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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_OSC
35 #include <linux/workqueue.h>
36 #include <lprocfs_status.h>
37 #include <lustre_debug.h>
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <lustre_ha.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
42 #include <lustre_net.h>
43 #include <lustre_obdo.h>
45 #include <obd_cksum.h>
46 #include <obd_class.h>
47 #include <lustre_osc.h>
49 #include "osc_internal.h"
51 atomic_t osc_pool_req_count;
52 unsigned int osc_reqpool_maxreqcount;
53 struct ptlrpc_request_pool *osc_rq_pool;
55 /* max memory used for request pool, unit is MB */
56 static unsigned int osc_reqpool_mem_max = 5;
57 module_param(osc_reqpool_mem_max, uint, 0444);
59 static int osc_idle_timeout = 20;
60 module_param(osc_idle_timeout, uint, 0644);
62 #define osc_grant_args osc_brw_async_args
64 struct osc_setattr_args {
66 obd_enqueue_update_f sa_upcall;
70 struct osc_fsync_args {
71 struct osc_object *fa_obj;
73 obd_enqueue_update_f fa_upcall;
77 struct osc_ladvise_args {
79 obd_enqueue_update_f la_upcall;
83 static void osc_release_ppga(struct brw_page **ppga, size_t count);
84 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
87 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
89 struct ost_body *body;
91 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
94 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
97 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
100 struct ptlrpc_request *req;
101 struct ost_body *body;
105 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
109 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
111 ptlrpc_request_free(req);
115 osc_pack_req_body(req, oa);
117 ptlrpc_request_set_replen(req);
119 rc = ptlrpc_queue_wait(req);
123 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
125 GOTO(out, rc = -EPROTO);
127 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
128 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
130 oa->o_blksize = cli_brw_size(exp->exp_obd);
131 oa->o_valid |= OBD_MD_FLBLKSZ;
135 ptlrpc_req_finished(req);
140 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
143 struct ptlrpc_request *req;
144 struct ost_body *body;
148 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
150 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
154 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
156 ptlrpc_request_free(req);
160 osc_pack_req_body(req, oa);
162 ptlrpc_request_set_replen(req);
164 rc = ptlrpc_queue_wait(req);
168 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
170 GOTO(out, rc = -EPROTO);
172 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
176 ptlrpc_req_finished(req);
181 static int osc_setattr_interpret(const struct lu_env *env,
182 struct ptlrpc_request *req, void *args, int rc)
184 struct osc_setattr_args *sa = args;
185 struct ost_body *body;
192 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
194 GOTO(out, rc = -EPROTO);
196 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
199 rc = sa->sa_upcall(sa->sa_cookie, rc);
203 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
204 obd_enqueue_update_f upcall, void *cookie,
205 struct ptlrpc_request_set *rqset)
207 struct ptlrpc_request *req;
208 struct osc_setattr_args *sa;
213 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
217 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
219 ptlrpc_request_free(req);
223 osc_pack_req_body(req, oa);
225 ptlrpc_request_set_replen(req);
227 /* do mds to ost setattr asynchronously */
229 /* Do not wait for response. */
230 ptlrpcd_add_req(req);
232 req->rq_interpret_reply = osc_setattr_interpret;
234 sa = ptlrpc_req_async_args(sa, req);
236 sa->sa_upcall = upcall;
237 sa->sa_cookie = cookie;
239 if (rqset == PTLRPCD_SET)
240 ptlrpcd_add_req(req);
242 ptlrpc_set_add_req(rqset, req);
248 static int osc_ladvise_interpret(const struct lu_env *env,
249 struct ptlrpc_request *req,
252 struct osc_ladvise_args *la = arg;
253 struct ost_body *body;
259 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
261 GOTO(out, rc = -EPROTO);
263 *la->la_oa = body->oa;
265 rc = la->la_upcall(la->la_cookie, rc);
270 * If rqset is NULL, do not wait for response. Upcall and cookie could also
271 * be NULL in this case
273 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
274 struct ladvise_hdr *ladvise_hdr,
275 obd_enqueue_update_f upcall, void *cookie,
276 struct ptlrpc_request_set *rqset)
278 struct ptlrpc_request *req;
279 struct ost_body *body;
280 struct osc_ladvise_args *la;
282 struct lu_ladvise *req_ladvise;
283 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
284 int num_advise = ladvise_hdr->lah_count;
285 struct ladvise_hdr *req_ladvise_hdr;
288 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
292 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
293 num_advise * sizeof(*ladvise));
294 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
296 ptlrpc_request_free(req);
299 req->rq_request_portal = OST_IO_PORTAL;
300 ptlrpc_at_set_req_timeout(req);
302 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
304 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
307 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
308 &RMF_OST_LADVISE_HDR);
309 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
311 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
312 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
313 ptlrpc_request_set_replen(req);
316 /* Do not wait for response. */
317 ptlrpcd_add_req(req);
321 req->rq_interpret_reply = osc_ladvise_interpret;
322 la = ptlrpc_req_async_args(la, req);
324 la->la_upcall = upcall;
325 la->la_cookie = cookie;
327 if (rqset == PTLRPCD_SET)
328 ptlrpcd_add_req(req);
330 ptlrpc_set_add_req(rqset, req);
335 static int osc_create(const struct lu_env *env, struct obd_export *exp,
338 struct ptlrpc_request *req;
339 struct ost_body *body;
344 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
345 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
347 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
349 GOTO(out, rc = -ENOMEM);
351 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
353 ptlrpc_request_free(req);
357 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
360 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
362 ptlrpc_request_set_replen(req);
364 rc = ptlrpc_queue_wait(req);
368 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
370 GOTO(out_req, rc = -EPROTO);
372 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
373 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
375 oa->o_blksize = cli_brw_size(exp->exp_obd);
376 oa->o_valid |= OBD_MD_FLBLKSZ;
378 CDEBUG(D_HA, "transno: %lld\n",
379 lustre_msg_get_transno(req->rq_repmsg));
381 ptlrpc_req_finished(req);
386 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
387 obd_enqueue_update_f upcall, void *cookie)
389 struct ptlrpc_request *req;
390 struct osc_setattr_args *sa;
391 struct obd_import *imp = class_exp2cliimp(exp);
392 struct ost_body *body;
397 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
401 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
403 ptlrpc_request_free(req);
407 osc_set_io_portal(req);
409 ptlrpc_at_set_req_timeout(req);
411 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
413 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
415 ptlrpc_request_set_replen(req);
417 req->rq_interpret_reply = osc_setattr_interpret;
418 sa = ptlrpc_req_async_args(sa, req);
420 sa->sa_upcall = upcall;
421 sa->sa_cookie = cookie;
423 ptlrpcd_add_req(req);
427 EXPORT_SYMBOL(osc_punch_send);
429 static int osc_sync_interpret(const struct lu_env *env,
430 struct ptlrpc_request *req, void *args, int rc)
432 struct osc_fsync_args *fa = args;
433 struct ost_body *body;
434 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
435 unsigned long valid = 0;
436 struct cl_object *obj;
442 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
444 CERROR("can't unpack ost_body\n");
445 GOTO(out, rc = -EPROTO);
448 *fa->fa_oa = body->oa;
449 obj = osc2cl(fa->fa_obj);
451 /* Update osc object's blocks attribute */
452 cl_object_attr_lock(obj);
453 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
454 attr->cat_blocks = body->oa.o_blocks;
459 cl_object_attr_update(env, obj, attr, valid);
460 cl_object_attr_unlock(obj);
463 rc = fa->fa_upcall(fa->fa_cookie, rc);
467 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
468 obd_enqueue_update_f upcall, void *cookie,
469 struct ptlrpc_request_set *rqset)
471 struct obd_export *exp = osc_export(obj);
472 struct ptlrpc_request *req;
473 struct ost_body *body;
474 struct osc_fsync_args *fa;
478 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
482 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
484 ptlrpc_request_free(req);
488 /* overload the size and blocks fields in the oa with start/end */
489 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
491 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
493 ptlrpc_request_set_replen(req);
494 req->rq_interpret_reply = osc_sync_interpret;
496 fa = ptlrpc_req_async_args(fa, req);
499 fa->fa_upcall = upcall;
500 fa->fa_cookie = cookie;
502 if (rqset == PTLRPCD_SET)
503 ptlrpcd_add_req(req);
505 ptlrpc_set_add_req(rqset, req);
510 /* Find and cancel locally locks matched by @mode in the resource found by
511 * @objid. Found locks are added into @cancel list. Returns the amount of
512 * locks added to @cancels list. */
513 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
514 struct list_head *cancels,
515 enum ldlm_mode mode, __u64 lock_flags)
517 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
518 struct ldlm_res_id res_id;
519 struct ldlm_resource *res;
523 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
524 * export) but disabled through procfs (flag in NS).
526 * This distinguishes from a case when ELC is not supported originally,
527 * when we still want to cancel locks in advance and just cancel them
528 * locally, without sending any RPC. */
529 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
532 ostid_build_res_name(&oa->o_oi, &res_id);
533 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
537 LDLM_RESOURCE_ADDREF(res);
538 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
539 lock_flags, 0, NULL);
540 LDLM_RESOURCE_DELREF(res);
541 ldlm_resource_putref(res);
545 static int osc_destroy_interpret(const struct lu_env *env,
546 struct ptlrpc_request *req, void *args, int rc)
548 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
550 atomic_dec(&cli->cl_destroy_in_flight);
551 wake_up(&cli->cl_destroy_waitq);
556 static int osc_can_send_destroy(struct client_obd *cli)
558 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
559 cli->cl_max_rpcs_in_flight) {
560 /* The destroy request can be sent */
563 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
564 cli->cl_max_rpcs_in_flight) {
566 * The counter has been modified between the two atomic
569 wake_up(&cli->cl_destroy_waitq);
574 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
577 struct client_obd *cli = &exp->exp_obd->u.cli;
578 struct ptlrpc_request *req;
579 struct ost_body *body;
580 struct list_head cancels = LIST_HEAD_INIT(cancels);
585 CDEBUG(D_INFO, "oa NULL\n");
589 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
590 LDLM_FL_DISCARD_DATA);
592 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
594 ldlm_lock_list_put(&cancels, l_bl_ast, count);
598 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
601 ptlrpc_request_free(req);
605 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
606 ptlrpc_at_set_req_timeout(req);
608 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
610 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
612 ptlrpc_request_set_replen(req);
614 req->rq_interpret_reply = osc_destroy_interpret;
615 if (!osc_can_send_destroy(cli)) {
616 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
619 * Wait until the number of on-going destroy RPCs drops
620 * under max_rpc_in_flight
622 rc = l_wait_event_exclusive(cli->cl_destroy_waitq,
623 osc_can_send_destroy(cli), &lwi);
625 ptlrpc_req_finished(req);
630 /* Do not wait for response */
631 ptlrpcd_add_req(req);
635 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
638 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
640 LASSERT(!(oa->o_valid & bits));
643 spin_lock(&cli->cl_loi_list_lock);
644 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
645 oa->o_dirty = cli->cl_dirty_grant;
647 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
648 if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
649 cli->cl_dirty_max_pages)) {
650 CERROR("dirty %lu - %lu > dirty_max %lu\n",
651 cli->cl_dirty_pages, cli->cl_dirty_transit,
652 cli->cl_dirty_max_pages);
654 } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
655 atomic_long_read(&obd_dirty_transit_pages) >
656 (long)(obd_max_dirty_pages + 1))) {
657 /* The atomic_read() allowing the atomic_inc() are
658 * not covered by a lock thus they may safely race and trip
659 * this CERROR() unless we add in a small fudge factor (+1). */
660 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
661 cli_name(cli), atomic_long_read(&obd_dirty_pages),
662 atomic_long_read(&obd_dirty_transit_pages),
663 obd_max_dirty_pages);
665 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
667 CERROR("dirty %lu - dirty_max %lu too big???\n",
668 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
671 unsigned long nrpages;
672 unsigned long undirty;
674 nrpages = cli->cl_max_pages_per_rpc;
675 nrpages *= cli->cl_max_rpcs_in_flight + 1;
676 nrpages = max(nrpages, cli->cl_dirty_max_pages);
677 undirty = nrpages << PAGE_SHIFT;
678 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
682 /* take extent tax into account when asking for more
684 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
685 cli->cl_max_extent_pages;
686 undirty += nrextents * cli->cl_grant_extent_tax;
688 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
689 * to add extent tax, etc.
691 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
692 ~(PTLRPC_MAX_BRW_SIZE * 4UL));
694 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
695 oa->o_dropped = cli->cl_lost_grant;
696 cli->cl_lost_grant = 0;
697 spin_unlock(&cli->cl_loi_list_lock);
698 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
699 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
702 void osc_update_next_shrink(struct client_obd *cli)
704 cli->cl_next_shrink_grant = ktime_get_seconds() +
705 cli->cl_grant_shrink_interval;
707 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
708 cli->cl_next_shrink_grant);
711 static void __osc_update_grant(struct client_obd *cli, u64 grant)
713 spin_lock(&cli->cl_loi_list_lock);
714 cli->cl_avail_grant += grant;
715 spin_unlock(&cli->cl_loi_list_lock);
718 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
720 if (body->oa.o_valid & OBD_MD_FLGRANT) {
721 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
722 __osc_update_grant(cli, body->oa.o_grant);
727 * grant thread data for shrinking space.
729 struct grant_thread_data {
730 struct list_head gtd_clients;
731 struct mutex gtd_mutex;
732 unsigned long gtd_stopped:1;
734 static struct grant_thread_data client_gtd;
736 static int osc_shrink_grant_interpret(const struct lu_env *env,
737 struct ptlrpc_request *req,
740 struct osc_grant_args *aa = args;
741 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
742 struct ost_body *body;
745 __osc_update_grant(cli, aa->aa_oa->o_grant);
749 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
751 osc_update_grant(cli, body);
753 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
759 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
761 spin_lock(&cli->cl_loi_list_lock);
762 oa->o_grant = cli->cl_avail_grant / 4;
763 cli->cl_avail_grant -= oa->o_grant;
764 spin_unlock(&cli->cl_loi_list_lock);
765 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
766 oa->o_valid |= OBD_MD_FLFLAGS;
769 oa->o_flags |= OBD_FL_SHRINK_GRANT;
770 osc_update_next_shrink(cli);
773 /* Shrink the current grant, either from some large amount to enough for a
774 * full set of in-flight RPCs, or if we have already shrunk to that limit
775 * then to enough for a single RPC. This avoids keeping more grant than
776 * needed, and avoids shrinking the grant piecemeal. */
777 static int osc_shrink_grant(struct client_obd *cli)
779 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
780 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
782 spin_lock(&cli->cl_loi_list_lock);
783 if (cli->cl_avail_grant <= target_bytes)
784 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
785 spin_unlock(&cli->cl_loi_list_lock);
787 return osc_shrink_grant_to_target(cli, target_bytes);
790 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
793 struct ost_body *body;
796 spin_lock(&cli->cl_loi_list_lock);
797 /* Don't shrink if we are already above or below the desired limit
798 * We don't want to shrink below a single RPC, as that will negatively
799 * impact block allocation and long-term performance. */
800 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
801 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
803 if (target_bytes >= cli->cl_avail_grant) {
804 spin_unlock(&cli->cl_loi_list_lock);
807 spin_unlock(&cli->cl_loi_list_lock);
813 osc_announce_cached(cli, &body->oa, 0);
815 spin_lock(&cli->cl_loi_list_lock);
816 if (target_bytes >= cli->cl_avail_grant) {
817 /* available grant has changed since target calculation */
818 spin_unlock(&cli->cl_loi_list_lock);
819 GOTO(out_free, rc = 0);
821 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
822 cli->cl_avail_grant = target_bytes;
823 spin_unlock(&cli->cl_loi_list_lock);
824 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
825 body->oa.o_valid |= OBD_MD_FLFLAGS;
826 body->oa.o_flags = 0;
828 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
829 osc_update_next_shrink(cli);
831 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
832 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
833 sizeof(*body), body, NULL);
835 __osc_update_grant(cli, body->oa.o_grant);
841 static int osc_should_shrink_grant(struct client_obd *client)
843 time64_t next_shrink = client->cl_next_shrink_grant;
845 if (client->cl_import == NULL)
848 if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
849 client->cl_import->imp_grant_shrink_disabled)
852 if (ktime_get_seconds() >= next_shrink - 5) {
853 /* Get the current RPC size directly, instead of going via:
854 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
855 * Keep comment here so that it can be found by searching. */
856 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
858 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
859 client->cl_avail_grant > brw_size)
862 osc_update_next_shrink(client);
867 #define GRANT_SHRINK_RPC_BATCH 100
869 static struct delayed_work work;
871 static void osc_grant_work_handler(struct work_struct *data)
873 struct client_obd *cli;
875 bool init_next_shrink = true;
876 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
879 mutex_lock(&client_gtd.gtd_mutex);
880 list_for_each_entry(cli, &client_gtd.gtd_clients,
882 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
883 osc_should_shrink_grant(cli)) {
884 osc_shrink_grant(cli);
888 if (!init_next_shrink) {
889 if (cli->cl_next_shrink_grant < next_shrink &&
890 cli->cl_next_shrink_grant > ktime_get_seconds())
891 next_shrink = cli->cl_next_shrink_grant;
893 init_next_shrink = false;
894 next_shrink = cli->cl_next_shrink_grant;
897 mutex_unlock(&client_gtd.gtd_mutex);
899 if (client_gtd.gtd_stopped == 1)
902 if (next_shrink > ktime_get_seconds()) {
903 time64_t delay = next_shrink - ktime_get_seconds();
905 schedule_delayed_work(&work, cfs_time_seconds(delay));
907 schedule_work(&work.work);
911 void osc_schedule_grant_work(void)
913 cancel_delayed_work_sync(&work);
914 schedule_work(&work.work);
918 * Start grant thread for returing grant to server for idle clients.
920 static int osc_start_grant_work(void)
922 client_gtd.gtd_stopped = 0;
923 mutex_init(&client_gtd.gtd_mutex);
924 INIT_LIST_HEAD(&client_gtd.gtd_clients);
926 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
927 schedule_work(&work.work);
932 static void osc_stop_grant_work(void)
934 client_gtd.gtd_stopped = 1;
935 cancel_delayed_work_sync(&work);
938 static void osc_add_grant_list(struct client_obd *client)
940 mutex_lock(&client_gtd.gtd_mutex);
941 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
942 mutex_unlock(&client_gtd.gtd_mutex);
945 static void osc_del_grant_list(struct client_obd *client)
947 if (list_empty(&client->cl_grant_chain))
950 mutex_lock(&client_gtd.gtd_mutex);
951 list_del_init(&client->cl_grant_chain);
952 mutex_unlock(&client_gtd.gtd_mutex);
955 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
958 * ocd_grant is the total grant amount we're expect to hold: if we've
959 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
960 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
963 * race is tolerable here: if we're evicted, but imp_state already
964 * left EVICTED state, then cl_dirty_pages must be 0 already.
966 spin_lock(&cli->cl_loi_list_lock);
967 cli->cl_avail_grant = ocd->ocd_grant;
968 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
969 cli->cl_avail_grant -= cli->cl_reserved_grant;
970 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
971 cli->cl_avail_grant -= cli->cl_dirty_grant;
973 cli->cl_avail_grant -=
974 cli->cl_dirty_pages << PAGE_SHIFT;
977 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
981 /* overhead for each extent insertion */
982 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
983 /* determine the appropriate chunk size used by osc_extent. */
984 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
985 ocd->ocd_grant_blkbits);
986 /* max_pages_per_rpc must be chunk aligned */
987 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
988 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
989 ~chunk_mask) & chunk_mask;
990 /* determine maximum extent size, in #pages */
991 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
992 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
993 if (cli->cl_max_extent_pages == 0)
994 cli->cl_max_extent_pages = 1;
996 cli->cl_grant_extent_tax = 0;
997 cli->cl_chunkbits = PAGE_SHIFT;
998 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1000 spin_unlock(&cli->cl_loi_list_lock);
1002 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1003 "chunk bits: %d cl_max_extent_pages: %d\n",
1005 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1006 cli->cl_max_extent_pages);
1008 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1009 osc_add_grant_list(cli);
1011 EXPORT_SYMBOL(osc_init_grant);
1013 /* We assume that the reason this OSC got a short read is because it read
1014 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1015 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1016 * this stripe never got written at or beyond this stripe offset yet. */
1017 static void handle_short_read(int nob_read, size_t page_count,
1018 struct brw_page **pga)
1023 /* skip bytes read OK */
1024 while (nob_read > 0) {
1025 LASSERT (page_count > 0);
1027 if (pga[i]->count > nob_read) {
1028 /* EOF inside this page */
1029 ptr = kmap(pga[i]->pg) +
1030 (pga[i]->off & ~PAGE_MASK);
1031 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1038 nob_read -= pga[i]->count;
1043 /* zero remaining pages */
1044 while (page_count-- > 0) {
1045 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1046 memset(ptr, 0, pga[i]->count);
1052 static int check_write_rcs(struct ptlrpc_request *req,
1053 int requested_nob, int niocount,
1054 size_t page_count, struct brw_page **pga)
1059 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1060 sizeof(*remote_rcs) *
1062 if (remote_rcs == NULL) {
1063 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1067 /* return error if any niobuf was in error */
1068 for (i = 0; i < niocount; i++) {
1069 if ((int)remote_rcs[i] < 0) {
1070 CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1071 i, remote_rcs[i], req);
1072 return remote_rcs[i];
1075 if (remote_rcs[i] != 0) {
1076 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1077 i, remote_rcs[i], req);
1081 if (req->rq_bulk != NULL &&
1082 req->rq_bulk->bd_nob_transferred != requested_nob) {
1083 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1084 req->rq_bulk->bd_nob_transferred, requested_nob);
1091 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1093 if (p1->flag != p2->flag) {
1094 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1095 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1096 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1098 /* warn if we try to combine flags that we don't know to be
1099 * safe to combine */
1100 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1101 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1102 "report this at https://jira.whamcloud.com/\n",
1103 p1->flag, p2->flag);
1108 return (p1->off + p1->count == p2->off);
1111 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1112 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1113 size_t pg_count, struct brw_page **pga,
1114 int opc, obd_dif_csum_fn *fn,
1118 struct ahash_request *req;
1119 /* Used Adler as the default checksum type on top of DIF tags */
1120 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1121 struct page *__page;
1122 unsigned char *buffer;
1124 unsigned int bufsize;
1126 int used_number = 0;
1132 LASSERT(pg_count > 0);
1134 __page = alloc_page(GFP_KERNEL);
1138 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1141 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1142 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1146 buffer = kmap(__page);
1147 guard_start = (__u16 *)buffer;
1148 guard_number = PAGE_SIZE / sizeof(*guard_start);
1149 while (nob > 0 && pg_count > 0) {
1150 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1152 /* corrupt the data before we compute the checksum, to
1153 * simulate an OST->client data error */
1154 if (unlikely(i == 0 && opc == OST_READ &&
1155 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1156 unsigned char *ptr = kmap(pga[i]->pg);
1157 int off = pga[i]->off & ~PAGE_MASK;
1159 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1164 * The left guard number should be able to hold checksums of a
1167 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1168 pga[i]->off & ~PAGE_MASK,
1170 guard_start + used_number,
1171 guard_number - used_number,
1177 used_number += used;
1178 if (used_number == guard_number) {
1179 cfs_crypto_hash_update_page(req, __page, 0,
1180 used_number * sizeof(*guard_start));
1184 nob -= pga[i]->count;
1192 if (used_number != 0)
1193 cfs_crypto_hash_update_page(req, __page, 0,
1194 used_number * sizeof(*guard_start));
1196 bufsize = sizeof(cksum);
1197 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1199 /* For sending we only compute the wrong checksum instead
1200 * of corrupting the data so it is still correct on a redo */
1201 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1206 __free_page(__page);
1209 #else /* !CONFIG_CRC_T10DIF */
1210 #define obd_dif_ip_fn NULL
1211 #define obd_dif_crc_fn NULL
1212 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum) \
1214 #endif /* CONFIG_CRC_T10DIF */
1216 static int osc_checksum_bulk(int nob, size_t pg_count,
1217 struct brw_page **pga, int opc,
1218 enum cksum_types cksum_type,
1222 struct ahash_request *req;
1223 unsigned int bufsize;
1224 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1226 LASSERT(pg_count > 0);
1228 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1230 CERROR("Unable to initialize checksum hash %s\n",
1231 cfs_crypto_hash_name(cfs_alg));
1232 return PTR_ERR(req);
1235 while (nob > 0 && pg_count > 0) {
1236 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1238 /* corrupt the data before we compute the checksum, to
1239 * simulate an OST->client data error */
1240 if (i == 0 && opc == OST_READ &&
1241 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1242 unsigned char *ptr = kmap(pga[i]->pg);
1243 int off = pga[i]->off & ~PAGE_MASK;
1245 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1248 cfs_crypto_hash_update_page(req, pga[i]->pg,
1249 pga[i]->off & ~PAGE_MASK,
1251 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1252 (int)(pga[i]->off & ~PAGE_MASK));
1254 nob -= pga[i]->count;
1259 bufsize = sizeof(*cksum);
1260 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1262 /* For sending we only compute the wrong checksum instead
1263 * of corrupting the data so it is still correct on a redo */
1264 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1270 static int osc_checksum_bulk_rw(const char *obd_name,
1271 enum cksum_types cksum_type,
1272 int nob, size_t pg_count,
1273 struct brw_page **pga, int opc,
1276 obd_dif_csum_fn *fn = NULL;
1277 int sector_size = 0;
1281 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1284 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1285 opc, fn, sector_size, check_sum);
1287 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1294 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1295 u32 page_count, struct brw_page **pga,
1296 struct ptlrpc_request **reqp, int resend)
1298 struct ptlrpc_request *req;
1299 struct ptlrpc_bulk_desc *desc;
1300 struct ost_body *body;
1301 struct obd_ioobj *ioobj;
1302 struct niobuf_remote *niobuf;
1303 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1304 struct osc_brw_async_args *aa;
1305 struct req_capsule *pill;
1306 struct brw_page *pg_prev;
1308 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1311 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1312 RETURN(-ENOMEM); /* Recoverable */
1313 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1314 RETURN(-EINVAL); /* Fatal */
1316 if ((cmd & OBD_BRW_WRITE) != 0) {
1318 req = ptlrpc_request_alloc_pool(cli->cl_import,
1320 &RQF_OST_BRW_WRITE);
1323 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1328 for (niocount = i = 1; i < page_count; i++) {
1329 if (!can_merge_pages(pga[i - 1], pga[i]))
1333 pill = &req->rq_pill;
1334 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1336 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1337 niocount * sizeof(*niobuf));
1339 for (i = 0; i < page_count; i++)
1340 short_io_size += pga[i]->count;
1342 /* Check if read/write is small enough to be a short io. */
1343 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1344 !imp_connect_shortio(cli->cl_import))
1347 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1348 opc == OST_READ ? 0 : short_io_size);
1349 if (opc == OST_READ)
1350 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1353 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1355 ptlrpc_request_free(req);
1358 osc_set_io_portal(req);
1360 ptlrpc_at_set_req_timeout(req);
1361 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1363 req->rq_no_retry_einprogress = 1;
1365 if (short_io_size != 0) {
1367 short_io_buf = NULL;
1371 desc = ptlrpc_prep_bulk_imp(req, page_count,
1372 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1373 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1374 PTLRPC_BULK_PUT_SINK) |
1375 PTLRPC_BULK_BUF_KIOV,
1377 &ptlrpc_bulk_kiov_pin_ops);
1380 GOTO(out, rc = -ENOMEM);
1381 /* NB request now owns desc and will free it when it gets freed */
1383 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1384 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1385 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1386 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1388 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1390 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1391 * and from_kgid(), because they are asynchronous. Fortunately, variable
1392 * oa contains valid o_uid and o_gid in these two operations.
1393 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1394 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1395 * other process logic */
1396 body->oa.o_uid = oa->o_uid;
1397 body->oa.o_gid = oa->o_gid;
1399 obdo_to_ioobj(oa, ioobj);
1400 ioobj->ioo_bufcnt = niocount;
1401 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1402 * that might be send for this request. The actual number is decided
1403 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1404 * "max - 1" for old client compatibility sending "0", and also so the
1405 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1407 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1409 ioobj_max_brw_set(ioobj, 0);
1411 if (short_io_size != 0) {
1412 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1413 body->oa.o_valid |= OBD_MD_FLFLAGS;
1414 body->oa.o_flags = 0;
1416 body->oa.o_flags |= OBD_FL_SHORT_IO;
1417 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1419 if (opc == OST_WRITE) {
1420 short_io_buf = req_capsule_client_get(pill,
1422 LASSERT(short_io_buf != NULL);
1426 LASSERT(page_count > 0);
1428 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1429 struct brw_page *pg = pga[i];
1430 int poff = pg->off & ~PAGE_MASK;
1432 LASSERT(pg->count > 0);
1433 /* make sure there is no gap in the middle of page array */
1434 LASSERTF(page_count == 1 ||
1435 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1436 ergo(i > 0 && i < page_count - 1,
1437 poff == 0 && pg->count == PAGE_SIZE) &&
1438 ergo(i == page_count - 1, poff == 0)),
1439 "i: %d/%d pg: %p off: %llu, count: %u\n",
1440 i, page_count, pg, pg->off, pg->count);
1441 LASSERTF(i == 0 || pg->off > pg_prev->off,
1442 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1443 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1445 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1446 pg_prev->pg, page_private(pg_prev->pg),
1447 pg_prev->pg->index, pg_prev->off);
1448 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1449 (pg->flag & OBD_BRW_SRVLOCK));
1450 if (short_io_size != 0 && opc == OST_WRITE) {
1451 unsigned char *ptr = ll_kmap_atomic(pg->pg, KM_USER0);
1453 LASSERT(short_io_size >= requested_nob + pg->count);
1454 memcpy(short_io_buf + requested_nob,
1457 ll_kunmap_atomic(ptr, KM_USER0);
1458 } else if (short_io_size == 0) {
1459 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1462 requested_nob += pg->count;
1464 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1466 niobuf->rnb_len += pg->count;
1468 niobuf->rnb_offset = pg->off;
1469 niobuf->rnb_len = pg->count;
1470 niobuf->rnb_flags = pg->flag;
1475 LASSERTF((void *)(niobuf - niocount) ==
1476 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1477 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1478 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1480 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1482 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1483 body->oa.o_valid |= OBD_MD_FLFLAGS;
1484 body->oa.o_flags = 0;
1486 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1489 if (osc_should_shrink_grant(cli))
1490 osc_shrink_grant_local(cli, &body->oa);
1492 /* size[REQ_REC_OFF] still sizeof (*body) */
1493 if (opc == OST_WRITE) {
1494 if (cli->cl_checksum &&
1495 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1496 /* store cl_cksum_type in a local variable since
1497 * it can be changed via lprocfs */
1498 enum cksum_types cksum_type = cli->cl_cksum_type;
1500 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1501 body->oa.o_flags = 0;
1503 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1505 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1507 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1508 requested_nob, page_count,
1512 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1516 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1519 /* save this in 'oa', too, for later checking */
1520 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1521 oa->o_flags |= obd_cksum_type_pack(obd_name,
1524 /* clear out the checksum flag, in case this is a
1525 * resend but cl_checksum is no longer set. b=11238 */
1526 oa->o_valid &= ~OBD_MD_FLCKSUM;
1528 oa->o_cksum = body->oa.o_cksum;
1529 /* 1 RC per niobuf */
1530 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1531 sizeof(__u32) * niocount);
1533 if (cli->cl_checksum &&
1534 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1535 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1536 body->oa.o_flags = 0;
1537 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1538 cli->cl_cksum_type);
1539 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1542 /* Client cksum has been already copied to wire obdo in previous
1543 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1544 * resent due to cksum error, this will allow Server to
1545 * check+dump pages on its side */
1547 ptlrpc_request_set_replen(req);
1549 aa = ptlrpc_req_async_args(aa, req);
1551 aa->aa_requested_nob = requested_nob;
1552 aa->aa_nio_count = niocount;
1553 aa->aa_page_count = page_count;
1557 INIT_LIST_HEAD(&aa->aa_oaps);
1560 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1561 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1562 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1563 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1567 ptlrpc_req_finished(req);
1571 char dbgcksum_file_name[PATH_MAX];
1573 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1574 struct brw_page **pga, __u32 server_cksum,
1582 /* will only keep dump of pages on first error for the same range in
1583 * file/fid, not during the resends/retries. */
1584 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1585 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1586 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1587 libcfs_debug_file_path_arr :
1588 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1589 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1590 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1591 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1593 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1594 client_cksum, server_cksum);
1595 filp = filp_open(dbgcksum_file_name,
1596 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1600 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1601 "checksum error: rc = %d\n", dbgcksum_file_name,
1604 CERROR("%s: can't open to dump pages with checksum "
1605 "error: rc = %d\n", dbgcksum_file_name, rc);
1609 for (i = 0; i < page_count; i++) {
1610 len = pga[i]->count;
1611 buf = kmap(pga[i]->pg);
1613 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1615 CERROR("%s: wanted to write %u but got %d "
1616 "error\n", dbgcksum_file_name, len, rc);
1621 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1622 dbgcksum_file_name, rc);
1627 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1629 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1630 filp_close(filp, NULL);
1634 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1635 __u32 client_cksum, __u32 server_cksum,
1636 struct osc_brw_async_args *aa)
1638 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1639 enum cksum_types cksum_type;
1640 obd_dif_csum_fn *fn = NULL;
1641 int sector_size = 0;
1646 if (server_cksum == client_cksum) {
1647 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1651 if (aa->aa_cli->cl_checksum_dump)
1652 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1653 server_cksum, client_cksum);
1655 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1658 switch (cksum_type) {
1659 case OBD_CKSUM_T10IP512:
1663 case OBD_CKSUM_T10IP4K:
1667 case OBD_CKSUM_T10CRC512:
1668 fn = obd_dif_crc_fn;
1671 case OBD_CKSUM_T10CRC4K:
1672 fn = obd_dif_crc_fn;
1680 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1681 aa->aa_page_count, aa->aa_ppga,
1682 OST_WRITE, fn, sector_size,
1685 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1686 aa->aa_ppga, OST_WRITE, cksum_type,
1690 msg = "failed to calculate the client write checksum";
1691 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1692 msg = "the server did not use the checksum type specified in "
1693 "the original request - likely a protocol problem";
1694 else if (new_cksum == server_cksum)
1695 msg = "changed on the client after we checksummed it - "
1696 "likely false positive due to mmap IO (bug 11742)";
1697 else if (new_cksum == client_cksum)
1698 msg = "changed in transit before arrival at OST";
1700 msg = "changed in transit AND doesn't match the original - "
1701 "likely false positive due to mmap IO (bug 11742)";
1703 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1704 DFID " object "DOSTID" extent [%llu-%llu], original "
1705 "client csum %x (type %x), server csum %x (type %x),"
1706 " client csum now %x\n",
1707 obd_name, msg, libcfs_nid2str(peer->nid),
1708 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1709 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1710 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1711 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1712 aa->aa_ppga[aa->aa_page_count - 1]->off +
1713 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1715 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1716 server_cksum, cksum_type, new_cksum);
1720 /* Note rc enters this function as number of bytes transferred */
1721 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1723 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1724 struct client_obd *cli = aa->aa_cli;
1725 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1726 const struct lnet_process_id *peer =
1727 &req->rq_import->imp_connection->c_peer;
1728 struct ost_body *body;
1729 u32 client_cksum = 0;
1733 if (rc < 0 && rc != -EDQUOT) {
1734 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1738 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1739 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1741 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1745 /* set/clear over quota flag for a uid/gid/projid */
1746 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1747 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1748 unsigned qid[LL_MAXQUOTAS] = {
1749 body->oa.o_uid, body->oa.o_gid,
1750 body->oa.o_projid };
1752 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1753 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1754 body->oa.o_valid, body->oa.o_flags);
1755 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1759 osc_update_grant(cli, body);
1764 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1765 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1767 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1769 CERROR("%s: unexpected positive size %d\n",
1774 if (req->rq_bulk != NULL &&
1775 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1778 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1779 check_write_checksum(&body->oa, peer, client_cksum,
1780 body->oa.o_cksum, aa))
1783 rc = check_write_rcs(req, aa->aa_requested_nob,
1784 aa->aa_nio_count, aa->aa_page_count,
1789 /* The rest of this function executes only for OST_READs */
1791 if (req->rq_bulk == NULL) {
1792 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1794 LASSERT(rc == req->rq_status);
1796 /* if unwrap_bulk failed, return -EAGAIN to retry */
1797 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1800 GOTO(out, rc = -EAGAIN);
1802 if (rc > aa->aa_requested_nob) {
1803 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
1804 rc, aa->aa_requested_nob);
1808 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1809 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
1810 rc, req->rq_bulk->bd_nob_transferred);
1814 if (req->rq_bulk == NULL) {
1816 int nob, pg_count, i = 0;
1819 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1820 pg_count = aa->aa_page_count;
1821 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1824 while (nob > 0 && pg_count > 0) {
1826 int count = aa->aa_ppga[i]->count > nob ?
1827 nob : aa->aa_ppga[i]->count;
1829 CDEBUG(D_CACHE, "page %p count %d\n",
1830 aa->aa_ppga[i]->pg, count);
1831 ptr = ll_kmap_atomic(aa->aa_ppga[i]->pg, KM_USER0);
1832 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1834 ll_kunmap_atomic((void *) ptr, KM_USER0);
1843 if (rc < aa->aa_requested_nob)
1844 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1846 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1847 static int cksum_counter;
1848 u32 server_cksum = body->oa.o_cksum;
1851 enum cksum_types cksum_type;
1852 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1853 body->oa.o_flags : 0;
1855 cksum_type = obd_cksum_type_unpack(o_flags);
1856 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1857 aa->aa_page_count, aa->aa_ppga,
1858 OST_READ, &client_cksum);
1862 if (req->rq_bulk != NULL &&
1863 peer->nid != req->rq_bulk->bd_sender) {
1865 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1868 if (server_cksum != client_cksum) {
1869 struct ost_body *clbody;
1870 u32 page_count = aa->aa_page_count;
1872 clbody = req_capsule_client_get(&req->rq_pill,
1874 if (cli->cl_checksum_dump)
1875 dump_all_bulk_pages(&clbody->oa, page_count,
1876 aa->aa_ppga, server_cksum,
1879 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1880 "%s%s%s inode "DFID" object "DOSTID
1881 " extent [%llu-%llu], client %x, "
1882 "server %x, cksum_type %x\n",
1884 libcfs_nid2str(peer->nid),
1886 clbody->oa.o_valid & OBD_MD_FLFID ?
1887 clbody->oa.o_parent_seq : 0ULL,
1888 clbody->oa.o_valid & OBD_MD_FLFID ?
1889 clbody->oa.o_parent_oid : 0,
1890 clbody->oa.o_valid & OBD_MD_FLFID ?
1891 clbody->oa.o_parent_ver : 0,
1892 POSTID(&body->oa.o_oi),
1893 aa->aa_ppga[0]->off,
1894 aa->aa_ppga[page_count-1]->off +
1895 aa->aa_ppga[page_count-1]->count - 1,
1896 client_cksum, server_cksum,
1899 aa->aa_oa->o_cksum = client_cksum;
1903 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1906 } else if (unlikely(client_cksum)) {
1907 static int cksum_missed;
1910 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1911 CERROR("%s: checksum %u requested from %s but not sent\n",
1912 obd_name, cksum_missed,
1913 libcfs_nid2str(peer->nid));
1919 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1920 aa->aa_oa, &body->oa);
1925 static int osc_brw_redo_request(struct ptlrpc_request *request,
1926 struct osc_brw_async_args *aa, int rc)
1928 struct ptlrpc_request *new_req;
1929 struct osc_brw_async_args *new_aa;
1930 struct osc_async_page *oap;
1933 /* The below message is checked in replay-ost-single.sh test_8ae*/
1934 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1935 "redo for recoverable error %d", rc);
1937 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1938 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1939 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1940 aa->aa_ppga, &new_req, 1);
1944 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1945 if (oap->oap_request != NULL) {
1946 LASSERTF(request == oap->oap_request,
1947 "request %p != oap_request %p\n",
1948 request, oap->oap_request);
1949 if (oap->oap_interrupted) {
1950 ptlrpc_req_finished(new_req);
1956 * New request takes over pga and oaps from old request.
1957 * Note that copying a list_head doesn't work, need to move it...
1960 new_req->rq_interpret_reply = request->rq_interpret_reply;
1961 new_req->rq_async_args = request->rq_async_args;
1962 new_req->rq_commit_cb = request->rq_commit_cb;
1963 /* cap resend delay to the current request timeout, this is similar to
1964 * what ptlrpc does (see after_reply()) */
1965 if (aa->aa_resends > new_req->rq_timeout)
1966 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1968 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1969 new_req->rq_generation_set = 1;
1970 new_req->rq_import_generation = request->rq_import_generation;
1972 new_aa = ptlrpc_req_async_args(new_aa, new_req);
1974 INIT_LIST_HEAD(&new_aa->aa_oaps);
1975 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1976 INIT_LIST_HEAD(&new_aa->aa_exts);
1977 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1978 new_aa->aa_resends = aa->aa_resends;
1980 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1981 if (oap->oap_request) {
1982 ptlrpc_req_finished(oap->oap_request);
1983 oap->oap_request = ptlrpc_request_addref(new_req);
1987 /* XXX: This code will run into problem if we're going to support
1988 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1989 * and wait for all of them to be finished. We should inherit request
1990 * set from old request. */
1991 ptlrpcd_add_req(new_req);
1993 DEBUG_REQ(D_INFO, new_req, "new request");
1998 * ugh, we want disk allocation on the target to happen in offset order. we'll
1999 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2000 * fine for our small page arrays and doesn't require allocation. its an
2001 * insertion sort that swaps elements that are strides apart, shrinking the
2002 * stride down until its '1' and the array is sorted.
2004 static void sort_brw_pages(struct brw_page **array, int num)
2007 struct brw_page *tmp;
2011 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2016 for (i = stride ; i < num ; i++) {
2019 while (j >= stride && array[j - stride]->off > tmp->off) {
2020 array[j] = array[j - stride];
2025 } while (stride > 1);
2028 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2030 LASSERT(ppga != NULL);
2031 OBD_FREE(ppga, sizeof(*ppga) * count);
2034 static int brw_interpret(const struct lu_env *env,
2035 struct ptlrpc_request *req, void *args, int rc)
2037 struct osc_brw_async_args *aa = args;
2038 struct osc_extent *ext;
2039 struct osc_extent *tmp;
2040 struct client_obd *cli = aa->aa_cli;
2041 unsigned long transferred = 0;
2045 rc = osc_brw_fini_request(req, rc);
2046 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2048 * When server returns -EINPROGRESS, client should always retry
2049 * regardless of the number of times the bulk was resent already.
2051 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2052 if (req->rq_import_generation !=
2053 req->rq_import->imp_generation) {
2054 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2055 ""DOSTID", rc = %d.\n",
2056 req->rq_import->imp_obd->obd_name,
2057 POSTID(&aa->aa_oa->o_oi), rc);
2058 } else if (rc == -EINPROGRESS ||
2059 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2060 rc = osc_brw_redo_request(req, aa, rc);
2062 CERROR("%s: too many resent retries for object: "
2063 "%llu:%llu, rc = %d.\n",
2064 req->rq_import->imp_obd->obd_name,
2065 POSTID(&aa->aa_oa->o_oi), rc);
2070 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2075 struct obdo *oa = aa->aa_oa;
2076 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2077 unsigned long valid = 0;
2078 struct cl_object *obj;
2079 struct osc_async_page *last;
2081 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2082 obj = osc2cl(last->oap_obj);
2084 cl_object_attr_lock(obj);
2085 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2086 attr->cat_blocks = oa->o_blocks;
2087 valid |= CAT_BLOCKS;
2089 if (oa->o_valid & OBD_MD_FLMTIME) {
2090 attr->cat_mtime = oa->o_mtime;
2093 if (oa->o_valid & OBD_MD_FLATIME) {
2094 attr->cat_atime = oa->o_atime;
2097 if (oa->o_valid & OBD_MD_FLCTIME) {
2098 attr->cat_ctime = oa->o_ctime;
2102 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2103 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2104 loff_t last_off = last->oap_count + last->oap_obj_off +
2107 /* Change file size if this is an out of quota or
2108 * direct IO write and it extends the file size */
2109 if (loi->loi_lvb.lvb_size < last_off) {
2110 attr->cat_size = last_off;
2113 /* Extend KMS if it's not a lockless write */
2114 if (loi->loi_kms < last_off &&
2115 oap2osc_page(last)->ops_srvlock == 0) {
2116 attr->cat_kms = last_off;
2122 cl_object_attr_update(env, obj, attr, valid);
2123 cl_object_attr_unlock(obj);
2125 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2128 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2129 osc_inc_unstable_pages(req);
2131 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2132 list_del_init(&ext->oe_link);
2133 osc_extent_finish(env, ext, 1,
2134 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2136 LASSERT(list_empty(&aa->aa_exts));
2137 LASSERT(list_empty(&aa->aa_oaps));
2139 transferred = (req->rq_bulk == NULL ? /* short io */
2140 aa->aa_requested_nob :
2141 req->rq_bulk->bd_nob_transferred);
2143 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2144 ptlrpc_lprocfs_brw(req, transferred);
2146 spin_lock(&cli->cl_loi_list_lock);
2147 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2148 * is called so we know whether to go to sync BRWs or wait for more
2149 * RPCs to complete */
2150 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2151 cli->cl_w_in_flight--;
2153 cli->cl_r_in_flight--;
2154 osc_wake_cache_waiters(cli);
2155 spin_unlock(&cli->cl_loi_list_lock);
2157 osc_io_unplug(env, cli, NULL);
2161 static void brw_commit(struct ptlrpc_request *req)
2163 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2164 * this called via the rq_commit_cb, I need to ensure
2165 * osc_dec_unstable_pages is still called. Otherwise unstable
2166 * pages may be leaked. */
2167 spin_lock(&req->rq_lock);
2168 if (likely(req->rq_unstable)) {
2169 req->rq_unstable = 0;
2170 spin_unlock(&req->rq_lock);
2172 osc_dec_unstable_pages(req);
2174 req->rq_committed = 1;
2175 spin_unlock(&req->rq_lock);
2180 * Build an RPC by the list of extent @ext_list. The caller must ensure
2181 * that the total pages in this list are NOT over max pages per RPC.
2182 * Extents in the list must be in OES_RPC state.
2184 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2185 struct list_head *ext_list, int cmd)
2187 struct ptlrpc_request *req = NULL;
2188 struct osc_extent *ext;
2189 struct brw_page **pga = NULL;
2190 struct osc_brw_async_args *aa = NULL;
2191 struct obdo *oa = NULL;
2192 struct osc_async_page *oap;
2193 struct osc_object *obj = NULL;
2194 struct cl_req_attr *crattr = NULL;
2195 loff_t starting_offset = OBD_OBJECT_EOF;
2196 loff_t ending_offset = 0;
2200 bool soft_sync = false;
2201 bool interrupted = false;
2202 bool ndelay = false;
2206 __u32 layout_version = 0;
2207 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
2208 struct ost_body *body;
2210 LASSERT(!list_empty(ext_list));
2212 /* add pages into rpc_list to build BRW rpc */
2213 list_for_each_entry(ext, ext_list, oe_link) {
2214 LASSERT(ext->oe_state == OES_RPC);
2215 mem_tight |= ext->oe_memalloc;
2216 grant += ext->oe_grants;
2217 page_count += ext->oe_nr_pages;
2218 layout_version = MAX(layout_version, ext->oe_layout_version);
2223 soft_sync = osc_over_unstable_soft_limit(cli);
2225 mpflag = cfs_memory_pressure_get_and_set();
2227 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2229 GOTO(out, rc = -ENOMEM);
2231 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2233 GOTO(out, rc = -ENOMEM);
2236 list_for_each_entry(ext, ext_list, oe_link) {
2237 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2239 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2241 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2242 pga[i] = &oap->oap_brw_page;
2243 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2246 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2247 if (starting_offset == OBD_OBJECT_EOF ||
2248 starting_offset > oap->oap_obj_off)
2249 starting_offset = oap->oap_obj_off;
2251 LASSERT(oap->oap_page_off == 0);
2252 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2253 ending_offset = oap->oap_obj_off +
2256 LASSERT(oap->oap_page_off + oap->oap_count ==
2258 if (oap->oap_interrupted)
2265 /* first page in the list */
2266 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2268 crattr = &osc_env_info(env)->oti_req_attr;
2269 memset(crattr, 0, sizeof(*crattr));
2270 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2271 crattr->cra_flags = ~0ULL;
2272 crattr->cra_page = oap2cl_page(oap);
2273 crattr->cra_oa = oa;
2274 cl_req_attr_set(env, osc2cl(obj), crattr);
2276 if (cmd == OBD_BRW_WRITE) {
2277 oa->o_grant_used = grant;
2278 if (layout_version > 0) {
2279 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2280 PFID(&oa->o_oi.oi_fid), layout_version);
2282 oa->o_layout_version = layout_version;
2283 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2287 sort_brw_pages(pga, page_count);
2288 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2290 CERROR("prep_req failed: %d\n", rc);
2294 req->rq_commit_cb = brw_commit;
2295 req->rq_interpret_reply = brw_interpret;
2296 req->rq_memalloc = mem_tight != 0;
2297 oap->oap_request = ptlrpc_request_addref(req);
2298 if (interrupted && !req->rq_intr)
2299 ptlrpc_mark_interrupted(req);
2301 req->rq_no_resend = req->rq_no_delay = 1;
2302 /* probably set a shorter timeout value.
2303 * to handle ETIMEDOUT in brw_interpret() correctly. */
2304 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2307 /* Need to update the timestamps after the request is built in case
2308 * we race with setattr (locally or in queue at OST). If OST gets
2309 * later setattr before earlier BRW (as determined by the request xid),
2310 * the OST will not use BRW timestamps. Sadly, there is no obvious
2311 * way to do this in a single call. bug 10150 */
2312 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2313 crattr->cra_oa = &body->oa;
2314 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2315 cl_req_attr_set(env, osc2cl(obj), crattr);
2316 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2318 aa = ptlrpc_req_async_args(aa, req);
2319 INIT_LIST_HEAD(&aa->aa_oaps);
2320 list_splice_init(&rpc_list, &aa->aa_oaps);
2321 INIT_LIST_HEAD(&aa->aa_exts);
2322 list_splice_init(ext_list, &aa->aa_exts);
2324 spin_lock(&cli->cl_loi_list_lock);
2325 starting_offset >>= PAGE_SHIFT;
2326 if (cmd == OBD_BRW_READ) {
2327 cli->cl_r_in_flight++;
2328 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2329 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2330 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2331 starting_offset + 1);
2333 cli->cl_w_in_flight++;
2334 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2335 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2336 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2337 starting_offset + 1);
2339 spin_unlock(&cli->cl_loi_list_lock);
2341 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2342 page_count, aa, cli->cl_r_in_flight,
2343 cli->cl_w_in_flight);
2344 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2346 ptlrpcd_add_req(req);
2352 cfs_memory_pressure_restore(mpflag);
2355 LASSERT(req == NULL);
2358 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2360 OBD_FREE(pga, sizeof(*pga) * page_count);
2361 /* this should happen rarely and is pretty bad, it makes the
2362 * pending list not follow the dirty order */
2363 while (!list_empty(ext_list)) {
2364 ext = list_entry(ext_list->next, struct osc_extent,
2366 list_del_init(&ext->oe_link);
2367 osc_extent_finish(env, ext, 0, rc);
2373 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2377 LASSERT(lock != NULL);
2379 lock_res_and_lock(lock);
2381 if (lock->l_ast_data == NULL)
2382 lock->l_ast_data = data;
2383 if (lock->l_ast_data == data)
2386 unlock_res_and_lock(lock);
2391 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2392 void *cookie, struct lustre_handle *lockh,
2393 enum ldlm_mode mode, __u64 *flags, bool speculative,
2396 bool intent = *flags & LDLM_FL_HAS_INTENT;
2400 /* The request was created before ldlm_cli_enqueue call. */
2401 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2402 struct ldlm_reply *rep;
2404 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2405 LASSERT(rep != NULL);
2407 rep->lock_policy_res1 =
2408 ptlrpc_status_ntoh(rep->lock_policy_res1);
2409 if (rep->lock_policy_res1)
2410 errcode = rep->lock_policy_res1;
2412 *flags |= LDLM_FL_LVB_READY;
2413 } else if (errcode == ELDLM_OK) {
2414 *flags |= LDLM_FL_LVB_READY;
2417 /* Call the update callback. */
2418 rc = (*upcall)(cookie, lockh, errcode);
2420 /* release the reference taken in ldlm_cli_enqueue() */
2421 if (errcode == ELDLM_LOCK_MATCHED)
2423 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2424 ldlm_lock_decref(lockh, mode);
2429 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2432 struct osc_enqueue_args *aa = args;
2433 struct ldlm_lock *lock;
2434 struct lustre_handle *lockh = &aa->oa_lockh;
2435 enum ldlm_mode mode = aa->oa_mode;
2436 struct ost_lvb *lvb = aa->oa_lvb;
2437 __u32 lvb_len = sizeof(*lvb);
2442 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2444 lock = ldlm_handle2lock(lockh);
2445 LASSERTF(lock != NULL,
2446 "lockh %#llx, req %p, aa %p - client evicted?\n",
2447 lockh->cookie, req, aa);
2449 /* Take an additional reference so that a blocking AST that
2450 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2451 * to arrive after an upcall has been executed by
2452 * osc_enqueue_fini(). */
2453 ldlm_lock_addref(lockh, mode);
2455 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2456 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2458 /* Let CP AST to grant the lock first. */
2459 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2461 if (aa->oa_speculative) {
2462 LASSERT(aa->oa_lvb == NULL);
2463 LASSERT(aa->oa_flags == NULL);
2464 aa->oa_flags = &flags;
2467 /* Complete obtaining the lock procedure. */
2468 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2469 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2471 /* Complete osc stuff. */
2472 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2473 aa->oa_flags, aa->oa_speculative, rc);
2475 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2477 ldlm_lock_decref(lockh, mode);
2478 LDLM_LOCK_PUT(lock);
2482 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2484 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2485 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2486 * other synchronous requests, however keeping some locks and trying to obtain
2487 * others may take a considerable amount of time in a case of ost failure; and
2488 * when other sync requests do not get released lock from a client, the client
2489 * is evicted from the cluster -- such scenarious make the life difficult, so
2490 * release locks just after they are obtained. */
2491 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2492 __u64 *flags, union ldlm_policy_data *policy,
2493 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2494 void *cookie, struct ldlm_enqueue_info *einfo,
2495 struct ptlrpc_request_set *rqset, int async,
2498 struct obd_device *obd = exp->exp_obd;
2499 struct lustre_handle lockh = { 0 };
2500 struct ptlrpc_request *req = NULL;
2501 int intent = *flags & LDLM_FL_HAS_INTENT;
2502 __u64 match_flags = *flags;
2503 enum ldlm_mode mode;
2507 /* Filesystem lock extents are extended to page boundaries so that
2508 * dealing with the page cache is a little smoother. */
2509 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2510 policy->l_extent.end |= ~PAGE_MASK;
2512 /* Next, search for already existing extent locks that will cover us */
2513 /* If we're trying to read, we also search for an existing PW lock. The
2514 * VFS and page cache already protect us locally, so lots of readers/
2515 * writers can share a single PW lock.
2517 * There are problems with conversion deadlocks, so instead of
2518 * converting a read lock to a write lock, we'll just enqueue a new
2521 * At some point we should cancel the read lock instead of making them
2522 * send us a blocking callback, but there are problems with canceling
2523 * locks out from other users right now, too. */
2524 mode = einfo->ei_mode;
2525 if (einfo->ei_mode == LCK_PR)
2527 /* Normal lock requests must wait for the LVB to be ready before
2528 * matching a lock; speculative lock requests do not need to,
2529 * because they will not actually use the lock. */
2531 match_flags |= LDLM_FL_LVB_READY;
2533 match_flags |= LDLM_FL_BLOCK_GRANTED;
2534 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2535 einfo->ei_type, policy, mode, &lockh, 0);
2537 struct ldlm_lock *matched;
2539 if (*flags & LDLM_FL_TEST_LOCK)
2542 matched = ldlm_handle2lock(&lockh);
2544 /* This DLM lock request is speculative, and does not
2545 * have an associated IO request. Therefore if there
2546 * is already a DLM lock, it wll just inform the
2547 * caller to cancel the request for this stripe.*/
2548 lock_res_and_lock(matched);
2549 if (ldlm_extent_equal(&policy->l_extent,
2550 &matched->l_policy_data.l_extent))
2554 unlock_res_and_lock(matched);
2556 ldlm_lock_decref(&lockh, mode);
2557 LDLM_LOCK_PUT(matched);
2559 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2560 *flags |= LDLM_FL_LVB_READY;
2562 /* We already have a lock, and it's referenced. */
2563 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2565 ldlm_lock_decref(&lockh, mode);
2566 LDLM_LOCK_PUT(matched);
2569 ldlm_lock_decref(&lockh, mode);
2570 LDLM_LOCK_PUT(matched);
2574 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2578 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2579 &RQF_LDLM_ENQUEUE_LVB);
2583 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2585 ptlrpc_request_free(req);
2589 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2591 ptlrpc_request_set_replen(req);
2594 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2595 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2597 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2598 sizeof(*lvb), LVB_T_OST, &lockh, async);
2601 struct osc_enqueue_args *aa;
2602 aa = ptlrpc_req_async_args(aa, req);
2604 aa->oa_mode = einfo->ei_mode;
2605 aa->oa_type = einfo->ei_type;
2606 lustre_handle_copy(&aa->oa_lockh, &lockh);
2607 aa->oa_upcall = upcall;
2608 aa->oa_cookie = cookie;
2609 aa->oa_speculative = speculative;
2611 aa->oa_flags = flags;
2614 /* speculative locks are essentially to enqueue
2615 * a DLM lock in advance, so we don't care
2616 * about the result of the enqueue. */
2618 aa->oa_flags = NULL;
2621 req->rq_interpret_reply = osc_enqueue_interpret;
2622 if (rqset == PTLRPCD_SET)
2623 ptlrpcd_add_req(req);
2625 ptlrpc_set_add_req(rqset, req);
2626 } else if (intent) {
2627 ptlrpc_req_finished(req);
2632 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2633 flags, speculative, rc);
2635 ptlrpc_req_finished(req);
2640 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2641 struct ldlm_res_id *res_id, enum ldlm_type type,
2642 union ldlm_policy_data *policy, enum ldlm_mode mode,
2643 __u64 *flags, struct osc_object *obj,
2644 struct lustre_handle *lockh, int unref)
2646 struct obd_device *obd = exp->exp_obd;
2647 __u64 lflags = *flags;
2651 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2654 /* Filesystem lock extents are extended to page boundaries so that
2655 * dealing with the page cache is a little smoother */
2656 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2657 policy->l_extent.end |= ~PAGE_MASK;
2659 /* Next, search for already existing extent locks that will cover us */
2660 /* If we're trying to read, we also search for an existing PW lock. The
2661 * VFS and page cache already protect us locally, so lots of readers/
2662 * writers can share a single PW lock. */
2666 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2667 res_id, type, policy, rc, lockh, unref);
2668 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2672 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2674 LASSERT(lock != NULL);
2675 if (osc_set_lock_data(lock, obj)) {
2676 lock_res_and_lock(lock);
2677 if (!ldlm_is_lvb_cached(lock)) {
2678 LASSERT(lock->l_ast_data == obj);
2679 osc_lock_lvb_update(env, obj, lock, NULL);
2680 ldlm_set_lvb_cached(lock);
2682 unlock_res_and_lock(lock);
2684 ldlm_lock_decref(lockh, rc);
2687 LDLM_LOCK_PUT(lock);
2692 static int osc_statfs_interpret(const struct lu_env *env,
2693 struct ptlrpc_request *req, void *args, int rc)
2695 struct osc_async_args *aa = args;
2696 struct obd_statfs *msfs;
2701 * The request has in fact never been sent due to issues at
2702 * a higher level (LOV). Exit immediately since the caller
2703 * is aware of the problem and takes care of the clean up.
2707 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2708 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2714 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2716 GOTO(out, rc = -EPROTO);
2718 *aa->aa_oi->oi_osfs = *msfs;
2720 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2725 static int osc_statfs_async(struct obd_export *exp,
2726 struct obd_info *oinfo, time64_t max_age,
2727 struct ptlrpc_request_set *rqset)
2729 struct obd_device *obd = class_exp2obd(exp);
2730 struct ptlrpc_request *req;
2731 struct osc_async_args *aa;
2735 if (obd->obd_osfs_age >= max_age) {
2737 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2738 obd->obd_name, &obd->obd_osfs,
2739 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2740 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2741 spin_lock(&obd->obd_osfs_lock);
2742 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2743 spin_unlock(&obd->obd_osfs_lock);
2744 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2745 if (oinfo->oi_cb_up)
2746 oinfo->oi_cb_up(oinfo, 0);
2751 /* We could possibly pass max_age in the request (as an absolute
2752 * timestamp or a "seconds.usec ago") so the target can avoid doing
2753 * extra calls into the filesystem if that isn't necessary (e.g.
2754 * during mount that would help a bit). Having relative timestamps
2755 * is not so great if request processing is slow, while absolute
2756 * timestamps are not ideal because they need time synchronization. */
2757 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2761 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2763 ptlrpc_request_free(req);
2766 ptlrpc_request_set_replen(req);
2767 req->rq_request_portal = OST_CREATE_PORTAL;
2768 ptlrpc_at_set_req_timeout(req);
2770 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2771 /* procfs requests not want stat in wait for avoid deadlock */
2772 req->rq_no_resend = 1;
2773 req->rq_no_delay = 1;
2776 req->rq_interpret_reply = osc_statfs_interpret;
2777 aa = ptlrpc_req_async_args(aa, req);
2780 ptlrpc_set_add_req(rqset, req);
2784 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2785 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2787 struct obd_device *obd = class_exp2obd(exp);
2788 struct obd_statfs *msfs;
2789 struct ptlrpc_request *req;
2790 struct obd_import *imp = NULL;
2795 /*Since the request might also come from lprocfs, so we need
2796 *sync this with client_disconnect_export Bug15684*/
2797 down_read(&obd->u.cli.cl_sem);
2798 if (obd->u.cli.cl_import)
2799 imp = class_import_get(obd->u.cli.cl_import);
2800 up_read(&obd->u.cli.cl_sem);
2804 /* We could possibly pass max_age in the request (as an absolute
2805 * timestamp or a "seconds.usec ago") so the target can avoid doing
2806 * extra calls into the filesystem if that isn't necessary (e.g.
2807 * during mount that would help a bit). Having relative timestamps
2808 * is not so great if request processing is slow, while absolute
2809 * timestamps are not ideal because they need time synchronization. */
2810 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2812 class_import_put(imp);
2817 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2819 ptlrpc_request_free(req);
2822 ptlrpc_request_set_replen(req);
2823 req->rq_request_portal = OST_CREATE_PORTAL;
2824 ptlrpc_at_set_req_timeout(req);
2826 if (flags & OBD_STATFS_NODELAY) {
2827 /* procfs requests not want stat in wait for avoid deadlock */
2828 req->rq_no_resend = 1;
2829 req->rq_no_delay = 1;
2832 rc = ptlrpc_queue_wait(req);
2836 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2838 GOTO(out, rc = -EPROTO);
2844 ptlrpc_req_finished(req);
2848 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2849 void *karg, void __user *uarg)
2851 struct obd_device *obd = exp->exp_obd;
2852 struct obd_ioctl_data *data = karg;
2856 if (!try_module_get(THIS_MODULE)) {
2857 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2858 module_name(THIS_MODULE));
2862 case OBD_IOC_CLIENT_RECOVER:
2863 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
2864 data->ioc_inlbuf1, 0);
2868 case IOC_OSC_SET_ACTIVE:
2869 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
2874 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
2875 obd->obd_name, cmd, current_comm(), rc);
2879 module_put(THIS_MODULE);
2883 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2884 u32 keylen, void *key, u32 vallen, void *val,
2885 struct ptlrpc_request_set *set)
2887 struct ptlrpc_request *req;
2888 struct obd_device *obd = exp->exp_obd;
2889 struct obd_import *imp = class_exp2cliimp(exp);
2894 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2896 if (KEY_IS(KEY_CHECKSUM)) {
2897 if (vallen != sizeof(int))
2899 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2903 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2904 sptlrpc_conf_client_adapt(obd);
2908 if (KEY_IS(KEY_FLUSH_CTX)) {
2909 sptlrpc_import_flush_my_ctx(imp);
2913 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2914 struct client_obd *cli = &obd->u.cli;
2915 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2916 long target = *(long *)val;
2918 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2923 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2926 /* We pass all other commands directly to OST. Since nobody calls osc
2927 methods directly and everybody is supposed to go through LOV, we
2928 assume lov checked invalid values for us.
2929 The only recognised values so far are evict_by_nid and mds_conn.
2930 Even if something bad goes through, we'd get a -EINVAL from OST
2933 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2934 &RQF_OST_SET_GRANT_INFO :
2939 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2940 RCL_CLIENT, keylen);
2941 if (!KEY_IS(KEY_GRANT_SHRINK))
2942 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2943 RCL_CLIENT, vallen);
2944 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2946 ptlrpc_request_free(req);
2950 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2951 memcpy(tmp, key, keylen);
2952 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2955 memcpy(tmp, val, vallen);
2957 if (KEY_IS(KEY_GRANT_SHRINK)) {
2958 struct osc_grant_args *aa;
2961 aa = ptlrpc_req_async_args(aa, req);
2962 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2964 ptlrpc_req_finished(req);
2967 *oa = ((struct ost_body *)val)->oa;
2969 req->rq_interpret_reply = osc_shrink_grant_interpret;
2972 ptlrpc_request_set_replen(req);
2973 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2974 LASSERT(set != NULL);
2975 ptlrpc_set_add_req(set, req);
2976 ptlrpc_check_set(NULL, set);
2978 ptlrpcd_add_req(req);
2983 EXPORT_SYMBOL(osc_set_info_async);
2985 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
2986 struct obd_device *obd, struct obd_uuid *cluuid,
2987 struct obd_connect_data *data, void *localdata)
2989 struct client_obd *cli = &obd->u.cli;
2991 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2995 spin_lock(&cli->cl_loi_list_lock);
2996 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2997 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
2998 /* restore ocd_grant_blkbits as client page bits */
2999 data->ocd_grant_blkbits = PAGE_SHIFT;
3000 grant += cli->cl_dirty_grant;
3002 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3004 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3005 lost_grant = cli->cl_lost_grant;
3006 cli->cl_lost_grant = 0;
3007 spin_unlock(&cli->cl_loi_list_lock);
3009 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3010 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3011 data->ocd_version, data->ocd_grant, lost_grant);
3016 EXPORT_SYMBOL(osc_reconnect);
3018 int osc_disconnect(struct obd_export *exp)
3020 struct obd_device *obd = class_exp2obd(exp);
3023 rc = client_disconnect_export(exp);
3025 * Initially we put del_shrink_grant before disconnect_export, but it
3026 * causes the following problem if setup (connect) and cleanup
3027 * (disconnect) are tangled together.
3028 * connect p1 disconnect p2
3029 * ptlrpc_connect_import
3030 * ............... class_manual_cleanup
3033 * ptlrpc_connect_interrupt
3035 * add this client to shrink list
3037 * Bang! grant shrink thread trigger the shrink. BUG18662
3039 osc_del_grant_list(&obd->u.cli);
3042 EXPORT_SYMBOL(osc_disconnect);
3044 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3045 struct hlist_node *hnode, void *arg)
3047 struct lu_env *env = arg;
3048 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3049 struct ldlm_lock *lock;
3050 struct osc_object *osc = NULL;
3054 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3055 if (lock->l_ast_data != NULL && osc == NULL) {
3056 osc = lock->l_ast_data;
3057 cl_object_get(osc2cl(osc));
3060 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3061 * by the 2nd round of ldlm_namespace_clean() call in
3062 * osc_import_event(). */
3063 ldlm_clear_cleaned(lock);
3068 osc_object_invalidate(env, osc);
3069 cl_object_put(env, osc2cl(osc));
3074 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3076 static int osc_import_event(struct obd_device *obd,
3077 struct obd_import *imp,
3078 enum obd_import_event event)
3080 struct client_obd *cli;
3084 LASSERT(imp->imp_obd == obd);
3087 case IMP_EVENT_DISCON: {
3089 spin_lock(&cli->cl_loi_list_lock);
3090 cli->cl_avail_grant = 0;
3091 cli->cl_lost_grant = 0;
3092 spin_unlock(&cli->cl_loi_list_lock);
3095 case IMP_EVENT_INACTIVE: {
3096 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3099 case IMP_EVENT_INVALIDATE: {
3100 struct ldlm_namespace *ns = obd->obd_namespace;
3104 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3106 env = cl_env_get(&refcheck);
3108 osc_io_unplug(env, &obd->u.cli, NULL);
3110 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3111 osc_ldlm_resource_invalidate,
3113 cl_env_put(env, &refcheck);
3115 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3120 case IMP_EVENT_ACTIVE: {
3121 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3124 case IMP_EVENT_OCD: {
3125 struct obd_connect_data *ocd = &imp->imp_connect_data;
3127 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3128 osc_init_grant(&obd->u.cli, ocd);
3131 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3132 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3134 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3137 case IMP_EVENT_DEACTIVATE: {
3138 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3141 case IMP_EVENT_ACTIVATE: {
3142 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3146 CERROR("Unknown import event %d\n", event);
3153 * Determine whether the lock can be canceled before replaying the lock
3154 * during recovery, see bug16774 for detailed information.
3156 * \retval zero the lock can't be canceled
3157 * \retval other ok to cancel
3159 static int osc_cancel_weight(struct ldlm_lock *lock)
3162 * Cancel all unused and granted extent lock.
3164 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3165 ldlm_is_granted(lock) &&
3166 osc_ldlm_weigh_ast(lock) == 0)
3172 static int brw_queue_work(const struct lu_env *env, void *data)
3174 struct client_obd *cli = data;
3176 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3178 osc_io_unplug(env, cli, NULL);
3182 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3184 struct client_obd *cli = &obd->u.cli;
3190 rc = ptlrpcd_addref();
3194 rc = client_obd_setup(obd, lcfg);
3196 GOTO(out_ptlrpcd, rc);
3199 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3200 if (IS_ERR(handler))
3201 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3202 cli->cl_writeback_work = handler;
3204 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3205 if (IS_ERR(handler))
3206 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3207 cli->cl_lru_work = handler;
3209 rc = osc_quota_setup(obd);
3211 GOTO(out_ptlrpcd_work, rc);
3213 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3214 osc_update_next_shrink(cli);
3219 if (cli->cl_writeback_work != NULL) {
3220 ptlrpcd_destroy_work(cli->cl_writeback_work);
3221 cli->cl_writeback_work = NULL;
3223 if (cli->cl_lru_work != NULL) {
3224 ptlrpcd_destroy_work(cli->cl_lru_work);
3225 cli->cl_lru_work = NULL;
3227 client_obd_cleanup(obd);
3232 EXPORT_SYMBOL(osc_setup_common);
3234 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3236 struct client_obd *cli = &obd->u.cli;
3244 rc = osc_setup_common(obd, lcfg);
3248 rc = osc_tunables_init(obd);
3253 * We try to control the total number of requests with a upper limit
3254 * osc_reqpool_maxreqcount. There might be some race which will cause
3255 * over-limit allocation, but it is fine.
3257 req_count = atomic_read(&osc_pool_req_count);
3258 if (req_count < osc_reqpool_maxreqcount) {
3259 adding = cli->cl_max_rpcs_in_flight + 2;
3260 if (req_count + adding > osc_reqpool_maxreqcount)
3261 adding = osc_reqpool_maxreqcount - req_count;
3263 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3264 atomic_add(added, &osc_pool_req_count);
3267 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3269 spin_lock(&osc_shrink_lock);
3270 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3271 spin_unlock(&osc_shrink_lock);
3272 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3273 cli->cl_import->imp_idle_debug = D_HA;
3278 int osc_precleanup_common(struct obd_device *obd)
3280 struct client_obd *cli = &obd->u.cli;
3284 * for echo client, export may be on zombie list, wait for
3285 * zombie thread to cull it, because cli.cl_import will be
3286 * cleared in client_disconnect_export():
3287 * class_export_destroy() -> obd_cleanup() ->
3288 * echo_device_free() -> echo_client_cleanup() ->
3289 * obd_disconnect() -> osc_disconnect() ->
3290 * client_disconnect_export()
3292 obd_zombie_barrier();
3293 if (cli->cl_writeback_work) {
3294 ptlrpcd_destroy_work(cli->cl_writeback_work);
3295 cli->cl_writeback_work = NULL;
3298 if (cli->cl_lru_work) {
3299 ptlrpcd_destroy_work(cli->cl_lru_work);
3300 cli->cl_lru_work = NULL;
3303 obd_cleanup_client_import(obd);
3306 EXPORT_SYMBOL(osc_precleanup_common);
3308 static int osc_precleanup(struct obd_device *obd)
3312 osc_precleanup_common(obd);
3314 ptlrpc_lprocfs_unregister_obd(obd);
3318 int osc_cleanup_common(struct obd_device *obd)
3320 struct client_obd *cli = &obd->u.cli;
3325 spin_lock(&osc_shrink_lock);
3326 list_del(&cli->cl_shrink_list);
3327 spin_unlock(&osc_shrink_lock);
3330 if (cli->cl_cache != NULL) {
3331 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3332 spin_lock(&cli->cl_cache->ccc_lru_lock);
3333 list_del_init(&cli->cl_lru_osc);
3334 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3335 cli->cl_lru_left = NULL;
3336 cl_cache_decref(cli->cl_cache);
3337 cli->cl_cache = NULL;
3340 /* free memory of osc quota cache */
3341 osc_quota_cleanup(obd);
3343 rc = client_obd_cleanup(obd);
3348 EXPORT_SYMBOL(osc_cleanup_common);
3350 static struct obd_ops osc_obd_ops = {
3351 .o_owner = THIS_MODULE,
3352 .o_setup = osc_setup,
3353 .o_precleanup = osc_precleanup,
3354 .o_cleanup = osc_cleanup_common,
3355 .o_add_conn = client_import_add_conn,
3356 .o_del_conn = client_import_del_conn,
3357 .o_connect = client_connect_import,
3358 .o_reconnect = osc_reconnect,
3359 .o_disconnect = osc_disconnect,
3360 .o_statfs = osc_statfs,
3361 .o_statfs_async = osc_statfs_async,
3362 .o_create = osc_create,
3363 .o_destroy = osc_destroy,
3364 .o_getattr = osc_getattr,
3365 .o_setattr = osc_setattr,
3366 .o_iocontrol = osc_iocontrol,
3367 .o_set_info_async = osc_set_info_async,
3368 .o_import_event = osc_import_event,
3369 .o_quotactl = osc_quotactl,
3372 static struct shrinker *osc_cache_shrinker;
3373 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
3374 DEFINE_SPINLOCK(osc_shrink_lock);
3376 #ifndef HAVE_SHRINKER_COUNT
3377 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3379 struct shrink_control scv = {
3380 .nr_to_scan = shrink_param(sc, nr_to_scan),
3381 .gfp_mask = shrink_param(sc, gfp_mask)
3383 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
3384 struct shrinker *shrinker = NULL;
3387 (void)osc_cache_shrink_scan(shrinker, &scv);
3389 return osc_cache_shrink_count(shrinker, &scv);
3393 static int __init osc_init(void)
3395 unsigned int reqpool_size;
3396 unsigned int reqsize;
3398 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3399 osc_cache_shrink_count, osc_cache_shrink_scan);
3402 /* print an address of _any_ initialized kernel symbol from this
3403 * module, to allow debugging with gdb that doesn't support data
3404 * symbols from modules.*/
3405 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3407 rc = lu_kmem_init(osc_caches);
3411 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3412 LUSTRE_OSC_NAME, &osc_device_type);
3416 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3418 /* This is obviously too much memory, only prevent overflow here */
3419 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3420 GOTO(out_type, rc = -EINVAL);
3422 reqpool_size = osc_reqpool_mem_max << 20;
3425 while (reqsize < OST_IO_MAXREQSIZE)
3426 reqsize = reqsize << 1;
3429 * We don't enlarge the request count in OSC pool according to
3430 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3431 * tried after normal allocation failed. So a small OSC pool won't
3432 * cause much performance degression in most of cases.
3434 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3436 atomic_set(&osc_pool_req_count, 0);
3437 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3438 ptlrpc_add_rqs_to_pool);
3440 if (osc_rq_pool == NULL)
3441 GOTO(out_type, rc = -ENOMEM);
3443 rc = osc_start_grant_work();
3445 GOTO(out_req_pool, rc);
3450 ptlrpc_free_rq_pool(osc_rq_pool);
3452 class_unregister_type(LUSTRE_OSC_NAME);
3454 lu_kmem_fini(osc_caches);
3459 static void __exit osc_exit(void)
3461 osc_stop_grant_work();
3462 remove_shrinker(osc_cache_shrinker);
3463 class_unregister_type(LUSTRE_OSC_NAME);
3464 lu_kmem_fini(osc_caches);
3465 ptlrpc_free_rq_pool(osc_rq_pool);
3468 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3469 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3470 MODULE_VERSION(LUSTRE_VERSION_STRING);
3471 MODULE_LICENSE("GPL");
3473 module_init(osc_init);
3474 module_exit(osc_exit);