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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.h
38 * Author: Eric Barton <eric@bartonsoftware.com>
41 #include <linux/module.h>
42 #include <linux/kernel.h>
44 #include <linux/string.h>
45 #include <linux/stat.h>
46 #include <linux/errno.h>
47 #include <linux/unistd.h>
48 #include <linux/uio.h>
49 #include <linux/uaccess.h>
54 #include <linux/file.h>
55 #include <linux/list.h>
56 #include <linux/kmod.h>
57 #include <linux/sysctl.h>
58 #include <linux/pci.h>
63 #include <rdma/rdma_cm.h>
64 #include <rdma/ib_cm.h>
65 #include <rdma/ib_verbs.h>
66 #include <rdma/ib_fmr_pool.h>
68 #define DEBUG_SUBSYSTEM S_LND
70 #include "../../../include/linux/libcfs/libcfs.h"
71 #include "../../../include/linux/lnet/lnet.h"
72 #include "../../../include/linux/lnet/lib-lnet.h"
74 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
75 /* # scheduler loops before reschedule */
76 #define IBLND_RESCHED 100
78 #define IBLND_N_SCHED 2
79 #define IBLND_N_SCHED_HIGH 4
82 int *kib_dev_failover; /* HCA failover */
83 unsigned int *kib_service; /* IB service number */
84 int *kib_min_reconnect_interval; /* first failed connection retry... */
85 int *kib_max_reconnect_interval; /* exponentially increasing to this */
86 int *kib_cksum; /* checksum kib_msg_t? */
87 int *kib_timeout; /* comms timeout (seconds) */
88 int *kib_keepalive; /* keepalive timeout (seconds) */
89 int *kib_ntx; /* # tx descs */
90 int *kib_credits; /* # concurrent sends */
91 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
92 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
93 int *kib_peercredits_hiw; /* # when eagerly to return credits */
94 int *kib_peertimeout; /* seconds to consider peer dead */
95 char **kib_default_ipif; /* default IPoIB interface */
97 int *kib_rnr_retry_count;
98 int *kib_concurrent_sends; /* send work queue sizing */
99 int *kib_ib_mtu; /* IB MTU */
100 int *kib_map_on_demand; /* map-on-demand if RD has more */
101 /* fragments than this value, 0 */
102 /* disable map-on-demand */
103 int *kib_fmr_pool_size; /* # FMRs in pool */
104 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
105 int *kib_fmr_cache; /* enable FMR pool cache? */
106 int *kib_require_priv_port; /* accept only privileged ports */
107 int *kib_use_priv_port; /* use privileged port for active connect */
108 int *kib_nscheds; /* # threads on each CPT */
111 extern kib_tunables_t kiblnd_tunables;
113 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
114 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
116 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
117 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t *) 0)->ibm_credits)) - 1) /* Max # of peer credits */
119 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
120 IBLND_MSG_QUEUE_SIZE_V1 : \
121 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
122 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
123 IBLND_CREDIT_HIGHWATER_V1 : \
124 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
126 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(&init_net, \
131 kiblnd_concurrent_sends_v1(void)
133 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
134 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
136 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
137 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
139 return *kiblnd_tunables.kib_concurrent_sends;
142 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
143 kiblnd_concurrent_sends_v1() : \
144 *kiblnd_tunables.kib_concurrent_sends)
145 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
146 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
147 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
149 #define IBLND_MSG_SIZE (4 << 10) /* max size of queued messages (inc hdr) */
150 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
151 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand ? \
152 *kiblnd_tunables.kib_map_on_demand : \
153 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
154 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
155 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
157 /************************/
158 /* derived constants... */
159 /* Pools (shared by connections on each CPT) */
160 /* These pools can grow at runtime, so don't need give a very large value */
161 #define IBLND_TX_POOL 256
162 #define IBLND_FMR_POOL 256
163 #define IBLND_FMR_POOL_FLUSH 192
165 #define IBLND_RX_MSGS(c) \
166 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
167 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
168 #define IBLND_RX_MSG_PAGES(c) \
169 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
171 /* WRs and CQEs (per connection) */
172 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
173 #define IBLND_SEND_WRS(c) \
174 ((c->ibc_max_frags + 1) * IBLND_CONCURRENT_SENDS(c->ibc_version))
175 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
179 /* o2iblnd can run over aliased interface */
181 #define KIB_IFNAME_SIZE IFALIASZ
183 #define KIB_IFNAME_SIZE 256
187 struct list_head ibd_list; /* chain on kib_devs */
188 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
189 __u32 ibd_ifip; /* IPoIB interface IP */
191 /* IPoIB interface name */
192 char ibd_ifname[KIB_IFNAME_SIZE];
193 int ibd_nnets; /* # nets extant */
195 unsigned long ibd_next_failover;
196 int ibd_failed_failover; /* # failover failures */
197 unsigned int ibd_failover; /* failover in progress */
198 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
199 struct list_head ibd_nets;
200 struct kib_hca_dev *ibd_hdev;
203 typedef struct kib_hca_dev {
204 struct rdma_cm_id *ibh_cmid; /* listener cmid */
205 struct ib_device *ibh_ibdev; /* IB device */
206 int ibh_page_shift; /* page shift of current HCA */
207 int ibh_page_size; /* page size of current HCA */
208 __u64 ibh_page_mask; /* page mask of current HCA */
209 int ibh_mr_shift; /* bits shift of max MR size */
210 __u64 ibh_mr_size; /* size of MR */
211 struct ib_mr *ibh_mrs; /* global MR */
212 struct ib_pd *ibh_pd; /* PD */
213 kib_dev_t *ibh_dev; /* owner */
214 atomic_t ibh_ref; /* refcount */
217 /** # of seconds to keep pool alive */
218 #define IBLND_POOL_DEADLINE 300
219 /** # of seconds to retry if allocation failed */
220 #define IBLND_POOL_RETRY 1
223 int ibp_npages; /* # pages */
224 struct page *ibp_pages[0]; /* page array */
230 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
231 int inc, struct kib_pool **pp_po);
232 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
233 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
234 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
238 #define IBLND_POOL_NAME_LEN 32
240 typedef struct kib_poolset {
241 spinlock_t ps_lock; /* serialize */
242 struct kib_net *ps_net; /* network it belongs to */
243 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
244 struct list_head ps_pool_list; /* list of pools */
245 struct list_head ps_failed_pool_list;/* failed pool list */
246 unsigned long ps_next_retry; /* time stamp for retry if */
247 /* failed to allocate */
248 int ps_increasing; /* is allocating new pool */
249 int ps_pool_size; /* new pool size */
250 int ps_cpt; /* CPT id */
252 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
253 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
254 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
255 kib_ps_node_fini_t ps_node_fini; /* finalize node */
258 typedef struct kib_pool {
259 struct list_head po_list; /* chain on pool list */
260 struct list_head po_free_list; /* pre-allocated node */
261 kib_poolset_t *po_owner; /* pool_set of this pool */
262 unsigned long po_deadline; /* deadline of this pool */
263 int po_allocated; /* # of elements in use */
264 int po_failed; /* pool is created on failed HCA */
265 int po_size; /* # of pre-allocated elements */
269 kib_poolset_t tps_poolset; /* pool-set */
270 __u64 tps_next_tx_cookie; /* cookie of TX */
274 kib_pool_t tpo_pool; /* pool */
275 struct kib_hca_dev *tpo_hdev; /* device for this pool */
276 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
277 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
281 spinlock_t fps_lock; /* serialize */
282 struct kib_net *fps_net; /* IB network */
283 struct list_head fps_pool_list; /* FMR pool list */
284 struct list_head fps_failed_pool_list;/* FMR pool list */
285 __u64 fps_version; /* validity stamp */
286 int fps_cpt; /* CPT id */
288 int fps_flush_trigger;
289 int fps_increasing; /* is allocating new pool */
290 unsigned long fps_next_retry; /* time stamp for retry if*/
291 /* failed to allocate */
295 struct list_head fpo_list; /* chain on pool list */
296 struct kib_hca_dev *fpo_hdev; /* device for this pool */
297 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
298 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
299 unsigned long fpo_deadline; /* deadline of this pool */
300 int fpo_failed; /* fmr pool is failed */
301 int fpo_map_count; /* # of mapped FMR */
305 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
306 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
309 typedef struct kib_net {
310 struct list_head ibn_list; /* chain on kib_dev_t::ibd_nets */
311 __u64 ibn_incarnation;/* my epoch */
312 int ibn_init; /* initialisation state */
313 int ibn_shutdown; /* shutting down? */
315 atomic_t ibn_npeers; /* # peers extant */
316 atomic_t ibn_nconns; /* # connections extant */
318 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
319 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
321 kib_dev_t *ibn_dev; /* underlying IB device */
324 #define KIB_THREAD_SHIFT 16
325 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
326 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
327 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
329 struct kib_sched_info {
330 spinlock_t ibs_lock; /* serialise */
331 wait_queue_head_t ibs_waitq; /* schedulers sleep here */
332 struct list_head ibs_conns; /* conns to check for rx completions */
333 int ibs_nthreads; /* number of scheduler threads */
334 int ibs_nthreads_max; /* max allowed scheduler threads */
335 int ibs_cpt; /* CPT id */
339 int kib_init; /* initialisation state */
340 int kib_shutdown; /* shut down? */
341 struct list_head kib_devs; /* IB devices extant */
342 struct list_head kib_failed_devs; /* list head of failed devices */
343 wait_queue_head_t kib_failover_waitq; /* schedulers sleep here */
344 atomic_t kib_nthreads; /* # live threads */
345 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
346 struct list_head *kib_peers; /* hash table of all my known peers */
347 int kib_peer_hash_size; /* size of kib_peers */
348 void *kib_connd; /* the connd task (serialisation assertions) */
349 struct list_head kib_connd_conns; /* connections to setup/teardown */
350 struct list_head kib_connd_zombies; /* connections with zero refcount */
351 /* connections to reconnect */
352 struct list_head kib_reconn_list;
353 /* peers wait for reconnection */
354 struct list_head kib_reconn_wait;
356 * The second that peers are pulled out from \a kib_reconn_wait
359 time64_t kib_reconn_sec;
361 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
362 spinlock_t kib_connd_lock; /* serialise */
363 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
364 struct kib_sched_info **kib_scheds; /* percpt data for schedulers */
367 #define IBLND_INIT_NOTHING 0
368 #define IBLND_INIT_DATA 1
369 #define IBLND_INIT_ALL 2
371 /************************************************************************
372 * IB Wire message format.
373 * These are sent in sender's byte order (i.e. receiver flips).
376 typedef struct kib_connparams {
377 __u16 ibcp_queue_depth;
378 __u16 ibcp_max_frags;
379 __u32 ibcp_max_msg_size;
380 } WIRE_ATTR kib_connparams_t;
383 lnet_hdr_t ibim_hdr; /* portals header */
384 char ibim_payload[0]; /* piggy-backed payload */
385 } WIRE_ATTR kib_immediate_msg_t;
388 __u32 rf_nob; /* # bytes this frag */
389 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
390 } WIRE_ATTR kib_rdma_frag_t;
393 __u32 rd_key; /* local/remote key */
394 __u32 rd_nfrags; /* # fragments */
395 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
396 } WIRE_ATTR kib_rdma_desc_t;
399 lnet_hdr_t ibprm_hdr; /* portals header */
400 __u64 ibprm_cookie; /* opaque completion cookie */
401 } WIRE_ATTR kib_putreq_msg_t;
404 __u64 ibpam_src_cookie; /* reflected completion cookie */
405 __u64 ibpam_dst_cookie; /* opaque completion cookie */
406 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
407 } WIRE_ATTR kib_putack_msg_t;
410 lnet_hdr_t ibgm_hdr; /* portals header */
411 __u64 ibgm_cookie; /* opaque completion cookie */
412 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
413 } WIRE_ATTR kib_get_msg_t;
416 __u64 ibcm_cookie; /* opaque completion cookie */
417 __s32 ibcm_status; /* < 0 failure: >= 0 length */
418 } WIRE_ATTR kib_completion_msg_t;
421 /* First 2 fields fixed FOR ALL TIME */
422 __u32 ibm_magic; /* I'm an ibnal message */
423 __u16 ibm_version; /* this is my version number */
425 __u8 ibm_type; /* msg type */
426 __u8 ibm_credits; /* returned credits */
427 __u32 ibm_nob; /* # bytes in whole message */
428 __u32 ibm_cksum; /* checksum (0 == no checksum) */
429 __u64 ibm_srcnid; /* sender's NID */
430 __u64 ibm_srcstamp; /* sender's incarnation */
431 __u64 ibm_dstnid; /* destination's NID */
432 __u64 ibm_dststamp; /* destination's incarnation */
435 kib_connparams_t connparams;
436 kib_immediate_msg_t immediate;
437 kib_putreq_msg_t putreq;
438 kib_putack_msg_t putack;
440 kib_completion_msg_t completion;
442 } WIRE_ATTR kib_msg_t;
444 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
446 #define IBLND_MSG_VERSION_1 0x11
447 #define IBLND_MSG_VERSION_2 0x12
448 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
450 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
451 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
452 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
453 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
454 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
455 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
456 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
457 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
458 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
459 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
462 __u32 ibr_magic; /* sender's magic */
463 __u16 ibr_version; /* sender's version */
464 __u8 ibr_why; /* reject reason */
465 __u8 ibr_padding; /* padding */
466 __u64 ibr_incarnation; /* incarnation of peer */
467 kib_connparams_t ibr_cp; /* connection parameters */
468 } WIRE_ATTR kib_rej_t;
470 /* connection rejection reasons */
471 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
472 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
473 #define IBLND_REJECT_FATAL 3 /* Anything else */
474 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
475 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
476 /* peer's rdma frags doesn't match mine */
477 #define IBLND_REJECT_RDMA_FRAGS 6
478 /* peer's msg queue size doesn't match mine */
479 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
481 /***********************************************************************/
483 typedef struct kib_rx /* receive message */
485 struct list_head rx_list; /* queue for attention */
486 struct kib_conn *rx_conn; /* owning conn */
487 int rx_nob; /* # bytes received (-1 while posted) */
488 enum ib_wc_status rx_status; /* completion status */
489 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
490 __u64 rx_msgaddr; /* message buffer (I/O addr) */
491 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap); /* for dma_unmap_single() */
492 struct ib_recv_wr rx_wrq; /* receive work item... */
493 struct ib_sge rx_sge; /* ...and its memory */
496 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
497 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
498 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
499 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give self back 1 reserved credit */
501 typedef struct kib_tx /* transmit message */
503 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
504 kib_tx_pool_t *tx_pool; /* pool I'm from */
505 struct kib_conn *tx_conn; /* owning conn */
506 short tx_sending; /* # tx callbacks outstanding */
507 short tx_queued; /* queued for sending */
508 short tx_waiting; /* waiting for peer */
509 int tx_status; /* LNET completion status */
510 unsigned long tx_deadline; /* completion deadline */
511 __u64 tx_cookie; /* completion cookie */
512 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
513 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
514 __u64 tx_msgaddr; /* message buffer (I/O addr) */
515 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap); /* for dma_unmap_single() */
516 int tx_nwrq; /* # send work items */
517 struct ib_rdma_wr *tx_wrq; /* send work items... */
518 struct ib_sge *tx_sge; /* ...and their memory */
519 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
520 int tx_nfrags; /* # entries in... */
521 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
522 __u64 *tx_pages; /* rdma phys page addrs */
523 kib_fmr_t fmr; /* FMR */
524 int tx_dmadir; /* dma direction */
527 typedef struct kib_connvars {
528 kib_msg_t cv_msg; /* connection-in-progress variables */
531 typedef struct kib_conn {
532 struct kib_sched_info *ibc_sched; /* scheduler information */
533 struct kib_peer *ibc_peer; /* owning peer */
534 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
535 struct list_head ibc_list; /* stash on peer's conn list */
536 struct list_head ibc_sched_list; /* schedule for attention */
537 __u16 ibc_version; /* version of connection */
538 /* reconnect later */
539 __u16 ibc_reconnect:1;
540 __u64 ibc_incarnation; /* which instance of the peer */
541 atomic_t ibc_refcount; /* # users */
542 int ibc_state; /* what's happening */
543 int ibc_nsends_posted; /* # uncompleted sends */
544 int ibc_noops_posted; /* # uncompleted NOOPs */
545 int ibc_credits; /* # credits I have */
546 int ibc_outstanding_credits; /* # credits to return */
547 int ibc_reserved_credits; /* # ACK/DONE msg credits */
548 int ibc_comms_error; /* set on comms error */
549 /* connections queue depth */
550 __u16 ibc_queue_depth;
551 /* connections max frags */
553 unsigned int ibc_nrx:16; /* receive buffers owned */
554 unsigned int ibc_scheduled:1; /* scheduled for attention */
555 unsigned int ibc_ready:1; /* CQ callback fired */
556 unsigned long ibc_last_send; /* time of last send */
557 struct list_head ibc_connd_list; /* link chain for */
558 /* kiblnd_check_conns only */
559 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
560 struct list_head ibc_tx_noops; /* IBLND_MSG_NOOPs for */
561 /* IBLND_MSG_VERSION_1 */
562 struct list_head ibc_tx_queue; /* sends that need a credit */
563 struct list_head ibc_tx_queue_nocred; /* sends that don't need a */
565 struct list_head ibc_tx_queue_rsrvd; /* sends that need to */
566 /* reserve an ACK/DONE msg */
567 struct list_head ibc_active_txs; /* active tx awaiting completion */
568 spinlock_t ibc_lock; /* serialise */
569 kib_rx_t *ibc_rxs; /* the rx descs */
570 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
572 struct rdma_cm_id *ibc_cmid; /* CM id */
573 struct ib_cq *ibc_cq; /* completion queue */
575 kib_connvars_t *ibc_connvars; /* in-progress connection state */
578 #define IBLND_CONN_INIT 0 /* being initialised */
579 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
580 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
581 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
582 #define IBLND_CONN_CLOSING 4 /* being closed */
583 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
585 typedef struct kib_peer {
586 struct list_head ibp_list; /* stash on global peer list */
587 lnet_nid_t ibp_nid; /* who's on the other end(s) */
588 lnet_ni_t *ibp_ni; /* LNet interface */
589 struct list_head ibp_conns; /* all active connections */
590 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
591 __u64 ibp_incarnation; /* incarnation of peer */
592 /* when (in jiffies) I was last alive */
593 unsigned long ibp_last_alive;
595 atomic_t ibp_refcount;
596 /* version of peer */
598 /* current passive connection attempts */
599 unsigned short ibp_accepting;
600 /* current active connection attempts */
601 unsigned short ibp_connecting;
602 /* reconnect this peer later */
603 unsigned short ibp_reconnecting:1;
604 /* # consecutive reconnection attempts to this peer */
605 unsigned int ibp_reconnected;
606 /* errno on closing this peer */
608 /* max map_on_demand */
610 /* max_peer_credits */
611 __u16 ibp_queue_depth;
614 extern kib_data_t kiblnd_data;
616 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
619 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
621 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
622 atomic_inc(&hdev->ibh_ref);
626 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
628 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
629 if (atomic_dec_and_test(&hdev->ibh_ref))
630 kiblnd_hdev_destroy(hdev);
634 kiblnd_dev_can_failover(kib_dev_t *dev)
636 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
639 if (!*kiblnd_tunables.kib_dev_failover) /* disabled */
642 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
645 return dev->ibd_can_failover;
648 #define kiblnd_conn_addref(conn) \
650 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
651 (conn), atomic_read(&(conn)->ibc_refcount)); \
652 atomic_inc(&(conn)->ibc_refcount); \
655 #define kiblnd_conn_decref(conn) \
657 unsigned long flags; \
659 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
660 (conn), atomic_read(&(conn)->ibc_refcount)); \
661 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
662 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
663 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
664 list_add_tail(&(conn)->ibc_list, \
665 &kiblnd_data.kib_connd_zombies); \
666 wake_up(&kiblnd_data.kib_connd_waitq); \
667 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
671 #define kiblnd_peer_addref(peer) \
673 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
674 (peer), libcfs_nid2str((peer)->ibp_nid), \
675 atomic_read(&(peer)->ibp_refcount)); \
676 atomic_inc(&(peer)->ibp_refcount); \
679 #define kiblnd_peer_decref(peer) \
681 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
682 (peer), libcfs_nid2str((peer)->ibp_nid), \
683 atomic_read(&(peer)->ibp_refcount)); \
684 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
685 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
686 kiblnd_destroy_peer(peer); \
690 kiblnd_peer_connecting(kib_peer_t *peer)
692 return peer->ibp_connecting ||
693 peer->ibp_reconnecting ||
698 kiblnd_peer_idle(kib_peer_t *peer)
700 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
703 static inline struct list_head *
704 kiblnd_nid2peerlist(lnet_nid_t nid)
707 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
709 return &kiblnd_data.kib_peers[hash];
713 kiblnd_peer_active(kib_peer_t *peer)
715 /* Am I in the peer hash table? */
716 return !list_empty(&peer->ibp_list);
719 static inline kib_conn_t *
720 kiblnd_get_conn_locked(kib_peer_t *peer)
722 LASSERT(!list_empty(&peer->ibp_conns));
724 /* just return the first connection */
725 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
729 kiblnd_send_keepalive(kib_conn_t *conn)
731 return (*kiblnd_tunables.kib_keepalive > 0) &&
732 cfs_time_after(jiffies, conn->ibc_last_send +
733 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
738 kiblnd_need_noop(kib_conn_t *conn)
740 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
742 if (conn->ibc_outstanding_credits <
743 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
744 !kiblnd_send_keepalive(conn))
745 return 0; /* No need to send NOOP */
747 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
748 if (!list_empty(&conn->ibc_tx_queue_nocred))
749 return 0; /* NOOP can be piggybacked */
751 /* No tx to piggyback NOOP onto or no credit to send a tx */
752 return (list_empty(&conn->ibc_tx_queue) ||
756 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
757 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
758 !conn->ibc_credits) /* no credit */
761 if (conn->ibc_credits == 1 && /* last credit reserved for */
762 !conn->ibc_outstanding_credits) /* giving back credits */
765 /* No tx to piggyback NOOP onto or no credit to send a tx */
766 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
770 kiblnd_abort_receives(kib_conn_t *conn)
772 ib_modify_qp(conn->ibc_cmid->qp,
773 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
776 static inline const char *
777 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
779 if (q == &conn->ibc_tx_queue)
782 if (q == &conn->ibc_tx_queue_rsrvd)
783 return "tx_queue_rsrvd";
785 if (q == &conn->ibc_tx_queue_nocred)
786 return "tx_queue_nocred";
788 if (q == &conn->ibc_active_txs)
795 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the */
796 /* lowest bits of the work request id to stash the work item type. */
798 #define IBLND_WID_INVAL 0
799 #define IBLND_WID_TX 1
800 #define IBLND_WID_RX 2
801 #define IBLND_WID_RDMA 3
802 #define IBLND_WID_MASK 3UL
805 kiblnd_ptr2wreqid(void *ptr, int type)
807 unsigned long lptr = (unsigned long)ptr;
809 LASSERT(!(lptr & IBLND_WID_MASK));
810 LASSERT(!(type & ~IBLND_WID_MASK));
811 return (__u64)(lptr | type);
815 kiblnd_wreqid2ptr(__u64 wreqid)
817 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
821 kiblnd_wreqid2type(__u64 wreqid)
823 return wreqid & IBLND_WID_MASK;
827 kiblnd_set_conn_state(kib_conn_t *conn, int state)
829 conn->ibc_state = state;
834 kiblnd_init_msg(kib_msg_t *msg, int type, int body_nob)
836 msg->ibm_type = type;
837 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
841 kiblnd_rd_size(kib_rdma_desc_t *rd)
846 for (i = size = 0; i < rd->rd_nfrags; i++)
847 size += rd->rd_frags[i].rf_nob;
853 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
855 return rd->rd_frags[index].rf_addr;
859 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
861 return rd->rd_frags[index].rf_nob;
865 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
871 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
873 if (nob < rd->rd_frags[index].rf_nob) {
874 rd->rd_frags[index].rf_addr += nob;
875 rd->rd_frags[index].rf_nob -= nob;
884 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
886 LASSERT(msgtype == IBLND_MSG_GET_REQ ||
887 msgtype == IBLND_MSG_PUT_ACK);
889 return msgtype == IBLND_MSG_GET_REQ ?
890 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
891 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
895 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
897 return ib_dma_mapping_error(dev, dma_addr);
900 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
901 void *msg, size_t size,
902 enum dma_data_direction direction)
904 return ib_dma_map_single(dev, msg, size, direction);
907 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
908 __u64 addr, size_t size,
909 enum dma_data_direction direction)
911 ib_dma_unmap_single(dev, addr, size, direction);
914 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
915 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
917 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
918 struct scatterlist *sg, int nents,
919 enum dma_data_direction direction)
921 return ib_dma_map_sg(dev, sg, nents, direction);
924 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
925 struct scatterlist *sg, int nents,
926 enum dma_data_direction direction)
928 ib_dma_unmap_sg(dev, sg, nents, direction);
931 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
932 struct scatterlist *sg)
934 return ib_sg_dma_address(dev, sg);
937 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
938 struct scatterlist *sg)
940 return ib_sg_dma_len(dev, sg);
943 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly */
944 /* right because OFED1.2 defines it as const, to use it we have to add */
945 /* (void *) cast to overcome "const" */
947 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
948 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
950 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
952 int negotiated_nfrags);
953 void kiblnd_map_rx_descs(kib_conn_t *conn);
954 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
955 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
956 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
958 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
959 int npages, __u64 iov, kib_fmr_t *fmr);
960 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
962 int kiblnd_tunables_init(void);
963 void kiblnd_tunables_fini(void);
965 int kiblnd_connd(void *arg);
966 int kiblnd_scheduler(void *arg);
967 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
968 int kiblnd_failover_thread(void *arg);
970 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
972 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
973 struct rdma_cm_event *event);
974 int kiblnd_translate_mtu(int value);
976 int kiblnd_dev_failover(kib_dev_t *dev);
977 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
978 void kiblnd_destroy_peer(kib_peer_t *peer);
979 bool kiblnd_reconnect_peer(kib_peer_t *peer);
980 void kiblnd_destroy_dev(kib_dev_t *dev);
981 void kiblnd_unlink_peer_locked(kib_peer_t *peer);
982 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid);
983 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
984 int version, __u64 incarnation);
985 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why);
987 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
988 int state, int version);
989 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
990 void kiblnd_close_conn(kib_conn_t *conn, int error);
991 void kiblnd_close_conn_locked(kib_conn_t *conn, int error);
993 void kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
994 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist,
997 void kiblnd_qp_event(struct ib_event *event, void *arg);
998 void kiblnd_cq_event(struct ib_event *event, void *arg);
999 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1001 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
1002 int credits, lnet_nid_t dstnid, __u64 dststamp);
1003 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1004 int kiblnd_post_rx(kib_rx_t *rx, int credit);
1006 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1007 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1008 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1009 unsigned int offset, unsigned int mlen, unsigned int rlen);