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, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <libcfs/libcfs.h>
41 #include <lustre_dlm.h>
42 #include <lustre_net.h>
43 #include <lustre/lustre_user.h>
44 #include <obd_cksum.h>
45 #include <lustre_ha.h>
46 #include <lprocfs_status.h>
47 #include <lustre_ioctl.h>
48 #include <lustre_debug.h>
49 #include <lustre_param.h>
50 #include <lustre_fid.h>
51 #include <obd_class.h>
52 #include "osc_internal.h"
53 #include "osc_cl_internal.h"
55 struct osc_brw_async_args {
61 struct brw_page **aa_ppga;
62 struct client_obd *aa_cli;
63 struct list_head aa_oaps;
64 struct list_head aa_exts;
65 struct cl_req *aa_clerq;
68 #define osc_grant_args osc_brw_async_args
70 struct osc_setattr_args {
72 obd_enqueue_update_f sa_upcall;
76 struct osc_fsync_args {
78 obd_enqueue_update_f fa_upcall;
82 struct osc_enqueue_args {
83 struct obd_export *oa_exp;
87 osc_enqueue_upcall_f oa_upcall;
89 struct ost_lvb *oa_lvb;
90 struct lustre_handle oa_lockh;
91 unsigned int oa_agl:1;
94 static void osc_release_ppga(struct brw_page **ppga, size_t count);
95 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
98 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
100 struct ost_body *body;
102 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
105 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
108 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
111 struct ptlrpc_request *req;
112 struct ost_body *body;
116 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
120 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
122 ptlrpc_request_free(req);
126 osc_pack_req_body(req, oa);
128 ptlrpc_request_set_replen(req);
130 rc = ptlrpc_queue_wait(req);
134 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
136 GOTO(out, rc = -EPROTO);
138 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
139 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
141 oa->o_blksize = cli_brw_size(exp->exp_obd);
142 oa->o_valid |= OBD_MD_FLBLKSZ;
146 ptlrpc_req_finished(req);
151 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
154 struct ptlrpc_request *req;
155 struct ost_body *body;
159 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
161 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
165 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
167 ptlrpc_request_free(req);
171 osc_pack_req_body(req, oa);
173 ptlrpc_request_set_replen(req);
175 rc = ptlrpc_queue_wait(req);
179 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
181 GOTO(out, rc = -EPROTO);
183 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
187 ptlrpc_req_finished(req);
192 static int osc_setattr_interpret(const struct lu_env *env,
193 struct ptlrpc_request *req,
194 struct osc_setattr_args *sa, int rc)
196 struct ost_body *body;
202 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
204 GOTO(out, rc = -EPROTO);
206 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
209 rc = sa->sa_upcall(sa->sa_cookie, rc);
213 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
214 obd_enqueue_update_f upcall, void *cookie,
215 struct ptlrpc_request_set *rqset)
217 struct ptlrpc_request *req;
218 struct osc_setattr_args *sa;
223 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
227 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
229 ptlrpc_request_free(req);
233 osc_pack_req_body(req, oa);
235 ptlrpc_request_set_replen(req);
237 /* do mds to ost setattr asynchronously */
239 /* Do not wait for response. */
240 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
242 req->rq_interpret_reply =
243 (ptlrpc_interpterer_t)osc_setattr_interpret;
245 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
246 sa = ptlrpc_req_async_args(req);
248 sa->sa_upcall = upcall;
249 sa->sa_cookie = cookie;
251 if (rqset == PTLRPCD_SET)
252 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
254 ptlrpc_set_add_req(rqset, req);
260 static int osc_create(const struct lu_env *env, struct obd_export *exp,
263 struct ptlrpc_request *req;
264 struct ost_body *body;
269 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
270 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
272 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
274 GOTO(out, rc = -ENOMEM);
276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
278 ptlrpc_request_free(req);
282 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
285 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
287 ptlrpc_request_set_replen(req);
289 rc = ptlrpc_queue_wait(req);
293 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
295 GOTO(out_req, rc = -EPROTO);
297 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
298 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
300 oa->o_blksize = cli_brw_size(exp->exp_obd);
301 oa->o_valid |= OBD_MD_FLBLKSZ;
303 CDEBUG(D_HA, "transno: "LPD64"\n",
304 lustre_msg_get_transno(req->rq_repmsg));
306 ptlrpc_req_finished(req);
311 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
312 obd_enqueue_update_f upcall, void *cookie,
313 struct ptlrpc_request_set *rqset)
315 struct ptlrpc_request *req;
316 struct osc_setattr_args *sa;
317 struct ost_body *body;
321 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
325 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
327 ptlrpc_request_free(req);
330 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
331 ptlrpc_at_set_req_timeout(req);
333 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
335 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
337 ptlrpc_request_set_replen(req);
339 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
340 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
341 sa = ptlrpc_req_async_args(req);
343 sa->sa_upcall = upcall;
344 sa->sa_cookie = cookie;
345 if (rqset == PTLRPCD_SET)
346 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
348 ptlrpc_set_add_req(rqset, req);
353 static int osc_sync_interpret(const struct lu_env *env,
354 struct ptlrpc_request *req,
357 struct osc_fsync_args *fa = arg;
358 struct ost_body *body;
364 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
366 CERROR ("can't unpack ost_body\n");
367 GOTO(out, rc = -EPROTO);
370 *fa->fa_oa = body->oa;
372 rc = fa->fa_upcall(fa->fa_cookie, rc);
376 int osc_sync_base(struct obd_export *exp, struct obdo *oa,
377 obd_enqueue_update_f upcall, void *cookie,
378 struct ptlrpc_request_set *rqset)
380 struct ptlrpc_request *req;
381 struct ost_body *body;
382 struct osc_fsync_args *fa;
386 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
390 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
392 ptlrpc_request_free(req);
396 /* overload the size and blocks fields in the oa with start/end */
397 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
399 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
401 ptlrpc_request_set_replen(req);
402 req->rq_interpret_reply = osc_sync_interpret;
404 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
405 fa = ptlrpc_req_async_args(req);
407 fa->fa_upcall = upcall;
408 fa->fa_cookie = cookie;
410 if (rqset == PTLRPCD_SET)
411 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
413 ptlrpc_set_add_req(rqset, req);
418 /* Find and cancel locally locks matched by @mode in the resource found by
419 * @objid. Found locks are added into @cancel list. Returns the amount of
420 * locks added to @cancels list. */
421 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
422 struct list_head *cancels,
423 ldlm_mode_t mode, __u64 lock_flags)
425 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
426 struct ldlm_res_id res_id;
427 struct ldlm_resource *res;
431 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
432 * export) but disabled through procfs (flag in NS).
434 * This distinguishes from a case when ELC is not supported originally,
435 * when we still want to cancel locks in advance and just cancel them
436 * locally, without sending any RPC. */
437 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
440 ostid_build_res_name(&oa->o_oi, &res_id);
441 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
445 LDLM_RESOURCE_ADDREF(res);
446 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
447 lock_flags, 0, NULL);
448 LDLM_RESOURCE_DELREF(res);
449 ldlm_resource_putref(res);
453 static int osc_destroy_interpret(const struct lu_env *env,
454 struct ptlrpc_request *req, void *data,
457 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
459 atomic_dec(&cli->cl_destroy_in_flight);
460 wake_up(&cli->cl_destroy_waitq);
464 static int osc_can_send_destroy(struct client_obd *cli)
466 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
467 cli->cl_max_rpcs_in_flight) {
468 /* The destroy request can be sent */
471 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
472 cli->cl_max_rpcs_in_flight) {
474 * The counter has been modified between the two atomic
477 wake_up(&cli->cl_destroy_waitq);
482 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
485 struct client_obd *cli = &exp->exp_obd->u.cli;
486 struct ptlrpc_request *req;
487 struct ost_body *body;
488 struct list_head cancels = LIST_HEAD_INIT(cancels);
493 CDEBUG(D_INFO, "oa NULL\n");
497 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
498 LDLM_FL_DISCARD_DATA);
500 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
502 ldlm_lock_list_put(&cancels, l_bl_ast, count);
506 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
509 ptlrpc_request_free(req);
513 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
514 ptlrpc_at_set_req_timeout(req);
516 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
518 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
520 ptlrpc_request_set_replen(req);
522 req->rq_interpret_reply = osc_destroy_interpret;
523 if (!osc_can_send_destroy(cli)) {
524 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
527 * Wait until the number of on-going destroy RPCs drops
528 * under max_rpc_in_flight
530 l_wait_event_exclusive(cli->cl_destroy_waitq,
531 osc_can_send_destroy(cli), &lwi);
534 /* Do not wait for response */
535 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
539 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
542 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
544 LASSERT(!(oa->o_valid & bits));
547 spin_lock(&cli->cl_loi_list_lock);
548 oa->o_dirty = cli->cl_dirty_pages << PAGE_CACHE_SHIFT;
549 if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
550 cli->cl_dirty_max_pages)) {
551 CERROR("dirty %lu - %lu > dirty_max %lu\n",
552 cli->cl_dirty_pages, cli->cl_dirty_transit,
553 cli->cl_dirty_max_pages);
555 } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
556 atomic_long_read(&obd_dirty_transit_pages) >
557 (obd_max_dirty_pages + 1))) {
558 /* The atomic_read() allowing the atomic_inc() are
559 * not covered by a lock thus they may safely race and trip
560 * this CERROR() unless we add in a small fudge factor (+1). */
561 CERROR("%s: dirty %ld - %ld > system dirty_max %lu\n",
562 cli->cl_import->imp_obd->obd_name,
563 atomic_long_read(&obd_dirty_pages),
564 atomic_long_read(&obd_dirty_transit_pages),
565 obd_max_dirty_pages);
567 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
569 CERROR("dirty %lu - dirty_max %lu too big???\n",
570 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
573 unsigned long max_in_flight = (cli->cl_max_pages_per_rpc <<
575 (cli->cl_max_rpcs_in_flight + 1);
576 oa->o_undirty = max(cli->cl_dirty_max_pages << PAGE_CACHE_SHIFT,
579 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
580 oa->o_dropped = cli->cl_lost_grant;
581 cli->cl_lost_grant = 0;
582 spin_unlock(&cli->cl_loi_list_lock);
583 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
584 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
588 void osc_update_next_shrink(struct client_obd *cli)
590 cli->cl_next_shrink_grant =
591 cfs_time_shift(cli->cl_grant_shrink_interval);
592 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
593 cli->cl_next_shrink_grant);
596 static void __osc_update_grant(struct client_obd *cli, u64 grant)
598 spin_lock(&cli->cl_loi_list_lock);
599 cli->cl_avail_grant += grant;
600 spin_unlock(&cli->cl_loi_list_lock);
603 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
605 if (body->oa.o_valid & OBD_MD_FLGRANT) {
606 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
607 __osc_update_grant(cli, body->oa.o_grant);
611 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
612 u32 keylen, void *key,
613 u32 vallen, void *val,
614 struct ptlrpc_request_set *set);
616 static int osc_shrink_grant_interpret(const struct lu_env *env,
617 struct ptlrpc_request *req,
620 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
621 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
622 struct ost_body *body;
625 __osc_update_grant(cli, oa->o_grant);
629 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
631 osc_update_grant(cli, body);
637 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
639 spin_lock(&cli->cl_loi_list_lock);
640 oa->o_grant = cli->cl_avail_grant / 4;
641 cli->cl_avail_grant -= oa->o_grant;
642 spin_unlock(&cli->cl_loi_list_lock);
643 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
644 oa->o_valid |= OBD_MD_FLFLAGS;
647 oa->o_flags |= OBD_FL_SHRINK_GRANT;
648 osc_update_next_shrink(cli);
651 /* Shrink the current grant, either from some large amount to enough for a
652 * full set of in-flight RPCs, or if we have already shrunk to that limit
653 * then to enough for a single RPC. This avoids keeping more grant than
654 * needed, and avoids shrinking the grant piecemeal. */
655 static int osc_shrink_grant(struct client_obd *cli)
657 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
658 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
660 spin_lock(&cli->cl_loi_list_lock);
661 if (cli->cl_avail_grant <= target_bytes)
662 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
663 spin_unlock(&cli->cl_loi_list_lock);
665 return osc_shrink_grant_to_target(cli, target_bytes);
668 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
671 struct ost_body *body;
674 spin_lock(&cli->cl_loi_list_lock);
675 /* Don't shrink if we are already above or below the desired limit
676 * We don't want to shrink below a single RPC, as that will negatively
677 * impact block allocation and long-term performance. */
678 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
679 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
681 if (target_bytes >= cli->cl_avail_grant) {
682 spin_unlock(&cli->cl_loi_list_lock);
685 spin_unlock(&cli->cl_loi_list_lock);
691 osc_announce_cached(cli, &body->oa, 0);
693 spin_lock(&cli->cl_loi_list_lock);
694 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
695 cli->cl_avail_grant = target_bytes;
696 spin_unlock(&cli->cl_loi_list_lock);
697 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
698 body->oa.o_valid |= OBD_MD_FLFLAGS;
699 body->oa.o_flags = 0;
701 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
702 osc_update_next_shrink(cli);
704 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
705 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
706 sizeof(*body), body, NULL);
708 __osc_update_grant(cli, body->oa.o_grant);
713 static int osc_should_shrink_grant(struct client_obd *client)
715 cfs_time_t time = cfs_time_current();
716 cfs_time_t next_shrink = client->cl_next_shrink_grant;
718 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
719 OBD_CONNECT_GRANT_SHRINK) == 0)
722 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
723 /* Get the current RPC size directly, instead of going via:
724 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
725 * Keep comment here so that it can be found by searching. */
726 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
728 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
729 client->cl_avail_grant > brw_size)
732 osc_update_next_shrink(client);
737 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
739 struct client_obd *client;
741 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
742 if (osc_should_shrink_grant(client))
743 osc_shrink_grant(client);
748 static int osc_add_shrink_grant(struct client_obd *client)
752 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
754 osc_grant_shrink_grant_cb, NULL,
755 &client->cl_grant_shrink_list);
757 CERROR("add grant client %s error %d\n",
758 client->cl_import->imp_obd->obd_name, rc);
761 CDEBUG(D_CACHE, "add grant client %s \n",
762 client->cl_import->imp_obd->obd_name);
763 osc_update_next_shrink(client);
767 static int osc_del_shrink_grant(struct client_obd *client)
769 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
773 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
776 * ocd_grant is the total grant amount we're expect to hold: if we've
777 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
778 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
781 * race is tolerable here: if we're evicted, but imp_state already
782 * left EVICTED state, then cl_dirty_pages must be 0 already.
784 spin_lock(&cli->cl_loi_list_lock);
785 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
786 cli->cl_avail_grant = ocd->ocd_grant;
788 cli->cl_avail_grant = ocd->ocd_grant -
789 (cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
791 if (cli->cl_avail_grant < 0) {
792 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
793 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
794 ocd->ocd_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
795 /* workaround for servers which do not have the patch from
797 cli->cl_avail_grant = ocd->ocd_grant;
800 /* determine the appropriate chunk size used by osc_extent. */
801 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
802 spin_unlock(&cli->cl_loi_list_lock);
804 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
805 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
806 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
808 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
809 list_empty(&cli->cl_grant_shrink_list))
810 osc_add_shrink_grant(cli);
813 /* We assume that the reason this OSC got a short read is because it read
814 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
815 * via the LOV, and it _knows_ it's reading inside the file, it's just that
816 * this stripe never got written at or beyond this stripe offset yet. */
817 static void handle_short_read(int nob_read, size_t page_count,
818 struct brw_page **pga)
823 /* skip bytes read OK */
824 while (nob_read > 0) {
825 LASSERT (page_count > 0);
827 if (pga[i]->count > nob_read) {
828 /* EOF inside this page */
829 ptr = kmap(pga[i]->pg) +
830 (pga[i]->off & ~PAGE_MASK);
831 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
838 nob_read -= pga[i]->count;
843 /* zero remaining pages */
844 while (page_count-- > 0) {
845 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
846 memset(ptr, 0, pga[i]->count);
852 static int check_write_rcs(struct ptlrpc_request *req,
853 int requested_nob, int niocount,
854 size_t page_count, struct brw_page **pga)
859 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
860 sizeof(*remote_rcs) *
862 if (remote_rcs == NULL) {
863 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
867 /* return error if any niobuf was in error */
868 for (i = 0; i < niocount; i++) {
869 if ((int)remote_rcs[i] < 0)
870 return(remote_rcs[i]);
872 if (remote_rcs[i] != 0) {
873 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
874 i, remote_rcs[i], req);
879 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
880 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
881 req->rq_bulk->bd_nob_transferred, requested_nob);
888 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
890 if (p1->flag != p2->flag) {
891 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
892 OBD_BRW_SYNC | OBD_BRW_ASYNC |
893 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
895 /* warn if we try to combine flags that we don't know to be
897 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
898 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
899 "report this at https://jira.hpdd.intel.com/\n",
905 return (p1->off + p1->count == p2->off);
908 static u32 osc_checksum_bulk(int nob, size_t pg_count,
909 struct brw_page **pga, int opc,
910 cksum_type_t cksum_type)
914 struct cfs_crypto_hash_desc *hdesc;
915 unsigned int bufsize;
917 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
919 LASSERT(pg_count > 0);
921 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
923 CERROR("Unable to initialize checksum hash %s\n",
924 cfs_crypto_hash_name(cfs_alg));
925 return PTR_ERR(hdesc);
928 while (nob > 0 && pg_count > 0) {
929 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
931 /* corrupt the data before we compute the checksum, to
932 * simulate an OST->client data error */
933 if (i == 0 && opc == OST_READ &&
934 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
935 unsigned char *ptr = kmap(pga[i]->pg);
936 int off = pga[i]->off & ~PAGE_MASK;
938 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
941 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
942 pga[i]->off & ~PAGE_MASK,
944 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
945 (int)(pga[i]->off & ~PAGE_MASK));
947 nob -= pga[i]->count;
952 bufsize = sizeof(cksum);
953 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
955 /* For sending we only compute the wrong checksum instead
956 * of corrupting the data so it is still correct on a redo */
957 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
964 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
965 u32 page_count, struct brw_page **pga,
966 struct ptlrpc_request **reqp, int resend)
968 struct ptlrpc_request *req;
969 struct ptlrpc_bulk_desc *desc;
970 struct ost_body *body;
971 struct obd_ioobj *ioobj;
972 struct niobuf_remote *niobuf;
973 int niocount, i, requested_nob, opc, rc;
974 struct osc_brw_async_args *aa;
975 struct req_capsule *pill;
976 struct brw_page *pg_prev;
979 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
980 RETURN(-ENOMEM); /* Recoverable */
981 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
982 RETURN(-EINVAL); /* Fatal */
984 if ((cmd & OBD_BRW_WRITE) != 0) {
986 req = ptlrpc_request_alloc_pool(cli->cl_import,
987 cli->cl_import->imp_rq_pool,
991 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
996 for (niocount = i = 1; i < page_count; i++) {
997 if (!can_merge_pages(pga[i - 1], pga[i]))
1001 pill = &req->rq_pill;
1002 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1004 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1005 niocount * sizeof(*niobuf));
1007 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1009 ptlrpc_request_free(req);
1012 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1013 ptlrpc_at_set_req_timeout(req);
1014 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1016 req->rq_no_retry_einprogress = 1;
1018 desc = ptlrpc_prep_bulk_imp(req, page_count,
1019 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1020 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1024 GOTO(out, rc = -ENOMEM);
1025 /* NB request now owns desc and will free it when it gets freed */
1027 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1028 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1029 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1030 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1032 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1034 obdo_to_ioobj(oa, ioobj);
1035 ioobj->ioo_bufcnt = niocount;
1036 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1037 * that might be send for this request. The actual number is decided
1038 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1039 * "max - 1" for old client compatibility sending "0", and also so the
1040 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1041 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1042 LASSERT(page_count > 0);
1044 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1045 struct brw_page *pg = pga[i];
1046 int poff = pg->off & ~PAGE_MASK;
1048 LASSERT(pg->count > 0);
1049 /* make sure there is no gap in the middle of page array */
1050 LASSERTF(page_count == 1 ||
1051 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1052 ergo(i > 0 && i < page_count - 1,
1053 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1054 ergo(i == page_count - 1, poff == 0)),
1055 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1056 i, page_count, pg, pg->off, pg->count);
1057 LASSERTF(i == 0 || pg->off > pg_prev->off,
1058 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1059 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1061 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1062 pg_prev->pg, page_private(pg_prev->pg),
1063 pg_prev->pg->index, pg_prev->off);
1064 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1065 (pg->flag & OBD_BRW_SRVLOCK));
1067 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1068 requested_nob += pg->count;
1070 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1072 niobuf->rnb_len += pg->count;
1074 niobuf->rnb_offset = pg->off;
1075 niobuf->rnb_len = pg->count;
1076 niobuf->rnb_flags = pg->flag;
1081 LASSERTF((void *)(niobuf - niocount) ==
1082 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1083 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1084 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1086 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1088 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1089 body->oa.o_valid |= OBD_MD_FLFLAGS;
1090 body->oa.o_flags = 0;
1092 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1095 if (osc_should_shrink_grant(cli))
1096 osc_shrink_grant_local(cli, &body->oa);
1098 /* size[REQ_REC_OFF] still sizeof (*body) */
1099 if (opc == OST_WRITE) {
1100 if (cli->cl_checksum &&
1101 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1102 /* store cl_cksum_type in a local variable since
1103 * it can be changed via lprocfs */
1104 cksum_type_t cksum_type = cli->cl_cksum_type;
1106 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1107 oa->o_flags &= OBD_FL_LOCAL_MASK;
1108 body->oa.o_flags = 0;
1110 body->oa.o_flags |= cksum_type_pack(cksum_type);
1111 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1112 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1116 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1118 /* save this in 'oa', too, for later checking */
1119 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1120 oa->o_flags |= cksum_type_pack(cksum_type);
1122 /* clear out the checksum flag, in case this is a
1123 * resend but cl_checksum is no longer set. b=11238 */
1124 oa->o_valid &= ~OBD_MD_FLCKSUM;
1126 oa->o_cksum = body->oa.o_cksum;
1127 /* 1 RC per niobuf */
1128 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1129 sizeof(__u32) * niocount);
1131 if (cli->cl_checksum &&
1132 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1133 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1134 body->oa.o_flags = 0;
1135 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1136 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1139 ptlrpc_request_set_replen(req);
1141 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1142 aa = ptlrpc_req_async_args(req);
1144 aa->aa_requested_nob = requested_nob;
1145 aa->aa_nio_count = niocount;
1146 aa->aa_page_count = page_count;
1150 INIT_LIST_HEAD(&aa->aa_oaps);
1153 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1154 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1155 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1156 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1160 ptlrpc_req_finished(req);
1164 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1165 __u32 client_cksum, __u32 server_cksum, int nob,
1166 size_t page_count, struct brw_page **pga,
1167 cksum_type_t client_cksum_type)
1171 cksum_type_t cksum_type;
1173 if (server_cksum == client_cksum) {
1174 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1178 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1180 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1183 if (cksum_type != client_cksum_type)
1184 msg = "the server did not use the checksum type specified in "
1185 "the original request - likely a protocol problem";
1186 else if (new_cksum == server_cksum)
1187 msg = "changed on the client after we checksummed it - "
1188 "likely false positive due to mmap IO (bug 11742)";
1189 else if (new_cksum == client_cksum)
1190 msg = "changed in transit before arrival at OST";
1192 msg = "changed in transit AND doesn't match the original - "
1193 "likely false positive due to mmap IO (bug 11742)";
1195 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1196 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1197 msg, libcfs_nid2str(peer->nid),
1198 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1199 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1200 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1201 POSTID(&oa->o_oi), pga[0]->off,
1202 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1203 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1204 "client csum now %x\n", client_cksum, client_cksum_type,
1205 server_cksum, cksum_type, new_cksum);
1209 /* Note rc enters this function as number of bytes transferred */
1210 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1212 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1213 const lnet_process_id_t *peer =
1214 &req->rq_import->imp_connection->c_peer;
1215 struct client_obd *cli = aa->aa_cli;
1216 struct ost_body *body;
1217 u32 client_cksum = 0;
1220 if (rc < 0 && rc != -EDQUOT) {
1221 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1225 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1226 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1228 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1232 /* set/clear over quota flag for a uid/gid */
1233 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1234 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1235 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1237 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1238 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1240 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1243 osc_update_grant(cli, body);
1248 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1249 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1251 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1253 CERROR("Unexpected +ve rc %d\n", rc);
1256 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1258 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1261 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1262 check_write_checksum(&body->oa, peer, client_cksum,
1263 body->oa.o_cksum, aa->aa_requested_nob,
1264 aa->aa_page_count, aa->aa_ppga,
1265 cksum_type_unpack(aa->aa_oa->o_flags)))
1268 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1269 aa->aa_page_count, aa->aa_ppga);
1273 /* The rest of this function executes only for OST_READs */
1275 /* if unwrap_bulk failed, return -EAGAIN to retry */
1276 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1278 GOTO(out, rc = -EAGAIN);
1280 if (rc > aa->aa_requested_nob) {
1281 CERROR("Unexpected rc %d (%d requested)\n", rc,
1282 aa->aa_requested_nob);
1286 if (rc != req->rq_bulk->bd_nob_transferred) {
1287 CERROR ("Unexpected rc %d (%d transferred)\n",
1288 rc, req->rq_bulk->bd_nob_transferred);
1292 if (rc < aa->aa_requested_nob)
1293 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1295 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1296 static int cksum_counter;
1297 u32 server_cksum = body->oa.o_cksum;
1300 cksum_type_t cksum_type;
1302 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1303 body->oa.o_flags : 0);
1304 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1305 aa->aa_ppga, OST_READ,
1308 if (peer->nid != req->rq_bulk->bd_sender) {
1310 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1313 if (server_cksum != client_cksum) {
1314 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1315 "%s%s%s inode "DFID" object "DOSTID
1316 " extent ["LPU64"-"LPU64"]\n",
1317 req->rq_import->imp_obd->obd_name,
1318 libcfs_nid2str(peer->nid),
1320 body->oa.o_valid & OBD_MD_FLFID ?
1321 body->oa.o_parent_seq : (__u64)0,
1322 body->oa.o_valid & OBD_MD_FLFID ?
1323 body->oa.o_parent_oid : 0,
1324 body->oa.o_valid & OBD_MD_FLFID ?
1325 body->oa.o_parent_ver : 0,
1326 POSTID(&body->oa.o_oi),
1327 aa->aa_ppga[0]->off,
1328 aa->aa_ppga[aa->aa_page_count-1]->off +
1329 aa->aa_ppga[aa->aa_page_count-1]->count -
1331 CERROR("client %x, server %x, cksum_type %x\n",
1332 client_cksum, server_cksum, cksum_type);
1334 aa->aa_oa->o_cksum = client_cksum;
1338 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1341 } else if (unlikely(client_cksum)) {
1342 static int cksum_missed;
1345 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1346 CERROR("Checksum %u requested from %s but not sent\n",
1347 cksum_missed, libcfs_nid2str(peer->nid));
1353 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1354 aa->aa_oa, &body->oa);
1359 static int osc_brw_redo_request(struct ptlrpc_request *request,
1360 struct osc_brw_async_args *aa, int rc)
1362 struct ptlrpc_request *new_req;
1363 struct osc_brw_async_args *new_aa;
1364 struct osc_async_page *oap;
1367 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1368 "redo for recoverable error %d", rc);
1370 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1371 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1372 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1373 aa->aa_ppga, &new_req, 1);
1377 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1378 if (oap->oap_request != NULL) {
1379 LASSERTF(request == oap->oap_request,
1380 "request %p != oap_request %p\n",
1381 request, oap->oap_request);
1382 if (oap->oap_interrupted) {
1383 ptlrpc_req_finished(new_req);
1388 /* New request takes over pga and oaps from old request.
1389 * Note that copying a list_head doesn't work, need to move it... */
1391 new_req->rq_interpret_reply = request->rq_interpret_reply;
1392 new_req->rq_async_args = request->rq_async_args;
1393 new_req->rq_commit_cb = request->rq_commit_cb;
1394 /* cap resend delay to the current request timeout, this is similar to
1395 * what ptlrpc does (see after_reply()) */
1396 if (aa->aa_resends > new_req->rq_timeout)
1397 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1399 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1400 new_req->rq_generation_set = 1;
1401 new_req->rq_import_generation = request->rq_import_generation;
1403 new_aa = ptlrpc_req_async_args(new_req);
1405 INIT_LIST_HEAD(&new_aa->aa_oaps);
1406 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1407 INIT_LIST_HEAD(&new_aa->aa_exts);
1408 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1409 new_aa->aa_resends = aa->aa_resends;
1411 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1412 if (oap->oap_request) {
1413 ptlrpc_req_finished(oap->oap_request);
1414 oap->oap_request = ptlrpc_request_addref(new_req);
1418 /* XXX: This code will run into problem if we're going to support
1419 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1420 * and wait for all of them to be finished. We should inherit request
1421 * set from old request. */
1422 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1424 DEBUG_REQ(D_INFO, new_req, "new request");
1429 * ugh, we want disk allocation on the target to happen in offset order. we'll
1430 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1431 * fine for our small page arrays and doesn't require allocation. its an
1432 * insertion sort that swaps elements that are strides apart, shrinking the
1433 * stride down until its '1' and the array is sorted.
1435 static void sort_brw_pages(struct brw_page **array, int num)
1438 struct brw_page *tmp;
1442 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1447 for (i = stride ; i < num ; i++) {
1450 while (j >= stride && array[j - stride]->off > tmp->off) {
1451 array[j] = array[j - stride];
1456 } while (stride > 1);
1459 static void osc_release_ppga(struct brw_page **ppga, size_t count)
1461 LASSERT(ppga != NULL);
1462 OBD_FREE(ppga, sizeof(*ppga) * count);
1465 static int brw_interpret(const struct lu_env *env,
1466 struct ptlrpc_request *req, void *data, int rc)
1468 struct osc_brw_async_args *aa = data;
1469 struct osc_extent *ext;
1470 struct osc_extent *tmp;
1471 struct client_obd *cli = aa->aa_cli;
1474 rc = osc_brw_fini_request(req, rc);
1475 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1476 /* When server return -EINPROGRESS, client should always retry
1477 * regardless of the number of times the bulk was resent already. */
1478 if (osc_recoverable_error(rc)) {
1479 if (req->rq_import_generation !=
1480 req->rq_import->imp_generation) {
1481 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1482 ""DOSTID", rc = %d.\n",
1483 req->rq_import->imp_obd->obd_name,
1484 POSTID(&aa->aa_oa->o_oi), rc);
1485 } else if (rc == -EINPROGRESS ||
1486 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1487 rc = osc_brw_redo_request(req, aa, rc);
1489 CERROR("%s: too many resent retries for object: "
1490 ""LPU64":"LPU64", rc = %d.\n",
1491 req->rq_import->imp_obd->obd_name,
1492 POSTID(&aa->aa_oa->o_oi), rc);
1497 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1502 struct obdo *oa = aa->aa_oa;
1503 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1504 unsigned long valid = 0;
1505 struct cl_object *obj;
1506 struct osc_async_page *last;
1508 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1509 obj = osc2cl(last->oap_obj);
1511 cl_object_attr_lock(obj);
1512 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1513 attr->cat_blocks = oa->o_blocks;
1514 valid |= CAT_BLOCKS;
1516 if (oa->o_valid & OBD_MD_FLMTIME) {
1517 attr->cat_mtime = oa->o_mtime;
1520 if (oa->o_valid & OBD_MD_FLATIME) {
1521 attr->cat_atime = oa->o_atime;
1524 if (oa->o_valid & OBD_MD_FLCTIME) {
1525 attr->cat_ctime = oa->o_ctime;
1529 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1530 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1531 loff_t last_off = last->oap_count + last->oap_obj_off +
1534 /* Change file size if this is an out of quota or
1535 * direct IO write and it extends the file size */
1536 if (loi->loi_lvb.lvb_size < last_off) {
1537 attr->cat_size = last_off;
1540 /* Extend KMS if it's not a lockless write */
1541 if (loi->loi_kms < last_off &&
1542 oap2osc_page(last)->ops_srvlock == 0) {
1543 attr->cat_kms = last_off;
1549 cl_object_attr_update(env, obj, attr, valid);
1550 cl_object_attr_unlock(obj);
1552 OBDO_FREE(aa->aa_oa);
1554 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
1555 osc_inc_unstable_pages(req);
1557 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1558 list_del_init(&ext->oe_link);
1559 osc_extent_finish(env, ext, 1, rc);
1561 LASSERT(list_empty(&aa->aa_exts));
1562 LASSERT(list_empty(&aa->aa_oaps));
1564 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1565 req->rq_bulk->bd_nob_transferred);
1566 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1567 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1569 spin_lock(&cli->cl_loi_list_lock);
1570 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1571 * is called so we know whether to go to sync BRWs or wait for more
1572 * RPCs to complete */
1573 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1574 cli->cl_w_in_flight--;
1576 cli->cl_r_in_flight--;
1577 osc_wake_cache_waiters(cli);
1578 spin_unlock(&cli->cl_loi_list_lock);
1580 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
1584 static void brw_commit(struct ptlrpc_request *req)
1586 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1587 * this called via the rq_commit_cb, I need to ensure
1588 * osc_dec_unstable_pages is still called. Otherwise unstable
1589 * pages may be leaked. */
1590 spin_lock(&req->rq_lock);
1591 if (likely(req->rq_unstable)) {
1592 req->rq_unstable = 0;
1593 spin_unlock(&req->rq_lock);
1595 osc_dec_unstable_pages(req);
1597 req->rq_committed = 1;
1598 spin_unlock(&req->rq_lock);
1603 * Build an RPC by the list of extent @ext_list. The caller must ensure
1604 * that the total pages in this list are NOT over max pages per RPC.
1605 * Extents in the list must be in OES_RPC state.
1607 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1608 struct list_head *ext_list, int cmd, pdl_policy_t pol)
1610 struct ptlrpc_request *req = NULL;
1611 struct osc_extent *ext;
1612 struct brw_page **pga = NULL;
1613 struct osc_brw_async_args *aa = NULL;
1614 struct obdo *oa = NULL;
1615 struct osc_async_page *oap;
1616 struct osc_async_page *tmp;
1617 struct cl_req *clerq = NULL;
1618 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
1620 struct cl_req_attr *crattr = NULL;
1621 loff_t starting_offset = OBD_OBJECT_EOF;
1622 loff_t ending_offset = 0;
1626 bool soft_sync = false;
1629 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1630 struct ost_body *body;
1632 LASSERT(!list_empty(ext_list));
1634 /* add pages into rpc_list to build BRW rpc */
1635 list_for_each_entry(ext, ext_list, oe_link) {
1636 LASSERT(ext->oe_state == OES_RPC);
1637 mem_tight |= ext->oe_memalloc;
1638 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1640 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1641 if (starting_offset == OBD_OBJECT_EOF ||
1642 starting_offset > oap->oap_obj_off)
1643 starting_offset = oap->oap_obj_off;
1645 LASSERT(oap->oap_page_off == 0);
1646 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1647 ending_offset = oap->oap_obj_off +
1650 LASSERT(oap->oap_page_off + oap->oap_count ==
1655 soft_sync = osc_over_unstable_soft_limit(cli);
1657 mpflag = cfs_memory_pressure_get_and_set();
1659 OBD_ALLOC(crattr, sizeof(*crattr));
1661 GOTO(out, rc = -ENOMEM);
1663 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1665 GOTO(out, rc = -ENOMEM);
1669 GOTO(out, rc = -ENOMEM);
1672 list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1673 struct cl_page *page = oap2cl_page(oap);
1674 if (clerq == NULL) {
1675 clerq = cl_req_alloc(env, page, crt,
1676 1 /* only 1-object rpcs for now */);
1678 GOTO(out, rc = PTR_ERR(clerq));
1681 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1683 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
1684 pga[i] = &oap->oap_brw_page;
1685 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1686 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1687 pga[i]->pg, page_index(oap->oap_page), oap,
1690 cl_req_page_add(env, clerq, page);
1693 /* always get the data for the obdo for the rpc */
1694 LASSERT(clerq != NULL);
1695 crattr->cra_oa = oa;
1696 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1698 rc = cl_req_prep(env, clerq);
1700 CERROR("cl_req_prep failed: %d\n", rc);
1704 sort_brw_pages(pga, page_count);
1705 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
1707 CERROR("prep_req failed: %d\n", rc);
1711 req->rq_commit_cb = brw_commit;
1712 req->rq_interpret_reply = brw_interpret;
1715 req->rq_memalloc = 1;
1717 /* Need to update the timestamps after the request is built in case
1718 * we race with setattr (locally or in queue at OST). If OST gets
1719 * later setattr before earlier BRW (as determined by the request xid),
1720 * the OST will not use BRW timestamps. Sadly, there is no obvious
1721 * way to do this in a single call. bug 10150 */
1722 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1723 crattr->cra_oa = &body->oa;
1724 cl_req_attr_set(env, clerq, crattr,
1725 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
1727 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1729 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1730 aa = ptlrpc_req_async_args(req);
1731 INIT_LIST_HEAD(&aa->aa_oaps);
1732 list_splice_init(&rpc_list, &aa->aa_oaps);
1733 INIT_LIST_HEAD(&aa->aa_exts);
1734 list_splice_init(ext_list, &aa->aa_exts);
1735 aa->aa_clerq = clerq;
1737 /* queued sync pages can be torn down while the pages
1738 * were between the pending list and the rpc */
1740 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1741 /* only one oap gets a request reference */
1744 if (oap->oap_interrupted && !req->rq_intr) {
1745 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1747 ptlrpc_mark_interrupted(req);
1751 tmp->oap_request = ptlrpc_request_addref(req);
1753 spin_lock(&cli->cl_loi_list_lock);
1754 starting_offset >>= PAGE_CACHE_SHIFT;
1755 if (cmd == OBD_BRW_READ) {
1756 cli->cl_r_in_flight++;
1757 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1758 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1759 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
1760 starting_offset + 1);
1762 cli->cl_w_in_flight++;
1763 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1764 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
1765 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
1766 starting_offset + 1);
1768 spin_unlock(&cli->cl_loi_list_lock);
1770 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
1771 page_count, aa, cli->cl_r_in_flight,
1772 cli->cl_w_in_flight);
1774 /* XXX: Maybe the caller can check the RPC bulk descriptor to
1775 * see which CPU/NUMA node the majority of pages were allocated
1776 * on, and try to assign the async RPC to the CPU core
1777 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
1779 * But on the other hand, we expect that multiple ptlrpcd
1780 * threads and the initial write sponsor can run in parallel,
1781 * especially when data checksum is enabled, which is CPU-bound
1782 * operation and single ptlrpcd thread cannot process in time.
1783 * So more ptlrpcd threads sharing BRW load
1784 * (with PDL_POLICY_ROUND) seems better.
1786 ptlrpcd_add_req(req, pol, -1);
1792 cfs_memory_pressure_restore(mpflag);
1795 OBD_FREE(crattr, sizeof(*crattr));
1798 LASSERT(req == NULL);
1803 OBD_FREE(pga, sizeof(*pga) * page_count);
1804 /* this should happen rarely and is pretty bad, it makes the
1805 * pending list not follow the dirty order */
1806 while (!list_empty(ext_list)) {
1807 ext = list_entry(ext_list->next, struct osc_extent,
1809 list_del_init(&ext->oe_link);
1810 osc_extent_finish(env, ext, 0, rc);
1812 if (clerq && !IS_ERR(clerq))
1813 cl_req_completion(env, clerq, rc);
1818 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
1819 struct ldlm_enqueue_info *einfo)
1821 void *data = einfo->ei_cbdata;
1824 LASSERT(lock != NULL);
1825 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
1826 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
1827 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
1828 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
1830 lock_res_and_lock(lock);
1832 if (lock->l_ast_data == NULL)
1833 lock->l_ast_data = data;
1834 if (lock->l_ast_data == data)
1837 unlock_res_and_lock(lock);
1842 static int osc_set_data_with_check(struct lustre_handle *lockh,
1843 struct ldlm_enqueue_info *einfo)
1845 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
1849 set = osc_set_lock_data_with_check(lock, einfo);
1850 LDLM_LOCK_PUT(lock);
1852 CERROR("lockh %p, data %p - client evicted?\n",
1853 lockh, einfo->ei_cbdata);
1857 static int osc_enqueue_fini(struct ptlrpc_request *req,
1858 osc_enqueue_upcall_f upcall, void *cookie,
1859 struct lustre_handle *lockh, ldlm_mode_t mode,
1860 __u64 *flags, int agl, int errcode)
1862 bool intent = *flags & LDLM_FL_HAS_INTENT;
1866 /* The request was created before ldlm_cli_enqueue call. */
1867 if (intent && errcode == ELDLM_LOCK_ABORTED) {
1868 struct ldlm_reply *rep;
1870 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1871 LASSERT(rep != NULL);
1873 rep->lock_policy_res1 =
1874 ptlrpc_status_ntoh(rep->lock_policy_res1);
1875 if (rep->lock_policy_res1)
1876 errcode = rep->lock_policy_res1;
1878 *flags |= LDLM_FL_LVB_READY;
1879 } else if (errcode == ELDLM_OK) {
1880 *flags |= LDLM_FL_LVB_READY;
1883 /* Call the update callback. */
1884 rc = (*upcall)(cookie, lockh, errcode);
1886 /* release the reference taken in ldlm_cli_enqueue() */
1887 if (errcode == ELDLM_LOCK_MATCHED)
1889 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
1890 ldlm_lock_decref(lockh, mode);
1895 static int osc_enqueue_interpret(const struct lu_env *env,
1896 struct ptlrpc_request *req,
1897 struct osc_enqueue_args *aa, int rc)
1899 struct ldlm_lock *lock;
1900 struct lustre_handle *lockh = &aa->oa_lockh;
1901 ldlm_mode_t mode = aa->oa_mode;
1902 struct ost_lvb *lvb = aa->oa_lvb;
1903 __u32 lvb_len = sizeof(*lvb);
1908 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
1910 lock = ldlm_handle2lock(lockh);
1911 LASSERTF(lock != NULL,
1912 "lockh "LPX64", req %p, aa %p - client evicted?\n",
1913 lockh->cookie, req, aa);
1915 /* Take an additional reference so that a blocking AST that
1916 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
1917 * to arrive after an upcall has been executed by
1918 * osc_enqueue_fini(). */
1919 ldlm_lock_addref(lockh, mode);
1921 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
1922 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
1924 /* Let CP AST to grant the lock first. */
1925 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
1928 LASSERT(aa->oa_lvb == NULL);
1929 LASSERT(aa->oa_flags == NULL);
1930 aa->oa_flags = &flags;
1933 /* Complete obtaining the lock procedure. */
1934 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
1935 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
1937 /* Complete osc stuff. */
1938 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
1939 aa->oa_flags, aa->oa_agl, rc);
1941 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
1943 ldlm_lock_decref(lockh, mode);
1944 LDLM_LOCK_PUT(lock);
1948 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
1950 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
1951 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
1952 * other synchronous requests, however keeping some locks and trying to obtain
1953 * others may take a considerable amount of time in a case of ost failure; and
1954 * when other sync requests do not get released lock from a client, the client
1955 * is evicted from the cluster -- such scenarious make the life difficult, so
1956 * release locks just after they are obtained. */
1957 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
1958 __u64 *flags, ldlm_policy_data_t *policy,
1959 struct ost_lvb *lvb, int kms_valid,
1960 osc_enqueue_upcall_f upcall, void *cookie,
1961 struct ldlm_enqueue_info *einfo,
1962 struct ptlrpc_request_set *rqset, int async, int agl)
1964 struct obd_device *obd = exp->exp_obd;
1965 struct lustre_handle lockh = { 0 };
1966 struct ptlrpc_request *req = NULL;
1967 int intent = *flags & LDLM_FL_HAS_INTENT;
1968 __u64 match_lvb = agl ? 0 : LDLM_FL_LVB_READY;
1973 /* Filesystem lock extents are extended to page boundaries so that
1974 * dealing with the page cache is a little smoother. */
1975 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
1976 policy->l_extent.end |= ~PAGE_MASK;
1979 * kms is not valid when either object is completely fresh (so that no
1980 * locks are cached), or object was evicted. In the latter case cached
1981 * lock cannot be used, because it would prime inode state with
1982 * potentially stale LVB.
1987 /* Next, search for already existing extent locks that will cover us */
1988 /* If we're trying to read, we also search for an existing PW lock. The
1989 * VFS and page cache already protect us locally, so lots of readers/
1990 * writers can share a single PW lock.
1992 * There are problems with conversion deadlocks, so instead of
1993 * converting a read lock to a write lock, we'll just enqueue a new
1996 * At some point we should cancel the read lock instead of making them
1997 * send us a blocking callback, but there are problems with canceling
1998 * locks out from other users right now, too. */
1999 mode = einfo->ei_mode;
2000 if (einfo->ei_mode == LCK_PR)
2002 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2003 einfo->ei_type, policy, mode, &lockh, 0);
2005 struct ldlm_lock *matched;
2007 if (*flags & LDLM_FL_TEST_LOCK)
2010 matched = ldlm_handle2lock(&lockh);
2012 /* AGL enqueues DLM locks speculatively. Therefore if
2013 * it already exists a DLM lock, it wll just inform the
2014 * caller to cancel the AGL process for this stripe. */
2015 ldlm_lock_decref(&lockh, mode);
2016 LDLM_LOCK_PUT(matched);
2018 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2019 *flags |= LDLM_FL_LVB_READY;
2021 /* We already have a lock, and it's referenced. */
2022 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2024 ldlm_lock_decref(&lockh, mode);
2025 LDLM_LOCK_PUT(matched);
2028 ldlm_lock_decref(&lockh, mode);
2029 LDLM_LOCK_PUT(matched);
2034 if (*flags & LDLM_FL_TEST_LOCK)
2038 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2039 &RQF_LDLM_ENQUEUE_LVB);
2043 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2045 ptlrpc_request_free(req);
2049 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2051 ptlrpc_request_set_replen(req);
2054 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2055 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2057 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2058 sizeof(*lvb), LVB_T_OST, &lockh, async);
2061 struct osc_enqueue_args *aa;
2062 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2063 aa = ptlrpc_req_async_args(req);
2065 aa->oa_mode = einfo->ei_mode;
2066 aa->oa_type = einfo->ei_type;
2067 lustre_handle_copy(&aa->oa_lockh, &lockh);
2068 aa->oa_upcall = upcall;
2069 aa->oa_cookie = cookie;
2072 aa->oa_flags = flags;
2075 /* AGL is essentially to enqueue an DLM lock
2076 * in advance, so we don't care about the
2077 * result of AGL enqueue. */
2079 aa->oa_flags = NULL;
2082 req->rq_interpret_reply =
2083 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2084 if (rqset == PTLRPCD_SET)
2085 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2087 ptlrpc_set_add_req(rqset, req);
2088 } else if (intent) {
2089 ptlrpc_req_finished(req);
2094 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2097 ptlrpc_req_finished(req);
2102 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2103 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2104 __u64 *flags, void *data, struct lustre_handle *lockh,
2107 struct obd_device *obd = exp->exp_obd;
2108 __u64 lflags = *flags;
2112 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2115 /* Filesystem lock extents are extended to page boundaries so that
2116 * dealing with the page cache is a little smoother */
2117 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2118 policy->l_extent.end |= ~PAGE_MASK;
2120 /* Next, search for already existing extent locks that will cover us */
2121 /* If we're trying to read, we also search for an existing PW lock. The
2122 * VFS and page cache already protect us locally, so lots of readers/
2123 * writers can share a single PW lock. */
2127 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2128 res_id, type, policy, rc, lockh, unref);
2131 if (!osc_set_data_with_check(lockh, data)) {
2132 if (!(lflags & LDLM_FL_TEST_LOCK))
2133 ldlm_lock_decref(lockh, rc);
2137 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2138 ldlm_lock_addref(lockh, LCK_PR);
2139 ldlm_lock_decref(lockh, LCK_PW);
2146 static int osc_statfs_interpret(const struct lu_env *env,
2147 struct ptlrpc_request *req,
2148 struct osc_async_args *aa, int rc)
2150 struct obd_statfs *msfs;
2154 /* The request has in fact never been sent
2155 * due to issues at a higher level (LOV).
2156 * Exit immediately since the caller is
2157 * aware of the problem and takes care
2158 * of the clean up */
2161 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2162 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2168 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2170 GOTO(out, rc = -EPROTO);
2173 *aa->aa_oi->oi_osfs = *msfs;
2175 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2179 static int osc_statfs_async(struct obd_export *exp,
2180 struct obd_info *oinfo, __u64 max_age,
2181 struct ptlrpc_request_set *rqset)
2183 struct obd_device *obd = class_exp2obd(exp);
2184 struct ptlrpc_request *req;
2185 struct osc_async_args *aa;
2189 /* We could possibly pass max_age in the request (as an absolute
2190 * timestamp or a "seconds.usec ago") so the target can avoid doing
2191 * extra calls into the filesystem if that isn't necessary (e.g.
2192 * during mount that would help a bit). Having relative timestamps
2193 * is not so great if request processing is slow, while absolute
2194 * timestamps are not ideal because they need time synchronization. */
2195 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2199 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2201 ptlrpc_request_free(req);
2204 ptlrpc_request_set_replen(req);
2205 req->rq_request_portal = OST_CREATE_PORTAL;
2206 ptlrpc_at_set_req_timeout(req);
2208 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2209 /* procfs requests not want stat in wait for avoid deadlock */
2210 req->rq_no_resend = 1;
2211 req->rq_no_delay = 1;
2214 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2215 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2216 aa = ptlrpc_req_async_args(req);
2219 ptlrpc_set_add_req(rqset, req);
2223 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2224 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2226 struct obd_device *obd = class_exp2obd(exp);
2227 struct obd_statfs *msfs;
2228 struct ptlrpc_request *req;
2229 struct obd_import *imp = NULL;
2233 /*Since the request might also come from lprocfs, so we need
2234 *sync this with client_disconnect_export Bug15684*/
2235 down_read(&obd->u.cli.cl_sem);
2236 if (obd->u.cli.cl_import)
2237 imp = class_import_get(obd->u.cli.cl_import);
2238 up_read(&obd->u.cli.cl_sem);
2242 /* We could possibly pass max_age in the request (as an absolute
2243 * timestamp or a "seconds.usec ago") so the target can avoid doing
2244 * extra calls into the filesystem if that isn't necessary (e.g.
2245 * during mount that would help a bit). Having relative timestamps
2246 * is not so great if request processing is slow, while absolute
2247 * timestamps are not ideal because they need time synchronization. */
2248 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2250 class_import_put(imp);
2255 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2257 ptlrpc_request_free(req);
2260 ptlrpc_request_set_replen(req);
2261 req->rq_request_portal = OST_CREATE_PORTAL;
2262 ptlrpc_at_set_req_timeout(req);
2264 if (flags & OBD_STATFS_NODELAY) {
2265 /* procfs requests not want stat in wait for avoid deadlock */
2266 req->rq_no_resend = 1;
2267 req->rq_no_delay = 1;
2270 rc = ptlrpc_queue_wait(req);
2274 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2276 GOTO(out, rc = -EPROTO);
2283 ptlrpc_req_finished(req);
2287 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2288 void *karg, void *uarg)
2290 struct obd_device *obd = exp->exp_obd;
2291 struct obd_ioctl_data *data = karg;
2295 if (!try_module_get(THIS_MODULE)) {
2296 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2297 module_name(THIS_MODULE));
2301 case OBD_IOC_CLIENT_RECOVER:
2302 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2303 data->ioc_inlbuf1, 0);
2307 case IOC_OSC_SET_ACTIVE:
2308 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2311 case OBD_IOC_PING_TARGET:
2312 err = ptlrpc_obd_ping(obd);
2315 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2316 cmd, current_comm());
2317 GOTO(out, err = -ENOTTY);
2320 module_put(THIS_MODULE);
2324 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2325 u32 keylen, void *key,
2326 u32 vallen, void *val,
2327 struct ptlrpc_request_set *set)
2329 struct ptlrpc_request *req;
2330 struct obd_device *obd = exp->exp_obd;
2331 struct obd_import *imp = class_exp2cliimp(exp);
2336 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2338 if (KEY_IS(KEY_CHECKSUM)) {
2339 if (vallen != sizeof(int))
2341 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2345 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2346 sptlrpc_conf_client_adapt(obd);
2350 if (KEY_IS(KEY_FLUSH_CTX)) {
2351 sptlrpc_import_flush_my_ctx(imp);
2355 if (KEY_IS(KEY_CACHE_SET)) {
2356 struct client_obd *cli = &obd->u.cli;
2358 LASSERT(cli->cl_cache == NULL); /* only once */
2359 cli->cl_cache = (struct cl_client_cache *)val;
2360 cl_cache_incref(cli->cl_cache);
2361 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2363 /* add this osc into entity list */
2364 LASSERT(list_empty(&cli->cl_lru_osc));
2365 spin_lock(&cli->cl_cache->ccc_lru_lock);
2366 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2367 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2372 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2373 struct client_obd *cli = &obd->u.cli;
2374 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2375 long target = *(long *)val;
2377 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2382 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2385 /* We pass all other commands directly to OST. Since nobody calls osc
2386 methods directly and everybody is supposed to go through LOV, we
2387 assume lov checked invalid values for us.
2388 The only recognised values so far are evict_by_nid and mds_conn.
2389 Even if something bad goes through, we'd get a -EINVAL from OST
2392 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2393 &RQF_OST_SET_GRANT_INFO :
2398 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2399 RCL_CLIENT, keylen);
2400 if (!KEY_IS(KEY_GRANT_SHRINK))
2401 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2402 RCL_CLIENT, vallen);
2403 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2405 ptlrpc_request_free(req);
2409 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2410 memcpy(tmp, key, keylen);
2411 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2414 memcpy(tmp, val, vallen);
2416 if (KEY_IS(KEY_GRANT_SHRINK)) {
2417 struct osc_grant_args *aa;
2420 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2421 aa = ptlrpc_req_async_args(req);
2424 ptlrpc_req_finished(req);
2427 *oa = ((struct ost_body *)val)->oa;
2429 req->rq_interpret_reply = osc_shrink_grant_interpret;
2432 ptlrpc_request_set_replen(req);
2433 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2434 LASSERT(set != NULL);
2435 ptlrpc_set_add_req(set, req);
2436 ptlrpc_check_set(NULL, set);
2438 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2443 static int osc_reconnect(const struct lu_env *env,
2444 struct obd_export *exp, struct obd_device *obd,
2445 struct obd_uuid *cluuid,
2446 struct obd_connect_data *data,
2449 struct client_obd *cli = &obd->u.cli;
2451 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2454 spin_lock(&cli->cl_loi_list_lock);
2455 data->ocd_grant = (cli->cl_avail_grant +
2456 (cli->cl_dirty_pages << PAGE_CACHE_SHIFT)) ?:
2457 2 * cli_brw_size(obd);
2458 lost_grant = cli->cl_lost_grant;
2459 cli->cl_lost_grant = 0;
2460 spin_unlock(&cli->cl_loi_list_lock);
2462 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
2463 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2464 data->ocd_version, data->ocd_grant, lost_grant);
2470 static int osc_disconnect(struct obd_export *exp)
2472 struct obd_device *obd = class_exp2obd(exp);
2475 rc = client_disconnect_export(exp);
2477 * Initially we put del_shrink_grant before disconnect_export, but it
2478 * causes the following problem if setup (connect) and cleanup
2479 * (disconnect) are tangled together.
2480 * connect p1 disconnect p2
2481 * ptlrpc_connect_import
2482 * ............... class_manual_cleanup
2485 * ptlrpc_connect_interrupt
2487 * add this client to shrink list
2489 * Bang! pinger trigger the shrink.
2490 * So the osc should be disconnected from the shrink list, after we
2491 * are sure the import has been destroyed. BUG18662
2493 if (obd->u.cli.cl_import == NULL)
2494 osc_del_shrink_grant(&obd->u.cli);
2498 static int osc_import_event(struct obd_device *obd,
2499 struct obd_import *imp,
2500 enum obd_import_event event)
2502 struct client_obd *cli;
2506 LASSERT(imp->imp_obd == obd);
2509 case IMP_EVENT_DISCON: {
2511 spin_lock(&cli->cl_loi_list_lock);
2512 cli->cl_avail_grant = 0;
2513 cli->cl_lost_grant = 0;
2514 spin_unlock(&cli->cl_loi_list_lock);
2517 case IMP_EVENT_INACTIVE: {
2518 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
2521 case IMP_EVENT_INVALIDATE: {
2522 struct ldlm_namespace *ns = obd->obd_namespace;
2526 env = cl_env_get(&refcheck);
2530 /* all pages go to failing rpcs due to the invalid
2532 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
2534 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2535 cl_env_put(env, &refcheck);
2540 case IMP_EVENT_ACTIVE: {
2541 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
2544 case IMP_EVENT_OCD: {
2545 struct obd_connect_data *ocd = &imp->imp_connect_data;
2547 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
2548 osc_init_grant(&obd->u.cli, ocd);
2551 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
2552 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
2554 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
2557 case IMP_EVENT_DEACTIVATE: {
2558 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
2561 case IMP_EVENT_ACTIVATE: {
2562 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
2566 CERROR("Unknown import event %d\n", event);
2573 * Determine whether the lock can be canceled before replaying the lock
2574 * during recovery, see bug16774 for detailed information.
2576 * \retval zero the lock can't be canceled
2577 * \retval other ok to cancel
2579 static int osc_cancel_weight(struct ldlm_lock *lock)
2582 * Cancel all unused and granted extent lock.
2584 if (lock->l_resource->lr_type == LDLM_EXTENT &&
2585 lock->l_granted_mode == lock->l_req_mode &&
2586 osc_ldlm_weigh_ast(lock) == 0)
2592 static int brw_queue_work(const struct lu_env *env, void *data)
2594 struct client_obd *cli = data;
2596 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2598 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2602 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2604 struct client_obd *cli = &obd->u.cli;
2605 struct obd_type *type;
2610 rc = ptlrpcd_addref();
2614 rc = client_obd_setup(obd, lcfg);
2616 GOTO(out_ptlrpcd, rc);
2618 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
2619 if (IS_ERR(handler))
2620 GOTO(out_client_setup, rc = PTR_ERR(handler));
2621 cli->cl_writeback_work = handler;
2623 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
2624 if (IS_ERR(handler))
2625 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
2626 cli->cl_lru_work = handler;
2628 rc = osc_quota_setup(obd);
2630 GOTO(out_ptlrpcd_work, rc);
2632 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
2634 #ifdef CONFIG_PROC_FS
2635 obd->obd_vars = lprocfs_osc_obd_vars;
2637 /* If this is true then both client (osc) and server (osp) are on the
2638 * same node. The osp layer if loaded first will register the osc proc
2639 * directory. In that case this obd_device will be attached its proc
2640 * tree to type->typ_procsym instead of obd->obd_type->typ_procroot. */
2641 type = class_search_type(LUSTRE_OSP_NAME);
2642 if (type && type->typ_procsym) {
2643 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
2645 obd->obd_vars, obd);
2646 if (IS_ERR(obd->obd_proc_entry)) {
2647 rc = PTR_ERR(obd->obd_proc_entry);
2648 CERROR("error %d setting up lprocfs for %s\n", rc,
2650 obd->obd_proc_entry = NULL;
2653 rc = lprocfs_obd_setup(obd);
2656 /* If the basic OSC proc tree construction succeeded then
2657 * lets do the rest. */
2659 lproc_osc_attach_seqstat(obd);
2660 sptlrpc_lprocfs_cliobd_attach(obd);
2661 ptlrpc_lprocfs_register_obd(obd);
2664 /* We need to allocate a few requests more, because
2665 * brw_interpret tries to create new requests before freeing
2666 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
2667 * reserved, but I'm afraid that might be too much wasted RAM
2668 * in fact, so 2 is just my guess and still should work. */
2669 cli->cl_import->imp_rq_pool =
2670 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
2672 ptlrpc_add_rqs_to_pool);
2674 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
2675 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
2679 if (cli->cl_writeback_work != NULL) {
2680 ptlrpcd_destroy_work(cli->cl_writeback_work);
2681 cli->cl_writeback_work = NULL;
2683 if (cli->cl_lru_work != NULL) {
2684 ptlrpcd_destroy_work(cli->cl_lru_work);
2685 cli->cl_lru_work = NULL;
2688 client_obd_cleanup(obd);
2694 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2700 case OBD_CLEANUP_EARLY: {
2701 struct obd_import *imp;
2702 imp = obd->u.cli.cl_import;
2703 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
2704 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
2705 ptlrpc_deactivate_import(imp);
2706 spin_lock(&imp->imp_lock);
2707 imp->imp_pingable = 0;
2708 spin_unlock(&imp->imp_lock);
2711 case OBD_CLEANUP_EXPORTS: {
2712 struct client_obd *cli = &obd->u.cli;
2714 * for echo client, export may be on zombie list, wait for
2715 * zombie thread to cull it, because cli.cl_import will be
2716 * cleared in client_disconnect_export():
2717 * class_export_destroy() -> obd_cleanup() ->
2718 * echo_device_free() -> echo_client_cleanup() ->
2719 * obd_disconnect() -> osc_disconnect() ->
2720 * client_disconnect_export()
2722 obd_zombie_barrier();
2723 if (cli->cl_writeback_work) {
2724 ptlrpcd_destroy_work(cli->cl_writeback_work);
2725 cli->cl_writeback_work = NULL;
2727 if (cli->cl_lru_work) {
2728 ptlrpcd_destroy_work(cli->cl_lru_work);
2729 cli->cl_lru_work = NULL;
2731 obd_cleanup_client_import(obd);
2732 ptlrpc_lprocfs_unregister_obd(obd);
2733 lprocfs_obd_cleanup(obd);
2740 int osc_cleanup(struct obd_device *obd)
2742 struct client_obd *cli = &obd->u.cli;
2748 if (cli->cl_cache != NULL) {
2749 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
2750 spin_lock(&cli->cl_cache->ccc_lru_lock);
2751 list_del_init(&cli->cl_lru_osc);
2752 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2753 cli->cl_lru_left = NULL;
2754 cl_cache_decref(cli->cl_cache);
2755 cli->cl_cache = NULL;
2758 /* free memory of osc quota cache */
2759 osc_quota_cleanup(obd);
2761 rc = client_obd_cleanup(obd);
2767 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
2769 int rc = class_process_proc_param(PARAM_OSC, obd->obd_vars, lcfg, obd);
2770 return rc > 0 ? 0: rc;
2773 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
2775 return osc_process_config_base(obd, buf);
2778 static struct obd_ops osc_obd_ops = {
2779 .o_owner = THIS_MODULE,
2780 .o_setup = osc_setup,
2781 .o_precleanup = osc_precleanup,
2782 .o_cleanup = osc_cleanup,
2783 .o_add_conn = client_import_add_conn,
2784 .o_del_conn = client_import_del_conn,
2785 .o_connect = client_connect_import,
2786 .o_reconnect = osc_reconnect,
2787 .o_disconnect = osc_disconnect,
2788 .o_statfs = osc_statfs,
2789 .o_statfs_async = osc_statfs_async,
2790 .o_create = osc_create,
2791 .o_destroy = osc_destroy,
2792 .o_getattr = osc_getattr,
2793 .o_setattr = osc_setattr,
2794 .o_iocontrol = osc_iocontrol,
2795 .o_set_info_async = osc_set_info_async,
2796 .o_import_event = osc_import_event,
2797 .o_process_config = osc_process_config,
2798 .o_quotactl = osc_quotactl,
2801 static int __init osc_init(void)
2803 bool enable_proc = true;
2804 struct obd_type *type;
2808 /* print an address of _any_ initialized kernel symbol from this
2809 * module, to allow debugging with gdb that doesn't support data
2810 * symbols from modules.*/
2811 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
2813 rc = lu_kmem_init(osc_caches);
2817 type = class_search_type(LUSTRE_OSP_NAME);
2818 if (type != NULL && type->typ_procsym != NULL)
2819 enable_proc = false;
2821 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
2822 LUSTRE_OSC_NAME, &osc_device_type);
2824 lu_kmem_fini(osc_caches);
2831 static void /*__exit*/ osc_exit(void)
2833 class_unregister_type(LUSTRE_OSC_NAME);
2834 lu_kmem_fini(osc_caches);
2837 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2838 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
2839 MODULE_VERSION(LUSTRE_VERSION_STRING);
2840 MODULE_LICENSE("GPL");
2842 module_init(osc_init);
2843 module_exit(osc_exit);