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
86 int *kib_max_reconnect_interval
; /* ...exponentially increasing
88 int *kib_cksum
; /* checksum kib_msg_t? */
89 int *kib_timeout
; /* comms timeout (seconds) */
90 int *kib_keepalive
; /* keepalive timeout (seconds) */
91 int *kib_ntx
; /* # tx descs */
92 int *kib_credits
; /* # concurrent sends */
93 int *kib_peertxcredits
; /* # concurrent sends to 1 peer */
94 int *kib_peerrtrcredits
; /* # per-peer router buffer
96 int *kib_peercredits_hiw
; /* # when eagerly to return
98 int *kib_peertimeout
; /* seconds to consider peer dead */
99 char **kib_default_ipif
; /* default IPoIB interface */
100 int *kib_retry_count
;
101 int *kib_rnr_retry_count
;
102 int *kib_concurrent_sends
; /* send work queue sizing */
103 int *kib_ib_mtu
; /* IB MTU */
104 int *kib_map_on_demand
; /* map-on-demand if RD has more
105 * fragments than this value, 0
106 * disable map-on-demand */
107 int *kib_pmr_pool_size
; /* # physical MR in pool */
108 int *kib_fmr_pool_size
; /* # FMRs in pool */
109 int *kib_fmr_flush_trigger
; /* When to trigger FMR flush */
110 int *kib_fmr_cache
; /* enable FMR pool cache? */
111 int *kib_require_priv_port
; /* accept only privileged ports */
112 int *kib_use_priv_port
; /* use privileged port for active
114 int *kib_nscheds
; /* # threads on each CPT */
117 extern kib_tunables_t kiblnd_tunables
;
119 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
120 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
122 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
123 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
125 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
126 IBLND_MSG_QUEUE_SIZE_V1 : \
127 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
128 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
129 IBLND_CREDIT_HIGHWATER_V1 : \
130 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
132 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
135 kiblnd_concurrent_sends_v1(void)
137 if (*kiblnd_tunables
.kib_concurrent_sends
> IBLND_MSG_QUEUE_SIZE_V1
* 2)
138 return IBLND_MSG_QUEUE_SIZE_V1
* 2;
140 if (*kiblnd_tunables
.kib_concurrent_sends
< IBLND_MSG_QUEUE_SIZE_V1
/ 2)
141 return IBLND_MSG_QUEUE_SIZE_V1
/ 2;
143 return *kiblnd_tunables
.kib_concurrent_sends
;
146 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
147 kiblnd_concurrent_sends_v1() : \
148 *kiblnd_tunables.kib_concurrent_sends)
149 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
150 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
151 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
153 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
154 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
155 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
156 *kiblnd_tunables.kib_map_on_demand : \
157 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
158 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
159 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
161 /************************/
162 /* derived constants... */
163 /* Pools (shared by connections on each CPT) */
164 /* These pools can grow at runtime, so don't need give a very large value */
165 #define IBLND_TX_POOL 256
166 #define IBLND_PMR_POOL 256
167 #define IBLND_FMR_POOL 256
168 #define IBLND_FMR_POOL_FLUSH 192
170 /* TX messages (shared by all connections) */
171 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
173 /* RX messages (per connection) */
174 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
175 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
176 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
178 /* WRs and CQEs (per connection) */
179 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
180 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
181 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
185 /* o2iblnd can run over aliased interface */
187 #define KIB_IFNAME_SIZE IFALIASZ
189 #define KIB_IFNAME_SIZE 256
193 struct list_head ibd_list
; /* chain on kib_devs */
194 struct list_head ibd_fail_list
; /* chain on kib_failed_devs */
195 __u32 ibd_ifip
; /* IPoIB interface IP */
197 /* IPoIB interface name */
198 char ibd_ifname
[KIB_IFNAME_SIZE
];
199 int ibd_nnets
; /* # nets extant */
201 unsigned long ibd_next_failover
;
202 int ibd_failed_failover
; /* # failover failures */
203 unsigned int ibd_failover
; /* failover in progress */
204 unsigned int ibd_can_failover
; /* IPoIB interface is a bonding
206 struct list_head ibd_nets
;
207 struct kib_hca_dev
*ibd_hdev
;
210 typedef struct kib_hca_dev
{
211 struct rdma_cm_id
*ibh_cmid
; /* listener cmid */
212 struct ib_device
*ibh_ibdev
; /* IB device */
213 int ibh_page_shift
; /* page shift of current HCA */
214 int ibh_page_size
; /* page size of current HCA */
215 __u64 ibh_page_mask
; /* page mask of current HCA */
216 int ibh_mr_shift
; /* bits shift of max MR size */
217 __u64 ibh_mr_size
; /* size of MR */
218 int ibh_nmrs
; /* # of global MRs */
219 struct ib_mr
**ibh_mrs
; /* global MR */
220 struct ib_pd
*ibh_pd
; /* PD */
221 kib_dev_t
*ibh_dev
; /* owner */
222 atomic_t ibh_ref
; /* refcount */
225 /** # of seconds to keep pool alive */
226 #define IBLND_POOL_DEADLINE 300
227 /** # of seconds to retry if allocation failed */
228 #define IBLND_POOL_RETRY 1
231 int ibp_npages
; /* # pages */
232 struct page
*ibp_pages
[0]; /* page array */
238 struct list_head pmr_list
; /* chain node */
239 struct ib_phys_buf
*pmr_ipb
; /* physical buffer */
240 struct ib_mr
*pmr_mr
; /* IB MR */
241 struct kib_pmr_pool
*pmr_pool
; /* owner of this MR */
242 __u64 pmr_iova
; /* Virtual I/O address */
243 int pmr_refcount
; /* reference count */
249 typedef int (*kib_ps_pool_create_t
)(struct kib_poolset
*ps
,
250 int inc
, struct kib_pool
**pp_po
);
251 typedef void (*kib_ps_pool_destroy_t
)(struct kib_pool
*po
);
252 typedef void (*kib_ps_node_init_t
)(struct kib_pool
*po
, struct list_head
*node
);
253 typedef void (*kib_ps_node_fini_t
)(struct kib_pool
*po
, struct list_head
*node
);
257 #define IBLND_POOL_NAME_LEN 32
259 typedef struct kib_poolset
{
260 spinlock_t ps_lock
; /* serialize */
261 struct kib_net
*ps_net
; /* network it belongs to */
262 char ps_name
[IBLND_POOL_NAME_LEN
]; /* pool set name */
263 struct list_head ps_pool_list
; /* list of pools */
264 struct list_head ps_failed_pool_list
;/* failed pool list */
265 unsigned long ps_next_retry
; /* time stamp for retry if
266 * failed to allocate */
267 int ps_increasing
; /* is allocating new pool */
268 int ps_pool_size
; /* new pool size */
269 int ps_cpt
; /* CPT id */
271 kib_ps_pool_create_t ps_pool_create
; /* create a new pool */
272 kib_ps_pool_destroy_t ps_pool_destroy
; /* destroy a pool */
273 kib_ps_node_init_t ps_node_init
; /* initialize new allocated
275 kib_ps_node_fini_t ps_node_fini
; /* finalize node */
278 typedef struct kib_pool
{
279 struct list_head po_list
; /* chain on pool list */
280 struct list_head po_free_list
; /* pre-allocated node */
281 kib_poolset_t
*po_owner
; /* pool_set of this pool */
282 unsigned long po_deadline
; /* deadline of this pool */
283 int po_allocated
; /* # of elements in use */
284 int po_failed
; /* pool is created on failed
286 int po_size
; /* # of pre-allocated elements */
290 kib_poolset_t tps_poolset
; /* pool-set */
291 __u64 tps_next_tx_cookie
; /* cookie of TX */
295 kib_pool_t tpo_pool
; /* pool */
296 struct kib_hca_dev
*tpo_hdev
; /* device for this pool */
297 struct kib_tx
*tpo_tx_descs
; /* all the tx descriptors */
298 kib_pages_t
*tpo_tx_pages
; /* premapped tx msg pages */
302 kib_poolset_t pps_poolset
; /* pool-set */
305 typedef struct kib_pmr_pool
{
306 struct kib_hca_dev
*ppo_hdev
; /* device for this pool */
307 kib_pool_t ppo_pool
; /* pool */
311 spinlock_t fps_lock
; /* serialize */
312 struct kib_net
*fps_net
; /* IB network */
313 struct list_head fps_pool_list
; /* FMR pool list */
314 struct list_head fps_failed_pool_list
;/* FMR pool list */
315 __u64 fps_version
; /* validity stamp */
316 int fps_cpt
; /* CPT id */
318 int fps_flush_trigger
;
319 int fps_increasing
; /* is allocating new pool */
320 unsigned long fps_next_retry
; /* time stamp for retry if
321 * failed to allocate */
325 struct list_head fpo_list
; /* chain on pool list */
326 struct kib_hca_dev
*fpo_hdev
; /* device for this pool */
327 kib_fmr_poolset_t
*fpo_owner
; /* owner of this pool */
328 struct ib_fmr_pool
*fpo_fmr_pool
; /* IB FMR pool */
329 unsigned long fpo_deadline
; /* deadline of this pool */
330 int fpo_failed
; /* fmr pool is failed */
331 int fpo_map_count
; /* # of mapped FMR */
335 struct ib_pool_fmr
*fmr_pfmr
; /* IB pool fmr */
336 kib_fmr_pool_t
*fmr_pool
; /* pool of FMR */
339 typedef struct kib_net
{
340 struct list_head ibn_list
; /* chain on kib_dev_t::ibd_nets */
341 __u64 ibn_incarnation
;/* my epoch */
342 int ibn_init
; /* initialisation state */
343 int ibn_shutdown
; /* shutting down? */
345 atomic_t ibn_npeers
; /* # peers extant */
346 atomic_t ibn_nconns
; /* # connections extant */
348 kib_tx_poolset_t
**ibn_tx_ps
; /* tx pool-set */
349 kib_fmr_poolset_t
**ibn_fmr_ps
; /* fmr pool-set */
350 kib_pmr_poolset_t
**ibn_pmr_ps
; /* pmr pool-set */
352 kib_dev_t
*ibn_dev
; /* underlying IB device */
355 #define KIB_THREAD_SHIFT 16
356 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
357 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
358 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
360 struct kib_sched_info
{
361 spinlock_t ibs_lock
; /* serialise */
362 wait_queue_head_t ibs_waitq
; /* schedulers sleep here */
363 struct list_head ibs_conns
; /* conns to check for rx completions */
364 int ibs_nthreads
; /* number of scheduler threads */
365 int ibs_nthreads_max
; /* max allowed scheduler threads */
366 int ibs_cpt
; /* CPT id */
370 int kib_init
; /* initialisation state */
371 int kib_shutdown
; /* shut down? */
372 struct list_head kib_devs
; /* IB devices extant */
373 struct list_head kib_failed_devs
; /* list head of failed
375 wait_queue_head_t kib_failover_waitq
; /* schedulers sleep here */
376 atomic_t kib_nthreads
; /* # live threads */
377 rwlock_t kib_global_lock
; /* stabilize net/dev/peer/conn
379 struct list_head
*kib_peers
; /* hash table of all my known
381 int kib_peer_hash_size
; /* size of kib_peers */
382 void *kib_connd
; /* the connd task
383 * (serialisation assertions)
385 struct list_head kib_connd_conns
; /* connections to
387 struct list_head kib_connd_zombies
; /* connections with zero
389 wait_queue_head_t kib_connd_waitq
; /* connection daemon sleeps
391 spinlock_t kib_connd_lock
; /* serialise */
392 struct ib_qp_attr kib_error_qpa
; /* QP->ERROR */
393 struct kib_sched_info
**kib_scheds
; /* percpt data for schedulers
397 #define IBLND_INIT_NOTHING 0
398 #define IBLND_INIT_DATA 1
399 #define IBLND_INIT_ALL 2
401 /************************************************************************
402 * IB Wire message format.
403 * These are sent in sender's byte order (i.e. receiver flips).
406 typedef struct kib_connparams
{
407 __u16 ibcp_queue_depth
;
408 __u16 ibcp_max_frags
;
409 __u32 ibcp_max_msg_size
;
410 } WIRE_ATTR kib_connparams_t
;
413 lnet_hdr_t ibim_hdr
; /* portals header */
414 char ibim_payload
[0]; /* piggy-backed payload */
415 } WIRE_ATTR kib_immediate_msg_t
;
418 __u32 rf_nob
; /* # bytes this frag */
419 __u64 rf_addr
; /* CAVEAT EMPTOR: misaligned!! */
420 } WIRE_ATTR kib_rdma_frag_t
;
423 __u32 rd_key
; /* local/remote key */
424 __u32 rd_nfrags
; /* # fragments */
425 kib_rdma_frag_t rd_frags
[0]; /* buffer frags */
426 } WIRE_ATTR kib_rdma_desc_t
;
429 lnet_hdr_t ibprm_hdr
; /* portals header */
430 __u64 ibprm_cookie
; /* opaque completion cookie */
431 } WIRE_ATTR kib_putreq_msg_t
;
434 __u64 ibpam_src_cookie
; /* reflected completion cookie */
435 __u64 ibpam_dst_cookie
; /* opaque completion cookie */
436 kib_rdma_desc_t ibpam_rd
; /* sender's sink buffer */
437 } WIRE_ATTR kib_putack_msg_t
;
440 lnet_hdr_t ibgm_hdr
; /* portals header */
441 __u64 ibgm_cookie
; /* opaque completion cookie */
442 kib_rdma_desc_t ibgm_rd
; /* rdma descriptor */
443 } WIRE_ATTR kib_get_msg_t
;
446 __u64 ibcm_cookie
; /* opaque completion cookie */
447 __s32 ibcm_status
; /* < 0 failure: >= 0 length */
448 } WIRE_ATTR kib_completion_msg_t
;
451 /* First 2 fields fixed FOR ALL TIME */
452 __u32 ibm_magic
; /* I'm an ibnal message */
453 __u16 ibm_version
; /* this is my version number */
455 __u8 ibm_type
; /* msg type */
456 __u8 ibm_credits
; /* returned credits */
457 __u32 ibm_nob
; /* # bytes in whole message */
458 __u32 ibm_cksum
; /* checksum (0 == no checksum) */
459 __u64 ibm_srcnid
; /* sender's NID */
460 __u64 ibm_srcstamp
; /* sender's incarnation */
461 __u64 ibm_dstnid
; /* destination's NID */
462 __u64 ibm_dststamp
; /* destination's incarnation */
465 kib_connparams_t connparams
;
466 kib_immediate_msg_t immediate
;
467 kib_putreq_msg_t putreq
;
468 kib_putack_msg_t putack
;
470 kib_completion_msg_t completion
;
472 } WIRE_ATTR kib_msg_t
;
474 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
476 #define IBLND_MSG_VERSION_1 0x11
477 #define IBLND_MSG_VERSION_2 0x12
478 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
480 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
481 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
482 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
483 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
484 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
485 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
486 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
487 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
488 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
489 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
492 __u32 ibr_magic
; /* sender's magic */
493 __u16 ibr_version
; /* sender's version */
494 __u8 ibr_why
; /* reject reason */
495 __u8 ibr_padding
; /* padding */
496 __u64 ibr_incarnation
; /* incarnation of peer */
497 kib_connparams_t ibr_cp
; /* connection parameters */
498 } WIRE_ATTR kib_rej_t
;
500 /* connection rejection reasons */
501 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
502 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
503 #define IBLND_REJECT_FATAL 3 /* Anything else */
504 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
505 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
506 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match
508 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't
511 /***********************************************************************/
513 typedef struct kib_rx
/* receive message */
515 struct list_head rx_list
; /* queue for attention */
516 struct kib_conn
*rx_conn
; /* owning conn */
517 int rx_nob
; /* # bytes received (-1 while
519 enum ib_wc_status rx_status
; /* completion status */
520 kib_msg_t
*rx_msg
; /* message buffer (host vaddr) */
521 __u64 rx_msgaddr
; /* message buffer (I/O addr) */
522 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap
); /* for dma_unmap_single() */
523 struct ib_recv_wr rx_wrq
; /* receive work item... */
524 struct ib_sge rx_sge
; /* ...and its memory */
527 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
528 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
529 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
530 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved
533 typedef struct kib_tx
/* transmit message */
535 struct list_head tx_list
; /* queue on idle_txs ibc_tx_queue
537 kib_tx_pool_t
*tx_pool
; /* pool I'm from */
538 struct kib_conn
*tx_conn
; /* owning conn */
539 short tx_sending
; /* # tx callbacks outstanding */
540 short tx_queued
; /* queued for sending */
541 short tx_waiting
; /* waiting for peer */
542 int tx_status
; /* LNET completion status */
543 unsigned long tx_deadline
; /* completion deadline */
544 __u64 tx_cookie
; /* completion cookie */
545 lnet_msg_t
*tx_lntmsg
[2]; /* lnet msgs to finalize on
547 kib_msg_t
*tx_msg
; /* message buffer (host vaddr) */
548 __u64 tx_msgaddr
; /* message buffer (I/O addr) */
549 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap
); /* for dma_unmap_single() */
550 int tx_nwrq
; /* # send work items */
551 struct ib_send_wr
*tx_wrq
; /* send work items... */
552 struct ib_sge
*tx_sge
; /* ...and their memory */
553 kib_rdma_desc_t
*tx_rd
; /* rdma descriptor */
554 int tx_nfrags
; /* # entries in... */
555 struct scatterlist
*tx_frags
; /* dma_map_sg descriptor */
556 __u64
*tx_pages
; /* rdma phys page addrs */
558 kib_phys_mr_t
*pmr
; /* MR for physical buffer */
559 kib_fmr_t fmr
; /* FMR */
561 int tx_dmadir
; /* dma direction */
564 typedef struct kib_connvars
{
565 kib_msg_t cv_msg
; /* connection-in-progress variables */
568 typedef struct kib_conn
{
569 struct kib_sched_info
*ibc_sched
; /* scheduler information */
570 struct kib_peer
*ibc_peer
; /* owning peer */
571 kib_hca_dev_t
*ibc_hdev
; /* HCA bound on */
572 struct list_head ibc_list
; /* stash on peer's conn
574 struct list_head ibc_sched_list
; /* schedule for attention */
575 __u16 ibc_version
; /* version of connection */
576 __u64 ibc_incarnation
; /* which instance of the
578 atomic_t ibc_refcount
; /* # users */
579 int ibc_state
; /* what's happening */
580 int ibc_nsends_posted
; /* # uncompleted sends */
581 int ibc_noops_posted
; /* # uncompleted NOOPs */
582 int ibc_credits
; /* # credits I have */
583 int ibc_outstanding_credits
; /* # credits to return */
584 int ibc_reserved_credits
; /* # ACK/DONE msg credits */
585 int ibc_comms_error
; /* set on comms error */
586 unsigned int ibc_nrx
:16; /* receive buffers owned */
587 unsigned int ibc_scheduled
:1; /* scheduled for attention
589 unsigned int ibc_ready
:1; /* CQ callback fired */
590 unsigned long ibc_last_send
; /* time of last send */
591 struct list_head ibc_connd_list
; /* link chain for
592 * kiblnd_check_conns only
594 struct list_head ibc_early_rxs
; /* rxs completed before
596 struct list_head ibc_tx_noops
; /* IBLND_MSG_NOOPs for
597 * IBLND_MSG_VERSION_1 */
598 struct list_head ibc_tx_queue
; /* sends that need a credit
600 struct list_head ibc_tx_queue_nocred
; /* sends that don't need a
602 struct list_head ibc_tx_queue_rsrvd
; /* sends that need to
603 * reserve an ACK/DONE msg
605 struct list_head ibc_active_txs
; /* active tx awaiting
607 spinlock_t ibc_lock
; /* serialise */
608 kib_rx_t
*ibc_rxs
; /* the rx descs */
609 kib_pages_t
*ibc_rx_pages
; /* premapped rx msg pages */
611 struct rdma_cm_id
*ibc_cmid
; /* CM id */
612 struct ib_cq
*ibc_cq
; /* completion queue */
614 kib_connvars_t
*ibc_connvars
; /* in-progress connection
618 #define IBLND_CONN_INIT 0 /* being initialised */
619 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
620 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
621 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
622 #define IBLND_CONN_CLOSING 4 /* being closed */
623 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
625 typedef struct kib_peer
{
626 struct list_head ibp_list
; /* stash on global peer list */
627 lnet_nid_t ibp_nid
; /* who's on the other end(s) */
628 lnet_ni_t
*ibp_ni
; /* LNet interface */
629 atomic_t ibp_refcount
; /* # users */
630 struct list_head ibp_conns
; /* all active connections */
631 struct list_head ibp_tx_queue
; /* msgs waiting for a conn */
632 __u16 ibp_version
; /* version of peer */
633 __u64 ibp_incarnation
; /* incarnation of peer */
634 int ibp_connecting
; /* current active connection attempts
636 int ibp_accepting
; /* current passive connection attempts
638 int ibp_error
; /* errno on closing this peer */
639 unsigned long ibp_last_alive
; /* when (in jiffies) I was last alive
643 extern kib_data_t kiblnd_data
;
645 extern void kiblnd_hdev_destroy(kib_hca_dev_t
*hdev
);
648 kiblnd_hdev_addref_locked(kib_hca_dev_t
*hdev
)
650 LASSERT (atomic_read(&hdev
->ibh_ref
) > 0);
651 atomic_inc(&hdev
->ibh_ref
);
655 kiblnd_hdev_decref(kib_hca_dev_t
*hdev
)
657 LASSERT (atomic_read(&hdev
->ibh_ref
) > 0);
658 if (atomic_dec_and_test(&hdev
->ibh_ref
))
659 kiblnd_hdev_destroy(hdev
);
663 kiblnd_dev_can_failover(kib_dev_t
*dev
)
665 if (!list_empty(&dev
->ibd_fail_list
)) /* already scheduled */
668 if (*kiblnd_tunables
.kib_dev_failover
== 0) /* disabled */
671 if (*kiblnd_tunables
.kib_dev_failover
> 1) /* force failover */
674 return dev
->ibd_can_failover
;
677 #define kiblnd_conn_addref(conn) \
679 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
680 (conn), atomic_read(&(conn)->ibc_refcount)); \
681 atomic_inc(&(conn)->ibc_refcount); \
684 #define kiblnd_conn_decref(conn) \
686 unsigned long flags; \
688 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
689 (conn), atomic_read(&(conn)->ibc_refcount)); \
690 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
691 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
692 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
693 list_add_tail(&(conn)->ibc_list, \
694 &kiblnd_data.kib_connd_zombies); \
695 wake_up(&kiblnd_data.kib_connd_waitq); \
696 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
700 #define kiblnd_peer_addref(peer) \
702 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
703 (peer), libcfs_nid2str((peer)->ibp_nid), \
704 atomic_read (&(peer)->ibp_refcount)); \
705 atomic_inc(&(peer)->ibp_refcount); \
708 #define kiblnd_peer_decref(peer) \
710 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
711 (peer), libcfs_nid2str((peer)->ibp_nid), \
712 atomic_read (&(peer)->ibp_refcount)); \
713 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
714 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
715 kiblnd_destroy_peer(peer); \
718 static inline struct list_head
*
719 kiblnd_nid2peerlist (lnet_nid_t nid
)
722 ((unsigned int)nid
) % kiblnd_data
.kib_peer_hash_size
;
724 return (&kiblnd_data
.kib_peers
[hash
]);
728 kiblnd_peer_active (kib_peer_t
*peer
)
730 /* Am I in the peer hash table? */
731 return (!list_empty(&peer
->ibp_list
));
734 static inline kib_conn_t
*
735 kiblnd_get_conn_locked (kib_peer_t
*peer
)
737 LASSERT (!list_empty(&peer
->ibp_conns
));
739 /* just return the first connection */
740 return list_entry(peer
->ibp_conns
.next
, kib_conn_t
, ibc_list
);
744 kiblnd_send_keepalive(kib_conn_t
*conn
)
746 return (*kiblnd_tunables
.kib_keepalive
> 0) &&
747 cfs_time_after(jiffies
, conn
->ibc_last_send
+
748 *kiblnd_tunables
.kib_keepalive
*HZ
);
752 kiblnd_need_noop(kib_conn_t
*conn
)
754 LASSERT (conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
756 if (conn
->ibc_outstanding_credits
<
757 IBLND_CREDITS_HIGHWATER(conn
->ibc_version
) &&
758 !kiblnd_send_keepalive(conn
))
759 return 0; /* No need to send NOOP */
761 if (IBLND_OOB_CAPABLE(conn
->ibc_version
)) {
762 if (!list_empty(&conn
->ibc_tx_queue_nocred
))
763 return 0; /* NOOP can be piggybacked */
765 /* No tx to piggyback NOOP onto or no credit to send a tx */
766 return (list_empty(&conn
->ibc_tx_queue
) ||
767 conn
->ibc_credits
== 0);
770 if (!list_empty(&conn
->ibc_tx_noops
) || /* NOOP already queued */
771 !list_empty(&conn
->ibc_tx_queue_nocred
) || /* piggyback NOOP */
772 conn
->ibc_credits
== 0) /* no credit */
775 if (conn
->ibc_credits
== 1 && /* last credit reserved for */
776 conn
->ibc_outstanding_credits
== 0) /* giving back credits */
779 /* No tx to piggyback NOOP onto or no credit to send a tx */
780 return (list_empty(&conn
->ibc_tx_queue
) || conn
->ibc_credits
== 1);
784 kiblnd_abort_receives(kib_conn_t
*conn
)
786 ib_modify_qp(conn
->ibc_cmid
->qp
,
787 &kiblnd_data
.kib_error_qpa
, IB_QP_STATE
);
790 static inline const char *
791 kiblnd_queue2str (kib_conn_t
*conn
, struct list_head
*q
)
793 if (q
== &conn
->ibc_tx_queue
)
796 if (q
== &conn
->ibc_tx_queue_rsrvd
)
797 return "tx_queue_rsrvd";
799 if (q
== &conn
->ibc_tx_queue_nocred
)
800 return "tx_queue_nocred";
802 if (q
== &conn
->ibc_active_txs
)
809 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
810 * lowest bits of the work request id to stash the work item type. */
812 #define IBLND_WID_TX 0
813 #define IBLND_WID_RDMA 1
814 #define IBLND_WID_RX 2
815 #define IBLND_WID_MASK 3UL
818 kiblnd_ptr2wreqid (void *ptr
, int type
)
820 unsigned long lptr
= (unsigned long)ptr
;
822 LASSERT ((lptr
& IBLND_WID_MASK
) == 0);
823 LASSERT ((type
& ~IBLND_WID_MASK
) == 0);
824 return (__u64
)(lptr
| type
);
828 kiblnd_wreqid2ptr (__u64 wreqid
)
830 return (void *)(((unsigned long)wreqid
) & ~IBLND_WID_MASK
);
834 kiblnd_wreqid2type (__u64 wreqid
)
836 return (wreqid
& IBLND_WID_MASK
);
840 kiblnd_set_conn_state (kib_conn_t
*conn
, int state
)
842 conn
->ibc_state
= state
;
847 kiblnd_init_msg (kib_msg_t
*msg
, int type
, int body_nob
)
849 msg
->ibm_type
= type
;
850 msg
->ibm_nob
= offsetof(kib_msg_t
, ibm_u
) + body_nob
;
854 kiblnd_rd_size (kib_rdma_desc_t
*rd
)
859 for (i
= size
= 0; i
< rd
->rd_nfrags
; i
++)
860 size
+= rd
->rd_frags
[i
].rf_nob
;
866 kiblnd_rd_frag_addr(kib_rdma_desc_t
*rd
, int index
)
868 return rd
->rd_frags
[index
].rf_addr
;
872 kiblnd_rd_frag_size(kib_rdma_desc_t
*rd
, int index
)
874 return rd
->rd_frags
[index
].rf_nob
;
878 kiblnd_rd_frag_key(kib_rdma_desc_t
*rd
, int index
)
884 kiblnd_rd_consume_frag(kib_rdma_desc_t
*rd
, int index
, __u32 nob
)
886 if (nob
< rd
->rd_frags
[index
].rf_nob
) {
887 rd
->rd_frags
[index
].rf_addr
+= nob
;
888 rd
->rd_frags
[index
].rf_nob
-= nob
;
897 kiblnd_rd_msg_size(kib_rdma_desc_t
*rd
, int msgtype
, int n
)
899 LASSERT (msgtype
== IBLND_MSG_GET_REQ
||
900 msgtype
== IBLND_MSG_PUT_ACK
);
902 return msgtype
== IBLND_MSG_GET_REQ
?
903 offsetof(kib_get_msg_t
, ibgm_rd
.rd_frags
[n
]) :
904 offsetof(kib_putack_msg_t
, ibpam_rd
.rd_frags
[n
]);
909 kiblnd_dma_mapping_error(struct ib_device
*dev
, u64 dma_addr
)
911 return ib_dma_mapping_error(dev
, dma_addr
);
914 static inline __u64
kiblnd_dma_map_single(struct ib_device
*dev
,
915 void *msg
, size_t size
,
916 enum dma_data_direction direction
)
918 return ib_dma_map_single(dev
, msg
, size
, direction
);
921 static inline void kiblnd_dma_unmap_single(struct ib_device
*dev
,
922 __u64 addr
, size_t size
,
923 enum dma_data_direction direction
)
925 ib_dma_unmap_single(dev
, addr
, size
, direction
);
928 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
929 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
931 static inline int kiblnd_dma_map_sg(struct ib_device
*dev
,
932 struct scatterlist
*sg
, int nents
,
933 enum dma_data_direction direction
)
935 return ib_dma_map_sg(dev
, sg
, nents
, direction
);
938 static inline void kiblnd_dma_unmap_sg(struct ib_device
*dev
,
939 struct scatterlist
*sg
, int nents
,
940 enum dma_data_direction direction
)
942 ib_dma_unmap_sg(dev
, sg
, nents
, direction
);
945 static inline __u64
kiblnd_sg_dma_address(struct ib_device
*dev
,
946 struct scatterlist
*sg
)
948 return ib_sg_dma_address(dev
, sg
);
951 static inline unsigned int kiblnd_sg_dma_len(struct ib_device
*dev
,
952 struct scatterlist
*sg
)
954 return ib_sg_dma_len(dev
, sg
);
957 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
958 * right because OFED1.2 defines it as const, to use it we have to add
959 * (void *) cast to overcome "const" */
961 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
962 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
965 struct ib_mr
*kiblnd_find_rd_dma_mr(kib_hca_dev_t
*hdev
,
966 kib_rdma_desc_t
*rd
);
967 struct ib_mr
*kiblnd_find_dma_mr(kib_hca_dev_t
*hdev
,
968 __u64 addr
, __u64 size
);
969 void kiblnd_map_rx_descs(kib_conn_t
*conn
);
970 void kiblnd_unmap_rx_descs(kib_conn_t
*conn
);
971 int kiblnd_map_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
,
972 kib_rdma_desc_t
*rd
, int nfrags
);
973 void kiblnd_unmap_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
);
974 void kiblnd_pool_free_node(kib_pool_t
*pool
, struct list_head
*node
);
975 struct list_head
*kiblnd_pool_alloc_node(kib_poolset_t
*ps
);
977 int kiblnd_fmr_pool_map(kib_fmr_poolset_t
*fps
, __u64
*pages
,
978 int npages
, __u64 iov
, kib_fmr_t
*fmr
);
979 void kiblnd_fmr_pool_unmap(kib_fmr_t
*fmr
, int status
);
981 int kiblnd_pmr_pool_map(kib_pmr_poolset_t
*pps
, kib_hca_dev_t
*hdev
,
982 kib_rdma_desc_t
*rd
, __u64
*iova
, kib_phys_mr_t
**pp_pmr
);
983 void kiblnd_pmr_pool_unmap(kib_phys_mr_t
*pmr
);
985 int kiblnd_startup (lnet_ni_t
*ni
);
986 void kiblnd_shutdown (lnet_ni_t
*ni
);
987 int kiblnd_ctl (lnet_ni_t
*ni
, unsigned int cmd
, void *arg
);
988 void kiblnd_query (struct lnet_ni
*ni
, lnet_nid_t nid
, unsigned long *when
);
990 int kiblnd_tunables_init(void);
991 void kiblnd_tunables_fini(void);
993 int kiblnd_connd (void *arg
);
994 int kiblnd_scheduler(void *arg
);
995 int kiblnd_thread_start(int (*fn
)(void *arg
), void *arg
, char *name
);
996 int kiblnd_failover_thread (void *arg
);
998 int kiblnd_alloc_pages(kib_pages_t
**pp
, int cpt
, int npages
);
999 void kiblnd_free_pages (kib_pages_t
*p
);
1001 int kiblnd_cm_callback(struct rdma_cm_id
*cmid
,
1002 struct rdma_cm_event
*event
);
1003 int kiblnd_translate_mtu(int value
);
1005 int kiblnd_dev_failover(kib_dev_t
*dev
);
1006 int kiblnd_create_peer (lnet_ni_t
*ni
, kib_peer_t
**peerp
, lnet_nid_t nid
);
1007 void kiblnd_destroy_peer (kib_peer_t
*peer
);
1008 void kiblnd_destroy_dev (kib_dev_t
*dev
);
1009 void kiblnd_unlink_peer_locked (kib_peer_t
*peer
);
1010 void kiblnd_peer_alive (kib_peer_t
*peer
);
1011 kib_peer_t
*kiblnd_find_peer_locked (lnet_nid_t nid
);
1012 void kiblnd_peer_connect_failed (kib_peer_t
*peer
, int active
, int error
);
1013 int kiblnd_close_stale_conns_locked (kib_peer_t
*peer
,
1014 int version
, __u64 incarnation
);
1015 int kiblnd_close_peer_conns_locked (kib_peer_t
*peer
, int why
);
1017 void kiblnd_connreq_done(kib_conn_t
*conn
, int status
);
1018 kib_conn_t
*kiblnd_create_conn (kib_peer_t
*peer
, struct rdma_cm_id
*cmid
,
1019 int state
, int version
);
1020 void kiblnd_destroy_conn (kib_conn_t
*conn
);
1021 void kiblnd_close_conn (kib_conn_t
*conn
, int error
);
1022 void kiblnd_close_conn_locked (kib_conn_t
*conn
, int error
);
1024 int kiblnd_init_rdma (kib_conn_t
*conn
, kib_tx_t
*tx
, int type
,
1025 int nob
, kib_rdma_desc_t
*dstrd
, __u64 dstcookie
);
1027 void kiblnd_launch_tx (lnet_ni_t
*ni
, kib_tx_t
*tx
, lnet_nid_t nid
);
1028 void kiblnd_queue_tx_locked (kib_tx_t
*tx
, kib_conn_t
*conn
);
1029 void kiblnd_queue_tx (kib_tx_t
*tx
, kib_conn_t
*conn
);
1030 void kiblnd_init_tx_msg (lnet_ni_t
*ni
, kib_tx_t
*tx
, int type
, int body_nob
);
1031 void kiblnd_txlist_done (lnet_ni_t
*ni
, struct list_head
*txlist
,
1033 void kiblnd_check_sends (kib_conn_t
*conn
);
1035 void kiblnd_qp_event(struct ib_event
*event
, void *arg
);
1036 void kiblnd_cq_event(struct ib_event
*event
, void *arg
);
1037 void kiblnd_cq_completion(struct ib_cq
*cq
, void *arg
);
1039 void kiblnd_pack_msg (lnet_ni_t
*ni
, kib_msg_t
*msg
, int version
,
1040 int credits
, lnet_nid_t dstnid
, __u64 dststamp
);
1041 int kiblnd_unpack_msg(kib_msg_t
*msg
, int nob
);
1042 int kiblnd_post_rx (kib_rx_t
*rx
, int credit
);
1044 int kiblnd_send(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
);
1045 int kiblnd_recv(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
, int delayed
,
1046 unsigned int niov
, struct kvec
*iov
, lnet_kiov_t
*kiov
,
1047 unsigned int offset
, unsigned int mlen
, unsigned int rlen
);