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) 2002, 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.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
41 #include "../include/obd_support.h"
42 #include "../include/obd_class.h"
43 #include "../include/lustre_lib.h"
44 #include "../include/lustre_ha.h"
45 #include "../include/lustre_import.h"
46 #include "../include/lustre_req_layout.h"
48 #include "ptlrpc_internal.h"
50 static int ptlrpc_send_new_req(struct ptlrpc_request
*req
);
51 static int ptlrpcd_check_work(struct ptlrpc_request
*req
);
54 * Initialize passed in client structure \a cl.
56 void ptlrpc_init_client(int req_portal
, int rep_portal
, char *name
,
57 struct ptlrpc_client
*cl
)
59 cl
->cli_request_portal
= req_portal
;
60 cl
->cli_reply_portal
= rep_portal
;
63 EXPORT_SYMBOL(ptlrpc_init_client
);
66 * Return PortalRPC connection for remote uud \a uuid
68 struct ptlrpc_connection
*ptlrpc_uuid_to_connection(struct obd_uuid
*uuid
)
70 struct ptlrpc_connection
*c
;
72 lnet_process_id_t peer
;
75 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
76 * before accessing its values. */
77 /* coverity[uninit_use_in_call] */
78 err
= ptlrpc_uuid_to_peer(uuid
, &peer
, &self
);
80 CNETERR("cannot find peer %s!\n", uuid
->uuid
);
84 c
= ptlrpc_connection_get(peer
, self
, uuid
);
86 memcpy(c
->c_remote_uuid
.uuid
,
87 uuid
->uuid
, sizeof(c
->c_remote_uuid
.uuid
));
90 CDEBUG(D_INFO
, "%s -> %p\n", uuid
->uuid
, c
);
94 EXPORT_SYMBOL(ptlrpc_uuid_to_connection
);
97 * Allocate and initialize new bulk descriptor on the sender.
98 * Returns pointer to the descriptor or NULL on error.
100 struct ptlrpc_bulk_desc
*ptlrpc_new_bulk(unsigned npages
, unsigned max_brw
,
101 unsigned type
, unsigned portal
)
103 struct ptlrpc_bulk_desc
*desc
;
106 OBD_ALLOC(desc
, offsetof(struct ptlrpc_bulk_desc
, bd_iov
[npages
]));
110 spin_lock_init(&desc
->bd_lock
);
111 init_waitqueue_head(&desc
->bd_waitq
);
112 desc
->bd_max_iov
= npages
;
113 desc
->bd_iov_count
= 0;
114 desc
->bd_portal
= portal
;
115 desc
->bd_type
= type
;
116 desc
->bd_md_count
= 0;
117 LASSERT(max_brw
> 0);
118 desc
->bd_md_max_brw
= min(max_brw
, PTLRPC_BULK_OPS_COUNT
);
119 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
120 * node. Negotiated ocd_brw_size will always be <= this number. */
121 for (i
= 0; i
< PTLRPC_BULK_OPS_COUNT
; i
++)
122 LNetInvalidateHandle(&desc
->bd_mds
[i
]);
128 * Prepare bulk descriptor for specified outgoing request \a req that
129 * can fit \a npages * pages. \a type is bulk type. \a portal is where
130 * the bulk to be sent. Used on client-side.
131 * Returns pointer to newly allocated initialized bulk descriptor or NULL on
134 struct ptlrpc_bulk_desc
*ptlrpc_prep_bulk_imp(struct ptlrpc_request
*req
,
135 unsigned npages
, unsigned max_brw
,
136 unsigned type
, unsigned portal
)
138 struct obd_import
*imp
= req
->rq_import
;
139 struct ptlrpc_bulk_desc
*desc
;
141 LASSERT(type
== BULK_PUT_SINK
|| type
== BULK_GET_SOURCE
);
142 desc
= ptlrpc_new_bulk(npages
, max_brw
, type
, portal
);
146 desc
->bd_import_generation
= req
->rq_import_generation
;
147 desc
->bd_import
= class_import_get(imp
);
150 desc
->bd_cbid
.cbid_fn
= client_bulk_callback
;
151 desc
->bd_cbid
.cbid_arg
= desc
;
153 /* This makes req own desc, and free it when she frees herself */
158 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp
);
161 * Add a page \a page to the bulk descriptor \a desc.
162 * Data to transfer in the page starts at offset \a pageoffset and
163 * amount of data to transfer from the page is \a len
165 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc
*desc
,
166 struct page
*page
, int pageoffset
, int len
, int pin
)
168 LASSERT(desc
->bd_iov_count
< desc
->bd_max_iov
);
169 LASSERT(page
!= NULL
);
170 LASSERT(pageoffset
>= 0);
172 LASSERT(pageoffset
+ len
<= PAGE_CACHE_SIZE
);
177 page_cache_get(page
);
179 ptlrpc_add_bulk_page(desc
, page
, pageoffset
, len
);
181 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page
);
184 * Uninitialize and free bulk descriptor \a desc.
185 * Works on bulk descriptors both from server and client side.
187 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc
*desc
, int unpin
)
191 LASSERT(desc
!= NULL
);
192 LASSERT(desc
->bd_iov_count
!= LI_POISON
); /* not freed already */
193 LASSERT(desc
->bd_md_count
== 0); /* network hands off */
194 LASSERT((desc
->bd_export
!= NULL
) ^ (desc
->bd_import
!= NULL
));
196 sptlrpc_enc_pool_put_pages(desc
);
199 class_export_put(desc
->bd_export
);
201 class_import_put(desc
->bd_import
);
204 for (i
= 0; i
< desc
->bd_iov_count
; i
++)
205 page_cache_release(desc
->bd_iov
[i
].kiov_page
);
208 OBD_FREE(desc
, offsetof(struct ptlrpc_bulk_desc
,
209 bd_iov
[desc
->bd_max_iov
]));
211 EXPORT_SYMBOL(__ptlrpc_free_bulk
);
214 * Set server timelimit for this req, i.e. how long are we willing to wait
215 * for reply before timing out this request.
217 void ptlrpc_at_set_req_timeout(struct ptlrpc_request
*req
)
223 LASSERT(req
->rq_import
);
226 /* non-AT settings */
228 * \a imp_server_timeout means this is reverse import and
229 * we send (currently only) ASTs to the client and cannot afford
230 * to wait too long for the reply, otherwise the other client
231 * (because of which we are sending this request) would
232 * timeout waiting for us
234 req
->rq_timeout
= req
->rq_import
->imp_server_timeout
?
235 obd_timeout
/ 2 : obd_timeout
;
237 at
= &req
->rq_import
->imp_at
;
238 idx
= import_at_get_index(req
->rq_import
,
239 req
->rq_request_portal
);
240 serv_est
= at_get(&at
->iat_service_estimate
[idx
]);
241 req
->rq_timeout
= at_est2timeout(serv_est
);
243 /* We could get even fancier here, using history to predict increased
246 /* Let the server know what this RPC timeout is by putting it in the
248 lustre_msg_set_timeout(req
->rq_reqmsg
, req
->rq_timeout
);
250 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout
);
252 /* Adjust max service estimate based on server value */
253 static void ptlrpc_at_adj_service(struct ptlrpc_request
*req
,
254 unsigned int serv_est
)
260 LASSERT(req
->rq_import
);
261 at
= &req
->rq_import
->imp_at
;
263 idx
= import_at_get_index(req
->rq_import
, req
->rq_request_portal
);
264 /* max service estimates are tracked on the server side,
265 so just keep minimal history here */
266 oldse
= at_measured(&at
->iat_service_estimate
[idx
], serv_est
);
268 CDEBUG(D_ADAPTTO
, "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
269 req
->rq_import
->imp_obd
->obd_name
, req
->rq_request_portal
,
270 oldse
, at_get(&at
->iat_service_estimate
[idx
]));
273 /* Expected network latency per remote node (secs) */
274 int ptlrpc_at_get_net_latency(struct ptlrpc_request
*req
)
276 return AT_OFF
? 0 : at_get(&req
->rq_import
->imp_at
.iat_net_latency
);
279 /* Adjust expected network latency */
280 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request
*req
,
281 unsigned int service_time
)
283 unsigned int nl
, oldnl
;
285 time_t now
= get_seconds();
287 LASSERT(req
->rq_import
);
288 at
= &req
->rq_import
->imp_at
;
290 /* Network latency is total time less server processing time */
291 nl
= max_t(int, now
- req
->rq_sent
- service_time
, 0) + 1/*st rounding*/;
292 if (service_time
> now
- req
->rq_sent
+ 3 /* bz16408 */)
293 CWARN("Reported service time %u > total measured time "
294 CFS_DURATION_T
"\n", service_time
,
295 cfs_time_sub(now
, req
->rq_sent
));
297 oldnl
= at_measured(&at
->iat_net_latency
, nl
);
299 CDEBUG(D_ADAPTTO
, "The network latency for %s (nid %s) has changed from %d to %d\n",
300 req
->rq_import
->imp_obd
->obd_name
,
302 &req
->rq_import
->imp_connection
->c_remote_uuid
),
303 oldnl
, at_get(&at
->iat_net_latency
));
306 static int unpack_reply(struct ptlrpc_request
*req
)
310 if (SPTLRPC_FLVR_POLICY(req
->rq_flvr
.sf_rpc
) != SPTLRPC_POLICY_NULL
) {
311 rc
= ptlrpc_unpack_rep_msg(req
, req
->rq_replen
);
313 DEBUG_REQ(D_ERROR
, req
, "unpack_rep failed: %d", rc
);
318 rc
= lustre_unpack_rep_ptlrpc_body(req
, MSG_PTLRPC_BODY_OFF
);
320 DEBUG_REQ(D_ERROR
, req
, "unpack ptlrpc body failed: %d", rc
);
327 * Handle an early reply message, called with the rq_lock held.
328 * If anything goes wrong just ignore it - same as if it never happened
330 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request
*req
)
332 struct ptlrpc_request
*early_req
;
337 spin_unlock(&req
->rq_lock
);
339 rc
= sptlrpc_cli_unwrap_early_reply(req
, &early_req
);
341 spin_lock(&req
->rq_lock
);
345 rc
= unpack_reply(early_req
);
347 /* Expecting to increase the service time estimate here */
348 ptlrpc_at_adj_service(req
,
349 lustre_msg_get_timeout(early_req
->rq_repmsg
));
350 ptlrpc_at_adj_net_latency(req
,
351 lustre_msg_get_service_time(early_req
->rq_repmsg
));
354 sptlrpc_cli_finish_early_reply(early_req
);
357 spin_lock(&req
->rq_lock
);
361 /* Adjust the local timeout for this req */
362 ptlrpc_at_set_req_timeout(req
);
364 spin_lock(&req
->rq_lock
);
365 olddl
= req
->rq_deadline
;
366 /* server assumes it now has rq_timeout from when it sent the
367 * early reply, so client should give it at least that long. */
368 req
->rq_deadline
= get_seconds() + req
->rq_timeout
+
369 ptlrpc_at_get_net_latency(req
);
371 DEBUG_REQ(D_ADAPTTO
, req
,
372 "Early reply #%d, new deadline in " CFS_DURATION_T
"s (" CFS_DURATION_T
"s)",
374 cfs_time_sub(req
->rq_deadline
, get_seconds()),
375 cfs_time_sub(req
->rq_deadline
, olddl
));
380 struct kmem_cache
*request_cache
;
382 int ptlrpc_request_cache_init(void)
384 request_cache
= kmem_cache_create("ptlrpc_cache",
385 sizeof(struct ptlrpc_request
),
386 0, SLAB_HWCACHE_ALIGN
, NULL
);
387 return request_cache
== NULL
? -ENOMEM
: 0;
390 void ptlrpc_request_cache_fini(void)
392 kmem_cache_destroy(request_cache
);
395 struct ptlrpc_request
*ptlrpc_request_cache_alloc(gfp_t flags
)
397 struct ptlrpc_request
*req
;
399 OBD_SLAB_ALLOC_PTR_GFP(req
, request_cache
, flags
);
403 void ptlrpc_request_cache_free(struct ptlrpc_request
*req
)
405 OBD_SLAB_FREE_PTR(req
, request_cache
);
409 * Wind down request pool \a pool.
410 * Frees all requests from the pool too
412 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool
*pool
)
414 struct list_head
*l
, *tmp
;
415 struct ptlrpc_request
*req
;
417 LASSERT(pool
!= NULL
);
419 spin_lock(&pool
->prp_lock
);
420 list_for_each_safe(l
, tmp
, &pool
->prp_req_list
) {
421 req
= list_entry(l
, struct ptlrpc_request
, rq_list
);
422 list_del(&req
->rq_list
);
423 LASSERT(req
->rq_reqbuf
);
424 LASSERT(req
->rq_reqbuf_len
== pool
->prp_rq_size
);
425 OBD_FREE_LARGE(req
->rq_reqbuf
, pool
->prp_rq_size
);
426 ptlrpc_request_cache_free(req
);
428 spin_unlock(&pool
->prp_lock
);
429 OBD_FREE(pool
, sizeof(*pool
));
431 EXPORT_SYMBOL(ptlrpc_free_rq_pool
);
434 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
436 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool
*pool
, int num_rq
)
441 while (size
< pool
->prp_rq_size
)
444 LASSERTF(list_empty(&pool
->prp_req_list
) ||
445 size
== pool
->prp_rq_size
,
446 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
447 pool
->prp_rq_size
, size
);
449 spin_lock(&pool
->prp_lock
);
450 pool
->prp_rq_size
= size
;
451 for (i
= 0; i
< num_rq
; i
++) {
452 struct ptlrpc_request
*req
;
453 struct lustre_msg
*msg
;
455 spin_unlock(&pool
->prp_lock
);
456 req
= ptlrpc_request_cache_alloc(GFP_NOFS
);
459 OBD_ALLOC_LARGE(msg
, size
);
461 ptlrpc_request_cache_free(req
);
464 req
->rq_reqbuf
= msg
;
465 req
->rq_reqbuf_len
= size
;
467 spin_lock(&pool
->prp_lock
);
468 list_add_tail(&req
->rq_list
, &pool
->prp_req_list
);
470 spin_unlock(&pool
->prp_lock
);
473 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool
);
476 * Create and initialize new request pool with given attributes:
477 * \a num_rq - initial number of requests to create for the pool
478 * \a msgsize - maximum message size possible for requests in thid pool
479 * \a populate_pool - function to be called when more requests need to be added
481 * Returns pointer to newly created pool or NULL on error.
483 struct ptlrpc_request_pool
*
484 ptlrpc_init_rq_pool(int num_rq
, int msgsize
,
485 void (*populate_pool
)(struct ptlrpc_request_pool
*, int))
487 struct ptlrpc_request_pool
*pool
;
489 OBD_ALLOC(pool
, sizeof(struct ptlrpc_request_pool
));
493 /* Request next power of two for the allocation, because internally
494 kernel would do exactly this */
496 spin_lock_init(&pool
->prp_lock
);
497 INIT_LIST_HEAD(&pool
->prp_req_list
);
498 pool
->prp_rq_size
= msgsize
+ SPTLRPC_MAX_PAYLOAD
;
499 pool
->prp_populate
= populate_pool
;
501 populate_pool(pool
, num_rq
);
503 if (list_empty(&pool
->prp_req_list
)) {
504 /* have not allocated a single request for the pool */
505 OBD_FREE(pool
, sizeof(struct ptlrpc_request_pool
));
510 EXPORT_SYMBOL(ptlrpc_init_rq_pool
);
513 * Fetches one request from pool \a pool
515 static struct ptlrpc_request
*
516 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool
*pool
)
518 struct ptlrpc_request
*request
;
519 struct lustre_msg
*reqbuf
;
524 spin_lock(&pool
->prp_lock
);
526 /* See if we have anything in a pool, and bail out if nothing,
527 * in writeout path, where this matters, this is safe to do, because
528 * nothing is lost in this case, and when some in-flight requests
529 * complete, this code will be called again. */
530 if (unlikely(list_empty(&pool
->prp_req_list
))) {
531 spin_unlock(&pool
->prp_lock
);
535 request
= list_entry(pool
->prp_req_list
.next
, struct ptlrpc_request
,
537 list_del_init(&request
->rq_list
);
538 spin_unlock(&pool
->prp_lock
);
540 LASSERT(request
->rq_reqbuf
);
541 LASSERT(request
->rq_pool
);
543 reqbuf
= request
->rq_reqbuf
;
544 memset(request
, 0, sizeof(*request
));
545 request
->rq_reqbuf
= reqbuf
;
546 request
->rq_reqbuf_len
= pool
->prp_rq_size
;
547 request
->rq_pool
= pool
;
553 * Returns freed \a request to pool.
555 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request
*request
)
557 struct ptlrpc_request_pool
*pool
= request
->rq_pool
;
559 spin_lock(&pool
->prp_lock
);
560 LASSERT(list_empty(&request
->rq_list
));
561 LASSERT(!request
->rq_receiving_reply
);
562 list_add_tail(&request
->rq_list
, &pool
->prp_req_list
);
563 spin_unlock(&pool
->prp_lock
);
566 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
567 __u32 version
, int opcode
,
568 int count
, __u32
*lengths
, char **bufs
,
569 struct ptlrpc_cli_ctx
*ctx
)
571 struct obd_import
*imp
= request
->rq_import
;
575 request
->rq_cli_ctx
= sptlrpc_cli_ctx_get(ctx
);
577 rc
= sptlrpc_req_get_ctx(request
);
582 sptlrpc_req_set_flavor(request
, opcode
);
584 rc
= lustre_pack_request(request
, imp
->imp_msg_magic
, count
,
587 LASSERT(!request
->rq_pool
);
591 lustre_msg_add_version(request
->rq_reqmsg
, version
);
592 request
->rq_send_state
= LUSTRE_IMP_FULL
;
593 request
->rq_type
= PTL_RPC_MSG_REQUEST
;
594 request
->rq_export
= NULL
;
596 request
->rq_req_cbid
.cbid_fn
= request_out_callback
;
597 request
->rq_req_cbid
.cbid_arg
= request
;
599 request
->rq_reply_cbid
.cbid_fn
= reply_in_callback
;
600 request
->rq_reply_cbid
.cbid_arg
= request
;
602 request
->rq_reply_deadline
= 0;
603 request
->rq_phase
= RQ_PHASE_NEW
;
604 request
->rq_next_phase
= RQ_PHASE_UNDEFINED
;
606 request
->rq_request_portal
= imp
->imp_client
->cli_request_portal
;
607 request
->rq_reply_portal
= imp
->imp_client
->cli_reply_portal
;
609 ptlrpc_at_set_req_timeout(request
);
611 spin_lock_init(&request
->rq_lock
);
612 INIT_LIST_HEAD(&request
->rq_list
);
613 INIT_LIST_HEAD(&request
->rq_timed_list
);
614 INIT_LIST_HEAD(&request
->rq_replay_list
);
615 INIT_LIST_HEAD(&request
->rq_ctx_chain
);
616 INIT_LIST_HEAD(&request
->rq_set_chain
);
617 INIT_LIST_HEAD(&request
->rq_history_list
);
618 INIT_LIST_HEAD(&request
->rq_exp_list
);
619 init_waitqueue_head(&request
->rq_reply_waitq
);
620 init_waitqueue_head(&request
->rq_set_waitq
);
621 request
->rq_xid
= ptlrpc_next_xid();
622 atomic_set(&request
->rq_refcount
, 1);
624 lustre_msg_set_opc(request
->rq_reqmsg
, opcode
);
628 sptlrpc_cli_ctx_put(request
->rq_cli_ctx
, 1);
630 class_import_put(imp
);
634 int ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
635 __u32 version
, int opcode
, char **bufs
,
636 struct ptlrpc_cli_ctx
*ctx
)
640 count
= req_capsule_filled_sizes(&request
->rq_pill
, RCL_CLIENT
);
641 return __ptlrpc_request_bufs_pack(request
, version
, opcode
, count
,
642 request
->rq_pill
.rc_area
[RCL_CLIENT
],
645 EXPORT_SYMBOL(ptlrpc_request_bufs_pack
);
648 * Pack request buffers for network transfer, performing necessary encryption
649 * steps if necessary.
651 int ptlrpc_request_pack(struct ptlrpc_request
*request
,
652 __u32 version
, int opcode
)
655 rc
= ptlrpc_request_bufs_pack(request
, version
, opcode
, NULL
, NULL
);
659 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
660 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
661 * have to send old ptlrpc_body to keep interoperability with these
664 * Only three kinds of server->client RPCs so far:
669 * XXX This should be removed whenever we drop the interoperability with
670 * the these old clients.
672 if (opcode
== LDLM_BL_CALLBACK
|| opcode
== LDLM_CP_CALLBACK
||
673 opcode
== LDLM_GL_CALLBACK
)
674 req_capsule_shrink(&request
->rq_pill
, &RMF_PTLRPC_BODY
,
675 sizeof(struct ptlrpc_body_v2
), RCL_CLIENT
);
679 EXPORT_SYMBOL(ptlrpc_request_pack
);
682 * Helper function to allocate new request on import \a imp
683 * and possibly using existing request from pool \a pool if provided.
684 * Returns allocated request structure with import field filled or
688 struct ptlrpc_request
*__ptlrpc_request_alloc(struct obd_import
*imp
,
689 struct ptlrpc_request_pool
*pool
)
691 struct ptlrpc_request
*request
= NULL
;
694 request
= ptlrpc_prep_req_from_pool(pool
);
697 request
= ptlrpc_request_cache_alloc(GFP_NOFS
);
700 LASSERTF((unsigned long)imp
> 0x1000, "%p", imp
);
701 LASSERT(imp
!= LP_POISON
);
702 LASSERTF((unsigned long)imp
->imp_client
> 0x1000, "%p",
704 LASSERT(imp
->imp_client
!= LP_POISON
);
706 request
->rq_import
= class_import_get(imp
);
708 CERROR("request allocation out of memory\n");
715 * Helper function for creating a request.
716 * Calls __ptlrpc_request_alloc to allocate new request structure and inits
717 * buffer structures according to capsule template \a format.
718 * Returns allocated request structure pointer or NULL on error.
720 static struct ptlrpc_request
*
721 ptlrpc_request_alloc_internal(struct obd_import
*imp
,
722 struct ptlrpc_request_pool
*pool
,
723 const struct req_format
*format
)
725 struct ptlrpc_request
*request
;
727 request
= __ptlrpc_request_alloc(imp
, pool
);
731 req_capsule_init(&request
->rq_pill
, request
, RCL_CLIENT
);
732 req_capsule_set(&request
->rq_pill
, format
);
737 * Allocate new request structure for import \a imp and initialize its
738 * buffer structure according to capsule template \a format.
740 struct ptlrpc_request
*ptlrpc_request_alloc(struct obd_import
*imp
,
741 const struct req_format
*format
)
743 return ptlrpc_request_alloc_internal(imp
, NULL
, format
);
745 EXPORT_SYMBOL(ptlrpc_request_alloc
);
748 * Allocate new request structure for import \a imp from pool \a pool and
749 * initialize its buffer structure according to capsule template \a format.
751 struct ptlrpc_request
*ptlrpc_request_alloc_pool(struct obd_import
*imp
,
752 struct ptlrpc_request_pool
*pool
,
753 const struct req_format
*format
)
755 return ptlrpc_request_alloc_internal(imp
, pool
, format
);
757 EXPORT_SYMBOL(ptlrpc_request_alloc_pool
);
760 * For requests not from pool, free memory of the request structure.
761 * For requests obtained from a pool earlier, return request back to pool.
763 void ptlrpc_request_free(struct ptlrpc_request
*request
)
765 if (request
->rq_pool
)
766 __ptlrpc_free_req_to_pool(request
);
768 ptlrpc_request_cache_free(request
);
770 EXPORT_SYMBOL(ptlrpc_request_free
);
773 * Allocate new request for operation \a opcode and immediately pack it for
775 * Only used for simple requests like OBD_PING where the only important
776 * part of the request is operation itself.
777 * Returns allocated request or NULL on error.
779 struct ptlrpc_request
*ptlrpc_request_alloc_pack(struct obd_import
*imp
,
780 const struct req_format
*format
,
781 __u32 version
, int opcode
)
783 struct ptlrpc_request
*req
= ptlrpc_request_alloc(imp
, format
);
787 rc
= ptlrpc_request_pack(req
, version
, opcode
);
789 ptlrpc_request_free(req
);
795 EXPORT_SYMBOL(ptlrpc_request_alloc_pack
);
798 * Prepare request (fetched from pool \a pool if not NULL) on import \a imp
799 * for operation \a opcode. Request would contain \a count buffers.
800 * Sizes of buffers are described in array \a lengths and buffers themselves
801 * are provided by a pointer \a bufs.
802 * Returns prepared request structure pointer or NULL on error.
804 struct ptlrpc_request
*
805 ptlrpc_prep_req_pool(struct obd_import
*imp
,
806 __u32 version
, int opcode
,
807 int count
, __u32
*lengths
, char **bufs
,
808 struct ptlrpc_request_pool
*pool
)
810 struct ptlrpc_request
*request
;
813 request
= __ptlrpc_request_alloc(imp
, pool
);
817 rc
= __ptlrpc_request_bufs_pack(request
, version
, opcode
, count
,
818 lengths
, bufs
, NULL
);
820 ptlrpc_request_free(request
);
825 EXPORT_SYMBOL(ptlrpc_prep_req_pool
);
828 * Same as ptlrpc_prep_req_pool, but without pool
830 struct ptlrpc_request
*
831 ptlrpc_prep_req(struct obd_import
*imp
, __u32 version
, int opcode
, int count
,
832 __u32
*lengths
, char **bufs
)
834 return ptlrpc_prep_req_pool(imp
, version
, opcode
, count
, lengths
, bufs
,
837 EXPORT_SYMBOL(ptlrpc_prep_req
);
840 * Allocate and initialize new request set structure.
841 * Returns a pointer to the newly allocated set structure or NULL on error.
843 struct ptlrpc_request_set
*ptlrpc_prep_set(void)
845 struct ptlrpc_request_set
*set
;
847 OBD_ALLOC(set
, sizeof(*set
));
850 atomic_set(&set
->set_refcount
, 1);
851 INIT_LIST_HEAD(&set
->set_requests
);
852 init_waitqueue_head(&set
->set_waitq
);
853 atomic_set(&set
->set_new_count
, 0);
854 atomic_set(&set
->set_remaining
, 0);
855 spin_lock_init(&set
->set_new_req_lock
);
856 INIT_LIST_HEAD(&set
->set_new_requests
);
857 INIT_LIST_HEAD(&set
->set_cblist
);
858 set
->set_max_inflight
= UINT_MAX
;
859 set
->set_producer
= NULL
;
860 set
->set_producer_arg
= NULL
;
865 EXPORT_SYMBOL(ptlrpc_prep_set
);
868 * Allocate and initialize new request set structure with flow control
869 * extension. This extension allows to control the number of requests in-flight
870 * for the whole set. A callback function to generate requests must be provided
871 * and the request set will keep the number of requests sent over the wire to
873 * Returns a pointer to the newly allocated set structure or NULL on error.
875 struct ptlrpc_request_set
*ptlrpc_prep_fcset(int max
, set_producer_func func
,
879 struct ptlrpc_request_set
*set
;
881 set
= ptlrpc_prep_set();
885 set
->set_max_inflight
= max
;
886 set
->set_producer
= func
;
887 set
->set_producer_arg
= arg
;
891 EXPORT_SYMBOL(ptlrpc_prep_fcset
);
894 * Wind down and free request set structure previously allocated with
896 * Ensures that all requests on the set have completed and removes
897 * all requests from the request list in a set.
898 * If any unsent request happen to be on the list, pretends that they got
899 * an error in flight and calls their completion handler.
901 void ptlrpc_set_destroy(struct ptlrpc_request_set
*set
)
903 struct list_head
*tmp
;
904 struct list_head
*next
;
908 /* Requests on the set should either all be completed, or all be new */
909 expected_phase
= (atomic_read(&set
->set_remaining
) == 0) ?
910 RQ_PHASE_COMPLETE
: RQ_PHASE_NEW
;
911 list_for_each(tmp
, &set
->set_requests
) {
912 struct ptlrpc_request
*req
=
913 list_entry(tmp
, struct ptlrpc_request
,
916 LASSERT(req
->rq_phase
== expected_phase
);
920 LASSERTF(atomic_read(&set
->set_remaining
) == 0 ||
921 atomic_read(&set
->set_remaining
) == n
, "%d / %d\n",
922 atomic_read(&set
->set_remaining
), n
);
924 list_for_each_safe(tmp
, next
, &set
->set_requests
) {
925 struct ptlrpc_request
*req
=
926 list_entry(tmp
, struct ptlrpc_request
,
928 list_del_init(&req
->rq_set_chain
);
930 LASSERT(req
->rq_phase
== expected_phase
);
932 if (req
->rq_phase
== RQ_PHASE_NEW
) {
933 ptlrpc_req_interpret(NULL
, req
, -EBADR
);
934 atomic_dec(&set
->set_remaining
);
937 spin_lock(&req
->rq_lock
);
939 req
->rq_invalid_rqset
= 0;
940 spin_unlock(&req
->rq_lock
);
942 ptlrpc_req_finished(req
);
945 LASSERT(atomic_read(&set
->set_remaining
) == 0);
947 ptlrpc_reqset_put(set
);
949 EXPORT_SYMBOL(ptlrpc_set_destroy
);
952 * Add a callback function \a fn to the set.
953 * This function would be called when all requests on this set are completed.
954 * The function will be passed \a data argument.
956 int ptlrpc_set_add_cb(struct ptlrpc_request_set
*set
,
957 set_interpreter_func fn
, void *data
)
959 struct ptlrpc_set_cbdata
*cbdata
;
961 OBD_ALLOC_PTR(cbdata
);
965 cbdata
->psc_interpret
= fn
;
966 cbdata
->psc_data
= data
;
967 list_add_tail(&cbdata
->psc_item
, &set
->set_cblist
);
971 EXPORT_SYMBOL(ptlrpc_set_add_cb
);
974 * Add a new request to the general purpose request set.
975 * Assumes request reference from the caller.
977 void ptlrpc_set_add_req(struct ptlrpc_request_set
*set
,
978 struct ptlrpc_request
*req
)
980 LASSERT(list_empty(&req
->rq_set_chain
));
982 /* The set takes over the caller's request reference */
983 list_add_tail(&req
->rq_set_chain
, &set
->set_requests
);
985 atomic_inc(&set
->set_remaining
);
986 req
->rq_queued_time
= cfs_time_current();
988 if (req
->rq_reqmsg
!= NULL
)
989 lustre_msg_set_jobid(req
->rq_reqmsg
, NULL
);
991 if (set
->set_producer
!= NULL
)
992 /* If the request set has a producer callback, the RPC must be
993 * sent straight away */
994 ptlrpc_send_new_req(req
);
996 EXPORT_SYMBOL(ptlrpc_set_add_req
);
999 * Add a request to a request with dedicated server thread
1000 * and wake the thread to make any necessary processing.
1001 * Currently only used for ptlrpcd.
1003 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl
*pc
,
1004 struct ptlrpc_request
*req
)
1006 struct ptlrpc_request_set
*set
= pc
->pc_set
;
1009 LASSERT(req
->rq_set
== NULL
);
1010 LASSERT(test_bit(LIOD_STOP
, &pc
->pc_flags
) == 0);
1012 spin_lock(&set
->set_new_req_lock
);
1014 * The set takes over the caller's request reference.
1017 req
->rq_queued_time
= cfs_time_current();
1018 list_add_tail(&req
->rq_set_chain
, &set
->set_new_requests
);
1019 count
= atomic_inc_return(&set
->set_new_count
);
1020 spin_unlock(&set
->set_new_req_lock
);
1022 /* Only need to call wakeup once for the first entry. */
1024 wake_up(&set
->set_waitq
);
1026 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1027 * guarantee the async RPC can be processed ASAP, we have
1028 * no other better choice. It maybe fixed in future. */
1029 for (i
= 0; i
< pc
->pc_npartners
; i
++)
1030 wake_up(&pc
->pc_partners
[i
]->pc_set
->set_waitq
);
1033 EXPORT_SYMBOL(ptlrpc_set_add_new_req
);
1036 * Based on the current state of the import, determine if the request
1037 * can be sent, is an error, or should be delayed.
1039 * Returns true if this request should be delayed. If false, and
1040 * *status is set, then the request can not be sent and *status is the
1041 * error code. If false and status is 0, then request can be sent.
1043 * The imp->imp_lock must be held.
1045 static int ptlrpc_import_delay_req(struct obd_import
*imp
,
1046 struct ptlrpc_request
*req
, int *status
)
1050 LASSERT(status
!= NULL
);
1053 if (req
->rq_ctx_init
|| req
->rq_ctx_fini
) {
1054 /* always allow ctx init/fini rpc go through */
1055 } else if (imp
->imp_state
== LUSTRE_IMP_NEW
) {
1056 DEBUG_REQ(D_ERROR
, req
, "Uninitialized import.");
1058 } else if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
1059 /* pings may safely race with umount */
1060 DEBUG_REQ(lustre_msg_get_opc(req
->rq_reqmsg
) == OBD_PING
?
1061 D_HA
: D_ERROR
, req
, "IMP_CLOSED ");
1063 } else if (ptlrpc_send_limit_expired(req
)) {
1064 /* probably doesn't need to be a D_ERROR after initial testing */
1065 DEBUG_REQ(D_ERROR
, req
, "send limit expired ");
1067 } else if (req
->rq_send_state
== LUSTRE_IMP_CONNECTING
&&
1068 imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
1069 /* allow CONNECT even if import is invalid */
1070 if (atomic_read(&imp
->imp_inval_count
) != 0) {
1071 DEBUG_REQ(D_ERROR
, req
, "invalidate in flight");
1074 } else if (imp
->imp_invalid
|| imp
->imp_obd
->obd_no_recov
) {
1075 if (!imp
->imp_deactive
)
1076 DEBUG_REQ(D_NET
, req
, "IMP_INVALID");
1077 *status
= -ESHUTDOWN
; /* bz 12940 */
1078 } else if (req
->rq_import_generation
!= imp
->imp_generation
) {
1079 DEBUG_REQ(D_ERROR
, req
, "req wrong generation:");
1081 } else if (req
->rq_send_state
!= imp
->imp_state
) {
1082 /* invalidate in progress - any requests should be drop */
1083 if (atomic_read(&imp
->imp_inval_count
) != 0) {
1084 DEBUG_REQ(D_ERROR
, req
, "invalidate in flight");
1086 } else if (imp
->imp_dlm_fake
|| req
->rq_no_delay
) {
1087 *status
= -EWOULDBLOCK
;
1088 } else if (req
->rq_allow_replay
&&
1089 (imp
->imp_state
== LUSTRE_IMP_REPLAY
||
1090 imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
||
1091 imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
||
1092 imp
->imp_state
== LUSTRE_IMP_RECOVER
)) {
1093 DEBUG_REQ(D_HA
, req
, "allow during recovery.\n");
1103 * Decide if the error message regarding provided request \a req
1104 * should be printed to the console or not.
1105 * Makes it's decision on request status and other properties.
1106 * Returns 1 to print error on the system console or 0 if not.
1108 static int ptlrpc_console_allow(struct ptlrpc_request
*req
)
1113 LASSERT(req
->rq_reqmsg
!= NULL
);
1114 opc
= lustre_msg_get_opc(req
->rq_reqmsg
);
1116 /* Suppress particular reconnect errors which are to be expected. No
1117 * errors are suppressed for the initial connection on an import */
1118 if ((lustre_handle_is_used(&req
->rq_import
->imp_remote_handle
)) &&
1119 (opc
== OST_CONNECT
|| opc
== MDS_CONNECT
|| opc
== MGS_CONNECT
)) {
1121 /* Suppress timed out reconnect requests */
1122 if (req
->rq_timedout
)
1125 /* Suppress unavailable/again reconnect requests */
1126 err
= lustre_msg_get_status(req
->rq_repmsg
);
1127 if (err
== -ENODEV
|| err
== -EAGAIN
)
1135 * Check request processing status.
1136 * Returns the status.
1138 static int ptlrpc_check_status(struct ptlrpc_request
*req
)
1142 err
= lustre_msg_get_status(req
->rq_repmsg
);
1143 if (lustre_msg_get_type(req
->rq_repmsg
) == PTL_RPC_MSG_ERR
) {
1144 struct obd_import
*imp
= req
->rq_import
;
1145 __u32 opc
= lustre_msg_get_opc(req
->rq_reqmsg
);
1146 if (ptlrpc_console_allow(req
))
1147 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s, operation %s failed with %d.\n",
1148 imp
->imp_obd
->obd_name
,
1150 imp
->imp_connection
->c_peer
.nid
),
1151 ll_opcode2str(opc
), err
);
1152 return err
< 0 ? err
: -EINVAL
;
1156 DEBUG_REQ(D_INFO
, req
, "status is %d", err
);
1157 } else if (err
> 0) {
1158 /* XXX: translate this error from net to host */
1159 DEBUG_REQ(D_INFO
, req
, "status is %d", err
);
1166 * save pre-versions of objects into request for replay.
1167 * Versions are obtained from server reply.
1170 static void ptlrpc_save_versions(struct ptlrpc_request
*req
)
1172 struct lustre_msg
*repmsg
= req
->rq_repmsg
;
1173 struct lustre_msg
*reqmsg
= req
->rq_reqmsg
;
1174 __u64
*versions
= lustre_msg_get_versions(repmsg
);
1176 if (lustre_msg_get_flags(req
->rq_reqmsg
) & MSG_REPLAY
)
1180 lustre_msg_set_versions(reqmsg
, versions
);
1181 CDEBUG(D_INFO
, "Client save versions [%#llx/%#llx]\n",
1182 versions
[0], versions
[1]);
1186 * Callback function called when client receives RPC reply for \a req.
1187 * Returns 0 on success or error code.
1188 * The return value would be assigned to req->rq_status by the caller
1189 * as request processing status.
1190 * This function also decides if the request needs to be saved for later replay.
1192 static int after_reply(struct ptlrpc_request
*req
)
1194 struct obd_import
*imp
= req
->rq_import
;
1195 struct obd_device
*obd
= req
->rq_import
->imp_obd
;
1197 struct timeval work_start
;
1200 LASSERT(obd
!= NULL
);
1201 /* repbuf must be unlinked */
1202 LASSERT(!req
->rq_receiving_reply
&& !req
->rq_reply_unlink
);
1204 if (req
->rq_reply_truncate
) {
1205 if (ptlrpc_no_resend(req
)) {
1206 DEBUG_REQ(D_ERROR
, req
, "reply buffer overflow, expected: %d, actual size: %d",
1207 req
->rq_nob_received
, req
->rq_repbuf_len
);
1211 sptlrpc_cli_free_repbuf(req
);
1212 /* Pass the required reply buffer size (include
1213 * space for early reply).
1214 * NB: no need to roundup because alloc_repbuf
1215 * will roundup it */
1216 req
->rq_replen
= req
->rq_nob_received
;
1217 req
->rq_nob_received
= 0;
1218 spin_lock(&req
->rq_lock
);
1220 spin_unlock(&req
->rq_lock
);
1225 * NB Until this point, the whole of the incoming message,
1226 * including buflens, status etc is in the sender's byte order.
1228 rc
= sptlrpc_cli_unwrap_reply(req
);
1230 DEBUG_REQ(D_ERROR
, req
, "unwrap reply failed (%d):", rc
);
1235 * Security layer unwrap might ask resend this request.
1240 rc
= unpack_reply(req
);
1244 /* retry indefinitely on EINPROGRESS */
1245 if (lustre_msg_get_status(req
->rq_repmsg
) == -EINPROGRESS
&&
1246 ptlrpc_no_resend(req
) == 0 && !req
->rq_no_retry_einprogress
) {
1247 time_t now
= get_seconds();
1249 DEBUG_REQ(D_RPCTRACE
, req
, "Resending request on EINPROGRESS");
1250 spin_lock(&req
->rq_lock
);
1252 spin_unlock(&req
->rq_lock
);
1253 req
->rq_nr_resend
++;
1255 /* allocate new xid to avoid reply reconstruction */
1256 if (!req
->rq_bulk
) {
1257 /* new xid is already allocated for bulk in
1258 * ptlrpc_check_set() */
1259 req
->rq_xid
= ptlrpc_next_xid();
1260 DEBUG_REQ(D_RPCTRACE
, req
, "Allocating new xid for resend on EINPROGRESS");
1263 /* Readjust the timeout for current conditions */
1264 ptlrpc_at_set_req_timeout(req
);
1265 /* delay resend to give a chance to the server to get ready.
1266 * The delay is increased by 1s on every resend and is capped to
1267 * the current request timeout (i.e. obd_timeout if AT is off,
1268 * or AT service time x 125% + 5s, see at_est2timeout) */
1269 if (req
->rq_nr_resend
> req
->rq_timeout
)
1270 req
->rq_sent
= now
+ req
->rq_timeout
;
1272 req
->rq_sent
= now
+ req
->rq_nr_resend
;
1277 do_gettimeofday(&work_start
);
1278 timediff
= cfs_timeval_sub(&work_start
, &req
->rq_arrival_time
, NULL
);
1279 if (obd
->obd_svc_stats
!= NULL
) {
1280 lprocfs_counter_add(obd
->obd_svc_stats
, PTLRPC_REQWAIT_CNTR
,
1282 ptlrpc_lprocfs_rpc_sent(req
, timediff
);
1285 if (lustre_msg_get_type(req
->rq_repmsg
) != PTL_RPC_MSG_REPLY
&&
1286 lustre_msg_get_type(req
->rq_repmsg
) != PTL_RPC_MSG_ERR
) {
1287 DEBUG_REQ(D_ERROR
, req
, "invalid packet received (type=%u)",
1288 lustre_msg_get_type(req
->rq_repmsg
));
1292 if (lustre_msg_get_opc(req
->rq_reqmsg
) != OBD_PING
)
1293 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP
, cfs_fail_val
);
1294 ptlrpc_at_adj_service(req
, lustre_msg_get_timeout(req
->rq_repmsg
));
1295 ptlrpc_at_adj_net_latency(req
,
1296 lustre_msg_get_service_time(req
->rq_repmsg
));
1298 rc
= ptlrpc_check_status(req
);
1299 imp
->imp_connect_error
= rc
;
1303 * Either we've been evicted, or the server has failed for
1304 * some reason. Try to reconnect, and if that fails, punt to
1307 if (ll_rpc_recoverable_error(rc
)) {
1308 if (req
->rq_send_state
!= LUSTRE_IMP_FULL
||
1309 imp
->imp_obd
->obd_no_recov
|| imp
->imp_dlm_fake
) {
1312 ptlrpc_request_handle_notconn(req
);
1317 * Let's look if server sent slv. Do it only for RPC with
1320 ldlm_cli_update_pool(req
);
1324 * Store transno in reqmsg for replay.
1326 if (!(lustre_msg_get_flags(req
->rq_reqmsg
) & MSG_REPLAY
)) {
1327 req
->rq_transno
= lustre_msg_get_transno(req
->rq_repmsg
);
1328 lustre_msg_set_transno(req
->rq_reqmsg
, req
->rq_transno
);
1331 if (imp
->imp_replayable
) {
1332 spin_lock(&imp
->imp_lock
);
1334 * No point in adding already-committed requests to the replay
1335 * list, we will just remove them immediately. b=9829
1337 if (req
->rq_transno
!= 0 &&
1339 lustre_msg_get_last_committed(req
->rq_repmsg
) ||
1341 /** version recovery */
1342 ptlrpc_save_versions(req
);
1343 ptlrpc_retain_replayable_request(req
, imp
);
1344 } else if (req
->rq_commit_cb
!= NULL
&&
1345 list_empty(&req
->rq_replay_list
)) {
1346 /* NB: don't call rq_commit_cb if it's already on
1347 * rq_replay_list, ptlrpc_free_committed() will call
1348 * it later, see LU-3618 for details */
1349 spin_unlock(&imp
->imp_lock
);
1350 req
->rq_commit_cb(req
);
1351 spin_lock(&imp
->imp_lock
);
1355 * Replay-enabled imports return commit-status information.
1357 if (lustre_msg_get_last_committed(req
->rq_repmsg
)) {
1358 imp
->imp_peer_committed_transno
=
1359 lustre_msg_get_last_committed(req
->rq_repmsg
);
1362 ptlrpc_free_committed(imp
);
1364 if (!list_empty(&imp
->imp_replay_list
)) {
1365 struct ptlrpc_request
*last
;
1367 last
= list_entry(imp
->imp_replay_list
.prev
,
1368 struct ptlrpc_request
,
1371 * Requests with rq_replay stay on the list even if no
1372 * commit is expected.
1374 if (last
->rq_transno
> imp
->imp_peer_committed_transno
)
1375 ptlrpc_pinger_commit_expected(imp
);
1378 spin_unlock(&imp
->imp_lock
);
1385 * Helper function to send request \a req over the network for the first time
1386 * Also adjusts request phase.
1387 * Returns 0 on success or error code.
1389 static int ptlrpc_send_new_req(struct ptlrpc_request
*req
)
1391 struct obd_import
*imp
= req
->rq_import
;
1394 LASSERT(req
->rq_phase
== RQ_PHASE_NEW
);
1395 if (req
->rq_sent
&& (req
->rq_sent
> get_seconds()) &&
1396 (!req
->rq_generation_set
||
1397 req
->rq_import_generation
== imp
->imp_generation
))
1400 ptlrpc_rqphase_move(req
, RQ_PHASE_RPC
);
1402 spin_lock(&imp
->imp_lock
);
1404 if (!req
->rq_generation_set
)
1405 req
->rq_import_generation
= imp
->imp_generation
;
1407 if (ptlrpc_import_delay_req(imp
, req
, &rc
)) {
1408 spin_lock(&req
->rq_lock
);
1409 req
->rq_waiting
= 1;
1410 spin_unlock(&req
->rq_lock
);
1412 DEBUG_REQ(D_HA
, req
, "req from PID %d waiting for recovery: (%s != %s)",
1413 lustre_msg_get_status(req
->rq_reqmsg
),
1414 ptlrpc_import_state_name(req
->rq_send_state
),
1415 ptlrpc_import_state_name(imp
->imp_state
));
1416 LASSERT(list_empty(&req
->rq_list
));
1417 list_add_tail(&req
->rq_list
, &imp
->imp_delayed_list
);
1418 atomic_inc(&req
->rq_import
->imp_inflight
);
1419 spin_unlock(&imp
->imp_lock
);
1424 spin_unlock(&imp
->imp_lock
);
1425 req
->rq_status
= rc
;
1426 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1430 LASSERT(list_empty(&req
->rq_list
));
1431 list_add_tail(&req
->rq_list
, &imp
->imp_sending_list
);
1432 atomic_inc(&req
->rq_import
->imp_inflight
);
1433 spin_unlock(&imp
->imp_lock
);
1435 lustre_msg_set_status(req
->rq_reqmsg
, current_pid());
1437 rc
= sptlrpc_req_refresh_ctx(req
, -1);
1440 req
->rq_status
= rc
;
1443 spin_lock(&req
->rq_lock
);
1444 req
->rq_wait_ctx
= 1;
1445 spin_unlock(&req
->rq_lock
);
1450 CDEBUG(D_RPCTRACE
, "Sending RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1452 imp
->imp_obd
->obd_uuid
.uuid
,
1453 lustre_msg_get_status(req
->rq_reqmsg
), req
->rq_xid
,
1454 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
),
1455 lustre_msg_get_opc(req
->rq_reqmsg
));
1457 rc
= ptl_send_rpc(req
, 0);
1459 DEBUG_REQ(D_HA
, req
, "send failed (%d); expect timeout", rc
);
1460 spin_lock(&req
->rq_lock
);
1461 req
->rq_net_err
= 1;
1462 spin_unlock(&req
->rq_lock
);
1468 static inline int ptlrpc_set_producer(struct ptlrpc_request_set
*set
)
1472 LASSERT(set
->set_producer
!= NULL
);
1474 remaining
= atomic_read(&set
->set_remaining
);
1476 /* populate the ->set_requests list with requests until we
1477 * reach the maximum number of RPCs in flight for this set */
1478 while (atomic_read(&set
->set_remaining
) < set
->set_max_inflight
) {
1479 rc
= set
->set_producer(set
, set
->set_producer_arg
);
1480 if (rc
== -ENOENT
) {
1481 /* no more RPC to produce */
1482 set
->set_producer
= NULL
;
1483 set
->set_producer_arg
= NULL
;
1488 return (atomic_read(&set
->set_remaining
) - remaining
);
1492 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1493 * and no more replies are expected.
1494 * (it is possible to get less replies than requests sent e.g. due to timed out
1495 * requests or requests that we had trouble to send out)
1497 * NOTE: This function contains a potential schedule point (cond_resched()).
1499 int ptlrpc_check_set(const struct lu_env
*env
, struct ptlrpc_request_set
*set
)
1501 struct list_head
*tmp
, *next
;
1502 struct list_head comp_reqs
;
1503 int force_timer_recalc
= 0;
1505 if (atomic_read(&set
->set_remaining
) == 0)
1508 INIT_LIST_HEAD(&comp_reqs
);
1509 list_for_each_safe(tmp
, next
, &set
->set_requests
) {
1510 struct ptlrpc_request
*req
=
1511 list_entry(tmp
, struct ptlrpc_request
,
1513 struct obd_import
*imp
= req
->rq_import
;
1514 int unregistered
= 0;
1517 /* This schedule point is mainly for the ptlrpcd caller of this
1518 * function. Most ptlrpc sets are not long-lived and unbounded
1519 * in length, but at the least the set used by the ptlrpcd is.
1520 * Since the processing time is unbounded, we need to insert an
1521 * explicit schedule point to make the thread well-behaved.
1525 if (req
->rq_phase
== RQ_PHASE_NEW
&&
1526 ptlrpc_send_new_req(req
)) {
1527 force_timer_recalc
= 1;
1530 /* delayed send - skip */
1531 if (req
->rq_phase
== RQ_PHASE_NEW
&& req
->rq_sent
)
1534 /* delayed resend - skip */
1535 if (req
->rq_phase
== RQ_PHASE_RPC
&& req
->rq_resend
&&
1536 req
->rq_sent
> get_seconds())
1539 if (!(req
->rq_phase
== RQ_PHASE_RPC
||
1540 req
->rq_phase
== RQ_PHASE_BULK
||
1541 req
->rq_phase
== RQ_PHASE_INTERPRET
||
1542 req
->rq_phase
== RQ_PHASE_UNREGISTERING
||
1543 req
->rq_phase
== RQ_PHASE_COMPLETE
)) {
1544 DEBUG_REQ(D_ERROR
, req
, "bad phase %x", req
->rq_phase
);
1548 if (req
->rq_phase
== RQ_PHASE_UNREGISTERING
) {
1549 LASSERT(req
->rq_next_phase
!= req
->rq_phase
);
1550 LASSERT(req
->rq_next_phase
!= RQ_PHASE_UNDEFINED
);
1553 * Skip processing until reply is unlinked. We
1554 * can't return to pool before that and we can't
1555 * call interpret before that. We need to make
1556 * sure that all rdma transfers finished and will
1557 * not corrupt any data.
1559 if (ptlrpc_client_recv_or_unlink(req
) ||
1560 ptlrpc_client_bulk_active(req
))
1564 * Turn fail_loc off to prevent it from looping
1567 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
1568 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
,
1571 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
)) {
1572 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
,
1577 * Move to next phase if reply was successfully
1580 ptlrpc_rqphase_move(req
, req
->rq_next_phase
);
1583 if (req
->rq_phase
== RQ_PHASE_COMPLETE
) {
1584 list_move_tail(&req
->rq_set_chain
, &comp_reqs
);
1588 if (req
->rq_phase
== RQ_PHASE_INTERPRET
)
1592 * Note that this also will start async reply unlink.
1594 if (req
->rq_net_err
&& !req
->rq_timedout
) {
1595 ptlrpc_expire_one_request(req
, 1);
1598 * Check if we still need to wait for unlink.
1600 if (ptlrpc_client_recv_or_unlink(req
) ||
1601 ptlrpc_client_bulk_active(req
))
1603 /* If there is no need to resend, fail it now. */
1604 if (req
->rq_no_resend
) {
1605 if (req
->rq_status
== 0)
1606 req
->rq_status
= -EIO
;
1607 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1615 spin_lock(&req
->rq_lock
);
1616 req
->rq_replied
= 0;
1617 spin_unlock(&req
->rq_lock
);
1618 if (req
->rq_status
== 0)
1619 req
->rq_status
= -EIO
;
1620 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1624 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1625 * so it sets rq_intr regardless of individual rpc
1626 * timeouts. The synchronous IO waiting path sets
1627 * rq_intr irrespective of whether ptlrpcd
1628 * has seen a timeout. Our policy is to only interpret
1629 * interrupted rpcs after they have timed out, so we
1630 * need to enforce that here.
1633 if (req
->rq_intr
&& (req
->rq_timedout
|| req
->rq_waiting
||
1634 req
->rq_wait_ctx
)) {
1635 req
->rq_status
= -EINTR
;
1636 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1640 if (req
->rq_phase
== RQ_PHASE_RPC
) {
1641 if (req
->rq_timedout
|| req
->rq_resend
||
1642 req
->rq_waiting
|| req
->rq_wait_ctx
) {
1645 if (!ptlrpc_unregister_reply(req
, 1))
1648 spin_lock(&imp
->imp_lock
);
1649 if (ptlrpc_import_delay_req(imp
, req
,
1651 /* put on delay list - only if we wait
1652 * recovery finished - before send */
1653 list_del_init(&req
->rq_list
);
1654 list_add_tail(&req
->rq_list
,
1657 spin_unlock(&imp
->imp_lock
);
1662 req
->rq_status
= status
;
1663 ptlrpc_rqphase_move(req
,
1664 RQ_PHASE_INTERPRET
);
1665 spin_unlock(&imp
->imp_lock
);
1668 if (ptlrpc_no_resend(req
) &&
1669 !req
->rq_wait_ctx
) {
1670 req
->rq_status
= -ENOTCONN
;
1671 ptlrpc_rqphase_move(req
,
1672 RQ_PHASE_INTERPRET
);
1673 spin_unlock(&imp
->imp_lock
);
1677 list_del_init(&req
->rq_list
);
1678 list_add_tail(&req
->rq_list
,
1679 &imp
->imp_sending_list
);
1681 spin_unlock(&imp
->imp_lock
);
1683 spin_lock(&req
->rq_lock
);
1684 req
->rq_waiting
= 0;
1685 spin_unlock(&req
->rq_lock
);
1687 if (req
->rq_timedout
|| req
->rq_resend
) {
1688 /* This is re-sending anyways,
1689 * let's mark req as resend. */
1690 spin_lock(&req
->rq_lock
);
1692 spin_unlock(&req
->rq_lock
);
1696 if (!ptlrpc_unregister_bulk(req
, 1))
1699 /* ensure previous bulk fails */
1700 old_xid
= req
->rq_xid
;
1701 req
->rq_xid
= ptlrpc_next_xid();
1702 CDEBUG(D_HA
, "resend bulk old x%llu new x%llu\n",
1703 old_xid
, req
->rq_xid
);
1707 * rq_wait_ctx is only touched by ptlrpcd,
1708 * so no lock is needed here.
1710 status
= sptlrpc_req_refresh_ctx(req
, -1);
1713 req
->rq_status
= status
;
1714 spin_lock(&req
->rq_lock
);
1715 req
->rq_wait_ctx
= 0;
1716 spin_unlock(&req
->rq_lock
);
1717 force_timer_recalc
= 1;
1719 spin_lock(&req
->rq_lock
);
1720 req
->rq_wait_ctx
= 1;
1721 spin_unlock(&req
->rq_lock
);
1726 spin_lock(&req
->rq_lock
);
1727 req
->rq_wait_ctx
= 0;
1728 spin_unlock(&req
->rq_lock
);
1731 rc
= ptl_send_rpc(req
, 0);
1733 DEBUG_REQ(D_HA
, req
,
1734 "send failed: rc = %d", rc
);
1735 force_timer_recalc
= 1;
1736 spin_lock(&req
->rq_lock
);
1737 req
->rq_net_err
= 1;
1738 spin_unlock(&req
->rq_lock
);
1741 /* need to reset the timeout */
1742 force_timer_recalc
= 1;
1745 spin_lock(&req
->rq_lock
);
1747 if (ptlrpc_client_early(req
)) {
1748 ptlrpc_at_recv_early_reply(req
);
1749 spin_unlock(&req
->rq_lock
);
1753 /* Still waiting for a reply? */
1754 if (ptlrpc_client_recv(req
)) {
1755 spin_unlock(&req
->rq_lock
);
1759 /* Did we actually receive a reply? */
1760 if (!ptlrpc_client_replied(req
)) {
1761 spin_unlock(&req
->rq_lock
);
1765 spin_unlock(&req
->rq_lock
);
1767 /* unlink from net because we are going to
1768 * swab in-place of reply buffer */
1769 unregistered
= ptlrpc_unregister_reply(req
, 1);
1773 req
->rq_status
= after_reply(req
);
1777 /* If there is no bulk associated with this request,
1778 * then we're done and should let the interpreter
1779 * process the reply. Similarly if the RPC returned
1780 * an error, and therefore the bulk will never arrive.
1782 if (req
->rq_bulk
== NULL
|| req
->rq_status
< 0) {
1783 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1787 ptlrpc_rqphase_move(req
, RQ_PHASE_BULK
);
1790 LASSERT(req
->rq_phase
== RQ_PHASE_BULK
);
1791 if (ptlrpc_client_bulk_active(req
))
1794 if (req
->rq_bulk
->bd_failure
) {
1795 /* The RPC reply arrived OK, but the bulk screwed
1796 * up! Dead weird since the server told us the RPC
1797 * was good after getting the REPLY for her GET or
1798 * the ACK for her PUT. */
1799 DEBUG_REQ(D_ERROR
, req
, "bulk transfer failed");
1800 req
->rq_status
= -EIO
;
1803 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1806 LASSERT(req
->rq_phase
== RQ_PHASE_INTERPRET
);
1808 /* This moves to "unregistering" phase we need to wait for
1810 if (!unregistered
&& !ptlrpc_unregister_reply(req
, 1)) {
1811 /* start async bulk unlink too */
1812 ptlrpc_unregister_bulk(req
, 1);
1816 if (!ptlrpc_unregister_bulk(req
, 1))
1819 /* When calling interpret receiving already should be
1821 LASSERT(!req
->rq_receiving_reply
);
1823 ptlrpc_req_interpret(env
, req
, req
->rq_status
);
1825 if (ptlrpcd_check_work(req
)) {
1826 atomic_dec(&set
->set_remaining
);
1829 ptlrpc_rqphase_move(req
, RQ_PHASE_COMPLETE
);
1831 CDEBUG(req
->rq_reqmsg
!= NULL
? D_RPCTRACE
: 0,
1832 "Completed RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1833 current_comm(), imp
->imp_obd
->obd_uuid
.uuid
,
1834 lustre_msg_get_status(req
->rq_reqmsg
), req
->rq_xid
,
1835 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
),
1836 lustre_msg_get_opc(req
->rq_reqmsg
));
1838 spin_lock(&imp
->imp_lock
);
1839 /* Request already may be not on sending or delaying list. This
1840 * may happen in the case of marking it erroneous for the case
1841 * ptlrpc_import_delay_req(req, status) find it impossible to
1842 * allow sending this rpc and returns *status != 0. */
1843 if (!list_empty(&req
->rq_list
)) {
1844 list_del_init(&req
->rq_list
);
1845 atomic_dec(&imp
->imp_inflight
);
1847 spin_unlock(&imp
->imp_lock
);
1849 atomic_dec(&set
->set_remaining
);
1850 wake_up_all(&imp
->imp_recovery_waitq
);
1852 if (set
->set_producer
) {
1853 /* produce a new request if possible */
1854 if (ptlrpc_set_producer(set
) > 0)
1855 force_timer_recalc
= 1;
1857 /* free the request that has just been completed
1858 * in order not to pollute set->set_requests */
1859 list_del_init(&req
->rq_set_chain
);
1860 spin_lock(&req
->rq_lock
);
1862 req
->rq_invalid_rqset
= 0;
1863 spin_unlock(&req
->rq_lock
);
1865 /* record rq_status to compute the final status later */
1866 if (req
->rq_status
!= 0)
1867 set
->set_rc
= req
->rq_status
;
1868 ptlrpc_req_finished(req
);
1870 list_move_tail(&req
->rq_set_chain
, &comp_reqs
);
1874 /* move completed request at the head of list so it's easier for
1875 * caller to find them */
1876 list_splice(&comp_reqs
, &set
->set_requests
);
1878 /* If we hit an error, we want to recover promptly. */
1879 return atomic_read(&set
->set_remaining
) == 0 || force_timer_recalc
;
1881 EXPORT_SYMBOL(ptlrpc_check_set
);
1884 * Time out request \a req. is \a async_unlink is set, that means do not wait
1885 * until LNet actually confirms network buffer unlinking.
1886 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1888 int ptlrpc_expire_one_request(struct ptlrpc_request
*req
, int async_unlink
)
1890 struct obd_import
*imp
= req
->rq_import
;
1893 spin_lock(&req
->rq_lock
);
1894 req
->rq_timedout
= 1;
1895 spin_unlock(&req
->rq_lock
);
1897 DEBUG_REQ(D_WARNING
, req
, "Request sent has %s: [sent "CFS_DURATION_T
1898 "/real "CFS_DURATION_T
"]",
1899 req
->rq_net_err
? "failed due to network error" :
1900 ((req
->rq_real_sent
== 0 ||
1901 time_before((unsigned long)req
->rq_real_sent
, (unsigned long)req
->rq_sent
) ||
1902 cfs_time_aftereq(req
->rq_real_sent
, req
->rq_deadline
)) ?
1903 "timed out for sent delay" : "timed out for slow reply"),
1904 req
->rq_sent
, req
->rq_real_sent
);
1906 if (imp
!= NULL
&& obd_debug_peer_on_timeout
)
1907 LNetCtl(IOC_LIBCFS_DEBUG_PEER
, &imp
->imp_connection
->c_peer
);
1909 ptlrpc_unregister_reply(req
, async_unlink
);
1910 ptlrpc_unregister_bulk(req
, async_unlink
);
1912 if (obd_dump_on_timeout
)
1913 libcfs_debug_dumplog();
1916 DEBUG_REQ(D_HA
, req
, "NULL import: already cleaned up?");
1920 atomic_inc(&imp
->imp_timeouts
);
1922 /* The DLM server doesn't want recovery run on its imports. */
1923 if (imp
->imp_dlm_fake
)
1926 /* If this request is for recovery or other primordial tasks,
1927 * then error it out here. */
1928 if (req
->rq_ctx_init
|| req
->rq_ctx_fini
||
1929 req
->rq_send_state
!= LUSTRE_IMP_FULL
||
1930 imp
->imp_obd
->obd_no_recov
) {
1931 DEBUG_REQ(D_RPCTRACE
, req
, "err -110, sent_state=%s (now=%s)",
1932 ptlrpc_import_state_name(req
->rq_send_state
),
1933 ptlrpc_import_state_name(imp
->imp_state
));
1934 spin_lock(&req
->rq_lock
);
1935 req
->rq_status
= -ETIMEDOUT
;
1937 spin_unlock(&req
->rq_lock
);
1941 /* if a request can't be resent we can't wait for an answer after
1943 if (ptlrpc_no_resend(req
)) {
1944 DEBUG_REQ(D_RPCTRACE
, req
, "TIMEOUT-NORESEND:");
1948 ptlrpc_fail_import(imp
, lustre_msg_get_conn_cnt(req
->rq_reqmsg
));
1954 * Time out all uncompleted requests in request set pointed by \a data
1955 * Callback used when waiting on sets with l_wait_event.
1958 int ptlrpc_expired_set(void *data
)
1960 struct ptlrpc_request_set
*set
= data
;
1961 struct list_head
*tmp
;
1962 time_t now
= get_seconds();
1964 LASSERT(set
!= NULL
);
1967 * A timeout expired. See which reqs it applies to...
1969 list_for_each(tmp
, &set
->set_requests
) {
1970 struct ptlrpc_request
*req
=
1971 list_entry(tmp
, struct ptlrpc_request
,
1974 /* don't expire request waiting for context */
1975 if (req
->rq_wait_ctx
)
1978 /* Request in-flight? */
1979 if (!((req
->rq_phase
== RQ_PHASE_RPC
&&
1980 !req
->rq_waiting
&& !req
->rq_resend
) ||
1981 (req
->rq_phase
== RQ_PHASE_BULK
)))
1984 if (req
->rq_timedout
|| /* already dealt with */
1985 req
->rq_deadline
> now
) /* not expired */
1988 /* Deal with this guy. Do it asynchronously to not block
1989 * ptlrpcd thread. */
1990 ptlrpc_expire_one_request(req
, 1);
1994 * When waiting for a whole set, we always break out of the
1995 * sleep so we can recalculate the timeout, or enable interrupts
1996 * if everyone's timed out.
2000 EXPORT_SYMBOL(ptlrpc_expired_set
);
2003 * Sets rq_intr flag in \a req under spinlock.
2005 void ptlrpc_mark_interrupted(struct ptlrpc_request
*req
)
2007 spin_lock(&req
->rq_lock
);
2009 spin_unlock(&req
->rq_lock
);
2011 EXPORT_SYMBOL(ptlrpc_mark_interrupted
);
2014 * Interrupts (sets interrupted flag) all uncompleted requests in
2015 * a set \a data. Callback for l_wait_event for interruptible waits.
2017 void ptlrpc_interrupted_set(void *data
)
2019 struct ptlrpc_request_set
*set
= data
;
2020 struct list_head
*tmp
;
2022 LASSERT(set
!= NULL
);
2023 CDEBUG(D_RPCTRACE
, "INTERRUPTED SET %p\n", set
);
2025 list_for_each(tmp
, &set
->set_requests
) {
2026 struct ptlrpc_request
*req
=
2027 list_entry(tmp
, struct ptlrpc_request
,
2030 if (req
->rq_phase
!= RQ_PHASE_RPC
&&
2031 req
->rq_phase
!= RQ_PHASE_UNREGISTERING
)
2034 ptlrpc_mark_interrupted(req
);
2037 EXPORT_SYMBOL(ptlrpc_interrupted_set
);
2040 * Get the smallest timeout in the set; this does NOT set a timeout.
2042 int ptlrpc_set_next_timeout(struct ptlrpc_request_set
*set
)
2044 struct list_head
*tmp
;
2045 time_t now
= get_seconds();
2047 struct ptlrpc_request
*req
;
2050 list_for_each(tmp
, &set
->set_requests
) {
2051 req
= list_entry(tmp
, struct ptlrpc_request
, rq_set_chain
);
2054 * Request in-flight?
2056 if (!(((req
->rq_phase
== RQ_PHASE_RPC
) && !req
->rq_waiting
) ||
2057 (req
->rq_phase
== RQ_PHASE_BULK
) ||
2058 (req
->rq_phase
== RQ_PHASE_NEW
)))
2062 * Already timed out.
2064 if (req
->rq_timedout
)
2070 if (req
->rq_wait_ctx
)
2073 if (req
->rq_phase
== RQ_PHASE_NEW
)
2074 deadline
= req
->rq_sent
;
2075 else if (req
->rq_phase
== RQ_PHASE_RPC
&& req
->rq_resend
)
2076 deadline
= req
->rq_sent
;
2078 deadline
= req
->rq_sent
+ req
->rq_timeout
;
2080 if (deadline
<= now
) /* actually expired already */
2081 timeout
= 1; /* ASAP */
2082 else if (timeout
== 0 || timeout
> deadline
- now
)
2083 timeout
= deadline
- now
;
2087 EXPORT_SYMBOL(ptlrpc_set_next_timeout
);
2090 * Send all unset request from the set and then wait until all
2091 * requests in the set complete (either get a reply, timeout, get an
2092 * error or otherwise be interrupted).
2093 * Returns 0 on success or error code otherwise.
2095 int ptlrpc_set_wait(struct ptlrpc_request_set
*set
)
2097 struct list_head
*tmp
;
2098 struct ptlrpc_request
*req
;
2099 struct l_wait_info lwi
;
2102 if (set
->set_producer
)
2103 (void)ptlrpc_set_producer(set
);
2105 list_for_each(tmp
, &set
->set_requests
) {
2106 req
= list_entry(tmp
, struct ptlrpc_request
,
2108 if (req
->rq_phase
== RQ_PHASE_NEW
)
2109 (void)ptlrpc_send_new_req(req
);
2112 if (list_empty(&set
->set_requests
))
2116 timeout
= ptlrpc_set_next_timeout(set
);
2118 /* wait until all complete, interrupted, or an in-flight
2120 CDEBUG(D_RPCTRACE
, "set %p going to sleep for %d seconds\n",
2123 if (timeout
== 0 && !cfs_signal_pending())
2125 * No requests are in-flight (ether timed out
2126 * or delayed), so we can allow interrupts.
2127 * We still want to block for a limited time,
2128 * so we allow interrupts during the timeout.
2130 lwi
= LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2132 ptlrpc_interrupted_set
, set
);
2135 * At least one request is in flight, so no
2136 * interrupts are allowed. Wait until all
2137 * complete, or an in-flight req times out.
2139 lwi
= LWI_TIMEOUT(cfs_time_seconds(timeout
? timeout
: 1),
2140 ptlrpc_expired_set
, set
);
2142 rc
= l_wait_event(set
->set_waitq
, ptlrpc_check_set(NULL
, set
), &lwi
);
2144 /* LU-769 - if we ignored the signal because it was already
2145 * pending when we started, we need to handle it now or we risk
2146 * it being ignored forever */
2147 if (rc
== -ETIMEDOUT
&& !lwi
.lwi_allow_intr
&&
2148 cfs_signal_pending()) {
2149 sigset_t blocked_sigs
=
2150 cfs_block_sigsinv(LUSTRE_FATAL_SIGS
);
2152 /* In fact we only interrupt for the "fatal" signals
2153 * like SIGINT or SIGKILL. We still ignore less
2154 * important signals since ptlrpc set is not easily
2155 * reentrant from userspace again */
2156 if (cfs_signal_pending())
2157 ptlrpc_interrupted_set(set
);
2158 cfs_restore_sigs(blocked_sigs
);
2161 LASSERT(rc
== 0 || rc
== -EINTR
|| rc
== -ETIMEDOUT
);
2163 /* -EINTR => all requests have been flagged rq_intr so next
2165 * -ETIMEDOUT => someone timed out. When all reqs have
2166 * timed out, signals are enabled allowing completion with
2168 * I don't really care if we go once more round the loop in
2169 * the error cases -eeb. */
2170 if (rc
== 0 && atomic_read(&set
->set_remaining
) == 0) {
2171 list_for_each(tmp
, &set
->set_requests
) {
2172 req
= list_entry(tmp
, struct ptlrpc_request
,
2174 spin_lock(&req
->rq_lock
);
2175 req
->rq_invalid_rqset
= 1;
2176 spin_unlock(&req
->rq_lock
);
2179 } while (rc
!= 0 || atomic_read(&set
->set_remaining
) != 0);
2181 LASSERT(atomic_read(&set
->set_remaining
) == 0);
2183 rc
= set
->set_rc
; /* rq_status of already freed requests if any */
2184 list_for_each(tmp
, &set
->set_requests
) {
2185 req
= list_entry(tmp
, struct ptlrpc_request
, rq_set_chain
);
2187 LASSERT(req
->rq_phase
== RQ_PHASE_COMPLETE
);
2188 if (req
->rq_status
!= 0)
2189 rc
= req
->rq_status
;
2192 if (set
->set_interpret
!= NULL
) {
2193 int (*interpreter
)(struct ptlrpc_request_set
*set
, void *, int) =
2195 rc
= interpreter (set
, set
->set_arg
, rc
);
2197 struct ptlrpc_set_cbdata
*cbdata
, *n
;
2200 list_for_each_entry_safe(cbdata
, n
,
2201 &set
->set_cblist
, psc_item
) {
2202 list_del_init(&cbdata
->psc_item
);
2203 err
= cbdata
->psc_interpret(set
, cbdata
->psc_data
, rc
);
2206 OBD_FREE_PTR(cbdata
);
2212 EXPORT_SYMBOL(ptlrpc_set_wait
);
2215 * Helper function for request freeing.
2216 * Called when request count reached zero and request needs to be freed.
2217 * Removes request from all sorts of sending/replay lists it might be on,
2218 * frees network buffers if any are present.
2219 * If \a locked is set, that means caller is already holding import imp_lock
2220 * and so we no longer need to reobtain it (for certain lists manipulations)
2222 static void __ptlrpc_free_req(struct ptlrpc_request
*request
, int locked
)
2224 if (request
== NULL
)
2226 LASSERTF(!request
->rq_receiving_reply
, "req %p\n", request
);
2227 LASSERTF(request
->rq_rqbd
== NULL
, "req %p\n", request
);/* client-side */
2228 LASSERTF(list_empty(&request
->rq_list
), "req %p\n", request
);
2229 LASSERTF(list_empty(&request
->rq_set_chain
), "req %p\n", request
);
2230 LASSERTF(list_empty(&request
->rq_exp_list
), "req %p\n", request
);
2231 LASSERTF(!request
->rq_replay
, "req %p\n", request
);
2233 req_capsule_fini(&request
->rq_pill
);
2235 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2236 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2237 if (request
->rq_import
!= NULL
) {
2239 spin_lock(&request
->rq_import
->imp_lock
);
2240 list_del_init(&request
->rq_replay_list
);
2242 spin_unlock(&request
->rq_import
->imp_lock
);
2244 LASSERTF(list_empty(&request
->rq_replay_list
), "req %p\n", request
);
2246 if (atomic_read(&request
->rq_refcount
) != 0) {
2247 DEBUG_REQ(D_ERROR
, request
,
2248 "freeing request with nonzero refcount");
2252 if (request
->rq_repbuf
!= NULL
)
2253 sptlrpc_cli_free_repbuf(request
);
2254 if (request
->rq_export
!= NULL
) {
2255 class_export_put(request
->rq_export
);
2256 request
->rq_export
= NULL
;
2258 if (request
->rq_import
!= NULL
) {
2259 class_import_put(request
->rq_import
);
2260 request
->rq_import
= NULL
;
2262 if (request
->rq_bulk
!= NULL
)
2263 ptlrpc_free_bulk_pin(request
->rq_bulk
);
2265 if (request
->rq_reqbuf
!= NULL
|| request
->rq_clrbuf
!= NULL
)
2266 sptlrpc_cli_free_reqbuf(request
);
2268 if (request
->rq_cli_ctx
)
2269 sptlrpc_req_put_ctx(request
, !locked
);
2271 if (request
->rq_pool
)
2272 __ptlrpc_free_req_to_pool(request
);
2274 ptlrpc_request_cache_free(request
);
2277 static int __ptlrpc_req_finished(struct ptlrpc_request
*request
, int locked
);
2279 * Drop one request reference. Must be called with import imp_lock held.
2280 * When reference count drops to zero, request is freed.
2282 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request
*request
)
2284 assert_spin_locked(&request
->rq_import
->imp_lock
);
2285 (void)__ptlrpc_req_finished(request
, 1);
2287 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock
);
2291 * Drops one reference count for request \a request.
2292 * \a locked set indicates that caller holds import imp_lock.
2293 * Frees the request when reference count reaches zero.
2295 static int __ptlrpc_req_finished(struct ptlrpc_request
*request
, int locked
)
2297 if (request
== NULL
)
2300 if (request
== LP_POISON
||
2301 request
->rq_reqmsg
== LP_POISON
) {
2302 CERROR("dereferencing freed request (bug 575)\n");
2307 DEBUG_REQ(D_INFO
, request
, "refcount now %u",
2308 atomic_read(&request
->rq_refcount
) - 1);
2310 if (atomic_dec_and_test(&request
->rq_refcount
)) {
2311 __ptlrpc_free_req(request
, locked
);
2319 * Drops one reference count for a request.
2321 void ptlrpc_req_finished(struct ptlrpc_request
*request
)
2323 __ptlrpc_req_finished(request
, 0);
2325 EXPORT_SYMBOL(ptlrpc_req_finished
);
2328 * Returns xid of a \a request
2330 __u64
ptlrpc_req_xid(struct ptlrpc_request
*request
)
2332 return request
->rq_xid
;
2334 EXPORT_SYMBOL(ptlrpc_req_xid
);
2337 * Disengage the client's reply buffer from the network
2338 * NB does _NOT_ unregister any client-side bulk.
2339 * IDEMPOTENT, but _not_ safe against concurrent callers.
2340 * The request owner (i.e. the thread doing the I/O) must call...
2341 * Returns 0 on success or 1 if unregistering cannot be made.
2343 int ptlrpc_unregister_reply(struct ptlrpc_request
*request
, int async
)
2346 wait_queue_head_t
*wq
;
2347 struct l_wait_info lwi
;
2352 LASSERT(!in_interrupt());
2355 * Let's setup deadline for reply unlink.
2357 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2358 async
&& request
->rq_reply_deadline
== 0)
2359 request
->rq_reply_deadline
= get_seconds()+LONG_UNLINK
;
2362 * Nothing left to do.
2364 if (!ptlrpc_client_recv_or_unlink(request
))
2367 LNetMDUnlink(request
->rq_reply_md_h
);
2370 * Let's check it once again.
2372 if (!ptlrpc_client_recv_or_unlink(request
))
2376 * Move to "Unregistering" phase as reply was not unlinked yet.
2378 ptlrpc_rqphase_move(request
, RQ_PHASE_UNREGISTERING
);
2381 * Do not wait for unlink to finish.
2387 * We have to l_wait_event() whatever the result, to give liblustre
2388 * a chance to run reply_in_callback(), and to make sure we've
2389 * unlinked before returning a req to the pool.
2391 if (request
->rq_set
!= NULL
)
2392 wq
= &request
->rq_set
->set_waitq
;
2394 wq
= &request
->rq_reply_waitq
;
2397 /* Network access will complete in finite time but the HUGE
2398 * timeout lets us CWARN for visibility of sluggish NALs */
2399 lwi
= LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK
),
2400 cfs_time_seconds(1), NULL
, NULL
);
2401 rc
= l_wait_event(*wq
, !ptlrpc_client_recv_or_unlink(request
),
2404 ptlrpc_rqphase_move(request
, request
->rq_next_phase
);
2408 LASSERT(rc
== -ETIMEDOUT
);
2409 DEBUG_REQ(D_WARNING
, request
,
2410 "Unexpectedly long timeout rvcng=%d unlnk=%d/%d",
2411 request
->rq_receiving_reply
,
2412 request
->rq_req_unlink
, request
->rq_reply_unlink
);
2416 EXPORT_SYMBOL(ptlrpc_unregister_reply
);
2418 static void ptlrpc_free_request(struct ptlrpc_request
*req
)
2420 spin_lock(&req
->rq_lock
);
2422 spin_unlock(&req
->rq_lock
);
2424 if (req
->rq_commit_cb
!= NULL
)
2425 req
->rq_commit_cb(req
);
2426 list_del_init(&req
->rq_replay_list
);
2428 __ptlrpc_req_finished(req
, 1);
2432 * the request is committed and dropped from the replay list of its import
2434 void ptlrpc_request_committed(struct ptlrpc_request
*req
, int force
)
2436 struct obd_import
*imp
= req
->rq_import
;
2438 spin_lock(&imp
->imp_lock
);
2439 if (list_empty(&req
->rq_replay_list
)) {
2440 spin_unlock(&imp
->imp_lock
);
2444 if (force
|| req
->rq_transno
<= imp
->imp_peer_committed_transno
)
2445 ptlrpc_free_request(req
);
2447 spin_unlock(&imp
->imp_lock
);
2449 EXPORT_SYMBOL(ptlrpc_request_committed
);
2452 * Iterates through replay_list on import and prunes
2453 * all requests have transno smaller than last_committed for the
2454 * import and don't have rq_replay set.
2455 * Since requests are sorted in transno order, stops when meeting first
2456 * transno bigger than last_committed.
2457 * caller must hold imp->imp_lock
2459 void ptlrpc_free_committed(struct obd_import
*imp
)
2461 struct ptlrpc_request
*req
, *saved
;
2462 struct ptlrpc_request
*last_req
= NULL
; /* temporary fire escape */
2463 bool skip_committed_list
= true;
2465 LASSERT(imp
!= NULL
);
2466 assert_spin_locked(&imp
->imp_lock
);
2468 if (imp
->imp_peer_committed_transno
== imp
->imp_last_transno_checked
&&
2469 imp
->imp_generation
== imp
->imp_last_generation_checked
) {
2470 CDEBUG(D_INFO
, "%s: skip recheck: last_committed %llu\n",
2471 imp
->imp_obd
->obd_name
, imp
->imp_peer_committed_transno
);
2474 CDEBUG(D_RPCTRACE
, "%s: committing for last_committed %llu gen %d\n",
2475 imp
->imp_obd
->obd_name
, imp
->imp_peer_committed_transno
,
2476 imp
->imp_generation
);
2478 if (imp
->imp_generation
!= imp
->imp_last_generation_checked
)
2479 skip_committed_list
= false;
2481 imp
->imp_last_transno_checked
= imp
->imp_peer_committed_transno
;
2482 imp
->imp_last_generation_checked
= imp
->imp_generation
;
2484 list_for_each_entry_safe(req
, saved
, &imp
->imp_replay_list
,
2486 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2487 LASSERT(req
!= last_req
);
2490 if (req
->rq_transno
== 0) {
2491 DEBUG_REQ(D_EMERG
, req
, "zero transno during replay");
2494 if (req
->rq_import_generation
< imp
->imp_generation
) {
2495 DEBUG_REQ(D_RPCTRACE
, req
, "free request with old gen");
2499 /* not yet committed */
2500 if (req
->rq_transno
> imp
->imp_peer_committed_transno
) {
2501 DEBUG_REQ(D_RPCTRACE
, req
, "stopping search");
2505 if (req
->rq_replay
) {
2506 DEBUG_REQ(D_RPCTRACE
, req
, "keeping (FL_REPLAY)");
2507 list_move_tail(&req
->rq_replay_list
,
2508 &imp
->imp_committed_list
);
2512 DEBUG_REQ(D_INFO
, req
, "commit (last_committed %llu)",
2513 imp
->imp_peer_committed_transno
);
2515 ptlrpc_free_request(req
);
2517 if (skip_committed_list
)
2520 list_for_each_entry_safe(req
, saved
, &imp
->imp_committed_list
,
2522 LASSERT(req
->rq_transno
!= 0);
2523 if (req
->rq_import_generation
< imp
->imp_generation
) {
2524 DEBUG_REQ(D_RPCTRACE
, req
, "free stale open request");
2525 ptlrpc_free_request(req
);
2530 void ptlrpc_cleanup_client(struct obd_import
*imp
)
2533 EXPORT_SYMBOL(ptlrpc_cleanup_client
);
2536 * Schedule previously sent request for resend.
2537 * For bulk requests we assign new xid (to avoid problems with
2538 * lost replies and therefore several transfers landing into same buffer
2539 * from different sending attempts).
2541 void ptlrpc_resend_req(struct ptlrpc_request
*req
)
2543 DEBUG_REQ(D_HA
, req
, "going to resend");
2544 spin_lock(&req
->rq_lock
);
2546 /* Request got reply but linked to the import list still.
2547 Let ptlrpc_check_set() to process it. */
2548 if (ptlrpc_client_replied(req
)) {
2549 spin_unlock(&req
->rq_lock
);
2550 DEBUG_REQ(D_HA
, req
, "it has reply, so skip it");
2554 lustre_msg_set_handle(req
->rq_reqmsg
, &(struct lustre_handle
){ 0 });
2555 req
->rq_status
= -EAGAIN
;
2558 req
->rq_net_err
= 0;
2559 req
->rq_timedout
= 0;
2561 __u64 old_xid
= req
->rq_xid
;
2563 /* ensure previous bulk fails */
2564 req
->rq_xid
= ptlrpc_next_xid();
2565 CDEBUG(D_HA
, "resend bulk old x%llu new x%llu\n",
2566 old_xid
, req
->rq_xid
);
2568 ptlrpc_client_wake_req(req
);
2569 spin_unlock(&req
->rq_lock
);
2571 EXPORT_SYMBOL(ptlrpc_resend_req
);
2573 /* XXX: this function and rq_status are currently unused */
2574 void ptlrpc_restart_req(struct ptlrpc_request
*req
)
2576 DEBUG_REQ(D_HA
, req
, "restarting (possibly-)completed request");
2577 req
->rq_status
= -ERESTARTSYS
;
2579 spin_lock(&req
->rq_lock
);
2580 req
->rq_restart
= 1;
2581 req
->rq_timedout
= 0;
2582 ptlrpc_client_wake_req(req
);
2583 spin_unlock(&req
->rq_lock
);
2585 EXPORT_SYMBOL(ptlrpc_restart_req
);
2588 * Grab additional reference on a request \a req
2590 struct ptlrpc_request
*ptlrpc_request_addref(struct ptlrpc_request
*req
)
2592 atomic_inc(&req
->rq_refcount
);
2595 EXPORT_SYMBOL(ptlrpc_request_addref
);
2598 * Add a request to import replay_list.
2599 * Must be called under imp_lock
2601 void ptlrpc_retain_replayable_request(struct ptlrpc_request
*req
,
2602 struct obd_import
*imp
)
2604 struct list_head
*tmp
;
2606 assert_spin_locked(&imp
->imp_lock
);
2608 if (req
->rq_transno
== 0) {
2609 DEBUG_REQ(D_EMERG
, req
, "saving request with zero transno");
2613 /* clear this for new requests that were resent as well
2614 as resent replayed requests. */
2615 lustre_msg_clear_flags(req
->rq_reqmsg
, MSG_RESENT
);
2617 /* don't re-add requests that have been replayed */
2618 if (!list_empty(&req
->rq_replay_list
))
2621 lustre_msg_add_flags(req
->rq_reqmsg
, MSG_REPLAY
);
2623 LASSERT(imp
->imp_replayable
);
2624 /* Balanced in ptlrpc_free_committed, usually. */
2625 ptlrpc_request_addref(req
);
2626 list_for_each_prev(tmp
, &imp
->imp_replay_list
) {
2627 struct ptlrpc_request
*iter
=
2628 list_entry(tmp
, struct ptlrpc_request
,
2631 /* We may have duplicate transnos if we create and then
2632 * open a file, or for closes retained if to match creating
2633 * opens, so use req->rq_xid as a secondary key.
2634 * (See bugs 684, 685, and 428.)
2635 * XXX no longer needed, but all opens need transnos!
2637 if (iter
->rq_transno
> req
->rq_transno
)
2640 if (iter
->rq_transno
== req
->rq_transno
) {
2641 LASSERT(iter
->rq_xid
!= req
->rq_xid
);
2642 if (iter
->rq_xid
> req
->rq_xid
)
2646 list_add(&req
->rq_replay_list
, &iter
->rq_replay_list
);
2650 list_add(&req
->rq_replay_list
, &imp
->imp_replay_list
);
2652 EXPORT_SYMBOL(ptlrpc_retain_replayable_request
);
2655 * Send request and wait until it completes.
2656 * Returns request processing status.
2658 int ptlrpc_queue_wait(struct ptlrpc_request
*req
)
2660 struct ptlrpc_request_set
*set
;
2663 LASSERT(req
->rq_set
== NULL
);
2664 LASSERT(!req
->rq_receiving_reply
);
2666 set
= ptlrpc_prep_set();
2668 CERROR("Unable to allocate ptlrpc set.");
2672 /* for distributed debugging */
2673 lustre_msg_set_status(req
->rq_reqmsg
, current_pid());
2675 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2676 ptlrpc_request_addref(req
);
2677 ptlrpc_set_add_req(set
, req
);
2678 rc
= ptlrpc_set_wait(set
);
2679 ptlrpc_set_destroy(set
);
2683 EXPORT_SYMBOL(ptlrpc_queue_wait
);
2685 struct ptlrpc_replay_async_args
{
2687 int praa_old_status
;
2691 * Callback used for replayed requests reply processing.
2692 * In case of successful reply calls registered request replay callback.
2693 * In case of error restart replay process.
2695 static int ptlrpc_replay_interpret(const struct lu_env
*env
,
2696 struct ptlrpc_request
*req
,
2699 struct ptlrpc_replay_async_args
*aa
= data
;
2700 struct obd_import
*imp
= req
->rq_import
;
2702 atomic_dec(&imp
->imp_replay_inflight
);
2704 if (!ptlrpc_client_replied(req
)) {
2705 CERROR("request replay timed out, restarting recovery\n");
2710 if (lustre_msg_get_type(req
->rq_repmsg
) == PTL_RPC_MSG_ERR
&&
2711 (lustre_msg_get_status(req
->rq_repmsg
) == -ENOTCONN
||
2712 lustre_msg_get_status(req
->rq_repmsg
) == -ENODEV
)) {
2713 rc
= lustre_msg_get_status(req
->rq_repmsg
);
2717 /** VBR: check version failure */
2718 if (lustre_msg_get_status(req
->rq_repmsg
) == -EOVERFLOW
) {
2719 /** replay was failed due to version mismatch */
2720 DEBUG_REQ(D_WARNING
, req
, "Version mismatch during replay\n");
2721 spin_lock(&imp
->imp_lock
);
2722 imp
->imp_vbr_failed
= 1;
2723 imp
->imp_no_lock_replay
= 1;
2724 spin_unlock(&imp
->imp_lock
);
2725 lustre_msg_set_status(req
->rq_repmsg
, aa
->praa_old_status
);
2727 /** The transno had better not change over replay. */
2728 LASSERTF(lustre_msg_get_transno(req
->rq_reqmsg
) ==
2729 lustre_msg_get_transno(req
->rq_repmsg
) ||
2730 lustre_msg_get_transno(req
->rq_repmsg
) == 0,
2732 lustre_msg_get_transno(req
->rq_reqmsg
),
2733 lustre_msg_get_transno(req
->rq_repmsg
));
2736 spin_lock(&imp
->imp_lock
);
2737 /** if replays by version then gap occur on server, no trust to locks */
2738 if (lustre_msg_get_flags(req
->rq_repmsg
) & MSG_VERSION_REPLAY
)
2739 imp
->imp_no_lock_replay
= 1;
2740 imp
->imp_last_replay_transno
= lustre_msg_get_transno(req
->rq_reqmsg
);
2741 spin_unlock(&imp
->imp_lock
);
2742 LASSERT(imp
->imp_last_replay_transno
);
2744 /* transaction number shouldn't be bigger than the latest replayed */
2745 if (req
->rq_transno
> lustre_msg_get_transno(req
->rq_reqmsg
)) {
2746 DEBUG_REQ(D_ERROR
, req
,
2747 "Reported transno %llu is bigger than the replayed one: %llu",
2749 lustre_msg_get_transno(req
->rq_reqmsg
));
2754 DEBUG_REQ(D_HA
, req
, "got rep");
2756 /* let the callback do fixups, possibly including in the request */
2757 if (req
->rq_replay_cb
)
2758 req
->rq_replay_cb(req
);
2760 if (ptlrpc_client_replied(req
) &&
2761 lustre_msg_get_status(req
->rq_repmsg
) != aa
->praa_old_status
) {
2762 DEBUG_REQ(D_ERROR
, req
, "status %d, old was %d",
2763 lustre_msg_get_status(req
->rq_repmsg
),
2764 aa
->praa_old_status
);
2766 /* Put it back for re-replay. */
2767 lustre_msg_set_status(req
->rq_repmsg
, aa
->praa_old_status
);
2771 * Errors while replay can set transno to 0, but
2772 * imp_last_replay_transno shouldn't be set to 0 anyway
2774 if (req
->rq_transno
== 0)
2775 CERROR("Transno is 0 during replay!\n");
2777 /* continue with recovery */
2778 rc
= ptlrpc_import_recovery_state_machine(imp
);
2780 req
->rq_send_state
= aa
->praa_old_state
;
2783 /* this replay failed, so restart recovery */
2784 ptlrpc_connect_import(imp
);
2790 * Prepares and queues request for replay.
2791 * Adds it to ptlrpcd queue for actual sending.
2792 * Returns 0 on success.
2794 int ptlrpc_replay_req(struct ptlrpc_request
*req
)
2796 struct ptlrpc_replay_async_args
*aa
;
2798 LASSERT(req
->rq_import
->imp_state
== LUSTRE_IMP_REPLAY
);
2800 LASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2801 aa
= ptlrpc_req_async_args(req
);
2802 memset(aa
, 0, sizeof(*aa
));
2804 /* Prepare request to be resent with ptlrpcd */
2805 aa
->praa_old_state
= req
->rq_send_state
;
2806 req
->rq_send_state
= LUSTRE_IMP_REPLAY
;
2807 req
->rq_phase
= RQ_PHASE_NEW
;
2808 req
->rq_next_phase
= RQ_PHASE_UNDEFINED
;
2810 aa
->praa_old_status
= lustre_msg_get_status(req
->rq_repmsg
);
2812 req
->rq_interpret_reply
= ptlrpc_replay_interpret
;
2813 /* Readjust the timeout for current conditions */
2814 ptlrpc_at_set_req_timeout(req
);
2816 /* Tell server the net_latency, so the server can calculate how long
2817 * it should wait for next replay */
2818 lustre_msg_set_service_time(req
->rq_reqmsg
,
2819 ptlrpc_at_get_net_latency(req
));
2820 DEBUG_REQ(D_HA
, req
, "REPLAY");
2822 atomic_inc(&req
->rq_import
->imp_replay_inflight
);
2823 ptlrpc_request_addref(req
); /* ptlrpcd needs a ref */
2825 ptlrpcd_add_req(req
, PDL_POLICY_LOCAL
, -1);
2828 EXPORT_SYMBOL(ptlrpc_replay_req
);
2831 * Aborts all in-flight request on import \a imp sending and delayed lists
2833 void ptlrpc_abort_inflight(struct obd_import
*imp
)
2835 struct list_head
*tmp
, *n
;
2837 /* Make sure that no new requests get processed for this import.
2838 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2839 * this flag and then putting requests on sending_list or delayed_list.
2841 spin_lock(&imp
->imp_lock
);
2843 /* XXX locking? Maybe we should remove each request with the list
2844 * locked? Also, how do we know if the requests on the list are
2845 * being freed at this time?
2847 list_for_each_safe(tmp
, n
, &imp
->imp_sending_list
) {
2848 struct ptlrpc_request
*req
=
2849 list_entry(tmp
, struct ptlrpc_request
, rq_list
);
2851 DEBUG_REQ(D_RPCTRACE
, req
, "inflight");
2853 spin_lock(&req
->rq_lock
);
2854 if (req
->rq_import_generation
< imp
->imp_generation
) {
2856 req
->rq_status
= -EIO
;
2857 ptlrpc_client_wake_req(req
);
2859 spin_unlock(&req
->rq_lock
);
2862 list_for_each_safe(tmp
, n
, &imp
->imp_delayed_list
) {
2863 struct ptlrpc_request
*req
=
2864 list_entry(tmp
, struct ptlrpc_request
, rq_list
);
2866 DEBUG_REQ(D_RPCTRACE
, req
, "aborting waiting req");
2868 spin_lock(&req
->rq_lock
);
2869 if (req
->rq_import_generation
< imp
->imp_generation
) {
2871 req
->rq_status
= -EIO
;
2872 ptlrpc_client_wake_req(req
);
2874 spin_unlock(&req
->rq_lock
);
2877 /* Last chance to free reqs left on the replay list, but we
2878 * will still leak reqs that haven't committed. */
2879 if (imp
->imp_replayable
)
2880 ptlrpc_free_committed(imp
);
2882 spin_unlock(&imp
->imp_lock
);
2884 EXPORT_SYMBOL(ptlrpc_abort_inflight
);
2887 * Abort all uncompleted requests in request set \a set
2889 void ptlrpc_abort_set(struct ptlrpc_request_set
*set
)
2891 struct list_head
*tmp
, *pos
;
2893 LASSERT(set
!= NULL
);
2895 list_for_each_safe(pos
, tmp
, &set
->set_requests
) {
2896 struct ptlrpc_request
*req
=
2897 list_entry(pos
, struct ptlrpc_request
,
2900 spin_lock(&req
->rq_lock
);
2901 if (req
->rq_phase
!= RQ_PHASE_RPC
) {
2902 spin_unlock(&req
->rq_lock
);
2907 req
->rq_status
= -EINTR
;
2908 ptlrpc_client_wake_req(req
);
2909 spin_unlock(&req
->rq_lock
);
2913 static __u64 ptlrpc_last_xid
;
2914 static spinlock_t ptlrpc_last_xid_lock
;
2917 * Initialize the XID for the node. This is common among all requests on
2918 * this node, and only requires the property that it is monotonically
2919 * increasing. It does not need to be sequential. Since this is also used
2920 * as the RDMA match bits, it is important that a single client NOT have
2921 * the same match bits for two different in-flight requests, hence we do
2922 * NOT want to have an XID per target or similar.
2924 * To avoid an unlikely collision between match bits after a client reboot
2925 * (which would deliver old data into the wrong RDMA buffer) initialize
2926 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2927 * If the time is clearly incorrect, we instead use a 62-bit random number.
2928 * In the worst case the random number will overflow 1M RPCs per second in
2929 * 9133 years, or permutations thereof.
2931 #define YEAR_2004 (1ULL << 30)
2932 void ptlrpc_init_xid(void)
2934 time_t now
= get_seconds();
2936 spin_lock_init(&ptlrpc_last_xid_lock
);
2937 if (now
< YEAR_2004
) {
2938 cfs_get_random_bytes(&ptlrpc_last_xid
, sizeof(ptlrpc_last_xid
));
2939 ptlrpc_last_xid
>>= 2;
2940 ptlrpc_last_xid
|= (1ULL << 61);
2942 ptlrpc_last_xid
= (__u64
)now
<< 20;
2945 /* Always need to be aligned to a power-of-two for multi-bulk BRW */
2946 CLASSERT((PTLRPC_BULK_OPS_COUNT
& (PTLRPC_BULK_OPS_COUNT
- 1)) == 0);
2947 ptlrpc_last_xid
&= PTLRPC_BULK_OPS_MASK
;
2951 * Increase xid and returns resulting new value to the caller.
2953 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2954 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2955 * itself uses the last bulk xid needed, so the server can determine the
2956 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2957 * xid must align to a power-of-two value.
2959 * This is assumed to be true due to the initial ptlrpc_last_xid
2960 * value also being initialized to a power-of-two value. LU-1431
2962 __u64
ptlrpc_next_xid(void)
2966 spin_lock(&ptlrpc_last_xid_lock
);
2967 next
= ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
2968 ptlrpc_last_xid
= next
;
2969 spin_unlock(&ptlrpc_last_xid_lock
);
2973 EXPORT_SYMBOL(ptlrpc_next_xid
);
2976 * Get a glimpse at what next xid value might have been.
2977 * Returns possible next xid.
2979 __u64
ptlrpc_sample_next_xid(void)
2981 #if BITS_PER_LONG == 32
2982 /* need to avoid possible word tearing on 32-bit systems */
2985 spin_lock(&ptlrpc_last_xid_lock
);
2986 next
= ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
2987 spin_unlock(&ptlrpc_last_xid_lock
);
2991 /* No need to lock, since returned value is racy anyways */
2992 return ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
2995 EXPORT_SYMBOL(ptlrpc_sample_next_xid
);
2998 * Functions for operating ptlrpc workers.
3000 * A ptlrpc work is a function which will be running inside ptlrpc context.
3001 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3003 * 1. after a work is created, it can be used many times, that is:
3004 * handler = ptlrpcd_alloc_work();
3005 * ptlrpcd_queue_work();
3007 * queue it again when necessary:
3008 * ptlrpcd_queue_work();
3009 * ptlrpcd_destroy_work();
3010 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3011 * it will only be queued once in any time. Also as its name implies, it may
3012 * have delay before it really runs by ptlrpcd thread.
3014 struct ptlrpc_work_async_args
{
3015 int (*cb
)(const struct lu_env
*, void *);
3019 static void ptlrpcd_add_work_req(struct ptlrpc_request
*req
)
3021 /* re-initialize the req */
3022 req
->rq_timeout
= obd_timeout
;
3023 req
->rq_sent
= get_seconds();
3024 req
->rq_deadline
= req
->rq_sent
+ req
->rq_timeout
;
3025 req
->rq_reply_deadline
= req
->rq_deadline
;
3026 req
->rq_phase
= RQ_PHASE_INTERPRET
;
3027 req
->rq_next_phase
= RQ_PHASE_COMPLETE
;
3028 req
->rq_xid
= ptlrpc_next_xid();
3029 req
->rq_import_generation
= req
->rq_import
->imp_generation
;
3031 ptlrpcd_add_req(req
, PDL_POLICY_ROUND
, -1);
3034 static int work_interpreter(const struct lu_env
*env
,
3035 struct ptlrpc_request
*req
, void *data
, int rc
)
3037 struct ptlrpc_work_async_args
*arg
= data
;
3039 LASSERT(ptlrpcd_check_work(req
));
3040 LASSERT(arg
->cb
!= NULL
);
3042 rc
= arg
->cb(env
, arg
->cbdata
);
3044 list_del_init(&req
->rq_set_chain
);
3047 if (atomic_dec_return(&req
->rq_refcount
) > 1) {
3048 atomic_set(&req
->rq_refcount
, 2);
3049 ptlrpcd_add_work_req(req
);
3054 static int worker_format
;
3056 static int ptlrpcd_check_work(struct ptlrpc_request
*req
)
3058 return req
->rq_pill
.rc_fmt
== (void *)&worker_format
;
3062 * Create a work for ptlrpc.
3064 void *ptlrpcd_alloc_work(struct obd_import
*imp
,
3065 int (*cb
)(const struct lu_env
*, void *), void *cbdata
)
3067 struct ptlrpc_request
*req
= NULL
;
3068 struct ptlrpc_work_async_args
*args
;
3073 return ERR_PTR(-EINVAL
);
3075 /* copy some code from deprecated fakereq. */
3076 req
= ptlrpc_request_cache_alloc(GFP_NOFS
);
3078 CERROR("ptlrpc: run out of memory!\n");
3079 return ERR_PTR(-ENOMEM
);
3082 req
->rq_send_state
= LUSTRE_IMP_FULL
;
3083 req
->rq_type
= PTL_RPC_MSG_REQUEST
;
3084 req
->rq_import
= class_import_get(imp
);
3085 req
->rq_export
= NULL
;
3086 req
->rq_interpret_reply
= work_interpreter
;
3087 /* don't want reply */
3088 req
->rq_receiving_reply
= 0;
3089 req
->rq_req_unlink
= req
->rq_reply_unlink
= 0;
3090 req
->rq_no_delay
= req
->rq_no_resend
= 1;
3091 req
->rq_pill
.rc_fmt
= (void *)&worker_format
;
3093 spin_lock_init(&req
->rq_lock
);
3094 INIT_LIST_HEAD(&req
->rq_list
);
3095 INIT_LIST_HEAD(&req
->rq_replay_list
);
3096 INIT_LIST_HEAD(&req
->rq_set_chain
);
3097 INIT_LIST_HEAD(&req
->rq_history_list
);
3098 INIT_LIST_HEAD(&req
->rq_exp_list
);
3099 init_waitqueue_head(&req
->rq_reply_waitq
);
3100 init_waitqueue_head(&req
->rq_set_waitq
);
3101 atomic_set(&req
->rq_refcount
, 1);
3103 CLASSERT(sizeof(*args
) <= sizeof(req
->rq_async_args
));
3104 args
= ptlrpc_req_async_args(req
);
3106 args
->cbdata
= cbdata
;
3110 EXPORT_SYMBOL(ptlrpcd_alloc_work
);
3112 void ptlrpcd_destroy_work(void *handler
)
3114 struct ptlrpc_request
*req
= handler
;
3117 ptlrpc_req_finished(req
);
3119 EXPORT_SYMBOL(ptlrpcd_destroy_work
);
3121 int ptlrpcd_queue_work(void *handler
)
3123 struct ptlrpc_request
*req
= handler
;
3126 * Check if the req is already being queued.
3128 * Here comes a trick: it lacks a way of checking if a req is being
3129 * processed reliably in ptlrpc. Here I have to use refcount of req
3130 * for this purpose. This is okay because the caller should use this
3131 * req as opaque data. - Jinshan
3133 LASSERT(atomic_read(&req
->rq_refcount
) > 0);
3134 if (atomic_inc_return(&req
->rq_refcount
) == 2)
3135 ptlrpcd_add_work_req(req
);
3138 EXPORT_SYMBOL(ptlrpcd_queue_work
);