1 // SPDX-License-Identifier: GPL-2.0
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 only,
9 * as published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful, but
12 * WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * General Public License version 2 for more details (a copy is included
15 * in the LICENSE file that accompanied this code).
17 * You should have received a copy of the GNU General Public License
18 * version 2 along with this program; If not, see
19 * http://www.gnu.org/licenses/gpl-2.0.html
24 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
25 * Use is subject to license terms.
27 * Copyright (c) 2011, 2015, Intel Corporation.
30 * This file is part of Lustre, http://www.lustre.org/
31 * Lustre is a trademark of Sun Microsystems, Inc.
33 * lnet/klnds/o2iblnd/o2iblnd.h
35 * Author: Eric Barton <eric@bartonsoftware.com>
38 #include <linux/module.h>
39 #include <linux/kernel.h>
41 #include <linux/string.h>
42 #include <linux/stat.h>
43 #include <linux/errno.h>
44 #include <linux/unistd.h>
45 #include <linux/uio.h>
46 #include <linux/uaccess.h>
51 #include <linux/file.h>
52 #include <linux/list.h>
53 #include <linux/kmod.h>
54 #include <linux/sysctl.h>
55 #include <linux/pci.h>
60 #include <rdma/rdma_cm.h>
61 #include <rdma/ib_cm.h>
62 #include <rdma/ib_verbs.h>
63 #include <rdma/ib_fmr_pool.h>
65 #define DEBUG_SUBSYSTEM S_LND
67 #include <linux/libcfs/libcfs.h>
68 #include <linux/lnet/lib-lnet.h>
70 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
71 /* # scheduler loops before reschedule */
72 #define IBLND_RESCHED 100
74 #define IBLND_N_SCHED 2
75 #define IBLND_N_SCHED_HIGH 4
78 int *kib_dev_failover; /* HCA failover */
79 unsigned int *kib_service; /* IB service number */
80 int *kib_min_reconnect_interval; /* first failed connection retry... */
81 int *kib_max_reconnect_interval; /* exponentially increasing to this */
82 int *kib_cksum; /* checksum struct kib_msg? */
83 int *kib_timeout; /* comms timeout (seconds) */
84 int *kib_keepalive; /* keepalive timeout (seconds) */
85 int *kib_ntx; /* # tx descs */
86 char **kib_default_ipif; /* default IPoIB interface */
88 int *kib_rnr_retry_count;
89 int *kib_ib_mtu; /* IB MTU */
90 int *kib_require_priv_port; /* accept only privileged ports */
91 int *kib_use_priv_port; /* use privileged port for active connect */
92 int *kib_nscheds; /* # threads on each CPT */
95 extern struct kib_tunables kiblnd_tunables;
97 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
98 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
100 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
101 #define IBLND_CREDITS_MAX ((typeof(((struct kib_msg *)0)->ibm_credits)) - 1) /* Max # of peer credits */
103 /* when eagerly to return credits */
104 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
105 IBLND_CREDIT_HIGHWATER_V1 : \
106 t->lnd_peercredits_hiw)
108 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
112 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
113 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
114 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
116 #define IBLND_FRAG_SHIFT (PAGE_SHIFT - 12) /* frag size on wire is in 4K units */
117 #define IBLND_MSG_SIZE (4 << 10) /* max size of queued messages (inc hdr) */
118 #define IBLND_MAX_RDMA_FRAGS (LNET_MAX_PAYLOAD >> 12)/* max # of fragments supported in 4K size */
120 /************************/
121 /* derived constants... */
122 /* Pools (shared by connections on each CPT) */
123 /* These pools can grow at runtime, so don't need give a very large value */
124 #define IBLND_TX_POOL 256
125 #define IBLND_FMR_POOL 256
126 #define IBLND_FMR_POOL_FLUSH 192
128 #define IBLND_RX_MSGS(c) \
129 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
130 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
131 #define IBLND_RX_MSG_PAGES(c) \
132 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
134 /* WRs and CQEs (per connection) */
135 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
136 #define IBLND_SEND_WRS(c) \
137 (((c->ibc_max_frags + 1) << IBLND_FRAG_SHIFT) * \
138 kiblnd_concurrent_sends(c->ibc_version, c->ibc_peer->ibp_ni))
139 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
143 /* o2iblnd can run over aliased interface */
145 #define KIB_IFNAME_SIZE IFALIASZ
147 #define KIB_IFNAME_SIZE 256
151 struct list_head ibd_list; /* chain on kib_devs */
152 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
153 __u32 ibd_ifip; /* IPoIB interface IP */
155 /* IPoIB interface name */
156 char ibd_ifname[KIB_IFNAME_SIZE];
157 int ibd_nnets; /* # nets extant */
159 unsigned long ibd_next_failover;
160 int ibd_failed_failover; /* # failover failures */
161 unsigned int ibd_failover; /* failover in progress */
162 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
163 struct list_head ibd_nets;
164 struct kib_hca_dev *ibd_hdev;
168 struct rdma_cm_id *ibh_cmid; /* listener cmid */
169 struct ib_device *ibh_ibdev; /* IB device */
170 int ibh_page_shift; /* page shift of current HCA */
171 int ibh_page_size; /* page size of current HCA */
172 __u64 ibh_page_mask; /* page mask of current HCA */
173 int ibh_mr_shift; /* bits shift of max MR size */
174 __u64 ibh_mr_size; /* size of MR */
175 struct ib_pd *ibh_pd; /* PD */
176 struct kib_dev *ibh_dev; /* owner */
177 atomic_t ibh_ref; /* refcount */
180 /** # of seconds to keep pool alive */
181 #define IBLND_POOL_DEADLINE 300
182 /** # of seconds to retry if allocation failed */
183 #define IBLND_POOL_RETRY 1
186 int ibp_npages; /* # pages */
187 struct page *ibp_pages[0]; /* page array */
193 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
194 int inc, struct kib_pool **pp_po);
195 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
196 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
197 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
201 #define IBLND_POOL_NAME_LEN 32
204 spinlock_t ps_lock; /* serialize */
205 struct kib_net *ps_net; /* network it belongs to */
206 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
207 struct list_head ps_pool_list; /* list of pools */
208 struct list_head ps_failed_pool_list;/* failed pool list */
209 unsigned long ps_next_retry; /* time stamp for retry if */
210 /* failed to allocate */
211 int ps_increasing; /* is allocating new pool */
212 int ps_pool_size; /* new pool size */
213 int ps_cpt; /* CPT id */
215 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
216 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
217 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
218 kib_ps_node_fini_t ps_node_fini; /* finalize node */
222 struct list_head po_list; /* chain on pool list */
223 struct list_head po_free_list; /* pre-allocated node */
224 struct kib_poolset *po_owner; /* pool_set of this pool */
225 unsigned long po_deadline; /* deadline of this pool */
226 int po_allocated; /* # of elements in use */
227 int po_failed; /* pool is created on failed HCA */
228 int po_size; /* # of pre-allocated elements */
231 struct kib_tx_poolset {
232 struct kib_poolset tps_poolset; /* pool-set */
233 __u64 tps_next_tx_cookie; /* cookie of TX */
237 struct kib_pool tpo_pool; /* pool */
238 struct kib_hca_dev *tpo_hdev; /* device for this pool */
239 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
240 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
243 struct kib_fmr_poolset {
244 spinlock_t fps_lock; /* serialize */
245 struct kib_net *fps_net; /* IB network */
246 struct list_head fps_pool_list; /* FMR pool list */
247 struct list_head fps_failed_pool_list;/* FMR pool list */
248 __u64 fps_version; /* validity stamp */
249 int fps_cpt; /* CPT id */
251 int fps_flush_trigger;
253 int fps_increasing; /* is allocating new pool */
254 unsigned long fps_next_retry; /* time stamp for retry if*/
255 /* failed to allocate */
258 struct kib_fast_reg_descriptor { /* For fast registration */
259 struct list_head frd_list;
260 struct ib_send_wr frd_inv_wr;
261 struct ib_reg_wr frd_fastreg_wr;
262 struct ib_mr *frd_mr;
266 struct kib_fmr_pool {
267 struct list_head fpo_list; /* chain on pool list */
268 struct kib_hca_dev *fpo_hdev; /* device for this pool */
269 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
272 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
274 struct { /* For fast registration */
275 struct list_head fpo_pool_list;
279 unsigned long fpo_deadline; /* deadline of this pool */
280 int fpo_failed; /* fmr pool is failed */
281 int fpo_map_count; /* # of mapped FMR */
286 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
287 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
288 struct kib_fast_reg_descriptor *fmr_frd;
293 struct list_head ibn_list; /* chain on struct kib_dev::ibd_nets */
294 __u64 ibn_incarnation;/* my epoch */
295 int ibn_init; /* initialisation state */
296 int ibn_shutdown; /* shutting down? */
298 atomic_t ibn_npeers; /* # peers extant */
299 atomic_t ibn_nconns; /* # connections extant */
301 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
302 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
304 struct kib_dev *ibn_dev; /* underlying IB device */
307 #define KIB_THREAD_SHIFT 16
308 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
309 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
310 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
312 struct kib_sched_info {
313 spinlock_t ibs_lock; /* serialise */
314 wait_queue_head_t ibs_waitq; /* schedulers sleep here */
315 struct list_head ibs_conns; /* conns to check for rx completions */
316 int ibs_nthreads; /* number of scheduler threads */
317 int ibs_nthreads_max; /* max allowed scheduler threads */
318 int ibs_cpt; /* CPT id */
322 int kib_init; /* initialisation state */
323 int kib_shutdown; /* shut down? */
324 struct list_head kib_devs; /* IB devices extant */
325 struct list_head kib_failed_devs; /* list head of failed devices */
326 wait_queue_head_t kib_failover_waitq; /* schedulers sleep here */
327 atomic_t kib_nthreads; /* # live threads */
328 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
329 struct list_head *kib_peers; /* hash table of all my known peers */
330 int kib_peer_hash_size; /* size of kib_peers */
331 void *kib_connd; /* the connd task (serialisation assertions) */
332 struct list_head kib_connd_conns; /* connections to setup/teardown */
333 struct list_head kib_connd_zombies; /* connections with zero refcount */
334 /* connections to reconnect */
335 struct list_head kib_reconn_list;
336 /* peers wait for reconnection */
337 struct list_head kib_reconn_wait;
339 * The second that peers are pulled out from \a kib_reconn_wait
342 time64_t kib_reconn_sec;
344 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
345 spinlock_t kib_connd_lock; /* serialise */
346 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
347 struct kib_sched_info **kib_scheds; /* percpt data for schedulers */
350 #define IBLND_INIT_NOTHING 0
351 #define IBLND_INIT_DATA 1
352 #define IBLND_INIT_ALL 2
354 /************************************************************************
355 * IB Wire message format.
356 * These are sent in sender's byte order (i.e. receiver flips).
359 struct kib_connparams {
360 __u16 ibcp_queue_depth;
361 __u16 ibcp_max_frags;
362 __u32 ibcp_max_msg_size;
365 struct kib_immediate_msg {
366 struct lnet_hdr ibim_hdr; /* portals header */
367 char ibim_payload[0]; /* piggy-backed payload */
370 struct kib_rdma_frag {
371 __u32 rf_nob; /* # bytes this frag */
372 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
375 struct kib_rdma_desc {
376 __u32 rd_key; /* local/remote key */
377 __u32 rd_nfrags; /* # fragments */
378 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
381 struct kib_putreq_msg {
382 struct lnet_hdr ibprm_hdr; /* portals header */
383 __u64 ibprm_cookie; /* opaque completion cookie */
386 struct kib_putack_msg {
387 __u64 ibpam_src_cookie; /* reflected completion cookie */
388 __u64 ibpam_dst_cookie; /* opaque completion cookie */
389 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
393 struct lnet_hdr ibgm_hdr; /* portals header */
394 __u64 ibgm_cookie; /* opaque completion cookie */
395 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
398 struct kib_completion_msg {
399 __u64 ibcm_cookie; /* opaque completion cookie */
400 __s32 ibcm_status; /* < 0 failure: >= 0 length */
404 /* First 2 fields fixed FOR ALL TIME */
405 __u32 ibm_magic; /* I'm an ibnal message */
406 __u16 ibm_version; /* this is my version number */
408 __u8 ibm_type; /* msg type */
409 __u8 ibm_credits; /* returned credits */
410 __u32 ibm_nob; /* # bytes in whole message */
411 __u32 ibm_cksum; /* checksum (0 == no checksum) */
412 __u64 ibm_srcnid; /* sender's NID */
413 __u64 ibm_srcstamp; /* sender's incarnation */
414 __u64 ibm_dstnid; /* destination's NID */
415 __u64 ibm_dststamp; /* destination's incarnation */
418 struct kib_connparams connparams;
419 struct kib_immediate_msg immediate;
420 struct kib_putreq_msg putreq;
421 struct kib_putack_msg putack;
422 struct kib_get_msg get;
423 struct kib_completion_msg completion;
427 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
429 #define IBLND_MSG_VERSION_1 0x11
430 #define IBLND_MSG_VERSION_2 0x12
431 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
433 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
434 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
435 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
436 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
437 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
438 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
439 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
440 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
441 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
442 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
445 __u32 ibr_magic; /* sender's magic */
446 __u16 ibr_version; /* sender's version */
447 __u8 ibr_why; /* reject reason */
448 __u8 ibr_padding; /* padding */
449 __u64 ibr_incarnation; /* incarnation of peer */
450 struct kib_connparams ibr_cp; /* connection parameters */
453 /* connection rejection reasons */
454 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
455 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
456 #define IBLND_REJECT_FATAL 3 /* Anything else */
457 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
458 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
459 /* peer's rdma frags doesn't match mine */
460 #define IBLND_REJECT_RDMA_FRAGS 6
461 /* peer's msg queue size doesn't match mine */
462 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
464 /***********************************************************************/
466 struct kib_rx { /* receive message */
467 struct list_head rx_list; /* queue for attention */
468 struct kib_conn *rx_conn; /* owning conn */
469 int rx_nob; /* # bytes received (-1 while posted) */
470 enum ib_wc_status rx_status; /* completion status */
471 struct kib_msg *rx_msg; /* message buffer (host vaddr) */
472 __u64 rx_msgaddr; /* message buffer (I/O addr) */
473 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap); /* for dma_unmap_single() */
474 struct ib_recv_wr rx_wrq; /* receive work item... */
475 struct ib_sge rx_sge; /* ...and its memory */
478 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
479 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
480 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
481 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give self back 1 reserved credit */
483 struct kib_tx { /* transmit message */
484 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
485 struct kib_tx_pool *tx_pool; /* pool I'm from */
486 struct kib_conn *tx_conn; /* owning conn */
487 short tx_sending; /* # tx callbacks outstanding */
488 short tx_queued; /* queued for sending */
489 short tx_waiting; /* waiting for peer */
490 int tx_status; /* LNET completion status */
491 unsigned long tx_deadline; /* completion deadline */
492 __u64 tx_cookie; /* completion cookie */
493 struct lnet_msg *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
494 struct kib_msg *tx_msg; /* message buffer (host vaddr) */
495 __u64 tx_msgaddr; /* message buffer (I/O addr) */
496 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap); /* for dma_unmap_single() */
497 int tx_nwrq; /* # send work items */
498 struct ib_rdma_wr *tx_wrq; /* send work items... */
499 struct ib_sge *tx_sge; /* ...and their memory */
500 struct kib_rdma_desc *tx_rd; /* rdma descriptor */
501 int tx_nfrags; /* # entries in... */
502 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
503 __u64 *tx_pages; /* rdma phys page addrs */
504 struct kib_fmr fmr; /* FMR */
505 int tx_dmadir; /* dma direction */
508 struct kib_connvars {
509 struct kib_msg cv_msg; /* connection-in-progress variables */
513 struct kib_sched_info *ibc_sched; /* scheduler information */
514 struct kib_peer *ibc_peer; /* owning peer */
515 struct kib_hca_dev *ibc_hdev; /* HCA bound on */
516 struct list_head ibc_list; /* stash on peer's conn list */
517 struct list_head ibc_sched_list; /* schedule for attention */
518 __u16 ibc_version; /* version of connection */
519 /* reconnect later */
520 __u16 ibc_reconnect:1;
521 __u64 ibc_incarnation; /* which instance of the peer */
522 atomic_t ibc_refcount; /* # users */
523 int ibc_state; /* what's happening */
524 int ibc_nsends_posted; /* # uncompleted sends */
525 int ibc_noops_posted; /* # uncompleted NOOPs */
526 int ibc_credits; /* # credits I have */
527 int ibc_outstanding_credits; /* # credits to return */
528 int ibc_reserved_credits; /* # ACK/DONE msg credits */
529 int ibc_comms_error; /* set on comms error */
530 /* connections queue depth */
531 __u16 ibc_queue_depth;
532 /* connections max frags */
534 unsigned int ibc_nrx:16; /* receive buffers owned */
535 unsigned int ibc_scheduled:1; /* scheduled for attention */
536 unsigned int ibc_ready:1; /* CQ callback fired */
537 unsigned long ibc_last_send; /* time of last send */
538 struct list_head ibc_connd_list; /* link chain for */
539 /* kiblnd_check_conns only */
540 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
541 struct list_head ibc_tx_noops; /* IBLND_MSG_NOOPs for */
542 /* IBLND_MSG_VERSION_1 */
543 struct list_head ibc_tx_queue; /* sends that need a credit */
544 struct list_head ibc_tx_queue_nocred; /* sends that don't need a */
546 struct list_head ibc_tx_queue_rsrvd; /* sends that need to */
547 /* reserve an ACK/DONE msg */
548 struct list_head ibc_active_txs; /* active tx awaiting completion */
549 spinlock_t ibc_lock; /* serialise */
550 struct kib_rx *ibc_rxs; /* the rx descs */
551 struct kib_pages *ibc_rx_pages; /* premapped rx msg pages */
553 struct rdma_cm_id *ibc_cmid; /* CM id */
554 struct ib_cq *ibc_cq; /* completion queue */
556 struct kib_connvars *ibc_connvars; /* in-progress connection state */
559 #define IBLND_CONN_INIT 0 /* being initialised */
560 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
561 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
562 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
563 #define IBLND_CONN_CLOSING 4 /* being closed */
564 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
567 struct list_head ibp_list; /* stash on global peer list */
568 lnet_nid_t ibp_nid; /* who's on the other end(s) */
569 struct lnet_ni *ibp_ni; /* LNet interface */
570 struct list_head ibp_conns; /* all active connections */
571 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
572 __u64 ibp_incarnation; /* incarnation of peer */
573 /* when (in jiffies) I was last alive */
574 unsigned long ibp_last_alive;
576 atomic_t ibp_refcount;
577 /* version of peer */
579 /* current passive connection attempts */
580 unsigned short ibp_accepting;
581 /* current active connection attempts */
582 unsigned short ibp_connecting;
583 /* reconnect this peer later */
584 unsigned short ibp_reconnecting:1;
585 /* counter of how many times we triggered a conn race */
586 unsigned char ibp_races;
587 /* # consecutive reconnection attempts to this peer */
588 unsigned int ibp_reconnected;
589 /* errno on closing this peer */
591 /* max map_on_demand */
593 /* max_peer_credits */
594 __u16 ibp_queue_depth;
597 extern struct kib_data kiblnd_data;
599 void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
601 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
603 /* max # of fragments configured by user */
605 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
607 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
610 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
611 mod = tunables->lnd_map_on_demand;
612 return mod ? mod : IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT;
616 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
618 return version == IBLND_MSG_VERSION_1 ?
619 (IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT) :
620 kiblnd_cfg_rdma_frags(ni);
624 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
626 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
627 int concurrent_sends;
629 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
630 concurrent_sends = tunables->lnd_concurrent_sends;
632 if (version == IBLND_MSG_VERSION_1) {
633 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
634 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
636 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
637 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
640 return concurrent_sends;
644 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
646 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
647 atomic_inc(&hdev->ibh_ref);
651 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
653 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
654 if (atomic_dec_and_test(&hdev->ibh_ref))
655 kiblnd_hdev_destroy(hdev);
659 kiblnd_dev_can_failover(struct kib_dev *dev)
661 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
664 if (!*kiblnd_tunables.kib_dev_failover) /* disabled */
667 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
670 return dev->ibd_can_failover;
673 #define kiblnd_conn_addref(conn) \
675 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
676 (conn), atomic_read(&(conn)->ibc_refcount)); \
677 atomic_inc(&(conn)->ibc_refcount); \
680 #define kiblnd_conn_decref(conn) \
682 unsigned long flags; \
684 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
685 (conn), atomic_read(&(conn)->ibc_refcount)); \
686 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
687 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
688 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
689 list_add_tail(&(conn)->ibc_list, \
690 &kiblnd_data.kib_connd_zombies); \
691 wake_up(&kiblnd_data.kib_connd_waitq); \
692 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
696 #define kiblnd_peer_addref(peer) \
698 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
699 (peer), libcfs_nid2str((peer)->ibp_nid), \
700 atomic_read(&(peer)->ibp_refcount)); \
701 atomic_inc(&(peer)->ibp_refcount); \
704 #define kiblnd_peer_decref(peer) \
706 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
707 (peer), libcfs_nid2str((peer)->ibp_nid), \
708 atomic_read(&(peer)->ibp_refcount)); \
709 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
710 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
711 kiblnd_destroy_peer(peer); \
715 kiblnd_peer_connecting(struct kib_peer *peer)
717 return peer->ibp_connecting ||
718 peer->ibp_reconnecting ||
723 kiblnd_peer_idle(struct kib_peer *peer)
725 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
728 static inline struct list_head *
729 kiblnd_nid2peerlist(lnet_nid_t nid)
732 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
734 return &kiblnd_data.kib_peers[hash];
738 kiblnd_peer_active(struct kib_peer *peer)
740 /* Am I in the peer hash table? */
741 return !list_empty(&peer->ibp_list);
744 static inline struct kib_conn *
745 kiblnd_get_conn_locked(struct kib_peer *peer)
747 LASSERT(!list_empty(&peer->ibp_conns));
749 /* just return the first connection */
750 return list_entry(peer->ibp_conns.next, struct kib_conn, ibc_list);
754 kiblnd_send_keepalive(struct kib_conn *conn)
756 return (*kiblnd_tunables.kib_keepalive > 0) &&
757 cfs_time_after(jiffies, conn->ibc_last_send +
758 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
763 kiblnd_need_noop(struct kib_conn *conn)
765 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
766 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
768 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
769 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
771 if (conn->ibc_outstanding_credits <
772 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
773 !kiblnd_send_keepalive(conn))
774 return 0; /* No need to send NOOP */
776 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
777 if (!list_empty(&conn->ibc_tx_queue_nocred))
778 return 0; /* NOOP can be piggybacked */
780 /* No tx to piggyback NOOP onto or no credit to send a tx */
781 return (list_empty(&conn->ibc_tx_queue) ||
785 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
786 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
787 !conn->ibc_credits) /* no credit */
790 if (conn->ibc_credits == 1 && /* last credit reserved for */
791 !conn->ibc_outstanding_credits) /* giving back credits */
794 /* No tx to piggyback NOOP onto or no credit to send a tx */
795 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
799 kiblnd_abort_receives(struct kib_conn *conn)
801 ib_modify_qp(conn->ibc_cmid->qp,
802 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
805 static inline const char *
806 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
808 if (q == &conn->ibc_tx_queue)
811 if (q == &conn->ibc_tx_queue_rsrvd)
812 return "tx_queue_rsrvd";
814 if (q == &conn->ibc_tx_queue_nocred)
815 return "tx_queue_nocred";
817 if (q == &conn->ibc_active_txs)
824 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the */
825 /* lowest bits of the work request id to stash the work item type. */
827 #define IBLND_WID_INVAL 0
828 #define IBLND_WID_TX 1
829 #define IBLND_WID_RX 2
830 #define IBLND_WID_RDMA 3
831 #define IBLND_WID_MR 4
832 #define IBLND_WID_MASK 7UL
835 kiblnd_ptr2wreqid(void *ptr, int type)
837 unsigned long lptr = (unsigned long)ptr;
839 LASSERT(!(lptr & IBLND_WID_MASK));
840 LASSERT(!(type & ~IBLND_WID_MASK));
841 return (__u64)(lptr | type);
845 kiblnd_wreqid2ptr(__u64 wreqid)
847 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
851 kiblnd_wreqid2type(__u64 wreqid)
853 return wreqid & IBLND_WID_MASK;
857 kiblnd_set_conn_state(struct kib_conn *conn, int state)
859 conn->ibc_state = state;
864 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
866 msg->ibm_type = type;
867 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
871 kiblnd_rd_size(struct kib_rdma_desc *rd)
876 for (i = size = 0; i < rd->rd_nfrags; i++)
877 size += rd->rd_frags[i].rf_nob;
883 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
885 return rd->rd_frags[index].rf_addr;
889 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
891 return rd->rd_frags[index].rf_nob;
895 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
901 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
903 if (nob < rd->rd_frags[index].rf_nob) {
904 rd->rd_frags[index].rf_addr += nob;
905 rd->rd_frags[index].rf_nob -= nob;
914 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
916 LASSERT(msgtype == IBLND_MSG_GET_REQ ||
917 msgtype == IBLND_MSG_PUT_ACK);
919 return msgtype == IBLND_MSG_GET_REQ ?
920 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
921 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
925 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
927 return ib_dma_mapping_error(dev, dma_addr);
930 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
931 void *msg, size_t size,
932 enum dma_data_direction direction)
934 return ib_dma_map_single(dev, msg, size, direction);
937 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
938 __u64 addr, size_t size,
939 enum dma_data_direction direction)
941 ib_dma_unmap_single(dev, addr, size, direction);
944 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
945 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
947 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
948 struct scatterlist *sg, int nents,
949 enum dma_data_direction direction)
951 return ib_dma_map_sg(dev, sg, nents, direction);
954 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
955 struct scatterlist *sg, int nents,
956 enum dma_data_direction direction)
958 ib_dma_unmap_sg(dev, sg, nents, direction);
961 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
962 struct scatterlist *sg)
964 return ib_sg_dma_address(dev, sg);
967 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
968 struct scatterlist *sg)
970 return ib_sg_dma_len(dev, sg);
973 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly */
974 /* right because OFED1.2 defines it as const, to use it we have to add */
975 /* (void *) cast to overcome "const" */
977 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
978 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
980 void kiblnd_map_rx_descs(struct kib_conn *conn);
981 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
982 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
983 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
985 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
986 struct kib_rdma_desc *rd, __u32 nob, __u64 iov,
987 struct kib_fmr *fmr);
988 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
990 int kiblnd_tunables_setup(struct lnet_ni *ni);
991 void kiblnd_tunables_init(void);
993 int kiblnd_connd(void *arg);
994 int kiblnd_scheduler(void *arg);
995 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
996 int kiblnd_failover_thread(void *arg);
998 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1000 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1001 struct rdma_cm_event *event);
1002 int kiblnd_translate_mtu(int value);
1004 int kiblnd_dev_failover(struct kib_dev *dev);
1005 int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer **peerp,
1007 void kiblnd_destroy_peer(struct kib_peer *peer);
1008 bool kiblnd_reconnect_peer(struct kib_peer *peer);
1009 void kiblnd_destroy_dev(struct kib_dev *dev);
1010 void kiblnd_unlink_peer_locked(struct kib_peer *peer);
1011 struct kib_peer *kiblnd_find_peer_locked(lnet_nid_t nid);
1012 int kiblnd_close_stale_conns_locked(struct kib_peer *peer,
1013 int version, __u64 incarnation);
1014 int kiblnd_close_peer_conns_locked(struct kib_peer *peer, int why);
1016 struct kib_conn *kiblnd_create_conn(struct kib_peer *peer,
1017 struct rdma_cm_id *cmid,
1018 int state, int version);
1019 void kiblnd_destroy_conn(struct kib_conn *conn);
1020 void kiblnd_close_conn(struct kib_conn *conn, int error);
1021 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1023 void kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid);
1024 void kiblnd_txlist_done(struct lnet_ni *ni, struct list_head *txlist,
1027 void kiblnd_qp_event(struct ib_event *event, void *arg);
1028 void kiblnd_cq_event(struct ib_event *event, void *arg);
1029 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1031 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
1032 int credits, lnet_nid_t dstnid, __u64 dststamp);
1033 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1034 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1036 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1037 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1038 int delayed, struct iov_iter *to, unsigned int rlen);