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/
32 #define DEBUG_SUBSYSTEM S_OSC
34 #include <linux/workqueue.h>
35 #include <libcfs/libcfs.h>
36 #include <linux/falloc.h>
37 #include <lprocfs_status.h>
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <lustre_ha.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
42 #include <lustre_net.h>
43 #include <lustre_obdo.h>
45 #include <obd_cksum.h>
46 #include <obd_class.h>
47 #include <lustre_osc.h>
48 #include <linux/falloc.h>
50 #include "osc_internal.h"
51 #include <lnet/lnet_rdma.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);
452 oa->o_falloc_mode = mode;
453 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
458 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_FALLOCATE);
460 ptlrpc_request_free(req);
464 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
467 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
469 ptlrpc_request_set_replen(req);
471 req->rq_interpret_reply = osc_setattr_interpret;
472 BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
473 sa = ptlrpc_req_async_args(sa, req);
475 sa->sa_upcall = upcall;
476 sa->sa_cookie = cookie;
478 ptlrpcd_add_req(req);
482 EXPORT_SYMBOL(osc_fallocate_base);
484 static int osc_sync_interpret(const struct lu_env *env,
485 struct ptlrpc_request *req, void *args, int rc)
487 struct osc_fsync_args *fa = args;
488 struct ost_body *body;
489 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
490 unsigned long valid = 0;
491 struct cl_object *obj;
497 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
499 CERROR("can't unpack ost_body\n");
500 GOTO(out, rc = -EPROTO);
503 *fa->fa_oa = body->oa;
504 obj = osc2cl(fa->fa_obj);
506 /* Update osc object's blocks attribute */
507 cl_object_attr_lock(obj);
508 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
509 attr->cat_blocks = body->oa.o_blocks;
514 cl_object_attr_update(env, obj, attr, valid);
515 cl_object_attr_unlock(obj);
518 rc = fa->fa_upcall(fa->fa_cookie, rc);
522 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
523 obd_enqueue_update_f upcall, void *cookie,
524 struct ptlrpc_request_set *rqset)
526 struct obd_export *exp = osc_export(obj);
527 struct ptlrpc_request *req;
528 struct ost_body *body;
529 struct osc_fsync_args *fa;
533 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
537 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
539 ptlrpc_request_free(req);
543 /* overload the size and blocks fields in the oa with start/end */
544 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
546 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
548 ptlrpc_request_set_replen(req);
549 req->rq_interpret_reply = osc_sync_interpret;
551 fa = ptlrpc_req_async_args(fa, req);
554 fa->fa_upcall = upcall;
555 fa->fa_cookie = cookie;
557 ptlrpc_set_add_req(rqset, req);
562 /* Find and cancel locally locks matched by @mode in the resource found by
563 * @objid. Found locks are added into @cancel list. Returns the amount of
564 * locks added to @cancels list. */
565 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
566 struct list_head *cancels,
567 enum ldlm_mode mode, __u64 lock_flags)
569 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
570 struct ldlm_res_id res_id;
571 struct ldlm_resource *res;
575 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
576 * export) but disabled through procfs (flag in NS).
578 * This distinguishes from a case when ELC is not supported originally,
579 * when we still want to cancel locks in advance and just cancel them
580 * locally, without sending any RPC. */
581 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
584 ostid_build_res_name(&oa->o_oi, &res_id);
585 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
589 LDLM_RESOURCE_ADDREF(res);
590 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
591 lock_flags, 0, NULL);
592 LDLM_RESOURCE_DELREF(res);
593 ldlm_resource_putref(res);
597 static int osc_destroy_interpret(const struct lu_env *env,
598 struct ptlrpc_request *req, void *args, int rc)
600 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
602 atomic_dec(&cli->cl_destroy_in_flight);
603 wake_up(&cli->cl_destroy_waitq);
608 static int osc_can_send_destroy(struct client_obd *cli)
610 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
611 cli->cl_max_rpcs_in_flight) {
612 /* The destroy request can be sent */
615 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
616 cli->cl_max_rpcs_in_flight) {
618 * The counter has been modified between the two atomic
621 wake_up(&cli->cl_destroy_waitq);
626 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
629 struct client_obd *cli = &exp->exp_obd->u.cli;
630 struct ptlrpc_request *req;
631 struct ost_body *body;
637 CDEBUG(D_INFO, "oa NULL\n");
641 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
642 LDLM_FL_DISCARD_DATA);
644 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
646 ldlm_lock_list_put(&cancels, l_bl_ast, count);
650 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
653 ptlrpc_request_free(req);
657 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
658 ptlrpc_at_set_req_timeout(req);
660 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
662 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
664 ptlrpc_request_set_replen(req);
666 req->rq_interpret_reply = osc_destroy_interpret;
667 if (!osc_can_send_destroy(cli)) {
669 * Wait until the number of on-going destroy RPCs drops
670 * under max_rpc_in_flight
672 rc = l_wait_event_abortable_exclusive(
673 cli->cl_destroy_waitq,
674 osc_can_send_destroy(cli));
676 ptlrpc_req_finished(req);
681 /* Do not wait for response */
682 ptlrpcd_add_req(req);
686 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
689 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
691 LASSERT(!(oa->o_valid & bits));
694 spin_lock(&cli->cl_loi_list_lock);
695 if (cli->cl_ocd_grant_param)
696 oa->o_dirty = cli->cl_dirty_grant;
698 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
699 if (unlikely(cli->cl_dirty_pages > cli->cl_dirty_max_pages)) {
700 CERROR("dirty %lu > dirty_max %lu\n",
702 cli->cl_dirty_max_pages);
704 } else if (unlikely(atomic_long_read(&obd_dirty_pages) >
705 (long)(obd_max_dirty_pages + 1))) {
706 /* The atomic_read() allowing the atomic_inc() are
707 * not covered by a lock thus they may safely race and trip
708 * this CERROR() unless we add in a small fudge factor (+1). */
709 CERROR("%s: dirty %ld > system dirty_max %ld\n",
710 cli_name(cli), atomic_long_read(&obd_dirty_pages),
711 obd_max_dirty_pages);
713 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
715 CERROR("dirty %lu - dirty_max %lu too big???\n",
716 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
719 unsigned long nrpages;
720 unsigned long undirty;
722 nrpages = cli->cl_max_pages_per_rpc;
723 nrpages *= cli->cl_max_rpcs_in_flight + 1;
724 nrpages = max(nrpages, cli->cl_dirty_max_pages);
725 undirty = nrpages << PAGE_SHIFT;
726 if (cli->cl_ocd_grant_param) {
729 /* take extent tax into account when asking for more
731 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
732 cli->cl_max_extent_pages;
733 undirty += nrextents * cli->cl_grant_extent_tax;
735 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
736 * to add extent tax, etc.
738 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
739 ~(PTLRPC_MAX_BRW_SIZE * 4UL));
741 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
742 /* o_dropped AKA o_misc is 32 bits, but cl_lost_grant is 64 bits */
743 if (cli->cl_lost_grant > INT_MAX) {
745 "%s: avoided o_dropped overflow: cl_lost_grant %lu\n",
746 cli_name(cli), cli->cl_lost_grant);
747 oa->o_dropped = INT_MAX;
749 oa->o_dropped = cli->cl_lost_grant;
751 cli->cl_lost_grant -= oa->o_dropped;
752 spin_unlock(&cli->cl_loi_list_lock);
753 CDEBUG(D_CACHE, "%s: dirty: %llu undirty: %u dropped %u grant: %llu"
754 " cl_lost_grant %lu\n", cli_name(cli), oa->o_dirty,
755 oa->o_undirty, oa->o_dropped, oa->o_grant, cli->cl_lost_grant);
758 void osc_update_next_shrink(struct client_obd *cli)
760 cli->cl_next_shrink_grant = ktime_get_seconds() +
761 cli->cl_grant_shrink_interval;
763 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
764 cli->cl_next_shrink_grant);
766 EXPORT_SYMBOL(osc_update_next_shrink);
768 static void __osc_update_grant(struct client_obd *cli, u64 grant)
770 spin_lock(&cli->cl_loi_list_lock);
771 cli->cl_avail_grant += grant;
772 spin_unlock(&cli->cl_loi_list_lock);
775 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
777 if (body->oa.o_valid & OBD_MD_FLGRANT) {
778 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
779 __osc_update_grant(cli, body->oa.o_grant);
784 * grant thread data for shrinking space.
786 struct grant_thread_data {
787 struct list_head gtd_clients;
788 struct mutex gtd_mutex;
789 unsigned long gtd_stopped:1;
791 static struct grant_thread_data client_gtd;
793 static int osc_shrink_grant_interpret(const struct lu_env *env,
794 struct ptlrpc_request *req,
797 struct osc_grant_args *aa = args;
798 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
799 struct ost_body *body;
802 __osc_update_grant(cli, aa->aa_oa->o_grant);
806 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
808 osc_update_grant(cli, body);
810 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
816 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
818 spin_lock(&cli->cl_loi_list_lock);
819 oa->o_grant = cli->cl_avail_grant / 4;
820 cli->cl_avail_grant -= oa->o_grant;
821 spin_unlock(&cli->cl_loi_list_lock);
822 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
823 oa->o_valid |= OBD_MD_FLFLAGS;
826 oa->o_flags |= OBD_FL_SHRINK_GRANT;
827 osc_update_next_shrink(cli);
830 /* Shrink the current grant, either from some large amount to enough for a
831 * full set of in-flight RPCs, or if we have already shrunk to that limit
832 * then to enough for a single RPC. This avoids keeping more grant than
833 * needed, and avoids shrinking the grant piecemeal. */
834 static int osc_shrink_grant(struct client_obd *cli)
836 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
837 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
839 spin_lock(&cli->cl_loi_list_lock);
840 if (cli->cl_avail_grant <= target_bytes)
841 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
842 spin_unlock(&cli->cl_loi_list_lock);
844 return osc_shrink_grant_to_target(cli, target_bytes);
847 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
850 struct ost_body *body;
853 spin_lock(&cli->cl_loi_list_lock);
854 /* Don't shrink if we are already above or below the desired limit
855 * We don't want to shrink below a single RPC, as that will negatively
856 * impact block allocation and long-term performance. */
857 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
858 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
860 if (target_bytes >= cli->cl_avail_grant) {
861 spin_unlock(&cli->cl_loi_list_lock);
864 spin_unlock(&cli->cl_loi_list_lock);
870 osc_announce_cached(cli, &body->oa, 0);
872 spin_lock(&cli->cl_loi_list_lock);
873 if (target_bytes >= cli->cl_avail_grant) {
874 /* available grant has changed since target calculation */
875 spin_unlock(&cli->cl_loi_list_lock);
876 GOTO(out_free, rc = 0);
878 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
879 cli->cl_avail_grant = target_bytes;
880 spin_unlock(&cli->cl_loi_list_lock);
881 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
882 body->oa.o_valid |= OBD_MD_FLFLAGS;
883 body->oa.o_flags = 0;
885 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
886 osc_update_next_shrink(cli);
888 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
889 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
890 sizeof(*body), body, NULL);
892 __osc_update_grant(cli, body->oa.o_grant);
898 static int osc_should_shrink_grant(struct client_obd *client)
900 time64_t next_shrink = client->cl_next_shrink_grant;
902 if (client->cl_import == NULL)
905 if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
906 client->cl_import->imp_grant_shrink_disabled) {
907 osc_update_next_shrink(client);
911 if (ktime_get_seconds() >= next_shrink - 5) {
912 /* Get the current RPC size directly, instead of going via:
913 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
914 * Keep comment here so that it can be found by searching. */
915 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
917 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
918 client->cl_avail_grant > brw_size)
921 osc_update_next_shrink(client);
926 #define GRANT_SHRINK_RPC_BATCH 100
928 static struct delayed_work work;
930 static void osc_grant_work_handler(struct work_struct *data)
932 struct client_obd *cli;
934 bool init_next_shrink = true;
935 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
938 mutex_lock(&client_gtd.gtd_mutex);
939 list_for_each_entry(cli, &client_gtd.gtd_clients,
941 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
942 osc_should_shrink_grant(cli)) {
943 osc_shrink_grant(cli);
947 if (!init_next_shrink) {
948 if (cli->cl_next_shrink_grant < next_shrink &&
949 cli->cl_next_shrink_grant > ktime_get_seconds())
950 next_shrink = cli->cl_next_shrink_grant;
952 init_next_shrink = false;
953 next_shrink = cli->cl_next_shrink_grant;
956 mutex_unlock(&client_gtd.gtd_mutex);
958 if (client_gtd.gtd_stopped == 1)
961 if (next_shrink > ktime_get_seconds()) {
962 time64_t delay = next_shrink - ktime_get_seconds();
964 schedule_delayed_work(&work, cfs_time_seconds(delay));
966 schedule_work(&work.work);
970 void osc_schedule_grant_work(void)
972 cancel_delayed_work_sync(&work);
973 schedule_work(&work.work);
975 EXPORT_SYMBOL(osc_schedule_grant_work);
978 * Start grant thread for returing grant to server for idle clients.
980 static int osc_start_grant_work(void)
982 client_gtd.gtd_stopped = 0;
983 mutex_init(&client_gtd.gtd_mutex);
984 INIT_LIST_HEAD(&client_gtd.gtd_clients);
986 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
987 schedule_work(&work.work);
992 static void osc_stop_grant_work(void)
994 client_gtd.gtd_stopped = 1;
995 cancel_delayed_work_sync(&work);
998 static void osc_add_grant_list(struct client_obd *client)
1000 mutex_lock(&client_gtd.gtd_mutex);
1001 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
1002 mutex_unlock(&client_gtd.gtd_mutex);
1005 static void osc_del_grant_list(struct client_obd *client)
1007 if (list_empty(&client->cl_grant_chain))
1010 mutex_lock(&client_gtd.gtd_mutex);
1011 list_del_init(&client->cl_grant_chain);
1012 mutex_unlock(&client_gtd.gtd_mutex);
1015 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1018 * ocd_grant is the total grant amount we're expect to hold: if we've
1019 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
1020 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
1023 * race is tolerable here: if we're evicted, but imp_state already
1024 * left EVICTED state, then cl_dirty_pages must be 0 already.
1026 spin_lock(&cli->cl_loi_list_lock);
1027 cli->cl_avail_grant = ocd->ocd_grant;
1028 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
1029 unsigned long consumed = cli->cl_reserved_grant;
1031 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
1032 consumed += cli->cl_dirty_grant;
1034 consumed += cli->cl_dirty_pages << PAGE_SHIFT;
1035 if (cli->cl_avail_grant < consumed) {
1036 CERROR("%s: granted %ld but already consumed %ld\n",
1037 cli_name(cli), cli->cl_avail_grant, consumed);
1038 cli->cl_avail_grant = 0;
1040 cli->cl_avail_grant -= consumed;
1044 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
1048 /* overhead for each extent insertion */
1049 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
1050 /* determine the appropriate chunk size used by osc_extent. */
1051 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
1052 ocd->ocd_grant_blkbits);
1053 /* max_pages_per_rpc must be chunk aligned */
1054 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
1055 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
1056 ~chunk_mask) & chunk_mask;
1057 /* determine maximum extent size, in #pages */
1058 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
1059 cli->cl_max_extent_pages = (size >> PAGE_SHIFT) ?: 1;
1060 cli->cl_ocd_grant_param = 1;
1062 cli->cl_ocd_grant_param = 0;
1063 cli->cl_grant_extent_tax = 0;
1064 cli->cl_chunkbits = PAGE_SHIFT;
1065 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1067 spin_unlock(&cli->cl_loi_list_lock);
1070 "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld. chunk bits: %d cl_max_extent_pages: %d\n",
1072 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1073 cli->cl_max_extent_pages);
1075 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1076 osc_add_grant_list(cli);
1078 EXPORT_SYMBOL(osc_init_grant);
1080 /* We assume that the reason this OSC got a short read is because it read
1081 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1082 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1083 * this stripe never got written at or beyond this stripe offset yet. */
1084 static void handle_short_read(int nob_read, size_t page_count,
1085 struct brw_page **pga)
1090 /* skip bytes read OK */
1091 while (nob_read > 0) {
1092 LASSERT (page_count > 0);
1094 if (pga[i]->count > nob_read) {
1095 /* EOF inside this page */
1096 ptr = kmap(pga[i]->pg) +
1097 (pga[i]->off & ~PAGE_MASK);
1098 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1105 nob_read -= pga[i]->count;
1110 /* zero remaining pages */
1111 while (page_count-- > 0) {
1112 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1113 memset(ptr, 0, pga[i]->count);
1119 static int check_write_rcs(struct ptlrpc_request *req,
1120 int requested_nob, int niocount,
1121 size_t page_count, struct brw_page **pga)
1126 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1127 sizeof(*remote_rcs) *
1129 if (remote_rcs == NULL) {
1130 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1134 /* return error if any niobuf was in error */
1135 for (i = 0; i < niocount; i++) {
1136 if ((int)remote_rcs[i] < 0) {
1137 CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1138 i, remote_rcs[i], req);
1139 return remote_rcs[i];
1142 if (remote_rcs[i] != 0) {
1143 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1144 i, remote_rcs[i], req);
1148 if (req->rq_bulk != NULL &&
1149 req->rq_bulk->bd_nob_transferred != requested_nob) {
1150 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1151 req->rq_bulk->bd_nob_transferred, requested_nob);
1158 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1160 if (p1->flag != p2->flag) {
1161 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1162 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1163 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC |
1164 OBD_BRW_SYS_RESOURCE);
1166 /* warn if we try to combine flags that we don't know to be
1167 * safe to combine */
1168 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1169 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1170 "report this at https://jira.whamcloud.com/\n",
1171 p1->flag, p2->flag);
1176 return (p1->off + p1->count == p2->off);
1179 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1180 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1181 size_t pg_count, struct brw_page **pga,
1182 int opc, obd_dif_csum_fn *fn,
1184 u32 *check_sum, bool resend)
1186 struct ahash_request *req;
1187 /* Used Adler as the default checksum type on top of DIF tags */
1188 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1189 struct page *__page;
1190 unsigned char *buffer;
1192 unsigned int bufsize;
1194 int used_number = 0;
1200 LASSERT(pg_count > 0);
1202 __page = alloc_page(GFP_KERNEL);
1206 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1209 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1210 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1214 buffer = kmap(__page);
1215 guard_start = (__u16 *)buffer;
1216 guard_number = PAGE_SIZE / sizeof(*guard_start);
1217 CDEBUG(D_PAGE | (resend ? D_HA : 0),
1218 "GRD tags per page=%u, resend=%u, bytes=%u, pages=%zu\n",
1219 guard_number, resend, nob, pg_count);
1221 while (nob > 0 && pg_count > 0) {
1222 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1224 /* corrupt the data before we compute the checksum, to
1225 * simulate an OST->client data error */
1226 if (unlikely(i == 0 && opc == OST_READ &&
1227 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1228 unsigned char *ptr = kmap(pga[i]->pg);
1229 int off = pga[i]->off & ~PAGE_MASK;
1231 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1236 * The left guard number should be able to hold checksums of a
1239 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1240 pga[i]->off & ~PAGE_MASK,
1242 guard_start + used_number,
1243 guard_number - used_number,
1246 if (unlikely(resend))
1247 CDEBUG(D_PAGE | D_HA,
1248 "pga[%u]: used %u off %llu+%u gen checksum: %*phN\n",
1249 i, used, pga[i]->off & ~PAGE_MASK, count,
1250 (int)(used * sizeof(*guard_start)),
1251 guard_start + used_number);
1255 used_number += used;
1256 if (used_number == guard_number) {
1257 cfs_crypto_hash_update_page(req, __page, 0,
1258 used_number * sizeof(*guard_start));
1262 nob -= pga[i]->count;
1270 if (used_number != 0)
1271 cfs_crypto_hash_update_page(req, __page, 0,
1272 used_number * sizeof(*guard_start));
1274 bufsize = sizeof(cksum);
1275 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1277 /* For sending we only compute the wrong checksum instead
1278 * of corrupting the data so it is still correct on a redo */
1279 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1284 __free_page(__page);
1287 #else /* !CONFIG_CRC_T10DIF */
1288 #define obd_dif_ip_fn NULL
1289 #define obd_dif_crc_fn NULL
1290 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum, re) \
1292 #endif /* CONFIG_CRC_T10DIF */
1294 static int osc_checksum_bulk(int nob, size_t pg_count,
1295 struct brw_page **pga, int opc,
1296 enum cksum_types cksum_type,
1300 struct ahash_request *req;
1301 unsigned int bufsize;
1302 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1304 LASSERT(pg_count > 0);
1306 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1308 CERROR("Unable to initialize checksum hash %s\n",
1309 cfs_crypto_hash_name(cfs_alg));
1310 return PTR_ERR(req);
1313 while (nob > 0 && pg_count > 0) {
1314 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1316 /* corrupt the data before we compute the checksum, to
1317 * simulate an OST->client data error */
1318 if (i == 0 && opc == OST_READ &&
1319 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1320 unsigned char *ptr = kmap(pga[i]->pg);
1321 int off = pga[i]->off & ~PAGE_MASK;
1323 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1326 cfs_crypto_hash_update_page(req, pga[i]->pg,
1327 pga[i]->off & ~PAGE_MASK,
1329 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1330 (int)(pga[i]->off & ~PAGE_MASK));
1332 nob -= pga[i]->count;
1337 bufsize = sizeof(*cksum);
1338 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1340 /* For sending we only compute the wrong checksum instead
1341 * of corrupting the data so it is still correct on a redo */
1342 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1348 static int osc_checksum_bulk_rw(const char *obd_name,
1349 enum cksum_types cksum_type,
1350 int nob, size_t pg_count,
1351 struct brw_page **pga, int opc,
1352 u32 *check_sum, bool resend)
1354 obd_dif_csum_fn *fn = NULL;
1355 int sector_size = 0;
1359 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1362 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1363 opc, fn, sector_size, check_sum,
1366 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1372 static inline void osc_release_bounce_pages(struct brw_page **pga,
1375 #ifdef HAVE_LUSTRE_CRYPTO
1378 for (i = 0; i < page_count; i++) {
1379 /* Bounce pages allocated by a call to
1380 * llcrypt_encrypt_pagecache_blocks() in osc_brw_prep_request()
1381 * are identified thanks to the PageChecked flag.
1383 if (PageChecked(pga[i]->pg))
1384 llcrypt_finalize_bounce_page(&pga[i]->pg);
1385 pga[i]->count -= pga[i]->bp_count_diff;
1386 pga[i]->off += pga[i]->bp_off_diff;
1392 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1393 u32 page_count, struct brw_page **pga,
1394 struct ptlrpc_request **reqp, int resend)
1396 struct ptlrpc_request *req;
1397 struct ptlrpc_bulk_desc *desc;
1398 struct ost_body *body;
1399 struct obd_ioobj *ioobj;
1400 struct niobuf_remote *niobuf;
1401 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1402 struct osc_brw_async_args *aa;
1403 struct req_capsule *pill;
1404 struct brw_page *pg_prev;
1406 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1407 struct inode *inode = NULL;
1408 bool directio = false;
1409 bool enable_checksum = true;
1413 inode = page2inode(pga[0]->pg);
1414 if (inode == NULL) {
1415 /* Try to get reference to inode from cl_page if we are
1416 * dealing with direct IO, as handled pages are not
1417 * actual page cache pages.
1419 struct osc_async_page *oap = brw_page2oap(pga[0]);
1420 struct cl_page *clpage = oap2cl_page(oap);
1422 inode = clpage->cp_inode;
1427 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1428 RETURN(-ENOMEM); /* Recoverable */
1429 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1430 RETURN(-EINVAL); /* Fatal */
1432 if ((cmd & OBD_BRW_WRITE) != 0) {
1434 req = ptlrpc_request_alloc_pool(cli->cl_import,
1436 &RQF_OST_BRW_WRITE);
1439 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1444 if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1445 for (i = 0; i < page_count; i++) {
1446 struct brw_page *pg = pga[i];
1447 struct page *data_page = NULL;
1448 bool retried = false;
1449 bool lockedbymyself;
1450 u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1451 struct address_space *map_orig = NULL;
1455 if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1456 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1457 LUSTRE_ENCRYPTION_UNIT_SIZE;
1458 /* The page can already be locked when we arrive here.
1459 * This is possible when cl_page_assume/vvp_page_assume
1460 * is stuck on wait_on_page_writeback with page lock
1461 * held. In this case there is no risk for the lock to
1462 * be released while we are doing our encryption
1463 * processing, because writeback against that page will
1464 * end in vvp_page_completion_write/cl_page_completion,
1465 * which means only once the page is fully processed.
1467 lockedbymyself = trylock_page(pg->pg);
1469 map_orig = pg->pg->mapping;
1470 pg->pg->mapping = inode->i_mapping;
1471 index_orig = pg->pg->index;
1472 pg->pg->index = pg->off >> PAGE_SHIFT;
1475 llcrypt_encrypt_pagecache_blocks(pg->pg,
1479 pg->pg->mapping = map_orig;
1480 pg->pg->index = index_orig;
1483 unlock_page(pg->pg);
1484 if (IS_ERR(data_page)) {
1485 rc = PTR_ERR(data_page);
1486 if (rc == -ENOMEM && !retried) {
1491 ptlrpc_request_free(req);
1494 /* Set PageChecked flag on bounce page for
1495 * disambiguation in osc_release_bounce_pages().
1497 SetPageChecked(data_page);
1499 /* there should be no gap in the middle of page array */
1500 if (i == page_count - 1) {
1501 struct osc_async_page *oap = brw_page2oap(pg);
1503 oa->o_size = oap->oap_count +
1504 oap->oap_obj_off + oap->oap_page_off;
1506 /* len is forced to nunits, and relative offset to 0
1507 * so store the old, clear text info
1509 pg->bp_count_diff = nunits - pg->count;
1511 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1512 pg->off = pg->off & PAGE_MASK;
1514 } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode)) {
1515 for (i = 0; i < page_count; i++) {
1516 struct brw_page *pg = pga[i];
1517 u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1519 if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1520 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1521 LUSTRE_ENCRYPTION_UNIT_SIZE;
1522 /* count/off are forced to cover the whole encryption
1523 * unit size so that all encrypted data is stored on the
1524 * OST, so adjust bp_{count,off}_diff for the size of
1527 pg->bp_count_diff = nunits - pg->count;
1529 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1530 pg->off = pg->off & PAGE_MASK;
1534 for (niocount = i = 1; i < page_count; i++) {
1535 if (!can_merge_pages(pga[i - 1], pga[i]))
1539 pill = &req->rq_pill;
1540 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1542 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1543 niocount * sizeof(*niobuf));
1545 for (i = 0; i < page_count; i++) {
1546 short_io_size += pga[i]->count;
1547 if (!inode || !IS_ENCRYPTED(inode)) {
1548 pga[i]->bp_count_diff = 0;
1549 pga[i]->bp_off_diff = 0;
1553 if (lnet_is_rdma_only_page(pga[0]->pg)) {
1554 enable_checksum = false;
1558 /* Check if read/write is small enough to be a short io. */
1559 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1560 !imp_connect_shortio(cli->cl_import))
1563 /* If this is an empty RPC to old server, just ignore it */
1564 if (!short_io_size && !pga[0]->pg) {
1565 ptlrpc_request_free(req);
1569 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1570 opc == OST_READ ? 0 : short_io_size);
1571 if (opc == OST_READ)
1572 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1575 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1577 ptlrpc_request_free(req);
1580 osc_set_io_portal(req);
1582 ptlrpc_at_set_req_timeout(req);
1583 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1585 req->rq_no_retry_einprogress = 1;
1587 if (short_io_size != 0) {
1589 short_io_buf = NULL;
1593 desc = ptlrpc_prep_bulk_imp(req, page_count,
1594 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1595 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1596 PTLRPC_BULK_PUT_SINK),
1598 &ptlrpc_bulk_kiov_pin_ops);
1601 GOTO(out, rc = -ENOMEM);
1602 /* NB request now owns desc and will free it when it gets freed */
1604 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1605 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1606 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1607 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1609 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1611 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1612 * and from_kgid(), because they are asynchronous. Fortunately, variable
1613 * oa contains valid o_uid and o_gid in these two operations.
1614 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1615 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1616 * other process logic */
1617 body->oa.o_uid = oa->o_uid;
1618 body->oa.o_gid = oa->o_gid;
1620 obdo_to_ioobj(oa, ioobj);
1621 ioobj->ioo_bufcnt = niocount;
1622 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1623 * that might be send for this request. The actual number is decided
1624 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1625 * "max - 1" for old client compatibility sending "0", and also so the
1626 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1628 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1630 ioobj_max_brw_set(ioobj, 0);
1632 if (short_io_size != 0) {
1633 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1634 body->oa.o_valid |= OBD_MD_FLFLAGS;
1635 body->oa.o_flags = 0;
1637 body->oa.o_flags |= OBD_FL_SHORT_IO;
1638 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1640 if (opc == OST_WRITE) {
1641 short_io_buf = req_capsule_client_get(pill,
1643 LASSERT(short_io_buf != NULL);
1647 LASSERT(page_count > 0);
1649 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1650 struct brw_page *pg = pga[i];
1651 int poff = pg->off & ~PAGE_MASK;
1653 LASSERT(pg->count > 0);
1654 /* make sure there is no gap in the middle of page array */
1655 LASSERTF(page_count == 1 ||
1656 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1657 ergo(i > 0 && i < page_count - 1,
1658 poff == 0 && pg->count == PAGE_SIZE) &&
1659 ergo(i == page_count - 1, poff == 0)),
1660 "i: %d/%d pg: %p off: %llu, count: %u\n",
1661 i, page_count, pg, pg->off, pg->count);
1662 LASSERTF(i == 0 || pg->off > pg_prev->off,
1663 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1664 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1666 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1667 pg_prev->pg, page_private(pg_prev->pg),
1668 pg_prev->pg->index, pg_prev->off);
1669 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1670 (pg->flag & OBD_BRW_SRVLOCK));
1671 if (short_io_size != 0 && opc == OST_WRITE) {
1672 unsigned char *ptr = kmap_atomic(pg->pg);
1674 LASSERT(short_io_size >= requested_nob + pg->count);
1675 memcpy(short_io_buf + requested_nob,
1679 } else if (short_io_size == 0) {
1680 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1683 requested_nob += pg->count;
1685 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1687 niobuf->rnb_len += pg->count;
1689 niobuf->rnb_offset = pg->off;
1690 niobuf->rnb_len = pg->count;
1691 niobuf->rnb_flags = pg->flag;
1696 LASSERTF((void *)(niobuf - niocount) ==
1697 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1698 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1699 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1701 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1703 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1704 body->oa.o_valid |= OBD_MD_FLFLAGS;
1705 body->oa.o_flags = 0;
1707 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1710 if (osc_should_shrink_grant(cli))
1711 osc_shrink_grant_local(cli, &body->oa);
1713 if (!cli->cl_checksum || sptlrpc_flavor_has_bulk(&req->rq_flvr))
1714 enable_checksum = false;
1716 /* size[REQ_REC_OFF] still sizeof (*body) */
1717 if (opc == OST_WRITE) {
1718 if (enable_checksum) {
1719 /* store cl_cksum_type in a local variable since
1720 * it can be changed via lprocfs */
1721 enum cksum_types cksum_type = cli->cl_cksum_type;
1723 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1724 body->oa.o_flags = 0;
1726 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1728 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1730 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1731 requested_nob, page_count,
1733 &body->oa.o_cksum, resend);
1735 CDEBUG(D_PAGE, "failed to checksum: rc = %d\n",
1739 CDEBUG(D_PAGE | (resend ? D_HA : 0),
1740 "checksum at write origin: %x (%x)\n",
1741 body->oa.o_cksum, cksum_type);
1743 /* save this in 'oa', too, for later checking */
1744 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1745 oa->o_flags |= obd_cksum_type_pack(obd_name,
1748 /* clear out the checksum flag, in case this is a
1749 * resend but cl_checksum is no longer set. b=11238 */
1750 oa->o_valid &= ~OBD_MD_FLCKSUM;
1752 oa->o_cksum = body->oa.o_cksum;
1753 /* 1 RC per niobuf */
1754 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1755 sizeof(__u32) * niocount);
1757 if (enable_checksum) {
1758 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1759 body->oa.o_flags = 0;
1760 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1761 cli->cl_cksum_type);
1762 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1765 /* Client cksum has been already copied to wire obdo in previous
1766 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1767 * resent due to cksum error, this will allow Server to
1768 * check+dump pages on its side */
1770 ptlrpc_request_set_replen(req);
1772 aa = ptlrpc_req_async_args(aa, req);
1774 aa->aa_requested_nob = requested_nob;
1775 aa->aa_nio_count = niocount;
1776 aa->aa_page_count = page_count;
1780 INIT_LIST_HEAD(&aa->aa_oaps);
1783 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1784 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1785 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1786 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1790 ptlrpc_req_finished(req);
1794 char dbgcksum_file_name[PATH_MAX];
1796 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1797 struct brw_page **pga, __u32 server_cksum,
1805 /* will only keep dump of pages on first error for the same range in
1806 * file/fid, not during the resends/retries. */
1807 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1808 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1809 (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1810 libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1811 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1812 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1813 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1815 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1816 client_cksum, server_cksum);
1817 CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1818 filp = filp_open(dbgcksum_file_name,
1819 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1823 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1824 "checksum error: rc = %d\n", dbgcksum_file_name,
1827 CERROR("%s: can't open to dump pages with checksum "
1828 "error: rc = %d\n", dbgcksum_file_name, rc);
1832 for (i = 0; i < page_count; i++) {
1833 len = pga[i]->count;
1834 buf = kmap(pga[i]->pg);
1836 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1838 CERROR("%s: wanted to write %u but got %d "
1839 "error\n", dbgcksum_file_name, len, rc);
1848 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1850 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1851 filp_close(filp, NULL);
1853 libcfs_debug_dumplog();
1857 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1858 __u32 client_cksum, __u32 server_cksum,
1859 struct osc_brw_async_args *aa)
1861 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1862 enum cksum_types cksum_type;
1863 obd_dif_csum_fn *fn = NULL;
1864 int sector_size = 0;
1869 if (server_cksum == client_cksum) {
1870 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1874 if (aa->aa_cli->cl_checksum_dump)
1875 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1876 server_cksum, client_cksum);
1878 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1881 switch (cksum_type) {
1882 case OBD_CKSUM_T10IP512:
1886 case OBD_CKSUM_T10IP4K:
1890 case OBD_CKSUM_T10CRC512:
1891 fn = obd_dif_crc_fn;
1894 case OBD_CKSUM_T10CRC4K:
1895 fn = obd_dif_crc_fn;
1903 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1904 aa->aa_page_count, aa->aa_ppga,
1905 OST_WRITE, fn, sector_size,
1908 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1909 aa->aa_ppga, OST_WRITE, cksum_type,
1913 msg = "failed to calculate the client write checksum";
1914 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1915 msg = "the server did not use the checksum type specified in "
1916 "the original request - likely a protocol problem";
1917 else if (new_cksum == server_cksum)
1918 msg = "changed on the client after we checksummed it - "
1919 "likely false positive due to mmap IO (bug 11742)";
1920 else if (new_cksum == client_cksum)
1921 msg = "changed in transit before arrival at OST";
1923 msg = "changed in transit AND doesn't match the original - "
1924 "likely false positive due to mmap IO (bug 11742)";
1926 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1927 DFID " object "DOSTID" extent [%llu-%llu], original "
1928 "client csum %x (type %x), server csum %x (type %x),"
1929 " client csum now %x\n",
1930 obd_name, msg, libcfs_nid2str(peer->nid),
1931 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1932 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1933 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1934 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1935 aa->aa_ppga[aa->aa_page_count - 1]->off +
1936 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1938 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1939 server_cksum, cksum_type, new_cksum);
1943 /* Note rc enters this function as number of bytes transferred */
1944 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1946 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1947 struct client_obd *cli = aa->aa_cli;
1948 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1949 const struct lnet_process_id *peer =
1950 &req->rq_import->imp_connection->c_peer;
1951 struct ost_body *body;
1952 u32 client_cksum = 0;
1953 struct inode *inode;
1954 unsigned int blockbits = 0, blocksize = 0;
1958 if (rc < 0 && rc != -EDQUOT) {
1959 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1963 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1964 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1966 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1970 /* set/clear over quota flag for a uid/gid/projid */
1971 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1972 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1973 unsigned qid[LL_MAXQUOTAS] = {
1974 body->oa.o_uid, body->oa.o_gid,
1975 body->oa.o_projid };
1977 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1978 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1979 body->oa.o_valid, body->oa.o_flags);
1980 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1984 osc_update_grant(cli, body);
1989 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1990 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1992 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1994 CERROR("%s: unexpected positive size %d\n",
1999 if (req->rq_bulk != NULL &&
2000 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
2003 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
2004 check_write_checksum(&body->oa, peer, client_cksum,
2005 body->oa.o_cksum, aa))
2008 rc = check_write_rcs(req, aa->aa_requested_nob,
2009 aa->aa_nio_count, aa->aa_page_count,
2014 /* The rest of this function executes only for OST_READs */
2016 if (req->rq_bulk == NULL) {
2017 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
2019 LASSERT(rc == req->rq_status);
2021 /* if unwrap_bulk failed, return -EAGAIN to retry */
2022 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
2025 GOTO(out, rc = -EAGAIN);
2027 if (rc > aa->aa_requested_nob) {
2028 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
2029 rc, aa->aa_requested_nob);
2033 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
2034 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
2035 rc, req->rq_bulk->bd_nob_transferred);
2039 if (req->rq_bulk == NULL) {
2041 int nob, pg_count, i = 0;
2044 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
2045 pg_count = aa->aa_page_count;
2046 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
2049 while (nob > 0 && pg_count > 0) {
2051 int count = aa->aa_ppga[i]->count > nob ?
2052 nob : aa->aa_ppga[i]->count;
2054 CDEBUG(D_CACHE, "page %p count %d\n",
2055 aa->aa_ppga[i]->pg, count);
2056 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
2057 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
2059 kunmap_atomic((void *) ptr);
2068 if (rc < aa->aa_requested_nob)
2069 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
2071 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2072 static int cksum_counter;
2073 u32 server_cksum = body->oa.o_cksum;
2077 enum cksum_types cksum_type;
2078 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
2079 body->oa.o_flags : 0;
2081 cksum_type = obd_cksum_type_unpack(o_flags);
2082 rc = osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2083 aa->aa_page_count, aa->aa_ppga,
2084 OST_READ, &client_cksum, false);
2088 if (req->rq_bulk != NULL &&
2089 peer->nid != req->rq_bulk->bd_sender) {
2091 router = libcfs_nid2str(req->rq_bulk->bd_sender);
2094 if (server_cksum != client_cksum) {
2095 struct ost_body *clbody;
2096 __u32 client_cksum2;
2097 u32 page_count = aa->aa_page_count;
2099 osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2100 page_count, aa->aa_ppga,
2101 OST_READ, &client_cksum2, true);
2102 clbody = req_capsule_client_get(&req->rq_pill,
2104 if (cli->cl_checksum_dump)
2105 dump_all_bulk_pages(&clbody->oa, page_count,
2106 aa->aa_ppga, server_cksum,
2109 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2110 "%s%s%s inode "DFID" object "DOSTID
2111 " extent [%llu-%llu], client %x/%x, "
2112 "server %x, cksum_type %x\n",
2114 libcfs_nid2str(peer->nid),
2116 clbody->oa.o_valid & OBD_MD_FLFID ?
2117 clbody->oa.o_parent_seq : 0ULL,
2118 clbody->oa.o_valid & OBD_MD_FLFID ?
2119 clbody->oa.o_parent_oid : 0,
2120 clbody->oa.o_valid & OBD_MD_FLFID ?
2121 clbody->oa.o_parent_ver : 0,
2122 POSTID(&body->oa.o_oi),
2123 aa->aa_ppga[0]->off,
2124 aa->aa_ppga[page_count-1]->off +
2125 aa->aa_ppga[page_count-1]->count - 1,
2126 client_cksum, client_cksum2,
2127 server_cksum, cksum_type);
2129 aa->aa_oa->o_cksum = client_cksum;
2133 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2136 } else if (unlikely(client_cksum)) {
2137 static int cksum_missed;
2140 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2141 CERROR("%s: checksum %u requested from %s but not sent\n",
2142 obd_name, cksum_missed,
2143 libcfs_nid2str(peer->nid));
2148 inode = page2inode(aa->aa_ppga[0]->pg);
2149 if (inode == NULL) {
2150 /* Try to get reference to inode from cl_page if we are
2151 * dealing with direct IO, as handled pages are not
2152 * actual page cache pages.
2154 struct osc_async_page *oap = brw_page2oap(aa->aa_ppga[0]);
2156 inode = oap2cl_page(oap)->cp_inode;
2158 blockbits = inode->i_blkbits;
2159 blocksize = 1 << blockbits;
2162 if (inode && IS_ENCRYPTED(inode)) {
2165 if (!llcrypt_has_encryption_key(inode)) {
2166 CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2169 for (idx = 0; idx < aa->aa_page_count; idx++) {
2170 struct brw_page *pg = aa->aa_ppga[idx];
2171 unsigned int offs = 0;
2173 while (offs < PAGE_SIZE) {
2174 /* do not decrypt if page is all 0s */
2175 if (memchr_inv(page_address(pg->pg) + offs, 0,
2176 LUSTRE_ENCRYPTION_UNIT_SIZE) == NULL) {
2177 /* if page is empty forward info to
2178 * upper layers (ll_io_zero_page) by
2179 * clearing PagePrivate2
2182 ClearPagePrivate2(pg->pg);
2187 /* This is direct IO case. Directly call
2188 * decrypt function that takes inode as
2189 * input parameter. Page does not need
2193 ((u64)(pg->off >> PAGE_SHIFT) <<
2194 (PAGE_SHIFT - blockbits)) +
2195 (offs >> blockbits);
2200 LUSTRE_ENCRYPTION_UNIT_SIZE;
2201 i += blocksize, lblk_num++) {
2203 llcrypt_decrypt_block_inplace(
2211 rc = llcrypt_decrypt_pagecache_blocks(
2213 LUSTRE_ENCRYPTION_UNIT_SIZE,
2219 offs += LUSTRE_ENCRYPTION_UNIT_SIZE;
2226 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2227 aa->aa_oa, &body->oa);
2232 static int osc_brw_redo_request(struct ptlrpc_request *request,
2233 struct osc_brw_async_args *aa, int rc)
2235 struct ptlrpc_request *new_req;
2236 struct osc_brw_async_args *new_aa;
2237 struct osc_async_page *oap;
2240 /* The below message is checked in replay-ost-single.sh test_8ae*/
2241 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2242 "redo for recoverable error %d", rc);
2244 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2245 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2246 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2247 aa->aa_ppga, &new_req, 1);
2251 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2252 if (oap->oap_request != NULL) {
2253 LASSERTF(request == oap->oap_request,
2254 "request %p != oap_request %p\n",
2255 request, oap->oap_request);
2259 * New request takes over pga and oaps from old request.
2260 * Note that copying a list_head doesn't work, need to move it...
2263 new_req->rq_interpret_reply = request->rq_interpret_reply;
2264 new_req->rq_async_args = request->rq_async_args;
2265 new_req->rq_commit_cb = request->rq_commit_cb;
2266 /* cap resend delay to the current request timeout, this is similar to
2267 * what ptlrpc does (see after_reply()) */
2268 if (aa->aa_resends > new_req->rq_timeout)
2269 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2271 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2272 new_req->rq_generation_set = 1;
2273 new_req->rq_import_generation = request->rq_import_generation;
2275 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2277 INIT_LIST_HEAD(&new_aa->aa_oaps);
2278 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2279 INIT_LIST_HEAD(&new_aa->aa_exts);
2280 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2281 new_aa->aa_resends = aa->aa_resends;
2283 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2284 if (oap->oap_request) {
2285 ptlrpc_req_finished(oap->oap_request);
2286 oap->oap_request = ptlrpc_request_addref(new_req);
2290 /* XXX: This code will run into problem if we're going to support
2291 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2292 * and wait for all of them to be finished. We should inherit request
2293 * set from old request. */
2294 ptlrpcd_add_req(new_req);
2296 DEBUG_REQ(D_INFO, new_req, "new request");
2301 * ugh, we want disk allocation on the target to happen in offset order. we'll
2302 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2303 * fine for our small page arrays and doesn't require allocation. its an
2304 * insertion sort that swaps elements that are strides apart, shrinking the
2305 * stride down until its '1' and the array is sorted.
2307 static void sort_brw_pages(struct brw_page **array, int num)
2310 struct brw_page *tmp;
2314 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2319 for (i = stride ; i < num ; i++) {
2322 while (j >= stride && array[j - stride]->off > tmp->off) {
2323 array[j] = array[j - stride];
2328 } while (stride > 1);
2331 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2333 LASSERT(ppga != NULL);
2334 OBD_FREE_PTR_ARRAY_LARGE(ppga, count);
2337 static int brw_interpret(const struct lu_env *env,
2338 struct ptlrpc_request *req, void *args, int rc)
2340 struct osc_brw_async_args *aa = args;
2341 struct osc_extent *ext;
2342 struct osc_extent *tmp;
2343 struct client_obd *cli = aa->aa_cli;
2344 unsigned long transferred = 0;
2348 rc = osc_brw_fini_request(req, rc);
2349 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2351 /* restore clear text pages */
2352 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2355 * When server returns -EINPROGRESS, client should always retry
2356 * regardless of the number of times the bulk was resent already.
2358 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2359 if (req->rq_import_generation !=
2360 req->rq_import->imp_generation) {
2361 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2362 ""DOSTID", rc = %d.\n",
2363 req->rq_import->imp_obd->obd_name,
2364 POSTID(&aa->aa_oa->o_oi), rc);
2365 } else if (rc == -EINPROGRESS ||
2366 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2367 rc = osc_brw_redo_request(req, aa, rc);
2369 CERROR("%s: too many resent retries for object: "
2370 "%llu:%llu, rc = %d.\n",
2371 req->rq_import->imp_obd->obd_name,
2372 POSTID(&aa->aa_oa->o_oi), rc);
2377 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2382 struct obdo *oa = aa->aa_oa;
2383 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2384 unsigned long valid = 0;
2385 struct cl_object *obj;
2386 struct osc_async_page *last;
2388 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2389 obj = osc2cl(last->oap_obj);
2391 cl_object_attr_lock(obj);
2392 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2393 attr->cat_blocks = oa->o_blocks;
2394 valid |= CAT_BLOCKS;
2396 if (oa->o_valid & OBD_MD_FLMTIME) {
2397 attr->cat_mtime = oa->o_mtime;
2400 if (oa->o_valid & OBD_MD_FLATIME) {
2401 attr->cat_atime = oa->o_atime;
2404 if (oa->o_valid & OBD_MD_FLCTIME) {
2405 attr->cat_ctime = oa->o_ctime;
2409 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2410 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2411 loff_t last_off = last->oap_count + last->oap_obj_off +
2414 /* Change file size if this is an out of quota or
2415 * direct IO write and it extends the file size */
2416 if (loi->loi_lvb.lvb_size < last_off) {
2417 attr->cat_size = last_off;
2420 /* Extend KMS if it's not a lockless write */
2421 if (loi->loi_kms < last_off &&
2422 oap2osc_page(last)->ops_srvlock == 0) {
2423 attr->cat_kms = last_off;
2429 cl_object_attr_update(env, obj, attr, valid);
2430 cl_object_attr_unlock(obj);
2432 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2435 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2436 osc_inc_unstable_pages(req);
2438 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2439 list_del_init(&ext->oe_link);
2440 osc_extent_finish(env, ext, 1,
2441 rc && req->rq_no_delay ? -EAGAIN : rc);
2443 LASSERT(list_empty(&aa->aa_exts));
2444 LASSERT(list_empty(&aa->aa_oaps));
2446 transferred = (req->rq_bulk == NULL ? /* short io */
2447 aa->aa_requested_nob :
2448 req->rq_bulk->bd_nob_transferred);
2450 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2451 ptlrpc_lprocfs_brw(req, transferred);
2453 spin_lock(&cli->cl_loi_list_lock);
2454 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2455 * is called so we know whether to go to sync BRWs or wait for more
2456 * RPCs to complete */
2457 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2458 cli->cl_w_in_flight--;
2460 cli->cl_r_in_flight--;
2461 osc_wake_cache_waiters(cli);
2462 spin_unlock(&cli->cl_loi_list_lock);
2464 osc_io_unplug(env, cli, NULL);
2468 static void brw_commit(struct ptlrpc_request *req)
2470 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2471 * this called via the rq_commit_cb, I need to ensure
2472 * osc_dec_unstable_pages is still called. Otherwise unstable
2473 * pages may be leaked. */
2474 spin_lock(&req->rq_lock);
2475 if (likely(req->rq_unstable)) {
2476 req->rq_unstable = 0;
2477 spin_unlock(&req->rq_lock);
2479 osc_dec_unstable_pages(req);
2481 req->rq_committed = 1;
2482 spin_unlock(&req->rq_lock);
2487 * Build an RPC by the list of extent @ext_list. The caller must ensure
2488 * that the total pages in this list are NOT over max pages per RPC.
2489 * Extents in the list must be in OES_RPC state.
2491 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2492 struct list_head *ext_list, int cmd)
2494 struct ptlrpc_request *req = NULL;
2495 struct osc_extent *ext;
2496 struct brw_page **pga = NULL;
2497 struct osc_brw_async_args *aa = NULL;
2498 struct obdo *oa = NULL;
2499 struct osc_async_page *oap;
2500 struct osc_object *obj = NULL;
2501 struct cl_req_attr *crattr = NULL;
2502 loff_t starting_offset = OBD_OBJECT_EOF;
2503 loff_t ending_offset = 0;
2504 /* '1' for consistency with code that checks !mpflag to restore */
2508 bool soft_sync = false;
2509 bool ndelay = false;
2513 __u32 layout_version = 0;
2514 LIST_HEAD(rpc_list);
2515 struct ost_body *body;
2517 LASSERT(!list_empty(ext_list));
2519 /* add pages into rpc_list to build BRW rpc */
2520 list_for_each_entry(ext, ext_list, oe_link) {
2521 LASSERT(ext->oe_state == OES_RPC);
2522 mem_tight |= ext->oe_memalloc;
2523 grant += ext->oe_grants;
2524 page_count += ext->oe_nr_pages;
2525 layout_version = max(layout_version, ext->oe_layout_version);
2530 soft_sync = osc_over_unstable_soft_limit(cli);
2532 mpflag = memalloc_noreclaim_save();
2534 OBD_ALLOC_PTR_ARRAY_LARGE(pga, page_count);
2536 GOTO(out, rc = -ENOMEM);
2538 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2540 GOTO(out, rc = -ENOMEM);
2543 list_for_each_entry(ext, ext_list, oe_link) {
2544 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2546 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2548 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2549 pga[i] = &oap->oap_brw_page;
2550 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2553 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2554 if (starting_offset == OBD_OBJECT_EOF ||
2555 starting_offset > oap->oap_obj_off)
2556 starting_offset = oap->oap_obj_off;
2558 LASSERT(oap->oap_page_off == 0);
2559 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2560 ending_offset = oap->oap_obj_off +
2563 LASSERT(oap->oap_page_off + oap->oap_count ==
2570 /* first page in the list */
2571 oap = list_first_entry(&rpc_list, typeof(*oap), oap_rpc_item);
2573 crattr = &osc_env_info(env)->oti_req_attr;
2574 memset(crattr, 0, sizeof(*crattr));
2575 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2576 crattr->cra_flags = ~0ULL;
2577 crattr->cra_page = oap2cl_page(oap);
2578 crattr->cra_oa = oa;
2579 cl_req_attr_set(env, osc2cl(obj), crattr);
2581 if (cmd == OBD_BRW_WRITE) {
2582 oa->o_grant_used = grant;
2583 if (layout_version > 0) {
2584 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2585 PFID(&oa->o_oi.oi_fid), layout_version);
2587 oa->o_layout_version = layout_version;
2588 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2592 sort_brw_pages(pga, page_count);
2593 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2595 CERROR("prep_req failed: %d\n", rc);
2599 req->rq_commit_cb = brw_commit;
2600 req->rq_interpret_reply = brw_interpret;
2601 req->rq_memalloc = mem_tight != 0;
2602 oap->oap_request = ptlrpc_request_addref(req);
2604 req->rq_no_resend = req->rq_no_delay = 1;
2605 /* probably set a shorter timeout value.
2606 * to handle ETIMEDOUT in brw_interpret() correctly. */
2607 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2610 /* Need to update the timestamps after the request is built in case
2611 * we race with setattr (locally or in queue at OST). If OST gets
2612 * later setattr before earlier BRW (as determined by the request xid),
2613 * the OST will not use BRW timestamps. Sadly, there is no obvious
2614 * way to do this in a single call. bug 10150 */
2615 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2616 crattr->cra_oa = &body->oa;
2617 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2618 cl_req_attr_set(env, osc2cl(obj), crattr);
2619 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2621 aa = ptlrpc_req_async_args(aa, req);
2622 INIT_LIST_HEAD(&aa->aa_oaps);
2623 list_splice_init(&rpc_list, &aa->aa_oaps);
2624 INIT_LIST_HEAD(&aa->aa_exts);
2625 list_splice_init(ext_list, &aa->aa_exts);
2627 spin_lock(&cli->cl_loi_list_lock);
2628 starting_offset >>= PAGE_SHIFT;
2629 if (cmd == OBD_BRW_READ) {
2630 cli->cl_r_in_flight++;
2631 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2632 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2633 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2634 starting_offset + 1);
2636 cli->cl_w_in_flight++;
2637 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2638 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2639 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2640 starting_offset + 1);
2642 spin_unlock(&cli->cl_loi_list_lock);
2644 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2645 page_count, aa, cli->cl_r_in_flight,
2646 cli->cl_w_in_flight);
2647 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2649 ptlrpcd_add_req(req);
2655 memalloc_noreclaim_restore(mpflag);
2658 LASSERT(req == NULL);
2661 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2663 osc_release_bounce_pages(pga, page_count);
2664 osc_release_ppga(pga, page_count);
2666 /* this should happen rarely and is pretty bad, it makes the
2667 * pending list not follow the dirty order
2669 while ((ext = list_first_entry_or_null(ext_list,
2671 oe_link)) != NULL) {
2672 list_del_init(&ext->oe_link);
2673 osc_extent_finish(env, ext, 0, rc);
2679 /* This is to refresh our lock in face of no RPCs. */
2680 void osc_send_empty_rpc(struct osc_object *osc, pgoff_t start)
2682 struct ptlrpc_request *req;
2684 struct brw_page bpg = { .off = start, .count = 1};
2685 struct brw_page *pga = &bpg;
2688 memset(&oa, 0, sizeof(oa));
2689 oa.o_oi = osc->oo_oinfo->loi_oi;
2690 oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
2691 /* For updated servers - don't do a read */
2692 oa.o_flags = OBD_FL_NORPC;
2694 rc = osc_brw_prep_request(OBD_BRW_READ, osc_cli(osc), &oa, 1, &pga,
2697 /* If we succeeded we ship it off, if not there's no point in doing
2698 * anything. Also no resends.
2699 * No interpret callback, no commit callback.
2702 req->rq_no_resend = 1;
2703 ptlrpcd_add_req(req);
2707 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2711 LASSERT(lock != NULL);
2713 lock_res_and_lock(lock);
2715 if (lock->l_ast_data == NULL)
2716 lock->l_ast_data = data;
2717 if (lock->l_ast_data == data)
2720 unlock_res_and_lock(lock);
2725 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2726 void *cookie, struct lustre_handle *lockh,
2727 enum ldlm_mode mode, __u64 *flags, bool speculative,
2730 bool intent = *flags & LDLM_FL_HAS_INTENT;
2734 /* The request was created before ldlm_cli_enqueue call. */
2735 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2736 struct ldlm_reply *rep;
2738 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2739 LASSERT(rep != NULL);
2741 rep->lock_policy_res1 =
2742 ptlrpc_status_ntoh(rep->lock_policy_res1);
2743 if (rep->lock_policy_res1)
2744 errcode = rep->lock_policy_res1;
2746 *flags |= LDLM_FL_LVB_READY;
2747 } else if (errcode == ELDLM_OK) {
2748 *flags |= LDLM_FL_LVB_READY;
2751 /* Call the update callback. */
2752 rc = (*upcall)(cookie, lockh, errcode);
2754 /* release the reference taken in ldlm_cli_enqueue() */
2755 if (errcode == ELDLM_LOCK_MATCHED)
2757 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2758 ldlm_lock_decref(lockh, mode);
2763 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2766 struct osc_enqueue_args *aa = args;
2767 struct ldlm_lock *lock;
2768 struct lustre_handle *lockh = &aa->oa_lockh;
2769 enum ldlm_mode mode = aa->oa_mode;
2770 struct ost_lvb *lvb = aa->oa_lvb;
2771 __u32 lvb_len = sizeof(*lvb);
2773 struct ldlm_enqueue_info einfo = {
2774 .ei_type = aa->oa_type,
2780 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2782 lock = ldlm_handle2lock(lockh);
2783 LASSERTF(lock != NULL,
2784 "lockh %#llx, req %p, aa %p - client evicted?\n",
2785 lockh->cookie, req, aa);
2787 /* Take an additional reference so that a blocking AST that
2788 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2789 * to arrive after an upcall has been executed by
2790 * osc_enqueue_fini(). */
2791 ldlm_lock_addref(lockh, mode);
2793 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2794 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2796 /* Let CP AST to grant the lock first. */
2797 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2799 if (aa->oa_speculative) {
2800 LASSERT(aa->oa_lvb == NULL);
2801 LASSERT(aa->oa_flags == NULL);
2802 aa->oa_flags = &flags;
2805 /* Complete obtaining the lock procedure. */
2806 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, &einfo, 1, aa->oa_flags,
2807 lvb, lvb_len, lockh, rc);
2808 /* Complete osc stuff. */
2809 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2810 aa->oa_flags, aa->oa_speculative, rc);
2812 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2814 ldlm_lock_decref(lockh, mode);
2815 LDLM_LOCK_PUT(lock);
2819 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2820 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2821 * other synchronous requests, however keeping some locks and trying to obtain
2822 * others may take a considerable amount of time in a case of ost failure; and
2823 * when other sync requests do not get released lock from a client, the client
2824 * is evicted from the cluster -- such scenarious make the life difficult, so
2825 * release locks just after they are obtained. */
2826 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2827 __u64 *flags, union ldlm_policy_data *policy,
2828 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2829 void *cookie, struct ldlm_enqueue_info *einfo,
2830 struct ptlrpc_request_set *rqset, int async,
2833 struct obd_device *obd = exp->exp_obd;
2834 struct lustre_handle lockh = { 0 };
2835 struct ptlrpc_request *req = NULL;
2836 int intent = *flags & LDLM_FL_HAS_INTENT;
2837 __u64 match_flags = *flags;
2838 enum ldlm_mode mode;
2842 /* Filesystem lock extents are extended to page boundaries so that
2843 * dealing with the page cache is a little smoother. */
2844 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2845 policy->l_extent.end |= ~PAGE_MASK;
2847 /* Next, search for already existing extent locks that will cover us */
2848 /* If we're trying to read, we also search for an existing PW lock. The
2849 * VFS and page cache already protect us locally, so lots of readers/
2850 * writers can share a single PW lock.
2852 * There are problems with conversion deadlocks, so instead of
2853 * converting a read lock to a write lock, we'll just enqueue a new
2856 * At some point we should cancel the read lock instead of making them
2857 * send us a blocking callback, but there are problems with canceling
2858 * locks out from other users right now, too. */
2859 mode = einfo->ei_mode;
2860 if (einfo->ei_mode == LCK_PR)
2862 /* Normal lock requests must wait for the LVB to be ready before
2863 * matching a lock; speculative lock requests do not need to,
2864 * because they will not actually use the lock. */
2866 match_flags |= LDLM_FL_LVB_READY;
2868 match_flags |= LDLM_FL_BLOCK_GRANTED;
2869 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2870 einfo->ei_type, policy, mode, &lockh);
2872 struct ldlm_lock *matched;
2874 if (*flags & LDLM_FL_TEST_LOCK)
2877 matched = ldlm_handle2lock(&lockh);
2879 /* This DLM lock request is speculative, and does not
2880 * have an associated IO request. Therefore if there
2881 * is already a DLM lock, it wll just inform the
2882 * caller to cancel the request for this stripe.*/
2883 lock_res_and_lock(matched);
2884 if (ldlm_extent_equal(&policy->l_extent,
2885 &matched->l_policy_data.l_extent))
2889 unlock_res_and_lock(matched);
2891 ldlm_lock_decref(&lockh, mode);
2892 LDLM_LOCK_PUT(matched);
2894 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2895 *flags |= LDLM_FL_LVB_READY;
2897 /* We already have a lock, and it's referenced. */
2898 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2900 ldlm_lock_decref(&lockh, mode);
2901 LDLM_LOCK_PUT(matched);
2904 ldlm_lock_decref(&lockh, mode);
2905 LDLM_LOCK_PUT(matched);
2909 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2912 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2913 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2915 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2916 sizeof(*lvb), LVB_T_OST, &lockh, async);
2919 struct osc_enqueue_args *aa;
2920 aa = ptlrpc_req_async_args(aa, req);
2922 aa->oa_mode = einfo->ei_mode;
2923 aa->oa_type = einfo->ei_type;
2924 lustre_handle_copy(&aa->oa_lockh, &lockh);
2925 aa->oa_upcall = upcall;
2926 aa->oa_cookie = cookie;
2927 aa->oa_speculative = speculative;
2929 aa->oa_flags = flags;
2932 /* speculative locks are essentially to enqueue
2933 * a DLM lock in advance, so we don't care
2934 * about the result of the enqueue. */
2936 aa->oa_flags = NULL;
2939 req->rq_interpret_reply = osc_enqueue_interpret;
2940 ptlrpc_set_add_req(rqset, req);
2945 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2946 flags, speculative, rc);
2951 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2952 struct ldlm_res_id *res_id, enum ldlm_type type,
2953 union ldlm_policy_data *policy, enum ldlm_mode mode,
2954 __u64 *flags, struct osc_object *obj,
2955 struct lustre_handle *lockh, enum ldlm_match_flags match_flags)
2957 struct obd_device *obd = exp->exp_obd;
2958 __u64 lflags = *flags;
2962 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2965 /* Filesystem lock extents are extended to page boundaries so that
2966 * dealing with the page cache is a little smoother */
2967 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2968 policy->l_extent.end |= ~PAGE_MASK;
2970 /* Next, search for already existing extent locks that will cover us */
2971 rc = ldlm_lock_match_with_skip(obd->obd_namespace, lflags, 0,
2972 res_id, type, policy, mode, lockh,
2974 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2978 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2980 LASSERT(lock != NULL);
2981 if (osc_set_lock_data(lock, obj)) {
2982 lock_res_and_lock(lock);
2983 if (!ldlm_is_lvb_cached(lock)) {
2984 LASSERT(lock->l_ast_data == obj);
2985 osc_lock_lvb_update(env, obj, lock, NULL);
2986 ldlm_set_lvb_cached(lock);
2988 unlock_res_and_lock(lock);
2990 ldlm_lock_decref(lockh, rc);
2993 LDLM_LOCK_PUT(lock);
2998 static int osc_statfs_interpret(const struct lu_env *env,
2999 struct ptlrpc_request *req, void *args, int rc)
3001 struct osc_async_args *aa = args;
3002 struct obd_statfs *msfs;
3007 * The request has in fact never been sent due to issues at
3008 * a higher level (LOV). Exit immediately since the caller
3009 * is aware of the problem and takes care of the clean up.
3013 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3014 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3020 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3022 GOTO(out, rc = -EPROTO);
3024 *aa->aa_oi->oi_osfs = *msfs;
3026 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3031 static int osc_statfs_async(struct obd_export *exp,
3032 struct obd_info *oinfo, time64_t max_age,
3033 struct ptlrpc_request_set *rqset)
3035 struct obd_device *obd = class_exp2obd(exp);
3036 struct ptlrpc_request *req;
3037 struct osc_async_args *aa;
3041 if (obd->obd_osfs_age >= max_age) {
3043 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
3044 obd->obd_name, &obd->obd_osfs,
3045 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
3046 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
3047 spin_lock(&obd->obd_osfs_lock);
3048 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
3049 spin_unlock(&obd->obd_osfs_lock);
3050 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
3051 if (oinfo->oi_cb_up)
3052 oinfo->oi_cb_up(oinfo, 0);
3057 /* We could possibly pass max_age in the request (as an absolute
3058 * timestamp or a "seconds.usec ago") so the target can avoid doing
3059 * extra calls into the filesystem if that isn't necessary (e.g.
3060 * during mount that would help a bit). Having relative timestamps
3061 * is not so great if request processing is slow, while absolute
3062 * timestamps are not ideal because they need time synchronization. */
3063 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3067 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3069 ptlrpc_request_free(req);
3072 ptlrpc_request_set_replen(req);
3073 req->rq_request_portal = OST_CREATE_PORTAL;
3074 ptlrpc_at_set_req_timeout(req);
3076 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3077 /* procfs requests not want stat in wait for avoid deadlock */
3078 req->rq_no_resend = 1;
3079 req->rq_no_delay = 1;
3082 req->rq_interpret_reply = osc_statfs_interpret;
3083 aa = ptlrpc_req_async_args(aa, req);
3086 ptlrpc_set_add_req(rqset, req);
3090 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
3091 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
3093 struct obd_device *obd = class_exp2obd(exp);
3094 struct obd_statfs *msfs;
3095 struct ptlrpc_request *req;
3096 struct obd_import *imp, *imp0;
3100 /*Since the request might also come from lprocfs, so we need
3101 *sync this with client_disconnect_export Bug15684
3103 with_imp_locked(obd, imp0, rc)
3104 imp = class_import_get(imp0);
3108 /* We could possibly pass max_age in the request (as an absolute
3109 * timestamp or a "seconds.usec ago") so the target can avoid doing
3110 * extra calls into the filesystem if that isn't necessary (e.g.
3111 * during mount that would help a bit). Having relative timestamps
3112 * is not so great if request processing is slow, while absolute
3113 * timestamps are not ideal because they need time synchronization. */
3114 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3116 class_import_put(imp);
3121 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3123 ptlrpc_request_free(req);
3126 ptlrpc_request_set_replen(req);
3127 req->rq_request_portal = OST_CREATE_PORTAL;
3128 ptlrpc_at_set_req_timeout(req);
3130 if (flags & OBD_STATFS_NODELAY) {
3131 /* procfs requests not want stat in wait for avoid deadlock */
3132 req->rq_no_resend = 1;
3133 req->rq_no_delay = 1;
3136 rc = ptlrpc_queue_wait(req);
3140 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3142 GOTO(out, rc = -EPROTO);
3148 ptlrpc_req_finished(req);
3152 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3153 void *karg, void __user *uarg)
3155 struct obd_device *obd = exp->exp_obd;
3156 struct obd_ioctl_data *data = karg;
3160 if (!try_module_get(THIS_MODULE)) {
3161 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3162 module_name(THIS_MODULE));
3166 case OBD_IOC_CLIENT_RECOVER:
3167 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3168 data->ioc_inlbuf1, 0);
3172 case IOC_OSC_SET_ACTIVE:
3173 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3178 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
3179 obd->obd_name, cmd, current->comm, rc);
3183 module_put(THIS_MODULE);
3187 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3188 u32 keylen, void *key, u32 vallen, void *val,
3189 struct ptlrpc_request_set *set)
3191 struct ptlrpc_request *req;
3192 struct obd_device *obd = exp->exp_obd;
3193 struct obd_import *imp = class_exp2cliimp(exp);
3198 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3200 if (KEY_IS(KEY_CHECKSUM)) {
3201 if (vallen != sizeof(int))
3203 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3207 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3208 sptlrpc_conf_client_adapt(obd);
3212 if (KEY_IS(KEY_FLUSH_CTX)) {
3213 sptlrpc_import_flush_my_ctx(imp);
3217 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3218 struct client_obd *cli = &obd->u.cli;
3219 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3220 long target = *(long *)val;
3222 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3227 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3230 /* We pass all other commands directly to OST. Since nobody calls osc
3231 methods directly and everybody is supposed to go through LOV, we
3232 assume lov checked invalid values for us.
3233 The only recognised values so far are evict_by_nid and mds_conn.
3234 Even if something bad goes through, we'd get a -EINVAL from OST
3237 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3238 &RQF_OST_SET_GRANT_INFO :
3243 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3244 RCL_CLIENT, keylen);
3245 if (!KEY_IS(KEY_GRANT_SHRINK))
3246 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3247 RCL_CLIENT, vallen);
3248 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3250 ptlrpc_request_free(req);
3254 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3255 memcpy(tmp, key, keylen);
3256 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3259 memcpy(tmp, val, vallen);
3261 if (KEY_IS(KEY_GRANT_SHRINK)) {
3262 struct osc_grant_args *aa;
3265 aa = ptlrpc_req_async_args(aa, req);
3266 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3268 ptlrpc_req_finished(req);
3271 *oa = ((struct ost_body *)val)->oa;
3273 req->rq_interpret_reply = osc_shrink_grant_interpret;
3276 ptlrpc_request_set_replen(req);
3277 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3278 LASSERT(set != NULL);
3279 ptlrpc_set_add_req(set, req);
3280 ptlrpc_check_set(NULL, set);
3282 ptlrpcd_add_req(req);
3287 EXPORT_SYMBOL(osc_set_info_async);
3289 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3290 struct obd_device *obd, struct obd_uuid *cluuid,
3291 struct obd_connect_data *data, void *localdata)
3293 struct client_obd *cli = &obd->u.cli;
3295 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3299 spin_lock(&cli->cl_loi_list_lock);
3300 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3301 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3302 /* restore ocd_grant_blkbits as client page bits */
3303 data->ocd_grant_blkbits = PAGE_SHIFT;
3304 grant += cli->cl_dirty_grant;
3306 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3308 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3309 lost_grant = cli->cl_lost_grant;
3310 cli->cl_lost_grant = 0;
3311 spin_unlock(&cli->cl_loi_list_lock);
3313 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3314 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3315 data->ocd_version, data->ocd_grant, lost_grant);
3320 EXPORT_SYMBOL(osc_reconnect);
3322 int osc_disconnect(struct obd_export *exp)
3324 struct obd_device *obd = class_exp2obd(exp);
3327 rc = client_disconnect_export(exp);
3329 * Initially we put del_shrink_grant before disconnect_export, but it
3330 * causes the following problem if setup (connect) and cleanup
3331 * (disconnect) are tangled together.
3332 * connect p1 disconnect p2
3333 * ptlrpc_connect_import
3334 * ............... class_manual_cleanup
3337 * ptlrpc_connect_interrupt
3339 * add this client to shrink list
3341 * Bang! grant shrink thread trigger the shrink. BUG18662
3343 osc_del_grant_list(&obd->u.cli);
3346 EXPORT_SYMBOL(osc_disconnect);
3348 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3349 struct hlist_node *hnode, void *arg)
3351 struct lu_env *env = arg;
3352 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3353 struct ldlm_lock *lock;
3354 struct osc_object *osc = NULL;
3358 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3359 if (lock->l_ast_data != NULL && osc == NULL) {
3360 osc = lock->l_ast_data;
3361 cl_object_get(osc2cl(osc));
3364 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3365 * by the 2nd round of ldlm_namespace_clean() call in
3366 * osc_import_event(). */
3367 ldlm_clear_cleaned(lock);
3372 osc_object_invalidate(env, osc);
3373 cl_object_put(env, osc2cl(osc));
3378 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3380 static int osc_import_event(struct obd_device *obd,
3381 struct obd_import *imp,
3382 enum obd_import_event event)
3384 struct client_obd *cli;
3388 LASSERT(imp->imp_obd == obd);
3391 case IMP_EVENT_DISCON: {
3393 spin_lock(&cli->cl_loi_list_lock);
3394 cli->cl_avail_grant = 0;
3395 cli->cl_lost_grant = 0;
3396 spin_unlock(&cli->cl_loi_list_lock);
3399 case IMP_EVENT_INACTIVE: {
3400 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3403 case IMP_EVENT_INVALIDATE: {
3404 struct ldlm_namespace *ns = obd->obd_namespace;
3408 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3410 env = cl_env_get(&refcheck);
3412 osc_io_unplug(env, &obd->u.cli, NULL);
3414 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3415 osc_ldlm_resource_invalidate,
3417 cl_env_put(env, &refcheck);
3419 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3424 case IMP_EVENT_ACTIVE: {
3425 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3428 case IMP_EVENT_OCD: {
3429 struct obd_connect_data *ocd = &imp->imp_connect_data;
3431 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3432 osc_init_grant(&obd->u.cli, ocd);
3435 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3436 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3438 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3441 case IMP_EVENT_DEACTIVATE: {
3442 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3445 case IMP_EVENT_ACTIVATE: {
3446 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3450 CERROR("Unknown import event %d\n", event);
3457 * Determine whether the lock can be canceled before replaying the lock
3458 * during recovery, see bug16774 for detailed information.
3460 * \retval zero the lock can't be canceled
3461 * \retval other ok to cancel
3463 static int osc_cancel_weight(struct ldlm_lock *lock)
3466 * Cancel all unused and granted extent lock.
3468 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3469 ldlm_is_granted(lock) &&
3470 osc_ldlm_weigh_ast(lock) == 0)
3476 static int brw_queue_work(const struct lu_env *env, void *data)
3478 struct client_obd *cli = data;
3480 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3482 osc_io_unplug(env, cli, NULL);
3486 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3488 struct client_obd *cli = &obd->u.cli;
3494 rc = ptlrpcd_addref();
3498 rc = client_obd_setup(obd, lcfg);
3500 GOTO(out_ptlrpcd, rc);
3503 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3504 if (IS_ERR(handler))
3505 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3506 cli->cl_writeback_work = handler;
3508 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3509 if (IS_ERR(handler))
3510 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3511 cli->cl_lru_work = handler;
3513 rc = osc_quota_setup(obd);
3515 GOTO(out_ptlrpcd_work, rc);
3517 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3518 cli->cl_root_squash = 0;
3519 osc_update_next_shrink(cli);
3524 if (cli->cl_writeback_work != NULL) {
3525 ptlrpcd_destroy_work(cli->cl_writeback_work);
3526 cli->cl_writeback_work = NULL;
3528 if (cli->cl_lru_work != NULL) {
3529 ptlrpcd_destroy_work(cli->cl_lru_work);
3530 cli->cl_lru_work = NULL;
3532 client_obd_cleanup(obd);
3537 EXPORT_SYMBOL(osc_setup_common);
3539 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3541 struct client_obd *cli = &obd->u.cli;
3549 rc = osc_setup_common(obd, lcfg);
3553 rc = osc_tunables_init(obd);
3558 * We try to control the total number of requests with a upper limit
3559 * osc_reqpool_maxreqcount. There might be some race which will cause
3560 * over-limit allocation, but it is fine.
3562 req_count = atomic_read(&osc_pool_req_count);
3563 if (req_count < osc_reqpool_maxreqcount) {
3564 adding = cli->cl_max_rpcs_in_flight + 2;
3565 if (req_count + adding > osc_reqpool_maxreqcount)
3566 adding = osc_reqpool_maxreqcount - req_count;
3568 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3569 atomic_add(added, &osc_pool_req_count);
3572 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3574 spin_lock(&osc_shrink_lock);
3575 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3576 spin_unlock(&osc_shrink_lock);
3577 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3578 cli->cl_import->imp_idle_debug = D_HA;
3583 int osc_precleanup_common(struct obd_device *obd)
3585 struct client_obd *cli = &obd->u.cli;
3589 * for echo client, export may be on zombie list, wait for
3590 * zombie thread to cull it, because cli.cl_import will be
3591 * cleared in client_disconnect_export():
3592 * class_export_destroy() -> obd_cleanup() ->
3593 * echo_device_free() -> echo_client_cleanup() ->
3594 * obd_disconnect() -> osc_disconnect() ->
3595 * client_disconnect_export()
3597 obd_zombie_barrier();
3598 if (cli->cl_writeback_work) {
3599 ptlrpcd_destroy_work(cli->cl_writeback_work);
3600 cli->cl_writeback_work = NULL;
3603 if (cli->cl_lru_work) {
3604 ptlrpcd_destroy_work(cli->cl_lru_work);
3605 cli->cl_lru_work = NULL;
3608 obd_cleanup_client_import(obd);
3611 EXPORT_SYMBOL(osc_precleanup_common);
3613 static int osc_precleanup(struct obd_device *obd)
3617 osc_precleanup_common(obd);
3619 ptlrpc_lprocfs_unregister_obd(obd);
3623 int osc_cleanup_common(struct obd_device *obd)
3625 struct client_obd *cli = &obd->u.cli;
3630 spin_lock(&osc_shrink_lock);
3631 list_del(&cli->cl_shrink_list);
3632 spin_unlock(&osc_shrink_lock);
3635 if (cli->cl_cache != NULL) {
3636 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3637 spin_lock(&cli->cl_cache->ccc_lru_lock);
3638 list_del_init(&cli->cl_lru_osc);
3639 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3640 cli->cl_lru_left = NULL;
3641 cl_cache_decref(cli->cl_cache);
3642 cli->cl_cache = NULL;
3645 /* free memory of osc quota cache */
3646 osc_quota_cleanup(obd);
3648 rc = client_obd_cleanup(obd);
3653 EXPORT_SYMBOL(osc_cleanup_common);
3655 static const struct obd_ops osc_obd_ops = {
3656 .o_owner = THIS_MODULE,
3657 .o_setup = osc_setup,
3658 .o_precleanup = osc_precleanup,
3659 .o_cleanup = osc_cleanup_common,
3660 .o_add_conn = client_import_add_conn,
3661 .o_del_conn = client_import_del_conn,
3662 .o_connect = client_connect_import,
3663 .o_reconnect = osc_reconnect,
3664 .o_disconnect = osc_disconnect,
3665 .o_statfs = osc_statfs,
3666 .o_statfs_async = osc_statfs_async,
3667 .o_create = osc_create,
3668 .o_destroy = osc_destroy,
3669 .o_getattr = osc_getattr,
3670 .o_setattr = osc_setattr,
3671 .o_iocontrol = osc_iocontrol,
3672 .o_set_info_async = osc_set_info_async,
3673 .o_import_event = osc_import_event,
3674 .o_quotactl = osc_quotactl,
3677 LIST_HEAD(osc_shrink_list);
3678 DEFINE_SPINLOCK(osc_shrink_lock);
3680 #ifdef HAVE_SHRINKER_COUNT
3681 static struct shrinker osc_cache_shrinker = {
3682 .count_objects = osc_cache_shrink_count,
3683 .scan_objects = osc_cache_shrink_scan,
3684 .seeks = DEFAULT_SEEKS,
3687 static int osc_cache_shrink(struct shrinker *shrinker,
3688 struct shrink_control *sc)
3690 (void)osc_cache_shrink_scan(shrinker, sc);
3692 return osc_cache_shrink_count(shrinker, sc);
3695 static struct shrinker osc_cache_shrinker = {
3696 .shrink = osc_cache_shrink,
3697 .seeks = DEFAULT_SEEKS,
3701 static int __init osc_init(void)
3703 unsigned int reqpool_size;
3704 unsigned int reqsize;
3708 /* print an address of _any_ initialized kernel symbol from this
3709 * module, to allow debugging with gdb that doesn't support data
3710 * symbols from modules.*/
3711 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3713 rc = lu_kmem_init(osc_caches);
3717 rc = class_register_type(&osc_obd_ops, NULL, true,
3718 LUSTRE_OSC_NAME, &osc_device_type);
3722 rc = register_shrinker(&osc_cache_shrinker);
3726 /* This is obviously too much memory, only prevent overflow here */
3727 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3728 GOTO(out_shrinker, rc = -EINVAL);
3730 reqpool_size = osc_reqpool_mem_max << 20;
3733 while (reqsize < OST_IO_MAXREQSIZE)
3734 reqsize = reqsize << 1;
3737 * We don't enlarge the request count in OSC pool according to
3738 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3739 * tried after normal allocation failed. So a small OSC pool won't
3740 * cause much performance degression in most of cases.
3742 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3744 atomic_set(&osc_pool_req_count, 0);
3745 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3746 ptlrpc_add_rqs_to_pool);
3748 if (osc_rq_pool == NULL)
3749 GOTO(out_shrinker, rc = -ENOMEM);
3751 rc = osc_start_grant_work();
3753 GOTO(out_req_pool, rc);
3758 ptlrpc_free_rq_pool(osc_rq_pool);
3760 unregister_shrinker(&osc_cache_shrinker);
3762 class_unregister_type(LUSTRE_OSC_NAME);
3764 lu_kmem_fini(osc_caches);
3769 static void __exit osc_exit(void)
3771 osc_stop_grant_work();
3772 unregister_shrinker(&osc_cache_shrinker);
3773 class_unregister_type(LUSTRE_OSC_NAME);
3774 lu_kmem_fini(osc_caches);
3775 ptlrpc_free_rq_pool(osc_rq_pool);
3778 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3779 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3780 MODULE_VERSION(LUSTRE_VERSION_STRING);
3781 MODULE_LICENSE("GPL");
3783 module_init(osc_init);
3784 module_exit(osc_exit);