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_ioctl_old.h>
43 #include <lustre_net.h>
44 #include <lustre_obdo.h>
45 #include <lustre_osc.h>
47 #include <obd_cksum.h>
48 #include <obd_class.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 static 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, &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]->bp_count > nob_read) {
1095 /* EOF inside this page */
1096 ptr = kmap(pga[i]->bp_page) +
1097 (pga[i]->bp_off & ~PAGE_MASK);
1098 memset(ptr + nob_read, 0, pga[i]->bp_count - nob_read);
1099 kunmap(pga[i]->bp_page);
1105 nob_read -= pga[i]->bp_count;
1110 /* zero remaining pages */
1111 while (page_count-- > 0) {
1112 ptr = kmap(pga[i]->bp_page) + (pga[i]->bp_off & ~PAGE_MASK);
1113 memset(ptr, 0, pga[i]->bp_count);
1114 kunmap(pga[i]->bp_page);
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->bp_flag != p2->bp_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->bp_flag & mask) != (p2->bp_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->bp_flag, p2->bp_flag);
1176 return (p1->bp_off + p1->bp_count == p2->bp_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;
1191 __be16 *guard_start;
1193 int used_number = 0;
1196 unsigned int bufsize = sizeof(cksum);
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 = (__be16 *)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 int off = pga[i]->bp_off & ~PAGE_MASK;
1223 unsigned int count =
1224 pga[i]->bp_count > nob ? nob : pga[i]->bp_count;
1225 int guards_needed = DIV_ROUND_UP(off + count, sector_size) -
1226 (off / sector_size);
1228 if (guards_needed > guard_number - used_number) {
1229 cfs_crypto_hash_update_page(req, __page, 0,
1230 used_number * sizeof(*guard_start));
1234 /* corrupt the data before we compute the checksum, to
1235 * simulate an OST->client data error */
1236 if (unlikely(i == 0 && opc == OST_READ &&
1237 CFS_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1238 unsigned char *ptr = kmap(pga[i]->bp_page);
1240 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1241 kunmap(pga[i]->bp_page);
1245 * The left guard number should be able to hold checksums of a
1248 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->bp_page,
1249 pga[i]->bp_off & ~PAGE_MASK,
1251 guard_start + used_number,
1252 guard_number - used_number,
1255 if (unlikely(resend))
1256 CDEBUG(D_PAGE | D_HA,
1257 "pga[%u]: used %u off %llu+%u gen checksum: %*phN\n",
1258 i, used, pga[i]->bp_off & ~PAGE_MASK, count,
1259 (int)(used * sizeof(*guard_start)),
1260 guard_start + used_number);
1264 used_number += used;
1265 nob -= pga[i]->bp_count;
1273 if (used_number != 0)
1274 cfs_crypto_hash_update_page(req, __page, 0,
1275 used_number * sizeof(*guard_start));
1278 rc2 = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1282 /* For sending we only compute the wrong checksum instead
1283 * of corrupting the data so it is still correct on a redo */
1284 if (opc == OST_WRITE &&
1285 CFS_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1291 __free_page(__page);
1294 #else /* !CONFIG_CRC_T10DIF */
1295 #define obd_dif_ip_fn NULL
1296 #define obd_dif_crc_fn NULL
1297 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum, re) \
1299 #endif /* CONFIG_CRC_T10DIF */
1301 static int osc_checksum_bulk(int nob, size_t pg_count,
1302 struct brw_page **pga, int opc,
1303 enum cksum_types cksum_type,
1307 struct ahash_request *req;
1308 unsigned int bufsize;
1309 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1311 LASSERT(pg_count > 0);
1313 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1315 CERROR("Unable to initialize checksum hash %s\n",
1316 cfs_crypto_hash_name(cfs_alg));
1317 return PTR_ERR(req);
1320 while (nob > 0 && pg_count > 0) {
1321 unsigned int count =
1322 pga[i]->bp_count > nob ? nob : pga[i]->bp_count;
1324 /* corrupt the data before we compute the checksum, to
1325 * simulate an OST->client data error */
1326 if (i == 0 && opc == OST_READ &&
1327 CFS_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1328 unsigned char *ptr = kmap(pga[i]->bp_page);
1329 int off = pga[i]->bp_off & ~PAGE_MASK;
1331 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1332 kunmap(pga[i]->bp_page);
1334 cfs_crypto_hash_update_page(req, pga[i]->bp_page,
1335 pga[i]->bp_off & ~PAGE_MASK,
1337 LL_CDEBUG_PAGE(D_PAGE, pga[i]->bp_page, "off %d\n",
1338 (int)(pga[i]->bp_off & ~PAGE_MASK));
1340 nob -= pga[i]->bp_count;
1345 bufsize = sizeof(*cksum);
1346 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1348 /* For sending we only compute the wrong checksum instead
1349 * of corrupting the data so it is still correct on a redo */
1350 if (opc == OST_WRITE && CFS_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1356 static int osc_checksum_bulk_rw(const char *obd_name,
1357 enum cksum_types cksum_type,
1358 int nob, size_t pg_count,
1359 struct brw_page **pga, int opc,
1360 u32 *check_sum, bool resend)
1362 obd_dif_csum_fn *fn = NULL;
1363 int sector_size = 0;
1367 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1370 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1371 opc, fn, sector_size, check_sum,
1374 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1380 #ifdef CONFIG_LL_ENCRYPTION
1382 * osc_encrypt_pagecache_blocks() - overlay to llcrypt_encrypt_pagecache_blocks
1383 * @srcpage: The locked pagecache page containing the block(s) to encrypt
1384 * @dstpage: The page to put encryption result
1385 * @len: Total size of the block(s) to encrypt. Must be a nonzero
1386 * multiple of the filesystem's block size.
1387 * @offs: Byte offset within @page of the first block to encrypt. Must be
1388 * a multiple of the filesystem's block size.
1389 * @gfp_flags: Memory allocation flags
1391 * This overlay function is necessary to be able to provide our own bounce page.
1393 static struct page *osc_encrypt_pagecache_blocks(struct page *srcpage,
1394 struct page *dstpage,
1400 const struct inode *inode = srcpage->mapping->host;
1401 const unsigned int blockbits = inode->i_blkbits;
1402 const unsigned int blocksize = 1 << blockbits;
1403 u64 lblk_num = ((u64)srcpage->index << (PAGE_SHIFT - blockbits)) +
1404 (offs >> blockbits);
1408 if (unlikely(!dstpage))
1409 return llcrypt_encrypt_pagecache_blocks(srcpage, len, offs,
1412 if (WARN_ON_ONCE(!PageLocked(srcpage)))
1413 return ERR_PTR(-EINVAL);
1415 if (WARN_ON_ONCE(len <= 0 || !IS_ALIGNED(len | offs, blocksize)))
1416 return ERR_PTR(-EINVAL);
1418 /* Set PagePrivate2 for disambiguation in
1419 * osc_finalize_bounce_page().
1420 * It means cipher page was not allocated by llcrypt.
1422 SetPagePrivate2(dstpage);
1424 for (i = offs; i < offs + len; i += blocksize, lblk_num++) {
1425 err = llcrypt_encrypt_block(inode, srcpage, dstpage, blocksize,
1426 i, lblk_num, gfp_flags);
1428 return ERR_PTR(err);
1430 SetPagePrivate(dstpage);
1431 set_page_private(dstpage, (unsigned long)srcpage);
1436 * osc_finalize_bounce_page() - overlay to llcrypt_finalize_bounce_page
1438 * This overlay function is necessary to handle bounce pages
1439 * allocated by ourselves.
1441 static inline void osc_finalize_bounce_page(struct page **pagep)
1443 struct page *page = *pagep;
1445 /* PagePrivate2 was set in osc_encrypt_pagecache_blocks
1446 * to indicate the cipher page was allocated by ourselves.
1447 * So we must not free it via llcrypt.
1449 if (unlikely(!page || !PagePrivate2(page)))
1450 return llcrypt_finalize_bounce_page(pagep);
1452 if (llcrypt_is_bounce_page(page)) {
1453 *pagep = llcrypt_pagecache_page(page);
1454 ClearPagePrivate2(page);
1455 set_page_private(page, (unsigned long)NULL);
1456 ClearPagePrivate(page);
1459 #else /* !CONFIG_LL_ENCRYPTION */
1460 #define osc_encrypt_pagecache_blocks(srcpage, dstpage, len, offs, gfp_flags) \
1461 llcrypt_encrypt_pagecache_blocks(srcpage, len, offs, gfp_flags)
1462 #define osc_finalize_bounce_page(page) llcrypt_finalize_bounce_page(page)
1465 static inline void osc_release_bounce_pages(struct brw_page **pga,
1468 #ifdef HAVE_LUSTRE_CRYPTO
1469 struct page **pa = NULL;
1475 #ifdef CONFIG_LL_ENCRYPTION
1476 if (PageChecked(pga[0]->bp_page)) {
1477 OBD_ALLOC_PTR_ARRAY_LARGE(pa, page_count);
1483 for (i = 0; i < page_count; i++) {
1484 /* Bounce pages used by osc_encrypt_pagecache_blocks()
1485 * called from osc_brw_prep_request()
1486 * are identified thanks to the PageChecked flag.
1488 if (PageChecked(pga[i]->bp_page)) {
1490 pa[j++] = pga[i]->bp_page;
1491 osc_finalize_bounce_page(&pga[i]->bp_page);
1493 pga[i]->bp_count -= pga[i]->bp_count_diff;
1494 pga[i]->bp_off += pga[i]->bp_off_diff;
1498 obd_pool_put_pages_array(pa, j);
1499 OBD_FREE_PTR_ARRAY_LARGE(pa, page_count);
1505 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1506 u32 page_count, struct brw_page **pga,
1507 struct ptlrpc_request **reqp, int resend)
1509 struct ptlrpc_request *req;
1510 struct ptlrpc_bulk_desc *desc;
1511 struct ost_body *body;
1512 struct obd_ioobj *ioobj;
1513 struct niobuf_remote *niobuf;
1514 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1515 struct osc_brw_async_args *aa;
1516 struct req_capsule *pill;
1517 struct brw_page *pg_prev;
1519 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1520 struct inode *inode = NULL;
1521 bool directio = false;
1523 bool enable_checksum = true;
1524 struct cl_page *clpage;
1527 if (pga[0]->bp_page) {
1528 clpage = oap2cl_page(brw_page2oap(pga[0]));
1529 inode = clpage->cp_inode;
1530 if (clpage->cp_type == CPT_TRANSIENT)
1533 if (CFS_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1534 RETURN(-ENOMEM); /* Recoverable */
1535 if (CFS_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1536 RETURN(-EINVAL); /* Fatal */
1538 if ((cmd & OBD_BRW_WRITE) != 0) {
1540 req = ptlrpc_request_alloc_pool(cli->cl_import,
1542 &RQF_OST_BRW_WRITE);
1545 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1550 if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode) &&
1551 llcrypt_has_encryption_key(inode)) {
1552 struct page **pa = NULL;
1554 #ifdef CONFIG_LL_ENCRYPTION
1555 OBD_ALLOC_PTR_ARRAY_LARGE(pa, page_count);
1557 ptlrpc_request_free(req);
1561 rc = obd_pool_get_pages_array(pa, page_count);
1563 CDEBUG(D_SEC, "failed to allocate from enc pool: %d\n",
1565 ptlrpc_request_free(req);
1570 for (i = 0; i < page_count; i++) {
1571 struct brw_page *brwpg = pga[i];
1572 struct page *data_page = NULL;
1573 bool retried = false;
1574 bool lockedbymyself;
1576 (brwpg->bp_off & ~PAGE_MASK) + brwpg->bp_count;
1577 struct address_space *map_orig = NULL;
1581 nunits = round_up(nunits, LUSTRE_ENCRYPTION_UNIT_SIZE);
1582 /* The page can already be locked when we arrive here.
1583 * This is possible when cl_page_assume/vvp_page_assume
1584 * is stuck on wait_on_page_writeback with page lock
1585 * held. In this case there is no risk for the lock to
1586 * be released while we are doing our encryption
1587 * processing, because writeback against that page will
1588 * end in vvp_page_completion_write/cl_page_completion,
1589 * which means only once the page is fully processed.
1591 lockedbymyself = trylock_page(brwpg->bp_page);
1593 map_orig = brwpg->bp_page->mapping;
1594 brwpg->bp_page->mapping = inode->i_mapping;
1595 index_orig = brwpg->bp_page->index;
1596 clpage = oap2cl_page(brw_page2oap(brwpg));
1597 brwpg->bp_page->index = clpage->cp_page_index;
1600 osc_encrypt_pagecache_blocks(brwpg->bp_page,
1605 brwpg->bp_page->mapping = map_orig;
1606 brwpg->bp_page->index = index_orig;
1609 unlock_page(brwpg->bp_page);
1610 if (IS_ERR(data_page)) {
1611 rc = PTR_ERR(data_page);
1612 if (rc == -ENOMEM && !retried) {
1618 obd_pool_put_pages_array(pa + i,
1620 OBD_FREE_PTR_ARRAY_LARGE(pa,
1623 ptlrpc_request_free(req);
1626 /* Set PageChecked flag on bounce page for
1627 * disambiguation in osc_release_bounce_pages().
1629 SetPageChecked(data_page);
1630 brwpg->bp_page = data_page;
1631 /* there should be no gap in the middle of page array */
1632 if (i == page_count - 1) {
1633 struct osc_async_page *oap =
1634 brw_page2oap(brwpg);
1636 oa->o_size = oap->oap_count +
1637 oap->oap_obj_off + oap->oap_page_off;
1639 /* len is forced to nunits, and relative offset to 0
1640 * so store the old, clear text info
1642 brwpg->bp_count_diff = nunits - brwpg->bp_count;
1643 brwpg->bp_count = nunits;
1644 brwpg->bp_off_diff = brwpg->bp_off & ~PAGE_MASK;
1645 brwpg->bp_off = brwpg->bp_off & PAGE_MASK;
1649 OBD_FREE_PTR_ARRAY_LARGE(pa, page_count);
1650 } else if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1651 struct osc_async_page *oap = brw_page2oap(pga[0]);
1652 struct cl_page *clpage = oap2cl_page(oap);
1653 struct cl_object *clobj = clpage->cp_obj;
1654 struct cl_attr attr = { 0 };
1658 env = cl_env_get(&refcheck);
1661 ptlrpc_request_free(req);
1665 cl_object_attr_lock(clobj);
1666 rc = cl_object_attr_get(env, clobj, &attr);
1667 cl_object_attr_unlock(clobj);
1668 cl_env_put(env, &refcheck);
1670 ptlrpc_request_free(req);
1674 oa->o_size = attr.cat_size;
1675 } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode) &&
1676 llcrypt_has_encryption_key(inode)) {
1677 for (i = 0; i < page_count; i++) {
1678 struct brw_page *pg = pga[i];
1679 u32 nunits = (pg->bp_off & ~PAGE_MASK) + pg->bp_count;
1681 nunits = round_up(nunits, LUSTRE_ENCRYPTION_UNIT_SIZE);
1682 /* count/off are forced to cover the whole encryption
1683 * unit size so that all encrypted data is stored on the
1684 * OST, so adjust bp_{count,off}_diff for the size of
1687 pg->bp_count_diff = nunits - pg->bp_count;
1688 pg->bp_count = nunits;
1689 pg->bp_off_diff = pg->bp_off & ~PAGE_MASK;
1690 pg->bp_off = pg->bp_off & PAGE_MASK;
1694 for (niocount = i = 1; i < page_count; i++) {
1695 if (!can_merge_pages(pga[i - 1], pga[i]))
1699 pill = &req->rq_pill;
1700 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1702 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1703 niocount * sizeof(*niobuf));
1705 for (i = 0; i < page_count; i++) {
1706 short_io_size += pga[i]->bp_count;
1707 if (!inode || !IS_ENCRYPTED(inode) ||
1708 !llcrypt_has_encryption_key(inode)) {
1709 pga[i]->bp_count_diff = 0;
1710 pga[i]->bp_off_diff = 0;
1714 if (brw_page2oap(pga[0])->oap_brw_flags & OBD_BRW_RDMA_ONLY) {
1715 enable_checksum = false;
1720 /* Check if read/write is small enough to be a short io. */
1721 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1722 !imp_connect_shortio(cli->cl_import))
1725 /* If this is an empty RPC to old server, just ignore it */
1726 if (!short_io_size && !pga[0]->bp_page) {
1727 ptlrpc_request_free(req);
1731 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1732 opc == OST_READ ? 0 : short_io_size);
1733 if (opc == OST_READ)
1734 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1737 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1739 ptlrpc_request_free(req);
1742 osc_set_io_portal(req);
1744 ptlrpc_at_set_req_timeout(req);
1745 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1747 req->rq_no_retry_einprogress = 1;
1749 if (short_io_size != 0) {
1751 short_io_buf = NULL;
1755 desc = ptlrpc_prep_bulk_imp(req, page_count,
1756 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1757 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1758 PTLRPC_BULK_PUT_SINK),
1760 &ptlrpc_bulk_kiov_pin_ops);
1763 GOTO(out, rc = -ENOMEM);
1764 /* NB request now owns desc and will free it when it gets freed */
1765 desc->bd_is_rdma = gpu;
1767 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1768 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1769 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1770 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1772 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1774 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1775 * and from_kgid(), because they are asynchronous. Fortunately, variable
1776 * oa contains valid o_uid and o_gid in these two operations.
1777 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1778 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1779 * other process logic */
1780 body->oa.o_uid = oa->o_uid;
1781 body->oa.o_gid = oa->o_gid;
1783 obdo_to_ioobj(oa, ioobj);
1784 ioobj->ioo_bufcnt = niocount;
1785 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1786 * that might be send for this request. The actual number is decided
1787 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1788 * "max - 1" for old client compatibility sending "0", and also so the
1789 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1791 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1793 ioobj_max_brw_set(ioobj, 0);
1795 if (inode && IS_ENCRYPTED(inode) &&
1796 llcrypt_has_encryption_key(inode) &&
1797 !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
1798 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1799 body->oa.o_valid |= OBD_MD_FLFLAGS;
1800 body->oa.o_flags = 0;
1802 body->oa.o_flags |= LUSTRE_ENCRYPT_FL;
1805 if (short_io_size != 0) {
1806 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1807 body->oa.o_valid |= OBD_MD_FLFLAGS;
1808 body->oa.o_flags = 0;
1810 body->oa.o_flags |= OBD_FL_SHORT_IO;
1811 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1813 if (opc == OST_WRITE) {
1814 short_io_buf = req_capsule_client_get(pill,
1816 LASSERT(short_io_buf != NULL);
1820 LASSERT(page_count > 0);
1822 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1823 struct brw_page *pg = pga[i];
1824 int poff = pg->bp_off & ~PAGE_MASK;
1826 LASSERT(pg->bp_count > 0);
1827 /* make sure there is no gap in the middle of page array */
1828 LASSERTF(page_count == 1 ||
1829 (ergo(i == 0, poff + pg->bp_count == PAGE_SIZE) &&
1830 ergo(i > 0 && i < page_count - 1,
1831 poff == 0 && pg->bp_count == PAGE_SIZE) &&
1832 ergo(i == page_count - 1, poff == 0)),
1833 "i: %d/%d pg: %px off: %llu, count: %u\n",
1834 i, page_count, pg, pg->bp_off, pg->bp_count);
1835 LASSERTF(i == 0 || pg->bp_off > pg_prev->bp_off,
1836 "i %d p_c %u pg %px [pri %lu ind %lu] off %llu prev_pg %px [pri %lu ind %lu] off %llu\n",
1838 pg->bp_page, page_private(pg->bp_page),
1839 pg->bp_page->index, pg->bp_off,
1840 pg_prev->bp_page, page_private(pg_prev->bp_page),
1841 pg_prev->bp_page->index, pg_prev->bp_off);
1842 LASSERT((pga[0]->bp_flag & OBD_BRW_SRVLOCK) ==
1843 (pg->bp_flag & OBD_BRW_SRVLOCK));
1844 if (short_io_size != 0 && opc == OST_WRITE) {
1845 unsigned char *ptr = kmap_atomic(pg->bp_page);
1847 LASSERT(short_io_size >= requested_nob + pg->bp_count);
1848 memcpy(short_io_buf + requested_nob,
1852 } else if (short_io_size == 0) {
1853 desc->bd_frag_ops->add_kiov_frag(desc, pg->bp_page, poff,
1856 requested_nob += pg->bp_count;
1858 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1860 niobuf->rnb_len += pg->bp_count;
1862 niobuf->rnb_offset = pg->bp_off;
1863 niobuf->rnb_len = pg->bp_count;
1864 niobuf->rnb_flags = pg->bp_flag;
1867 if (CFS_FAIL_CHECK(OBD_FAIL_OSC_MARK_COMPRESSED))
1868 niobuf->rnb_flags |= OBD_BRW_COMPRESSED;
1871 LASSERTF((void *)(niobuf - niocount) ==
1872 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1873 "want %px - real %px\n",
1874 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1875 (void *)(niobuf - niocount));
1877 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1879 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1880 body->oa.o_valid |= OBD_MD_FLFLAGS;
1881 body->oa.o_flags = 0;
1883 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1886 if (osc_should_shrink_grant(cli))
1887 osc_shrink_grant_local(cli, &body->oa);
1889 if (!cli->cl_checksum || sptlrpc_flavor_has_bulk(&req->rq_flvr))
1890 enable_checksum = false;
1892 /* size[REQ_REC_OFF] still sizeof (*body) */
1893 if (opc == OST_WRITE) {
1894 if (enable_checksum) {
1895 /* store cl_cksum_type in a local variable since
1896 * it can be changed via lprocfs */
1897 enum cksum_types cksum_type = cli->cl_cksum_type;
1899 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1900 body->oa.o_flags = 0;
1902 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1904 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1906 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1907 requested_nob, page_count,
1909 &body->oa.o_cksum, resend);
1911 CDEBUG(D_PAGE, "failed to checksum: rc = %d\n",
1915 CDEBUG(D_PAGE | (resend ? D_HA : 0),
1916 "checksum at write origin: %x (%x)\n",
1917 body->oa.o_cksum, cksum_type);
1919 /* save this in 'oa', too, for later checking */
1920 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1921 oa->o_flags |= obd_cksum_type_pack(obd_name,
1924 /* clear out the checksum flag, in case this is a
1925 * resend but cl_checksum is no longer set. b=11238 */
1926 oa->o_valid &= ~OBD_MD_FLCKSUM;
1928 oa->o_cksum = body->oa.o_cksum;
1929 /* 1 RC per niobuf */
1930 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1931 sizeof(__u32) * niocount);
1933 if (enable_checksum) {
1934 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1935 body->oa.o_flags = 0;
1936 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1937 cli->cl_cksum_type);
1938 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1941 /* Client cksum has been already copied to wire obdo in previous
1942 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1943 * resent due to cksum error, this will allow Server to
1944 * check+dump pages on its side */
1946 ptlrpc_request_set_replen(req);
1948 aa = ptlrpc_req_async_args(aa, req);
1950 aa->aa_requested_nob = requested_nob;
1951 aa->aa_nio_count = niocount;
1952 aa->aa_page_count = page_count;
1956 INIT_LIST_HEAD(&aa->aa_oaps);
1959 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1960 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1961 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1962 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1966 ptlrpc_req_finished(req);
1970 char dbgcksum_file_name[PATH_MAX];
1972 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1973 struct brw_page **pga, __u32 server_cksum,
1981 /* will only keep dump of pages on first error for the same range in
1982 * file/fid, not during the resends/retries. */
1983 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1984 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1985 (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1986 libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1987 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1988 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1989 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1991 pga[page_count-1]->bp_off + pga[page_count-1]->bp_count - 1,
1992 client_cksum, server_cksum);
1993 CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1994 filp = filp_open(dbgcksum_file_name,
1995 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1999 CDEBUG(D_INFO, "%s: can't open to dump pages with "
2000 "checksum error: rc = %d\n", dbgcksum_file_name,
2003 CERROR("%s: can't open to dump pages with checksum "
2004 "error: rc = %d\n", dbgcksum_file_name, rc);
2008 for (i = 0; i < page_count; i++) {
2009 len = pga[i]->bp_count;
2010 buf = kmap(pga[i]->bp_page);
2012 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
2014 CERROR("%s: wanted to write %u but got %d "
2015 "error\n", dbgcksum_file_name, len, rc);
2021 kunmap(pga[i]->bp_page);
2024 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
2026 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
2027 filp_close(filp, NULL);
2029 libcfs_debug_dumplog();
2033 check_write_checksum(struct obdo *oa, const struct lnet_processid *peer,
2034 __u32 client_cksum, __u32 server_cksum,
2035 struct osc_brw_async_args *aa)
2037 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
2038 enum cksum_types cksum_type;
2039 obd_dif_csum_fn *fn = NULL;
2040 int sector_size = 0;
2045 if (server_cksum == client_cksum) {
2046 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2050 if (aa->aa_cli->cl_checksum_dump)
2051 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
2052 server_cksum, client_cksum);
2054 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
2057 switch (cksum_type) {
2058 case OBD_CKSUM_T10IP512:
2062 case OBD_CKSUM_T10IP4K:
2066 case OBD_CKSUM_T10CRC512:
2067 fn = obd_dif_crc_fn;
2070 case OBD_CKSUM_T10CRC4K:
2071 fn = obd_dif_crc_fn;
2079 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
2080 aa->aa_page_count, aa->aa_ppga,
2081 OST_WRITE, fn, sector_size,
2084 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
2085 aa->aa_ppga, OST_WRITE, cksum_type,
2089 msg = "failed to calculate the client write checksum";
2090 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
2091 msg = "the server did not use the checksum type specified in "
2092 "the original request - likely a protocol problem";
2093 else if (new_cksum == server_cksum)
2094 msg = "changed on the client after we checksummed it - "
2095 "likely false positive due to mmap IO (bug 11742)";
2096 else if (new_cksum == client_cksum)
2097 msg = "changed in transit before arrival at OST";
2099 msg = "changed in transit AND doesn't match the original - "
2100 "likely false positive due to mmap IO (bug 11742)";
2102 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
2103 DFID " object "DOSTID" extent [%llu-%llu], original "
2104 "client csum %x (type %x), server csum %x (type %x),"
2105 " client csum now %x\n",
2106 obd_name, msg, libcfs_nidstr(&peer->nid),
2107 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
2108 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
2109 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
2110 POSTID(&oa->o_oi), aa->aa_ppga[0]->bp_off,
2111 aa->aa_ppga[aa->aa_page_count - 1]->bp_off +
2112 aa->aa_ppga[aa->aa_page_count-1]->bp_count - 1,
2114 obd_cksum_type_unpack(aa->aa_oa->o_flags),
2115 server_cksum, cksum_type, new_cksum);
2119 /* Note rc enters this function as number of bytes transferred */
2120 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
2122 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
2123 struct client_obd *cli = aa->aa_cli;
2124 const char *obd_name = cli->cl_import->imp_obd->obd_name;
2125 const struct lnet_processid *peer =
2126 &req->rq_import->imp_connection->c_peer;
2127 struct ost_body *body;
2128 u32 client_cksum = 0;
2129 struct inode *inode = NULL;
2130 unsigned int blockbits = 0, blocksize = 0;
2131 struct cl_page *clpage;
2135 if (rc < 0 && rc != -EDQUOT) {
2136 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
2140 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
2141 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2143 DEBUG_REQ(D_INFO, req, "cannot unpack body");
2147 /* set/clear over quota flag for a uid/gid/projid */
2148 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
2149 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
2150 unsigned qid[LL_MAXQUOTAS] = {
2151 body->oa.o_uid, body->oa.o_gid,
2152 body->oa.o_projid };
2154 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
2155 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
2156 body->oa.o_valid, body->oa.o_flags);
2157 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
2161 osc_update_grant(cli, body);
2166 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
2167 client_cksum = aa->aa_oa->o_cksum; /* save for later */
2169 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2171 CERROR("%s: unexpected positive size %d\n",
2176 if (req->rq_bulk != NULL &&
2177 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
2180 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
2181 check_write_checksum(&body->oa, peer, client_cksum,
2182 body->oa.o_cksum, aa))
2185 rc = check_write_rcs(req, aa->aa_requested_nob,
2186 aa->aa_nio_count, aa->aa_page_count,
2191 /* The rest of this function executes only for OST_READs */
2193 if (req->rq_bulk == NULL) {
2194 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
2196 LASSERT(rc == req->rq_status);
2198 /* if unwrap_bulk failed, return -EAGAIN to retry */
2199 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
2202 GOTO(out, rc = -EAGAIN);
2204 if (rc > aa->aa_requested_nob) {
2205 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
2206 rc, aa->aa_requested_nob);
2210 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
2211 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
2212 rc, req->rq_bulk->bd_nob_transferred);
2216 if (req->rq_bulk == NULL) {
2218 int nob, pg_count, i = 0;
2221 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
2222 pg_count = aa->aa_page_count;
2223 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
2226 while (nob > 0 && pg_count > 0) {
2228 int count = aa->aa_ppga[i]->bp_count > nob ?
2229 nob : aa->aa_ppga[i]->bp_count;
2231 CDEBUG(D_CACHE, "page %p count %d\n",
2232 aa->aa_ppga[i]->bp_page, count);
2233 ptr = kmap_atomic(aa->aa_ppga[i]->bp_page);
2234 memcpy(ptr + (aa->aa_ppga[i]->bp_off & ~PAGE_MASK), buf,
2236 kunmap_atomic((void *) ptr);
2245 if (rc < aa->aa_requested_nob)
2246 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
2248 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2249 static int cksum_counter;
2250 u32 server_cksum = body->oa.o_cksum;
2254 enum cksum_types cksum_type;
2255 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
2256 body->oa.o_flags : 0;
2258 cksum_type = obd_cksum_type_unpack(o_flags);
2259 rc = osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2260 aa->aa_page_count, aa->aa_ppga,
2261 OST_READ, &client_cksum, false);
2265 if (req->rq_bulk != NULL &&
2266 !nid_same(&peer->nid, &req->rq_bulk->bd_sender)) {
2268 router = libcfs_nidstr(&req->rq_bulk->bd_sender);
2271 if (server_cksum != client_cksum) {
2272 struct ost_body *clbody;
2273 __u32 client_cksum2;
2274 u32 page_count = aa->aa_page_count;
2276 osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2277 page_count, aa->aa_ppga,
2278 OST_READ, &client_cksum2, true);
2279 clbody = req_capsule_client_get(&req->rq_pill,
2281 if (cli->cl_checksum_dump)
2282 dump_all_bulk_pages(&clbody->oa, page_count,
2283 aa->aa_ppga, server_cksum,
2286 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2287 "%s%s%s inode "DFID" object "DOSTID
2288 " extent [%llu-%llu], client %x/%x, "
2289 "server %x, cksum_type %x\n",
2291 libcfs_nidstr(&peer->nid),
2293 clbody->oa.o_valid & OBD_MD_FLFID ?
2294 clbody->oa.o_parent_seq : 0ULL,
2295 clbody->oa.o_valid & OBD_MD_FLFID ?
2296 clbody->oa.o_parent_oid : 0,
2297 clbody->oa.o_valid & OBD_MD_FLFID ?
2298 clbody->oa.o_parent_ver : 0,
2299 POSTID(&body->oa.o_oi),
2300 aa->aa_ppga[0]->bp_off,
2301 aa->aa_ppga[page_count-1]->bp_off +
2302 aa->aa_ppga[page_count-1]->bp_count - 1,
2303 client_cksum, client_cksum2,
2304 server_cksum, cksum_type);
2306 aa->aa_oa->o_cksum = client_cksum;
2310 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2313 } else if (unlikely(client_cksum)) {
2314 static int cksum_missed;
2317 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2318 CERROR("%s: checksum %u requested from %s but not sent\n",
2319 obd_name, cksum_missed,
2320 libcfs_nidstr(&peer->nid));
2325 /* get the inode from the first cl_page */
2326 clpage = oap2cl_page(brw_page2oap(aa->aa_ppga[0]));
2327 inode = clpage->cp_inode;
2328 if (clpage->cp_type == CPT_TRANSIENT && inode) {
2329 blockbits = inode->i_blkbits;
2330 blocksize = 1 << blockbits;
2332 if (inode && IS_ENCRYPTED(inode)) {
2335 if (!llcrypt_has_encryption_key(inode)) {
2336 CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2339 for (idx = 0; idx < aa->aa_page_count; idx++) {
2340 struct brw_page *brwpg = aa->aa_ppga[idx];
2341 unsigned int offs = 0;
2343 while (offs < PAGE_SIZE) {
2344 /* do not decrypt if page is all 0s */
2345 if (memchr_inv(page_address(brwpg->bp_page) + offs,
2346 0, LUSTRE_ENCRYPTION_UNIT_SIZE) == NULL) {
2347 /* if page is empty forward info to
2348 * upper layers (ll_io_zero_page) by
2349 * clearing PagePrivate2
2352 ClearPagePrivate2(brwpg->bp_page);
2357 /* This is direct IO case. Directly call
2358 * decrypt function that takes inode as
2359 * input parameter. Page does not need
2366 oap2cl_page(brw_page2oap(brwpg));
2368 ((u64)(clpage->cp_page_index) <<
2369 (PAGE_SHIFT - blockbits)) +
2370 (offs >> blockbits);
2373 LUSTRE_ENCRYPTION_UNIT_SIZE;
2374 i += blocksize, lblk_num++) {
2376 llcrypt_decrypt_block_inplace(
2377 inode, brwpg->bp_page,
2384 rc = llcrypt_decrypt_pagecache_blocks(
2386 LUSTRE_ENCRYPTION_UNIT_SIZE,
2392 offs += LUSTRE_ENCRYPTION_UNIT_SIZE;
2399 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2400 aa->aa_oa, &body->oa);
2405 static int osc_brw_redo_request(struct ptlrpc_request *request,
2406 struct osc_brw_async_args *aa, int rc)
2408 struct ptlrpc_request *new_req;
2409 struct osc_brw_async_args *new_aa;
2412 /* The below message is checked in replay-ost-single.sh test_8ae*/
2413 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2414 "redo for recoverable error %d", rc);
2416 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2417 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2418 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2419 aa->aa_ppga, &new_req, 1);
2424 LASSERTF(request == aa->aa_request,
2425 "request %p != aa_request %p\n",
2426 request, aa->aa_request);
2428 * New request takes over pga and oaps from old request.
2429 * Note that copying a list_head doesn't work, need to move it...
2432 new_req->rq_interpret_reply = request->rq_interpret_reply;
2433 new_req->rq_async_args = request->rq_async_args;
2434 new_req->rq_commit_cb = request->rq_commit_cb;
2435 /* cap resend delay to the current request timeout, this is similar to
2436 * what ptlrpc does (see after_reply()) */
2437 if (aa->aa_resends > new_req->rq_timeout)
2438 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2440 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2441 new_req->rq_generation_set = 1;
2442 new_req->rq_import_generation = request->rq_import_generation;
2444 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2446 INIT_LIST_HEAD(&new_aa->aa_oaps);
2447 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2448 INIT_LIST_HEAD(&new_aa->aa_exts);
2449 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2450 new_aa->aa_resends = aa->aa_resends;
2452 if (aa->aa_request) {
2453 ptlrpc_req_finished(aa->aa_request);
2454 new_aa->aa_request = ptlrpc_request_addref(new_req);
2457 /* XXX: This code will run into problem if we're going to support
2458 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2459 * and wait for all of them to be finished. We should inherit request
2460 * set from old request. */
2461 ptlrpcd_add_req(new_req);
2463 DEBUG_REQ(D_INFO, new_req, "new request");
2468 * ugh, we want disk allocation on the target to happen in offset order. we'll
2469 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2470 * fine for our small page arrays and doesn't require allocation. its an
2471 * insertion sort that swaps elements that are strides apart, shrinking the
2472 * stride down until its '1' and the array is sorted.
2474 static void sort_brw_pages(struct brw_page **array, int num)
2477 struct brw_page *tmp;
2481 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2486 for (i = stride ; i < num ; i++) {
2489 while (j >= stride && array[j - stride]->bp_off > tmp->bp_off) {
2490 array[j] = array[j - stride];
2495 } while (stride > 1);
2498 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2500 LASSERT(ppga != NULL);
2501 OBD_FREE_PTR_ARRAY_LARGE(ppga, count);
2504 /* this is trying to propogate async writeback errors back up to the
2505 * application. As an async write fails we record the error code for later if
2506 * the app does an fsync. As long as errors persist we force future rpcs to be
2507 * sync so that the app can get a sync error and break the cycle of queueing
2508 * pages for which writeback will fail.
2510 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2517 ar->ar_force_sync = 1;
2518 ar->ar_min_xid = ptlrpc_sample_next_xid();
2523 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2524 ar->ar_force_sync = 0;
2527 static int brw_interpret(const struct lu_env *env,
2528 struct ptlrpc_request *req, void *args, int rc)
2530 struct osc_brw_async_args *aa = args;
2531 struct client_obd *cli = aa->aa_cli;
2532 unsigned long transferred = 0;
2533 struct cl_object *obj = NULL;
2534 struct osc_async_page *last;
2535 struct osc_extent *ext;
2536 struct osc_extent *tmp;
2537 struct lov_oinfo *loi;
2541 ext = list_first_entry(&aa->aa_exts, struct osc_extent, oe_link);
2543 rc = osc_brw_fini_request(req, rc);
2544 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2546 /* restore clear text pages */
2547 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2550 * When server returns -EINPROGRESS, client should always retry
2551 * regardless of the number of times the bulk was resent already.
2553 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2554 if (req->rq_import_generation !=
2555 req->rq_import->imp_generation) {
2556 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2557 ""DOSTID", rc = %d.\n",
2558 req->rq_import->imp_obd->obd_name,
2559 POSTID(&aa->aa_oa->o_oi), rc);
2560 } else if (rc == -EINPROGRESS ||
2561 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2562 rc = osc_brw_redo_request(req, aa, rc);
2564 CERROR("%s: too many resent retries for object: "
2565 "%llu:%llu, rc = %d.\n",
2566 req->rq_import->imp_obd->obd_name,
2567 POSTID(&aa->aa_oa->o_oi), rc);
2572 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2576 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2577 obj = osc2cl(ext->oe_obj);
2578 loi = cl2osc(obj)->oo_oinfo;
2581 struct obdo *oa = aa->aa_oa;
2582 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2583 unsigned long valid = 0;
2585 cl_object_attr_lock(obj);
2586 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2587 attr->cat_blocks = oa->o_blocks;
2588 valid |= CAT_BLOCKS;
2590 if (oa->o_valid & OBD_MD_FLMTIME) {
2591 attr->cat_mtime = oa->o_mtime;
2594 if (oa->o_valid & OBD_MD_FLATIME) {
2595 attr->cat_atime = oa->o_atime;
2598 if (oa->o_valid & OBD_MD_FLCTIME) {
2599 attr->cat_ctime = oa->o_ctime;
2603 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2604 loff_t last_off = last->oap_count + last->oap_obj_off +
2607 /* Change file size if this is an out of quota or
2608 * direct IO write and it extends the file size */
2609 if (loi->loi_lvb.lvb_size < last_off) {
2610 attr->cat_size = last_off;
2613 /* Extend KMS if it's not a lockless write */
2614 if (loi->loi_kms < last_off &&
2615 oap2osc_page(last)->ops_srvlock == 0) {
2616 attr->cat_kms = last_off;
2622 cl_object_attr_update(env, obj, attr, valid);
2623 cl_object_attr_unlock(obj);
2625 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2628 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0) {
2629 osc_inc_unstable_pages(req);
2631 * If req->rq_committed is set, it means that the dirty pages
2632 * have already committed into the stable storage on OSTs
2633 * (i.e. Direct I/O).
2635 if (!req->rq_committed)
2636 cl_object_dirty_for_sync(env, cl_object_top(obj));
2639 if (aa->aa_request) {
2640 __u64 xid = ptlrpc_req_xid(req);
2642 ptlrpc_req_finished(req);
2643 if (xid && lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2644 spin_lock(&cli->cl_loi_list_lock);
2645 osc_process_ar(&cli->cl_ar, xid, rc);
2646 osc_process_ar(&loi->loi_ar, xid, rc);
2647 spin_unlock(&cli->cl_loi_list_lock);
2650 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2651 list_del_init(&ext->oe_link);
2652 osc_extent_finish(env, ext, 1,
2653 rc && req->rq_no_delay ? -EAGAIN : rc);
2655 LASSERT(list_empty(&aa->aa_exts));
2656 LASSERT(list_empty(&aa->aa_oaps));
2658 transferred = (req->rq_bulk == NULL ? /* short io */
2659 aa->aa_requested_nob :
2660 req->rq_bulk->bd_nob_transferred);
2662 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2663 ptlrpc_lprocfs_brw(req, transferred);
2665 spin_lock(&cli->cl_loi_list_lock);
2666 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2667 * is called so we know whether to go to sync BRWs or wait for more
2668 * RPCs to complete */
2669 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2670 cli->cl_w_in_flight--;
2672 cli->cl_r_in_flight--;
2673 osc_wake_cache_waiters(cli);
2674 spin_unlock(&cli->cl_loi_list_lock);
2676 osc_io_unplug(env, cli, NULL);
2680 static void brw_commit(struct ptlrpc_request *req)
2682 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2683 * this called via the rq_commit_cb, I need to ensure
2684 * osc_dec_unstable_pages is still called. Otherwise unstable
2685 * pages may be leaked. */
2686 spin_lock(&req->rq_lock);
2687 if (likely(req->rq_unstable)) {
2688 req->rq_unstable = 0;
2689 spin_unlock(&req->rq_lock);
2691 osc_dec_unstable_pages(req);
2693 req->rq_committed = 1;
2694 spin_unlock(&req->rq_lock);
2699 * Build an RPC by the list of extent @ext_list. The caller must ensure
2700 * that the total pages in this list are NOT over max pages per RPC.
2701 * Extents in the list must be in OES_RPC state.
2703 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2704 struct list_head *ext_list, int cmd)
2706 struct ptlrpc_request *req = NULL;
2707 struct osc_extent *ext;
2708 struct brw_page **pga = NULL;
2709 struct osc_brw_async_args *aa = NULL;
2710 struct obdo *oa = NULL;
2711 struct osc_async_page *oap;
2712 struct osc_object *obj = NULL;
2713 struct cl_req_attr *crattr = NULL;
2714 loff_t starting_offset = OBD_OBJECT_EOF;
2715 loff_t ending_offset = 0;
2716 /* '1' for consistency with code that checks !mpflag to restore */
2720 bool soft_sync = false;
2721 bool ndelay = false;
2725 __u32 layout_version = 0;
2726 LIST_HEAD(rpc_list);
2727 struct ost_body *body;
2729 LASSERT(!list_empty(ext_list));
2731 /* add pages into rpc_list to build BRW rpc */
2732 list_for_each_entry(ext, ext_list, oe_link) {
2733 LASSERT(ext->oe_state == OES_RPC);
2734 mem_tight |= ext->oe_memalloc;
2735 grant += ext->oe_grants;
2736 page_count += ext->oe_nr_pages;
2737 layout_version = max(layout_version, ext->oe_layout_version);
2742 soft_sync = osc_over_unstable_soft_limit(cli);
2744 mpflag = memalloc_noreclaim_save();
2746 OBD_ALLOC_PTR_ARRAY_LARGE(pga, page_count);
2748 GOTO(out, rc = -ENOMEM);
2750 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2752 GOTO(out, rc = -ENOMEM);
2755 list_for_each_entry(ext, ext_list, oe_link) {
2756 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2758 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2760 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2761 pga[i] = &oap->oap_brw_page;
2762 pga[i]->bp_off = oap->oap_obj_off + oap->oap_page_off;
2765 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2766 if (starting_offset == OBD_OBJECT_EOF ||
2767 starting_offset > oap->oap_obj_off) {
2768 starting_offset = oap->oap_obj_off;
2770 CDEBUG(D_CACHE, "page i:%d, oap->oap_obj_off %llu, oap->oap_page_off %u\n",
2771 i, oap->oap_obj_off, oap->oap_page_off);
2772 LASSERT(oap->oap_page_off == 0);
2774 if (ending_offset < oap->oap_obj_off + oap->oap_count) {
2775 ending_offset = oap->oap_obj_off +
2778 LASSERT(oap->oap_page_off + oap->oap_count ==
2786 /* first page in the list */
2787 oap = list_first_entry(&rpc_list, typeof(*oap), oap_rpc_item);
2789 crattr = &osc_env_info(env)->oti_req_attr;
2790 memset(crattr, 0, sizeof(*crattr));
2791 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2792 crattr->cra_flags = ~0ULL;
2793 crattr->cra_page = oap2cl_page(oap);
2794 crattr->cra_oa = oa;
2795 cl_req_attr_set(env, osc2cl(obj), crattr);
2797 if (cmd == OBD_BRW_WRITE) {
2798 oa->o_grant_used = grant;
2799 if (layout_version > 0) {
2800 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2801 PFID(&oa->o_oi.oi_fid), layout_version);
2803 oa->o_layout_version = layout_version;
2804 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2808 sort_brw_pages(pga, page_count);
2809 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2811 CERROR("prep_req failed: %d\n", rc);
2815 req->rq_commit_cb = brw_commit;
2816 req->rq_interpret_reply = brw_interpret;
2817 req->rq_memalloc = mem_tight != 0;
2819 req->rq_no_resend = req->rq_no_delay = 1;
2820 /* probably set a shorter timeout value.
2821 * to handle ETIMEDOUT in brw_interpret() correctly. */
2822 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2825 /* Need to update the timestamps after the request is built in case
2826 * we race with setattr (locally or in queue at OST). If OST gets
2827 * later setattr before earlier BRW (as determined by the request xid),
2828 * the OST will not use BRW timestamps. Sadly, there is no obvious
2829 * way to do this in a single call. bug 10150 */
2830 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2831 crattr->cra_oa = &body->oa;
2832 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2833 cl_req_attr_set(env, osc2cl(obj), crattr);
2834 lustre_msg_set_uid_gid(req->rq_reqmsg, &crattr->cra_uid,
2836 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2838 aa = ptlrpc_req_async_args(aa, req);
2839 INIT_LIST_HEAD(&aa->aa_oaps);
2840 list_splice_init(&rpc_list, &aa->aa_oaps);
2841 INIT_LIST_HEAD(&aa->aa_exts);
2842 list_splice_init(ext_list, &aa->aa_exts);
2843 aa->aa_request = ptlrpc_request_addref(req);
2845 spin_lock(&cli->cl_loi_list_lock);
2846 starting_offset >>= PAGE_SHIFT;
2847 ending_offset >>= PAGE_SHIFT;
2848 if (cmd == OBD_BRW_READ) {
2849 cli->cl_r_in_flight++;
2850 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2851 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2852 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2853 starting_offset + 1);
2855 cli->cl_w_in_flight++;
2856 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2857 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2858 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2859 starting_offset + 1);
2861 spin_unlock(&cli->cl_loi_list_lock);
2863 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2864 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2865 if (libcfs_debug & D_IOTRACE) {
2868 fid.f_seq = crattr->cra_oa->o_parent_seq;
2869 fid.f_oid = crattr->cra_oa->o_parent_oid;
2870 fid.f_ver = crattr->cra_oa->o_parent_ver;
2872 DFID": %d %s pages, start %lld, end %lld, now %ur/%uw in flight\n",
2873 PFID(&fid), page_count,
2874 cmd == OBD_BRW_READ ? "read" : "write", starting_offset,
2875 ending_offset, cli->cl_r_in_flight, cli->cl_w_in_flight);
2877 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2879 ptlrpcd_add_req(req);
2885 memalloc_noreclaim_restore(mpflag);
2888 LASSERT(req == NULL);
2891 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2893 osc_release_bounce_pages(pga, page_count);
2894 osc_release_ppga(pga, page_count);
2896 /* this should happen rarely and is pretty bad, it makes the
2897 * pending list not follow the dirty order
2899 while ((ext = list_first_entry_or_null(ext_list,
2901 oe_link)) != NULL) {
2902 list_del_init(&ext->oe_link);
2903 osc_extent_finish(env, ext, 0, rc);
2909 /* This is to refresh our lock in face of no RPCs. */
2910 void osc_send_empty_rpc(struct osc_object *osc, pgoff_t start)
2912 struct ptlrpc_request *req;
2914 struct brw_page bpg = { .bp_off = start, .bp_count = 1};
2915 struct brw_page *pga = &bpg;
2918 memset(&oa, 0, sizeof(oa));
2919 oa.o_oi = osc->oo_oinfo->loi_oi;
2920 oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
2921 /* For updated servers - don't do a read */
2922 oa.o_flags = OBD_FL_NORPC;
2924 rc = osc_brw_prep_request(OBD_BRW_READ, osc_cli(osc), &oa, 1, &pga,
2927 /* If we succeeded we ship it off, if not there's no point in doing
2928 * anything. Also no resends.
2929 * No interpret callback, no commit callback.
2932 req->rq_no_resend = 1;
2933 ptlrpcd_add_req(req);
2937 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2941 LASSERT(lock != NULL);
2943 lock_res_and_lock(lock);
2945 if (lock->l_ast_data == NULL)
2946 lock->l_ast_data = data;
2947 if (lock->l_ast_data == data)
2950 unlock_res_and_lock(lock);
2955 static int osc_enqueue_fini(struct ptlrpc_request *req,
2956 osc_enqueue_upcall_f upcall,
2957 void *cookie, struct lustre_handle *lockh,
2958 enum ldlm_mode mode, __u64 *flags,
2959 bool speculative, int errcode)
2961 bool intent = *flags & LDLM_FL_HAS_INTENT;
2965 /* The request was created before ldlm_cli_enqueue call. */
2966 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2967 struct ldlm_reply *rep;
2969 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2970 LASSERT(rep != NULL);
2972 rep->lock_policy_res1 =
2973 ptlrpc_status_ntoh(rep->lock_policy_res1);
2974 if (rep->lock_policy_res1)
2975 errcode = rep->lock_policy_res1;
2977 *flags |= LDLM_FL_LVB_READY;
2978 } else if (errcode == ELDLM_OK) {
2979 *flags |= LDLM_FL_LVB_READY;
2982 /* Call the update callback. */
2983 rc = (*upcall)(cookie, lockh, errcode);
2985 /* release the reference taken in ldlm_cli_enqueue() */
2986 if (errcode == ELDLM_LOCK_MATCHED)
2988 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2989 ldlm_lock_decref(lockh, mode);
2994 static int osc_enqueue_interpret(const struct lu_env *env,
2995 struct ptlrpc_request *req,
2998 struct osc_enqueue_args *aa = args;
2999 struct ldlm_lock *lock;
3000 struct lustre_handle *lockh = &aa->oa_lockh;
3001 enum ldlm_mode mode = aa->oa_mode;
3002 struct ost_lvb *lvb = aa->oa_lvb;
3003 __u32 lvb_len = sizeof(*lvb);
3005 struct ldlm_enqueue_info einfo = {
3006 .ei_type = aa->oa_type,
3012 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3014 lock = ldlm_handle2lock(lockh);
3015 LASSERTF(lock != NULL,
3016 "lockh %#llx, req %px, aa %px - client evicted?\n",
3017 lockh->cookie, req, aa);
3019 /* Take an additional reference so that a blocking AST that
3020 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3021 * to arrive after an upcall has been executed by
3022 * osc_enqueue_fini(). */
3023 ldlm_lock_addref(lockh, mode);
3025 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
3026 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
3028 /* Let CP AST to grant the lock first. */
3029 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3031 if (aa->oa_speculative) {
3032 LASSERT(aa->oa_lvb == NULL);
3033 LASSERT(aa->oa_flags == NULL);
3034 aa->oa_flags = &flags;
3037 /* Complete obtaining the lock procedure. */
3038 rc = ldlm_cli_enqueue_fini(aa->oa_exp, &req->rq_pill, &einfo, 1,
3039 aa->oa_flags, lvb, lvb_len, lockh, rc,
3041 /* Complete osc stuff. */
3042 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
3043 aa->oa_flags, aa->oa_speculative, rc);
3045 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3047 ldlm_lock_decref(lockh, mode);
3048 LDLM_LOCK_PUT(lock);
3052 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3053 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3054 * other synchronous requests, however keeping some locks and trying to obtain
3055 * others may take a considerable amount of time in a case of ost failure; and
3056 * when other sync requests do not get released lock from a client, the client
3057 * is evicted from the cluster -- such scenarious make the life difficult, so
3058 * release locks just after they are obtained. */
3059 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3060 __u64 *flags, union ldlm_policy_data *policy,
3061 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
3062 void *cookie, struct ldlm_enqueue_info *einfo,
3063 struct ptlrpc_request_set *rqset, int async,
3066 struct obd_device *obd = exp->exp_obd;
3067 struct lustre_handle lockh = { 0 };
3068 struct ptlrpc_request *req = NULL;
3069 int intent = *flags & LDLM_FL_HAS_INTENT;
3070 __u64 search_flags = *flags;
3071 __u64 match_flags = 0;
3072 enum ldlm_mode mode;
3076 /* Filesystem lock extents are extended to page boundaries so that
3077 * dealing with the page cache is a little smoother. */
3078 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
3079 policy->l_extent.end |= ~PAGE_MASK;
3081 /* Next, search for already existing extent locks that will cover us */
3082 /* If we're trying to read, we also search for an existing PW lock. The
3083 * VFS and page cache already protect us locally, so lots of readers/
3084 * writers can share a single PW lock.
3086 * There are problems with conversion deadlocks, so instead of
3087 * converting a read lock to a write lock, we'll just enqueue a new
3090 * At some point we should cancel the read lock instead of making them
3091 * send us a blocking callback, but there are problems with canceling
3092 * locks out from other users right now, too. */
3093 mode = einfo->ei_mode;
3094 if (einfo->ei_mode == LCK_PR)
3096 /* Normal lock requests must wait for the LVB to be ready before
3097 * matching a lock; speculative lock requests do not need to,
3098 * because they will not actually use the lock. */
3100 search_flags |= LDLM_FL_LVB_READY;
3102 search_flags |= LDLM_FL_BLOCK_GRANTED;
3103 if (mode == LCK_GROUP)
3104 match_flags = LDLM_MATCH_GROUP;
3105 mode = ldlm_lock_match_with_skip(obd->obd_namespace, search_flags, 0,
3106 res_id, einfo->ei_type, policy, mode,
3107 &lockh, match_flags);
3109 struct ldlm_lock *matched;
3111 if (*flags & LDLM_FL_TEST_LOCK)
3114 matched = ldlm_handle2lock(&lockh);
3116 /* This DLM lock request is speculative, and does not
3117 * have an associated IO request. Therefore if there
3118 * is already a DLM lock, it wll just inform the
3119 * caller to cancel the request for this stripe.*/
3120 lock_res_and_lock(matched);
3121 if (ldlm_extent_equal(&policy->l_extent,
3122 &matched->l_policy_data.l_extent))
3126 unlock_res_and_lock(matched);
3128 ldlm_lock_decref(&lockh, mode);
3129 LDLM_LOCK_PUT(matched);
3131 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
3132 *flags |= LDLM_FL_LVB_READY;
3134 /* We already have a lock, and it's referenced. */
3135 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
3137 ldlm_lock_decref(&lockh, mode);
3138 LDLM_LOCK_PUT(matched);
3141 ldlm_lock_decref(&lockh, mode);
3142 LDLM_LOCK_PUT(matched);
3146 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
3149 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3150 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3152 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3153 sizeof(*lvb), LVB_T_OST, &lockh, async);
3156 struct osc_enqueue_args *aa;
3157 aa = ptlrpc_req_async_args(aa, req);
3159 aa->oa_mode = einfo->ei_mode;
3160 aa->oa_type = einfo->ei_type;
3161 lustre_handle_copy(&aa->oa_lockh, &lockh);
3162 aa->oa_upcall = upcall;
3163 aa->oa_cookie = cookie;
3164 aa->oa_speculative = speculative;
3166 aa->oa_flags = flags;
3169 /* speculative locks are essentially to enqueue
3170 * a DLM lock in advance, so we don't care
3171 * about the result of the enqueue. */
3173 aa->oa_flags = NULL;
3176 req->rq_interpret_reply = osc_enqueue_interpret;
3177 ptlrpc_set_add_req(rqset, req);
3182 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
3183 flags, speculative, rc);
3188 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
3189 struct ldlm_res_id *res_id, enum ldlm_type type,
3190 union ldlm_policy_data *policy, enum ldlm_mode mode,
3191 __u64 *flags, struct osc_object *obj,
3192 struct lustre_handle *lockh, enum ldlm_match_flags match_flags)
3194 struct obd_device *obd = exp->exp_obd;
3195 __u64 lflags = *flags;
3199 if (CFS_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3202 /* Filesystem lock extents are extended to page boundaries so that
3203 * dealing with the page cache is a little smoother */
3204 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
3205 policy->l_extent.end |= ~PAGE_MASK;
3207 /* Next, search for already existing extent locks that will cover us */
3208 rc = ldlm_lock_match_with_skip(obd->obd_namespace, lflags, 0,
3209 res_id, type, policy, mode, lockh,
3211 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
3215 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3217 LASSERT(lock != NULL);
3218 if (osc_set_lock_data(lock, obj)) {
3219 lock_res_and_lock(lock);
3220 if (!ldlm_is_lvb_cached(lock)) {
3221 LASSERT(lock->l_ast_data == obj);
3222 osc_lock_lvb_update(env, obj, lock, NULL);
3223 ldlm_set_lvb_cached(lock);
3225 unlock_res_and_lock(lock);
3227 ldlm_lock_decref(lockh, rc);
3230 LDLM_LOCK_PUT(lock);
3235 static int osc_statfs_interpret(const struct lu_env *env,
3236 struct ptlrpc_request *req, void *args, int rc)
3238 struct osc_async_args *aa = args;
3239 struct obd_statfs *msfs;
3244 * The request has in fact never been sent due to issues at
3245 * a higher level (LOV). Exit immediately since the caller
3246 * is aware of the problem and takes care of the clean up.
3250 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3251 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3257 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3259 GOTO(out, rc = -EPROTO);
3261 *aa->aa_oi->oi_osfs = *msfs;
3263 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3268 static int osc_statfs_async(struct obd_export *exp,
3269 struct obd_info *oinfo, time64_t max_age,
3270 struct ptlrpc_request_set *rqset)
3272 struct obd_device *obd = class_exp2obd(exp);
3273 struct ptlrpc_request *req;
3274 struct osc_async_args *aa;
3278 if (obd->obd_osfs_age >= max_age) {
3280 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
3281 obd->obd_name, &obd->obd_osfs,
3282 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
3283 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
3284 spin_lock(&obd->obd_osfs_lock);
3285 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
3286 spin_unlock(&obd->obd_osfs_lock);
3287 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
3288 if (oinfo->oi_cb_up)
3289 oinfo->oi_cb_up(oinfo, 0);
3294 /* We could possibly pass max_age in the request (as an absolute
3295 * timestamp or a "seconds.usec ago") so the target can avoid doing
3296 * extra calls into the filesystem if that isn't necessary (e.g.
3297 * during mount that would help a bit). Having relative timestamps
3298 * is not so great if request processing is slow, while absolute
3299 * timestamps are not ideal because they need time synchronization. */
3300 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3304 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3306 ptlrpc_request_free(req);
3309 ptlrpc_request_set_replen(req);
3310 req->rq_request_portal = OST_CREATE_PORTAL;
3311 ptlrpc_at_set_req_timeout(req);
3313 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3314 /* procfs requests not want stat in wait for avoid deadlock */
3315 req->rq_no_resend = 1;
3316 req->rq_no_delay = 1;
3319 req->rq_interpret_reply = osc_statfs_interpret;
3320 aa = ptlrpc_req_async_args(aa, req);
3323 ptlrpc_set_add_req(rqset, req);
3327 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
3328 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
3330 struct obd_device *obd = class_exp2obd(exp);
3331 struct obd_statfs *msfs;
3332 struct ptlrpc_request *req;
3333 struct obd_import *imp, *imp0;
3337 /*Since the request might also come from lprocfs, so we need
3338 *sync this with client_disconnect_export Bug15684
3340 with_imp_locked(obd, imp0, rc)
3341 imp = class_import_get(imp0);
3345 /* We could possibly pass max_age in the request (as an absolute
3346 * timestamp or a "seconds.usec ago") so the target can avoid doing
3347 * extra calls into the filesystem if that isn't necessary (e.g.
3348 * during mount that would help a bit). Having relative timestamps
3349 * is not so great if request processing is slow, while absolute
3350 * timestamps are not ideal because they need time synchronization. */
3351 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3353 class_import_put(imp);
3358 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3360 ptlrpc_request_free(req);
3363 ptlrpc_request_set_replen(req);
3364 req->rq_request_portal = OST_CREATE_PORTAL;
3365 ptlrpc_at_set_req_timeout(req);
3367 if (flags & OBD_STATFS_NODELAY) {
3368 /* procfs requests not want stat in wait for avoid deadlock */
3369 req->rq_no_resend = 1;
3370 req->rq_no_delay = 1;
3373 rc = ptlrpc_queue_wait(req);
3377 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3379 GOTO(out, rc = -EPROTO);
3385 ptlrpc_req_finished(req);
3389 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3390 void *karg, void __user *uarg)
3392 struct obd_device *obd = exp->exp_obd;
3393 struct obd_ioctl_data *data;
3397 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
3398 obd->obd_name, cmd, len, karg, uarg);
3400 if (!try_module_get(THIS_MODULE)) {
3401 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3402 module_name(THIS_MODULE));
3407 case OBD_IOC_CLIENT_RECOVER:
3408 if (unlikely(karg == NULL)) {
3409 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL",
3414 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3415 data->ioc_inlbuf1, 0);
3419 case OBD_IOC_GETATTR:
3420 if (unlikely(karg == NULL)) {
3421 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL",
3426 rc = obd_getattr(NULL, exp, &data->ioc_obdo1);
3428 #ifdef IOC_OSC_SET_ACTIVE
3429 case_OBD_IOC_DEPRECATED_FT(IOC_OSC_SET_ACTIVE, obd->obd_name, 2, 17);
3431 case OBD_IOC_SET_ACTIVE:
3432 if (unlikely(karg == NULL)) {
3433 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL",
3438 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3442 rc = OBD_IOC_DEBUG(D_IOCTL, obd->obd_name, cmd, "unrecognized",
3447 module_put(THIS_MODULE);
3451 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3452 u32 keylen, void *key, u32 vallen, void *val,
3453 struct ptlrpc_request_set *set)
3455 struct ptlrpc_request *req;
3456 struct obd_device *obd = exp->exp_obd;
3457 struct obd_import *imp = class_exp2cliimp(exp);
3462 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3464 if (KEY_IS(KEY_CHECKSUM)) {
3465 if (vallen != sizeof(int))
3467 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3471 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3472 sptlrpc_conf_client_adapt(obd);
3476 if (KEY_IS(KEY_FLUSH_CTX)) {
3477 sptlrpc_import_flush_my_ctx(imp);
3481 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3482 struct client_obd *cli = &obd->u.cli;
3483 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3484 long target = *(long *)val;
3486 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3491 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3495 * We pass all other commands directly to OST. Since nobody calls osc
3496 * methods directly and everybody is supposed to go through LOV, we
3497 * assume lov checked invalid values for us.
3498 * The only recognised values so far are evict_by_nid and mds_conn.
3499 * Even if something bad goes through, we'd get a -EINVAL from OST
3503 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3504 &RQF_OST_SET_GRANT_INFO :
3509 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3510 RCL_CLIENT, keylen);
3511 if (!KEY_IS(KEY_GRANT_SHRINK))
3512 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3513 RCL_CLIENT, vallen);
3514 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3516 ptlrpc_request_free(req);
3520 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3521 memcpy(tmp, key, keylen);
3522 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3525 memcpy(tmp, val, vallen);
3527 if (KEY_IS(KEY_GRANT_SHRINK)) {
3528 struct osc_grant_args *aa;
3531 aa = ptlrpc_req_async_args(aa, req);
3532 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3534 ptlrpc_req_finished(req);
3537 *oa = ((struct ost_body *)val)->oa;
3539 req->rq_interpret_reply = osc_shrink_grant_interpret;
3542 ptlrpc_request_set_replen(req);
3543 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3544 LASSERT(set != NULL);
3545 ptlrpc_set_add_req(set, req);
3546 ptlrpc_check_set(NULL, set);
3548 ptlrpcd_add_req(req);
3553 EXPORT_SYMBOL(osc_set_info_async);
3555 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3556 struct obd_device *obd, struct obd_uuid *cluuid,
3557 struct obd_connect_data *data, void *localdata)
3559 struct client_obd *cli = &obd->u.cli;
3561 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3565 spin_lock(&cli->cl_loi_list_lock);
3566 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3567 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3568 /* restore ocd_grant_blkbits as client page bits */
3569 data->ocd_grant_blkbits = PAGE_SHIFT;
3570 grant += cli->cl_dirty_grant;
3572 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3574 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3575 lost_grant = cli->cl_lost_grant;
3576 cli->cl_lost_grant = 0;
3577 spin_unlock(&cli->cl_loi_list_lock);
3579 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3580 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3581 data->ocd_version, data->ocd_grant, lost_grant);
3586 EXPORT_SYMBOL(osc_reconnect);
3588 int osc_disconnect(struct obd_export *exp)
3590 struct obd_device *obd = class_exp2obd(exp);
3593 rc = client_disconnect_export(exp);
3595 * Initially we put del_shrink_grant before disconnect_export, but it
3596 * causes the following problem if setup (connect) and cleanup
3597 * (disconnect) are tangled together.
3598 * connect p1 disconnect p2
3599 * ptlrpc_connect_import
3600 * ............... class_manual_cleanup
3603 * ptlrpc_connect_interrupt
3605 * add this client to shrink list
3607 * Bang! grant shrink thread trigger the shrink. BUG18662
3609 osc_del_grant_list(&obd->u.cli);
3612 EXPORT_SYMBOL(osc_disconnect);
3614 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3615 struct hlist_node *hnode, void *arg)
3617 struct lu_env *env = arg;
3618 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3619 struct ldlm_lock *lock;
3620 struct osc_object *osc = NULL;
3624 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3625 if (lock->l_ast_data != NULL && osc == NULL) {
3626 osc = lock->l_ast_data;
3627 cl_object_get(osc2cl(osc));
3630 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3631 * by the 2nd round of ldlm_namespace_clean() call in
3632 * osc_import_event(). */
3633 ldlm_clear_cleaned(lock);
3638 osc_object_invalidate(env, osc);
3639 cl_object_put(env, osc2cl(osc));
3644 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3646 static int osc_import_event(struct obd_device *obd, struct obd_import *imp,
3647 enum obd_import_event event)
3649 struct client_obd *cli;
3653 if (WARN_ON_ONCE(!obd || !imp || imp->imp_obd != obd))
3657 case IMP_EVENT_DISCON: {
3661 spin_lock(&cli->cl_loi_list_lock);
3662 cli->cl_avail_grant = 0;
3663 cli->cl_lost_grant = 0;
3664 spin_unlock(&cli->cl_loi_list_lock);
3667 case IMP_EVENT_INACTIVE: {
3668 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3671 case IMP_EVENT_INVALIDATE: {
3672 struct ldlm_namespace *ns = obd->obd_namespace;
3676 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3678 env = cl_env_get(&refcheck);
3680 osc_io_unplug(env, &obd->u.cli, NULL);
3682 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3683 osc_ldlm_resource_invalidate,
3685 cl_env_put(env, &refcheck);
3687 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3693 case IMP_EVENT_ACTIVE: {
3694 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3697 case IMP_EVENT_OCD: {
3698 struct obd_connect_data *ocd = &imp->imp_connect_data;
3700 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3701 osc_init_grant(&obd->u.cli, ocd);
3704 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3705 imp->imp_client->cli_request_portal =
3708 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3711 case IMP_EVENT_DEACTIVATE: {
3712 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3715 case IMP_EVENT_ACTIVATE: {
3716 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3720 CERROR("%s: Unknown import event %d: rc = %d\n",
3721 obd->obd_name, event, -EINVAL);
3728 * Determine whether the lock can be canceled before replaying the lock
3729 * during recovery, see bug16774 for detailed information.
3731 * \retval zero the lock can't be canceled
3732 * \retval other ok to cancel
3734 static int osc_cancel_weight(struct ldlm_lock *lock)
3737 * Cancel all unused and granted extent lock.
3739 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3740 ldlm_is_granted(lock) &&
3741 osc_ldlm_weigh_ast(lock) == 0)
3747 static int brw_queue_work(const struct lu_env *env, void *data)
3749 struct client_obd *cli = data;
3751 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3753 osc_io_unplug(env, cli, NULL);
3757 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3759 struct client_obd *cli = &obd->u.cli;
3765 rc = ptlrpcd_addref();
3769 rc = client_obd_setup(obd, lcfg);
3771 GOTO(out_ptlrpcd, rc);
3774 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3775 if (IS_ERR(handler))
3776 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3777 cli->cl_writeback_work = handler;
3779 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3780 if (IS_ERR(handler))
3781 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3782 cli->cl_lru_work = handler;
3784 rc = osc_quota_setup(obd);
3786 GOTO(out_ptlrpcd_work, rc);
3788 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3789 cli->cl_root_squash = 0;
3790 osc_update_next_shrink(cli);
3795 if (cli->cl_writeback_work != NULL) {
3796 ptlrpcd_destroy_work(cli->cl_writeback_work);
3797 cli->cl_writeback_work = NULL;
3799 if (cli->cl_lru_work != NULL) {
3800 ptlrpcd_destroy_work(cli->cl_lru_work);
3801 cli->cl_lru_work = NULL;
3803 client_obd_cleanup(obd);
3808 EXPORT_SYMBOL(osc_setup_common);
3810 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3812 struct client_obd *cli = &obd->u.cli;
3820 rc = osc_setup_common(obd, lcfg);
3824 rc = osc_tunables_init(obd);
3829 * We try to control the total number of requests with a upper limit
3830 * osc_reqpool_maxreqcount. There might be some race which will cause
3831 * over-limit allocation, but it is fine.
3833 req_count = atomic_read(&osc_pool_req_count);
3834 if (req_count < osc_reqpool_maxreqcount) {
3835 adding = cli->cl_max_rpcs_in_flight + 2;
3836 if (req_count + adding > osc_reqpool_maxreqcount)
3837 adding = osc_reqpool_maxreqcount - req_count;
3839 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3840 atomic_add(added, &osc_pool_req_count);
3843 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3845 spin_lock(&osc_shrink_lock);
3846 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3847 spin_unlock(&osc_shrink_lock);
3848 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3849 cli->cl_import->imp_idle_debug = D_HA;
3854 int osc_precleanup_common(struct obd_device *obd)
3856 struct client_obd *cli = &obd->u.cli;
3860 * for echo client, export may be on zombie list, wait for
3861 * zombie thread to cull it, because cli.cl_import will be
3862 * cleared in client_disconnect_export():
3863 * class_export_destroy() -> obd_cleanup() ->
3864 * echo_device_free() -> echo_client_cleanup() ->
3865 * obd_disconnect() -> osc_disconnect() ->
3866 * client_disconnect_export()
3868 obd_zombie_barrier();
3869 if (cli->cl_writeback_work) {
3870 ptlrpcd_destroy_work(cli->cl_writeback_work);
3871 cli->cl_writeback_work = NULL;
3874 if (cli->cl_lru_work) {
3875 ptlrpcd_destroy_work(cli->cl_lru_work);
3876 cli->cl_lru_work = NULL;
3879 obd_cleanup_client_import(obd);
3882 EXPORT_SYMBOL(osc_precleanup_common);
3884 static int osc_precleanup(struct obd_device *obd)
3888 osc_precleanup_common(obd);
3890 ptlrpc_lprocfs_unregister_obd(obd);
3894 int osc_cleanup_common(struct obd_device *obd)
3896 struct client_obd *cli = &obd->u.cli;
3901 spin_lock(&osc_shrink_lock);
3902 list_del(&cli->cl_shrink_list);
3903 spin_unlock(&osc_shrink_lock);
3906 if (cli->cl_cache != NULL) {
3907 LASSERT(refcount_read(&cli->cl_cache->ccc_users) > 0);
3908 spin_lock(&cli->cl_cache->ccc_lru_lock);
3909 list_del_init(&cli->cl_lru_osc);
3910 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3911 cli->cl_lru_left = NULL;
3912 cl_cache_decref(cli->cl_cache);
3913 cli->cl_cache = NULL;
3916 /* free memory of osc quota cache */
3917 osc_quota_cleanup(obd);
3919 rc = client_obd_cleanup(obd);
3924 EXPORT_SYMBOL(osc_cleanup_common);
3926 static const struct obd_ops osc_obd_ops = {
3927 .o_owner = THIS_MODULE,
3928 .o_setup = osc_setup,
3929 .o_precleanup = osc_precleanup,
3930 .o_cleanup = osc_cleanup_common,
3931 .o_add_conn = client_import_add_conn,
3932 .o_del_conn = client_import_del_conn,
3933 .o_connect = client_connect_import,
3934 .o_reconnect = osc_reconnect,
3935 .o_disconnect = osc_disconnect,
3936 .o_statfs = osc_statfs,
3937 .o_statfs_async = osc_statfs_async,
3938 .o_create = osc_create,
3939 .o_destroy = osc_destroy,
3940 .o_getattr = osc_getattr,
3941 .o_setattr = osc_setattr,
3942 .o_iocontrol = osc_iocontrol,
3943 .o_set_info_async = osc_set_info_async,
3944 .o_import_event = osc_import_event,
3945 .o_quotactl = osc_quotactl,
3948 LIST_HEAD(osc_shrink_list);
3949 DEFINE_SPINLOCK(osc_shrink_lock);
3950 bool osc_page_cache_shrink_enabled = true;
3952 #ifdef HAVE_SHRINKER_COUNT
3953 static struct ll_shrinker_ops osc_cache_sh_ops = {
3954 .count_objects = osc_cache_shrink_count,
3955 .scan_objects = osc_cache_shrink_scan,
3956 .seeks = DEFAULT_SEEKS,
3959 static int osc_cache_shrink(struct shrinker *shrinker,
3960 struct shrink_control *sc)
3962 (void)osc_cache_shrink_scan(shrinker, sc);
3964 return osc_cache_shrink_count(shrinker, sc);
3967 static struct ll_shrinker_ops osc_cache_sh_ops = {
3968 .shrink = osc_cache_shrink,
3969 .seeks = DEFAULT_SEEKS,
3973 static struct shrinker *osc_cache_shrinker;
3975 static int __init osc_init(void)
3977 unsigned int reqpool_size;
3978 unsigned int reqsize;
3982 /* print an address of _any_ initialized kernel symbol from this
3983 * module, to allow debugging with gdb that doesn't support data
3984 * symbols from modules.*/
3985 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3987 rc = libcfs_setup();
3991 rc = lu_kmem_init(osc_caches);
3995 osc_cache_shrinker = ll_shrinker_create(&osc_cache_sh_ops, 0,
3997 if (IS_ERR(osc_cache_shrinker))
3998 GOTO(out_kmem, rc = PTR_ERR(osc_cache_shrinker));
4000 /* This is obviously too much memory, only prevent overflow here */
4001 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
4002 GOTO(out_shrinker, rc = -EINVAL);
4004 reqpool_size = osc_reqpool_mem_max << 20;
4007 while (reqsize < OST_IO_MAXREQSIZE)
4008 reqsize = reqsize << 1;
4011 * We don't enlarge the request count in OSC pool according to
4012 * cl_max_rpcs_in_flight. The allocation from the pool will only be
4013 * tried after normal allocation failed. So a small OSC pool won't
4014 * cause much performance degression in most of cases.
4016 osc_reqpool_maxreqcount = reqpool_size / reqsize;
4018 atomic_set(&osc_pool_req_count, 0);
4019 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
4020 ptlrpc_add_rqs_to_pool);
4022 if (osc_rq_pool == NULL)
4023 GOTO(out_shrinker, rc = -ENOMEM);
4025 rc = osc_start_grant_work();
4027 GOTO(out_req_pool, rc);
4029 rc = class_register_type(&osc_obd_ops, NULL, true,
4030 LUSTRE_OSC_NAME, &osc_device_type);
4032 GOTO(out_stop_grant, rc);
4037 osc_stop_grant_work();
4039 ptlrpc_free_rq_pool(osc_rq_pool);
4041 shrinker_free(osc_cache_shrinker);
4043 lu_kmem_fini(osc_caches);
4048 static void __exit osc_exit(void)
4050 class_unregister_type(LUSTRE_OSC_NAME);
4051 ptlrpc_free_rq_pool(osc_rq_pool);
4052 osc_stop_grant_work();
4053 shrinker_free(osc_cache_shrinker);
4054 lu_kmem_fini(osc_caches);
4057 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4058 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4059 MODULE_VERSION(LUSTRE_VERSION_STRING);
4060 MODULE_LICENSE("GPL");
4062 module_init(osc_init);
4063 module_exit(osc_exit);