4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd.h
34 * Author: Eric Barton <eric@bartonsoftware.com>
37 #include <linux/module.h>
38 #include <linux/kernel.h>
40 #if defined(NEED_LOCKDEP_IS_HELD_DISCARD_CONST) \
41 && defined(CONFIG_LOCKDEP) \
42 && defined(lockdep_is_held)
43 #undef lockdep_is_held
44 #define lockdep_is_held(lock) \
45 lock_is_held((struct lockdep_map *)&(lock)->dep_map)
48 #ifdef HAVE_COMPAT_RDMA
49 #include <linux/compat-2.6.h>
51 #ifdef LINUX_3_17_COMPAT_H
52 #undef NEED_KTIME_GET_REAL_NS
55 /* MOFED has its own bitmap_alloc backport */
56 #define HAVE_BITMAP_ALLOC 1
60 #include <linux/kthread.h>
62 #include <linux/string.h>
63 #include <linux/stat.h>
64 #include <linux/errno.h>
65 #include <linux/unistd.h>
66 #include <linux/uio.h>
68 #include <asm/uaccess.h>
71 #include <linux/init.h>
73 #include <linux/file.h>
74 #include <linux/stat.h>
75 #include <linux/list.h>
76 #include <linux/kmod.h>
77 #include <linux/sysctl.h>
78 #include <linux/pci.h>
83 #include <rdma/rdma_cm.h>
84 #include <rdma/ib_cm.h>
85 #include <rdma/ib_verbs.h>
86 #ifdef HAVE_FMR_POOL_API
87 #include <rdma/ib_fmr_pool.h>
90 #define DEBUG_SUBSYSTEM S_LND
92 #include <libcfs/libcfs.h>
93 #include <lnet/lib-lnet.h>
95 #define IBLND_PEER_HASH_SIZE 101 /* # peer_ni lists */
97 #define IBLND_N_SCHED 2
98 #define IBLND_N_SCHED_HIGH 4
100 struct kib_tunables {
101 int *kib_dev_failover; /* HCA failover */
102 unsigned int *kib_service; /* IB service number */
103 int *kib_cksum; /* checksum struct kib_msg? */
104 int *kib_timeout; /* comms timeout (seconds) */
105 int *kib_keepalive; /* keepalive timeout (seconds) */
106 char **kib_default_ipif; /* default IPoIB interface */
107 int *kib_retry_count;
108 int *kib_rnr_retry_count;
109 int *kib_ib_mtu; /* IB MTU */
110 int *kib_require_priv_port;/* accept only privileged ports */
111 int *kib_use_priv_port; /* use privileged port for active connect */
112 /* # threads on each CPT */
114 int *kib_wrq_sge; /* # sg elements per wrq */
115 int *kib_use_fastreg_gaps; /* enable discontiguous fastreg fragment support */
118 extern struct kib_tunables kiblnd_tunables;
120 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
121 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
123 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer_ni credits */
124 #define IBLND_CREDITS_MAX ((typeof(((struct kib_msg *) 0)->ibm_credits)) - 1) /* Max # of peer_ni credits */
126 /* when eagerly to return credits */
127 #define IBLND_CREDITS_HIGHWATER(t, conn) ((conn->ibc_version) == IBLND_MSG_VERSION_1 ? \
128 IBLND_CREDIT_HIGHWATER_V1 : \
129 min(t->lnd_peercredits_hiw, (__u32)conn->ibc_queue_depth - 1))
131 #ifdef HAVE_RDMA_CREATE_ID_5ARG
132 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
133 rdma_create_id((ns) ? (ns) : &init_net, cb, dev, ps, qpt)
135 # ifdef HAVE_RDMA_CREATE_ID_4ARG
136 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
137 rdma_create_id(cb, dev, ps, qpt)
139 # define kiblnd_rdma_create_id(ns, cb, dev, ps, qpt) \
140 rdma_create_id(cb, dev, ps)
144 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
145 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
146 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
148 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
149 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
151 /************************/
152 /* derived constants... */
153 /* Pools (shared by connections on each CPT) */
154 /* These pools can grow at runtime, so don't need give a very large value */
155 #define IBLND_TX_POOL 256
156 #define IBLND_FMR_POOL 256
157 #define IBLND_FMR_POOL_FLUSH 192
159 /* RX messages (per connection) */
160 #define IBLND_RX_MSGS(c) \
161 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
162 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
163 #define IBLND_RX_MSG_PAGES(c) \
164 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
166 /* WRs and CQEs (per connection) */
167 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
169 /* 2 = LNet msg + Transfer chain */
170 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + kiblnd_send_wrs(c))
174 /* o2iblnd can run over aliased interface */
176 #define KIB_IFNAME_SIZE IFALIASZ
178 #define KIB_IFNAME_SIZE 256
182 IBLND_DEV_CAPS_FASTREG_ENABLED = BIT(0),
183 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT = BIT(1),
184 #ifdef HAVE_FMR_POOL_API
185 IBLND_DEV_CAPS_FMR_ENABLED = BIT(2),
190 struct list_head ibd_list; /* chain on kib_devs */
191 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
192 __u32 ibd_ifip; /* IPoIB interface IP */
193 /** IPoIB interface name */
194 char ibd_ifname[KIB_IFNAME_SIZE];
195 int ibd_nnets; /* # nets extant */
197 time64_t ibd_next_failover;
198 /* # failover failures */
199 int ibd_failed_failover;
200 /* failover in progress */
201 unsigned int ibd_failover;
202 /* IPoIB interface is a bonding master */
203 unsigned int ibd_can_failover;
204 struct list_head ibd_nets;
205 struct kib_hca_dev *ibd_hdev;
206 enum kib_dev_caps ibd_dev_caps;
210 struct rdma_cm_id *ibh_cmid; /* listener cmid */
211 struct ib_device *ibh_ibdev; /* IB device */
212 int ibh_page_shift; /* page shift of current HCA */
213 int ibh_page_size; /* page size of current HCA */
214 __u64 ibh_page_mask; /* page mask of current HCA */
215 __u64 ibh_mr_size; /* size of MR */
216 int ibh_max_qp_wr; /* maximum work requests size */
217 #ifdef HAVE_IB_GET_DMA_MR
218 struct ib_mr *ibh_mrs; /* global MR */
220 struct ib_pd *ibh_pd; /* PD */
221 u8 ibh_port; /* port number */
222 struct ib_event_handler
223 ibh_event_handler; /* IB event handler */
224 int ibh_state; /* device status */
225 #define IBLND_DEV_PORT_DOWN 0
226 #define IBLND_DEV_PORT_ACTIVE 1
227 #define IBLND_DEV_FATAL 2
228 struct kib_dev *ibh_dev; /* owner */
229 atomic_t ibh_ref; /* refcount */
232 /** # of seconds to keep pool alive */
233 #define IBLND_POOL_DEADLINE 300
234 /** # of seconds to retry if allocation failed */
235 #define IBLND_POOL_RETRY 1
238 int ibp_npages; /* # pages */
239 struct page *ibp_pages[0]; /* page array */
245 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
246 int inc, struct kib_pool **pp_po);
247 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
248 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
249 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
253 #define IBLND_POOL_NAME_LEN 32
258 /* network it belongs to */
259 struct kib_net *ps_net;
261 char ps_name[IBLND_POOL_NAME_LEN];
263 struct list_head ps_pool_list;
264 /* failed pool list */
265 struct list_head ps_failed_pool_list;
266 /* time stamp for retry if failed to allocate */
267 time64_t ps_next_retry;
268 /* is allocating new pool */
275 /* create a new pool */
276 kib_ps_pool_create_t ps_pool_create;
278 kib_ps_pool_destroy_t ps_pool_destroy;
279 /* initialize new allocated node */
280 kib_ps_node_init_t ps_node_init;
282 kib_ps_node_fini_t ps_node_fini;
286 /* chain on pool list */
287 struct list_head po_list;
288 /* pre-allocated node */
289 struct list_head po_free_list;
290 /* pool_set of this pool */
291 struct kib_poolset *po_owner;
292 /* deadline of this pool */
293 time64_t po_deadline;
294 /* # of elements in use */
296 /* pool is created on failed HCA */
298 /* # of pre-allocated elements */
302 struct kib_tx_poolset {
303 struct kib_poolset tps_poolset; /* pool-set */
304 __u64 tps_next_tx_cookie; /* cookie of TX */
308 struct kib_pool tpo_pool; /* pool */
309 struct kib_hca_dev *tpo_hdev; /* device for this pool */
310 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
311 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
314 struct kib_fmr_poolset {
315 spinlock_t fps_lock; /* serialize */
316 struct kib_net *fps_net; /* IB network */
317 struct list_head fps_pool_list; /* FMR pool list */
318 struct list_head fps_failed_pool_list; /* FMR pool list */
319 __u64 fps_version; /* validity stamp */
320 int fps_cpt; /* CPT id */
322 int fps_flush_trigger;
324 /* is allocating new pool */
326 /* time stamp for retry if failed to allocate */
327 time64_t fps_next_retry;
330 #ifndef HAVE_IB_RDMA_WR
332 struct ib_send_wr wr;
336 struct kib_fast_reg_descriptor { /* For fast registration */
337 struct list_head frd_list;
338 struct ib_rdma_wr frd_inv_wr;
339 #ifdef HAVE_IB_MAP_MR_SG
340 struct ib_reg_wr frd_fastreg_wr;
342 struct ib_rdma_wr frd_fastreg_wr;
343 struct ib_fast_reg_page_list *frd_frpl;
345 struct ib_mr *frd_mr;
349 struct kib_fmr_pool {
350 struct list_head fpo_list; /* chain on pool list */
351 struct kib_hca_dev *fpo_hdev; /* device for this pool */
352 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
353 #ifdef HAVE_FMR_POOL_API
356 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
359 struct { /* For fast registration */
360 struct list_head fpo_pool_list;
363 #ifdef HAVE_FMR_POOL_API
365 bool fpo_is_fmr; /* True if FMR pools allocated */
367 time64_t fpo_deadline; /* deadline of this pool */
368 int fpo_failed; /* fmr pool is failed */
369 int fpo_map_count; /* # of mapped FMR */
373 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
374 #ifdef HAVE_FMR_POOL_API
375 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
376 #endif /* HAVE_FMR_POOL_API */
377 struct kib_fast_reg_descriptor *fmr_frd;
381 #ifdef HAVE_FMR_POOL_API
383 #ifdef HAVE_ORACLE_OFED_EXTENSIONS
384 #define kib_fmr_pool_map(pool, pgs, n, iov) \
385 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov), NULL)
387 #define kib_fmr_pool_map(pool, pgs, n, iov) \
388 ib_fmr_pool_map_phys((pool), (pgs), (n), (iov))
391 #endif /* HAVE_FMR_POOL_API */
394 /* chain on struct kib_dev::ibd_nets */
395 struct list_head ibn_list;
396 __u64 ibn_incarnation;/* my epoch */
397 int ibn_init; /* initialisation state */
398 int ibn_shutdown; /* shutting down? */
400 atomic_t ibn_npeers; /* # peers extant */
401 atomic_t ibn_nconns; /* # connections extant */
403 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
404 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
406 struct kib_dev *ibn_dev; /* underlying IB device */
407 struct lnet_ni *ibn_ni; /* LNet interface */
410 #define KIB_THREAD_SHIFT 16
411 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
412 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
413 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
415 struct kib_sched_info {
418 /* schedulers sleep here */
419 wait_queue_head_t ibs_waitq;
420 /* conns to check for rx completions */
421 struct list_head ibs_conns;
422 /* number of scheduler threads */
424 /* max allowed scheduler threads */
425 int ibs_nthreads_max;
426 int ibs_cpt; /* CPT id */
430 int kib_init; /* initialisation state */
431 int kib_shutdown; /* shut down? */
432 struct list_head kib_devs; /* IB devices extant */
433 /* list head of failed devices */
434 struct list_head kib_failed_devs;
435 /* schedulers sleep here */
436 wait_queue_head_t kib_failover_waitq;
437 atomic_t kib_nthreads; /* # live threads */
438 /* stabilize net/dev/peer_ni/conn ops */
439 rwlock_t kib_global_lock;
440 /* hash table of all my known peers */
441 struct list_head *kib_peers;
442 /* size of kib_peers */
443 int kib_peer_hash_size;
444 /* the connd task (serialisation assertions) */
446 /* connections to setup/teardown */
447 struct list_head kib_connd_conns;
448 /* connections with zero refcount */
449 struct list_head kib_connd_zombies;
450 /* connections to reconnect */
451 struct list_head kib_reconn_list;
452 /* peers wait for reconnection */
453 struct list_head kib_reconn_wait;
454 /* connections wait for completion */
455 struct list_head kib_connd_waits;
457 * The second that peers are pulled out from \a kib_reconn_wait
460 time64_t kib_reconn_sec;
461 /* connection daemon sleeps here */
462 wait_queue_head_t kib_connd_waitq;
463 spinlock_t kib_connd_lock; /* serialise */
464 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
465 /* percpt data for schedulers */
466 struct kib_sched_info **kib_scheds;
469 #define IBLND_INIT_NOTHING 0
470 #define IBLND_INIT_DATA 1
471 #define IBLND_INIT_ALL 2
473 /************************************************************************
474 * IB Wire message format.
475 * These are sent in sender's byte order (i.e. receiver flips).
478 struct kib_connparams {
479 __u16 ibcp_queue_depth;
480 __u16 ibcp_max_frags;
481 __u32 ibcp_max_msg_size;
484 struct kib_immediate_msg {
485 struct lnet_hdr ibim_hdr; /* portals header */
486 char ibim_payload[0];/* piggy-backed payload */
489 struct kib_rdma_frag {
490 __u32 rf_nob; /* # bytes this frag */
491 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
494 struct kib_rdma_desc {
495 __u32 rd_key; /* local/remote key */
496 __u32 rd_nfrags; /* # fragments */
497 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
500 struct kib_putreq_msg {
501 struct lnet_hdr ibprm_hdr; /* portals header */
502 __u64 ibprm_cookie; /* opaque completion cookie */
505 struct kib_putack_msg {
506 __u64 ibpam_src_cookie; /* reflected completion cookie */
507 __u64 ibpam_dst_cookie; /* opaque completion cookie */
508 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
512 struct lnet_hdr ibgm_hdr; /* portals header */
513 __u64 ibgm_cookie; /* opaque completion cookie */
514 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
517 struct kib_completion_msg {
518 __u64 ibcm_cookie; /* opaque completion cookie */
519 __s32 ibcm_status; /* < 0 failure: >= 0 length */
523 /* First 2 fields fixed FOR ALL TIME */
524 __u32 ibm_magic; /* I'm an ibnal message */
525 __u16 ibm_version; /* this is my version number */
527 __u8 ibm_type; /* msg type */
528 __u8 ibm_credits; /* returned credits */
529 __u32 ibm_nob; /* # bytes in whole message */
530 __u32 ibm_cksum; /* checksum (0 == no checksum) */
531 __u64 ibm_srcnid; /* sender's NID */
532 __u64 ibm_srcstamp; /* sender's incarnation */
533 __u64 ibm_dstnid; /* destination's NID */
534 __u64 ibm_dststamp; /* destination's incarnation */
537 struct kib_connparams connparams;
538 struct kib_immediate_msg immediate;
539 struct kib_putreq_msg putreq;
540 struct kib_putack_msg putack;
541 struct kib_get_msg get;
542 struct kib_completion_msg completion;
546 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
548 #define IBLND_MSG_VERSION_1 0x11
549 #define IBLND_MSG_VERSION_2 0x12
550 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
552 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
553 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
554 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
555 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
556 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
557 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
558 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
559 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
560 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
561 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
564 __u32 ibr_magic; /* sender's magic */
565 __u16 ibr_version; /* sender's version */
566 __u8 ibr_why; /* reject reason */
567 __u8 ibr_padding; /* padding */
568 __u64 ibr_incarnation; /* incarnation of peer_ni */
569 struct kib_connparams ibr_cp; /* connection parameters */
572 /* connection rejection reasons */
573 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
574 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
575 #define IBLND_REJECT_FATAL 3 /* Anything else */
577 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
578 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
580 /* peer_ni's rdma frags doesn't match mine */
581 #define IBLND_REJECT_RDMA_FRAGS 6
582 /* peer_ni's msg queue size doesn't match mine */
583 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
584 #define IBLND_REJECT_INVALID_SRV_ID 8
586 /***********************************************************************/
588 struct kib_rx { /* receive message */
589 /* queue for attention */
590 struct list_head rx_list;
592 struct kib_conn *rx_conn;
593 /* # bytes received (-1 while posted) */
595 /* message buffer (host vaddr) */
596 struct kib_msg *rx_msg;
597 /* message buffer (I/O addr) */
599 /* for dma_unmap_single() */
600 DEFINE_DMA_UNMAP_ADDR(rx_msgunmap);
601 /* receive work item... */
602 struct ib_recv_wr rx_wrq;
603 /* ...and its memory */
604 struct ib_sge rx_sge;
607 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
608 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
609 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
610 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
612 struct kib_tx { /* transmit message */
613 /* queue on idle_txs ibc_tx_queue etc. */
614 struct list_head tx_list;
616 struct kib_tx_pool *tx_pool;
618 struct kib_conn *tx_conn;
619 /* # tx callbacks outstanding */
621 /* queued for sending */
623 /* waiting for peer_ni */
625 /* LNET completion status */
627 /* health status of the transmit */
628 enum lnet_msg_hstatus tx_hstatus;
629 /* completion deadline */
631 /* completion cookie */
633 /* lnet msgs to finalize on completion */
634 struct lnet_msg *tx_lntmsg[2];
635 /* message buffer (host vaddr) */
636 struct kib_msg *tx_msg;
637 /* message buffer (I/O addr) */
639 /* for dma_unmap_single() */
640 DEFINE_DMA_UNMAP_ADDR(tx_msgunmap);
641 /** sge for tx_msgaddr */
642 struct ib_sge tx_msgsge;
643 /* # send work items */
645 /* # used scatter/gather elements */
647 /* send work items... */
648 struct ib_rdma_wr *tx_wrq;
649 /* ...and their memory */
650 struct ib_sge *tx_sge;
651 /* rdma descriptor */
652 struct kib_rdma_desc *tx_rd;
653 /* # entries in... */
655 /* dma_map_sg descriptor */
656 struct scatterlist *tx_frags;
657 /* rdma phys page addrs */
659 /* gaps in fragments */
662 struct kib_fmr tx_fmr;
667 struct kib_connvars {
668 /* connection-in-progress variables */
669 struct kib_msg cv_msg;
673 /* scheduler information */
674 struct kib_sched_info *ibc_sched;
676 struct kib_peer_ni *ibc_peer;
678 struct kib_hca_dev *ibc_hdev;
679 /* stash on peer_ni's conn list */
680 struct list_head ibc_list;
681 /* schedule for attention */
682 struct list_head ibc_sched_list;
683 /* version of connection */
685 /* reconnect later */
686 __u16 ibc_reconnect:1;
687 /* which instance of the peer */
688 __u64 ibc_incarnation;
690 atomic_t ibc_refcount;
691 /* what's happening */
693 /* # uncompleted sends */
694 int ibc_nsends_posted;
695 /* # uncompleted NOOPs */
696 int ibc_noops_posted;
697 /* # credits I have */
699 /* # credits to return */
700 int ibc_outstanding_credits;
701 /* # ACK/DONE msg credits */
702 int ibc_reserved_credits;
703 /* set on comms error */
705 /* connections queue depth */
706 __u16 ibc_queue_depth;
707 /* connections max frags */
709 /* count of timeout txs waiting on cq */
711 /* receive buffers owned */
712 unsigned int ibc_nrx:16;
713 /* scheduled for attention */
714 unsigned int ibc_scheduled:1;
715 /* CQ callback fired */
716 unsigned int ibc_ready:1;
717 /* time of last send */
718 ktime_t ibc_last_send;
719 /** link chain for kiblnd_check_conns only */
720 struct list_head ibc_connd_list;
721 /** rxs completed before ESTABLISHED */
722 struct list_head ibc_early_rxs;
723 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
724 struct list_head ibc_tx_noops;
725 /* sends that need a credit */
726 struct list_head ibc_tx_queue;
727 /* sends that don't need a credit */
728 struct list_head ibc_tx_queue_nocred;
729 /* sends that need to reserve an ACK/DONE msg */
730 struct list_head ibc_tx_queue_rsrvd;
731 /* active tx awaiting completion */
732 struct list_head ibc_active_txs;
733 /* zombie tx awaiting done */
734 struct list_head ibc_zombie_txs;
738 struct kib_rx *ibc_rxs;
739 /* premapped rx msg pages */
740 struct kib_pages *ibc_rx_pages;
743 struct rdma_cm_id *ibc_cmid;
744 /* completion queue */
745 struct ib_cq *ibc_cq;
747 /* in-progress connection state */
748 struct kib_connvars *ibc_connvars;
751 #define IBLND_CONN_INIT 0 /* being initialised */
752 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
753 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
754 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
755 #define IBLND_CONN_CLOSING 4 /* being closed */
756 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
759 /* stash on global peer_ni list */
760 struct list_head ibp_list;
761 /* who's on the other end(s) */
764 struct lnet_ni *ibp_ni;
765 /* all active connections */
766 struct list_head ibp_conns;
767 /* next connection to send on for round robin */
768 struct kib_conn *ibp_next_conn;
769 /* msgs waiting for a conn */
770 struct list_head ibp_tx_queue;
771 /* incarnation of peer_ni */
772 __u64 ibp_incarnation;
773 /* when (in seconds) I was last alive */
774 time64_t ibp_last_alive;
776 atomic_t ibp_refcount;
777 /* version of peer_ni */
779 /* current passive connection attempts */
780 unsigned short ibp_accepting;
781 /* current active connection attempts */
782 unsigned short ibp_connecting;
783 /* reconnect this peer_ni later */
784 unsigned char ibp_reconnecting;
785 /* counter of how many times we triggered a conn race */
786 unsigned char ibp_races;
787 /* # consecutive reconnection attempts to this peer */
788 unsigned int ibp_reconnected;
789 /* number of total active retries */
790 unsigned int ibp_retries;
791 /* errno on closing this peer_ni */
793 /* max map_on_demand */
795 /* max_peer_credits */
796 __u16 ibp_queue_depth;
797 /* reduced value which allows conn to be created if max fails */
798 __u16 ibp_queue_depth_mod;
801 #ifndef HAVE_IB_INC_RKEY
803 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
804 * for calculating a new rkey for type 2 memory windows.
805 * @rkey - the rkey to increment.
807 static inline u32 ib_inc_rkey(u32 rkey)
809 const u32 mask = 0x000000ff;
810 return ((rkey + 1) & mask) | (rkey & ~mask);
814 extern struct kib_data kiblnd_data;
816 extern void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
818 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
820 static inline int kiblnd_timeout(void)
822 return *kiblnd_tunables.kib_timeout ? *kiblnd_tunables.kib_timeout :
823 lnet_get_lnd_timeout();
827 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
829 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
830 int concurrent_sends;
832 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
833 concurrent_sends = tunables->lnd_concurrent_sends;
835 if (version == IBLND_MSG_VERSION_1) {
836 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
837 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
839 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
840 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
843 return concurrent_sends;
847 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
849 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
850 atomic_inc(&hdev->ibh_ref);
854 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
856 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
857 if (atomic_dec_and_test(&hdev->ibh_ref))
858 kiblnd_hdev_destroy(hdev);
862 kiblnd_dev_can_failover(struct kib_dev *dev)
864 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
867 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
870 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
873 return dev->ibd_can_failover;
876 #define kiblnd_conn_addref(conn) \
878 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
879 (conn), atomic_read(&(conn)->ibc_refcount)); \
880 atomic_inc(&(conn)->ibc_refcount); \
883 #define kiblnd_conn_decref(conn) \
885 unsigned long flags; \
887 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
888 (conn), atomic_read(&(conn)->ibc_refcount)); \
889 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
890 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
891 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
892 list_add_tail(&(conn)->ibc_list, \
893 &kiblnd_data.kib_connd_zombies); \
894 wake_up(&kiblnd_data.kib_connd_waitq); \
895 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
899 #define kiblnd_peer_addref(peer_ni) \
901 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
902 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
903 atomic_read (&(peer_ni)->ibp_refcount)); \
904 atomic_inc(&(peer_ni)->ibp_refcount); \
907 #define kiblnd_peer_decref(peer_ni) \
909 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
910 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
911 atomic_read (&(peer_ni)->ibp_refcount)); \
912 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
913 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
914 kiblnd_destroy_peer(peer_ni); \
918 kiblnd_peer_connecting(struct kib_peer_ni *peer_ni)
920 return peer_ni->ibp_connecting != 0 ||
921 peer_ni->ibp_reconnecting != 0 ||
922 peer_ni->ibp_accepting != 0;
926 kiblnd_peer_idle(struct kib_peer_ni *peer_ni)
928 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
931 static inline struct list_head *
932 kiblnd_nid2peerlist (lnet_nid_t nid)
935 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
937 return &kiblnd_data.kib_peers[hash];
941 kiblnd_peer_active(struct kib_peer_ni *peer_ni)
943 /* Am I in the peer_ni hash table? */
944 return !list_empty(&peer_ni->ibp_list);
947 static inline struct kib_conn *
948 kiblnd_get_conn_locked(struct kib_peer_ni *peer_ni)
950 struct list_head *next;
952 LASSERT(!list_empty(&peer_ni->ibp_conns));
954 /* Advance to next connection, be sure to skip the head node */
955 if (!peer_ni->ibp_next_conn ||
956 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
957 next = peer_ni->ibp_conns.next;
959 next = peer_ni->ibp_next_conn->ibc_list.next;
960 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
962 return peer_ni->ibp_next_conn;
966 kiblnd_send_keepalive(struct kib_conn *conn)
968 s64 keepalive_ns = *kiblnd_tunables.kib_keepalive * NSEC_PER_SEC;
970 return (*kiblnd_tunables.kib_keepalive > 0) &&
971 ktime_after(ktime_get(),
972 ktime_add_ns(conn->ibc_last_send, keepalive_ns));
976 kiblnd_need_noop(struct kib_conn *conn)
978 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
979 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
981 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
982 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
984 if (conn->ibc_outstanding_credits <
985 IBLND_CREDITS_HIGHWATER(tunables, conn) &&
986 !kiblnd_send_keepalive(conn))
987 return 0; /* No need to send NOOP */
989 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
990 if (!list_empty(&conn->ibc_tx_queue_nocred))
991 return 0; /* NOOP can be piggybacked */
993 /* No tx to piggyback NOOP onto or no credit to send a tx */
994 return (list_empty(&conn->ibc_tx_queue) ||
995 conn->ibc_credits == 0);
998 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
999 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
1000 conn->ibc_credits == 0) /* no credit */
1003 if (conn->ibc_credits == 1 && /* last credit reserved for */
1004 conn->ibc_outstanding_credits == 0) /* giving back credits */
1007 /* No tx to piggyback NOOP onto or no credit to send a tx */
1008 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
1012 kiblnd_abort_receives(struct kib_conn *conn)
1014 ib_modify_qp(conn->ibc_cmid->qp,
1015 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
1018 static inline const char *
1019 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
1021 if (q == &conn->ibc_tx_queue)
1024 if (q == &conn->ibc_tx_queue_rsrvd)
1025 return "tx_queue_rsrvd";
1027 if (q == &conn->ibc_tx_queue_nocred)
1028 return "tx_queue_nocred";
1030 if (q == &conn->ibc_active_txs)
1031 return "active_txs";
1037 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1038 * lowest bits of the work request id to stash the work item type. */
1040 #define IBLND_WID_INVAL 0
1041 #define IBLND_WID_TX 1
1042 #define IBLND_WID_RX 2
1043 #define IBLND_WID_RDMA 3
1044 #define IBLND_WID_MR 4
1045 #define IBLND_WID_MASK 7UL
1048 kiblnd_ptr2wreqid (void *ptr, int type)
1050 unsigned long lptr = (unsigned long)ptr;
1052 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1053 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1054 return (__u64)(lptr | type);
1057 static inline void *
1058 kiblnd_wreqid2ptr (__u64 wreqid)
1060 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1064 kiblnd_wreqid2type (__u64 wreqid)
1066 return (wreqid & IBLND_WID_MASK);
1070 kiblnd_set_conn_state(struct kib_conn *conn, int state)
1072 conn->ibc_state = state;
1077 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
1079 msg->ibm_type = type;
1080 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1084 kiblnd_rd_size(struct kib_rdma_desc *rd)
1089 for (i = size = 0; i < rd->rd_nfrags; i++)
1090 size += rd->rd_frags[i].rf_nob;
1096 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
1098 return rd->rd_frags[index].rf_addr;
1102 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
1104 return rd->rd_frags[index].rf_nob;
1108 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
1114 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
1116 if (nob < rd->rd_frags[index].rf_nob) {
1117 rd->rd_frags[index].rf_addr += nob;
1118 rd->rd_frags[index].rf_nob -= nob;
1127 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
1129 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1130 msgtype == IBLND_MSG_PUT_ACK);
1132 return msgtype == IBLND_MSG_GET_REQ ?
1133 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
1134 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
1138 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1140 return ib_dma_mapping_error(dev, dma_addr);
1143 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1144 void *msg, size_t size,
1145 enum dma_data_direction direction)
1147 return ib_dma_map_single(dev, msg, size, direction);
1150 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1151 __u64 addr, size_t size,
1152 enum dma_data_direction direction)
1154 ib_dma_unmap_single(dev, addr, size, direction);
1157 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1158 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1160 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1161 struct scatterlist *sg, int nents,
1162 enum dma_data_direction direction)
1164 return ib_dma_map_sg(dev, sg, nents, direction);
1167 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1168 struct scatterlist *sg, int nents,
1169 enum dma_data_direction direction)
1171 ib_dma_unmap_sg(dev, sg, nents, direction);
1174 #ifndef HAVE_IB_SG_DMA_ADDRESS
1175 #include <linux/scatterlist.h>
1176 #define ib_sg_dma_address(dev, sg) sg_dma_address(sg)
1177 #define ib_sg_dma_len(dev, sg) sg_dma_len(sg)
1180 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1181 struct scatterlist *sg)
1183 return ib_sg_dma_address(dev, sg);
1186 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1187 struct scatterlist *sg)
1189 return ib_sg_dma_len(dev, sg);
1192 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1193 * right because OFED1.2 defines it as const, to use it we have to add
1194 * (void *) cast to overcome "const" */
1196 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1197 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1199 void kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs);
1200 void kiblnd_map_rx_descs(struct kib_conn *conn);
1201 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
1202 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
1203 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
1205 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1206 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1207 struct kib_fmr *fmr);
1208 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
1210 int kiblnd_tunables_setup(struct lnet_ni *ni);
1211 int kiblnd_tunables_init(void);
1213 int kiblnd_connd (void *arg);
1214 int kiblnd_scheduler(void *arg);
1215 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1216 int kiblnd_failover_thread (void *arg);
1218 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1220 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1221 struct rdma_cm_event *event);
1222 int kiblnd_translate_mtu(int value);
1224 int kiblnd_dev_failover(struct kib_dev *dev, struct net *ns);
1225 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
1227 void kiblnd_destroy_peer(struct kib_peer_ni *peer);
1228 bool kiblnd_reconnect_peer(struct kib_peer_ni *peer);
1229 void kiblnd_destroy_dev(struct kib_dev *dev);
1230 void kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni);
1231 struct kib_peer_ni *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1232 int kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1233 int version, u64 incarnation);
1234 int kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why);
1236 struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
1237 struct rdma_cm_id *cmid,
1238 int state, int version);
1239 void kiblnd_destroy_conn(struct kib_conn *conn);
1240 void kiblnd_close_conn(struct kib_conn *conn, int error);
1241 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1243 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1244 void kiblnd_txlist_done(struct list_head *txlist, int status,
1245 enum lnet_msg_hstatus hstatus);
1247 void kiblnd_qp_event(struct ib_event *event, void *arg);
1248 void kiblnd_cq_event(struct ib_event *event, void *arg);
1249 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1251 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1252 int credits, lnet_nid_t dstnid, __u64 dststamp);
1253 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1254 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1256 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1257 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1258 int delayed, unsigned int niov,
1259 struct bio_vec *kiov, unsigned int offset, unsigned int mlen,