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, 2015, 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 #include <linux/string.h>
41 #include <linux/stat.h>
42 #include <linux/errno.h>
43 #include <linux/unistd.h>
44 #include <linux/uio.h>
45 #include <linux/uaccess.h>
50 #include <linux/file.h>
51 #include <linux/list.h>
52 #include <linux/kmod.h>
53 #include <linux/sysctl.h>
54 #include <linux/pci.h>
59 #include <rdma/rdma_cm.h>
60 #include <rdma/ib_cm.h>
61 #include <rdma/ib_verbs.h>
62 #include <rdma/ib_fmr_pool.h>
64 #define DEBUG_SUBSYSTEM S_LND
66 #include "../../../include/linux/libcfs/libcfs.h"
67 #include "../../../include/linux/lnet/lnet.h"
68 #include "../../../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_mr *ibh_mrs; /* global MR */
176 struct ib_pd *ibh_pd; /* PD */
177 struct kib_dev *ibh_dev; /* owner */
178 atomic_t ibh_ref; /* refcount */
181 /** # of seconds to keep pool alive */
182 #define IBLND_POOL_DEADLINE 300
183 /** # of seconds to retry if allocation failed */
184 #define IBLND_POOL_RETRY 1
187 int ibp_npages; /* # pages */
188 struct page *ibp_pages[0]; /* page array */
194 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
195 int inc, struct kib_pool **pp_po);
196 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
197 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
198 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
202 #define IBLND_POOL_NAME_LEN 32
205 spinlock_t ps_lock; /* serialize */
206 struct kib_net *ps_net; /* network it belongs to */
207 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
208 struct list_head ps_pool_list; /* list of pools */
209 struct list_head ps_failed_pool_list;/* failed pool list */
210 unsigned long ps_next_retry; /* time stamp for retry if */
211 /* failed to allocate */
212 int ps_increasing; /* is allocating new pool */
213 int ps_pool_size; /* new pool size */
214 int ps_cpt; /* CPT id */
216 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
217 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
218 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
219 kib_ps_node_fini_t ps_node_fini; /* finalize node */
223 struct list_head po_list; /* chain on pool list */
224 struct list_head po_free_list; /* pre-allocated node */
225 struct kib_poolset *po_owner; /* pool_set of this pool */
226 unsigned long po_deadline; /* deadline of this pool */
227 int po_allocated; /* # of elements in use */
228 int po_failed; /* pool is created on failed HCA */
229 int po_size; /* # of pre-allocated elements */
232 struct kib_tx_poolset {
233 struct kib_poolset tps_poolset; /* pool-set */
234 __u64 tps_next_tx_cookie; /* cookie of TX */
238 struct kib_pool tpo_pool; /* pool */
239 struct kib_hca_dev *tpo_hdev; /* device for this pool */
240 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
241 struct kib_pages *tpo_tx_pages; /* premapped tx msg pages */
244 struct kib_fmr_poolset {
245 spinlock_t fps_lock; /* serialize */
246 struct kib_net *fps_net; /* IB network */
247 struct list_head fps_pool_list; /* FMR pool list */
248 struct list_head fps_failed_pool_list;/* FMR pool list */
249 __u64 fps_version; /* validity stamp */
250 int fps_cpt; /* CPT id */
252 int fps_flush_trigger;
254 int fps_increasing; /* is allocating new pool */
255 unsigned long fps_next_retry; /* time stamp for retry if*/
256 /* failed to allocate */
259 struct kib_fast_reg_descriptor { /* For fast registration */
260 struct list_head frd_list;
261 struct ib_send_wr frd_inv_wr;
262 struct ib_reg_wr frd_fastreg_wr;
263 struct ib_mr *frd_mr;
267 struct kib_fmr_pool {
268 struct list_head fpo_list; /* chain on pool list */
269 struct kib_hca_dev *fpo_hdev; /* device for this pool */
270 struct kib_fmr_poolset *fpo_owner; /* owner of this pool */
273 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
275 struct { /* For fast registration */
276 struct list_head fpo_pool_list;
280 unsigned long fpo_deadline; /* deadline of this pool */
281 int fpo_failed; /* fmr pool is failed */
282 int fpo_map_count; /* # of mapped FMR */
287 struct kib_fmr_pool *fmr_pool; /* pool of FMR */
288 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
289 struct kib_fast_reg_descriptor *fmr_frd;
294 struct list_head ibn_list; /* chain on struct kib_dev::ibd_nets */
295 __u64 ibn_incarnation;/* my epoch */
296 int ibn_init; /* initialisation state */
297 int ibn_shutdown; /* shutting down? */
299 atomic_t ibn_npeers; /* # peers extant */
300 atomic_t ibn_nconns; /* # connections extant */
302 struct kib_tx_poolset **ibn_tx_ps; /* tx pool-set */
303 struct kib_fmr_poolset **ibn_fmr_ps; /* fmr pool-set */
305 struct kib_dev *ibn_dev; /* underlying IB device */
308 #define KIB_THREAD_SHIFT 16
309 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
310 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
311 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
313 struct kib_sched_info {
314 spinlock_t ibs_lock; /* serialise */
315 wait_queue_head_t ibs_waitq; /* schedulers sleep here */
316 struct list_head ibs_conns; /* conns to check for rx completions */
317 int ibs_nthreads; /* number of scheduler threads */
318 int ibs_nthreads_max; /* max allowed scheduler threads */
319 int ibs_cpt; /* CPT id */
323 int kib_init; /* initialisation state */
324 int kib_shutdown; /* shut down? */
325 struct list_head kib_devs; /* IB devices extant */
326 struct list_head kib_failed_devs; /* list head of failed devices */
327 wait_queue_head_t kib_failover_waitq; /* schedulers sleep here */
328 atomic_t kib_nthreads; /* # live threads */
329 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
330 struct list_head *kib_peers; /* hash table of all my known peers */
331 int kib_peer_hash_size; /* size of kib_peers */
332 void *kib_connd; /* the connd task (serialisation assertions) */
333 struct list_head kib_connd_conns; /* connections to setup/teardown */
334 struct list_head kib_connd_zombies; /* connections with zero refcount */
335 /* connections to reconnect */
336 struct list_head kib_reconn_list;
337 /* peers wait for reconnection */
338 struct list_head kib_reconn_wait;
340 * The second that peers are pulled out from \a kib_reconn_wait
343 time64_t kib_reconn_sec;
345 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
346 spinlock_t kib_connd_lock; /* serialise */
347 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
348 struct kib_sched_info **kib_scheds; /* percpt data for schedulers */
351 #define IBLND_INIT_NOTHING 0
352 #define IBLND_INIT_DATA 1
353 #define IBLND_INIT_ALL 2
355 /************************************************************************
356 * IB Wire message format.
357 * These are sent in sender's byte order (i.e. receiver flips).
360 struct kib_connparams {
361 __u16 ibcp_queue_depth;
362 __u16 ibcp_max_frags;
363 __u32 ibcp_max_msg_size;
366 struct kib_immediate_msg {
367 lnet_hdr_t ibim_hdr; /* portals header */
368 char ibim_payload[0]; /* piggy-backed payload */
371 struct kib_rdma_frag {
372 __u32 rf_nob; /* # bytes this frag */
373 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
376 struct kib_rdma_desc {
377 __u32 rd_key; /* local/remote key */
378 __u32 rd_nfrags; /* # fragments */
379 struct kib_rdma_frag rd_frags[0]; /* buffer frags */
382 struct kib_putreq_msg {
383 lnet_hdr_t ibprm_hdr; /* portals header */
384 __u64 ibprm_cookie; /* opaque completion cookie */
387 struct kib_putack_msg {
388 __u64 ibpam_src_cookie; /* reflected completion cookie */
389 __u64 ibpam_dst_cookie; /* opaque completion cookie */
390 struct kib_rdma_desc ibpam_rd; /* sender's sink buffer */
394 lnet_hdr_t ibgm_hdr; /* portals header */
395 __u64 ibgm_cookie; /* opaque completion cookie */
396 struct kib_rdma_desc ibgm_rd; /* rdma descriptor */
399 struct kib_completion_msg {
400 __u64 ibcm_cookie; /* opaque completion cookie */
401 __s32 ibcm_status; /* < 0 failure: >= 0 length */
405 /* First 2 fields fixed FOR ALL TIME */
406 __u32 ibm_magic; /* I'm an ibnal message */
407 __u16 ibm_version; /* this is my version number */
409 __u8 ibm_type; /* msg type */
410 __u8 ibm_credits; /* returned credits */
411 __u32 ibm_nob; /* # bytes in whole message */
412 __u32 ibm_cksum; /* checksum (0 == no checksum) */
413 __u64 ibm_srcnid; /* sender's NID */
414 __u64 ibm_srcstamp; /* sender's incarnation */
415 __u64 ibm_dstnid; /* destination's NID */
416 __u64 ibm_dststamp; /* destination's incarnation */
419 struct kib_connparams connparams;
420 struct kib_immediate_msg immediate;
421 struct kib_putreq_msg putreq;
422 struct kib_putack_msg putack;
423 struct kib_get_msg get;
424 struct kib_completion_msg completion;
428 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
430 #define IBLND_MSG_VERSION_1 0x11
431 #define IBLND_MSG_VERSION_2 0x12
432 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
434 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
435 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
436 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
437 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
438 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
439 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
440 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
441 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
442 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
443 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
446 __u32 ibr_magic; /* sender's magic */
447 __u16 ibr_version; /* sender's version */
448 __u8 ibr_why; /* reject reason */
449 __u8 ibr_padding; /* padding */
450 __u64 ibr_incarnation; /* incarnation of peer */
451 struct kib_connparams ibr_cp; /* connection parameters */
454 /* connection rejection reasons */
455 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
456 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
457 #define IBLND_REJECT_FATAL 3 /* Anything else */
458 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
459 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
460 /* peer's rdma frags doesn't match mine */
461 #define IBLND_REJECT_RDMA_FRAGS 6
462 /* peer's msg queue size doesn't match mine */
463 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
465 /***********************************************************************/
467 struct kib_rx { /* receive message */
468 struct list_head rx_list; /* queue for attention */
469 struct kib_conn *rx_conn; /* owning conn */
470 int rx_nob; /* # bytes received (-1 while posted) */
471 enum ib_wc_status rx_status; /* completion status */
472 struct kib_msg *rx_msg; /* message buffer (host vaddr) */
473 __u64 rx_msgaddr; /* message buffer (I/O addr) */
474 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap); /* for dma_unmap_single() */
475 struct ib_recv_wr rx_wrq; /* receive work item... */
476 struct ib_sge rx_sge; /* ...and its memory */
479 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
480 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
481 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
482 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give self back 1 reserved credit */
484 struct kib_tx { /* transmit message */
485 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
486 struct kib_tx_pool *tx_pool; /* pool I'm from */
487 struct kib_conn *tx_conn; /* owning conn */
488 short tx_sending; /* # tx callbacks outstanding */
489 short tx_queued; /* queued for sending */
490 short tx_waiting; /* waiting for peer */
491 int tx_status; /* LNET completion status */
492 unsigned long tx_deadline; /* completion deadline */
493 __u64 tx_cookie; /* completion cookie */
494 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
495 struct kib_msg *tx_msg; /* message buffer (host vaddr) */
496 __u64 tx_msgaddr; /* message buffer (I/O addr) */
497 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap); /* for dma_unmap_single() */
498 int tx_nwrq; /* # send work items */
499 struct ib_rdma_wr *tx_wrq; /* send work items... */
500 struct ib_sge *tx_sge; /* ...and their memory */
501 struct kib_rdma_desc *tx_rd; /* rdma descriptor */
502 int tx_nfrags; /* # entries in... */
503 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
504 __u64 *tx_pages; /* rdma phys page addrs */
505 struct kib_fmr fmr; /* FMR */
506 int tx_dmadir; /* dma direction */
509 struct kib_connvars {
510 struct kib_msg cv_msg; /* connection-in-progress variables */
514 struct kib_sched_info *ibc_sched; /* scheduler information */
515 struct kib_peer *ibc_peer; /* owning peer */
516 struct kib_hca_dev *ibc_hdev; /* HCA bound on */
517 struct list_head ibc_list; /* stash on peer's conn list */
518 struct list_head ibc_sched_list; /* schedule for attention */
519 __u16 ibc_version; /* version of connection */
520 /* reconnect later */
521 __u16 ibc_reconnect:1;
522 __u64 ibc_incarnation; /* which instance of the peer */
523 atomic_t ibc_refcount; /* # users */
524 int ibc_state; /* what's happening */
525 int ibc_nsends_posted; /* # uncompleted sends */
526 int ibc_noops_posted; /* # uncompleted NOOPs */
527 int ibc_credits; /* # credits I have */
528 int ibc_outstanding_credits; /* # credits to return */
529 int ibc_reserved_credits; /* # ACK/DONE msg credits */
530 int ibc_comms_error; /* set on comms error */
531 /* connections queue depth */
532 __u16 ibc_queue_depth;
533 /* connections max frags */
535 unsigned int ibc_nrx:16; /* receive buffers owned */
536 unsigned int ibc_scheduled:1; /* scheduled for attention */
537 unsigned int ibc_ready:1; /* CQ callback fired */
538 unsigned long ibc_last_send; /* time of last send */
539 struct list_head ibc_connd_list; /* link chain for */
540 /* kiblnd_check_conns only */
541 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
542 struct list_head ibc_tx_noops; /* IBLND_MSG_NOOPs for */
543 /* IBLND_MSG_VERSION_1 */
544 struct list_head ibc_tx_queue; /* sends that need a credit */
545 struct list_head ibc_tx_queue_nocred; /* sends that don't need a */
547 struct list_head ibc_tx_queue_rsrvd; /* sends that need to */
548 /* reserve an ACK/DONE msg */
549 struct list_head ibc_active_txs; /* active tx awaiting completion */
550 spinlock_t ibc_lock; /* serialise */
551 struct kib_rx *ibc_rxs; /* the rx descs */
552 struct kib_pages *ibc_rx_pages; /* premapped rx msg pages */
554 struct rdma_cm_id *ibc_cmid; /* CM id */
555 struct ib_cq *ibc_cq; /* completion queue */
557 struct kib_connvars *ibc_connvars; /* in-progress connection state */
560 #define IBLND_CONN_INIT 0 /* being initialised */
561 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
562 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
563 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
564 #define IBLND_CONN_CLOSING 4 /* being closed */
565 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
568 struct list_head ibp_list; /* stash on global peer list */
569 lnet_nid_t ibp_nid; /* who's on the other end(s) */
570 lnet_ni_t *ibp_ni; /* LNet interface */
571 struct list_head ibp_conns; /* all active connections */
572 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
573 __u64 ibp_incarnation; /* incarnation of peer */
574 /* when (in jiffies) I was last alive */
575 unsigned long ibp_last_alive;
577 atomic_t ibp_refcount;
578 /* version of peer */
580 /* current passive connection attempts */
581 unsigned short ibp_accepting;
582 /* current active connection attempts */
583 unsigned short ibp_connecting;
584 /* reconnect this peer later */
585 unsigned short ibp_reconnecting:1;
586 /* counter of how many times we triggered a conn race */
587 unsigned char ibp_races;
588 /* # consecutive reconnection attempts to this peer */
589 unsigned int ibp_reconnected;
590 /* errno on closing this peer */
592 /* max map_on_demand */
594 /* max_peer_credits */
595 __u16 ibp_queue_depth;
598 extern struct kib_data kiblnd_data;
600 void kiblnd_hdev_destroy(struct kib_hca_dev *hdev);
602 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
604 /* max # of fragments configured by user */
606 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
608 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
611 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
612 mod = tunables->lnd_map_on_demand;
613 return mod ? mod : IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT;
617 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
619 return version == IBLND_MSG_VERSION_1 ?
620 (IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT) :
621 kiblnd_cfg_rdma_frags(ni);
625 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
627 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
628 int concurrent_sends;
630 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
631 concurrent_sends = tunables->lnd_concurrent_sends;
633 if (version == IBLND_MSG_VERSION_1) {
634 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
635 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
637 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
638 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
641 return concurrent_sends;
645 kiblnd_hdev_addref_locked(struct kib_hca_dev *hdev)
647 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
648 atomic_inc(&hdev->ibh_ref);
652 kiblnd_hdev_decref(struct kib_hca_dev *hdev)
654 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
655 if (atomic_dec_and_test(&hdev->ibh_ref))
656 kiblnd_hdev_destroy(hdev);
660 kiblnd_dev_can_failover(struct kib_dev *dev)
662 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
665 if (!*kiblnd_tunables.kib_dev_failover) /* disabled */
668 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
671 return dev->ibd_can_failover;
674 #define kiblnd_conn_addref(conn) \
676 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
677 (conn), atomic_read(&(conn)->ibc_refcount)); \
678 atomic_inc(&(conn)->ibc_refcount); \
681 #define kiblnd_conn_decref(conn) \
683 unsigned long flags; \
685 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
686 (conn), atomic_read(&(conn)->ibc_refcount)); \
687 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
688 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
689 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
690 list_add_tail(&(conn)->ibc_list, \
691 &kiblnd_data.kib_connd_zombies); \
692 wake_up(&kiblnd_data.kib_connd_waitq); \
693 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
697 #define kiblnd_peer_addref(peer) \
699 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
700 (peer), libcfs_nid2str((peer)->ibp_nid), \
701 atomic_read(&(peer)->ibp_refcount)); \
702 atomic_inc(&(peer)->ibp_refcount); \
705 #define kiblnd_peer_decref(peer) \
707 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
708 (peer), libcfs_nid2str((peer)->ibp_nid), \
709 atomic_read(&(peer)->ibp_refcount)); \
710 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
711 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
712 kiblnd_destroy_peer(peer); \
716 kiblnd_peer_connecting(struct kib_peer *peer)
718 return peer->ibp_connecting ||
719 peer->ibp_reconnecting ||
724 kiblnd_peer_idle(struct kib_peer *peer)
726 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
729 static inline struct list_head *
730 kiblnd_nid2peerlist(lnet_nid_t nid)
733 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
735 return &kiblnd_data.kib_peers[hash];
739 kiblnd_peer_active(struct kib_peer *peer)
741 /* Am I in the peer hash table? */
742 return !list_empty(&peer->ibp_list);
745 static inline struct kib_conn *
746 kiblnd_get_conn_locked(struct kib_peer *peer)
748 LASSERT(!list_empty(&peer->ibp_conns));
750 /* just return the first connection */
751 return list_entry(peer->ibp_conns.next, struct kib_conn, ibc_list);
755 kiblnd_send_keepalive(struct kib_conn *conn)
757 return (*kiblnd_tunables.kib_keepalive > 0) &&
758 cfs_time_after(jiffies, conn->ibc_last_send +
759 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
764 kiblnd_need_noop(struct kib_conn *conn)
766 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
767 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
769 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
770 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
772 if (conn->ibc_outstanding_credits <
773 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
774 !kiblnd_send_keepalive(conn))
775 return 0; /* No need to send NOOP */
777 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
778 if (!list_empty(&conn->ibc_tx_queue_nocred))
779 return 0; /* NOOP can be piggybacked */
781 /* No tx to piggyback NOOP onto or no credit to send a tx */
782 return (list_empty(&conn->ibc_tx_queue) ||
786 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
787 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
788 !conn->ibc_credits) /* no credit */
791 if (conn->ibc_credits == 1 && /* last credit reserved for */
792 !conn->ibc_outstanding_credits) /* giving back credits */
795 /* No tx to piggyback NOOP onto or no credit to send a tx */
796 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
800 kiblnd_abort_receives(struct kib_conn *conn)
802 ib_modify_qp(conn->ibc_cmid->qp,
803 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
806 static inline const char *
807 kiblnd_queue2str(struct kib_conn *conn, struct list_head *q)
809 if (q == &conn->ibc_tx_queue)
812 if (q == &conn->ibc_tx_queue_rsrvd)
813 return "tx_queue_rsrvd";
815 if (q == &conn->ibc_tx_queue_nocred)
816 return "tx_queue_nocred";
818 if (q == &conn->ibc_active_txs)
825 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the */
826 /* lowest bits of the work request id to stash the work item type. */
828 #define IBLND_WID_INVAL 0
829 #define IBLND_WID_TX 1
830 #define IBLND_WID_RX 2
831 #define IBLND_WID_RDMA 3
832 #define IBLND_WID_MR 4
833 #define IBLND_WID_MASK 7UL
836 kiblnd_ptr2wreqid(void *ptr, int type)
838 unsigned long lptr = (unsigned long)ptr;
840 LASSERT(!(lptr & IBLND_WID_MASK));
841 LASSERT(!(type & ~IBLND_WID_MASK));
842 return (__u64)(lptr | type);
846 kiblnd_wreqid2ptr(__u64 wreqid)
848 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
852 kiblnd_wreqid2type(__u64 wreqid)
854 return wreqid & IBLND_WID_MASK;
858 kiblnd_set_conn_state(struct kib_conn *conn, int state)
860 conn->ibc_state = state;
865 kiblnd_init_msg(struct kib_msg *msg, int type, int body_nob)
867 msg->ibm_type = type;
868 msg->ibm_nob = offsetof(struct kib_msg, ibm_u) + body_nob;
872 kiblnd_rd_size(struct kib_rdma_desc *rd)
877 for (i = size = 0; i < rd->rd_nfrags; i++)
878 size += rd->rd_frags[i].rf_nob;
884 kiblnd_rd_frag_addr(struct kib_rdma_desc *rd, int index)
886 return rd->rd_frags[index].rf_addr;
890 kiblnd_rd_frag_size(struct kib_rdma_desc *rd, int index)
892 return rd->rd_frags[index].rf_nob;
896 kiblnd_rd_frag_key(struct kib_rdma_desc *rd, int index)
902 kiblnd_rd_consume_frag(struct kib_rdma_desc *rd, int index, __u32 nob)
904 if (nob < rd->rd_frags[index].rf_nob) {
905 rd->rd_frags[index].rf_addr += nob;
906 rd->rd_frags[index].rf_nob -= nob;
915 kiblnd_rd_msg_size(struct kib_rdma_desc *rd, int msgtype, int n)
917 LASSERT(msgtype == IBLND_MSG_GET_REQ ||
918 msgtype == IBLND_MSG_PUT_ACK);
920 return msgtype == IBLND_MSG_GET_REQ ?
921 offsetof(struct kib_get_msg, ibgm_rd.rd_frags[n]) :
922 offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[n]);
926 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
928 return ib_dma_mapping_error(dev, dma_addr);
931 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
932 void *msg, size_t size,
933 enum dma_data_direction direction)
935 return ib_dma_map_single(dev, msg, size, direction);
938 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
939 __u64 addr, size_t size,
940 enum dma_data_direction direction)
942 ib_dma_unmap_single(dev, addr, size, direction);
945 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
946 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
948 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
949 struct scatterlist *sg, int nents,
950 enum dma_data_direction direction)
952 return ib_dma_map_sg(dev, sg, nents, direction);
955 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
956 struct scatterlist *sg, int nents,
957 enum dma_data_direction direction)
959 ib_dma_unmap_sg(dev, sg, nents, direction);
962 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
963 struct scatterlist *sg)
965 return ib_sg_dma_address(dev, sg);
968 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
969 struct scatterlist *sg)
971 return ib_sg_dma_len(dev, sg);
974 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly */
975 /* right because OFED1.2 defines it as const, to use it we have to add */
976 /* (void *) cast to overcome "const" */
978 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
979 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
981 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd,
982 int negotiated_nfrags);
983 void kiblnd_map_rx_descs(struct kib_conn *conn);
984 void kiblnd_unmap_rx_descs(struct kib_conn *conn);
985 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node);
986 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps);
988 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
989 struct kib_rdma_desc *rd, __u32 nob, __u64 iov,
990 struct kib_fmr *fmr);
991 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status);
993 int kiblnd_tunables_setup(struct lnet_ni *ni);
994 void kiblnd_tunables_init(void);
996 int kiblnd_connd(void *arg);
997 int kiblnd_scheduler(void *arg);
998 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
999 int kiblnd_failover_thread(void *arg);
1001 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages);
1003 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1004 struct rdma_cm_event *event);
1005 int kiblnd_translate_mtu(int value);
1007 int kiblnd_dev_failover(struct kib_dev *dev);
1008 int kiblnd_create_peer(lnet_ni_t *ni, struct kib_peer **peerp, lnet_nid_t nid);
1009 void kiblnd_destroy_peer(struct kib_peer *peer);
1010 bool kiblnd_reconnect_peer(struct kib_peer *peer);
1011 void kiblnd_destroy_dev(struct kib_dev *dev);
1012 void kiblnd_unlink_peer_locked(struct kib_peer *peer);
1013 struct kib_peer *kiblnd_find_peer_locked(lnet_nid_t nid);
1014 int kiblnd_close_stale_conns_locked(struct kib_peer *peer,
1015 int version, __u64 incarnation);
1016 int kiblnd_close_peer_conns_locked(struct kib_peer *peer, int why);
1018 struct kib_conn *kiblnd_create_conn(struct kib_peer *peer,
1019 struct rdma_cm_id *cmid,
1020 int state, int version);
1021 void kiblnd_destroy_conn(struct kib_conn *conn);
1022 void kiblnd_close_conn(struct kib_conn *conn, int error);
1023 void kiblnd_close_conn_locked(struct kib_conn *conn, int error);
1025 void kiblnd_launch_tx(lnet_ni_t *ni, struct kib_tx *tx, lnet_nid_t nid);
1026 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist,
1029 void kiblnd_qp_event(struct ib_event *event, void *arg);
1030 void kiblnd_cq_event(struct ib_event *event, void *arg);
1031 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1033 void kiblnd_pack_msg(lnet_ni_t *ni, struct kib_msg *msg, int version,
1034 int credits, lnet_nid_t dstnid, __u64 dststamp);
1035 int kiblnd_unpack_msg(struct kib_msg *msg, int nob);
1036 int kiblnd_post_rx(struct kib_rx *rx, int credit);
1038 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1039 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1040 struct iov_iter *to, unsigned int rlen);