG_BYTES (IBNAL_RX_MSGS * IBNAL_MSG_SIZE) +#define IBNAL_RX_MSG_PAGES ((IBNAL_RX_MSG_BYTES + PAGE_SIZE - 1)/PAGE_SIZE) - -/* we may have up to 2 completions per transmit + - 1 completion per receive, per connection */ -#define IBNAL_CQ_ENTRIES ((2*IBNAL_TX_MSGS) + \ - (IBNAL_RX_MSGS * IBNAL_CONCURRENT_PEERS)) - -#define IBNAL_RDMA_BASE 0x0eeb0000 -#define IBNAL_FMR 0 -#define IBNAL_WHOLE_MEM 1 -#define IBNAL_CKSUM 0 -//#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_PROCESS -#define IBNAL_CALLBACK_CTXT IB_CQ_CALLBACK_INTERRUPT - -/* XXX I have no idea. */ -#define IBNAL_STARTING_PSN 1 +#define IBNAL_CQ_ENTRIES() (IBNAL_TX_MSGS() * (1 + IBNAL_MAX_RDMA_FRAGS) + \ + (IBNAL_RX_MSGS * *kibnal_tunables.kib_concurrent_peers)) typedef struct { - int kib_io_timeout; /* comms timeout (seconds) */ - int *kib_min_reconnect_interval; /* first failed connection retry... */ - int *kib_max_reconnect_interval; /* ...exponentially increasing to this */ + char **kib_hca_basename; /* HCA base name */ + char **kib_ipif_basename; /* IPoIB interface base name */ + char **kib_service_name; /* global service name */ + unsigned int *kib_service_number; /* global service number */ + int *kib_min_reconnect_interval; /* min connect retry seconds... */ + int *kib_max_reconnect_interval; /* max connect retry seconds */ + int *kib_concurrent_peers; /* max # peers */ + int *kib_cksum; /* checksum kib_msg_t? */ + int *kib_timeout; /* comms timeout (seconds) */ + int *kib_ntx; /* # tx descs */ + int *kib_credits; /* # concurrent sends */ + int *kib_peercredits; /* # concurrent sends to 1 peer */ + int *kib_sd_retries; /* # concurrent sends to 1 peer */ +#if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM struct ctl_table_header *kib_sysctl; /* sysctl interface */ +#endif } kib_tunables_t; -/* some of these have specific types in the stack that just map back - * to the uFOO types, like IB_{L,R}_KEY. */ +/* NB The Infinicon stack has specific typedefs for some things + * (e.g. IB_{L,R}_KEY), that just map back to __u32 etc */ typedef struct { int ibp_npages; /* # pages */ - int ibp_mapped; /* mapped? */ - __u64 ibp_vaddr; /* mapped region vaddr */ - __u32 ibp_lkey; /* mapped region lkey */ - __u32 ibp_rkey; /* mapped region rkey */ - IB_HANDLE ibp_handle; /* mapped region handle */ struct page *ibp_pages[0]; } kib_pages_t; @@ -172,31 +175,29 @@ typedef struct __u64 kib_incarnation; /* which one am I */ int kib_shutdown; /* shut down? */ atomic_t kib_nthreads; /* # live threads */ - lnet_ni_t *kib_ni; /* _the_ nal instance */ + lnet_ni_t *kib_ni; /* _the_ iib instance */ - __u64 kib_service_id; /* service number I listen on */ __u64 kib_port_guid; /* my GUID (lo 64 of GID)*/ __u16 kib_port_pkey; /* my pkey, whatever that is */ - struct semaphore kib_nid_mutex; /* serialise NID ops */ - struct semaphore kib_nid_signal; /* signal completion */ - IB_HANDLE kib_cep; /* connection end point */ + struct semaphore kib_listener_signal; /* signal completion */ + IB_HANDLE kib_listener_cep; /* connection end point */ rwlock_t kib_global_lock; /* stabilize peer/conn ops */ + int kib_ready; /* CQ callback fired */ + int kib_checking_cq; /* a scheduler is checking the CQ */ struct list_head *kib_peers; /* hash table of all my known peers */ int kib_peer_hash_size; /* size of kib_peers */ atomic_t kib_npeers; /* # peers extant */ atomic_t kib_nconns; /* # connections extant */ + struct list_head kib_connd_zombies; /* connections to free */ struct list_head kib_connd_conns; /* connections to progress */ struct list_head kib_connd_peers; /* peers waiting for a connection */ - wait_queue_head_t kib_connd_waitq; /* connection daemons sleep here */ - unsigned long kib_connd_waketime; /* when connd will wake */ + wait_queue_head_t kib_connd_waitq; /* connection daemon sleep here */ spinlock_t kib_connd_lock; /* serialise */ wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */ - struct list_head kib_sched_txq; /* tx requiring attention */ - struct list_head kib_sched_rxq; /* rx requiring attention */ spinlock_t kib_sched_lock; /* serialise */ struct kib_tx *kib_tx_descs; /* all the tx descriptors */ @@ -211,15 +212,15 @@ typedef struct IB_HANDLE kib_pd; /* protection domain */ IB_HANDLE kib_sd; /* SD handle */ IB_HANDLE kib_cq; /* completion queue */ - kib_md_t kib_md; /* full-mem registration */ - - void *kib_listen_handle; /* where I listen for connections */ + kib_md_t kib_whole_mem; /* whole-mem registration */ IBT_INTERFACE_UNION kib_interfaces; /* The Infinicon IBT interface */ + int kib_hca_idx; /* my HCA number */ uint64 kib_hca_guids[8]; /* all the HCA guids */ IB_CA_ATTRIBUTES kib_hca_attrs; /* where to get HCA attrs */ - FABRIC_OPERATION_DATA kib_fabopdata; /* (un)advertise service record */ + + COMMAND_CONTROL_PARAMETERS kib_sdretry; /* control SD query retries */ } kib_data_t; #define IBNAL_INIT_NOTHING 0 @@ -227,14 +228,12 @@ typedef struct #define IBNAL_INIT_LIB 2 #define IBNAL_INIT_HCA 3 #define IBNAL_INIT_PORTATTRS 4 -#define IBNAL_INIT_PORT 5 -#define IBNAL_INIT_SD 6 -#define IBNAL_INIT_PD 7 -#define IBNAL_INIT_FMR 8 -#define IBNAL_INIT_MR 9 -#define IBNAL_INIT_TXD 10 -#define IBNAL_INIT_CQ 11 -#define IBNAL_INIT_ALL 12 +#define IBNAL_INIT_SD 5 +#define IBNAL_INIT_PD 6 +#define IBNAL_INIT_MD 7 +#define IBNAL_INIT_TXD 8 +#define IBNAL_INIT_CQ 9 +#define IBNAL_INIT_ALL 10 /************************************************************************ * Wire message structs. @@ -243,13 +242,12 @@ typedef struct * private data and SM service info), is LE on the wire. */ -/* also kib_md_t above */ - -typedef struct +typedef struct kib_connparams { - __u32 rd_nob; /* # of bytes */ - __u64 rd_addr; /* remote io vaddr */ -} WIRE_ATTR kib_rdma_desc_t; + __u32 ibcp_queue_depth; + __u32 ibcp_max_msg_size; + __u32 ibcp_max_frags; +} WIRE_ATTR kib_connparams_t; typedef struct { @@ -257,21 +255,47 @@ typedef struct char ibim_payload[0]; /* piggy-backed payload */ } WIRE_ATTR kib_immediate_msg_t; -/* these arrays serve two purposes during rdma. they are built on the passive - * side and sent to the active side as remote arguments. On the active side - * the descs are used as a data structure on the way to local gather items. - * the different roles result in split local/remote meaning of desc->rd_key */ +#if IBNAL_USE_FMR typedef struct { - lnet_hdr_t ibrm_hdr; /* portals header */ - __u64 ibrm_cookie; /* opaque completion cookie */ - __u32 ibrm_num_descs; /* how many descs */ - __u32 rd_key; /* remote key */ - kib_rdma_desc_t ibrm_desc[0]; /* where to suck/blow */ -} WIRE_ATTR kib_rdma_msg_t; + __u64 rd_addr; /* IO VMA address */ + __u32 rd_nob; /* # of bytes */ + __u32 rd_key; /* remote key */ +} WIRE_ATTR kib_rdma_desc_t; +#else +typedef struct +{ + __u32 rf_nob; /* # of bytes */ + __u64 rf_addr; /* remote io vaddr */ +} WIRE_ATTR kib_rdma_frag_t; + +typedef struct +{ + __u32 rd_key; /* local/remote key */ + __u32 rd_nfrag; /* # fragments */ + kib_rdma_frag_t rd_frags[0]; /* buffer frags */ +} WIRE_ATTR kib_rdma_desc_t; +#endif -#define kib_rdma_msg_len(num_descs) \ - offsetof(kib_msg_t, ibm_u.rdma.ibrm_desc[num_descs]) +typedef struct +{ + lnet_hdr_t ibprm_hdr; /* LNET header */ + __u64 ibprm_cookie; /* opaque completion cookie */ +} WIRE_ATTR kib_putreq_msg_t; + +typedef struct +{ + __u64 ibpam_src_cookie; /* reflected completion cookie */ + __u64 ibpam_dst_cookie; /* opaque completion cookie */ + kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */ +} WIRE_ATTR kib_putack_msg_t; + +typedef struct +{ + lnet_hdr_t ibgm_hdr; /* LNET header */ + __u64 ibgm_cookie; /* opaque completion cookie */ + kib_rdma_desc_t ibgm_rd; /* sender's sink buffer */ +} WIRE_ATTR kib_get_msg_t; typedef struct { @@ -281,17 +305,26 @@ typedef struct typedef struct { - __u32 ibm_magic; /* I'm an openibnal message */ - __u16 ibm_version; /* this is my version number */ - __u8 ibm_type; /* msg type */ - __u8 ibm_credits; /* returned credits */ -#if IBNAL_CKSUM - __u32 ibm_nob; - __u32 ibm_cksum; -#endif + /* First 2 fields fixed FOR ALL TIME */ + __u32 ibm_magic; /* I'm an openibnal message */ + __u16 ibm_version; /* this is my version number */ + + __u8 ibm_type; /* msg type */ + __u8 ibm_credits; /* returned credits */ + __u32 ibm_nob; /* # bytes in whole message */ + __u32 ibm_cksum; /* checksum (0 == no checksum) */ + __u64 ibm_srcnid; /* sender's NID */ + __u64 ibm_srcstamp; /* sender's incarnation */ + __u64 ibm_dstnid; /* destination's NID */ + __u64 ibm_dststamp; /* destination's incarnation */ + __u64 ibm_seq; /* sequence number */ + union { + kib_connparams_t connparams; kib_immediate_msg_t immediate; - kib_rdma_msg_t rdma; + kib_putreq_msg_t putreq; + kib_putack_msg_t putack; + kib_get_msg_t get; kib_completion_msg_t completion; } WIRE_ATTR ibm_u; } WIRE_ATTR kib_msg_t; @@ -299,12 +332,16 @@ typedef struct #define IBNAL_MSG_MAGIC 0x0be91b91 /* unique magic */ #define IBNAL_MSG_VERSION 1 /* current protocol version */ +#define IBNAL_MSG_CONNREQ 0xc0 /* connection request */ +#define IBNAL_MSG_CONNACK 0xc1 /* connection acknowledge */ #define IBNAL_MSG_NOOP 0xd0 /* nothing (just credits) */ -#define IBNAL_MSG_IMMEDIATE 0xd1 /* portals hdr + payload */ -#define IBNAL_MSG_PUT_RDMA 0xd2 /* portals PUT hdr + source rdma desc */ -#define IBNAL_MSG_PUT_DONE 0xd3 /* signal PUT rdma completion */ -#define IBNAL_MSG_GET_RDMA 0xd4 /* portals GET hdr + sink rdma desc */ -#define IBNAL_MSG_GET_DONE 0xd5 /* signal GET rdma completion */ +#define IBNAL_MSG_IMMEDIATE 0xd1 /* immediate */ +#define IBNAL_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */ +#define IBNAL_MSG_PUT_NAK 0xd3 /* completion (sink->src) */ +#define IBNAL_MSG_PUT_ACK 0xd4 /* putack (sink->src) */ +#define IBNAL_MSG_PUT_DONE 0xd5 /* completion (src->sink) */ +#define IBNAL_MSG_GET_REQ 0xd6 /* getreq (sink->src) */ +#define IBNAL_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */ /***********************************************************************/ @@ -312,9 +349,9 @@ typedef struct kib_rx /* receive message */ { struct list_head rx_list; /* queue for attention */ struct kib_conn *rx_conn; /* owning conn */ - int rx_rdma; /* RDMA completion posted? */ + int rx_responded; /* responded to peer? */ int rx_nob; /* # bytes received (-1 while posted) */ - __u64 rx_vaddr; /* pre-mapped buffer (hca vaddr) */ + __u64 rx_hca_msg; /* pre-mapped buffer (hca vaddr) */ kib_msg_t *rx_msg; /* pre-mapped buffer (host vaddr) */ IB_WORK_REQ rx_wrq; IB_LOCAL_DATASEGMENT rx_gl; /* and its memory */ @@ -326,93 +363,68 @@ typedef struct kib_tx /* transmit message */ struct kib_conn *tx_conn; /* owning conn */ int tx_mapped; /* mapped for RDMA? */ int tx_sending; /* # tx callbacks outstanding */ + int tx_queued; /* queued for sending */ + int tx_waiting; /* waiting for peer */ int tx_status; /* completion status */ unsigned long tx_deadline; /* completion deadline */ - int tx_passive_rdma; /* peer sucks/blows */ - int tx_passive_rdma_wait; /* waiting for peer to complete */ - __u64 tx_passive_rdma_cookie; /* completion cookie */ + __u64 tx_cookie; /* completion cookie */ lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */ - kib_md_t tx_md; /* RDMA mapping (active/passive) */ - __u64 tx_vaddr; /* pre-mapped buffer (hca vaddr) */ kib_msg_t *tx_msg; /* pre-mapped buffer (host vaddr) */ - int tx_nsp; /* # send work items */ - IB_WORK_REQ tx_wrq[IBNAL_TX_MAX_SG]; /* send work items... */ - IB_LOCAL_DATASEGMENT tx_gl[IBNAL_TX_MAX_SG]; /* ...and their memory */ + __u64 tx_hca_msg; /* pre-mapped buffer (HCA vaddr) */ + int tx_nwrq; /* # send work items */ +#if IBNAL_USE_FMR + IB_WORK_REQ tx_wrq[2]; /* send work items... */ + IB_LOCAL_DATASEGMENT tx_gl[2]; /* ...and their memory */ + kib_rdma_desc_t tx_rd[1]; /* rdma descriptor */ + kib_md_t tx_md; /* mapping */ + __u64 *tx_pages; /* page phys addrs */ +#else + IB_WORK_REQ *tx_wrq; /* send work items... */ + IB_LOCAL_DATASEGMENT *tx_gl; /* ...and their memory */ + kib_rdma_desc_t *tx_rd; /* rdma descriptor (src buffers) */ +#endif } kib_tx_t; -#define KIB_TX_UNMAPPED 0 -#define KIB_TX_MAPPED 1 -#define KIB_TX_MAPPED_FMR 2 - -typedef struct kib_wire_connreq -{ - __u32 wcr_magic; /* I'm an openibnal connreq */ - __u16 wcr_version; /* this is my version number */ - __u16 wcr_queue_depth; /* this is my receive queue size */ - __u64 wcr_nid; /* peer's NID */ - __u64 wcr_incarnation; /* peer's incarnation */ -} kib_wire_connreq_t; - -typedef struct kib_gid -{ - __u64 hi, lo; -} kib_gid_t; - -typedef struct kib_connreq +typedef struct { - /* connection-in-progress */ - struct kib_conn *cr_conn; - kib_wire_connreq_t cr_wcr; - __u64 cr_tid; - IB_SERVICE_RECORD cr_service; - kib_gid_t cr_gid; - IB_PATH_RECORD cr_path; - CM_REQUEST_INFO cr_cmreq; - CM_CONN_INFO cr_discarded; -} kib_connreq_t; + /* scratchpad during connection establishment */ + IB_QP_ATTRIBUTES_QUERY cv_qpattrs; + QUERY cv_query; + IB_SERVICE_RECORD cv_svcrec; + IB_PATH_RECORD cv_path; + CM_CONN_INFO cv_cmci; +} kib_connvars_t; typedef struct kib_conn { struct kib_peer *ibc_peer; /* owning peer */ struct list_head ibc_list; /* stash on peer's conn list */ __u64 ibc_incarnation; /* which instance of the peer */ + __u64 ibc_txseq; /* tx sequence number */ + __u64 ibc_rxseq; /* rx sequence number */ atomic_t ibc_refcount; /* # users */ int ibc_state; /* what's happening */ atomic_t ibc_nob; /* # bytes buffered */ int ibc_nsends_posted; /* # uncompleted sends */ int ibc_credits; /* # credits I have */ int ibc_outstanding_credits; /* # credits to return */ - int ibc_rcvd_disconnect;/* received discon request */ - int ibc_sent_disconnect;/* sent discon request */ + struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */ struct list_head ibc_tx_queue; /* send queue */ struct list_head ibc_active_txs; /* active tx awaiting completion */ spinlock_t ibc_lock; /* serialise */ kib_rx_t *ibc_rxs; /* the rx descs */ kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */ IB_HANDLE ibc_qp; /* queue pair */ - IB_HANDLE ibc_cep; /* connection ID? */ - IB_QP_ATTRIBUTES_QUERY ibc_qp_attrs; /* QP attrs */ - kib_connreq_t *ibc_connreq; /* connection request state */ + IB_HANDLE ibc_cep; /* CM endpoint */ + kib_connvars_t *ibc_cvars; /* connection scratchpad */ } kib_conn_t; #define IBNAL_CONN_INIT_NOTHING 0 /* initial state */ #define IBNAL_CONN_INIT_QP 1 /* ibc_qp set up */ #define IBNAL_CONN_CONNECTING 2 /* started to connect */ #define IBNAL_CONN_ESTABLISHED 3 /* connection established */ -#define IBNAL_CONN_SEND_DREQ 4 /* to send disconnect req */ -#define IBNAL_CONN_DREQ 5 /* sent disconnect req */ -#define IBNAL_CONN_DREP 6 /* sent disconnect rep */ -#define IBNAL_CONN_DISCONNECTED 7 /* no more QP or CM traffic */ - -#define KIB_ASSERT_CONN_STATE(conn, state) do { \ - LASSERTF((conn)->ibc_state == state, "%d\n", conn->ibc_state); \ -} while (0) - -#define KIB_ASSERT_CONN_STATE_RANGE(conn, low, high) do { \ - LASSERTF(low <= high, "%d %d\n", low, high); \ - LASSERTF((conn)->ibc_state >= low && (conn)->ibc_state <= high, \ - "%d\n", conn->ibc_state); \ -} while (0) +#define IBNAL_CONN_DISCONNECTING 4 /* to send disconnect req */ +#define IBNAL_CONN_DISCONNECTED 5 /* no more QP or CM traffic */ typedef struct kib_peer { @@ -711,26 +723,47 @@ static inline int wrq_signals_completion(IB_WORK_REQ *wrq) /* these are purposely avoiding using local vars so they don't increase * stack consumption. */ -#define kib_peer_addref(peer) do { \ - LASSERTF(atomic_read(&peer->ibp_refcount) > 0, "%d\n", \ - atomic_read(&peer->ibp_refcount)); \ - CDEBUG(D_NET, "++peer[%p] -> %s (%d)\n", \ - peer, libcfs_nid2str(peer->ibp_nid), \ - atomic_read (&peer->ibp_refcount)); \ - atomic_inc(&peer->ibp_refcount); \ +#define kibnal_conn_addref(conn) \ +do { \ + CDEBUG(D_NET, "conn[%p] (%d)++\n", \ + (conn), atomic_read(&(conn)->ibc_refcount)); \ + LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \ + atomic_inc(&(conn)->ibc_refcount); \ +} while (0) + +#define kibnal_conn_decref(conn) \ +do { \ + unsigned long flags; \ + \ + CDEBUG(D_NET, "conn[%p] (%d)--\n", \ + (conn), atomic_read(&(conn)->ibc_refcount)); \ + LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \ + if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \ + spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags); \ + list_add_tail(&(conn)->ibc_list, \ + &kibnal_data.kib_connd_zombies); \ + wake_up(&kibnal_data.kib_connd_waitq); \ + spin_unlock_irqrestore(&kibnal_data.kib_connd_lock, flags); \ + } \ +} while (0) + +#define kibnal_peer_addref(peer) \ +do { \ + CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \ + (peer), libcfs_nid2str((peer)->ibp_nid), \ + atomic_read (&(peer)->ibp_refcount)); \ + LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \ + atomic_inc(&(peer)->ibp_refcount); \ } while (0) -#define kib_peer_decref(peer) do { \ - LASSERTF(atomic_read(&peer->ibp_refcount) > 0, "%d\n", \ - atomic_read(&peer->ibp_refcount)); \ - CDEBUG(D_NET, "--peer[%p] -> %s (%d)\n", \ - peer, libcfs_nid2str(peer->ibp_nid), \ - atomic_read (&peer->ibp_refcount)); \ - if (atomic_dec_and_test (&peer->ibp_refcount)) { \ - CDEBUG (D_NET, "destroying peer %s %p\n", \ - libcfs_nid2str(peer->ibp_nid), peer); \ - kibnal_destroy_peer (peer); \ - } \ +#define kibnal_peer_decref(peer) \ +do { \ + CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \ + (peer), libcfs_nid2str((peer)->ibp_nid), \ + atomic_read (&(peer)->ibp_refcount)); \ + LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \ + if (atomic_dec_and_test(&(peer)->ibp_refcount)) \ + kibnal_destroy_peer(peer); \ } while (0) /******************************************************************************/ @@ -755,11 +788,18 @@ kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn) { /* CAVEAT EMPTOR: tx takes caller's ref on conn */ - LASSERT (tx->tx_nsp > 0); /* work items set up */ - LASSERT (tx->tx_conn == NULL); /* only set here */ + LASSERT (tx->tx_nwrq > 0); /* work items set up */ + LASSERT (!tx->tx_queued); /* not queued for sending already */ - tx->tx_conn = conn; - tx->tx_deadline = jiffies + kibnal_tunables.kib_io_timeout * HZ; + if (tx->tx_conn == NULL) { + kibnal_conn_addref(conn); + tx->tx_conn = conn; + } else { + LASSERT (tx->tx_conn == conn); + LASSERT (tx->tx_msg->ibm_type == IBNAL_MSG_PUT_DONE); + } + tx->tx_queued = 1; + tx->tx_deadline = jiffies + (*kibnal_tunables.kib_timeout * HZ); list_add_tail(&tx->tx_list, &conn->ibc_tx_queue); } @@ -780,13 +820,14 @@ kibnal_service_nid_field(IB_SERVICE_RECORD *srv) return (__u64 *)srv->ServiceData8; } - static inline void kibnal_set_service_keys(IB_SERVICE_RECORD *srv, lnet_nid_t nid) { - LASSERT (strlen(IBNAL_SERVICE_NAME) < sizeof(srv->ServiceName)); + char *svc_name = *kibnal_tunables.kib_service_name; + + LASSERT (strlen(svc_name) < sizeof(srv->ServiceName)); memset (srv->ServiceName, 0, sizeof(srv->ServiceName)); - strcpy (srv->ServiceName, IBNAL_SERVICE_NAME); + strcpy (srv->ServiceName, svc_name); *kibnal_service_nid_field(srv) = cpu_to_le64(nid); } @@ -813,88 +854,120 @@ kibnal_show_rdma_attr (kib_conn_t *conn) } #endif -#if CONFIG_X86 static inline __u64 kibnal_page2phys (struct page *p) { - __u64 page_number = p - mem_map; - - return (page_number << PAGE_SHIFT); + return page_to_phys(p); } -#else -# error "no page->phys" -#endif -/* CAVEAT EMPTOR: - * We rely on tx/rx descriptor alignment to allow us to use the lowest bit - * of the work request id as a flag to determine if the completion is for a - * transmit or a receive. It seems that that the CQ entry's 'op' field - * isn't always set correctly on completions that occur after QP teardown. */ + +/* CAVEAT EMPTOR: We rely on tx/rx descriptor alignment to allow us to use the + * lowest 2 bits of the work request id to stash the work item type (the op + * field is not valid when the wc completes in error). */ + +#define IBNAL_WID_TX 0 +#define IBNAL_WID_RX 1 +#define IBNAL_WID_RDMA 2 +#define IBNAL_WID_MASK 3UL static inline __u64 -kibnal_ptr2wreqid (void *ptr, int isrx) +kibnal_ptr2wreqid (void *ptr, int type) { unsigned long lptr = (unsigned long)ptr; - LASSERT ((lptr & 1) == 0); - return (__u64)(lptr | (isrx ? 1 : 0)); + LASSERT ((lptr & IBNAL_WID_MASK) == 0); + LASSERT ((type & ~IBNAL_WID_MASK) == 0); + return (__u64)(lptr | type); } static inline void * kibnal_wreqid2ptr (__u64 wreqid) { - return (void *)(((unsigned long)wreqid) & ~1UL); + return (void *)(((unsigned long)wreqid) & ~IBNAL_WID_MASK); } static inline int -kibnal_wreqid_is_rx (__u64 wreqid) +kibnal_wreqid2type (__u64 wreqid) +{ + return (wreqid & IBNAL_WID_MASK); +} + +static inline void +kibnal_set_conn_state (kib_conn_t *conn, int state) { - return (wreqid & 1) != 0; + conn->ibc_state = state; + mb(); } +#if IBNAL_USE_FMR + static inline int -kibnal_whole_mem(void) -{ - return kibnal_data.kib_md.md_handle != NULL; -} - -extern int kibnal_startup (lnet_ni_t *ni); -extern void kibnal_shutdown (lnet_ni_t *ni); -extern int kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg); -int kibnal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg); -extern int kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, - int delayed, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, - unsigned int offset, unsigned int mlen, unsigned int rlen); - -extern kib_peer_t *kibnal_create_peer (lnet_nid_t nid); -extern void kibnal_destroy_peer (kib_peer_t *peer); -extern int kibnal_del_peer (lnet_nid_t nid); -extern kib_peer_t *kibnal_find_peer_locked (lnet_nid_t nid); -extern void kibnal_unlink_peer_locked (kib_peer_t *peer); -extern int kibnal_close_stale_conns_locked (kib_peer_t *peer, - __u64 incarnation); -extern kib_conn_t *kibnal_create_conn (void); -extern void kibnal_put_conn (kib_conn_t *conn); -extern void kibnal_destroy_conn (kib_conn_t *conn); -void kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg); +kibnal_rd_size (kib_rdma_desc_t *rd) +{ + return rd->rd_nob; +} -extern int kibnal_alloc_pages (kib_pages_t **pp, int npages, int access); -extern void kibnal_free_pages (kib_pages_t *p); - -extern void kibnal_check_sends (kib_conn_t *conn); -extern void kibnal_close_conn_locked (kib_conn_t *conn, int error); -extern void kibnal_destroy_conn (kib_conn_t *conn); -extern int kibnal_thread_start (int (*fn)(void *arg), void *arg); -extern int kibnal_scheduler(void *arg); -extern int kibnal_connd (void *arg); -extern void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob); -extern void kibnal_close_conn (kib_conn_t *conn, int why); -extern void kibnal_start_active_rdma (int type, int status, - kib_rx_t *rx, lnet_msg_t *lntmsg, - unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, - unsigned int offset, unsigned int nob); - -void kibnal_ca_async_callback (void *ca_arg, IB_EVENT_RECORD *ev); -void kibnal_ca_callback (void *ca_arg, void *cq_arg); +#else +static inline int +kibnal_rd_size (kib_rdma_desc_t *rd) +{ + int i; + int size; + + for (i = size = 0; i < rd->rd_nfrag; i++) + size += rd->rd_frags[i].rf_nob; + + return size; +} +#endif + +int kibnal_startup (lnet_ni_t *ni); +void kibnal_shutdown (lnet_ni_t *ni); +int kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg); +int kibnal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg); +int kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, + int delayed, unsigned int niov, + struct iovec *iov, lnet_kiov_t *kiov, + unsigned int offset, unsigned int mlen, unsigned int rlen); +void kibnal_init_msg(kib_msg_t *msg, int type, int body_nob); +void kibnal_pack_msg(kib_msg_t *msg, int credits, lnet_nid_t dstnid, + __u64 dststamp, __u64 seq); +void kibnal_pack_connmsg(kib_msg_t *msg, int nob, int type, + lnet_nid_t dstnid, __u64 dststamp); +int kibnal_unpack_msg(kib_msg_t *msg, int nob); +IB_HANDLE kibnal_create_cep(lnet_nid_t nid); +int kibnal_create_peer (kib_peer_t **peerp, lnet_nid_t nid); +void kibnal_destroy_peer (kib_peer_t *peer); +kib_peer_t *kibnal_find_peer_locked (lnet_nid_t nid); +int kibnal_del_peer (lnet_nid_t nid); +void kibnal_unlink_peer_locked (kib_peer_t *peer); +int kibnal_add_persistent_peer (lnet_nid_t nid); +int kibnal_close_stale_conns_locked (kib_peer_t *peer, + __u64 incarnation); +int kibnal_conn_rts(kib_conn_t *conn, + __u32 qpn, __u8 resp_res, __u8 init_depth, __u32 psn); +kib_conn_t *kibnal_create_conn (lnet_nid_t nid); +void kibnal_destroy_conn (kib_conn_t *conn); +void kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg); +int kibnal_alloc_pages (kib_pages_t **pp, int npages); +void kibnal_free_pages (kib_pages_t *p); +void kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn); +int kibnal_post_receives (kib_conn_t *conn); +int kibnal_init_rdma (kib_tx_t *tx, int type, int nob, + kib_rdma_desc_t *dstrd, __u64 dstcookie); +void kibnal_check_sends (kib_conn_t *conn); +void kibnal_close_conn_locked (kib_conn_t *conn, int error); +int kibnal_thread_start (int (*fn)(void *arg), void *arg); +int kibnal_scheduler(void *arg); +int kibnal_connd (void *arg); +void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob); +void kibnal_close_conn (kib_conn_t *conn, int why); +void kibnal_start_active_rdma (int type, int status, + kib_rx_t *rx, lnet_msg_t *lntmsg, + unsigned int niov, + struct iovec *iov, lnet_kiov_t *kiov, + unsigned int offset, unsigned int nob); +void kibnal_hca_async_callback (void *hca_arg, IB_EVENT_RECORD *ev); +void kibnal_hca_callback (void *hca_arg, void *cq_arg); +int kibnal_tunables_init (void); +void kibnal_tunables_fini (void); diff --git a/lnet/klnds/iiblnd/iiblnd_cb.c b/lnet/klnds/iiblnd/iiblnd_cb.c index 4ed6bd3..472840a 100644 --- a/lnet/klnds/iiblnd/iiblnd_cb.c +++ b/lnet/klnds/iiblnd/iiblnd_cb.c @@ -23,317 +23,407 @@ #include "iiblnd.h" -/* - * LIB functions follow - * - */ -static void -kibnal_schedule_tx_done (kib_tx_t *tx) +void +hexdump(char *string, void *ptr, int len) { - unsigned long flags; + unsigned char *c = ptr; + int i; - spin_lock_irqsave (&kibnal_data.kib_sched_lock, flags); + return; - list_add_tail(&tx->tx_list, &kibnal_data.kib_sched_txq); - wake_up (&kibnal_data.kib_sched_waitq); + if (len < 0 || len > 2048) { + printk("XXX what the hell? %d\n",len); + return; + } - spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); + printk("%d bytes of '%s' from 0x%p\n", len, string, ptr); + + for (i = 0; i < len;) { + printk("%02x",*(c++)); + i++; + if (!(i & 15)) { + printk("\n"); + } else if (!(i&1)) { + printk(" "); + } + } + + if(len & 15) { + printk("\n"); + } } -static void +void kibnal_tx_done (kib_tx_t *tx) { - unsigned long flags; - int i; - FSTATUS frc; + int rc = tx->tx_status; + int i; + FSTATUS frc; - LASSERT (tx->tx_sending == 0); /* mustn't be awaiting callback */ - LASSERT (!tx->tx_passive_rdma_wait); /* mustn't be awaiting RDMA */ + LASSERT (!in_interrupt()); + LASSERT (!tx->tx_queued); /* mustn't be queued for sending */ + LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */ + LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */ - switch (tx->tx_mapped) { - default: - LBUG(); - - case KIB_TX_UNMAPPED: - break; - - case KIB_TX_MAPPED: - if (in_interrupt()) { - /* can't deregister memory in IRQ context... */ - kibnal_schedule_tx_done(tx); - return; - } - frc = iibt_deregister_memory(tx->tx_md.md_handle); - LASSERT (frc == FSUCCESS); - tx->tx_mapped = KIB_TX_UNMAPPED; - break; - -#if IBNAL_FMR - case KIB_TX_MAPPED_FMR: - if (in_interrupt() && tx->tx_status != 0) { - /* can't flush FMRs in IRQ context... */ - kibnal_schedule_tx_done(tx); - return; - } - - rc = ib_fmr_deregister(tx->tx_md.md_handle.fmr); - LASSERT (rc == 0); - - if (tx->tx_status != 0) - ib_fmr_pool_force_flush(kibnal_data.kib_fmr_pool); - tx->tx_mapped = KIB_TX_UNMAPPED; - break; +#if IBNAL_USE_FMR + /* Handle unmapping if required */ #endif - } - for (i = 0; i < 2; i++) { /* tx may have up to 2 ptlmsgs to finalise */ if (tx->tx_lntmsg[i] == NULL) continue; - lnet_finalize (kibnal_data.kib_ni, tx->tx_lntmsg[i], - tx->tx_status); + lnet_finalize (kibnal_data.kib_ni, tx->tx_lntmsg[i], rc); tx->tx_lntmsg[i] = NULL; } if (tx->tx_conn != NULL) { - kibnal_put_conn (tx->tx_conn); + kibnal_conn_decref(tx->tx_conn); tx->tx_conn = NULL; } - tx->tx_nsp = 0; - tx->tx_passive_rdma = 0; + tx->tx_nwrq = 0; tx->tx_status = 0; - spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags); + spin_lock(&kibnal_data.kib_tx_lock); - list_add_tail (&tx->tx_list, &kibnal_data.kib_idle_txs); + list_add (&tx->tx_list, &kibnal_data.kib_idle_txs); - spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); + spin_unlock(&kibnal_data.kib_tx_lock); } -static kib_tx_t * +kib_tx_t * kibnal_get_idle_tx (void) { - unsigned long flags; - kib_tx_t *tx = NULL; + kib_tx_t *tx; - spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags); + spin_lock(&kibnal_data.kib_tx_lock); if (list_empty (&kibnal_data.kib_idle_txs)) { - spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); + spin_unlock(&kibnal_data.kib_tx_lock); return NULL; } - + tx = list_entry (kibnal_data.kib_idle_txs.next, kib_tx_t, tx_list); list_del (&tx->tx_list); - /* Allocate a new passive RDMA completion cookie. It might not be - * needed, but we've got a lock right now and we're unlikely to + /* Allocate a new completion cookie. It might not be needed, + * but we've got a lock right now and we're unlikely to * wrap... */ - tx->tx_passive_rdma_cookie = kibnal_data.kib_next_tx_cookie++; + tx->tx_cookie = kibnal_data.kib_next_tx_cookie++; - spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags); + spin_unlock(&kibnal_data.kib_tx_lock); - LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); - LASSERT (tx->tx_nsp == 0); + LASSERT (tx->tx_nwrq == 0); + LASSERT (!tx->tx_queued); LASSERT (tx->tx_sending == 0); + LASSERT (!tx->tx_waiting); LASSERT (tx->tx_status == 0); LASSERT (tx->tx_conn == NULL); - LASSERT (!tx->tx_passive_rdma); - LASSERT (!tx->tx_passive_rdma_wait); LASSERT (tx->tx_lntmsg[0] == NULL); LASSERT (tx->tx_lntmsg[1] == NULL); - RETURN(tx); + return tx; } -static void -kibnal_complete_passive_rdma(kib_conn_t *conn, __u64 cookie, int status) +int +kibnal_post_rx (kib_rx_t *rx, int credit) { - struct list_head *ttmp; - unsigned long flags; - int idle; - - spin_lock_irqsave (&conn->ibc_lock, flags); + kib_conn_t *conn = rx->rx_conn; + int rc = 0; + unsigned long flags; + FSTATUS frc; - list_for_each (ttmp, &conn->ibc_active_txs) { - kib_tx_t *tx = list_entry(ttmp, kib_tx_t, tx_list); + LASSERT (!in_interrupt()); + + rx->rx_gl = (IB_LOCAL_DATASEGMENT) { + .Address = rx->rx_hca_msg, + .Lkey = kibnal_data.kib_whole_mem.md_lkey, + .Length = IBNAL_MSG_SIZE, + }; - LASSERT (tx->tx_passive_rdma || - !tx->tx_passive_rdma_wait); + rx->rx_wrq = (IB_WORK_REQ) { + .WorkReqId = kibnal_ptr2wreqid(rx, IBNAL_WID_RX), + .MessageLen = IBNAL_MSG_SIZE, + .DSList = &rx->rx_gl, + .DSListDepth = 1, + .Operation = WROpRecv, + }; - LASSERT (tx->tx_passive_rdma_wait || - tx->tx_sending != 0); + LASSERT (conn->ibc_state >= IBNAL_CONN_CONNECTING); + LASSERT (rx->rx_nob >= 0); /* not posted */ - if (!tx->tx_passive_rdma_wait || - tx->tx_passive_rdma_cookie != cookie) - continue; + CDEBUG(D_NET, "posting rx [%d %x "LPX64"]\n", + rx->rx_wrq.DSList->Length, + rx->rx_wrq.DSList->Lkey, + rx->rx_wrq.DSList->Address); - CDEBUG(D_NET, "Complete %p "LPD64": %d\n", tx, cookie, status); + if (conn->ibc_state > IBNAL_CONN_ESTABLISHED) { + /* No more posts for this rx; so lose its ref */ + kibnal_conn_decref(conn); + return 0; + } + + rx->rx_nob = -1; /* flag posted */ + mb(); - /* XXX Set mlength of REPLY here */ + frc = iibt_postrecv(conn->ibc_qp, &rx->rx_wrq); + if (frc == FSUCCESS) { + if (credit) { + spin_lock(&conn->ibc_lock); + conn->ibc_outstanding_credits++; + spin_unlock(&conn->ibc_lock); - tx->tx_status = status; - tx->tx_passive_rdma_wait = 0; - idle = (tx->tx_sending == 0); + kibnal_check_sends(conn); + } + return 0; + } + + CERROR ("post rx -> %s failed %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + rc = -EIO; + kibnal_close_conn(rx->rx_conn, rc); + /* No more posts for this rx; so lose its ref */ + kibnal_conn_decref(conn); + return rc; +} - if (idle) - list_del (&tx->tx_list); +int +kibnal_post_receives (kib_conn_t *conn) +{ + int i; + int rc; - spin_unlock_irqrestore (&conn->ibc_lock, flags); + LASSERT (conn->ibc_state == IBNAL_CONN_CONNECTING); - /* I could be racing with tx callbacks. It's whoever - * _makes_ tx idle that frees it */ - if (idle) - kibnal_tx_done (tx); - return; + for (i = 0; i < IBNAL_RX_MSGS; i++) { + /* +1 ref for rx desc. This ref remains until kibnal_post_rx + * fails (i.e. actual failure or we're disconnecting) */ + kibnal_conn_addref(conn); + rc = kibnal_post_rx (&conn->ibc_rxs[i], 0); + if (rc != 0) + return rc; } - - spin_unlock_irqrestore (&conn->ibc_lock, flags); - CERROR ("Unmatched (late?) RDMA completion "LPX64" from %s\n", - cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + return 0; } -static __u32 -kibnal_lkey(kib_pages_t *ibp) +kib_tx_t * +kibnal_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie) { - if (kibnal_whole_mem()) - return kibnal_data.kib_md.md_lkey; + struct list_head *tmp; + + list_for_each(tmp, &conn->ibc_active_txs) { + kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list); + + LASSERT (!tx->tx_queued); + LASSERT (tx->tx_sending != 0 || tx->tx_waiting); + + if (tx->tx_cookie != cookie) + continue; - return ibp->ibp_lkey; + if (tx->tx_waiting && + tx->tx_msg->ibm_type == txtype) + return tx; + + CWARN("Bad completion: %swaiting, type %x (wanted %x)\n", + tx->tx_waiting ? "" : "NOT ", + tx->tx_msg->ibm_type, txtype); + } + return NULL; } -static void -kibnal_post_rx (kib_rx_t *rx, int do_credits) +void +kibnal_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie) { - kib_conn_t *conn = rx->rx_conn; - int rc = 0; - unsigned long flags; - FSTATUS frc; - ENTRY; + kib_tx_t *tx; + int idle; - rx->rx_gl = (IB_LOCAL_DATASEGMENT) { - .Address = rx->rx_vaddr, - .Length = IBNAL_MSG_SIZE, - .Lkey = kibnal_lkey(conn->ibc_rx_pages), - }; + spin_lock(&conn->ibc_lock); - rx->rx_wrq = (IB_WORK_REQ) { - .Operation = WROpRecv, - .DSListDepth = 1, - .MessageLen = IBNAL_MSG_SIZE, - .WorkReqId = kibnal_ptr2wreqid(rx, 1), - .DSList = &rx->rx_gl, - }; - - KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_ESTABLISHED, - IBNAL_CONN_DREP); - LASSERT (rx->rx_nob >= 0); - rx->rx_nob = -1; /* flag posted */ - mb(); + tx = kibnal_find_waiting_tx_locked(conn, txtype, cookie); + if (tx == NULL) { + spin_unlock(&conn->ibc_lock); - if (conn->ibc_state != IBNAL_CONN_ESTABLISHED) - rc = -ECONNABORTED; - else { - frc = iibt_postrecv(conn->ibc_qp, &rx->rx_wrq); - if (frc != FSUCCESS) { - CDEBUG(D_NET, "post failed %d\n", frc); - rc = -EINVAL; - } - CDEBUG(D_NET, "posted rx %p\n", &rx->rx_wrq); + CWARN("Unmatched completion type %x cookie "LPX64" from %s\n", + txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + kibnal_close_conn (conn, -EPROTO); + return; } - if (rc == 0) { - if (do_credits) { - spin_lock_irqsave(&conn->ibc_lock, flags); - conn->ibc_outstanding_credits++; - spin_unlock_irqrestore(&conn->ibc_lock, flags); + if (tx->tx_status == 0) { /* success so far */ + if (status < 0) { /* failed? */ + tx->tx_status = status; + } else if (txtype == IBNAL_MSG_GET_REQ) { + /* XXX layering violation: set REPLY data length */ + LASSERT (tx->tx_lntmsg[1] != NULL); + LASSERT (tx->tx_lntmsg[1]->msg_ev.type == + LNET_EVENT_REPLY); - kibnal_check_sends(conn); + tx->tx_lntmsg[1]->msg_ev.mlength = status; } - EXIT; - return; } + + tx->tx_waiting = 0; - if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { - CERROR ("Error posting receive -> %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); - kibnal_close_conn (rx->rx_conn, rc); - } else { - CDEBUG (D_NET, "Error posting receive -> %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); - } + idle = !tx->tx_queued && (tx->tx_sending == 0); + if (idle) + list_del(&tx->tx_list); - /* Drop rx's ref */ - kibnal_put_conn (conn); - EXIT; + spin_unlock(&conn->ibc_lock); + + if (idle) + kibnal_tx_done(tx); } -#if IBNAL_CKSUM -static inline __u32 kibnal_cksum (void *ptr, int nob) +void +kibnal_send_completion (kib_conn_t *conn, int type, int status, __u64 cookie) { - char *c = ptr; - __u32 sum = 0; - - while (nob-- > 0) - sum = ((sum << 1) | (sum >> 31)) + *c++; + kib_tx_t *tx = kibnal_get_idle_tx(); + + if (tx == NULL) { + CERROR("Can't get tx for completion %x for %s\n", + type, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + return; + } + + tx->tx_msg->ibm_u.completion.ibcm_status = status; + tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie; + kibnal_init_tx_msg(tx, type, sizeof(kib_completion_msg_t)); - return (sum); + kibnal_queue_tx(tx, conn); } -#endif -static void hexdump(char *string, void *ptr, int len) +void +kibnal_handle_rx (kib_rx_t *rx) { - unsigned char *c = ptr; - int i; + kib_msg_t *msg = rx->rx_msg; + kib_conn_t *conn = rx->rx_conn; + int credits = msg->ibm_credits; + kib_tx_t *tx; + int rc = 0; + int repost = 1; + int rc2; - return; + LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED); - if (len < 0 || len > 2048) { - printk("XXX what the hell? %d\n",len); - return; + CDEBUG (D_NET, "Received %x[%d] from %s\n", + msg->ibm_type, credits, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + + if (credits != 0) { + /* Have I received credits that will let me send? */ + spin_lock(&conn->ibc_lock); + conn->ibc_credits += credits; + spin_unlock(&conn->ibc_lock); + + kibnal_check_sends(conn); } - printk("%d bytes of '%s' from 0x%p\n", len, string, ptr); + /* clear flag so GET_REQ can see if it caused a REPLY */ + rx->rx_responded = 0; - for (i = 0; i < len;) { - printk("%02x",*(c++)); - i++; - if (!(i & 15)) { - printk("\n"); - } else if (!(i&1)) { - printk(" "); + switch (msg->ibm_type) { + default: + CERROR("Bad IBNAL message type %x from %s\n", + msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + rc = -EPROTO; + break; + + case IBNAL_MSG_NOOP: + break; + + case IBNAL_MSG_IMMEDIATE: + rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.immediate.ibim_hdr, + msg->ibm_srcnid, rx); + repost = rc < 0; /* repost on error */ + break; + + case IBNAL_MSG_PUT_REQ: + rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.putreq.ibprm_hdr, + msg->ibm_srcnid, rx); + repost = rc < 0; /* repost on error */ + break; + + case IBNAL_MSG_PUT_NAK: + CWARN ("PUT_NACK from %s\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); + kibnal_handle_completion(conn, IBNAL_MSG_PUT_REQ, + msg->ibm_u.completion.ibcm_status, + msg->ibm_u.completion.ibcm_cookie); + break; + + case IBNAL_MSG_PUT_ACK: + spin_lock(&conn->ibc_lock); + tx = kibnal_find_waiting_tx_locked(conn, IBNAL_MSG_PUT_REQ, + msg->ibm_u.putack.ibpam_src_cookie); + if (tx != NULL) + list_del(&tx->tx_list); + spin_unlock(&conn->ibc_lock); + + if (tx == NULL) { + CERROR("Unmatched PUT_ACK from %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + rc = -EPROTO; + break; } - } - if(len & 15) { - printk("\n"); + LASSERT (tx->tx_waiting); + /* CAVEAT EMPTOR: I could be racing with tx_complete, but... + * (a) I can overwrite tx_msg since my peer has received it! + * (b) tx_waiting set tells tx_complete() it's not done. */ + + tx->tx_nwrq = 0; /* overwrite PUT_REQ */ + + rc2 = kibnal_init_rdma(tx, IBNAL_MSG_PUT_DONE, + kibnal_rd_size(&msg->ibm_u.putack.ibpam_rd), + &msg->ibm_u.putack.ibpam_rd, + msg->ibm_u.putack.ibpam_dst_cookie); + if (rc2 < 0) + CERROR("Can't setup rdma for PUT to %s: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); + + spin_lock(&conn->ibc_lock); + if (tx->tx_status == 0 && rc2 < 0) + tx->tx_status = rc2; + tx->tx_waiting = 0; /* clear waiting and queue atomically */ + kibnal_queue_tx_locked(tx, conn); + spin_unlock(&conn->ibc_lock); + break; + + case IBNAL_MSG_PUT_DONE: + kibnal_handle_completion(conn, IBNAL_MSG_PUT_ACK, + msg->ibm_u.completion.ibcm_status, + msg->ibm_u.completion.ibcm_cookie); + break; + + case IBNAL_MSG_GET_REQ: + rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.get.ibgm_hdr, + msg->ibm_srcnid, rx); + repost = rc < 0; /* repost on error */ + break; + + case IBNAL_MSG_GET_DONE: + kibnal_handle_completion(conn, IBNAL_MSG_GET_REQ, + msg->ibm_u.completion.ibcm_status, + msg->ibm_u.completion.ibcm_cookie); + break; } + + if (rc < 0) /* protocol error */ + kibnal_close_conn(conn, rc); + + if (repost) + kibnal_post_rx(rx, 1); } -static void -kibnal_rx_callback (IB_WORK_COMPLETION *wc) +void +kibnal_rx_complete (IB_WORK_COMPLETION *wc, __u64 rxseq) { kib_rx_t *rx = (kib_rx_t *)kibnal_wreqid2ptr(wc->WorkReqId); + int nob = wc->Length; kib_msg_t *msg = rx->rx_msg; kib_conn_t *conn = rx->rx_conn; - int nob = wc->Length; - const int base_nob = offsetof(kib_msg_t, ibm_u); - int credits; - int flipped; unsigned long flags; - __u32 i; -#if IBNAL_CKSUM - __u32 msg_cksum; - __u32 computed_cksum; -#endif - - /* we set the QP to erroring after we've finished disconnecting, - * maybe we should do so sooner. */ - KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_ESTABLISHED, - IBNAL_CONN_DISCONNECTED); + int rc; CDEBUG(D_NET, "rx %p conn %p\n", rx, conn); LASSERT (rx->rx_nob < 0); /* was posted */ @@ -343,7 +433,7 @@ kibnal_rx_callback (IB_WORK_COMPLETION *wc) /* receives complete with error in any case after we've started * disconnecting */ if (conn->ibc_state > IBNAL_CONN_ESTABLISHED) - goto failed; + goto ignore; if (wc->Status != WRStatusSuccess) { CERROR("Rx from %s failed: %d\n", @@ -351,277 +441,257 @@ kibnal_rx_callback (IB_WORK_COMPLETION *wc) goto failed; } - if (nob < base_nob) { - CERROR ("Short rx from %s: %d < expected %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - nob, base_nob); + rc = kibnal_unpack_msg(msg, nob); + if (rc != 0) { + CERROR ("Error %d unpacking rx from %s\n", + rc, libcfs_nid2str(conn->ibc_peer->ibp_nid)); goto failed; } - rx->rx_nob = nob; + rx->rx_nob = nob; /* Now I know nob > 0 */ mb(); - hexdump("rx", rx->rx_msg, sizeof(kib_msg_t)); - - /* Receiver does any byte flipping if necessary... */ - - if (msg->ibm_magic == IBNAL_MSG_MAGIC) { - flipped = 0; - } else { - if (msg->ibm_magic != __swab32(IBNAL_MSG_MAGIC)) { - CERROR ("Unrecognised magic: %08x from %s\n", - msg->ibm_magic, - libcfs_nid2str(conn->ibc_peer->ibp_nid)); - goto failed; - } - flipped = 1; - __swab16s (&msg->ibm_version); - LASSERT (sizeof(msg->ibm_type) == 1); - LASSERT (sizeof(msg->ibm_credits) == 1); - } - - if (msg->ibm_version != IBNAL_MSG_VERSION) { - CERROR ("Incompatible msg version %d (%d expected)\n", - msg->ibm_version, IBNAL_MSG_VERSION); + if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid || + msg->ibm_dstnid != kibnal_data.kib_ni->ni_nid || + msg->ibm_srcstamp != conn->ibc_incarnation || + msg->ibm_dststamp != kibnal_data.kib_incarnation) { + CERROR ("Stale rx from %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); goto failed; } -#if IBNAL_CKSUM - if (nob != msg->ibm_nob) { - CERROR ("Unexpected # bytes %d (%d expected)\n", nob, msg->ibm_nob); + if (msg->ibm_seq != rxseq) { + CERROR ("Out-of-sequence rx from %s" + ": got "LPD64" but expected "LPD64"\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + msg->ibm_seq, rxseq); goto failed; } - msg_cksum = le32_to_cpu(msg->ibm_cksum); - msg->ibm_cksum = 0; - computed_cksum = kibnal_cksum (msg, nob); - - if (msg_cksum != computed_cksum) { - CERROR ("Checksum failure %d: (%d expected)\n", - computed_cksum, msg_cksum); -// goto failed; - } - CDEBUG(D_NET, "cksum %x, nob %d\n", computed_cksum, nob); -#endif - - /* Have I received credits that will let me send? */ - credits = msg->ibm_credits; - if (credits != 0) { - spin_lock_irqsave(&conn->ibc_lock, flags); - conn->ibc_credits += credits; - spin_unlock_irqrestore(&conn->ibc_lock, flags); - - kibnal_check_sends(conn); - } - - switch (msg->ibm_type) { - case IBNAL_MSG_NOOP: - kibnal_post_rx (rx, 1); - return; + /* racing with connection establishment/teardown! */ - case IBNAL_MSG_IMMEDIATE: - if (nob < base_nob + sizeof (kib_immediate_msg_t)) { - CERROR ("Short IMMEDIATE from %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - nob); - goto failed; - } - break; - - case IBNAL_MSG_PUT_RDMA: - case IBNAL_MSG_GET_RDMA: - if (nob < base_nob + sizeof (kib_rdma_msg_t)) { - CERROR ("Short RDMA msg from %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), nob); - goto failed; - } - if (flipped) - __swab32(msg->ibm_u.rdma.ibrm_num_descs); - - CDEBUG(D_NET, "%d RDMA: cookie "LPX64":\n", - msg->ibm_type, msg->ibm_u.rdma.ibrm_cookie); - - if ((msg->ibm_u.rdma.ibrm_num_descs > LNET_MAX_IOV) || - (kib_rdma_msg_len(msg->ibm_u.rdma.ibrm_num_descs) > - min(nob, IBNAL_MSG_SIZE))) { - CERROR ("num_descs %d too large\n", - msg->ibm_u.rdma.ibrm_num_descs); - goto failed; - } - - if (flipped) { - __swab32(msg->ibm_u.rdma.rd_key); - } - - for(i = 0; i < msg->ibm_u.rdma.ibrm_num_descs; i++) { - kib_rdma_desc_t *desc = &msg->ibm_u.rdma.ibrm_desc[i]; - - if (flipped) { - __swab32(desc->rd_nob); - __swab64(desc->rd_addr); - } - - CDEBUG(D_NET, " key %x, " "addr "LPX64", nob %u\n", - msg->ibm_u.rdma.rd_key, desc->rd_addr, desc->rd_nob); - } - break; - - case IBNAL_MSG_PUT_DONE: - case IBNAL_MSG_GET_DONE: - if (nob < base_nob + sizeof (kib_completion_msg_t)) { - CERROR ("Short COMPLETION msg from %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), nob); - goto failed; + if (conn->ibc_state < IBNAL_CONN_ESTABLISHED) { + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + /* must check holding global lock to eliminate race */ + if (conn->ibc_state < IBNAL_CONN_ESTABLISHED) { + list_add_tail(&rx->rx_list, &conn->ibc_early_rxs); + write_unlock_irqrestore(&kibnal_data.kib_global_lock, + flags); + return; } - if (flipped) - __swab32s(&msg->ibm_u.completion.ibcm_status); - - CDEBUG(D_NET, "%d DONE: cookie "LPX64", status %d\n", - msg->ibm_type, msg->ibm_u.completion.ibcm_cookie, - msg->ibm_u.completion.ibcm_status); - - kibnal_complete_passive_rdma (conn, - msg->ibm_u.completion.ibcm_cookie, - msg->ibm_u.completion.ibcm_status); - kibnal_post_rx (rx, 1); - return; - - default: - CERROR ("Can't parse type from %s: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), msg->ibm_type); - goto failed; + write_unlock_irqrestore(&kibnal_data.kib_global_lock, + flags); } - - /* schedule for kibnal_rx() in thread context */ - spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); - - list_add_tail (&rx->rx_list, &kibnal_data.kib_sched_rxq); - wake_up (&kibnal_data.kib_sched_waitq); - - spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); + kibnal_handle_rx(rx); return; failed: CDEBUG(D_NET, "rx %p conn %p\n", rx, conn); - kibnal_close_conn(conn, -ECONNABORTED); - + kibnal_close_conn(conn, -EIO); + ignore: /* Don't re-post rx & drop its ref on conn */ - kibnal_put_conn(conn); -} - -void -kibnal_rx (kib_rx_t *rx) -{ - int rc = 0; - kib_msg_t *msg = rx->rx_msg; - - /* Clear flag so I can detect if I've sent an RDMA completion */ - rx->rx_rdma = 0; - - switch (msg->ibm_type) { - case IBNAL_MSG_GET_RDMA: - rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.rdma.ibrm_hdr, - rx->rx_conn->ibc_peer->ibp_nid, rx); - break; - - case IBNAL_MSG_PUT_RDMA: - rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.rdma.ibrm_hdr, - rx->rx_conn->ibc_peer->ibp_nid, rx); - break; - - case IBNAL_MSG_IMMEDIATE: - rc = lnet_parse(kibnal_data.kib_ni, &msg->ibm_u.immediate.ibim_hdr, - rx->rx_conn->ibc_peer->ibp_nid, rx); - break; - - default: - LBUG(); - break; - } - - if (rc < 0) { - kibnal_close_conn(rx->rx_conn, rc); - kibnal_post_rx (rx, 1); - } + kibnal_conn_decref(conn); } -static struct page * +struct page * kibnal_kvaddr_to_page (unsigned long vaddr) { struct page *page; if (vaddr >= VMALLOC_START && - vaddr < VMALLOC_END) + vaddr < VMALLOC_END) { page = vmalloc_to_page ((void *)vaddr); + LASSERT (page != NULL); + return page; + } #if CONFIG_HIGHMEM - else if (vaddr >= PKMAP_BASE && - vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) - page = vmalloc_to_page ((void *)vaddr); - /* in 2.4 ^ just walks the page tables */ + if (vaddr >= PKMAP_BASE && + vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) { + /* No highmem pages only used for bulk (kiov) I/O */ + CERROR("find page for address in highmem\n"); + LBUG(); + } #endif - else - page = virt_to_page (vaddr); - - if (!VALID_PAGE (page)) - page = NULL; - + page = virt_to_page (vaddr); + LASSERT (page != NULL); return page; } -static void -kibnal_fill_ibrm(kib_tx_t *tx, struct page *page, unsigned long page_offset, - unsigned long len, int active) +#if !IBNAL_USE_FMR +int +kibnal_append_rdfrag(kib_rdma_desc_t *rd, int active, struct page *page, + unsigned long page_offset, unsigned long len) { - kib_rdma_msg_t *ibrm = &tx->tx_msg->ibm_u.rdma; - kib_rdma_desc_t *desc; + kib_rdma_frag_t *frag = &rd->rd_frags[rd->rd_nfrag]; - LASSERTF(ibrm->ibrm_num_descs < LNET_MAX_IOV, "%u\n", - ibrm->ibrm_num_descs); + if (rd->rd_nfrag >= IBNAL_MAX_RDMA_FRAGS) { + CERROR ("Too many RDMA fragments\n"); + return -EMSGSIZE; + } - desc = &ibrm->ibrm_desc[ibrm->ibrm_num_descs]; - if (active) - ibrm->rd_key = kibnal_data.kib_md.md_lkey; - else - ibrm->rd_key = kibnal_data.kib_md.md_rkey; - desc->rd_nob = len; /*PAGE_SIZE - kiov->kiov_offset; */ - desc->rd_addr = kibnal_page2phys(page) + page_offset + - kibnal_data.kib_md.md_addr; + if (active) { + if (rd->rd_nfrag == 0) + rd->rd_key = kibnal_data.kib_whole_mem.md_lkey; + } else { + if (rd->rd_nfrag == 0) + rd->rd_key = kibnal_data.kib_whole_mem.md_rkey; + } + + frag->rf_nob = len; + frag->rf_addr = kibnal_data.kib_whole_mem.md_addr + + kibnal_page2phys(page) + page_offset; - ibrm->ibrm_num_descs++; + CDEBUG(D_NET,"map key %x frag [%d]["LPX64" for %d]\n", + rd->rd_key, rd->rd_nfrag, frag->rf_addr, frag->rf_nob); + + rd->rd_nfrag++; + return 0; } -static int -kibnal_map_rdma_iov(kib_tx_t *tx, unsigned long vaddr, int nob, int active) +int +kibnal_setup_rd_iov(kib_tx_t *tx, kib_rdma_desc_t *rd, int active, + unsigned int niov, struct iovec *iov, int offset, int nob) + { - struct page *page; - int page_offset, len; + int fragnob; + int rc; + unsigned long vaddr; + struct page *page; + int page_offset; + + LASSERT (nob > 0); + LASSERT (niov > 0); + LASSERT ((rd != tx->tx_rd) == !active); + + while (offset >= iov->iov_len) { + offset -= iov->iov_len; + niov--; + iov++; + LASSERT (niov > 0); + } - while (nob > 0) { + rd->rd_nfrag = 0; + do { + LASSERT (niov > 0); + + vaddr = ((unsigned long)iov->iov_base) + offset; + page_offset = vaddr & (PAGE_SIZE - 1); page = kibnal_kvaddr_to_page(vaddr); - if (page == NULL) + if (page == NULL) { + CERROR ("Can't find page\n"); return -EFAULT; + } - page_offset = vaddr & (PAGE_SIZE - 1); - len = min(nob, (int)PAGE_SIZE - page_offset); - - kibnal_fill_ibrm(tx, page, page_offset, len, active); - nob -= len; - vaddr += len; + fragnob = min((int)(iov->iov_len - offset), nob); + fragnob = min(fragnob, (int)PAGE_SIZE - page_offset); + + rc = kibnal_append_rdfrag(rd, active, page, + page_offset, fragnob); + if (rc != 0) + return rc; + + if (offset + fragnob < iov->iov_len) { + offset += fragnob; + } else { + offset = 0; + iov++; + niov--; + } + nob -= fragnob; + } while (nob > 0); + + return 0; +} + +int +kibnal_setup_rd_kiov (kib_tx_t *tx, kib_rdma_desc_t *rd, int active, + int nkiov, lnet_kiov_t *kiov, int offset, int nob) +{ + int fragnob; + int rc; + + CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob); + + LASSERT (nob > 0); + LASSERT (nkiov > 0); + LASSERT ((rd != tx->tx_rd) == !active); + + while (offset >= kiov->kiov_len) { + offset -= kiov->kiov_len; + nkiov--; + kiov++; + LASSERT (nkiov > 0); } + + rd->rd_nfrag = 0; + do { + LASSERT (nkiov > 0); + fragnob = min((int)(kiov->kiov_len - offset), nob); + + rc = kibnal_append_rdfrag(rd, active, kiov->kiov_page, + kiov->kiov_offset + offset, + fragnob); + if (rc != 0) + return rc; + + offset = 0; + kiov++; + nkiov--; + nob -= fragnob; + } while (nob > 0); + return 0; } +#else +int +kibnal_map_tx (kib_tx_t *tx, kib_rdma_desc_t *rd, int active, + int npages, unsigned long page_offset, int nob) +{ + IB_ACCESS_CONTROL access = {0,}; + FSTATUS frc; -static int -kibnal_map_iov (kib_tx_t *tx, IB_ACCESS_CONTROL access, - unsigned int niov, struct iovec *iov, int offset, int nob, int active) + LASSERT ((rd != tx->tx_rd) == !active); + LASSERT (!tx->tx_md.md_active); + LASSERT (tx->tx_md.md_fmrcount > 0); + LASSERT (page_offset < PAGE_SIZE); + LASSERT (npages >= (1 + ((page_offset + nob - 1)>>PAGE_SHIFT))); + LASSERT (npages <= LNET_MAX_IOV); + + if (!active) { + // access.s.MWBindable = 1; + access.s.LocalWrite = 1; + access.s.RdmaWrite = 1; + } + + /* Map the memory described by tx->tx_pages + frc = iibt_register_physical_memory(kibnal_data.kib_hca, + IBNAL_RDMA_BASE, + tx->tx_pages, npages, + page_offset, + kibnal_data.kib_pd, + access, + &tx->tx_md.md_handle, + &tx->tx_md.md_addr, + &tx->tx_md.md_lkey, + &tx->tx_md.md_rkey); + */ + return -EINVAL; +} + +int +kibnal_setup_rd_iov (kib_tx_t *tx, kib_rdma_desc_t *rd, int active, + unsigned int niov, struct iovec *iov, int offset, int nob) { - void *vaddr; - FSTATUS frc; + int resid; + int fragnob; + struct page *page; + int npages; + unsigned long page_offset; + unsigned long vaddr; LASSERT (nob > 0); LASSERT (niov > 0); - LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); while (offset >= iov->iov_len) { offset -= iov->iov_len; @@ -635,54 +705,47 @@ kibnal_map_iov (kib_tx_t *tx, IB_ACCESS_CONTROL access, return (-EMSGSIZE); } - /* our large contiguous iov could be backed by multiple physical - * pages. */ - if (kibnal_whole_mem()) { - int rc; - tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 0; - rc = kibnal_map_rdma_iov(tx, (unsigned long)iov->iov_base + - offset, nob, active); - if (rc != 0) { - CERROR ("Can't map iov: %d\n", rc); - return rc; + vaddr = ((unsigned long)iov->iov_base) + offset; + + page_offset = vaddr & (PAGE_SIZE - 1); + resid = nob; + npages = 0; + + do { + LASSERT (npages < LNET_MAX_IOV); + + page = kibnal_kvaddr_to_page(vaddr); + if (page == NULL) { + CERROR("Can't find page for %lu\n", vaddr); + return -EFAULT; } - return 0; - } - vaddr = (void *)(((unsigned long)iov->iov_base) + offset); - tx->tx_md.md_addr = (__u64)((unsigned long)vaddr); + tx->tx_pages[npages++] = kibnal_page2phys(page); - frc = iibt_register_memory(kibnal_data.kib_hca, vaddr, nob, - kibnal_data.kib_pd, access, - &tx->tx_md.md_handle, &tx->tx_md.md_lkey, - &tx->tx_md.md_rkey); - if (frc != 0) { - CERROR ("Can't map vaddr %p: %d\n", vaddr, frc); - return -EINVAL; - } + fragnob = PAGE_SIZE - (vaddr & (PAGE_SIZE - 1)); + vaddr += fragnob; + resid -= fragnob; - tx->tx_mapped = KIB_TX_MAPPED; - return (0); + } while (resid > 0); + + return kibnal_map_tx(tx, rd, active, npages, page_offset, nob); } -static int -kibnal_map_kiov (kib_tx_t *tx, IB_ACCESS_CONTROL access, - int nkiov, lnet_kiov_t *kiov, - int offset, int nob, int active) +int +kibnal_setup_rd_kiov (kib_tx_t *tx, kib_rdma_desc_t *rd, int active, + int nkiov, lnet_kiov_t *kiov, int offset, int nob) { - __u64 *phys = NULL; - int page_offset; - int nphys; - int resid; - int phys_size = 0; - FSTATUS frc; - int i, rc = 0; - + int resid; + int npages; + unsigned long page_offset; + CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob); LASSERT (nob > 0); LASSERT (nkiov > 0); - LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED); + LASSERT (nkiov <= LNET_MAX_IOV); + LASSERT (!tx->tx_md.md_active); + LASSERT ((rd != tx->tx_rd) == !active); while (offset >= kiov->kiov_len) { offset -= kiov->kiov_len; @@ -692,122 +755,36 @@ kibnal_map_kiov (kib_tx_t *tx, IB_ACCESS_CONTROL access, } page_offset = kiov->kiov_offset + offset; - nphys = 1; - - if (!kibnal_whole_mem()) { - phys_size = nkiov * sizeof (*phys); - LIBCFS_ALLOC(phys, phys_size); - if (phys == NULL) { - CERROR ("Can't allocate tmp phys\n"); - return (-ENOMEM); - } - - phys[0] = kibnal_page2phys(kiov->kiov_page); - } else { - tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 0; - kibnal_fill_ibrm(tx, kiov->kiov_page, kiov->kiov_offset, - kiov->kiov_len, active); - } - - resid = nob - (kiov->kiov_len - offset); + + resid = offset + nob; + npages = 0; - while (resid > 0) { - kiov++; - nkiov--; + do { + LASSERT (npages < LNET_MAX_IOV); LASSERT (nkiov > 0); - if (kiov->kiov_offset != 0 || - ((resid > PAGE_SIZE) && - kiov->kiov_len < PAGE_SIZE)) { + if ((npages > 0 && kiov->kiov_offset != 0) || + (resid > kiov->kiov_len && + (kiov->kiov_offset + kiov->kiov_len) != PAGE_SIZE)) { /* Can't have gaps */ CERROR ("Can't make payload contiguous in I/O VM:" - "page %d, offset %d, len %d \n", nphys, - kiov->kiov_offset, kiov->kiov_len); - - for (i = -nphys; i < nkiov; i++) - { - CERROR("kiov[%d] %p +%d for %d\n", - i, kiov[i].kiov_page, kiov[i].kiov_offset, kiov[i].kiov_len); - } + "page %d, offset %d, len %d \n", + npages, kiov->kiov_offset, kiov->kiov_len); - rc = -EINVAL; - goto out; + return -EINVAL; } - if (nphys == LNET_MAX_IOV) { - CERROR ("payload too big (%d)\n", nphys); - rc = -EMSGSIZE; - goto out; - } - - if (!kibnal_whole_mem()) { - LASSERT (nphys * sizeof (*phys) < phys_size); - phys[nphys] = kibnal_page2phys(kiov->kiov_page); - } else { - if (kib_rdma_msg_len(nphys) > IBNAL_MSG_SIZE) { - CERROR ("payload too big (%d)\n", nphys); - rc = -EMSGSIZE; - goto out; - } - kibnal_fill_ibrm(tx, kiov->kiov_page, - kiov->kiov_offset, kiov->kiov_len, - active); - } - - nphys ++; - resid -= PAGE_SIZE; - } - - if (kibnal_whole_mem()) - goto out; - -#if 0 - CWARN ("nphys %d, nob %d, page_offset %d\n", nphys, nob, page_offset); - for (i = 0; i < nphys; i++) - CWARN (" [%d] "LPX64"\n", i, phys[i]); -#endif - -#if IBNAL_FMR -#error "iibnal hasn't learned about FMR yet" - rc = ib_fmr_register_physical (kibnal_data.kib_fmr_pool, - phys, nphys, - &tx->tx_md.md_addr, - page_offset, - &tx->tx_md.md_handle.fmr, - &tx->tx_md.md_lkey, - &tx->tx_md.md_rkey); -#else - frc = iibt_register_physical_memory(kibnal_data.kib_hca, - IBNAL_RDMA_BASE, - phys, nphys, - 0, /* offset */ - kibnal_data.kib_pd, - access, - &tx->tx_md.md_handle, - &tx->tx_md.md_addr, - &tx->tx_md.md_lkey, - &tx->tx_md.md_rkey); -#endif - if (frc == FSUCCESS) { - CDEBUG(D_NET, "Mapped %d pages %d bytes @ offset %d: lkey %x, rkey %x\n", - nphys, nob, page_offset, tx->tx_md.md_lkey, tx->tx_md.md_rkey); -#if IBNAL_FMR - tx->tx_mapped = KIB_TX_MAPPED_FMR; -#else - tx->tx_mapped = KIB_TX_MAPPED; -#endif - } else { - CERROR ("Can't map phys: %d\n", frc); - rc = -EFAULT; - } + tx->tx_pages[npages++] = kibnal_page2phys(kiov->kiov_page); + resid -= kiov->kiov_len; + kiov++; + nkiov--; + } while (resid > 0); - out: - if (phys != NULL) - LIBCFS_FREE(phys, phys_size); - return (rc); + return kibnal_map_tx(tx, rd, active, npages, page_offset, nob); } +#endif -static kib_conn_t * +kib_conn_t * kibnal_find_conn_locked (kib_peer_t *peer) { struct list_head *tmp; @@ -823,111 +800,123 @@ kibnal_find_conn_locked (kib_peer_t *peer) void kibnal_check_sends (kib_conn_t *conn) { - unsigned long flags; kib_tx_t *tx; + FSTATUS frc; int rc; - int i; int done; + int i; int nwork; - ENTRY; - spin_lock_irqsave (&conn->ibc_lock, flags); + LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED); + + spin_lock(&conn->ibc_lock); LASSERT (conn->ibc_nsends_posted <= IBNAL_MSG_QUEUE_SIZE); if (list_empty(&conn->ibc_tx_queue) && conn->ibc_outstanding_credits >= IBNAL_CREDIT_HIGHWATER) { - spin_unlock_irqrestore(&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); tx = kibnal_get_idle_tx(); if (tx != NULL) kibnal_init_tx_msg(tx, IBNAL_MSG_NOOP, 0); - spin_lock_irqsave(&conn->ibc_lock, flags); + spin_lock(&conn->ibc_lock); - if (tx != NULL) { - atomic_inc(&conn->ibc_refcount); + if (tx != NULL) kibnal_queue_tx_locked(tx, conn); - } } while (!list_empty (&conn->ibc_tx_queue)) { tx = list_entry (conn->ibc_tx_queue.next, kib_tx_t, tx_list); + LASSERT (tx->tx_queued); /* We rely on this for QP sizing */ - LASSERT (tx->tx_nsp > 0 && tx->tx_nsp <= IBNAL_TX_MAX_SG); + LASSERT (tx->tx_nwrq > 0 && tx->tx_nwrq <= 1 + IBNAL_MAX_RDMA_FRAGS); LASSERT (conn->ibc_outstanding_credits >= 0); LASSERT (conn->ibc_outstanding_credits <= IBNAL_MSG_QUEUE_SIZE); LASSERT (conn->ibc_credits >= 0); LASSERT (conn->ibc_credits <= IBNAL_MSG_QUEUE_SIZE); - /* Not on ibc_rdma_queue */ - LASSERT (!tx->tx_passive_rdma_wait); - - if (conn->ibc_nsends_posted == IBNAL_MSG_QUEUE_SIZE) - GOTO(out, 0); - - if (conn->ibc_credits == 0) /* no credits */ - GOTO(out, 1); + if (conn->ibc_nsends_posted == IBNAL_MSG_QUEUE_SIZE) { + CDEBUG(D_NET, "%s: posted enough\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + break; + } + + if (conn->ibc_credits == 0) { /* no credits */ + CDEBUG(D_NET, "%s: no credits\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + break; + } if (conn->ibc_credits == 1 && /* last credit reserved for */ - conn->ibc_outstanding_credits == 0) /* giving back credits */ - GOTO(out, 2); - + conn->ibc_outstanding_credits == 0) { /* giving back credits */ + CDEBUG(D_NET, "%s: not using last credit\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + break; + } + list_del (&tx->tx_list); + tx->tx_queued = 0; + + /* NB don't drop ibc_lock before bumping tx_sending */ if (tx->tx_msg->ibm_type == IBNAL_MSG_NOOP && (!list_empty(&conn->ibc_tx_queue) || conn->ibc_outstanding_credits < IBNAL_CREDIT_HIGHWATER)) { /* redundant NOOP */ - spin_unlock_irqrestore(&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); kibnal_tx_done(tx); - spin_lock_irqsave(&conn->ibc_lock, flags); + spin_lock(&conn->ibc_lock); + CDEBUG(D_NET, "%s: redundant noop\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); continue; } - tx->tx_msg->ibm_credits = conn->ibc_outstanding_credits; - conn->ibc_outstanding_credits = 0; + kibnal_pack_msg(tx->tx_msg, conn->ibc_outstanding_credits, + conn->ibc_peer->ibp_nid, conn->ibc_incarnation, + conn->ibc_txseq); + conn->ibc_txseq++; + conn->ibc_outstanding_credits = 0; conn->ibc_nsends_posted++; conn->ibc_credits--; - /* we only get a tx completion for the final rdma op */ - tx->tx_sending = min(tx->tx_nsp, 2); - tx->tx_passive_rdma_wait = tx->tx_passive_rdma; + /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA + * PUT. If so, it was first queued here as a PUT_REQ, sent and + * stashed on ibc_active_txs, matched by an incoming PUT_ACK, + * and then re-queued here. It's (just) possible that + * tx_sending is non-zero if we've not done the tx_complete() from + * the first send; hence the ++ rather than = below. */ + tx->tx_sending++; + list_add (&tx->tx_list, &conn->ibc_active_txs); -#if IBNAL_CKSUM - tx->tx_msg->ibm_cksum = 0; - tx->tx_msg->ibm_cksum = kibnal_cksum(tx->tx_msg, tx->tx_msg->ibm_nob); - CDEBUG(D_NET, "cksum %x, nob %d\n", tx->tx_msg->ibm_cksum, tx->tx_msg->ibm_nob); -#endif - spin_unlock_irqrestore (&conn->ibc_lock, flags); - /* NB the gap between removing tx from the queue and sending it - * allows message re-ordering to occur */ + /* Drop the lock while I send (this can re-order sends) */ + spin_unlock(&conn->ibc_lock); - LASSERT (tx->tx_nsp > 0); + LASSERT (tx->tx_nwrq > 0); rc = -ECONNABORTED; + frc = FSUCCESS; nwork = 0; if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { - tx->tx_status = 0; /* Driver only accepts 1 item at a time */ - for (i = 0; i < tx->tx_nsp; i++) { - hexdump("tx", tx->tx_msg, sizeof(kib_msg_t)); - rc = iibt_postsend(conn->ibc_qp, - &tx->tx_wrq[i]); - if (rc != 0) + for (i = 0; i < tx->tx_nwrq; i++) { + frc = iibt_postsend(conn->ibc_qp, + &tx->tx_wrq[i]); + if (frc != FSUCCESS) { + rc = -EIO; break; - if (wrq_signals_completion(&tx->tx_wrq[i])) - nwork++; + } CDEBUG(D_NET, "posted tx wrq %p\n", &tx->tx_wrq[i]); } } - spin_lock_irqsave (&conn->ibc_lock, flags); + spin_lock(&conn->ibc_lock); if (rc != 0) { /* NB credits are transferred in the actual * message, which can only be the last work item */ @@ -936,18 +925,18 @@ kibnal_check_sends (kib_conn_t *conn) conn->ibc_nsends_posted--; tx->tx_status = rc; - tx->tx_passive_rdma_wait = 0; - tx->tx_sending -= tx->tx_nsp - nwork; - + tx->tx_waiting = 0; + tx->tx_sending--; + done = (tx->tx_sending == 0); if (done) list_del (&tx->tx_list); - spin_unlock_irqrestore (&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) - CERROR ("Error %d posting transmit to %s\n", rc, - libcfs_nid2str(conn->ibc_peer->ibp_nid)); + CERROR ("Error %d posting transmit to %s\n", + frc, libcfs_nid2str(conn->ibc_peer->ibp_nid)); else CDEBUG (D_NET, "Error %d posting transmit to %s\n", rc, libcfs_nid2str(conn->ibc_peer->ibp_nid)); @@ -958,140 +947,87 @@ kibnal_check_sends (kib_conn_t *conn) kibnal_tx_done (tx); return; } - } - EXIT; -out: - spin_unlock_irqrestore (&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); } -static void -kibnal_tx_callback (IB_WORK_COMPLETION *wc) +void +kibnal_tx_complete (IB_WORK_COMPLETION *wc) { kib_tx_t *tx = (kib_tx_t *)kibnal_wreqid2ptr(wc->WorkReqId); - kib_conn_t *conn; - unsigned long flags; + kib_conn_t *conn = tx->tx_conn; + int failed = wc->Status != WRStatusSuccess; int idle; - conn = tx->tx_conn; - LASSERT (conn != NULL); - LASSERT (tx->tx_sending != 0); + CDEBUG(D_NET, "tx %p conn %p sending %d nwrq %d status %d\n", + tx, conn, tx->tx_sending, tx->tx_nwrq, wc->Status); + + LASSERT (tx->tx_sending > 0); - spin_lock_irqsave(&conn->ibc_lock, flags); + if (failed && + tx->tx_status == 0 && + conn->ibc_state == IBNAL_CONN_ESTABLISHED) + CERROR("tx -> %s type %x cookie "LPX64 + "sending %d waiting %d: failed %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + tx->tx_msg->ibm_type, tx->tx_cookie, + tx->tx_sending, tx->tx_waiting, wc->Status); - CDEBUG(D_NET, "conn %p tx %p [%d/%d]: %d\n", conn, tx, - tx->tx_sending, tx->tx_nsp, wc->Status); + spin_lock(&conn->ibc_lock); /* I could be racing with rdma completion. Whoever makes 'tx' idle - * gets to free it, which also drops its ref on 'conn'. If it's - * not me, then I take an extra ref on conn so it can't disappear - * under me. */ + * gets to free it, which also drops its ref on 'conn'. */ tx->tx_sending--; + conn->ibc_nsends_posted--; + + if (failed) { + tx->tx_waiting = 0; + tx->tx_status = -EIO; + } + idle = (tx->tx_sending == 0) && /* This is the final callback */ - (!tx->tx_passive_rdma_wait); /* Not waiting for RDMA completion */ + !tx->tx_waiting && /* Not waiting for peer */ + !tx->tx_queued; /* Not re-queued (PUT_DONE) */ if (idle) list_del(&tx->tx_list); - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); - - if (tx->tx_sending == 0) - conn->ibc_nsends_posted--; - - if (wc->Status != WRStatusSuccess && - tx->tx_status == 0) - tx->tx_status = -ECONNABORTED; - - spin_unlock_irqrestore(&conn->ibc_lock, flags); - - if (idle) - kibnal_tx_done (tx); + kibnal_conn_addref(conn); /* 1 ref for me.... */ - if (wc->Status != WRStatusSuccess) { - CERROR ("Tx completion to %s failed: %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - wc->Status); - kibnal_close_conn (conn, -ENETDOWN); - } else { - /* can I shovel some more sends out the door? */ - kibnal_check_sends(conn); - } + spin_unlock(&conn->ibc_lock); - kibnal_put_conn (conn); -} - -void -kibnal_ca_async_callback (void *ca_arg, IB_EVENT_RECORD *ev) -{ - /* XXX flesh out. this seems largely for async errors */ - CERROR("type: %d code: %u\n", ev->EventType, ev->EventCode); -} - -void -kibnal_ca_callback (void *ca_arg, void *cq_arg) -{ - IB_HANDLE cq = *(IB_HANDLE *)cq_arg; - IB_HANDLE ca = *(IB_HANDLE *)ca_arg; - IB_WORK_COMPLETION wc; - int armed = 0; - - CDEBUG(D_NET, "ca %p cq %p\n", ca, cq); - - for(;;) { - while (iibt_cq_poll(cq, &wc) == FSUCCESS) { - - /* We will need to rearm the CQ to avoid a potential race. */ - armed = 0; - - if (kibnal_wreqid_is_rx(wc.WorkReqId)) - kibnal_rx_callback(&wc); - else - kibnal_tx_callback(&wc); - } - if (armed) - return; - if (iibt_cq_rearm(cq, CQEventSelNextWC) != FSUCCESS) { - CERROR("rearm failed?\n"); - return; - } - armed = 1; - } + if (idle) + kibnal_tx_done (tx); + + if (failed) + kibnal_close_conn (conn, -EIO); + else + kibnal_check_sends(conn); + + kibnal_conn_decref(conn); /* ...until here */ } void kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob) { - IB_LOCAL_DATASEGMENT *gl = &tx->tx_gl[tx->tx_nsp]; - IB_WORK_REQ *wrq = &tx->tx_wrq[tx->tx_nsp]; - int fence; - int nob = offsetof (kib_msg_t, ibm_u) + body_nob; + IB_LOCAL_DATASEGMENT *gl = &tx->tx_gl[tx->tx_nwrq]; + IB_WORK_REQ *wrq = &tx->tx_wrq[tx->tx_nwrq]; + int nob = offsetof (kib_msg_t, ibm_u) + body_nob; - LASSERT (tx->tx_nsp >= 0 && - tx->tx_nsp < sizeof(tx->tx_wrq)/sizeof(tx->tx_wrq[0])); + LASSERT (tx->tx_nwrq >= 0 && + tx->tx_nwrq < (1 + IBNAL_MAX_RDMA_FRAGS)); LASSERT (nob <= IBNAL_MSG_SIZE); - - tx->tx_msg->ibm_magic = IBNAL_MSG_MAGIC; - tx->tx_msg->ibm_version = IBNAL_MSG_VERSION; - tx->tx_msg->ibm_type = type; -#if IBNAL_CKSUM - tx->tx_msg->ibm_nob = nob; -#endif - /* Fence the message if it's bundled with an RDMA read */ - fence = (tx->tx_nsp > 0) && - (type == IBNAL_MSG_PUT_DONE); + + kibnal_init_msg(tx->tx_msg, type, body_nob); *gl = (IB_LOCAL_DATASEGMENT) { - .Address = tx->tx_vaddr, + .Address = tx->tx_hca_msg, .Length = IBNAL_MSG_SIZE, - .Lkey = kibnal_lkey(kibnal_data.kib_tx_pages), + .Lkey = kibnal_data.kib_whole_mem.md_lkey, }; - wrq->WorkReqId = kibnal_ptr2wreqid(tx, 0); + wrq->WorkReqId = kibnal_ptr2wreqid(tx, IBNAL_WID_TX); wrq->Operation = WROpSend; wrq->DSList = gl; wrq->DSListDepth = 1; @@ -1100,365 +1036,263 @@ kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob) wrq->Req.SendRC.Options.s.SolicitedEvent = 1; wrq->Req.SendRC.Options.s.SignaledCompletion = 1; wrq->Req.SendRC.Options.s.ImmediateData = 0; - wrq->Req.SendRC.Options.s.Fence = fence; - - tx->tx_nsp++; + wrq->Req.SendRC.Options.s.Fence = 0; + /* fence only needed on RDMA reads */ + + tx->tx_nwrq++; } -static void -kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn) +int +kibnal_init_rdma (kib_tx_t *tx, int type, int nob, + kib_rdma_desc_t *dstrd, __u64 dstcookie) { - unsigned long flags; + kib_msg_t *ibmsg = tx->tx_msg; + kib_rdma_desc_t *srcrd = tx->tx_rd; + IB_LOCAL_DATASEGMENT *gl; + IB_WORK_REQ *wrq; + int rc; - spin_lock_irqsave(&conn->ibc_lock, flags); +#if IBNAL_USE_FMR + LASSERT (tx->tx_nwrq == 0); - kibnal_queue_tx_locked (tx, conn); - - spin_unlock_irqrestore(&conn->ibc_lock, flags); - - kibnal_check_sends(conn); -} + gl = &tx->tx_gl[0]; + gl->Length = nob; + gl->Address = srcrd->rd_addr; + gl->Lkey = srcrd->rd_key; -static void -kibnal_launch_tx (kib_tx_t *tx, lnet_nid_t nid) -{ - unsigned long flags; - kib_peer_t *peer; - kib_conn_t *conn; - rwlock_t *g_lock = &kibnal_data.kib_global_lock; + wrq = &tx->tx_wrq[0]; - /* If I get here, I've committed to send, so I complete the tx with - * failure on any problems */ - - LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */ - LASSERT (tx->tx_nsp > 0); /* work items have been set up */ + wrq->WorkReqId = kibnal_ptr2wreqid(tx, IBNAL_WID_RDMA); + wrq->Operation = WROpRdmaWrite; + wrq->DSList = gl; + wrq->DSListDepth = 1; + wrq->MessageLen = nob; - read_lock_irqsave(g_lock, flags); - - peer = kibnal_find_peer_locked (nid); - if (peer == NULL) { - read_unlock_irqrestore(g_lock, flags); - tx->tx_status = -EHOSTUNREACH; - kibnal_tx_done (tx); - return; - } + wrq->Req.SendRC.ImmediateData = 0; + wrq->Req.SendRC.Options.s.SolicitedEvent = 0; + wrq->Req.SendRC.Options.s.SignaledCompletion = 0; + wrq->Req.SendRC.Options.s.ImmediateData = 0; + wrq->Req.SendRC.Options.s.Fence = 0; - conn = kibnal_find_conn_locked (peer); - if (conn != NULL) { - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */ - read_unlock_irqrestore(g_lock, flags); - - kibnal_queue_tx (tx, conn); - return; - } - - /* Making one or more connections; I'll need a write lock... */ - read_unlock(g_lock); - write_lock(g_lock); + wrq->Req.SendRC.RemoteDS.Address = dstrd->rd_addr; + wrq->Req.SendRC.RemoteDS.Rkey = dstrd->rd_key; - peer = kibnal_find_peer_locked (nid); - if (peer == NULL) { - write_unlock_irqrestore (g_lock, flags); - tx->tx_status = -EHOSTUNREACH; - kibnal_tx_done (tx); - return; - } + tx->tx_nwrq = 1; + rc = nob; +#else + /* CAVEAT EMPTOR: this 'consumes' the frags in 'dstrd' */ + int resid = nob; + kib_rdma_frag_t *srcfrag; + int srcidx; + kib_rdma_frag_t *dstfrag; + int dstidx; + int wrknob; - conn = kibnal_find_conn_locked (peer); - if (conn != NULL) { - /* Connection exists; queue message on it */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */ - write_unlock_irqrestore (g_lock, flags); + /* Called by scheduler */ + LASSERT (!in_interrupt()); + + LASSERT (type == IBNAL_MSG_GET_DONE || + type == IBNAL_MSG_PUT_DONE); + + srcidx = dstidx = 0; + srcfrag = &srcrd->rd_frags[0]; + dstfrag = &dstrd->rd_frags[0]; + rc = resid; + + while (resid > 0) { + if (srcidx >= srcrd->rd_nfrag) { + CERROR("Src buffer exhausted: %d frags\n", srcidx); + rc = -EPROTO; + break; + } - kibnal_queue_tx (tx, conn); - return; - } + if (dstidx == dstrd->rd_nfrag) { + CERROR("Dst buffer exhausted: %d frags\n", dstidx); + rc = -EPROTO; + break; + } - if (peer->ibp_connecting == 0) { - if (!(peer->ibp_reconnect_interval == 0 || /* first attempt */ - time_after_eq(jiffies, peer->ibp_reconnect_time))) { - write_unlock_irqrestore (g_lock, flags); - tx->tx_status = -EHOSTUNREACH; - kibnal_tx_done (tx); - return; + if (tx->tx_nwrq == IBNAL_MAX_RDMA_FRAGS) { + CERROR("RDMA too fragmented: %d/%d src %d/%d dst frags\n", + srcidx, srcrd->rd_nfrag, + dstidx, dstrd->rd_nfrag); + rc = -EMSGSIZE; + break; } - - peer->ibp_connecting = 1; - kib_peer_addref(peer); /* extra ref for connd */ - - spin_lock (&kibnal_data.kib_connd_lock); - - list_add_tail (&peer->ibp_connd_list, - &kibnal_data.kib_connd_peers); - wake_up (&kibnal_data.kib_connd_waitq); - - spin_unlock (&kibnal_data.kib_connd_lock); - } - - /* A connection is being established; queue the message... */ - list_add_tail (&tx->tx_list, &peer->ibp_tx_queue); - write_unlock_irqrestore (g_lock, flags); -} + wrknob = MIN(MIN(srcfrag->rf_nob, dstfrag->rf_nob), resid); -static int -kibnal_start_passive_rdma (int type, int may_block, lnet_msg_t *lntmsg) -{ - lnet_nid_t nid = lntmsg->msg_target.nid; - int nob = lntmsg->msg_md->md_length; - kib_tx_t *tx; - kib_msg_t *ibmsg; - int rc; - IB_ACCESS_CONTROL access = {0,}; - - LASSERT (type == IBNAL_MSG_PUT_RDMA || type == IBNAL_MSG_GET_RDMA); - LASSERT (nob > 0); - LASSERT (!in_interrupt()); /* Mapping could block */ + gl = &tx->tx_gl[tx->tx_nwrq]; + gl->Length = wrknob; + gl->Address = srcfrag->rf_addr; + gl->Lkey = srcrd->rd_key; - access.s.MWBindable = 1; - access.s.LocalWrite = 1; - access.s.RdmaRead = 1; - access.s.RdmaWrite = 1; + wrq = &tx->tx_wrq[tx->tx_nwrq]; - tx = kibnal_get_idle_tx (); - if (tx == NULL) { - CERROR("Can't allocate %s txd for %s\n", - (type == IBNAL_MSG_PUT_RDMA) ? "PUT/REPLY" : "GET", - libcfs_nid2str(nid)); - return -ENOMEM; - } + wrq->WorkReqId = kibnal_ptr2wreqid(tx, IBNAL_WID_RDMA); + wrq->Operation = WROpRdmaWrite; + wrq->DSList = gl; + wrq->DSListDepth = 1; + wrq->MessageLen = nob; - if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) - rc = kibnal_map_iov (tx, access, - lntmsg->msg_md->md_niov, - lntmsg->msg_md->md_iov.iov, - 0, nob, 0); - else - rc = kibnal_map_kiov (tx, access, - lntmsg->msg_md->md_niov, - lntmsg->msg_md->md_iov.kiov, - 0, nob, 0); + wrq->Req.SendRC.ImmediateData = 0; + wrq->Req.SendRC.Options.s.SolicitedEvent = 0; + wrq->Req.SendRC.Options.s.SignaledCompletion = 0; + wrq->Req.SendRC.Options.s.ImmediateData = 0; + wrq->Req.SendRC.Options.s.Fence = 0; - if (rc != 0) { - CERROR ("Can't map RDMA for %s: %d\n", - libcfs_nid2str(nid), rc); - goto failed; - } - - if (type == IBNAL_MSG_GET_RDMA) { - /* reply gets finalized when tx completes */ - tx->tx_lntmsg[1] = lnet_create_reply_msg(kibnal_data.kib_ni, - lntmsg); - if (tx->tx_lntmsg[1] == NULL) { - CERROR ("Can't create reply for GET -> %s\n", - libcfs_nid2str(nid)); - rc = -ENOMEM; - goto failed; + wrq->Req.SendRC.RemoteDS.Address = dstfrag->rf_addr; + wrq->Req.SendRC.RemoteDS.Rkey = dstrd->rd_key; + + resid -= wrknob; + if (wrknob < srcfrag->rf_nob) { + srcfrag->rf_addr += wrknob; + srcfrag->rf_nob -= wrknob; + } else { + srcfrag++; + srcidx++; + } + + if (wrknob < dstfrag->rf_nob) { + dstfrag->rf_addr += wrknob; + dstfrag->rf_nob -= wrknob; + } else { + dstfrag++; + dstidx++; } + + tx->tx_nwrq++; } + + if (rc < 0) /* no RDMA if completing with failure */ + tx->tx_nwrq = 0; +#endif - tx->tx_passive_rdma = 1; + ibmsg->ibm_u.completion.ibcm_status = rc; + ibmsg->ibm_u.completion.ibcm_cookie = dstcookie; + kibnal_init_tx_msg(tx, type, sizeof (kib_completion_msg_t)); - ibmsg = tx->tx_msg; + return rc; +} - ibmsg->ibm_u.rdma.ibrm_hdr = lntmsg->msg_hdr; - ibmsg->ibm_u.rdma.ibrm_cookie = tx->tx_passive_rdma_cookie; - /* map_kiov alrady filled the rdma descs for the whole_mem case */ - if (!kibnal_whole_mem()) { - ibmsg->ibm_u.rdma.rd_key = tx->tx_md.md_rkey; - ibmsg->ibm_u.rdma.ibrm_desc[0].rd_addr = tx->tx_md.md_addr; - ibmsg->ibm_u.rdma.ibrm_desc[0].rd_nob = nob; - ibmsg->ibm_u.rdma.ibrm_num_descs = 1; - } +void +kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn) +{ + spin_lock(&conn->ibc_lock); + kibnal_queue_tx_locked (tx, conn); + spin_unlock(&conn->ibc_lock); + + kibnal_check_sends(conn); +} - kibnal_init_tx_msg (tx, type, - kib_rdma_msg_len(ibmsg->ibm_u.rdma.ibrm_num_descs)); +void +kibnal_schedule_active_connect_locked (kib_peer_t *peer) +{ + /* Called holding kib_global_lock exclusive */ + unsigned long flags; - CDEBUG(D_NET, "Passive: %p cookie "LPX64", key %x, addr " - LPX64", nob %d\n", - tx, tx->tx_passive_rdma_cookie, tx->tx_md.md_rkey, - tx->tx_md.md_addr, nob); - - /* lntmsg gets finalized when tx completes. */ - tx->tx_lntmsg[0] = lntmsg; + peer->ibp_connecting++; /* I'm connecting */ + kibnal_peer_addref(peer); /* extra ref for connd */ - kibnal_launch_tx(tx, nid); - return (0); + spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags); - failed: - tx->tx_status = rc; - kibnal_tx_done (tx); - return (-EIO); + list_add_tail (&peer->ibp_connd_list, &kibnal_data.kib_connd_peers); + wake_up (&kibnal_data.kib_connd_waitq); + + spin_unlock_irqrestore(&kibnal_data.kib_connd_lock, flags); } void -kibnal_start_active_rdma (int type, int status, - kib_rx_t *rx, lnet_msg_t *lntmsg, - unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, - unsigned int offset, unsigned int nob) +kibnal_launch_tx (kib_tx_t *tx, lnet_nid_t nid) { - kib_msg_t *rxmsg = rx->rx_msg; - kib_msg_t *txmsg; - kib_tx_t *tx; - IB_ACCESS_CONTROL access = {0,}; - IB_WR_OP rdma_op; - int rc; - __u32 i; + kib_peer_t *peer; + kib_conn_t *conn; + unsigned long flags; + rwlock_t *g_lock = &kibnal_data.kib_global_lock; + int retry; + int rc; - CDEBUG(D_NET, "type %d, status %d, niov %d, offset %d, nob %d\n", - type, status, niov, offset, nob); + /* If I get here, I've committed to send, so I complete the tx with + * failure on any problems */ + + LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */ + LASSERT (tx->tx_nwrq > 0); /* work items have been set up */ - /* Called by scheduler */ - LASSERT (!in_interrupt ()); + for (retry = 0; ; retry = 1) { + read_lock_irqsave(g_lock, flags); + + peer = kibnal_find_peer_locked (nid); + if (peer != NULL) { + conn = kibnal_find_conn_locked (peer); + if (conn != NULL) { + kibnal_conn_addref(conn); /* 1 ref for me... */ + read_unlock_irqrestore(g_lock, flags); + + kibnal_queue_tx (tx, conn); + kibnal_conn_decref(conn); /* ...to here */ + return; + } + } + + /* Making one or more connections; I'll need a write lock... */ + read_unlock(g_lock); + write_lock(g_lock); - /* Either all pages or all vaddrs */ - LASSERT (!(kiov != NULL && iov != NULL)); + peer = kibnal_find_peer_locked (nid); + if (peer != NULL) + break; - /* No data if we're completing with failure */ - LASSERT (status == 0 || nob == 0); + write_unlock_irqrestore(g_lock, flags); - LASSERT (type == IBNAL_MSG_GET_DONE || - type == IBNAL_MSG_PUT_DONE); + if (retry) { + CERROR("Can't find peer %s\n", libcfs_nid2str(nid)); - /* Flag I'm completing the RDMA. Even if I fail to send the - * completion message, I will have tried my best so further - * attempts shouldn't be tried. */ - LASSERT (!rx->rx_rdma); - rx->rx_rdma = 1; + tx->tx_status = -EHOSTUNREACH; + tx->tx_waiting = 0; + kibnal_tx_done (tx); + return; + } - if (type == IBNAL_MSG_GET_DONE) { - rdma_op = WROpRdmaWrite; - LASSERT (rxmsg->ibm_type == IBNAL_MSG_GET_RDMA); - } else { - access.s.LocalWrite = 1; - rdma_op = WROpRdmaRead; - LASSERT (rxmsg->ibm_type == IBNAL_MSG_PUT_RDMA); + rc = kibnal_add_persistent_peer(nid); + if (rc != 0) { + CERROR("Can't add peer %s: %d\n", + libcfs_nid2str(nid), rc); + + tx->tx_status = -EHOSTUNREACH; + tx->tx_waiting = 0; + kibnal_tx_done (tx); + return; + } } - tx = kibnal_get_idle_tx (); - if (tx == NULL) { - CERROR ("tx descs exhausted on RDMA from %s" - " completing locally with failure\n", - libcfs_nid2str(rx->rx_conn->ibc_peer->ibp_nid)); - lnet_finalize (kibnal_data.kib_ni, lntmsg, -ENOMEM); + conn = kibnal_find_conn_locked (peer); + if (conn != NULL) { + /* Connection exists; queue message on it */ + kibnal_conn_addref(conn); /* 1 ref for me... */ + write_unlock_irqrestore(g_lock, flags); + + kibnal_queue_tx (tx, conn); + kibnal_conn_decref(conn); /* ...until here */ return; } - LASSERT (tx->tx_nsp == 0); - - if (nob == 0) - GOTO(init_tx, 0); - - /* We actually need to transfer some data (the transfer - * size could get truncated to zero when the incoming - * message is matched) */ - if (kiov != NULL) - rc = kibnal_map_kiov (tx, access, niov, kiov, offset, nob, 1); - else - rc = kibnal_map_iov (tx, access, niov, iov, offset, nob, 1); - - if (rc != 0) { - CERROR ("Can't map RDMA -> %s: %d\n", - libcfs_nid2str(rx->rx_conn->ibc_peer->ibp_nid), rc); - /* We'll skip the RDMA and complete with failure. */ - status = rc; - nob = 0; - GOTO(init_tx, rc); - } - - if (!kibnal_whole_mem()) { - tx->tx_msg->ibm_u.rdma.rd_key = tx->tx_md.md_lkey; - tx->tx_msg->ibm_u.rdma.ibrm_desc[0].rd_addr = tx->tx_md.md_addr; - tx->tx_msg->ibm_u.rdma.ibrm_desc[0].rd_nob = nob; - tx->tx_msg->ibm_u.rdma.ibrm_num_descs = 1; - } - - /* XXX ugh. different page-sized hosts. */ - if (tx->tx_msg->ibm_u.rdma.ibrm_num_descs != - rxmsg->ibm_u.rdma.ibrm_num_descs) { - CERROR("tx descs (%u) != rx descs (%u)\n", - tx->tx_msg->ibm_u.rdma.ibrm_num_descs, - rxmsg->ibm_u.rdma.ibrm_num_descs); - /* We'll skip the RDMA and complete with failure. */ - status = rc; - nob = 0; - GOTO(init_tx, rc); - } - - /* map_kiov filled in the rdma descs which describe our side of the - * rdma transfer. */ - /* ibrm_num_descs was verified in rx_callback */ - for(i = 0; i < rxmsg->ibm_u.rdma.ibrm_num_descs; i++) { - kib_rdma_desc_t *ldesc, *rdesc; /* local, remote */ - IB_LOCAL_DATASEGMENT *ds = &tx->tx_gl[i]; - IB_WORK_REQ *wrq = &tx->tx_wrq[i]; - - ldesc = &tx->tx_msg->ibm_u.rdma.ibrm_desc[i]; - rdesc = &rxmsg->ibm_u.rdma.ibrm_desc[i]; - - ds->Address = ldesc->rd_addr; - ds->Length = ldesc->rd_nob; - ds->Lkey = tx->tx_msg->ibm_u.rdma.rd_key; - - memset(wrq, 0, sizeof(*wrq)); - wrq->WorkReqId = kibnal_ptr2wreqid(tx, 0); - wrq->Operation = rdma_op; - wrq->DSList = ds; - wrq->DSListDepth = 1; - wrq->MessageLen = ds->Length; - wrq->Req.SendRC.ImmediateData = 0; - wrq->Req.SendRC.Options.s.SolicitedEvent = 0; - wrq->Req.SendRC.Options.s.SignaledCompletion = 0; - wrq->Req.SendRC.Options.s.ImmediateData = 0; - wrq->Req.SendRC.Options.s.Fence = 0; - wrq->Req.SendRC.RemoteDS.Address = rdesc->rd_addr; - wrq->Req.SendRC.RemoteDS.Rkey = rxmsg->ibm_u.rdma.rd_key; - /* only the last rdma post triggers tx completion */ - if (i == rxmsg->ibm_u.rdma.ibrm_num_descs - 1) - wrq->Req.SendRC.Options.s.SignaledCompletion = 1; + if (peer->ibp_connecting == 0) { + if (!(peer->ibp_reconnect_interval == 0 || /* first attempt */ + time_after_eq(jiffies, peer->ibp_reconnect_time))) { + write_unlock_irqrestore(g_lock, flags); + tx->tx_status = -EHOSTUNREACH; + tx->tx_waiting = 0; + kibnal_tx_done (tx); + return; + } - tx->tx_nsp++; + kibnal_schedule_active_connect_locked(peer); } - -init_tx: - txmsg = tx->tx_msg; - - txmsg->ibm_u.completion.ibcm_cookie = rxmsg->ibm_u.rdma.ibrm_cookie; - txmsg->ibm_u.completion.ibcm_status = status; - kibnal_init_tx_msg(tx, type, sizeof (kib_completion_msg_t)); + /* A connection is being established; queue the message... */ + list_add_tail (&tx->tx_list, &peer->ibp_tx_queue); - if (status == 0 && nob != 0) { - LASSERT (tx->tx_nsp > 1); - /* RDMA: lntmsg gets finalized when the tx completes. This - * is after the completion message has been sent, which in - * turn is after the RDMA has finished. */ - tx->tx_lntmsg[0] = lntmsg; - } else { - LASSERT (tx->tx_nsp == 1); - /* No RDMA: local completion happens now! */ - CDEBUG(D_WARNING,"No data: immediate completion\n"); - lnet_finalize (kibnal_data.kib_ni, lntmsg, - status == 0 ? 0 : -EIO); - } - - /* +1 ref for this tx... */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - rx->rx_conn, rx->rx_conn->ibc_state, - libcfs_nid2str(rx->rx_conn->ibc_peer->ibp_nid), - atomic_read (&rx->rx_conn->ibc_refcount)); - atomic_inc (&rx->rx_conn->ibc_refcount); - /* ...and queue it up */ - kibnal_queue_tx(tx, rx->rx_conn); + write_unlock_irqrestore(g_lock, flags); } int @@ -1469,7 +1303,7 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) lnet_process_id_t target = lntmsg->msg_target; int target_is_router = lntmsg->msg_target_is_router; int routing = lntmsg->msg_routing; - unsigned int payload_niov = lntmsg->msg_niov; + unsigned int payload_niov = lntmsg->msg_niov; struct iovec *payload_iov = lntmsg->msg_iov; lnet_kiov_t *payload_kiov = lntmsg->msg_kiov; unsigned int payload_offset = lntmsg->msg_offset; @@ -1477,17 +1311,18 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) kib_msg_t *ibmsg; kib_tx_t *tx; int nob; + int rc; /* NB 'private' is different depending on what we're sending.... */ - CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n", + CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n", payload_nob, payload_niov, libcfs_id2str(target)); LASSERT (payload_nob == 0 || payload_niov > 0); LASSERT (payload_niov <= LNET_MAX_IOV); - /* Thread context if we're sending payload */ - LASSERT (!in_interrupt() || payload_nob == 0); + /* Thread context */ + LASSERT (!in_interrupt()); /* payload is either all vaddrs or all pages */ LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); @@ -1503,37 +1338,132 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) case LNET_MSG_GET: if (routing || target_is_router) break; /* send IMMEDIATE */ - + /* is the REPLY message too small for RDMA? */ nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]); if (nob <= IBNAL_MSG_SIZE) break; /* send IMMEDIATE */ - return kibnal_start_passive_rdma(IBNAL_MSG_GET_RDMA, 1, lntmsg); + tx = kibnal_get_idle_tx(); + if (tx == NULL) { + CERROR("Can allocate txd for GET to %s: \n", + libcfs_nid2str(target.nid)); + return -ENOMEM; + } + + ibmsg = tx->tx_msg; + ibmsg->ibm_u.get.ibgm_hdr = *hdr; + ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie; + + if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) + rc = kibnal_setup_rd_iov(tx, &ibmsg->ibm_u.get.ibgm_rd, + 0, + lntmsg->msg_md->md_niov, + lntmsg->msg_md->md_iov.iov, + 0, lntmsg->msg_md->md_length); + else + rc = kibnal_setup_rd_kiov(tx, &ibmsg->ibm_u.get.ibgm_rd, + 0, + lntmsg->msg_md->md_niov, + lntmsg->msg_md->md_iov.kiov, + 0, lntmsg->msg_md->md_length); + if (rc != 0) { + CERROR("Can't setup GET sink for %s: %d\n", + libcfs_nid2str(target.nid), rc); + kibnal_tx_done(tx); + return -EIO; + } + +#if IBNAL_USE_FMR + nob = sizeof(kib_get_msg_t); +#else + { + int n = ibmsg->ibm_u.get.ibgm_rd.rd_nfrag; + + nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]); + } +#endif + kibnal_init_tx_msg(tx, IBNAL_MSG_GET_REQ, nob); + + tx->tx_lntmsg[1] = lnet_create_reply_msg(kibnal_data.kib_ni, + lntmsg); + if (tx->tx_lntmsg[1] == NULL) { + CERROR("Can't create reply for GET -> %s\n", + libcfs_nid2str(target.nid)); + kibnal_tx_done(tx); + return -EIO; + } + + tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg[0,1] on completion */ + tx->tx_waiting = 1; /* waiting for GET_DONE */ + kibnal_launch_tx(tx, target.nid); + return 0; case LNET_MSG_REPLY: { /* reply's 'private' is the incoming receive */ kib_rx_t *rx = private; - LASSERT (routing || rx != NULL); + LASSERT(routing || rx != NULL); - /* RDMA reply expected? */ if (!routing && rx->rx_msg->ibm_type != IBNAL_MSG_IMMEDIATE) { - /* Incoming message consistent with RDMA */ - if (rx->rx_msg->ibm_type != IBNAL_MSG_GET_RDMA) { - CERROR ("REPLY to %s bad ibm type %d!!!\n", - libcfs_nid2str(target.nid), - rx->rx_msg->ibm_type); - return (-EIO); + /* Incoming message consistent with RDMA? */ + if (rx->rx_msg->ibm_type != IBNAL_MSG_GET_REQ) { + CERROR("REPLY to %s bad msg type %x!!!\n", + libcfs_nid2str(target.nid), + rx->rx_msg->ibm_type); + return -EIO; } - kibnal_start_active_rdma(IBNAL_MSG_GET_DONE, 0, - rx, lntmsg, payload_niov, - payload_iov, payload_kiov, - payload_offset, payload_nob); - return (0); + /* NB handle_rx() will send GET_NAK when I return to + * it from here, unless I set rx_responded! */ + + tx = kibnal_get_idle_tx(); + if (tx == NULL) { + CERROR("Can't get tx for REPLY to %s\n", + libcfs_nid2str(target.nid)); + return -ENOMEM; + } + + if (payload_nob == 0) + rc = 0; + else if (payload_kiov == NULL) + rc = kibnal_setup_rd_iov( + tx, tx->tx_rd, 1, + payload_niov, payload_iov, + payload_offset, payload_nob); + else + rc = kibnal_setup_rd_kiov( + tx, tx->tx_rd, 1, + payload_niov, payload_kiov, + payload_offset, payload_nob); + if (rc != 0) { + CERROR("Can't setup GET src for %s: %d\n", + libcfs_nid2str(target.nid), rc); + kibnal_tx_done(tx); + return -EIO; + } + + rc = kibnal_init_rdma(tx, IBNAL_MSG_GET_DONE, + payload_nob, + &rx->rx_msg->ibm_u.get.ibgm_rd, + rx->rx_msg->ibm_u.get.ibgm_cookie); + if (rc < 0) { + CERROR("Can't setup rdma for GET from %s: %d\n", + libcfs_nid2str(target.nid), rc); + } else if (rc == 0) { + /* No RDMA: local completion may happen now! */ + lnet_finalize (kibnal_data.kib_ni, lntmsg, 0); + } else { + /* RDMA: lnet_finalize(lntmsg) when it + * completes */ + tx->tx_lntmsg[0] = lntmsg; + } + + kibnal_queue_tx(tx, rx->rx_conn); + rx->rx_responded = 1; + return (rc >= 0) ? 0 : -EIO; } - /* Fall through to handle like PUT */ + /* fall through to handle like PUT */ } case LNET_MSG_PUT: @@ -1541,73 +1471,103 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]); if (nob <= IBNAL_MSG_SIZE) break; /* send IMMEDIATE */ - - return kibnal_start_passive_rdma(IBNAL_MSG_PUT_RDMA, - !(routing || type == LNET_MSG_REPLY), - lntmsg); + + tx = kibnal_get_idle_tx(); + if (tx == NULL) { + CERROR("Can't allocate %s txd for %s\n", + type == LNET_MSG_PUT ? "PUT" : "REPLY", + libcfs_nid2str(target.nid)); + return -ENOMEM; + } + + if (payload_kiov == NULL) + rc = kibnal_setup_rd_iov(tx, tx->tx_rd, 1, + payload_niov, payload_iov, + payload_offset, payload_nob); + else + rc = kibnal_setup_rd_kiov(tx, tx->tx_rd, 1, + payload_niov, payload_kiov, + payload_offset, payload_nob); + if (rc != 0) { + CERROR("Can't setup PUT src for %s: %d\n", + libcfs_nid2str(target.nid), rc); + kibnal_tx_done(tx); + return -EIO; + } + + ibmsg = tx->tx_msg; + ibmsg->ibm_u.putreq.ibprm_hdr = *hdr; + ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie; + kibnal_init_tx_msg(tx, IBNAL_MSG_PUT_REQ, sizeof(kib_putreq_msg_t)); + + tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */ + tx->tx_waiting = 1; /* waiting for PUT_{ACK,NAK} */ + kibnal_launch_tx(tx, target.nid); + return 0; } - /* Send IMMEDIATE */ + /* send IMMEDIATE */ + + LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]) + <= IBNAL_MSG_SIZE); tx = kibnal_get_idle_tx(); if (tx == NULL) { - CERROR ("Can't send %d to %s: tx descs exhausted%s\n", - type, libcfs_nid2str(target.nid), - in_interrupt() ? " (intr)" : ""); - return (-ENOMEM); + CERROR ("Can't send %d to %s: tx descs exhausted\n", + type, libcfs_nid2str(target.nid)); + return -ENOMEM; } ibmsg = tx->tx_msg; ibmsg->ibm_u.immediate.ibim_hdr = *hdr; if (payload_kiov != NULL) - lnet_copy_kiov2flat(IBNAL_MSG_SIZE, ibmsg, + lnet_copy_kiov2flat(IBNAL_MSG_SIZE, ibmsg, offsetof(kib_msg_t, ibm_u.immediate.ibim_payload), - payload_niov, payload_kiov, + payload_niov, payload_kiov, payload_offset, payload_nob); else - lnet_copy_iov2flat(IBNAL_MSG_SIZE, ibmsg, + lnet_copy_iov2flat(IBNAL_MSG_SIZE, ibmsg, offsetof(kib_msg_t, ibm_u.immediate.ibim_payload), - payload_niov, payload_iov, + payload_niov, payload_iov, payload_offset, payload_nob); - kibnal_init_tx_msg (tx, IBNAL_MSG_IMMEDIATE, - offsetof(kib_immediate_msg_t, - ibim_payload[payload_nob])); - - /* lntmsg gets finalized when tx completes */ - tx->tx_lntmsg[0] = lntmsg; + nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]); + kibnal_init_tx_msg (tx, IBNAL_MSG_IMMEDIATE, nob); + tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */ kibnal_launch_tx(tx, target.nid); - return (0); + return 0; } int -kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, - int delayed, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, +kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, + unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen, unsigned int rlen) { kib_rx_t *rx = private; kib_msg_t *rxmsg = rx->rx_msg; - int msg_nob; + kib_conn_t *conn = rx->rx_conn; + kib_tx_t *tx; + kib_msg_t *txmsg; + int nob; int rc = 0; - + LASSERT (mlen <= rlen); - LASSERT (!in_interrupt ()); + LASSERT (!in_interrupt()); /* Either all pages or all vaddrs */ LASSERT (!(kiov != NULL && iov != NULL)); switch (rxmsg->ibm_type) { default: LBUG(); - + case IBNAL_MSG_IMMEDIATE: - msg_nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]); - if (msg_nob > rx->rx_nob) { + nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]); + if (nob > rx->rx_nob) { CERROR ("Immediate message from %s too big: %d(%d)\n", libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid), - msg_nob, rx->rx_nob); + nob, rx->rx_nob); rc = -EPROTO; break; } @@ -1618,41 +1578,84 @@ kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, offsetof(kib_msg_t, ibm_u.immediate.ibim_payload), mlen); else - lnet_copy_flat2iov(niov, iov, offset, - IBNAL_MSG_SIZE, rxmsg, - offsetof(kib_msg_t, ibm_u.immediate.ibim_payload), - mlen); + lnet_copy_flat2iov(niov, iov, offset, + IBNAL_MSG_SIZE, rxmsg, + offsetof(kib_msg_t, ibm_u.immediate.ibim_payload), + mlen); + lnet_finalize (ni, lntmsg, 0); + break; + + case IBNAL_MSG_PUT_REQ: + if (mlen == 0) { + lnet_finalize(ni, lntmsg, 0); + kibnal_send_completion(rx->rx_conn, IBNAL_MSG_PUT_NAK, 0, + rxmsg->ibm_u.putreq.ibprm_cookie); + break; + } + + tx = kibnal_get_idle_tx(); + if (tx == NULL) { + CERROR("Can't allocate tx for %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + /* Not replying will break the connection */ + rc = -ENOMEM; + break; + } + + txmsg = tx->tx_msg; + if (kiov == NULL) + rc = kibnal_setup_rd_iov(tx, + &txmsg->ibm_u.putack.ibpam_rd, + 0, + niov, iov, offset, mlen); + else + rc = kibnal_setup_rd_kiov(tx, + &txmsg->ibm_u.putack.ibpam_rd, + 0, + niov, kiov, offset, mlen); + if (rc != 0) { + CERROR("Can't setup PUT sink for %s: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), rc); + kibnal_tx_done(tx); + /* tell peer it's over */ + kibnal_send_completion(rx->rx_conn, IBNAL_MSG_PUT_NAK, rc, + rxmsg->ibm_u.putreq.ibprm_cookie); + break; + } - lnet_finalize (ni, lntmsg, 0); + txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie; + txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie; +#if IBNAL_USE_FMR + nob = sizeof(kib_putack_msg_t); +#else + { + int n = tx->tx_msg->ibm_u.putack.ibpam_rd.rd_nfrag; + + nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]); + } +#endif + kibnal_init_tx_msg(tx, IBNAL_MSG_PUT_ACK, nob); + + tx->tx_lntmsg[0] = lntmsg; /* finalise lntmsg on completion */ + tx->tx_waiting = 1; /* waiting for PUT_DONE */ + kibnal_queue_tx(tx, conn); break; - case IBNAL_MSG_GET_RDMA: + case IBNAL_MSG_GET_REQ: LASSERT (lntmsg == NULL); /* no need to finalise */ - if (!rx->rx_rdma) { + if (!rx->rx_responded) { /* GET didn't match anything */ - kibnal_start_active_rdma (IBNAL_MSG_GET_DONE, -ENODATA, - rx, NULL, 0, NULL, NULL, 0, 0); + kibnal_send_completion(rx->rx_conn, IBNAL_MSG_GET_DONE, + -ENODATA, + rxmsg->ibm_u.get.ibgm_cookie); } break; - - case IBNAL_MSG_PUT_RDMA: - kibnal_start_active_rdma (IBNAL_MSG_PUT_DONE, 0, rx, lntmsg, - niov, iov, kiov, offset, mlen); - break; } kibnal_post_rx(rx, 1); return rc; } -/***************************************************************************** - * the rest of this file concerns connection management. active connetions - * start with connect_peer, passive connections start with passive_callback. - * active disconnects start with conn_close, cm_callback starts passive - * disconnects and contains the guts of how the disconnect state machine - * progresses. - *****************************************************************************/ - int kibnal_thread_start (int (*fn)(void *arg), void *arg) { @@ -1665,45 +1668,82 @@ kibnal_thread_start (int (*fn)(void *arg), void *arg) return (0); } -static void +void kibnal_thread_fini (void) { atomic_dec (&kibnal_data.kib_nthreads); } -/* this can be called by anyone at any time to close a connection. if - * the connection is still established it heads to the connd to start - * the disconnection in a safe context. It has no effect if called - * on a connection that is already disconnecting */ void -kibnal_close_conn_locked (kib_conn_t *conn, int error) +kibnal_schedule_conn (kib_conn_t *conn) { - /* This just does the immmediate housekeeping, and schedules the - * connection for the connd to finish off. - * Caller holds kib_global_lock exclusively in irq context */ - kib_peer_t *peer = conn->ibc_peer; + unsigned long flags; - KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_CONNECTING, - IBNAL_CONN_DISCONNECTED); + kibnal_conn_addref(conn); /* ++ref for connd */ + + spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags); - if (conn->ibc_state > IBNAL_CONN_ESTABLISHED) - return; /* already disconnecting */ + list_add_tail (&conn->ibc_list, &kibnal_data.kib_connd_conns); + wake_up (&kibnal_data.kib_connd_waitq); + + spin_unlock_irqrestore(&kibnal_data.kib_connd_lock, flags); +} - CDEBUG (error == 0 ? D_NET : D_ERROR, - "closing conn to %s: error %d\n", - libcfs_nid2str(peer->ibp_nid), error); +void +kibnal_close_conn_locked (kib_conn_t *conn, int error) +{ + /* This just does the immediate housekeeping to start shutdown of an + * established connection. 'error' is zero for a normal shutdown. + * Caller holds kib_global_lock exclusively in irq context */ + kib_peer_t *peer = conn->ibc_peer; + + LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED); - if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) { - /* kib_connd_conns takes ibc_list's ref */ - list_del (&conn->ibc_list); + if (conn->ibc_state != IBNAL_CONN_ESTABLISHED) + return; /* already being handled */ + + /* NB Can't take ibc_lock here (could be in IRQ context), without + * risking deadlock, so access to ibc_{tx_queue,active_txs} is racey */ + + if (error == 0 && + list_empty(&conn->ibc_tx_queue) && + list_empty(&conn->ibc_active_txs)) { + CDEBUG(D_NET, "closing conn to %s" + " rx# "LPD64" tx# "LPD64"\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_txseq, conn->ibc_rxseq); } else { - /* new ref for kib_connd_conns */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); + CERROR("Closing conn to %s: error %d%s%s" + " rx# "LPD64" tx# "LPD64"\n", + libcfs_nid2str(peer->ibp_nid), error, + list_empty(&conn->ibc_tx_queue) ? "" : "(sending)", + list_empty(&conn->ibc_active_txs) ? "" : "(waiting)", + conn->ibc_txseq, conn->ibc_rxseq); +#if 0 + /* can't skip down the queue without holding ibc_lock (see above) */ + list_for_each(tmp, &conn->ibc_tx_queue) { + kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list); + + CERROR(" queued tx type %x cookie "LPX64 + " sending %d waiting %d ticks %ld/%d\n", + tx->tx_msg->ibm_type, tx->tx_cookie, + tx->tx_sending, tx->tx_waiting, + (long)(tx->tx_deadline - jiffies), HZ); + } + + list_for_each(tmp, &conn->ibc_active_txs) { + kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list); + + CERROR(" active tx type %x cookie "LPX64 + " sending %d waiting %d ticks %ld/%d\n", + tx->tx_msg->ibm_type, tx->tx_cookie, + tx->tx_sending, tx->tx_waiting, + (long)(tx->tx_deadline - jiffies), HZ); + } +#endif } + + list_del (&conn->ibc_list); if (list_empty (&peer->ibp_conns) && /* no more conns */ peer->ibp_persistence == 0 && /* non-persistent peer */ @@ -1711,29 +1751,119 @@ kibnal_close_conn_locked (kib_conn_t *conn, int error) kibnal_unlink_peer_locked (peer); } - conn->ibc_state = IBNAL_CONN_SEND_DREQ; + kibnal_set_conn_state(conn, IBNAL_CONN_DISCONNECTING); - spin_lock (&kibnal_data.kib_connd_lock); + kibnal_schedule_conn(conn); + kibnal_conn_decref(conn); /* lose ibc_list's ref */ +} - list_add_tail (&conn->ibc_list, &kibnal_data.kib_connd_conns); - wake_up (&kibnal_data.kib_connd_waitq); +void +kibnal_close_conn (kib_conn_t *conn, int error) +{ + unsigned long flags; + + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + + kibnal_close_conn_locked (conn, error); + + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); +} + +void +kibnal_handle_early_rxs(kib_conn_t *conn) +{ + unsigned long flags; + kib_rx_t *rx; + + LASSERT (!in_interrupt()); + LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED); + + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + while (!list_empty(&conn->ibc_early_rxs)) { + rx = list_entry(conn->ibc_early_rxs.next, + kib_rx_t, rx_list); + list_del(&rx->rx_list); + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); - spin_unlock (&kibnal_data.kib_connd_lock); + kibnal_handle_rx(rx); + + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + } + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); } void -kibnal_close_conn (kib_conn_t *conn, int error) +kibnal_conn_disconnected(kib_conn_t *conn) { - unsigned long flags; + static IB_QP_ATTRIBUTES_MODIFY qpam = {.RequestState = QPStateError}; - write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + LIST_HEAD (zombies); + struct list_head *tmp; + struct list_head *nxt; + kib_tx_t *tx; + FSTATUS frc; + int done; - kibnal_close_conn_locked (conn, error); + LASSERT (conn->ibc_state >= IBNAL_CONN_INIT_QP); + + kibnal_set_conn_state(conn, IBNAL_CONN_DISCONNECTED); + + /* move QP to error state to make posted work items complete */ + frc = iibt_qp_modify(conn->ibc_qp, &qpam, NULL); + if (frc != FSUCCESS) + CERROR("can't move qp state to error: %d\n", frc); + + spin_lock(&conn->ibc_lock); + + /* Complete all tx descs not waiting for sends to complete. + * NB we should be safe from RDMA now that the QP has changed state */ + + list_for_each_safe (tmp, nxt, &conn->ibc_tx_queue) { + tx = list_entry (tmp, kib_tx_t, tx_list); + + LASSERT (tx->tx_queued); + + tx->tx_status = -ECONNABORTED; + tx->tx_queued = 0; + tx->tx_waiting = 0; + + if (tx->tx_sending != 0) + continue; + + list_del (&tx->tx_list); + list_add (&tx->tx_list, &zombies); + } + + list_for_each_safe (tmp, nxt, &conn->ibc_active_txs) { + tx = list_entry (tmp, kib_tx_t, tx_list); + + LASSERT (!tx->tx_queued); + LASSERT (tx->tx_waiting || + tx->tx_sending != 0); + + tx->tx_status = -ECONNABORTED; + tx->tx_waiting = 0; + + if (tx->tx_sending != 0) + continue; + + list_del (&tx->tx_list); + list_add (&tx->tx_list, &zombies); + } - write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + spin_unlock(&conn->ibc_lock); + + while (!list_empty(&zombies)) { + tx = list_entry (zombies.next, kib_tx_t, tx_list); + + list_del(&tx->tx_list); + kibnal_tx_done (tx); + } + + kibnal_handle_early_rxs(conn); } -static void +void kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc) { LIST_HEAD (zombies); @@ -1742,13 +1872,14 @@ kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc) LASSERT (rc != 0); - write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); LASSERT (peer->ibp_connecting != 0); peer->ibp_connecting--; if (peer->ibp_connecting != 0) { - /* another connection attempt under way (loopback?)... */ + /* another connection attempt under way (e.g. STALE on first + * attempt)... */ write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); return; } @@ -1763,18 +1894,12 @@ kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc) MIN(peer->ibp_reconnect_interval, *kibnal_tunables.kib_max_reconnect_interval); - peer->ibp_reconnect_time = jiffies + + peer->ibp_reconnect_time = jiffies + peer->ibp_reconnect_interval * HZ; - - /* Take peer's blocked blocked transmits; I'll complete - * them with error */ - while (!list_empty (&peer->ibp_tx_queue)) { - tx = list_entry (peer->ibp_tx_queue.next, - kib_tx_t, tx_list); - - list_del (&tx->tx_list); - list_add_tail (&tx->tx_list, &zombies); - } + + /* Take peer's blocked transmits to complete with error */ + list_add(&zombies, &peer->ibp_tx_queue); + list_del_init(&peer->ibp_tx_queue); if (kibnal_peer_active(peer) && (peer->ibp_persistence == 0)) { @@ -1786,169 +1911,274 @@ kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc) LASSERT (list_empty(&peer->ibp_tx_queue)); } - write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); - - if (!list_empty (&zombies)) - CERROR ("Deleting messages for %s: connection failed\n", - libcfs_nid2str(peer->ibp_nid)); + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); - while (!list_empty (&zombies)) { + if (list_empty (&zombies)) + return; + + CERROR ("Deleting messages for %s: connection failed\n", + libcfs_nid2str(peer->ibp_nid)); + do { tx = list_entry (zombies.next, kib_tx_t, tx_list); list_del (&tx->tx_list); /* complete now */ tx->tx_status = -EHOSTUNREACH; kibnal_tx_done (tx); - } + } while (!list_empty (&zombies)); } -static void +void kibnal_connreq_done (kib_conn_t *conn, int active, int status) { - int state = conn->ibc_state; kib_peer_t *peer = conn->ibc_peer; + struct list_head txs; kib_tx_t *tx; unsigned long flags; int i; - /* passive connection has no connreq & vice versa */ - LASSERTF(!active == !(conn->ibc_connreq != NULL), - "%d %p\n", active, conn->ibc_connreq); - if (active) { - LIBCFS_FREE (conn->ibc_connreq, sizeof (*conn->ibc_connreq)); - conn->ibc_connreq = NULL; + LASSERT (conn->ibc_state >= IBNAL_CONN_INIT_QP); + LASSERT (conn->ibc_state < IBNAL_CONN_ESTABLISHED); + LASSERT (peer->ibp_connecting > 0); + + LIBCFS_FREE(conn->ibc_cvars, sizeof(*conn->ibc_cvars)); + conn->ibc_cvars = NULL; + + if (status != 0) { + /* failed to establish connection */ + kibnal_peer_connect_failed(conn->ibc_peer, active, status); + kibnal_conn_disconnected(conn); + kibnal_conn_decref(conn); /* Lose CM's ref */ + return; } - write_lock_irqsave (&kibnal_data.kib_global_lock, flags); + /* connection established */ + LASSERT(conn->ibc_state == IBNAL_CONN_CONNECTING); + kibnal_set_conn_state(conn, IBNAL_CONN_ESTABLISHED); - LASSERT (peer->ibp_connecting != 0); - - if (status == 0) { - /* connection established... */ - KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_CONNECTING); - conn->ibc_state = IBNAL_CONN_ESTABLISHED; - - if (!kibnal_peer_active(peer)) { - /* ...but peer deleted meantime */ - status = -ECONNABORTED; - } - } else { - KIB_ASSERT_CONN_STATE_RANGE(conn, IBNAL_CONN_INIT_QP, - IBNAL_CONN_CONNECTING); + CDEBUG(D_WARNING, "Connection %p -> %s ESTABLISHED\n", + conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + + write_lock_irqsave(&kibnal_data.kib_global_lock, flags); + + if (!kibnal_peer_active(peer)) { + /* peer has been deleted */ + kibnal_close_conn_locked(conn, -ECONNABORTED); + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + + kibnal_peer_connect_failed(conn->ibc_peer, active, + -ECONNABORTED); + kibnal_conn_decref(conn); /* lose CM's ref */ + return; } + + peer->ibp_connecting--; + peer->ibp_reconnect_interval = 0; /* OK to reconnect at any time */ - if (status == 0) { - /* Everything worked! */ + /* Add conn to peer's list and nuke any dangling conns from a different + * peer instance... */ + kibnal_conn_addref(conn); /* +1 ref for ibc_list */ + kibnal_close_stale_conns_locked(peer, conn->ibc_incarnation); - peer->ibp_connecting--; + /* grab txs blocking for a conn */ + list_add(&txs, &peer->ibp_tx_queue); + list_del_init(&peer->ibp_tx_queue); - /* +1 ref for ibc_list; caller(== CM)'s ref remains until - * the IB_CM_IDLE callback */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); - list_add (&conn->ibc_list, &peer->ibp_conns); - - peer->ibp_reconnect_interval = 0; /* OK to reconnect at any time */ + write_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); + + /* Schedule blocked txs */ + spin_lock (&conn->ibc_lock); + while (!list_empty (&txs)) { + tx = list_entry (txs.next, kib_tx_t, tx_list); + list_del (&tx->tx_list); - /* post blocked sends to the new connection */ - spin_lock (&conn->ibc_lock); - - while (!list_empty (&peer->ibp_tx_queue)) { - tx = list_entry (peer->ibp_tx_queue.next, - kib_tx_t, tx_list); - - list_del (&tx->tx_list); - - /* +1 ref for each tx */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); - kibnal_queue_tx_locked (tx, conn); + kibnal_queue_tx_locked (tx, conn); + } + spin_unlock (&conn->ibc_lock); + kibnal_check_sends (conn); +} + +void +kibnal_reject (lnet_nid_t nid, IB_HANDLE cep, int reason) +{ + static CM_REJECT_INFO msgs[] = {{.Reason = RC_USER_REJ}, + {.Reason = RC_NO_RESOURCES}}; + const int nmsg = sizeof(msgs)/sizeof(msgs[0]); + CM_REJECT_INFO *msg; + FSTATUS frc; + + for (msg = &msgs[0]; msg < &msgs[nmsg]; msg++) + if (msg->Reason == reason) + break; + + LASSERT (msg < &msgs[nmsg]); + + frc = iibt_cm_reject(cep, msg); + if (frc != FSUCCESS) + CERROR("Error %d rejecting %s\n", frc, libcfs_nid2str(nid)); +} + +void +kibnal_check_connreject(kib_conn_t *conn, int active, CM_REJECT_INFO *rej) +{ + kib_peer_t *peer = conn->ibc_peer; + unsigned long flags; + FSTATUS frc; + + if (rej->Reason != RC_STALE_CONN) { + CERROR("%s connection to %s rejected: %d\n", + active ? "Active" : "Passive", + libcfs_nid2str(peer->ibp_nid), rej->Reason); + } else { + if (!active) { + CERROR("Connection to %s rejected (stale QP)\n", + libcfs_nid2str(peer->ibp_nid)); + } else { + CWARN("Connection to %s rejected (stale QP): " + "retrying...\n", libcfs_nid2str(peer->ibp_nid)); + + /* retry from scratch to allocate a new conn + * which will use a different QP */ + write_lock_irqsave(&kibnal_data.kib_global_lock, + flags); + kibnal_schedule_active_connect_locked(peer); + write_unlock_irqrestore(&kibnal_data.kib_global_lock, + flags); } - - spin_unlock (&conn->ibc_lock); - /* Nuke any dangling conns from a different peer instance... */ - kibnal_close_stale_conns_locked (conn->ibc_peer, - conn->ibc_incarnation); + /* An FCM_DISCONNECTED callback is still outstanding: give it a + * ref since kibnal_connreq_done() drops the CM's ref on conn + * on failure */ + kibnal_conn_addref(conn); + } + + kibnal_connreq_done(conn, 1, -ECONNRESET); +} - write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); +void +kibnal_cm_disconnect_callback(kib_conn_t *conn, CM_CONN_INFO *info) +{ + CDEBUG(D_NET, "status 0x%x\n", info->Status); - /* queue up all the receives */ - for (i = 0; i < IBNAL_RX_MSGS; i++) { - /* +1 ref for rx desc */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_inc (&conn->ibc_refcount); + LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED); - CDEBUG(D_NET, "RX[%d] %p->%p - "LPX64"\n", - i, &conn->ibc_rxs[i], conn->ibc_rxs[i].rx_msg, - conn->ibc_rxs[i].rx_vaddr); + switch (info->Status) { + default: + LBUG(); + break; - kibnal_post_rx (&conn->ibc_rxs[i], 0); - } + case FCM_DISCONNECT_REPLY: + /* You can't get this if you set TIMEWAIT */ + CERROR("Unexpected FCM_DISCONNECT_REPLY for %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + LBUG(); + break; + + case FCM_DISCONNECT_REQUEST: + /* Schedule conn to iibt_cm_disconnect() if it wasn't already */ + kibnal_close_conn (conn, 0); + break; - kibnal_check_sends (conn); - return; + case FCM_DISCONNECTED: + CDEBUG(D_NET, "Connection %p -> %s disconnected.\n", + conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + kibnal_conn_decref(conn); /* Lose CM's ref */ + break; } +} - /* connection failed */ - if (state == IBNAL_CONN_CONNECTING) { - /* schedule for connd to close */ - kibnal_close_conn_locked (conn, status); - } else { - /* Don't have a CM comm_id; just wait for refs to drain */ - conn->ibc_state = IBNAL_CONN_DISCONNECTED; - } +void +kibnal_cm_passive_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +{ + kib_conn_t *conn = arg; - write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); + CDEBUG(D_NET, "status 0x%x\n", info->Status); + kibnal_set_conn_state(conn, IBNAL_CONN_CONNECTING); + + /* Established Connection Notifier */ + switch (info->Status) { + default: + CERROR("Unexpected status %d on Connection %p -> %s\n", + info->Status, conn, + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + LBUG(); + break; + + case FCM_CONNECT_TIMEOUT: + kibnal_connreq_done(conn, 0, -ETIMEDOUT); + break; + + case FCM_CONNECT_REJECT: + kibnal_check_connreject(conn, 0, &info->Info.Reject); + break; - kibnal_peer_connect_failed (conn->ibc_peer, active, status); + case FCM_CONNECT_ESTABLISHED: + kibnal_connreq_done(conn, 0, 0); + break; - /* If we didn't establish the connection we don't have to pass - * through the disconnect protocol before dropping the CM ref */ - if (state < IBNAL_CONN_CONNECTING) - kibnal_put_conn (conn); + case FCM_DISCONNECT_REQUEST: + case FCM_DISCONNECT_REPLY: + case FCM_DISCONNECTED: + kibnal_cm_disconnect_callback(conn, info); + break; + } } -static int -kibnal_accept (kib_conn_t **connp, IB_HANDLE *cep, - lnet_nid_t nid, __u64 incarnation, int queue_depth) +int +kibnal_accept (kib_conn_t **connp, kib_msg_t *msg, int nob) { - kib_conn_t *conn = kibnal_create_conn(); + lnet_nid_t nid; + kib_conn_t *conn; kib_peer_t *peer; kib_peer_t *peer2; unsigned long flags; + int rc; - if (conn == NULL) - return (-ENOMEM); + rc = kibnal_unpack_msg(msg, nob); + if (rc != 0) { + CERROR("Error %d unpacking connreq\n", rc); + return -EPROTO; + } - if (queue_depth != IBNAL_MSG_QUEUE_SIZE) { - CERROR("Can't accept %s: bad queue depth %d (%d expected)\n", + nid = msg->ibm_srcnid; + + if (msg->ibm_type != IBNAL_MSG_CONNREQ) { + CERROR("Can't accept %s: bad request type %d (%d expected)\n", + libcfs_nid2str(nid), msg->ibm_type, IBNAL_MSG_CONNREQ); + return -EPROTO; + } + + if (msg->ibm_dstnid != kibnal_data.kib_ni->ni_nid) { + CERROR("Can't accept %s: bad dst NID %s (%s expected)\n", libcfs_nid2str(nid), - queue_depth, IBNAL_MSG_QUEUE_SIZE); - atomic_dec (&conn->ibc_refcount); - kibnal_destroy_conn(conn); - return (-EPROTO); + libcfs_nid2str(msg->ibm_dstnid), + libcfs_nid2str(kibnal_data.kib_ni->ni_nid)); + return -EPROTO; } + if (msg->ibm_u.connparams.ibcp_queue_depth != IBNAL_MSG_QUEUE_SIZE || + msg->ibm_u.connparams.ibcp_max_msg_size > IBNAL_MSG_SIZE || + msg->ibm_u.connparams.ibcp_max_frags > IBNAL_MAX_RDMA_FRAGS) { + CERROR("Reject %s: q %d sz %d frag %d, (%d %d %d expected)\n", + libcfs_nid2str(nid), + msg->ibm_u.connparams.ibcp_queue_depth, + msg->ibm_u.connparams.ibcp_max_msg_size, + msg->ibm_u.connparams.ibcp_max_frags, + IBNAL_MSG_QUEUE_SIZE, + IBNAL_MSG_SIZE, + IBNAL_MAX_RDMA_FRAGS); + return -EPROTO; + } + + conn = kibnal_create_conn(nid); + if (conn == NULL) + return -ENOMEM; + /* assume 'nid' is a new peer */ - peer = kibnal_create_peer (nid); - if (peer == NULL) { - CDEBUG(D_NET, "--conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(conn->ibc_peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); - atomic_dec (&conn->ibc_refcount); - kibnal_destroy_conn(conn); - return (-ENOMEM); + rc = kibnal_create_peer(&peer, nid); + if (rc != 0) { + kibnal_conn_decref(conn); + return rc; } write_lock_irqsave (&kibnal_data.kib_global_lock, flags); @@ -1958,459 +2188,234 @@ kibnal_accept (kib_conn_t **connp, IB_HANDLE *cep, /* peer table takes my ref on peer */ list_add_tail (&peer->ibp_list, kibnal_nid2peerlist(nid)); } else { - kib_peer_decref (peer); + kibnal_peer_decref(peer); peer = peer2; } - kib_peer_addref(peer); /* +1 ref for conn */ + kibnal_peer_addref(peer); /* +1 ref for conn */ peer->ibp_connecting++; - write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); - + kibnal_set_conn_state(conn, IBNAL_CONN_CONNECTING); conn->ibc_peer = peer; - conn->ibc_state = IBNAL_CONN_CONNECTING; - /* conn->ibc_cep is set when cm_accept is called */ - conn->ibc_incarnation = incarnation; + conn->ibc_incarnation = msg->ibm_srcstamp; conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE; - *connp = conn; - return (0); -} - -static void kibnal_set_qp_state(IB_HANDLE *qp, IB_QP_STATE state) -{ - IB_QP_ATTRIBUTES_MODIFY modify_attr = {0,}; - FSTATUS frc; - - modify_attr.RequestState = state; - - frc = iibt_qp_modify(qp, &modify_attr, NULL); - if (frc != FSUCCESS) - CERROR("couldn't set qp state to %d, error %d\n", state, frc); -} - -static void kibnal_flush_pending(kib_conn_t *conn) -{ - LIST_HEAD (zombies); - struct list_head *tmp; - struct list_head *nxt; - kib_tx_t *tx; - unsigned long flags; - int done; - - /* NB we wait until the connection has closed before completing - * outstanding passive RDMAs so we can be sure the network can't - * touch the mapped memory any more. */ - KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_DISCONNECTED); - - /* set the QP to the error state so that we get flush callbacks - * on our posted receives which can then drop their conn refs */ - kibnal_set_qp_state(conn->ibc_qp, QPStateError); - - spin_lock_irqsave (&conn->ibc_lock, flags); - - /* grab passive RDMAs not waiting for the tx callback */ - list_for_each_safe (tmp, nxt, &conn->ibc_active_txs) { - tx = list_entry (tmp, kib_tx_t, tx_list); - - LASSERT (tx->tx_passive_rdma || - !tx->tx_passive_rdma_wait); - - LASSERT (tx->tx_passive_rdma_wait || - tx->tx_sending != 0); - - /* still waiting for tx callback? */ - if (!tx->tx_passive_rdma_wait) - continue; + write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags); - tx->tx_status = -ECONNABORTED; - tx->tx_passive_rdma_wait = 0; - done = (tx->tx_sending == 0); + *connp = conn; + return (0); +} - if (!done) - continue; +void +kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +{ - list_del (&tx->tx_list); - list_add (&tx->tx_list, &zombies); - } + CM_REQUEST_INFO *req = &info->Info.Request; + CM_REPLY_INFO *rep; + kib_conn_t *conn; + FSTATUS frc; + int rc; + + LASSERT(arg == NULL); /* no conn yet for passive */ - /* grab all blocked transmits */ - list_for_each_safe (tmp, nxt, &conn->ibc_tx_queue) { - tx = list_entry (tmp, kib_tx_t, tx_list); - - list_del (&tx->tx_list); - list_add (&tx->tx_list, &zombies); + CDEBUG(D_NET, "%x\n", info->Status); + + if (info->Status == FCM_CONNECT_CANCEL) { + up(&kibnal_data.kib_listener_signal); + return; } - spin_unlock_irqrestore (&conn->ibc_lock, flags); + LASSERT (info->Status == FCM_CONNECT_REQUEST); - while (!list_empty(&zombies)) { - tx = list_entry (zombies.next, kib_tx_t, tx_list); + rc = kibnal_accept(&conn, (kib_msg_t *)req->PrivateData, + CM_REQUEST_INFO_USER_LEN); + if (rc != 0) { + kibnal_reject(LNET_NID_ANY, cep, + (rc == -EPROTO) ? RC_USER_REJ : RC_NO_RESOURCES); + return; + } - list_del(&tx->tx_list); - kibnal_tx_done (tx); + conn->ibc_cvars->cv_path = req->PathInfo.Path; + + rc = kibnal_conn_rts(conn, + req->CEPInfo.QPN, + req->CEPInfo.OfferedInitiatorDepth, + req->CEPInfo.OfferedResponderResources, + req->CEPInfo.StartingPSN); + if (rc != 0) { + kibnal_reject(conn->ibc_peer->ibp_nid, cep, RC_NO_RESOURCES); + kibnal_connreq_done(conn, 0, -ECONNABORTED); + return; } -} -static void -kibnal_reject (IB_HANDLE cep, uint16_t reason) -{ - CM_REJECT_INFO *rej; + memset(&conn->ibc_cvars->cv_cmci, 0, sizeof(conn->ibc_cvars->cv_cmci)); + rep = &conn->ibc_cvars->cv_cmci.Info.Reply; - LIBCFS_ALLOC(rej, sizeof(*rej)); - if (rej == NULL) /* LIBCFS_ALLOC() will CERROR on failure */ - return; + rep->QPN = conn->ibc_cvars->cv_qpattrs.QPNumber; + rep->QKey = conn->ibc_cvars->cv_qpattrs.Qkey; + rep->StartingPSN = conn->ibc_cvars->cv_qpattrs.RecvPSN; + rep->EndToEndFlowControl = conn->ibc_cvars->cv_qpattrs.FlowControl; + rep->ArbInitiatorDepth = conn->ibc_cvars->cv_qpattrs.InitiatorDepth; + rep->ArbResponderResources = conn->ibc_cvars->cv_qpattrs.ResponderResources; + rep->TargetAckDelay = kibnal_data.kib_hca_attrs.LocalCaAckDelay; + rep->FailoverAccepted = IBNAL_FAILOVER_ACCEPTED; + rep->RnRRetryCount = req->CEPInfo.RnrRetryCount; + + CLASSERT (CM_REPLY_INFO_USER_LEN >= + offsetof(kib_msg_t, ibm_u) + sizeof(kib_connparams_t)); - rej->Reason = reason; - iibt_cm_reject(cep, rej); - LIBCFS_FREE(rej, sizeof(*rej)); -} + kibnal_pack_connmsg((kib_msg_t *)rep->PrivateData, + CM_REPLY_INFO_USER_LEN, + IBNAL_MSG_CONNACK, + conn->ibc_peer->ibp_nid, conn->ibc_incarnation); -static FSTATUS -kibnal_qp_rts(IB_HANDLE qp_handle, __u32 qpn, __u8 resp_res, - IB_PATH_RECORD *path, __u8 init_depth, __u32 send_psn) -{ - IB_QP_ATTRIBUTES_MODIFY modify_attr; - FSTATUS frc; - ENTRY; - - modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { - .RequestState = QPStateReadyToRecv, - .RecvPSN = IBNAL_STARTING_PSN, - .DestQPNumber = qpn, - .ResponderResources = resp_res, - .MinRnrTimer = UsecToRnrNakTimer(2000), /* 20 ms */ - .Attrs = (IB_QP_ATTR_RECVPSN | - IB_QP_ATTR_DESTQPNUMBER | - IB_QP_ATTR_RESPONDERRESOURCES | - IB_QP_ATTR_DESTAV | - IB_QP_ATTR_PATHMTU | - IB_QP_ATTR_MINRNRTIMER), - }; - GetAVFromPath(0, path, &modify_attr.PathMTU, NULL, - &modify_attr.DestAV); - - frc = iibt_qp_modify(qp_handle, &modify_attr, NULL); - if (frc != FSUCCESS) - RETURN(frc); - - modify_attr = (IB_QP_ATTRIBUTES_MODIFY) { - .RequestState = QPStateReadyToSend, - .FlowControl = TRUE, - .InitiatorDepth = init_depth, - .SendPSN = send_psn, - .LocalAckTimeout = path->PktLifeTime + 2, /* 2 or 1? */ - .RetryCount = IBNAL_RETRY, - .RnrRetryCount = IBNAL_RNR_RETRY, - .Attrs = (IB_QP_ATTR_FLOWCONTROL | - IB_QP_ATTR_INITIATORDEPTH | - IB_QP_ATTR_SENDPSN | - IB_QP_ATTR_LOCALACKTIMEOUT | - IB_QP_ATTR_RETRYCOUNT | - IB_QP_ATTR_RNRRETRYCOUNT), - }; + LASSERT (conn->ibc_cep == NULL); + kibnal_set_conn_state(conn, IBNAL_CONN_CONNECTING); + + frc = iibt_cm_accept(cep, + &conn->ibc_cvars->cv_cmci, + NULL, + kibnal_cm_passive_callback, conn, + &conn->ibc_cep); - frc = iibt_qp_modify(qp_handle, &modify_attr, NULL); - RETURN(frc); + if (frc == FSUCCESS || frc == FPENDING) + return; + + CERROR("iibt_cm_accept(%s) failed: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + kibnal_connreq_done(conn, 0, -ECONNABORTED); } -static void -kibnal_connect_reply (IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +void +kibnal_check_connreply(kib_conn_t *conn, CM_REPLY_INFO *rep) { - IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; - kib_conn_t *conn = arg; - kib_wire_connreq_t *wcr; - CM_REPLY_INFO *rep = &info->Info.Reply; - uint16_t reason; - FSTATUS frc; + kib_msg_t *msg = (kib_msg_t *)rep->PrivateData; + lnet_nid_t nid = conn->ibc_peer->ibp_nid; + FSTATUS frc; + int rc; - wcr = (kib_wire_connreq_t *)info->Info.Reply.PrivateData; + rc = kibnal_unpack_msg(msg, CM_REPLY_INFO_USER_LEN); + if (rc != 0) { + CERROR ("Error %d unpacking connack from %s\n", + rc, libcfs_nid2str(nid)); + kibnal_reject(nid, conn->ibc_cep, RC_USER_REJ); + kibnal_connreq_done(conn, 1, -EPROTO); + return; + } + + if (msg->ibm_type != IBNAL_MSG_CONNACK) { + CERROR("Bad connack request type %d (%d expected) from %s\n", + msg->ibm_type, IBNAL_MSG_CONNREQ, + libcfs_nid2str(msg->ibm_srcnid)); + kibnal_reject(nid, conn->ibc_cep, RC_USER_REJ); + kibnal_connreq_done(conn, 1, -EPROTO); + return; + } - if (wcr->wcr_magic != cpu_to_le32(IBNAL_MSG_MAGIC)) { - CERROR ("Can't connect %s: bad magic %08x\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - le32_to_cpu(wcr->wcr_magic)); - GOTO(reject, reason = RC_USER_REJ); + if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid || + msg->ibm_dstnid != kibnal_data.kib_ni->ni_nid || + msg->ibm_dststamp != kibnal_data.kib_incarnation) { + CERROR("Stale connack from %s(%s): %s(%s), "LPX64"("LPX64")\n", + libcfs_nid2str(msg->ibm_srcnid), + libcfs_nid2str(conn->ibc_peer->ibp_nid), + libcfs_nid2str(msg->ibm_dstnid), + libcfs_nid2str(kibnal_data.kib_ni->ni_nid), + msg->ibm_dststamp, kibnal_data.kib_incarnation); + kibnal_reject(nid, conn->ibc_cep, RC_USER_REJ); + kibnal_connreq_done(conn, 1, -EPROTO); + return; } - if (wcr->wcr_version != cpu_to_le16(IBNAL_MSG_VERSION)) { - CERROR ("Can't connect %s: bad version %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - le16_to_cpu(wcr->wcr_magic)); - GOTO(reject, reason = RC_USER_REJ); - } - - if (wcr->wcr_queue_depth != cpu_to_le16(IBNAL_MSG_QUEUE_SIZE)) { - CERROR ("Can't connect %s: bad queue depth %d\n", - libcfs_nid2str(conn->ibc_peer->ibp_nid), - le16_to_cpu(wcr->wcr_queue_depth)); - GOTO(reject, reason = RC_USER_REJ); + if (msg->ibm_u.connparams.ibcp_queue_depth != IBNAL_MSG_QUEUE_SIZE || + msg->ibm_u.connparams.ibcp_max_msg_size > IBNAL_MSG_SIZE || + msg->ibm_u.connparams.ibcp_max_frags > IBNAL_MAX_RDMA_FRAGS) { + CERROR("Reject %s: q %d sz %d frag %d, (%d %d %d expected)\n", + libcfs_nid2str(msg->ibm_srcnid), + msg->ibm_u.connparams.ibcp_queue_depth, + msg->ibm_u.connparams.ibcp_max_msg_size, + msg->ibm_u.connparams.ibcp_max_frags, + IBNAL_MSG_QUEUE_SIZE, + IBNAL_MSG_SIZE, + IBNAL_MAX_RDMA_FRAGS); + kibnal_reject(nid, conn->ibc_cep, RC_USER_REJ); + kibnal_connreq_done(conn, 1, -EPROTO); + return; } - if (le64_to_cpu(wcr->wcr_nid) != conn->ibc_peer->ibp_nid) { - CERROR ("Unexpected NID %s from %s\n", - libcfs_nid2str(le64_to_cpu(wcr->wcr_nid)), - libcfs_nid2str(conn->ibc_peer->ibp_nid)); - GOTO(reject, reason = RC_USER_REJ); - } - CDEBUG(D_NET, "Connection %p -> %s REP_RECEIVED.\n", conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - conn->ibc_incarnation = le64_to_cpu(wcr->wcr_incarnation); + conn->ibc_incarnation = msg->ibm_srcstamp; conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE; - frc = kibnal_qp_rts(conn->ibc_qp, rep->QPN, - min_t(__u8, rep->ArbInitiatorDepth, - ca_attr->MaxQPResponderResources), - &conn->ibc_connreq->cr_path, - min_t(__u8, rep->ArbResponderResources, - ca_attr->MaxQPInitiatorDepth), - rep->StartingPSN); - if (frc != FSUCCESS) { - CERROR("Connection %p -> %s QP RTS/RTR failed: %d\n", - conn, libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); - GOTO(reject, reason = RC_NO_QP); - } - - /* the callback arguments are ignored for an active accept */ - conn->ibc_connreq->cr_discarded.Status = FSUCCESS; - frc = iibt_cm_accept(cep, &conn->ibc_connreq->cr_discarded, - NULL, NULL, NULL, NULL); - if (frc != FCM_CONNECT_ESTABLISHED) { - CERROR("Connection %p -> %s CMAccept failed: %d\n", - conn, libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); - kibnal_connreq_done (conn, 1, -ECONNABORTED); - /* XXX don't call reject after accept fails? */ + rc = kibnal_conn_rts(conn, + rep->QPN, + rep->ArbInitiatorDepth, + rep->ArbResponderResources, + rep->StartingPSN); + if (rc != 0) { + kibnal_reject(nid, conn->ibc_cep, RC_NO_RESOURCES); + kibnal_connreq_done(conn, 1, -EIO); return; } - CDEBUG(D_NET, "Connection %p -> %s Established\n", - conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + memset(&conn->ibc_cvars->cv_cmci, 0, sizeof(conn->ibc_cvars->cv_cmci)); + + frc = iibt_cm_accept(conn->ibc_cep, + &conn->ibc_cvars->cv_cmci, + NULL, NULL, NULL, NULL); + + if (frc == FCM_CONNECT_ESTABLISHED) + kibnal_connreq_done(conn, 1, 0); - kibnal_connreq_done (conn, 1, 0); - return; -reject: - kibnal_reject(cep, reason); - kibnal_connreq_done (conn, 1, -EPROTO); + CERROR("Connection %p -> %s CMAccept failed: %d\n", + conn, libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + kibnal_connreq_done(conn, 1, -ECONNABORTED); } -/* ib_cm.h has a wealth of information on the CM procedures */ -static void -kibnal_cm_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) +void +kibnal_cm_active_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) { kib_conn_t *conn = arg; CDEBUG(D_NET, "status 0x%x\n", info->Status); - /* Established Connection Notifier */ switch (info->Status) { default: - CERROR("unknown status %d on Connection %p -> %s\n", + CERROR("unknown status %d on Connection %p -> %s\n", info->Status, conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); LBUG(); break; - case FCM_CONNECT_REPLY: - kibnal_connect_reply(cep, info, arg); + case FCM_CONNECT_TIMEOUT: + kibnal_connreq_done(conn, 1, -ETIMEDOUT); + break; + + case FCM_CONNECT_REJECT: + kibnal_check_connreject(conn, 1, &info->Info.Reject); break; - case FCM_DISCONNECT_REQUEST: - /* XXX lock around these state management bits? */ - if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) - kibnal_close_conn (conn, 0); - conn->ibc_state = IBNAL_CONN_DREP; - iibt_cm_disconnect(conn->ibc_cep, NULL, NULL); + case FCM_CONNECT_REPLY: + kibnal_check_connreply(conn, &info->Info.Reply); break; - /* these both guarantee that no more cm callbacks will occur */ - case FCM_DISCONNECTED: /* aka FCM_DISCONNECT_TIMEOUT */ + case FCM_DISCONNECT_REQUEST: case FCM_DISCONNECT_REPLY: - CDEBUG(D_NET, "Connection %p -> %s disconnect done.\n", - conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - - conn->ibc_state = IBNAL_CONN_DISCONNECTED; - kibnal_flush_pending(conn); - kibnal_put_conn(conn); /* Lose CM's ref */ + case FCM_DISCONNECTED: + kibnal_cm_disconnect_callback(conn, info); break; } - - return; -} - -static int -kibnal_set_cm_flags(IB_HANDLE cep) -{ - FSTATUS frc; - uint32 value = 1; - - frc = iibt_cm_modify_cep(cep, CM_FLAG_TIMEWAIT_CALLBACK, - (char *)&value, sizeof(value), 0); - if (frc != FSUCCESS) { - CERROR("error setting timeout callback: %d\n", frc); - return -1; - } - -#if 0 - frc = iibt_cm_modify_cep(cep, CM_FLAG_ASYNC_ACCEPT, (char *)&value, - sizeof(value), 0); - if (frc != FSUCCESS) { - CERROR("error setting async accept: %d\n", frc); - return -1; - } -#endif - - return 0; } void -kibnal_listen_callback(IB_HANDLE cep, CM_CONN_INFO *info, void *arg) -{ - IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; - IB_QP_ATTRIBUTES_QUERY *query; - CM_REQUEST_INFO *req; - CM_CONN_INFO *rep = NULL, *rcv = NULL; - kib_wire_connreq_t *wcr; - kib_conn_t *conn = NULL; - uint16_t reason = 0; - FSTATUS frc; - int rc = 0; - - LASSERT(cep); - LASSERT(info); - LASSERT(arg == NULL); /* no conn yet for passive */ - - CDEBUG(D_NET, "status 0x%x\n", info->Status); - - req = &info->Info.Request; - wcr = (kib_wire_connreq_t *)req->PrivateData; - - CDEBUG(D_NET, "%d from %s\n", info->Status, - libcfs_nid2str(le64_to_cpu(wcr->wcr_nid))); - - if (info->Status == FCM_CONNECT_CANCEL) - return; - - LASSERT (info->Status == FCM_CONNECT_REQUEST); - - if (wcr->wcr_magic != cpu_to_le32(IBNAL_MSG_MAGIC)) { - CERROR ("Can't accept: bad magic %08x\n", - le32_to_cpu(wcr->wcr_magic)); - GOTO(out, reason = RC_USER_REJ); - } - - if (wcr->wcr_version != cpu_to_le16(IBNAL_MSG_VERSION)) { - CERROR ("Can't accept: bad version %d\n", - le16_to_cpu(wcr->wcr_magic)); - GOTO(out, reason = RC_USER_REJ); - } - - rc = kibnal_accept(&conn, cep, - le64_to_cpu(wcr->wcr_nid), - le64_to_cpu(wcr->wcr_incarnation), - le16_to_cpu(wcr->wcr_queue_depth)); - if (rc != 0) { - CERROR ("Can't accept %s: %d\n", - libcfs_nid2str(le64_to_cpu(wcr->wcr_nid)), rc); - GOTO(out, reason = RC_NO_RESOURCES); - } - - frc = kibnal_qp_rts(conn->ibc_qp, req->CEPInfo.QPN, - min_t(__u8, req->CEPInfo.OfferedInitiatorDepth, - ca_attr->MaxQPResponderResources), - &req->PathInfo.Path, - min_t(__u8, req->CEPInfo.OfferedResponderResources, - ca_attr->MaxQPInitiatorDepth), - req->CEPInfo.StartingPSN); - - if (frc != FSUCCESS) { - CERROR ("Can't mark QP RTS/RTR %s: %d\n", - libcfs_nid2str(le64_to_cpu(wcr->wcr_nid)), frc); - GOTO(out, reason = RC_NO_QP); - } - - frc = iibt_qp_query(conn->ibc_qp, &conn->ibc_qp_attrs, NULL); - if (frc != FSUCCESS) { - CERROR ("Couldn't query qp attributes %s: %d\n", - libcfs_nid2str(le64_to_cpu(wcr->wcr_nid)), frc); - GOTO(out, reason = RC_NO_QP); - } - query = &conn->ibc_qp_attrs; - - LIBCFS_ALLOC(rep, sizeof(*rep)); - LIBCFS_ALLOC(rcv, sizeof(*rcv)); - if (rep == NULL || rcv == NULL) { - if (rep) LIBCFS_FREE(rep, sizeof(*rep)); - if (rcv) LIBCFS_FREE(rcv, sizeof(*rcv)); - CERROR ("can't allocate reply and receive buffers\n"); - GOTO(out, reason = RC_INSUFFICIENT_RESP_RES); - } - - /* don't try to deref this into the incoming wcr :) */ - wcr = (kib_wire_connreq_t *)rep->Info.Reply.PrivateData; - - rep->Info.Reply = (CM_REPLY_INFO) { - .QPN = query->QPNumber, - .QKey = query->Qkey, - .StartingPSN = query->RecvPSN, - .EndToEndFlowControl = query->FlowControl, - /* XXX Hmm. */ - .ArbInitiatorDepth = query->InitiatorDepth, - .ArbResponderResources = query->ResponderResources, - .TargetAckDelay = 0, - .FailoverAccepted = 0, - .RnRRetryCount = req->CEPInfo.RnrRetryCount, - }; - - *wcr = (kib_wire_connreq_t) { - .wcr_magic = cpu_to_le32(IBNAL_MSG_MAGIC), - .wcr_version = cpu_to_le16(IBNAL_MSG_VERSION), - .wcr_queue_depth = cpu_to_le32(IBNAL_MSG_QUEUE_SIZE), - .wcr_nid = cpu_to_le64(kibnal_data.kib_ni->ni_nid), - .wcr_incarnation = cpu_to_le64(kibnal_data.kib_incarnation), - }; - - frc = iibt_cm_accept(cep, rep, rcv, kibnal_cm_callback, conn, - &conn->ibc_cep); - - LIBCFS_FREE(rep, sizeof(*rep)); - LIBCFS_FREE(rcv, sizeof(*rcv)); - - if (frc != FCM_CONNECT_ESTABLISHED) { - /* XXX it seems we don't call reject after this point? */ - CERROR("iibt_cm_accept() failed: %d, aborting\n", frc); - rc = -ECONNABORTED; - goto out; - } - - if (kibnal_set_cm_flags(conn->ibc_cep)) { - rc = -ECONNABORTED; - goto out; - } - - CDEBUG(D_WARNING, "Connection %p -> %s ESTABLISHED.\n", - conn, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - -out: - if (reason) { - kibnal_reject(cep, reason); - rc = -ECONNABORTED; - } - if (conn != NULL) - kibnal_connreq_done(conn, 0, rc); - - return; -} - -static void dump_path_records(PATH_RESULTS *results) { IB_PATH_RECORD *path; int i; - for(i = 0; i < results->NumPathRecords; i++) { + for (i = 0; i < results->NumPathRecords; i++) { path = &results->PathRecords[i]; CDEBUG(D_NET, "%d: sgid "LPX64":"LPX64" dgid " LPX64":"LPX64" pkey %x\n", @@ -2423,113 +2428,88 @@ dump_path_records(PATH_RESULTS *results) } } -static void -kibnal_pathreq_callback (void *arg, QUERY *query, - QUERY_RESULT_VALUES *query_res) +void +kibnal_pathreq_callback (void *arg, QUERY *qry, + QUERY_RESULT_VALUES *qrslt) { - IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; - kib_conn_t *conn = arg; - PATH_RESULTS *path; - FSTATUS frc; - lnet_nid_t nid; + IB_CA_ATTRIBUTES *ca_attr = &kibnal_data.kib_hca_attrs; + kib_conn_t *conn = arg; + CM_REQUEST_INFO *req = &conn->ibc_cvars->cv_cmci.Info.Request; + PATH_RESULTS *path = (PATH_RESULTS *)qrslt->QueryResult; + FSTATUS frc; - if (query_res->Status != FSUCCESS || query_res->ResultDataSize == 0) { - CERROR ("status %d data size %d\n", query_res->Status, - query_res->ResultDataSize); - kibnal_connreq_done (conn, 1, -EINVAL); + if (qrslt->Status != FSUCCESS || + qrslt->ResultDataSize < sizeof(*path)) { + CERROR ("pathreq %s failed: status %d data size %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + qrslt->Status, qrslt->ResultDataSize); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); return; } - path = (PATH_RESULTS *)query_res->QueryResult; - if (path->NumPathRecords < 1) { - CERROR ("expected path records: %d\n", path->NumPathRecords); - kibnal_connreq_done (conn, 1, -EINVAL); + CERROR ("pathreq %s failed: no path records\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); return; } - dump_path_records(path); + //dump_path_records(path); + conn->ibc_cvars->cv_path = path->PathRecords[0]; - /* just using the first. this is probably a horrible idea. */ - conn->ibc_connreq->cr_path = path->PathRecords[0]; + LASSERT (conn->ibc_cep == NULL); - conn->ibc_cep = iibt_cm_create_cep(CM_RC_TYPE); + conn->ibc_cep = kibnal_create_cep(conn->ibc_peer->ibp_nid); if (conn->ibc_cep == NULL) { - CERROR ("Can't create CEP\n"); - kibnal_connreq_done (conn, 1, -EINVAL); - return; - } - - if (kibnal_set_cm_flags(conn->ibc_cep)) { - kibnal_connreq_done (conn, 1, -EINVAL); + kibnal_connreq_done(conn, 1, -ENOMEM); return; } - conn->ibc_connreq->cr_wcr = (kib_wire_connreq_t) { - .wcr_magic = cpu_to_le32(IBNAL_MSG_MAGIC), - .wcr_version = cpu_to_le16(IBNAL_MSG_VERSION), - .wcr_queue_depth = cpu_to_le16(IBNAL_MSG_QUEUE_SIZE), - .wcr_nid = cpu_to_le64(kibnal_data.kib_ni->ni_nid), - .wcr_incarnation = cpu_to_le64(kibnal_data.kib_incarnation), - }; - - conn->ibc_connreq->cr_cmreq = (CM_REQUEST_INFO) { - .SID = conn->ibc_connreq->cr_service.RID.ServiceID, - .CEPInfo = (CM_CEP_INFO) { - .CaGUID = kibnal_data.kib_hca_guids[0], - .EndToEndFlowControl = FALSE, - .PortGUID = conn->ibc_connreq->cr_path.SGID.Type.Global.InterfaceID, - .RetryCount = IBNAL_RETRY, - .RnrRetryCount = IBNAL_RNR_RETRY, - .AckTimeout = IBNAL_ACK_TIMEOUT, - .StartingPSN = IBNAL_STARTING_PSN, - .QPN = conn->ibc_qp_attrs.QPNumber, - .QKey = conn->ibc_qp_attrs.Qkey, - .OfferedResponderResources = ca_attr->MaxQPResponderResources, - .OfferedInitiatorDepth = ca_attr->MaxQPInitiatorDepth, - }, - .PathInfo = (CM_CEP_PATHINFO) { - .bSubnetLocal = TRUE, - .Path = conn->ibc_connreq->cr_path, - }, - }; + memset(req, 0, sizeof(*req)); + req->SID = conn->ibc_cvars->cv_svcrec.RID.ServiceID; + req->CEPInfo.CaGUID = kibnal_data.kib_hca_guids[kibnal_data.kib_hca_idx]; + req->CEPInfo.EndToEndFlowControl = IBNAL_EE_FLOW; + req->CEPInfo.PortGUID = conn->ibc_cvars->cv_path.SGID.Type.Global.InterfaceID; + req->CEPInfo.RetryCount = IBNAL_RETRY; + req->CEPInfo.RnrRetryCount = IBNAL_RNR_RETRY; + req->CEPInfo.AckTimeout = IBNAL_ACK_TIMEOUT; + req->CEPInfo.StartingPSN = IBNAL_STARTING_PSN; + req->CEPInfo.QPN = conn->ibc_cvars->cv_qpattrs.QPNumber; + req->CEPInfo.QKey = conn->ibc_cvars->cv_qpattrs.Qkey; + req->CEPInfo.OfferedResponderResources = ca_attr->MaxQPResponderResources; + req->CEPInfo.OfferedInitiatorDepth = ca_attr->MaxQPInitiatorDepth; + req->PathInfo.bSubnetLocal = IBNAL_LOCAL_SUB; + req->PathInfo.Path = conn->ibc_cvars->cv_path; + + CLASSERT (CM_REQUEST_INFO_USER_LEN >= + offsetof(kib_msg_t, ibm_u) + sizeof(kib_connparams_t)); + + kibnal_pack_connmsg((kib_msg_t *)req->PrivateData, + CM_REQUEST_INFO_USER_LEN, + IBNAL_MSG_CONNREQ, + conn->ibc_peer->ibp_nid, 0); -#if 0 - /* XXX set timeout just like SDP!!!*/ - conn->ibc_connreq->cr_path.packet_life = 13; -#endif /* Flag I'm getting involved with the CM... */ - conn->ibc_state = IBNAL_CONN_CONNECTING; - - nid = *kibnal_service_nid_field(&conn->ibc_connreq->cr_service); - - CDEBUG(D_NET, "Connecting to, service id "LPX64", on %s\n", - conn->ibc_connreq->cr_service.RID.ServiceID, - libcfs_nid2str(nid)); + kibnal_set_conn_state(conn, IBNAL_CONN_CONNECTING); - memset(conn->ibc_connreq->cr_cmreq.PrivateData, 0, - CM_REQUEST_INFO_USER_LEN); - memcpy(conn->ibc_connreq->cr_cmreq.PrivateData, - &conn->ibc_connreq->cr_wcr, sizeof(conn->ibc_connreq->cr_wcr)); - - /* kibnal_cm_callback gets my conn ref */ - frc = iibt_cm_connect(conn->ibc_cep, &conn->ibc_connreq->cr_cmreq, - kibnal_cm_callback, conn); - if (frc != FPENDING && frc != FSUCCESS) { - CERROR ("Connect: %d\n", frc); - /* Back out state change as connect failed */ - conn->ibc_state = IBNAL_CONN_INIT_QP; - kibnal_connreq_done (conn, 1, -EINVAL); - } + /* cm callback gets my conn ref */ + frc = iibt_cm_connect(conn->ibc_cep, req, + kibnal_cm_active_callback, conn); + if (frc == FPENDING || frc == FSUCCESS) + return; + + CERROR ("Connect %s failed: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); } -static void -dump_service_records(SERVICE_RECORD_RESULTS *results) +void +kibnal_dump_service_records(SERVICE_RECORD_RESULTS *results) { IB_SERVICE_RECORD *svc; int i; - for(i = 0; i < results->NumServiceRecords; i++) { + for (i = 0; i < results->NumServiceRecords; i++) { svc = &results->ServiceRecords[i]; CDEBUG(D_NET, "%d: sid "LPX64" gid "LPX64":"LPX64" pkey %x\n", i, @@ -2540,76 +2520,71 @@ dump_service_records(SERVICE_RECORD_RESULTS *results) } } - -static void -kibnal_service_get_callback (void *arg, QUERY *query, - QUERY_RESULT_VALUES *query_res) +void +kibnal_service_get_callback (void *arg, QUERY *qry, + QUERY_RESULT_VALUES *qrslt) { - kib_conn_t *conn = arg; - SERVICE_RECORD_RESULTS *svc; - COMMAND_CONTROL_PARAMETERS sd_params; - QUERY path_query; - FSTATUS frc; - lnet_nid_t nid; - - if (query_res->Status != FSUCCESS || query_res->ResultDataSize == 0) { - CERROR ("status %d data size %d\n", query_res->Status, - query_res->ResultDataSize); - kibnal_connreq_done (conn, 1, -EINVAL); + kib_conn_t *conn = arg; + SERVICE_RECORD_RESULTS *svc; + FSTATUS frc; + lnet_nid_t nid; + + if (qrslt->Status != FSUCCESS || + qrslt->ResultDataSize < sizeof(*svc)) { + CERROR ("Lookup %s failed: status %d data size %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), + qrslt->Status, qrslt->ResultDataSize); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); return; } - svc = (SERVICE_RECORD_RESULTS *)query_res->QueryResult; - + svc = (SERVICE_RECORD_RESULTS *)qrslt->QueryResult; if (svc->NumServiceRecords < 1) { - CERROR ("%d service records\n", svc->NumServiceRecords); - kibnal_connreq_done (conn, 1, -EINVAL); + CERROR ("lookup %s failed: no service records\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); return; } - dump_service_records(svc); - - conn->ibc_connreq->cr_service = svc->ServiceRecords[0]; - nid = *kibnal_service_nid_field(&conn->ibc_connreq->cr_service); - - CDEBUG(D_NET, "Got status %d, service id "LPX64", on %s\n", - query_res->Status , conn->ibc_connreq->cr_service.RID.ServiceID, - libcfs_nid2str(nid)); + //kibnal_dump_service_records(svc); + conn->ibc_cvars->cv_svcrec = svc->ServiceRecords[0]; - memset(&path_query, 0, sizeof(path_query)); - path_query.InputType = InputTypePortGuidPair; - path_query.OutputType = OutputTypePathRecord; - path_query.InputValue.PortGuidPair.SourcePortGuid = kibnal_data.kib_port_guid; - path_query.InputValue.PortGuidPair.DestPortGuid = conn->ibc_connreq->cr_service.RID.ServiceGID.Type.Global.InterfaceID; + qry = &conn->ibc_cvars->cv_query; + memset(qry, 0, sizeof(*qry)); - memset(&sd_params, 0, sizeof(sd_params)); - sd_params.RetryCount = IBNAL_RETRY; - sd_params.Timeout = 10 * 1000; /* wait 10 seconds */ + qry->OutputType = OutputTypePathRecord; + qry->InputType = InputTypePortGuidPair; - /* kibnal_service_get_callback gets my conn ref */ + qry->InputValue.PortGuidPair.SourcePortGuid = + kibnal_data.kib_port_guid; + qry->InputValue.PortGuidPair.DestPortGuid = + conn->ibc_cvars->cv_svcrec.RID.ServiceGID.Type.Global.InterfaceID; + /* kibnal_pathreq_callback gets my conn ref */ frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, kibnal_data.kib_port_guid, - &path_query, + qry, kibnal_pathreq_callback, - &sd_params, conn); + &kibnal_data.kib_sdretry, + conn); if (frc == FPENDING) return; - CERROR ("Path record request failed: %d\n", frc); - kibnal_connreq_done (conn, 1, -EINVAL); + CERROR ("pathreq %s failed: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); } -static void +void kibnal_connect_peer (kib_peer_t *peer) { - COMMAND_CONTROL_PARAMETERS sd_params; - QUERY query; - FSTATUS frc; - kib_conn_t *conn = kibnal_create_conn(); + QUERY *qry; + FSTATUS frc; + kib_conn_t *conn; LASSERT (peer->ibp_connecting != 0); + conn = kibnal_create_conn(peer->ibp_nid); if (conn == NULL) { CERROR ("Can't allocate conn\n"); kibnal_peer_connect_failed (peer, 1, -ENOMEM); @@ -2617,59 +2592,49 @@ kibnal_connect_peer (kib_peer_t *peer) } conn->ibc_peer = peer; - kib_peer_addref(peer); - - LIBCFS_ALLOC (conn->ibc_connreq, sizeof (*conn->ibc_connreq)); - if (conn->ibc_connreq == NULL) { - CERROR ("Can't allocate connreq\n"); - kibnal_connreq_done (conn, 1, -ENOMEM); - return; - } + kibnal_peer_addref(peer); - memset(conn->ibc_connreq, 0, sizeof (*conn->ibc_connreq)); + qry = &conn->ibc_cvars->cv_query; + memset(qry, 0, sizeof(*qry)); - kibnal_set_service_keys(&conn->ibc_connreq->cr_service, peer->ibp_nid); + qry->OutputType = OutputTypeServiceRecord; + qry->InputType = InputTypeServiceRecord; - memset(&query, 0, sizeof(query)); - query.InputType = InputTypeServiceRecord; - query.OutputType = OutputTypeServiceRecord; - query.InputValue.ServiceRecordValue.ServiceRecord = conn->ibc_connreq->cr_service; - query.InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK; - - memset(&sd_params, 0, sizeof(sd_params)); - sd_params.RetryCount = IBNAL_RETRY; - sd_params.Timeout = 10 * 1000; /* wait 10 seconds */ + qry->InputValue.ServiceRecordValue.ComponentMask = + KIBNAL_SERVICE_KEY_MASK; + kibnal_set_service_keys( + &qry->InputValue.ServiceRecordValue.ServiceRecord, + peer->ibp_nid); /* kibnal_service_get_callback gets my conn ref */ frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd, kibnal_data.kib_port_guid, - &query, - kibnal_service_get_callback, - &sd_params, conn); + qry, + kibnal_service_get_callback, + &kibnal_data.kib_sdretry, + conn); if (frc == FPENDING) return; - CERROR ("iibt_sd_query_port_fabric_information(): %d\n", frc); - kibnal_connreq_done (conn, 1, frc); + CERROR("Lookup %s failed: %d\n", libcfs_nid2str(peer->ibp_nid), frc); + kibnal_connreq_done(conn, 1, -EHOSTUNREACH); } -static int +int kibnal_conn_timed_out (kib_conn_t *conn) { kib_tx_t *tx; struct list_head *ttmp; - unsigned long flags; - spin_lock_irqsave (&conn->ibc_lock, flags); + spin_lock(&conn->ibc_lock); list_for_each (ttmp, &conn->ibc_tx_queue) { tx = list_entry (ttmp, kib_tx_t, tx_list); - LASSERT (!tx->tx_passive_rdma_wait); - LASSERT (tx->tx_sending == 0); + LASSERT (tx->tx_queued); if (time_after_eq (jiffies, tx->tx_deadline)) { - spin_unlock_irqrestore (&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); return 1; } } @@ -2677,24 +2642,21 @@ kibnal_conn_timed_out (kib_conn_t *conn) list_for_each (ttmp, &conn->ibc_active_txs) { tx = list_entry (ttmp, kib_tx_t, tx_list); - LASSERT (tx->tx_passive_rdma || - !tx->tx_passive_rdma_wait); - - LASSERT (tx->tx_passive_rdma_wait || + LASSERT (!tx->tx_queued); + LASSERT (tx->tx_waiting || tx->tx_sending != 0); if (time_after_eq (jiffies, tx->tx_deadline)) { - spin_unlock_irqrestore (&conn->ibc_lock, flags); + spin_unlock(&conn->ibc_lock); return 1; } } - spin_unlock_irqrestore (&conn->ibc_lock, flags); - + spin_unlock(&conn->ibc_lock); return 0; } -static void +void kibnal_check_conns (int idx) { struct list_head *peers = &kibnal_data.kib_peers[idx]; @@ -2716,7 +2678,7 @@ kibnal_check_conns (int idx) list_for_each (ctmp, &peer->ibp_conns) { conn = list_entry (ctmp, kib_conn_t, ibc_list); - KIB_ASSERT_CONN_STATE(conn, IBNAL_CONN_ESTABLISHED); + LASSERT (conn->ibc_state == IBNAL_CONN_ESTABLISHED); /* In case we have enough credits to return via a * NOOP, but there were no non-blocking tx descs @@ -2725,13 +2687,13 @@ kibnal_check_conns (int idx) if (!kibnal_conn_timed_out(conn)) continue; + + /* Handle timeout by closing the whole connection. We + * can only be sure RDMA activity has ceased once the + * QP has been modified. */ - CDEBUG(D_NET, "++conn[%p] state %d -> %s (%d)\n", - conn, conn->ibc_state, - libcfs_nid2str(peer->ibp_nid), - atomic_read (&conn->ibc_refcount)); + kibnal_conn_addref(conn); /* 1 ref for me... */ - atomic_inc (&conn->ibc_refcount); read_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); @@ -2739,7 +2701,7 @@ kibnal_check_conns (int idx) libcfs_nid2str(peer->ibp_nid)); kibnal_close_conn (conn, -ETIMEDOUT); - kibnal_put_conn (conn); + kibnal_conn_decref(conn); /* ...until here */ /* start again now I've dropped the lock */ goto again; @@ -2749,37 +2711,32 @@ kibnal_check_conns (int idx) read_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); } -static void -kib_connd_handle_state(kib_conn_t *conn) +void +kibnal_disconnect_conn (kib_conn_t *conn) { - FSTATUS frc; - - switch (conn->ibc_state) { - /* all refs have gone, free and be done with it */ - case IBNAL_CONN_DISCONNECTED: - kibnal_destroy_conn (conn); - return; /* avoid put_conn */ - - case IBNAL_CONN_SEND_DREQ: - frc = iibt_cm_disconnect(conn->ibc_cep, NULL, NULL); - if (frc != FSUCCESS) /* XXX do real things */ - CERROR("disconnect failed: %d\n", frc); - conn->ibc_state = IBNAL_CONN_DREQ; - break; + FSTATUS frc; - /* a callback got to the conn before we did */ - case IBNAL_CONN_DREP: - break; - - default: - CERROR ("Bad conn %p state: %d\n", conn, - conn->ibc_state); - LBUG(); - break; - } + LASSERT (conn->ibc_state == IBNAL_CONN_DISCONNECTING); - /* drop ref from close_conn */ - kibnal_put_conn(conn); + kibnal_conn_disconnected(conn); + + frc = iibt_cm_disconnect(conn->ibc_cep, NULL, NULL); + switch (frc) { + case FSUCCESS: + break; + + case FINSUFFICIENT_RESOURCES: + CERROR("ENOMEM disconnecting %s\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid)); + /* This might cause the module to become unloadable since the + * FCM_DISCONNECTED callback is still outstanding */ + break; + + default: + CERROR("Unexpected error disconnecting %s: %d\n", + libcfs_nid2str(conn->ibc_peer->ibp_nid), frc); + LBUG(); + } } int @@ -2791,6 +2748,7 @@ kibnal_connd (void *arg) kib_peer_t *peer; int timeout; int i; + int did_something; int peer_index = 0; unsigned long deadline = jiffies; @@ -2799,19 +2757,34 @@ kibnal_connd (void *arg) init_waitqueue_entry (&wait, current); - spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags); + + while (!kibnal_data.kib_shutdown) { + did_something = 0; + + if (!list_empty (&kibnal_data.kib_connd_zombies)) { + conn = list_entry (kibnal_data.kib_connd_zombies.next, + kib_conn_t, ibc_list); + list_del (&conn->ibc_list); + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + did_something = 1; + + kibnal_destroy_conn(conn); + + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + } - for (;;) { if (!list_empty (&kibnal_data.kib_connd_conns)) { conn = list_entry (kibnal_data.kib_connd_conns.next, kib_conn_t, ibc_list); list_del (&conn->ibc_list); - spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); - kib_connd_handle_state(conn); + did_something = 1; + kibnal_disconnect_conn(conn); + kibnal_conn_decref(conn); + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); - continue; } if (!list_empty (&kibnal_data.kib_connd_peers)) { @@ -2820,26 +2793,22 @@ kibnal_connd (void *arg) list_del_init (&peer->ibp_connd_list); spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + did_something = 1; kibnal_connect_peer (peer); - kib_peer_decref (peer); + kibnal_peer_decref (peer); spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); } - /* shut down and nobody left to reap... */ - if (kibnal_data.kib_shutdown && - atomic_read(&kibnal_data.kib_nconns) == 0) - break; - - spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); - /* careful with the jiffy wrap... */ while ((timeout = (int)(deadline - jiffies)) <= 0) { const int n = 4; const int p = 1; int chunk = kibnal_data.kib_peer_hash_size; + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); + /* Time to check for RDMA timeouts on a few more * peers: I do checks every 'p' seconds on a * proportion of the peer table and I need to check @@ -2848,9 +2817,9 @@ kibnal_connd (void *arg) * connection within (n+1)/n times the timeout * interval. */ - if (kibnal_tunables.kib_io_timeout > n * p) + if (*kibnal_tunables.kib_timeout > n * p) chunk = (chunk * n * p) / - kibnal_tunables.kib_io_timeout; + *kibnal_tunables.kib_timeout; if (chunk == 0) chunk = 1; @@ -2861,9 +2830,14 @@ kibnal_connd (void *arg) } deadline += p * HZ; + spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + did_something = 1; } - kibnal_data.kib_connd_waketime = jiffies + timeout; + if (did_something) + continue; + + spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags); set_current_state (TASK_INTERRUPTIBLE); add_wait_queue (&kibnal_data.kib_connd_waitq, &wait); @@ -2885,78 +2859,152 @@ kibnal_connd (void *arg) return (0); } + +void +kibnal_hca_async_callback (void *hca_arg, IB_EVENT_RECORD *ev) +{ + /* XXX flesh out. this seems largely for async errors */ + CERROR("type: %d code: %u\n", ev->EventType, ev->EventCode); +} + +void +kibnal_hca_callback (void *hca_arg, void *cq_arg) +{ + unsigned long flags; + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + kibnal_data.kib_ready = 1; + wake_up(&kibnal_data.kib_sched_waitq); + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); +} + int kibnal_scheduler(void *arg) { - long id = (long)arg; - char name[16]; - kib_rx_t *rx; - kib_tx_t *tx; - unsigned long flags; - int rc; - int counter = 0; - int did_something; + long id = (long)arg; + wait_queue_t wait; + char name[16]; + FSTATUS frc; + FSTATUS frc2; + IB_WORK_COMPLETION wc; + kib_rx_t *rx; + kib_tx_t *tx; + unsigned long flags; + int rc; + int did_something; + __u64 rxseq = 0; + int busy_loops = 0; snprintf(name, sizeof(name), "kibnal_sd_%02ld", id); libcfs_daemonize(name); libcfs_blockallsigs(); - spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + init_waitqueue_entry(&wait, current); - for (;;) { - did_something = 0; + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); - while (!list_empty(&kibnal_data.kib_sched_txq)) { - tx = list_entry(kibnal_data.kib_sched_txq.next, - kib_tx_t, tx_list); - list_del(&tx->tx_list); + while (!kibnal_data.kib_shutdown) { + if (busy_loops++ >= IBNAL_RESCHED) { spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); - kibnal_tx_done(tx); - spin_lock_irqsave(&kibnal_data.kib_sched_lock, - flags); + our_cond_resched(); + busy_loops = 0; + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); } - if (!list_empty(&kibnal_data.kib_sched_rxq)) { - rx = list_entry(kibnal_data.kib_sched_rxq.next, - kib_rx_t, rx_list); - list_del(&rx->rx_list); + if (kibnal_data.kib_ready && + !kibnal_data.kib_checking_cq) { + /* take ownership of completion polling */ + kibnal_data.kib_checking_cq = 1; + /* Assume I'll exhaust the CQ */ + kibnal_data.kib_ready = 0; spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); + + frc = iibt_cq_poll(kibnal_data.kib_cq, &wc); + if (frc == FNOT_DONE) { + /* CQ empty */ + frc2 = iibt_cq_rearm(kibnal_data.kib_cq, + CQEventSelNextWC); + LASSERT (frc2 == FSUCCESS); + } + + if (frc == FSUCCESS && + kibnal_wreqid2type(wc.WorkReqId) == IBNAL_WID_RX) { + rx = (kib_rx_t *)kibnal_wreqid2ptr(wc.WorkReqId); + + /* Grab the RX sequence number NOW before + * anyone else can get an RX completion */ + rxseq = rx->rx_conn->ibc_rxseq++; + } + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + /* give up ownership of completion polling */ + kibnal_data.kib_checking_cq = 0; - kibnal_rx(rx); + if (frc == FNOT_DONE) + continue; - did_something = 1; - spin_lock_irqsave(&kibnal_data.kib_sched_lock, - flags); - } + LASSERT (frc == FSUCCESS); + /* Assume there's more: get another scheduler to check + * while I handle this completion... */ - /* shut down and no receives to complete... */ - if (kibnal_data.kib_shutdown && - atomic_read(&kibnal_data.kib_nconns) == 0) - break; + kibnal_data.kib_ready = 1; + wake_up(&kibnal_data.kib_sched_waitq); - /* nothing to do or hogging CPU */ - if (!did_something || counter++ == IBNAL_RESCHED) { spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); - counter = 0; - - if (!did_something) { - rc = wait_event_interruptible( - kibnal_data.kib_sched_waitq, - !list_empty(&kibnal_data.kib_sched_txq) || - !list_empty(&kibnal_data.kib_sched_rxq) || - (kibnal_data.kib_shutdown && - atomic_read (&kibnal_data.kib_nconns) == 0)); - } else { - our_cond_resched(); - } - spin_lock_irqsave(&kibnal_data.kib_sched_lock, - flags); + switch (kibnal_wreqid2type(wc.WorkReqId)) { + case IBNAL_WID_RX: + kibnal_rx_complete(&wc, rxseq); + break; + + case IBNAL_WID_TX: + kibnal_tx_complete(&wc); + break; + + case IBNAL_WID_RDMA: + /* We only get RDMA completion notification if + * it fails. So we just ignore them completely + * because... + * + * 1) If an RDMA fails, all subsequent work + * items, including the final SEND will fail + * too, so I'm still guaranteed to notice that + * this connection is hosed. + * + * 2) It's positively dangerous to look inside + * the tx descriptor obtained from an RDMA work + * item. As soon as I drop the kib_sched_lock, + * I give a scheduler on another CPU a chance + * to get the final SEND completion, so the tx + * descriptor can get freed as I inspect it. */ + CERROR ("RDMA failed: %d\n", wc.Status); + break; + + default: + LBUG(); + } + + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); + continue; } + + /* Nothing to do; sleep... */ + + set_current_state(TASK_INTERRUPTIBLE); + add_wait_queue(&kibnal_data.kib_sched_waitq, &wait); + spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, + flags); + + schedule(); + + remove_wait_queue(&kibnal_data.kib_sched_waitq, &wait); + set_current_state(TASK_RUNNING); + spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags); } spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags); diff --git a/lnet/klnds/iiblnd/iiblnd_modparams.c b/lnet/klnds/iiblnd/iiblnd_modparams.c new file mode 100644 index 0000000..ab9f0d3 --- /dev/null +++ b/lnet/klnds/iiblnd/iiblnd_modparams.c @@ -0,0 +1,160 @@ +/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- + * vim:expandtab:shiftwidth=8:tabstop=8: + * + * Copyright (C) 2004 Cluster File Systems, Inc. + * Author: Eric Barton + * + * This file is part of Lustre, http://www.lustre.org. + * + * Lustre is free software; you can redistribute it and/or + * modify it under the terms of version 2 of the GNU General Public + * License as published by the Free Software Foundation. + * + * Lustre is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with Lustre; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + */ + +#include "iiblnd.h" + +static char *ipif_basename = IBNAL_IPIF_BASENAME; +CFS_MODULE_PARM(ipif_basename, "s", charp, 0444, + "IPoIB interface base name"); + +static char *service_name = IBNAL_SERVICE_NAME; +CFS_MODULE_PARM(service_name, "s", charp, 0444, + "IB service name"); + +static int service_number = IBNAL_SERVICE_NUMBER; +CFS_MODULE_PARM(service_number, "i", int, 0444, + "IB service number"); + +static int min_reconnect_interval = IBNAL_MIN_RECONNECT_INTERVAL; +CFS_MODULE_PARM(min_reconnect_interval, "i", int, 0644, + "minimum connection retry interval (seconds)"); + +static int max_reconnect_interval = IBNAL_MAX_RECONNECT_INTERVAL; +CFS_MODULE_PARM(max_reconnect_interval, "i", int, 0644, + "maximum connection retry interval (seconds)"); + +static int concurrent_peers = IBNAL_CONCURRENT_PEERS; +CFS_MODULE_PARM(concurrent_peers, "i", int, 0444, + "maximum number of peers that may connect"); + +static int cksum = IBNAL_CKSUM; +CFS_MODULE_PARM(cksum, "i", int, 0644, + "set non-zero to enable message (not RDMA) checksums"); + +static int timeout = IBNAL_TIMEOUT; +CFS_MODULE_PARM(timeout, "i", int, 0644, + "timeout (seconds)"); + +static int ntx = IBNAL_NTX; +CFS_MODULE_PARM(ntx, "i", int, 0444, + "# of message descriptors"); + +static int credits = IBNAL_CREDITS; +CFS_MODULE_PARM(credits, "i", int, 0444, + "# concurrent sends"); + +static int peer_credits = IBNAL_PEERCREDITS; +CFS_MODULE_PARM(peer_credits, "i", int, 0444, + "# concurrent sends to 1 peer"); + +static int sd_retries = IBNAL_SD_RETRIES; +CFS_MODULE_PARM(sa_retries, "i", int, 0444, + "# times to retry SD queries"); + +kib_tunables_t kibnal_tunables = { + .kib_ipif_basename = &ipif_basename, + .kib_service_name = &service_name, + .kib_service_number = &service_number, + .kib_min_reconnect_interval = &min_reconnect_interval, + .kib_max_reconnect_interval = &max_reconnect_interval, + .kib_concurrent_peers = &concurrent_peers, + .kib_cksum = &cksum, + .kib_timeout = &timeout, + .kib_ntx = &ntx, + .kib_credits = &credits, + .kib_peercredits = &peer_credits, + .kib_sd_retries = &sd_retries, +}; + +#if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM + +/* NB max_size specified for proc_dostring entries only needs to be big enough + * not to truncate the printout; it only needs to be the actual size of the + * string buffer if we allow writes (and we don't) */ + +static ctl_table kibnal_ctl_table[] = { + {1, "ipif_basename", &ipif_basename, + 1024, 0444, NULL, &proc_dostring}, + {2, "service_name", &service_name, + 1024, 0444, NULL, &proc_dostring}, + {3, "service_number", &service_number, + sizeof(int), 0444, NULL, &proc_dointvec}, + {4, "min_reconnect_interval", &min_reconnect_interval, + sizeof(int), 0644, NULL, &proc_dointvec}, + {5, "max_reconnect_interval", &max_reconnect_interval, + sizeof(int), 0644, NULL, &proc_dointvec}, + {6, "concurrent_peers", &concurrent_peers, + sizeof(int), 0444, NULL, &proc_dointvec}, + {7, "cksum", &cksum, + sizeof(int), 0644, NULL, &proc_dointvec}, + {8, "timeout", &timeout, + sizeof(int), 0644, NULL, &proc_dointvec}, + {9, "ntx", &ntx, + sizeof(int), 0444, NULL, &proc_dointvec}, + {10, "credits", &credits, + sizeof(int), 0444, NULL, &proc_dointvec}, + {11, "peer_credits", &peer_credits, + sizeof(int), 0444, NULL, &proc_dointvec}, + {12, "sd_retries", &sd_retries, + sizeof(int), 0444, NULL, &proc_dointvec}, + {0} +}; + +static ctl_table kibnal_top_ctl_table[] = { + {203, "openibnal", NULL, 0, 0555, kibnal_ctl_table}, + {0} +}; + +int +kibnal_tunables_init () +{ + kibnal_tunables.kib_sysctl = + register_sysctl_table(kibnal_top_ctl_table, 0); + + if (kibnal_tunables.kib_sysctl == NULL) + CWARN("Can't setup /proc tunables\n"); + + return 0; +} + +void +kibnal_tunables_fini () +{ + if (kibnal_tunables.kib_sysctl != NULL) + unregister_sysctl_table(kibnal_tunables.kib_sysctl); +} + +#else + +int +kibnal_tunables_init () +{ + return 0; +} + +void +kibnal_tunables_fini () +{ +} + +#endif diff --git a/lnet/klnds/openiblnd/openiblnd.c b/lnet/klnds/openiblnd/openiblnd.c index 54310c2..7a9da5e 100644 --- a/lnet/klnds/openiblnd/openiblnd.c +++ b/lnet/klnds/openiblnd/openiblnd.c @@ -1672,7 +1672,7 @@ kibnal_module_init (void) } MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel OpenIB NAL v1.00"); +MODULE_DESCRIPTION("Kernel OpenIB LND v1.00"); MODULE_LICENSE("GPL"); module_init(kibnal_module_init); diff --git a/lnet/klnds/ralnd/ralnd.c b/lnet/klnds/ralnd/ralnd.c index a590115..b9324f9 100644 --- a/lnet/klnds/ralnd/ralnd.c +++ b/lnet/klnds/ralnd/ralnd.c @@ -1645,7 +1645,7 @@ kranal_module_init (void) } MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel RapidArray NAL v0.01"); +MODULE_DESCRIPTION("Kernel RapidArray LND v0.01"); MODULE_LICENSE("GPL"); module_init(kranal_module_init); diff --git a/lnet/klnds/socklnd/socklnd.c b/lnet/klnds/socklnd/socklnd.c index 3009a84..c77096a 100644 --- a/lnet/klnds/socklnd/socklnd.c +++ b/lnet/klnds/socklnd/socklnd.c @@ -2148,8 +2148,10 @@ ksocknal_enumerate_interfaces(ksock_net_t *net) int up; __u32 ip; __u32 mask; + if (!strcmp(names[i], "lo")) /* skip the loopback IF */ continue; + rc = libcfs_ipif_query(names[i], &up, &ip, &mask); if (rc != 0) { CWARN("Can't get interface %s info: %d\n", @@ -2283,7 +2285,7 @@ ksocknal_module_init (void) } MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel TCP Socket NAL v1.0.0"); +MODULE_DESCRIPTION("Kernel TCP Socket LND v1.0.0"); MODULE_LICENSE("GPL"); cfs_module(ksocknal, "1.0.0", ksocknal_module_init, ksocknal_module_fini); diff --git a/lnet/klnds/viblnd/viblnd.c b/lnet/klnds/viblnd/viblnd.c index 14baf37..c4ad266 100644 --- a/lnet/klnds/viblnd/viblnd.c +++ b/lnet/klnds/viblnd/viblnd.c @@ -1145,8 +1145,8 @@ kibnal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) switch(cmd) { case IOC_LIBCFS_GET_PEER: { lnet_nid_t nid = 0; - __u32 ip = 0; - int share_count = 0; + __u32 ip = 0; + int share_count = 0; rc = kibnal_get_peer_info(data->ioc_count, &nid, &ip, &share_count); @@ -1498,8 +1498,6 @@ kibnal_shutdown (lnet_ni_t *ni) LASSERT (list_empty (&kibnal_data.kib_peers[i])); } LASSERT (atomic_read (&kibnal_data.kib_nconns) == 0); - LASSERT (list_empty (&kibnal_data.kib_sched_rxq)); - LASSERT (list_empty (&kibnal_data.kib_sched_txq)); LASSERT (list_empty (&kibnal_data.kib_connd_zombies)); LASSERT (list_empty (&kibnal_data.kib_connd_conns)); LASSERT (list_empty (&kibnal_data.kib_connd_pcreqs)); @@ -1655,8 +1653,6 @@ kibnal_startup (lnet_ni_t *ni) init_waitqueue_head (&kibnal_data.kib_connd_waitq); spin_lock_init (&kibnal_data.kib_sched_lock); - INIT_LIST_HEAD (&kibnal_data.kib_sched_txq); - INIT_LIST_HEAD (&kibnal_data.kib_sched_rxq); init_waitqueue_head (&kibnal_data.kib_sched_waitq); spin_lock_init (&kibnal_data.kib_tx_lock); @@ -1898,7 +1894,7 @@ kibnal_module_init (void) } MODULE_AUTHOR("Cluster File Systems, Inc. "); -MODULE_DESCRIPTION("Kernel Voltaire IB NAL v1.00"); +MODULE_DESCRIPTION("Kernel Voltaire IB LND v1.00"); MODULE_LICENSE("GPL"); module_init(kibnal_module_init); diff --git a/lnet/klnds/viblnd/viblnd.h b/lnet/klnds/viblnd/viblnd.h index a894934..9fcac45 100644 --- a/lnet/klnds/viblnd/viblnd.h +++ b/lnet/klnds/viblnd/viblnd.h @@ -223,7 +223,6 @@ typedef struct cm_cep_handle_t kib_listen_handle; /* IB listen handle */ rwlock_t kib_global_lock; /* stabilize peer/conn ops */ - spinlock_t kib_vverbs_lock; /* serialize vverbs calls */ int kib_ready; /* CQ callback fired */ int kib_checking_cq; /* a scheduler is checking the CQ */ @@ -241,8 +240,6 @@ typedef struct spinlock_t kib_connd_lock; /* serialise */ wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */ - struct list_head kib_sched_txq; /* tx requiring attention */ - struct list_head kib_sched_rxq; /* rx requiring attention */ spinlock_t kib_sched_lock; /* serialise */ struct kib_tx *kib_tx_descs; /* all the tx descriptors */ @@ -298,7 +295,7 @@ typedef struct kib_tx /* transmit message */ int tx_status; /* completion status */ unsigned long tx_deadline; /* completion deadline */ __u64 tx_cookie; /* completion cookie */ - lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */ + lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */ vv_l_key_t tx_lkey; /* local key for message buffer */ kib_msg_t *tx_msg; /* message buffer (host vaddr) */ int tx_nwrq; /* # send work items */ @@ -446,12 +443,6 @@ extern int kibnal_init_rdma(kib_tx_t *tx, int type, int nob, extern int kibnal_tunables_init(void); extern void kibnal_tunables_fini(void); -static inline int -wrq_signals_completion (vv_wr_t *wrq) -{ - return wrq->completion_notification != 0; -} - #define kibnal_conn_addref(conn) \ do { \ CDEBUG(D_NET, "conn[%p] (%d)++\n", \ diff --git a/lnet/klnds/viblnd/viblnd_cb.c b/lnet/klnds/viblnd/viblnd_cb.c index 05e3bd1..1c83872 100644 --- a/lnet/klnds/viblnd/viblnd_cb.c +++ b/lnet/klnds/viblnd/viblnd_cb.c @@ -82,7 +82,6 @@ kibnal_get_idle_tx (void) spin_lock(&kibnal_data.kib_tx_lock); - /* "normal" descriptor is free */ if (list_empty (&kibnal_data.kib_idle_txs)) { spin_unlock(&kibnal_data.kib_tx_lock); return NULL; @@ -412,7 +411,7 @@ kibnal_rx_complete (kib_rx_t *rx, vv_comp_status_t vvrc, int nob, __u64 rxseq) unsigned long flags; int rc; - CDEBUG (D_NET, "rx %p conn %p\n", rx, conn); + CDEBUG(D_NET, "rx %p conn %p\n", rx, conn); LASSERT (rx->rx_nob < 0); /* was posted */ rx->rx_nob = 0; /* isn't now */ @@ -834,7 +833,7 @@ void kibnal_check_sends (kib_conn_t *conn) { kib_tx_t *tx; - vv_return_t vvrc; + vv_return_t vvrc; int rc; int done; @@ -1370,7 +1369,7 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) /* NB 'private' is different depending on what we're sending.... */ - CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n", + CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n", payload_nob, payload_niov, libcfs_id2str(target)); LASSERT (payload_nob == 0 || payload_niov > 0); @@ -1527,7 +1526,6 @@ kibnal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) if (nob <= IBNAL_MSG_SIZE) break; /* send IMMEDIATE */ - /* may block if caller is app thread */ tx = kibnal_get_idle_tx(); if (tx == NULL) { CERROR("Can't allocate %s txd for %s\n", @@ -1610,7 +1608,6 @@ kibnal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, int rc = 0; LASSERT (mlen <= rlen); - LASSERT (mlen >= 0); LASSERT (!in_interrupt()); /* Either all pages or all vaddrs */ LASSERT (!(kiov != NULL && iov != NULL)); @@ -1749,7 +1746,7 @@ kibnal_schedule_conn (kib_conn_t *conn) void kibnal_close_conn_locked (kib_conn_t *conn, int error) { - /* This just does the immmediate housekeeping. 'error' is zero for a + /* This just does the immediate housekeeping. 'error' is zero for a * normal shutdown which can happen only after the connection has been * established. If the connection is established, schedule the * connection to be finished off by the connd. Otherwise the connd is @@ -1782,7 +1779,6 @@ kibnal_close_conn_locked (kib_conn_t *conn, int error) list_empty(&conn->ibc_tx_queue) ? "" : "(sending)", list_empty(&conn->ibc_active_txs) ? "" : "(waiting)", conn->ibc_txseq, conn->ibc_rxseq); - #if 0 /* can't skip down the queue without holding ibc_lock (see above) */ list_for_each(tmp, &conn->ibc_tx_queue) { @@ -3012,7 +3008,7 @@ kibnal_check_conns (int idx) kibnal_conn_addref(conn); /* 1 ref for me... */ - read_unlock_irqrestore(&kibnal_data.kib_global_lock, + read_unlock_irqrestore(&kibnal_data.kib_global_lock, flags); CERROR("Timed out RDMA with %s\n", @@ -3095,7 +3091,7 @@ kibnal_connd (void *arg) init_waitqueue_entry (&wait, current); kibnal_data.kib_connd = current; - spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags); + spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags); while (!kibnal_data.kib_shutdown) { diff --git a/lnet/lnet/api-ni.c b/lnet/lnet/api-ni.c index abc0782..7b4d613 100644 --- a/lnet/lnet/api-ni.c +++ b/lnet/lnet/api-ni.c @@ -184,7 +184,7 @@ lnet_get_portals_compatibility(void) return 0; } -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD void lnet_init_locks(void) { @@ -1044,7 +1044,7 @@ lnet_startup_lndnis (void) the_lnet.ln_eqwaitni = ni; } } else { -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD LCONSOLE_ERROR("LND %s not supported in a " "single-threaded runtime\n", libcfs_lnd2str(lnd_type)); @@ -1096,7 +1096,7 @@ lnet_startup_lndnis (void) } #ifndef __KERNEL__ -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD extern lnd_t the_ptllnd; # else extern lnd_t the_tcplnd; @@ -1128,7 +1128,7 @@ LNetInit(void) #else /* Register all LNDs that have been loaded * NB the order here determines default 'networks=' order */ -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD lnet_register_lnd(&the_ptllnd); # else lnet_register_lnd(&the_tcplnd); diff --git a/lnet/lnet/config.c b/lnet/lnet/config.c index a8ebb08..46ff65a 100644 --- a/lnet/lnet/config.c +++ b/lnet/lnet/config.c @@ -309,6 +309,7 @@ lnet_new_text_buf (int str_len) lnet_text_buf_t *ltb; int nob; + /* NB allocate space for the terminating 0 */ nob = offsetof(lnet_text_buf_t, ltb_text[str_len + 1]); if (nob > LNET_SINGLE_TEXTBUF_NOB) { /* _way_ conservative for "route net gateway..." */ @@ -326,6 +327,7 @@ lnet_new_text_buf (int str_len) return NULL; ltb->ltb_size = nob; + ltb->ltb_text[0] = 0; lnet_tbnob += nob; return ltb; } @@ -389,7 +391,7 @@ lnet_str2tbs_sep (struct list_head *tbs, char *str) nob = sep - str; if (nob > 0) { - ltb = lnet_new_text_buf(nob + 1); + ltb = lnet_new_text_buf(nob); if (ltb == NULL) { lnet_free_text_bufs(&pending); return -1; @@ -435,7 +437,7 @@ lnet_expand1tb (struct list_head *list, LASSERT (*sep1 == '['); LASSERT (*sep2 == ']'); - ltb = lnet_new_text_buf(len1 + itemlen + len2 + 1); + ltb = lnet_new_text_buf(len1 + itemlen + len2); if (ltb == NULL) return -ENOMEM; diff --git a/lnet/lnet/lib-eq.c b/lnet/lnet/lib-eq.c index 2c5390a..cafcb86 100644 --- a/lnet/lnet/lib-eq.c +++ b/lnet/lnet/lib-eq.c @@ -179,7 +179,7 @@ LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms, cfs_waitlink_t wl; cfs_time_t now; #else -# if !LNET_SINGLE_THREADED +# if HAVE_LIBPTHREAD struct timeval then; struct timeval now; struct timespec ts; @@ -246,7 +246,7 @@ LNetEQPoll (lnet_handle_eq_t *eventqs, int neq, int timeout_ms, continue; } -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD LNET_UNLOCK(); return -ENOENT; # else diff --git a/lnet/lnet/lib-msg.c b/lnet/lnet/lib-msg.c index 21cecab..45c2afe 100644 --- a/lnet/lnet/lib-msg.c +++ b/lnet/lnet/lib-msg.c @@ -57,7 +57,7 @@ lnet_enq_event_locked (lnet_eq_t *eq, lnet_event_t *ev) if (cfs_waitq_active(&the_lnet.ln_waitq)) cfs_waitq_broadcast(&the_lnet.ln_waitq); #else -# if LNET_SINGLE_THREADED +# if !HAVE_LIBPTHREAD /* LNetEQPoll() calls into _the_ LND to wait for action */ # else /* Wake anyone waiting in LNetEQPoll() */ diff --git a/lnet/lnet/peer.c b/lnet/lnet/peer.c index c6c8b83..9fb4f35 100644 --- a/lnet/lnet/peer.c +++ b/lnet/lnet/peer.c @@ -162,6 +162,8 @@ lnet_nid2peer_locked(lnet_peer_t **lpp, lnet_nid_t nid) do_gettimeofday (&now); + memset(lp, 0, sizeof(*lp)); /* zero counters etc */ + CFS_INIT_LIST_HEAD(&lp->lp_txq); CFS_INIT_LIST_HEAD(&lp->lp_rtrq); diff --git a/lnet/ulnds/ptllnd/ptllnd.c b/lnet/ulnds/ptllnd/ptllnd.c index a414461..9872d18 100644 --- a/lnet/ulnds/ptllnd/ptllnd.c +++ b/lnet/ulnds/ptllnd/ptllnd.c @@ -437,11 +437,11 @@ ptllnd_startup (lnet_ni_t *ni) */ ni->ni_nid = ptl2lnetnid(ni,plni->plni_portals_id.nid); - PJK_UT_MSG("ptl pid=" FMT_NID "\n",plni->plni_portals_id.pid); + PJK_UT_MSG("ptl pid=" FMT_PID "\n",plni->plni_portals_id.pid); PJK_UT_MSG("ptl nid=" FMT_NID "\n",plni->plni_portals_id.nid); PJK_UT_MSG("lnet nid=" LPX64 " (passed back)\n",ni->ni_nid); - CDEBUG(D_INFO,"ptl pid=" FMT_NID "\n",plni->plni_portals_id.pid); + CDEBUG(D_INFO,"ptl pid=" FMT_PID "\n",plni->plni_portals_id.pid); CDEBUG(D_INFO,"ptl nid=" FMT_NID "\n",plni->plni_portals_id.nid); CDEBUG(D_INFO,"lnet nid=" LPX64 "\n",ni->ni_nid); diff --git a/lnet/ulnds/socklnd/procapi.c b/lnet/ulnds/socklnd/procapi.c index 7397da1..bf9e33a 100644 --- a/lnet/ulnds/socklnd/procapi.c +++ b/lnet/ulnds/socklnd/procapi.c @@ -46,7 +46,7 @@ # include #endif -#if LNET_SINGLE_THREADED +#if !HAVE_LIBPTHREAD # error "This LND requires a multi-threaded runtime" #endif -- 1.8.3.1