#ifndef EXPORT_SYMTAB
# define EXPORT_SYMTAB
#endif
-
+#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
+#endif
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/kmod.h>
#include <linux/sysctl.h>
-#define DEBUG_SUBSYSTEM S_NAL
+#define DEBUG_SUBSYSTEM S_LND
-#include <linux/kp30.h>
-#include <portals/p30.h>
-#include <portals/lib-p30.h>
-#include <portals/nal.h>
+#include <libcfs/kp30.h>
+#include <lnet/lnet.h>
+#include <lnet/lib-lnet.h>
#include <linux/iba/ibt.h>
#error Invalid GCC version. Must use GCC >= 3.2.3
#endif
-#define IBNAL_SERVICE_NAME "iibnal"
-#define IBNAL_SERVICE_NUMBER 0x11b9a1
-
-#if CONFIG_SMP
+#ifdef 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_USE_FMR 0 /* map on demand v. use whole mem mapping */
+#define KIBLND_DETAILED_DEBUG 0
-#define IBNAL_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
+/* 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
-#define IBNAL_MSG_QUEUE_SIZE 8 /* # messages/RDMAs in-flight */
-#define IBNAL_CREDIT_HIGHWATER 7 /* when to eagerly return credits */
+/* 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_NTX_NBLK 128 /* # reserved tx descs */
-
-#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 + IBNAL_NTX_NBLK)
-#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 (PTL_MD_MAX_IOV + 1)
+#if IBNAL_USE_FMR
+# define IBNAL_MAX_RDMA_FRAGS 1
+# define IBNAL_CONCURRENT_SENDS IBNAL_RX_MSGS
+#else
+# define IBNAL_MAX_RDMA_FRAGS LNET_MAX_IOV
+# define IBNAL_CONCURRENT_SENDS IBNAL_MSG_QUEUE_SIZE
+#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)
-
-
-/* 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
+#define IBNAL_RX_MSGS (IBNAL_MSG_QUEUE_SIZE * 2)
+#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)
-/* 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
+typedef struct
{
- int kib_io_timeout; /* comms timeout (seconds) */
- struct ctl_table_header *kib_sysctl; /* sysctl interface */
+ 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_keepalive; /* keepalive 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 */
+ int *kib_concurrent_sends; /* send work queue sizing */
+#if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
+ cfs_sysctl_table_header_t *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;
__u32 md_lkey;
__u32 md_rkey;
__u64 md_addr;
-} kib_md_t __attribute__((packed));
-
-typedef struct
+} kib_md_t;
+
+typedef struct
{
int kib_init; /* initialisation state */
__u64 kib_incarnation; /* which one am I */
int kib_shutdown; /* shut down? */
atomic_t kib_nthreads; /* # live threads */
+ 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 */
- ptl_nid_t kib_nid; /* my NID */
- 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 */
kib_pages_t *kib_tx_pages; /* premapped tx msg pages */
struct list_head kib_idle_txs; /* idle tx descriptors */
- struct list_head kib_idle_nblk_txs; /* idle reserved tx descriptors */
- wait_queue_head_t kib_idle_tx_waitq; /* block here for tx descriptor */
__u64 kib_next_tx_cookie; /* RDMA completion cookie */
spinlock_t kib_tx_lock; /* serialise */
-
+
IB_HANDLE kib_hca; /* The HCA */
int kib_port; /* port on the device */
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 */
- 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
#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.
* private data and SM service info), is LE on the wire.
*/
-/* also kib_md_t above */
+typedef struct kib_connparams
+{
+ __u32 ibcp_queue_depth;
+ __u32 ibcp_max_msg_size;
+ __u32 ibcp_max_frags;
+} WIRE_ATTR kib_connparams_t;
typedef struct
{
- __u32 rd_key; /* remote key */
- __u32 rd_nob; /* # of bytes */
- __u64 rd_addr; /* remote io vaddr */
-} kib_rdma_desc_t __attribute__((packed));
+ lnet_hdr_t ibim_hdr; /* portals header */
+ char ibim_payload[0]; /* piggy-backed payload */
+} WIRE_ATTR kib_immediate_msg_t;
+#if IBNAL_USE_FMR
typedef struct
{
- ptl_hdr_t ibim_hdr; /* portals header */
- char ibim_payload[0]; /* piggy-backed payload */
-} kib_immediate_msg_t __attribute__((packed));
+ __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;
-/* 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 */
typedef struct
{
- ptl_hdr_t ibrm_hdr; /* portals header */
- __u64 ibrm_cookie; /* opaque completion cookie */
- __u32 ibrm_num_descs; /* how many descs */
- kib_rdma_desc_t ibrm_desc[0]; /* where to suck/blow */
-} kib_rdma_msg_t __attribute__((packed));
+ __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
{
__u64 ibcm_cookie; /* opaque completion cookie */
__u32 ibcm_status; /* completion status */
-} kib_completion_msg_t __attribute__((packed));
+} WIRE_ATTR kib_completion_msg_t;
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;
- } ibm_u __attribute__((packed));
-} kib_msg_t __attribute__((packed));
+ } WIRE_ATTR ibm_u;
+} WIRE_ATTR kib_msg_t;
-#define IBNAL_MSG_MAGIC 0x0be91b91 /* unique magic */
-#define IBNAL_MSG_VERSION 1 /* current protocol version */
+#define IBNAL_MSG_MAGIC LNET_PROTO_IIB_MAGIC /* unique magic */
+#define IBNAL_MSG_VERSION 2 /* current protocol version */
+#define IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD 1 /* previous 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) */
+
+/* connection rejection reasons */
+#define IBNAL_REJECT_CONN_RACE 0 /* You lost connection race */
+#define IBNAL_REJECT_NO_RESOURCES 1 /* Out of memory/conns etc */
+#define IBNAL_REJECT_FATAL 2 /* Anything else */
/***********************************************************************/
{
struct list_head rx_list; /* queue for attention */
struct kib_conn *rx_conn; /* owning conn */
- int rx_rdma; /* RDMA completion posted? */
- int rx_posted; /* posted? */
- __u64 rx_vaddr; /* pre-mapped buffer (hca vaddr) */
+ int rx_nob; /* # bytes received (-1 while posted) */
+ __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 it's memory */
+ IB_WORK_REQ2 rx_wrq;
+ IB_LOCAL_DATASEGMENT rx_gl; /* and its memory */
} kib_rx_t;
typedef struct kib_tx /* transmit message */
{
struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
- int tx_isnblk; /* I'm reserved for non-blocking sends */
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 */
- lib_msg_t *tx_libmsg[2]; /* lib msgs to finalize on completion */
- kib_md_t tx_md; /* RDMA mapping (active/passive) */
- __u64 tx_vaddr; /* pre-mapped buffer (hca vaddr) */
+ __u64 tx_cookie; /* completion cookie */
+ lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
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_REQ2 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_REQ2 *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;
- CM_REJECT_INFO cr_rej_info;
-} 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 */
+ __u32 ibc_version; /* peer protocol version */
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 */
+ int ibc_reserved_credits; /* # credits for ACK/DONE msgs */
+ unsigned long ibc_last_send; /* time of last send */
+ struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
+ struct list_head ibc_tx_queue_nocred; /* sends that don't need a cred */
+ struct list_head ibc_tx_queue_rsrvd; /* sends that need a reserved cred */
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 IBNAL_CONN_DISCONNECTING 4 /* to send disconnect req */
+#define IBNAL_CONN_DISCONNECTED 5 /* 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)
+/* types of connection */
+#define IBNAL_CONN_ACTIVE 0 /* active connect */
+#define IBNAL_CONN_PASSIVE 1 /* passive connect */
+#define IBNAL_CONN_WAITING 2 /* waiting for connect */
typedef struct kib_peer
{
struct list_head ibp_list; /* stash on global peer list */
struct list_head ibp_connd_list; /* schedule on kib_connd_peers */
- ptl_nid_t ibp_nid; /* who's on the other end(s) */
+ lnet_nid_t ibp_nid; /* who's on the other end(s) */
atomic_t ibp_refcount; /* # users */
int ibp_persistence; /* "known" peer refs */
+ int ibp_version; /* protocol version */
struct list_head ibp_conns; /* all active connections */
struct list_head ibp_tx_queue; /* msgs waiting for a conn */
- int ibp_connecting; /* connecting+accepting */
+ int ibp_connecting; /* active connects in progress */
+ int ibp_accepting; /* passive connects in progress */
+ int ibp_passivewait; /* waiting for peer to connect */
+ unsigned long ibp_passivewait_deadline; /* when passive wait must complete */
unsigned long ibp_reconnect_time; /* when reconnect may be attempted */
unsigned long ibp_reconnect_interval; /* exponential backoff */
+ int ibp_error; /* errno on closing this peer */
+ cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
} kib_peer_t;
-extern lib_nal_t kibnal_lib;
extern kib_data_t kibnal_data;
extern kib_tunables_t kibnal_tunables;
/******************************************************************************/
-/* Infinicon IBT interface wrappers */
-#define IIBT_IF (kibnal_data.kib_interfaces.ver2)
-
-static inline FSTATUS
-iibt_get_hca_guids(uint32 *hca_count, EUI64 *hca_guid_list)
-{
- return IIBT_IF.GetCaGuids(hca_count, hca_guid_list);
-}
-
-static inline FSTATUS
-iibt_open_hca(EUI64 hca_guid,
- IB_COMPLETION_CALLBACK completion_callback,
- IB_ASYNC_EVENT_CALLBACK async_event_callback,
- void *arg,
- IB_HANDLE *handle)
-{
- return IIBT_IF.Vpi.OpenCA(hca_guid, completion_callback,
- async_event_callback, arg, handle);
-}
-
-static inline FSTATUS
-iibt_query_hca(IB_HANDLE hca_handle, IB_CA_ATTRIBUTES *attrs, void **argp)
-{
- return IIBT_IF.Vpi.QueryCA(hca_handle, attrs, argp);
-}
-
-static inline FSTATUS
-iibt_close_hca(IB_HANDLE hca_handle)
-{
- return IIBT_IF.Vpi.CloseCA(hca_handle);
-}
-
-static inline FSTATUS
-iibt_pd_allocate(IB_HANDLE hca_handle, __u32 max_avs, IB_HANDLE *pd_handle)
-{
- return IIBT_IF.Vpi.AllocatePD(hca_handle, max_avs, pd_handle);
-}
-
-static inline FSTATUS
-iibt_pd_free(IB_HANDLE pd_handle)
-{
- return IIBT_IF.Vpi.FreePD(pd_handle);
-}
-
-static inline FSTATUS
-iibt_register_physical_memory(IB_HANDLE hca_handle,
- IB_VIRT_ADDR requested_io_va,
- void *phys_buffers, uint64 nphys_buffers,
- uint32 io_va_offset, IB_HANDLE pd_handle,
- IB_ACCESS_CONTROL access,
- IB_HANDLE *mem_handle,
- IB_VIRT_ADDR *actual_io_va,
- IB_L_KEY *lkey, IB_R_KEY *rkey)
-{
- return IIBT_IF.Vpi.RegisterPhysMemRegion(hca_handle, requested_io_va,
- phys_buffers, nphys_buffers,
- io_va_offset, pd_handle,
- access,
- mem_handle, actual_io_va,
- lkey, rkey);
-}
-
-static inline FSTATUS
-iibt_register_contig_physical_memory(IB_HANDLE hca_handle,
- IB_VIRT_ADDR requested_io_va,
- IB_MR_PHYS_BUFFER *phys_buffers,
- uint64 nphys_buffers,
- uint32 io_va_offset, IB_HANDLE pd_handle,
- IB_ACCESS_CONTROL access,
- IB_HANDLE *mem_handle,
- IB_VIRT_ADDR *actual_io_va,
- IB_L_KEY *lkey, IB_R_KEY *rkey)
-{
- return IIBT_IF.Vpi.RegisterContigPhysMemRegion(hca_handle,
- requested_io_va,
- phys_buffers,
- nphys_buffers,
- io_va_offset, pd_handle,
- access,
- mem_handle, actual_io_va,
- lkey, rkey);
-}
-
-static inline FSTATUS
-iibt_register_memory(IB_HANDLE hca_handle,
- void *virt_addr, unsigned int length,
- IB_HANDLE pd_handle,
- IB_ACCESS_CONTROL access,
- IB_HANDLE *mem_handle,
- IB_L_KEY *lkey, IB_R_KEY *rkey)
-{
- return IIBT_IF.Vpi.RegisterMemRegion(hca_handle,
- virt_addr, length,
- pd_handle,
- access,
- mem_handle,
- lkey, rkey);
-}
-
-static inline FSTATUS
-iibt_deregister_memory(IB_HANDLE mem_handle)
-{
- return IIBT_IF.Vpi.DeregisterMemRegion(mem_handle);
-}
-
-static inline FSTATUS
-iibt_cq_create(IB_HANDLE hca_handle, uint32 requested_size,
- void *arg, IB_HANDLE *cq_handle, uint32 *actual_size)
-{
- return IIBT_IF.Vpi.CreateCQ(hca_handle, requested_size,
- arg, cq_handle, actual_size);
-}
-
-static inline FSTATUS
-iibt_cq_poll(IB_HANDLE cq_handle, IB_WORK_COMPLETION *wc)
-{
- return IIBT_IF.Vpi.PollCQ(cq_handle, wc);
-}
-
-static inline FSTATUS
-iibt_cq_rearm(IB_HANDLE cq_handle, IB_CQ_EVENT_SELECT select)
-{
- return IIBT_IF.Vpi.RearmCQ(cq_handle, select);
-}
-
-static inline FSTATUS
-iibt_cq_destroy(IB_HANDLE cq_handle)
-{
- return IIBT_IF.Vpi.DestroyCQ(cq_handle);
-}
-
-static inline FSTATUS
-iibt_qp_create(IB_HANDLE hca_handle, IB_QP_ATTRIBUTES_CREATE *create_attr,
- void *arg, IB_HANDLE *cq_handle,
- IB_QP_ATTRIBUTES_QUERY *query_attr)
-{
- return IIBT_IF.Vpi.CreateQP(hca_handle, create_attr, arg, cq_handle,
- query_attr);
-}
-
-static inline FSTATUS
-iibt_qp_query(IB_HANDLE qp_handle, IB_QP_ATTRIBUTES_QUERY *query_attr,
- void **arg_ptr)
-{
- return IIBT_IF.Vpi.QueryQP(qp_handle, query_attr, arg_ptr);
-}
-
-static inline FSTATUS
-iibt_qp_modify(IB_HANDLE qp_handle, IB_QP_ATTRIBUTES_MODIFY *modify_attr,
- IB_QP_ATTRIBUTES_QUERY *query_attr)
-{
- return IIBT_IF.Vpi.ModifyQP(qp_handle, modify_attr, query_attr);
-}
-
-static inline FSTATUS
-iibt_qp_destroy(IB_HANDLE qp_handle)
-{
- return IIBT_IF.Vpi.DestroyQP(qp_handle);
-}
-
-static inline FSTATUS
-iibt_postrecv(IB_HANDLE qp_handle, IB_WORK_REQ *work_req)
-{
- return IIBT_IF.Vpi.PostRecv(qp_handle, work_req);
-}
-
-static inline FSTATUS
-iibt_postsend(IB_HANDLE qp_handle, IB_WORK_REQ *work_req)
-{
- return IIBT_IF.Vpi.PostSend(qp_handle, work_req);
-}
-
-static inline FSTATUS
-iibt_sd_register(IB_HANDLE *sd_handle, CLIENT_CONTROL_PARAMETERS *p)
-{
- return IIBT_IF.Sdi.Register(sd_handle, p);
-}
-
-static inline FSTATUS
-iibt_sd_deregister(IB_HANDLE sd_handle)
-{
- return IIBT_IF.Sdi.Deregister(sd_handle);
-}
-
-static inline FSTATUS
-iibt_sd_port_fabric_operation(IB_HANDLE sd_handle, EUI64 port_guid,
- FABRIC_OPERATION_DATA *fod,
- PFABRIC_OPERATION_CALLBACK callback,
- COMMAND_CONTROL_PARAMETERS *p, void *arg)
-{
- return IIBT_IF.Sdi.PortFabricOperation(sd_handle, port_guid,
- fod, callback, p, arg);
-}
-
-static inline FSTATUS
-iibt_sd_query_port_fabric_information(IB_HANDLE sd_handle, EUI64 port_guid,
- QUERY *qry,
- PQUERY_CALLBACK callback,
- COMMAND_CONTROL_PARAMETERS *p, void *arg)
-{
- return IIBT_IF.Sdi.QueryPortFabricInformation(sd_handle, port_guid,
- qry, callback, p, arg);
-}
-
-static inline IB_HANDLE
-iibt_cm_create_cep(CM_CEP_TYPE type)
-{
- return IIBT_IF.Cmi.CmCreateCEP(type);
-}
-
-static inline FSTATUS
-iibt_cm_modify_cep(IB_HANDLE cep, uint32 attr, char* value, uint32 len,
- uint32 offset)
-{
- return IIBT_IF.Cmi.CmModifyCEP(cep, attr, value, len, offset);
-}
-
-static inline FSTATUS
-iibt_cm_destroy_cep(IB_HANDLE cep_handle)
-{
- return IIBT_IF.Cmi.CmDestroyCEP(cep_handle);
-}
-
-static inline FSTATUS
-iibt_cm_listen(IB_HANDLE cep, CM_LISTEN_INFO *info,
- PFN_CM_CALLBACK callback, void *arg)
-{
- return IIBT_IF.Cmi.CmListen(cep, info, callback, arg);
-}
-
-static inline FSTATUS
-iibt_cm_cancel(IB_HANDLE cep)
-{
- return IIBT_IF.Cmi.CmCancel(cep);
-}
-
-static inline FSTATUS
-iibt_cm_accept(IB_HANDLE cep,
- CM_CONN_INFO *send_info, CM_CONN_INFO *recv_info,
- PFN_CM_CALLBACK callback, void *arg,
- IB_HANDLE *new_cep)
-{
- return IIBT_IF.Cmi.CmAccept(cep,
- send_info, recv_info,
- callback, arg, new_cep);
-}
-
-static inline FSTATUS
-iibt_cm_reject(IB_HANDLE cep, CM_REJECT_INFO *rej)
-{
- return IIBT_IF.Cmi.CmReject(cep, rej);
-}
-
-static inline FSTATUS
-iibt_cm_disconnect(IB_HANDLE cep, CM_DREQUEST_INFO *req,
- CM_DREPLY_INFO *reply)
-{
- return IIBT_IF.Cmi.CmDisconnect(cep, req, reply);
-}
-
-static inline FSTATUS
-iibt_cm_connect (IB_HANDLE cep, CM_REQUEST_INFO *req,
- PFN_CM_CALLBACK callback, void *arg)
-{
- return IIBT_IF.Cmi.CmConnect (cep, req, callback, arg);
-}
-
-static inline int wrq_signals_completion(IB_WORK_REQ *wrq)
-{
- return wrq->Req.SendRC.Options.s.SignaledCompletion == 1;
-}
-
-
-/******************************************************************************/
/* 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] -> "LPX64" (%d)\n", \
- peer, 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] -> "LPX64" (%d)\n", \
- peer, peer->ibp_nid, atomic_read (&peer->ibp_refcount)); \
- if (atomic_dec_and_test (&peer->ibp_refcount)) { \
- CDEBUG (D_NET, "destroying peer "LPX64" %p\n", \
- 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)
/******************************************************************************/
static inline struct list_head *
-kibnal_nid2peerlist (ptl_nid_t nid)
+kibnal_nid2peerlist (lnet_nid_t nid)
{
unsigned int hash = ((unsigned int)nid) % kibnal_data.kib_peer_hash_size;
-
+
return (&kibnal_data.kib_peers [hash]);
}
return (!list_empty(&peer->ibp_list));
}
+static inline int
+kibnal_peer_connecting(kib_peer_t *peer)
+{
+ /* Am I expecting a connection to materialise? */
+ return (peer->ibp_connecting != 0 ||
+ peer->ibp_accepting != 0 ||
+ peer->ibp_passivewait);
+}
+
static inline void
kibnal_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
{
- /* CAVEAT EMPTOR: tx takes caller's ref on conn */
+ struct list_head *q;
+
+ LASSERT (tx->tx_nwrq > 0); /* work items set up */
+ LASSERT (!tx->tx_queued); /* not queued for sending already */
+
+ tx->tx_queued = 1;
+ tx->tx_deadline = jiffies + (*kibnal_tunables.kib_timeout * HZ);
+
+ if (tx->tx_conn == NULL) {
+ kibnal_conn_addref(conn);
+ tx->tx_conn = conn;
+ LASSERT (tx->tx_msg->ibm_type != IBNAL_MSG_PUT_DONE);
+ } else {
+ LASSERT (tx->tx_conn == conn);
+ LASSERT (tx->tx_msg->ibm_type == IBNAL_MSG_PUT_DONE);
+ }
- LASSERT (tx->tx_nsp > 0); /* work items set up */
- LASSERT (tx->tx_conn == NULL); /* only set here */
+ if (conn->ibc_version == IBNAL_MSG_VERSION_RDMAREPLYNOTRSRVD) {
+ /* All messages have simple credit control */
+ q = &conn->ibc_tx_queue;
+ } else {
+ LASSERT (conn->ibc_version == IBNAL_MSG_VERSION);
+
+ switch (tx->tx_msg->ibm_type) {
+ case IBNAL_MSG_PUT_REQ:
+ case IBNAL_MSG_GET_REQ:
+ /* RDMA request: reserve a buffer for the RDMA reply
+ * before sending */
+ q = &conn->ibc_tx_queue_rsrvd;
+ break;
+
+ case IBNAL_MSG_PUT_NAK:
+ case IBNAL_MSG_PUT_ACK:
+ case IBNAL_MSG_PUT_DONE:
+ case IBNAL_MSG_GET_DONE:
+ /* RDMA reply/completion: no credits; peer has reserved
+ * a reply buffer */
+ q = &conn->ibc_tx_queue_nocred;
+ break;
+
+ case IBNAL_MSG_NOOP:
+ case IBNAL_MSG_IMMEDIATE:
+ /* Otherwise: consume a credit before sending */
+ q = &conn->ibc_tx_queue;
+ break;
+
+ default:
+ LBUG();
+ q = NULL;
+ }
+ }
+
+ list_add_tail(&tx->tx_list, q);
+}
- tx->tx_conn = conn;
- tx->tx_deadline = jiffies + kibnal_tunables.kib_io_timeout * HZ;
- list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
+static inline int
+kibnal_send_keepalive(kib_conn_t *conn)
+{
+ return (*kibnal_tunables.kib_keepalive > 0) &&
+ time_after(jiffies, conn->ibc_last_send +
+ *kibnal_tunables.kib_keepalive*HZ);
}
#define KIBNAL_SERVICE_KEY_MASK (IB_SERVICE_RECORD_COMP_SERVICENAME | \
return (__u64 *)srv->ServiceData8;
}
-
static inline void
-kibnal_set_service_keys(IB_SERVICE_RECORD *srv, ptl_nid_t nid)
+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);
}
-#if 0
-static inline void
-kibnal_show_rdma_attr (kib_conn_t *conn)
-{
- struct ib_qp_attribute qp_attr;
- int rc;
-
- memset (&qp_attr, 0, sizeof(qp_attr));
- rc = ib_qp_query(conn->ibc_qp, &qp_attr);
- if (rc != 0) {
- CERROR ("Can't get qp attrs: %d\n", rc);
- return;
- }
-
- CWARN ("RDMA CAPABILITY: write %s read %s\n",
- (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ?
- (qp_attr.enable_rdma_write ? "enabled" : "disabled") : "invalid",
- (qp_attr.valid_fields & TS_IB_QP_ATTRIBUTE_RDMA_ATOMIC_ENABLE) ?
- (qp_attr.enable_rdma_read ? "enabled" : "disabled") : "invalid");
-}
-#endif
-
-#if CONFIG_X86
-static inline __u64
-kibnal_page2phys (struct page *p)
-{
- __u64 page_number = p - mem_map;
-
- return (page_number << PAGE_SHIFT);
-}
-#else
-# error "no page->phys"
-#endif
+/* 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). */
-/* 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. */
+#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 & 1) != 0;
+ return (wreqid & IBNAL_WID_MASK);
}
+static inline void
+kibnal_set_conn_state (kib_conn_t *conn, int state)
+{
+ CDEBUG(D_NET,"%p state %d\n", conn, state);
+ conn->ibc_state = state;
+ mb();
+}
+
+#if IBNAL_USE_FMR
+
static inline int
-kibnal_whole_mem(void)
+kibnal_rd_size (kib_rdma_desc_t *rd)
{
- return kibnal_data.kib_md.md_handle != NULL;
+ return rd->rd_nob;
}
-extern kib_peer_t *kibnal_create_peer (ptl_nid_t nid);
-extern void kibnal_destroy_peer (kib_peer_t *peer);
-extern int kibnal_del_peer (ptl_nid_t nid, int single_share);
-extern kib_peer_t *kibnal_find_peer_locked (ptl_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);
+#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
-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, lib_msg_t *libmsg,
- unsigned int niov,
- struct iovec *iov, ptl_kiov_t *kiov,
- size_t offset, size_t nob);
-
-void kibnal_ca_async_callback (void *ca_arg, IB_EVENT_RECORD *ev);
-void kibnal_ca_callback (void *ca_arg, void *cq_arg);
+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_eager_recv (lnet_ni_t *ni, void *private,
+ lnet_msg_t *lntmsg, void **new_private);
+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, __u32 version, int credits,
+ lnet_nid_t dstnid, __u64 dststamp, __u64 seq);
+void kibnal_pack_connmsg(kib_msg_t *msg, __u32 version, int nob, int type,
+ lnet_nid_t dstnid, __u64 dststamp);
+int kibnal_unpack_msg(kib_msg_t *msg, __u32 expected_version, 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_peer_alive (kib_peer_t *peer);
+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, int proto_version);
+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);
+void kibnal_txlist_done (struct list_head *txlist, int status);
+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);