4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_OSC
35 #include <lprocfs_status.h>
36 #include <lustre_debug.h>
37 #include <lustre_dlm.h>
38 #include <lustre_fid.h>
39 #include <lustre_ha.h>
40 #include <uapi/linux/lustre/lustre_ioctl.h>
41 #include <lustre_net.h>
42 #include <lustre_obdo.h>
43 #include <uapi/linux/lustre/lustre_param.h>
45 #include <obd_cksum.h>
46 #include <obd_class.h>
47 #include <lustre_osc.h>
49 #include "osc_internal.h"
51 atomic_t osc_pool_req_count;
52 unsigned int osc_reqpool_maxreqcount;
53 struct ptlrpc_request_pool *osc_rq_pool;
55 /* max memory used for request pool, unit is MB */
56 static unsigned int osc_reqpool_mem_max = 5;
57 module_param(osc_reqpool_mem_max, uint, 0444);
59 #define osc_grant_args osc_brw_async_args
61 struct osc_setattr_args {
63 obd_enqueue_update_f sa_upcall;
67 struct osc_fsync_args {
68 struct osc_object *fa_obj;
70 obd_enqueue_update_f fa_upcall;
74 struct osc_ladvise_args {
76 obd_enqueue_update_f la_upcall;
80 static void osc_release_ppga(struct brw_page **ppga, size_t count);
81 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
84 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
86 struct ost_body *body;
88 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
91 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
94 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
97 struct ptlrpc_request *req;
98 struct ost_body *body;
102 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
106 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
108 ptlrpc_request_free(req);
112 osc_pack_req_body(req, oa);
114 ptlrpc_request_set_replen(req);
116 rc = ptlrpc_queue_wait(req);
120 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
122 GOTO(out, rc = -EPROTO);
124 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
125 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
127 oa->o_blksize = cli_brw_size(exp->exp_obd);
128 oa->o_valid |= OBD_MD_FLBLKSZ;
132 ptlrpc_req_finished(req);
137 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
140 struct ptlrpc_request *req;
141 struct ost_body *body;
145 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
147 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
151 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
153 ptlrpc_request_free(req);
157 osc_pack_req_body(req, oa);
159 ptlrpc_request_set_replen(req);
161 rc = ptlrpc_queue_wait(req);
165 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
167 GOTO(out, rc = -EPROTO);
169 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
173 ptlrpc_req_finished(req);
178 static int osc_setattr_interpret(const struct lu_env *env,
179 struct ptlrpc_request *req,
180 struct osc_setattr_args *sa, int rc)
182 struct ost_body *body;
188 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
190 GOTO(out, rc = -EPROTO);
192 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
195 rc = sa->sa_upcall(sa->sa_cookie, rc);
199 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
200 obd_enqueue_update_f upcall, void *cookie,
201 struct ptlrpc_request_set *rqset)
203 struct ptlrpc_request *req;
204 struct osc_setattr_args *sa;
209 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
213 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
215 ptlrpc_request_free(req);
219 osc_pack_req_body(req, oa);
221 ptlrpc_request_set_replen(req);
223 /* do mds to ost setattr asynchronously */
225 /* Do not wait for response. */
226 ptlrpcd_add_req(req);
228 req->rq_interpret_reply =
229 (ptlrpc_interpterer_t)osc_setattr_interpret;
231 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
232 sa = ptlrpc_req_async_args(req);
234 sa->sa_upcall = upcall;
235 sa->sa_cookie = cookie;
237 if (rqset == PTLRPCD_SET)
238 ptlrpcd_add_req(req);
240 ptlrpc_set_add_req(rqset, req);
246 static int osc_ladvise_interpret(const struct lu_env *env,
247 struct ptlrpc_request *req,
250 struct osc_ladvise_args *la = arg;
251 struct ost_body *body;
257 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
259 GOTO(out, rc = -EPROTO);
261 *la->la_oa = body->oa;
263 rc = la->la_upcall(la->la_cookie, rc);
268 * If rqset is NULL, do not wait for response. Upcall and cookie could also
269 * be NULL in this case
271 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
272 struct ladvise_hdr *ladvise_hdr,
273 obd_enqueue_update_f upcall, void *cookie,
274 struct ptlrpc_request_set *rqset)
276 struct ptlrpc_request *req;
277 struct ost_body *body;
278 struct osc_ladvise_args *la;
280 struct lu_ladvise *req_ladvise;
281 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
282 int num_advise = ladvise_hdr->lah_count;
283 struct ladvise_hdr *req_ladvise_hdr;
286 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
290 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
291 num_advise * sizeof(*ladvise));
292 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
294 ptlrpc_request_free(req);
297 req->rq_request_portal = OST_IO_PORTAL;
298 ptlrpc_at_set_req_timeout(req);
300 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
302 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
305 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
306 &RMF_OST_LADVISE_HDR);
307 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
309 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
310 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
311 ptlrpc_request_set_replen(req);
314 /* Do not wait for response. */
315 ptlrpcd_add_req(req);
319 req->rq_interpret_reply = osc_ladvise_interpret;
320 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
321 la = ptlrpc_req_async_args(req);
323 la->la_upcall = upcall;
324 la->la_cookie = cookie;
326 if (rqset == PTLRPCD_SET)
327 ptlrpcd_add_req(req);
329 ptlrpc_set_add_req(rqset, req);
334 static int osc_create(const struct lu_env *env, struct obd_export *exp,
337 struct ptlrpc_request *req;
338 struct ost_body *body;
343 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
344 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
346 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
348 GOTO(out, rc = -ENOMEM);
350 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
352 ptlrpc_request_free(req);
356 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
359 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
361 ptlrpc_request_set_replen(req);
363 rc = ptlrpc_queue_wait(req);
367 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
369 GOTO(out_req, rc = -EPROTO);
371 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
372 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
374 oa->o_blksize = cli_brw_size(exp->exp_obd);
375 oa->o_valid |= OBD_MD_FLBLKSZ;
377 CDEBUG(D_HA, "transno: %lld\n",
378 lustre_msg_get_transno(req->rq_repmsg));
380 ptlrpc_req_finished(req);
385 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
386 obd_enqueue_update_f upcall, void *cookie)
388 struct ptlrpc_request *req;
389 struct osc_setattr_args *sa;
390 struct obd_import *imp = class_exp2cliimp(exp);
391 struct ost_body *body;
396 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
400 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
402 ptlrpc_request_free(req);
406 osc_set_io_portal(req);
408 ptlrpc_at_set_req_timeout(req);
410 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
412 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
414 ptlrpc_request_set_replen(req);
416 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
417 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
418 sa = ptlrpc_req_async_args(req);
420 sa->sa_upcall = upcall;
421 sa->sa_cookie = cookie;
423 ptlrpcd_add_req(req);
427 EXPORT_SYMBOL(osc_punch_send);
429 static int osc_sync_interpret(const struct lu_env *env,
430 struct ptlrpc_request *req,
433 struct osc_fsync_args *fa = arg;
434 struct ost_body *body;
435 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
436 unsigned long valid = 0;
437 struct cl_object *obj;
443 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
445 CERROR("can't unpack ost_body\n");
446 GOTO(out, rc = -EPROTO);
449 *fa->fa_oa = body->oa;
450 obj = osc2cl(fa->fa_obj);
452 /* Update osc object's blocks attribute */
453 cl_object_attr_lock(obj);
454 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
455 attr->cat_blocks = body->oa.o_blocks;
460 cl_object_attr_update(env, obj, attr, valid);
461 cl_object_attr_unlock(obj);
464 rc = fa->fa_upcall(fa->fa_cookie, rc);
468 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
469 obd_enqueue_update_f upcall, void *cookie,
470 struct ptlrpc_request_set *rqset)
472 struct obd_export *exp = osc_export(obj);
473 struct ptlrpc_request *req;
474 struct ost_body *body;
475 struct osc_fsync_args *fa;
479 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
483 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
485 ptlrpc_request_free(req);
489 /* overload the size and blocks fields in the oa with start/end */
490 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
492 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
494 ptlrpc_request_set_replen(req);
495 req->rq_interpret_reply = osc_sync_interpret;
497 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
498 fa = ptlrpc_req_async_args(req);
501 fa->fa_upcall = upcall;
502 fa->fa_cookie = cookie;
504 if (rqset == PTLRPCD_SET)
505 ptlrpcd_add_req(req);
507 ptlrpc_set_add_req(rqset, req);
512 /* Find and cancel locally locks matched by @mode in the resource found by
513 * @objid. Found locks are added into @cancel list. Returns the amount of
514 * locks added to @cancels list. */
515 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
516 struct list_head *cancels,
517 enum ldlm_mode mode, __u64 lock_flags)
519 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
520 struct ldlm_res_id res_id;
521 struct ldlm_resource *res;
525 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
526 * export) but disabled through procfs (flag in NS).
528 * This distinguishes from a case when ELC is not supported originally,
529 * when we still want to cancel locks in advance and just cancel them
530 * locally, without sending any RPC. */
531 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
534 ostid_build_res_name(&oa->o_oi, &res_id);
535 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
539 LDLM_RESOURCE_ADDREF(res);
540 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
541 lock_flags, 0, NULL);
542 LDLM_RESOURCE_DELREF(res);
543 ldlm_resource_putref(res);
547 static int osc_destroy_interpret(const struct lu_env *env,
548 struct ptlrpc_request *req, void *data,
551 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
553 atomic_dec(&cli->cl_destroy_in_flight);
554 wake_up(&cli->cl_destroy_waitq);
558 static int osc_can_send_destroy(struct client_obd *cli)
560 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
561 cli->cl_max_rpcs_in_flight) {
562 /* The destroy request can be sent */
565 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
566 cli->cl_max_rpcs_in_flight) {
568 * The counter has been modified between the two atomic
571 wake_up(&cli->cl_destroy_waitq);
576 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
579 struct client_obd *cli = &exp->exp_obd->u.cli;
580 struct ptlrpc_request *req;
581 struct ost_body *body;
582 struct list_head cancels = LIST_HEAD_INIT(cancels);
587 CDEBUG(D_INFO, "oa NULL\n");
591 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
592 LDLM_FL_DISCARD_DATA);
594 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
596 ldlm_lock_list_put(&cancels, l_bl_ast, count);
600 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
603 ptlrpc_request_free(req);
607 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
608 ptlrpc_at_set_req_timeout(req);
610 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
612 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
614 ptlrpc_request_set_replen(req);
616 req->rq_interpret_reply = osc_destroy_interpret;
617 if (!osc_can_send_destroy(cli)) {
618 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
621 * Wait until the number of on-going destroy RPCs drops
622 * under max_rpc_in_flight
624 rc = l_wait_event_exclusive(cli->cl_destroy_waitq,
625 osc_can_send_destroy(cli), &lwi);
627 ptlrpc_req_finished(req);
632 /* Do not wait for response */
633 ptlrpcd_add_req(req);
637 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
640 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
642 LASSERT(!(oa->o_valid & bits));
645 spin_lock(&cli->cl_loi_list_lock);
646 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
647 oa->o_dirty = cli->cl_dirty_grant;
649 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
650 if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
651 cli->cl_dirty_max_pages)) {
652 CERROR("dirty %lu - %lu > dirty_max %lu\n",
653 cli->cl_dirty_pages, cli->cl_dirty_transit,
654 cli->cl_dirty_max_pages);
656 } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
657 atomic_long_read(&obd_dirty_transit_pages) >
658 (long)(obd_max_dirty_pages + 1))) {
659 /* The atomic_read() allowing the atomic_inc() are
660 * not covered by a lock thus they may safely race and trip
661 * this CERROR() unless we add in a small fudge factor (+1). */
662 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
663 cli_name(cli), atomic_long_read(&obd_dirty_pages),
664 atomic_long_read(&obd_dirty_transit_pages),
665 obd_max_dirty_pages);
667 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
669 CERROR("dirty %lu - dirty_max %lu too big???\n",
670 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
673 unsigned long nrpages;
675 nrpages = cli->cl_max_pages_per_rpc;
676 nrpages *= cli->cl_max_rpcs_in_flight + 1;
677 nrpages = max(nrpages, cli->cl_dirty_max_pages);
678 oa->o_undirty = nrpages << PAGE_SHIFT;
679 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
683 /* take extent tax into account when asking for more
685 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
686 cli->cl_max_extent_pages;
687 oa->o_undirty += nrextents * cli->cl_grant_extent_tax;
690 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
691 oa->o_dropped = cli->cl_lost_grant;
692 cli->cl_lost_grant = 0;
693 spin_unlock(&cli->cl_loi_list_lock);
694 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
695 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
698 void osc_update_next_shrink(struct client_obd *cli)
700 cli->cl_next_shrink_grant = ktime_get_seconds() +
701 cli->cl_grant_shrink_interval;
703 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
704 cli->cl_next_shrink_grant);
707 static void __osc_update_grant(struct client_obd *cli, u64 grant)
709 spin_lock(&cli->cl_loi_list_lock);
710 cli->cl_avail_grant += grant;
711 spin_unlock(&cli->cl_loi_list_lock);
714 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
716 if (body->oa.o_valid & OBD_MD_FLGRANT) {
717 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
718 __osc_update_grant(cli, body->oa.o_grant);
722 static int osc_shrink_grant_interpret(const struct lu_env *env,
723 struct ptlrpc_request *req,
726 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
727 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
728 struct ost_body *body;
731 __osc_update_grant(cli, oa->o_grant);
735 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
737 osc_update_grant(cli, body);
743 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
745 spin_lock(&cli->cl_loi_list_lock);
746 oa->o_grant = cli->cl_avail_grant / 4;
747 cli->cl_avail_grant -= oa->o_grant;
748 spin_unlock(&cli->cl_loi_list_lock);
749 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
750 oa->o_valid |= OBD_MD_FLFLAGS;
753 oa->o_flags |= OBD_FL_SHRINK_GRANT;
754 osc_update_next_shrink(cli);
757 /* Shrink the current grant, either from some large amount to enough for a
758 * full set of in-flight RPCs, or if we have already shrunk to that limit
759 * then to enough for a single RPC. This avoids keeping more grant than
760 * needed, and avoids shrinking the grant piecemeal. */
761 static int osc_shrink_grant(struct client_obd *cli)
763 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
764 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
766 spin_lock(&cli->cl_loi_list_lock);
767 if (cli->cl_avail_grant <= target_bytes)
768 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
769 spin_unlock(&cli->cl_loi_list_lock);
771 return osc_shrink_grant_to_target(cli, target_bytes);
774 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
777 struct ost_body *body;
780 spin_lock(&cli->cl_loi_list_lock);
781 /* Don't shrink if we are already above or below the desired limit
782 * We don't want to shrink below a single RPC, as that will negatively
783 * impact block allocation and long-term performance. */
784 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
785 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
787 if (target_bytes >= cli->cl_avail_grant) {
788 spin_unlock(&cli->cl_loi_list_lock);
791 spin_unlock(&cli->cl_loi_list_lock);
797 osc_announce_cached(cli, &body->oa, 0);
799 spin_lock(&cli->cl_loi_list_lock);
800 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
801 cli->cl_avail_grant = target_bytes;
802 spin_unlock(&cli->cl_loi_list_lock);
803 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
804 body->oa.o_valid |= OBD_MD_FLFLAGS;
805 body->oa.o_flags = 0;
807 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
808 osc_update_next_shrink(cli);
810 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
811 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
812 sizeof(*body), body, NULL);
814 __osc_update_grant(cli, body->oa.o_grant);
819 static int osc_should_shrink_grant(struct client_obd *client)
821 time64_t next_shrink = client->cl_next_shrink_grant;
823 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
824 OBD_CONNECT_GRANT_SHRINK) == 0)
827 if (ktime_get_seconds() >= next_shrink - 5) {
828 /* Get the current RPC size directly, instead of going via:
829 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
830 * Keep comment here so that it can be found by searching. */
831 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
833 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
834 client->cl_avail_grant > brw_size)
837 osc_update_next_shrink(client);
842 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
844 struct client_obd *client;
846 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
847 if (osc_should_shrink_grant(client))
848 osc_shrink_grant(client);
853 static int osc_add_shrink_grant(struct client_obd *client)
857 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
859 osc_grant_shrink_grant_cb, NULL,
860 &client->cl_grant_shrink_list);
862 CERROR("add grant client %s error %d\n", cli_name(client), rc);
865 CDEBUG(D_CACHE, "add grant client %s\n", cli_name(client));
866 osc_update_next_shrink(client);
870 static int osc_del_shrink_grant(struct client_obd *client)
872 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
876 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
879 * ocd_grant is the total grant amount we're expect to hold: if we've
880 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
881 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
884 * race is tolerable here: if we're evicted, but imp_state already
885 * left EVICTED state, then cl_dirty_pages must be 0 already.
887 spin_lock(&cli->cl_loi_list_lock);
888 cli->cl_avail_grant = ocd->ocd_grant;
889 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
890 cli->cl_avail_grant -= cli->cl_reserved_grant;
891 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
892 cli->cl_avail_grant -= cli->cl_dirty_grant;
894 cli->cl_avail_grant -=
895 cli->cl_dirty_pages << PAGE_SHIFT;
898 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
902 /* overhead for each extent insertion */
903 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
904 /* determine the appropriate chunk size used by osc_extent. */
905 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
906 ocd->ocd_grant_blkbits);
907 /* max_pages_per_rpc must be chunk aligned */
908 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
909 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
910 ~chunk_mask) & chunk_mask;
911 /* determine maximum extent size, in #pages */
912 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
913 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
914 if (cli->cl_max_extent_pages == 0)
915 cli->cl_max_extent_pages = 1;
917 cli->cl_grant_extent_tax = 0;
918 cli->cl_chunkbits = PAGE_SHIFT;
919 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
921 spin_unlock(&cli->cl_loi_list_lock);
923 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
924 "chunk bits: %d cl_max_extent_pages: %d\n",
926 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
927 cli->cl_max_extent_pages);
929 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
930 list_empty(&cli->cl_grant_shrink_list))
931 osc_add_shrink_grant(cli);
933 EXPORT_SYMBOL(osc_init_grant);
935 /* We assume that the reason this OSC got a short read is because it read
936 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
937 * via the LOV, and it _knows_ it's reading inside the file, it's just that
938 * this stripe never got written at or beyond this stripe offset yet. */
939 static void handle_short_read(int nob_read, size_t page_count,
940 struct brw_page **pga)
945 /* skip bytes read OK */
946 while (nob_read > 0) {
947 LASSERT (page_count > 0);
949 if (pga[i]->count > nob_read) {
950 /* EOF inside this page */
951 ptr = kmap(pga[i]->pg) +
952 (pga[i]->off & ~PAGE_MASK);
953 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
960 nob_read -= pga[i]->count;
965 /* zero remaining pages */
966 while (page_count-- > 0) {
967 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
968 memset(ptr, 0, pga[i]->count);
974 static int check_write_rcs(struct ptlrpc_request *req,
975 int requested_nob, int niocount,
976 size_t page_count, struct brw_page **pga)
981 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
982 sizeof(*remote_rcs) *
984 if (remote_rcs == NULL) {
985 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
989 /* return error if any niobuf was in error */
990 for (i = 0; i < niocount; i++) {
991 if ((int)remote_rcs[i] < 0)
992 return(remote_rcs[i]);
994 if (remote_rcs[i] != 0) {
995 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
996 i, remote_rcs[i], req);
1000 if (req->rq_bulk != NULL &&
1001 req->rq_bulk->bd_nob_transferred != requested_nob) {
1002 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1003 req->rq_bulk->bd_nob_transferred, requested_nob);
1010 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1012 if (p1->flag != p2->flag) {
1013 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1014 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1015 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1017 /* warn if we try to combine flags that we don't know to be
1018 * safe to combine */
1019 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1020 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1021 "report this at https://jira.hpdd.intel.com/\n",
1022 p1->flag, p2->flag);
1027 return (p1->off + p1->count == p2->off);
1030 static int osc_checksum_bulk(int nob, size_t pg_count,
1031 struct brw_page **pga, int opc,
1032 enum cksum_types cksum_type,
1036 struct cfs_crypto_hash_desc *hdesc;
1037 unsigned int bufsize;
1038 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1040 LASSERT(pg_count > 0);
1042 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1043 if (IS_ERR(hdesc)) {
1044 CERROR("Unable to initialize checksum hash %s\n",
1045 cfs_crypto_hash_name(cfs_alg));
1046 return PTR_ERR(hdesc);
1049 while (nob > 0 && pg_count > 0) {
1050 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1052 /* corrupt the data before we compute the checksum, to
1053 * simulate an OST->client data error */
1054 if (i == 0 && opc == OST_READ &&
1055 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1056 unsigned char *ptr = kmap(pga[i]->pg);
1057 int off = pga[i]->off & ~PAGE_MASK;
1059 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1062 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1063 pga[i]->off & ~PAGE_MASK,
1065 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1066 (int)(pga[i]->off & ~PAGE_MASK));
1068 nob -= pga[i]->count;
1073 bufsize = sizeof(*cksum);
1074 cfs_crypto_hash_final(hdesc, (unsigned char *)cksum, &bufsize);
1076 /* For sending we only compute the wrong checksum instead
1077 * of corrupting the data so it is still correct on a redo */
1078 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1085 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1086 u32 page_count, struct brw_page **pga,
1087 struct ptlrpc_request **reqp, int resend)
1089 struct ptlrpc_request *req;
1090 struct ptlrpc_bulk_desc *desc;
1091 struct ost_body *body;
1092 struct obd_ioobj *ioobj;
1093 struct niobuf_remote *niobuf;
1094 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1095 struct osc_brw_async_args *aa;
1096 struct req_capsule *pill;
1097 struct brw_page *pg_prev;
1101 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1102 RETURN(-ENOMEM); /* Recoverable */
1103 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1104 RETURN(-EINVAL); /* Fatal */
1106 if ((cmd & OBD_BRW_WRITE) != 0) {
1108 req = ptlrpc_request_alloc_pool(cli->cl_import,
1110 &RQF_OST_BRW_WRITE);
1113 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1118 for (niocount = i = 1; i < page_count; i++) {
1119 if (!can_merge_pages(pga[i - 1], pga[i]))
1123 pill = &req->rq_pill;
1124 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1126 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1127 niocount * sizeof(*niobuf));
1129 for (i = 0; i < page_count; i++)
1130 short_io_size += pga[i]->count;
1132 /* Check if we can do a short io. */
1133 if (!(short_io_size <= cli->cl_short_io_bytes && niocount == 1 &&
1134 imp_connect_shortio(cli->cl_import)))
1137 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1138 opc == OST_READ ? 0 : short_io_size);
1139 if (opc == OST_READ)
1140 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1143 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1145 ptlrpc_request_free(req);
1148 osc_set_io_portal(req);
1150 ptlrpc_at_set_req_timeout(req);
1151 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1153 req->rq_no_retry_einprogress = 1;
1155 if (short_io_size != 0) {
1157 short_io_buf = NULL;
1161 desc = ptlrpc_prep_bulk_imp(req, page_count,
1162 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1163 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1164 PTLRPC_BULK_PUT_SINK) |
1165 PTLRPC_BULK_BUF_KIOV,
1167 &ptlrpc_bulk_kiov_pin_ops);
1170 GOTO(out, rc = -ENOMEM);
1171 /* NB request now owns desc and will free it when it gets freed */
1173 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1174 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1175 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1176 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1178 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1180 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1181 * and from_kgid(), because they are asynchronous. Fortunately, variable
1182 * oa contains valid o_uid and o_gid in these two operations.
1183 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1184 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1185 * other process logic */
1186 body->oa.o_uid = oa->o_uid;
1187 body->oa.o_gid = oa->o_gid;
1189 obdo_to_ioobj(oa, ioobj);
1190 ioobj->ioo_bufcnt = niocount;
1191 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1192 * that might be send for this request. The actual number is decided
1193 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1194 * "max - 1" for old client compatibility sending "0", and also so the
1195 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1197 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1199 ioobj_max_brw_set(ioobj, 0);
1201 if (short_io_size != 0) {
1202 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1203 body->oa.o_valid |= OBD_MD_FLFLAGS;
1204 body->oa.o_flags = 0;
1206 body->oa.o_flags |= OBD_FL_SHORT_IO;
1207 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1209 if (opc == OST_WRITE) {
1210 short_io_buf = req_capsule_client_get(pill,
1212 LASSERT(short_io_buf != NULL);
1216 LASSERT(page_count > 0);
1218 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1219 struct brw_page *pg = pga[i];
1220 int poff = pg->off & ~PAGE_MASK;
1222 LASSERT(pg->count > 0);
1223 /* make sure there is no gap in the middle of page array */
1224 LASSERTF(page_count == 1 ||
1225 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1226 ergo(i > 0 && i < page_count - 1,
1227 poff == 0 && pg->count == PAGE_SIZE) &&
1228 ergo(i == page_count - 1, poff == 0)),
1229 "i: %d/%d pg: %p off: %llu, count: %u\n",
1230 i, page_count, pg, pg->off, pg->count);
1231 LASSERTF(i == 0 || pg->off > pg_prev->off,
1232 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1233 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1235 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1236 pg_prev->pg, page_private(pg_prev->pg),
1237 pg_prev->pg->index, pg_prev->off);
1238 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1239 (pg->flag & OBD_BRW_SRVLOCK));
1240 if (short_io_size != 0 && opc == OST_WRITE) {
1241 unsigned char *ptr = ll_kmap_atomic(pg->pg, KM_USER0);
1243 LASSERT(short_io_size >= requested_nob + pg->count);
1244 memcpy(short_io_buf + requested_nob,
1247 ll_kunmap_atomic(ptr, KM_USER0);
1248 } else if (short_io_size == 0) {
1249 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1252 requested_nob += pg->count;
1254 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1256 niobuf->rnb_len += pg->count;
1258 niobuf->rnb_offset = pg->off;
1259 niobuf->rnb_len = pg->count;
1260 niobuf->rnb_flags = pg->flag;
1265 LASSERTF((void *)(niobuf - niocount) ==
1266 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1267 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1268 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1270 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1272 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1273 body->oa.o_valid |= OBD_MD_FLFLAGS;
1274 body->oa.o_flags = 0;
1276 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1279 if (osc_should_shrink_grant(cli))
1280 osc_shrink_grant_local(cli, &body->oa);
1282 /* size[REQ_REC_OFF] still sizeof (*body) */
1283 if (opc == OST_WRITE) {
1284 if (cli->cl_checksum &&
1285 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1286 /* store cl_cksum_type in a local variable since
1287 * it can be changed via lprocfs */
1288 enum cksum_types cksum_type = cli->cl_cksum_type;
1290 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1291 body->oa.o_flags = 0;
1293 body->oa.o_flags |= cksum_type_pack(cksum_type);
1294 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1296 rc = osc_checksum_bulk(requested_nob, page_count,
1297 pga, OST_WRITE, cksum_type,
1300 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1304 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1307 /* save this in 'oa', too, for later checking */
1308 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1309 oa->o_flags |= cksum_type_pack(cksum_type);
1311 /* clear out the checksum flag, in case this is a
1312 * resend but cl_checksum is no longer set. b=11238 */
1313 oa->o_valid &= ~OBD_MD_FLCKSUM;
1315 oa->o_cksum = body->oa.o_cksum;
1316 /* 1 RC per niobuf */
1317 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1318 sizeof(__u32) * niocount);
1320 if (cli->cl_checksum &&
1321 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1322 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1323 body->oa.o_flags = 0;
1324 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1325 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1328 /* Client cksum has been already copied to wire obdo in previous
1329 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1330 * resent due to cksum error, this will allow Server to
1331 * check+dump pages on its side */
1333 ptlrpc_request_set_replen(req);
1335 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1336 aa = ptlrpc_req_async_args(req);
1338 aa->aa_requested_nob = requested_nob;
1339 aa->aa_nio_count = niocount;
1340 aa->aa_page_count = page_count;
1344 INIT_LIST_HEAD(&aa->aa_oaps);
1347 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1348 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1349 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1350 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1354 ptlrpc_req_finished(req);
1358 char dbgcksum_file_name[PATH_MAX];
1360 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1361 struct brw_page **pga, __u32 server_cksum,
1370 /* will only keep dump of pages on first error for the same range in
1371 * file/fid, not during the resends/retries. */
1372 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1373 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1374 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1375 libcfs_debug_file_path_arr :
1376 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1377 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1378 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1379 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1381 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1382 client_cksum, server_cksum);
1383 filp = filp_open(dbgcksum_file_name,
1384 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1388 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1389 "checksum error: rc = %d\n", dbgcksum_file_name,
1392 CERROR("%s: can't open to dump pages with checksum "
1393 "error: rc = %d\n", dbgcksum_file_name, rc);
1399 for (i = 0; i < page_count; i++) {
1400 len = pga[i]->count;
1401 buf = kmap(pga[i]->pg);
1403 rc = vfs_write(filp, (__force const char __user *)buf,
1406 CERROR("%s: wanted to write %u but got %d "
1407 "error\n", dbgcksum_file_name, len, rc);
1412 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1413 dbgcksum_file_name, rc);
1419 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1421 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1422 filp_close(filp, NULL);
1427 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1428 __u32 client_cksum, __u32 server_cksum,
1429 struct osc_brw_async_args *aa)
1433 enum cksum_types cksum_type;
1436 if (server_cksum == client_cksum) {
1437 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1441 if (aa->aa_cli->cl_checksum_dump)
1442 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1443 server_cksum, client_cksum);
1445 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1447 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1448 aa->aa_ppga, OST_WRITE, cksum_type,
1452 msg = "failed to calculate the client write checksum";
1453 else if (cksum_type != cksum_type_unpack(aa->aa_oa->o_flags))
1454 msg = "the server did not use the checksum type specified in "
1455 "the original request - likely a protocol problem";
1456 else if (new_cksum == server_cksum)
1457 msg = "changed on the client after we checksummed it - "
1458 "likely false positive due to mmap IO (bug 11742)";
1459 else if (new_cksum == client_cksum)
1460 msg = "changed in transit before arrival at OST";
1462 msg = "changed in transit AND doesn't match the original - "
1463 "likely false positive due to mmap IO (bug 11742)";
1465 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1466 DFID " object "DOSTID" extent [%llu-%llu], original "
1467 "client csum %x (type %x), server csum %x (type %x),"
1468 " client csum now %x\n",
1469 aa->aa_cli->cl_import->imp_obd->obd_name,
1470 msg, libcfs_nid2str(peer->nid),
1471 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1472 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1473 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1474 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1475 aa->aa_ppga[aa->aa_page_count - 1]->off +
1476 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1477 client_cksum, cksum_type_unpack(aa->aa_oa->o_flags),
1478 server_cksum, cksum_type, new_cksum);
1482 /* Note rc enters this function as number of bytes transferred */
1483 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1485 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1486 const struct lnet_process_id *peer =
1487 &req->rq_import->imp_connection->c_peer;
1488 struct client_obd *cli = aa->aa_cli;
1489 struct ost_body *body;
1490 u32 client_cksum = 0;
1493 if (rc < 0 && rc != -EDQUOT) {
1494 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1498 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1499 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1501 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1505 /* set/clear over quota flag for a uid/gid/projid */
1506 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1507 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1508 unsigned qid[LL_MAXQUOTAS] = {
1509 body->oa.o_uid, body->oa.o_gid,
1510 body->oa.o_projid };
1511 CDEBUG(D_QUOTA, "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1512 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1513 body->oa.o_valid, body->oa.o_flags);
1514 osc_quota_setdq(cli, qid, body->oa.o_valid,
1518 osc_update_grant(cli, body);
1523 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1524 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1526 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1528 CERROR("Unexpected +ve rc %d\n", rc);
1532 if (req->rq_bulk != NULL &&
1533 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1536 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1537 check_write_checksum(&body->oa, peer, client_cksum,
1538 body->oa.o_cksum, aa))
1541 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1542 aa->aa_page_count, aa->aa_ppga);
1546 /* The rest of this function executes only for OST_READs */
1548 if (req->rq_bulk == NULL) {
1549 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1551 LASSERT(rc == req->rq_status);
1553 /* if unwrap_bulk failed, return -EAGAIN to retry */
1554 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1557 GOTO(out, rc = -EAGAIN);
1559 if (rc > aa->aa_requested_nob) {
1560 CERROR("Unexpected rc %d (%d requested)\n", rc,
1561 aa->aa_requested_nob);
1565 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1566 CERROR ("Unexpected rc %d (%d transferred)\n",
1567 rc, req->rq_bulk->bd_nob_transferred);
1571 if (req->rq_bulk == NULL) {
1573 int nob, pg_count, i = 0;
1576 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1577 pg_count = aa->aa_page_count;
1578 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1581 while (nob > 0 && pg_count > 0) {
1583 int count = aa->aa_ppga[i]->count > nob ?
1584 nob : aa->aa_ppga[i]->count;
1586 CDEBUG(D_CACHE, "page %p count %d\n",
1587 aa->aa_ppga[i]->pg, count);
1588 ptr = ll_kmap_atomic(aa->aa_ppga[i]->pg, KM_USER0);
1589 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1591 ll_kunmap_atomic((void *) ptr, KM_USER0);
1600 if (rc < aa->aa_requested_nob)
1601 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1603 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1604 static int cksum_counter;
1605 u32 server_cksum = body->oa.o_cksum;
1608 enum cksum_types cksum_type;
1610 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1611 body->oa.o_flags : 0);
1612 rc = osc_checksum_bulk(rc, aa->aa_page_count, aa->aa_ppga,
1613 OST_READ, cksum_type, &client_cksum);
1616 "failed to calculate checksum, rc = %d\n", rc);
1619 if (req->rq_bulk != NULL &&
1620 peer->nid != req->rq_bulk->bd_sender) {
1622 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1625 if (server_cksum != client_cksum) {
1626 struct ost_body *clbody;
1627 u32 page_count = aa->aa_page_count;
1629 clbody = req_capsule_client_get(&req->rq_pill,
1631 if (cli->cl_checksum_dump)
1632 dump_all_bulk_pages(&clbody->oa, page_count,
1633 aa->aa_ppga, server_cksum,
1636 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1637 "%s%s%s inode "DFID" object "DOSTID
1638 " extent [%llu-%llu], client %x, "
1639 "server %x, cksum_type %x\n",
1640 req->rq_import->imp_obd->obd_name,
1641 libcfs_nid2str(peer->nid),
1643 clbody->oa.o_valid & OBD_MD_FLFID ?
1644 clbody->oa.o_parent_seq : 0ULL,
1645 clbody->oa.o_valid & OBD_MD_FLFID ?
1646 clbody->oa.o_parent_oid : 0,
1647 clbody->oa.o_valid & OBD_MD_FLFID ?
1648 clbody->oa.o_parent_ver : 0,
1649 POSTID(&body->oa.o_oi),
1650 aa->aa_ppga[0]->off,
1651 aa->aa_ppga[page_count-1]->off +
1652 aa->aa_ppga[page_count-1]->count - 1,
1653 client_cksum, server_cksum,
1656 aa->aa_oa->o_cksum = client_cksum;
1660 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1663 } else if (unlikely(client_cksum)) {
1664 static int cksum_missed;
1667 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1668 CERROR("Checksum %u requested from %s but not sent\n",
1669 cksum_missed, libcfs_nid2str(peer->nid));
1675 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1676 aa->aa_oa, &body->oa);
1681 static int osc_brw_redo_request(struct ptlrpc_request *request,
1682 struct osc_brw_async_args *aa, int rc)
1684 struct ptlrpc_request *new_req;
1685 struct osc_brw_async_args *new_aa;
1686 struct osc_async_page *oap;
1689 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1690 "redo for recoverable error %d", rc);
1692 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1693 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1694 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1695 aa->aa_ppga, &new_req, 1);
1699 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1700 if (oap->oap_request != NULL) {
1701 LASSERTF(request == oap->oap_request,
1702 "request %p != oap_request %p\n",
1703 request, oap->oap_request);
1704 if (oap->oap_interrupted) {
1705 ptlrpc_req_finished(new_req);
1710 /* New request takes over pga and oaps from old request.
1711 * Note that copying a list_head doesn't work, need to move it... */
1713 new_req->rq_interpret_reply = request->rq_interpret_reply;
1714 new_req->rq_async_args = request->rq_async_args;
1715 new_req->rq_commit_cb = request->rq_commit_cb;
1716 /* cap resend delay to the current request timeout, this is similar to
1717 * what ptlrpc does (see after_reply()) */
1718 if (aa->aa_resends > new_req->rq_timeout)
1719 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1721 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1722 new_req->rq_generation_set = 1;
1723 new_req->rq_import_generation = request->rq_import_generation;
1725 new_aa = ptlrpc_req_async_args(new_req);
1727 INIT_LIST_HEAD(&new_aa->aa_oaps);
1728 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1729 INIT_LIST_HEAD(&new_aa->aa_exts);
1730 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1731 new_aa->aa_resends = aa->aa_resends;
1733 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1734 if (oap->oap_request) {
1735 ptlrpc_req_finished(oap->oap_request);
1736 oap->oap_request = ptlrpc_request_addref(new_req);
1740 /* XXX: This code will run into problem if we're going to support
1741 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1742 * and wait for all of them to be finished. We should inherit request
1743 * set from old request. */
1744 ptlrpcd_add_req(new_req);
1746 DEBUG_REQ(D_INFO, new_req, "new request");
1751 * ugh, we want disk allocation on the target to happen in offset order. we'll
1752 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1753 * fine for our small page arrays and doesn't require allocation. its an
1754 * insertion sort that swaps elements that are strides apart, shrinking the
1755 * stride down until its '1' and the array is sorted.
1757 static void sort_brw_pages(struct brw_page **array, int num)
1760 struct brw_page *tmp;
1764 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1769 for (i = stride ; i < num ; i++) {
1772 while (j >= stride && array[j - stride]->off > tmp->off) {
1773 array[j] = array[j - stride];
1778 } while (stride > 1);
1781 static void osc_release_ppga(struct brw_page **ppga, size_t count)
1783 LASSERT(ppga != NULL);
1784 OBD_FREE(ppga, sizeof(*ppga) * count);
1787 static int brw_interpret(const struct lu_env *env,
1788 struct ptlrpc_request *req, void *data, int rc)
1790 struct osc_brw_async_args *aa = data;
1791 struct osc_extent *ext;
1792 struct osc_extent *tmp;
1793 struct client_obd *cli = aa->aa_cli;
1794 unsigned long transferred = 0;
1797 rc = osc_brw_fini_request(req, rc);
1798 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1799 /* When server return -EINPROGRESS, client should always retry
1800 * regardless of the number of times the bulk was resent already. */
1801 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
1802 if (req->rq_import_generation !=
1803 req->rq_import->imp_generation) {
1804 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1805 ""DOSTID", rc = %d.\n",
1806 req->rq_import->imp_obd->obd_name,
1807 POSTID(&aa->aa_oa->o_oi), rc);
1808 } else if (rc == -EINPROGRESS ||
1809 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1810 rc = osc_brw_redo_request(req, aa, rc);
1812 CERROR("%s: too many resent retries for object: "
1813 "%llu:%llu, rc = %d.\n",
1814 req->rq_import->imp_obd->obd_name,
1815 POSTID(&aa->aa_oa->o_oi), rc);
1820 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1825 struct obdo *oa = aa->aa_oa;
1826 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1827 unsigned long valid = 0;
1828 struct cl_object *obj;
1829 struct osc_async_page *last;
1831 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1832 obj = osc2cl(last->oap_obj);
1834 cl_object_attr_lock(obj);
1835 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1836 attr->cat_blocks = oa->o_blocks;
1837 valid |= CAT_BLOCKS;
1839 if (oa->o_valid & OBD_MD_FLMTIME) {
1840 attr->cat_mtime = oa->o_mtime;
1843 if (oa->o_valid & OBD_MD_FLATIME) {
1844 attr->cat_atime = oa->o_atime;
1847 if (oa->o_valid & OBD_MD_FLCTIME) {
1848 attr->cat_ctime = oa->o_ctime;
1852 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1853 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1854 loff_t last_off = last->oap_count + last->oap_obj_off +
1857 /* Change file size if this is an out of quota or
1858 * direct IO write and it extends the file size */
1859 if (loi->loi_lvb.lvb_size < last_off) {
1860 attr->cat_size = last_off;
1863 /* Extend KMS if it's not a lockless write */
1864 if (loi->loi_kms < last_off &&
1865 oap2osc_page(last)->ops_srvlock == 0) {
1866 attr->cat_kms = last_off;
1872 cl_object_attr_update(env, obj, attr, valid);
1873 cl_object_attr_unlock(obj);
1875 OBDO_FREE(aa->aa_oa);
1877 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
1878 osc_inc_unstable_pages(req);
1880 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1881 list_del_init(&ext->oe_link);
1882 osc_extent_finish(env, ext, 1,
1883 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
1885 LASSERT(list_empty(&aa->aa_exts));
1886 LASSERT(list_empty(&aa->aa_oaps));
1888 transferred = (req->rq_bulk == NULL ? /* short io */
1889 aa->aa_requested_nob :
1890 req->rq_bulk->bd_nob_transferred);
1892 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1893 ptlrpc_lprocfs_brw(req, transferred);
1895 spin_lock(&cli->cl_loi_list_lock);
1896 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1897 * is called so we know whether to go to sync BRWs or wait for more
1898 * RPCs to complete */
1899 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1900 cli->cl_w_in_flight--;
1902 cli->cl_r_in_flight--;
1903 osc_wake_cache_waiters(cli);
1904 spin_unlock(&cli->cl_loi_list_lock);
1906 osc_io_unplug(env, cli, NULL);
1910 static void brw_commit(struct ptlrpc_request *req)
1912 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1913 * this called via the rq_commit_cb, I need to ensure
1914 * osc_dec_unstable_pages is still called. Otherwise unstable
1915 * pages may be leaked. */
1916 spin_lock(&req->rq_lock);
1917 if (likely(req->rq_unstable)) {
1918 req->rq_unstable = 0;
1919 spin_unlock(&req->rq_lock);
1921 osc_dec_unstable_pages(req);
1923 req->rq_committed = 1;
1924 spin_unlock(&req->rq_lock);
1929 * Build an RPC by the list of extent @ext_list. The caller must ensure
1930 * that the total pages in this list are NOT over max pages per RPC.
1931 * Extents in the list must be in OES_RPC state.
1933 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1934 struct list_head *ext_list, int cmd)
1936 struct ptlrpc_request *req = NULL;
1937 struct osc_extent *ext;
1938 struct brw_page **pga = NULL;
1939 struct osc_brw_async_args *aa = NULL;
1940 struct obdo *oa = NULL;
1941 struct osc_async_page *oap;
1942 struct osc_object *obj = NULL;
1943 struct cl_req_attr *crattr = NULL;
1944 loff_t starting_offset = OBD_OBJECT_EOF;
1945 loff_t ending_offset = 0;
1949 bool soft_sync = false;
1950 bool interrupted = false;
1951 bool ndelay = false;
1955 __u32 layout_version = 0;
1956 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1957 struct ost_body *body;
1959 LASSERT(!list_empty(ext_list));
1961 /* add pages into rpc_list to build BRW rpc */
1962 list_for_each_entry(ext, ext_list, oe_link) {
1963 LASSERT(ext->oe_state == OES_RPC);
1964 mem_tight |= ext->oe_memalloc;
1965 grant += ext->oe_grants;
1966 page_count += ext->oe_nr_pages;
1967 layout_version = MAX(layout_version, ext->oe_layout_version);
1972 soft_sync = osc_over_unstable_soft_limit(cli);
1974 mpflag = cfs_memory_pressure_get_and_set();
1976 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1978 GOTO(out, rc = -ENOMEM);
1982 GOTO(out, rc = -ENOMEM);
1985 list_for_each_entry(ext, ext_list, oe_link) {
1986 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1988 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1990 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
1991 pga[i] = &oap->oap_brw_page;
1992 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1995 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1996 if (starting_offset == OBD_OBJECT_EOF ||
1997 starting_offset > oap->oap_obj_off)
1998 starting_offset = oap->oap_obj_off;
2000 LASSERT(oap->oap_page_off == 0);
2001 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2002 ending_offset = oap->oap_obj_off +
2005 LASSERT(oap->oap_page_off + oap->oap_count ==
2007 if (oap->oap_interrupted)
2014 /* first page in the list */
2015 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2017 crattr = &osc_env_info(env)->oti_req_attr;
2018 memset(crattr, 0, sizeof(*crattr));
2019 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2020 crattr->cra_flags = ~0ULL;
2021 crattr->cra_page = oap2cl_page(oap);
2022 crattr->cra_oa = oa;
2023 cl_req_attr_set(env, osc2cl(obj), crattr);
2025 if (cmd == OBD_BRW_WRITE) {
2026 oa->o_grant_used = grant;
2027 if (layout_version > 0) {
2028 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2029 PFID(&oa->o_oi.oi_fid), layout_version);
2031 oa->o_layout_version = layout_version;
2032 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2036 sort_brw_pages(pga, page_count);
2037 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2039 CERROR("prep_req failed: %d\n", rc);
2043 req->rq_commit_cb = brw_commit;
2044 req->rq_interpret_reply = brw_interpret;
2045 req->rq_memalloc = mem_tight != 0;
2046 oap->oap_request = ptlrpc_request_addref(req);
2047 if (interrupted && !req->rq_intr)
2048 ptlrpc_mark_interrupted(req);
2050 req->rq_no_resend = req->rq_no_delay = 1;
2051 /* probably set a shorter timeout value.
2052 * to handle ETIMEDOUT in brw_interpret() correctly. */
2053 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2056 /* Need to update the timestamps after the request is built in case
2057 * we race with setattr (locally or in queue at OST). If OST gets
2058 * later setattr before earlier BRW (as determined by the request xid),
2059 * the OST will not use BRW timestamps. Sadly, there is no obvious
2060 * way to do this in a single call. bug 10150 */
2061 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2062 crattr->cra_oa = &body->oa;
2063 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2064 cl_req_attr_set(env, osc2cl(obj), crattr);
2065 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2067 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2068 aa = ptlrpc_req_async_args(req);
2069 INIT_LIST_HEAD(&aa->aa_oaps);
2070 list_splice_init(&rpc_list, &aa->aa_oaps);
2071 INIT_LIST_HEAD(&aa->aa_exts);
2072 list_splice_init(ext_list, &aa->aa_exts);
2074 spin_lock(&cli->cl_loi_list_lock);
2075 starting_offset >>= PAGE_SHIFT;
2076 if (cmd == OBD_BRW_READ) {
2077 cli->cl_r_in_flight++;
2078 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2079 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2080 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2081 starting_offset + 1);
2083 cli->cl_w_in_flight++;
2084 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2085 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2086 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2087 starting_offset + 1);
2089 spin_unlock(&cli->cl_loi_list_lock);
2091 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
2092 page_count, aa, cli->cl_r_in_flight,
2093 cli->cl_w_in_flight);
2094 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2096 ptlrpcd_add_req(req);
2102 cfs_memory_pressure_restore(mpflag);
2105 LASSERT(req == NULL);
2110 OBD_FREE(pga, sizeof(*pga) * page_count);
2111 /* this should happen rarely and is pretty bad, it makes the
2112 * pending list not follow the dirty order */
2113 while (!list_empty(ext_list)) {
2114 ext = list_entry(ext_list->next, struct osc_extent,
2116 list_del_init(&ext->oe_link);
2117 osc_extent_finish(env, ext, 0, rc);
2123 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2127 LASSERT(lock != NULL);
2129 lock_res_and_lock(lock);
2131 if (lock->l_ast_data == NULL)
2132 lock->l_ast_data = data;
2133 if (lock->l_ast_data == data)
2136 unlock_res_and_lock(lock);
2141 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2142 void *cookie, struct lustre_handle *lockh,
2143 enum ldlm_mode mode, __u64 *flags, bool speculative,
2146 bool intent = *flags & LDLM_FL_HAS_INTENT;
2150 /* The request was created before ldlm_cli_enqueue call. */
2151 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2152 struct ldlm_reply *rep;
2154 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2155 LASSERT(rep != NULL);
2157 rep->lock_policy_res1 =
2158 ptlrpc_status_ntoh(rep->lock_policy_res1);
2159 if (rep->lock_policy_res1)
2160 errcode = rep->lock_policy_res1;
2162 *flags |= LDLM_FL_LVB_READY;
2163 } else if (errcode == ELDLM_OK) {
2164 *flags |= LDLM_FL_LVB_READY;
2167 /* Call the update callback. */
2168 rc = (*upcall)(cookie, lockh, errcode);
2170 /* release the reference taken in ldlm_cli_enqueue() */
2171 if (errcode == ELDLM_LOCK_MATCHED)
2173 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2174 ldlm_lock_decref(lockh, mode);
2179 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2180 struct osc_enqueue_args *aa, int rc)
2182 struct ldlm_lock *lock;
2183 struct lustre_handle *lockh = &aa->oa_lockh;
2184 enum ldlm_mode mode = aa->oa_mode;
2185 struct ost_lvb *lvb = aa->oa_lvb;
2186 __u32 lvb_len = sizeof(*lvb);
2191 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2193 lock = ldlm_handle2lock(lockh);
2194 LASSERTF(lock != NULL,
2195 "lockh %#llx, req %p, aa %p - client evicted?\n",
2196 lockh->cookie, req, aa);
2198 /* Take an additional reference so that a blocking AST that
2199 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2200 * to arrive after an upcall has been executed by
2201 * osc_enqueue_fini(). */
2202 ldlm_lock_addref(lockh, mode);
2204 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2205 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2207 /* Let CP AST to grant the lock first. */
2208 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2210 if (aa->oa_speculative) {
2211 LASSERT(aa->oa_lvb == NULL);
2212 LASSERT(aa->oa_flags == NULL);
2213 aa->oa_flags = &flags;
2216 /* Complete obtaining the lock procedure. */
2217 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2218 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2220 /* Complete osc stuff. */
2221 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2222 aa->oa_flags, aa->oa_speculative, rc);
2224 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2226 ldlm_lock_decref(lockh, mode);
2227 LDLM_LOCK_PUT(lock);
2231 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2233 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2234 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2235 * other synchronous requests, however keeping some locks and trying to obtain
2236 * others may take a considerable amount of time in a case of ost failure; and
2237 * when other sync requests do not get released lock from a client, the client
2238 * is evicted from the cluster -- such scenarious make the life difficult, so
2239 * release locks just after they are obtained. */
2240 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2241 __u64 *flags, union ldlm_policy_data *policy,
2242 struct ost_lvb *lvb, int kms_valid,
2243 osc_enqueue_upcall_f upcall, void *cookie,
2244 struct ldlm_enqueue_info *einfo,
2245 struct ptlrpc_request_set *rqset, int async,
2248 struct obd_device *obd = exp->exp_obd;
2249 struct lustre_handle lockh = { 0 };
2250 struct ptlrpc_request *req = NULL;
2251 int intent = *flags & LDLM_FL_HAS_INTENT;
2252 __u64 match_flags = *flags;
2253 enum ldlm_mode mode;
2257 /* Filesystem lock extents are extended to page boundaries so that
2258 * dealing with the page cache is a little smoother. */
2259 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2260 policy->l_extent.end |= ~PAGE_MASK;
2263 * kms is not valid when either object is completely fresh (so that no
2264 * locks are cached), or object was evicted. In the latter case cached
2265 * lock cannot be used, because it would prime inode state with
2266 * potentially stale LVB.
2271 /* Next, search for already existing extent locks that will cover us */
2272 /* If we're trying to read, we also search for an existing PW lock. The
2273 * VFS and page cache already protect us locally, so lots of readers/
2274 * writers can share a single PW lock.
2276 * There are problems with conversion deadlocks, so instead of
2277 * converting a read lock to a write lock, we'll just enqueue a new
2280 * At some point we should cancel the read lock instead of making them
2281 * send us a blocking callback, but there are problems with canceling
2282 * locks out from other users right now, too. */
2283 mode = einfo->ei_mode;
2284 if (einfo->ei_mode == LCK_PR)
2286 /* Normal lock requests must wait for the LVB to be ready before
2287 * matching a lock; speculative lock requests do not need to,
2288 * because they will not actually use the lock. */
2290 match_flags |= LDLM_FL_LVB_READY;
2292 match_flags |= LDLM_FL_BLOCK_GRANTED;
2293 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2294 einfo->ei_type, policy, mode, &lockh, 0);
2296 struct ldlm_lock *matched;
2298 if (*flags & LDLM_FL_TEST_LOCK)
2301 matched = ldlm_handle2lock(&lockh);
2303 /* This DLM lock request is speculative, and does not
2304 * have an associated IO request. Therefore if there
2305 * is already a DLM lock, it wll just inform the
2306 * caller to cancel the request for this stripe.*/
2307 lock_res_and_lock(matched);
2308 if (ldlm_extent_equal(&policy->l_extent,
2309 &matched->l_policy_data.l_extent))
2313 unlock_res_and_lock(matched);
2315 ldlm_lock_decref(&lockh, mode);
2316 LDLM_LOCK_PUT(matched);
2318 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2319 *flags |= LDLM_FL_LVB_READY;
2321 /* We already have a lock, and it's referenced. */
2322 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2324 ldlm_lock_decref(&lockh, mode);
2325 LDLM_LOCK_PUT(matched);
2328 ldlm_lock_decref(&lockh, mode);
2329 LDLM_LOCK_PUT(matched);
2334 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2338 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2339 &RQF_LDLM_ENQUEUE_LVB);
2343 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2345 ptlrpc_request_free(req);
2349 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2351 ptlrpc_request_set_replen(req);
2354 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2355 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2357 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2358 sizeof(*lvb), LVB_T_OST, &lockh, async);
2361 struct osc_enqueue_args *aa;
2362 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2363 aa = ptlrpc_req_async_args(req);
2365 aa->oa_mode = einfo->ei_mode;
2366 aa->oa_type = einfo->ei_type;
2367 lustre_handle_copy(&aa->oa_lockh, &lockh);
2368 aa->oa_upcall = upcall;
2369 aa->oa_cookie = cookie;
2370 aa->oa_speculative = speculative;
2372 aa->oa_flags = flags;
2375 /* speculative locks are essentially to enqueue
2376 * a DLM lock in advance, so we don't care
2377 * about the result of the enqueue. */
2379 aa->oa_flags = NULL;
2382 req->rq_interpret_reply =
2383 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2384 if (rqset == PTLRPCD_SET)
2385 ptlrpcd_add_req(req);
2387 ptlrpc_set_add_req(rqset, req);
2388 } else if (intent) {
2389 ptlrpc_req_finished(req);
2394 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2395 flags, speculative, rc);
2397 ptlrpc_req_finished(req);
2402 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2403 enum ldlm_type type, union ldlm_policy_data *policy,
2404 enum ldlm_mode mode, __u64 *flags, void *data,
2405 struct lustre_handle *lockh, int unref)
2407 struct obd_device *obd = exp->exp_obd;
2408 __u64 lflags = *flags;
2412 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2415 /* Filesystem lock extents are extended to page boundaries so that
2416 * dealing with the page cache is a little smoother */
2417 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2418 policy->l_extent.end |= ~PAGE_MASK;
2420 /* Next, search for already existing extent locks that will cover us */
2421 /* If we're trying to read, we also search for an existing PW lock. The
2422 * VFS and page cache already protect us locally, so lots of readers/
2423 * writers can share a single PW lock. */
2427 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2428 res_id, type, policy, rc, lockh, unref);
2429 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2433 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2435 LASSERT(lock != NULL);
2436 if (!osc_set_lock_data(lock, data)) {
2437 ldlm_lock_decref(lockh, rc);
2440 LDLM_LOCK_PUT(lock);
2445 static int osc_statfs_interpret(const struct lu_env *env,
2446 struct ptlrpc_request *req,
2447 struct osc_async_args *aa, int rc)
2449 struct obd_statfs *msfs;
2453 /* The request has in fact never been sent
2454 * due to issues at a higher level (LOV).
2455 * Exit immediately since the caller is
2456 * aware of the problem and takes care
2457 * of the clean up */
2460 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2461 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2467 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2469 GOTO(out, rc = -EPROTO);
2472 *aa->aa_oi->oi_osfs = *msfs;
2474 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2478 static int osc_statfs_async(struct obd_export *exp,
2479 struct obd_info *oinfo, time64_t max_age,
2480 struct ptlrpc_request_set *rqset)
2482 struct obd_device *obd = class_exp2obd(exp);
2483 struct ptlrpc_request *req;
2484 struct osc_async_args *aa;
2488 /* We could possibly pass max_age in the request (as an absolute
2489 * timestamp or a "seconds.usec ago") so the target can avoid doing
2490 * extra calls into the filesystem if that isn't necessary (e.g.
2491 * during mount that would help a bit). Having relative timestamps
2492 * is not so great if request processing is slow, while absolute
2493 * timestamps are not ideal because they need time synchronization. */
2494 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2498 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2500 ptlrpc_request_free(req);
2503 ptlrpc_request_set_replen(req);
2504 req->rq_request_portal = OST_CREATE_PORTAL;
2505 ptlrpc_at_set_req_timeout(req);
2507 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2508 /* procfs requests not want stat in wait for avoid deadlock */
2509 req->rq_no_resend = 1;
2510 req->rq_no_delay = 1;
2513 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2514 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2515 aa = ptlrpc_req_async_args(req);
2518 ptlrpc_set_add_req(rqset, req);
2522 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2523 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2525 struct obd_device *obd = class_exp2obd(exp);
2526 struct obd_statfs *msfs;
2527 struct ptlrpc_request *req;
2528 struct obd_import *imp = NULL;
2532 /*Since the request might also come from lprocfs, so we need
2533 *sync this with client_disconnect_export Bug15684*/
2534 down_read(&obd->u.cli.cl_sem);
2535 if (obd->u.cli.cl_import)
2536 imp = class_import_get(obd->u.cli.cl_import);
2537 up_read(&obd->u.cli.cl_sem);
2541 /* We could possibly pass max_age in the request (as an absolute
2542 * timestamp or a "seconds.usec ago") so the target can avoid doing
2543 * extra calls into the filesystem if that isn't necessary (e.g.
2544 * during mount that would help a bit). Having relative timestamps
2545 * is not so great if request processing is slow, while absolute
2546 * timestamps are not ideal because they need time synchronization. */
2547 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2549 class_import_put(imp);
2554 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2556 ptlrpc_request_free(req);
2559 ptlrpc_request_set_replen(req);
2560 req->rq_request_portal = OST_CREATE_PORTAL;
2561 ptlrpc_at_set_req_timeout(req);
2563 if (flags & OBD_STATFS_NODELAY) {
2564 /* procfs requests not want stat in wait for avoid deadlock */
2565 req->rq_no_resend = 1;
2566 req->rq_no_delay = 1;
2569 rc = ptlrpc_queue_wait(req);
2573 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2575 GOTO(out, rc = -EPROTO);
2582 ptlrpc_req_finished(req);
2586 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2587 void *karg, void __user *uarg)
2589 struct obd_device *obd = exp->exp_obd;
2590 struct obd_ioctl_data *data = karg;
2594 if (!try_module_get(THIS_MODULE)) {
2595 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2596 module_name(THIS_MODULE));
2600 case OBD_IOC_CLIENT_RECOVER:
2601 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2602 data->ioc_inlbuf1, 0);
2606 case IOC_OSC_SET_ACTIVE:
2607 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2610 case OBD_IOC_PING_TARGET:
2611 err = ptlrpc_obd_ping(obd);
2614 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2615 cmd, current_comm());
2616 GOTO(out, err = -ENOTTY);
2619 module_put(THIS_MODULE);
2623 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2624 u32 keylen, void *key, u32 vallen, void *val,
2625 struct ptlrpc_request_set *set)
2627 struct ptlrpc_request *req;
2628 struct obd_device *obd = exp->exp_obd;
2629 struct obd_import *imp = class_exp2cliimp(exp);
2634 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2636 if (KEY_IS(KEY_CHECKSUM)) {
2637 if (vallen != sizeof(int))
2639 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2643 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2644 sptlrpc_conf_client_adapt(obd);
2648 if (KEY_IS(KEY_FLUSH_CTX)) {
2649 sptlrpc_import_flush_my_ctx(imp);
2653 if (KEY_IS(KEY_CACHE_SET)) {
2654 struct client_obd *cli = &obd->u.cli;
2656 LASSERT(cli->cl_cache == NULL); /* only once */
2657 cli->cl_cache = (struct cl_client_cache *)val;
2658 cl_cache_incref(cli->cl_cache);
2659 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2661 /* add this osc into entity list */
2662 LASSERT(list_empty(&cli->cl_lru_osc));
2663 spin_lock(&cli->cl_cache->ccc_lru_lock);
2664 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2665 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2670 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2671 struct client_obd *cli = &obd->u.cli;
2672 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2673 long target = *(long *)val;
2675 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2680 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2683 /* We pass all other commands directly to OST. Since nobody calls osc
2684 methods directly and everybody is supposed to go through LOV, we
2685 assume lov checked invalid values for us.
2686 The only recognised values so far are evict_by_nid and mds_conn.
2687 Even if something bad goes through, we'd get a -EINVAL from OST
2690 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2691 &RQF_OST_SET_GRANT_INFO :
2696 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2697 RCL_CLIENT, keylen);
2698 if (!KEY_IS(KEY_GRANT_SHRINK))
2699 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2700 RCL_CLIENT, vallen);
2701 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2703 ptlrpc_request_free(req);
2707 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2708 memcpy(tmp, key, keylen);
2709 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2712 memcpy(tmp, val, vallen);
2714 if (KEY_IS(KEY_GRANT_SHRINK)) {
2715 struct osc_grant_args *aa;
2718 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2719 aa = ptlrpc_req_async_args(req);
2722 ptlrpc_req_finished(req);
2725 *oa = ((struct ost_body *)val)->oa;
2727 req->rq_interpret_reply = osc_shrink_grant_interpret;
2730 ptlrpc_request_set_replen(req);
2731 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2732 LASSERT(set != NULL);
2733 ptlrpc_set_add_req(set, req);
2734 ptlrpc_check_set(NULL, set);
2736 ptlrpcd_add_req(req);
2741 EXPORT_SYMBOL(osc_set_info_async);
2743 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
2744 struct obd_device *obd, struct obd_uuid *cluuid,
2745 struct obd_connect_data *data, void *localdata)
2747 struct client_obd *cli = &obd->u.cli;
2749 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2753 spin_lock(&cli->cl_loi_list_lock);
2754 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2755 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM)
2756 grant += cli->cl_dirty_grant;
2758 grant += cli->cl_dirty_pages << PAGE_SHIFT;
2759 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
2760 lost_grant = cli->cl_lost_grant;
2761 cli->cl_lost_grant = 0;
2762 spin_unlock(&cli->cl_loi_list_lock);
2764 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
2765 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2766 data->ocd_version, data->ocd_grant, lost_grant);
2771 EXPORT_SYMBOL(osc_reconnect);
2773 int osc_disconnect(struct obd_export *exp)
2775 struct obd_device *obd = class_exp2obd(exp);
2778 rc = client_disconnect_export(exp);
2780 * Initially we put del_shrink_grant before disconnect_export, but it
2781 * causes the following problem if setup (connect) and cleanup
2782 * (disconnect) are tangled together.
2783 * connect p1 disconnect p2
2784 * ptlrpc_connect_import
2785 * ............... class_manual_cleanup
2788 * ptlrpc_connect_interrupt
2790 * add this client to shrink list
2792 * Bang! pinger trigger the shrink.
2793 * So the osc should be disconnected from the shrink list, after we
2794 * are sure the import has been destroyed. BUG18662
2796 if (obd->u.cli.cl_import == NULL)
2797 osc_del_shrink_grant(&obd->u.cli);
2800 EXPORT_SYMBOL(osc_disconnect);
2802 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2803 struct hlist_node *hnode, void *arg)
2805 struct lu_env *env = arg;
2806 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2807 struct ldlm_lock *lock;
2808 struct osc_object *osc = NULL;
2812 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2813 if (lock->l_ast_data != NULL && osc == NULL) {
2814 osc = lock->l_ast_data;
2815 cl_object_get(osc2cl(osc));
2818 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
2819 * by the 2nd round of ldlm_namespace_clean() call in
2820 * osc_import_event(). */
2821 ldlm_clear_cleaned(lock);
2826 osc_object_invalidate(env, osc);
2827 cl_object_put(env, osc2cl(osc));
2832 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
2834 static int osc_import_event(struct obd_device *obd,
2835 struct obd_import *imp,
2836 enum obd_import_event event)
2838 struct client_obd *cli;
2842 LASSERT(imp->imp_obd == obd);
2845 case IMP_EVENT_DISCON: {
2847 spin_lock(&cli->cl_loi_list_lock);
2848 cli->cl_avail_grant = 0;
2849 cli->cl_lost_grant = 0;
2850 spin_unlock(&cli->cl_loi_list_lock);
2853 case IMP_EVENT_INACTIVE: {
2854 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2857 case IMP_EVENT_INVALIDATE: {
2858 struct ldlm_namespace *ns = obd->obd_namespace;
2862 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2864 env = cl_env_get(&refcheck);
2866 osc_io_unplug(env, &obd->u.cli, NULL);
2868 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2869 osc_ldlm_resource_invalidate,
2871 cl_env_put(env, &refcheck);
2873 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2878 case IMP_EVENT_ACTIVE: {
2879 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2882 case IMP_EVENT_OCD: {
2883 struct obd_connect_data *ocd = &imp->imp_connect_data;
2885 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
2886 osc_init_grant(&obd->u.cli, ocd);
2889 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
2890 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
2892 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2895 case IMP_EVENT_DEACTIVATE: {
2896 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
2899 case IMP_EVENT_ACTIVATE: {
2900 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
2904 CERROR("Unknown import event %d\n", event);
2911 * Determine whether the lock can be canceled before replaying the lock
2912 * during recovery, see bug16774 for detailed information.
2914 * \retval zero the lock can't be canceled
2915 * \retval other ok to cancel
2917 static int osc_cancel_weight(struct ldlm_lock *lock)
2920 * Cancel all unused and granted extent lock.
2922 if (lock->l_resource->lr_type == LDLM_EXTENT &&
2923 lock->l_granted_mode == lock->l_req_mode &&
2924 osc_ldlm_weigh_ast(lock) == 0)
2930 static int brw_queue_work(const struct lu_env *env, void *data)
2932 struct client_obd *cli = data;
2934 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2936 osc_io_unplug(env, cli, NULL);
2940 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
2942 struct client_obd *cli = &obd->u.cli;
2948 rc = ptlrpcd_addref();
2952 rc = client_obd_setup(obd, lcfg);
2954 GOTO(out_ptlrpcd, rc);
2957 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
2958 if (IS_ERR(handler))
2959 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
2960 cli->cl_writeback_work = handler;
2962 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
2963 if (IS_ERR(handler))
2964 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
2965 cli->cl_lru_work = handler;
2967 rc = osc_quota_setup(obd);
2969 GOTO(out_ptlrpcd_work, rc);
2971 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
2973 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
2977 if (cli->cl_writeback_work != NULL) {
2978 ptlrpcd_destroy_work(cli->cl_writeback_work);
2979 cli->cl_writeback_work = NULL;
2981 if (cli->cl_lru_work != NULL) {
2982 ptlrpcd_destroy_work(cli->cl_lru_work);
2983 cli->cl_lru_work = NULL;
2985 client_obd_cleanup(obd);
2990 EXPORT_SYMBOL(osc_setup_common);
2992 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2994 struct client_obd *cli = &obd->u.cli;
2995 struct obd_type *type;
3003 rc = osc_setup_common(obd, lcfg);
3007 #ifdef CONFIG_PROC_FS
3008 obd->obd_vars = lprocfs_osc_obd_vars;
3010 /* If this is true then both client (osc) and server (osp) are on the
3011 * same node. The osp layer if loaded first will register the osc proc
3012 * directory. In that case this obd_device will be attached its proc
3013 * tree to type->typ_procsym instead of obd->obd_type->typ_procroot.
3015 type = class_search_type(LUSTRE_OSP_NAME);
3016 if (type && type->typ_procsym) {
3017 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
3019 obd->obd_vars, obd);
3020 if (IS_ERR(obd->obd_proc_entry)) {
3021 rc = PTR_ERR(obd->obd_proc_entry);
3022 CERROR("error %d setting up lprocfs for %s\n", rc,
3024 obd->obd_proc_entry = NULL;
3028 rc = lprocfs_obd_setup(obd, false);
3030 /* If the basic OSC proc tree construction succeeded then
3033 lproc_osc_attach_seqstat(obd);
3034 sptlrpc_lprocfs_cliobd_attach(obd);
3035 ptlrpc_lprocfs_register_obd(obd);
3039 * We try to control the total number of requests with a upper limit
3040 * osc_reqpool_maxreqcount. There might be some race which will cause
3041 * over-limit allocation, but it is fine.
3043 req_count = atomic_read(&osc_pool_req_count);
3044 if (req_count < osc_reqpool_maxreqcount) {
3045 adding = cli->cl_max_rpcs_in_flight + 2;
3046 if (req_count + adding > osc_reqpool_maxreqcount)
3047 adding = osc_reqpool_maxreqcount - req_count;
3049 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3050 atomic_add(added, &osc_pool_req_count);
3053 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3054 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3056 spin_lock(&osc_shrink_lock);
3057 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3058 spin_unlock(&osc_shrink_lock);
3063 int osc_precleanup_common(struct obd_device *obd)
3065 struct client_obd *cli = &obd->u.cli;
3069 * for echo client, export may be on zombie list, wait for
3070 * zombie thread to cull it, because cli.cl_import will be
3071 * cleared in client_disconnect_export():
3072 * class_export_destroy() -> obd_cleanup() ->
3073 * echo_device_free() -> echo_client_cleanup() ->
3074 * obd_disconnect() -> osc_disconnect() ->
3075 * client_disconnect_export()
3077 obd_zombie_barrier();
3078 if (cli->cl_writeback_work) {
3079 ptlrpcd_destroy_work(cli->cl_writeback_work);
3080 cli->cl_writeback_work = NULL;
3083 if (cli->cl_lru_work) {
3084 ptlrpcd_destroy_work(cli->cl_lru_work);
3085 cli->cl_lru_work = NULL;
3088 obd_cleanup_client_import(obd);
3091 EXPORT_SYMBOL(osc_precleanup_common);
3093 static int osc_precleanup(struct obd_device *obd)
3097 osc_precleanup_common(obd);
3099 ptlrpc_lprocfs_unregister_obd(obd);
3103 int osc_cleanup_common(struct obd_device *obd)
3105 struct client_obd *cli = &obd->u.cli;
3110 spin_lock(&osc_shrink_lock);
3111 list_del(&cli->cl_shrink_list);
3112 spin_unlock(&osc_shrink_lock);
3115 if (cli->cl_cache != NULL) {
3116 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3117 spin_lock(&cli->cl_cache->ccc_lru_lock);
3118 list_del_init(&cli->cl_lru_osc);
3119 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3120 cli->cl_lru_left = NULL;
3121 cl_cache_decref(cli->cl_cache);
3122 cli->cl_cache = NULL;
3125 /* free memory of osc quota cache */
3126 osc_quota_cleanup(obd);
3128 rc = client_obd_cleanup(obd);
3133 EXPORT_SYMBOL(osc_cleanup_common);
3135 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3137 int rc = class_process_proc_param(PARAM_OSC, obd->obd_vars, lcfg, obd);
3138 return rc > 0 ? 0: rc;
3141 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
3143 return osc_process_config_base(obd, buf);
3146 static struct obd_ops osc_obd_ops = {
3147 .o_owner = THIS_MODULE,
3148 .o_setup = osc_setup,
3149 .o_precleanup = osc_precleanup,
3150 .o_cleanup = osc_cleanup_common,
3151 .o_add_conn = client_import_add_conn,
3152 .o_del_conn = client_import_del_conn,
3153 .o_connect = client_connect_import,
3154 .o_reconnect = osc_reconnect,
3155 .o_disconnect = osc_disconnect,
3156 .o_statfs = osc_statfs,
3157 .o_statfs_async = osc_statfs_async,
3158 .o_create = osc_create,
3159 .o_destroy = osc_destroy,
3160 .o_getattr = osc_getattr,
3161 .o_setattr = osc_setattr,
3162 .o_iocontrol = osc_iocontrol,
3163 .o_set_info_async = osc_set_info_async,
3164 .o_import_event = osc_import_event,
3165 .o_process_config = osc_process_config,
3166 .o_quotactl = osc_quotactl,
3169 static struct shrinker *osc_cache_shrinker;
3170 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
3171 DEFINE_SPINLOCK(osc_shrink_lock);
3173 #ifndef HAVE_SHRINKER_COUNT
3174 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3176 struct shrink_control scv = {
3177 .nr_to_scan = shrink_param(sc, nr_to_scan),
3178 .gfp_mask = shrink_param(sc, gfp_mask)
3180 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
3181 struct shrinker *shrinker = NULL;
3184 (void)osc_cache_shrink_scan(shrinker, &scv);
3186 return osc_cache_shrink_count(shrinker, &scv);
3190 static int __init osc_init(void)
3192 bool enable_proc = true;
3193 struct obd_type *type;
3194 unsigned int reqpool_size;
3195 unsigned int reqsize;
3197 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3198 osc_cache_shrink_count, osc_cache_shrink_scan);
3201 /* print an address of _any_ initialized kernel symbol from this
3202 * module, to allow debugging with gdb that doesn't support data
3203 * symbols from modules.*/
3204 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3206 rc = lu_kmem_init(osc_caches);
3210 type = class_search_type(LUSTRE_OSP_NAME);
3211 if (type != NULL && type->typ_procsym != NULL)
3212 enable_proc = false;
3214 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
3215 LUSTRE_OSC_NAME, &osc_device_type);
3219 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3221 /* This is obviously too much memory, only prevent overflow here */
3222 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3223 GOTO(out_type, rc = -EINVAL);
3225 reqpool_size = osc_reqpool_mem_max << 20;
3228 while (reqsize < OST_IO_MAXREQSIZE)
3229 reqsize = reqsize << 1;
3232 * We don't enlarge the request count in OSC pool according to
3233 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3234 * tried after normal allocation failed. So a small OSC pool won't
3235 * cause much performance degression in most of cases.
3237 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3239 atomic_set(&osc_pool_req_count, 0);
3240 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3241 ptlrpc_add_rqs_to_pool);
3243 if (osc_rq_pool != NULL)
3247 class_unregister_type(LUSTRE_OSC_NAME);
3249 lu_kmem_fini(osc_caches);
3254 static void __exit osc_exit(void)
3256 remove_shrinker(osc_cache_shrinker);
3257 class_unregister_type(LUSTRE_OSC_NAME);
3258 lu_kmem_fini(osc_caches);
3259 ptlrpc_free_rq_pool(osc_rq_pool);
3262 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3263 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3264 MODULE_VERSION(LUSTRE_VERSION_STRING);
3265 MODULE_LICENSE("GPL");
3267 module_init(osc_init);
3268 module_exit(osc_exit);