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_min_reconnect_interval; /* first failed connection retry... */
92 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
93 int *kib_cksum; /* checksum struct kib_msg? */
94 int *kib_timeout; /* comms timeout (seconds) */
95 int *kib_keepalive; /* keepalive timeout (seconds) */
96 int *kib_ntx; /* # tx descs */
97 char **kib_default_ipif; /* default IPoIB interface */
99 int *kib_rnr_retry_count;
100 int *kib_ib_mtu; /* IB MTU */
101 int *kib_require_priv_port;/* accept only privileged ports */
102 int *kib_use_priv_port; /* use privileged port for active connect */
103 /* # threads on each CPT */
105 int *kib_wrq_sge; /* # sg elements per wrq */
106 int *kib_use_fastreg_gaps; /* enable discontiguous fastreg fragment support */
109 extern struct kib_tunables kiblnd_tunables;
111 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
112 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
114 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer_ni credits */
115 #define IBLND_CREDITS_MAX ((typeof(((struct kib_msg *) 0)->ibm_credits)) - 1) /* Max # of peer_ni credits */
117 /* when eagerly to return credits */
118 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
119 IBLND_CREDIT_HIGHWATER_V1 : \
120 t->lnd_peercredits_hiw)
122 #ifdef HAVE_RDMA_CREATE_ID_5ARG
123 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
127 # ifdef HAVE_RDMA_CREATE_ID_4ARG
128 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \
131 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
135 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
136 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
137 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
139 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
140 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
142 /************************/
143 /* derived constants... */
144 /* Pools (shared by connections on each CPT) */
145 /* These pools can grow at runtime, so don't need give a very large value */
146 #define IBLND_TX_POOL 256
147 #define IBLND_FMR_POOL 256
148 #define IBLND_FMR_POOL_FLUSH 192
150 /* RX messages (per connection) */
151 #define IBLND_RX_MSGS(c) \
152 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
153 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
154 #define IBLND_RX_MSG_PAGES(c) \
155 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
157 /* WRs and CQEs (per connection) */
158 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
160 /* 2 = LNet msg + Transfer chain */
161 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + kiblnd_send_wrs(c))
165 /* o2iblnd can run over aliased interface */
167 #define KIB_IFNAME_SIZE IFALIASZ
169 #define KIB_IFNAME_SIZE 256
173 IBLND_DEV_CAPS_FASTREG_ENABLED = BIT(0),
174 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT = BIT(1),
175 IBLND_DEV_CAPS_FMR_ENABLED = BIT(2),
179 struct list_head ibd_list; /* chain on kib_devs */
180 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
181 __u32 ibd_ifip; /* IPoIB interface IP */
182 /** IPoIB interface name */
183 char ibd_ifname[KIB_IFNAME_SIZE];
184 int ibd_nnets; /* # nets extant */
186 time64_t ibd_next_failover;
187 /* # failover failures */
188 int ibd_failed_failover;
189 /* failover in progress */
190 unsigned int ibd_failover;
191 /* IPoIB interface is a bonding master */
192 unsigned int ibd_can_failover;
193 struct list_head ibd_nets;
194 struct kib_hca_dev *ibd_hdev;
195 enum kib_dev_caps ibd_dev_caps;
199 struct rdma_cm_id *ibh_cmid; /* listener cmid */
200 struct ib_device *ibh_ibdev; /* IB device */
201 int ibh_page_shift; /* page shift of current HCA */
202 int ibh_page_size; /* page size of current HCA */
203 __u64 ibh_page_mask; /* page mask of current HCA */
204 int ibh_mr_shift; /* bits shift of max MR size */
205 __u64 ibh_mr_size; /* size of MR */
206 #ifdef HAVE_IB_GET_DMA_MR
207 struct ib_mr *ibh_mrs; /* global MR */
209 struct ib_pd *ibh_pd; /* PD */
210 struct kib_dev *ibh_dev; /* owner */
211 atomic_t ibh_ref; /* refcount */
214 /** # of seconds to keep pool alive */
215 #define IBLND_POOL_DEADLINE 300
216 /** # of seconds to retry if allocation failed */
217 #define IBLND_POOL_RETRY 1
220 int ibp_npages; /* # pages */
221 struct page *ibp_pages[0]; /* page array */
227 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
228 int inc, struct kib_pool **pp_po);
229 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
230 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
231 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
235 #define IBLND_POOL_NAME_LEN 32
240 /* network it belongs to */
241 struct kib_net *ps_net;
243 char ps_name[IBLND_POOL_NAME_LEN];
245 struct list_head ps_pool_list;
246 /* failed pool list */
247 struct list_head ps_failed_pool_list;
248 /* time stamp for retry if failed to allocate */
249 time64_t ps_next_retry;
250 /* is allocating new pool */
257 /* create a new pool */
258 kib_ps_pool_create_t ps_pool_create;
260 kib_ps_pool_destroy_t ps_pool_destroy;
261 /* initialize new allocated node */
262 kib_ps_node_init_t ps_node_init;
264 kib_ps_node_fini_t ps_node_fini;
268 /* chain on pool list */
269 struct list_head po_list;
270 /* pre-allocated node */
271 struct list_head po_free_list;
272 /* pool_set of this pool */
273 struct kib_poolset *po_owner;
274 /* deadline of this pool */
275 time64_t po_deadline;
276 /* # of elements in use */
278 /* pool is created on failed HCA */
280 /* # of pre-allocated elements */
284 struct kib_tx_poolset {
285 struct kib_poolset tps_poolset; /* pool-set */
286 __u64 tps_next_tx_cookie; /* cookie of TX */
290 struct kib_pool tpo_pool; /* pool */
291 struct kib_hca_dev *tpo_hdev; /* device for this pool */
292 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
293 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
296 struct kib_fmr_poolset {
297 spinlock_t fps_lock; /* serialize */
298 struct kib_net *fps_net; /* IB network */
299 struct list_head fps_pool_list; /* FMR pool list */
300 struct list_head fps_failed_pool_list; /* FMR pool list */
301 __u64 fps_version; /* validity stamp */
302 int fps_cpt; /* CPT id */
304 int fps_flush_trigger;
306 /* is allocating new pool */
308 /* time stamp for retry if failed to allocate */
309 time64_t fps_next_retry;
312 #ifndef HAVE_IB_RDMA_WR
314 struct ib_send_wr wr;
318 struct kib_fast_reg_descriptor { /* For fast registration */
319 struct list_head frd_list;
320 struct ib_rdma_wr frd_inv_wr;
321 #ifdef HAVE_IB_MAP_MR_SG
322 struct ib_reg_wr frd_fastreg_wr;
324 struct ib_rdma_wr frd_fastreg_wr;
325 struct ib_fast_reg_page_list *frd_frpl;
327 struct ib_mr *frd_mr;
331 struct kib_fmr_pool {
332 struct list_head fpo_list; /* chain on pool list */
333 struct kib_hca_dev *fpo_hdev; /* device for this pool */
334 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
337 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
339 struct { /* For fast registration */
340 struct list_head fpo_pool_list;
344 time64_t fpo_deadline; /* deadline of this pool */
345 int fpo_failed; /* fmr pool is failed */
346 int fpo_map_count; /* # of mapped FMR */
347 bool fpo_is_fmr; /* True if FMR pools allocated */
351 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
352 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
353 struct kib_fast_reg_descriptor *fmr_frd;
357 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
358 #define kib_fmr_pool_map(pool, pgs, n, iov) \
359 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
361 #define kib_fmr_pool_map(pool, pgs, n, iov) \
362 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
366 /* chain on struct kib_dev::ibd_nets */
367 struct list_head ibn_list;
368 __u64 ibn_incarnation;/* my epoch */
369 int ibn_init; /* initialisation state */
370 int ibn_shutdown; /* shutting down? */
372 atomic_t ibn_npeers; /* # peers extant */
373 atomic_t ibn_nconns; /* # connections extant */
375 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
376 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
378 struct kib_dev *ibn_dev; /* underlying IB device */
381 #define KIB_THREAD_SHIFT 16
382 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
383 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
384 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
386 struct kib_sched_info {
389 /* schedulers sleep here */
390 wait_queue_head_t ibs_waitq;
391 /* conns to check for rx completions */
392 struct list_head ibs_conns;
393 /* number of scheduler threads */
395 /* max allowed scheduler threads */
396 int ibs_nthreads_max;
397 int ibs_cpt; /* CPT id */
401 int kib_init; /* initialisation state */
402 int kib_shutdown; /* shut down? */
403 struct list_head kib_devs; /* IB devices extant */
404 /* list head of failed devices */
405 struct list_head kib_failed_devs;
406 /* schedulers sleep here */
407 wait_queue_head_t kib_failover_waitq;
408 atomic_t kib_nthreads; /* # live threads */
409 /* stabilize net/dev/peer_ni/conn ops */
410 rwlock_t kib_global_lock;
411 /* hash table of all my known peers */
412 struct list_head *kib_peers;
413 /* size of kib_peers */
414 int kib_peer_hash_size;
415 /* the connd task (serialisation assertions) */
417 /* connections to setup/teardown */
418 struct list_head kib_connd_conns;
419 /* connections with zero refcount */
420 struct list_head kib_connd_zombies;
421 /* connections to reconnect */
422 struct list_head kib_reconn_list;
423 /* peers wait for reconnection */
424 struct list_head kib_reconn_wait;
426 * The second that peers are pulled out from \a kib_reconn_wait
429 time64_t kib_reconn_sec;
430 /* connection daemon sleeps here */
431 wait_queue_head_t kib_connd_waitq;
432 spinlock_t kib_connd_lock; /* serialise */
433 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
434 /* percpt data for schedulers */
435 struct kib_sched_info **kib_scheds;
438 #define IBLND_INIT_NOTHING 0
439 #define IBLND_INIT_DATA 1
440 #define IBLND_INIT_ALL 2
442 /************************************************************************
443 * IB Wire message format.
444 * These are sent in sender's byte order (i.e. receiver flips).
447 struct kib_connparams {
448 __u16 ibcp_queue_depth;
449 __u16 ibcp_max_frags;
450 __u32 ibcp_max_msg_size;
453 struct kib_immediate_msg {
454 struct lnet_hdr ibim_hdr; /* portals header */
455 char ibim_payload[0];/* piggy-backed payload */
458 struct kib_rdma_frag {
459 __u32 rf_nob; /* # bytes this frag */
460 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
463 struct kib_rdma_desc {
464 __u32 rd_key; /* local/remote key */
465 __u32 rd_nfrags; /* # fragments */
466 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
469 struct kib_putreq_msg {
470 struct lnet_hdr ibprm_hdr; /* portals header */
471 __u64 ibprm_cookie; /* opaque completion cookie */
474 struct kib_putack_msg {
475 __u64 ibpam_src_cookie; /* reflected completion cookie */
476 __u64 ibpam_dst_cookie; /* opaque completion cookie */
477 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
481 struct lnet_hdr ibgm_hdr; /* portals header */
482 __u64 ibgm_cookie; /* opaque completion cookie */
483 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
486 struct kib_completion_msg {
487 __u64 ibcm_cookie; /* opaque completion cookie */
488 __s32 ibcm_status; /* < 0 failure: >= 0 length */
492 /* First 2 fields fixed FOR ALL TIME */
493 __u32 ibm_magic; /* I'm an ibnal message */
494 __u16 ibm_version; /* this is my version number */
496 __u8 ibm_type; /* msg type */
497 __u8 ibm_credits; /* returned credits */
498 __u32 ibm_nob; /* # bytes in whole message */
499 __u32 ibm_cksum; /* checksum (0 == no checksum) */
500 __u64 ibm_srcnid; /* sender's NID */
501 __u64 ibm_srcstamp; /* sender's incarnation */
502 __u64 ibm_dstnid; /* destination's NID */
503 __u64 ibm_dststamp; /* destination's incarnation */
506 struct kib_connparams connparams;
507 struct kib_immediate_msg immediate;
508 struct kib_putreq_msg putreq;
509 struct kib_putack_msg putack;
510 struct kib_get_msg get;
511 struct kib_completion_msg completion;
515 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
517 #define IBLND_MSG_VERSION_1 0x11
518 #define IBLND_MSG_VERSION_2 0x12
519 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
521 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
522 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
523 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
524 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
525 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
526 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
527 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
528 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
529 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
530 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
533 __u32 ibr_magic; /* sender's magic */
534 __u16 ibr_version; /* sender's version */
535 __u8 ibr_why; /* reject reason */
536 __u8 ibr_padding; /* padding */
537 __u64 ibr_incarnation; /* incarnation of peer_ni */
538 struct kib_connparams ibr_cp; /* connection parameters */
541 /* connection rejection reasons */
542 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
543 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
544 #define IBLND_REJECT_FATAL 3 /* Anything else */
546 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
547 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
549 /* peer_ni's rdma frags doesn't match mine */
550 #define IBLND_REJECT_RDMA_FRAGS 6
551 /* peer_ni's msg queue size doesn't match mine */
552 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
553 #define IBLND_REJECT_INVALID_SRV_ID 8
555 /***********************************************************************/
557 struct kib_rx { /* receive message */
558 /* queue for attention */
559 struct list_head rx_list;
561 struct kib_conn *rx_conn;
562 /* # bytes received (-1 while posted) */
564 /* completion status */
565 enum ib_wc_status rx_status;
566 /* message buffer (host vaddr) */
567 struct kib_msg *rx_msg;
568 /* message buffer (I/O addr) */
570 /* for dma_unmap_single() */
571 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
572 /* receive work item... */
573 struct ib_recv_wr rx_wrq;
574 /* ...and its memory */
575 struct ib_sge rx_sge;
578 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
579 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
580 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
581 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
583 struct kib_tx { /* transmit message */
584 /* queue on idle_txs ibc_tx_queue etc. */
585 struct list_head tx_list;
587 struct kib_tx_pool *tx_pool;
589 struct kib_conn *tx_conn;
590 /* # tx callbacks outstanding */
592 /* queued for sending */
594 /* waiting for peer_ni */
596 /* LNET completion status */
598 /* health status of the transmit */
599 enum lnet_msg_hstatus tx_hstatus;
600 /* completion deadline */
602 /* completion cookie */
604 /* lnet msgs to finalize on completion */
605 struct lnet_msg *tx_lntmsg[2];
606 /* message buffer (host vaddr) */
607 struct kib_msg *tx_msg;
608 /* message buffer (I/O addr) */
610 /* for dma_unmap_single() */
611 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
612 /** sge for tx_msgaddr */
613 struct ib_sge tx_msgsge;
614 /* # send work items */
616 /* # used scatter/gather elements */
618 /* send work items... */
619 struct ib_rdma_wr *tx_wrq;
620 /* ...and their memory */
621 struct ib_sge *tx_sge;
622 /* rdma descriptor */
623 struct kib_rdma_desc *tx_rd;
624 /* # entries in... */
626 /* dma_map_sg descriptor */
627 struct scatterlist *tx_frags;
628 /* rdma phys page addrs */
630 /* gaps in fragments */
633 struct kib_fmr tx_fmr;
638 struct kib_connvars {
639 /* connection-in-progress variables */
640 struct kib_msg cv_msg;
644 /* scheduler information */
645 struct kib_sched_info *ibc_sched;
647 struct kib_peer_ni *ibc_peer;
649 struct kib_hca_dev *ibc_hdev;
650 /* stash on peer_ni's conn list */
651 struct list_head ibc_list;
652 /* schedule for attention */
653 struct list_head ibc_sched_list;
654 /* version of connection */
656 /* reconnect later */
657 __u16 ibc_reconnect:1;
658 /* which instance of the peer */
659 __u64 ibc_incarnation;
661 atomic_t ibc_refcount;
662 /* what's happening */
664 /* # uncompleted sends */
665 int ibc_nsends_posted;
666 /* # uncompleted NOOPs */
667 int ibc_noops_posted;
668 /* # credits I have */
670 /* # credits to return */
671 int ibc_outstanding_credits;
672 /* # ACK/DONE msg credits */
673 int ibc_reserved_credits;
674 /* set on comms error */
676 /* connections queue depth */
677 __u16 ibc_queue_depth;
678 /* connections max frags */
680 /* receive buffers owned */
681 unsigned int ibc_nrx:16;
682 /* scheduled for attention */
683 unsigned int ibc_scheduled:1;
684 /* CQ callback fired */
685 unsigned int ibc_ready:1;
686 /* time of last send */
687 ktime_t ibc_last_send;
688 /** link chain for kiblnd_check_conns only */
689 struct list_head ibc_connd_list;
690 /** rxs completed before ESTABLISHED */
691 struct list_head ibc_early_rxs;
692 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
693 struct list_head ibc_tx_noops;
694 /* sends that need a credit */
695 struct list_head ibc_tx_queue;
696 /* sends that don't need a credit */
697 struct list_head ibc_tx_queue_nocred;
698 /* sends that need to reserve an ACK/DONE msg */
699 struct list_head ibc_tx_queue_rsrvd;
700 /* active tx awaiting completion */
701 struct list_head ibc_active_txs;
702 /* zombie tx awaiting done */
703 struct list_head ibc_zombie_txs;
707 struct kib_rx *ibc_rxs;
708 /* premapped rx msg pages */
709 struct kib_pages *ibc_rx_pages;
712 struct rdma_cm_id *ibc_cmid;
713 /* completion queue */
714 struct ib_cq *ibc_cq;
716 /* in-progress connection state */
717 struct kib_connvars *ibc_connvars;
720 #define IBLND_CONN_INIT 0 /* being initialised */
721 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
722 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
723 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
724 #define IBLND_CONN_CLOSING 4 /* being closed */
725 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
728 /* stash on global peer_ni list */
729 struct list_head ibp_list;
730 /* who's on the other end(s) */
733 struct lnet_ni *ibp_ni;
734 /* all active connections */
735 struct list_head ibp_conns;
736 /* next connection to send on for round robin */
737 struct kib_conn *ibp_next_conn;
738 /* msgs waiting for a conn */
739 struct list_head ibp_tx_queue;
740 /* incarnation of peer_ni */
741 __u64 ibp_incarnation;
742 /* when (in seconds) I was last alive */
743 time64_t ibp_last_alive;
745 atomic_t ibp_refcount;
746 /* version of peer_ni */
748 /* current passive connection attempts */
749 unsigned short ibp_accepting;
750 /* current active connection attempts */
751 unsigned short ibp_connecting;
752 /* reconnect this peer_ni later */
753 unsigned char ibp_reconnecting;
754 /* counter of how many times we triggered a conn race */
755 unsigned char ibp_races;
756 /* # consecutive reconnection attempts to this peer */
757 unsigned int ibp_reconnected;
758 /* errno on closing this peer_ni */
760 /* max map_on_demand */
762 /* max_peer_credits */
763 __u16 ibp_queue_depth;
766 #ifndef HAVE_IB_INC_RKEY
768 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
769 * for calculating a new rkey for type 2 memory windows.
770 * @rkey - the rkey to increment.
772 static inline u32 ib_inc_rkey(u32 rkey)
774 const u32 mask = 0x000000ff;
775 return ((rkey + 1) & mask) | (rkey & ~mask);
779 extern struct kib_data kiblnd_data;
781 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
783 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
786 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
788 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
789 int concurrent_sends;
791 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
792 concurrent_sends = tunables->lnd_concurrent_sends;
794 if (version == IBLND_MSG_VERSION_1) {
795 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
796 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
798 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
799 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
802 return concurrent_sends;
806 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
808 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
809 atomic_inc(&hdev->ibh_ref);
813 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
815 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
816 if (atomic_dec_and_test(&hdev->ibh_ref))
817 kiblnd_hdev_destroy(hdev);
821 kiblnd_dev_can_failover(struct kib_dev *dev)
823 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
826 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
829 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
832 return dev->ibd_can_failover;
835 #define kiblnd_conn_addref(conn) \
837 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
838 (conn), atomic_read(&(conn)->ibc_refcount)); \
839 atomic_inc(&(conn)->ibc_refcount); \
842 #define kiblnd_conn_decref(conn) \
844 unsigned long flags; \
846 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
847 (conn), atomic_read(&(conn)->ibc_refcount)); \
848 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
849 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
850 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
851 list_add_tail(&(conn)->ibc_list, \
852 &kiblnd_data.kib_connd_zombies); \
853 wake_up(&kiblnd_data.kib_connd_waitq); \
854 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
858 #define kiblnd_peer_addref(peer_ni) \
860 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
861 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
862 atomic_read (&(peer_ni)->ibp_refcount)); \
863 atomic_inc(&(peer_ni)->ibp_refcount); \
866 #define kiblnd_peer_decref(peer_ni) \
868 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
869 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
870 atomic_read (&(peer_ni)->ibp_refcount)); \
871 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
872 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
873 kiblnd_destroy_peer(peer_ni); \
877 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
879 return peer_ni->ibp_connecting != 0 ||
880 peer_ni->ibp_reconnecting != 0 ||
881 peer_ni->ibp_accepting != 0;
885 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
887 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
890 static inline struct list_head *
891 kiblnd_nid2peerlist (lnet_nid_t nid)
894 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
896 return &kiblnd_data.kib_peers[hash];
900 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
902 /* Am I in the peer_ni hash table? */
903 return !list_empty(&peer_ni->ibp_list);
906 static inline struct kib_conn *
907 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
909 struct list_head *next;
911 LASSERT(!list_empty(&peer_ni->ibp_conns));
913 /* Advance to next connection, be sure to skip the head node */
914 if (!peer_ni->ibp_next_conn ||
915 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
916 next = peer_ni->ibp_conns.next;
918 next = peer_ni->ibp_next_conn->ibc_list.next;
919 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
921 return peer_ni->ibp_next_conn;
925 kiblnd_send_keepalive(struct kib_conn *conn)
927 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
929 return (*kiblnd_tunables.kib_keepalive > 0) &&
930 ktime_after(ktime_get(),
931 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
935 kiblnd_need_noop(struct kib_conn *conn)
937 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
938 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
940 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
941 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
943 if (conn->ibc_outstanding_credits <
944 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
945 !kiblnd_send_keepalive(conn))
946 return 0; /* No need to send NOOP */
948 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
949 if (!list_empty(&conn->ibc_tx_queue_nocred))
950 return 0; /* NOOP can be piggybacked */
952 /* No tx to piggyback NOOP onto or no credit to send a tx */
953 return (list_empty(&conn->ibc_tx_queue) ||
954 conn->ibc_credits == 0);
957 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
958 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
959 conn->ibc_credits == 0) /* no credit */
962 if (conn->ibc_credits == 1 && /* last credit reserved for */
963 conn->ibc_outstanding_credits == 0) /* giving back credits */
966 /* No tx to piggyback NOOP onto or no credit to send a tx */
967 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
971 kiblnd_abort_receives(struct kib_conn *conn)
973 ib_modify_qp(conn->ibc_cmid->qp,
974 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
977 static inline const char *
978 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
980 if (q == &conn->ibc_tx_queue)
983 if (q == &conn->ibc_tx_queue_rsrvd)
984 return "tx_queue_rsrvd";
986 if (q == &conn->ibc_tx_queue_nocred)
987 return "tx_queue_nocred";
989 if (q == &conn->ibc_active_txs)
996 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
997 * lowest bits of the work request id to stash the work item type. */
999 #define IBLND_WID_INVAL 0
1000 #define IBLND_WID_TX 1
1001 #define IBLND_WID_RX 2
1002 #define IBLND_WID_RDMA 3
1003 #define IBLND_WID_MR 4
1004 #define IBLND_WID_MASK 7UL
1007 kiblnd_ptr2wreqid (void *ptr, int type)
1009 unsigned long lptr = (unsigned long)ptr;
1011 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1012 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1013 return (__u64)(lptr | type);
1016 static inline void *
1017 kiblnd_wreqid2ptr (__u64 wreqid)
1019 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1023 kiblnd_wreqid2type (__u64 wreqid)
1025 return (wreqid & IBLND_WID_MASK);
1029 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1031 conn->ibc_state = state;
1036 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1038 msg->ibm_type = type;
1039 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1043 kiblnd_rd_size(struct kib_rdma_desc *rd)
1048 for (i = size = 0; i < rd->rd_nfrags; i++)
1049 size += rd->rd_frags[i].rf_nob;
1055 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1057 return rd->rd_frags[index].rf_addr;
1061 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1063 return rd->rd_frags[index].rf_nob;
1067 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1073 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1075 if (nob < rd->rd_frags[index].rf_nob) {
1076 rd->rd_frags[index].rf_addr += nob;
1077 rd->rd_frags[index].rf_nob -= nob;
1086 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1088 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1089 msgtype == IBLND_MSG_PUT_ACK);
1091 return msgtype == IBLND_MSG_GET_REQ ?
1092 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1093 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1097 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1099 return ib_dma_mapping_error(dev, dma_addr);
1102 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1103 void *msg, size_t size,
1104 enum dma_data_direction direction)
1106 return ib_dma_map_single(dev, msg, size, direction);
1109 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1110 __u64 addr, size_t size,
1111 enum dma_data_direction direction)
1113 ib_dma_unmap_single(dev, addr, size, direction);
1116 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1117 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1119 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1120 struct scatterlist *sg, int nents,
1121 enum dma_data_direction direction)
1123 return ib_dma_map_sg(dev, sg, nents, direction);
1126 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1127 struct scatterlist *sg, int nents,
1128 enum dma_data_direction direction)
1130 ib_dma_unmap_sg(dev, sg, nents, direction);
1133 #ifndef HAVE_IB_SG_DMA_ADDRESS
1134 #include <linux/scatterlist.h>
1135 #define ib_sg_dma_address(dev, sg) sg_dma_address((dev), (sg))
1136 #define ib_sg_dma_len(dev, sg) sg_dma_len((dev), (sg))
1139 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1140 struct scatterlist *sg)
1142 return ib_sg_dma_address(dev, sg);
1145 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1146 struct scatterlist *sg)
1148 return ib_sg_dma_len(dev, sg);
1151 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1152 * right because OFED1.2 defines it as const, to use it we have to add
1153 * (void *) cast to overcome "const" */
1155 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1156 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1158 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1159 void kiblnd_map_rx_descs(struct kib_conn *conn);
1160 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1161 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1162 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1164 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1165 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1166 struct kib_fmr *fmr);
1167 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1169 int kiblnd_tunables_setup(struct lnet_ni *ni);
1170 int kiblnd_tunables_init(void);
1172 int kiblnd_connd (void *arg);
1173 int kiblnd_scheduler(void *arg);
1174 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1175 int kiblnd_failover_thread (void *arg);
1177 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1179 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1180 struct rdma_cm_event *event);
1181 int kiblnd_translate_mtu(int value);
1183 int kiblnd_dev_failover(struct kib_dev *dev);
1184 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1186 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1187 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1188 void kiblnd_destroy_dev(struct kib_dev *dev);
1189 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1190 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1191 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1192 int version, u64 incarnation);
1193 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1195 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1196 struct rdma_cm_id *cmid,
1197 int state, int version);
1198 void kiblnd_destroy_conn(struct kib_conn *conn);
1199 void kiblnd_close_conn(struct kib_conn *conn, int error);
1200 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1202 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1203 void kiblnd_txlist_done(struct list_head *txlist, int status,
1204 enum lnet_msg_hstatus hstatus);
1206 void kiblnd_qp_event(struct ib_event *event, void *arg);
1207 void kiblnd_cq_event(struct ib_event *event, void *arg);
1208 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1210 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1211 int credits, lnet_nid_t dstnid, __u64 dststamp);
1212 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1213 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1215 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1216 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1217 int delayed, unsigned int niov, struct kvec *iov,
1218 lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,