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 #ifdef HAVE_COMPAT_RDMA
38 #include <linux/compat-2.6.h>
40 #ifdef LINUX_3_17_COMPAT_H
41 #undef NEED_KTIME_GET_REAL_NS
46 #include <linux/module.h>
47 #include <linux/kernel.h>
48 #include <linux/kthread.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/unistd.h>
54 #include <linux/uio.h>
56 #include <asm/uaccess.h>
59 #include <linux/init.h>
61 #include <linux/file.h>
62 #include <linux/stat.h>
63 #include <linux/list.h>
64 #include <linux/kmod.h>
65 #include <linux/sysctl.h>
66 #include <linux/pci.h>
71 #include <rdma/rdma_cm.h>
72 #include <rdma/ib_cm.h>
73 #include <rdma/ib_verbs.h>
74 #include <rdma/ib_fmr_pool.h>
76 #define DEBUG_SUBSYSTEM S_LND
78 #include <libcfs/libcfs.h>
79 #include <lnet/lib-lnet.h>
81 #define IBLND_PEER_HASH_SIZE 101 /* # peer_ni lists */
82 /* # scheduler loops before reschedule */
83 #define IBLND_RESCHED 100
85 #define IBLND_N_SCHED 2
86 #define IBLND_N_SCHED_HIGH 4
89 int *kib_dev_failover; /* HCA failover */
90 unsigned int *kib_service; /* IB service number */
91 int *kib_cksum; /* checksum struct kib_msg? */
92 int *kib_timeout; /* comms timeout (seconds) */
93 int *kib_keepalive; /* keepalive timeout (seconds) */
94 char **kib_default_ipif; /* default IPoIB interface */
96 int *kib_rnr_retry_count;
97 int *kib_ib_mtu; /* IB MTU */
98 int *kib_require_priv_port;/* accept only privileged ports */
99 int *kib_use_priv_port; /* use privileged port for active connect */
100 /* # threads on each CPT */
102 int *kib_wrq_sge; /* # sg elements per wrq */
103 int *kib_use_fastreg_gaps; /* enable discontiguous fastreg fragment support */
106 extern struct kib_tunables kiblnd_tunables;
108 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
109 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
111 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer_ni credits */
112 #define IBLND_CREDITS_MAX ((typeof(((struct kib_msg *) 0)->ibm_credits)) - 1) /* Max # of peer_ni credits */
114 /* when eagerly to return credits */
115 #define IBLND_CREDITS_HIGHWATER(t, conn) ((conn->ibc_version) == IBLND_MSG_VERSION_1 ? \
116 IBLND_CREDIT_HIGHWATER_V1 : \
117 min(t->lnd_peercredits_hiw, (__u32)conn->ibc_queue_depth - 1))
119 #ifdef HAVE_RDMA_CREATE_ID_5ARG
120 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) rdma_create_id(ns, cb, \
124 # ifdef HAVE_RDMA_CREATE_ID_4ARG
125 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) rdma_create_id(cb, dev, \
128 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) rdma_create_id(cb, dev, \
133 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
134 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
135 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
137 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
138 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
140 /************************/
141 /* derived constants... */
142 /* Pools (shared by connections on each CPT) */
143 /* These pools can grow at runtime, so don't need give a very large value */
144 #define IBLND_TX_POOL 256
145 #define IBLND_FMR_POOL 256
146 #define IBLND_FMR_POOL_FLUSH 192
148 /* RX messages (per connection) */
149 #define IBLND_RX_MSGS(c) \
150 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
151 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
152 #define IBLND_RX_MSG_PAGES(c) \
153 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
155 /* WRs and CQEs (per connection) */
156 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
158 /* 2 = LNet msg + Transfer chain */
159 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + kiblnd_send_wrs(c))
163 /* o2iblnd can run over aliased interface */
165 #define KIB_IFNAME_SIZE IFALIASZ
167 #define KIB_IFNAME_SIZE 256
171 IBLND_DEV_CAPS_FASTREG_ENABLED = BIT(0),
172 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT = BIT(1),
173 IBLND_DEV_CAPS_FMR_ENABLED = BIT(2),
177 struct list_head ibd_list; /* chain on kib_devs */
178 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
179 __u32 ibd_ifip; /* IPoIB interface IP */
180 /** IPoIB interface name */
181 char ibd_ifname[KIB_IFNAME_SIZE];
182 int ibd_nnets; /* # nets extant */
184 time64_t ibd_next_failover;
185 /* # failover failures */
186 int ibd_failed_failover;
187 /* failover in progress */
188 unsigned int ibd_failover;
189 /* IPoIB interface is a bonding master */
190 unsigned int ibd_can_failover;
191 struct list_head ibd_nets;
192 struct kib_hca_dev *ibd_hdev;
193 enum kib_dev_caps ibd_dev_caps;
197 struct rdma_cm_id *ibh_cmid; /* listener cmid */
198 struct ib_device *ibh_ibdev; /* IB device */
199 int ibh_page_shift; /* page shift of current HCA */
200 int ibh_page_size; /* page size of current HCA */
201 __u64 ibh_page_mask; /* page mask of current HCA */
202 __u64 ibh_mr_size; /* size of MR */
203 int ibh_max_qp_wr; /* maximum work requests size */
204 #ifdef HAVE_IB_GET_DMA_MR
205 struct ib_mr *ibh_mrs; /* global MR */
207 struct ib_pd *ibh_pd; /* PD */
208 struct kib_dev *ibh_dev; /* owner */
209 atomic_t ibh_ref; /* refcount */
212 /** # of seconds to keep pool alive */
213 #define IBLND_POOL_DEADLINE 300
214 /** # of seconds to retry if allocation failed */
215 #define IBLND_POOL_RETRY 1
218 int ibp_npages; /* # pages */
219 struct page *ibp_pages[0]; /* page array */
225 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
226 int inc, struct kib_pool **pp_po);
227 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
228 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
229 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
233 #define IBLND_POOL_NAME_LEN 32
238 /* network it belongs to */
239 struct kib_net *ps_net;
241 char ps_name[IBLND_POOL_NAME_LEN];
243 struct list_head ps_pool_list;
244 /* failed pool list */
245 struct list_head ps_failed_pool_list;
246 /* time stamp for retry if failed to allocate */
247 time64_t ps_next_retry;
248 /* is allocating new pool */
255 /* create a new pool */
256 kib_ps_pool_create_t ps_pool_create;
258 kib_ps_pool_destroy_t ps_pool_destroy;
259 /* initialize new allocated node */
260 kib_ps_node_init_t ps_node_init;
262 kib_ps_node_fini_t ps_node_fini;
266 /* chain on pool list */
267 struct list_head po_list;
268 /* pre-allocated node */
269 struct list_head po_free_list;
270 /* pool_set of this pool */
271 struct kib_poolset *po_owner;
272 /* deadline of this pool */
273 time64_t po_deadline;
274 /* # of elements in use */
276 /* pool is created on failed HCA */
278 /* # of pre-allocated elements */
282 struct kib_tx_poolset {
283 struct kib_poolset tps_poolset; /* pool-set */
284 __u64 tps_next_tx_cookie; /* cookie of TX */
288 struct kib_pool tpo_pool; /* pool */
289 struct kib_hca_dev *tpo_hdev; /* device for this pool */
290 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
291 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
294 struct kib_fmr_poolset {
295 spinlock_t fps_lock; /* serialize */
296 struct kib_net *fps_net; /* IB network */
297 struct list_head fps_pool_list; /* FMR pool list */
298 struct list_head fps_failed_pool_list; /* FMR pool list */
299 __u64 fps_version; /* validity stamp */
300 int fps_cpt; /* CPT id */
302 int fps_flush_trigger;
304 /* is allocating new pool */
306 /* time stamp for retry if failed to allocate */
307 time64_t fps_next_retry;
310 #ifndef HAVE_IB_RDMA_WR
312 struct ib_send_wr wr;
316 struct kib_fast_reg_descriptor { /* For fast registration */
317 struct list_head frd_list;
318 struct ib_rdma_wr frd_inv_wr;
319 #ifdef HAVE_IB_MAP_MR_SG
320 struct ib_reg_wr frd_fastreg_wr;
322 struct ib_rdma_wr frd_fastreg_wr;
323 struct ib_fast_reg_page_list *frd_frpl;
325 struct ib_mr *frd_mr;
329 struct kib_fmr_pool {
330 struct list_head fpo_list; /* chain on pool list */
331 struct kib_hca_dev *fpo_hdev; /* device for this pool */
332 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
335 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
337 struct { /* For fast registration */
338 struct list_head fpo_pool_list;
342 time64_t fpo_deadline; /* deadline of this pool */
343 int fpo_failed; /* fmr pool is failed */
344 int fpo_map_count; /* # of mapped FMR */
345 bool fpo_is_fmr; /* True if FMR pools allocated */
349 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
350 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
351 struct kib_fast_reg_descriptor *fmr_frd;
355 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
356 #define kib_fmr_pool_map(pool, pgs, n, iov) \
357 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
359 #define kib_fmr_pool_map(pool, pgs, n, iov) \
360 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
364 /* chain on struct kib_dev::ibd_nets */
365 struct list_head ibn_list;
366 __u64 ibn_incarnation;/* my epoch */
367 int ibn_init; /* initialisation state */
368 int ibn_shutdown; /* shutting down? */
370 atomic_t ibn_npeers; /* # peers extant */
371 atomic_t ibn_nconns; /* # connections extant */
373 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
374 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
376 struct kib_dev *ibn_dev; /* underlying IB device */
379 #define KIB_THREAD_SHIFT 16
380 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
381 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
382 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
384 struct kib_sched_info {
387 /* schedulers sleep here */
388 wait_queue_head_t ibs_waitq;
389 /* conns to check for rx completions */
390 struct list_head ibs_conns;
391 /* number of scheduler threads */
393 /* max allowed scheduler threads */
394 int ibs_nthreads_max;
395 int ibs_cpt; /* CPT id */
399 int kib_init; /* initialisation state */
400 int kib_shutdown; /* shut down? */
401 struct list_head kib_devs; /* IB devices extant */
402 /* list head of failed devices */
403 struct list_head kib_failed_devs;
404 /* schedulers sleep here */
405 wait_queue_head_t kib_failover_waitq;
406 atomic_t kib_nthreads; /* # live threads */
407 /* stabilize net/dev/peer_ni/conn ops */
408 rwlock_t kib_global_lock;
409 /* hash table of all my known peers */
410 struct list_head *kib_peers;
411 /* size of kib_peers */
412 int kib_peer_hash_size;
413 /* the connd task (serialisation assertions) */
415 /* connections to setup/teardown */
416 struct list_head kib_connd_conns;
417 /* connections with zero refcount */
418 struct list_head kib_connd_zombies;
419 /* connections to reconnect */
420 struct list_head kib_reconn_list;
421 /* peers wait for reconnection */
422 struct list_head kib_reconn_wait;
424 * The second that peers are pulled out from \a kib_reconn_wait
427 time64_t kib_reconn_sec;
428 /* connection daemon sleeps here */
429 wait_queue_head_t kib_connd_waitq;
430 spinlock_t kib_connd_lock; /* serialise */
431 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
432 /* percpt data for schedulers */
433 struct kib_sched_info **kib_scheds;
436 #define IBLND_INIT_NOTHING 0
437 #define IBLND_INIT_DATA 1
438 #define IBLND_INIT_ALL 2
440 /************************************************************************
441 * IB Wire message format.
442 * These are sent in sender's byte order (i.e. receiver flips).
445 struct kib_connparams {
446 __u16 ibcp_queue_depth;
447 __u16 ibcp_max_frags;
448 __u32 ibcp_max_msg_size;
451 struct kib_immediate_msg {
452 struct lnet_hdr ibim_hdr; /* portals header */
453 char ibim_payload[0];/* piggy-backed payload */
456 struct kib_rdma_frag {
457 __u32 rf_nob; /* # bytes this frag */
458 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
461 struct kib_rdma_desc {
462 __u32 rd_key; /* local/remote key */
463 __u32 rd_nfrags; /* # fragments */
464 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
467 struct kib_putreq_msg {
468 struct lnet_hdr ibprm_hdr; /* portals header */
469 __u64 ibprm_cookie; /* opaque completion cookie */
472 struct kib_putack_msg {
473 __u64 ibpam_src_cookie; /* reflected completion cookie */
474 __u64 ibpam_dst_cookie; /* opaque completion cookie */
475 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
479 struct lnet_hdr ibgm_hdr; /* portals header */
480 __u64 ibgm_cookie; /* opaque completion cookie */
481 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
484 struct kib_completion_msg {
485 __u64 ibcm_cookie; /* opaque completion cookie */
486 __s32 ibcm_status; /* < 0 failure: >= 0 length */
490 /* First 2 fields fixed FOR ALL TIME */
491 __u32 ibm_magic; /* I'm an ibnal message */
492 __u16 ibm_version; /* this is my version number */
494 __u8 ibm_type; /* msg type */
495 __u8 ibm_credits; /* returned credits */
496 __u32 ibm_nob; /* # bytes in whole message */
497 __u32 ibm_cksum; /* checksum (0 == no checksum) */
498 __u64 ibm_srcnid; /* sender's NID */
499 __u64 ibm_srcstamp; /* sender's incarnation */
500 __u64 ibm_dstnid; /* destination's NID */
501 __u64 ibm_dststamp; /* destination's incarnation */
504 struct kib_connparams connparams;
505 struct kib_immediate_msg immediate;
506 struct kib_putreq_msg putreq;
507 struct kib_putack_msg putack;
508 struct kib_get_msg get;
509 struct kib_completion_msg completion;
513 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
515 #define IBLND_MSG_VERSION_1 0x11
516 #define IBLND_MSG_VERSION_2 0x12
517 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
519 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
520 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
521 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
522 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
523 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
524 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
525 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
526 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
527 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
528 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
531 __u32 ibr_magic; /* sender's magic */
532 __u16 ibr_version; /* sender's version */
533 __u8 ibr_why; /* reject reason */
534 __u8 ibr_padding; /* padding */
535 __u64 ibr_incarnation; /* incarnation of peer_ni */
536 struct kib_connparams ibr_cp; /* connection parameters */
539 /* connection rejection reasons */
540 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
541 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
542 #define IBLND_REJECT_FATAL 3 /* Anything else */
544 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
545 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
547 /* peer_ni's rdma frags doesn't match mine */
548 #define IBLND_REJECT_RDMA_FRAGS 6
549 /* peer_ni's msg queue size doesn't match mine */
550 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
551 #define IBLND_REJECT_INVALID_SRV_ID 8
553 /***********************************************************************/
555 struct kib_rx { /* receive message */
556 /* queue for attention */
557 struct list_head rx_list;
559 struct kib_conn *rx_conn;
560 /* # bytes received (-1 while posted) */
562 /* message buffer (host vaddr) */
563 struct kib_msg *rx_msg;
564 /* message buffer (I/O addr) */
566 /* for dma_unmap_single() */
567 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
568 /* receive work item... */
569 struct ib_recv_wr rx_wrq;
570 /* ...and its memory */
571 struct ib_sge rx_sge;
574 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
575 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
576 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
577 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
579 struct kib_tx { /* transmit message */
580 /* queue on idle_txs ibc_tx_queue etc. */
581 struct list_head tx_list;
583 struct kib_tx_pool *tx_pool;
585 struct kib_conn *tx_conn;
586 /* # tx callbacks outstanding */
588 /* queued for sending */
590 /* waiting for peer_ni */
592 /* LNET completion status */
594 /* health status of the transmit */
595 enum lnet_msg_hstatus tx_hstatus;
596 /* completion deadline */
598 /* completion cookie */
600 /* lnet msgs to finalize on completion */
601 struct lnet_msg *tx_lntmsg[2];
602 /* message buffer (host vaddr) */
603 struct kib_msg *tx_msg;
604 /* message buffer (I/O addr) */
606 /* for dma_unmap_single() */
607 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
608 /** sge for tx_msgaddr */
609 struct ib_sge tx_msgsge;
610 /* # send work items */
612 /* # used scatter/gather elements */
614 /* send work items... */
615 struct ib_rdma_wr *tx_wrq;
616 /* ...and their memory */
617 struct ib_sge *tx_sge;
618 /* rdma descriptor */
619 struct kib_rdma_desc *tx_rd;
620 /* # entries in... */
622 /* dma_map_sg descriptor */
623 struct scatterlist *tx_frags;
624 /* rdma phys page addrs */
626 /* gaps in fragments */
629 struct kib_fmr tx_fmr;
634 struct kib_connvars {
635 /* connection-in-progress variables */
636 struct kib_msg cv_msg;
640 /* scheduler information */
641 struct kib_sched_info *ibc_sched;
643 struct kib_peer_ni *ibc_peer;
645 struct kib_hca_dev *ibc_hdev;
646 /* stash on peer_ni's conn list */
647 struct list_head ibc_list;
648 /* schedule for attention */
649 struct list_head ibc_sched_list;
650 /* version of connection */
652 /* reconnect later */
653 __u16 ibc_reconnect:1;
654 /* which instance of the peer */
655 __u64 ibc_incarnation;
657 atomic_t ibc_refcount;
658 /* what's happening */
660 /* # uncompleted sends */
661 int ibc_nsends_posted;
662 /* # uncompleted NOOPs */
663 int ibc_noops_posted;
664 /* # credits I have */
666 /* # credits to return */
667 int ibc_outstanding_credits;
668 /* # ACK/DONE msg credits */
669 int ibc_reserved_credits;
670 /* set on comms error */
672 /* connections queue depth */
673 __u16 ibc_queue_depth;
674 /* connections max frags */
676 /* receive buffers owned */
677 unsigned int ibc_nrx:16;
678 /* scheduled for attention */
679 unsigned int ibc_scheduled:1;
680 /* CQ callback fired */
681 unsigned int ibc_ready:1;
682 /* time of last send */
683 ktime_t ibc_last_send;
684 /** link chain for kiblnd_check_conns only */
685 struct list_head ibc_connd_list;
686 /** rxs completed before ESTABLISHED */
687 struct list_head ibc_early_rxs;
688 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
689 struct list_head ibc_tx_noops;
690 /* sends that need a credit */
691 struct list_head ibc_tx_queue;
692 /* sends that don't need a credit */
693 struct list_head ibc_tx_queue_nocred;
694 /* sends that need to reserve an ACK/DONE msg */
695 struct list_head ibc_tx_queue_rsrvd;
696 /* active tx awaiting completion */
697 struct list_head ibc_active_txs;
698 /* zombie tx awaiting done */
699 struct list_head ibc_zombie_txs;
703 struct kib_rx *ibc_rxs;
704 /* premapped rx msg pages */
705 struct kib_pages *ibc_rx_pages;
708 struct rdma_cm_id *ibc_cmid;
709 /* completion queue */
710 struct ib_cq *ibc_cq;
712 /* in-progress connection state */
713 struct kib_connvars *ibc_connvars;
716 #define IBLND_CONN_INIT 0 /* being initialised */
717 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
718 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
719 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
720 #define IBLND_CONN_CLOSING 4 /* being closed */
721 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
724 /* stash on global peer_ni list */
725 struct list_head ibp_list;
726 /* who's on the other end(s) */
729 struct lnet_ni *ibp_ni;
730 /* all active connections */
731 struct list_head ibp_conns;
732 /* next connection to send on for round robin */
733 struct kib_conn *ibp_next_conn;
734 /* msgs waiting for a conn */
735 struct list_head ibp_tx_queue;
736 /* incarnation of peer_ni */
737 __u64 ibp_incarnation;
738 /* when (in seconds) I was last alive */
739 time64_t ibp_last_alive;
741 atomic_t ibp_refcount;
742 /* version of peer_ni */
744 /* current passive connection attempts */
745 unsigned short ibp_accepting;
746 /* current active connection attempts */
747 unsigned short ibp_connecting;
748 /* reconnect this peer_ni later */
749 unsigned char ibp_reconnecting;
750 /* counter of how many times we triggered a conn race */
751 unsigned char ibp_races;
752 /* # consecutive reconnection attempts to this peer */
753 unsigned int ibp_reconnected;
754 /* errno on closing this peer_ni */
756 /* max map_on_demand */
758 /* max_peer_credits */
759 __u16 ibp_queue_depth;
762 #ifndef HAVE_IB_INC_RKEY
764 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
765 * for calculating a new rkey for type 2 memory windows.
766 * @rkey - the rkey to increment.
768 static inline u32 ib_inc_rkey(u32 rkey)
770 const u32 mask = 0x000000ff;
771 return ((rkey + 1) & mask) | (rkey & ~mask);
775 extern struct kib_data kiblnd_data;
777 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
779 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
782 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
784 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
785 int concurrent_sends;
787 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
788 concurrent_sends = tunables->lnd_concurrent_sends;
790 if (version == IBLND_MSG_VERSION_1) {
791 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
792 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
794 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
795 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
798 return concurrent_sends;
802 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
804 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
805 atomic_inc(&hdev->ibh_ref);
809 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
811 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
812 if (atomic_dec_and_test(&hdev->ibh_ref))
813 kiblnd_hdev_destroy(hdev);
817 kiblnd_dev_can_failover(struct kib_dev *dev)
819 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
822 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
825 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
828 return dev->ibd_can_failover;
831 #define kiblnd_conn_addref(conn) \
833 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
834 (conn), atomic_read(&(conn)->ibc_refcount)); \
835 atomic_inc(&(conn)->ibc_refcount); \
838 #define kiblnd_conn_decref(conn) \
840 unsigned long flags; \
842 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
843 (conn), atomic_read(&(conn)->ibc_refcount)); \
844 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
845 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
846 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
847 list_add_tail(&(conn)->ibc_list, \
848 &kiblnd_data.kib_connd_zombies); \
849 wake_up(&kiblnd_data.kib_connd_waitq); \
850 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
854 #define kiblnd_peer_addref(peer_ni) \
856 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
857 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
858 atomic_read (&(peer_ni)->ibp_refcount)); \
859 atomic_inc(&(peer_ni)->ibp_refcount); \
862 #define kiblnd_peer_decref(peer_ni) \
864 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
865 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
866 atomic_read (&(peer_ni)->ibp_refcount)); \
867 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
868 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
869 kiblnd_destroy_peer(peer_ni); \
873 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
875 return peer_ni->ibp_connecting != 0 ||
876 peer_ni->ibp_reconnecting != 0 ||
877 peer_ni->ibp_accepting != 0;
881 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
883 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
886 static inline struct list_head *
887 kiblnd_nid2peerlist (lnet_nid_t nid)
890 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
892 return &kiblnd_data.kib_peers[hash];
896 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
898 /* Am I in the peer_ni hash table? */
899 return !list_empty(&peer_ni->ibp_list);
902 static inline struct kib_conn *
903 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
905 struct list_head *next;
907 LASSERT(!list_empty(&peer_ni->ibp_conns));
909 /* Advance to next connection, be sure to skip the head node */
910 if (!peer_ni->ibp_next_conn ||
911 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
912 next = peer_ni->ibp_conns.next;
914 next = peer_ni->ibp_next_conn->ibc_list.next;
915 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
917 return peer_ni->ibp_next_conn;
921 kiblnd_send_keepalive(struct kib_conn *conn)
923 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
925 return (*kiblnd_tunables.kib_keepalive > 0) &&
926 ktime_after(ktime_get(),
927 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
931 kiblnd_need_noop(struct kib_conn *conn)
933 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
934 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
936 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
937 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
939 if (conn->ibc_outstanding_credits <
940 IBLND_CREDITS_HIGHWATER(tunables, conn) &&
941 !kiblnd_send_keepalive(conn))
942 return 0; /* No need to send NOOP */
944 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
945 if (!list_empty(&conn->ibc_tx_queue_nocred))
946 return 0; /* NOOP can be piggybacked */
948 /* No tx to piggyback NOOP onto or no credit to send a tx */
949 return (list_empty(&conn->ibc_tx_queue) ||
950 conn->ibc_credits == 0);
953 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
954 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
955 conn->ibc_credits == 0) /* no credit */
958 if (conn->ibc_credits == 1 && /* last credit reserved for */
959 conn->ibc_outstanding_credits == 0) /* giving back credits */
962 /* No tx to piggyback NOOP onto or no credit to send a tx */
963 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
967 kiblnd_abort_receives(struct kib_conn *conn)
969 ib_modify_qp(conn->ibc_cmid->qp,
970 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
973 static inline const char *
974 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
976 if (q == &conn->ibc_tx_queue)
979 if (q == &conn->ibc_tx_queue_rsrvd)
980 return "tx_queue_rsrvd";
982 if (q == &conn->ibc_tx_queue_nocred)
983 return "tx_queue_nocred";
985 if (q == &conn->ibc_active_txs)
992 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
993 * lowest bits of the work request id to stash the work item type. */
995 #define IBLND_WID_INVAL 0
996 #define IBLND_WID_TX 1
997 #define IBLND_WID_RX 2
998 #define IBLND_WID_RDMA 3
999 #define IBLND_WID_MR 4
1000 #define IBLND_WID_MASK 7UL
1003 kiblnd_ptr2wreqid (void *ptr, int type)
1005 unsigned long lptr = (unsigned long)ptr;
1007 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1008 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1009 return (__u64)(lptr | type);
1012 static inline void *
1013 kiblnd_wreqid2ptr (__u64 wreqid)
1015 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1019 kiblnd_wreqid2type (__u64 wreqid)
1021 return (wreqid & IBLND_WID_MASK);
1025 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1027 conn->ibc_state = state;
1032 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1034 msg->ibm_type = type;
1035 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1039 kiblnd_rd_size(struct kib_rdma_desc *rd)
1044 for (i = size = 0; i < rd->rd_nfrags; i++)
1045 size += rd->rd_frags[i].rf_nob;
1051 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1053 return rd->rd_frags[index].rf_addr;
1057 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1059 return rd->rd_frags[index].rf_nob;
1063 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1069 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1071 if (nob < rd->rd_frags[index].rf_nob) {
1072 rd->rd_frags[index].rf_addr += nob;
1073 rd->rd_frags[index].rf_nob -= nob;
1082 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1084 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1085 msgtype == IBLND_MSG_PUT_ACK);
1087 return msgtype == IBLND_MSG_GET_REQ ?
1088 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1089 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1093 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1095 return ib_dma_mapping_error(dev, dma_addr);
1098 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1099 void *msg, size_t size,
1100 enum dma_data_direction direction)
1102 return ib_dma_map_single(dev, msg, size, direction);
1105 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1106 __u64 addr, size_t size,
1107 enum dma_data_direction direction)
1109 ib_dma_unmap_single(dev, addr, size, direction);
1112 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1113 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1115 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1116 struct scatterlist *sg, int nents,
1117 enum dma_data_direction direction)
1119 return ib_dma_map_sg(dev, sg, nents, direction);
1122 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1123 struct scatterlist *sg, int nents,
1124 enum dma_data_direction direction)
1126 ib_dma_unmap_sg(dev, sg, nents, direction);
1129 #ifndef HAVE_IB_SG_DMA_ADDRESS
1130 #include <linux/scatterlist.h>
1131 #define ib_sg_dma_address(dev, sg) sg_dma_address(sg)
1132 #define ib_sg_dma_len(dev, sg) sg_dma_len(sg)
1135 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1136 struct scatterlist *sg)
1138 return ib_sg_dma_address(dev, sg);
1141 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1142 struct scatterlist *sg)
1144 return ib_sg_dma_len(dev, sg);
1147 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1148 * right because OFED1.2 defines it as const, to use it we have to add
1149 * (void *) cast to overcome "const" */
1151 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1152 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1154 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1155 void kiblnd_map_rx_descs(struct kib_conn *conn);
1156 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1157 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1158 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1160 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1161 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1162 struct kib_fmr *fmr);
1163 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1165 int kiblnd_tunables_setup(struct lnet_ni *ni);
1166 int kiblnd_tunables_init(void);
1168 int kiblnd_connd (void *arg);
1169 int kiblnd_scheduler(void *arg);
1170 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1171 int kiblnd_failover_thread (void *arg);
1173 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1175 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1176 struct rdma_cm_event *event);
1177 int kiblnd_translate_mtu(int value);
1179 int kiblnd_dev_failover(struct kib_dev *dev, struct net *ns);
1180 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1182 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1183 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1184 void kiblnd_destroy_dev(struct kib_dev *dev);
1185 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1186 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1187 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1188 int version, u64 incarnation);
1189 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1191 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1192 struct rdma_cm_id *cmid,
1193 int state, int version);
1194 void kiblnd_destroy_conn(struct kib_conn *conn);
1195 void kiblnd_close_conn(struct kib_conn *conn, int error);
1196 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1198 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1199 void kiblnd_txlist_done(struct list_head *txlist, int status,
1200 enum lnet_msg_hstatus hstatus);
1202 void kiblnd_qp_event(struct ib_event *event, void *arg);
1203 void kiblnd_cq_event(struct ib_event *event, void *arg);
1204 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1206 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1207 int credits, lnet_nid_t dstnid, __u64 dststamp);
1208 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1209 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1211 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1212 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1213 int delayed, unsigned int niov, struct kvec *iov,
1214 lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,