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 desc
= kzalloc(offsetof(struct ptlrpc_bulk_desc
, bd_iov
[npages
]),
111 spin_lock_init(&desc
->bd_lock
);
112 init_waitqueue_head(&desc
->bd_waitq
);
113 desc
->bd_max_iov
= npages
;
114 desc
->bd_iov_count
= 0;
115 desc
->bd_portal
= portal
;
116 desc
->bd_type
= type
;
117 desc
->bd_md_count
= 0;
118 LASSERT(max_brw
> 0);
119 desc
->bd_md_max_brw
= min(max_brw
, PTLRPC_BULK_OPS_COUNT
);
120 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
121 * node. Negotiated ocd_brw_size will always be <= this number. */
122 for (i
= 0; i
< PTLRPC_BULK_OPS_COUNT
; i
++)
123 LNetInvalidateHandle(&desc
->bd_mds
[i
]);
129 * Prepare bulk descriptor for specified outgoing request \a req that
130 * can fit \a npages * pages. \a type is bulk type. \a portal is where
131 * the bulk to be sent. Used on client-side.
132 * Returns pointer to newly allocated initialized bulk descriptor or NULL on
135 struct ptlrpc_bulk_desc
*ptlrpc_prep_bulk_imp(struct ptlrpc_request
*req
,
136 unsigned npages
, unsigned max_brw
,
137 unsigned type
, unsigned portal
)
139 struct obd_import
*imp
= req
->rq_import
;
140 struct ptlrpc_bulk_desc
*desc
;
142 LASSERT(type
== BULK_PUT_SINK
|| type
== BULK_GET_SOURCE
);
143 desc
= ptlrpc_new_bulk(npages
, max_brw
, type
, portal
);
147 desc
->bd_import_generation
= req
->rq_import_generation
;
148 desc
->bd_import
= class_import_get(imp
);
151 desc
->bd_cbid
.cbid_fn
= client_bulk_callback
;
152 desc
->bd_cbid
.cbid_arg
= desc
;
154 /* This makes req own desc, and free it when she frees herself */
159 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp
);
162 * Add a page \a page to the bulk descriptor \a desc.
163 * Data to transfer in the page starts at offset \a pageoffset and
164 * amount of data to transfer from the page is \a len
166 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc
*desc
,
167 struct page
*page
, int pageoffset
, int len
, int pin
)
169 LASSERT(desc
->bd_iov_count
< desc
->bd_max_iov
);
170 LASSERT(page
!= NULL
);
171 LASSERT(pageoffset
>= 0);
173 LASSERT(pageoffset
+ len
<= PAGE_CACHE_SIZE
);
178 page_cache_get(page
);
180 ptlrpc_add_bulk_page(desc
, page
, pageoffset
, len
);
182 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page
);
185 * Uninitialize and free bulk descriptor \a desc.
186 * Works on bulk descriptors both from server and client side.
188 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc
*desc
, int unpin
)
192 LASSERT(desc
!= NULL
);
193 LASSERT(desc
->bd_iov_count
!= LI_POISON
); /* not freed already */
194 LASSERT(desc
->bd_md_count
== 0); /* network hands off */
195 LASSERT((desc
->bd_export
!= NULL
) ^ (desc
->bd_import
!= NULL
));
197 sptlrpc_enc_pool_put_pages(desc
);
200 class_export_put(desc
->bd_export
);
202 class_import_put(desc
->bd_import
);
205 for (i
= 0; i
< desc
->bd_iov_count
; i
++)
206 page_cache_release(desc
->bd_iov
[i
].kiov_page
);
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
);
289 if (service_time
> now
- req
->rq_sent
+ 3) {
290 /* bz16408, however, this can also happen if early reply
291 * is lost and client RPC is expired and resent, early reply
292 * or reply of original RPC can still be fit in reply buffer
293 * of resent RPC, now client is measuring time from the
294 * resent time, but server sent back service time of original
297 CDEBUG((lustre_msg_get_flags(req
->rq_reqmsg
) & MSG_RESENT
) ?
298 D_ADAPTTO
: D_WARNING
,
299 "Reported service time %u > total measured time "
300 CFS_DURATION_T
"\n", service_time
,
301 cfs_time_sub(now
, req
->rq_sent
));
305 /* Network latency is total time less server processing time */
306 nl
= max_t(int, now
- req
->rq_sent
-
307 service_time
, 0) + 1; /* st rounding */
308 at
= &req
->rq_import
->imp_at
;
310 oldnl
= at_measured(&at
->iat_net_latency
, nl
);
312 CDEBUG(D_ADAPTTO
, "The network latency for %s (nid %s) has changed from %d to %d\n",
313 req
->rq_import
->imp_obd
->obd_name
,
315 &req
->rq_import
->imp_connection
->c_remote_uuid
),
316 oldnl
, at_get(&at
->iat_net_latency
));
319 static int unpack_reply(struct ptlrpc_request
*req
)
323 if (SPTLRPC_FLVR_POLICY(req
->rq_flvr
.sf_rpc
) != SPTLRPC_POLICY_NULL
) {
324 rc
= ptlrpc_unpack_rep_msg(req
, req
->rq_replen
);
326 DEBUG_REQ(D_ERROR
, req
, "unpack_rep failed: %d", rc
);
331 rc
= lustre_unpack_rep_ptlrpc_body(req
, MSG_PTLRPC_BODY_OFF
);
333 DEBUG_REQ(D_ERROR
, req
, "unpack ptlrpc body failed: %d", rc
);
340 * Handle an early reply message, called with the rq_lock held.
341 * If anything goes wrong just ignore it - same as if it never happened
343 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request
*req
)
345 struct ptlrpc_request
*early_req
;
350 spin_unlock(&req
->rq_lock
);
352 rc
= sptlrpc_cli_unwrap_early_reply(req
, &early_req
);
354 spin_lock(&req
->rq_lock
);
358 rc
= unpack_reply(early_req
);
360 /* Expecting to increase the service time estimate here */
361 ptlrpc_at_adj_service(req
,
362 lustre_msg_get_timeout(early_req
->rq_repmsg
));
363 ptlrpc_at_adj_net_latency(req
,
364 lustre_msg_get_service_time(early_req
->rq_repmsg
));
367 sptlrpc_cli_finish_early_reply(early_req
);
370 spin_lock(&req
->rq_lock
);
374 /* Adjust the local timeout for this req */
375 ptlrpc_at_set_req_timeout(req
);
377 spin_lock(&req
->rq_lock
);
378 olddl
= req
->rq_deadline
;
379 /* server assumes it now has rq_timeout from when it sent the
380 * early reply, so client should give it at least that long. */
381 req
->rq_deadline
= get_seconds() + req
->rq_timeout
+
382 ptlrpc_at_get_net_latency(req
);
384 DEBUG_REQ(D_ADAPTTO
, req
,
385 "Early reply #%d, new deadline in " CFS_DURATION_T
"s (" CFS_DURATION_T
"s)",
387 cfs_time_sub(req
->rq_deadline
, get_seconds()),
388 cfs_time_sub(req
->rq_deadline
, olddl
));
393 static struct kmem_cache
*request_cache
;
395 int ptlrpc_request_cache_init(void)
397 request_cache
= kmem_cache_create("ptlrpc_cache",
398 sizeof(struct ptlrpc_request
),
399 0, SLAB_HWCACHE_ALIGN
, NULL
);
400 return request_cache
== NULL
? -ENOMEM
: 0;
403 void ptlrpc_request_cache_fini(void)
405 kmem_cache_destroy(request_cache
);
408 struct ptlrpc_request
*ptlrpc_request_cache_alloc(gfp_t flags
)
410 struct ptlrpc_request
*req
;
412 OBD_SLAB_ALLOC_PTR_GFP(req
, request_cache
, flags
);
416 void ptlrpc_request_cache_free(struct ptlrpc_request
*req
)
418 OBD_SLAB_FREE_PTR(req
, request_cache
);
422 * Wind down request pool \a pool.
423 * Frees all requests from the pool too
425 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool
*pool
)
427 struct list_head
*l
, *tmp
;
428 struct ptlrpc_request
*req
;
430 LASSERT(pool
!= NULL
);
432 spin_lock(&pool
->prp_lock
);
433 list_for_each_safe(l
, tmp
, &pool
->prp_req_list
) {
434 req
= list_entry(l
, struct ptlrpc_request
, rq_list
);
435 list_del(&req
->rq_list
);
436 LASSERT(req
->rq_reqbuf
);
437 LASSERT(req
->rq_reqbuf_len
== pool
->prp_rq_size
);
438 kvfree(req
->rq_reqbuf
);
439 ptlrpc_request_cache_free(req
);
441 spin_unlock(&pool
->prp_lock
);
444 EXPORT_SYMBOL(ptlrpc_free_rq_pool
);
447 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
449 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool
*pool
, int num_rq
)
454 while (size
< pool
->prp_rq_size
)
457 LASSERTF(list_empty(&pool
->prp_req_list
) ||
458 size
== pool
->prp_rq_size
,
459 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
460 pool
->prp_rq_size
, size
);
462 spin_lock(&pool
->prp_lock
);
463 pool
->prp_rq_size
= size
;
464 for (i
= 0; i
< num_rq
; i
++) {
465 struct ptlrpc_request
*req
;
466 struct lustre_msg
*msg
;
468 spin_unlock(&pool
->prp_lock
);
469 req
= ptlrpc_request_cache_alloc(GFP_NOFS
);
472 msg
= libcfs_kvzalloc(size
, GFP_NOFS
);
474 ptlrpc_request_cache_free(req
);
477 req
->rq_reqbuf
= msg
;
478 req
->rq_reqbuf_len
= size
;
480 spin_lock(&pool
->prp_lock
);
481 list_add_tail(&req
->rq_list
, &pool
->prp_req_list
);
483 spin_unlock(&pool
->prp_lock
);
485 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool
);
488 * Create and initialize new request pool with given attributes:
489 * \a num_rq - initial number of requests to create for the pool
490 * \a msgsize - maximum message size possible for requests in thid pool
491 * \a populate_pool - function to be called when more requests need to be added
493 * Returns pointer to newly created pool or NULL on error.
495 struct ptlrpc_request_pool
*
496 ptlrpc_init_rq_pool(int num_rq
, int msgsize
,
497 void (*populate_pool
)(struct ptlrpc_request_pool
*, int))
499 struct ptlrpc_request_pool
*pool
;
501 pool
= kzalloc(sizeof(struct ptlrpc_request_pool
), GFP_NOFS
);
505 /* Request next power of two for the allocation, because internally
506 kernel would do exactly this */
508 spin_lock_init(&pool
->prp_lock
);
509 INIT_LIST_HEAD(&pool
->prp_req_list
);
510 pool
->prp_rq_size
= msgsize
+ SPTLRPC_MAX_PAYLOAD
;
511 pool
->prp_populate
= populate_pool
;
513 populate_pool(pool
, num_rq
);
515 if (list_empty(&pool
->prp_req_list
)) {
516 /* have not allocated a single request for the pool */
522 EXPORT_SYMBOL(ptlrpc_init_rq_pool
);
525 * Fetches one request from pool \a pool
527 static struct ptlrpc_request
*
528 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool
*pool
)
530 struct ptlrpc_request
*request
;
531 struct lustre_msg
*reqbuf
;
536 spin_lock(&pool
->prp_lock
);
538 /* See if we have anything in a pool, and bail out if nothing,
539 * in writeout path, where this matters, this is safe to do, because
540 * nothing is lost in this case, and when some in-flight requests
541 * complete, this code will be called again. */
542 if (unlikely(list_empty(&pool
->prp_req_list
))) {
543 spin_unlock(&pool
->prp_lock
);
547 request
= list_entry(pool
->prp_req_list
.next
, struct ptlrpc_request
,
549 list_del_init(&request
->rq_list
);
550 spin_unlock(&pool
->prp_lock
);
552 LASSERT(request
->rq_reqbuf
);
553 LASSERT(request
->rq_pool
);
555 reqbuf
= request
->rq_reqbuf
;
556 memset(request
, 0, sizeof(*request
));
557 request
->rq_reqbuf
= reqbuf
;
558 request
->rq_reqbuf_len
= pool
->prp_rq_size
;
559 request
->rq_pool
= pool
;
565 * Returns freed \a request to pool.
567 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request
*request
)
569 struct ptlrpc_request_pool
*pool
= request
->rq_pool
;
571 spin_lock(&pool
->prp_lock
);
572 LASSERT(list_empty(&request
->rq_list
));
573 LASSERT(!request
->rq_receiving_reply
);
574 list_add_tail(&request
->rq_list
, &pool
->prp_req_list
);
575 spin_unlock(&pool
->prp_lock
);
578 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
579 __u32 version
, int opcode
,
580 int count
, __u32
*lengths
, char **bufs
,
581 struct ptlrpc_cli_ctx
*ctx
)
583 struct obd_import
*imp
= request
->rq_import
;
587 request
->rq_cli_ctx
= sptlrpc_cli_ctx_get(ctx
);
589 rc
= sptlrpc_req_get_ctx(request
);
594 sptlrpc_req_set_flavor(request
, opcode
);
596 rc
= lustre_pack_request(request
, imp
->imp_msg_magic
, count
,
599 LASSERT(!request
->rq_pool
);
603 lustre_msg_add_version(request
->rq_reqmsg
, version
);
604 request
->rq_send_state
= LUSTRE_IMP_FULL
;
605 request
->rq_type
= PTL_RPC_MSG_REQUEST
;
606 request
->rq_export
= NULL
;
608 request
->rq_req_cbid
.cbid_fn
= request_out_callback
;
609 request
->rq_req_cbid
.cbid_arg
= request
;
611 request
->rq_reply_cbid
.cbid_fn
= reply_in_callback
;
612 request
->rq_reply_cbid
.cbid_arg
= request
;
614 request
->rq_reply_deadline
= 0;
615 request
->rq_phase
= RQ_PHASE_NEW
;
616 request
->rq_next_phase
= RQ_PHASE_UNDEFINED
;
618 request
->rq_request_portal
= imp
->imp_client
->cli_request_portal
;
619 request
->rq_reply_portal
= imp
->imp_client
->cli_reply_portal
;
621 ptlrpc_at_set_req_timeout(request
);
623 spin_lock_init(&request
->rq_lock
);
624 INIT_LIST_HEAD(&request
->rq_list
);
625 INIT_LIST_HEAD(&request
->rq_timed_list
);
626 INIT_LIST_HEAD(&request
->rq_replay_list
);
627 INIT_LIST_HEAD(&request
->rq_ctx_chain
);
628 INIT_LIST_HEAD(&request
->rq_set_chain
);
629 INIT_LIST_HEAD(&request
->rq_history_list
);
630 INIT_LIST_HEAD(&request
->rq_exp_list
);
631 init_waitqueue_head(&request
->rq_reply_waitq
);
632 init_waitqueue_head(&request
->rq_set_waitq
);
633 request
->rq_xid
= ptlrpc_next_xid();
634 atomic_set(&request
->rq_refcount
, 1);
636 lustre_msg_set_opc(request
->rq_reqmsg
, opcode
);
640 sptlrpc_cli_ctx_put(request
->rq_cli_ctx
, 1);
642 class_import_put(imp
);
646 int ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
647 __u32 version
, int opcode
, char **bufs
,
648 struct ptlrpc_cli_ctx
*ctx
)
652 count
= req_capsule_filled_sizes(&request
->rq_pill
, RCL_CLIENT
);
653 return __ptlrpc_request_bufs_pack(request
, version
, opcode
, count
,
654 request
->rq_pill
.rc_area
[RCL_CLIENT
],
657 EXPORT_SYMBOL(ptlrpc_request_bufs_pack
);
660 * Pack request buffers for network transfer, performing necessary encryption
661 * steps if necessary.
663 int ptlrpc_request_pack(struct ptlrpc_request
*request
,
664 __u32 version
, int opcode
)
667 rc
= ptlrpc_request_bufs_pack(request
, version
, opcode
, NULL
, NULL
);
671 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
672 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
673 * have to send old ptlrpc_body to keep interoperability with these
676 * Only three kinds of server->client RPCs so far:
681 * XXX This should be removed whenever we drop the interoperability with
682 * the these old clients.
684 if (opcode
== LDLM_BL_CALLBACK
|| opcode
== LDLM_CP_CALLBACK
||
685 opcode
== LDLM_GL_CALLBACK
)
686 req_capsule_shrink(&request
->rq_pill
, &RMF_PTLRPC_BODY
,
687 sizeof(struct ptlrpc_body_v2
), RCL_CLIENT
);
691 EXPORT_SYMBOL(ptlrpc_request_pack
);
694 * Helper function to allocate new request on import \a imp
695 * and possibly using existing request from pool \a pool if provided.
696 * Returns allocated request structure with import field filled or
700 struct ptlrpc_request
*__ptlrpc_request_alloc(struct obd_import
*imp
,
701 struct ptlrpc_request_pool
*pool
)
703 struct ptlrpc_request
*request
= NULL
;
706 request
= ptlrpc_prep_req_from_pool(pool
);
709 request
= ptlrpc_request_cache_alloc(GFP_NOFS
);
712 LASSERTF((unsigned long)imp
> 0x1000, "%p", imp
);
713 LASSERT(imp
!= LP_POISON
);
714 LASSERTF((unsigned long)imp
->imp_client
> 0x1000, "%p",
716 LASSERT(imp
->imp_client
!= LP_POISON
);
718 request
->rq_import
= class_import_get(imp
);
720 CERROR("request allocation out of memory\n");
727 * Helper function for creating a request.
728 * Calls __ptlrpc_request_alloc to allocate new request structure and inits
729 * buffer structures according to capsule template \a format.
730 * Returns allocated request structure pointer or NULL on error.
732 static struct ptlrpc_request
*
733 ptlrpc_request_alloc_internal(struct obd_import
*imp
,
734 struct ptlrpc_request_pool
*pool
,
735 const struct req_format
*format
)
737 struct ptlrpc_request
*request
;
739 request
= __ptlrpc_request_alloc(imp
, pool
);
743 req_capsule_init(&request
->rq_pill
, request
, RCL_CLIENT
);
744 req_capsule_set(&request
->rq_pill
, format
);
749 * Allocate new request structure for import \a imp and initialize its
750 * buffer structure according to capsule template \a format.
752 struct ptlrpc_request
*ptlrpc_request_alloc(struct obd_import
*imp
,
753 const struct req_format
*format
)
755 return ptlrpc_request_alloc_internal(imp
, NULL
, format
);
757 EXPORT_SYMBOL(ptlrpc_request_alloc
);
760 * Allocate new request structure for import \a imp from pool \a pool and
761 * initialize its buffer structure according to capsule template \a format.
763 struct ptlrpc_request
*ptlrpc_request_alloc_pool(struct obd_import
*imp
,
764 struct ptlrpc_request_pool
*pool
,
765 const struct req_format
*format
)
767 return ptlrpc_request_alloc_internal(imp
, pool
, format
);
769 EXPORT_SYMBOL(ptlrpc_request_alloc_pool
);
772 * For requests not from pool, free memory of the request structure.
773 * For requests obtained from a pool earlier, return request back to pool.
775 void ptlrpc_request_free(struct ptlrpc_request
*request
)
777 if (request
->rq_pool
)
778 __ptlrpc_free_req_to_pool(request
);
780 ptlrpc_request_cache_free(request
);
782 EXPORT_SYMBOL(ptlrpc_request_free
);
785 * Allocate new request for operation \a opcode and immediately pack it for
787 * Only used for simple requests like OBD_PING where the only important
788 * part of the request is operation itself.
789 * Returns allocated request or NULL on error.
791 struct ptlrpc_request
*ptlrpc_request_alloc_pack(struct obd_import
*imp
,
792 const struct req_format
*format
,
793 __u32 version
, int opcode
)
795 struct ptlrpc_request
*req
= ptlrpc_request_alloc(imp
, format
);
799 rc
= ptlrpc_request_pack(req
, version
, opcode
);
801 ptlrpc_request_free(req
);
807 EXPORT_SYMBOL(ptlrpc_request_alloc_pack
);
810 * Prepare request (fetched from pool \a pool if not NULL) on import \a imp
811 * for operation \a opcode. Request would contain \a count buffers.
812 * Sizes of buffers are described in array \a lengths and buffers themselves
813 * are provided by a pointer \a bufs.
814 * Returns prepared request structure pointer or NULL on error.
816 struct ptlrpc_request
*
817 ptlrpc_prep_req_pool(struct obd_import
*imp
,
818 __u32 version
, int opcode
,
819 int count
, __u32
*lengths
, char **bufs
,
820 struct ptlrpc_request_pool
*pool
)
822 struct ptlrpc_request
*request
;
825 request
= __ptlrpc_request_alloc(imp
, pool
);
829 rc
= __ptlrpc_request_bufs_pack(request
, version
, opcode
, count
,
830 lengths
, bufs
, NULL
);
832 ptlrpc_request_free(request
);
837 EXPORT_SYMBOL(ptlrpc_prep_req_pool
);
840 * Same as ptlrpc_prep_req_pool, but without pool
842 struct ptlrpc_request
*
843 ptlrpc_prep_req(struct obd_import
*imp
, __u32 version
, int opcode
, int count
,
844 __u32
*lengths
, char **bufs
)
846 return ptlrpc_prep_req_pool(imp
, version
, opcode
, count
, lengths
, bufs
,
849 EXPORT_SYMBOL(ptlrpc_prep_req
);
852 * Allocate and initialize new request set structure.
853 * Returns a pointer to the newly allocated set structure or NULL on error.
855 struct ptlrpc_request_set
*ptlrpc_prep_set(void)
857 struct ptlrpc_request_set
*set
;
859 set
= kzalloc(sizeof(*set
), GFP_NOFS
);
862 atomic_set(&set
->set_refcount
, 1);
863 INIT_LIST_HEAD(&set
->set_requests
);
864 init_waitqueue_head(&set
->set_waitq
);
865 atomic_set(&set
->set_new_count
, 0);
866 atomic_set(&set
->set_remaining
, 0);
867 spin_lock_init(&set
->set_new_req_lock
);
868 INIT_LIST_HEAD(&set
->set_new_requests
);
869 INIT_LIST_HEAD(&set
->set_cblist
);
870 set
->set_max_inflight
= UINT_MAX
;
871 set
->set_producer
= NULL
;
872 set
->set_producer_arg
= NULL
;
877 EXPORT_SYMBOL(ptlrpc_prep_set
);
880 * Allocate and initialize new request set structure with flow control
881 * extension. This extension allows to control the number of requests in-flight
882 * for the whole set. A callback function to generate requests must be provided
883 * and the request set will keep the number of requests sent over the wire to
885 * Returns a pointer to the newly allocated set structure or NULL on error.
887 struct ptlrpc_request_set
*ptlrpc_prep_fcset(int max
, set_producer_func func
,
891 struct ptlrpc_request_set
*set
;
893 set
= ptlrpc_prep_set();
897 set
->set_max_inflight
= max
;
898 set
->set_producer
= func
;
899 set
->set_producer_arg
= arg
;
903 EXPORT_SYMBOL(ptlrpc_prep_fcset
);
906 * Wind down and free request set structure previously allocated with
908 * Ensures that all requests on the set have completed and removes
909 * all requests from the request list in a set.
910 * If any unsent request happen to be on the list, pretends that they got
911 * an error in flight and calls their completion handler.
913 void ptlrpc_set_destroy(struct ptlrpc_request_set
*set
)
915 struct list_head
*tmp
;
916 struct list_head
*next
;
920 /* Requests on the set should either all be completed, or all be new */
921 expected_phase
= (atomic_read(&set
->set_remaining
) == 0) ?
922 RQ_PHASE_COMPLETE
: RQ_PHASE_NEW
;
923 list_for_each(tmp
, &set
->set_requests
) {
924 struct ptlrpc_request
*req
=
925 list_entry(tmp
, struct ptlrpc_request
,
928 LASSERT(req
->rq_phase
== expected_phase
);
932 LASSERTF(atomic_read(&set
->set_remaining
) == 0 ||
933 atomic_read(&set
->set_remaining
) == n
, "%d / %d\n",
934 atomic_read(&set
->set_remaining
), n
);
936 list_for_each_safe(tmp
, next
, &set
->set_requests
) {
937 struct ptlrpc_request
*req
=
938 list_entry(tmp
, struct ptlrpc_request
,
940 list_del_init(&req
->rq_set_chain
);
942 LASSERT(req
->rq_phase
== expected_phase
);
944 if (req
->rq_phase
== RQ_PHASE_NEW
) {
945 ptlrpc_req_interpret(NULL
, req
, -EBADR
);
946 atomic_dec(&set
->set_remaining
);
949 spin_lock(&req
->rq_lock
);
951 req
->rq_invalid_rqset
= 0;
952 spin_unlock(&req
->rq_lock
);
954 ptlrpc_req_finished(req
);
957 LASSERT(atomic_read(&set
->set_remaining
) == 0);
959 ptlrpc_reqset_put(set
);
961 EXPORT_SYMBOL(ptlrpc_set_destroy
);
964 * Add a callback function \a fn to the set.
965 * This function would be called when all requests on this set are completed.
966 * The function will be passed \a data argument.
968 int ptlrpc_set_add_cb(struct ptlrpc_request_set
*set
,
969 set_interpreter_func fn
, void *data
)
971 struct ptlrpc_set_cbdata
*cbdata
;
973 cbdata
= kzalloc(sizeof(*cbdata
), GFP_NOFS
);
977 cbdata
->psc_interpret
= fn
;
978 cbdata
->psc_data
= data
;
979 list_add_tail(&cbdata
->psc_item
, &set
->set_cblist
);
983 EXPORT_SYMBOL(ptlrpc_set_add_cb
);
986 * Add a new request to the general purpose request set.
987 * Assumes request reference from the caller.
989 void ptlrpc_set_add_req(struct ptlrpc_request_set
*set
,
990 struct ptlrpc_request
*req
)
992 LASSERT(list_empty(&req
->rq_set_chain
));
994 /* The set takes over the caller's request reference */
995 list_add_tail(&req
->rq_set_chain
, &set
->set_requests
);
997 atomic_inc(&set
->set_remaining
);
998 req
->rq_queued_time
= cfs_time_current();
1000 if (req
->rq_reqmsg
!= NULL
)
1001 lustre_msg_set_jobid(req
->rq_reqmsg
, NULL
);
1003 if (set
->set_producer
!= NULL
)
1004 /* If the request set has a producer callback, the RPC must be
1005 * sent straight away */
1006 ptlrpc_send_new_req(req
);
1008 EXPORT_SYMBOL(ptlrpc_set_add_req
);
1011 * Add a request to a request with dedicated server thread
1012 * and wake the thread to make any necessary processing.
1013 * Currently only used for ptlrpcd.
1015 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl
*pc
,
1016 struct ptlrpc_request
*req
)
1018 struct ptlrpc_request_set
*set
= pc
->pc_set
;
1021 LASSERT(req
->rq_set
== NULL
);
1022 LASSERT(test_bit(LIOD_STOP
, &pc
->pc_flags
) == 0);
1024 spin_lock(&set
->set_new_req_lock
);
1026 * The set takes over the caller's request reference.
1029 req
->rq_queued_time
= cfs_time_current();
1030 list_add_tail(&req
->rq_set_chain
, &set
->set_new_requests
);
1031 count
= atomic_inc_return(&set
->set_new_count
);
1032 spin_unlock(&set
->set_new_req_lock
);
1034 /* Only need to call wakeup once for the first entry. */
1036 wake_up(&set
->set_waitq
);
1038 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1039 * guarantee the async RPC can be processed ASAP, we have
1040 * no other better choice. It maybe fixed in future. */
1041 for (i
= 0; i
< pc
->pc_npartners
; i
++)
1042 wake_up(&pc
->pc_partners
[i
]->pc_set
->set_waitq
);
1045 EXPORT_SYMBOL(ptlrpc_set_add_new_req
);
1048 * Based on the current state of the import, determine if the request
1049 * can be sent, is an error, or should be delayed.
1051 * Returns true if this request should be delayed. If false, and
1052 * *status is set, then the request can not be sent and *status is the
1053 * error code. If false and status is 0, then request can be sent.
1055 * The imp->imp_lock must be held.
1057 static int ptlrpc_import_delay_req(struct obd_import
*imp
,
1058 struct ptlrpc_request
*req
, int *status
)
1062 LASSERT(status
!= NULL
);
1065 if (req
->rq_ctx_init
|| req
->rq_ctx_fini
) {
1066 /* always allow ctx init/fini rpc go through */
1067 } else if (imp
->imp_state
== LUSTRE_IMP_NEW
) {
1068 DEBUG_REQ(D_ERROR
, req
, "Uninitialized import.");
1070 } else if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
1071 /* pings may safely race with umount */
1072 DEBUG_REQ(lustre_msg_get_opc(req
->rq_reqmsg
) == OBD_PING
?
1073 D_HA
: D_ERROR
, req
, "IMP_CLOSED ");
1075 } else if (ptlrpc_send_limit_expired(req
)) {
1076 /* probably doesn't need to be a D_ERROR after initial testing */
1077 DEBUG_REQ(D_ERROR
, req
, "send limit expired ");
1079 } else if (req
->rq_send_state
== LUSTRE_IMP_CONNECTING
&&
1080 imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
1081 /* allow CONNECT even if import is invalid */
1082 if (atomic_read(&imp
->imp_inval_count
) != 0) {
1083 DEBUG_REQ(D_ERROR
, req
, "invalidate in flight");
1086 } else if (imp
->imp_invalid
|| imp
->imp_obd
->obd_no_recov
) {
1087 if (!imp
->imp_deactive
)
1088 DEBUG_REQ(D_NET
, req
, "IMP_INVALID");
1089 *status
= -ESHUTDOWN
; /* bz 12940 */
1090 } else if (req
->rq_import_generation
!= imp
->imp_generation
) {
1091 DEBUG_REQ(D_ERROR
, req
, "req wrong generation:");
1093 } else if (req
->rq_send_state
!= imp
->imp_state
) {
1094 /* invalidate in progress - any requests should be drop */
1095 if (atomic_read(&imp
->imp_inval_count
) != 0) {
1096 DEBUG_REQ(D_ERROR
, req
, "invalidate in flight");
1098 } else if (imp
->imp_dlm_fake
|| req
->rq_no_delay
) {
1099 *status
= -EWOULDBLOCK
;
1100 } else if (req
->rq_allow_replay
&&
1101 (imp
->imp_state
== LUSTRE_IMP_REPLAY
||
1102 imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
||
1103 imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
||
1104 imp
->imp_state
== LUSTRE_IMP_RECOVER
)) {
1105 DEBUG_REQ(D_HA
, req
, "allow during recovery.\n");
1115 * Decide if the error message regarding provided request \a req
1116 * should be printed to the console or not.
1117 * Makes it's decision on request status and other properties.
1118 * Returns 1 to print error on the system console or 0 if not.
1120 static int ptlrpc_console_allow(struct ptlrpc_request
*req
)
1125 LASSERT(req
->rq_reqmsg
!= NULL
);
1126 opc
= lustre_msg_get_opc(req
->rq_reqmsg
);
1128 /* Suppress particular reconnect errors which are to be expected. No
1129 * errors are suppressed for the initial connection on an import */
1130 if ((lustre_handle_is_used(&req
->rq_import
->imp_remote_handle
)) &&
1131 (opc
== OST_CONNECT
|| opc
== MDS_CONNECT
|| opc
== MGS_CONNECT
)) {
1133 /* Suppress timed out reconnect requests */
1134 if (req
->rq_timedout
)
1137 /* Suppress unavailable/again reconnect requests */
1138 err
= lustre_msg_get_status(req
->rq_repmsg
);
1139 if (err
== -ENODEV
|| err
== -EAGAIN
)
1147 * Check request processing status.
1148 * Returns the status.
1150 static int ptlrpc_check_status(struct ptlrpc_request
*req
)
1154 err
= lustre_msg_get_status(req
->rq_repmsg
);
1155 if (lustre_msg_get_type(req
->rq_repmsg
) == PTL_RPC_MSG_ERR
) {
1156 struct obd_import
*imp
= req
->rq_import
;
1157 __u32 opc
= lustre_msg_get_opc(req
->rq_reqmsg
);
1158 if (ptlrpc_console_allow(req
))
1159 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s, operation %s failed with %d.\n",
1160 imp
->imp_obd
->obd_name
,
1162 imp
->imp_connection
->c_peer
.nid
),
1163 ll_opcode2str(opc
), err
);
1164 return err
< 0 ? err
: -EINVAL
;
1168 DEBUG_REQ(D_INFO
, req
, "status is %d", err
);
1169 } else if (err
> 0) {
1170 /* XXX: translate this error from net to host */
1171 DEBUG_REQ(D_INFO
, req
, "status is %d", err
);
1178 * save pre-versions of objects into request for replay.
1179 * Versions are obtained from server reply.
1182 static void ptlrpc_save_versions(struct ptlrpc_request
*req
)
1184 struct lustre_msg
*repmsg
= req
->rq_repmsg
;
1185 struct lustre_msg
*reqmsg
= req
->rq_reqmsg
;
1186 __u64
*versions
= lustre_msg_get_versions(repmsg
);
1188 if (lustre_msg_get_flags(req
->rq_reqmsg
) & MSG_REPLAY
)
1192 lustre_msg_set_versions(reqmsg
, versions
);
1193 CDEBUG(D_INFO
, "Client save versions [%#llx/%#llx]\n",
1194 versions
[0], versions
[1]);
1198 * Callback function called when client receives RPC reply for \a req.
1199 * Returns 0 on success or error code.
1200 * The return value would be assigned to req->rq_status by the caller
1201 * as request processing status.
1202 * This function also decides if the request needs to be saved for later replay.
1204 static int after_reply(struct ptlrpc_request
*req
)
1206 struct obd_import
*imp
= req
->rq_import
;
1207 struct obd_device
*obd
= req
->rq_import
->imp_obd
;
1209 struct timeval work_start
;
1212 LASSERT(obd
!= NULL
);
1213 /* repbuf must be unlinked */
1214 LASSERT(!req
->rq_receiving_reply
&& !req
->rq_reply_unlink
);
1216 if (req
->rq_reply_truncate
) {
1217 if (ptlrpc_no_resend(req
)) {
1218 DEBUG_REQ(D_ERROR
, req
, "reply buffer overflow, expected: %d, actual size: %d",
1219 req
->rq_nob_received
, req
->rq_repbuf_len
);
1223 sptlrpc_cli_free_repbuf(req
);
1224 /* Pass the required reply buffer size (include
1225 * space for early reply).
1226 * NB: no need to roundup because alloc_repbuf
1227 * will roundup it */
1228 req
->rq_replen
= req
->rq_nob_received
;
1229 req
->rq_nob_received
= 0;
1230 spin_lock(&req
->rq_lock
);
1232 spin_unlock(&req
->rq_lock
);
1237 * NB Until this point, the whole of the incoming message,
1238 * including buflens, status etc is in the sender's byte order.
1240 rc
= sptlrpc_cli_unwrap_reply(req
);
1242 DEBUG_REQ(D_ERROR
, req
, "unwrap reply failed (%d):", rc
);
1247 * Security layer unwrap might ask resend this request.
1252 rc
= unpack_reply(req
);
1256 /* retry indefinitely on EINPROGRESS */
1257 if (lustre_msg_get_status(req
->rq_repmsg
) == -EINPROGRESS
&&
1258 ptlrpc_no_resend(req
) == 0 && !req
->rq_no_retry_einprogress
) {
1259 time_t now
= get_seconds();
1261 DEBUG_REQ(D_RPCTRACE
, req
, "Resending request on EINPROGRESS");
1262 spin_lock(&req
->rq_lock
);
1264 spin_unlock(&req
->rq_lock
);
1265 req
->rq_nr_resend
++;
1267 /* allocate new xid to avoid reply reconstruction */
1268 if (!req
->rq_bulk
) {
1269 /* new xid is already allocated for bulk in
1270 * ptlrpc_check_set() */
1271 req
->rq_xid
= ptlrpc_next_xid();
1272 DEBUG_REQ(D_RPCTRACE
, req
, "Allocating new xid for resend on EINPROGRESS");
1275 /* Readjust the timeout for current conditions */
1276 ptlrpc_at_set_req_timeout(req
);
1277 /* delay resend to give a chance to the server to get ready.
1278 * The delay is increased by 1s on every resend and is capped to
1279 * the current request timeout (i.e. obd_timeout if AT is off,
1280 * or AT service time x 125% + 5s, see at_est2timeout) */
1281 if (req
->rq_nr_resend
> req
->rq_timeout
)
1282 req
->rq_sent
= now
+ req
->rq_timeout
;
1284 req
->rq_sent
= now
+ req
->rq_nr_resend
;
1289 do_gettimeofday(&work_start
);
1290 timediff
= cfs_timeval_sub(&work_start
, &req
->rq_arrival_time
, NULL
);
1291 if (obd
->obd_svc_stats
!= NULL
) {
1292 lprocfs_counter_add(obd
->obd_svc_stats
, PTLRPC_REQWAIT_CNTR
,
1294 ptlrpc_lprocfs_rpc_sent(req
, timediff
);
1297 if (lustre_msg_get_type(req
->rq_repmsg
) != PTL_RPC_MSG_REPLY
&&
1298 lustre_msg_get_type(req
->rq_repmsg
) != PTL_RPC_MSG_ERR
) {
1299 DEBUG_REQ(D_ERROR
, req
, "invalid packet received (type=%u)",
1300 lustre_msg_get_type(req
->rq_repmsg
));
1304 if (lustre_msg_get_opc(req
->rq_reqmsg
) != OBD_PING
)
1305 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP
, cfs_fail_val
);
1306 ptlrpc_at_adj_service(req
, lustre_msg_get_timeout(req
->rq_repmsg
));
1307 ptlrpc_at_adj_net_latency(req
,
1308 lustre_msg_get_service_time(req
->rq_repmsg
));
1310 rc
= ptlrpc_check_status(req
);
1311 imp
->imp_connect_error
= rc
;
1315 * Either we've been evicted, or the server has failed for
1316 * some reason. Try to reconnect, and if that fails, punt to
1319 if (ll_rpc_recoverable_error(rc
)) {
1320 if (req
->rq_send_state
!= LUSTRE_IMP_FULL
||
1321 imp
->imp_obd
->obd_no_recov
|| imp
->imp_dlm_fake
) {
1324 ptlrpc_request_handle_notconn(req
);
1329 * Let's look if server sent slv. Do it only for RPC with
1332 ldlm_cli_update_pool(req
);
1336 * Store transno in reqmsg for replay.
1338 if (!(lustre_msg_get_flags(req
->rq_reqmsg
) & MSG_REPLAY
)) {
1339 req
->rq_transno
= lustre_msg_get_transno(req
->rq_repmsg
);
1340 lustre_msg_set_transno(req
->rq_reqmsg
, req
->rq_transno
);
1343 if (imp
->imp_replayable
) {
1344 spin_lock(&imp
->imp_lock
);
1346 * No point in adding already-committed requests to the replay
1347 * list, we will just remove them immediately. b=9829
1349 if (req
->rq_transno
!= 0 &&
1351 lustre_msg_get_last_committed(req
->rq_repmsg
) ||
1353 /** version recovery */
1354 ptlrpc_save_versions(req
);
1355 ptlrpc_retain_replayable_request(req
, imp
);
1356 } else if (req
->rq_commit_cb
!= NULL
&&
1357 list_empty(&req
->rq_replay_list
)) {
1358 /* NB: don't call rq_commit_cb if it's already on
1359 * rq_replay_list, ptlrpc_free_committed() will call
1360 * it later, see LU-3618 for details */
1361 spin_unlock(&imp
->imp_lock
);
1362 req
->rq_commit_cb(req
);
1363 spin_lock(&imp
->imp_lock
);
1367 * Replay-enabled imports return commit-status information.
1369 if (lustre_msg_get_last_committed(req
->rq_repmsg
)) {
1370 imp
->imp_peer_committed_transno
=
1371 lustre_msg_get_last_committed(req
->rq_repmsg
);
1374 ptlrpc_free_committed(imp
);
1376 if (!list_empty(&imp
->imp_replay_list
)) {
1377 struct ptlrpc_request
*last
;
1379 last
= list_entry(imp
->imp_replay_list
.prev
,
1380 struct ptlrpc_request
,
1383 * Requests with rq_replay stay on the list even if no
1384 * commit is expected.
1386 if (last
->rq_transno
> imp
->imp_peer_committed_transno
)
1387 ptlrpc_pinger_commit_expected(imp
);
1390 spin_unlock(&imp
->imp_lock
);
1397 * Helper function to send request \a req over the network for the first time
1398 * Also adjusts request phase.
1399 * Returns 0 on success or error code.
1401 static int ptlrpc_send_new_req(struct ptlrpc_request
*req
)
1403 struct obd_import
*imp
= req
->rq_import
;
1406 LASSERT(req
->rq_phase
== RQ_PHASE_NEW
);
1407 if (req
->rq_sent
&& (req
->rq_sent
> get_seconds()) &&
1408 (!req
->rq_generation_set
||
1409 req
->rq_import_generation
== imp
->imp_generation
))
1412 ptlrpc_rqphase_move(req
, RQ_PHASE_RPC
);
1414 spin_lock(&imp
->imp_lock
);
1416 if (!req
->rq_generation_set
)
1417 req
->rq_import_generation
= imp
->imp_generation
;
1419 if (ptlrpc_import_delay_req(imp
, req
, &rc
)) {
1420 spin_lock(&req
->rq_lock
);
1421 req
->rq_waiting
= 1;
1422 spin_unlock(&req
->rq_lock
);
1424 DEBUG_REQ(D_HA
, req
, "req from PID %d waiting for recovery: (%s != %s)",
1425 lustre_msg_get_status(req
->rq_reqmsg
),
1426 ptlrpc_import_state_name(req
->rq_send_state
),
1427 ptlrpc_import_state_name(imp
->imp_state
));
1428 LASSERT(list_empty(&req
->rq_list
));
1429 list_add_tail(&req
->rq_list
, &imp
->imp_delayed_list
);
1430 atomic_inc(&req
->rq_import
->imp_inflight
);
1431 spin_unlock(&imp
->imp_lock
);
1436 spin_unlock(&imp
->imp_lock
);
1437 req
->rq_status
= rc
;
1438 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1442 LASSERT(list_empty(&req
->rq_list
));
1443 list_add_tail(&req
->rq_list
, &imp
->imp_sending_list
);
1444 atomic_inc(&req
->rq_import
->imp_inflight
);
1445 spin_unlock(&imp
->imp_lock
);
1447 lustre_msg_set_status(req
->rq_reqmsg
, current_pid());
1449 rc
= sptlrpc_req_refresh_ctx(req
, -1);
1452 req
->rq_status
= rc
;
1455 spin_lock(&req
->rq_lock
);
1456 req
->rq_wait_ctx
= 1;
1457 spin_unlock(&req
->rq_lock
);
1461 CDEBUG(D_RPCTRACE
, "Sending RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1463 imp
->imp_obd
->obd_uuid
.uuid
,
1464 lustre_msg_get_status(req
->rq_reqmsg
), req
->rq_xid
,
1465 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
),
1466 lustre_msg_get_opc(req
->rq_reqmsg
));
1468 rc
= ptl_send_rpc(req
, 0);
1470 DEBUG_REQ(D_HA
, req
, "send failed (%d); expect timeout", rc
);
1471 spin_lock(&req
->rq_lock
);
1472 req
->rq_net_err
= 1;
1473 spin_unlock(&req
->rq_lock
);
1479 static inline int ptlrpc_set_producer(struct ptlrpc_request_set
*set
)
1483 LASSERT(set
->set_producer
!= NULL
);
1485 remaining
= atomic_read(&set
->set_remaining
);
1487 /* populate the ->set_requests list with requests until we
1488 * reach the maximum number of RPCs in flight for this set */
1489 while (atomic_read(&set
->set_remaining
) < set
->set_max_inflight
) {
1490 rc
= set
->set_producer(set
, set
->set_producer_arg
);
1491 if (rc
== -ENOENT
) {
1492 /* no more RPC to produce */
1493 set
->set_producer
= NULL
;
1494 set
->set_producer_arg
= NULL
;
1499 return (atomic_read(&set
->set_remaining
) - remaining
);
1503 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1504 * and no more replies are expected.
1505 * (it is possible to get less replies than requests sent e.g. due to timed out
1506 * requests or requests that we had trouble to send out)
1508 * NOTE: This function contains a potential schedule point (cond_resched()).
1510 int ptlrpc_check_set(const struct lu_env
*env
, struct ptlrpc_request_set
*set
)
1512 struct list_head
*tmp
, *next
;
1513 struct list_head comp_reqs
;
1514 int force_timer_recalc
= 0;
1516 if (atomic_read(&set
->set_remaining
) == 0)
1519 INIT_LIST_HEAD(&comp_reqs
);
1520 list_for_each_safe(tmp
, next
, &set
->set_requests
) {
1521 struct ptlrpc_request
*req
=
1522 list_entry(tmp
, struct ptlrpc_request
,
1524 struct obd_import
*imp
= req
->rq_import
;
1525 int unregistered
= 0;
1528 /* This schedule point is mainly for the ptlrpcd caller of this
1529 * function. Most ptlrpc sets are not long-lived and unbounded
1530 * in length, but at the least the set used by the ptlrpcd is.
1531 * Since the processing time is unbounded, we need to insert an
1532 * explicit schedule point to make the thread well-behaved.
1536 if (req
->rq_phase
== RQ_PHASE_NEW
&&
1537 ptlrpc_send_new_req(req
)) {
1538 force_timer_recalc
= 1;
1541 /* delayed send - skip */
1542 if (req
->rq_phase
== RQ_PHASE_NEW
&& req
->rq_sent
)
1545 /* delayed resend - skip */
1546 if (req
->rq_phase
== RQ_PHASE_RPC
&& req
->rq_resend
&&
1547 req
->rq_sent
> get_seconds())
1550 if (!(req
->rq_phase
== RQ_PHASE_RPC
||
1551 req
->rq_phase
== RQ_PHASE_BULK
||
1552 req
->rq_phase
== RQ_PHASE_INTERPRET
||
1553 req
->rq_phase
== RQ_PHASE_UNREGISTERING
||
1554 req
->rq_phase
== RQ_PHASE_COMPLETE
)) {
1555 DEBUG_REQ(D_ERROR
, req
, "bad phase %x", req
->rq_phase
);
1559 if (req
->rq_phase
== RQ_PHASE_UNREGISTERING
) {
1560 LASSERT(req
->rq_next_phase
!= req
->rq_phase
);
1561 LASSERT(req
->rq_next_phase
!= RQ_PHASE_UNDEFINED
);
1564 * Skip processing until reply is unlinked. We
1565 * can't return to pool before that and we can't
1566 * call interpret before that. We need to make
1567 * sure that all rdma transfers finished and will
1568 * not corrupt any data.
1570 if (ptlrpc_client_recv_or_unlink(req
) ||
1571 ptlrpc_client_bulk_active(req
))
1575 * Turn fail_loc off to prevent it from looping
1578 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
1579 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
,
1582 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
)) {
1583 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
,
1588 * Move to next phase if reply was successfully
1591 ptlrpc_rqphase_move(req
, req
->rq_next_phase
);
1594 if (req
->rq_phase
== RQ_PHASE_COMPLETE
) {
1595 list_move_tail(&req
->rq_set_chain
, &comp_reqs
);
1599 if (req
->rq_phase
== RQ_PHASE_INTERPRET
)
1603 * Note that this also will start async reply unlink.
1605 if (req
->rq_net_err
&& !req
->rq_timedout
) {
1606 ptlrpc_expire_one_request(req
, 1);
1609 * Check if we still need to wait for unlink.
1611 if (ptlrpc_client_recv_or_unlink(req
) ||
1612 ptlrpc_client_bulk_active(req
))
1614 /* If there is no need to resend, fail it now. */
1615 if (req
->rq_no_resend
) {
1616 if (req
->rq_status
== 0)
1617 req
->rq_status
= -EIO
;
1618 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1626 spin_lock(&req
->rq_lock
);
1627 req
->rq_replied
= 0;
1628 spin_unlock(&req
->rq_lock
);
1629 if (req
->rq_status
== 0)
1630 req
->rq_status
= -EIO
;
1631 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1635 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1636 * so it sets rq_intr regardless of individual rpc
1637 * timeouts. The synchronous IO waiting path sets
1638 * rq_intr irrespective of whether ptlrpcd
1639 * has seen a timeout. Our policy is to only interpret
1640 * interrupted rpcs after they have timed out, so we
1641 * need to enforce that here.
1644 if (req
->rq_intr
&& (req
->rq_timedout
|| req
->rq_waiting
||
1645 req
->rq_wait_ctx
)) {
1646 req
->rq_status
= -EINTR
;
1647 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1651 if (req
->rq_phase
== RQ_PHASE_RPC
) {
1652 if (req
->rq_timedout
|| req
->rq_resend
||
1653 req
->rq_waiting
|| req
->rq_wait_ctx
) {
1656 if (!ptlrpc_unregister_reply(req
, 1))
1659 spin_lock(&imp
->imp_lock
);
1660 if (ptlrpc_import_delay_req(imp
, req
,
1662 /* put on delay list - only if we wait
1663 * recovery finished - before send */
1664 list_del_init(&req
->rq_list
);
1665 list_add_tail(&req
->rq_list
,
1668 spin_unlock(&imp
->imp_lock
);
1673 req
->rq_status
= status
;
1674 ptlrpc_rqphase_move(req
,
1675 RQ_PHASE_INTERPRET
);
1676 spin_unlock(&imp
->imp_lock
);
1679 if (ptlrpc_no_resend(req
) &&
1680 !req
->rq_wait_ctx
) {
1681 req
->rq_status
= -ENOTCONN
;
1682 ptlrpc_rqphase_move(req
,
1683 RQ_PHASE_INTERPRET
);
1684 spin_unlock(&imp
->imp_lock
);
1688 list_del_init(&req
->rq_list
);
1689 list_add_tail(&req
->rq_list
,
1690 &imp
->imp_sending_list
);
1692 spin_unlock(&imp
->imp_lock
);
1694 spin_lock(&req
->rq_lock
);
1695 req
->rq_waiting
= 0;
1696 spin_unlock(&req
->rq_lock
);
1698 if (req
->rq_timedout
|| req
->rq_resend
) {
1699 /* This is re-sending anyways,
1700 * let's mark req as resend. */
1701 spin_lock(&req
->rq_lock
);
1703 spin_unlock(&req
->rq_lock
);
1707 if (!ptlrpc_unregister_bulk(req
, 1))
1710 /* ensure previous bulk fails */
1711 old_xid
= req
->rq_xid
;
1712 req
->rq_xid
= ptlrpc_next_xid();
1713 CDEBUG(D_HA
, "resend bulk old x%llu new x%llu\n",
1714 old_xid
, req
->rq_xid
);
1718 * rq_wait_ctx is only touched by ptlrpcd,
1719 * so no lock is needed here.
1721 status
= sptlrpc_req_refresh_ctx(req
, -1);
1724 req
->rq_status
= status
;
1725 spin_lock(&req
->rq_lock
);
1726 req
->rq_wait_ctx
= 0;
1727 spin_unlock(&req
->rq_lock
);
1728 force_timer_recalc
= 1;
1730 spin_lock(&req
->rq_lock
);
1731 req
->rq_wait_ctx
= 1;
1732 spin_unlock(&req
->rq_lock
);
1737 spin_lock(&req
->rq_lock
);
1738 req
->rq_wait_ctx
= 0;
1739 spin_unlock(&req
->rq_lock
);
1742 rc
= ptl_send_rpc(req
, 0);
1744 DEBUG_REQ(D_HA
, req
,
1745 "send failed: rc = %d", rc
);
1746 force_timer_recalc
= 1;
1747 spin_lock(&req
->rq_lock
);
1748 req
->rq_net_err
= 1;
1749 spin_unlock(&req
->rq_lock
);
1752 /* need to reset the timeout */
1753 force_timer_recalc
= 1;
1756 spin_lock(&req
->rq_lock
);
1758 if (ptlrpc_client_early(req
)) {
1759 ptlrpc_at_recv_early_reply(req
);
1760 spin_unlock(&req
->rq_lock
);
1764 /* Still waiting for a reply? */
1765 if (ptlrpc_client_recv(req
)) {
1766 spin_unlock(&req
->rq_lock
);
1770 /* Did we actually receive a reply? */
1771 if (!ptlrpc_client_replied(req
)) {
1772 spin_unlock(&req
->rq_lock
);
1776 spin_unlock(&req
->rq_lock
);
1778 /* unlink from net because we are going to
1779 * swab in-place of reply buffer */
1780 unregistered
= ptlrpc_unregister_reply(req
, 1);
1784 req
->rq_status
= after_reply(req
);
1788 /* If there is no bulk associated with this request,
1789 * then we're done and should let the interpreter
1790 * process the reply. Similarly if the RPC returned
1791 * an error, and therefore the bulk will never arrive.
1793 if (req
->rq_bulk
== NULL
|| req
->rq_status
< 0) {
1794 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1798 ptlrpc_rqphase_move(req
, RQ_PHASE_BULK
);
1801 LASSERT(req
->rq_phase
== RQ_PHASE_BULK
);
1802 if (ptlrpc_client_bulk_active(req
))
1805 if (req
->rq_bulk
->bd_failure
) {
1806 /* The RPC reply arrived OK, but the bulk screwed
1807 * up! Dead weird since the server told us the RPC
1808 * was good after getting the REPLY for her GET or
1809 * the ACK for her PUT. */
1810 DEBUG_REQ(D_ERROR
, req
, "bulk transfer failed");
1811 req
->rq_status
= -EIO
;
1814 ptlrpc_rqphase_move(req
, RQ_PHASE_INTERPRET
);
1817 LASSERT(req
->rq_phase
== RQ_PHASE_INTERPRET
);
1819 /* This moves to "unregistering" phase we need to wait for
1821 if (!unregistered
&& !ptlrpc_unregister_reply(req
, 1)) {
1822 /* start async bulk unlink too */
1823 ptlrpc_unregister_bulk(req
, 1);
1827 if (!ptlrpc_unregister_bulk(req
, 1))
1830 /* When calling interpret receiving already should be
1832 LASSERT(!req
->rq_receiving_reply
);
1834 ptlrpc_req_interpret(env
, req
, req
->rq_status
);
1836 if (ptlrpcd_check_work(req
)) {
1837 atomic_dec(&set
->set_remaining
);
1840 ptlrpc_rqphase_move(req
, RQ_PHASE_COMPLETE
);
1842 CDEBUG(req
->rq_reqmsg
!= NULL
? D_RPCTRACE
: 0,
1843 "Completed RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1844 current_comm(), imp
->imp_obd
->obd_uuid
.uuid
,
1845 lustre_msg_get_status(req
->rq_reqmsg
), req
->rq_xid
,
1846 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
),
1847 lustre_msg_get_opc(req
->rq_reqmsg
));
1849 spin_lock(&imp
->imp_lock
);
1850 /* Request already may be not on sending or delaying list. This
1851 * may happen in the case of marking it erroneous for the case
1852 * ptlrpc_import_delay_req(req, status) find it impossible to
1853 * allow sending this rpc and returns *status != 0. */
1854 if (!list_empty(&req
->rq_list
)) {
1855 list_del_init(&req
->rq_list
);
1856 atomic_dec(&imp
->imp_inflight
);
1858 spin_unlock(&imp
->imp_lock
);
1860 atomic_dec(&set
->set_remaining
);
1861 wake_up_all(&imp
->imp_recovery_waitq
);
1863 if (set
->set_producer
) {
1864 /* produce a new request if possible */
1865 if (ptlrpc_set_producer(set
) > 0)
1866 force_timer_recalc
= 1;
1868 /* free the request that has just been completed
1869 * in order not to pollute set->set_requests */
1870 list_del_init(&req
->rq_set_chain
);
1871 spin_lock(&req
->rq_lock
);
1873 req
->rq_invalid_rqset
= 0;
1874 spin_unlock(&req
->rq_lock
);
1876 /* record rq_status to compute the final status later */
1877 if (req
->rq_status
!= 0)
1878 set
->set_rc
= req
->rq_status
;
1879 ptlrpc_req_finished(req
);
1881 list_move_tail(&req
->rq_set_chain
, &comp_reqs
);
1885 /* move completed request at the head of list so it's easier for
1886 * caller to find them */
1887 list_splice(&comp_reqs
, &set
->set_requests
);
1889 /* If we hit an error, we want to recover promptly. */
1890 return atomic_read(&set
->set_remaining
) == 0 || force_timer_recalc
;
1892 EXPORT_SYMBOL(ptlrpc_check_set
);
1895 * Time out request \a req. is \a async_unlink is set, that means do not wait
1896 * until LNet actually confirms network buffer unlinking.
1897 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1899 int ptlrpc_expire_one_request(struct ptlrpc_request
*req
, int async_unlink
)
1901 struct obd_import
*imp
= req
->rq_import
;
1904 spin_lock(&req
->rq_lock
);
1905 req
->rq_timedout
= 1;
1906 spin_unlock(&req
->rq_lock
);
1908 DEBUG_REQ(D_WARNING
, req
, "Request sent has %s: [sent "CFS_DURATION_T
1909 "/real "CFS_DURATION_T
"]",
1910 req
->rq_net_err
? "failed due to network error" :
1911 ((req
->rq_real_sent
== 0 ||
1912 time_before((unsigned long)req
->rq_real_sent
, (unsigned long)req
->rq_sent
) ||
1913 cfs_time_aftereq(req
->rq_real_sent
, req
->rq_deadline
)) ?
1914 "timed out for sent delay" : "timed out for slow reply"),
1915 req
->rq_sent
, req
->rq_real_sent
);
1917 if (imp
!= NULL
&& obd_debug_peer_on_timeout
)
1918 LNetCtl(IOC_LIBCFS_DEBUG_PEER
, &imp
->imp_connection
->c_peer
);
1920 ptlrpc_unregister_reply(req
, async_unlink
);
1921 ptlrpc_unregister_bulk(req
, async_unlink
);
1923 if (obd_dump_on_timeout
)
1924 libcfs_debug_dumplog();
1927 DEBUG_REQ(D_HA
, req
, "NULL import: already cleaned up?");
1931 atomic_inc(&imp
->imp_timeouts
);
1933 /* The DLM server doesn't want recovery run on its imports. */
1934 if (imp
->imp_dlm_fake
)
1937 /* If this request is for recovery or other primordial tasks,
1938 * then error it out here. */
1939 if (req
->rq_ctx_init
|| req
->rq_ctx_fini
||
1940 req
->rq_send_state
!= LUSTRE_IMP_FULL
||
1941 imp
->imp_obd
->obd_no_recov
) {
1942 DEBUG_REQ(D_RPCTRACE
, req
, "err -110, sent_state=%s (now=%s)",
1943 ptlrpc_import_state_name(req
->rq_send_state
),
1944 ptlrpc_import_state_name(imp
->imp_state
));
1945 spin_lock(&req
->rq_lock
);
1946 req
->rq_status
= -ETIMEDOUT
;
1948 spin_unlock(&req
->rq_lock
);
1952 /* if a request can't be resent we can't wait for an answer after
1954 if (ptlrpc_no_resend(req
)) {
1955 DEBUG_REQ(D_RPCTRACE
, req
, "TIMEOUT-NORESEND:");
1959 ptlrpc_fail_import(imp
, lustre_msg_get_conn_cnt(req
->rq_reqmsg
));
1965 * Time out all uncompleted requests in request set pointed by \a data
1966 * Callback used when waiting on sets with l_wait_event.
1969 int ptlrpc_expired_set(void *data
)
1971 struct ptlrpc_request_set
*set
= data
;
1972 struct list_head
*tmp
;
1973 time_t now
= get_seconds();
1975 LASSERT(set
!= NULL
);
1978 * A timeout expired. See which reqs it applies to...
1980 list_for_each(tmp
, &set
->set_requests
) {
1981 struct ptlrpc_request
*req
=
1982 list_entry(tmp
, struct ptlrpc_request
,
1985 /* don't expire request waiting for context */
1986 if (req
->rq_wait_ctx
)
1989 /* Request in-flight? */
1990 if (!((req
->rq_phase
== RQ_PHASE_RPC
&&
1991 !req
->rq_waiting
&& !req
->rq_resend
) ||
1992 (req
->rq_phase
== RQ_PHASE_BULK
)))
1995 if (req
->rq_timedout
|| /* already dealt with */
1996 req
->rq_deadline
> now
) /* not expired */
1999 /* Deal with this guy. Do it asynchronously to not block
2000 * ptlrpcd thread. */
2001 ptlrpc_expire_one_request(req
, 1);
2005 * When waiting for a whole set, we always break out of the
2006 * sleep so we can recalculate the timeout, or enable interrupts
2007 * if everyone's timed out.
2011 EXPORT_SYMBOL(ptlrpc_expired_set
);
2014 * Sets rq_intr flag in \a req under spinlock.
2016 void ptlrpc_mark_interrupted(struct ptlrpc_request
*req
)
2018 spin_lock(&req
->rq_lock
);
2020 spin_unlock(&req
->rq_lock
);
2022 EXPORT_SYMBOL(ptlrpc_mark_interrupted
);
2025 * Interrupts (sets interrupted flag) all uncompleted requests in
2026 * a set \a data. Callback for l_wait_event for interruptible waits.
2028 void ptlrpc_interrupted_set(void *data
)
2030 struct ptlrpc_request_set
*set
= data
;
2031 struct list_head
*tmp
;
2033 LASSERT(set
!= NULL
);
2034 CDEBUG(D_RPCTRACE
, "INTERRUPTED SET %p\n", set
);
2036 list_for_each(tmp
, &set
->set_requests
) {
2037 struct ptlrpc_request
*req
=
2038 list_entry(tmp
, struct ptlrpc_request
,
2041 if (req
->rq_phase
!= RQ_PHASE_RPC
&&
2042 req
->rq_phase
!= RQ_PHASE_UNREGISTERING
)
2045 ptlrpc_mark_interrupted(req
);
2048 EXPORT_SYMBOL(ptlrpc_interrupted_set
);
2051 * Get the smallest timeout in the set; this does NOT set a timeout.
2053 int ptlrpc_set_next_timeout(struct ptlrpc_request_set
*set
)
2055 struct list_head
*tmp
;
2056 time_t now
= get_seconds();
2058 struct ptlrpc_request
*req
;
2061 list_for_each(tmp
, &set
->set_requests
) {
2062 req
= list_entry(tmp
, struct ptlrpc_request
, rq_set_chain
);
2065 * Request in-flight?
2067 if (!(((req
->rq_phase
== RQ_PHASE_RPC
) && !req
->rq_waiting
) ||
2068 (req
->rq_phase
== RQ_PHASE_BULK
) ||
2069 (req
->rq_phase
== RQ_PHASE_NEW
)))
2073 * Already timed out.
2075 if (req
->rq_timedout
)
2081 if (req
->rq_wait_ctx
)
2084 if (req
->rq_phase
== RQ_PHASE_NEW
)
2085 deadline
= req
->rq_sent
;
2086 else if (req
->rq_phase
== RQ_PHASE_RPC
&& req
->rq_resend
)
2087 deadline
= req
->rq_sent
;
2089 deadline
= req
->rq_sent
+ req
->rq_timeout
;
2091 if (deadline
<= now
) /* actually expired already */
2092 timeout
= 1; /* ASAP */
2093 else if (timeout
== 0 || timeout
> deadline
- now
)
2094 timeout
= deadline
- now
;
2098 EXPORT_SYMBOL(ptlrpc_set_next_timeout
);
2101 * Send all unset request from the set and then wait until all
2102 * requests in the set complete (either get a reply, timeout, get an
2103 * error or otherwise be interrupted).
2104 * Returns 0 on success or error code otherwise.
2106 int ptlrpc_set_wait(struct ptlrpc_request_set
*set
)
2108 struct list_head
*tmp
;
2109 struct ptlrpc_request
*req
;
2110 struct l_wait_info lwi
;
2113 if (set
->set_producer
)
2114 (void)ptlrpc_set_producer(set
);
2116 list_for_each(tmp
, &set
->set_requests
) {
2117 req
= list_entry(tmp
, struct ptlrpc_request
,
2119 if (req
->rq_phase
== RQ_PHASE_NEW
)
2120 (void)ptlrpc_send_new_req(req
);
2123 if (list_empty(&set
->set_requests
))
2127 timeout
= ptlrpc_set_next_timeout(set
);
2129 /* wait until all complete, interrupted, or an in-flight
2131 CDEBUG(D_RPCTRACE
, "set %p going to sleep for %d seconds\n",
2134 if (timeout
== 0 && !cfs_signal_pending())
2136 * No requests are in-flight (ether timed out
2137 * or delayed), so we can allow interrupts.
2138 * We still want to block for a limited time,
2139 * so we allow interrupts during the timeout.
2141 lwi
= LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2143 ptlrpc_interrupted_set
, set
);
2146 * At least one request is in flight, so no
2147 * interrupts are allowed. Wait until all
2148 * complete, or an in-flight req times out.
2150 lwi
= LWI_TIMEOUT(cfs_time_seconds(timeout
? timeout
: 1),
2151 ptlrpc_expired_set
, set
);
2153 rc
= l_wait_event(set
->set_waitq
, ptlrpc_check_set(NULL
, set
), &lwi
);
2155 /* LU-769 - if we ignored the signal because it was already
2156 * pending when we started, we need to handle it now or we risk
2157 * it being ignored forever */
2158 if (rc
== -ETIMEDOUT
&& !lwi
.lwi_allow_intr
&&
2159 cfs_signal_pending()) {
2160 sigset_t blocked_sigs
=
2161 cfs_block_sigsinv(LUSTRE_FATAL_SIGS
);
2163 /* In fact we only interrupt for the "fatal" signals
2164 * like SIGINT or SIGKILL. We still ignore less
2165 * important signals since ptlrpc set is not easily
2166 * reentrant from userspace again */
2167 if (cfs_signal_pending())
2168 ptlrpc_interrupted_set(set
);
2169 cfs_restore_sigs(blocked_sigs
);
2172 LASSERT(rc
== 0 || rc
== -EINTR
|| rc
== -ETIMEDOUT
);
2174 /* -EINTR => all requests have been flagged rq_intr so next
2176 * -ETIMEDOUT => someone timed out. When all reqs have
2177 * timed out, signals are enabled allowing completion with
2179 * I don't really care if we go once more round the loop in
2180 * the error cases -eeb. */
2181 if (rc
== 0 && atomic_read(&set
->set_remaining
) == 0) {
2182 list_for_each(tmp
, &set
->set_requests
) {
2183 req
= list_entry(tmp
, struct ptlrpc_request
,
2185 spin_lock(&req
->rq_lock
);
2186 req
->rq_invalid_rqset
= 1;
2187 spin_unlock(&req
->rq_lock
);
2190 } while (rc
!= 0 || atomic_read(&set
->set_remaining
) != 0);
2192 LASSERT(atomic_read(&set
->set_remaining
) == 0);
2194 rc
= set
->set_rc
; /* rq_status of already freed requests if any */
2195 list_for_each(tmp
, &set
->set_requests
) {
2196 req
= list_entry(tmp
, struct ptlrpc_request
, rq_set_chain
);
2198 LASSERT(req
->rq_phase
== RQ_PHASE_COMPLETE
);
2199 if (req
->rq_status
!= 0)
2200 rc
= req
->rq_status
;
2203 if (set
->set_interpret
!= NULL
) {
2204 int (*interpreter
)(struct ptlrpc_request_set
*set
, void *, int) =
2206 rc
= interpreter(set
, set
->set_arg
, rc
);
2208 struct ptlrpc_set_cbdata
*cbdata
, *n
;
2211 list_for_each_entry_safe(cbdata
, n
,
2212 &set
->set_cblist
, psc_item
) {
2213 list_del_init(&cbdata
->psc_item
);
2214 err
= cbdata
->psc_interpret(set
, cbdata
->psc_data
, rc
);
2223 EXPORT_SYMBOL(ptlrpc_set_wait
);
2226 * Helper function for request freeing.
2227 * Called when request count reached zero and request needs to be freed.
2228 * Removes request from all sorts of sending/replay lists it might be on,
2229 * frees network buffers if any are present.
2230 * If \a locked is set, that means caller is already holding import imp_lock
2231 * and so we no longer need to reobtain it (for certain lists manipulations)
2233 static void __ptlrpc_free_req(struct ptlrpc_request
*request
, int locked
)
2235 if (request
== NULL
)
2237 LASSERTF(!request
->rq_receiving_reply
, "req %p\n", request
);
2238 LASSERTF(request
->rq_rqbd
== NULL
, "req %p\n", request
);/* client-side */
2239 LASSERTF(list_empty(&request
->rq_list
), "req %p\n", request
);
2240 LASSERTF(list_empty(&request
->rq_set_chain
), "req %p\n", request
);
2241 LASSERTF(list_empty(&request
->rq_exp_list
), "req %p\n", request
);
2242 LASSERTF(!request
->rq_replay
, "req %p\n", request
);
2244 req_capsule_fini(&request
->rq_pill
);
2246 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2247 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2248 if (request
->rq_import
!= NULL
) {
2250 spin_lock(&request
->rq_import
->imp_lock
);
2251 list_del_init(&request
->rq_replay_list
);
2253 spin_unlock(&request
->rq_import
->imp_lock
);
2255 LASSERTF(list_empty(&request
->rq_replay_list
), "req %p\n", request
);
2257 if (atomic_read(&request
->rq_refcount
) != 0) {
2258 DEBUG_REQ(D_ERROR
, request
,
2259 "freeing request with nonzero refcount");
2263 if (request
->rq_repbuf
!= NULL
)
2264 sptlrpc_cli_free_repbuf(request
);
2265 if (request
->rq_export
!= NULL
) {
2266 class_export_put(request
->rq_export
);
2267 request
->rq_export
= NULL
;
2269 if (request
->rq_import
!= NULL
) {
2270 class_import_put(request
->rq_import
);
2271 request
->rq_import
= NULL
;
2273 if (request
->rq_bulk
!= NULL
)
2274 ptlrpc_free_bulk_pin(request
->rq_bulk
);
2276 if (request
->rq_reqbuf
!= NULL
|| request
->rq_clrbuf
!= NULL
)
2277 sptlrpc_cli_free_reqbuf(request
);
2279 if (request
->rq_cli_ctx
)
2280 sptlrpc_req_put_ctx(request
, !locked
);
2282 if (request
->rq_pool
)
2283 __ptlrpc_free_req_to_pool(request
);
2285 ptlrpc_request_cache_free(request
);
2288 static int __ptlrpc_req_finished(struct ptlrpc_request
*request
, int locked
);
2290 * Drop one request reference. Must be called with import imp_lock held.
2291 * When reference count drops to zero, request is freed.
2293 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request
*request
)
2295 assert_spin_locked(&request
->rq_import
->imp_lock
);
2296 (void)__ptlrpc_req_finished(request
, 1);
2298 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock
);
2302 * Drops one reference count for request \a request.
2303 * \a locked set indicates that caller holds import imp_lock.
2304 * Frees the request when reference count reaches zero.
2306 static int __ptlrpc_req_finished(struct ptlrpc_request
*request
, int locked
)
2308 if (request
== NULL
)
2311 if (request
== LP_POISON
||
2312 request
->rq_reqmsg
== LP_POISON
) {
2313 CERROR("dereferencing freed request (bug 575)\n");
2318 DEBUG_REQ(D_INFO
, request
, "refcount now %u",
2319 atomic_read(&request
->rq_refcount
) - 1);
2321 if (atomic_dec_and_test(&request
->rq_refcount
)) {
2322 __ptlrpc_free_req(request
, locked
);
2330 * Drops one reference count for a request.
2332 void ptlrpc_req_finished(struct ptlrpc_request
*request
)
2334 __ptlrpc_req_finished(request
, 0);
2336 EXPORT_SYMBOL(ptlrpc_req_finished
);
2339 * Returns xid of a \a request
2341 __u64
ptlrpc_req_xid(struct ptlrpc_request
*request
)
2343 return request
->rq_xid
;
2345 EXPORT_SYMBOL(ptlrpc_req_xid
);
2348 * Disengage the client's reply buffer from the network
2349 * NB does _NOT_ unregister any client-side bulk.
2350 * IDEMPOTENT, but _not_ safe against concurrent callers.
2351 * The request owner (i.e. the thread doing the I/O) must call...
2352 * Returns 0 on success or 1 if unregistering cannot be made.
2354 int ptlrpc_unregister_reply(struct ptlrpc_request
*request
, int async
)
2357 wait_queue_head_t
*wq
;
2358 struct l_wait_info lwi
;
2363 LASSERT(!in_interrupt());
2366 * Let's setup deadline for reply unlink.
2368 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2369 async
&& request
->rq_reply_deadline
== 0)
2370 request
->rq_reply_deadline
= get_seconds()+LONG_UNLINK
;
2373 * Nothing left to do.
2375 if (!ptlrpc_client_recv_or_unlink(request
))
2378 LNetMDUnlink(request
->rq_reply_md_h
);
2381 * Let's check it once again.
2383 if (!ptlrpc_client_recv_or_unlink(request
))
2387 * Move to "Unregistering" phase as reply was not unlinked yet.
2389 ptlrpc_rqphase_move(request
, RQ_PHASE_UNREGISTERING
);
2392 * Do not wait for unlink to finish.
2398 * We have to l_wait_event() whatever the result, to give liblustre
2399 * a chance to run reply_in_callback(), and to make sure we've
2400 * unlinked before returning a req to the pool.
2402 if (request
->rq_set
!= NULL
)
2403 wq
= &request
->rq_set
->set_waitq
;
2405 wq
= &request
->rq_reply_waitq
;
2408 /* Network access will complete in finite time but the HUGE
2409 * timeout lets us CWARN for visibility of sluggish NALs */
2410 lwi
= LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK
),
2411 cfs_time_seconds(1), NULL
, NULL
);
2412 rc
= l_wait_event(*wq
, !ptlrpc_client_recv_or_unlink(request
),
2415 ptlrpc_rqphase_move(request
, request
->rq_next_phase
);
2419 LASSERT(rc
== -ETIMEDOUT
);
2420 DEBUG_REQ(D_WARNING
, request
,
2421 "Unexpectedly long timeout rvcng=%d unlnk=%d/%d",
2422 request
->rq_receiving_reply
,
2423 request
->rq_req_unlink
, request
->rq_reply_unlink
);
2427 EXPORT_SYMBOL(ptlrpc_unregister_reply
);
2429 static void ptlrpc_free_request(struct ptlrpc_request
*req
)
2431 spin_lock(&req
->rq_lock
);
2433 spin_unlock(&req
->rq_lock
);
2435 if (req
->rq_commit_cb
!= NULL
)
2436 req
->rq_commit_cb(req
);
2437 list_del_init(&req
->rq_replay_list
);
2439 __ptlrpc_req_finished(req
, 1);
2443 * the request is committed and dropped from the replay list of its import
2445 void ptlrpc_request_committed(struct ptlrpc_request
*req
, int force
)
2447 struct obd_import
*imp
= req
->rq_import
;
2449 spin_lock(&imp
->imp_lock
);
2450 if (list_empty(&req
->rq_replay_list
)) {
2451 spin_unlock(&imp
->imp_lock
);
2455 if (force
|| req
->rq_transno
<= imp
->imp_peer_committed_transno
)
2456 ptlrpc_free_request(req
);
2458 spin_unlock(&imp
->imp_lock
);
2460 EXPORT_SYMBOL(ptlrpc_request_committed
);
2463 * Iterates through replay_list on import and prunes
2464 * all requests have transno smaller than last_committed for the
2465 * import and don't have rq_replay set.
2466 * Since requests are sorted in transno order, stops when meeting first
2467 * transno bigger than last_committed.
2468 * caller must hold imp->imp_lock
2470 void ptlrpc_free_committed(struct obd_import
*imp
)
2472 struct ptlrpc_request
*req
, *saved
;
2473 struct ptlrpc_request
*last_req
= NULL
; /* temporary fire escape */
2474 bool skip_committed_list
= true;
2476 LASSERT(imp
!= NULL
);
2477 assert_spin_locked(&imp
->imp_lock
);
2479 if (imp
->imp_peer_committed_transno
== imp
->imp_last_transno_checked
&&
2480 imp
->imp_generation
== imp
->imp_last_generation_checked
) {
2481 CDEBUG(D_INFO
, "%s: skip recheck: last_committed %llu\n",
2482 imp
->imp_obd
->obd_name
, imp
->imp_peer_committed_transno
);
2485 CDEBUG(D_RPCTRACE
, "%s: committing for last_committed %llu gen %d\n",
2486 imp
->imp_obd
->obd_name
, imp
->imp_peer_committed_transno
,
2487 imp
->imp_generation
);
2489 if (imp
->imp_generation
!= imp
->imp_last_generation_checked
)
2490 skip_committed_list
= false;
2492 imp
->imp_last_transno_checked
= imp
->imp_peer_committed_transno
;
2493 imp
->imp_last_generation_checked
= imp
->imp_generation
;
2495 list_for_each_entry_safe(req
, saved
, &imp
->imp_replay_list
,
2497 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2498 LASSERT(req
!= last_req
);
2501 if (req
->rq_transno
== 0) {
2502 DEBUG_REQ(D_EMERG
, req
, "zero transno during replay");
2505 if (req
->rq_import_generation
< imp
->imp_generation
) {
2506 DEBUG_REQ(D_RPCTRACE
, req
, "free request with old gen");
2510 /* not yet committed */
2511 if (req
->rq_transno
> imp
->imp_peer_committed_transno
) {
2512 DEBUG_REQ(D_RPCTRACE
, req
, "stopping search");
2516 if (req
->rq_replay
) {
2517 DEBUG_REQ(D_RPCTRACE
, req
, "keeping (FL_REPLAY)");
2518 list_move_tail(&req
->rq_replay_list
,
2519 &imp
->imp_committed_list
);
2523 DEBUG_REQ(D_INFO
, req
, "commit (last_committed %llu)",
2524 imp
->imp_peer_committed_transno
);
2526 ptlrpc_free_request(req
);
2528 if (skip_committed_list
)
2531 list_for_each_entry_safe(req
, saved
, &imp
->imp_committed_list
,
2533 LASSERT(req
->rq_transno
!= 0);
2534 if (req
->rq_import_generation
< imp
->imp_generation
) {
2535 DEBUG_REQ(D_RPCTRACE
, req
, "free stale open request");
2536 ptlrpc_free_request(req
);
2541 void ptlrpc_cleanup_client(struct obd_import
*imp
)
2544 EXPORT_SYMBOL(ptlrpc_cleanup_client
);
2547 * Schedule previously sent request for resend.
2548 * For bulk requests we assign new xid (to avoid problems with
2549 * lost replies and therefore several transfers landing into same buffer
2550 * from different sending attempts).
2552 void ptlrpc_resend_req(struct ptlrpc_request
*req
)
2554 DEBUG_REQ(D_HA
, req
, "going to resend");
2555 spin_lock(&req
->rq_lock
);
2557 /* Request got reply but linked to the import list still.
2558 Let ptlrpc_check_set() to process it. */
2559 if (ptlrpc_client_replied(req
)) {
2560 spin_unlock(&req
->rq_lock
);
2561 DEBUG_REQ(D_HA
, req
, "it has reply, so skip it");
2565 lustre_msg_set_handle(req
->rq_reqmsg
, &(struct lustre_handle
){ 0 });
2566 req
->rq_status
= -EAGAIN
;
2569 req
->rq_net_err
= 0;
2570 req
->rq_timedout
= 0;
2572 __u64 old_xid
= req
->rq_xid
;
2574 /* ensure previous bulk fails */
2575 req
->rq_xid
= ptlrpc_next_xid();
2576 CDEBUG(D_HA
, "resend bulk old x%llu new x%llu\n",
2577 old_xid
, req
->rq_xid
);
2579 ptlrpc_client_wake_req(req
);
2580 spin_unlock(&req
->rq_lock
);
2582 EXPORT_SYMBOL(ptlrpc_resend_req
);
2584 /* XXX: this function and rq_status are currently unused */
2585 void ptlrpc_restart_req(struct ptlrpc_request
*req
)
2587 DEBUG_REQ(D_HA
, req
, "restarting (possibly-)completed request");
2588 req
->rq_status
= -ERESTARTSYS
;
2590 spin_lock(&req
->rq_lock
);
2591 req
->rq_restart
= 1;
2592 req
->rq_timedout
= 0;
2593 ptlrpc_client_wake_req(req
);
2594 spin_unlock(&req
->rq_lock
);
2596 EXPORT_SYMBOL(ptlrpc_restart_req
);
2599 * Grab additional reference on a request \a req
2601 struct ptlrpc_request
*ptlrpc_request_addref(struct ptlrpc_request
*req
)
2603 atomic_inc(&req
->rq_refcount
);
2606 EXPORT_SYMBOL(ptlrpc_request_addref
);
2609 * Add a request to import replay_list.
2610 * Must be called under imp_lock
2612 void ptlrpc_retain_replayable_request(struct ptlrpc_request
*req
,
2613 struct obd_import
*imp
)
2615 struct list_head
*tmp
;
2617 assert_spin_locked(&imp
->imp_lock
);
2619 if (req
->rq_transno
== 0) {
2620 DEBUG_REQ(D_EMERG
, req
, "saving request with zero transno");
2624 /* clear this for new requests that were resent as well
2625 as resent replayed requests. */
2626 lustre_msg_clear_flags(req
->rq_reqmsg
, MSG_RESENT
);
2628 /* don't re-add requests that have been replayed */
2629 if (!list_empty(&req
->rq_replay_list
))
2632 lustre_msg_add_flags(req
->rq_reqmsg
, MSG_REPLAY
);
2634 LASSERT(imp
->imp_replayable
);
2635 /* Balanced in ptlrpc_free_committed, usually. */
2636 ptlrpc_request_addref(req
);
2637 list_for_each_prev(tmp
, &imp
->imp_replay_list
) {
2638 struct ptlrpc_request
*iter
=
2639 list_entry(tmp
, struct ptlrpc_request
,
2642 /* We may have duplicate transnos if we create and then
2643 * open a file, or for closes retained if to match creating
2644 * opens, so use req->rq_xid as a secondary key.
2645 * (See bugs 684, 685, and 428.)
2646 * XXX no longer needed, but all opens need transnos!
2648 if (iter
->rq_transno
> req
->rq_transno
)
2651 if (iter
->rq_transno
== req
->rq_transno
) {
2652 LASSERT(iter
->rq_xid
!= req
->rq_xid
);
2653 if (iter
->rq_xid
> req
->rq_xid
)
2657 list_add(&req
->rq_replay_list
, &iter
->rq_replay_list
);
2661 list_add(&req
->rq_replay_list
, &imp
->imp_replay_list
);
2663 EXPORT_SYMBOL(ptlrpc_retain_replayable_request
);
2666 * Send request and wait until it completes.
2667 * Returns request processing status.
2669 int ptlrpc_queue_wait(struct ptlrpc_request
*req
)
2671 struct ptlrpc_request_set
*set
;
2674 LASSERT(req
->rq_set
== NULL
);
2675 LASSERT(!req
->rq_receiving_reply
);
2677 set
= ptlrpc_prep_set();
2679 CERROR("Unable to allocate ptlrpc set.");
2683 /* for distributed debugging */
2684 lustre_msg_set_status(req
->rq_reqmsg
, current_pid());
2686 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2687 ptlrpc_request_addref(req
);
2688 ptlrpc_set_add_req(set
, req
);
2689 rc
= ptlrpc_set_wait(set
);
2690 ptlrpc_set_destroy(set
);
2694 EXPORT_SYMBOL(ptlrpc_queue_wait
);
2696 struct ptlrpc_replay_async_args
{
2698 int praa_old_status
;
2702 * Callback used for replayed requests reply processing.
2703 * In case of successful reply calls registered request replay callback.
2704 * In case of error restart replay process.
2706 static int ptlrpc_replay_interpret(const struct lu_env
*env
,
2707 struct ptlrpc_request
*req
,
2710 struct ptlrpc_replay_async_args
*aa
= data
;
2711 struct obd_import
*imp
= req
->rq_import
;
2713 atomic_dec(&imp
->imp_replay_inflight
);
2715 if (!ptlrpc_client_replied(req
)) {
2716 CERROR("request replay timed out, restarting recovery\n");
2721 if (lustre_msg_get_type(req
->rq_repmsg
) == PTL_RPC_MSG_ERR
&&
2722 (lustre_msg_get_status(req
->rq_repmsg
) == -ENOTCONN
||
2723 lustre_msg_get_status(req
->rq_repmsg
) == -ENODEV
)) {
2724 rc
= lustre_msg_get_status(req
->rq_repmsg
);
2728 /** VBR: check version failure */
2729 if (lustre_msg_get_status(req
->rq_repmsg
) == -EOVERFLOW
) {
2730 /** replay was failed due to version mismatch */
2731 DEBUG_REQ(D_WARNING
, req
, "Version mismatch during replay\n");
2732 spin_lock(&imp
->imp_lock
);
2733 imp
->imp_vbr_failed
= 1;
2734 imp
->imp_no_lock_replay
= 1;
2735 spin_unlock(&imp
->imp_lock
);
2736 lustre_msg_set_status(req
->rq_repmsg
, aa
->praa_old_status
);
2738 /** The transno had better not change over replay. */
2739 LASSERTF(lustre_msg_get_transno(req
->rq_reqmsg
) ==
2740 lustre_msg_get_transno(req
->rq_repmsg
) ||
2741 lustre_msg_get_transno(req
->rq_repmsg
) == 0,
2743 lustre_msg_get_transno(req
->rq_reqmsg
),
2744 lustre_msg_get_transno(req
->rq_repmsg
));
2747 spin_lock(&imp
->imp_lock
);
2748 /** if replays by version then gap occur on server, no trust to locks */
2749 if (lustre_msg_get_flags(req
->rq_repmsg
) & MSG_VERSION_REPLAY
)
2750 imp
->imp_no_lock_replay
= 1;
2751 imp
->imp_last_replay_transno
= lustre_msg_get_transno(req
->rq_reqmsg
);
2752 spin_unlock(&imp
->imp_lock
);
2753 LASSERT(imp
->imp_last_replay_transno
);
2755 /* transaction number shouldn't be bigger than the latest replayed */
2756 if (req
->rq_transno
> lustre_msg_get_transno(req
->rq_reqmsg
)) {
2757 DEBUG_REQ(D_ERROR
, req
,
2758 "Reported transno %llu is bigger than the replayed one: %llu",
2760 lustre_msg_get_transno(req
->rq_reqmsg
));
2765 DEBUG_REQ(D_HA
, req
, "got rep");
2767 /* let the callback do fixups, possibly including in the request */
2768 if (req
->rq_replay_cb
)
2769 req
->rq_replay_cb(req
);
2771 if (ptlrpc_client_replied(req
) &&
2772 lustre_msg_get_status(req
->rq_repmsg
) != aa
->praa_old_status
) {
2773 DEBUG_REQ(D_ERROR
, req
, "status %d, old was %d",
2774 lustre_msg_get_status(req
->rq_repmsg
),
2775 aa
->praa_old_status
);
2777 /* Put it back for re-replay. */
2778 lustre_msg_set_status(req
->rq_repmsg
, aa
->praa_old_status
);
2782 * Errors while replay can set transno to 0, but
2783 * imp_last_replay_transno shouldn't be set to 0 anyway
2785 if (req
->rq_transno
== 0)
2786 CERROR("Transno is 0 during replay!\n");
2788 /* continue with recovery */
2789 rc
= ptlrpc_import_recovery_state_machine(imp
);
2791 req
->rq_send_state
= aa
->praa_old_state
;
2794 /* this replay failed, so restart recovery */
2795 ptlrpc_connect_import(imp
);
2801 * Prepares and queues request for replay.
2802 * Adds it to ptlrpcd queue for actual sending.
2803 * Returns 0 on success.
2805 int ptlrpc_replay_req(struct ptlrpc_request
*req
)
2807 struct ptlrpc_replay_async_args
*aa
;
2809 LASSERT(req
->rq_import
->imp_state
== LUSTRE_IMP_REPLAY
);
2811 LASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2812 aa
= ptlrpc_req_async_args(req
);
2813 memset(aa
, 0, sizeof(*aa
));
2815 /* Prepare request to be resent with ptlrpcd */
2816 aa
->praa_old_state
= req
->rq_send_state
;
2817 req
->rq_send_state
= LUSTRE_IMP_REPLAY
;
2818 req
->rq_phase
= RQ_PHASE_NEW
;
2819 req
->rq_next_phase
= RQ_PHASE_UNDEFINED
;
2821 aa
->praa_old_status
= lustre_msg_get_status(req
->rq_repmsg
);
2823 req
->rq_interpret_reply
= ptlrpc_replay_interpret
;
2824 /* Readjust the timeout for current conditions */
2825 ptlrpc_at_set_req_timeout(req
);
2827 /* Tell server the net_latency, so the server can calculate how long
2828 * it should wait for next replay */
2829 lustre_msg_set_service_time(req
->rq_reqmsg
,
2830 ptlrpc_at_get_net_latency(req
));
2831 DEBUG_REQ(D_HA
, req
, "REPLAY");
2833 atomic_inc(&req
->rq_import
->imp_replay_inflight
);
2834 ptlrpc_request_addref(req
); /* ptlrpcd needs a ref */
2836 ptlrpcd_add_req(req
, PDL_POLICY_LOCAL
, -1);
2839 EXPORT_SYMBOL(ptlrpc_replay_req
);
2842 * Aborts all in-flight request on import \a imp sending and delayed lists
2844 void ptlrpc_abort_inflight(struct obd_import
*imp
)
2846 struct list_head
*tmp
, *n
;
2848 /* Make sure that no new requests get processed for this import.
2849 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2850 * this flag and then putting requests on sending_list or delayed_list.
2852 spin_lock(&imp
->imp_lock
);
2854 /* XXX locking? Maybe we should remove each request with the list
2855 * locked? Also, how do we know if the requests on the list are
2856 * being freed at this time?
2858 list_for_each_safe(tmp
, n
, &imp
->imp_sending_list
) {
2859 struct ptlrpc_request
*req
=
2860 list_entry(tmp
, struct ptlrpc_request
, rq_list
);
2862 DEBUG_REQ(D_RPCTRACE
, req
, "inflight");
2864 spin_lock(&req
->rq_lock
);
2865 if (req
->rq_import_generation
< imp
->imp_generation
) {
2867 req
->rq_status
= -EIO
;
2868 ptlrpc_client_wake_req(req
);
2870 spin_unlock(&req
->rq_lock
);
2873 list_for_each_safe(tmp
, n
, &imp
->imp_delayed_list
) {
2874 struct ptlrpc_request
*req
=
2875 list_entry(tmp
, struct ptlrpc_request
, rq_list
);
2877 DEBUG_REQ(D_RPCTRACE
, req
, "aborting waiting req");
2879 spin_lock(&req
->rq_lock
);
2880 if (req
->rq_import_generation
< imp
->imp_generation
) {
2882 req
->rq_status
= -EIO
;
2883 ptlrpc_client_wake_req(req
);
2885 spin_unlock(&req
->rq_lock
);
2888 /* Last chance to free reqs left on the replay list, but we
2889 * will still leak reqs that haven't committed. */
2890 if (imp
->imp_replayable
)
2891 ptlrpc_free_committed(imp
);
2893 spin_unlock(&imp
->imp_lock
);
2895 EXPORT_SYMBOL(ptlrpc_abort_inflight
);
2898 * Abort all uncompleted requests in request set \a set
2900 void ptlrpc_abort_set(struct ptlrpc_request_set
*set
)
2902 struct list_head
*tmp
, *pos
;
2904 LASSERT(set
!= NULL
);
2906 list_for_each_safe(pos
, tmp
, &set
->set_requests
) {
2907 struct ptlrpc_request
*req
=
2908 list_entry(pos
, struct ptlrpc_request
,
2911 spin_lock(&req
->rq_lock
);
2912 if (req
->rq_phase
!= RQ_PHASE_RPC
) {
2913 spin_unlock(&req
->rq_lock
);
2918 req
->rq_status
= -EINTR
;
2919 ptlrpc_client_wake_req(req
);
2920 spin_unlock(&req
->rq_lock
);
2924 static __u64 ptlrpc_last_xid
;
2925 static spinlock_t ptlrpc_last_xid_lock
;
2928 * Initialize the XID for the node. This is common among all requests on
2929 * this node, and only requires the property that it is monotonically
2930 * increasing. It does not need to be sequential. Since this is also used
2931 * as the RDMA match bits, it is important that a single client NOT have
2932 * the same match bits for two different in-flight requests, hence we do
2933 * NOT want to have an XID per target or similar.
2935 * To avoid an unlikely collision between match bits after a client reboot
2936 * (which would deliver old data into the wrong RDMA buffer) initialize
2937 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2938 * If the time is clearly incorrect, we instead use a 62-bit random number.
2939 * In the worst case the random number will overflow 1M RPCs per second in
2940 * 9133 years, or permutations thereof.
2942 #define YEAR_2004 (1ULL << 30)
2943 void ptlrpc_init_xid(void)
2945 time_t now
= get_seconds();
2947 spin_lock_init(&ptlrpc_last_xid_lock
);
2948 if (now
< YEAR_2004
) {
2949 cfs_get_random_bytes(&ptlrpc_last_xid
, sizeof(ptlrpc_last_xid
));
2950 ptlrpc_last_xid
>>= 2;
2951 ptlrpc_last_xid
|= (1ULL << 61);
2953 ptlrpc_last_xid
= (__u64
)now
<< 20;
2956 /* Always need to be aligned to a power-of-two for multi-bulk BRW */
2957 CLASSERT((PTLRPC_BULK_OPS_COUNT
& (PTLRPC_BULK_OPS_COUNT
- 1)) == 0);
2958 ptlrpc_last_xid
&= PTLRPC_BULK_OPS_MASK
;
2962 * Increase xid and returns resulting new value to the caller.
2964 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2965 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2966 * itself uses the last bulk xid needed, so the server can determine the
2967 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2968 * xid must align to a power-of-two value.
2970 * This is assumed to be true due to the initial ptlrpc_last_xid
2971 * value also being initialized to a power-of-two value. LU-1431
2973 __u64
ptlrpc_next_xid(void)
2977 spin_lock(&ptlrpc_last_xid_lock
);
2978 next
= ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
2979 ptlrpc_last_xid
= next
;
2980 spin_unlock(&ptlrpc_last_xid_lock
);
2984 EXPORT_SYMBOL(ptlrpc_next_xid
);
2987 * Get a glimpse at what next xid value might have been.
2988 * Returns possible next xid.
2990 __u64
ptlrpc_sample_next_xid(void)
2992 #if BITS_PER_LONG == 32
2993 /* need to avoid possible word tearing on 32-bit systems */
2996 spin_lock(&ptlrpc_last_xid_lock
);
2997 next
= ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
2998 spin_unlock(&ptlrpc_last_xid_lock
);
3002 /* No need to lock, since returned value is racy anyways */
3003 return ptlrpc_last_xid
+ PTLRPC_BULK_OPS_COUNT
;
3006 EXPORT_SYMBOL(ptlrpc_sample_next_xid
);
3009 * Functions for operating ptlrpc workers.
3011 * A ptlrpc work is a function which will be running inside ptlrpc context.
3012 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3014 * 1. after a work is created, it can be used many times, that is:
3015 * handler = ptlrpcd_alloc_work();
3016 * ptlrpcd_queue_work();
3018 * queue it again when necessary:
3019 * ptlrpcd_queue_work();
3020 * ptlrpcd_destroy_work();
3021 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3022 * it will only be queued once in any time. Also as its name implies, it may
3023 * have delay before it really runs by ptlrpcd thread.
3025 struct ptlrpc_work_async_args
{
3026 int (*cb
)(const struct lu_env
*, void *);
3030 static void ptlrpcd_add_work_req(struct ptlrpc_request
*req
)
3032 /* re-initialize the req */
3033 req
->rq_timeout
= obd_timeout
;
3034 req
->rq_sent
= get_seconds();
3035 req
->rq_deadline
= req
->rq_sent
+ req
->rq_timeout
;
3036 req
->rq_reply_deadline
= req
->rq_deadline
;
3037 req
->rq_phase
= RQ_PHASE_INTERPRET
;
3038 req
->rq_next_phase
= RQ_PHASE_COMPLETE
;
3039 req
->rq_xid
= ptlrpc_next_xid();
3040 req
->rq_import_generation
= req
->rq_import
->imp_generation
;
3042 ptlrpcd_add_req(req
, PDL_POLICY_ROUND
, -1);
3045 static int work_interpreter(const struct lu_env
*env
,
3046 struct ptlrpc_request
*req
, void *data
, int rc
)
3048 struct ptlrpc_work_async_args
*arg
= data
;
3050 LASSERT(ptlrpcd_check_work(req
));
3051 LASSERT(arg
->cb
!= NULL
);
3053 rc
= arg
->cb(env
, arg
->cbdata
);
3055 list_del_init(&req
->rq_set_chain
);
3058 if (atomic_dec_return(&req
->rq_refcount
) > 1) {
3059 atomic_set(&req
->rq_refcount
, 2);
3060 ptlrpcd_add_work_req(req
);
3065 static int worker_format
;
3067 static int ptlrpcd_check_work(struct ptlrpc_request
*req
)
3069 return req
->rq_pill
.rc_fmt
== (void *)&worker_format
;
3073 * Create a work for ptlrpc.
3075 void *ptlrpcd_alloc_work(struct obd_import
*imp
,
3076 int (*cb
)(const struct lu_env
*, void *), void *cbdata
)
3078 struct ptlrpc_request
*req
= NULL
;
3079 struct ptlrpc_work_async_args
*args
;
3084 return ERR_PTR(-EINVAL
);
3086 /* copy some code from deprecated fakereq. */
3087 req
= ptlrpc_request_cache_alloc(GFP_NOFS
);
3089 CERROR("ptlrpc: run out of memory!\n");
3090 return ERR_PTR(-ENOMEM
);
3093 req
->rq_send_state
= LUSTRE_IMP_FULL
;
3094 req
->rq_type
= PTL_RPC_MSG_REQUEST
;
3095 req
->rq_import
= class_import_get(imp
);
3096 req
->rq_export
= NULL
;
3097 req
->rq_interpret_reply
= work_interpreter
;
3098 /* don't want reply */
3099 req
->rq_receiving_reply
= 0;
3100 req
->rq_req_unlink
= req
->rq_reply_unlink
= 0;
3101 req
->rq_no_delay
= req
->rq_no_resend
= 1;
3102 req
->rq_pill
.rc_fmt
= (void *)&worker_format
;
3104 spin_lock_init(&req
->rq_lock
);
3105 INIT_LIST_HEAD(&req
->rq_list
);
3106 INIT_LIST_HEAD(&req
->rq_replay_list
);
3107 INIT_LIST_HEAD(&req
->rq_set_chain
);
3108 INIT_LIST_HEAD(&req
->rq_history_list
);
3109 INIT_LIST_HEAD(&req
->rq_exp_list
);
3110 init_waitqueue_head(&req
->rq_reply_waitq
);
3111 init_waitqueue_head(&req
->rq_set_waitq
);
3112 atomic_set(&req
->rq_refcount
, 1);
3114 CLASSERT(sizeof(*args
) <= sizeof(req
->rq_async_args
));
3115 args
= ptlrpc_req_async_args(req
);
3117 args
->cbdata
= cbdata
;
3121 EXPORT_SYMBOL(ptlrpcd_alloc_work
);
3123 void ptlrpcd_destroy_work(void *handler
)
3125 struct ptlrpc_request
*req
= handler
;
3128 ptlrpc_req_finished(req
);
3130 EXPORT_SYMBOL(ptlrpcd_destroy_work
);
3132 int ptlrpcd_queue_work(void *handler
)
3134 struct ptlrpc_request
*req
= handler
;
3137 * Check if the req is already being queued.
3139 * Here comes a trick: it lacks a way of checking if a req is being
3140 * processed reliably in ptlrpc. Here I have to use refcount of req
3141 * for this purpose. This is okay because the caller should use this
3142 * req as opaque data. - Jinshan
3144 LASSERT(atomic_read(&req
->rq_refcount
) > 0);
3145 if (atomic_inc_return(&req
->rq_refcount
) == 2)
3146 ptlrpcd_add_work_req(req
);
3149 EXPORT_SYMBOL(ptlrpcd_queue_work
);