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, 2014, 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 #ifdef HAVE_COMPAT_RDMA
42 #include <linux/compat-2.6.h>
45 #include <linux/version.h>
46 #include <linux/module.h>
47 #include <linux/kernel.h>
48 #include <linux/kthread.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/unistd.h>
54 #include <linux/uio.h>
56 #include <asm/uaccess.h>
59 #include <linux/init.h>
61 #include <linux/file.h>
62 #include <linux/stat.h>
63 #include <linux/list.h>
64 #include <linux/kmod.h>
65 #include <linux/sysctl.h>
66 #include <linux/pci.h>
67 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
68 #include <linux/pci-dma.h>
74 #include <rdma/rdma_cm.h>
75 #include <rdma/ib_cm.h>
76 #include <rdma/ib_verbs.h>
77 #include <rdma/ib_fmr_pool.h>
79 #define DEBUG_SUBSYSTEM S_LND
81 #include <libcfs/libcfs.h>
82 #include <lnet/lnet.h>
83 #include <lnet/lib-lnet.h>
85 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
86 /* # scheduler loops before reschedule */
87 #define IBLND_RESCHED 100
89 #define IBLND_N_SCHED 2
90 #define IBLND_N_SCHED_HIGH 4
94 int *kib_dev_failover; /* HCA failover */
95 unsigned int *kib_service; /* IB service number */
96 int *kib_min_reconnect_interval; /* first failed connection retry... */
97 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
98 int *kib_cksum; /* checksum kib_msg_t? */
99 int *kib_timeout; /* comms timeout (seconds) */
100 int *kib_keepalive; /* keepalive timeout (seconds) */
101 int *kib_ntx; /* # tx descs */
102 int *kib_credits; /* # concurrent sends */
103 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
104 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
105 int *kib_peercredits_hiw; /* # when eagerly to return credits */
106 int *kib_peertimeout; /* seconds to consider peer dead */
107 char **kib_default_ipif; /* default IPoIB interface */
108 int *kib_retry_count;
109 int *kib_rnr_retry_count;
110 int *kib_concurrent_sends; /* send work queue sizing */
111 int *kib_ib_mtu; /* IB MTU */
112 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
113 * than this value, 0 disable map-on-demand */
114 int *kib_fmr_pool_size; /* # FMRs in pool */
115 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
116 int *kib_fmr_cache; /* enable FMR pool cache? */
117 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
118 struct ctl_table_header *kib_sysctl; /* sysctl interface */
120 int *kib_require_priv_port;/* accept only privileged ports */
121 int *kib_use_priv_port; /* use privileged port for active connect */
122 /* # threads on each CPT */
126 extern kib_tunables_t kiblnd_tunables;
128 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
129 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
131 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
132 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
134 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
135 IBLND_MSG_QUEUE_SIZE_V1 : \
136 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
137 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
138 IBLND_CREDIT_HIGHWATER_V1 : \
139 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
141 #ifdef HAVE_RDMA_CREATE_ID_4ARG
142 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
144 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
148 kiblnd_concurrent_sends_v1(void)
150 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
151 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
153 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
154 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
156 return *kiblnd_tunables.kib_concurrent_sends;
159 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
160 kiblnd_concurrent_sends_v1() : \
161 *kiblnd_tunables.kib_concurrent_sends)
162 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
163 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
164 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
166 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
167 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
168 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
169 *kiblnd_tunables.kib_map_on_demand : \
170 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
171 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
172 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
174 /************************/
175 /* derived constants... */
176 /* Pools (shared by connections on each CPT) */
177 /* These pools can grow at runtime, so don't need give a very large value */
178 #define IBLND_TX_POOL 256
179 #define IBLND_FMR_POOL 256
180 #define IBLND_FMR_POOL_FLUSH 192
182 /* TX messages (shared by all connections) */
183 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
185 /* RX messages (per connection) */
186 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
187 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
188 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
190 /* WRs and CQEs (per connection) */
191 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
192 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
193 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
197 /* o2iblnd can run over aliased interface */
199 #define KIB_IFNAME_SIZE IFALIASZ
201 #define KIB_IFNAME_SIZE 256
206 struct list_head ibd_list; /* chain on kib_devs */
207 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
208 __u32 ibd_ifip; /* IPoIB interface IP */
209 /** IPoIB interface name */
210 char ibd_ifname[KIB_IFNAME_SIZE];
211 int ibd_nnets; /* # nets extant */
213 cfs_time_t ibd_next_failover;
214 /* # failover failures */
215 int ibd_failed_failover;
216 /* failover in progress */
217 unsigned int ibd_failover;
218 /* IPoIB interface is a bonding master */
219 unsigned int ibd_can_failover;
220 struct list_head ibd_nets;
221 struct kib_hca_dev *ibd_hdev;
224 typedef struct kib_hca_dev
226 struct rdma_cm_id *ibh_cmid; /* listener cmid */
227 struct ib_device *ibh_ibdev; /* IB device */
228 int ibh_page_shift; /* page shift of current HCA */
229 int ibh_page_size; /* page size of current HCA */
230 __u64 ibh_page_mask; /* page mask of current HCA */
231 int ibh_mr_shift; /* bits shift of max MR size */
232 __u64 ibh_mr_size; /* size of MR */
233 struct ib_mr *ibh_mrs; /* global MR */
234 struct ib_pd *ibh_pd; /* PD */
235 kib_dev_t *ibh_dev; /* owner */
236 atomic_t ibh_ref; /* refcount */
239 /** # of seconds to keep pool alive */
240 #define IBLND_POOL_DEADLINE 300
241 /** # of seconds to retry if allocation failed */
242 #define IBLND_POOL_RETRY 1
246 int ibp_npages; /* # pages */
247 struct page *ibp_pages[0]; /* page array */
253 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
254 int inc, struct kib_pool **pp_po);
255 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
256 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
257 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
261 #define IBLND_POOL_NAME_LEN 32
263 typedef struct kib_poolset
267 /* network it belongs to */
268 struct kib_net *ps_net;
270 char ps_name[IBLND_POOL_NAME_LEN];
272 struct list_head ps_pool_list;
273 /* failed pool list */
274 struct list_head ps_failed_pool_list;
275 /* time stamp for retry if failed to allocate */
276 cfs_time_t ps_next_retry;
277 /* is allocating new pool */
284 /* create a new pool */
285 kib_ps_pool_create_t ps_pool_create;
287 kib_ps_pool_destroy_t ps_pool_destroy;
288 /* initialize new allocated node */
289 kib_ps_node_init_t ps_node_init;
291 kib_ps_node_fini_t ps_node_fini;
294 typedef struct kib_pool
296 /* chain on pool list */
297 struct list_head po_list;
298 /* pre-allocated node */
299 struct list_head po_free_list;
300 /* pool_set of this pool */
301 kib_poolset_t *po_owner;
302 /* deadline of this pool */
303 cfs_time_t po_deadline;
304 /* # of elements in use */
306 /* pool is created on failed HCA */
308 /* # of pre-allocated elements */
313 kib_poolset_t tps_poolset; /* pool-set */
314 __u64 tps_next_tx_cookie; /* cookie of TX */
318 kib_pool_t tpo_pool; /* pool */
319 struct kib_hca_dev *tpo_hdev; /* device for this pool */
320 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
321 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
326 spinlock_t fps_lock; /* serialize */
327 struct kib_net *fps_net; /* IB network */
328 struct list_head fps_pool_list; /* FMR pool list */
329 struct list_head fps_failed_pool_list; /* FMR pool list */
330 __u64 fps_version; /* validity stamp */
331 int fps_cpt; /* CPT id */
333 int fps_flush_trigger;
334 /* is allocating new pool */
336 /* time stamp for retry if failed to allocate */
337 cfs_time_t fps_next_retry;
342 struct list_head fpo_list; /* chain on pool list */
343 struct kib_hca_dev *fpo_hdev; /* device for this pool */
344 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
345 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
346 cfs_time_t fpo_deadline; /* deadline of this pool */
347 int fpo_failed; /* fmr pool is failed */
348 int fpo_map_count; /* # of mapped FMR */
352 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
353 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
356 typedef struct kib_net
358 /* chain on kib_dev_t::ibd_nets */
359 struct list_head ibn_list;
360 __u64 ibn_incarnation;/* my epoch */
361 int ibn_init; /* initialisation state */
362 int ibn_shutdown; /* shutting down? */
364 atomic_t ibn_npeers; /* # peers extant */
365 atomic_t ibn_nconns; /* # connections extant */
367 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
368 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
370 kib_dev_t *ibn_dev; /* underlying IB device */
373 #define KIB_THREAD_SHIFT 16
374 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
375 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
376 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
378 struct kib_sched_info {
381 /* schedulers sleep here */
382 wait_queue_head_t ibs_waitq;
383 /* conns to check for rx completions */
384 struct list_head ibs_conns;
385 /* number of scheduler threads */
387 /* max allowed scheduler threads */
388 int ibs_nthreads_max;
389 int ibs_cpt; /* CPT id */
394 int kib_init; /* initialisation state */
395 int kib_shutdown; /* shut down? */
396 struct list_head kib_devs; /* IB devices extant */
397 /* list head of failed devices */
398 struct list_head kib_failed_devs;
399 /* schedulers sleep here */
400 wait_queue_head_t kib_failover_waitq;
401 atomic_t kib_nthreads; /* # live threads */
402 /* stabilize net/dev/peer/conn ops */
403 rwlock_t kib_global_lock;
404 /* hash table of all my known peers */
405 struct list_head *kib_peers;
406 /* size of kib_peers */
407 int kib_peer_hash_size;
408 /* the connd task (serialisation assertions) */
410 /* connections to setup/teardown */
411 struct list_head kib_connd_conns;
412 /* connections with zero refcount */
413 struct list_head kib_connd_zombies;
414 /* connection daemon sleeps here */
415 wait_queue_head_t kib_connd_waitq;
416 spinlock_t kib_connd_lock; /* serialise */
417 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
418 /* percpt data for schedulers */
419 struct kib_sched_info **kib_scheds;
422 #define IBLND_INIT_NOTHING 0
423 #define IBLND_INIT_DATA 1
424 #define IBLND_INIT_ALL 2
426 /************************************************************************
427 * IB Wire message format.
428 * These are sent in sender's byte order (i.e. receiver flips).
431 typedef struct kib_connparams
433 __u16 ibcp_queue_depth;
434 __u16 ibcp_max_frags;
435 __u32 ibcp_max_msg_size;
436 } WIRE_ATTR kib_connparams_t;
440 lnet_hdr_t ibim_hdr; /* portals header */
441 char ibim_payload[0]; /* piggy-backed payload */
442 } WIRE_ATTR kib_immediate_msg_t;
446 __u32 rf_nob; /* # bytes this frag */
447 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
448 } WIRE_ATTR kib_rdma_frag_t;
452 __u32 rd_key; /* local/remote key */
453 __u32 rd_nfrags; /* # fragments */
454 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
455 } WIRE_ATTR kib_rdma_desc_t;
459 lnet_hdr_t ibprm_hdr; /* portals header */
460 __u64 ibprm_cookie; /* opaque completion cookie */
461 } WIRE_ATTR kib_putreq_msg_t;
465 __u64 ibpam_src_cookie; /* reflected completion cookie */
466 __u64 ibpam_dst_cookie; /* opaque completion cookie */
467 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
468 } WIRE_ATTR kib_putack_msg_t;
472 lnet_hdr_t ibgm_hdr; /* portals header */
473 __u64 ibgm_cookie; /* opaque completion cookie */
474 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
475 } WIRE_ATTR kib_get_msg_t;
479 __u64 ibcm_cookie; /* opaque completion cookie */
480 __s32 ibcm_status; /* < 0 failure: >= 0 length */
481 } WIRE_ATTR kib_completion_msg_t;
485 /* First 2 fields fixed FOR ALL TIME */
486 __u32 ibm_magic; /* I'm an ibnal message */
487 __u16 ibm_version; /* this is my version number */
489 __u8 ibm_type; /* msg type */
490 __u8 ibm_credits; /* returned credits */
491 __u32 ibm_nob; /* # bytes in whole message */
492 __u32 ibm_cksum; /* checksum (0 == no checksum) */
493 __u64 ibm_srcnid; /* sender's NID */
494 __u64 ibm_srcstamp; /* sender's incarnation */
495 __u64 ibm_dstnid; /* destination's NID */
496 __u64 ibm_dststamp; /* destination's incarnation */
499 kib_connparams_t connparams;
500 kib_immediate_msg_t immediate;
501 kib_putreq_msg_t putreq;
502 kib_putack_msg_t putack;
504 kib_completion_msg_t completion;
506 } WIRE_ATTR kib_msg_t;
508 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
510 #define IBLND_MSG_VERSION_1 0x11
511 #define IBLND_MSG_VERSION_2 0x12
512 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
514 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
515 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
516 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
517 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
518 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
519 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
520 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
521 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
522 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
523 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
526 __u32 ibr_magic; /* sender's magic */
527 __u16 ibr_version; /* sender's version */
528 __u8 ibr_why; /* reject reason */
529 __u8 ibr_padding; /* padding */
530 __u64 ibr_incarnation; /* incarnation of peer */
531 kib_connparams_t ibr_cp; /* connection parameters */
532 } WIRE_ATTR kib_rej_t;
534 /* connection rejection reasons */
535 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
536 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
537 #define IBLND_REJECT_FATAL 3 /* Anything else */
539 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
540 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
542 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
543 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
545 /***********************************************************************/
547 typedef struct kib_rx /* receive message */
549 /* queue for attention */
550 struct list_head rx_list;
552 struct kib_conn *rx_conn;
553 /* # bytes received (-1 while posted) */
555 /* completion status */
556 enum ib_wc_status rx_status;
557 /* message buffer (host vaddr) */
559 /* message buffer (I/O addr) */
561 /* for dma_unmap_single() */
562 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
563 /* receive work item... */
564 struct ib_recv_wr rx_wrq;
565 /* ...and its memory */
566 struct ib_sge rx_sge;
569 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
570 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
571 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
572 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
574 typedef struct kib_tx /* transmit message */
576 /* queue on idle_txs ibc_tx_queue etc. */
577 struct list_head tx_list;
579 kib_tx_pool_t *tx_pool;
581 struct kib_conn *tx_conn;
582 /* # tx callbacks outstanding */
584 /* queued for sending */
586 /* waiting for peer */
588 /* LNET completion status */
590 /* completion deadline */
591 unsigned long tx_deadline;
592 /* completion cookie */
594 /* lnet msgs to finalize on completion */
595 lnet_msg_t *tx_lntmsg[2];
596 /* message buffer (host vaddr) */
598 /* message buffer (I/O addr) */
600 /* for dma_unmap_single() */
601 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
602 /* # send work items */
604 /* send work items... */
605 struct ib_send_wr *tx_wrq;
606 /* ...and their memory */
607 struct ib_sge *tx_sge;
608 /* rdma descriptor */
609 kib_rdma_desc_t *tx_rd;
610 /* # entries in... */
612 /* dma_map_sg descriptor */
613 struct scatterlist *tx_frags;
614 /* rdma phys page addrs */
622 typedef struct kib_connvars
624 /* connection-in-progress variables */
628 typedef struct kib_conn
630 /* scheduler information */
631 struct kib_sched_info *ibc_sched;
633 struct kib_peer *ibc_peer;
635 kib_hca_dev_t *ibc_hdev;
636 /* stash on peer's conn list */
637 struct list_head ibc_list;
638 /* schedule for attention */
639 struct list_head ibc_sched_list;
640 /* version of connection */
642 /* which instance of the peer */
643 __u64 ibc_incarnation;
645 atomic_t ibc_refcount;
646 /* what's happening */
648 /* # uncompleted sends */
649 int ibc_nsends_posted;
650 /* # uncompleted NOOPs */
651 int ibc_noops_posted;
652 /* # credits I have */
654 /* # credits to return */
655 int ibc_outstanding_credits;
656 /* # ACK/DONE msg credits */
657 int ibc_reserved_credits;
658 /* set on comms error */
660 /* receive buffers owned */
661 unsigned short ibc_nrx;
662 /** rejected by connection race */
663 unsigned short ibc_conn_race:1;
664 /* scheduled for attention */
665 unsigned short ibc_scheduled:1;
666 /* CQ callback fired */
667 unsigned short ibc_ready:1;
668 /* time of last send */
669 unsigned long ibc_last_send;
670 /** link chain for kiblnd_check_conns only */
671 struct list_head ibc_connd_list;
672 /** rxs completed before ESTABLISHED */
673 struct list_head ibc_early_rxs;
674 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
675 struct list_head ibc_tx_noops;
676 /* sends that need a credit */
677 struct list_head ibc_tx_queue;
678 /* sends that don't need a credit */
679 struct list_head ibc_tx_queue_nocred;
680 /* sends that need to reserve an ACK/DONE msg */
681 struct list_head ibc_tx_queue_rsrvd;
682 /* active tx awaiting completion */
683 struct list_head ibc_active_txs;
688 /* premapped rx msg pages */
689 kib_pages_t *ibc_rx_pages;
692 struct rdma_cm_id *ibc_cmid;
693 /* completion queue */
694 struct ib_cq *ibc_cq;
696 /* in-progress connection state */
697 kib_connvars_t *ibc_connvars;
700 #define IBLND_CONN_INIT 0 /* being initialised */
701 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
702 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
703 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
704 #define IBLND_CONN_CLOSING 4 /* being closed */
705 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
707 typedef struct kib_peer
709 /* stash on global peer list */
710 struct list_head ibp_list;
711 /* who's on the other end(s) */
716 atomic_t ibp_refcount;
717 /* all active connections */
718 struct list_head ibp_conns;
719 /* msgs waiting for a conn */
720 struct list_head ibp_tx_queue;
721 /* version of peer */
723 /* incarnation of peer */
724 __u64 ibp_incarnation;
725 /* current active connection attempts */
727 /* current passive connection attempts */
729 /* errno on closing this peer */
731 /* when (in jiffies) I was last alive */
732 cfs_time_t ibp_last_alive;
735 extern kib_data_t kiblnd_data;
737 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
740 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
742 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
743 atomic_inc(&hdev->ibh_ref);
747 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
749 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
750 if (atomic_dec_and_test(&hdev->ibh_ref))
751 kiblnd_hdev_destroy(hdev);
755 kiblnd_dev_can_failover(kib_dev_t *dev)
757 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
760 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
763 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
766 return dev->ibd_can_failover;
769 #define kiblnd_conn_addref(conn) \
771 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
772 (conn), atomic_read(&(conn)->ibc_refcount)); \
773 atomic_inc(&(conn)->ibc_refcount); \
776 #define kiblnd_conn_decref(conn) \
778 unsigned long flags; \
780 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
781 (conn), atomic_read(&(conn)->ibc_refcount)); \
782 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
783 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
784 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
785 list_add_tail(&(conn)->ibc_list, \
786 &kiblnd_data.kib_connd_zombies); \
787 wake_up(&kiblnd_data.kib_connd_waitq); \
788 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
792 #define kiblnd_peer_addref(peer) \
794 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
795 (peer), libcfs_nid2str((peer)->ibp_nid), \
796 atomic_read (&(peer)->ibp_refcount)); \
797 atomic_inc(&(peer)->ibp_refcount); \
800 #define kiblnd_peer_decref(peer) \
802 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
803 (peer), libcfs_nid2str((peer)->ibp_nid), \
804 atomic_read (&(peer)->ibp_refcount)); \
805 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
806 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
807 kiblnd_destroy_peer(peer); \
810 static inline struct list_head *
811 kiblnd_nid2peerlist (lnet_nid_t nid)
814 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
816 return &kiblnd_data.kib_peers[hash];
820 kiblnd_peer_active (kib_peer_t *peer)
822 /* Am I in the peer hash table? */
823 return !list_empty(&peer->ibp_list);
826 static inline kib_conn_t *
827 kiblnd_get_conn_locked (kib_peer_t *peer)
829 LASSERT(!list_empty(&peer->ibp_conns));
831 /* just return the first connection */
832 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
836 kiblnd_send_keepalive(kib_conn_t *conn)
838 return (*kiblnd_tunables.kib_keepalive > 0) &&
839 cfs_time_after(jiffies, conn->ibc_last_send +
840 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
845 kiblnd_need_noop(kib_conn_t *conn)
847 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
849 if (conn->ibc_outstanding_credits <
850 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
851 !kiblnd_send_keepalive(conn))
852 return 0; /* No need to send NOOP */
854 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
855 if (!list_empty(&conn->ibc_tx_queue_nocred))
856 return 0; /* NOOP can be piggybacked */
858 /* No tx to piggyback NOOP onto or no credit to send a tx */
859 return (list_empty(&conn->ibc_tx_queue) ||
860 conn->ibc_credits == 0);
863 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
864 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
865 conn->ibc_credits == 0) /* no credit */
868 if (conn->ibc_credits == 1 && /* last credit reserved for */
869 conn->ibc_outstanding_credits == 0) /* giving back credits */
872 /* No tx to piggyback NOOP onto or no credit to send a tx */
873 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
877 kiblnd_abort_receives(kib_conn_t *conn)
879 ib_modify_qp(conn->ibc_cmid->qp,
880 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
883 static inline const char *
884 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
886 if (q == &conn->ibc_tx_queue)
889 if (q == &conn->ibc_tx_queue_rsrvd)
890 return "tx_queue_rsrvd";
892 if (q == &conn->ibc_tx_queue_nocred)
893 return "tx_queue_nocred";
895 if (q == &conn->ibc_active_txs)
902 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
903 * lowest bits of the work request id to stash the work item type. */
905 #define IBLND_WID_TX 0
906 #define IBLND_WID_RDMA 1
907 #define IBLND_WID_RX 2
908 #define IBLND_WID_MASK 3UL
911 kiblnd_ptr2wreqid (void *ptr, int type)
913 unsigned long lptr = (unsigned long)ptr;
915 LASSERT ((lptr & IBLND_WID_MASK) == 0);
916 LASSERT ((type & ~IBLND_WID_MASK) == 0);
917 return (__u64)(lptr | type);
921 kiblnd_wreqid2ptr (__u64 wreqid)
923 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
927 kiblnd_wreqid2type (__u64 wreqid)
929 return (wreqid & IBLND_WID_MASK);
933 kiblnd_set_conn_state (kib_conn_t *conn, int state)
935 conn->ibc_state = state;
940 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
942 msg->ibm_type = type;
943 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
947 kiblnd_rd_size (kib_rdma_desc_t *rd)
952 for (i = size = 0; i < rd->rd_nfrags; i++)
953 size += rd->rd_frags[i].rf_nob;
959 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
961 return rd->rd_frags[index].rf_addr;
965 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
967 return rd->rd_frags[index].rf_nob;
971 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
977 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
979 if (nob < rd->rd_frags[index].rf_nob) {
980 rd->rd_frags[index].rf_addr += nob;
981 rd->rd_frags[index].rf_nob -= nob;
990 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
992 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
993 msgtype == IBLND_MSG_PUT_ACK);
995 return msgtype == IBLND_MSG_GET_REQ ?
996 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
997 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1001 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1003 return ib_dma_mapping_error(dev, dma_addr);
1006 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1007 void *msg, size_t size,
1008 enum dma_data_direction direction)
1010 return ib_dma_map_single(dev, msg, size, direction);
1013 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1014 __u64 addr, size_t size,
1015 enum dma_data_direction direction)
1017 ib_dma_unmap_single(dev, addr, size, direction);
1020 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1021 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1023 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1024 struct scatterlist *sg, int nents,
1025 enum dma_data_direction direction)
1027 return ib_dma_map_sg(dev, sg, nents, direction);
1030 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1031 struct scatterlist *sg, int nents,
1032 enum dma_data_direction direction)
1034 ib_dma_unmap_sg(dev, sg, nents, direction);
1037 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1038 struct scatterlist *sg)
1040 return ib_sg_dma_address(dev, sg);
1043 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1044 struct scatterlist *sg)
1046 return ib_sg_dma_len(dev, sg);
1049 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1050 * right because OFED1.2 defines it as const, to use it we have to add
1051 * (void *) cast to overcome "const" */
1053 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1054 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1056 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1057 kib_rdma_desc_t *rd);
1058 void kiblnd_map_rx_descs(kib_conn_t *conn);
1059 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1060 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1061 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1063 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1064 int npages, __u64 iov, kib_fmr_t *fmr);
1065 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1067 int kiblnd_tunables_init(void);
1068 void kiblnd_tunables_fini(void);
1070 int kiblnd_connd (void *arg);
1071 int kiblnd_scheduler(void *arg);
1072 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1073 int kiblnd_failover_thread (void *arg);
1075 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1077 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1078 struct rdma_cm_event *event);
1079 int kiblnd_translate_mtu(int value);
1081 int kiblnd_dev_failover(kib_dev_t *dev);
1082 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1083 void kiblnd_destroy_peer (kib_peer_t *peer);
1084 void kiblnd_connect_peer(kib_peer_t *peer);
1085 void kiblnd_destroy_dev (kib_dev_t *dev);
1086 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1087 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1088 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1089 int version, __u64 incarnation);
1090 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1092 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1093 int state, int version);
1094 void kiblnd_destroy_conn (kib_conn_t *conn);
1095 void kiblnd_close_conn (kib_conn_t *conn, int error);
1096 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1098 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1099 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1101 void kiblnd_qp_event(struct ib_event *event, void *arg);
1102 void kiblnd_cq_event(struct ib_event *event, void *arg);
1103 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1105 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1106 int credits, lnet_nid_t dstnid, __u64 dststamp);
1107 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1108 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1110 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1111 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1112 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1113 unsigned int offset, unsigned int mlen, unsigned int rlen);