X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd_cb.c;h=4cff9eda91bb90686e2ab946f7c3906e7aebab7a;hp=7175f0d777c7089fc3f442d27d0c1fa5f6c3d7f6;hb=4c4c327b25f3414f20a9ae600e7311f1aa3a866d;hpb=e5d5d8d09a4be22ab225bac0db3f4abc0e24b296 diff --git a/lnet/klnds/gnilnd/gnilnd_cb.c b/lnet/klnds/gnilnd/gnilnd_cb.c index 7175f0d..4cff9ed 100644 --- a/lnet/klnds/gnilnd/gnilnd_cb.c +++ b/lnet/klnds/gnilnd/gnilnd_cb.c @@ -26,6 +26,7 @@ #include #include +#include #include "gnilnd.h" /* this is useful when needed to debug wire corruption. */ @@ -81,7 +82,6 @@ kgnilnd_schedule_device(kgn_device_t *dev) if (!already_live) { wake_up_all(&dev->gnd_waitq); } - return; } void kgnilnd_schedule_device_timer(unsigned long arg) @@ -151,7 +151,7 @@ kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent) * as scheduled */ int -_kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld) +_kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held) { kgn_device_t *dev = conn->gnc_device; int sched; @@ -184,10 +184,11 @@ _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refhe conn, sched); CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line); - - spin_lock(&dev->gnd_lock); + if (!lock_held) + spin_lock(&dev->gnd_lock); list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns); - spin_unlock(&dev->gnd_lock); + if (!lock_held) + spin_unlock(&dev->gnd_lock); set_mb(conn->gnc_last_sched_ask, jiffies); rc = 1; } else { @@ -197,6 +198,23 @@ _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refhe /* make sure thread(s) going to process conns - but let it make * separate decision from conn schedule */ + if (!lock_held) + kgnilnd_schedule_device(dev); + return rc; +} + +int +_kgnilnd_schedule_delay_conn(kgn_conn_t *conn) +{ + kgn_device_t *dev = conn->gnc_device; + int rc = 0; + spin_lock(&dev->gnd_lock); + if (list_empty(&conn->gnc_delaylist)) { + list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns); + rc = 1; + } + spin_unlock(&dev->gnd_lock); + kgnilnd_schedule_device(dev); return rc; } @@ -237,7 +255,7 @@ kgnilnd_free_tx(kgn_tx_t *tx) /* Only free the buffer if we used it */ if (tx->tx_buffer_copy != NULL) { - vfree(tx->tx_buffer_copy); + kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length); tx->tx_buffer_copy = NULL; CDEBUG(D_MALLOC, "vfreed buffer2\n"); } @@ -302,7 +320,7 @@ kgnilnd_cksum(void *ptr, size_t nob) return sum; } -inline __u16 +__u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob) { @@ -474,7 +492,7 @@ kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_ni LBUG(); } /* only allow NAK on error and truncate to zero */ - LASSERTF(error <= 0, "error %d conn 0x%p, cookie "LPU64"\n", + LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n", error, conn, cookie); tx = kgnilnd_new_tx_msg(nak_type, source); @@ -490,9 +508,9 @@ kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_ni } int -kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *iov, - lnet_kiov_t *kiov, unsigned int offset, unsigned int nob) - +kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, + struct kvec *iov, lnet_kiov_t *kiov, + unsigned int offset, unsigned int nob) { kgn_msg_t *msg = &tx->tx_msg; int i; @@ -507,7 +525,8 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io } else if (kiov != NULL) { if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) { - niov = ((nob + offset + PAGE_SIZE - 1) / PAGE_SIZE); + niov = ((nob + offset + kiov->kiov_offset + PAGE_SIZE - 1) / + PAGE_SIZE); } LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE, @@ -605,7 +624,7 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io int kgnilnd_setup_virt_buffer(kgn_tx_t *tx, - unsigned int niov, struct iovec *iov, + unsigned int niov, struct kvec *iov, unsigned int offset, unsigned int nob) { @@ -745,7 +764,7 @@ error: static inline int kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, + struct kvec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob) { int rc; @@ -773,7 +792,7 @@ kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov, * transfer. */ static void -kgnilnd_parse_lnet_rdma(lnet_msg_t *lntmsg, unsigned int *niov, +kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov, unsigned int *offset, unsigned int *nob, lnet_kiov_t **kiov, int put_len) { @@ -800,7 +819,7 @@ kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len) { unsigned int niov, offset, nob; lnet_kiov_t *kiov; - lnet_msg_t *lntmsg = tx->tx_lntmsg[0]; + struct lnet_msg *lntmsg = tx->tx_lntmsg[0]; int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1); GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) || @@ -852,7 +871,7 @@ kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len) __u16 cksum; unsigned int niov, offset, nob; lnet_kiov_t *kiov; - lnet_msg_t *lntmsg = tx->tx_lntmsg[0]; + struct lnet_msg *lntmsg = tx->tx_lntmsg[0]; int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump; /* we can only match certain requests */ @@ -946,7 +965,7 @@ kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx) if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK || tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) { atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out); - GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to "LPD64"", + GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld", bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out)); } @@ -997,7 +1016,7 @@ kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx) atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out); LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0, "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out)); - GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to "LPD64"", + GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld", bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out)); } @@ -1056,7 +1075,7 @@ kgnilnd_map_buffer(kgn_tx_t *tx) * GART resource, etc starvation handling */ if (rrc != GNI_RC_SUCCESS) { GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d " - "phys %u pp %u, virt %u nob "LPU64"", + "phys %u pp %u, virt %u nob %llu", tx->tx_phys_npages, dev->gnd_id, dev->gnd_map_nphys, dev->gnd_map_physnop, dev->gnd_map_nvirt, dev->gnd_map_virtnob); @@ -1073,7 +1092,7 @@ kgnilnd_map_buffer(kgn_tx_t *tx) NULL, flags, &tx->tx_map_key); if (rrc != GNI_RC_SUCCESS) { GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d " - "phys %u pp %u, virt %u nob "LPU64"", + "phys %u pp %u, virt %u nob %llu", tx->tx_nob, dev->gnd_id, dev->gnd_map_nphys, dev->gnd_map_physnop, dev->gnd_map_nvirt, dev->gnd_map_virtnob); @@ -1131,8 +1150,10 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error) int hold_timeout = 0; /* code below relies on +1 relationship ... */ - CLASSERT(GNILND_BUF_PHYS_MAPPED == (GNILND_BUF_PHYS_UNMAPPED + 1)); - CLASSERT(GNILND_BUF_VIRT_MAPPED == (GNILND_BUF_VIRT_UNMAPPED + 1)); + BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED != + (GNILND_BUF_PHYS_UNMAPPED + 1)); + BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED != + (GNILND_BUF_VIRT_UNMAPPED + 1)); switch (tx->tx_buftype) { default: @@ -1164,6 +1185,7 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error) * verified peer notification - the theory is that * a TX error can be communicated in all other cases */ if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED && + error != -GNILND_NOPURG && kgnilnd_check_purgatory_conn(tx->tx_conn)) { kgnilnd_add_purgatory_tx(tx); @@ -1173,7 +1195,7 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error) hold_timeout = GNILND_TIMEOUT2DEADMAN; GNIDBG_TX(D_NET, tx, - "dev %p delaying MDD release for %dms key "LPX64"."LPX64"", + "dev %p delaying MDD release for %dms key %#llx.%#llx", tx->tx_conn->gnc_device, hold_timeout, tx->tx_map_key.qword1, tx->tx_map_key.qword2); } @@ -1196,9 +1218,9 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error) void kgnilnd_tx_done(kgn_tx_t *tx, int completion) { - lnet_msg_t *lntmsg0, *lntmsg1; + struct lnet_msg *lntmsg0, *lntmsg1; int status0, status1; - lnet_ni_t *ni = NULL; + struct lnet_ni *ni = NULL; kgn_conn_t *conn = tx->tx_conn; LASSERT(!in_interrupt()); @@ -1263,10 +1285,10 @@ kgnilnd_tx_done(kgn_tx_t *tx, int completion) * could free up lnet credits, resulting in a call chain back into * the LND via kgnilnd_send and friends */ - lnet_finalize(ni, lntmsg0, status0); + lnet_finalize(lntmsg0, status0); if (lntmsg1 != NULL) { - lnet_finalize(ni, lntmsg1, status1); + lnet_finalize(lntmsg1, status1); } } @@ -1341,70 +1363,35 @@ search_again: return 0; } -static inline int -kgnilnd_tx_should_retry(kgn_conn_t *conn, kgn_tx_t *tx) +static inline void +kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx) { - int max_retrans = *kgnilnd_tunables.kgn_max_retransmits; int log_retrans; - int log_retrans_level; - - /* I need kgni credits to send this. Replace tx at the head of the - * fmaq and I'll get rescheduled when credits appear */ - tx->tx_state = 0; - tx->tx_retrans++; - conn->gnc_tx_retrans++; - log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0) || - (tx->tx_retrans > (max_retrans / 2))); - log_retrans_level = tx->tx_retrans < (max_retrans / 2) ? D_NET : D_NETERROR; - /* Decision time - either error, warn or just retransmit */ + log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0)); /* we don't care about TX timeout - it could be that the network is slower * or throttled. We'll keep retranmitting - so if the network is so slow * that we fill up our mailbox, we'll keep trying to resend that msg * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating * that he hasn't send us any traffic in return */ - - if (tx->tx_retrans > max_retrans) { - /* this means we are not backing off the retransmits - * in a healthy manner and are likely chewing up the - * CPU cycles quite badly */ - GNIDBG_TOMSG(D_ERROR, &tx->tx_msg, - "SOFTWARE BUG: too many retransmits (%d) for tx id %x " - "conn 0x%p->%s\n", - tx->tx_retrans, tx->tx_id, conn, - libcfs_nid2str(conn->gnc_peer->gnp_nid)); - - /* yes - double errors to help debug this condition */ - GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg, "connection dead. " - "unable to send to %s for %lu secs (%d tries)", - libcfs_nid2str(tx->tx_conn->gnc_peer->gnp_nid), - cfs_duration_sec(jiffies - tx->tx_cred_wait), - tx->tx_retrans); - - kgnilnd_close_conn(conn, -ETIMEDOUT); - - /* caller should terminate */ - RETURN(0); - } else { - /* some reasonable throttling of the debug message */ - if (log_retrans) { - unsigned long now = jiffies; - /* XXX Nic: Mystical TX debug here... */ - GNIDBG_SMSG_CREDS(log_retrans_level, conn); - GNIDBG_TOMSG(log_retrans_level, &tx->tx_msg, - "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus" - " last_msg %uus/%uus last_cq %uus/%uus", - conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), - tx->tx_id, tx->tx_retrans, - jiffies_to_usecs(now - tx->tx_cred_wait), - jiffies_to_usecs(now - conn->gnc_last_tx), - jiffies_to_usecs(now - conn->gnc_last_rx), - jiffies_to_usecs(now - conn->gnc_last_tx_cq), - jiffies_to_usecs(now - conn->gnc_last_rx_cq)); - } - /* caller should retry */ - RETURN(1); + + /* some reasonable throttling of the debug message */ + if (log_retrans) { + unsigned long now = jiffies; + /* XXX Nic: Mystical TX debug here... */ + /* We expect retransmissions so only log when D_NET is enabled */ + GNIDBG_SMSG_CREDS(D_NET, conn); + GNIDBG_TOMSG(D_NET, &tx->tx_msg, + "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus" + " last_msg %uus/%uus last_cq %uus/%uus", + conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), + tx->tx_id, tx->tx_retrans, + jiffies_to_usecs(now - tx->tx_cred_wait), + jiffies_to_usecs(now - conn->gnc_last_tx), + jiffies_to_usecs(now - conn->gnc_last_rx), + jiffies_to_usecs(now - conn->gnc_last_tx_cq), + jiffies_to_usecs(now - conn->gnc_last_rx_cq)); } } @@ -1417,7 +1404,6 @@ kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, { kgn_conn_t *conn = tx->tx_conn; kgn_msg_t *msg = &tx->tx_msg; - int retry_send; gni_return_t rrc; unsigned long newest_last_rx, timeout; unsigned long now; @@ -1471,7 +1457,6 @@ kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, */ msg->gnm_connstamp = conn->gnc_my_connstamp; msg->gnm_payload_len = immediatenob; - kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex); msg->gnm_seq = atomic_read(&conn->gnc_tx_seq); /* always init here - kgn_checksum is a /sys module tunable @@ -1528,9 +1513,11 @@ kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, return 0; case GNI_RC_NOT_DONE: - /* XXX Nic: We need to figure out how to track this - * - there are bound to be good reasons for it, - * but we want to know when it happens */ + /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently + * 1: out of mbox credits + * 2: out of mbox payload credits + * 3: On Aries out of dla credits + */ kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex); kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex); /* We'll handle this error inline - makes the calling logic much more @@ -1541,31 +1528,36 @@ kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, return -EAGAIN; } - retry_send = kgnilnd_tx_should_retry(conn, tx); - if (retry_send) { - /* add to head of list for the state and retries */ - spin_lock(state_lock); - kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0); - spin_unlock(state_lock); - - /* We only reschedule for a certain number of retries, then - * we will wait for the CQ events indicating a release of SMSG - * credits */ - if (tx->tx_retrans < (*kgnilnd_tunables.kgn_max_retransmits/4)) { - kgnilnd_schedule_conn(conn); - return 0; - } else { - /* CQ event coming in signifies either TX completed or - * RX receive. Either of these *could* free up credits - * in the SMSG mbox and we should try sending again */ - GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend", - tx->tx_conn->gnc_cqid); - /* use +ve return code to let upper layers know they - * should stop looping on sends */ - return EAGAIN; - } + /* I need kgni credits to send this. Replace tx at the head of the + * fmaq and I'll get rescheduled when credits appear. Reset the tx_state + * and bump retrans counts since we are requeueing the tx. + */ + tx->tx_state = 0; + tx->tx_retrans++; + conn->gnc_tx_retrans++; + + kgnilnd_tx_log_retrans(conn, tx); + /* add to head of list for the state and retries */ + spin_lock(state_lock); + kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0); + spin_unlock(state_lock); + + /* We only reschedule for a certain number of retries, then + * we will wait for the CQ events indicating a release of SMSG + * credits */ + if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) { + kgnilnd_schedule_conn(conn); + return 0; } else { - return -EAGAIN; + /* CQ event coming in signifies either TX completed or + * RX receive. Either of these *could* free up credits + * in the SMSG mbox and we should try sending again */ + GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend", + tx->tx_conn->gnc_cqid); + kgnilnd_schedule_delay_conn(conn); + /* use +ve return code to let upper layers know they + * should stop looping on sends */ + return EAGAIN; } default: /* handle bad retcode gracefully */ @@ -1586,6 +1578,7 @@ kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob, timestamp = jiffies; kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex); + kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex); /* delay in jiffies - we are really concerned only with things that * result in a schedule() or really holding this off for long times . * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */ @@ -1630,7 +1623,8 @@ kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob rc = 0; } else { atomic_inc(&conn->gnc_device->gnd_fast_try); - rc = kgnilnd_gl_mutex_trylock(&conn->gnc_device->gnd_cq_mutex); + rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex, + &conn->gnc_smsg_mutex); } if (!rc) { rc = -EAGAIN; @@ -1653,7 +1647,7 @@ kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob } /* lets us know if we can push this RDMA through now */ -inline int +static int kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx) { long bytes_left; @@ -1767,7 +1761,7 @@ kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx) } void -kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target) +kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target) { kgn_peer_t *peer; kgn_peer_t *new_peer = NULL; @@ -1802,7 +1796,7 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target) } /* don't create a connection if the peer is marked down */ - if (peer->gnp_down == GNILND_RCA_NODE_DOWN) { + if (peer->gnp_state != GNILND_PEER_UP) { read_unlock(&kgnilnd_data.kgn_peer_conn_lock); rc = -ENETRESET; GOTO(no_peer, rc); @@ -1841,7 +1835,7 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target) kgnilnd_add_peer_locked(target->nid, new_peer, &peer); /* don't create a connection if the peer is not up */ - if (peer->gnp_down != GNILND_RCA_NODE_UP) { + if (peer->gnp_state != GNILND_PEER_UP) { write_unlock(&kgnilnd_data.kgn_peer_conn_lock); rc = -ENETRESET; GOTO(no_peer, rc); @@ -1921,11 +1915,11 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3; if (tx->tx_offset) - kgnilnd_admin_addref(kgnilnd_data.kgn_rev_offset); + atomic_inc(&kgnilnd_data.kgn_rev_offset); if ((nob + tx->tx_offset) & 3) { desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3))); - kgnilnd_admin_addref(kgnilnd_data.kgn_rev_length); + atomic_inc(&kgnilnd_data.kgn_rev_length); } else { desc_nob = (nob + tx->tx_offset); } @@ -1933,7 +1927,7 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, if (tx->tx_buffer_copy == NULL) { /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data * and require at most we allocate a few extra bytes. */ - tx->tx_buffer_copy = vmalloc(desc_nob); + tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob); if (!tx->tx_buffer_copy) { /* allocation of buffer failed nak the rdma */ @@ -1941,11 +1935,12 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, kgnilnd_tx_done(tx, -EFAULT); return 0; } - kgnilnd_admin_addref(kgnilnd_data.kgn_rev_copy_buff); + atomic_inc(&kgnilnd_data.kgn_rev_copy_buff); rc = kgnilnd_mem_register(conn->gnc_device->gnd_handle, (__u64)tx->tx_buffer_copy, desc_nob, NULL, GNI_MEM_READWRITE, &tx->tx_buffer_copy_map_key); if (rc != GNI_RC_SUCCESS) { /* Registration Failed nak rdma and kill the tx. */ - vfree(tx->tx_buffer_copy); + kgnilnd_vfree(tx->tx_buffer_copy, + desc_nob); tx->tx_buffer_copy = NULL; kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid); kgnilnd_tx_done(tx, -EFAULT); @@ -1967,8 +1962,10 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key; tx->tx_rdma_desc.length = desc_nob; tx->tx_nob_rdma = nob; - if (*kgnilnd_tunables.kgn_bte_dlvr_mode) - tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_dlvr_mode; + if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode) + tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode; + if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode) + tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode; /* prep final completion message */ kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid); tx->tx_msg.gnm_u.completion.gncm_cookie = cookie; @@ -1987,7 +1984,7 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, tx, conn, conn->gnc_close_sent); GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode " - "0x%x cookie:"LPX64, + "0x%x cookie:%#llx", type, conn, tx->tx_rdma_desc.dlvr_mode, cookie); /* set CQ dedicated for RDMA */ @@ -2009,7 +2006,7 @@ kgnilnd_rdma(kgn_tx_t *tx, int type, kgnilnd_unmap_buffer(tx, 0); if (tx->tx_buffer_copy != NULL) { - vfree(tx->tx_buffer_copy); + kgnilnd_vfree(tx->tx_buffer_copy, desc_nob); tx->tx_buffer_copy = NULL; } @@ -2074,7 +2071,7 @@ kgnilnd_release_msg(kgn_conn_t *conn) LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc); GNIDBG_SMSG_CREDS(D_NET, conn); - return; + kgnilnd_schedule_conn(conn); } void @@ -2098,20 +2095,18 @@ kgnilnd_consume_rx(kgn_rx_t *rx) kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx); CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n", sizeof(*rx), rx); - - return; } int -kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) +kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg) { - lnet_hdr_t *hdr = &lntmsg->msg_hdr; + struct lnet_hdr *hdr = &lntmsg->msg_hdr; int type = lntmsg->msg_type; - lnet_process_id_t target = lntmsg->msg_target; + struct lnet_process_id target = lntmsg->msg_target; int target_is_router = lntmsg->msg_target_is_router; int routing = lntmsg->msg_routing; unsigned int niov = lntmsg->msg_niov; - struct iovec *iov = lntmsg->msg_iov; + struct kvec *iov = lntmsg->msg_iov; lnet_kiov_t *kiov = lntmsg->msg_kiov; unsigned int offset = lntmsg->msg_offset; unsigned int nob = lntmsg->msg_len; @@ -2275,12 +2270,12 @@ out: } void -kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen) +kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen) { kgn_conn_t *conn = rx->grx_conn; kgn_msg_t *rxmsg = rx->grx_msg; unsigned int niov = lntmsg->msg_niov; - struct iovec *iov = lntmsg->msg_iov; + struct kvec *iov = lntmsg->msg_iov; lnet_kiov_t *kiov = lntmsg->msg_kiov; unsigned int offset = lntmsg->msg_offset; unsigned int nob = lntmsg->msg_len; @@ -2332,11 +2327,11 @@ kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen) kgnilnd_tx_done(tx, rc); kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid); failed_0: - lnet_finalize(ni, lntmsg, rc); + lnet_finalize(lntmsg, rc); } int -kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, +kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg, void **new_private) { kgn_rx_t *rx = private; @@ -2374,7 +2369,7 @@ kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, CERROR("Couldnt find matching peer %p or conn %p / %p\n", peer, conn, found_conn); if (found_conn) { - CERROR("Unexpected connstamp "LPX64"("LPX64" expected)" + CERROR("Unexpected connstamp %#llx(%#llx expected)" " from %s", rxmsg->gnm_connstamp, found_conn->gnc_peer_connstamp, libcfs_nid2str(peer->gnp_nid)); @@ -2427,9 +2422,9 @@ kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, } int -kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, +kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg, int delayed, unsigned int niov, - struct iovec *iov, lnet_kiov_t *kiov, + struct kvec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen, unsigned int rlen) { kgn_rx_t *rx = private; @@ -2458,7 +2453,7 @@ kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, /* someone closed the conn after we copied this out, nuke it */ kgnilnd_consume_rx(rx); - lnet_finalize(ni, lntmsg, conn->gnc_error); + lnet_finalize(lntmsg, conn->gnc_error); RETURN(0); } read_unlock(&kgnilnd_data.kgn_peer_conn_lock); @@ -2532,14 +2527,14 @@ kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, &rxmsg[1], 0, mlen); kgnilnd_consume_rx(rx); - lnet_finalize(ni, lntmsg, 0); + lnet_finalize(lntmsg, 0); RETURN(0); case GNILND_MSG_PUT_REQ: /* LNET wants to truncate or drop transaction, sending NAK */ if (mlen == 0) { kgnilnd_consume_rx(rx); - lnet_finalize(ni, lntmsg, 0); + lnet_finalize(lntmsg, 0); /* only error if lntmsg == NULL, otherwise we are just * short circuiting the rdma process of 0 bytes */ @@ -2598,7 +2593,7 @@ nak_put_req: /* LNET wants to truncate or drop transaction, sending NAK */ if (mlen == 0) { kgnilnd_consume_rx(rx); - lnet_finalize(ni, lntmsg, 0); + lnet_finalize(lntmsg, 0); /* only error if lntmsg == NULL, otherwise we are just * short circuiting the rdma process of 0 bytes */ @@ -2668,7 +2663,7 @@ nak_get_req_rev: /* LNET wants to truncate or drop transaction, sending NAK */ if (mlen == 0) { kgnilnd_consume_rx(rx); - lnet_finalize(ni, lntmsg, 0); + lnet_finalize(lntmsg, 0); /* only error if lntmsg == NULL, otherwise we are just * short circuiting the rdma process of 0 bytes */ @@ -2743,7 +2738,7 @@ kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn) if (time_after_eq(now, newest_last_rx + timeout)) { uint32_t level = D_CONSOLE|D_NETERROR; - if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) { + if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) { level = D_NET; } GNIDBG_CONN(level, conn, @@ -2805,7 +2800,7 @@ kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie, peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_interval); - timeout = cfs_time_seconds(MAX(*kgnilnd_tunables.kgn_timeout, + timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout, GNILND_MIN_TIMEOUT)); conn = kgnilnd_find_conn_locked(peer); @@ -2819,6 +2814,14 @@ kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie, conn->gnc_close_recvd = GNILND_CLOSE_INJECT1; conn->gnc_peer_error = -ETIMEDOUT; } + + if (*kgnilnd_tunables.kgn_to_reconn_disable && + rc == -ETIMEDOUT) { + peer->gnp_state = GNILND_PEER_TIMED_OUT; + CDEBUG(D_WARNING, "%s conn timed out, will " + "reconnect upon request from peer\n", + libcfs_nid2str(conn->gnc_peer->gnp_nid)); + } /* Once we mark closed, any of the scheduler threads could * get it and move through before we hit the fail loc code */ kgnilnd_close_conn_locked(conn, rc); @@ -2862,7 +2865,7 @@ kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie, /* Don't reconnect if we are still trying to clear out old conns. * This prevents us sending traffic on the new mbox before ensuring we are done * with the old one */ - reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) && + reconnect = (peer->gnp_state == GNILND_PEER_UP) && (atomic_read(&peer->gnp_dirty_eps) == 0); /* fast reconnect after a timeout */ @@ -2881,8 +2884,9 @@ kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie, CDEBUG(D_NET, "starting connect to %s\n", libcfs_nid2str(peer->gnp_nid)); - LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, "Peer was idle and we" - "have a write_lock, state issue %d\n", peer->gnp_connecting); + LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, + "Peer was idle and we have a write_lock, state issue %d\n", + peer->gnp_connecting); peer->gnp_connecting = GNILND_PEER_CONNECT; kgnilnd_peer_addref(peer); /* extra ref for connd */ @@ -2972,8 +2976,6 @@ kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie, } } } - - return; } void @@ -2981,11 +2983,8 @@ kgnilnd_reaper_check(int idx) { struct list_head *peers = &kgnilnd_data.kgn_peers[idx]; struct list_head *ctmp, *ctmpN; - struct list_head geriatrics; - struct list_head souls; - - INIT_LIST_HEAD(&geriatrics); - INIT_LIST_HEAD(&souls); + LIST_HEAD(geriatrics); + LIST_HEAD(souls); write_lock(&kgnilnd_data.kgn_peer_conn_lock); @@ -3070,7 +3069,7 @@ kgnilnd_reaper(void *arg) next_check_time); mod_timer(&timer, (long) jiffies + timeout); - /* check flag variables before comitting */ + /* check flag variables before committing */ if (!kgnilnd_data.kgn_shutdown && !kgnilnd_data.kgn_quiesce_trigger) { CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n", @@ -3126,7 +3125,7 @@ int kgnilnd_recv_bte_get(kgn_tx_t *tx) { unsigned niov, offset, nob; lnet_kiov_t *kiov; - lnet_msg_t *lntmsg = tx->tx_lntmsg[0]; + struct lnet_msg *lntmsg = tx->tx_lntmsg[0]; kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma); if (kiov != NULL) { @@ -3190,7 +3189,7 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev) "this is bad, somehow our credits didn't protect us" " from CQ overrun\n"); LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST, - "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc, + "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc, event_data, GNI_CQ_GET_TYPE(event_data)); rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data, @@ -3245,6 +3244,11 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev) spin_unlock(&conn->gnc_list_lock); kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex); + if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) { + event_data = 1LL << 48; + rc = 1; + } + if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) { atomic_inc(&dev->gnd_rdma_ntx); atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes); @@ -3299,7 +3303,7 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev) -EFAULT, rcookie, tx->tx_msg.gnm_srcnid); - kgnilnd_tx_done(tx, -EFAULT); + kgnilnd_tx_done(tx, -GNILND_NOPURG); kgnilnd_close_conn(conn, -ECOMM); } @@ -3319,6 +3323,7 @@ kgnilnd_check_fma_send_cq(kgn_device_t *dev) kgn_conn_t *conn = NULL; int queued_fma, saw_reply, rc; long num_processed = 0; + struct list_head *ctmp, *ctmpN; for (;;) { /* make sure we don't keep looping if we need to reset */ @@ -3338,9 +3343,25 @@ kgnilnd_check_fma_send_cq(kgn_device_t *dev) if (rrc == GNI_RC_NOT_DONE) { CDEBUG(D_INFO, - "SMSG send CQ %d not ready (data "LPX64") " + "SMSG send CQ %d not ready (data %#llx) " "processed %ld\n", dev->gnd_id, event_data, num_processed); + + if (num_processed > 0) { + spin_lock(&dev->gnd_lock); + if (!list_empty(&dev->gnd_delay_conns)) { + list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) { + conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist); + list_del_init(&conn->gnc_delaylist); + CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn); + kgnilnd_schedule_conn_nolock(conn); + } + spin_unlock(&dev->gnd_lock); + kgnilnd_schedule_device(dev); + } else { + spin_unlock(&dev->gnd_lock); + } + } return num_processed; } @@ -3351,7 +3372,7 @@ kgnilnd_check_fma_send_cq(kgn_device_t *dev) "this is bad, somehow our credits didn't " "protect us from CQ overrun\n"); LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG, - "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc, + "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc, event_data, GNI_CQ_GET_TYPE(event_data)); /* if SMSG couldn't handle an error, time for conn to die */ @@ -3365,7 +3386,7 @@ kgnilnd_check_fma_send_cq(kgn_device_t *dev) if (conn == NULL) { /* Conn was destroyed? */ CDEBUG(D_NET, - "SMSG CQID lookup "LPX64" failed\n", + "SMSG CQID lookup %#llx failed\n", GNI_CQ_GET_INST_ID(event_data)); write_unlock(&kgnilnd_data.kgn_peer_conn_lock); continue; @@ -3493,7 +3514,7 @@ kgnilnd_check_fma_rcv_cq(kgn_device_t *dev) kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex); if (rrc == GNI_RC_NOT_DONE) { - CDEBUG(D_INFO, "SMSG RX CQ %d empty data "LPX64" " + CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx " "processed %ld\n", dev->gnd_id, event_data, num_processed); return num_processed; @@ -3510,14 +3531,13 @@ kgnilnd_check_fma_rcv_cq(kgn_device_t *dev) /* set overrun too */ event_data |= (1UL << 63); LASSERTF(GNI_CQ_OVERRUN(event_data), - "(1UL << 63) is no longer the bit to" - "set to indicate CQ_OVERRUN\n"); + "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n"); } } /* sender should get error event too and take care of failed transaction by re-transmitting */ if (rrc == GNI_RC_TRANSACTION_ERROR) { - CDEBUG(D_NET, "SMSG RX CQ error "LPX64"\n", event_data); + CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data); continue; } @@ -3526,12 +3546,12 @@ kgnilnd_check_fma_rcv_cq(kgn_device_t *dev) conn = kgnilnd_cqid2conn_locked( GNI_CQ_GET_INST_ID(event_data)); if (conn == NULL) { - CDEBUG(D_NET, "SMSG RX CQID lookup "LPU64" " - "failed, dropping event "LPX64"\n", + CDEBUG(D_NET, "SMSG RX CQID lookup %llu " + "failed, dropping event %#llx\n", GNI_CQ_GET_INST_ID(event_data), event_data); } else { - CDEBUG(D_NET, "SMSG RX: CQID "LPU64" " + CDEBUG(D_NET, "SMSG RX: CQID %llu " "conn %p->%s\n", GNI_CQ_GET_INST_ID(event_data), conn, conn->gnc_peer ? @@ -3791,7 +3811,7 @@ kgnilnd_process_fmaq(kgn_conn_t *conn) GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0, "tx with zero id", NULL); - CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie "LPX64"\n", + CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n", tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type), tx->tx_msg.gnm_type, tx->tx_id.txe_cookie); @@ -3989,8 +4009,8 @@ _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie) GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) && (tx->tx_id.txe_cookie = cookie)), "conn 0x%p->%s tx_ref_table hosed: wanted " - "txe_cookie "LPX64" txe_idx %d " - "found tx %p cookie "LPX64" txe_idx %d\n", + "txe_cookie %#llx txe_idx %d " + "found tx %p cookie %#llx txe_idx %d\n", conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), cookie, ev_id.txe_idx, tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx); @@ -4004,7 +4024,7 @@ _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie) tx->tx_state, GNILND_TX_WAITING_REPLY, libcfs_nid2str(conn->gnc_peer->gnp_nid)); } else { - CWARN("Unmatched reply %02x, or %02x/"LPX64" from %s\n", + CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n", type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid)); } return tx; @@ -4032,7 +4052,7 @@ kgnilnd_complete_tx(kgn_tx_t *tx, int rc) tx->tx_state &= ~GNILND_TX_WAITING_REPLY; if (rc == -EFAULT) { - CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16"LPF64"u physnop %8d buffertype %#8x MemHandle "LPX64"."LPX64"x\n", + CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n", rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2); if(*kgnilnd_tunables.kgn_efault_lbug) { @@ -4279,7 +4299,7 @@ kgnilnd_check_fma_rx(kgn_conn_t *conn) } if (msg->gnm_connstamp != conn->gnc_peer_connstamp) { - GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp "LPX64"("LPX64 + GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx" " expected) from %s", msg->gnm_connstamp, conn->gnc_peer_connstamp, libcfs_nid2str(peer->gnp_nid)); @@ -4836,7 +4856,7 @@ kgnilnd_process_mapped_tx(kgn_device_t *dev) } else { GNIDBG_TX(log_retrans_level, tx, "transient map failure #%d %d pages/%d bytes phys %u@%u " - "virt %u@"LPU64" " + "virt %u@%llu " "nq_map %d mdd# %d/%d GART %ld", dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob, dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE, @@ -4881,6 +4901,12 @@ kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline) conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist); list_del_init(&conn->gnc_schedlist); + /* + * Since we are processing conn now, we don't need to be on the delaylist any longer. + */ + + if (!list_empty(&conn->gnc_delaylist)) + list_del_init(&conn->gnc_delaylist); spin_unlock(&dev->gnd_lock); conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS); @@ -4907,7 +4933,7 @@ kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline) kgnilnd_conn_decref(conn); up_write(&dev->gnd_conn_sem); } else if (rc != 1) { - kgnilnd_conn_decref(conn); + kgnilnd_conn_decref(conn); } /* clear this so that scheduler thread doesn't spin */ found_work = 0; @@ -4958,7 +4984,7 @@ kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline) kgnilnd_conn_decref(conn); up_write(&dev->gnd_conn_sem); } else if (rc != 1) { - kgnilnd_conn_decref(conn); + kgnilnd_conn_decref(conn); } /* check list again with lock held */