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
44 /* MOFED has its own bitmap_alloc backport */
45 #define HAVE_BITMAP_ALLOC 1
49 #include <linux/module.h>
50 #include <linux/kernel.h>
51 #include <linux/kthread.h>
53 #include <linux/string.h>
54 #include <linux/stat.h>
55 #include <linux/errno.h>
56 #include <linux/unistd.h>
57 #include <linux/uio.h>
59 #include <asm/uaccess.h>
62 #include <linux/init.h>
64 #include <linux/file.h>
65 #include <linux/stat.h>
66 #include <linux/list.h>
67 #include <linux/kmod.h>
68 #include <linux/sysctl.h>
69 #include <linux/pci.h>
74 #include <rdma/rdma_cm.h>
75 #include <rdma/ib_cm.h>
76 #include <rdma/ib_verbs.h>
77 #include <rdma/ib_fmr_pool.h>
79 #define DEBUG_SUBSYSTEM S_LND
81 #include <libcfs/libcfs.h>
82 #include <lnet/lib-lnet.h>
84 #define IBLND_PEER_HASH_SIZE 101 /* # peer_ni lists */
85 /* # scheduler loops before reschedule */
86 #define IBLND_RESCHED 100
88 #define IBLND_N_SCHED 2
89 #define IBLND_N_SCHED_HIGH 4
92 int *kib_dev_failover; /* HCA failover */
93 unsigned int *kib_service; /* IB service number */
94 int *kib_cksum; /* checksum struct kib_msg? */
95 int *kib_timeout; /* comms timeout (seconds) */
96 int *kib_keepalive; /* keepalive timeout (seconds) */
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, conn) ((conn->ibc_version) == IBLND_MSG_VERSION_1 ? \
119 IBLND_CREDIT_HIGHWATER_V1 : \
120 min(t->lnd_peercredits_hiw, (__u32)conn->ibc_queue_depth - 1))
122 #ifdef HAVE_RDMA_CREATE_ID_5ARG
123 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
124 rdma_create_id((ns) ? (ns) : &init_net, cb, dev, ps, qpt)
126 # ifdef HAVE_RDMA_CREATE_ID_4ARG
127 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
128 rdma_create_id(cb, dev, ps, qpt)
130 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
131 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 __u64 ibh_mr_size; /* size of MR */
205 int ibh_max_qp_wr; /* maximum work requests size */
206 #ifdef HAVE_IB_GET_DMA_MR
207 struct ib_mr *ibh_mrs; /* global MR */
209 struct ib_pd *ibh_pd; /* PD */
210 u8 ibh_port; /* port number */
211 struct ib_event_handler
212 ibh_event_handler; /* IB event handler */
213 int ibh_state; /* device status */
214 #define IBLND_DEV_PORT_DOWN 0
215 #define IBLND_DEV_PORT_ACTIVE 1
216 #define IBLND_DEV_FATAL 2
217 struct kib_dev *ibh_dev; /* owner */
218 atomic_t ibh_ref; /* refcount */
221 /** # of seconds to keep pool alive */
222 #define IBLND_POOL_DEADLINE 300
223 /** # of seconds to retry if allocation failed */
224 #define IBLND_POOL_RETRY 1
227 int ibp_npages; /* # pages */
228 struct page *ibp_pages[0]; /* page array */
234 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
235 int inc, struct kib_pool **pp_po);
236 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
237 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
238 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
242 #define IBLND_POOL_NAME_LEN 32
247 /* network it belongs to */
248 struct kib_net *ps_net;
250 char ps_name[IBLND_POOL_NAME_LEN];
252 struct list_head ps_pool_list;
253 /* failed pool list */
254 struct list_head ps_failed_pool_list;
255 /* time stamp for retry if failed to allocate */
256 time64_t ps_next_retry;
257 /* is allocating new pool */
264 /* create a new pool */
265 kib_ps_pool_create_t ps_pool_create;
267 kib_ps_pool_destroy_t ps_pool_destroy;
268 /* initialize new allocated node */
269 kib_ps_node_init_t ps_node_init;
271 kib_ps_node_fini_t ps_node_fini;
275 /* chain on pool list */
276 struct list_head po_list;
277 /* pre-allocated node */
278 struct list_head po_free_list;
279 /* pool_set of this pool */
280 struct kib_poolset *po_owner;
281 /* deadline of this pool */
282 time64_t po_deadline;
283 /* # of elements in use */
285 /* pool is created on failed HCA */
287 /* # of pre-allocated elements */
291 struct kib_tx_poolset {
292 struct kib_poolset tps_poolset; /* pool-set */
293 __u64 tps_next_tx_cookie; /* cookie of TX */
297 struct kib_pool tpo_pool; /* pool */
298 struct kib_hca_dev *tpo_hdev; /* device for this pool */
299 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
300 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
303 struct kib_fmr_poolset {
304 spinlock_t fps_lock; /* serialize */
305 struct kib_net *fps_net; /* IB network */
306 struct list_head fps_pool_list; /* FMR pool list */
307 struct list_head fps_failed_pool_list; /* FMR pool list */
308 __u64 fps_version; /* validity stamp */
309 int fps_cpt; /* CPT id */
311 int fps_flush_trigger;
313 /* is allocating new pool */
315 /* time stamp for retry if failed to allocate */
316 time64_t fps_next_retry;
319 #ifndef HAVE_IB_RDMA_WR
321 struct ib_send_wr wr;
325 struct kib_fast_reg_descriptor { /* For fast registration */
326 struct list_head frd_list;
327 struct ib_rdma_wr frd_inv_wr;
328 #ifdef HAVE_IB_MAP_MR_SG
329 struct ib_reg_wr frd_fastreg_wr;
331 struct ib_rdma_wr frd_fastreg_wr;
332 struct ib_fast_reg_page_list *frd_frpl;
334 struct ib_mr *frd_mr;
338 struct kib_fmr_pool {
339 struct list_head fpo_list; /* chain on pool list */
340 struct kib_hca_dev *fpo_hdev; /* device for this pool */
341 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
344 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
346 struct { /* For fast registration */
347 struct list_head fpo_pool_list;
351 time64_t fpo_deadline; /* deadline of this pool */
352 int fpo_failed; /* fmr pool is failed */
353 int fpo_map_count; /* # of mapped FMR */
354 bool fpo_is_fmr; /* True if FMR pools allocated */
358 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
359 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
360 struct kib_fast_reg_descriptor *fmr_frd;
364 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
365 #define kib_fmr_pool_map(pool, pgs, n, iov) \
366 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
368 #define kib_fmr_pool_map(pool, pgs, n, iov) \
369 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
373 /* chain on struct kib_dev::ibd_nets */
374 struct list_head ibn_list;
375 __u64 ibn_incarnation;/* my epoch */
376 int ibn_init; /* initialisation state */
377 int ibn_shutdown; /* shutting down? */
379 atomic_t ibn_npeers; /* # peers extant */
380 atomic_t ibn_nconns; /* # connections extant */
382 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
383 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
385 struct kib_dev *ibn_dev; /* underlying IB device */
386 struct lnet_ni *ibn_ni; /* LNet interface */
389 #define KIB_THREAD_SHIFT 16
390 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
391 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
392 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
394 struct kib_sched_info {
397 /* schedulers sleep here */
398 wait_queue_head_t ibs_waitq;
399 /* conns to check for rx completions */
400 struct list_head ibs_conns;
401 /* number of scheduler threads */
403 /* max allowed scheduler threads */
404 int ibs_nthreads_max;
405 int ibs_cpt; /* CPT id */
409 int kib_init; /* initialisation state */
410 int kib_shutdown; /* shut down? */
411 struct list_head kib_devs; /* IB devices extant */
412 /* list head of failed devices */
413 struct list_head kib_failed_devs;
414 /* schedulers sleep here */
415 wait_queue_head_t kib_failover_waitq;
416 atomic_t kib_nthreads; /* # live threads */
417 /* stabilize net/dev/peer_ni/conn ops */
418 rwlock_t kib_global_lock;
419 /* hash table of all my known peers */
420 struct list_head *kib_peers;
421 /* size of kib_peers */
422 int kib_peer_hash_size;
423 /* the connd task (serialisation assertions) */
425 /* connections to setup/teardown */
426 struct list_head kib_connd_conns;
427 /* connections with zero refcount */
428 struct list_head kib_connd_zombies;
429 /* connections to reconnect */
430 struct list_head kib_reconn_list;
431 /* peers wait for reconnection */
432 struct list_head kib_reconn_wait;
434 * The second that peers are pulled out from \a kib_reconn_wait
437 time64_t kib_reconn_sec;
438 /* connection daemon sleeps here */
439 wait_queue_head_t kib_connd_waitq;
440 spinlock_t kib_connd_lock; /* serialise */
441 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
442 /* percpt data for schedulers */
443 struct kib_sched_info **kib_scheds;
446 #define IBLND_INIT_NOTHING 0
447 #define IBLND_INIT_DATA 1
448 #define IBLND_INIT_ALL 2
450 /************************************************************************
451 * IB Wire message format.
452 * These are sent in sender's byte order (i.e. receiver flips).
455 struct kib_connparams {
456 __u16 ibcp_queue_depth;
457 __u16 ibcp_max_frags;
458 __u32 ibcp_max_msg_size;
461 struct kib_immediate_msg {
462 struct lnet_hdr ibim_hdr; /* portals header */
463 char ibim_payload[0];/* piggy-backed payload */
466 struct kib_rdma_frag {
467 __u32 rf_nob; /* # bytes this frag */
468 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
471 struct kib_rdma_desc {
472 __u32 rd_key; /* local/remote key */
473 __u32 rd_nfrags; /* # fragments */
474 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
477 struct kib_putreq_msg {
478 struct lnet_hdr ibprm_hdr; /* portals header */
479 __u64 ibprm_cookie; /* opaque completion cookie */
482 struct kib_putack_msg {
483 __u64 ibpam_src_cookie; /* reflected completion cookie */
484 __u64 ibpam_dst_cookie; /* opaque completion cookie */
485 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
489 struct lnet_hdr ibgm_hdr; /* portals header */
490 __u64 ibgm_cookie; /* opaque completion cookie */
491 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
494 struct kib_completion_msg {
495 __u64 ibcm_cookie; /* opaque completion cookie */
496 __s32 ibcm_status; /* < 0 failure: >= 0 length */
500 /* First 2 fields fixed FOR ALL TIME */
501 __u32 ibm_magic; /* I'm an ibnal message */
502 __u16 ibm_version; /* this is my version number */
504 __u8 ibm_type; /* msg type */
505 __u8 ibm_credits; /* returned credits */
506 __u32 ibm_nob; /* # bytes in whole message */
507 __u32 ibm_cksum; /* checksum (0 == no checksum) */
508 __u64 ibm_srcnid; /* sender's NID */
509 __u64 ibm_srcstamp; /* sender's incarnation */
510 __u64 ibm_dstnid; /* destination's NID */
511 __u64 ibm_dststamp; /* destination's incarnation */
514 struct kib_connparams connparams;
515 struct kib_immediate_msg immediate;
516 struct kib_putreq_msg putreq;
517 struct kib_putack_msg putack;
518 struct kib_get_msg get;
519 struct kib_completion_msg completion;
523 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
525 #define IBLND_MSG_VERSION_1 0x11
526 #define IBLND_MSG_VERSION_2 0x12
527 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
529 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
530 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
531 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
532 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
533 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
534 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
535 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
536 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
537 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
538 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
541 __u32 ibr_magic; /* sender's magic */
542 __u16 ibr_version; /* sender's version */
543 __u8 ibr_why; /* reject reason */
544 __u8 ibr_padding; /* padding */
545 __u64 ibr_incarnation; /* incarnation of peer_ni */
546 struct kib_connparams ibr_cp; /* connection parameters */
549 /* connection rejection reasons */
550 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
551 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
552 #define IBLND_REJECT_FATAL 3 /* Anything else */
554 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
555 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
557 /* peer_ni's rdma frags doesn't match mine */
558 #define IBLND_REJECT_RDMA_FRAGS 6
559 /* peer_ni's msg queue size doesn't match mine */
560 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
561 #define IBLND_REJECT_INVALID_SRV_ID 8
563 /***********************************************************************/
565 struct kib_rx { /* receive message */
566 /* queue for attention */
567 struct list_head rx_list;
569 struct kib_conn *rx_conn;
570 /* # bytes received (-1 while posted) */
572 /* message buffer (host vaddr) */
573 struct kib_msg *rx_msg;
574 /* message buffer (I/O addr) */
576 /* for dma_unmap_single() */
577 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
578 /* receive work item... */
579 struct ib_recv_wr rx_wrq;
580 /* ...and its memory */
581 struct ib_sge rx_sge;
584 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
585 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
586 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
587 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
589 struct kib_tx { /* transmit message */
590 /* queue on idle_txs ibc_tx_queue etc. */
591 struct list_head tx_list;
593 struct kib_tx_pool *tx_pool;
595 struct kib_conn *tx_conn;
596 /* # tx callbacks outstanding */
598 /* queued for sending */
600 /* waiting for peer_ni */
602 /* LNET completion status */
604 /* health status of the transmit */
605 enum lnet_msg_hstatus tx_hstatus;
606 /* completion deadline */
608 /* completion cookie */
610 /* lnet msgs to finalize on completion */
611 struct lnet_msg *tx_lntmsg[2];
612 /* message buffer (host vaddr) */
613 struct kib_msg *tx_msg;
614 /* message buffer (I/O addr) */
616 /* for dma_unmap_single() */
617 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
618 /** sge for tx_msgaddr */
619 struct ib_sge tx_msgsge;
620 /* # send work items */
622 /* # used scatter/gather elements */
624 /* send work items... */
625 struct ib_rdma_wr *tx_wrq;
626 /* ...and their memory */
627 struct ib_sge *tx_sge;
628 /* rdma descriptor */
629 struct kib_rdma_desc *tx_rd;
630 /* # entries in... */
632 /* dma_map_sg descriptor */
633 struct scatterlist *tx_frags;
634 /* rdma phys page addrs */
636 /* gaps in fragments */
639 struct kib_fmr tx_fmr;
644 struct kib_connvars {
645 /* connection-in-progress variables */
646 struct kib_msg cv_msg;
650 /* scheduler information */
651 struct kib_sched_info *ibc_sched;
653 struct kib_peer_ni *ibc_peer;
655 struct kib_hca_dev *ibc_hdev;
656 /* stash on peer_ni's conn list */
657 struct list_head ibc_list;
658 /* schedule for attention */
659 struct list_head ibc_sched_list;
660 /* version of connection */
662 /* reconnect later */
663 __u16 ibc_reconnect:1;
664 /* which instance of the peer */
665 __u64 ibc_incarnation;
667 atomic_t ibc_refcount;
668 /* what's happening */
670 /* # uncompleted sends */
671 int ibc_nsends_posted;
672 /* # uncompleted NOOPs */
673 int ibc_noops_posted;
674 /* # credits I have */
676 /* # credits to return */
677 int ibc_outstanding_credits;
678 /* # ACK/DONE msg credits */
679 int ibc_reserved_credits;
680 /* set on comms error */
682 /* connections queue depth */
683 __u16 ibc_queue_depth;
684 /* connections max frags */
686 /* receive buffers owned */
687 unsigned int ibc_nrx:16;
688 /* scheduled for attention */
689 unsigned int ibc_scheduled:1;
690 /* CQ callback fired */
691 unsigned int ibc_ready:1;
692 /* time of last send */
693 ktime_t ibc_last_send;
694 /** link chain for kiblnd_check_conns only */
695 struct list_head ibc_connd_list;
696 /** rxs completed before ESTABLISHED */
697 struct list_head ibc_early_rxs;
698 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
699 struct list_head ibc_tx_noops;
700 /* sends that need a credit */
701 struct list_head ibc_tx_queue;
702 /* sends that don't need a credit */
703 struct list_head ibc_tx_queue_nocred;
704 /* sends that need to reserve an ACK/DONE msg */
705 struct list_head ibc_tx_queue_rsrvd;
706 /* active tx awaiting completion */
707 struct list_head ibc_active_txs;
708 /* zombie tx awaiting done */
709 struct list_head ibc_zombie_txs;
713 struct kib_rx *ibc_rxs;
714 /* premapped rx msg pages */
715 struct kib_pages *ibc_rx_pages;
718 struct rdma_cm_id *ibc_cmid;
719 /* completion queue */
720 struct ib_cq *ibc_cq;
722 /* in-progress connection state */
723 struct kib_connvars *ibc_connvars;
726 #define IBLND_CONN_INIT 0 /* being initialised */
727 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
728 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
729 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
730 #define IBLND_CONN_CLOSING 4 /* being closed */
731 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
734 /* stash on global peer_ni list */
735 struct list_head ibp_list;
736 /* who's on the other end(s) */
739 struct lnet_ni *ibp_ni;
740 /* all active connections */
741 struct list_head ibp_conns;
742 /* next connection to send on for round robin */
743 struct kib_conn *ibp_next_conn;
744 /* msgs waiting for a conn */
745 struct list_head ibp_tx_queue;
746 /* incarnation of peer_ni */
747 __u64 ibp_incarnation;
748 /* when (in seconds) I was last alive */
749 time64_t ibp_last_alive;
751 atomic_t ibp_refcount;
752 /* version of peer_ni */
754 /* current passive connection attempts */
755 unsigned short ibp_accepting;
756 /* current active connection attempts */
757 unsigned short ibp_connecting;
758 /* reconnect this peer_ni later */
759 unsigned char ibp_reconnecting;
760 /* counter of how many times we triggered a conn race */
761 unsigned char ibp_races;
762 /* # consecutive reconnection attempts to this peer */
763 unsigned int ibp_reconnected;
764 /* errno on closing this peer_ni */
766 /* max map_on_demand */
768 /* max_peer_credits */
769 __u16 ibp_queue_depth;
772 #ifndef HAVE_IB_INC_RKEY
774 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
775 * for calculating a new rkey for type 2 memory windows.
776 * @rkey - the rkey to increment.
778 static inline u32 ib_inc_rkey(u32 rkey)
780 const u32 mask = 0x000000ff;
781 return ((rkey + 1) & mask) | (rkey & ~mask);
785 extern struct kib_data kiblnd_data;
787 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
789 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
791 static inline int kiblnd_timeout(void)
793 return *kiblnd_tunables.kib_timeout ? *kiblnd_tunables.kib_timeout :
794 lnet_get_lnd_timeout();
798 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
800 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
801 int concurrent_sends;
803 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
804 concurrent_sends = tunables->lnd_concurrent_sends;
806 if (version == IBLND_MSG_VERSION_1) {
807 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
808 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
810 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
811 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
814 return concurrent_sends;
818 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
820 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
821 atomic_inc(&hdev->ibh_ref);
825 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
827 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
828 if (atomic_dec_and_test(&hdev->ibh_ref))
829 kiblnd_hdev_destroy(hdev);
833 kiblnd_dev_can_failover(struct kib_dev *dev)
835 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
838 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
841 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
844 return dev->ibd_can_failover;
847 #define kiblnd_conn_addref(conn) \
849 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
850 (conn), atomic_read(&(conn)->ibc_refcount)); \
851 atomic_inc(&(conn)->ibc_refcount); \
854 #define kiblnd_conn_decref(conn) \
856 unsigned long flags; \
858 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
859 (conn), atomic_read(&(conn)->ibc_refcount)); \
860 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
861 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
862 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
863 list_add_tail(&(conn)->ibc_list, \
864 &kiblnd_data.kib_connd_zombies); \
865 wake_up(&kiblnd_data.kib_connd_waitq); \
866 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
870 #define kiblnd_peer_addref(peer_ni) \
872 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
873 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
874 atomic_read (&(peer_ni)->ibp_refcount)); \
875 atomic_inc(&(peer_ni)->ibp_refcount); \
878 #define kiblnd_peer_decref(peer_ni) \
880 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
881 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
882 atomic_read (&(peer_ni)->ibp_refcount)); \
883 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
884 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
885 kiblnd_destroy_peer(peer_ni); \
889 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
891 return peer_ni->ibp_connecting != 0 ||
892 peer_ni->ibp_reconnecting != 0 ||
893 peer_ni->ibp_accepting != 0;
897 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
899 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
902 static inline struct list_head *
903 kiblnd_nid2peerlist (lnet_nid_t nid)
906 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
908 return &kiblnd_data.kib_peers[hash];
912 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
914 /* Am I in the peer_ni hash table? */
915 return !list_empty(&peer_ni->ibp_list);
918 static inline struct kib_conn *
919 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
921 struct list_head *next;
923 LASSERT(!list_empty(&peer_ni->ibp_conns));
925 /* Advance to next connection, be sure to skip the head node */
926 if (!peer_ni->ibp_next_conn ||
927 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
928 next = peer_ni->ibp_conns.next;
930 next = peer_ni->ibp_next_conn->ibc_list.next;
931 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
933 return peer_ni->ibp_next_conn;
937 kiblnd_send_keepalive(struct kib_conn *conn)
939 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
941 return (*kiblnd_tunables.kib_keepalive > 0) &&
942 ktime_after(ktime_get(),
943 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
947 kiblnd_need_noop(struct kib_conn *conn)
949 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
950 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
952 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
953 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
955 if (conn->ibc_outstanding_credits <
956 IBLND_CREDITS_HIGHWATER(tunables, conn) &&
957 !kiblnd_send_keepalive(conn))
958 return 0; /* No need to send NOOP */
960 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
961 if (!list_empty(&conn->ibc_tx_queue_nocred))
962 return 0; /* NOOP can be piggybacked */
964 /* No tx to piggyback NOOP onto or no credit to send a tx */
965 return (list_empty(&conn->ibc_tx_queue) ||
966 conn->ibc_credits == 0);
969 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
970 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
971 conn->ibc_credits == 0) /* no credit */
974 if (conn->ibc_credits == 1 && /* last credit reserved for */
975 conn->ibc_outstanding_credits == 0) /* giving back credits */
978 /* No tx to piggyback NOOP onto or no credit to send a tx */
979 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
983 kiblnd_abort_receives(struct kib_conn *conn)
985 ib_modify_qp(conn->ibc_cmid->qp,
986 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
989 static inline const char *
990 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
992 if (q == &conn->ibc_tx_queue)
995 if (q == &conn->ibc_tx_queue_rsrvd)
996 return "tx_queue_rsrvd";
998 if (q == &conn->ibc_tx_queue_nocred)
999 return "tx_queue_nocred";
1001 if (q == &conn->ibc_active_txs)
1002 return "active_txs";
1008 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1009 * lowest bits of the work request id to stash the work item type. */
1011 #define IBLND_WID_INVAL 0
1012 #define IBLND_WID_TX 1
1013 #define IBLND_WID_RX 2
1014 #define IBLND_WID_RDMA 3
1015 #define IBLND_WID_MR 4
1016 #define IBLND_WID_MASK 7UL
1019 kiblnd_ptr2wreqid (void *ptr, int type)
1021 unsigned long lptr = (unsigned long)ptr;
1023 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1024 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1025 return (__u64)(lptr | type);
1028 static inline void *
1029 kiblnd_wreqid2ptr (__u64 wreqid)
1031 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1035 kiblnd_wreqid2type (__u64 wreqid)
1037 return (wreqid & IBLND_WID_MASK);
1041 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1043 conn->ibc_state = state;
1048 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1050 msg->ibm_type = type;
1051 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1055 kiblnd_rd_size(struct kib_rdma_desc *rd)
1060 for (i = size = 0; i < rd->rd_nfrags; i++)
1061 size += rd->rd_frags[i].rf_nob;
1067 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1069 return rd->rd_frags[index].rf_addr;
1073 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1075 return rd->rd_frags[index].rf_nob;
1079 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1085 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1087 if (nob < rd->rd_frags[index].rf_nob) {
1088 rd->rd_frags[index].rf_addr += nob;
1089 rd->rd_frags[index].rf_nob -= nob;
1098 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1100 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1101 msgtype == IBLND_MSG_PUT_ACK);
1103 return msgtype == IBLND_MSG_GET_REQ ?
1104 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1105 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1109 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1111 return ib_dma_mapping_error(dev, dma_addr);
1114 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1115 void *msg, size_t size,
1116 enum dma_data_direction direction)
1118 return ib_dma_map_single(dev, msg, size, direction);
1121 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1122 __u64 addr, size_t size,
1123 enum dma_data_direction direction)
1125 ib_dma_unmap_single(dev, addr, size, direction);
1128 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1129 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1131 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1132 struct scatterlist *sg, int nents,
1133 enum dma_data_direction direction)
1135 return ib_dma_map_sg(dev, sg, nents, direction);
1138 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1139 struct scatterlist *sg, int nents,
1140 enum dma_data_direction direction)
1142 ib_dma_unmap_sg(dev, sg, nents, direction);
1145 #ifndef HAVE_IB_SG_DMA_ADDRESS
1146 #include <linux/scatterlist.h>
1147 #define ib_sg_dma_address(dev, sg) sg_dma_address(sg)
1148 #define ib_sg_dma_len(dev, sg) sg_dma_len(sg)
1151 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1152 struct scatterlist *sg)
1154 return ib_sg_dma_address(dev, sg);
1157 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1158 struct scatterlist *sg)
1160 return ib_sg_dma_len(dev, sg);
1163 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1164 * right because OFED1.2 defines it as const, to use it we have to add
1165 * (void *) cast to overcome "const" */
1167 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1168 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1170 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1171 void kiblnd_map_rx_descs(struct kib_conn *conn);
1172 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1173 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1174 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1176 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1177 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1178 struct kib_fmr *fmr);
1179 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1181 int kiblnd_tunables_setup(struct lnet_ni *ni);
1182 int kiblnd_tunables_init(void);
1184 int kiblnd_connd (void *arg);
1185 int kiblnd_scheduler(void *arg);
1186 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1187 int kiblnd_failover_thread (void *arg);
1189 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1191 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1192 struct rdma_cm_event *event);
1193 int kiblnd_translate_mtu(int value);
1195 int kiblnd_dev_failover(struct kib_dev *dev, struct net *ns);
1196 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1198 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1199 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1200 void kiblnd_destroy_dev(struct kib_dev *dev);
1201 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1202 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1203 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1204 int version, u64 incarnation);
1205 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1207 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1208 struct rdma_cm_id *cmid,
1209 int state, int version);
1210 void kiblnd_destroy_conn(struct kib_conn *conn);
1211 void kiblnd_close_conn(struct kib_conn *conn, int error);
1212 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1214 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1215 void kiblnd_txlist_done(struct list_head *txlist, int status,
1216 enum lnet_msg_hstatus hstatus);
1218 void kiblnd_qp_event(struct ib_event *event, void *arg);
1219 void kiblnd_cq_event(struct ib_event *event, void *arg);
1220 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1222 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1223 int credits, lnet_nid_t dstnid, __u64 dststamp);
1224 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1225 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1227 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1228 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1229 int delayed, unsigned int niov,
1230 struct bio_vec *kiov, unsigned int offset, unsigned int mlen,