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(ns, cb, dev, ps, qpt) rdma_create_id(ns, cb, \
127 # ifdef HAVE_RDMA_CREATE_ID_4ARG
128 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) rdma_create_id(cb, dev, \
131 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) rdma_create_id(cb, dev, \
136 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
137 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
138 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
140 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
141 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
143 /************************/
144 /* derived constants... */
145 /* Pools (shared by connections on each CPT) */
146 /* These pools can grow at runtime, so don't need give a very large value */
147 #define IBLND_TX_POOL 256
148 #define IBLND_FMR_POOL 256
149 #define IBLND_FMR_POOL_FLUSH 192
151 /* RX messages (per connection) */
152 #define IBLND_RX_MSGS(c) \
153 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
154 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
155 #define IBLND_RX_MSG_PAGES(c) \
156 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
158 /* WRs and CQEs (per connection) */
159 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
161 /* 2 = LNet msg + Transfer chain */
162 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + kiblnd_send_wrs(c))
166 /* o2iblnd can run over aliased interface */
168 #define KIB_IFNAME_SIZE IFALIASZ
170 #define KIB_IFNAME_SIZE 256
174 IBLND_DEV_CAPS_FASTREG_ENABLED = BIT(0),
175 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT = BIT(1),
176 IBLND_DEV_CAPS_FMR_ENABLED = BIT(2),
180 struct list_head ibd_list; /* chain on kib_devs */
181 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
182 __u32 ibd_ifip; /* IPoIB interface IP */
183 /** IPoIB interface name */
184 char ibd_ifname[KIB_IFNAME_SIZE];
185 int ibd_nnets; /* # nets extant */
187 time64_t ibd_next_failover;
188 /* # failover failures */
189 int ibd_failed_failover;
190 /* failover in progress */
191 unsigned int ibd_failover;
192 /* IPoIB interface is a bonding master */
193 unsigned int ibd_can_failover;
194 struct list_head ibd_nets;
195 struct kib_hca_dev *ibd_hdev;
196 enum kib_dev_caps ibd_dev_caps;
200 struct rdma_cm_id *ibh_cmid; /* listener cmid */
201 struct ib_device *ibh_ibdev; /* IB device */
202 int ibh_page_shift; /* page shift of current HCA */
203 int ibh_page_size; /* page size of current HCA */
204 __u64 ibh_page_mask; /* page mask of current HCA */
205 int ibh_mr_shift; /* bits shift of max MR size */
206 __u64 ibh_mr_size; /* size of MR */
207 #ifdef HAVE_IB_GET_DMA_MR
208 struct ib_mr *ibh_mrs; /* global MR */
210 struct ib_pd *ibh_pd; /* PD */
211 struct kib_dev *ibh_dev; /* owner */
212 atomic_t ibh_ref; /* refcount */
215 /** # of seconds to keep pool alive */
216 #define IBLND_POOL_DEADLINE 300
217 /** # of seconds to retry if allocation failed */
218 #define IBLND_POOL_RETRY 1
221 int ibp_npages; /* # pages */
222 struct page *ibp_pages[0]; /* page array */
228 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
229 int inc, struct kib_pool **pp_po);
230 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
231 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
232 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
236 #define IBLND_POOL_NAME_LEN 32
241 /* network it belongs to */
242 struct kib_net *ps_net;
244 char ps_name[IBLND_POOL_NAME_LEN];
246 struct list_head ps_pool_list;
247 /* failed pool list */
248 struct list_head ps_failed_pool_list;
249 /* time stamp for retry if failed to allocate */
250 time64_t ps_next_retry;
251 /* is allocating new pool */
258 /* create a new pool */
259 kib_ps_pool_create_t ps_pool_create;
261 kib_ps_pool_destroy_t ps_pool_destroy;
262 /* initialize new allocated node */
263 kib_ps_node_init_t ps_node_init;
265 kib_ps_node_fini_t ps_node_fini;
269 /* chain on pool list */
270 struct list_head po_list;
271 /* pre-allocated node */
272 struct list_head po_free_list;
273 /* pool_set of this pool */
274 struct kib_poolset *po_owner;
275 /* deadline of this pool */
276 time64_t po_deadline;
277 /* # of elements in use */
279 /* pool is created on failed HCA */
281 /* # of pre-allocated elements */
285 struct kib_tx_poolset {
286 struct kib_poolset tps_poolset; /* pool-set */
287 __u64 tps_next_tx_cookie; /* cookie of TX */
291 struct kib_pool tpo_pool; /* pool */
292 struct kib_hca_dev *tpo_hdev; /* device for this pool */
293 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
294 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
297 struct kib_fmr_poolset {
298 spinlock_t fps_lock; /* serialize */
299 struct kib_net *fps_net; /* IB network */
300 struct list_head fps_pool_list; /* FMR pool list */
301 struct list_head fps_failed_pool_list; /* FMR pool list */
302 __u64 fps_version; /* validity stamp */
303 int fps_cpt; /* CPT id */
305 int fps_flush_trigger;
307 /* is allocating new pool */
309 /* time stamp for retry if failed to allocate */
310 time64_t fps_next_retry;
313 #ifndef HAVE_IB_RDMA_WR
315 struct ib_send_wr wr;
319 struct kib_fast_reg_descriptor { /* For fast registration */
320 struct list_head frd_list;
321 struct ib_rdma_wr frd_inv_wr;
322 #ifdef HAVE_IB_MAP_MR_SG
323 struct ib_reg_wr frd_fastreg_wr;
325 struct ib_rdma_wr frd_fastreg_wr;
326 struct ib_fast_reg_page_list *frd_frpl;
328 struct ib_mr *frd_mr;
332 struct kib_fmr_pool {
333 struct list_head fpo_list; /* chain on pool list */
334 struct kib_hca_dev *fpo_hdev; /* device for this pool */
335 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
338 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
340 struct { /* For fast registration */
341 struct list_head fpo_pool_list;
345 time64_t fpo_deadline; /* deadline of this pool */
346 int fpo_failed; /* fmr pool is failed */
347 int fpo_map_count; /* # of mapped FMR */
348 bool fpo_is_fmr; /* True if FMR pools allocated */
352 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
353 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
354 struct kib_fast_reg_descriptor *fmr_frd;
358 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
359 #define kib_fmr_pool_map(pool, pgs, n, iov) \
360 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
362 #define kib_fmr_pool_map(pool, pgs, n, iov) \
363 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
367 /* chain on struct kib_dev::ibd_nets */
368 struct list_head ibn_list;
369 __u64 ibn_incarnation;/* my epoch */
370 int ibn_init; /* initialisation state */
371 int ibn_shutdown; /* shutting down? */
373 atomic_t ibn_npeers; /* # peers extant */
374 atomic_t ibn_nconns; /* # connections extant */
376 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
377 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
379 struct kib_dev *ibn_dev; /* underlying IB device */
382 #define KIB_THREAD_SHIFT 16
383 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
384 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
385 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
387 struct kib_sched_info {
390 /* schedulers sleep here */
391 wait_queue_head_t ibs_waitq;
392 /* conns to check for rx completions */
393 struct list_head ibs_conns;
394 /* number of scheduler threads */
396 /* max allowed scheduler threads */
397 int ibs_nthreads_max;
398 int ibs_cpt; /* CPT id */
402 int kib_init; /* initialisation state */
403 int kib_shutdown; /* shut down? */
404 struct list_head kib_devs; /* IB devices extant */
405 /* list head of failed devices */
406 struct list_head kib_failed_devs;
407 /* schedulers sleep here */
408 wait_queue_head_t kib_failover_waitq;
409 atomic_t kib_nthreads; /* # live threads */
410 /* stabilize net/dev/peer_ni/conn ops */
411 rwlock_t kib_global_lock;
412 /* hash table of all my known peers */
413 struct list_head *kib_peers;
414 /* size of kib_peers */
415 int kib_peer_hash_size;
416 /* the connd task (serialisation assertions) */
418 /* connections to setup/teardown */
419 struct list_head kib_connd_conns;
420 /* connections with zero refcount */
421 struct list_head kib_connd_zombies;
422 /* connections to reconnect */
423 struct list_head kib_reconn_list;
424 /* peers wait for reconnection */
425 struct list_head kib_reconn_wait;
427 * The second that peers are pulled out from \a kib_reconn_wait
430 time64_t kib_reconn_sec;
431 /* connection daemon sleeps here */
432 wait_queue_head_t kib_connd_waitq;
433 spinlock_t kib_connd_lock; /* serialise */
434 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
435 /* percpt data for schedulers */
436 struct kib_sched_info **kib_scheds;
439 #define IBLND_INIT_NOTHING 0
440 #define IBLND_INIT_DATA 1
441 #define IBLND_INIT_ALL 2
443 /************************************************************************
444 * IB Wire message format.
445 * These are sent in sender's byte order (i.e. receiver flips).
448 struct kib_connparams {
449 __u16 ibcp_queue_depth;
450 __u16 ibcp_max_frags;
451 __u32 ibcp_max_msg_size;
454 struct kib_immediate_msg {
455 struct lnet_hdr ibim_hdr; /* portals header */
456 char ibim_payload[0];/* piggy-backed payload */
459 struct kib_rdma_frag {
460 __u32 rf_nob; /* # bytes this frag */
461 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
464 struct kib_rdma_desc {
465 __u32 rd_key; /* local/remote key */
466 __u32 rd_nfrags; /* # fragments */
467 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
470 struct kib_putreq_msg {
471 struct lnet_hdr ibprm_hdr; /* portals header */
472 __u64 ibprm_cookie; /* opaque completion cookie */
475 struct kib_putack_msg {
476 __u64 ibpam_src_cookie; /* reflected completion cookie */
477 __u64 ibpam_dst_cookie; /* opaque completion cookie */
478 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
482 struct lnet_hdr ibgm_hdr; /* portals header */
483 __u64 ibgm_cookie; /* opaque completion cookie */
484 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
487 struct kib_completion_msg {
488 __u64 ibcm_cookie; /* opaque completion cookie */
489 __s32 ibcm_status; /* < 0 failure: >= 0 length */
493 /* First 2 fields fixed FOR ALL TIME */
494 __u32 ibm_magic; /* I'm an ibnal message */
495 __u16 ibm_version; /* this is my version number */
497 __u8 ibm_type; /* msg type */
498 __u8 ibm_credits; /* returned credits */
499 __u32 ibm_nob; /* # bytes in whole message */
500 __u32 ibm_cksum; /* checksum (0 == no checksum) */
501 __u64 ibm_srcnid; /* sender's NID */
502 __u64 ibm_srcstamp; /* sender's incarnation */
503 __u64 ibm_dstnid; /* destination's NID */
504 __u64 ibm_dststamp; /* destination's incarnation */
507 struct kib_connparams connparams;
508 struct kib_immediate_msg immediate;
509 struct kib_putreq_msg putreq;
510 struct kib_putack_msg putack;
511 struct kib_get_msg get;
512 struct kib_completion_msg completion;
516 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
518 #define IBLND_MSG_VERSION_1 0x11
519 #define IBLND_MSG_VERSION_2 0x12
520 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
522 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
523 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
524 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
525 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
526 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
527 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
528 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
529 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
530 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
531 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
534 __u32 ibr_magic; /* sender's magic */
535 __u16 ibr_version; /* sender's version */
536 __u8 ibr_why; /* reject reason */
537 __u8 ibr_padding; /* padding */
538 __u64 ibr_incarnation; /* incarnation of peer_ni */
539 struct kib_connparams ibr_cp; /* connection parameters */
542 /* connection rejection reasons */
543 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
544 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
545 #define IBLND_REJECT_FATAL 3 /* Anything else */
547 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
548 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
550 /* peer_ni's rdma frags doesn't match mine */
551 #define IBLND_REJECT_RDMA_FRAGS 6
552 /* peer_ni's msg queue size doesn't match mine */
553 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
554 #define IBLND_REJECT_INVALID_SRV_ID 8
556 /***********************************************************************/
558 struct kib_rx { /* receive message */
559 /* queue for attention */
560 struct list_head rx_list;
562 struct kib_conn *rx_conn;
563 /* # bytes received (-1 while posted) */
565 /* completion status */
566 enum ib_wc_status rx_status;
567 /* message buffer (host vaddr) */
568 struct kib_msg *rx_msg;
569 /* message buffer (I/O addr) */
571 /* for dma_unmap_single() */
572 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
573 /* receive work item... */
574 struct ib_recv_wr rx_wrq;
575 /* ...and its memory */
576 struct ib_sge rx_sge;
579 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
580 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
581 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
582 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
584 struct kib_tx { /* transmit message */
585 /* queue on idle_txs ibc_tx_queue etc. */
586 struct list_head tx_list;
588 struct kib_tx_pool *tx_pool;
590 struct kib_conn *tx_conn;
591 /* # tx callbacks outstanding */
593 /* queued for sending */
595 /* waiting for peer_ni */
597 /* LNET completion status */
599 /* health status of the transmit */
600 enum lnet_msg_hstatus tx_hstatus;
601 /* completion deadline */
603 /* completion cookie */
605 /* lnet msgs to finalize on completion */
606 struct lnet_msg *tx_lntmsg[2];
607 /* message buffer (host vaddr) */
608 struct kib_msg *tx_msg;
609 /* message buffer (I/O addr) */
611 /* for dma_unmap_single() */
612 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
613 /** sge for tx_msgaddr */
614 struct ib_sge tx_msgsge;
615 /* # send work items */
617 /* # used scatter/gather elements */
619 /* send work items... */
620 struct ib_rdma_wr *tx_wrq;
621 /* ...and their memory */
622 struct ib_sge *tx_sge;
623 /* rdma descriptor */
624 struct kib_rdma_desc *tx_rd;
625 /* # entries in... */
627 /* dma_map_sg descriptor */
628 struct scatterlist *tx_frags;
629 /* rdma phys page addrs */
631 /* gaps in fragments */
634 struct kib_fmr tx_fmr;
639 struct kib_connvars {
640 /* connection-in-progress variables */
641 struct kib_msg cv_msg;
645 /* scheduler information */
646 struct kib_sched_info *ibc_sched;
648 struct kib_peer_ni *ibc_peer;
650 struct kib_hca_dev *ibc_hdev;
651 /* stash on peer_ni's conn list */
652 struct list_head ibc_list;
653 /* schedule for attention */
654 struct list_head ibc_sched_list;
655 /* version of connection */
657 /* reconnect later */
658 __u16 ibc_reconnect:1;
659 /* which instance of the peer */
660 __u64 ibc_incarnation;
662 atomic_t ibc_refcount;
663 /* what's happening */
665 /* # uncompleted sends */
666 int ibc_nsends_posted;
667 /* # uncompleted NOOPs */
668 int ibc_noops_posted;
669 /* # credits I have */
671 /* # credits to return */
672 int ibc_outstanding_credits;
673 /* # ACK/DONE msg credits */
674 int ibc_reserved_credits;
675 /* set on comms error */
677 /* connections queue depth */
678 __u16 ibc_queue_depth;
679 /* connections max frags */
681 /* receive buffers owned */
682 unsigned int ibc_nrx:16;
683 /* scheduled for attention */
684 unsigned int ibc_scheduled:1;
685 /* CQ callback fired */
686 unsigned int ibc_ready:1;
687 /* time of last send */
688 ktime_t ibc_last_send;
689 /** link chain for kiblnd_check_conns only */
690 struct list_head ibc_connd_list;
691 /** rxs completed before ESTABLISHED */
692 struct list_head ibc_early_rxs;
693 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
694 struct list_head ibc_tx_noops;
695 /* sends that need a credit */
696 struct list_head ibc_tx_queue;
697 /* sends that don't need a credit */
698 struct list_head ibc_tx_queue_nocred;
699 /* sends that need to reserve an ACK/DONE msg */
700 struct list_head ibc_tx_queue_rsrvd;
701 /* active tx awaiting completion */
702 struct list_head ibc_active_txs;
703 /* zombie tx awaiting done */
704 struct list_head ibc_zombie_txs;
708 struct kib_rx *ibc_rxs;
709 /* premapped rx msg pages */
710 struct kib_pages *ibc_rx_pages;
713 struct rdma_cm_id *ibc_cmid;
714 /* completion queue */
715 struct ib_cq *ibc_cq;
717 /* in-progress connection state */
718 struct kib_connvars *ibc_connvars;
721 #define IBLND_CONN_INIT 0 /* being initialised */
722 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
723 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
724 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
725 #define IBLND_CONN_CLOSING 4 /* being closed */
726 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
729 /* stash on global peer_ni list */
730 struct list_head ibp_list;
731 /* who's on the other end(s) */
734 struct lnet_ni *ibp_ni;
735 /* all active connections */
736 struct list_head ibp_conns;
737 /* next connection to send on for round robin */
738 struct kib_conn *ibp_next_conn;
739 /* msgs waiting for a conn */
740 struct list_head ibp_tx_queue;
741 /* incarnation of peer_ni */
742 __u64 ibp_incarnation;
743 /* when (in seconds) I was last alive */
744 time64_t ibp_last_alive;
746 atomic_t ibp_refcount;
747 /* version of peer_ni */
749 /* current passive connection attempts */
750 unsigned short ibp_accepting;
751 /* current active connection attempts */
752 unsigned short ibp_connecting;
753 /* reconnect this peer_ni later */
754 unsigned char ibp_reconnecting;
755 /* counter of how many times we triggered a conn race */
756 unsigned char ibp_races;
757 /* # consecutive reconnection attempts to this peer */
758 unsigned int ibp_reconnected;
759 /* errno on closing this peer_ni */
761 /* max map_on_demand */
763 /* max_peer_credits */
764 __u16 ibp_queue_depth;
767 #ifndef HAVE_IB_INC_RKEY
769 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
770 * for calculating a new rkey for type 2 memory windows.
771 * @rkey - the rkey to increment.
773 static inline u32 ib_inc_rkey(u32 rkey)
775 const u32 mask = 0x000000ff;
776 return ((rkey + 1) & mask) | (rkey & ~mask);
780 extern struct kib_data kiblnd_data;
782 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
784 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
787 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
789 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
790 int concurrent_sends;
792 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
793 concurrent_sends = tunables->lnd_concurrent_sends;
795 if (version == IBLND_MSG_VERSION_1) {
796 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
797 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
799 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
800 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
803 return concurrent_sends;
807 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
809 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
810 atomic_inc(&hdev->ibh_ref);
814 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
816 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
817 if (atomic_dec_and_test(&hdev->ibh_ref))
818 kiblnd_hdev_destroy(hdev);
822 kiblnd_dev_can_failover(struct kib_dev *dev)
824 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
827 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
830 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
833 return dev->ibd_can_failover;
836 #define kiblnd_conn_addref(conn) \
838 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
839 (conn), atomic_read(&(conn)->ibc_refcount)); \
840 atomic_inc(&(conn)->ibc_refcount); \
843 #define kiblnd_conn_decref(conn) \
845 unsigned long flags; \
847 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
848 (conn), atomic_read(&(conn)->ibc_refcount)); \
849 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
850 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
851 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
852 list_add_tail(&(conn)->ibc_list, \
853 &kiblnd_data.kib_connd_zombies); \
854 wake_up(&kiblnd_data.kib_connd_waitq); \
855 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
859 #define kiblnd_peer_addref(peer_ni) \
861 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
862 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
863 atomic_read (&(peer_ni)->ibp_refcount)); \
864 atomic_inc(&(peer_ni)->ibp_refcount); \
867 #define kiblnd_peer_decref(peer_ni) \
869 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
870 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
871 atomic_read (&(peer_ni)->ibp_refcount)); \
872 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
873 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
874 kiblnd_destroy_peer(peer_ni); \
878 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
880 return peer_ni->ibp_connecting != 0 ||
881 peer_ni->ibp_reconnecting != 0 ||
882 peer_ni->ibp_accepting != 0;
886 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
888 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
891 static inline struct list_head *
892 kiblnd_nid2peerlist (lnet_nid_t nid)
895 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
897 return &kiblnd_data.kib_peers[hash];
901 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
903 /* Am I in the peer_ni hash table? */
904 return !list_empty(&peer_ni->ibp_list);
907 static inline struct kib_conn *
908 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
910 struct list_head *next;
912 LASSERT(!list_empty(&peer_ni->ibp_conns));
914 /* Advance to next connection, be sure to skip the head node */
915 if (!peer_ni->ibp_next_conn ||
916 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
917 next = peer_ni->ibp_conns.next;
919 next = peer_ni->ibp_next_conn->ibc_list.next;
920 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
922 return peer_ni->ibp_next_conn;
926 kiblnd_send_keepalive(struct kib_conn *conn)
928 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
930 return (*kiblnd_tunables.kib_keepalive > 0) &&
931 ktime_after(ktime_get(),
932 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
936 kiblnd_need_noop(struct kib_conn *conn)
938 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
939 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
941 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
942 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
944 if (conn->ibc_outstanding_credits <
945 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
946 !kiblnd_send_keepalive(conn))
947 return 0; /* No need to send NOOP */
949 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
950 if (!list_empty(&conn->ibc_tx_queue_nocred))
951 return 0; /* NOOP can be piggybacked */
953 /* No tx to piggyback NOOP onto or no credit to send a tx */
954 return (list_empty(&conn->ibc_tx_queue) ||
955 conn->ibc_credits == 0);
958 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
959 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
960 conn->ibc_credits == 0) /* no credit */
963 if (conn->ibc_credits == 1 && /* last credit reserved for */
964 conn->ibc_outstanding_credits == 0) /* giving back credits */
967 /* No tx to piggyback NOOP onto or no credit to send a tx */
968 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
972 kiblnd_abort_receives(struct kib_conn *conn)
974 ib_modify_qp(conn->ibc_cmid->qp,
975 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
978 static inline const char *
979 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
981 if (q == &conn->ibc_tx_queue)
984 if (q == &conn->ibc_tx_queue_rsrvd)
985 return "tx_queue_rsrvd";
987 if (q == &conn->ibc_tx_queue_nocred)
988 return "tx_queue_nocred";
990 if (q == &conn->ibc_active_txs)
997 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
998 * lowest bits of the work request id to stash the work item type. */
1000 #define IBLND_WID_INVAL 0
1001 #define IBLND_WID_TX 1
1002 #define IBLND_WID_RX 2
1003 #define IBLND_WID_RDMA 3
1004 #define IBLND_WID_MR 4
1005 #define IBLND_WID_MASK 7UL
1008 kiblnd_ptr2wreqid (void *ptr, int type)
1010 unsigned long lptr = (unsigned long)ptr;
1012 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1013 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1014 return (__u64)(lptr | type);
1017 static inline void *
1018 kiblnd_wreqid2ptr (__u64 wreqid)
1020 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1024 kiblnd_wreqid2type (__u64 wreqid)
1026 return (wreqid & IBLND_WID_MASK);
1030 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1032 conn->ibc_state = state;
1037 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1039 msg->ibm_type = type;
1040 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1044 kiblnd_rd_size(struct kib_rdma_desc *rd)
1049 for (i = size = 0; i < rd->rd_nfrags; i++)
1050 size += rd->rd_frags[i].rf_nob;
1056 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1058 return rd->rd_frags[index].rf_addr;
1062 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1064 return rd->rd_frags[index].rf_nob;
1068 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1074 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1076 if (nob < rd->rd_frags[index].rf_nob) {
1077 rd->rd_frags[index].rf_addr += nob;
1078 rd->rd_frags[index].rf_nob -= nob;
1087 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1089 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1090 msgtype == IBLND_MSG_PUT_ACK);
1092 return msgtype == IBLND_MSG_GET_REQ ?
1093 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1094 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1098 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1100 return ib_dma_mapping_error(dev, dma_addr);
1103 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1104 void *msg, size_t size,
1105 enum dma_data_direction direction)
1107 return ib_dma_map_single(dev, msg, size, direction);
1110 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1111 __u64 addr, size_t size,
1112 enum dma_data_direction direction)
1114 ib_dma_unmap_single(dev, addr, size, direction);
1117 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1118 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1120 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1121 struct scatterlist *sg, int nents,
1122 enum dma_data_direction direction)
1124 return ib_dma_map_sg(dev, sg, nents, direction);
1127 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1128 struct scatterlist *sg, int nents,
1129 enum dma_data_direction direction)
1131 ib_dma_unmap_sg(dev, sg, nents, direction);
1134 #ifndef HAVE_IB_SG_DMA_ADDRESS
1135 #include <linux/scatterlist.h>
1136 #define ib_sg_dma_address(dev, sg) sg_dma_address((dev), (sg))
1137 #define ib_sg_dma_len(dev, sg) sg_dma_len((dev), (sg))
1140 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1141 struct scatterlist *sg)
1143 return ib_sg_dma_address(dev, sg);
1146 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1147 struct scatterlist *sg)
1149 return ib_sg_dma_len(dev, sg);
1152 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1153 * right because OFED1.2 defines it as const, to use it we have to add
1154 * (void *) cast to overcome "const" */
1156 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1157 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1159 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1160 void kiblnd_map_rx_descs(struct kib_conn *conn);
1161 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1162 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1163 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1165 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1166 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1167 struct kib_fmr *fmr);
1168 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1170 int kiblnd_tunables_setup(struct lnet_ni *ni);
1171 int kiblnd_tunables_init(void);
1173 int kiblnd_connd (void *arg);
1174 int kiblnd_scheduler(void *arg);
1175 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1176 int kiblnd_failover_thread (void *arg);
1178 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1180 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1181 struct rdma_cm_event *event);
1182 int kiblnd_translate_mtu(int value);
1184 int kiblnd_dev_failover(struct kib_dev *dev, struct net *ns);
1185 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1187 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1188 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1189 void kiblnd_destroy_dev(struct kib_dev *dev);
1190 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1191 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1192 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1193 int version, u64 incarnation);
1194 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1196 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1197 struct rdma_cm_id *cmid,
1198 int state, int version);
1199 void kiblnd_destroy_conn(struct kib_conn *conn);
1200 void kiblnd_close_conn(struct kib_conn *conn, int error);
1201 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1203 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1204 void kiblnd_txlist_done(struct list_head *txlist, int status,
1205 enum lnet_msg_hstatus hstatus);
1207 void kiblnd_qp_event(struct ib_event *event, void *arg);
1208 void kiblnd_cq_event(struct ib_event *event, void *arg);
1209 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1211 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1212 int credits, lnet_nid_t dstnid, __u64 dststamp);
1213 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1214 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1216 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1217 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1218 int delayed, unsigned int niov, struct kvec *iov,
1219 lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,