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, 2012, 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 #define DEBUG_SUBSYSTEM S_LND
65 #include "../../../include/linux/libcfs/libcfs.h"
66 #include "../../../include/linux/lnet/lnet.h"
67 #include "../../../include/linux/lnet/lib-lnet.h"
69 #include <rdma/rdma_cm.h>
70 #include <rdma/ib_cm.h>
71 #include <rdma/ib_verbs.h>
72 #include <rdma/ib_fmr_pool.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 != 0 ? \
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 /* TX messages (shared by all connections) */
166 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
168 /* RX messages (per connection) */
169 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
170 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
171 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
173 /* WRs and CQEs (per connection) */
174 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
175 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
176 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
180 /* o2iblnd can run over aliased interface */
182 #define KIB_IFNAME_SIZE IFALIASZ
184 #define KIB_IFNAME_SIZE 256
188 struct list_head ibd_list
; /* chain on kib_devs */
189 struct list_head ibd_fail_list
; /* chain on kib_failed_devs */
190 __u32 ibd_ifip
; /* IPoIB interface IP */
192 /* IPoIB interface name */
193 char ibd_ifname
[KIB_IFNAME_SIZE
];
194 int ibd_nnets
; /* # nets extant */
196 unsigned long ibd_next_failover
;
197 int ibd_failed_failover
; /* # failover failures */
198 unsigned int ibd_failover
; /* failover in progress */
199 unsigned int ibd_can_failover
; /* IPoIB interface is a bonding master */
200 struct list_head ibd_nets
;
201 struct kib_hca_dev
*ibd_hdev
;
204 typedef struct kib_hca_dev
{
205 struct rdma_cm_id
*ibh_cmid
; /* listener cmid */
206 struct ib_device
*ibh_ibdev
; /* IB device */
207 int ibh_page_shift
; /* page shift of current HCA */
208 int ibh_page_size
; /* page size of current HCA */
209 __u64 ibh_page_mask
; /* page mask of current HCA */
210 int ibh_mr_shift
; /* bits shift of max MR size */
211 __u64 ibh_mr_size
; /* size of MR */
212 int ibh_nmrs
; /* # of global MRs */
213 struct ib_mr
**ibh_mrs
; /* global MR */
214 struct ib_pd
*ibh_pd
; /* PD */
215 kib_dev_t
*ibh_dev
; /* owner */
216 atomic_t ibh_ref
; /* refcount */
219 /** # of seconds to keep pool alive */
220 #define IBLND_POOL_DEADLINE 300
221 /** # of seconds to retry if allocation failed */
222 #define IBLND_POOL_RETRY 1
225 int ibp_npages
; /* # pages */
226 struct page
*ibp_pages
[0]; /* page array */
232 typedef int (*kib_ps_pool_create_t
)(struct kib_poolset
*ps
,
233 int inc
, struct kib_pool
**pp_po
);
234 typedef void (*kib_ps_pool_destroy_t
)(struct kib_pool
*po
);
235 typedef void (*kib_ps_node_init_t
)(struct kib_pool
*po
, struct list_head
*node
);
236 typedef void (*kib_ps_node_fini_t
)(struct kib_pool
*po
, struct list_head
*node
);
240 #define IBLND_POOL_NAME_LEN 32
242 typedef struct kib_poolset
{
243 spinlock_t ps_lock
; /* serialize */
244 struct kib_net
*ps_net
; /* network it belongs to */
245 char ps_name
[IBLND_POOL_NAME_LEN
]; /* pool set name */
246 struct list_head ps_pool_list
; /* list of pools */
247 struct list_head ps_failed_pool_list
;/* failed pool list */
248 unsigned long ps_next_retry
; /* time stamp for retry if */
249 /* failed to allocate */
250 int ps_increasing
; /* is allocating new pool */
251 int ps_pool_size
; /* new pool size */
252 int ps_cpt
; /* CPT id */
254 kib_ps_pool_create_t ps_pool_create
; /* create a new pool */
255 kib_ps_pool_destroy_t ps_pool_destroy
; /* destroy a pool */
256 kib_ps_node_init_t ps_node_init
; /* initialize new allocated node */
257 kib_ps_node_fini_t ps_node_fini
; /* finalize node */
260 typedef struct kib_pool
{
261 struct list_head po_list
; /* chain on pool list */
262 struct list_head po_free_list
; /* pre-allocated node */
263 kib_poolset_t
*po_owner
; /* pool_set of this pool */
264 unsigned long po_deadline
; /* deadline of this pool */
265 int po_allocated
; /* # of elements in use */
266 int po_failed
; /* pool is created on failed HCA */
267 int po_size
; /* # of pre-allocated elements */
271 kib_poolset_t tps_poolset
; /* pool-set */
272 __u64 tps_next_tx_cookie
; /* cookie of TX */
276 kib_pool_t tpo_pool
; /* pool */
277 struct kib_hca_dev
*tpo_hdev
; /* device for this pool */
278 struct kib_tx
*tpo_tx_descs
; /* all the tx descriptors */
279 kib_pages_t
*tpo_tx_pages
; /* premapped tx msg pages */
283 spinlock_t fps_lock
; /* serialize */
284 struct kib_net
*fps_net
; /* IB network */
285 struct list_head fps_pool_list
; /* FMR pool list */
286 struct list_head fps_failed_pool_list
;/* FMR pool list */
287 __u64 fps_version
; /* validity stamp */
288 int fps_cpt
; /* CPT id */
290 int fps_flush_trigger
;
291 int fps_increasing
; /* is allocating new pool */
292 unsigned long fps_next_retry
; /* time stamp for retry if*/
293 /* failed to allocate */
297 struct list_head fpo_list
; /* chain on pool list */
298 struct kib_hca_dev
*fpo_hdev
; /* device for this pool */
299 kib_fmr_poolset_t
*fpo_owner
; /* owner of this pool */
300 struct ib_fmr_pool
*fpo_fmr_pool
; /* IB FMR pool */
301 unsigned long fpo_deadline
; /* deadline of this pool */
302 int fpo_failed
; /* fmr pool is failed */
303 int fpo_map_count
; /* # of mapped FMR */
307 struct ib_pool_fmr
*fmr_pfmr
; /* IB pool fmr */
308 kib_fmr_pool_t
*fmr_pool
; /* pool of FMR */
311 typedef struct kib_net
{
312 struct list_head ibn_list
; /* chain on kib_dev_t::ibd_nets */
313 __u64 ibn_incarnation
;/* my epoch */
314 int ibn_init
; /* initialisation state */
315 int ibn_shutdown
; /* shutting down? */
317 atomic_t ibn_npeers
; /* # peers extant */
318 atomic_t ibn_nconns
; /* # connections extant */
320 kib_tx_poolset_t
**ibn_tx_ps
; /* tx pool-set */
321 kib_fmr_poolset_t
**ibn_fmr_ps
; /* fmr pool-set */
323 kib_dev_t
*ibn_dev
; /* underlying IB device */
326 #define KIB_THREAD_SHIFT 16
327 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
328 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
329 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
331 struct kib_sched_info
{
332 spinlock_t ibs_lock
; /* serialise */
333 wait_queue_head_t ibs_waitq
; /* schedulers sleep here */
334 struct list_head ibs_conns
; /* conns to check for rx completions */
335 int ibs_nthreads
; /* number of scheduler threads */
336 int ibs_nthreads_max
; /* max allowed scheduler threads */
337 int ibs_cpt
; /* CPT id */
341 int kib_init
; /* initialisation state */
342 int kib_shutdown
; /* shut down? */
343 struct list_head kib_devs
; /* IB devices extant */
344 struct list_head kib_failed_devs
; /* list head of failed devices */
345 wait_queue_head_t kib_failover_waitq
; /* schedulers sleep here */
346 atomic_t kib_nthreads
; /* # live threads */
347 rwlock_t kib_global_lock
; /* stabilize net/dev/peer/conn ops */
348 struct list_head
*kib_peers
; /* hash table of all my known peers */
349 int kib_peer_hash_size
; /* size of kib_peers */
350 void *kib_connd
; /* the connd task (serialisation assertions) */
351 struct list_head kib_connd_conns
; /* connections to setup/teardown */
352 struct list_head kib_connd_zombies
; /* connections with zero refcount */
353 wait_queue_head_t kib_connd_waitq
; /* connection daemon sleeps here */
354 spinlock_t kib_connd_lock
; /* serialise */
355 struct ib_qp_attr kib_error_qpa
; /* QP->ERROR */
356 struct kib_sched_info
**kib_scheds
; /* percpt data for schedulers */
359 #define IBLND_INIT_NOTHING 0
360 #define IBLND_INIT_DATA 1
361 #define IBLND_INIT_ALL 2
363 /************************************************************************
364 * IB Wire message format.
365 * These are sent in sender's byte order (i.e. receiver flips).
368 typedef struct kib_connparams
{
369 __u16 ibcp_queue_depth
;
370 __u16 ibcp_max_frags
;
371 __u32 ibcp_max_msg_size
;
372 } WIRE_ATTR kib_connparams_t
;
375 lnet_hdr_t ibim_hdr
; /* portals header */
376 char ibim_payload
[0]; /* piggy-backed payload */
377 } WIRE_ATTR kib_immediate_msg_t
;
380 __u32 rf_nob
; /* # bytes this frag */
381 __u64 rf_addr
; /* CAVEAT EMPTOR: misaligned!! */
382 } WIRE_ATTR kib_rdma_frag_t
;
385 __u32 rd_key
; /* local/remote key */
386 __u32 rd_nfrags
; /* # fragments */
387 kib_rdma_frag_t rd_frags
[0]; /* buffer frags */
388 } WIRE_ATTR kib_rdma_desc_t
;
391 lnet_hdr_t ibprm_hdr
; /* portals header */
392 __u64 ibprm_cookie
; /* opaque completion cookie */
393 } WIRE_ATTR kib_putreq_msg_t
;
396 __u64 ibpam_src_cookie
; /* reflected completion cookie */
397 __u64 ibpam_dst_cookie
; /* opaque completion cookie */
398 kib_rdma_desc_t ibpam_rd
; /* sender's sink buffer */
399 } WIRE_ATTR kib_putack_msg_t
;
402 lnet_hdr_t ibgm_hdr
; /* portals header */
403 __u64 ibgm_cookie
; /* opaque completion cookie */
404 kib_rdma_desc_t ibgm_rd
; /* rdma descriptor */
405 } WIRE_ATTR kib_get_msg_t
;
408 __u64 ibcm_cookie
; /* opaque completion cookie */
409 __s32 ibcm_status
; /* < 0 failure: >= 0 length */
410 } WIRE_ATTR kib_completion_msg_t
;
413 /* First 2 fields fixed FOR ALL TIME */
414 __u32 ibm_magic
; /* I'm an ibnal message */
415 __u16 ibm_version
; /* this is my version number */
417 __u8 ibm_type
; /* msg type */
418 __u8 ibm_credits
; /* returned credits */
419 __u32 ibm_nob
; /* # bytes in whole message */
420 __u32 ibm_cksum
; /* checksum (0 == no checksum) */
421 __u64 ibm_srcnid
; /* sender's NID */
422 __u64 ibm_srcstamp
; /* sender's incarnation */
423 __u64 ibm_dstnid
; /* destination's NID */
424 __u64 ibm_dststamp
; /* destination's incarnation */
427 kib_connparams_t connparams
;
428 kib_immediate_msg_t immediate
;
429 kib_putreq_msg_t putreq
;
430 kib_putack_msg_t putack
;
432 kib_completion_msg_t completion
;
434 } WIRE_ATTR kib_msg_t
;
436 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
438 #define IBLND_MSG_VERSION_1 0x11
439 #define IBLND_MSG_VERSION_2 0x12
440 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
442 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
443 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
444 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
445 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
446 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
447 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
448 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
449 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
450 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
451 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
454 __u32 ibr_magic
; /* sender's magic */
455 __u16 ibr_version
; /* sender's version */
456 __u8 ibr_why
; /* reject reason */
457 __u8 ibr_padding
; /* padding */
458 __u64 ibr_incarnation
; /* incarnation of peer */
459 kib_connparams_t ibr_cp
; /* connection parameters */
460 } WIRE_ATTR kib_rej_t
;
462 /* connection rejection reasons */
463 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
464 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
465 #define IBLND_REJECT_FATAL 3 /* Anything else */
466 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
467 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
468 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match */
470 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't */
473 /***********************************************************************/
475 typedef struct kib_rx
/* receive message */
477 struct list_head rx_list
; /* queue for attention */
478 struct kib_conn
*rx_conn
; /* owning conn */
479 int rx_nob
; /* # bytes received (-1 while posted) */
480 enum ib_wc_status rx_status
; /* completion status */
481 kib_msg_t
*rx_msg
; /* message buffer (host vaddr) */
482 __u64 rx_msgaddr
; /* message buffer (I/O addr) */
483 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap
); /* for dma_unmap_single() */
484 struct ib_recv_wr rx_wrq
; /* receive work item... */
485 struct ib_sge rx_sge
; /* ...and its memory */
488 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
489 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
490 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
491 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give self back 1 reserved credit */
493 typedef struct kib_tx
/* transmit message */
495 struct list_head tx_list
; /* queue on idle_txs ibc_tx_queue etc. */
496 kib_tx_pool_t
*tx_pool
; /* pool I'm from */
497 struct kib_conn
*tx_conn
; /* owning conn */
498 short tx_sending
; /* # tx callbacks outstanding */
499 short tx_queued
; /* queued for sending */
500 short tx_waiting
; /* waiting for peer */
501 int tx_status
; /* LNET completion status */
502 unsigned long tx_deadline
; /* completion deadline */
503 __u64 tx_cookie
; /* completion cookie */
504 lnet_msg_t
*tx_lntmsg
[2]; /* lnet msgs to finalize on completion */
505 kib_msg_t
*tx_msg
; /* message buffer (host vaddr) */
506 __u64 tx_msgaddr
; /* message buffer (I/O addr) */
507 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap
); /* for dma_unmap_single() */
508 int tx_nwrq
; /* # send work items */
509 struct ib_rdma_wr
*tx_wrq
; /* send work items... */
510 struct ib_sge
*tx_sge
; /* ...and their memory */
511 kib_rdma_desc_t
*tx_rd
; /* rdma descriptor */
512 int tx_nfrags
; /* # entries in... */
513 struct scatterlist
*tx_frags
; /* dma_map_sg descriptor */
514 __u64
*tx_pages
; /* rdma phys page addrs */
515 kib_fmr_t fmr
; /* FMR */
516 int tx_dmadir
; /* dma direction */
519 typedef struct kib_connvars
{
520 kib_msg_t cv_msg
; /* connection-in-progress variables */
523 typedef struct kib_conn
{
524 struct kib_sched_info
*ibc_sched
; /* scheduler information */
525 struct kib_peer
*ibc_peer
; /* owning peer */
526 kib_hca_dev_t
*ibc_hdev
; /* HCA bound on */
527 struct list_head ibc_list
; /* stash on peer's conn list */
528 struct list_head ibc_sched_list
; /* schedule for attention */
529 __u16 ibc_version
; /* version of connection */
530 __u64 ibc_incarnation
; /* which instance of the peer */
531 atomic_t ibc_refcount
; /* # users */
532 int ibc_state
; /* what's happening */
533 int ibc_nsends_posted
; /* # uncompleted sends */
534 int ibc_noops_posted
; /* # uncompleted NOOPs */
535 int ibc_credits
; /* # credits I have */
536 int ibc_outstanding_credits
; /* # credits to return */
537 int ibc_reserved_credits
; /* # ACK/DONE msg credits */
538 int ibc_comms_error
; /* set on comms error */
539 unsigned int ibc_nrx
:16; /* receive buffers owned */
540 unsigned int ibc_scheduled
:1; /* scheduled for attention */
541 unsigned int ibc_ready
:1; /* CQ callback fired */
542 unsigned long ibc_last_send
; /* time of last send */
543 struct list_head ibc_connd_list
; /* link chain for */
544 /* kiblnd_check_conns only */
545 struct list_head ibc_early_rxs
; /* rxs completed before ESTABLISHED */
546 struct list_head ibc_tx_noops
; /* IBLND_MSG_NOOPs for */
547 /* IBLND_MSG_VERSION_1 */
548 struct list_head ibc_tx_queue
; /* sends that need a credit */
549 struct list_head ibc_tx_queue_nocred
; /* sends that don't need a */
551 struct list_head ibc_tx_queue_rsrvd
; /* sends that need to */
552 /* reserve an ACK/DONE msg */
553 struct list_head ibc_active_txs
; /* active tx awaiting completion */
554 spinlock_t ibc_lock
; /* serialise */
555 kib_rx_t
*ibc_rxs
; /* the rx descs */
556 kib_pages_t
*ibc_rx_pages
; /* premapped rx msg pages */
558 struct rdma_cm_id
*ibc_cmid
; /* CM id */
559 struct ib_cq
*ibc_cq
; /* completion queue */
561 kib_connvars_t
*ibc_connvars
; /* in-progress connection state */
564 #define IBLND_CONN_INIT 0 /* being initialised */
565 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
566 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
567 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
568 #define IBLND_CONN_CLOSING 4 /* being closed */
569 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
571 typedef struct kib_peer
{
572 struct list_head ibp_list
; /* stash on global peer list */
573 lnet_nid_t ibp_nid
; /* who's on the other end(s) */
574 lnet_ni_t
*ibp_ni
; /* LNet interface */
575 atomic_t ibp_refcount
; /* # users */
576 struct list_head ibp_conns
; /* all active connections */
577 struct list_head ibp_tx_queue
; /* msgs waiting for a conn */
578 __u16 ibp_version
; /* version of peer */
579 __u64 ibp_incarnation
; /* incarnation of peer */
580 int ibp_connecting
; /* current active connection attempts
582 int ibp_accepting
; /* current passive connection attempts
584 int ibp_error
; /* errno on closing this peer */
585 unsigned long ibp_last_alive
; /* when (in jiffies) I was last alive
589 extern kib_data_t kiblnd_data
;
591 void kiblnd_hdev_destroy(kib_hca_dev_t
*hdev
);
594 kiblnd_hdev_addref_locked(kib_hca_dev_t
*hdev
)
596 LASSERT(atomic_read(&hdev
->ibh_ref
) > 0);
597 atomic_inc(&hdev
->ibh_ref
);
601 kiblnd_hdev_decref(kib_hca_dev_t
*hdev
)
603 LASSERT(atomic_read(&hdev
->ibh_ref
) > 0);
604 if (atomic_dec_and_test(&hdev
->ibh_ref
))
605 kiblnd_hdev_destroy(hdev
);
609 kiblnd_dev_can_failover(kib_dev_t
*dev
)
611 if (!list_empty(&dev
->ibd_fail_list
)) /* already scheduled */
614 if (*kiblnd_tunables
.kib_dev_failover
== 0) /* disabled */
617 if (*kiblnd_tunables
.kib_dev_failover
> 1) /* force failover */
620 return dev
->ibd_can_failover
;
623 #define kiblnd_conn_addref(conn) \
625 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
626 (conn), atomic_read(&(conn)->ibc_refcount)); \
627 atomic_inc(&(conn)->ibc_refcount); \
630 #define kiblnd_conn_decref(conn) \
632 unsigned long flags; \
634 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
635 (conn), atomic_read(&(conn)->ibc_refcount)); \
636 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
637 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
638 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
639 list_add_tail(&(conn)->ibc_list, \
640 &kiblnd_data.kib_connd_zombies); \
641 wake_up(&kiblnd_data.kib_connd_waitq); \
642 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
646 #define kiblnd_peer_addref(peer) \
648 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
649 (peer), libcfs_nid2str((peer)->ibp_nid), \
650 atomic_read(&(peer)->ibp_refcount)); \
651 atomic_inc(&(peer)->ibp_refcount); \
654 #define kiblnd_peer_decref(peer) \
656 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
657 (peer), libcfs_nid2str((peer)->ibp_nid), \
658 atomic_read(&(peer)->ibp_refcount)); \
659 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
660 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
661 kiblnd_destroy_peer(peer); \
664 static inline struct list_head
*
665 kiblnd_nid2peerlist(lnet_nid_t nid
)
668 ((unsigned int)nid
) % kiblnd_data
.kib_peer_hash_size
;
670 return &kiblnd_data
.kib_peers
[hash
];
674 kiblnd_peer_active(kib_peer_t
*peer
)
676 /* Am I in the peer hash table? */
677 return !list_empty(&peer
->ibp_list
);
680 static inline kib_conn_t
*
681 kiblnd_get_conn_locked(kib_peer_t
*peer
)
683 LASSERT(!list_empty(&peer
->ibp_conns
));
685 /* just return the first connection */
686 return list_entry(peer
->ibp_conns
.next
, kib_conn_t
, ibc_list
);
690 kiblnd_send_keepalive(kib_conn_t
*conn
)
692 return (*kiblnd_tunables
.kib_keepalive
> 0) &&
693 cfs_time_after(jiffies
, conn
->ibc_last_send
+
694 *kiblnd_tunables
.kib_keepalive
*HZ
);
698 kiblnd_need_noop(kib_conn_t
*conn
)
700 LASSERT(conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
702 if (conn
->ibc_outstanding_credits
<
703 IBLND_CREDITS_HIGHWATER(conn
->ibc_version
) &&
704 !kiblnd_send_keepalive(conn
))
705 return 0; /* No need to send NOOP */
707 if (IBLND_OOB_CAPABLE(conn
->ibc_version
)) {
708 if (!list_empty(&conn
->ibc_tx_queue_nocred
))
709 return 0; /* NOOP can be piggybacked */
711 /* No tx to piggyback NOOP onto or no credit to send a tx */
712 return (list_empty(&conn
->ibc_tx_queue
) ||
713 conn
->ibc_credits
== 0);
716 if (!list_empty(&conn
->ibc_tx_noops
) || /* NOOP already queued */
717 !list_empty(&conn
->ibc_tx_queue_nocred
) || /* piggyback NOOP */
718 conn
->ibc_credits
== 0) /* no credit */
721 if (conn
->ibc_credits
== 1 && /* last credit reserved for */
722 conn
->ibc_outstanding_credits
== 0) /* giving back credits */
725 /* No tx to piggyback NOOP onto or no credit to send a tx */
726 return (list_empty(&conn
->ibc_tx_queue
) || conn
->ibc_credits
== 1);
730 kiblnd_abort_receives(kib_conn_t
*conn
)
732 ib_modify_qp(conn
->ibc_cmid
->qp
,
733 &kiblnd_data
.kib_error_qpa
, IB_QP_STATE
);
736 static inline const char *
737 kiblnd_queue2str(kib_conn_t
*conn
, struct list_head
*q
)
739 if (q
== &conn
->ibc_tx_queue
)
742 if (q
== &conn
->ibc_tx_queue_rsrvd
)
743 return "tx_queue_rsrvd";
745 if (q
== &conn
->ibc_tx_queue_nocred
)
746 return "tx_queue_nocred";
748 if (q
== &conn
->ibc_active_txs
)
755 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the */
756 /* lowest bits of the work request id to stash the work item type. */
758 #define IBLND_WID_TX 0
759 #define IBLND_WID_RDMA 1
760 #define IBLND_WID_RX 2
761 #define IBLND_WID_MASK 3UL
764 kiblnd_ptr2wreqid(void *ptr
, int type
)
766 unsigned long lptr
= (unsigned long)ptr
;
768 LASSERT((lptr
& IBLND_WID_MASK
) == 0);
769 LASSERT((type
& ~IBLND_WID_MASK
) == 0);
770 return (__u64
)(lptr
| type
);
774 kiblnd_wreqid2ptr(__u64 wreqid
)
776 return (void *)(((unsigned long)wreqid
) & ~IBLND_WID_MASK
);
780 kiblnd_wreqid2type(__u64 wreqid
)
782 return wreqid
& IBLND_WID_MASK
;
786 kiblnd_set_conn_state(kib_conn_t
*conn
, int state
)
788 conn
->ibc_state
= state
;
793 kiblnd_init_msg(kib_msg_t
*msg
, int type
, int body_nob
)
795 msg
->ibm_type
= type
;
796 msg
->ibm_nob
= offsetof(kib_msg_t
, ibm_u
) + body_nob
;
800 kiblnd_rd_size(kib_rdma_desc_t
*rd
)
805 for (i
= size
= 0; i
< rd
->rd_nfrags
; i
++)
806 size
+= rd
->rd_frags
[i
].rf_nob
;
812 kiblnd_rd_frag_addr(kib_rdma_desc_t
*rd
, int index
)
814 return rd
->rd_frags
[index
].rf_addr
;
818 kiblnd_rd_frag_size(kib_rdma_desc_t
*rd
, int index
)
820 return rd
->rd_frags
[index
].rf_nob
;
824 kiblnd_rd_frag_key(kib_rdma_desc_t
*rd
, int index
)
830 kiblnd_rd_consume_frag(kib_rdma_desc_t
*rd
, int index
, __u32 nob
)
832 if (nob
< rd
->rd_frags
[index
].rf_nob
) {
833 rd
->rd_frags
[index
].rf_addr
+= nob
;
834 rd
->rd_frags
[index
].rf_nob
-= nob
;
843 kiblnd_rd_msg_size(kib_rdma_desc_t
*rd
, int msgtype
, int n
)
845 LASSERT(msgtype
== IBLND_MSG_GET_REQ
||
846 msgtype
== IBLND_MSG_PUT_ACK
);
848 return msgtype
== IBLND_MSG_GET_REQ
?
849 offsetof(kib_get_msg_t
, ibgm_rd
.rd_frags
[n
]) :
850 offsetof(kib_putack_msg_t
, ibpam_rd
.rd_frags
[n
]);
854 kiblnd_dma_mapping_error(struct ib_device
*dev
, u64 dma_addr
)
856 return ib_dma_mapping_error(dev
, dma_addr
);
859 static inline __u64
kiblnd_dma_map_single(struct ib_device
*dev
,
860 void *msg
, size_t size
,
861 enum dma_data_direction direction
)
863 return ib_dma_map_single(dev
, msg
, size
, direction
);
866 static inline void kiblnd_dma_unmap_single(struct ib_device
*dev
,
867 __u64 addr
, size_t size
,
868 enum dma_data_direction direction
)
870 ib_dma_unmap_single(dev
, addr
, size
, direction
);
873 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
874 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
876 static inline int kiblnd_dma_map_sg(struct ib_device
*dev
,
877 struct scatterlist
*sg
, int nents
,
878 enum dma_data_direction direction
)
880 return ib_dma_map_sg(dev
, sg
, nents
, direction
);
883 static inline void kiblnd_dma_unmap_sg(struct ib_device
*dev
,
884 struct scatterlist
*sg
, int nents
,
885 enum dma_data_direction direction
)
887 ib_dma_unmap_sg(dev
, sg
, nents
, direction
);
890 static inline __u64
kiblnd_sg_dma_address(struct ib_device
*dev
,
891 struct scatterlist
*sg
)
893 return ib_sg_dma_address(dev
, sg
);
896 static inline unsigned int kiblnd_sg_dma_len(struct ib_device
*dev
,
897 struct scatterlist
*sg
)
899 return ib_sg_dma_len(dev
, sg
);
902 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly */
903 /* right because OFED1.2 defines it as const, to use it we have to add */
904 /* (void *) cast to overcome "const" */
906 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
907 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
909 struct ib_mr
*kiblnd_find_rd_dma_mr(kib_hca_dev_t
*hdev
,
910 kib_rdma_desc_t
*rd
);
911 struct ib_mr
*kiblnd_find_dma_mr(kib_hca_dev_t
*hdev
,
912 __u64 addr
, __u64 size
);
913 void kiblnd_map_rx_descs(kib_conn_t
*conn
);
914 void kiblnd_unmap_rx_descs(kib_conn_t
*conn
);
915 void kiblnd_pool_free_node(kib_pool_t
*pool
, struct list_head
*node
);
916 struct list_head
*kiblnd_pool_alloc_node(kib_poolset_t
*ps
);
918 int kiblnd_fmr_pool_map(kib_fmr_poolset_t
*fps
, __u64
*pages
,
919 int npages
, __u64 iov
, kib_fmr_t
*fmr
);
920 void kiblnd_fmr_pool_unmap(kib_fmr_t
*fmr
, int status
);
922 int kiblnd_startup(lnet_ni_t
*ni
);
923 void kiblnd_shutdown(lnet_ni_t
*ni
);
924 int kiblnd_ctl(lnet_ni_t
*ni
, unsigned int cmd
, void *arg
);
925 void kiblnd_query(struct lnet_ni
*ni
, lnet_nid_t nid
, unsigned long *when
);
927 int kiblnd_tunables_init(void);
928 void kiblnd_tunables_fini(void);
930 int kiblnd_connd(void *arg
);
931 int kiblnd_scheduler(void *arg
);
932 int kiblnd_thread_start(int (*fn
)(void *arg
), void *arg
, char *name
);
933 int kiblnd_failover_thread(void *arg
);
935 int kiblnd_alloc_pages(kib_pages_t
**pp
, int cpt
, int npages
);
936 void kiblnd_free_pages(kib_pages_t
*p
);
938 int kiblnd_cm_callback(struct rdma_cm_id
*cmid
,
939 struct rdma_cm_event
*event
);
940 int kiblnd_translate_mtu(int value
);
942 int kiblnd_dev_failover(kib_dev_t
*dev
);
943 int kiblnd_create_peer(lnet_ni_t
*ni
, kib_peer_t
**peerp
, lnet_nid_t nid
);
944 void kiblnd_destroy_peer(kib_peer_t
*peer
);
945 void kiblnd_destroy_dev(kib_dev_t
*dev
);
946 void kiblnd_unlink_peer_locked(kib_peer_t
*peer
);
947 void kiblnd_peer_alive(kib_peer_t
*peer
);
948 kib_peer_t
*kiblnd_find_peer_locked(lnet_nid_t nid
);
949 void kiblnd_peer_connect_failed(kib_peer_t
*peer
, int active
, int error
);
950 int kiblnd_close_stale_conns_locked(kib_peer_t
*peer
,
951 int version
, __u64 incarnation
);
952 int kiblnd_close_peer_conns_locked(kib_peer_t
*peer
, int why
);
954 void kiblnd_connreq_done(kib_conn_t
*conn
, int status
);
955 kib_conn_t
*kiblnd_create_conn(kib_peer_t
*peer
, struct rdma_cm_id
*cmid
,
956 int state
, int version
);
957 void kiblnd_destroy_conn(kib_conn_t
*conn
);
958 void kiblnd_close_conn(kib_conn_t
*conn
, int error
);
959 void kiblnd_close_conn_locked(kib_conn_t
*conn
, int error
);
961 int kiblnd_init_rdma(kib_conn_t
*conn
, kib_tx_t
*tx
, int type
,
962 int nob
, kib_rdma_desc_t
*dstrd
, __u64 dstcookie
);
964 void kiblnd_launch_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
, lnet_nid_t nid
);
965 void kiblnd_queue_tx_locked(kib_tx_t
*tx
, kib_conn_t
*conn
);
966 void kiblnd_queue_tx(kib_tx_t
*tx
, kib_conn_t
*conn
);
967 void kiblnd_init_tx_msg(lnet_ni_t
*ni
, kib_tx_t
*tx
, int type
, int body_nob
);
968 void kiblnd_txlist_done(lnet_ni_t
*ni
, struct list_head
*txlist
,
970 void kiblnd_check_sends (kib_conn_t
*conn
);
972 void kiblnd_qp_event(struct ib_event
*event
, void *arg
);
973 void kiblnd_cq_event(struct ib_event
*event
, void *arg
);
974 void kiblnd_cq_completion(struct ib_cq
*cq
, void *arg
);
976 void kiblnd_pack_msg(lnet_ni_t
*ni
, kib_msg_t
*msg
, int version
,
977 int credits
, lnet_nid_t dstnid
, __u64 dststamp
);
978 int kiblnd_unpack_msg(kib_msg_t
*msg
, int nob
);
979 int kiblnd_post_rx(kib_rx_t
*rx
, int credit
);
981 int kiblnd_send(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
);
982 int kiblnd_recv(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
, int delayed
,
983 unsigned int niov
, struct kvec
*iov
, lnet_kiov_t
*kiov
,
984 unsigned int offset
, unsigned int mlen
, unsigned int rlen
);