4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd.h
34 * Author: Eric Barton <eric@bartonsoftware.com>
37 #include <linux/module.h>
38 #include <linux/kernel.h>
40 #if defined(NEED_LOCKDEP_IS_HELD_DISCARD_CONST) \
41 && defined(CONFIG_LOCKDEP) \
42 && defined(lockdep_is_held)
43 #undef lockdep_is_held
44 #define lockdep_is_held(lock) \
45 lock_is_held((struct lockdep_map *)&(lock)->dep_map)
48 #ifdef HAVE_COMPAT_RDMA
49 #include <linux/compat-2.6.h>
51 #ifdef LINUX_3_17_COMPAT_H
52 #undef NEED_KTIME_GET_REAL_NS
55 /* MOFED has its own bitmap_alloc backport */
56 #define HAVE_BITMAP_ALLOC 1
60 #include <linux/kthread.h>
62 #include <linux/string.h>
63 #include <linux/stat.h>
64 #include <linux/errno.h>
65 #include <linux/unistd.h>
66 #include <linux/uio.h>
68 #include <asm/uaccess.h>
71 #include <linux/init.h>
73 #include <linux/file.h>
74 #include <linux/stat.h>
75 #include <linux/list.h>
76 #include <linux/kmod.h>
77 #include <linux/sysctl.h>
78 #include <linux/pci.h>
83 #include <rdma/rdma_cm.h>
84 #include <rdma/ib_cm.h>
85 #include <rdma/ib_verbs.h>
86 #include <rdma/ib_fmr_pool.h>
88 #define DEBUG_SUBSYSTEM S_LND
90 #include <libcfs/libcfs.h>
91 #include <lnet/lib-lnet.h>
93 #define IBLND_PEER_HASH_SIZE 101 /* # peer_ni lists */
95 #define IBLND_N_SCHED 2
96 #define IBLND_N_SCHED_HIGH 4
99 int *kib_dev_failover; /* HCA failover */
100 unsigned int *kib_service; /* IB service number */
101 int *kib_cksum; /* checksum struct kib_msg? */
102 int *kib_timeout; /* comms timeout (seconds) */
103 int *kib_keepalive; /* keepalive timeout (seconds) */
104 char **kib_default_ipif; /* default IPoIB interface */
105 int *kib_retry_count;
106 int *kib_rnr_retry_count;
107 int *kib_ib_mtu; /* IB MTU */
108 int *kib_require_priv_port;/* accept only privileged ports */
109 int *kib_use_priv_port; /* use privileged port for active connect */
110 /* # threads on each CPT */
112 int *kib_wrq_sge; /* # sg elements per wrq */
113 int *kib_use_fastreg_gaps; /* enable discontiguous fastreg fragment support */
116 extern struct kib_tunables kiblnd_tunables;
118 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
119 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
121 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer_ni credits */
122 #define IBLND_CREDITS_MAX ((typeof(((struct kib_msg *) 0)->ibm_credits)) - 1) /* Max # of peer_ni credits */
124 /* when eagerly to return credits */
125 #define IBLND_CREDITS_HIGHWATER(t, conn) ((conn->ibc_version) == IBLND_MSG_VERSION_1 ? \
126 IBLND_CREDIT_HIGHWATER_V1 : \
127 min(t->lnd_peercredits_hiw, (__u32)conn->ibc_queue_depth - 1))
129 #ifdef HAVE_RDMA_CREATE_ID_5ARG
130 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
131 rdma_create_id((ns) ? (ns) : &init_net, cb, dev, ps, qpt)
133 # ifdef HAVE_RDMA_CREATE_ID_4ARG
134 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
135 rdma_create_id(cb, dev, ps, qpt)
137 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
138 rdma_create_id(cb, dev, ps)
142 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
143 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
144 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
146 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
147 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
149 /************************/
150 /* derived constants... */
151 /* Pools (shared by connections on each CPT) */
152 /* These pools can grow at runtime, so don't need give a very large value */
153 #define IBLND_TX_POOL 256
154 #define IBLND_FMR_POOL 256
155 #define IBLND_FMR_POOL_FLUSH 192
157 /* RX messages (per connection) */
158 #define IBLND_RX_MSGS(c) \
159 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
160 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
161 #define IBLND_RX_MSG_PAGES(c) \
162 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
164 /* WRs and CQEs (per connection) */
165 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
167 /* 2 = LNet msg + Transfer chain */
168 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + kiblnd_send_wrs(c))
172 /* o2iblnd can run over aliased interface */
174 #define KIB_IFNAME_SIZE IFALIASZ
176 #define KIB_IFNAME_SIZE 256
180 IBLND_DEV_CAPS_FASTREG_ENABLED = BIT(0),
181 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT = BIT(1),
182 IBLND_DEV_CAPS_FMR_ENABLED = BIT(2),
186 struct list_head ibd_list; /* chain on kib_devs */
187 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
188 __u32 ibd_ifip; /* IPoIB interface IP */
189 /** IPoIB interface name */
190 char ibd_ifname[KIB_IFNAME_SIZE];
191 int ibd_nnets; /* # nets extant */
193 time64_t ibd_next_failover;
194 /* # failover failures */
195 int ibd_failed_failover;
196 /* failover in progress */
197 unsigned int ibd_failover;
198 /* IPoIB interface is a bonding master */
199 unsigned int ibd_can_failover;
200 struct list_head ibd_nets;
201 struct kib_hca_dev *ibd_hdev;
202 enum kib_dev_caps ibd_dev_caps;
206 struct rdma_cm_id *ibh_cmid; /* listener cmid */
207 struct ib_device *ibh_ibdev; /* IB device */
208 int ibh_page_shift; /* page shift of current HCA */
209 int ibh_page_size; /* page size of current HCA */
210 __u64 ibh_page_mask; /* page mask of current HCA */
211 __u64 ibh_mr_size; /* size of MR */
212 int ibh_max_qp_wr; /* maximum work requests size */
213 #ifdef HAVE_IB_GET_DMA_MR
214 struct ib_mr *ibh_mrs; /* global MR */
216 struct ib_pd *ibh_pd; /* PD */
217 u8 ibh_port; /* port number */
218 struct ib_event_handler
219 ibh_event_handler; /* IB event handler */
220 int ibh_state; /* device status */
221 #define IBLND_DEV_PORT_DOWN 0
222 #define IBLND_DEV_PORT_ACTIVE 1
223 #define IBLND_DEV_FATAL 2
224 struct kib_dev *ibh_dev; /* owner */
225 atomic_t ibh_ref; /* refcount */
228 /** # of seconds to keep pool alive */
229 #define IBLND_POOL_DEADLINE 300
230 /** # of seconds to retry if allocation failed */
231 #define IBLND_POOL_RETRY 1
234 int ibp_npages; /* # pages */
235 struct page *ibp_pages[0]; /* page array */
241 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
242 int inc, struct kib_pool **pp_po);
243 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
244 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
245 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
249 #define IBLND_POOL_NAME_LEN 32
254 /* network it belongs to */
255 struct kib_net *ps_net;
257 char ps_name[IBLND_POOL_NAME_LEN];
259 struct list_head ps_pool_list;
260 /* failed pool list */
261 struct list_head ps_failed_pool_list;
262 /* time stamp for retry if failed to allocate */
263 time64_t ps_next_retry;
264 /* is allocating new pool */
271 /* create a new pool */
272 kib_ps_pool_create_t ps_pool_create;
274 kib_ps_pool_destroy_t ps_pool_destroy;
275 /* initialize new allocated node */
276 kib_ps_node_init_t ps_node_init;
278 kib_ps_node_fini_t ps_node_fini;
282 /* chain on pool list */
283 struct list_head po_list;
284 /* pre-allocated node */
285 struct list_head po_free_list;
286 /* pool_set of this pool */
287 struct kib_poolset *po_owner;
288 /* deadline of this pool */
289 time64_t po_deadline;
290 /* # of elements in use */
292 /* pool is created on failed HCA */
294 /* # of pre-allocated elements */
298 struct kib_tx_poolset {
299 struct kib_poolset tps_poolset; /* pool-set */
300 __u64 tps_next_tx_cookie; /* cookie of TX */
304 struct kib_pool tpo_pool; /* pool */
305 struct kib_hca_dev *tpo_hdev; /* device for this pool */
306 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
307 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
310 struct kib_fmr_poolset {
311 spinlock_t fps_lock; /* serialize */
312 struct kib_net *fps_net; /* IB network */
313 struct list_head fps_pool_list; /* FMR pool list */
314 struct list_head fps_failed_pool_list; /* FMR pool list */
315 __u64 fps_version; /* validity stamp */
316 int fps_cpt; /* CPT id */
318 int fps_flush_trigger;
320 /* is allocating new pool */
322 /* time stamp for retry if failed to allocate */
323 time64_t fps_next_retry;
326 #ifndef HAVE_IB_RDMA_WR
328 struct ib_send_wr wr;
332 struct kib_fast_reg_descriptor { /* For fast registration */
333 struct list_head frd_list;
334 struct ib_rdma_wr frd_inv_wr;
335 #ifdef HAVE_IB_MAP_MR_SG
336 struct ib_reg_wr frd_fastreg_wr;
338 struct ib_rdma_wr frd_fastreg_wr;
339 struct ib_fast_reg_page_list *frd_frpl;
341 struct ib_mr *frd_mr;
345 struct kib_fmr_pool {
346 struct list_head fpo_list; /* chain on pool list */
347 struct kib_hca_dev *fpo_hdev; /* device for this pool */
348 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
351 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
353 struct { /* For fast registration */
354 struct list_head fpo_pool_list;
358 time64_t fpo_deadline; /* deadline of this pool */
359 int fpo_failed; /* fmr pool is failed */
360 int fpo_map_count; /* # of mapped FMR */
361 bool fpo_is_fmr; /* True if FMR pools allocated */
365 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
366 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
367 struct kib_fast_reg_descriptor *fmr_frd;
371 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
372 #define kib_fmr_pool_map(pool, pgs, n, iov) \
373 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
375 #define kib_fmr_pool_map(pool, pgs, n, iov) \
376 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
380 /* chain on struct kib_dev::ibd_nets */
381 struct list_head ibn_list;
382 __u64 ibn_incarnation;/* my epoch */
383 int ibn_init; /* initialisation state */
384 int ibn_shutdown; /* shutting down? */
386 atomic_t ibn_npeers; /* # peers extant */
387 atomic_t ibn_nconns; /* # connections extant */
389 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
390 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
392 struct kib_dev *ibn_dev; /* underlying IB device */
393 struct lnet_ni *ibn_ni; /* LNet interface */
396 #define KIB_THREAD_SHIFT 16
397 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
398 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
399 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
401 struct kib_sched_info {
404 /* schedulers sleep here */
405 wait_queue_head_t ibs_waitq;
406 /* conns to check for rx completions */
407 struct list_head ibs_conns;
408 /* number of scheduler threads */
410 /* max allowed scheduler threads */
411 int ibs_nthreads_max;
412 int ibs_cpt; /* CPT id */
416 int kib_init; /* initialisation state */
417 int kib_shutdown; /* shut down? */
418 struct list_head kib_devs; /* IB devices extant */
419 /* list head of failed devices */
420 struct list_head kib_failed_devs;
421 /* schedulers sleep here */
422 wait_queue_head_t kib_failover_waitq;
423 atomic_t kib_nthreads; /* # live threads */
424 /* stabilize net/dev/peer_ni/conn ops */
425 rwlock_t kib_global_lock;
426 /* hash table of all my known peers */
427 struct list_head *kib_peers;
428 /* size of kib_peers */
429 int kib_peer_hash_size;
430 /* the connd task (serialisation assertions) */
432 /* connections to setup/teardown */
433 struct list_head kib_connd_conns;
434 /* connections with zero refcount */
435 struct list_head kib_connd_zombies;
436 /* connections to reconnect */
437 struct list_head kib_reconn_list;
438 /* peers wait for reconnection */
439 struct list_head kib_reconn_wait;
441 * The second that peers are pulled out from \a kib_reconn_wait
444 time64_t kib_reconn_sec;
445 /* connection daemon sleeps here */
446 wait_queue_head_t kib_connd_waitq;
447 spinlock_t kib_connd_lock; /* serialise */
448 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
449 /* percpt data for schedulers */
450 struct kib_sched_info **kib_scheds;
453 #define IBLND_INIT_NOTHING 0
454 #define IBLND_INIT_DATA 1
455 #define IBLND_INIT_ALL 2
457 /************************************************************************
458 * IB Wire message format.
459 * These are sent in sender's byte order (i.e. receiver flips).
462 struct kib_connparams {
463 __u16 ibcp_queue_depth;
464 __u16 ibcp_max_frags;
465 __u32 ibcp_max_msg_size;
468 struct kib_immediate_msg {
469 struct lnet_hdr ibim_hdr; /* portals header */
470 char ibim_payload[0];/* piggy-backed payload */
473 struct kib_rdma_frag {
474 __u32 rf_nob; /* # bytes this frag */
475 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
478 struct kib_rdma_desc {
479 __u32 rd_key; /* local/remote key */
480 __u32 rd_nfrags; /* # fragments */
481 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
484 struct kib_putreq_msg {
485 struct lnet_hdr ibprm_hdr; /* portals header */
486 __u64 ibprm_cookie; /* opaque completion cookie */
489 struct kib_putack_msg {
490 __u64 ibpam_src_cookie; /* reflected completion cookie */
491 __u64 ibpam_dst_cookie; /* opaque completion cookie */
492 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
496 struct lnet_hdr ibgm_hdr; /* portals header */
497 __u64 ibgm_cookie; /* opaque completion cookie */
498 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
501 struct kib_completion_msg {
502 __u64 ibcm_cookie; /* opaque completion cookie */
503 __s32 ibcm_status; /* < 0 failure: >= 0 length */
507 /* First 2 fields fixed FOR ALL TIME */
508 __u32 ibm_magic; /* I'm an ibnal message */
509 __u16 ibm_version; /* this is my version number */
511 __u8 ibm_type; /* msg type */
512 __u8 ibm_credits; /* returned credits */
513 __u32 ibm_nob; /* # bytes in whole message */
514 __u32 ibm_cksum; /* checksum (0 == no checksum) */
515 __u64 ibm_srcnid; /* sender's NID */
516 __u64 ibm_srcstamp; /* sender's incarnation */
517 __u64 ibm_dstnid; /* destination's NID */
518 __u64 ibm_dststamp; /* destination's incarnation */
521 struct kib_connparams connparams;
522 struct kib_immediate_msg immediate;
523 struct kib_putreq_msg putreq;
524 struct kib_putack_msg putack;
525 struct kib_get_msg get;
526 struct kib_completion_msg completion;
530 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
532 #define IBLND_MSG_VERSION_1 0x11
533 #define IBLND_MSG_VERSION_2 0x12
534 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
536 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
537 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
538 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
539 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
540 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
541 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
542 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
543 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
544 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
545 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
548 __u32 ibr_magic; /* sender's magic */
549 __u16 ibr_version; /* sender's version */
550 __u8 ibr_why; /* reject reason */
551 __u8 ibr_padding; /* padding */
552 __u64 ibr_incarnation; /* incarnation of peer_ni */
553 struct kib_connparams ibr_cp; /* connection parameters */
556 /* connection rejection reasons */
557 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
558 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
559 #define IBLND_REJECT_FATAL 3 /* Anything else */
561 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
562 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
564 /* peer_ni's rdma frags doesn't match mine */
565 #define IBLND_REJECT_RDMA_FRAGS 6
566 /* peer_ni's msg queue size doesn't match mine */
567 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
568 #define IBLND_REJECT_INVALID_SRV_ID 8
570 /***********************************************************************/
572 struct kib_rx { /* receive message */
573 /* queue for attention */
574 struct list_head rx_list;
576 struct kib_conn *rx_conn;
577 /* # bytes received (-1 while posted) */
579 /* message buffer (host vaddr) */
580 struct kib_msg *rx_msg;
581 /* message buffer (I/O addr) */
583 /* for dma_unmap_single() */
584 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
585 /* receive work item... */
586 struct ib_recv_wr rx_wrq;
587 /* ...and its memory */
588 struct ib_sge rx_sge;
591 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
592 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
593 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
594 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
596 struct kib_tx { /* transmit message */
597 /* queue on idle_txs ibc_tx_queue etc. */
598 struct list_head tx_list;
600 struct kib_tx_pool *tx_pool;
602 struct kib_conn *tx_conn;
603 /* # tx callbacks outstanding */
605 /* queued for sending */
607 /* waiting for peer_ni */
609 /* LNET completion status */
611 /* health status of the transmit */
612 enum lnet_msg_hstatus tx_hstatus;
613 /* completion deadline */
615 /* completion cookie */
617 /* lnet msgs to finalize on completion */
618 struct lnet_msg *tx_lntmsg[2];
619 /* message buffer (host vaddr) */
620 struct kib_msg *tx_msg;
621 /* message buffer (I/O addr) */
623 /* for dma_unmap_single() */
624 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
625 /** sge for tx_msgaddr */
626 struct ib_sge tx_msgsge;
627 /* # send work items */
629 /* # used scatter/gather elements */
631 /* send work items... */
632 struct ib_rdma_wr *tx_wrq;
633 /* ...and their memory */
634 struct ib_sge *tx_sge;
635 /* rdma descriptor */
636 struct kib_rdma_desc *tx_rd;
637 /* # entries in... */
639 /* dma_map_sg descriptor */
640 struct scatterlist *tx_frags;
641 /* rdma phys page addrs */
643 /* gaps in fragments */
646 struct kib_fmr tx_fmr;
651 struct kib_connvars {
652 /* connection-in-progress variables */
653 struct kib_msg cv_msg;
657 /* scheduler information */
658 struct kib_sched_info *ibc_sched;
660 struct kib_peer_ni *ibc_peer;
662 struct kib_hca_dev *ibc_hdev;
663 /* stash on peer_ni's conn list */
664 struct list_head ibc_list;
665 /* schedule for attention */
666 struct list_head ibc_sched_list;
667 /* version of connection */
669 /* reconnect later */
670 __u16 ibc_reconnect:1;
671 /* which instance of the peer */
672 __u64 ibc_incarnation;
674 atomic_t ibc_refcount;
675 /* what's happening */
677 /* # uncompleted sends */
678 int ibc_nsends_posted;
679 /* # uncompleted NOOPs */
680 int ibc_noops_posted;
681 /* # credits I have */
683 /* # credits to return */
684 int ibc_outstanding_credits;
685 /* # ACK/DONE msg credits */
686 int ibc_reserved_credits;
687 /* set on comms error */
689 /* connections queue depth */
690 __u16 ibc_queue_depth;
691 /* connections max frags */
693 /* receive buffers owned */
694 unsigned int ibc_nrx:16;
695 /* scheduled for attention */
696 unsigned int ibc_scheduled:1;
697 /* CQ callback fired */
698 unsigned int ibc_ready:1;
699 /* time of last send */
700 ktime_t ibc_last_send;
701 /** link chain for kiblnd_check_conns only */
702 struct list_head ibc_connd_list;
703 /** rxs completed before ESTABLISHED */
704 struct list_head ibc_early_rxs;
705 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
706 struct list_head ibc_tx_noops;
707 /* sends that need a credit */
708 struct list_head ibc_tx_queue;
709 /* sends that don't need a credit */
710 struct list_head ibc_tx_queue_nocred;
711 /* sends that need to reserve an ACK/DONE msg */
712 struct list_head ibc_tx_queue_rsrvd;
713 /* active tx awaiting completion */
714 struct list_head ibc_active_txs;
715 /* zombie tx awaiting done */
716 struct list_head ibc_zombie_txs;
720 struct kib_rx *ibc_rxs;
721 /* premapped rx msg pages */
722 struct kib_pages *ibc_rx_pages;
725 struct rdma_cm_id *ibc_cmid;
726 /* completion queue */
727 struct ib_cq *ibc_cq;
729 /* in-progress connection state */
730 struct kib_connvars *ibc_connvars;
733 #define IBLND_CONN_INIT 0 /* being initialised */
734 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
735 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
736 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
737 #define IBLND_CONN_CLOSING 4 /* being closed */
738 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
741 /* stash on global peer_ni list */
742 struct list_head ibp_list;
743 /* who's on the other end(s) */
746 struct lnet_ni *ibp_ni;
747 /* all active connections */
748 struct list_head ibp_conns;
749 /* next connection to send on for round robin */
750 struct kib_conn *ibp_next_conn;
751 /* msgs waiting for a conn */
752 struct list_head ibp_tx_queue;
753 /* incarnation of peer_ni */
754 __u64 ibp_incarnation;
755 /* when (in seconds) I was last alive */
756 time64_t ibp_last_alive;
758 atomic_t ibp_refcount;
759 /* version of peer_ni */
761 /* current passive connection attempts */
762 unsigned short ibp_accepting;
763 /* current active connection attempts */
764 unsigned short ibp_connecting;
765 /* reconnect this peer_ni later */
766 unsigned char ibp_reconnecting;
767 /* counter of how many times we triggered a conn race */
768 unsigned char ibp_races;
769 /* # consecutive reconnection attempts to this peer */
770 unsigned int ibp_reconnected;
771 /* number of total active retries */
772 unsigned int ibp_retries;
773 /* errno on closing this peer_ni */
775 /* max map_on_demand */
777 /* max_peer_credits */
778 __u16 ibp_queue_depth;
781 #ifndef HAVE_IB_INC_RKEY
783 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
784 * for calculating a new rkey for type 2 memory windows.
785 * @rkey - the rkey to increment.
787 static inline u32 ib_inc_rkey(u32 rkey)
789 const u32 mask = 0x000000ff;
790 return ((rkey + 1) & mask) | (rkey & ~mask);
794 extern struct kib_data kiblnd_data;
796 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
798 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
800 static inline int kiblnd_timeout(void)
802 return *kiblnd_tunables.kib_timeout ? *kiblnd_tunables.kib_timeout :
803 lnet_get_lnd_timeout();
807 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
809 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
810 int concurrent_sends;
812 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
813 concurrent_sends = tunables->lnd_concurrent_sends;
815 if (version == IBLND_MSG_VERSION_1) {
816 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
817 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
819 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
820 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
823 return concurrent_sends;
827 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
829 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
830 atomic_inc(&hdev->ibh_ref);
834 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
836 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
837 if (atomic_dec_and_test(&hdev->ibh_ref))
838 kiblnd_hdev_destroy(hdev);
842 kiblnd_dev_can_failover(struct kib_dev *dev)
844 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
847 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
850 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
853 return dev->ibd_can_failover;
856 #define kiblnd_conn_addref(conn) \
858 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
859 (conn), atomic_read(&(conn)->ibc_refcount)); \
860 atomic_inc(&(conn)->ibc_refcount); \
863 #define kiblnd_conn_decref(conn) \
865 unsigned long flags; \
867 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
868 (conn), atomic_read(&(conn)->ibc_refcount)); \
869 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
870 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
871 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
872 list_add_tail(&(conn)->ibc_list, \
873 &kiblnd_data.kib_connd_zombies); \
874 wake_up(&kiblnd_data.kib_connd_waitq); \
875 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
879 #define kiblnd_peer_addref(peer_ni) \
881 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
882 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
883 atomic_read (&(peer_ni)->ibp_refcount)); \
884 atomic_inc(&(peer_ni)->ibp_refcount); \
887 #define kiblnd_peer_decref(peer_ni) \
889 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
890 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
891 atomic_read (&(peer_ni)->ibp_refcount)); \
892 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
893 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
894 kiblnd_destroy_peer(peer_ni); \
898 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
900 return peer_ni->ibp_connecting != 0 ||
901 peer_ni->ibp_reconnecting != 0 ||
902 peer_ni->ibp_accepting != 0;
906 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
908 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
911 static inline struct list_head *
912 kiblnd_nid2peerlist (lnet_nid_t nid)
915 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
917 return &kiblnd_data.kib_peers[hash];
921 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
923 /* Am I in the peer_ni hash table? */
924 return !list_empty(&peer_ni->ibp_list);
927 static inline struct kib_conn *
928 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
930 struct list_head *next;
932 LASSERT(!list_empty(&peer_ni->ibp_conns));
934 /* Advance to next connection, be sure to skip the head node */
935 if (!peer_ni->ibp_next_conn ||
936 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
937 next = peer_ni->ibp_conns.next;
939 next = peer_ni->ibp_next_conn->ibc_list.next;
940 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
942 return peer_ni->ibp_next_conn;
946 kiblnd_send_keepalive(struct kib_conn *conn)
948 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
950 return (*kiblnd_tunables.kib_keepalive > 0) &&
951 ktime_after(ktime_get(),
952 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
956 kiblnd_need_noop(struct kib_conn *conn)
958 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
959 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
961 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
962 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
964 if (conn->ibc_outstanding_credits <
965 IBLND_CREDITS_HIGHWATER(tunables, conn) &&
966 !kiblnd_send_keepalive(conn))
967 return 0; /* No need to send NOOP */
969 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
970 if (!list_empty(&conn->ibc_tx_queue_nocred))
971 return 0; /* NOOP can be piggybacked */
973 /* No tx to piggyback NOOP onto or no credit to send a tx */
974 return (list_empty(&conn->ibc_tx_queue) ||
975 conn->ibc_credits == 0);
978 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
979 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
980 conn->ibc_credits == 0) /* no credit */
983 if (conn->ibc_credits == 1 && /* last credit reserved for */
984 conn->ibc_outstanding_credits == 0) /* giving back credits */
987 /* No tx to piggyback NOOP onto or no credit to send a tx */
988 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
992 kiblnd_abort_receives(struct kib_conn *conn)
994 ib_modify_qp(conn->ibc_cmid->qp,
995 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
998 static inline const char *
999 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
1001 if (q == &conn->ibc_tx_queue)
1004 if (q == &conn->ibc_tx_queue_rsrvd)
1005 return "tx_queue_rsrvd";
1007 if (q == &conn->ibc_tx_queue_nocred)
1008 return "tx_queue_nocred";
1010 if (q == &conn->ibc_active_txs)
1011 return "active_txs";
1017 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1018 * lowest bits of the work request id to stash the work item type. */
1020 #define IBLND_WID_INVAL 0
1021 #define IBLND_WID_TX 1
1022 #define IBLND_WID_RX 2
1023 #define IBLND_WID_RDMA 3
1024 #define IBLND_WID_MR 4
1025 #define IBLND_WID_MASK 7UL
1028 kiblnd_ptr2wreqid (void *ptr, int type)
1030 unsigned long lptr = (unsigned long)ptr;
1032 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1033 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1034 return (__u64)(lptr | type);
1037 static inline void *
1038 kiblnd_wreqid2ptr (__u64 wreqid)
1040 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1044 kiblnd_wreqid2type (__u64 wreqid)
1046 return (wreqid & IBLND_WID_MASK);
1050 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1052 conn->ibc_state = state;
1057 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1059 msg->ibm_type = type;
1060 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1064 kiblnd_rd_size(struct kib_rdma_desc *rd)
1069 for (i = size = 0; i < rd->rd_nfrags; i++)
1070 size += rd->rd_frags[i].rf_nob;
1076 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1078 return rd->rd_frags[index].rf_addr;
1082 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1084 return rd->rd_frags[index].rf_nob;
1088 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1094 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1096 if (nob < rd->rd_frags[index].rf_nob) {
1097 rd->rd_frags[index].rf_addr += nob;
1098 rd->rd_frags[index].rf_nob -= nob;
1107 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1109 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1110 msgtype == IBLND_MSG_PUT_ACK);
1112 return msgtype == IBLND_MSG_GET_REQ ?
1113 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1114 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1118 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1120 return ib_dma_mapping_error(dev, dma_addr);
1123 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1124 void *msg, size_t size,
1125 enum dma_data_direction direction)
1127 return ib_dma_map_single(dev, msg, size, direction);
1130 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1131 __u64 addr, size_t size,
1132 enum dma_data_direction direction)
1134 ib_dma_unmap_single(dev, addr, size, direction);
1137 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1138 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1140 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1141 struct scatterlist *sg, int nents,
1142 enum dma_data_direction direction)
1144 return ib_dma_map_sg(dev, sg, nents, direction);
1147 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1148 struct scatterlist *sg, int nents,
1149 enum dma_data_direction direction)
1151 ib_dma_unmap_sg(dev, sg, nents, direction);
1154 #ifndef HAVE_IB_SG_DMA_ADDRESS
1155 #include <linux/scatterlist.h>
1156 #define ib_sg_dma_address(dev, sg) sg_dma_address(sg)
1157 #define ib_sg_dma_len(dev, sg) sg_dma_len(sg)
1160 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1161 struct scatterlist *sg)
1163 return ib_sg_dma_address(dev, sg);
1166 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1167 struct scatterlist *sg)
1169 return ib_sg_dma_len(dev, sg);
1172 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1173 * right because OFED1.2 defines it as const, to use it we have to add
1174 * (void *) cast to overcome "const" */
1176 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1177 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1179 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1180 void kiblnd_map_rx_descs(struct kib_conn *conn);
1181 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1182 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1183 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1185 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1186 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1187 struct kib_fmr *fmr);
1188 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1190 int kiblnd_tunables_setup(struct lnet_ni *ni);
1191 int kiblnd_tunables_init(void);
1193 int kiblnd_connd (void *arg);
1194 int kiblnd_scheduler(void *arg);
1195 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1196 int kiblnd_failover_thread (void *arg);
1198 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1200 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1201 struct rdma_cm_event *event);
1202 int kiblnd_translate_mtu(int value);
1204 int kiblnd_dev_failover(struct kib_dev *dev, struct net *ns);
1205 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1207 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1208 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1209 void kiblnd_destroy_dev(struct kib_dev *dev);
1210 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1211 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1212 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1213 int version, u64 incarnation);
1214 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1216 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1217 struct rdma_cm_id *cmid,
1218 int state, int version);
1219 void kiblnd_destroy_conn(struct kib_conn *conn);
1220 void kiblnd_close_conn(struct kib_conn *conn, int error);
1221 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1223 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1224 void kiblnd_txlist_done(struct list_head *txlist, int status,
1225 enum lnet_msg_hstatus hstatus);
1227 void kiblnd_qp_event(struct ib_event *event, void *arg);
1228 void kiblnd_cq_event(struct ib_event *event, void *arg);
1229 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1231 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1232 int credits, lnet_nid_t dstnid, __u64 dststamp);
1233 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1234 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1236 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1237 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1238 int delayed, unsigned int niov,
1239 struct bio_vec *kiov, unsigned int offset, unsigned int mlen,