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 <libcfs/libcfs.h>
37 #include <linux/falloc.h>
38 #include <lprocfs_status.h>
39 #include <lustre_dlm.h>
40 #include <lustre_fid.h>
41 #include <lustre_ha.h>
42 #include <uapi/linux/lustre/lustre_ioctl.h>
43 #include <lustre_net.h>
44 #include <lustre_obdo.h>
46 #include <obd_cksum.h>
47 #include <obd_class.h>
48 #include <lustre_osc.h>
49 #include <linux/falloc.h>
51 #include "osc_internal.h"
53 atomic_t osc_pool_req_count;
54 unsigned int osc_reqpool_maxreqcount;
55 struct ptlrpc_request_pool *osc_rq_pool;
57 /* max memory used for request pool, unit is MB */
58 static unsigned int osc_reqpool_mem_max = 5;
59 module_param(osc_reqpool_mem_max, uint, 0444);
61 static int osc_idle_timeout = 20;
62 module_param(osc_idle_timeout, uint, 0644);
64 #define osc_grant_args osc_brw_async_args
66 struct osc_setattr_args {
68 obd_enqueue_update_f sa_upcall;
72 struct osc_fsync_args {
73 struct osc_object *fa_obj;
75 obd_enqueue_update_f fa_upcall;
79 struct osc_ladvise_args {
81 obd_enqueue_update_f la_upcall;
85 static void osc_release_ppga(struct brw_page **ppga, size_t count);
86 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
89 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
91 struct ost_body *body;
93 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
96 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
99 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
102 struct ptlrpc_request *req;
103 struct ost_body *body;
107 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
111 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
113 ptlrpc_request_free(req);
117 osc_pack_req_body(req, oa);
119 ptlrpc_request_set_replen(req);
121 rc = ptlrpc_queue_wait(req);
125 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
127 GOTO(out, rc = -EPROTO);
129 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
130 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
132 oa->o_blksize = cli_brw_size(exp->exp_obd);
133 oa->o_valid |= OBD_MD_FLBLKSZ;
137 ptlrpc_req_finished(req);
142 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
145 struct ptlrpc_request *req;
146 struct ost_body *body;
150 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
152 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
156 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
158 ptlrpc_request_free(req);
162 osc_pack_req_body(req, oa);
164 ptlrpc_request_set_replen(req);
166 rc = ptlrpc_queue_wait(req);
170 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
172 GOTO(out, rc = -EPROTO);
174 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
178 ptlrpc_req_finished(req);
183 static int osc_setattr_interpret(const struct lu_env *env,
184 struct ptlrpc_request *req, void *args, int rc)
186 struct osc_setattr_args *sa = args;
187 struct ost_body *body;
194 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
196 GOTO(out, rc = -EPROTO);
198 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
201 rc = sa->sa_upcall(sa->sa_cookie, rc);
205 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
206 obd_enqueue_update_f upcall, void *cookie,
207 struct ptlrpc_request_set *rqset)
209 struct ptlrpc_request *req;
210 struct osc_setattr_args *sa;
215 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
219 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
221 ptlrpc_request_free(req);
225 osc_pack_req_body(req, oa);
227 ptlrpc_request_set_replen(req);
229 /* do mds to ost setattr asynchronously */
231 /* Do not wait for response. */
232 ptlrpcd_add_req(req);
234 req->rq_interpret_reply = osc_setattr_interpret;
236 sa = ptlrpc_req_async_args(sa, req);
238 sa->sa_upcall = upcall;
239 sa->sa_cookie = cookie;
241 ptlrpc_set_add_req(rqset, req);
247 static int osc_ladvise_interpret(const struct lu_env *env,
248 struct ptlrpc_request *req,
251 struct osc_ladvise_args *la = arg;
252 struct ost_body *body;
258 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
260 GOTO(out, rc = -EPROTO);
262 *la->la_oa = body->oa;
264 rc = la->la_upcall(la->la_cookie, rc);
269 * If rqset is NULL, do not wait for response. Upcall and cookie could also
270 * be NULL in this case
272 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
273 struct ladvise_hdr *ladvise_hdr,
274 obd_enqueue_update_f upcall, void *cookie,
275 struct ptlrpc_request_set *rqset)
277 struct ptlrpc_request *req;
278 struct ost_body *body;
279 struct osc_ladvise_args *la;
281 struct lu_ladvise *req_ladvise;
282 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
283 int num_advise = ladvise_hdr->lah_count;
284 struct ladvise_hdr *req_ladvise_hdr;
287 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
291 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
292 num_advise * sizeof(*ladvise));
293 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
295 ptlrpc_request_free(req);
298 req->rq_request_portal = OST_IO_PORTAL;
299 ptlrpc_at_set_req_timeout(req);
301 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
303 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
306 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
307 &RMF_OST_LADVISE_HDR);
308 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
310 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
311 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
312 ptlrpc_request_set_replen(req);
315 /* Do not wait for response. */
316 ptlrpcd_add_req(req);
320 req->rq_interpret_reply = osc_ladvise_interpret;
321 la = ptlrpc_req_async_args(la, req);
323 la->la_upcall = upcall;
324 la->la_cookie = cookie;
326 ptlrpc_set_add_req(rqset, req);
331 static int osc_create(const struct lu_env *env, struct obd_export *exp,
334 struct ptlrpc_request *req;
335 struct ost_body *body;
340 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
341 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
343 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
345 GOTO(out, rc = -ENOMEM);
347 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
349 ptlrpc_request_free(req);
353 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
356 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
358 ptlrpc_request_set_replen(req);
360 rc = ptlrpc_queue_wait(req);
364 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
366 GOTO(out_req, rc = -EPROTO);
368 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
369 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
371 oa->o_blksize = cli_brw_size(exp->exp_obd);
372 oa->o_valid |= OBD_MD_FLBLKSZ;
374 CDEBUG(D_HA, "transno: %lld\n",
375 lustre_msg_get_transno(req->rq_repmsg));
377 ptlrpc_req_finished(req);
382 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
383 obd_enqueue_update_f upcall, void *cookie)
385 struct ptlrpc_request *req;
386 struct osc_setattr_args *sa;
387 struct obd_import *imp = class_exp2cliimp(exp);
388 struct ost_body *body;
393 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
397 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
399 ptlrpc_request_free(req);
403 osc_set_io_portal(req);
405 ptlrpc_at_set_req_timeout(req);
407 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
409 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
411 ptlrpc_request_set_replen(req);
413 req->rq_interpret_reply = osc_setattr_interpret;
414 sa = ptlrpc_req_async_args(sa, req);
416 sa->sa_upcall = upcall;
417 sa->sa_cookie = cookie;
419 ptlrpcd_add_req(req);
423 EXPORT_SYMBOL(osc_punch_send);
426 * osc_fallocate_base() - Handles fallocate request.
428 * @exp: Export structure
429 * @oa: Attributes passed to OSS from client (obdo structure)
430 * @upcall: Primary & supplementary group information
431 * @cookie: Exclusive identifier
432 * @rqset: Request list.
433 * @mode: Operation done on given range.
435 * osc_fallocate_base() - Handles fallocate requests only. Only block
436 * allocation or standard preallocate operation is supported currently.
437 * Other mode flags is not supported yet. ftruncate(2) or truncate(2)
438 * is supported via SETATTR request.
440 * Return: Non-zero on failure and O on success.
442 int osc_fallocate_base(struct obd_export *exp, struct obdo *oa,
443 obd_enqueue_update_f upcall, void *cookie, int mode)
445 struct ptlrpc_request *req;
446 struct osc_setattr_args *sa;
447 struct ost_body *body;
448 struct obd_import *imp = class_exp2cliimp(exp);
453 * Only mode == 0 (which is standard prealloc) is supported now.
454 * Punch is not supported yet.
456 if (mode & ~FALLOC_FL_KEEP_SIZE)
458 oa->o_falloc_mode = mode;
460 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
465 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_FALLOCATE);
467 ptlrpc_request_free(req);
471 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
474 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
476 ptlrpc_request_set_replen(req);
478 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
479 BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
480 sa = ptlrpc_req_async_args(sa, req);
482 sa->sa_upcall = upcall;
483 sa->sa_cookie = cookie;
485 ptlrpcd_add_req(req);
490 static int osc_sync_interpret(const struct lu_env *env,
491 struct ptlrpc_request *req, void *args, int rc)
493 struct osc_fsync_args *fa = args;
494 struct ost_body *body;
495 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
496 unsigned long valid = 0;
497 struct cl_object *obj;
503 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
505 CERROR("can't unpack ost_body\n");
506 GOTO(out, rc = -EPROTO);
509 *fa->fa_oa = body->oa;
510 obj = osc2cl(fa->fa_obj);
512 /* Update osc object's blocks attribute */
513 cl_object_attr_lock(obj);
514 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
515 attr->cat_blocks = body->oa.o_blocks;
520 cl_object_attr_update(env, obj, attr, valid);
521 cl_object_attr_unlock(obj);
524 rc = fa->fa_upcall(fa->fa_cookie, rc);
528 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
529 obd_enqueue_update_f upcall, void *cookie,
530 struct ptlrpc_request_set *rqset)
532 struct obd_export *exp = osc_export(obj);
533 struct ptlrpc_request *req;
534 struct ost_body *body;
535 struct osc_fsync_args *fa;
539 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
543 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
545 ptlrpc_request_free(req);
549 /* overload the size and blocks fields in the oa with start/end */
550 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
552 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
554 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = osc_sync_interpret;
557 fa = ptlrpc_req_async_args(fa, req);
560 fa->fa_upcall = upcall;
561 fa->fa_cookie = cookie;
563 ptlrpc_set_add_req(rqset, req);
568 /* Find and cancel locally locks matched by @mode in the resource found by
569 * @objid. Found locks are added into @cancel list. Returns the amount of
570 * locks added to @cancels list. */
571 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
572 struct list_head *cancels,
573 enum ldlm_mode mode, __u64 lock_flags)
575 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
576 struct ldlm_res_id res_id;
577 struct ldlm_resource *res;
581 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
582 * export) but disabled through procfs (flag in NS).
584 * This distinguishes from a case when ELC is not supported originally,
585 * when we still want to cancel locks in advance and just cancel them
586 * locally, without sending any RPC. */
587 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
590 ostid_build_res_name(&oa->o_oi, &res_id);
591 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
595 LDLM_RESOURCE_ADDREF(res);
596 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
597 lock_flags, 0, NULL);
598 LDLM_RESOURCE_DELREF(res);
599 ldlm_resource_putref(res);
603 static int osc_destroy_interpret(const struct lu_env *env,
604 struct ptlrpc_request *req, void *args, int rc)
606 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
608 atomic_dec(&cli->cl_destroy_in_flight);
609 wake_up(&cli->cl_destroy_waitq);
614 static int osc_can_send_destroy(struct client_obd *cli)
616 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
617 cli->cl_max_rpcs_in_flight) {
618 /* The destroy request can be sent */
621 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
622 cli->cl_max_rpcs_in_flight) {
624 * The counter has been modified between the two atomic
627 wake_up(&cli->cl_destroy_waitq);
632 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
635 struct client_obd *cli = &exp->exp_obd->u.cli;
636 struct ptlrpc_request *req;
637 struct ost_body *body;
643 CDEBUG(D_INFO, "oa NULL\n");
647 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
648 LDLM_FL_DISCARD_DATA);
650 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
652 ldlm_lock_list_put(&cancels, l_bl_ast, count);
656 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
659 ptlrpc_request_free(req);
663 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
664 ptlrpc_at_set_req_timeout(req);
666 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
668 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
670 ptlrpc_request_set_replen(req);
672 req->rq_interpret_reply = osc_destroy_interpret;
673 if (!osc_can_send_destroy(cli)) {
675 * Wait until the number of on-going destroy RPCs drops
676 * under max_rpc_in_flight
678 rc = l_wait_event_abortable_exclusive(
679 cli->cl_destroy_waitq,
680 osc_can_send_destroy(cli));
682 ptlrpc_req_finished(req);
687 /* Do not wait for response */
688 ptlrpcd_add_req(req);
692 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
695 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
697 LASSERT(!(oa->o_valid & bits));
700 spin_lock(&cli->cl_loi_list_lock);
701 if (cli->cl_ocd_grant_param)
702 oa->o_dirty = cli->cl_dirty_grant;
704 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
705 if (unlikely(cli->cl_dirty_pages > cli->cl_dirty_max_pages)) {
706 CERROR("dirty %lu > dirty_max %lu\n",
708 cli->cl_dirty_max_pages);
710 } else if (unlikely(atomic_long_read(&obd_dirty_pages) >
711 (long)(obd_max_dirty_pages + 1))) {
712 /* The atomic_read() allowing the atomic_inc() are
713 * not covered by a lock thus they may safely race and trip
714 * this CERROR() unless we add in a small fudge factor (+1). */
715 CERROR("%s: dirty %ld > system dirty_max %ld\n",
716 cli_name(cli), atomic_long_read(&obd_dirty_pages),
717 obd_max_dirty_pages);
719 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
721 CERROR("dirty %lu - dirty_max %lu too big???\n",
722 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
725 unsigned long nrpages;
726 unsigned long undirty;
728 nrpages = cli->cl_max_pages_per_rpc;
729 nrpages *= cli->cl_max_rpcs_in_flight + 1;
730 nrpages = max(nrpages, cli->cl_dirty_max_pages);
731 undirty = nrpages << PAGE_SHIFT;
732 if (cli->cl_ocd_grant_param) {
735 /* take extent tax into account when asking for more
737 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
738 cli->cl_max_extent_pages;
739 undirty += nrextents * cli->cl_grant_extent_tax;
741 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
742 * to add extent tax, etc.
744 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
745 ~(PTLRPC_MAX_BRW_SIZE * 4UL));
747 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
748 /* o_dropped AKA o_misc is 32 bits, but cl_lost_grant is 64 bits */
749 if (cli->cl_lost_grant > INT_MAX) {
751 "%s: avoided o_dropped overflow: cl_lost_grant %lu\n",
752 cli_name(cli), cli->cl_lost_grant);
753 oa->o_dropped = INT_MAX;
755 oa->o_dropped = cli->cl_lost_grant;
757 cli->cl_lost_grant -= oa->o_dropped;
758 spin_unlock(&cli->cl_loi_list_lock);
759 CDEBUG(D_CACHE, "%s: dirty: %llu undirty: %u dropped %u grant: %llu"
760 " cl_lost_grant %lu\n", cli_name(cli), oa->o_dirty,
761 oa->o_undirty, oa->o_dropped, oa->o_grant, cli->cl_lost_grant);
764 void osc_update_next_shrink(struct client_obd *cli)
766 cli->cl_next_shrink_grant = ktime_get_seconds() +
767 cli->cl_grant_shrink_interval;
769 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
770 cli->cl_next_shrink_grant);
773 static void __osc_update_grant(struct client_obd *cli, u64 grant)
775 spin_lock(&cli->cl_loi_list_lock);
776 cli->cl_avail_grant += grant;
777 spin_unlock(&cli->cl_loi_list_lock);
780 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
782 if (body->oa.o_valid & OBD_MD_FLGRANT) {
783 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
784 __osc_update_grant(cli, body->oa.o_grant);
789 * grant thread data for shrinking space.
791 struct grant_thread_data {
792 struct list_head gtd_clients;
793 struct mutex gtd_mutex;
794 unsigned long gtd_stopped:1;
796 static struct grant_thread_data client_gtd;
798 static int osc_shrink_grant_interpret(const struct lu_env *env,
799 struct ptlrpc_request *req,
802 struct osc_grant_args *aa = args;
803 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
804 struct ost_body *body;
807 __osc_update_grant(cli, aa->aa_oa->o_grant);
811 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
813 osc_update_grant(cli, body);
815 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
821 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
823 spin_lock(&cli->cl_loi_list_lock);
824 oa->o_grant = cli->cl_avail_grant / 4;
825 cli->cl_avail_grant -= oa->o_grant;
826 spin_unlock(&cli->cl_loi_list_lock);
827 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
828 oa->o_valid |= OBD_MD_FLFLAGS;
831 oa->o_flags |= OBD_FL_SHRINK_GRANT;
832 osc_update_next_shrink(cli);
835 /* Shrink the current grant, either from some large amount to enough for a
836 * full set of in-flight RPCs, or if we have already shrunk to that limit
837 * then to enough for a single RPC. This avoids keeping more grant than
838 * needed, and avoids shrinking the grant piecemeal. */
839 static int osc_shrink_grant(struct client_obd *cli)
841 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
842 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
844 spin_lock(&cli->cl_loi_list_lock);
845 if (cli->cl_avail_grant <= target_bytes)
846 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
847 spin_unlock(&cli->cl_loi_list_lock);
849 return osc_shrink_grant_to_target(cli, target_bytes);
852 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
855 struct ost_body *body;
858 spin_lock(&cli->cl_loi_list_lock);
859 /* Don't shrink if we are already above or below the desired limit
860 * We don't want to shrink below a single RPC, as that will negatively
861 * impact block allocation and long-term performance. */
862 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
863 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
865 if (target_bytes >= cli->cl_avail_grant) {
866 spin_unlock(&cli->cl_loi_list_lock);
869 spin_unlock(&cli->cl_loi_list_lock);
875 osc_announce_cached(cli, &body->oa, 0);
877 spin_lock(&cli->cl_loi_list_lock);
878 if (target_bytes >= cli->cl_avail_grant) {
879 /* available grant has changed since target calculation */
880 spin_unlock(&cli->cl_loi_list_lock);
881 GOTO(out_free, rc = 0);
883 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
884 cli->cl_avail_grant = target_bytes;
885 spin_unlock(&cli->cl_loi_list_lock);
886 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
887 body->oa.o_valid |= OBD_MD_FLFLAGS;
888 body->oa.o_flags = 0;
890 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
891 osc_update_next_shrink(cli);
893 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
894 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
895 sizeof(*body), body, NULL);
897 __osc_update_grant(cli, body->oa.o_grant);
903 static int osc_should_shrink_grant(struct client_obd *client)
905 time64_t next_shrink = client->cl_next_shrink_grant;
907 if (client->cl_import == NULL)
910 if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
911 client->cl_import->imp_grant_shrink_disabled) {
912 osc_update_next_shrink(client);
916 if (ktime_get_seconds() >= next_shrink - 5) {
917 /* Get the current RPC size directly, instead of going via:
918 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
919 * Keep comment here so that it can be found by searching. */
920 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
922 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
923 client->cl_avail_grant > brw_size)
926 osc_update_next_shrink(client);
931 #define GRANT_SHRINK_RPC_BATCH 100
933 static struct delayed_work work;
935 static void osc_grant_work_handler(struct work_struct *data)
937 struct client_obd *cli;
939 bool init_next_shrink = true;
940 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
943 mutex_lock(&client_gtd.gtd_mutex);
944 list_for_each_entry(cli, &client_gtd.gtd_clients,
946 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
947 osc_should_shrink_grant(cli)) {
948 osc_shrink_grant(cli);
952 if (!init_next_shrink) {
953 if (cli->cl_next_shrink_grant < next_shrink &&
954 cli->cl_next_shrink_grant > ktime_get_seconds())
955 next_shrink = cli->cl_next_shrink_grant;
957 init_next_shrink = false;
958 next_shrink = cli->cl_next_shrink_grant;
961 mutex_unlock(&client_gtd.gtd_mutex);
963 if (client_gtd.gtd_stopped == 1)
966 if (next_shrink > ktime_get_seconds()) {
967 time64_t delay = next_shrink - ktime_get_seconds();
969 schedule_delayed_work(&work, cfs_time_seconds(delay));
971 schedule_work(&work.work);
975 void osc_schedule_grant_work(void)
977 cancel_delayed_work_sync(&work);
978 schedule_work(&work.work);
982 * Start grant thread for returing grant to server for idle clients.
984 static int osc_start_grant_work(void)
986 client_gtd.gtd_stopped = 0;
987 mutex_init(&client_gtd.gtd_mutex);
988 INIT_LIST_HEAD(&client_gtd.gtd_clients);
990 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
991 schedule_work(&work.work);
996 static void osc_stop_grant_work(void)
998 client_gtd.gtd_stopped = 1;
999 cancel_delayed_work_sync(&work);
1002 static void osc_add_grant_list(struct client_obd *client)
1004 mutex_lock(&client_gtd.gtd_mutex);
1005 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
1006 mutex_unlock(&client_gtd.gtd_mutex);
1009 static void osc_del_grant_list(struct client_obd *client)
1011 if (list_empty(&client->cl_grant_chain))
1014 mutex_lock(&client_gtd.gtd_mutex);
1015 list_del_init(&client->cl_grant_chain);
1016 mutex_unlock(&client_gtd.gtd_mutex);
1019 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1022 * ocd_grant is the total grant amount we're expect to hold: if we've
1023 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
1024 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
1027 * race is tolerable here: if we're evicted, but imp_state already
1028 * left EVICTED state, then cl_dirty_pages must be 0 already.
1030 spin_lock(&cli->cl_loi_list_lock);
1031 cli->cl_avail_grant = ocd->ocd_grant;
1032 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
1033 unsigned long consumed = cli->cl_reserved_grant;
1035 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
1036 consumed += cli->cl_dirty_grant;
1038 consumed += cli->cl_dirty_pages << PAGE_SHIFT;
1039 if (cli->cl_avail_grant < consumed) {
1040 CERROR("%s: granted %ld but already consumed %ld\n",
1041 cli_name(cli), cli->cl_avail_grant, consumed);
1042 cli->cl_avail_grant = 0;
1044 cli->cl_avail_grant -= consumed;
1048 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
1052 /* overhead for each extent insertion */
1053 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
1054 /* determine the appropriate chunk size used by osc_extent. */
1055 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
1056 ocd->ocd_grant_blkbits);
1057 /* max_pages_per_rpc must be chunk aligned */
1058 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
1059 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
1060 ~chunk_mask) & chunk_mask;
1061 /* determine maximum extent size, in #pages */
1062 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
1063 cli->cl_max_extent_pages = (size >> PAGE_SHIFT) ?: 1;
1064 cli->cl_ocd_grant_param = 1;
1066 cli->cl_ocd_grant_param = 0;
1067 cli->cl_grant_extent_tax = 0;
1068 cli->cl_chunkbits = PAGE_SHIFT;
1069 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1071 spin_unlock(&cli->cl_loi_list_lock);
1074 "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld. chunk bits: %d cl_max_extent_pages: %d\n",
1076 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1077 cli->cl_max_extent_pages);
1079 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1080 osc_add_grant_list(cli);
1082 EXPORT_SYMBOL(osc_init_grant);
1084 /* We assume that the reason this OSC got a short read is because it read
1085 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1086 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1087 * this stripe never got written at or beyond this stripe offset yet. */
1088 static void handle_short_read(int nob_read, size_t page_count,
1089 struct brw_page **pga)
1094 /* skip bytes read OK */
1095 while (nob_read > 0) {
1096 LASSERT (page_count > 0);
1098 if (pga[i]->count > nob_read) {
1099 /* EOF inside this page */
1100 ptr = kmap(pga[i]->pg) +
1101 (pga[i]->off & ~PAGE_MASK);
1102 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1109 nob_read -= pga[i]->count;
1114 /* zero remaining pages */
1115 while (page_count-- > 0) {
1116 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1117 memset(ptr, 0, pga[i]->count);
1123 static int check_write_rcs(struct ptlrpc_request *req,
1124 int requested_nob, int niocount,
1125 size_t page_count, struct brw_page **pga)
1130 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1131 sizeof(*remote_rcs) *
1133 if (remote_rcs == NULL) {
1134 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1138 /* return error if any niobuf was in error */
1139 for (i = 0; i < niocount; i++) {
1140 if ((int)remote_rcs[i] < 0) {
1141 CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1142 i, remote_rcs[i], req);
1143 return remote_rcs[i];
1146 if (remote_rcs[i] != 0) {
1147 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1148 i, remote_rcs[i], req);
1152 if (req->rq_bulk != NULL &&
1153 req->rq_bulk->bd_nob_transferred != requested_nob) {
1154 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1155 req->rq_bulk->bd_nob_transferred, requested_nob);
1162 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1164 if (p1->flag != p2->flag) {
1165 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1166 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1167 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1169 /* warn if we try to combine flags that we don't know to be
1170 * safe to combine */
1171 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1172 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1173 "report this at https://jira.whamcloud.com/\n",
1174 p1->flag, p2->flag);
1179 return (p1->off + p1->count == p2->off);
1182 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1183 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1184 size_t pg_count, struct brw_page **pga,
1185 int opc, obd_dif_csum_fn *fn,
1189 struct ahash_request *req;
1190 /* Used Adler as the default checksum type on top of DIF tags */
1191 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1192 struct page *__page;
1193 unsigned char *buffer;
1195 unsigned int bufsize;
1197 int used_number = 0;
1203 LASSERT(pg_count > 0);
1205 __page = alloc_page(GFP_KERNEL);
1209 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1212 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1213 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1217 buffer = kmap(__page);
1218 guard_start = (__u16 *)buffer;
1219 guard_number = PAGE_SIZE / sizeof(*guard_start);
1220 while (nob > 0 && pg_count > 0) {
1221 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1223 /* corrupt the data before we compute the checksum, to
1224 * simulate an OST->client data error */
1225 if (unlikely(i == 0 && opc == OST_READ &&
1226 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1227 unsigned char *ptr = kmap(pga[i]->pg);
1228 int off = pga[i]->off & ~PAGE_MASK;
1230 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1235 * The left guard number should be able to hold checksums of a
1238 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1239 pga[i]->off & ~PAGE_MASK,
1241 guard_start + used_number,
1242 guard_number - used_number,
1248 used_number += used;
1249 if (used_number == guard_number) {
1250 cfs_crypto_hash_update_page(req, __page, 0,
1251 used_number * sizeof(*guard_start));
1255 nob -= pga[i]->count;
1263 if (used_number != 0)
1264 cfs_crypto_hash_update_page(req, __page, 0,
1265 used_number * sizeof(*guard_start));
1267 bufsize = sizeof(cksum);
1268 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1270 /* For sending we only compute the wrong checksum instead
1271 * of corrupting the data so it is still correct on a redo */
1272 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1277 __free_page(__page);
1280 #else /* !CONFIG_CRC_T10DIF */
1281 #define obd_dif_ip_fn NULL
1282 #define obd_dif_crc_fn NULL
1283 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum) \
1285 #endif /* CONFIG_CRC_T10DIF */
1287 static int osc_checksum_bulk(int nob, size_t pg_count,
1288 struct brw_page **pga, int opc,
1289 enum cksum_types cksum_type,
1293 struct ahash_request *req;
1294 unsigned int bufsize;
1295 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1297 LASSERT(pg_count > 0);
1299 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1301 CERROR("Unable to initialize checksum hash %s\n",
1302 cfs_crypto_hash_name(cfs_alg));
1303 return PTR_ERR(req);
1306 while (nob > 0 && pg_count > 0) {
1307 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1309 /* corrupt the data before we compute the checksum, to
1310 * simulate an OST->client data error */
1311 if (i == 0 && opc == OST_READ &&
1312 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1313 unsigned char *ptr = kmap(pga[i]->pg);
1314 int off = pga[i]->off & ~PAGE_MASK;
1316 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1319 cfs_crypto_hash_update_page(req, pga[i]->pg,
1320 pga[i]->off & ~PAGE_MASK,
1322 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1323 (int)(pga[i]->off & ~PAGE_MASK));
1325 nob -= pga[i]->count;
1330 bufsize = sizeof(*cksum);
1331 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1333 /* For sending we only compute the wrong checksum instead
1334 * of corrupting the data so it is still correct on a redo */
1335 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1341 static int osc_checksum_bulk_rw(const char *obd_name,
1342 enum cksum_types cksum_type,
1343 int nob, size_t pg_count,
1344 struct brw_page **pga, int opc,
1347 obd_dif_csum_fn *fn = NULL;
1348 int sector_size = 0;
1352 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1355 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1356 opc, fn, sector_size, check_sum);
1358 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1364 static inline void osc_release_bounce_pages(struct brw_page **pga,
1367 #ifdef HAVE_LUSTRE_CRYPTO
1370 for (i = 0; i < page_count; i++) {
1371 /* Bounce pages allocated by a call to
1372 * llcrypt_encrypt_pagecache_blocks() in osc_brw_prep_request()
1373 * are identified thanks to the PageChecked flag.
1375 if (PageChecked(pga[i]->pg))
1376 llcrypt_finalize_bounce_page(&pga[i]->pg);
1377 pga[i]->count -= pga[i]->bp_count_diff;
1378 pga[i]->off += pga[i]->bp_off_diff;
1384 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1385 u32 page_count, struct brw_page **pga,
1386 struct ptlrpc_request **reqp, int resend)
1388 struct ptlrpc_request *req;
1389 struct ptlrpc_bulk_desc *desc;
1390 struct ost_body *body;
1391 struct obd_ioobj *ioobj;
1392 struct niobuf_remote *niobuf;
1393 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1394 struct osc_brw_async_args *aa;
1395 struct req_capsule *pill;
1396 struct brw_page *pg_prev;
1398 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1399 struct inode *inode;
1400 bool directio = false;
1403 inode = page2inode(pga[0]->pg);
1404 if (inode == NULL) {
1405 /* Try to get reference to inode from cl_page if we are
1406 * dealing with direct IO, as handled pages are not
1407 * actual page cache pages.
1409 struct osc_async_page *oap = brw_page2oap(pga[0]);
1410 struct cl_page *clpage = oap2cl_page(oap);
1412 inode = clpage->cp_inode;
1416 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1417 RETURN(-ENOMEM); /* Recoverable */
1418 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1419 RETURN(-EINVAL); /* Fatal */
1421 if ((cmd & OBD_BRW_WRITE) != 0) {
1423 req = ptlrpc_request_alloc_pool(cli->cl_import,
1425 &RQF_OST_BRW_WRITE);
1428 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1433 if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1434 for (i = 0; i < page_count; i++) {
1435 struct brw_page *pg = pga[i];
1436 struct page *data_page = NULL;
1437 bool retried = false;
1438 bool lockedbymyself;
1439 u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1440 struct address_space *map_orig = NULL;
1444 if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1445 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1446 LUSTRE_ENCRYPTION_UNIT_SIZE;
1447 /* The page can already be locked when we arrive here.
1448 * This is possible when cl_page_assume/vvp_page_assume
1449 * is stuck on wait_on_page_writeback with page lock
1450 * held. In this case there is no risk for the lock to
1451 * be released while we are doing our encryption
1452 * processing, because writeback against that page will
1453 * end in vvp_page_completion_write/cl_page_completion,
1454 * which means only once the page is fully processed.
1456 lockedbymyself = trylock_page(pg->pg);
1458 map_orig = pg->pg->mapping;
1459 pg->pg->mapping = inode->i_mapping;
1460 index_orig = pg->pg->index;
1461 pg->pg->index = pg->off >> PAGE_SHIFT;
1464 llcrypt_encrypt_pagecache_blocks(pg->pg,
1468 pg->pg->mapping = map_orig;
1469 pg->pg->index = index_orig;
1472 unlock_page(pg->pg);
1473 if (IS_ERR(data_page)) {
1474 rc = PTR_ERR(data_page);
1475 if (rc == -ENOMEM && !retried) {
1480 ptlrpc_request_free(req);
1483 /* Set PageChecked flag on bounce page for
1484 * disambiguation in osc_release_bounce_pages().
1486 SetPageChecked(data_page);
1488 /* there should be no gap in the middle of page array */
1489 if (i == page_count - 1) {
1490 struct osc_async_page *oap = brw_page2oap(pg);
1492 oa->o_size = oap->oap_count +
1493 oap->oap_obj_off + oap->oap_page_off;
1495 /* len is forced to nunits, and relative offset to 0
1496 * so store the old, clear text info
1498 pg->bp_count_diff = nunits - pg->count;
1500 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1501 pg->off = pg->off & PAGE_MASK;
1503 } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode)) {
1504 for (i = 0; i < page_count; i++) {
1505 struct brw_page *pg = pga[i];
1506 u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1508 if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1509 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1510 LUSTRE_ENCRYPTION_UNIT_SIZE;
1511 /* count/off are forced to cover the whole encryption
1512 * unit size so that all encrypted data is stored on the
1513 * OST, so adjust bp_{count,off}_diff for the size of
1516 pg->bp_count_diff = nunits - pg->count;
1518 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1519 pg->off = pg->off & PAGE_MASK;
1523 for (niocount = i = 1; i < page_count; i++) {
1524 if (!can_merge_pages(pga[i - 1], pga[i]))
1528 pill = &req->rq_pill;
1529 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1531 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1532 niocount * sizeof(*niobuf));
1534 for (i = 0; i < page_count; i++) {
1535 short_io_size += pga[i]->count;
1536 if (!inode || !IS_ENCRYPTED(inode)) {
1537 pga[i]->bp_count_diff = 0;
1538 pga[i]->bp_off_diff = 0;
1542 /* Check if read/write is small enough to be a short io. */
1543 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1544 !imp_connect_shortio(cli->cl_import))
1547 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1548 opc == OST_READ ? 0 : short_io_size);
1549 if (opc == OST_READ)
1550 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1553 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1555 ptlrpc_request_free(req);
1558 osc_set_io_portal(req);
1560 ptlrpc_at_set_req_timeout(req);
1561 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1563 req->rq_no_retry_einprogress = 1;
1565 if (short_io_size != 0) {
1567 short_io_buf = NULL;
1571 desc = ptlrpc_prep_bulk_imp(req, page_count,
1572 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1573 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1574 PTLRPC_BULK_PUT_SINK),
1576 &ptlrpc_bulk_kiov_pin_ops);
1579 GOTO(out, rc = -ENOMEM);
1580 /* NB request now owns desc and will free it when it gets freed */
1582 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1583 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1584 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1585 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1587 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1589 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1590 * and from_kgid(), because they are asynchronous. Fortunately, variable
1591 * oa contains valid o_uid and o_gid in these two operations.
1592 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1593 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1594 * other process logic */
1595 body->oa.o_uid = oa->o_uid;
1596 body->oa.o_gid = oa->o_gid;
1598 obdo_to_ioobj(oa, ioobj);
1599 ioobj->ioo_bufcnt = niocount;
1600 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1601 * that might be send for this request. The actual number is decided
1602 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1603 * "max - 1" for old client compatibility sending "0", and also so the
1604 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1606 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1608 ioobj_max_brw_set(ioobj, 0);
1610 if (short_io_size != 0) {
1611 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1612 body->oa.o_valid |= OBD_MD_FLFLAGS;
1613 body->oa.o_flags = 0;
1615 body->oa.o_flags |= OBD_FL_SHORT_IO;
1616 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1618 if (opc == OST_WRITE) {
1619 short_io_buf = req_capsule_client_get(pill,
1621 LASSERT(short_io_buf != NULL);
1625 LASSERT(page_count > 0);
1627 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1628 struct brw_page *pg = pga[i];
1629 int poff = pg->off & ~PAGE_MASK;
1631 LASSERT(pg->count > 0);
1632 /* make sure there is no gap in the middle of page array */
1633 LASSERTF(page_count == 1 ||
1634 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1635 ergo(i > 0 && i < page_count - 1,
1636 poff == 0 && pg->count == PAGE_SIZE) &&
1637 ergo(i == page_count - 1, poff == 0)),
1638 "i: %d/%d pg: %p off: %llu, count: %u\n",
1639 i, page_count, pg, pg->off, pg->count);
1640 LASSERTF(i == 0 || pg->off > pg_prev->off,
1641 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1642 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1644 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1645 pg_prev->pg, page_private(pg_prev->pg),
1646 pg_prev->pg->index, pg_prev->off);
1647 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1648 (pg->flag & OBD_BRW_SRVLOCK));
1649 if (short_io_size != 0 && opc == OST_WRITE) {
1650 unsigned char *ptr = kmap_atomic(pg->pg);
1652 LASSERT(short_io_size >= requested_nob + pg->count);
1653 memcpy(short_io_buf + requested_nob,
1657 } else if (short_io_size == 0) {
1658 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1661 requested_nob += pg->count;
1663 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1665 niobuf->rnb_len += pg->count;
1667 niobuf->rnb_offset = pg->off;
1668 niobuf->rnb_len = pg->count;
1669 niobuf->rnb_flags = pg->flag;
1674 LASSERTF((void *)(niobuf - niocount) ==
1675 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1676 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1677 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1679 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1681 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1682 body->oa.o_valid |= OBD_MD_FLFLAGS;
1683 body->oa.o_flags = 0;
1685 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1688 if (osc_should_shrink_grant(cli))
1689 osc_shrink_grant_local(cli, &body->oa);
1691 /* size[REQ_REC_OFF] still sizeof (*body) */
1692 if (opc == OST_WRITE) {
1693 if (cli->cl_checksum &&
1694 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1695 /* store cl_cksum_type in a local variable since
1696 * it can be changed via lprocfs */
1697 enum cksum_types cksum_type = cli->cl_cksum_type;
1699 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1700 body->oa.o_flags = 0;
1702 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1704 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1706 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1707 requested_nob, page_count,
1711 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1715 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1718 /* save this in 'oa', too, for later checking */
1719 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1720 oa->o_flags |= obd_cksum_type_pack(obd_name,
1723 /* clear out the checksum flag, in case this is a
1724 * resend but cl_checksum is no longer set. b=11238 */
1725 oa->o_valid &= ~OBD_MD_FLCKSUM;
1727 oa->o_cksum = body->oa.o_cksum;
1728 /* 1 RC per niobuf */
1729 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1730 sizeof(__u32) * niocount);
1732 if (cli->cl_checksum &&
1733 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1734 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1735 body->oa.o_flags = 0;
1736 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1737 cli->cl_cksum_type);
1738 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1741 /* Client cksum has been already copied to wire obdo in previous
1742 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1743 * resent due to cksum error, this will allow Server to
1744 * check+dump pages on its side */
1746 ptlrpc_request_set_replen(req);
1748 aa = ptlrpc_req_async_args(aa, req);
1750 aa->aa_requested_nob = requested_nob;
1751 aa->aa_nio_count = niocount;
1752 aa->aa_page_count = page_count;
1756 INIT_LIST_HEAD(&aa->aa_oaps);
1759 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1760 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1761 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1762 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1766 ptlrpc_req_finished(req);
1770 char dbgcksum_file_name[PATH_MAX];
1772 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1773 struct brw_page **pga, __u32 server_cksum,
1781 /* will only keep dump of pages on first error for the same range in
1782 * file/fid, not during the resends/retries. */
1783 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1784 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1785 (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1786 libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1787 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1788 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1789 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1791 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1792 client_cksum, server_cksum);
1793 filp = filp_open(dbgcksum_file_name,
1794 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1798 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1799 "checksum error: rc = %d\n", dbgcksum_file_name,
1802 CERROR("%s: can't open to dump pages with checksum "
1803 "error: rc = %d\n", dbgcksum_file_name, rc);
1807 for (i = 0; i < page_count; i++) {
1808 len = pga[i]->count;
1809 buf = kmap(pga[i]->pg);
1811 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1813 CERROR("%s: wanted to write %u but got %d "
1814 "error\n", dbgcksum_file_name, len, rc);
1819 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1820 dbgcksum_file_name, rc);
1825 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1827 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1828 filp_close(filp, NULL);
1832 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1833 __u32 client_cksum, __u32 server_cksum,
1834 struct osc_brw_async_args *aa)
1836 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1837 enum cksum_types cksum_type;
1838 obd_dif_csum_fn *fn = NULL;
1839 int sector_size = 0;
1844 if (server_cksum == client_cksum) {
1845 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1849 if (aa->aa_cli->cl_checksum_dump)
1850 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1851 server_cksum, client_cksum);
1853 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1856 switch (cksum_type) {
1857 case OBD_CKSUM_T10IP512:
1861 case OBD_CKSUM_T10IP4K:
1865 case OBD_CKSUM_T10CRC512:
1866 fn = obd_dif_crc_fn;
1869 case OBD_CKSUM_T10CRC4K:
1870 fn = obd_dif_crc_fn;
1878 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1879 aa->aa_page_count, aa->aa_ppga,
1880 OST_WRITE, fn, sector_size,
1883 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1884 aa->aa_ppga, OST_WRITE, cksum_type,
1888 msg = "failed to calculate the client write checksum";
1889 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1890 msg = "the server did not use the checksum type specified in "
1891 "the original request - likely a protocol problem";
1892 else if (new_cksum == server_cksum)
1893 msg = "changed on the client after we checksummed it - "
1894 "likely false positive due to mmap IO (bug 11742)";
1895 else if (new_cksum == client_cksum)
1896 msg = "changed in transit before arrival at OST";
1898 msg = "changed in transit AND doesn't match the original - "
1899 "likely false positive due to mmap IO (bug 11742)";
1901 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1902 DFID " object "DOSTID" extent [%llu-%llu], original "
1903 "client csum %x (type %x), server csum %x (type %x),"
1904 " client csum now %x\n",
1905 obd_name, msg, libcfs_nid2str(peer->nid),
1906 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1907 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1908 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1909 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1910 aa->aa_ppga[aa->aa_page_count - 1]->off +
1911 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1913 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1914 server_cksum, cksum_type, new_cksum);
1918 /* Note rc enters this function as number of bytes transferred */
1919 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1921 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1922 struct client_obd *cli = aa->aa_cli;
1923 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1924 const struct lnet_process_id *peer =
1925 &req->rq_import->imp_connection->c_peer;
1926 struct ost_body *body;
1927 u32 client_cksum = 0;
1928 struct inode *inode;
1929 unsigned int blockbits = 0, blocksize = 0;
1933 if (rc < 0 && rc != -EDQUOT) {
1934 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1938 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1939 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1941 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1945 /* set/clear over quota flag for a uid/gid/projid */
1946 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1947 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1948 unsigned qid[LL_MAXQUOTAS] = {
1949 body->oa.o_uid, body->oa.o_gid,
1950 body->oa.o_projid };
1952 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1953 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1954 body->oa.o_valid, body->oa.o_flags);
1955 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1959 osc_update_grant(cli, body);
1964 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1965 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1967 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1969 CERROR("%s: unexpected positive size %d\n",
1974 if (req->rq_bulk != NULL &&
1975 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1978 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1979 check_write_checksum(&body->oa, peer, client_cksum,
1980 body->oa.o_cksum, aa))
1983 rc = check_write_rcs(req, aa->aa_requested_nob,
1984 aa->aa_nio_count, aa->aa_page_count,
1989 /* The rest of this function executes only for OST_READs */
1991 if (req->rq_bulk == NULL) {
1992 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1994 LASSERT(rc == req->rq_status);
1996 /* if unwrap_bulk failed, return -EAGAIN to retry */
1997 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
2000 GOTO(out, rc = -EAGAIN);
2002 if (rc > aa->aa_requested_nob) {
2003 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
2004 rc, aa->aa_requested_nob);
2008 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
2009 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
2010 rc, req->rq_bulk->bd_nob_transferred);
2014 if (req->rq_bulk == NULL) {
2016 int nob, pg_count, i = 0;
2019 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
2020 pg_count = aa->aa_page_count;
2021 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
2024 while (nob > 0 && pg_count > 0) {
2026 int count = aa->aa_ppga[i]->count > nob ?
2027 nob : aa->aa_ppga[i]->count;
2029 CDEBUG(D_CACHE, "page %p count %d\n",
2030 aa->aa_ppga[i]->pg, count);
2031 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
2032 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
2034 kunmap_atomic((void *) ptr);
2043 if (rc < aa->aa_requested_nob)
2044 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
2046 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2047 static int cksum_counter;
2048 u32 server_cksum = body->oa.o_cksum;
2051 enum cksum_types cksum_type;
2052 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
2053 body->oa.o_flags : 0;
2055 cksum_type = obd_cksum_type_unpack(o_flags);
2056 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
2057 aa->aa_page_count, aa->aa_ppga,
2058 OST_READ, &client_cksum);
2062 if (req->rq_bulk != NULL &&
2063 peer->nid != req->rq_bulk->bd_sender) {
2065 router = libcfs_nid2str(req->rq_bulk->bd_sender);
2068 if (server_cksum != client_cksum) {
2069 struct ost_body *clbody;
2070 u32 page_count = aa->aa_page_count;
2072 clbody = req_capsule_client_get(&req->rq_pill,
2074 if (cli->cl_checksum_dump)
2075 dump_all_bulk_pages(&clbody->oa, page_count,
2076 aa->aa_ppga, server_cksum,
2079 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2080 "%s%s%s inode "DFID" object "DOSTID
2081 " extent [%llu-%llu], client %x, "
2082 "server %x, cksum_type %x\n",
2084 libcfs_nid2str(peer->nid),
2086 clbody->oa.o_valid & OBD_MD_FLFID ?
2087 clbody->oa.o_parent_seq : 0ULL,
2088 clbody->oa.o_valid & OBD_MD_FLFID ?
2089 clbody->oa.o_parent_oid : 0,
2090 clbody->oa.o_valid & OBD_MD_FLFID ?
2091 clbody->oa.o_parent_ver : 0,
2092 POSTID(&body->oa.o_oi),
2093 aa->aa_ppga[0]->off,
2094 aa->aa_ppga[page_count-1]->off +
2095 aa->aa_ppga[page_count-1]->count - 1,
2096 client_cksum, server_cksum,
2099 aa->aa_oa->o_cksum = client_cksum;
2103 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2106 } else if (unlikely(client_cksum)) {
2107 static int cksum_missed;
2110 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2111 CERROR("%s: checksum %u requested from %s but not sent\n",
2112 obd_name, cksum_missed,
2113 libcfs_nid2str(peer->nid));
2118 inode = page2inode(aa->aa_ppga[0]->pg);
2119 if (inode == NULL) {
2120 /* Try to get reference to inode from cl_page if we are
2121 * dealing with direct IO, as handled pages are not
2122 * actual page cache pages.
2124 struct osc_async_page *oap = brw_page2oap(aa->aa_ppga[0]);
2126 inode = oap2cl_page(oap)->cp_inode;
2128 blockbits = inode->i_blkbits;
2129 blocksize = 1 << blockbits;
2132 if (inode && IS_ENCRYPTED(inode)) {
2135 if (!llcrypt_has_encryption_key(inode)) {
2136 CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2139 for (idx = 0; idx < aa->aa_page_count; idx++) {
2140 struct brw_page *pg = aa->aa_ppga[idx];
2141 unsigned int offs = 0;
2143 while (offs < PAGE_SIZE) {
2144 /* do not decrypt if page is all 0s */
2145 if (memchr_inv(page_address(pg->pg) + offs, 0,
2146 LUSTRE_ENCRYPTION_UNIT_SIZE) == NULL) {
2147 /* if page is empty forward info to
2148 * upper layers (ll_io_zero_page) by
2149 * clearing PagePrivate2
2152 ClearPagePrivate2(pg->pg);
2157 /* This is direct IO case. Directly call
2158 * decrypt function that takes inode as
2159 * input parameter. Page does not need
2163 ((u64)(pg->off >> PAGE_SHIFT) <<
2164 (PAGE_SHIFT - blockbits)) +
2165 (offs >> blockbits);
2170 LUSTRE_ENCRYPTION_UNIT_SIZE;
2171 i += blocksize, lblk_num++) {
2173 llcrypt_decrypt_block_inplace(
2181 rc = llcrypt_decrypt_pagecache_blocks(
2183 LUSTRE_ENCRYPTION_UNIT_SIZE,
2189 offs += LUSTRE_ENCRYPTION_UNIT_SIZE;
2196 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2197 aa->aa_oa, &body->oa);
2202 static int osc_brw_redo_request(struct ptlrpc_request *request,
2203 struct osc_brw_async_args *aa, int rc)
2205 struct ptlrpc_request *new_req;
2206 struct osc_brw_async_args *new_aa;
2207 struct osc_async_page *oap;
2210 /* The below message is checked in replay-ost-single.sh test_8ae*/
2211 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2212 "redo for recoverable error %d", rc);
2214 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2215 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2216 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2217 aa->aa_ppga, &new_req, 1);
2221 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2222 if (oap->oap_request != NULL) {
2223 LASSERTF(request == oap->oap_request,
2224 "request %p != oap_request %p\n",
2225 request, oap->oap_request);
2229 * New request takes over pga and oaps from old request.
2230 * Note that copying a list_head doesn't work, need to move it...
2233 new_req->rq_interpret_reply = request->rq_interpret_reply;
2234 new_req->rq_async_args = request->rq_async_args;
2235 new_req->rq_commit_cb = request->rq_commit_cb;
2236 /* cap resend delay to the current request timeout, this is similar to
2237 * what ptlrpc does (see after_reply()) */
2238 if (aa->aa_resends > new_req->rq_timeout)
2239 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2241 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2242 new_req->rq_generation_set = 1;
2243 new_req->rq_import_generation = request->rq_import_generation;
2245 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2247 INIT_LIST_HEAD(&new_aa->aa_oaps);
2248 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2249 INIT_LIST_HEAD(&new_aa->aa_exts);
2250 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2251 new_aa->aa_resends = aa->aa_resends;
2253 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2254 if (oap->oap_request) {
2255 ptlrpc_req_finished(oap->oap_request);
2256 oap->oap_request = ptlrpc_request_addref(new_req);
2260 /* XXX: This code will run into problem if we're going to support
2261 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2262 * and wait for all of them to be finished. We should inherit request
2263 * set from old request. */
2264 ptlrpcd_add_req(new_req);
2266 DEBUG_REQ(D_INFO, new_req, "new request");
2271 * ugh, we want disk allocation on the target to happen in offset order. we'll
2272 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2273 * fine for our small page arrays and doesn't require allocation. its an
2274 * insertion sort that swaps elements that are strides apart, shrinking the
2275 * stride down until its '1' and the array is sorted.
2277 static void sort_brw_pages(struct brw_page **array, int num)
2280 struct brw_page *tmp;
2284 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2289 for (i = stride ; i < num ; i++) {
2292 while (j >= stride && array[j - stride]->off > tmp->off) {
2293 array[j] = array[j - stride];
2298 } while (stride > 1);
2301 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2303 LASSERT(ppga != NULL);
2304 OBD_FREE_PTR_ARRAY_LARGE(ppga, count);
2307 static int brw_interpret(const struct lu_env *env,
2308 struct ptlrpc_request *req, void *args, int rc)
2310 struct osc_brw_async_args *aa = args;
2311 struct osc_extent *ext;
2312 struct osc_extent *tmp;
2313 struct client_obd *cli = aa->aa_cli;
2314 unsigned long transferred = 0;
2318 rc = osc_brw_fini_request(req, rc);
2319 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2321 /* restore clear text pages */
2322 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2325 * When server returns -EINPROGRESS, client should always retry
2326 * regardless of the number of times the bulk was resent already.
2328 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2329 if (req->rq_import_generation !=
2330 req->rq_import->imp_generation) {
2331 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2332 ""DOSTID", rc = %d.\n",
2333 req->rq_import->imp_obd->obd_name,
2334 POSTID(&aa->aa_oa->o_oi), rc);
2335 } else if (rc == -EINPROGRESS ||
2336 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2337 rc = osc_brw_redo_request(req, aa, rc);
2339 CERROR("%s: too many resent retries for object: "
2340 "%llu:%llu, rc = %d.\n",
2341 req->rq_import->imp_obd->obd_name,
2342 POSTID(&aa->aa_oa->o_oi), rc);
2347 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2352 struct obdo *oa = aa->aa_oa;
2353 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2354 unsigned long valid = 0;
2355 struct cl_object *obj;
2356 struct osc_async_page *last;
2358 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2359 obj = osc2cl(last->oap_obj);
2361 cl_object_attr_lock(obj);
2362 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2363 attr->cat_blocks = oa->o_blocks;
2364 valid |= CAT_BLOCKS;
2366 if (oa->o_valid & OBD_MD_FLMTIME) {
2367 attr->cat_mtime = oa->o_mtime;
2370 if (oa->o_valid & OBD_MD_FLATIME) {
2371 attr->cat_atime = oa->o_atime;
2374 if (oa->o_valid & OBD_MD_FLCTIME) {
2375 attr->cat_ctime = oa->o_ctime;
2379 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2380 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2381 loff_t last_off = last->oap_count + last->oap_obj_off +
2384 /* Change file size if this is an out of quota or
2385 * direct IO write and it extends the file size */
2386 if (loi->loi_lvb.lvb_size < last_off) {
2387 attr->cat_size = last_off;
2390 /* Extend KMS if it's not a lockless write */
2391 if (loi->loi_kms < last_off &&
2392 oap2osc_page(last)->ops_srvlock == 0) {
2393 attr->cat_kms = last_off;
2399 cl_object_attr_update(env, obj, attr, valid);
2400 cl_object_attr_unlock(obj);
2402 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2405 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2406 osc_inc_unstable_pages(req);
2408 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2409 list_del_init(&ext->oe_link);
2410 osc_extent_finish(env, ext, 1,
2411 rc && req->rq_no_delay ? -EAGAIN : rc);
2413 LASSERT(list_empty(&aa->aa_exts));
2414 LASSERT(list_empty(&aa->aa_oaps));
2416 transferred = (req->rq_bulk == NULL ? /* short io */
2417 aa->aa_requested_nob :
2418 req->rq_bulk->bd_nob_transferred);
2420 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2421 ptlrpc_lprocfs_brw(req, transferred);
2423 spin_lock(&cli->cl_loi_list_lock);
2424 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2425 * is called so we know whether to go to sync BRWs or wait for more
2426 * RPCs to complete */
2427 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2428 cli->cl_w_in_flight--;
2430 cli->cl_r_in_flight--;
2431 osc_wake_cache_waiters(cli);
2432 spin_unlock(&cli->cl_loi_list_lock);
2434 osc_io_unplug(env, cli, NULL);
2438 static void brw_commit(struct ptlrpc_request *req)
2440 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2441 * this called via the rq_commit_cb, I need to ensure
2442 * osc_dec_unstable_pages is still called. Otherwise unstable
2443 * pages may be leaked. */
2444 spin_lock(&req->rq_lock);
2445 if (likely(req->rq_unstable)) {
2446 req->rq_unstable = 0;
2447 spin_unlock(&req->rq_lock);
2449 osc_dec_unstable_pages(req);
2451 req->rq_committed = 1;
2452 spin_unlock(&req->rq_lock);
2457 * Build an RPC by the list of extent @ext_list. The caller must ensure
2458 * that the total pages in this list are NOT over max pages per RPC.
2459 * Extents in the list must be in OES_RPC state.
2461 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2462 struct list_head *ext_list, int cmd)
2464 struct ptlrpc_request *req = NULL;
2465 struct osc_extent *ext;
2466 struct brw_page **pga = NULL;
2467 struct osc_brw_async_args *aa = NULL;
2468 struct obdo *oa = NULL;
2469 struct osc_async_page *oap;
2470 struct osc_object *obj = NULL;
2471 struct cl_req_attr *crattr = NULL;
2472 loff_t starting_offset = OBD_OBJECT_EOF;
2473 loff_t ending_offset = 0;
2474 /* '1' for consistency with code that checks !mpflag to restore */
2478 bool soft_sync = false;
2479 bool ndelay = false;
2483 __u32 layout_version = 0;
2484 LIST_HEAD(rpc_list);
2485 struct ost_body *body;
2487 LASSERT(!list_empty(ext_list));
2489 /* add pages into rpc_list to build BRW rpc */
2490 list_for_each_entry(ext, ext_list, oe_link) {
2491 LASSERT(ext->oe_state == OES_RPC);
2492 mem_tight |= ext->oe_memalloc;
2493 grant += ext->oe_grants;
2494 page_count += ext->oe_nr_pages;
2495 layout_version = max(layout_version, ext->oe_layout_version);
2500 soft_sync = osc_over_unstable_soft_limit(cli);
2502 mpflag = memalloc_noreclaim_save();
2504 OBD_ALLOC_PTR_ARRAY_LARGE(pga, page_count);
2506 GOTO(out, rc = -ENOMEM);
2508 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2510 GOTO(out, rc = -ENOMEM);
2513 list_for_each_entry(ext, ext_list, oe_link) {
2514 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2516 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2518 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2519 pga[i] = &oap->oap_brw_page;
2520 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2523 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2524 if (starting_offset == OBD_OBJECT_EOF ||
2525 starting_offset > oap->oap_obj_off)
2526 starting_offset = oap->oap_obj_off;
2528 LASSERT(oap->oap_page_off == 0);
2529 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2530 ending_offset = oap->oap_obj_off +
2533 LASSERT(oap->oap_page_off + oap->oap_count ==
2540 /* first page in the list */
2541 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2543 crattr = &osc_env_info(env)->oti_req_attr;
2544 memset(crattr, 0, sizeof(*crattr));
2545 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2546 crattr->cra_flags = ~0ULL;
2547 crattr->cra_page = oap2cl_page(oap);
2548 crattr->cra_oa = oa;
2549 cl_req_attr_set(env, osc2cl(obj), crattr);
2551 if (cmd == OBD_BRW_WRITE) {
2552 oa->o_grant_used = grant;
2553 if (layout_version > 0) {
2554 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2555 PFID(&oa->o_oi.oi_fid), layout_version);
2557 oa->o_layout_version = layout_version;
2558 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2562 sort_brw_pages(pga, page_count);
2563 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2565 CERROR("prep_req failed: %d\n", rc);
2569 req->rq_commit_cb = brw_commit;
2570 req->rq_interpret_reply = brw_interpret;
2571 req->rq_memalloc = mem_tight != 0;
2572 oap->oap_request = ptlrpc_request_addref(req);
2574 req->rq_no_resend = req->rq_no_delay = 1;
2575 /* probably set a shorter timeout value.
2576 * to handle ETIMEDOUT in brw_interpret() correctly. */
2577 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2580 /* Need to update the timestamps after the request is built in case
2581 * we race with setattr (locally or in queue at OST). If OST gets
2582 * later setattr before earlier BRW (as determined by the request xid),
2583 * the OST will not use BRW timestamps. Sadly, there is no obvious
2584 * way to do this in a single call. bug 10150 */
2585 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2586 crattr->cra_oa = &body->oa;
2587 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2588 cl_req_attr_set(env, osc2cl(obj), crattr);
2589 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2591 aa = ptlrpc_req_async_args(aa, req);
2592 INIT_LIST_HEAD(&aa->aa_oaps);
2593 list_splice_init(&rpc_list, &aa->aa_oaps);
2594 INIT_LIST_HEAD(&aa->aa_exts);
2595 list_splice_init(ext_list, &aa->aa_exts);
2597 spin_lock(&cli->cl_loi_list_lock);
2598 starting_offset >>= PAGE_SHIFT;
2599 if (cmd == OBD_BRW_READ) {
2600 cli->cl_r_in_flight++;
2601 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2602 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2603 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2604 starting_offset + 1);
2606 cli->cl_w_in_flight++;
2607 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2608 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2609 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2610 starting_offset + 1);
2612 spin_unlock(&cli->cl_loi_list_lock);
2614 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2615 page_count, aa, cli->cl_r_in_flight,
2616 cli->cl_w_in_flight);
2617 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2619 ptlrpcd_add_req(req);
2625 memalloc_noreclaim_restore(mpflag);
2628 LASSERT(req == NULL);
2631 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2633 osc_release_bounce_pages(pga, page_count);
2634 osc_release_ppga(pga, page_count);
2636 /* this should happen rarely and is pretty bad, it makes the
2637 * pending list not follow the dirty order */
2638 while (!list_empty(ext_list)) {
2639 ext = list_entry(ext_list->next, struct osc_extent,
2641 list_del_init(&ext->oe_link);
2642 osc_extent_finish(env, ext, 0, rc);
2648 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2652 LASSERT(lock != NULL);
2654 lock_res_and_lock(lock);
2656 if (lock->l_ast_data == NULL)
2657 lock->l_ast_data = data;
2658 if (lock->l_ast_data == data)
2661 unlock_res_and_lock(lock);
2666 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2667 void *cookie, struct lustre_handle *lockh,
2668 enum ldlm_mode mode, __u64 *flags, bool speculative,
2671 bool intent = *flags & LDLM_FL_HAS_INTENT;
2675 /* The request was created before ldlm_cli_enqueue call. */
2676 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2677 struct ldlm_reply *rep;
2679 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2680 LASSERT(rep != NULL);
2682 rep->lock_policy_res1 =
2683 ptlrpc_status_ntoh(rep->lock_policy_res1);
2684 if (rep->lock_policy_res1)
2685 errcode = rep->lock_policy_res1;
2687 *flags |= LDLM_FL_LVB_READY;
2688 } else if (errcode == ELDLM_OK) {
2689 *flags |= LDLM_FL_LVB_READY;
2692 /* Call the update callback. */
2693 rc = (*upcall)(cookie, lockh, errcode);
2695 /* release the reference taken in ldlm_cli_enqueue() */
2696 if (errcode == ELDLM_LOCK_MATCHED)
2698 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2699 ldlm_lock_decref(lockh, mode);
2704 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2707 struct osc_enqueue_args *aa = args;
2708 struct ldlm_lock *lock;
2709 struct lustre_handle *lockh = &aa->oa_lockh;
2710 enum ldlm_mode mode = aa->oa_mode;
2711 struct ost_lvb *lvb = aa->oa_lvb;
2712 __u32 lvb_len = sizeof(*lvb);
2714 struct ldlm_enqueue_info einfo = {
2715 .ei_type = aa->oa_type,
2721 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2723 lock = ldlm_handle2lock(lockh);
2724 LASSERTF(lock != NULL,
2725 "lockh %#llx, req %p, aa %p - client evicted?\n",
2726 lockh->cookie, req, aa);
2728 /* Take an additional reference so that a blocking AST that
2729 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2730 * to arrive after an upcall has been executed by
2731 * osc_enqueue_fini(). */
2732 ldlm_lock_addref(lockh, mode);
2734 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2735 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2737 /* Let CP AST to grant the lock first. */
2738 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2740 if (aa->oa_speculative) {
2741 LASSERT(aa->oa_lvb == NULL);
2742 LASSERT(aa->oa_flags == NULL);
2743 aa->oa_flags = &flags;
2746 /* Complete obtaining the lock procedure. */
2747 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, &einfo, 1, aa->oa_flags,
2748 lvb, lvb_len, lockh, rc);
2749 /* Complete osc stuff. */
2750 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2751 aa->oa_flags, aa->oa_speculative, rc);
2753 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2755 ldlm_lock_decref(lockh, mode);
2756 LDLM_LOCK_PUT(lock);
2760 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2761 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2762 * other synchronous requests, however keeping some locks and trying to obtain
2763 * others may take a considerable amount of time in a case of ost failure; and
2764 * when other sync requests do not get released lock from a client, the client
2765 * is evicted from the cluster -- such scenarious make the life difficult, so
2766 * release locks just after they are obtained. */
2767 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2768 __u64 *flags, union ldlm_policy_data *policy,
2769 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2770 void *cookie, struct ldlm_enqueue_info *einfo,
2771 struct ptlrpc_request_set *rqset, int async,
2774 struct obd_device *obd = exp->exp_obd;
2775 struct lustre_handle lockh = { 0 };
2776 struct ptlrpc_request *req = NULL;
2777 int intent = *flags & LDLM_FL_HAS_INTENT;
2778 __u64 match_flags = *flags;
2779 enum ldlm_mode mode;
2783 /* Filesystem lock extents are extended to page boundaries so that
2784 * dealing with the page cache is a little smoother. */
2785 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2786 policy->l_extent.end |= ~PAGE_MASK;
2788 /* Next, search for already existing extent locks that will cover us */
2789 /* If we're trying to read, we also search for an existing PW lock. The
2790 * VFS and page cache already protect us locally, so lots of readers/
2791 * writers can share a single PW lock.
2793 * There are problems with conversion deadlocks, so instead of
2794 * converting a read lock to a write lock, we'll just enqueue a new
2797 * At some point we should cancel the read lock instead of making them
2798 * send us a blocking callback, but there are problems with canceling
2799 * locks out from other users right now, too. */
2800 mode = einfo->ei_mode;
2801 if (einfo->ei_mode == LCK_PR)
2803 /* Normal lock requests must wait for the LVB to be ready before
2804 * matching a lock; speculative lock requests do not need to,
2805 * because they will not actually use the lock. */
2807 match_flags |= LDLM_FL_LVB_READY;
2809 match_flags |= LDLM_FL_BLOCK_GRANTED;
2810 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2811 einfo->ei_type, policy, mode, &lockh);
2813 struct ldlm_lock *matched;
2815 if (*flags & LDLM_FL_TEST_LOCK)
2818 matched = ldlm_handle2lock(&lockh);
2820 /* This DLM lock request is speculative, and does not
2821 * have an associated IO request. Therefore if there
2822 * is already a DLM lock, it wll just inform the
2823 * caller to cancel the request for this stripe.*/
2824 lock_res_and_lock(matched);
2825 if (ldlm_extent_equal(&policy->l_extent,
2826 &matched->l_policy_data.l_extent))
2830 unlock_res_and_lock(matched);
2832 ldlm_lock_decref(&lockh, mode);
2833 LDLM_LOCK_PUT(matched);
2835 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2836 *flags |= LDLM_FL_LVB_READY;
2838 /* We already have a lock, and it's referenced. */
2839 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2841 ldlm_lock_decref(&lockh, mode);
2842 LDLM_LOCK_PUT(matched);
2845 ldlm_lock_decref(&lockh, mode);
2846 LDLM_LOCK_PUT(matched);
2850 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2853 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2854 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2856 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2857 sizeof(*lvb), LVB_T_OST, &lockh, async);
2860 struct osc_enqueue_args *aa;
2861 aa = ptlrpc_req_async_args(aa, req);
2863 aa->oa_mode = einfo->ei_mode;
2864 aa->oa_type = einfo->ei_type;
2865 lustre_handle_copy(&aa->oa_lockh, &lockh);
2866 aa->oa_upcall = upcall;
2867 aa->oa_cookie = cookie;
2868 aa->oa_speculative = speculative;
2870 aa->oa_flags = flags;
2873 /* speculative locks are essentially to enqueue
2874 * a DLM lock in advance, so we don't care
2875 * about the result of the enqueue. */
2877 aa->oa_flags = NULL;
2880 req->rq_interpret_reply = osc_enqueue_interpret;
2881 ptlrpc_set_add_req(rqset, req);
2886 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2887 flags, speculative, rc);
2892 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2893 struct ldlm_res_id *res_id, enum ldlm_type type,
2894 union ldlm_policy_data *policy, enum ldlm_mode mode,
2895 __u64 *flags, struct osc_object *obj,
2896 struct lustre_handle *lockh, enum ldlm_match_flags match_flags)
2898 struct obd_device *obd = exp->exp_obd;
2899 __u64 lflags = *flags;
2903 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2906 /* Filesystem lock extents are extended to page boundaries so that
2907 * dealing with the page cache is a little smoother */
2908 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2909 policy->l_extent.end |= ~PAGE_MASK;
2911 /* Next, search for already existing extent locks that will cover us */
2912 rc = ldlm_lock_match_with_skip(obd->obd_namespace, lflags, 0,
2913 res_id, type, policy, mode, lockh,
2915 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2919 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2921 LASSERT(lock != NULL);
2922 if (osc_set_lock_data(lock, obj)) {
2923 lock_res_and_lock(lock);
2924 if (!ldlm_is_lvb_cached(lock)) {
2925 LASSERT(lock->l_ast_data == obj);
2926 osc_lock_lvb_update(env, obj, lock, NULL);
2927 ldlm_set_lvb_cached(lock);
2929 unlock_res_and_lock(lock);
2931 ldlm_lock_decref(lockh, rc);
2934 LDLM_LOCK_PUT(lock);
2939 static int osc_statfs_interpret(const struct lu_env *env,
2940 struct ptlrpc_request *req, void *args, int rc)
2942 struct osc_async_args *aa = args;
2943 struct obd_statfs *msfs;
2948 * The request has in fact never been sent due to issues at
2949 * a higher level (LOV). Exit immediately since the caller
2950 * is aware of the problem and takes care of the clean up.
2954 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2955 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2961 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2963 GOTO(out, rc = -EPROTO);
2965 *aa->aa_oi->oi_osfs = *msfs;
2967 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2972 static int osc_statfs_async(struct obd_export *exp,
2973 struct obd_info *oinfo, time64_t max_age,
2974 struct ptlrpc_request_set *rqset)
2976 struct obd_device *obd = class_exp2obd(exp);
2977 struct ptlrpc_request *req;
2978 struct osc_async_args *aa;
2982 if (obd->obd_osfs_age >= max_age) {
2984 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2985 obd->obd_name, &obd->obd_osfs,
2986 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2987 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2988 spin_lock(&obd->obd_osfs_lock);
2989 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2990 spin_unlock(&obd->obd_osfs_lock);
2991 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2992 if (oinfo->oi_cb_up)
2993 oinfo->oi_cb_up(oinfo, 0);
2998 /* We could possibly pass max_age in the request (as an absolute
2999 * timestamp or a "seconds.usec ago") so the target can avoid doing
3000 * extra calls into the filesystem if that isn't necessary (e.g.
3001 * during mount that would help a bit). Having relative timestamps
3002 * is not so great if request processing is slow, while absolute
3003 * timestamps are not ideal because they need time synchronization. */
3004 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3008 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3010 ptlrpc_request_free(req);
3013 ptlrpc_request_set_replen(req);
3014 req->rq_request_portal = OST_CREATE_PORTAL;
3015 ptlrpc_at_set_req_timeout(req);
3017 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3018 /* procfs requests not want stat in wait for avoid deadlock */
3019 req->rq_no_resend = 1;
3020 req->rq_no_delay = 1;
3023 req->rq_interpret_reply = osc_statfs_interpret;
3024 aa = ptlrpc_req_async_args(aa, req);
3027 ptlrpc_set_add_req(rqset, req);
3031 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
3032 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
3034 struct obd_device *obd = class_exp2obd(exp);
3035 struct obd_statfs *msfs;
3036 struct ptlrpc_request *req;
3037 struct obd_import *imp, *imp0;
3041 /*Since the request might also come from lprocfs, so we need
3042 *sync this with client_disconnect_export Bug15684
3044 with_imp_locked(obd, imp0, rc)
3045 imp = class_import_get(imp0);
3049 /* We could possibly pass max_age in the request (as an absolute
3050 * timestamp or a "seconds.usec ago") so the target can avoid doing
3051 * extra calls into the filesystem if that isn't necessary (e.g.
3052 * during mount that would help a bit). Having relative timestamps
3053 * is not so great if request processing is slow, while absolute
3054 * timestamps are not ideal because they need time synchronization. */
3055 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3057 class_import_put(imp);
3062 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3064 ptlrpc_request_free(req);
3067 ptlrpc_request_set_replen(req);
3068 req->rq_request_portal = OST_CREATE_PORTAL;
3069 ptlrpc_at_set_req_timeout(req);
3071 if (flags & OBD_STATFS_NODELAY) {
3072 /* procfs requests not want stat in wait for avoid deadlock */
3073 req->rq_no_resend = 1;
3074 req->rq_no_delay = 1;
3077 rc = ptlrpc_queue_wait(req);
3081 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3083 GOTO(out, rc = -EPROTO);
3089 ptlrpc_req_finished(req);
3093 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3094 void *karg, void __user *uarg)
3096 struct obd_device *obd = exp->exp_obd;
3097 struct obd_ioctl_data *data = karg;
3101 if (!try_module_get(THIS_MODULE)) {
3102 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3103 module_name(THIS_MODULE));
3107 case OBD_IOC_CLIENT_RECOVER:
3108 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3109 data->ioc_inlbuf1, 0);
3113 case IOC_OSC_SET_ACTIVE:
3114 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3119 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
3120 obd->obd_name, cmd, current->comm, rc);
3124 module_put(THIS_MODULE);
3128 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3129 u32 keylen, void *key, u32 vallen, void *val,
3130 struct ptlrpc_request_set *set)
3132 struct ptlrpc_request *req;
3133 struct obd_device *obd = exp->exp_obd;
3134 struct obd_import *imp = class_exp2cliimp(exp);
3139 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3141 if (KEY_IS(KEY_CHECKSUM)) {
3142 if (vallen != sizeof(int))
3144 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3148 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3149 sptlrpc_conf_client_adapt(obd);
3153 if (KEY_IS(KEY_FLUSH_CTX)) {
3154 sptlrpc_import_flush_my_ctx(imp);
3158 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3159 struct client_obd *cli = &obd->u.cli;
3160 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3161 long target = *(long *)val;
3163 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3168 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3171 /* We pass all other commands directly to OST. Since nobody calls osc
3172 methods directly and everybody is supposed to go through LOV, we
3173 assume lov checked invalid values for us.
3174 The only recognised values so far are evict_by_nid and mds_conn.
3175 Even if something bad goes through, we'd get a -EINVAL from OST
3178 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3179 &RQF_OST_SET_GRANT_INFO :
3184 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3185 RCL_CLIENT, keylen);
3186 if (!KEY_IS(KEY_GRANT_SHRINK))
3187 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3188 RCL_CLIENT, vallen);
3189 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3191 ptlrpc_request_free(req);
3195 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3196 memcpy(tmp, key, keylen);
3197 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3200 memcpy(tmp, val, vallen);
3202 if (KEY_IS(KEY_GRANT_SHRINK)) {
3203 struct osc_grant_args *aa;
3206 aa = ptlrpc_req_async_args(aa, req);
3207 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3209 ptlrpc_req_finished(req);
3212 *oa = ((struct ost_body *)val)->oa;
3214 req->rq_interpret_reply = osc_shrink_grant_interpret;
3217 ptlrpc_request_set_replen(req);
3218 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3219 LASSERT(set != NULL);
3220 ptlrpc_set_add_req(set, req);
3221 ptlrpc_check_set(NULL, set);
3223 ptlrpcd_add_req(req);
3228 EXPORT_SYMBOL(osc_set_info_async);
3230 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3231 struct obd_device *obd, struct obd_uuid *cluuid,
3232 struct obd_connect_data *data, void *localdata)
3234 struct client_obd *cli = &obd->u.cli;
3236 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3240 spin_lock(&cli->cl_loi_list_lock);
3241 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3242 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3243 /* restore ocd_grant_blkbits as client page bits */
3244 data->ocd_grant_blkbits = PAGE_SHIFT;
3245 grant += cli->cl_dirty_grant;
3247 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3249 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3250 lost_grant = cli->cl_lost_grant;
3251 cli->cl_lost_grant = 0;
3252 spin_unlock(&cli->cl_loi_list_lock);
3254 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3255 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3256 data->ocd_version, data->ocd_grant, lost_grant);
3261 EXPORT_SYMBOL(osc_reconnect);
3263 int osc_disconnect(struct obd_export *exp)
3265 struct obd_device *obd = class_exp2obd(exp);
3268 rc = client_disconnect_export(exp);
3270 * Initially we put del_shrink_grant before disconnect_export, but it
3271 * causes the following problem if setup (connect) and cleanup
3272 * (disconnect) are tangled together.
3273 * connect p1 disconnect p2
3274 * ptlrpc_connect_import
3275 * ............... class_manual_cleanup
3278 * ptlrpc_connect_interrupt
3280 * add this client to shrink list
3282 * Bang! grant shrink thread trigger the shrink. BUG18662
3284 osc_del_grant_list(&obd->u.cli);
3287 EXPORT_SYMBOL(osc_disconnect);
3289 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3290 struct hlist_node *hnode, void *arg)
3292 struct lu_env *env = arg;
3293 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3294 struct ldlm_lock *lock;
3295 struct osc_object *osc = NULL;
3299 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3300 if (lock->l_ast_data != NULL && osc == NULL) {
3301 osc = lock->l_ast_data;
3302 cl_object_get(osc2cl(osc));
3305 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3306 * by the 2nd round of ldlm_namespace_clean() call in
3307 * osc_import_event(). */
3308 ldlm_clear_cleaned(lock);
3313 osc_object_invalidate(env, osc);
3314 cl_object_put(env, osc2cl(osc));
3319 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3321 static int osc_import_event(struct obd_device *obd,
3322 struct obd_import *imp,
3323 enum obd_import_event event)
3325 struct client_obd *cli;
3329 LASSERT(imp->imp_obd == obd);
3332 case IMP_EVENT_DISCON: {
3334 spin_lock(&cli->cl_loi_list_lock);
3335 cli->cl_avail_grant = 0;
3336 cli->cl_lost_grant = 0;
3337 spin_unlock(&cli->cl_loi_list_lock);
3340 case IMP_EVENT_INACTIVE: {
3341 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3344 case IMP_EVENT_INVALIDATE: {
3345 struct ldlm_namespace *ns = obd->obd_namespace;
3349 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3351 env = cl_env_get(&refcheck);
3353 osc_io_unplug(env, &obd->u.cli, NULL);
3355 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3356 osc_ldlm_resource_invalidate,
3358 cl_env_put(env, &refcheck);
3360 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3365 case IMP_EVENT_ACTIVE: {
3366 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3369 case IMP_EVENT_OCD: {
3370 struct obd_connect_data *ocd = &imp->imp_connect_data;
3372 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3373 osc_init_grant(&obd->u.cli, ocd);
3376 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3377 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3379 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3382 case IMP_EVENT_DEACTIVATE: {
3383 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3386 case IMP_EVENT_ACTIVATE: {
3387 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3391 CERROR("Unknown import event %d\n", event);
3398 * Determine whether the lock can be canceled before replaying the lock
3399 * during recovery, see bug16774 for detailed information.
3401 * \retval zero the lock can't be canceled
3402 * \retval other ok to cancel
3404 static int osc_cancel_weight(struct ldlm_lock *lock)
3407 * Cancel all unused and granted extent lock.
3409 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3410 ldlm_is_granted(lock) &&
3411 osc_ldlm_weigh_ast(lock) == 0)
3417 static int brw_queue_work(const struct lu_env *env, void *data)
3419 struct client_obd *cli = data;
3421 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3423 osc_io_unplug(env, cli, NULL);
3427 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3429 struct client_obd *cli = &obd->u.cli;
3435 rc = ptlrpcd_addref();
3439 rc = client_obd_setup(obd, lcfg);
3441 GOTO(out_ptlrpcd, rc);
3444 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3445 if (IS_ERR(handler))
3446 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3447 cli->cl_writeback_work = handler;
3449 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3450 if (IS_ERR(handler))
3451 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3452 cli->cl_lru_work = handler;
3454 rc = osc_quota_setup(obd);
3456 GOTO(out_ptlrpcd_work, rc);
3458 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3459 osc_update_next_shrink(cli);
3464 if (cli->cl_writeback_work != NULL) {
3465 ptlrpcd_destroy_work(cli->cl_writeback_work);
3466 cli->cl_writeback_work = NULL;
3468 if (cli->cl_lru_work != NULL) {
3469 ptlrpcd_destroy_work(cli->cl_lru_work);
3470 cli->cl_lru_work = NULL;
3472 client_obd_cleanup(obd);
3477 EXPORT_SYMBOL(osc_setup_common);
3479 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3481 struct client_obd *cli = &obd->u.cli;
3489 rc = osc_setup_common(obd, lcfg);
3493 rc = osc_tunables_init(obd);
3498 * We try to control the total number of requests with a upper limit
3499 * osc_reqpool_maxreqcount. There might be some race which will cause
3500 * over-limit allocation, but it is fine.
3502 req_count = atomic_read(&osc_pool_req_count);
3503 if (req_count < osc_reqpool_maxreqcount) {
3504 adding = cli->cl_max_rpcs_in_flight + 2;
3505 if (req_count + adding > osc_reqpool_maxreqcount)
3506 adding = osc_reqpool_maxreqcount - req_count;
3508 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3509 atomic_add(added, &osc_pool_req_count);
3512 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3514 spin_lock(&osc_shrink_lock);
3515 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3516 spin_unlock(&osc_shrink_lock);
3517 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3518 cli->cl_import->imp_idle_debug = D_HA;
3523 int osc_precleanup_common(struct obd_device *obd)
3525 struct client_obd *cli = &obd->u.cli;
3529 * for echo client, export may be on zombie list, wait for
3530 * zombie thread to cull it, because cli.cl_import will be
3531 * cleared in client_disconnect_export():
3532 * class_export_destroy() -> obd_cleanup() ->
3533 * echo_device_free() -> echo_client_cleanup() ->
3534 * obd_disconnect() -> osc_disconnect() ->
3535 * client_disconnect_export()
3537 obd_zombie_barrier();
3538 if (cli->cl_writeback_work) {
3539 ptlrpcd_destroy_work(cli->cl_writeback_work);
3540 cli->cl_writeback_work = NULL;
3543 if (cli->cl_lru_work) {
3544 ptlrpcd_destroy_work(cli->cl_lru_work);
3545 cli->cl_lru_work = NULL;
3548 obd_cleanup_client_import(obd);
3551 EXPORT_SYMBOL(osc_precleanup_common);
3553 static int osc_precleanup(struct obd_device *obd)
3557 osc_precleanup_common(obd);
3559 ptlrpc_lprocfs_unregister_obd(obd);
3563 int osc_cleanup_common(struct obd_device *obd)
3565 struct client_obd *cli = &obd->u.cli;
3570 spin_lock(&osc_shrink_lock);
3571 list_del(&cli->cl_shrink_list);
3572 spin_unlock(&osc_shrink_lock);
3575 if (cli->cl_cache != NULL) {
3576 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3577 spin_lock(&cli->cl_cache->ccc_lru_lock);
3578 list_del_init(&cli->cl_lru_osc);
3579 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3580 cli->cl_lru_left = NULL;
3581 cl_cache_decref(cli->cl_cache);
3582 cli->cl_cache = NULL;
3585 /* free memory of osc quota cache */
3586 osc_quota_cleanup(obd);
3588 rc = client_obd_cleanup(obd);
3593 EXPORT_SYMBOL(osc_cleanup_common);
3595 static const struct obd_ops osc_obd_ops = {
3596 .o_owner = THIS_MODULE,
3597 .o_setup = osc_setup,
3598 .o_precleanup = osc_precleanup,
3599 .o_cleanup = osc_cleanup_common,
3600 .o_add_conn = client_import_add_conn,
3601 .o_del_conn = client_import_del_conn,
3602 .o_connect = client_connect_import,
3603 .o_reconnect = osc_reconnect,
3604 .o_disconnect = osc_disconnect,
3605 .o_statfs = osc_statfs,
3606 .o_statfs_async = osc_statfs_async,
3607 .o_create = osc_create,
3608 .o_destroy = osc_destroy,
3609 .o_getattr = osc_getattr,
3610 .o_setattr = osc_setattr,
3611 .o_iocontrol = osc_iocontrol,
3612 .o_set_info_async = osc_set_info_async,
3613 .o_import_event = osc_import_event,
3614 .o_quotactl = osc_quotactl,
3617 LIST_HEAD(osc_shrink_list);
3618 DEFINE_SPINLOCK(osc_shrink_lock);
3620 #ifdef HAVE_SHRINKER_COUNT
3621 static struct shrinker osc_cache_shrinker = {
3622 .count_objects = osc_cache_shrink_count,
3623 .scan_objects = osc_cache_shrink_scan,
3624 .seeks = DEFAULT_SEEKS,
3627 static int osc_cache_shrink(struct shrinker *shrinker,
3628 struct shrink_control *sc)
3630 (void)osc_cache_shrink_scan(shrinker, sc);
3632 return osc_cache_shrink_count(shrinker, sc);
3635 static struct shrinker osc_cache_shrinker = {
3636 .shrink = osc_cache_shrink,
3637 .seeks = DEFAULT_SEEKS,
3641 static int __init osc_init(void)
3643 unsigned int reqpool_size;
3644 unsigned int reqsize;
3648 /* print an address of _any_ initialized kernel symbol from this
3649 * module, to allow debugging with gdb that doesn't support data
3650 * symbols from modules.*/
3651 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3653 rc = lu_kmem_init(osc_caches);
3657 rc = class_register_type(&osc_obd_ops, NULL, true,
3658 LUSTRE_OSC_NAME, &osc_device_type);
3662 rc = register_shrinker(&osc_cache_shrinker);
3666 /* This is obviously too much memory, only prevent overflow here */
3667 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3668 GOTO(out_shrinker, rc = -EINVAL);
3670 reqpool_size = osc_reqpool_mem_max << 20;
3673 while (reqsize < OST_IO_MAXREQSIZE)
3674 reqsize = reqsize << 1;
3677 * We don't enlarge the request count in OSC pool according to
3678 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3679 * tried after normal allocation failed. So a small OSC pool won't
3680 * cause much performance degression in most of cases.
3682 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3684 atomic_set(&osc_pool_req_count, 0);
3685 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3686 ptlrpc_add_rqs_to_pool);
3688 if (osc_rq_pool == NULL)
3689 GOTO(out_shrinker, rc = -ENOMEM);
3691 rc = osc_start_grant_work();
3693 GOTO(out_req_pool, rc);
3698 ptlrpc_free_rq_pool(osc_rq_pool);
3700 unregister_shrinker(&osc_cache_shrinker);
3702 class_unregister_type(LUSTRE_OSC_NAME);
3704 lu_kmem_fini(osc_caches);
3709 static void __exit osc_exit(void)
3711 osc_stop_grant_work();
3712 unregister_shrinker(&osc_cache_shrinker);
3713 class_unregister_type(LUSTRE_OSC_NAME);
3714 lu_kmem_fini(osc_caches);
3715 ptlrpc_free_rq_pool(osc_rq_pool);
3718 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3719 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3720 MODULE_VERSION(LUSTRE_VERSION_STRING);
3721 MODULE_LICENSE("GPL");
3723 module_init(osc_init);
3724 module_exit(osc_exit);