4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_OSC
35 #include <linux/workqueue.h>
36 #include <lprocfs_status.h>
37 #include <lustre_debug.h>
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <lustre_ha.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
42 #include <lustre_net.h>
43 #include <lustre_obdo.h>
45 #include <obd_cksum.h>
46 #include <obd_class.h>
47 #include <lustre_osc.h>
49 #include "osc_internal.h"
51 atomic_t osc_pool_req_count;
52 unsigned int osc_reqpool_maxreqcount;
53 struct ptlrpc_request_pool *osc_rq_pool;
55 /* max memory used for request pool, unit is MB */
56 static unsigned int osc_reqpool_mem_max = 5;
57 module_param(osc_reqpool_mem_max, uint, 0444);
59 static int osc_idle_timeout = 20;
60 module_param(osc_idle_timeout, uint, 0644);
62 #define osc_grant_args osc_brw_async_args
64 struct osc_setattr_args {
66 obd_enqueue_update_f sa_upcall;
70 struct osc_fsync_args {
71 struct osc_object *fa_obj;
73 obd_enqueue_update_f fa_upcall;
77 struct osc_ladvise_args {
79 obd_enqueue_update_f la_upcall;
83 static void osc_release_ppga(struct brw_page **ppga, size_t count);
84 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
87 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
89 struct ost_body *body;
91 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
94 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
97 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
100 struct ptlrpc_request *req;
101 struct ost_body *body;
105 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
109 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
111 ptlrpc_request_free(req);
115 osc_pack_req_body(req, oa);
117 ptlrpc_request_set_replen(req);
119 rc = ptlrpc_queue_wait(req);
123 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
125 GOTO(out, rc = -EPROTO);
127 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
128 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
130 oa->o_blksize = cli_brw_size(exp->exp_obd);
131 oa->o_valid |= OBD_MD_FLBLKSZ;
135 ptlrpc_req_finished(req);
140 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
143 struct ptlrpc_request *req;
144 struct ost_body *body;
148 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
150 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
154 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
156 ptlrpc_request_free(req);
160 osc_pack_req_body(req, oa);
162 ptlrpc_request_set_replen(req);
164 rc = ptlrpc_queue_wait(req);
168 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
170 GOTO(out, rc = -EPROTO);
172 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
176 ptlrpc_req_finished(req);
181 static int osc_setattr_interpret(const struct lu_env *env,
182 struct ptlrpc_request *req, void *args, int rc)
184 struct osc_setattr_args *sa = args;
185 struct ost_body *body;
192 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
194 GOTO(out, rc = -EPROTO);
196 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
199 rc = sa->sa_upcall(sa->sa_cookie, rc);
203 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
204 obd_enqueue_update_f upcall, void *cookie,
205 struct ptlrpc_request_set *rqset)
207 struct ptlrpc_request *req;
208 struct osc_setattr_args *sa;
213 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
217 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
219 ptlrpc_request_free(req);
223 osc_pack_req_body(req, oa);
225 ptlrpc_request_set_replen(req);
227 /* do mds to ost setattr asynchronously */
229 /* Do not wait for response. */
230 ptlrpcd_add_req(req);
232 req->rq_interpret_reply = osc_setattr_interpret;
234 sa = ptlrpc_req_async_args(sa, req);
236 sa->sa_upcall = upcall;
237 sa->sa_cookie = cookie;
239 if (rqset == PTLRPCD_SET)
240 ptlrpcd_add_req(req);
242 ptlrpc_set_add_req(rqset, req);
248 static int osc_ladvise_interpret(const struct lu_env *env,
249 struct ptlrpc_request *req,
252 struct osc_ladvise_args *la = arg;
253 struct ost_body *body;
259 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
261 GOTO(out, rc = -EPROTO);
263 *la->la_oa = body->oa;
265 rc = la->la_upcall(la->la_cookie, rc);
270 * If rqset is NULL, do not wait for response. Upcall and cookie could also
271 * be NULL in this case
273 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
274 struct ladvise_hdr *ladvise_hdr,
275 obd_enqueue_update_f upcall, void *cookie,
276 struct ptlrpc_request_set *rqset)
278 struct ptlrpc_request *req;
279 struct ost_body *body;
280 struct osc_ladvise_args *la;
282 struct lu_ladvise *req_ladvise;
283 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
284 int num_advise = ladvise_hdr->lah_count;
285 struct ladvise_hdr *req_ladvise_hdr;
288 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
292 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
293 num_advise * sizeof(*ladvise));
294 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
296 ptlrpc_request_free(req);
299 req->rq_request_portal = OST_IO_PORTAL;
300 ptlrpc_at_set_req_timeout(req);
302 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
304 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
307 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
308 &RMF_OST_LADVISE_HDR);
309 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
311 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
312 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
313 ptlrpc_request_set_replen(req);
316 /* Do not wait for response. */
317 ptlrpcd_add_req(req);
321 req->rq_interpret_reply = osc_ladvise_interpret;
322 la = ptlrpc_req_async_args(la, req);
324 la->la_upcall = upcall;
325 la->la_cookie = cookie;
327 if (rqset == PTLRPCD_SET)
328 ptlrpcd_add_req(req);
330 ptlrpc_set_add_req(rqset, req);
335 static int osc_create(const struct lu_env *env, struct obd_export *exp,
338 struct ptlrpc_request *req;
339 struct ost_body *body;
344 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
345 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
347 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
349 GOTO(out, rc = -ENOMEM);
351 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
353 ptlrpc_request_free(req);
357 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
360 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
362 ptlrpc_request_set_replen(req);
364 rc = ptlrpc_queue_wait(req);
368 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
370 GOTO(out_req, rc = -EPROTO);
372 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
373 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
375 oa->o_blksize = cli_brw_size(exp->exp_obd);
376 oa->o_valid |= OBD_MD_FLBLKSZ;
378 CDEBUG(D_HA, "transno: %lld\n",
379 lustre_msg_get_transno(req->rq_repmsg));
381 ptlrpc_req_finished(req);
386 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
387 obd_enqueue_update_f upcall, void *cookie)
389 struct ptlrpc_request *req;
390 struct osc_setattr_args *sa;
391 struct obd_import *imp = class_exp2cliimp(exp);
392 struct ost_body *body;
397 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
401 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
403 ptlrpc_request_free(req);
407 osc_set_io_portal(req);
409 ptlrpc_at_set_req_timeout(req);
411 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
413 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
415 ptlrpc_request_set_replen(req);
417 req->rq_interpret_reply = osc_setattr_interpret;
418 sa = ptlrpc_req_async_args(sa, req);
420 sa->sa_upcall = upcall;
421 sa->sa_cookie = cookie;
423 ptlrpcd_add_req(req);
427 EXPORT_SYMBOL(osc_punch_send);
429 static int osc_sync_interpret(const struct lu_env *env,
430 struct ptlrpc_request *req, void *args, int rc)
432 struct osc_fsync_args *fa = args;
433 struct ost_body *body;
434 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
435 unsigned long valid = 0;
436 struct cl_object *obj;
442 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
444 CERROR("can't unpack ost_body\n");
445 GOTO(out, rc = -EPROTO);
448 *fa->fa_oa = body->oa;
449 obj = osc2cl(fa->fa_obj);
451 /* Update osc object's blocks attribute */
452 cl_object_attr_lock(obj);
453 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
454 attr->cat_blocks = body->oa.o_blocks;
459 cl_object_attr_update(env, obj, attr, valid);
460 cl_object_attr_unlock(obj);
463 rc = fa->fa_upcall(fa->fa_cookie, rc);
467 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
468 obd_enqueue_update_f upcall, void *cookie,
469 struct ptlrpc_request_set *rqset)
471 struct obd_export *exp = osc_export(obj);
472 struct ptlrpc_request *req;
473 struct ost_body *body;
474 struct osc_fsync_args *fa;
478 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
482 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
484 ptlrpc_request_free(req);
488 /* overload the size and blocks fields in the oa with start/end */
489 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
491 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
493 ptlrpc_request_set_replen(req);
494 req->rq_interpret_reply = osc_sync_interpret;
496 fa = ptlrpc_req_async_args(fa, req);
499 fa->fa_upcall = upcall;
500 fa->fa_cookie = cookie;
502 if (rqset == PTLRPCD_SET)
503 ptlrpcd_add_req(req);
505 ptlrpc_set_add_req(rqset, req);
510 /* Find and cancel locally locks matched by @mode in the resource found by
511 * @objid. Found locks are added into @cancel list. Returns the amount of
512 * locks added to @cancels list. */
513 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
514 struct list_head *cancels,
515 enum ldlm_mode mode, __u64 lock_flags)
517 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
518 struct ldlm_res_id res_id;
519 struct ldlm_resource *res;
523 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
524 * export) but disabled through procfs (flag in NS).
526 * This distinguishes from a case when ELC is not supported originally,
527 * when we still want to cancel locks in advance and just cancel them
528 * locally, without sending any RPC. */
529 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
532 ostid_build_res_name(&oa->o_oi, &res_id);
533 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
537 LDLM_RESOURCE_ADDREF(res);
538 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
539 lock_flags, 0, NULL);
540 LDLM_RESOURCE_DELREF(res);
541 ldlm_resource_putref(res);
545 static int osc_destroy_interpret(const struct lu_env *env,
546 struct ptlrpc_request *req, void *args, int rc)
548 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
550 atomic_dec(&cli->cl_destroy_in_flight);
551 wake_up(&cli->cl_destroy_waitq);
556 static int osc_can_send_destroy(struct client_obd *cli)
558 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
559 cli->cl_max_rpcs_in_flight) {
560 /* The destroy request can be sent */
563 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
564 cli->cl_max_rpcs_in_flight) {
566 * The counter has been modified between the two atomic
569 wake_up(&cli->cl_destroy_waitq);
574 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
577 struct client_obd *cli = &exp->exp_obd->u.cli;
578 struct ptlrpc_request *req;
579 struct ost_body *body;
580 struct list_head cancels = LIST_HEAD_INIT(cancels);
585 CDEBUG(D_INFO, "oa NULL\n");
589 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
590 LDLM_FL_DISCARD_DATA);
592 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
594 ldlm_lock_list_put(&cancels, l_bl_ast, count);
598 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
601 ptlrpc_request_free(req);
605 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
606 ptlrpc_at_set_req_timeout(req);
608 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
610 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
612 ptlrpc_request_set_replen(req);
614 req->rq_interpret_reply = osc_destroy_interpret;
615 if (!osc_can_send_destroy(cli)) {
616 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
619 * Wait until the number of on-going destroy RPCs drops
620 * under max_rpc_in_flight
622 rc = l_wait_event_exclusive(cli->cl_destroy_waitq,
623 osc_can_send_destroy(cli), &lwi);
625 ptlrpc_req_finished(req);
630 /* Do not wait for response */
631 ptlrpcd_add_req(req);
635 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
638 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
640 LASSERT(!(oa->o_valid & bits));
643 spin_lock(&cli->cl_loi_list_lock);
644 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
645 oa->o_dirty = cli->cl_dirty_grant;
647 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
648 if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
649 cli->cl_dirty_max_pages)) {
650 CERROR("dirty %lu - %lu > dirty_max %lu\n",
651 cli->cl_dirty_pages, cli->cl_dirty_transit,
652 cli->cl_dirty_max_pages);
654 } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
655 atomic_long_read(&obd_dirty_transit_pages) >
656 (long)(obd_max_dirty_pages + 1))) {
657 /* The atomic_read() allowing the atomic_inc() are
658 * not covered by a lock thus they may safely race and trip
659 * this CERROR() unless we add in a small fudge factor (+1). */
660 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
661 cli_name(cli), atomic_long_read(&obd_dirty_pages),
662 atomic_long_read(&obd_dirty_transit_pages),
663 obd_max_dirty_pages);
665 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
667 CERROR("dirty %lu - dirty_max %lu too big???\n",
668 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
671 unsigned long nrpages;
672 unsigned long undirty;
674 nrpages = cli->cl_max_pages_per_rpc;
675 nrpages *= cli->cl_max_rpcs_in_flight + 1;
676 nrpages = max(nrpages, cli->cl_dirty_max_pages);
677 undirty = nrpages << PAGE_SHIFT;
678 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
682 /* take extent tax into account when asking for more
684 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
685 cli->cl_max_extent_pages;
686 undirty += nrextents * cli->cl_grant_extent_tax;
688 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
689 * to add extent tax, etc.
691 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
692 ~(PTLRPC_MAX_BRW_SIZE * 4UL));
694 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
695 oa->o_dropped = cli->cl_lost_grant;
696 cli->cl_lost_grant = 0;
697 spin_unlock(&cli->cl_loi_list_lock);
698 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
699 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
702 void osc_update_next_shrink(struct client_obd *cli)
704 cli->cl_next_shrink_grant = ktime_get_seconds() +
705 cli->cl_grant_shrink_interval;
707 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
708 cli->cl_next_shrink_grant);
711 static void __osc_update_grant(struct client_obd *cli, u64 grant)
713 spin_lock(&cli->cl_loi_list_lock);
714 cli->cl_avail_grant += grant;
715 spin_unlock(&cli->cl_loi_list_lock);
718 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
720 if (body->oa.o_valid & OBD_MD_FLGRANT) {
721 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
722 __osc_update_grant(cli, body->oa.o_grant);
727 * grant thread data for shrinking space.
729 struct grant_thread_data {
730 struct list_head gtd_clients;
731 struct mutex gtd_mutex;
732 unsigned long gtd_stopped:1;
734 static struct grant_thread_data client_gtd;
736 static int osc_shrink_grant_interpret(const struct lu_env *env,
737 struct ptlrpc_request *req,
740 struct osc_grant_args *aa = args;
741 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
742 struct ost_body *body;
745 __osc_update_grant(cli, aa->aa_oa->o_grant);
749 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
751 osc_update_grant(cli, body);
753 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
758 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
760 spin_lock(&cli->cl_loi_list_lock);
761 oa->o_grant = cli->cl_avail_grant / 4;
762 cli->cl_avail_grant -= oa->o_grant;
763 spin_unlock(&cli->cl_loi_list_lock);
764 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
765 oa->o_valid |= OBD_MD_FLFLAGS;
768 oa->o_flags |= OBD_FL_SHRINK_GRANT;
769 osc_update_next_shrink(cli);
772 /* Shrink the current grant, either from some large amount to enough for a
773 * full set of in-flight RPCs, or if we have already shrunk to that limit
774 * then to enough for a single RPC. This avoids keeping more grant than
775 * needed, and avoids shrinking the grant piecemeal. */
776 static int osc_shrink_grant(struct client_obd *cli)
778 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
779 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
781 spin_lock(&cli->cl_loi_list_lock);
782 if (cli->cl_avail_grant <= target_bytes)
783 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
784 spin_unlock(&cli->cl_loi_list_lock);
786 return osc_shrink_grant_to_target(cli, target_bytes);
789 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
792 struct ost_body *body;
795 spin_lock(&cli->cl_loi_list_lock);
796 /* Don't shrink if we are already above or below the desired limit
797 * We don't want to shrink below a single RPC, as that will negatively
798 * impact block allocation and long-term performance. */
799 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
800 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
802 if (target_bytes >= cli->cl_avail_grant) {
803 spin_unlock(&cli->cl_loi_list_lock);
806 spin_unlock(&cli->cl_loi_list_lock);
812 osc_announce_cached(cli, &body->oa, 0);
814 spin_lock(&cli->cl_loi_list_lock);
815 if (target_bytes >= cli->cl_avail_grant) {
816 /* available grant has changed since target calculation */
817 spin_unlock(&cli->cl_loi_list_lock);
818 GOTO(out_free, rc = 0);
820 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
821 cli->cl_avail_grant = target_bytes;
822 spin_unlock(&cli->cl_loi_list_lock);
823 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
824 body->oa.o_valid |= OBD_MD_FLFLAGS;
825 body->oa.o_flags = 0;
827 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
828 osc_update_next_shrink(cli);
830 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
831 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
832 sizeof(*body), body, NULL);
834 __osc_update_grant(cli, body->oa.o_grant);
840 static int osc_should_shrink_grant(struct client_obd *client)
842 time64_t next_shrink = client->cl_next_shrink_grant;
844 if (client->cl_import == NULL)
847 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
848 OBD_CONNECT_GRANT_SHRINK) == 0)
851 if (ktime_get_seconds() >= next_shrink - 5) {
852 /* Get the current RPC size directly, instead of going via:
853 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
854 * Keep comment here so that it can be found by searching. */
855 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
857 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
858 client->cl_avail_grant > brw_size)
861 osc_update_next_shrink(client);
866 #define GRANT_SHRINK_RPC_BATCH 100
868 static struct delayed_work work;
870 static void osc_grant_work_handler(struct work_struct *data)
872 struct client_obd *cli;
874 bool init_next_shrink = true;
875 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
878 mutex_lock(&client_gtd.gtd_mutex);
879 list_for_each_entry(cli, &client_gtd.gtd_clients,
881 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
882 osc_should_shrink_grant(cli)) {
883 osc_shrink_grant(cli);
887 if (!init_next_shrink) {
888 if (cli->cl_next_shrink_grant < next_shrink &&
889 cli->cl_next_shrink_grant > ktime_get_seconds())
890 next_shrink = cli->cl_next_shrink_grant;
892 init_next_shrink = false;
893 next_shrink = cli->cl_next_shrink_grant;
896 mutex_unlock(&client_gtd.gtd_mutex);
898 if (client_gtd.gtd_stopped == 1)
901 if (next_shrink > ktime_get_seconds()) {
902 time64_t delay = next_shrink - ktime_get_seconds();
904 schedule_delayed_work(&work, cfs_time_seconds(delay));
906 schedule_work(&work.work);
910 void osc_schedule_grant_work(void)
912 cancel_delayed_work_sync(&work);
913 schedule_work(&work.work);
917 * Start grant thread for returing grant to server for idle clients.
919 static int osc_start_grant_work(void)
921 client_gtd.gtd_stopped = 0;
922 mutex_init(&client_gtd.gtd_mutex);
923 INIT_LIST_HEAD(&client_gtd.gtd_clients);
925 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
926 schedule_work(&work.work);
931 static void osc_stop_grant_work(void)
933 client_gtd.gtd_stopped = 1;
934 cancel_delayed_work_sync(&work);
937 static void osc_add_grant_list(struct client_obd *client)
939 mutex_lock(&client_gtd.gtd_mutex);
940 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
941 mutex_unlock(&client_gtd.gtd_mutex);
944 static void osc_del_grant_list(struct client_obd *client)
946 if (list_empty(&client->cl_grant_chain))
949 mutex_lock(&client_gtd.gtd_mutex);
950 list_del_init(&client->cl_grant_chain);
951 mutex_unlock(&client_gtd.gtd_mutex);
954 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
957 * ocd_grant is the total grant amount we're expect to hold: if we've
958 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
959 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
962 * race is tolerable here: if we're evicted, but imp_state already
963 * left EVICTED state, then cl_dirty_pages must be 0 already.
965 spin_lock(&cli->cl_loi_list_lock);
966 cli->cl_avail_grant = ocd->ocd_grant;
967 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
968 cli->cl_avail_grant -= cli->cl_reserved_grant;
969 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
970 cli->cl_avail_grant -= cli->cl_dirty_grant;
972 cli->cl_avail_grant -=
973 cli->cl_dirty_pages << PAGE_SHIFT;
976 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
980 /* overhead for each extent insertion */
981 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
982 /* determine the appropriate chunk size used by osc_extent. */
983 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
984 ocd->ocd_grant_blkbits);
985 /* max_pages_per_rpc must be chunk aligned */
986 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
987 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
988 ~chunk_mask) & chunk_mask;
989 /* determine maximum extent size, in #pages */
990 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
991 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
992 if (cli->cl_max_extent_pages == 0)
993 cli->cl_max_extent_pages = 1;
995 cli->cl_grant_extent_tax = 0;
996 cli->cl_chunkbits = PAGE_SHIFT;
997 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
999 spin_unlock(&cli->cl_loi_list_lock);
1001 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1002 "chunk bits: %d cl_max_extent_pages: %d\n",
1004 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1005 cli->cl_max_extent_pages);
1007 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1008 osc_add_grant_list(cli);
1010 EXPORT_SYMBOL(osc_init_grant);
1012 /* We assume that the reason this OSC got a short read is because it read
1013 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1014 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1015 * this stripe never got written at or beyond this stripe offset yet. */
1016 static void handle_short_read(int nob_read, size_t page_count,
1017 struct brw_page **pga)
1022 /* skip bytes read OK */
1023 while (nob_read > 0) {
1024 LASSERT (page_count > 0);
1026 if (pga[i]->count > nob_read) {
1027 /* EOF inside this page */
1028 ptr = kmap(pga[i]->pg) +
1029 (pga[i]->off & ~PAGE_MASK);
1030 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1037 nob_read -= pga[i]->count;
1042 /* zero remaining pages */
1043 while (page_count-- > 0) {
1044 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1045 memset(ptr, 0, pga[i]->count);
1051 static int check_write_rcs(struct ptlrpc_request *req,
1052 int requested_nob, int niocount,
1053 size_t page_count, struct brw_page **pga)
1058 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1059 sizeof(*remote_rcs) *
1061 if (remote_rcs == NULL) {
1062 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1066 /* return error if any niobuf was in error */
1067 for (i = 0; i < niocount; i++) {
1068 if ((int)remote_rcs[i] < 0) {
1069 CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1070 i, remote_rcs[i], req);
1071 return remote_rcs[i];
1074 if (remote_rcs[i] != 0) {
1075 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1076 i, remote_rcs[i], req);
1080 if (req->rq_bulk != NULL &&
1081 req->rq_bulk->bd_nob_transferred != requested_nob) {
1082 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1083 req->rq_bulk->bd_nob_transferred, requested_nob);
1090 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1092 if (p1->flag != p2->flag) {
1093 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1094 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1095 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1097 /* warn if we try to combine flags that we don't know to be
1098 * safe to combine */
1099 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1100 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1101 "report this at https://jira.whamcloud.com/\n",
1102 p1->flag, p2->flag);
1107 return (p1->off + p1->count == p2->off);
1110 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1111 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1112 size_t pg_count, struct brw_page **pga,
1113 int opc, obd_dif_csum_fn *fn,
1117 struct ahash_request *req;
1118 /* Used Adler as the default checksum type on top of DIF tags */
1119 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1120 struct page *__page;
1121 unsigned char *buffer;
1123 unsigned int bufsize;
1125 int used_number = 0;
1131 LASSERT(pg_count > 0);
1133 __page = alloc_page(GFP_KERNEL);
1137 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1140 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1141 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1145 buffer = kmap(__page);
1146 guard_start = (__u16 *)buffer;
1147 guard_number = PAGE_SIZE / sizeof(*guard_start);
1148 while (nob > 0 && pg_count > 0) {
1149 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1151 /* corrupt the data before we compute the checksum, to
1152 * simulate an OST->client data error */
1153 if (unlikely(i == 0 && opc == OST_READ &&
1154 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1155 unsigned char *ptr = kmap(pga[i]->pg);
1156 int off = pga[i]->off & ~PAGE_MASK;
1158 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1163 * The left guard number should be able to hold checksums of a
1166 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1167 pga[i]->off & ~PAGE_MASK,
1169 guard_start + used_number,
1170 guard_number - used_number,
1176 used_number += used;
1177 if (used_number == guard_number) {
1178 cfs_crypto_hash_update_page(req, __page, 0,
1179 used_number * sizeof(*guard_start));
1183 nob -= pga[i]->count;
1191 if (used_number != 0)
1192 cfs_crypto_hash_update_page(req, __page, 0,
1193 used_number * sizeof(*guard_start));
1195 bufsize = sizeof(cksum);
1196 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1198 /* For sending we only compute the wrong checksum instead
1199 * of corrupting the data so it is still correct on a redo */
1200 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1205 __free_page(__page);
1208 #else /* !CONFIG_CRC_T10DIF */
1209 #define obd_dif_ip_fn NULL
1210 #define obd_dif_crc_fn NULL
1211 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum) \
1213 #endif /* CONFIG_CRC_T10DIF */
1215 static int osc_checksum_bulk(int nob, size_t pg_count,
1216 struct brw_page **pga, int opc,
1217 enum cksum_types cksum_type,
1221 struct ahash_request *req;
1222 unsigned int bufsize;
1223 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1225 LASSERT(pg_count > 0);
1227 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1229 CERROR("Unable to initialize checksum hash %s\n",
1230 cfs_crypto_hash_name(cfs_alg));
1231 return PTR_ERR(req);
1234 while (nob > 0 && pg_count > 0) {
1235 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1237 /* corrupt the data before we compute the checksum, to
1238 * simulate an OST->client data error */
1239 if (i == 0 && opc == OST_READ &&
1240 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1241 unsigned char *ptr = kmap(pga[i]->pg);
1242 int off = pga[i]->off & ~PAGE_MASK;
1244 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1247 cfs_crypto_hash_update_page(req, pga[i]->pg,
1248 pga[i]->off & ~PAGE_MASK,
1250 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1251 (int)(pga[i]->off & ~PAGE_MASK));
1253 nob -= pga[i]->count;
1258 bufsize = sizeof(*cksum);
1259 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1261 /* For sending we only compute the wrong checksum instead
1262 * of corrupting the data so it is still correct on a redo */
1263 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1269 static int osc_checksum_bulk_rw(const char *obd_name,
1270 enum cksum_types cksum_type,
1271 int nob, size_t pg_count,
1272 struct brw_page **pga, int opc,
1275 obd_dif_csum_fn *fn = NULL;
1276 int sector_size = 0;
1280 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1283 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1284 opc, fn, sector_size, check_sum);
1286 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1293 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1294 u32 page_count, struct brw_page **pga,
1295 struct ptlrpc_request **reqp, int resend)
1297 struct ptlrpc_request *req;
1298 struct ptlrpc_bulk_desc *desc;
1299 struct ost_body *body;
1300 struct obd_ioobj *ioobj;
1301 struct niobuf_remote *niobuf;
1302 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1303 struct osc_brw_async_args *aa;
1304 struct req_capsule *pill;
1305 struct brw_page *pg_prev;
1307 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1310 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1311 RETURN(-ENOMEM); /* Recoverable */
1312 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1313 RETURN(-EINVAL); /* Fatal */
1315 if ((cmd & OBD_BRW_WRITE) != 0) {
1317 req = ptlrpc_request_alloc_pool(cli->cl_import,
1319 &RQF_OST_BRW_WRITE);
1322 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1327 for (niocount = i = 1; i < page_count; i++) {
1328 if (!can_merge_pages(pga[i - 1], pga[i]))
1332 pill = &req->rq_pill;
1333 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1335 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1336 niocount * sizeof(*niobuf));
1338 for (i = 0; i < page_count; i++)
1339 short_io_size += pga[i]->count;
1341 /* Check if read/write is small enough to be a short io. */
1342 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1343 !imp_connect_shortio(cli->cl_import))
1346 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1347 opc == OST_READ ? 0 : short_io_size);
1348 if (opc == OST_READ)
1349 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1352 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1354 ptlrpc_request_free(req);
1357 osc_set_io_portal(req);
1359 ptlrpc_at_set_req_timeout(req);
1360 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1362 req->rq_no_retry_einprogress = 1;
1364 if (short_io_size != 0) {
1366 short_io_buf = NULL;
1370 desc = ptlrpc_prep_bulk_imp(req, page_count,
1371 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1372 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1373 PTLRPC_BULK_PUT_SINK) |
1374 PTLRPC_BULK_BUF_KIOV,
1376 &ptlrpc_bulk_kiov_pin_ops);
1379 GOTO(out, rc = -ENOMEM);
1380 /* NB request now owns desc and will free it when it gets freed */
1382 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1383 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1384 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1385 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1387 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1389 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1390 * and from_kgid(), because they are asynchronous. Fortunately, variable
1391 * oa contains valid o_uid and o_gid in these two operations.
1392 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1393 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1394 * other process logic */
1395 body->oa.o_uid = oa->o_uid;
1396 body->oa.o_gid = oa->o_gid;
1398 obdo_to_ioobj(oa, ioobj);
1399 ioobj->ioo_bufcnt = niocount;
1400 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1401 * that might be send for this request. The actual number is decided
1402 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1403 * "max - 1" for old client compatibility sending "0", and also so the
1404 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1406 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1408 ioobj_max_brw_set(ioobj, 0);
1410 if (short_io_size != 0) {
1411 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1412 body->oa.o_valid |= OBD_MD_FLFLAGS;
1413 body->oa.o_flags = 0;
1415 body->oa.o_flags |= OBD_FL_SHORT_IO;
1416 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1418 if (opc == OST_WRITE) {
1419 short_io_buf = req_capsule_client_get(pill,
1421 LASSERT(short_io_buf != NULL);
1425 LASSERT(page_count > 0);
1427 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1428 struct brw_page *pg = pga[i];
1429 int poff = pg->off & ~PAGE_MASK;
1431 LASSERT(pg->count > 0);
1432 /* make sure there is no gap in the middle of page array */
1433 LASSERTF(page_count == 1 ||
1434 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1435 ergo(i > 0 && i < page_count - 1,
1436 poff == 0 && pg->count == PAGE_SIZE) &&
1437 ergo(i == page_count - 1, poff == 0)),
1438 "i: %d/%d pg: %p off: %llu, count: %u\n",
1439 i, page_count, pg, pg->off, pg->count);
1440 LASSERTF(i == 0 || pg->off > pg_prev->off,
1441 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1442 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1444 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1445 pg_prev->pg, page_private(pg_prev->pg),
1446 pg_prev->pg->index, pg_prev->off);
1447 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1448 (pg->flag & OBD_BRW_SRVLOCK));
1449 if (short_io_size != 0 && opc == OST_WRITE) {
1450 unsigned char *ptr = ll_kmap_atomic(pg->pg, KM_USER0);
1452 LASSERT(short_io_size >= requested_nob + pg->count);
1453 memcpy(short_io_buf + requested_nob,
1456 ll_kunmap_atomic(ptr, KM_USER0);
1457 } else if (short_io_size == 0) {
1458 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1461 requested_nob += pg->count;
1463 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1465 niobuf->rnb_len += pg->count;
1467 niobuf->rnb_offset = pg->off;
1468 niobuf->rnb_len = pg->count;
1469 niobuf->rnb_flags = pg->flag;
1474 LASSERTF((void *)(niobuf - niocount) ==
1475 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1476 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1477 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1479 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1481 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1482 body->oa.o_valid |= OBD_MD_FLFLAGS;
1483 body->oa.o_flags = 0;
1485 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1488 if (osc_should_shrink_grant(cli))
1489 osc_shrink_grant_local(cli, &body->oa);
1491 /* size[REQ_REC_OFF] still sizeof (*body) */
1492 if (opc == OST_WRITE) {
1493 if (cli->cl_checksum &&
1494 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1495 /* store cl_cksum_type in a local variable since
1496 * it can be changed via lprocfs */
1497 enum cksum_types cksum_type = cli->cl_cksum_type;
1499 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1500 body->oa.o_flags = 0;
1502 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1504 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1506 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1507 requested_nob, page_count,
1511 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1515 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1518 /* save this in 'oa', too, for later checking */
1519 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1520 oa->o_flags |= obd_cksum_type_pack(obd_name,
1523 /* clear out the checksum flag, in case this is a
1524 * resend but cl_checksum is no longer set. b=11238 */
1525 oa->o_valid &= ~OBD_MD_FLCKSUM;
1527 oa->o_cksum = body->oa.o_cksum;
1528 /* 1 RC per niobuf */
1529 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1530 sizeof(__u32) * niocount);
1532 if (cli->cl_checksum &&
1533 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1534 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1535 body->oa.o_flags = 0;
1536 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1537 cli->cl_cksum_type);
1538 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1541 /* Client cksum has been already copied to wire obdo in previous
1542 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1543 * resent due to cksum error, this will allow Server to
1544 * check+dump pages on its side */
1546 ptlrpc_request_set_replen(req);
1548 aa = ptlrpc_req_async_args(aa, req);
1550 aa->aa_requested_nob = requested_nob;
1551 aa->aa_nio_count = niocount;
1552 aa->aa_page_count = page_count;
1556 INIT_LIST_HEAD(&aa->aa_oaps);
1559 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1560 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1561 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1562 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1566 ptlrpc_req_finished(req);
1570 char dbgcksum_file_name[PATH_MAX];
1572 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1573 struct brw_page **pga, __u32 server_cksum,
1581 /* will only keep dump of pages on first error for the same range in
1582 * file/fid, not during the resends/retries. */
1583 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1584 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1585 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1586 libcfs_debug_file_path_arr :
1587 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1588 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1589 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1590 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1592 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1593 client_cksum, server_cksum);
1594 filp = filp_open(dbgcksum_file_name,
1595 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1599 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1600 "checksum error: rc = %d\n", dbgcksum_file_name,
1603 CERROR("%s: can't open to dump pages with checksum "
1604 "error: rc = %d\n", dbgcksum_file_name, rc);
1608 for (i = 0; i < page_count; i++) {
1609 len = pga[i]->count;
1610 buf = kmap(pga[i]->pg);
1612 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1614 CERROR("%s: wanted to write %u but got %d "
1615 "error\n", dbgcksum_file_name, len, rc);
1620 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1621 dbgcksum_file_name, rc);
1626 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1628 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1629 filp_close(filp, NULL);
1634 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1635 __u32 client_cksum, __u32 server_cksum,
1636 struct osc_brw_async_args *aa)
1638 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1639 enum cksum_types cksum_type;
1640 obd_dif_csum_fn *fn = NULL;
1641 int sector_size = 0;
1646 if (server_cksum == client_cksum) {
1647 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1651 if (aa->aa_cli->cl_checksum_dump)
1652 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1653 server_cksum, client_cksum);
1655 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1658 switch (cksum_type) {
1659 case OBD_CKSUM_T10IP512:
1663 case OBD_CKSUM_T10IP4K:
1667 case OBD_CKSUM_T10CRC512:
1668 fn = obd_dif_crc_fn;
1671 case OBD_CKSUM_T10CRC4K:
1672 fn = obd_dif_crc_fn;
1680 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1681 aa->aa_page_count, aa->aa_ppga,
1682 OST_WRITE, fn, sector_size,
1685 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1686 aa->aa_ppga, OST_WRITE, cksum_type,
1690 msg = "failed to calculate the client write checksum";
1691 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1692 msg = "the server did not use the checksum type specified in "
1693 "the original request - likely a protocol problem";
1694 else if (new_cksum == server_cksum)
1695 msg = "changed on the client after we checksummed it - "
1696 "likely false positive due to mmap IO (bug 11742)";
1697 else if (new_cksum == client_cksum)
1698 msg = "changed in transit before arrival at OST";
1700 msg = "changed in transit AND doesn't match the original - "
1701 "likely false positive due to mmap IO (bug 11742)";
1703 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1704 DFID " object "DOSTID" extent [%llu-%llu], original "
1705 "client csum %x (type %x), server csum %x (type %x),"
1706 " client csum now %x\n",
1707 obd_name, msg, libcfs_nid2str(peer->nid),
1708 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1709 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1710 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1711 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1712 aa->aa_ppga[aa->aa_page_count - 1]->off +
1713 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1715 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1716 server_cksum, cksum_type, new_cksum);
1720 /* Note rc enters this function as number of bytes transferred */
1721 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1723 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1724 struct client_obd *cli = aa->aa_cli;
1725 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1726 const struct lnet_process_id *peer =
1727 &req->rq_import->imp_connection->c_peer;
1728 struct ost_body *body;
1729 u32 client_cksum = 0;
1732 if (rc < 0 && rc != -EDQUOT) {
1733 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1737 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1738 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1740 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1744 /* set/clear over quota flag for a uid/gid/projid */
1745 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1746 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1747 unsigned qid[LL_MAXQUOTAS] = {
1748 body->oa.o_uid, body->oa.o_gid,
1749 body->oa.o_projid };
1750 CDEBUG(D_QUOTA, "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1751 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1752 body->oa.o_valid, body->oa.o_flags);
1753 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1757 osc_update_grant(cli, body);
1762 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1763 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1765 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1767 CERROR("Unexpected +ve rc %d\n", rc);
1771 if (req->rq_bulk != NULL &&
1772 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1775 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1776 check_write_checksum(&body->oa, peer, client_cksum,
1777 body->oa.o_cksum, aa))
1780 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1781 aa->aa_page_count, aa->aa_ppga);
1785 /* The rest of this function executes only for OST_READs */
1787 if (req->rq_bulk == NULL) {
1788 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1790 LASSERT(rc == req->rq_status);
1792 /* if unwrap_bulk failed, return -EAGAIN to retry */
1793 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1796 GOTO(out, rc = -EAGAIN);
1798 if (rc > aa->aa_requested_nob) {
1799 CERROR("Unexpected rc %d (%d requested)\n", rc,
1800 aa->aa_requested_nob);
1804 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1805 CERROR ("Unexpected rc %d (%d transferred)\n",
1806 rc, req->rq_bulk->bd_nob_transferred);
1810 if (req->rq_bulk == NULL) {
1812 int nob, pg_count, i = 0;
1815 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1816 pg_count = aa->aa_page_count;
1817 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1820 while (nob > 0 && pg_count > 0) {
1822 int count = aa->aa_ppga[i]->count > nob ?
1823 nob : aa->aa_ppga[i]->count;
1825 CDEBUG(D_CACHE, "page %p count %d\n",
1826 aa->aa_ppga[i]->pg, count);
1827 ptr = ll_kmap_atomic(aa->aa_ppga[i]->pg, KM_USER0);
1828 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1830 ll_kunmap_atomic((void *) ptr, KM_USER0);
1839 if (rc < aa->aa_requested_nob)
1840 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1842 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1843 static int cksum_counter;
1844 u32 server_cksum = body->oa.o_cksum;
1847 enum cksum_types cksum_type;
1848 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1849 body->oa.o_flags : 0;
1851 cksum_type = obd_cksum_type_unpack(o_flags);
1852 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1853 aa->aa_page_count, aa->aa_ppga,
1854 OST_READ, &client_cksum);
1858 if (req->rq_bulk != NULL &&
1859 peer->nid != req->rq_bulk->bd_sender) {
1861 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1864 if (server_cksum != client_cksum) {
1865 struct ost_body *clbody;
1866 u32 page_count = aa->aa_page_count;
1868 clbody = req_capsule_client_get(&req->rq_pill,
1870 if (cli->cl_checksum_dump)
1871 dump_all_bulk_pages(&clbody->oa, page_count,
1872 aa->aa_ppga, server_cksum,
1875 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1876 "%s%s%s inode "DFID" object "DOSTID
1877 " extent [%llu-%llu], client %x, "
1878 "server %x, cksum_type %x\n",
1880 libcfs_nid2str(peer->nid),
1882 clbody->oa.o_valid & OBD_MD_FLFID ?
1883 clbody->oa.o_parent_seq : 0ULL,
1884 clbody->oa.o_valid & OBD_MD_FLFID ?
1885 clbody->oa.o_parent_oid : 0,
1886 clbody->oa.o_valid & OBD_MD_FLFID ?
1887 clbody->oa.o_parent_ver : 0,
1888 POSTID(&body->oa.o_oi),
1889 aa->aa_ppga[0]->off,
1890 aa->aa_ppga[page_count-1]->off +
1891 aa->aa_ppga[page_count-1]->count - 1,
1892 client_cksum, server_cksum,
1895 aa->aa_oa->o_cksum = client_cksum;
1899 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1902 } else if (unlikely(client_cksum)) {
1903 static int cksum_missed;
1906 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1907 CERROR("Checksum %u requested from %s but not sent\n",
1908 cksum_missed, libcfs_nid2str(peer->nid));
1914 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1915 aa->aa_oa, &body->oa);
1920 static int osc_brw_redo_request(struct ptlrpc_request *request,
1921 struct osc_brw_async_args *aa, int rc)
1923 struct ptlrpc_request *new_req;
1924 struct osc_brw_async_args *new_aa;
1925 struct osc_async_page *oap;
1928 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1929 "redo for recoverable error %d", rc);
1931 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1932 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1933 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1934 aa->aa_ppga, &new_req, 1);
1938 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1939 if (oap->oap_request != NULL) {
1940 LASSERTF(request == oap->oap_request,
1941 "request %p != oap_request %p\n",
1942 request, oap->oap_request);
1943 if (oap->oap_interrupted) {
1944 ptlrpc_req_finished(new_req);
1950 * New request takes over pga and oaps from old request.
1951 * Note that copying a list_head doesn't work, need to move it...
1954 new_req->rq_interpret_reply = request->rq_interpret_reply;
1955 new_req->rq_async_args = request->rq_async_args;
1956 new_req->rq_commit_cb = request->rq_commit_cb;
1957 /* cap resend delay to the current request timeout, this is similar to
1958 * what ptlrpc does (see after_reply()) */
1959 if (aa->aa_resends > new_req->rq_timeout)
1960 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1962 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1963 new_req->rq_generation_set = 1;
1964 new_req->rq_import_generation = request->rq_import_generation;
1966 new_aa = ptlrpc_req_async_args(new_aa, new_req);
1968 INIT_LIST_HEAD(&new_aa->aa_oaps);
1969 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1970 INIT_LIST_HEAD(&new_aa->aa_exts);
1971 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1972 new_aa->aa_resends = aa->aa_resends;
1974 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1975 if (oap->oap_request) {
1976 ptlrpc_req_finished(oap->oap_request);
1977 oap->oap_request = ptlrpc_request_addref(new_req);
1981 /* XXX: This code will run into problem if we're going to support
1982 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1983 * and wait for all of them to be finished. We should inherit request
1984 * set from old request. */
1985 ptlrpcd_add_req(new_req);
1987 DEBUG_REQ(D_INFO, new_req, "new request");
1992 * ugh, we want disk allocation on the target to happen in offset order. we'll
1993 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1994 * fine for our small page arrays and doesn't require allocation. its an
1995 * insertion sort that swaps elements that are strides apart, shrinking the
1996 * stride down until its '1' and the array is sorted.
1998 static void sort_brw_pages(struct brw_page **array, int num)
2001 struct brw_page *tmp;
2005 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2010 for (i = stride ; i < num ; i++) {
2013 while (j >= stride && array[j - stride]->off > tmp->off) {
2014 array[j] = array[j - stride];
2019 } while (stride > 1);
2022 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2024 LASSERT(ppga != NULL);
2025 OBD_FREE(ppga, sizeof(*ppga) * count);
2028 static int brw_interpret(const struct lu_env *env,
2029 struct ptlrpc_request *req, void *args, int rc)
2031 struct osc_brw_async_args *aa = args;
2032 struct osc_extent *ext;
2033 struct osc_extent *tmp;
2034 struct client_obd *cli = aa->aa_cli;
2035 unsigned long transferred = 0;
2039 rc = osc_brw_fini_request(req, rc);
2040 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2042 * When server returns -EINPROGRESS, client should always retry
2043 * regardless of the number of times the bulk was resent already.
2045 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2046 if (req->rq_import_generation !=
2047 req->rq_import->imp_generation) {
2048 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2049 ""DOSTID", rc = %d.\n",
2050 req->rq_import->imp_obd->obd_name,
2051 POSTID(&aa->aa_oa->o_oi), rc);
2052 } else if (rc == -EINPROGRESS ||
2053 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2054 rc = osc_brw_redo_request(req, aa, rc);
2056 CERROR("%s: too many resent retries for object: "
2057 "%llu:%llu, rc = %d.\n",
2058 req->rq_import->imp_obd->obd_name,
2059 POSTID(&aa->aa_oa->o_oi), rc);
2064 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2069 struct obdo *oa = aa->aa_oa;
2070 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2071 unsigned long valid = 0;
2072 struct cl_object *obj;
2073 struct osc_async_page *last;
2075 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2076 obj = osc2cl(last->oap_obj);
2078 cl_object_attr_lock(obj);
2079 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2080 attr->cat_blocks = oa->o_blocks;
2081 valid |= CAT_BLOCKS;
2083 if (oa->o_valid & OBD_MD_FLMTIME) {
2084 attr->cat_mtime = oa->o_mtime;
2087 if (oa->o_valid & OBD_MD_FLATIME) {
2088 attr->cat_atime = oa->o_atime;
2091 if (oa->o_valid & OBD_MD_FLCTIME) {
2092 attr->cat_ctime = oa->o_ctime;
2096 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2097 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2098 loff_t last_off = last->oap_count + last->oap_obj_off +
2101 /* Change file size if this is an out of quota or
2102 * direct IO write and it extends the file size */
2103 if (loi->loi_lvb.lvb_size < last_off) {
2104 attr->cat_size = last_off;
2107 /* Extend KMS if it's not a lockless write */
2108 if (loi->loi_kms < last_off &&
2109 oap2osc_page(last)->ops_srvlock == 0) {
2110 attr->cat_kms = last_off;
2116 cl_object_attr_update(env, obj, attr, valid);
2117 cl_object_attr_unlock(obj);
2119 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2121 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2122 osc_inc_unstable_pages(req);
2124 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2125 list_del_init(&ext->oe_link);
2126 osc_extent_finish(env, ext, 1,
2127 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2129 LASSERT(list_empty(&aa->aa_exts));
2130 LASSERT(list_empty(&aa->aa_oaps));
2132 transferred = (req->rq_bulk == NULL ? /* short io */
2133 aa->aa_requested_nob :
2134 req->rq_bulk->bd_nob_transferred);
2136 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2137 ptlrpc_lprocfs_brw(req, transferred);
2139 spin_lock(&cli->cl_loi_list_lock);
2140 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2141 * is called so we know whether to go to sync BRWs or wait for more
2142 * RPCs to complete */
2143 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2144 cli->cl_w_in_flight--;
2146 cli->cl_r_in_flight--;
2147 osc_wake_cache_waiters(cli);
2148 spin_unlock(&cli->cl_loi_list_lock);
2150 osc_io_unplug(env, cli, NULL);
2154 static void brw_commit(struct ptlrpc_request *req)
2156 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2157 * this called via the rq_commit_cb, I need to ensure
2158 * osc_dec_unstable_pages is still called. Otherwise unstable
2159 * pages may be leaked. */
2160 spin_lock(&req->rq_lock);
2161 if (likely(req->rq_unstable)) {
2162 req->rq_unstable = 0;
2163 spin_unlock(&req->rq_lock);
2165 osc_dec_unstable_pages(req);
2167 req->rq_committed = 1;
2168 spin_unlock(&req->rq_lock);
2173 * Build an RPC by the list of extent @ext_list. The caller must ensure
2174 * that the total pages in this list are NOT over max pages per RPC.
2175 * Extents in the list must be in OES_RPC state.
2177 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2178 struct list_head *ext_list, int cmd)
2180 struct ptlrpc_request *req = NULL;
2181 struct osc_extent *ext;
2182 struct brw_page **pga = NULL;
2183 struct osc_brw_async_args *aa = NULL;
2184 struct obdo *oa = NULL;
2185 struct osc_async_page *oap;
2186 struct osc_object *obj = NULL;
2187 struct cl_req_attr *crattr = NULL;
2188 loff_t starting_offset = OBD_OBJECT_EOF;
2189 loff_t ending_offset = 0;
2193 bool soft_sync = false;
2194 bool interrupted = false;
2195 bool ndelay = false;
2199 __u32 layout_version = 0;
2200 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
2201 struct ost_body *body;
2203 LASSERT(!list_empty(ext_list));
2205 /* add pages into rpc_list to build BRW rpc */
2206 list_for_each_entry(ext, ext_list, oe_link) {
2207 LASSERT(ext->oe_state == OES_RPC);
2208 mem_tight |= ext->oe_memalloc;
2209 grant += ext->oe_grants;
2210 page_count += ext->oe_nr_pages;
2211 layout_version = MAX(layout_version, ext->oe_layout_version);
2216 soft_sync = osc_over_unstable_soft_limit(cli);
2218 mpflag = cfs_memory_pressure_get_and_set();
2220 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2222 GOTO(out, rc = -ENOMEM);
2224 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2226 GOTO(out, rc = -ENOMEM);
2229 list_for_each_entry(ext, ext_list, oe_link) {
2230 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2232 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2234 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2235 pga[i] = &oap->oap_brw_page;
2236 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2239 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2240 if (starting_offset == OBD_OBJECT_EOF ||
2241 starting_offset > oap->oap_obj_off)
2242 starting_offset = oap->oap_obj_off;
2244 LASSERT(oap->oap_page_off == 0);
2245 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2246 ending_offset = oap->oap_obj_off +
2249 LASSERT(oap->oap_page_off + oap->oap_count ==
2251 if (oap->oap_interrupted)
2258 /* first page in the list */
2259 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2261 crattr = &osc_env_info(env)->oti_req_attr;
2262 memset(crattr, 0, sizeof(*crattr));
2263 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2264 crattr->cra_flags = ~0ULL;
2265 crattr->cra_page = oap2cl_page(oap);
2266 crattr->cra_oa = oa;
2267 cl_req_attr_set(env, osc2cl(obj), crattr);
2269 if (cmd == OBD_BRW_WRITE) {
2270 oa->o_grant_used = grant;
2271 if (layout_version > 0) {
2272 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2273 PFID(&oa->o_oi.oi_fid), layout_version);
2275 oa->o_layout_version = layout_version;
2276 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2280 sort_brw_pages(pga, page_count);
2281 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2283 CERROR("prep_req failed: %d\n", rc);
2287 req->rq_commit_cb = brw_commit;
2288 req->rq_interpret_reply = brw_interpret;
2289 req->rq_memalloc = mem_tight != 0;
2290 oap->oap_request = ptlrpc_request_addref(req);
2291 if (interrupted && !req->rq_intr)
2292 ptlrpc_mark_interrupted(req);
2294 req->rq_no_resend = req->rq_no_delay = 1;
2295 /* probably set a shorter timeout value.
2296 * to handle ETIMEDOUT in brw_interpret() correctly. */
2297 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2300 /* Need to update the timestamps after the request is built in case
2301 * we race with setattr (locally or in queue at OST). If OST gets
2302 * later setattr before earlier BRW (as determined by the request xid),
2303 * the OST will not use BRW timestamps. Sadly, there is no obvious
2304 * way to do this in a single call. bug 10150 */
2305 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2306 crattr->cra_oa = &body->oa;
2307 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2308 cl_req_attr_set(env, osc2cl(obj), crattr);
2309 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2311 aa = ptlrpc_req_async_args(aa, req);
2312 INIT_LIST_HEAD(&aa->aa_oaps);
2313 list_splice_init(&rpc_list, &aa->aa_oaps);
2314 INIT_LIST_HEAD(&aa->aa_exts);
2315 list_splice_init(ext_list, &aa->aa_exts);
2317 spin_lock(&cli->cl_loi_list_lock);
2318 starting_offset >>= PAGE_SHIFT;
2319 if (cmd == OBD_BRW_READ) {
2320 cli->cl_r_in_flight++;
2321 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2322 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2323 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2324 starting_offset + 1);
2326 cli->cl_w_in_flight++;
2327 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2328 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2329 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2330 starting_offset + 1);
2332 spin_unlock(&cli->cl_loi_list_lock);
2334 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
2335 page_count, aa, cli->cl_r_in_flight,
2336 cli->cl_w_in_flight);
2337 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2339 ptlrpcd_add_req(req);
2345 cfs_memory_pressure_restore(mpflag);
2348 LASSERT(req == NULL);
2351 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2353 OBD_FREE(pga, sizeof(*pga) * page_count);
2354 /* this should happen rarely and is pretty bad, it makes the
2355 * pending list not follow the dirty order */
2356 while (!list_empty(ext_list)) {
2357 ext = list_entry(ext_list->next, struct osc_extent,
2359 list_del_init(&ext->oe_link);
2360 osc_extent_finish(env, ext, 0, rc);
2366 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2370 LASSERT(lock != NULL);
2372 lock_res_and_lock(lock);
2374 if (lock->l_ast_data == NULL)
2375 lock->l_ast_data = data;
2376 if (lock->l_ast_data == data)
2379 unlock_res_and_lock(lock);
2384 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2385 void *cookie, struct lustre_handle *lockh,
2386 enum ldlm_mode mode, __u64 *flags, bool speculative,
2389 bool intent = *flags & LDLM_FL_HAS_INTENT;
2393 /* The request was created before ldlm_cli_enqueue call. */
2394 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2395 struct ldlm_reply *rep;
2397 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2398 LASSERT(rep != NULL);
2400 rep->lock_policy_res1 =
2401 ptlrpc_status_ntoh(rep->lock_policy_res1);
2402 if (rep->lock_policy_res1)
2403 errcode = rep->lock_policy_res1;
2405 *flags |= LDLM_FL_LVB_READY;
2406 } else if (errcode == ELDLM_OK) {
2407 *flags |= LDLM_FL_LVB_READY;
2410 /* Call the update callback. */
2411 rc = (*upcall)(cookie, lockh, errcode);
2413 /* release the reference taken in ldlm_cli_enqueue() */
2414 if (errcode == ELDLM_LOCK_MATCHED)
2416 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2417 ldlm_lock_decref(lockh, mode);
2422 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2425 struct osc_enqueue_args *aa = args;
2426 struct ldlm_lock *lock;
2427 struct lustre_handle *lockh = &aa->oa_lockh;
2428 enum ldlm_mode mode = aa->oa_mode;
2429 struct ost_lvb *lvb = aa->oa_lvb;
2430 __u32 lvb_len = sizeof(*lvb);
2435 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2437 lock = ldlm_handle2lock(lockh);
2438 LASSERTF(lock != NULL,
2439 "lockh %#llx, req %p, aa %p - client evicted?\n",
2440 lockh->cookie, req, aa);
2442 /* Take an additional reference so that a blocking AST that
2443 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2444 * to arrive after an upcall has been executed by
2445 * osc_enqueue_fini(). */
2446 ldlm_lock_addref(lockh, mode);
2448 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2449 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2451 /* Let CP AST to grant the lock first. */
2452 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2454 if (aa->oa_speculative) {
2455 LASSERT(aa->oa_lvb == NULL);
2456 LASSERT(aa->oa_flags == NULL);
2457 aa->oa_flags = &flags;
2460 /* Complete obtaining the lock procedure. */
2461 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2462 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2464 /* Complete osc stuff. */
2465 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2466 aa->oa_flags, aa->oa_speculative, rc);
2468 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2470 ldlm_lock_decref(lockh, mode);
2471 LDLM_LOCK_PUT(lock);
2475 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2477 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2478 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2479 * other synchronous requests, however keeping some locks and trying to obtain
2480 * others may take a considerable amount of time in a case of ost failure; and
2481 * when other sync requests do not get released lock from a client, the client
2482 * is evicted from the cluster -- such scenarious make the life difficult, so
2483 * release locks just after they are obtained. */
2484 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2485 __u64 *flags, union ldlm_policy_data *policy,
2486 struct ost_lvb *lvb, int kms_valid,
2487 osc_enqueue_upcall_f upcall, void *cookie,
2488 struct ldlm_enqueue_info *einfo,
2489 struct ptlrpc_request_set *rqset, int async,
2492 struct obd_device *obd = exp->exp_obd;
2493 struct lustre_handle lockh = { 0 };
2494 struct ptlrpc_request *req = NULL;
2495 int intent = *flags & LDLM_FL_HAS_INTENT;
2496 __u64 match_flags = *flags;
2497 enum ldlm_mode mode;
2501 /* Filesystem lock extents are extended to page boundaries so that
2502 * dealing with the page cache is a little smoother. */
2503 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2504 policy->l_extent.end |= ~PAGE_MASK;
2507 * kms is not valid when either object is completely fresh (so that no
2508 * locks are cached), or object was evicted. In the latter case cached
2509 * lock cannot be used, because it would prime inode state with
2510 * potentially stale LVB.
2515 /* Next, search for already existing extent locks that will cover us */
2516 /* If we're trying to read, we also search for an existing PW lock. The
2517 * VFS and page cache already protect us locally, so lots of readers/
2518 * writers can share a single PW lock.
2520 * There are problems with conversion deadlocks, so instead of
2521 * converting a read lock to a write lock, we'll just enqueue a new
2524 * At some point we should cancel the read lock instead of making them
2525 * send us a blocking callback, but there are problems with canceling
2526 * locks out from other users right now, too. */
2527 mode = einfo->ei_mode;
2528 if (einfo->ei_mode == LCK_PR)
2530 /* Normal lock requests must wait for the LVB to be ready before
2531 * matching a lock; speculative lock requests do not need to,
2532 * because they will not actually use the lock. */
2534 match_flags |= LDLM_FL_LVB_READY;
2536 match_flags |= LDLM_FL_BLOCK_GRANTED;
2537 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2538 einfo->ei_type, policy, mode, &lockh, 0);
2540 struct ldlm_lock *matched;
2542 if (*flags & LDLM_FL_TEST_LOCK)
2545 matched = ldlm_handle2lock(&lockh);
2547 /* This DLM lock request is speculative, and does not
2548 * have an associated IO request. Therefore if there
2549 * is already a DLM lock, it wll just inform the
2550 * caller to cancel the request for this stripe.*/
2551 lock_res_and_lock(matched);
2552 if (ldlm_extent_equal(&policy->l_extent,
2553 &matched->l_policy_data.l_extent))
2557 unlock_res_and_lock(matched);
2559 ldlm_lock_decref(&lockh, mode);
2560 LDLM_LOCK_PUT(matched);
2562 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2563 *flags |= LDLM_FL_LVB_READY;
2565 /* We already have a lock, and it's referenced. */
2566 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2568 ldlm_lock_decref(&lockh, mode);
2569 LDLM_LOCK_PUT(matched);
2572 ldlm_lock_decref(&lockh, mode);
2573 LDLM_LOCK_PUT(matched);
2578 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2582 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2583 &RQF_LDLM_ENQUEUE_LVB);
2587 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2589 ptlrpc_request_free(req);
2593 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2595 ptlrpc_request_set_replen(req);
2598 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2599 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2601 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2602 sizeof(*lvb), LVB_T_OST, &lockh, async);
2605 struct osc_enqueue_args *aa;
2606 aa = ptlrpc_req_async_args(aa, req);
2608 aa->oa_mode = einfo->ei_mode;
2609 aa->oa_type = einfo->ei_type;
2610 lustre_handle_copy(&aa->oa_lockh, &lockh);
2611 aa->oa_upcall = upcall;
2612 aa->oa_cookie = cookie;
2613 aa->oa_speculative = speculative;
2615 aa->oa_flags = flags;
2618 /* speculative locks are essentially to enqueue
2619 * a DLM lock in advance, so we don't care
2620 * about the result of the enqueue. */
2622 aa->oa_flags = NULL;
2625 req->rq_interpret_reply = osc_enqueue_interpret;
2626 if (rqset == PTLRPCD_SET)
2627 ptlrpcd_add_req(req);
2629 ptlrpc_set_add_req(rqset, req);
2630 } else if (intent) {
2631 ptlrpc_req_finished(req);
2636 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2637 flags, speculative, rc);
2639 ptlrpc_req_finished(req);
2644 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2645 enum ldlm_type type, union ldlm_policy_data *policy,
2646 enum ldlm_mode mode, __u64 *flags, void *data,
2647 struct lustre_handle *lockh, int unref)
2649 struct obd_device *obd = exp->exp_obd;
2650 __u64 lflags = *flags;
2654 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2657 /* Filesystem lock extents are extended to page boundaries so that
2658 * dealing with the page cache is a little smoother */
2659 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2660 policy->l_extent.end |= ~PAGE_MASK;
2662 /* Next, search for already existing extent locks that will cover us */
2663 /* If we're trying to read, we also search for an existing PW lock. The
2664 * VFS and page cache already protect us locally, so lots of readers/
2665 * writers can share a single PW lock. */
2669 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2670 res_id, type, policy, rc, lockh, unref);
2671 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2675 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2677 LASSERT(lock != NULL);
2678 if (!osc_set_lock_data(lock, data)) {
2679 ldlm_lock_decref(lockh, rc);
2682 LDLM_LOCK_PUT(lock);
2687 static int osc_statfs_interpret(const struct lu_env *env,
2688 struct ptlrpc_request *req, void *args, int rc)
2690 struct osc_async_args *aa = args;
2691 struct obd_statfs *msfs;
2696 * The request has in fact never been sent due to issues at
2697 * a higher level (LOV). Exit immediately since the caller
2698 * is aware of the problem and takes care of the clean up.
2702 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2703 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2709 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2711 GOTO(out, rc = -EPROTO);
2713 *aa->aa_oi->oi_osfs = *msfs;
2715 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2720 static int osc_statfs_async(struct obd_export *exp,
2721 struct obd_info *oinfo, time64_t max_age,
2722 struct ptlrpc_request_set *rqset)
2724 struct obd_device *obd = class_exp2obd(exp);
2725 struct ptlrpc_request *req;
2726 struct osc_async_args *aa;
2730 if (obd->obd_osfs_age >= max_age) {
2732 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2733 obd->obd_name, &obd->obd_osfs,
2734 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2735 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2736 spin_lock(&obd->obd_osfs_lock);
2737 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2738 spin_unlock(&obd->obd_osfs_lock);
2739 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2740 if (oinfo->oi_cb_up)
2741 oinfo->oi_cb_up(oinfo, 0);
2746 /* We could possibly pass max_age in the request (as an absolute
2747 * timestamp or a "seconds.usec ago") so the target can avoid doing
2748 * extra calls into the filesystem if that isn't necessary (e.g.
2749 * during mount that would help a bit). Having relative timestamps
2750 * is not so great if request processing is slow, while absolute
2751 * timestamps are not ideal because they need time synchronization. */
2752 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2756 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2758 ptlrpc_request_free(req);
2761 ptlrpc_request_set_replen(req);
2762 req->rq_request_portal = OST_CREATE_PORTAL;
2763 ptlrpc_at_set_req_timeout(req);
2765 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2766 /* procfs requests not want stat in wait for avoid deadlock */
2767 req->rq_no_resend = 1;
2768 req->rq_no_delay = 1;
2771 req->rq_interpret_reply = osc_statfs_interpret;
2772 aa = ptlrpc_req_async_args(aa, req);
2775 ptlrpc_set_add_req(rqset, req);
2779 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2780 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2782 struct obd_device *obd = class_exp2obd(exp);
2783 struct obd_statfs *msfs;
2784 struct ptlrpc_request *req;
2785 struct obd_import *imp = NULL;
2790 /*Since the request might also come from lprocfs, so we need
2791 *sync this with client_disconnect_export Bug15684*/
2792 down_read(&obd->u.cli.cl_sem);
2793 if (obd->u.cli.cl_import)
2794 imp = class_import_get(obd->u.cli.cl_import);
2795 up_read(&obd->u.cli.cl_sem);
2799 /* We could possibly pass max_age in the request (as an absolute
2800 * timestamp or a "seconds.usec ago") so the target can avoid doing
2801 * extra calls into the filesystem if that isn't necessary (e.g.
2802 * during mount that would help a bit). Having relative timestamps
2803 * is not so great if request processing is slow, while absolute
2804 * timestamps are not ideal because they need time synchronization. */
2805 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2807 class_import_put(imp);
2812 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2814 ptlrpc_request_free(req);
2817 ptlrpc_request_set_replen(req);
2818 req->rq_request_portal = OST_CREATE_PORTAL;
2819 ptlrpc_at_set_req_timeout(req);
2821 if (flags & OBD_STATFS_NODELAY) {
2822 /* procfs requests not want stat in wait for avoid deadlock */
2823 req->rq_no_resend = 1;
2824 req->rq_no_delay = 1;
2827 rc = ptlrpc_queue_wait(req);
2831 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2833 GOTO(out, rc = -EPROTO);
2839 ptlrpc_req_finished(req);
2843 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2844 void *karg, void __user *uarg)
2846 struct obd_device *obd = exp->exp_obd;
2847 struct obd_ioctl_data *data = karg;
2851 if (!try_module_get(THIS_MODULE)) {
2852 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2853 module_name(THIS_MODULE));
2857 case OBD_IOC_CLIENT_RECOVER:
2858 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
2859 data->ioc_inlbuf1, 0);
2863 case IOC_OSC_SET_ACTIVE:
2864 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
2869 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
2870 obd->obd_name, cmd, current_comm(), rc);
2874 module_put(THIS_MODULE);
2878 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2879 u32 keylen, void *key, u32 vallen, void *val,
2880 struct ptlrpc_request_set *set)
2882 struct ptlrpc_request *req;
2883 struct obd_device *obd = exp->exp_obd;
2884 struct obd_import *imp = class_exp2cliimp(exp);
2889 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2891 if (KEY_IS(KEY_CHECKSUM)) {
2892 if (vallen != sizeof(int))
2894 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2898 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2899 sptlrpc_conf_client_adapt(obd);
2903 if (KEY_IS(KEY_FLUSH_CTX)) {
2904 sptlrpc_import_flush_my_ctx(imp);
2908 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2909 struct client_obd *cli = &obd->u.cli;
2910 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2911 long target = *(long *)val;
2913 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2918 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2921 /* We pass all other commands directly to OST. Since nobody calls osc
2922 methods directly and everybody is supposed to go through LOV, we
2923 assume lov checked invalid values for us.
2924 The only recognised values so far are evict_by_nid and mds_conn.
2925 Even if something bad goes through, we'd get a -EINVAL from OST
2928 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2929 &RQF_OST_SET_GRANT_INFO :
2934 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2935 RCL_CLIENT, keylen);
2936 if (!KEY_IS(KEY_GRANT_SHRINK))
2937 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2938 RCL_CLIENT, vallen);
2939 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2941 ptlrpc_request_free(req);
2945 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2946 memcpy(tmp, key, keylen);
2947 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2950 memcpy(tmp, val, vallen);
2952 if (KEY_IS(KEY_GRANT_SHRINK)) {
2953 struct osc_grant_args *aa;
2956 aa = ptlrpc_req_async_args(aa, req);
2957 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2959 ptlrpc_req_finished(req);
2962 *oa = ((struct ost_body *)val)->oa;
2964 req->rq_interpret_reply = osc_shrink_grant_interpret;
2967 ptlrpc_request_set_replen(req);
2968 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2969 LASSERT(set != NULL);
2970 ptlrpc_set_add_req(set, req);
2971 ptlrpc_check_set(NULL, set);
2973 ptlrpcd_add_req(req);
2978 EXPORT_SYMBOL(osc_set_info_async);
2980 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
2981 struct obd_device *obd, struct obd_uuid *cluuid,
2982 struct obd_connect_data *data, void *localdata)
2984 struct client_obd *cli = &obd->u.cli;
2986 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2990 spin_lock(&cli->cl_loi_list_lock);
2991 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2992 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
2993 /* restore ocd_grant_blkbits as client page bits */
2994 data->ocd_grant_blkbits = PAGE_SHIFT;
2995 grant += cli->cl_dirty_grant;
2997 grant += cli->cl_dirty_pages << PAGE_SHIFT;
2999 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3000 lost_grant = cli->cl_lost_grant;
3001 cli->cl_lost_grant = 0;
3002 spin_unlock(&cli->cl_loi_list_lock);
3004 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3005 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3006 data->ocd_version, data->ocd_grant, lost_grant);
3011 EXPORT_SYMBOL(osc_reconnect);
3013 int osc_disconnect(struct obd_export *exp)
3015 struct obd_device *obd = class_exp2obd(exp);
3018 rc = client_disconnect_export(exp);
3020 * Initially we put del_shrink_grant before disconnect_export, but it
3021 * causes the following problem if setup (connect) and cleanup
3022 * (disconnect) are tangled together.
3023 * connect p1 disconnect p2
3024 * ptlrpc_connect_import
3025 * ............... class_manual_cleanup
3028 * ptlrpc_connect_interrupt
3030 * add this client to shrink list
3032 * Bang! grant shrink thread trigger the shrink. BUG18662
3034 osc_del_grant_list(&obd->u.cli);
3037 EXPORT_SYMBOL(osc_disconnect);
3039 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3040 struct hlist_node *hnode, void *arg)
3042 struct lu_env *env = arg;
3043 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3044 struct ldlm_lock *lock;
3045 struct osc_object *osc = NULL;
3049 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3050 if (lock->l_ast_data != NULL && osc == NULL) {
3051 osc = lock->l_ast_data;
3052 cl_object_get(osc2cl(osc));
3055 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3056 * by the 2nd round of ldlm_namespace_clean() call in
3057 * osc_import_event(). */
3058 ldlm_clear_cleaned(lock);
3063 osc_object_invalidate(env, osc);
3064 cl_object_put(env, osc2cl(osc));
3069 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3071 static int osc_import_event(struct obd_device *obd,
3072 struct obd_import *imp,
3073 enum obd_import_event event)
3075 struct client_obd *cli;
3079 LASSERT(imp->imp_obd == obd);
3082 case IMP_EVENT_DISCON: {
3084 spin_lock(&cli->cl_loi_list_lock);
3085 cli->cl_avail_grant = 0;
3086 cli->cl_lost_grant = 0;
3087 spin_unlock(&cli->cl_loi_list_lock);
3090 case IMP_EVENT_INACTIVE: {
3091 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3094 case IMP_EVENT_INVALIDATE: {
3095 struct ldlm_namespace *ns = obd->obd_namespace;
3099 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3101 env = cl_env_get(&refcheck);
3103 osc_io_unplug(env, &obd->u.cli, NULL);
3105 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3106 osc_ldlm_resource_invalidate,
3108 cl_env_put(env, &refcheck);
3110 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3115 case IMP_EVENT_ACTIVE: {
3116 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3119 case IMP_EVENT_OCD: {
3120 struct obd_connect_data *ocd = &imp->imp_connect_data;
3122 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3123 osc_init_grant(&obd->u.cli, ocd);
3126 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3127 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3129 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3132 case IMP_EVENT_DEACTIVATE: {
3133 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3136 case IMP_EVENT_ACTIVATE: {
3137 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3141 CERROR("Unknown import event %d\n", event);
3148 * Determine whether the lock can be canceled before replaying the lock
3149 * during recovery, see bug16774 for detailed information.
3151 * \retval zero the lock can't be canceled
3152 * \retval other ok to cancel
3154 static int osc_cancel_weight(struct ldlm_lock *lock)
3157 * Cancel all unused and granted extent lock.
3159 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3160 ldlm_is_granted(lock) &&
3161 osc_ldlm_weigh_ast(lock) == 0)
3167 static int brw_queue_work(const struct lu_env *env, void *data)
3169 struct client_obd *cli = data;
3171 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3173 osc_io_unplug(env, cli, NULL);
3177 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3179 struct client_obd *cli = &obd->u.cli;
3185 rc = ptlrpcd_addref();
3189 rc = client_obd_setup(obd, lcfg);
3191 GOTO(out_ptlrpcd, rc);
3194 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3195 if (IS_ERR(handler))
3196 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3197 cli->cl_writeback_work = handler;
3199 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3200 if (IS_ERR(handler))
3201 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3202 cli->cl_lru_work = handler;
3204 rc = osc_quota_setup(obd);
3206 GOTO(out_ptlrpcd_work, rc);
3208 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3209 osc_update_next_shrink(cli);
3214 if (cli->cl_writeback_work != NULL) {
3215 ptlrpcd_destroy_work(cli->cl_writeback_work);
3216 cli->cl_writeback_work = NULL;
3218 if (cli->cl_lru_work != NULL) {
3219 ptlrpcd_destroy_work(cli->cl_lru_work);
3220 cli->cl_lru_work = NULL;
3222 client_obd_cleanup(obd);
3227 EXPORT_SYMBOL(osc_setup_common);
3229 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3231 struct client_obd *cli = &obd->u.cli;
3239 rc = osc_setup_common(obd, lcfg);
3243 rc = osc_tunables_init(obd);
3248 * We try to control the total number of requests with a upper limit
3249 * osc_reqpool_maxreqcount. There might be some race which will cause
3250 * over-limit allocation, but it is fine.
3252 req_count = atomic_read(&osc_pool_req_count);
3253 if (req_count < osc_reqpool_maxreqcount) {
3254 adding = cli->cl_max_rpcs_in_flight + 2;
3255 if (req_count + adding > osc_reqpool_maxreqcount)
3256 adding = osc_reqpool_maxreqcount - req_count;
3258 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3259 atomic_add(added, &osc_pool_req_count);
3262 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3264 spin_lock(&osc_shrink_lock);
3265 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3266 spin_unlock(&osc_shrink_lock);
3267 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3268 cli->cl_import->imp_idle_debug = D_HA;
3273 int osc_precleanup_common(struct obd_device *obd)
3275 struct client_obd *cli = &obd->u.cli;
3279 * for echo client, export may be on zombie list, wait for
3280 * zombie thread to cull it, because cli.cl_import will be
3281 * cleared in client_disconnect_export():
3282 * class_export_destroy() -> obd_cleanup() ->
3283 * echo_device_free() -> echo_client_cleanup() ->
3284 * obd_disconnect() -> osc_disconnect() ->
3285 * client_disconnect_export()
3287 obd_zombie_barrier();
3288 if (cli->cl_writeback_work) {
3289 ptlrpcd_destroy_work(cli->cl_writeback_work);
3290 cli->cl_writeback_work = NULL;
3293 if (cli->cl_lru_work) {
3294 ptlrpcd_destroy_work(cli->cl_lru_work);
3295 cli->cl_lru_work = NULL;
3298 obd_cleanup_client_import(obd);
3301 EXPORT_SYMBOL(osc_precleanup_common);
3303 static int osc_precleanup(struct obd_device *obd)
3307 osc_precleanup_common(obd);
3309 ptlrpc_lprocfs_unregister_obd(obd);
3313 int osc_cleanup_common(struct obd_device *obd)
3315 struct client_obd *cli = &obd->u.cli;
3320 spin_lock(&osc_shrink_lock);
3321 list_del(&cli->cl_shrink_list);
3322 spin_unlock(&osc_shrink_lock);
3325 if (cli->cl_cache != NULL) {
3326 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3327 spin_lock(&cli->cl_cache->ccc_lru_lock);
3328 list_del_init(&cli->cl_lru_osc);
3329 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3330 cli->cl_lru_left = NULL;
3331 cl_cache_decref(cli->cl_cache);
3332 cli->cl_cache = NULL;
3335 /* free memory of osc quota cache */
3336 osc_quota_cleanup(obd);
3338 rc = client_obd_cleanup(obd);
3343 EXPORT_SYMBOL(osc_cleanup_common);
3345 static struct obd_ops osc_obd_ops = {
3346 .o_owner = THIS_MODULE,
3347 .o_setup = osc_setup,
3348 .o_precleanup = osc_precleanup,
3349 .o_cleanup = osc_cleanup_common,
3350 .o_add_conn = client_import_add_conn,
3351 .o_del_conn = client_import_del_conn,
3352 .o_connect = client_connect_import,
3353 .o_reconnect = osc_reconnect,
3354 .o_disconnect = osc_disconnect,
3355 .o_statfs = osc_statfs,
3356 .o_statfs_async = osc_statfs_async,
3357 .o_create = osc_create,
3358 .o_destroy = osc_destroy,
3359 .o_getattr = osc_getattr,
3360 .o_setattr = osc_setattr,
3361 .o_iocontrol = osc_iocontrol,
3362 .o_set_info_async = osc_set_info_async,
3363 .o_import_event = osc_import_event,
3364 .o_quotactl = osc_quotactl,
3367 static struct shrinker *osc_cache_shrinker;
3368 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
3369 DEFINE_SPINLOCK(osc_shrink_lock);
3371 #ifndef HAVE_SHRINKER_COUNT
3372 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3374 struct shrink_control scv = {
3375 .nr_to_scan = shrink_param(sc, nr_to_scan),
3376 .gfp_mask = shrink_param(sc, gfp_mask)
3378 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
3379 struct shrinker *shrinker = NULL;
3382 (void)osc_cache_shrink_scan(shrinker, &scv);
3384 return osc_cache_shrink_count(shrinker, &scv);
3388 static int __init osc_init(void)
3390 unsigned int reqpool_size;
3391 unsigned int reqsize;
3393 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3394 osc_cache_shrink_count, osc_cache_shrink_scan);
3397 /* print an address of _any_ initialized kernel symbol from this
3398 * module, to allow debugging with gdb that doesn't support data
3399 * symbols from modules.*/
3400 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3402 rc = lu_kmem_init(osc_caches);
3406 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3407 LUSTRE_OSC_NAME, &osc_device_type);
3411 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3413 /* This is obviously too much memory, only prevent overflow here */
3414 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3415 GOTO(out_type, rc = -EINVAL);
3417 reqpool_size = osc_reqpool_mem_max << 20;
3420 while (reqsize < OST_IO_MAXREQSIZE)
3421 reqsize = reqsize << 1;
3424 * We don't enlarge the request count in OSC pool according to
3425 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3426 * tried after normal allocation failed. So a small OSC pool won't
3427 * cause much performance degression in most of cases.
3429 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3431 atomic_set(&osc_pool_req_count, 0);
3432 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3433 ptlrpc_add_rqs_to_pool);
3435 if (osc_rq_pool == NULL)
3436 GOTO(out_type, rc = -ENOMEM);
3438 rc = osc_start_grant_work();
3440 GOTO(out_req_pool, rc);
3445 ptlrpc_free_rq_pool(osc_rq_pool);
3447 class_unregister_type(LUSTRE_OSC_NAME);
3449 lu_kmem_fini(osc_caches);
3454 static void __exit osc_exit(void)
3456 osc_stop_grant_work();
3457 remove_shrinker(osc_cache_shrinker);
3458 class_unregister_type(LUSTRE_OSC_NAME);
3459 lu_kmem_fini(osc_caches);
3460 ptlrpc_free_rq_pool(osc_rq_pool);
3463 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3464 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3465 MODULE_VERSION(LUSTRE_VERSION_STRING);
3466 MODULE_LICENSE("GPL");
3468 module_init(osc_init);
3469 module_exit(osc_exit);