1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.h
38 * Author: Eric Barton <eric@bartonsoftware.com>
42 # define EXPORT_SYMTAB
44 #ifndef AUTOCONF_INCLUDED
45 #include <linux/config.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/smp_lock.h>
54 #include <linux/unistd.h>
55 #include <linux/uio.h>
57 #include <asm/system.h>
58 #include <asm/uaccess.h>
61 #include <linux/init.h>
63 #include <linux/file.h>
64 #include <linux/stat.h>
65 #include <linux/list.h>
66 #include <linux/kmod.h>
67 #include <linux/sysctl.h>
68 #include <linux/random.h>
69 #include <linux/pci.h>
74 #define DEBUG_SUBSYSTEM S_LND
76 #include <libcfs/libcfs.h>
77 #include <lnet/lnet.h>
78 #include <lnet/lib-lnet.h>
79 #include <lnet/lnet-sysctl.h>
85 #include <rdma/rdma_cm.h>
86 #include <rdma/ib_cm.h>
87 #include <rdma/ib_verbs.h>
88 #include <rdma/ib_fmr_pool.h>
90 /* tunables fixed at compile time */
92 # define IBLND_N_SCHED num_online_cpus() /* # schedulers */
94 # define IBLND_N_SCHED 1 /* # schedulers */
97 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
98 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
102 unsigned int *kib_service; /* IB service number */
103 int *kib_min_reconnect_interval; /* first failed connection retry... */
104 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
105 int *kib_cksum; /* checksum kib_msg_t? */
106 int *kib_timeout; /* comms timeout (seconds) */
107 int *kib_keepalive; /* keepalive timeout (seconds) */
108 int *kib_ntx; /* # tx descs */
109 int *kib_credits; /* # concurrent sends */
110 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
111 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
112 int *kib_peercredits_hiw; /* # when eagerly to return credits */
113 int *kib_peertimeout; /* seconds to consider peer dead */
114 char **kib_default_ipif; /* default IPoIB interface */
115 int *kib_retry_count;
116 int *kib_rnr_retry_count;
117 int *kib_concurrent_sends; /* send work queue sizing */
118 int *kib_ib_mtu; /* IB MTU */
119 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
120 * than this value, 0 disable map-on-demand */
121 int *kib_pmr_pool_size; /* # physical MR in pool */
122 int *kib_fmr_pool_size; /* # FMRs in pool */
123 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
124 int *kib_fmr_cache; /* enable FMR pool cache? */
125 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
126 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
130 extern kib_tunables_t kiblnd_tunables;
132 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
133 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
135 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
136 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
138 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
139 IBLND_MSG_QUEUE_SIZE_V1 : \
140 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
141 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
142 IBLND_CREDIT_HIGHWATER_V1 : \
143 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
146 kiblnd_concurrent_sends_v1(void)
148 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
149 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
151 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
152 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
154 return *kiblnd_tunables.kib_concurrent_sends;
157 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
158 kiblnd_concurrent_sends_v1() : \
159 *kiblnd_tunables.kib_concurrent_sends)
160 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
161 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
162 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
164 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
165 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
166 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
167 *kiblnd_tunables.kib_map_on_demand : \
168 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
169 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
170 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
172 /************************/
173 /* derived constants... */
175 /* TX messages (shared by all connections) */
176 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
177 #define IBLND_TX_MSG_BYTES() (IBLND_TX_MSGS() * IBLND_MSG_SIZE)
178 #define IBLND_TX_MSG_PAGES() ((IBLND_TX_MSG_BYTES() + PAGE_SIZE - 1) / PAGE_SIZE)
180 /* RX messages (per connection) */
181 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
182 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
183 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
185 /* WRs and CQEs (per connection) */
186 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
187 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
188 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
192 struct ib_device *ibp_device; /* device for mapping */
193 int ibp_npages; /* # pages */
194 struct page *ibp_pages[0];
198 spinlock_t ibmp_lock; /* serialize */
199 int ibmp_allocated; /* MR in use */
200 struct list_head ibmp_free_list; /* pre-allocated MR */
201 } kib_phys_mr_pool_t;
204 struct list_head ibpm_link; /* link node */
205 struct ib_mr *ibpm_mr; /* MR */
206 __u64 ibpm_iova; /* Virtual I/O address */
207 int ibpm_refcount; /* reference count */
212 struct list_head ibd_list; /* chain on kib_devs */
213 __u32 ibd_ifip; /* IPoIB interface IP */
214 char ibd_ifname[32]; /* IPoIB interface name */
215 int ibd_nnets; /* # nets extant */
217 struct rdma_cm_id *ibd_cmid; /* IB listener (bound to 1 device) */
218 struct ib_pd *ibd_pd; /* PD for the device */
219 int ibd_page_shift; /* page shift of current HCA */
220 int ibd_page_size; /* page size of current HCA */
221 __u64 ibd_page_mask; /* page mask of current HCA */
222 int ibd_mr_shift; /* bits shift of max MR size */
223 __u64 ibd_mr_size; /* size of MR */
225 int ibd_nmrs; /* # of global MRs */
226 struct ib_mr **ibd_mrs; /* MR for non RDMA I/O */
231 __u64 ibn_incarnation; /* my epoch */
232 int ibn_init; /* initialisation state */
233 int ibn_shutdown; /* shutting down? */
235 atomic_t ibn_npeers; /* # peers extant */
236 atomic_t ibn_nconns; /* # connections extant */
238 __u64 ibn_tx_next_cookie; /* RDMA completion cookie */
239 struct kib_tx *ibn_tx_descs; /* all the tx descriptors */
240 kib_pages_t *ibn_tx_pages; /* premapped tx msg pages */
241 struct list_head ibn_idle_txs; /* idle tx descriptors */
242 spinlock_t ibn_tx_lock; /* serialise */
244 struct ib_fmr_pool *ibn_fmrpool; /* FMR pool for RDMA I/O */
245 kib_phys_mr_pool_t *ibn_pmrpool; /* Physical MR pool for RDMA I/O */
247 kib_dev_t *ibn_dev; /* underlying IB device */
252 int kib_init; /* initialisation state */
253 int kib_shutdown; /* shut down? */
254 struct list_head kib_devs; /* IB devices extant */
255 atomic_t kib_nthreads; /* # live threads */
256 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
258 struct list_head *kib_peers; /* hash table of all my known peers */
259 int kib_peer_hash_size; /* size of kib_peers */
261 void *kib_connd; /* the connd task (serialisation assertions) */
262 struct list_head kib_connd_conns; /* connections to setup/teardown */
263 struct list_head kib_connd_zombies; /* connections with zero refcount */
264 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
265 spinlock_t kib_connd_lock; /* serialise */
267 wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */
268 struct list_head kib_sched_conns; /* conns to check for rx completions */
269 spinlock_t kib_sched_lock; /* serialise */
271 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
274 #define IBLND_INIT_NOTHING 0
275 #define IBLND_INIT_DATA 1
276 #define IBLND_INIT_ALL 2
278 /************************************************************************
279 * IB Wire message format.
280 * These are sent in sender's byte order (i.e. receiver flips).
283 typedef struct kib_connparams
285 __u16 ibcp_queue_depth;
286 __u16 ibcp_max_frags;
287 __u32 ibcp_max_msg_size;
288 } WIRE_ATTR kib_connparams_t;
292 lnet_hdr_t ibim_hdr; /* portals header */
293 char ibim_payload[0]; /* piggy-backed payload */
294 } WIRE_ATTR kib_immediate_msg_t;
298 __u32 rf_nob; /* # bytes this frag */
299 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
300 } WIRE_ATTR kib_rdma_frag_t;
304 __u32 rd_key; /* local/remote key */
305 __u32 rd_nfrags; /* # fragments */
306 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
307 } WIRE_ATTR kib_rdma_desc_t;
311 lnet_hdr_t ibprm_hdr; /* portals header */
312 __u64 ibprm_cookie; /* opaque completion cookie */
313 } WIRE_ATTR kib_putreq_msg_t;
317 __u64 ibpam_src_cookie; /* reflected completion cookie */
318 __u64 ibpam_dst_cookie; /* opaque completion cookie */
319 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
320 } WIRE_ATTR kib_putack_msg_t;
324 lnet_hdr_t ibgm_hdr; /* portals header */
325 __u64 ibgm_cookie; /* opaque completion cookie */
326 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
327 } WIRE_ATTR kib_get_msg_t;
331 __u64 ibcm_cookie; /* opaque completion cookie */
332 __s32 ibcm_status; /* < 0 failure: >= 0 length */
333 } WIRE_ATTR kib_completion_msg_t;
337 /* First 2 fields fixed FOR ALL TIME */
338 __u32 ibm_magic; /* I'm an openibnal message */
339 __u16 ibm_version; /* this is my version number */
341 __u8 ibm_type; /* msg type */
342 __u8 ibm_credits; /* returned credits */
343 __u32 ibm_nob; /* # bytes in whole message */
344 __u32 ibm_cksum; /* checksum (0 == no checksum) */
345 __u64 ibm_srcnid; /* sender's NID */
346 __u64 ibm_srcstamp; /* sender's incarnation */
347 __u64 ibm_dstnid; /* destination's NID */
348 __u64 ibm_dststamp; /* destination's incarnation */
351 kib_connparams_t connparams;
352 kib_immediate_msg_t immediate;
353 kib_putreq_msg_t putreq;
354 kib_putack_msg_t putack;
356 kib_completion_msg_t completion;
358 } WIRE_ATTR kib_msg_t;
360 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
362 #define IBLND_MSG_VERSION_1 0x11
363 #define IBLND_MSG_VERSION_2 0x12
364 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
366 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
367 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
368 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
369 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
370 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
371 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
372 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
373 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
374 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
375 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
378 __u32 ibr_magic; /* sender's magic */
379 __u16 ibr_version; /* sender's version */
380 __u8 ibr_why; /* reject reason */
381 __u8 ibr_padding; /* padding */
382 __u64 ibr_incarnation; /* incarnation of peer */
383 kib_connparams_t ibr_cp; /* connection parameters */
384 } WIRE_ATTR kib_rej_t;
386 /* connection rejection reasons */
387 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
388 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
389 #define IBLND_REJECT_FATAL 3 /* Anything else */
391 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
392 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
394 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
395 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
397 /***********************************************************************/
399 typedef struct kib_rx /* receive message */
401 struct list_head rx_list; /* queue for attention */
402 struct kib_conn *rx_conn; /* owning conn */
403 int rx_nob; /* # bytes received (-1 while posted) */
404 enum ib_wc_status rx_status; /* completion status */
405 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
406 __u64 rx_msgaddr; /* message buffer (I/O addr) */
407 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
408 struct ib_recv_wr rx_wrq; /* receive work item... */
409 struct ib_sge rx_sge; /* ...and its memory */
412 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
413 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
414 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
415 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
417 typedef struct kib_tx /* transmit message */
419 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
420 struct kib_conn *tx_conn; /* owning conn */
421 int tx_sending; /* # tx callbacks outstanding */
422 int tx_queued; /* queued for sending */
423 int tx_waiting; /* waiting for peer */
424 int tx_status; /* LNET completion status */
425 unsigned long tx_deadline; /* completion deadline */
426 __u64 tx_cookie; /* completion cookie */
427 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
428 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
429 __u64 tx_msgaddr; /* message buffer (I/O addr) */
430 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
431 int tx_nwrq; /* # send work items */
432 struct ib_send_wr *tx_wrq; /* send work items... */
433 struct ib_sge *tx_sge; /* ...and their memory */
434 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
435 int tx_nfrags; /* # entries in... */
436 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
437 struct ib_phys_buf *tx_ipb; /* physical buffer (for iWARP) */
438 __u64 *tx_pages; /* rdma phys page addrs */
440 kib_phys_mr_t *pmr; /* MR for physical buffer */
441 struct ib_pool_fmr *fmr; /* rdma mapping (mapped if != NULL) */
443 int tx_dmadir; /* dma direction */
446 typedef struct kib_connvars
448 /* connection-in-progress variables */
452 typedef struct kib_conn
454 struct kib_peer *ibc_peer; /* owning peer */
455 struct list_head ibc_list; /* stash on peer's conn list */
456 struct list_head ibc_sched_list; /* schedule for attention */
457 __u16 ibc_version; /* version of connection */
458 __u64 ibc_incarnation; /* which instance of the peer */
459 atomic_t ibc_refcount; /* # users */
460 int ibc_state; /* what's happening */
461 int ibc_nsends_posted; /* # uncompleted sends */
462 int ibc_noops_posted; /* # uncompleted NOOPs */
463 int ibc_credits; /* # credits I have */
464 int ibc_outstanding_credits; /* # credits to return */
465 int ibc_reserved_credits;/* # ACK/DONE msg credits */
466 int ibc_comms_error; /* set on comms error */
467 int ibc_nrx:16; /* receive buffers owned */
468 int ibc_scheduled:1; /* scheduled for attention */
469 int ibc_ready:1; /* CQ callback fired */
470 unsigned long ibc_last_send; /* time of last send */
471 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
472 struct list_head ibc_tx_queue; /* sends that need a credit */
473 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
474 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
475 struct list_head ibc_active_txs; /* active tx awaiting completion */
476 spinlock_t ibc_lock; /* serialise */
477 kib_rx_t *ibc_rxs; /* the rx descs */
478 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
480 struct rdma_cm_id *ibc_cmid; /* CM id */
481 struct ib_cq *ibc_cq; /* completion queue */
483 kib_connvars_t *ibc_connvars; /* in-progress connection state */
486 #define IBLND_CONN_INIT 0 /* being intialised */
487 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
488 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
489 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
490 #define IBLND_CONN_CLOSING 4 /* being closed */
491 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
493 typedef struct kib_peer
495 struct list_head ibp_list; /* stash on global peer list */
496 lnet_nid_t ibp_nid; /* who's on the other end(s) */
497 lnet_ni_t *ibp_ni; /* LNet interface */
498 atomic_t ibp_refcount; /* # users */
499 struct list_head ibp_conns; /* all active connections */
500 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
501 __u16 ibp_version; /* version of peer */
502 __u64 ibp_incarnation; /* incarnation of peer */
503 int ibp_connecting; /* current active connection attempts */
504 int ibp_accepting; /* current passive connection attempts */
505 int ibp_error; /* errno on closing this peer */
506 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
509 extern kib_data_t kiblnd_data;
511 #define kiblnd_conn_addref(conn) \
513 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
514 (conn), atomic_read(&(conn)->ibc_refcount)); \
515 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
516 atomic_inc(&(conn)->ibc_refcount); \
519 #define kiblnd_conn_decref(conn) \
521 unsigned long flags; \
523 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
524 (conn), atomic_read(&(conn)->ibc_refcount)); \
525 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
526 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
527 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
528 list_add_tail(&(conn)->ibc_list, \
529 &kiblnd_data.kib_connd_zombies); \
530 wake_up(&kiblnd_data.kib_connd_waitq); \
531 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); \
535 #define kiblnd_peer_addref(peer) \
537 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
538 (peer), libcfs_nid2str((peer)->ibp_nid), \
539 atomic_read (&(peer)->ibp_refcount)); \
540 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
541 atomic_inc(&(peer)->ibp_refcount); \
544 #define kiblnd_peer_decref(peer) \
546 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
547 (peer), libcfs_nid2str((peer)->ibp_nid), \
548 atomic_read (&(peer)->ibp_refcount)); \
549 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
550 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
551 kiblnd_destroy_peer(peer); \
554 static inline struct list_head *
555 kiblnd_nid2peerlist (lnet_nid_t nid)
557 unsigned int hash = ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
559 return (&kiblnd_data.kib_peers [hash]);
563 kiblnd_peer_active (kib_peer_t *peer)
565 /* Am I in the peer hash table? */
566 return (!list_empty(&peer->ibp_list));
569 static inline kib_conn_t *
570 kiblnd_get_conn_locked (kib_peer_t *peer)
572 LASSERT (!list_empty(&peer->ibp_conns));
574 /* just return the first connection */
575 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
579 kiblnd_send_keepalive(kib_conn_t *conn)
581 return (*kiblnd_tunables.kib_keepalive > 0) &&
582 time_after(jiffies, conn->ibc_last_send +
583 *kiblnd_tunables.kib_keepalive*HZ);
587 kiblnd_send_noop(kib_conn_t *conn)
589 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
591 if (conn->ibc_outstanding_credits <
592 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
593 !kiblnd_send_keepalive(conn))
594 return 0; /* No need to send NOOP */
596 if (!list_empty(&conn->ibc_tx_queue_nocred))
597 return 0; /* NOOP can be piggybacked */
599 if (!IBLND_OOB_CAPABLE(conn->ibc_version))
600 return list_empty(&conn->ibc_tx_queue); /* can't piggyback? */
602 /* No tx to piggyback NOOP onto or no credit to send a tx */
603 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
607 kiblnd_abort_receives(kib_conn_t *conn)
609 ib_modify_qp(conn->ibc_cmid->qp,
610 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
613 static inline const char *
614 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
616 if (q == &conn->ibc_tx_queue)
619 if (q == &conn->ibc_tx_queue_rsrvd)
620 return "tx_queue_rsrvd";
622 if (q == &conn->ibc_tx_queue_nocred)
623 return "tx_queue_nocred";
625 if (q == &conn->ibc_active_txs)
632 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
633 * lowest bits of the work request id to stash the work item type. */
635 #define IBLND_WID_TX 0
636 #define IBLND_WID_RDMA 1
637 #define IBLND_WID_RX 2
638 #define IBLND_WID_MASK 3UL
641 kiblnd_ptr2wreqid (void *ptr, int type)
643 unsigned long lptr = (unsigned long)ptr;
645 LASSERT ((lptr & IBLND_WID_MASK) == 0);
646 LASSERT ((type & ~IBLND_WID_MASK) == 0);
647 return (__u64)(lptr | type);
651 kiblnd_wreqid2ptr (__u64 wreqid)
653 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
657 kiblnd_wreqid2type (__u64 wreqid)
659 return (wreqid & IBLND_WID_MASK);
663 kiblnd_set_conn_state (kib_conn_t *conn, int state)
665 conn->ibc_state = state;
670 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
672 msg->ibm_type = type;
673 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
677 kiblnd_rd_size (kib_rdma_desc_t *rd)
682 for (i = size = 0; i < rd->rd_nfrags; i++)
683 size += rd->rd_frags[i].rf_nob;
689 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
691 return rd->rd_frags[index].rf_addr;
695 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
697 return rd->rd_frags[index].rf_nob;
701 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
707 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
709 if (nob < rd->rd_frags[index].rf_nob) {
710 rd->rd_frags[index].rf_addr += nob;
711 rd->rd_frags[index].rf_nob -= nob;
720 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
722 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
723 msgtype == IBLND_MSG_PUT_ACK);
725 return msgtype == IBLND_MSG_GET_REQ ?
726 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
727 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
730 #ifdef HAVE_OFED_IB_DMA_MAP
733 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
735 return ib_dma_mapping_error(dev, dma_addr);
738 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
739 void *msg, size_t size,
740 enum dma_data_direction direction)
742 return ib_dma_map_single(dev, msg, size, direction);
745 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
746 __u64 addr, size_t size,
747 enum dma_data_direction direction)
749 ib_dma_unmap_single(dev, addr, size, direction);
752 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
753 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
755 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
756 struct scatterlist *sg, int nents,
757 enum dma_data_direction direction)
759 return ib_dma_map_sg(dev, sg, nents, direction);
762 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
763 struct scatterlist *sg, int nents,
764 enum dma_data_direction direction)
766 ib_dma_unmap_sg(dev, sg, nents, direction);
769 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
770 struct scatterlist *sg)
772 return ib_sg_dma_address(dev, sg);
775 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
776 struct scatterlist *sg)
778 return ib_sg_dma_len(dev, sg);
781 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
782 * right because OFED1.2 defines it as const, to use it we have to add
783 * (void *) cast to overcome "const" */
785 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
786 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
791 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
793 return dma_mapping_error(dma_addr);
796 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
797 void *msg, size_t size,
798 enum dma_data_direction direction)
800 return dma_map_single(dev->dma_device, msg, size, direction);
803 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
804 dma_addr_t addr, size_t size,
805 enum dma_data_direction direction)
807 dma_unmap_single(dev->dma_device, addr, size, direction);
810 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
811 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
813 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
814 struct scatterlist *sg, int nents,
815 enum dma_data_direction direction)
817 return dma_map_sg(dev->dma_device, sg, nents, direction);
820 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
821 struct scatterlist *sg, int nents,
822 enum dma_data_direction direction)
824 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
828 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
829 struct scatterlist *sg)
831 return sg_dma_address(sg);
835 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
836 struct scatterlist *sg)
838 return sg_dma_len(sg);
841 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
842 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
846 struct ib_mr *kiblnd_find_rd_dma_mr(kib_net_t *net,
847 kib_rdma_desc_t *rd);
848 struct ib_mr *kiblnd_find_dma_mr(kib_net_t *net,
849 __u64 addr, __u64 size);
850 void kiblnd_map_rx_descs(kib_conn_t *conn);
851 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
852 void kiblnd_map_tx_descs (lnet_ni_t *ni);
853 void kiblnd_unmap_tx_descs(lnet_ni_t *ni);
854 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
855 kib_rdma_desc_t *rd, int nfrags);
856 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
857 kib_phys_mr_t *kiblnd_phys_mr_map(kib_net_t *net, kib_rdma_desc_t *rd,
858 struct ib_phys_buf *ipb, __u64 *iova);
859 void kiblnd_phys_mr_unmap(kib_net_t *net, kib_phys_mr_t *pmr);
861 int kiblnd_startup (lnet_ni_t *ni);
862 void kiblnd_shutdown (lnet_ni_t *ni);
863 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
864 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, time_t *when);
866 int kiblnd_tunables_init(void);
867 void kiblnd_tunables_fini(void);
869 int kiblnd_connd (void *arg);
870 int kiblnd_scheduler(void *arg);
871 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
873 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
874 void kiblnd_free_pages (kib_pages_t *p);
876 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
877 struct rdma_cm_event *event);
878 int kiblnd_translate_mtu(int value);
880 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
881 void kiblnd_destroy_peer (kib_peer_t *peer);
882 void kiblnd_destroy_dev (kib_dev_t *dev);
883 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
884 void kiblnd_peer_alive (kib_peer_t *peer);
885 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
886 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
887 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
888 int version, __u64 incarnation);
889 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
891 void kiblnd_connreq_done(kib_conn_t *conn, int status);
892 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
893 int state, int version);
894 void kiblnd_destroy_conn (kib_conn_t *conn);
895 void kiblnd_close_conn (kib_conn_t *conn, int error);
896 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
898 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
899 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
901 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
902 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
903 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
904 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
905 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status);
906 void kiblnd_check_sends (kib_conn_t *conn);
908 void kiblnd_qp_event(struct ib_event *event, void *arg);
909 void kiblnd_cq_event(struct ib_event *event, void *arg);
910 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
912 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
913 int credits, lnet_nid_t dstnid, __u64 dststamp);
914 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
915 int kiblnd_post_rx (kib_rx_t *rx, int credit);
917 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
918 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
919 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
920 unsigned int offset, unsigned int mlen, unsigned int rlen);