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, 2013, 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/version.h>
42 #include <linux/module.h>
43 #include <linux/kernel.h>
45 #include <linux/string.h>
46 #include <linux/stat.h>
47 #include <linux/errno.h>
48 #include <linux/unistd.h>
49 #include <linux/uio.h>
51 #include <asm/uaccess.h>
54 #include <linux/init.h>
56 #include <linux/file.h>
57 #include <linux/stat.h>
58 #include <linux/list.h>
59 #include <linux/kmod.h>
60 #include <linux/sysctl.h>
61 #include <linux/pci.h>
62 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
63 #include <linux/pci-dma.h>
69 #ifdef HAVE_COMPAT_RDMA
70 #include <linux/compat-2.6.h>
72 #include <rdma/rdma_cm.h>
73 #include <rdma/ib_cm.h>
74 #include <rdma/ib_verbs.h>
75 #include <rdma/ib_fmr_pool.h>
77 #define DEBUG_SUBSYSTEM S_LND
79 #include <libcfs/libcfs.h>
80 #include <lnet/lnet.h>
81 #include <lnet/lib-lnet.h>
83 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
84 /* # scheduler loops before reschedule */
85 #define IBLND_RESCHED 100
87 #define IBLND_N_SCHED 2
88 #define IBLND_N_SCHED_HIGH 4
92 int *kib_dev_failover; /* HCA failover */
93 unsigned int *kib_service; /* IB service number */
94 int *kib_min_reconnect_interval; /* first failed connection retry... */
95 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
96 int *kib_cksum; /* checksum kib_msg_t? */
97 int *kib_timeout; /* comms timeout (seconds) */
98 int *kib_keepalive; /* keepalive timeout (seconds) */
99 int *kib_ntx; /* # tx descs */
100 int *kib_credits; /* # concurrent sends */
101 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
102 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
103 int *kib_peercredits_hiw; /* # when eagerly to return credits */
104 int *kib_peertimeout; /* seconds to consider peer dead */
105 char **kib_default_ipif; /* default IPoIB interface */
106 int *kib_retry_count;
107 int *kib_rnr_retry_count;
108 int *kib_concurrent_sends; /* send work queue sizing */
109 int *kib_ib_mtu; /* IB MTU */
110 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
111 * than this value, 0 disable map-on-demand */
112 int *kib_pmr_pool_size; /* # physical MR in pool */
113 int *kib_fmr_pool_size; /* # FMRs in pool */
114 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
115 int *kib_fmr_cache; /* enable FMR pool cache? */
116 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
117 struct ctl_table_header *kib_sysctl; /* sysctl interface */
119 int *kib_require_priv_port;/* accept only privileged ports */
120 int *kib_use_priv_port; /* use privileged port for active connect */
121 /* # threads on each CPT */
125 extern kib_tunables_t kiblnd_tunables;
127 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
128 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
130 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
131 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
133 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
134 IBLND_MSG_QUEUE_SIZE_V1 : \
135 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
136 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
137 IBLND_CREDIT_HIGHWATER_V1 : \
138 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
140 #ifdef HAVE_RDMA_CREATE_ID_4ARG
141 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
143 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
147 kiblnd_concurrent_sends_v1(void)
149 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
150 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
152 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
153 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
155 return *kiblnd_tunables.kib_concurrent_sends;
158 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
159 kiblnd_concurrent_sends_v1() : \
160 *kiblnd_tunables.kib_concurrent_sends)
161 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
162 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
163 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
165 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
166 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
167 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
168 *kiblnd_tunables.kib_map_on_demand : \
169 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
170 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
171 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
173 /************************/
174 /* derived constants... */
175 /* Pools (shared by connections on each CPT) */
176 /* These pools can grow at runtime, so don't need give a very large value */
177 #define IBLND_TX_POOL 256
178 #define IBLND_PMR_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 int ibh_nmrs; /* # of global MRs */
234 struct ib_mr **ibh_mrs; /* global MR */
235 struct ib_pd *ibh_pd; /* PD */
236 kib_dev_t *ibh_dev; /* owner */
237 atomic_t ibh_ref; /* refcount */
240 /** # of seconds to keep pool alive */
241 #define IBLND_POOL_DEADLINE 300
242 /** # of seconds to retry if allocation failed */
243 #define IBLND_POOL_RETRY 1
247 int ibp_npages; /* # pages */
248 struct page *ibp_pages[0]; /* page array */
254 struct list_head pmr_list; /* chain node */
255 struct ib_phys_buf *pmr_ipb; /* physical buffer */
256 struct ib_mr *pmr_mr; /* IB MR */
257 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
258 __u64 pmr_iova; /* Virtual I/O address */
259 int pmr_refcount; /* reference count */
265 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
266 int inc, struct kib_pool **pp_po);
267 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
268 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
269 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
273 #define IBLND_POOL_NAME_LEN 32
275 typedef struct kib_poolset
279 /* network it belongs to */
280 struct kib_net *ps_net;
282 char ps_name[IBLND_POOL_NAME_LEN];
284 struct list_head ps_pool_list;
285 /* failed pool list */
286 struct list_head ps_failed_pool_list;
287 /* time stamp for retry if failed to allocate */
288 cfs_time_t ps_next_retry;
289 /* is allocating new pool */
296 /* create a new pool */
297 kib_ps_pool_create_t ps_pool_create;
299 kib_ps_pool_destroy_t ps_pool_destroy;
300 /* initialize new allocated node */
301 kib_ps_node_init_t ps_node_init;
303 kib_ps_node_fini_t ps_node_fini;
306 typedef struct kib_pool
308 /* chain on pool list */
309 struct list_head po_list;
310 /* pre-allocated node */
311 struct list_head po_free_list;
312 /* pool_set of this pool */
313 kib_poolset_t *po_owner;
314 /* deadline of this pool */
315 cfs_time_t po_deadline;
316 /* # of elements in use */
318 /* pool is created on failed HCA */
320 /* # of pre-allocated elements */
325 kib_poolset_t tps_poolset; /* pool-set */
326 __u64 tps_next_tx_cookie; /* cookie of TX */
330 kib_pool_t tpo_pool; /* pool */
331 struct kib_hca_dev *tpo_hdev; /* device for this pool */
332 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
333 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
337 kib_poolset_t pps_poolset; /* pool-set */
340 typedef struct kib_pmr_pool {
341 struct kib_hca_dev *ppo_hdev; /* device for this pool */
342 kib_pool_t ppo_pool; /* pool */
347 spinlock_t fps_lock; /* serialize */
348 struct kib_net *fps_net; /* IB network */
349 struct list_head fps_pool_list; /* FMR pool list */
350 struct list_head fps_failed_pool_list; /* FMR pool list */
351 __u64 fps_version; /* validity stamp */
352 int fps_cpt; /* CPT id */
354 int fps_flush_trigger;
355 /* is allocating new pool */
357 /* time stamp for retry if failed to allocate */
358 cfs_time_t fps_next_retry;
363 struct list_head fpo_list; /* chain on pool list */
364 struct kib_hca_dev *fpo_hdev; /* device for this pool */
365 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
366 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
367 cfs_time_t fpo_deadline; /* deadline of this pool */
368 int fpo_failed; /* fmr pool is failed */
369 int fpo_map_count; /* # of mapped FMR */
373 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
374 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
377 typedef struct kib_net
379 /* chain on kib_dev_t::ibd_nets */
380 struct list_head ibn_list;
381 __u64 ibn_incarnation;/* my epoch */
382 int ibn_init; /* initialisation state */
383 int ibn_shutdown; /* shutting down? */
385 atomic_t ibn_npeers; /* # peers extant */
386 atomic_t ibn_nconns; /* # connections extant */
388 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
389 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
390 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
392 kib_dev_t *ibn_dev; /* underlying IB device */
395 #define KIB_THREAD_SHIFT 16
396 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
397 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
398 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
400 struct kib_sched_info {
403 /* schedulers sleep here */
404 wait_queue_head_t ibs_waitq;
405 /* conns to check for rx completions */
406 struct list_head ibs_conns;
407 /* number of scheduler threads */
409 /* max allowed scheduler threads */
410 int ibs_nthreads_max;
411 int ibs_cpt; /* CPT id */
416 int kib_init; /* initialisation state */
417 int kib_shutdown; /* shut down? */
418 struct list_head kib_devs; /* IB devices extant */
419 /* list head of failed devices */
420 struct list_head kib_failed_devs;
421 /* schedulers sleep here */
422 wait_queue_head_t kib_failover_waitq;
423 atomic_t kib_nthreads; /* # live threads */
424 /* stabilize net/dev/peer/conn ops */
425 rwlock_t kib_global_lock;
426 /* hash table of all my known peers */
427 struct list_head *kib_peers;
428 /* size of kib_peers */
429 int kib_peer_hash_size;
430 /* the connd task (serialisation assertions) */
432 /* connections to setup/teardown */
433 struct list_head kib_connd_conns;
434 /* connections with zero refcount */
435 struct list_head kib_connd_zombies;
436 /* connection daemon sleeps here */
437 wait_queue_head_t kib_connd_waitq;
438 spinlock_t kib_connd_lock; /* serialise */
439 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
440 /* percpt data for schedulers */
441 struct kib_sched_info **kib_scheds;
444 #define IBLND_INIT_NOTHING 0
445 #define IBLND_INIT_DATA 1
446 #define IBLND_INIT_ALL 2
448 /************************************************************************
449 * IB Wire message format.
450 * These are sent in sender's byte order (i.e. receiver flips).
453 typedef struct kib_connparams
455 __u16 ibcp_queue_depth;
456 __u16 ibcp_max_frags;
457 __u32 ibcp_max_msg_size;
458 } WIRE_ATTR kib_connparams_t;
462 lnet_hdr_t ibim_hdr; /* portals header */
463 char ibim_payload[0]; /* piggy-backed payload */
464 } WIRE_ATTR kib_immediate_msg_t;
468 __u32 rf_nob; /* # bytes this frag */
469 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
470 } WIRE_ATTR kib_rdma_frag_t;
474 __u32 rd_key; /* local/remote key */
475 __u32 rd_nfrags; /* # fragments */
476 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
477 } WIRE_ATTR kib_rdma_desc_t;
481 lnet_hdr_t ibprm_hdr; /* portals header */
482 __u64 ibprm_cookie; /* opaque completion cookie */
483 } WIRE_ATTR kib_putreq_msg_t;
487 __u64 ibpam_src_cookie; /* reflected completion cookie */
488 __u64 ibpam_dst_cookie; /* opaque completion cookie */
489 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
490 } WIRE_ATTR kib_putack_msg_t;
494 lnet_hdr_t ibgm_hdr; /* portals header */
495 __u64 ibgm_cookie; /* opaque completion cookie */
496 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
497 } WIRE_ATTR kib_get_msg_t;
501 __u64 ibcm_cookie; /* opaque completion cookie */
502 __s32 ibcm_status; /* < 0 failure: >= 0 length */
503 } WIRE_ATTR kib_completion_msg_t;
507 /* First 2 fields fixed FOR ALL TIME */
508 __u32 ibm_magic; /* I'm an ibnal message */
509 __u16 ibm_version; /* this is my version number */
511 __u8 ibm_type; /* msg type */
512 __u8 ibm_credits; /* returned credits */
513 __u32 ibm_nob; /* # bytes in whole message */
514 __u32 ibm_cksum; /* checksum (0 == no checksum) */
515 __u64 ibm_srcnid; /* sender's NID */
516 __u64 ibm_srcstamp; /* sender's incarnation */
517 __u64 ibm_dstnid; /* destination's NID */
518 __u64 ibm_dststamp; /* destination's incarnation */
521 kib_connparams_t connparams;
522 kib_immediate_msg_t immediate;
523 kib_putreq_msg_t putreq;
524 kib_putack_msg_t putack;
526 kib_completion_msg_t completion;
528 } WIRE_ATTR kib_msg_t;
530 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
532 #define IBLND_MSG_VERSION_1 0x11
533 #define IBLND_MSG_VERSION_2 0x12
534 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
536 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
537 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
538 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
539 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
540 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
541 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
542 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
543 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
544 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
545 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
548 __u32 ibr_magic; /* sender's magic */
549 __u16 ibr_version; /* sender's version */
550 __u8 ibr_why; /* reject reason */
551 __u8 ibr_padding; /* padding */
552 __u64 ibr_incarnation; /* incarnation of peer */
553 kib_connparams_t ibr_cp; /* connection parameters */
554 } WIRE_ATTR kib_rej_t;
556 /* connection rejection reasons */
557 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
558 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
559 #define IBLND_REJECT_FATAL 3 /* Anything else */
561 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
562 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
564 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
565 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
567 /***********************************************************************/
569 typedef struct kib_rx /* receive message */
571 /* queue for attention */
572 struct list_head rx_list;
574 struct kib_conn *rx_conn;
575 /* # bytes received (-1 while posted) */
577 /* completion status */
578 enum ib_wc_status rx_status;
579 /* message buffer (host vaddr) */
581 /* message buffer (I/O addr) */
583 /* for dma_unmap_single() */
584 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
585 /* receive work item... */
586 struct ib_recv_wr rx_wrq;
587 /* ...and its memory */
588 struct ib_sge rx_sge;
591 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
592 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
593 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
594 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
596 typedef struct kib_tx /* transmit message */
598 /* queue on idle_txs ibc_tx_queue etc. */
599 struct list_head tx_list;
601 kib_tx_pool_t *tx_pool;
603 struct kib_conn *tx_conn;
604 /* # tx callbacks outstanding */
606 /* queued for sending */
608 /* waiting for peer */
610 /* LNET completion status */
612 /* completion deadline */
613 unsigned long tx_deadline;
614 /* completion cookie */
616 /* lnet msgs to finalize on completion */
617 lnet_msg_t *tx_lntmsg[2];
618 /* message buffer (host vaddr) */
620 /* message buffer (I/O addr) */
622 /* for dma_unmap_single() */
623 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
624 /* # send work items */
626 /* send work items... */
627 struct ib_send_wr *tx_wrq;
628 /* ...and their memory */
629 struct ib_sge *tx_sge;
630 /* rdma descriptor */
631 kib_rdma_desc_t *tx_rd;
632 /* # entries in... */
634 /* dma_map_sg descriptor */
635 struct scatterlist *tx_frags;
636 /* rdma phys page addrs */
639 /* MR for physical buffer */
648 typedef struct kib_connvars
650 /* connection-in-progress variables */
654 typedef struct kib_conn
656 /* scheduler information */
657 struct kib_sched_info *ibc_sched;
659 struct kib_peer *ibc_peer;
661 kib_hca_dev_t *ibc_hdev;
662 /* stash on peer's conn list */
663 struct list_head ibc_list;
664 /* schedule for attention */
665 struct list_head ibc_sched_list;
666 /* version of connection */
668 /* which instance of the peer */
669 __u64 ibc_incarnation;
671 atomic_t ibc_refcount;
672 /* what's happening */
674 /* # uncompleted sends */
675 int ibc_nsends_posted;
676 /* # uncompleted NOOPs */
677 int ibc_noops_posted;
678 /* # credits I have */
680 /* # credits to return */
681 int ibc_outstanding_credits;
682 /* # ACK/DONE msg credits */
683 int ibc_reserved_credits;
684 /* set on comms error */
686 /* receive buffers owned */
687 unsigned int ibc_nrx:16;
688 /* scheduled for attention */
689 unsigned int ibc_scheduled:1;
690 /* CQ callback fired */
691 unsigned int ibc_ready:1;
692 /* time of last send */
693 unsigned long ibc_last_send;
694 /** link chain for kiblnd_check_conns only */
695 struct list_head ibc_connd_list;
696 /** rxs completed before ESTABLISHED */
697 struct list_head ibc_early_rxs;
698 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
699 struct list_head ibc_tx_noops;
700 /* sends that need a credit */
701 struct list_head ibc_tx_queue;
702 /* sends that don't need a credit */
703 struct list_head ibc_tx_queue_nocred;
704 /* sends that need to reserve an ACK/DONE msg */
705 struct list_head ibc_tx_queue_rsrvd;
706 /* active tx awaiting completion */
707 struct list_head ibc_active_txs;
712 /* premapped rx msg pages */
713 kib_pages_t *ibc_rx_pages;
716 struct rdma_cm_id *ibc_cmid;
717 /* completion queue */
718 struct ib_cq *ibc_cq;
720 /* in-progress connection state */
721 kib_connvars_t *ibc_connvars;
724 #define IBLND_CONN_INIT 0 /* being initialised */
725 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
726 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
727 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
728 #define IBLND_CONN_CLOSING 4 /* being closed */
729 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
731 typedef struct kib_peer
733 /* stash on global peer list */
734 struct list_head ibp_list;
735 /* who's on the other end(s) */
740 atomic_t ibp_refcount;
741 /* all active connections */
742 struct list_head ibp_conns;
743 /* msgs waiting for a conn */
744 struct list_head ibp_tx_queue;
745 /* version of peer */
747 /* incarnation of peer */
748 __u64 ibp_incarnation;
749 /* current active connection attempts */
751 /* current passive connection attempts */
753 /* errno on closing this peer */
755 /* when (in jiffies) I was last alive */
756 cfs_time_t ibp_last_alive;
759 extern kib_data_t kiblnd_data;
761 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
764 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
766 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
767 atomic_inc(&hdev->ibh_ref);
771 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
773 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
774 if (atomic_dec_and_test(&hdev->ibh_ref))
775 kiblnd_hdev_destroy(hdev);
779 kiblnd_dev_can_failover(kib_dev_t *dev)
781 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
784 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
787 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
790 return dev->ibd_can_failover;
793 #define kiblnd_conn_addref(conn) \
795 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
796 (conn), atomic_read(&(conn)->ibc_refcount)); \
797 atomic_inc(&(conn)->ibc_refcount); \
800 #define kiblnd_conn_decref(conn) \
802 unsigned long flags; \
804 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
805 (conn), atomic_read(&(conn)->ibc_refcount)); \
806 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
807 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
808 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
809 list_add_tail(&(conn)->ibc_list, \
810 &kiblnd_data.kib_connd_zombies); \
811 wake_up(&kiblnd_data.kib_connd_waitq); \
812 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
816 #define kiblnd_peer_addref(peer) \
818 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
819 (peer), libcfs_nid2str((peer)->ibp_nid), \
820 atomic_read (&(peer)->ibp_refcount)); \
821 atomic_inc(&(peer)->ibp_refcount); \
824 #define kiblnd_peer_decref(peer) \
826 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
827 (peer), libcfs_nid2str((peer)->ibp_nid), \
828 atomic_read (&(peer)->ibp_refcount)); \
829 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
830 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
831 kiblnd_destroy_peer(peer); \
834 static inline struct list_head *
835 kiblnd_nid2peerlist (lnet_nid_t nid)
838 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
840 return &kiblnd_data.kib_peers[hash];
844 kiblnd_peer_active (kib_peer_t *peer)
846 /* Am I in the peer hash table? */
847 return !list_empty(&peer->ibp_list);
850 static inline kib_conn_t *
851 kiblnd_get_conn_locked (kib_peer_t *peer)
853 LASSERT(!list_empty(&peer->ibp_conns));
855 /* just return the first connection */
856 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
860 kiblnd_send_keepalive(kib_conn_t *conn)
862 return (*kiblnd_tunables.kib_keepalive > 0) &&
863 cfs_time_after(jiffies, conn->ibc_last_send +
864 *kiblnd_tunables.kib_keepalive*HZ);
868 kiblnd_need_noop(kib_conn_t *conn)
870 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
872 if (conn->ibc_outstanding_credits <
873 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
874 !kiblnd_send_keepalive(conn))
875 return 0; /* No need to send NOOP */
877 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
878 if (!list_empty(&conn->ibc_tx_queue_nocred))
879 return 0; /* NOOP can be piggybacked */
881 /* No tx to piggyback NOOP onto or no credit to send a tx */
882 return (list_empty(&conn->ibc_tx_queue) ||
883 conn->ibc_credits == 0);
886 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
887 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
888 conn->ibc_credits == 0) /* no credit */
891 if (conn->ibc_credits == 1 && /* last credit reserved for */
892 conn->ibc_outstanding_credits == 0) /* giving back credits */
895 /* No tx to piggyback NOOP onto or no credit to send a tx */
896 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
900 kiblnd_abort_receives(kib_conn_t *conn)
902 ib_modify_qp(conn->ibc_cmid->qp,
903 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
906 static inline const char *
907 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
909 if (q == &conn->ibc_tx_queue)
912 if (q == &conn->ibc_tx_queue_rsrvd)
913 return "tx_queue_rsrvd";
915 if (q == &conn->ibc_tx_queue_nocred)
916 return "tx_queue_nocred";
918 if (q == &conn->ibc_active_txs)
925 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
926 * lowest bits of the work request id to stash the work item type. */
928 #define IBLND_WID_TX 0
929 #define IBLND_WID_RDMA 1
930 #define IBLND_WID_RX 2
931 #define IBLND_WID_MASK 3UL
934 kiblnd_ptr2wreqid (void *ptr, int type)
936 unsigned long lptr = (unsigned long)ptr;
938 LASSERT ((lptr & IBLND_WID_MASK) == 0);
939 LASSERT ((type & ~IBLND_WID_MASK) == 0);
940 return (__u64)(lptr | type);
944 kiblnd_wreqid2ptr (__u64 wreqid)
946 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
950 kiblnd_wreqid2type (__u64 wreqid)
952 return (wreqid & IBLND_WID_MASK);
956 kiblnd_set_conn_state (kib_conn_t *conn, int state)
958 conn->ibc_state = state;
963 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
965 msg->ibm_type = type;
966 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
970 kiblnd_rd_size (kib_rdma_desc_t *rd)
975 for (i = size = 0; i < rd->rd_nfrags; i++)
976 size += rd->rd_frags[i].rf_nob;
982 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
984 return rd->rd_frags[index].rf_addr;
988 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
990 return rd->rd_frags[index].rf_nob;
994 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1000 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1002 if (nob < rd->rd_frags[index].rf_nob) {
1003 rd->rd_frags[index].rf_addr += nob;
1004 rd->rd_frags[index].rf_nob -= nob;
1013 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1015 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1016 msgtype == IBLND_MSG_PUT_ACK);
1018 return msgtype == IBLND_MSG_GET_REQ ?
1019 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1020 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1023 #ifdef HAVE_OFED_IB_DMA_MAP
1026 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1028 return ib_dma_mapping_error(dev, dma_addr);
1031 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1032 void *msg, size_t size,
1033 enum dma_data_direction direction)
1035 return ib_dma_map_single(dev, msg, size, direction);
1038 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1039 __u64 addr, size_t size,
1040 enum dma_data_direction direction)
1042 ib_dma_unmap_single(dev, addr, size, direction);
1045 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1046 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1048 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1049 struct scatterlist *sg, int nents,
1050 enum dma_data_direction direction)
1052 return ib_dma_map_sg(dev, sg, nents, direction);
1055 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1056 struct scatterlist *sg, int nents,
1057 enum dma_data_direction direction)
1059 ib_dma_unmap_sg(dev, sg, nents, direction);
1062 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1063 struct scatterlist *sg)
1065 return ib_sg_dma_address(dev, sg);
1068 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1069 struct scatterlist *sg)
1071 return ib_sg_dma_len(dev, sg);
1074 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1075 * right because OFED1.2 defines it as const, to use it we have to add
1076 * (void *) cast to overcome "const" */
1078 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1079 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1084 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
1086 return dma_mapping_error(dma_addr);
1089 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
1090 void *msg, size_t size,
1091 enum dma_data_direction direction)
1093 return dma_map_single(dev->dma_device, msg, size, direction);
1096 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1097 dma_addr_t addr, size_t size,
1098 enum dma_data_direction direction)
1100 dma_unmap_single(dev->dma_device, addr, size, direction);
1103 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
1104 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
1106 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1107 struct scatterlist *sg, int nents,
1108 enum dma_data_direction direction)
1110 return dma_map_sg(dev->dma_device, sg, nents, direction);
1113 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1114 struct scatterlist *sg, int nents,
1115 enum dma_data_direction direction)
1117 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1121 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1122 struct scatterlist *sg)
1124 return sg_dma_address(sg);
1128 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1129 struct scatterlist *sg)
1131 return sg_dma_len(sg);
1134 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1135 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1139 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1140 kib_rdma_desc_t *rd);
1141 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1142 __u64 addr, __u64 size);
1143 void kiblnd_map_rx_descs(kib_conn_t *conn);
1144 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1145 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1146 kib_rdma_desc_t *rd, int nfrags);
1147 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1148 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1149 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1151 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1152 int npages, __u64 iov, kib_fmr_t *fmr);
1153 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1155 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1156 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1157 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1159 int kiblnd_startup (lnet_ni_t *ni);
1160 void kiblnd_shutdown (lnet_ni_t *ni);
1161 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1162 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1164 int kiblnd_tunables_init(void);
1165 void kiblnd_tunables_fini(void);
1167 int kiblnd_connd (void *arg);
1168 int kiblnd_scheduler(void *arg);
1169 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1170 int kiblnd_failover_thread (void *arg);
1172 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1173 void kiblnd_free_pages (kib_pages_t *p);
1175 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1176 struct rdma_cm_event *event);
1177 int kiblnd_translate_mtu(int value);
1179 int kiblnd_dev_failover(kib_dev_t *dev);
1180 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1181 void kiblnd_destroy_peer (kib_peer_t *peer);
1182 void kiblnd_destroy_dev (kib_dev_t *dev);
1183 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1184 void kiblnd_peer_alive (kib_peer_t *peer);
1185 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1186 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1187 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1188 int version, __u64 incarnation);
1189 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1191 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1192 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1193 int state, int version);
1194 void kiblnd_destroy_conn (kib_conn_t *conn);
1195 void kiblnd_close_conn (kib_conn_t *conn, int error);
1196 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1198 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1199 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1201 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1202 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1203 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1204 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1205 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1206 void kiblnd_check_sends (kib_conn_t *conn);
1208 void kiblnd_qp_event(struct ib_event *event, void *arg);
1209 void kiblnd_cq_event(struct ib_event *event, void *arg);
1210 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1212 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1213 int credits, lnet_nid_t dstnid, __u64 dststamp);
1214 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1215 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1217 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1218 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1219 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1220 unsigned int offset, unsigned int mlen, unsigned int rlen);