4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_OSC
35 #include <linux/workqueue.h>
36 #include <lprocfs_status.h>
37 #include <lustre_debug.h>
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <lustre_ha.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
42 #include <lustre_net.h>
43 #include <lustre_obdo.h>
44 #include <uapi/linux/lustre/lustre_param.h>
46 #include <obd_cksum.h>
47 #include <obd_class.h>
48 #include <lustre_osc.h>
50 #include "osc_internal.h"
52 atomic_t osc_pool_req_count;
53 unsigned int osc_reqpool_maxreqcount;
54 struct ptlrpc_request_pool *osc_rq_pool;
56 /* max memory used for request pool, unit is MB */
57 static unsigned int osc_reqpool_mem_max = 5;
58 module_param(osc_reqpool_mem_max, uint, 0444);
60 static int osc_idle_timeout = 20;
61 module_param(osc_idle_timeout, uint, 0644);
63 #define osc_grant_args osc_brw_async_args
65 struct osc_setattr_args {
67 obd_enqueue_update_f sa_upcall;
71 struct osc_fsync_args {
72 struct osc_object *fa_obj;
74 obd_enqueue_update_f fa_upcall;
78 struct osc_ladvise_args {
80 obd_enqueue_update_f la_upcall;
84 static void osc_release_ppga(struct brw_page **ppga, size_t count);
85 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
88 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
90 struct ost_body *body;
92 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
95 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
98 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
101 struct ptlrpc_request *req;
102 struct ost_body *body;
106 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
110 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
112 ptlrpc_request_free(req);
116 osc_pack_req_body(req, oa);
118 ptlrpc_request_set_replen(req);
120 rc = ptlrpc_queue_wait(req);
124 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
126 GOTO(out, rc = -EPROTO);
128 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
129 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
131 oa->o_blksize = cli_brw_size(exp->exp_obd);
132 oa->o_valid |= OBD_MD_FLBLKSZ;
136 ptlrpc_req_finished(req);
141 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
144 struct ptlrpc_request *req;
145 struct ost_body *body;
149 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
151 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
155 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
157 ptlrpc_request_free(req);
161 osc_pack_req_body(req, oa);
163 ptlrpc_request_set_replen(req);
165 rc = ptlrpc_queue_wait(req);
169 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
171 GOTO(out, rc = -EPROTO);
173 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
177 ptlrpc_req_finished(req);
182 static int osc_setattr_interpret(const struct lu_env *env,
183 struct ptlrpc_request *req,
184 struct osc_setattr_args *sa, int rc)
186 struct ost_body *body;
192 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
194 GOTO(out, rc = -EPROTO);
196 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
199 rc = sa->sa_upcall(sa->sa_cookie, rc);
203 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
204 obd_enqueue_update_f upcall, void *cookie,
205 struct ptlrpc_request_set *rqset)
207 struct ptlrpc_request *req;
208 struct osc_setattr_args *sa;
213 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
217 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
219 ptlrpc_request_free(req);
223 osc_pack_req_body(req, oa);
225 ptlrpc_request_set_replen(req);
227 /* do mds to ost setattr asynchronously */
229 /* Do not wait for response. */
230 ptlrpcd_add_req(req);
232 req->rq_interpret_reply =
233 (ptlrpc_interpterer_t)osc_setattr_interpret;
235 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
236 sa = ptlrpc_req_async_args(req);
238 sa->sa_upcall = upcall;
239 sa->sa_cookie = cookie;
241 if (rqset == PTLRPCD_SET)
242 ptlrpcd_add_req(req);
244 ptlrpc_set_add_req(rqset, req);
250 static int osc_ladvise_interpret(const struct lu_env *env,
251 struct ptlrpc_request *req,
254 struct osc_ladvise_args *la = arg;
255 struct ost_body *body;
261 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
263 GOTO(out, rc = -EPROTO);
265 *la->la_oa = body->oa;
267 rc = la->la_upcall(la->la_cookie, rc);
272 * If rqset is NULL, do not wait for response. Upcall and cookie could also
273 * be NULL in this case
275 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
276 struct ladvise_hdr *ladvise_hdr,
277 obd_enqueue_update_f upcall, void *cookie,
278 struct ptlrpc_request_set *rqset)
280 struct ptlrpc_request *req;
281 struct ost_body *body;
282 struct osc_ladvise_args *la;
284 struct lu_ladvise *req_ladvise;
285 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
286 int num_advise = ladvise_hdr->lah_count;
287 struct ladvise_hdr *req_ladvise_hdr;
290 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
294 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
295 num_advise * sizeof(*ladvise));
296 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
298 ptlrpc_request_free(req);
301 req->rq_request_portal = OST_IO_PORTAL;
302 ptlrpc_at_set_req_timeout(req);
304 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
306 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
309 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
310 &RMF_OST_LADVISE_HDR);
311 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
313 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
314 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
315 ptlrpc_request_set_replen(req);
318 /* Do not wait for response. */
319 ptlrpcd_add_req(req);
323 req->rq_interpret_reply = osc_ladvise_interpret;
324 CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
325 la = ptlrpc_req_async_args(req);
327 la->la_upcall = upcall;
328 la->la_cookie = cookie;
330 if (rqset == PTLRPCD_SET)
331 ptlrpcd_add_req(req);
333 ptlrpc_set_add_req(rqset, req);
338 static int osc_create(const struct lu_env *env, struct obd_export *exp,
341 struct ptlrpc_request *req;
342 struct ost_body *body;
347 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
348 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
350 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
352 GOTO(out, rc = -ENOMEM);
354 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
356 ptlrpc_request_free(req);
360 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
363 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
365 ptlrpc_request_set_replen(req);
367 rc = ptlrpc_queue_wait(req);
371 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
373 GOTO(out_req, rc = -EPROTO);
375 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
376 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
378 oa->o_blksize = cli_brw_size(exp->exp_obd);
379 oa->o_valid |= OBD_MD_FLBLKSZ;
381 CDEBUG(D_HA, "transno: %lld\n",
382 lustre_msg_get_transno(req->rq_repmsg));
384 ptlrpc_req_finished(req);
389 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
390 obd_enqueue_update_f upcall, void *cookie)
392 struct ptlrpc_request *req;
393 struct osc_setattr_args *sa;
394 struct obd_import *imp = class_exp2cliimp(exp);
395 struct ost_body *body;
400 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
404 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
406 ptlrpc_request_free(req);
410 osc_set_io_portal(req);
412 ptlrpc_at_set_req_timeout(req);
414 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
416 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
418 ptlrpc_request_set_replen(req);
420 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
421 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
422 sa = ptlrpc_req_async_args(req);
424 sa->sa_upcall = upcall;
425 sa->sa_cookie = cookie;
427 ptlrpcd_add_req(req);
431 EXPORT_SYMBOL(osc_punch_send);
433 static int osc_sync_interpret(const struct lu_env *env,
434 struct ptlrpc_request *req,
437 struct osc_fsync_args *fa = arg;
438 struct ost_body *body;
439 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
440 unsigned long valid = 0;
441 struct cl_object *obj;
447 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
449 CERROR("can't unpack ost_body\n");
450 GOTO(out, rc = -EPROTO);
453 *fa->fa_oa = body->oa;
454 obj = osc2cl(fa->fa_obj);
456 /* Update osc object's blocks attribute */
457 cl_object_attr_lock(obj);
458 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
459 attr->cat_blocks = body->oa.o_blocks;
464 cl_object_attr_update(env, obj, attr, valid);
465 cl_object_attr_unlock(obj);
468 rc = fa->fa_upcall(fa->fa_cookie, rc);
472 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
473 obd_enqueue_update_f upcall, void *cookie,
474 struct ptlrpc_request_set *rqset)
476 struct obd_export *exp = osc_export(obj);
477 struct ptlrpc_request *req;
478 struct ost_body *body;
479 struct osc_fsync_args *fa;
483 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
487 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
489 ptlrpc_request_free(req);
493 /* overload the size and blocks fields in the oa with start/end */
494 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
496 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
498 ptlrpc_request_set_replen(req);
499 req->rq_interpret_reply = osc_sync_interpret;
501 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
502 fa = ptlrpc_req_async_args(req);
505 fa->fa_upcall = upcall;
506 fa->fa_cookie = cookie;
508 if (rqset == PTLRPCD_SET)
509 ptlrpcd_add_req(req);
511 ptlrpc_set_add_req(rqset, req);
516 /* Find and cancel locally locks matched by @mode in the resource found by
517 * @objid. Found locks are added into @cancel list. Returns the amount of
518 * locks added to @cancels list. */
519 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
520 struct list_head *cancels,
521 enum ldlm_mode mode, __u64 lock_flags)
523 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
524 struct ldlm_res_id res_id;
525 struct ldlm_resource *res;
529 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
530 * export) but disabled through procfs (flag in NS).
532 * This distinguishes from a case when ELC is not supported originally,
533 * when we still want to cancel locks in advance and just cancel them
534 * locally, without sending any RPC. */
535 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
538 ostid_build_res_name(&oa->o_oi, &res_id);
539 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
543 LDLM_RESOURCE_ADDREF(res);
544 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
545 lock_flags, 0, NULL);
546 LDLM_RESOURCE_DELREF(res);
547 ldlm_resource_putref(res);
551 static int osc_destroy_interpret(const struct lu_env *env,
552 struct ptlrpc_request *req, void *data,
555 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
557 atomic_dec(&cli->cl_destroy_in_flight);
558 wake_up(&cli->cl_destroy_waitq);
562 static int osc_can_send_destroy(struct client_obd *cli)
564 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
565 cli->cl_max_rpcs_in_flight) {
566 /* The destroy request can be sent */
569 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
570 cli->cl_max_rpcs_in_flight) {
572 * The counter has been modified between the two atomic
575 wake_up(&cli->cl_destroy_waitq);
580 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
583 struct client_obd *cli = &exp->exp_obd->u.cli;
584 struct ptlrpc_request *req;
585 struct ost_body *body;
586 struct list_head cancels = LIST_HEAD_INIT(cancels);
591 CDEBUG(D_INFO, "oa NULL\n");
595 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
596 LDLM_FL_DISCARD_DATA);
598 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
600 ldlm_lock_list_put(&cancels, l_bl_ast, count);
604 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
607 ptlrpc_request_free(req);
611 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
612 ptlrpc_at_set_req_timeout(req);
614 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
616 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
618 ptlrpc_request_set_replen(req);
620 req->rq_interpret_reply = osc_destroy_interpret;
621 if (!osc_can_send_destroy(cli)) {
622 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
625 * Wait until the number of on-going destroy RPCs drops
626 * under max_rpc_in_flight
628 rc = l_wait_event_exclusive(cli->cl_destroy_waitq,
629 osc_can_send_destroy(cli), &lwi);
631 ptlrpc_req_finished(req);
636 /* Do not wait for response */
637 ptlrpcd_add_req(req);
641 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
644 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
646 LASSERT(!(oa->o_valid & bits));
649 spin_lock(&cli->cl_loi_list_lock);
650 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
651 oa->o_dirty = cli->cl_dirty_grant;
653 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
654 if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
655 cli->cl_dirty_max_pages)) {
656 CERROR("dirty %lu - %lu > dirty_max %lu\n",
657 cli->cl_dirty_pages, cli->cl_dirty_transit,
658 cli->cl_dirty_max_pages);
660 } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
661 atomic_long_read(&obd_dirty_transit_pages) >
662 (long)(obd_max_dirty_pages + 1))) {
663 /* The atomic_read() allowing the atomic_inc() are
664 * not covered by a lock thus they may safely race and trip
665 * this CERROR() unless we add in a small fudge factor (+1). */
666 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
667 cli_name(cli), atomic_long_read(&obd_dirty_pages),
668 atomic_long_read(&obd_dirty_transit_pages),
669 obd_max_dirty_pages);
671 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
673 CERROR("dirty %lu - dirty_max %lu too big???\n",
674 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
677 unsigned long nrpages;
678 unsigned long undirty;
680 nrpages = cli->cl_max_pages_per_rpc;
681 nrpages *= cli->cl_max_rpcs_in_flight + 1;
682 nrpages = max(nrpages, cli->cl_dirty_max_pages);
683 undirty = nrpages << PAGE_SHIFT;
684 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
688 /* take extent tax into account when asking for more
690 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
691 cli->cl_max_extent_pages;
692 undirty += nrextents * cli->cl_grant_extent_tax;
694 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
695 * to add extent tax, etc.
697 oa->o_undirty = min(undirty, OBD_MAX_GRANT -
698 (PTLRPC_MAX_BRW_PAGES << PAGE_SHIFT)*4UL);
700 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
701 oa->o_dropped = cli->cl_lost_grant;
702 cli->cl_lost_grant = 0;
703 spin_unlock(&cli->cl_loi_list_lock);
704 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
705 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
708 void osc_update_next_shrink(struct client_obd *cli)
710 cli->cl_next_shrink_grant = ktime_get_seconds() +
711 cli->cl_grant_shrink_interval;
713 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
714 cli->cl_next_shrink_grant);
717 static void __osc_update_grant(struct client_obd *cli, u64 grant)
719 spin_lock(&cli->cl_loi_list_lock);
720 cli->cl_avail_grant += grant;
721 spin_unlock(&cli->cl_loi_list_lock);
724 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
726 if (body->oa.o_valid & OBD_MD_FLGRANT) {
727 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
728 __osc_update_grant(cli, body->oa.o_grant);
733 * grant thread data for shrinking space.
735 struct grant_thread_data {
736 struct list_head gtd_clients;
737 struct mutex gtd_mutex;
738 unsigned long gtd_stopped:1;
740 static struct grant_thread_data client_gtd;
742 static int osc_shrink_grant_interpret(const struct lu_env *env,
743 struct ptlrpc_request *req,
746 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
747 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
748 struct ost_body *body;
751 __osc_update_grant(cli, oa->o_grant);
755 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
757 osc_update_grant(cli, body);
759 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
763 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
765 spin_lock(&cli->cl_loi_list_lock);
766 oa->o_grant = cli->cl_avail_grant / 4;
767 cli->cl_avail_grant -= oa->o_grant;
768 spin_unlock(&cli->cl_loi_list_lock);
769 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
770 oa->o_valid |= OBD_MD_FLFLAGS;
773 oa->o_flags |= OBD_FL_SHRINK_GRANT;
774 osc_update_next_shrink(cli);
777 /* Shrink the current grant, either from some large amount to enough for a
778 * full set of in-flight RPCs, or if we have already shrunk to that limit
779 * then to enough for a single RPC. This avoids keeping more grant than
780 * needed, and avoids shrinking the grant piecemeal. */
781 static int osc_shrink_grant(struct client_obd *cli)
783 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
784 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
786 spin_lock(&cli->cl_loi_list_lock);
787 if (cli->cl_avail_grant <= target_bytes)
788 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
789 spin_unlock(&cli->cl_loi_list_lock);
791 return osc_shrink_grant_to_target(cli, target_bytes);
794 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
797 struct ost_body *body;
800 spin_lock(&cli->cl_loi_list_lock);
801 /* Don't shrink if we are already above or below the desired limit
802 * We don't want to shrink below a single RPC, as that will negatively
803 * impact block allocation and long-term performance. */
804 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
805 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
807 if (target_bytes >= cli->cl_avail_grant) {
808 spin_unlock(&cli->cl_loi_list_lock);
811 spin_unlock(&cli->cl_loi_list_lock);
817 osc_announce_cached(cli, &body->oa, 0);
819 spin_lock(&cli->cl_loi_list_lock);
820 if (target_bytes >= cli->cl_avail_grant) {
821 /* available grant has changed since target calculation */
822 spin_unlock(&cli->cl_loi_list_lock);
823 GOTO(out_free, rc = 0);
825 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
826 cli->cl_avail_grant = target_bytes;
827 spin_unlock(&cli->cl_loi_list_lock);
828 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
829 body->oa.o_valid |= OBD_MD_FLFLAGS;
830 body->oa.o_flags = 0;
832 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
833 osc_update_next_shrink(cli);
835 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
836 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
837 sizeof(*body), body, NULL);
839 __osc_update_grant(cli, body->oa.o_grant);
845 static int osc_should_shrink_grant(struct client_obd *client)
847 time64_t next_shrink = client->cl_next_shrink_grant;
849 if (client->cl_import == NULL)
852 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
853 OBD_CONNECT_GRANT_SHRINK) == 0)
856 if (ktime_get_seconds() >= next_shrink - 5) {
857 /* Get the current RPC size directly, instead of going via:
858 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
859 * Keep comment here so that it can be found by searching. */
860 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
862 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
863 client->cl_avail_grant > brw_size)
866 osc_update_next_shrink(client);
871 #define GRANT_SHRINK_RPC_BATCH 100
873 static struct delayed_work work;
875 static void osc_grant_work_handler(struct work_struct *data)
877 struct client_obd *cli;
879 bool init_next_shrink = true;
880 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
883 mutex_lock(&client_gtd.gtd_mutex);
884 list_for_each_entry(cli, &client_gtd.gtd_clients,
886 if (++rpc_sent < GRANT_SHRINK_RPC_BATCH &&
887 osc_should_shrink_grant(cli))
888 osc_shrink_grant(cli);
890 if (!init_next_shrink) {
891 if (cli->cl_next_shrink_grant < next_shrink &&
892 cli->cl_next_shrink_grant > ktime_get_seconds())
893 next_shrink = cli->cl_next_shrink_grant;
895 init_next_shrink = false;
896 next_shrink = cli->cl_next_shrink_grant;
899 mutex_unlock(&client_gtd.gtd_mutex);
901 if (client_gtd.gtd_stopped == 1)
904 if (next_shrink > ktime_get_seconds())
905 schedule_delayed_work(&work, msecs_to_jiffies(
906 (next_shrink - ktime_get_seconds()) *
909 schedule_work(&work.work);
913 * Start grant thread for returing grant to server for idle clients.
915 static int osc_start_grant_work(void)
917 client_gtd.gtd_stopped = 0;
918 mutex_init(&client_gtd.gtd_mutex);
919 INIT_LIST_HEAD(&client_gtd.gtd_clients);
921 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
922 schedule_work(&work.work);
927 static void osc_stop_grant_work(void)
929 client_gtd.gtd_stopped = 1;
930 cancel_delayed_work_sync(&work);
933 static void osc_add_grant_list(struct client_obd *client)
935 mutex_lock(&client_gtd.gtd_mutex);
936 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
937 mutex_unlock(&client_gtd.gtd_mutex);
940 static void osc_del_grant_list(struct client_obd *client)
942 if (list_empty(&client->cl_grant_chain))
945 mutex_lock(&client_gtd.gtd_mutex);
946 list_del_init(&client->cl_grant_chain);
947 mutex_unlock(&client_gtd.gtd_mutex);
950 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
953 * ocd_grant is the total grant amount we're expect to hold: if we've
954 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
955 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
958 * race is tolerable here: if we're evicted, but imp_state already
959 * left EVICTED state, then cl_dirty_pages must be 0 already.
961 spin_lock(&cli->cl_loi_list_lock);
962 cli->cl_avail_grant = ocd->ocd_grant;
963 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
964 cli->cl_avail_grant -= cli->cl_reserved_grant;
965 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
966 cli->cl_avail_grant -= cli->cl_dirty_grant;
968 cli->cl_avail_grant -=
969 cli->cl_dirty_pages << PAGE_SHIFT;
972 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
976 /* overhead for each extent insertion */
977 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
978 /* determine the appropriate chunk size used by osc_extent. */
979 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
980 ocd->ocd_grant_blkbits);
981 /* max_pages_per_rpc must be chunk aligned */
982 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
983 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
984 ~chunk_mask) & chunk_mask;
985 /* determine maximum extent size, in #pages */
986 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
987 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
988 if (cli->cl_max_extent_pages == 0)
989 cli->cl_max_extent_pages = 1;
991 cli->cl_grant_extent_tax = 0;
992 cli->cl_chunkbits = PAGE_SHIFT;
993 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
995 spin_unlock(&cli->cl_loi_list_lock);
997 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
998 "chunk bits: %d cl_max_extent_pages: %d\n",
1000 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1001 cli->cl_max_extent_pages);
1003 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1004 osc_add_grant_list(cli);
1006 EXPORT_SYMBOL(osc_init_grant);
1008 /* We assume that the reason this OSC got a short read is because it read
1009 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1010 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1011 * this stripe never got written at or beyond this stripe offset yet. */
1012 static void handle_short_read(int nob_read, size_t page_count,
1013 struct brw_page **pga)
1018 /* skip bytes read OK */
1019 while (nob_read > 0) {
1020 LASSERT (page_count > 0);
1022 if (pga[i]->count > nob_read) {
1023 /* EOF inside this page */
1024 ptr = kmap(pga[i]->pg) +
1025 (pga[i]->off & ~PAGE_MASK);
1026 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1033 nob_read -= pga[i]->count;
1038 /* zero remaining pages */
1039 while (page_count-- > 0) {
1040 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1041 memset(ptr, 0, pga[i]->count);
1047 static int check_write_rcs(struct ptlrpc_request *req,
1048 int requested_nob, int niocount,
1049 size_t page_count, struct brw_page **pga)
1054 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1055 sizeof(*remote_rcs) *
1057 if (remote_rcs == NULL) {
1058 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1062 /* return error if any niobuf was in error */
1063 for (i = 0; i < niocount; i++) {
1064 if ((int)remote_rcs[i] < 0)
1065 return(remote_rcs[i]);
1067 if (remote_rcs[i] != 0) {
1068 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1069 i, remote_rcs[i], req);
1073 if (req->rq_bulk != NULL &&
1074 req->rq_bulk->bd_nob_transferred != requested_nob) {
1075 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1076 req->rq_bulk->bd_nob_transferred, requested_nob);
1083 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1085 if (p1->flag != p2->flag) {
1086 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1087 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1088 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1090 /* warn if we try to combine flags that we don't know to be
1091 * safe to combine */
1092 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1093 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1094 "report this at https://jira.whamcloud.com/\n",
1095 p1->flag, p2->flag);
1100 return (p1->off + p1->count == p2->off);
1103 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1104 size_t pg_count, struct brw_page **pga,
1105 int opc, obd_dif_csum_fn *fn,
1109 struct ahash_request *req;
1110 /* Used Adler as the default checksum type on top of DIF tags */
1111 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1112 struct page *__page;
1113 unsigned char *buffer;
1115 unsigned int bufsize;
1117 int used_number = 0;
1123 LASSERT(pg_count > 0);
1125 __page = alloc_page(GFP_KERNEL);
1129 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1132 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1133 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1137 buffer = kmap(__page);
1138 guard_start = (__u16 *)buffer;
1139 guard_number = PAGE_SIZE / sizeof(*guard_start);
1140 while (nob > 0 && pg_count > 0) {
1141 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1143 /* corrupt the data before we compute the checksum, to
1144 * simulate an OST->client data error */
1145 if (unlikely(i == 0 && opc == OST_READ &&
1146 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1147 unsigned char *ptr = kmap(pga[i]->pg);
1148 int off = pga[i]->off & ~PAGE_MASK;
1150 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1155 * The left guard number should be able to hold checksums of a
1158 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1159 pga[i]->off & ~PAGE_MASK,
1161 guard_start + used_number,
1162 guard_number - used_number,
1168 used_number += used;
1169 if (used_number == guard_number) {
1170 cfs_crypto_hash_update_page(req, __page, 0,
1171 used_number * sizeof(*guard_start));
1175 nob -= pga[i]->count;
1183 if (used_number != 0)
1184 cfs_crypto_hash_update_page(req, __page, 0,
1185 used_number * sizeof(*guard_start));
1187 bufsize = sizeof(cksum);
1188 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1190 /* For sending we only compute the wrong checksum instead
1191 * of corrupting the data so it is still correct on a redo */
1192 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1197 __free_page(__page);
1201 static int osc_checksum_bulk(int nob, size_t pg_count,
1202 struct brw_page **pga, int opc,
1203 enum cksum_types cksum_type,
1207 struct ahash_request *req;
1208 unsigned int bufsize;
1209 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1211 LASSERT(pg_count > 0);
1213 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1215 CERROR("Unable to initialize checksum hash %s\n",
1216 cfs_crypto_hash_name(cfs_alg));
1217 return PTR_ERR(req);
1220 while (nob > 0 && pg_count > 0) {
1221 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1223 /* corrupt the data before we compute the checksum, to
1224 * simulate an OST->client data error */
1225 if (i == 0 && opc == OST_READ &&
1226 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1227 unsigned char *ptr = kmap(pga[i]->pg);
1228 int off = pga[i]->off & ~PAGE_MASK;
1230 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1233 cfs_crypto_hash_update_page(req, pga[i]->pg,
1234 pga[i]->off & ~PAGE_MASK,
1236 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1237 (int)(pga[i]->off & ~PAGE_MASK));
1239 nob -= pga[i]->count;
1244 bufsize = sizeof(*cksum);
1245 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1247 /* For sending we only compute the wrong checksum instead
1248 * of corrupting the data so it is still correct on a redo */
1249 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1255 static int osc_checksum_bulk_rw(const char *obd_name,
1256 enum cksum_types cksum_type,
1257 int nob, size_t pg_count,
1258 struct brw_page **pga, int opc,
1261 obd_dif_csum_fn *fn = NULL;
1262 int sector_size = 0;
1266 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1269 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1270 opc, fn, sector_size, check_sum);
1272 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1279 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1280 u32 page_count, struct brw_page **pga,
1281 struct ptlrpc_request **reqp, int resend)
1283 struct ptlrpc_request *req;
1284 struct ptlrpc_bulk_desc *desc;
1285 struct ost_body *body;
1286 struct obd_ioobj *ioobj;
1287 struct niobuf_remote *niobuf;
1288 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1289 struct osc_brw_async_args *aa;
1290 struct req_capsule *pill;
1291 struct brw_page *pg_prev;
1293 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1296 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1297 RETURN(-ENOMEM); /* Recoverable */
1298 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1299 RETURN(-EINVAL); /* Fatal */
1301 if ((cmd & OBD_BRW_WRITE) != 0) {
1303 req = ptlrpc_request_alloc_pool(cli->cl_import,
1305 &RQF_OST_BRW_WRITE);
1308 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1313 for (niocount = i = 1; i < page_count; i++) {
1314 if (!can_merge_pages(pga[i - 1], pga[i]))
1318 pill = &req->rq_pill;
1319 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1321 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1322 niocount * sizeof(*niobuf));
1324 for (i = 0; i < page_count; i++)
1325 short_io_size += pga[i]->count;
1327 /* Check if read/write is small enough to be a short io. */
1328 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1329 !imp_connect_shortio(cli->cl_import))
1332 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1333 opc == OST_READ ? 0 : short_io_size);
1334 if (opc == OST_READ)
1335 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1338 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1340 ptlrpc_request_free(req);
1343 osc_set_io_portal(req);
1345 ptlrpc_at_set_req_timeout(req);
1346 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1348 req->rq_no_retry_einprogress = 1;
1350 if (short_io_size != 0) {
1352 short_io_buf = NULL;
1356 desc = ptlrpc_prep_bulk_imp(req, page_count,
1357 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1358 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1359 PTLRPC_BULK_PUT_SINK) |
1360 PTLRPC_BULK_BUF_KIOV,
1362 &ptlrpc_bulk_kiov_pin_ops);
1365 GOTO(out, rc = -ENOMEM);
1366 /* NB request now owns desc and will free it when it gets freed */
1368 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1369 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1370 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1371 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1373 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1375 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1376 * and from_kgid(), because they are asynchronous. Fortunately, variable
1377 * oa contains valid o_uid and o_gid in these two operations.
1378 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1379 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1380 * other process logic */
1381 body->oa.o_uid = oa->o_uid;
1382 body->oa.o_gid = oa->o_gid;
1384 obdo_to_ioobj(oa, ioobj);
1385 ioobj->ioo_bufcnt = niocount;
1386 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1387 * that might be send for this request. The actual number is decided
1388 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1389 * "max - 1" for old client compatibility sending "0", and also so the
1390 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1392 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1394 ioobj_max_brw_set(ioobj, 0);
1396 if (short_io_size != 0) {
1397 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1398 body->oa.o_valid |= OBD_MD_FLFLAGS;
1399 body->oa.o_flags = 0;
1401 body->oa.o_flags |= OBD_FL_SHORT_IO;
1402 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1404 if (opc == OST_WRITE) {
1405 short_io_buf = req_capsule_client_get(pill,
1407 LASSERT(short_io_buf != NULL);
1411 LASSERT(page_count > 0);
1413 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1414 struct brw_page *pg = pga[i];
1415 int poff = pg->off & ~PAGE_MASK;
1417 LASSERT(pg->count > 0);
1418 /* make sure there is no gap in the middle of page array */
1419 LASSERTF(page_count == 1 ||
1420 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1421 ergo(i > 0 && i < page_count - 1,
1422 poff == 0 && pg->count == PAGE_SIZE) &&
1423 ergo(i == page_count - 1, poff == 0)),
1424 "i: %d/%d pg: %p off: %llu, count: %u\n",
1425 i, page_count, pg, pg->off, pg->count);
1426 LASSERTF(i == 0 || pg->off > pg_prev->off,
1427 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1428 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1430 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1431 pg_prev->pg, page_private(pg_prev->pg),
1432 pg_prev->pg->index, pg_prev->off);
1433 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1434 (pg->flag & OBD_BRW_SRVLOCK));
1435 if (short_io_size != 0 && opc == OST_WRITE) {
1436 unsigned char *ptr = ll_kmap_atomic(pg->pg, KM_USER0);
1438 LASSERT(short_io_size >= requested_nob + pg->count);
1439 memcpy(short_io_buf + requested_nob,
1442 ll_kunmap_atomic(ptr, KM_USER0);
1443 } else if (short_io_size == 0) {
1444 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1447 requested_nob += pg->count;
1449 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1451 niobuf->rnb_len += pg->count;
1453 niobuf->rnb_offset = pg->off;
1454 niobuf->rnb_len = pg->count;
1455 niobuf->rnb_flags = pg->flag;
1460 LASSERTF((void *)(niobuf - niocount) ==
1461 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1462 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1463 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1465 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1467 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1468 body->oa.o_valid |= OBD_MD_FLFLAGS;
1469 body->oa.o_flags = 0;
1471 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1474 if (osc_should_shrink_grant(cli))
1475 osc_shrink_grant_local(cli, &body->oa);
1477 /* size[REQ_REC_OFF] still sizeof (*body) */
1478 if (opc == OST_WRITE) {
1479 if (cli->cl_checksum &&
1480 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1481 /* store cl_cksum_type in a local variable since
1482 * it can be changed via lprocfs */
1483 enum cksum_types cksum_type = cli->cl_cksum_type;
1485 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1486 body->oa.o_flags = 0;
1488 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1490 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1492 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1493 requested_nob, page_count,
1497 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1501 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1504 /* save this in 'oa', too, for later checking */
1505 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1506 oa->o_flags |= obd_cksum_type_pack(obd_name,
1509 /* clear out the checksum flag, in case this is a
1510 * resend but cl_checksum is no longer set. b=11238 */
1511 oa->o_valid &= ~OBD_MD_FLCKSUM;
1513 oa->o_cksum = body->oa.o_cksum;
1514 /* 1 RC per niobuf */
1515 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1516 sizeof(__u32) * niocount);
1518 if (cli->cl_checksum &&
1519 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1520 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1521 body->oa.o_flags = 0;
1522 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1523 cli->cl_cksum_type);
1524 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1527 /* Client cksum has been already copied to wire obdo in previous
1528 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1529 * resent due to cksum error, this will allow Server to
1530 * check+dump pages on its side */
1532 ptlrpc_request_set_replen(req);
1534 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1535 aa = ptlrpc_req_async_args(req);
1537 aa->aa_requested_nob = requested_nob;
1538 aa->aa_nio_count = niocount;
1539 aa->aa_page_count = page_count;
1543 INIT_LIST_HEAD(&aa->aa_oaps);
1546 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1547 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1548 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1549 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1553 ptlrpc_req_finished(req);
1557 char dbgcksum_file_name[PATH_MAX];
1559 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1560 struct brw_page **pga, __u32 server_cksum,
1568 /* will only keep dump of pages on first error for the same range in
1569 * file/fid, not during the resends/retries. */
1570 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1571 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1572 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1573 libcfs_debug_file_path_arr :
1574 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1575 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1576 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1577 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1579 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1580 client_cksum, server_cksum);
1581 filp = filp_open(dbgcksum_file_name,
1582 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1586 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1587 "checksum error: rc = %d\n", dbgcksum_file_name,
1590 CERROR("%s: can't open to dump pages with checksum "
1591 "error: rc = %d\n", dbgcksum_file_name, rc);
1595 for (i = 0; i < page_count; i++) {
1596 len = pga[i]->count;
1597 buf = kmap(pga[i]->pg);
1599 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1601 CERROR("%s: wanted to write %u but got %d "
1602 "error\n", dbgcksum_file_name, len, rc);
1607 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1608 dbgcksum_file_name, rc);
1613 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1615 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1616 filp_close(filp, NULL);
1621 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1622 __u32 client_cksum, __u32 server_cksum,
1623 struct osc_brw_async_args *aa)
1625 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1626 enum cksum_types cksum_type;
1627 obd_dif_csum_fn *fn = NULL;
1628 int sector_size = 0;
1634 if (server_cksum == client_cksum) {
1635 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1639 if (aa->aa_cli->cl_checksum_dump)
1640 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1641 server_cksum, client_cksum);
1643 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1646 switch (cksum_type) {
1647 case OBD_CKSUM_T10IP512:
1652 case OBD_CKSUM_T10IP4K:
1657 case OBD_CKSUM_T10CRC512:
1659 fn = obd_dif_crc_fn;
1662 case OBD_CKSUM_T10CRC4K:
1664 fn = obd_dif_crc_fn;
1672 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1680 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1681 aa->aa_ppga, OST_WRITE, cksum_type,
1685 msg = "failed to calculate the client write checksum";
1686 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1687 msg = "the server did not use the checksum type specified in "
1688 "the original request - likely a protocol problem";
1689 else if (new_cksum == server_cksum)
1690 msg = "changed on the client after we checksummed it - "
1691 "likely false positive due to mmap IO (bug 11742)";
1692 else if (new_cksum == client_cksum)
1693 msg = "changed in transit before arrival at OST";
1695 msg = "changed in transit AND doesn't match the original - "
1696 "likely false positive due to mmap IO (bug 11742)";
1698 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1699 DFID " object "DOSTID" extent [%llu-%llu], original "
1700 "client csum %x (type %x), server csum %x (type %x),"
1701 " client csum now %x\n",
1702 obd_name, msg, libcfs_nid2str(peer->nid),
1703 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1704 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1705 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1706 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1707 aa->aa_ppga[aa->aa_page_count - 1]->off +
1708 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1710 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1711 server_cksum, cksum_type, new_cksum);
1715 /* Note rc enters this function as number of bytes transferred */
1716 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1718 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1719 struct client_obd *cli = aa->aa_cli;
1720 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1721 const struct lnet_process_id *peer =
1722 &req->rq_import->imp_connection->c_peer;
1723 struct ost_body *body;
1724 u32 client_cksum = 0;
1727 if (rc < 0 && rc != -EDQUOT) {
1728 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1732 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1733 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1735 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1739 /* set/clear over quota flag for a uid/gid/projid */
1740 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1741 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1742 unsigned qid[LL_MAXQUOTAS] = {
1743 body->oa.o_uid, body->oa.o_gid,
1744 body->oa.o_projid };
1745 CDEBUG(D_QUOTA, "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1746 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1747 body->oa.o_valid, body->oa.o_flags);
1748 osc_quota_setdq(cli, qid, body->oa.o_valid,
1752 osc_update_grant(cli, body);
1757 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1758 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1760 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1762 CERROR("Unexpected +ve rc %d\n", rc);
1766 if (req->rq_bulk != NULL &&
1767 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1770 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1771 check_write_checksum(&body->oa, peer, client_cksum,
1772 body->oa.o_cksum, aa))
1775 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1776 aa->aa_page_count, aa->aa_ppga);
1780 /* The rest of this function executes only for OST_READs */
1782 if (req->rq_bulk == NULL) {
1783 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1785 LASSERT(rc == req->rq_status);
1787 /* if unwrap_bulk failed, return -EAGAIN to retry */
1788 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1791 GOTO(out, rc = -EAGAIN);
1793 if (rc > aa->aa_requested_nob) {
1794 CERROR("Unexpected rc %d (%d requested)\n", rc,
1795 aa->aa_requested_nob);
1799 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1800 CERROR ("Unexpected rc %d (%d transferred)\n",
1801 rc, req->rq_bulk->bd_nob_transferred);
1805 if (req->rq_bulk == NULL) {
1807 int nob, pg_count, i = 0;
1810 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1811 pg_count = aa->aa_page_count;
1812 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1815 while (nob > 0 && pg_count > 0) {
1817 int count = aa->aa_ppga[i]->count > nob ?
1818 nob : aa->aa_ppga[i]->count;
1820 CDEBUG(D_CACHE, "page %p count %d\n",
1821 aa->aa_ppga[i]->pg, count);
1822 ptr = ll_kmap_atomic(aa->aa_ppga[i]->pg, KM_USER0);
1823 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1825 ll_kunmap_atomic((void *) ptr, KM_USER0);
1834 if (rc < aa->aa_requested_nob)
1835 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1837 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1838 static int cksum_counter;
1839 u32 server_cksum = body->oa.o_cksum;
1842 enum cksum_types cksum_type;
1843 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1844 body->oa.o_flags : 0;
1846 cksum_type = obd_cksum_type_unpack(o_flags);
1847 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1848 aa->aa_page_count, aa->aa_ppga,
1849 OST_READ, &client_cksum);
1853 if (req->rq_bulk != NULL &&
1854 peer->nid != req->rq_bulk->bd_sender) {
1856 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1859 if (server_cksum != client_cksum) {
1860 struct ost_body *clbody;
1861 u32 page_count = aa->aa_page_count;
1863 clbody = req_capsule_client_get(&req->rq_pill,
1865 if (cli->cl_checksum_dump)
1866 dump_all_bulk_pages(&clbody->oa, page_count,
1867 aa->aa_ppga, server_cksum,
1870 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1871 "%s%s%s inode "DFID" object "DOSTID
1872 " extent [%llu-%llu], client %x, "
1873 "server %x, cksum_type %x\n",
1875 libcfs_nid2str(peer->nid),
1877 clbody->oa.o_valid & OBD_MD_FLFID ?
1878 clbody->oa.o_parent_seq : 0ULL,
1879 clbody->oa.o_valid & OBD_MD_FLFID ?
1880 clbody->oa.o_parent_oid : 0,
1881 clbody->oa.o_valid & OBD_MD_FLFID ?
1882 clbody->oa.o_parent_ver : 0,
1883 POSTID(&body->oa.o_oi),
1884 aa->aa_ppga[0]->off,
1885 aa->aa_ppga[page_count-1]->off +
1886 aa->aa_ppga[page_count-1]->count - 1,
1887 client_cksum, server_cksum,
1890 aa->aa_oa->o_cksum = client_cksum;
1894 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1897 } else if (unlikely(client_cksum)) {
1898 static int cksum_missed;
1901 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1902 CERROR("Checksum %u requested from %s but not sent\n",
1903 cksum_missed, libcfs_nid2str(peer->nid));
1909 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1910 aa->aa_oa, &body->oa);
1915 static int osc_brw_redo_request(struct ptlrpc_request *request,
1916 struct osc_brw_async_args *aa, int rc)
1918 struct ptlrpc_request *new_req;
1919 struct osc_brw_async_args *new_aa;
1920 struct osc_async_page *oap;
1923 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1924 "redo for recoverable error %d", rc);
1926 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1927 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1928 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1929 aa->aa_ppga, &new_req, 1);
1933 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1934 if (oap->oap_request != NULL) {
1935 LASSERTF(request == oap->oap_request,
1936 "request %p != oap_request %p\n",
1937 request, oap->oap_request);
1938 if (oap->oap_interrupted) {
1939 ptlrpc_req_finished(new_req);
1944 /* New request takes over pga and oaps from old request.
1945 * Note that copying a list_head doesn't work, need to move it... */
1947 new_req->rq_interpret_reply = request->rq_interpret_reply;
1948 new_req->rq_async_args = request->rq_async_args;
1949 new_req->rq_commit_cb = request->rq_commit_cb;
1950 /* cap resend delay to the current request timeout, this is similar to
1951 * what ptlrpc does (see after_reply()) */
1952 if (aa->aa_resends > new_req->rq_timeout)
1953 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1955 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1956 new_req->rq_generation_set = 1;
1957 new_req->rq_import_generation = request->rq_import_generation;
1959 new_aa = ptlrpc_req_async_args(new_req);
1961 INIT_LIST_HEAD(&new_aa->aa_oaps);
1962 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1963 INIT_LIST_HEAD(&new_aa->aa_exts);
1964 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1965 new_aa->aa_resends = aa->aa_resends;
1967 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1968 if (oap->oap_request) {
1969 ptlrpc_req_finished(oap->oap_request);
1970 oap->oap_request = ptlrpc_request_addref(new_req);
1974 /* XXX: This code will run into problem if we're going to support
1975 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1976 * and wait for all of them to be finished. We should inherit request
1977 * set from old request. */
1978 ptlrpcd_add_req(new_req);
1980 DEBUG_REQ(D_INFO, new_req, "new request");
1985 * ugh, we want disk allocation on the target to happen in offset order. we'll
1986 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1987 * fine for our small page arrays and doesn't require allocation. its an
1988 * insertion sort that swaps elements that are strides apart, shrinking the
1989 * stride down until its '1' and the array is sorted.
1991 static void sort_brw_pages(struct brw_page **array, int num)
1994 struct brw_page *tmp;
1998 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2003 for (i = stride ; i < num ; i++) {
2006 while (j >= stride && array[j - stride]->off > tmp->off) {
2007 array[j] = array[j - stride];
2012 } while (stride > 1);
2015 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2017 LASSERT(ppga != NULL);
2018 OBD_FREE(ppga, sizeof(*ppga) * count);
2021 static int brw_interpret(const struct lu_env *env,
2022 struct ptlrpc_request *req, void *data, int rc)
2024 struct osc_brw_async_args *aa = data;
2025 struct osc_extent *ext;
2026 struct osc_extent *tmp;
2027 struct client_obd *cli = aa->aa_cli;
2028 unsigned long transferred = 0;
2031 rc = osc_brw_fini_request(req, rc);
2032 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2033 /* When server return -EINPROGRESS, client should always retry
2034 * regardless of the number of times the bulk was resent already. */
2035 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2036 if (req->rq_import_generation !=
2037 req->rq_import->imp_generation) {
2038 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2039 ""DOSTID", rc = %d.\n",
2040 req->rq_import->imp_obd->obd_name,
2041 POSTID(&aa->aa_oa->o_oi), rc);
2042 } else if (rc == -EINPROGRESS ||
2043 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2044 rc = osc_brw_redo_request(req, aa, rc);
2046 CERROR("%s: too many resent retries for object: "
2047 "%llu:%llu, rc = %d.\n",
2048 req->rq_import->imp_obd->obd_name,
2049 POSTID(&aa->aa_oa->o_oi), rc);
2054 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2059 struct obdo *oa = aa->aa_oa;
2060 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2061 unsigned long valid = 0;
2062 struct cl_object *obj;
2063 struct osc_async_page *last;
2065 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2066 obj = osc2cl(last->oap_obj);
2068 cl_object_attr_lock(obj);
2069 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2070 attr->cat_blocks = oa->o_blocks;
2071 valid |= CAT_BLOCKS;
2073 if (oa->o_valid & OBD_MD_FLMTIME) {
2074 attr->cat_mtime = oa->o_mtime;
2077 if (oa->o_valid & OBD_MD_FLATIME) {
2078 attr->cat_atime = oa->o_atime;
2081 if (oa->o_valid & OBD_MD_FLCTIME) {
2082 attr->cat_ctime = oa->o_ctime;
2086 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2087 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2088 loff_t last_off = last->oap_count + last->oap_obj_off +
2091 /* Change file size if this is an out of quota or
2092 * direct IO write and it extends the file size */
2093 if (loi->loi_lvb.lvb_size < last_off) {
2094 attr->cat_size = last_off;
2097 /* Extend KMS if it's not a lockless write */
2098 if (loi->loi_kms < last_off &&
2099 oap2osc_page(last)->ops_srvlock == 0) {
2100 attr->cat_kms = last_off;
2106 cl_object_attr_update(env, obj, attr, valid);
2107 cl_object_attr_unlock(obj);
2109 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2111 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2112 osc_inc_unstable_pages(req);
2114 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2115 list_del_init(&ext->oe_link);
2116 osc_extent_finish(env, ext, 1,
2117 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2119 LASSERT(list_empty(&aa->aa_exts));
2120 LASSERT(list_empty(&aa->aa_oaps));
2122 transferred = (req->rq_bulk == NULL ? /* short io */
2123 aa->aa_requested_nob :
2124 req->rq_bulk->bd_nob_transferred);
2126 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2127 ptlrpc_lprocfs_brw(req, transferred);
2129 spin_lock(&cli->cl_loi_list_lock);
2130 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2131 * is called so we know whether to go to sync BRWs or wait for more
2132 * RPCs to complete */
2133 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2134 cli->cl_w_in_flight--;
2136 cli->cl_r_in_flight--;
2137 osc_wake_cache_waiters(cli);
2138 spin_unlock(&cli->cl_loi_list_lock);
2140 osc_io_unplug(env, cli, NULL);
2144 static void brw_commit(struct ptlrpc_request *req)
2146 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2147 * this called via the rq_commit_cb, I need to ensure
2148 * osc_dec_unstable_pages is still called. Otherwise unstable
2149 * pages may be leaked. */
2150 spin_lock(&req->rq_lock);
2151 if (likely(req->rq_unstable)) {
2152 req->rq_unstable = 0;
2153 spin_unlock(&req->rq_lock);
2155 osc_dec_unstable_pages(req);
2157 req->rq_committed = 1;
2158 spin_unlock(&req->rq_lock);
2163 * Build an RPC by the list of extent @ext_list. The caller must ensure
2164 * that the total pages in this list are NOT over max pages per RPC.
2165 * Extents in the list must be in OES_RPC state.
2167 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2168 struct list_head *ext_list, int cmd)
2170 struct ptlrpc_request *req = NULL;
2171 struct osc_extent *ext;
2172 struct brw_page **pga = NULL;
2173 struct osc_brw_async_args *aa = NULL;
2174 struct obdo *oa = NULL;
2175 struct osc_async_page *oap;
2176 struct osc_object *obj = NULL;
2177 struct cl_req_attr *crattr = NULL;
2178 loff_t starting_offset = OBD_OBJECT_EOF;
2179 loff_t ending_offset = 0;
2183 bool soft_sync = false;
2184 bool interrupted = false;
2185 bool ndelay = false;
2189 __u32 layout_version = 0;
2190 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
2191 struct ost_body *body;
2193 LASSERT(!list_empty(ext_list));
2195 /* add pages into rpc_list to build BRW rpc */
2196 list_for_each_entry(ext, ext_list, oe_link) {
2197 LASSERT(ext->oe_state == OES_RPC);
2198 mem_tight |= ext->oe_memalloc;
2199 grant += ext->oe_grants;
2200 page_count += ext->oe_nr_pages;
2201 layout_version = MAX(layout_version, ext->oe_layout_version);
2206 soft_sync = osc_over_unstable_soft_limit(cli);
2208 mpflag = cfs_memory_pressure_get_and_set();
2210 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2212 GOTO(out, rc = -ENOMEM);
2214 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2216 GOTO(out, rc = -ENOMEM);
2219 list_for_each_entry(ext, ext_list, oe_link) {
2220 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2222 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2224 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2225 pga[i] = &oap->oap_brw_page;
2226 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2229 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2230 if (starting_offset == OBD_OBJECT_EOF ||
2231 starting_offset > oap->oap_obj_off)
2232 starting_offset = oap->oap_obj_off;
2234 LASSERT(oap->oap_page_off == 0);
2235 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2236 ending_offset = oap->oap_obj_off +
2239 LASSERT(oap->oap_page_off + oap->oap_count ==
2241 if (oap->oap_interrupted)
2248 /* first page in the list */
2249 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2251 crattr = &osc_env_info(env)->oti_req_attr;
2252 memset(crattr, 0, sizeof(*crattr));
2253 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2254 crattr->cra_flags = ~0ULL;
2255 crattr->cra_page = oap2cl_page(oap);
2256 crattr->cra_oa = oa;
2257 cl_req_attr_set(env, osc2cl(obj), crattr);
2259 if (cmd == OBD_BRW_WRITE) {
2260 oa->o_grant_used = grant;
2261 if (layout_version > 0) {
2262 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2263 PFID(&oa->o_oi.oi_fid), layout_version);
2265 oa->o_layout_version = layout_version;
2266 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2270 sort_brw_pages(pga, page_count);
2271 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2273 CERROR("prep_req failed: %d\n", rc);
2277 req->rq_commit_cb = brw_commit;
2278 req->rq_interpret_reply = brw_interpret;
2279 req->rq_memalloc = mem_tight != 0;
2280 oap->oap_request = ptlrpc_request_addref(req);
2281 if (interrupted && !req->rq_intr)
2282 ptlrpc_mark_interrupted(req);
2284 req->rq_no_resend = req->rq_no_delay = 1;
2285 /* probably set a shorter timeout value.
2286 * to handle ETIMEDOUT in brw_interpret() correctly. */
2287 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2290 /* Need to update the timestamps after the request is built in case
2291 * we race with setattr (locally or in queue at OST). If OST gets
2292 * later setattr before earlier BRW (as determined by the request xid),
2293 * the OST will not use BRW timestamps. Sadly, there is no obvious
2294 * way to do this in a single call. bug 10150 */
2295 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2296 crattr->cra_oa = &body->oa;
2297 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2298 cl_req_attr_set(env, osc2cl(obj), crattr);
2299 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2301 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2302 aa = ptlrpc_req_async_args(req);
2303 INIT_LIST_HEAD(&aa->aa_oaps);
2304 list_splice_init(&rpc_list, &aa->aa_oaps);
2305 INIT_LIST_HEAD(&aa->aa_exts);
2306 list_splice_init(ext_list, &aa->aa_exts);
2308 spin_lock(&cli->cl_loi_list_lock);
2309 starting_offset >>= PAGE_SHIFT;
2310 if (cmd == OBD_BRW_READ) {
2311 cli->cl_r_in_flight++;
2312 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2313 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2314 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2315 starting_offset + 1);
2317 cli->cl_w_in_flight++;
2318 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2319 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2320 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2321 starting_offset + 1);
2323 spin_unlock(&cli->cl_loi_list_lock);
2325 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
2326 page_count, aa, cli->cl_r_in_flight,
2327 cli->cl_w_in_flight);
2328 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2330 ptlrpcd_add_req(req);
2336 cfs_memory_pressure_restore(mpflag);
2339 LASSERT(req == NULL);
2342 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2344 OBD_FREE(pga, sizeof(*pga) * page_count);
2345 /* this should happen rarely and is pretty bad, it makes the
2346 * pending list not follow the dirty order */
2347 while (!list_empty(ext_list)) {
2348 ext = list_entry(ext_list->next, struct osc_extent,
2350 list_del_init(&ext->oe_link);
2351 osc_extent_finish(env, ext, 0, rc);
2357 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2361 LASSERT(lock != NULL);
2363 lock_res_and_lock(lock);
2365 if (lock->l_ast_data == NULL)
2366 lock->l_ast_data = data;
2367 if (lock->l_ast_data == data)
2370 unlock_res_and_lock(lock);
2375 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2376 void *cookie, struct lustre_handle *lockh,
2377 enum ldlm_mode mode, __u64 *flags, bool speculative,
2380 bool intent = *flags & LDLM_FL_HAS_INTENT;
2384 /* The request was created before ldlm_cli_enqueue call. */
2385 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2386 struct ldlm_reply *rep;
2388 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2389 LASSERT(rep != NULL);
2391 rep->lock_policy_res1 =
2392 ptlrpc_status_ntoh(rep->lock_policy_res1);
2393 if (rep->lock_policy_res1)
2394 errcode = rep->lock_policy_res1;
2396 *flags |= LDLM_FL_LVB_READY;
2397 } else if (errcode == ELDLM_OK) {
2398 *flags |= LDLM_FL_LVB_READY;
2401 /* Call the update callback. */
2402 rc = (*upcall)(cookie, lockh, errcode);
2404 /* release the reference taken in ldlm_cli_enqueue() */
2405 if (errcode == ELDLM_LOCK_MATCHED)
2407 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2408 ldlm_lock_decref(lockh, mode);
2413 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2414 struct osc_enqueue_args *aa, int rc)
2416 struct ldlm_lock *lock;
2417 struct lustre_handle *lockh = &aa->oa_lockh;
2418 enum ldlm_mode mode = aa->oa_mode;
2419 struct ost_lvb *lvb = aa->oa_lvb;
2420 __u32 lvb_len = sizeof(*lvb);
2425 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2427 lock = ldlm_handle2lock(lockh);
2428 LASSERTF(lock != NULL,
2429 "lockh %#llx, req %p, aa %p - client evicted?\n",
2430 lockh->cookie, req, aa);
2432 /* Take an additional reference so that a blocking AST that
2433 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2434 * to arrive after an upcall has been executed by
2435 * osc_enqueue_fini(). */
2436 ldlm_lock_addref(lockh, mode);
2438 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2439 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2441 /* Let CP AST to grant the lock first. */
2442 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2444 if (aa->oa_speculative) {
2445 LASSERT(aa->oa_lvb == NULL);
2446 LASSERT(aa->oa_flags == NULL);
2447 aa->oa_flags = &flags;
2450 /* Complete obtaining the lock procedure. */
2451 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2452 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2454 /* Complete osc stuff. */
2455 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2456 aa->oa_flags, aa->oa_speculative, rc);
2458 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2460 ldlm_lock_decref(lockh, mode);
2461 LDLM_LOCK_PUT(lock);
2465 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2467 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2468 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2469 * other synchronous requests, however keeping some locks and trying to obtain
2470 * others may take a considerable amount of time in a case of ost failure; and
2471 * when other sync requests do not get released lock from a client, the client
2472 * is evicted from the cluster -- such scenarious make the life difficult, so
2473 * release locks just after they are obtained. */
2474 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2475 __u64 *flags, union ldlm_policy_data *policy,
2476 struct ost_lvb *lvb, int kms_valid,
2477 osc_enqueue_upcall_f upcall, void *cookie,
2478 struct ldlm_enqueue_info *einfo,
2479 struct ptlrpc_request_set *rqset, int async,
2482 struct obd_device *obd = exp->exp_obd;
2483 struct lustre_handle lockh = { 0 };
2484 struct ptlrpc_request *req = NULL;
2485 int intent = *flags & LDLM_FL_HAS_INTENT;
2486 __u64 match_flags = *flags;
2487 enum ldlm_mode mode;
2491 /* Filesystem lock extents are extended to page boundaries so that
2492 * dealing with the page cache is a little smoother. */
2493 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2494 policy->l_extent.end |= ~PAGE_MASK;
2497 * kms is not valid when either object is completely fresh (so that no
2498 * locks are cached), or object was evicted. In the latter case cached
2499 * lock cannot be used, because it would prime inode state with
2500 * potentially stale LVB.
2505 /* Next, search for already existing extent locks that will cover us */
2506 /* If we're trying to read, we also search for an existing PW lock. The
2507 * VFS and page cache already protect us locally, so lots of readers/
2508 * writers can share a single PW lock.
2510 * There are problems with conversion deadlocks, so instead of
2511 * converting a read lock to a write lock, we'll just enqueue a new
2514 * At some point we should cancel the read lock instead of making them
2515 * send us a blocking callback, but there are problems with canceling
2516 * locks out from other users right now, too. */
2517 mode = einfo->ei_mode;
2518 if (einfo->ei_mode == LCK_PR)
2520 /* Normal lock requests must wait for the LVB to be ready before
2521 * matching a lock; speculative lock requests do not need to,
2522 * because they will not actually use the lock. */
2524 match_flags |= LDLM_FL_LVB_READY;
2526 match_flags |= LDLM_FL_BLOCK_GRANTED;
2527 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2528 einfo->ei_type, policy, mode, &lockh, 0);
2530 struct ldlm_lock *matched;
2532 if (*flags & LDLM_FL_TEST_LOCK)
2535 matched = ldlm_handle2lock(&lockh);
2537 /* This DLM lock request is speculative, and does not
2538 * have an associated IO request. Therefore if there
2539 * is already a DLM lock, it wll just inform the
2540 * caller to cancel the request for this stripe.*/
2541 lock_res_and_lock(matched);
2542 if (ldlm_extent_equal(&policy->l_extent,
2543 &matched->l_policy_data.l_extent))
2547 unlock_res_and_lock(matched);
2549 ldlm_lock_decref(&lockh, mode);
2550 LDLM_LOCK_PUT(matched);
2552 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2553 *flags |= LDLM_FL_LVB_READY;
2555 /* We already have a lock, and it's referenced. */
2556 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2558 ldlm_lock_decref(&lockh, mode);
2559 LDLM_LOCK_PUT(matched);
2562 ldlm_lock_decref(&lockh, mode);
2563 LDLM_LOCK_PUT(matched);
2568 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2572 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2573 &RQF_LDLM_ENQUEUE_LVB);
2577 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2579 ptlrpc_request_free(req);
2583 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2585 ptlrpc_request_set_replen(req);
2588 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2589 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2591 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2592 sizeof(*lvb), LVB_T_OST, &lockh, async);
2595 struct osc_enqueue_args *aa;
2596 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2597 aa = ptlrpc_req_async_args(req);
2599 aa->oa_mode = einfo->ei_mode;
2600 aa->oa_type = einfo->ei_type;
2601 lustre_handle_copy(&aa->oa_lockh, &lockh);
2602 aa->oa_upcall = upcall;
2603 aa->oa_cookie = cookie;
2604 aa->oa_speculative = speculative;
2606 aa->oa_flags = flags;
2609 /* speculative locks are essentially to enqueue
2610 * a DLM lock in advance, so we don't care
2611 * about the result of the enqueue. */
2613 aa->oa_flags = NULL;
2616 req->rq_interpret_reply =
2617 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2618 if (rqset == PTLRPCD_SET)
2619 ptlrpcd_add_req(req);
2621 ptlrpc_set_add_req(rqset, req);
2622 } else if (intent) {
2623 ptlrpc_req_finished(req);
2628 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2629 flags, speculative, rc);
2631 ptlrpc_req_finished(req);
2636 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2637 enum ldlm_type type, union ldlm_policy_data *policy,
2638 enum ldlm_mode mode, __u64 *flags, void *data,
2639 struct lustre_handle *lockh, int unref)
2641 struct obd_device *obd = exp->exp_obd;
2642 __u64 lflags = *flags;
2646 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2649 /* Filesystem lock extents are extended to page boundaries so that
2650 * dealing with the page cache is a little smoother */
2651 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2652 policy->l_extent.end |= ~PAGE_MASK;
2654 /* Next, search for already existing extent locks that will cover us */
2655 /* If we're trying to read, we also search for an existing PW lock. The
2656 * VFS and page cache already protect us locally, so lots of readers/
2657 * writers can share a single PW lock. */
2661 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2662 res_id, type, policy, rc, lockh, unref);
2663 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2667 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2669 LASSERT(lock != NULL);
2670 if (!osc_set_lock_data(lock, data)) {
2671 ldlm_lock_decref(lockh, rc);
2674 LDLM_LOCK_PUT(lock);
2679 static int osc_statfs_interpret(const struct lu_env *env,
2680 struct ptlrpc_request *req,
2681 struct osc_async_args *aa, int rc)
2683 struct obd_statfs *msfs;
2687 /* The request has in fact never been sent
2688 * due to issues at a higher level (LOV).
2689 * Exit immediately since the caller is
2690 * aware of the problem and takes care
2691 * of the clean up */
2694 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2695 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2701 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2703 GOTO(out, rc = -EPROTO);
2706 *aa->aa_oi->oi_osfs = *msfs;
2708 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2712 static int osc_statfs_async(struct obd_export *exp,
2713 struct obd_info *oinfo, time64_t max_age,
2714 struct ptlrpc_request_set *rqset)
2716 struct obd_device *obd = class_exp2obd(exp);
2717 struct ptlrpc_request *req;
2718 struct osc_async_args *aa;
2722 /* We could possibly pass max_age in the request (as an absolute
2723 * timestamp or a "seconds.usec ago") so the target can avoid doing
2724 * extra calls into the filesystem if that isn't necessary (e.g.
2725 * during mount that would help a bit). Having relative timestamps
2726 * is not so great if request processing is slow, while absolute
2727 * timestamps are not ideal because they need time synchronization. */
2728 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2732 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2734 ptlrpc_request_free(req);
2737 ptlrpc_request_set_replen(req);
2738 req->rq_request_portal = OST_CREATE_PORTAL;
2739 ptlrpc_at_set_req_timeout(req);
2741 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2742 /* procfs requests not want stat in wait for avoid deadlock */
2743 req->rq_no_resend = 1;
2744 req->rq_no_delay = 1;
2747 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2748 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2749 aa = ptlrpc_req_async_args(req);
2752 ptlrpc_set_add_req(rqset, req);
2756 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2757 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2759 struct obd_device *obd = class_exp2obd(exp);
2760 struct obd_statfs *msfs;
2761 struct ptlrpc_request *req;
2762 struct obd_import *imp = NULL;
2767 /*Since the request might also come from lprocfs, so we need
2768 *sync this with client_disconnect_export Bug15684*/
2769 down_read(&obd->u.cli.cl_sem);
2770 if (obd->u.cli.cl_import)
2771 imp = class_import_get(obd->u.cli.cl_import);
2772 up_read(&obd->u.cli.cl_sem);
2776 /* We could possibly pass max_age in the request (as an absolute
2777 * timestamp or a "seconds.usec ago") so the target can avoid doing
2778 * extra calls into the filesystem if that isn't necessary (e.g.
2779 * during mount that would help a bit). Having relative timestamps
2780 * is not so great if request processing is slow, while absolute
2781 * timestamps are not ideal because they need time synchronization. */
2782 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2784 class_import_put(imp);
2789 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2791 ptlrpc_request_free(req);
2794 ptlrpc_request_set_replen(req);
2795 req->rq_request_portal = OST_CREATE_PORTAL;
2796 ptlrpc_at_set_req_timeout(req);
2798 if (flags & OBD_STATFS_NODELAY) {
2799 /* procfs requests not want stat in wait for avoid deadlock */
2800 req->rq_no_resend = 1;
2801 req->rq_no_delay = 1;
2804 rc = ptlrpc_queue_wait(req);
2808 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2810 GOTO(out, rc = -EPROTO);
2816 ptlrpc_req_finished(req);
2820 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2821 void *karg, void __user *uarg)
2823 struct obd_device *obd = exp->exp_obd;
2824 struct obd_ioctl_data *data = karg;
2828 if (!try_module_get(THIS_MODULE)) {
2829 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2830 module_name(THIS_MODULE));
2834 case OBD_IOC_CLIENT_RECOVER:
2835 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2836 data->ioc_inlbuf1, 0);
2840 case IOC_OSC_SET_ACTIVE:
2841 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2844 case OBD_IOC_PING_TARGET:
2845 err = ptlrpc_obd_ping(obd);
2848 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2849 cmd, current_comm());
2850 GOTO(out, err = -ENOTTY);
2853 module_put(THIS_MODULE);
2857 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2858 u32 keylen, void *key, u32 vallen, void *val,
2859 struct ptlrpc_request_set *set)
2861 struct ptlrpc_request *req;
2862 struct obd_device *obd = exp->exp_obd;
2863 struct obd_import *imp = class_exp2cliimp(exp);
2868 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2870 if (KEY_IS(KEY_CHECKSUM)) {
2871 if (vallen != sizeof(int))
2873 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2877 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2878 sptlrpc_conf_client_adapt(obd);
2882 if (KEY_IS(KEY_FLUSH_CTX)) {
2883 sptlrpc_import_flush_my_ctx(imp);
2887 if (KEY_IS(KEY_CACHE_SET)) {
2888 struct client_obd *cli = &obd->u.cli;
2890 LASSERT(cli->cl_cache == NULL); /* only once */
2891 cli->cl_cache = (struct cl_client_cache *)val;
2892 cl_cache_incref(cli->cl_cache);
2893 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2895 /* add this osc into entity list */
2896 LASSERT(list_empty(&cli->cl_lru_osc));
2897 spin_lock(&cli->cl_cache->ccc_lru_lock);
2898 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2899 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2904 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2905 struct client_obd *cli = &obd->u.cli;
2906 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2907 long target = *(long *)val;
2909 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2914 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2917 /* We pass all other commands directly to OST. Since nobody calls osc
2918 methods directly and everybody is supposed to go through LOV, we
2919 assume lov checked invalid values for us.
2920 The only recognised values so far are evict_by_nid and mds_conn.
2921 Even if something bad goes through, we'd get a -EINVAL from OST
2924 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2925 &RQF_OST_SET_GRANT_INFO :
2930 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2931 RCL_CLIENT, keylen);
2932 if (!KEY_IS(KEY_GRANT_SHRINK))
2933 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2934 RCL_CLIENT, vallen);
2935 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2937 ptlrpc_request_free(req);
2941 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2942 memcpy(tmp, key, keylen);
2943 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2946 memcpy(tmp, val, vallen);
2948 if (KEY_IS(KEY_GRANT_SHRINK)) {
2949 struct osc_grant_args *aa;
2952 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2953 aa = ptlrpc_req_async_args(req);
2954 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2956 ptlrpc_req_finished(req);
2959 *oa = ((struct ost_body *)val)->oa;
2961 req->rq_interpret_reply = osc_shrink_grant_interpret;
2964 ptlrpc_request_set_replen(req);
2965 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2966 LASSERT(set != NULL);
2967 ptlrpc_set_add_req(set, req);
2968 ptlrpc_check_set(NULL, set);
2970 ptlrpcd_add_req(req);
2975 EXPORT_SYMBOL(osc_set_info_async);
2977 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
2978 struct obd_device *obd, struct obd_uuid *cluuid,
2979 struct obd_connect_data *data, void *localdata)
2981 struct client_obd *cli = &obd->u.cli;
2983 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2987 spin_lock(&cli->cl_loi_list_lock);
2988 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2989 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM)
2990 grant += cli->cl_dirty_grant;
2992 grant += cli->cl_dirty_pages << PAGE_SHIFT;
2993 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
2994 lost_grant = cli->cl_lost_grant;
2995 cli->cl_lost_grant = 0;
2996 spin_unlock(&cli->cl_loi_list_lock);
2998 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
2999 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3000 data->ocd_version, data->ocd_grant, lost_grant);
3005 EXPORT_SYMBOL(osc_reconnect);
3007 int osc_disconnect(struct obd_export *exp)
3009 struct obd_device *obd = class_exp2obd(exp);
3012 rc = client_disconnect_export(exp);
3014 * Initially we put del_shrink_grant before disconnect_export, but it
3015 * causes the following problem if setup (connect) and cleanup
3016 * (disconnect) are tangled together.
3017 * connect p1 disconnect p2
3018 * ptlrpc_connect_import
3019 * ............... class_manual_cleanup
3022 * ptlrpc_connect_interrupt
3024 * add this client to shrink list
3026 * Bang! grant shrink thread trigger the shrink. BUG18662
3028 osc_del_grant_list(&obd->u.cli);
3031 EXPORT_SYMBOL(osc_disconnect);
3033 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3034 struct hlist_node *hnode, void *arg)
3036 struct lu_env *env = arg;
3037 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3038 struct ldlm_lock *lock;
3039 struct osc_object *osc = NULL;
3043 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3044 if (lock->l_ast_data != NULL && osc == NULL) {
3045 osc = lock->l_ast_data;
3046 cl_object_get(osc2cl(osc));
3049 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3050 * by the 2nd round of ldlm_namespace_clean() call in
3051 * osc_import_event(). */
3052 ldlm_clear_cleaned(lock);
3057 osc_object_invalidate(env, osc);
3058 cl_object_put(env, osc2cl(osc));
3063 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3065 static int osc_import_event(struct obd_device *obd,
3066 struct obd_import *imp,
3067 enum obd_import_event event)
3069 struct client_obd *cli;
3073 LASSERT(imp->imp_obd == obd);
3076 case IMP_EVENT_DISCON: {
3078 spin_lock(&cli->cl_loi_list_lock);
3079 cli->cl_avail_grant = 0;
3080 cli->cl_lost_grant = 0;
3081 spin_unlock(&cli->cl_loi_list_lock);
3084 case IMP_EVENT_INACTIVE: {
3085 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3088 case IMP_EVENT_INVALIDATE: {
3089 struct ldlm_namespace *ns = obd->obd_namespace;
3093 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3095 env = cl_env_get(&refcheck);
3097 osc_io_unplug(env, &obd->u.cli, NULL);
3099 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3100 osc_ldlm_resource_invalidate,
3102 cl_env_put(env, &refcheck);
3104 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3109 case IMP_EVENT_ACTIVE: {
3110 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3113 case IMP_EVENT_OCD: {
3114 struct obd_connect_data *ocd = &imp->imp_connect_data;
3116 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3117 osc_init_grant(&obd->u.cli, ocd);
3120 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3121 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3123 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3126 case IMP_EVENT_DEACTIVATE: {
3127 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3130 case IMP_EVENT_ACTIVATE: {
3131 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3135 CERROR("Unknown import event %d\n", event);
3142 * Determine whether the lock can be canceled before replaying the lock
3143 * during recovery, see bug16774 for detailed information.
3145 * \retval zero the lock can't be canceled
3146 * \retval other ok to cancel
3148 static int osc_cancel_weight(struct ldlm_lock *lock)
3151 * Cancel all unused and granted extent lock.
3153 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3154 lock->l_granted_mode == lock->l_req_mode &&
3155 osc_ldlm_weigh_ast(lock) == 0)
3161 static int brw_queue_work(const struct lu_env *env, void *data)
3163 struct client_obd *cli = data;
3165 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3167 osc_io_unplug(env, cli, NULL);
3171 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3173 struct client_obd *cli = &obd->u.cli;
3179 rc = ptlrpcd_addref();
3183 rc = client_obd_setup(obd, lcfg);
3185 GOTO(out_ptlrpcd, rc);
3188 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3189 if (IS_ERR(handler))
3190 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3191 cli->cl_writeback_work = handler;
3193 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3194 if (IS_ERR(handler))
3195 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3196 cli->cl_lru_work = handler;
3198 rc = osc_quota_setup(obd);
3200 GOTO(out_ptlrpcd_work, rc);
3202 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3203 osc_update_next_shrink(cli);
3208 if (cli->cl_writeback_work != NULL) {
3209 ptlrpcd_destroy_work(cli->cl_writeback_work);
3210 cli->cl_writeback_work = NULL;
3212 if (cli->cl_lru_work != NULL) {
3213 ptlrpcd_destroy_work(cli->cl_lru_work);
3214 cli->cl_lru_work = NULL;
3216 client_obd_cleanup(obd);
3221 EXPORT_SYMBOL(osc_setup_common);
3223 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3225 struct client_obd *cli = &obd->u.cli;
3233 rc = osc_setup_common(obd, lcfg);
3237 rc = osc_tunables_init(obd);
3242 * We try to control the total number of requests with a upper limit
3243 * osc_reqpool_maxreqcount. There might be some race which will cause
3244 * over-limit allocation, but it is fine.
3246 req_count = atomic_read(&osc_pool_req_count);
3247 if (req_count < osc_reqpool_maxreqcount) {
3248 adding = cli->cl_max_rpcs_in_flight + 2;
3249 if (req_count + adding > osc_reqpool_maxreqcount)
3250 adding = osc_reqpool_maxreqcount - req_count;
3252 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3253 atomic_add(added, &osc_pool_req_count);
3256 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3258 spin_lock(&osc_shrink_lock);
3259 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3260 spin_unlock(&osc_shrink_lock);
3261 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3262 cli->cl_import->imp_idle_debug = D_HA;
3267 int osc_precleanup_common(struct obd_device *obd)
3269 struct client_obd *cli = &obd->u.cli;
3273 * for echo client, export may be on zombie list, wait for
3274 * zombie thread to cull it, because cli.cl_import will be
3275 * cleared in client_disconnect_export():
3276 * class_export_destroy() -> obd_cleanup() ->
3277 * echo_device_free() -> echo_client_cleanup() ->
3278 * obd_disconnect() -> osc_disconnect() ->
3279 * client_disconnect_export()
3281 obd_zombie_barrier();
3282 if (cli->cl_writeback_work) {
3283 ptlrpcd_destroy_work(cli->cl_writeback_work);
3284 cli->cl_writeback_work = NULL;
3287 if (cli->cl_lru_work) {
3288 ptlrpcd_destroy_work(cli->cl_lru_work);
3289 cli->cl_lru_work = NULL;
3292 obd_cleanup_client_import(obd);
3295 EXPORT_SYMBOL(osc_precleanup_common);
3297 static int osc_precleanup(struct obd_device *obd)
3301 osc_precleanup_common(obd);
3303 ptlrpc_lprocfs_unregister_obd(obd);
3307 int osc_cleanup_common(struct obd_device *obd)
3309 struct client_obd *cli = &obd->u.cli;
3314 spin_lock(&osc_shrink_lock);
3315 list_del(&cli->cl_shrink_list);
3316 spin_unlock(&osc_shrink_lock);
3319 if (cli->cl_cache != NULL) {
3320 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3321 spin_lock(&cli->cl_cache->ccc_lru_lock);
3322 list_del_init(&cli->cl_lru_osc);
3323 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3324 cli->cl_lru_left = NULL;
3325 cl_cache_decref(cli->cl_cache);
3326 cli->cl_cache = NULL;
3329 /* free memory of osc quota cache */
3330 osc_quota_cleanup(obd);
3332 rc = client_obd_cleanup(obd);
3337 EXPORT_SYMBOL(osc_cleanup_common);
3339 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3341 ssize_t count = class_modify_config(lcfg, PARAM_OSC,
3342 &obd->obd_kset.kobj);
3343 return count > 0 ? 0 : count;
3346 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
3348 return osc_process_config_base(obd, buf);
3351 static struct obd_ops osc_obd_ops = {
3352 .o_owner = THIS_MODULE,
3353 .o_setup = osc_setup,
3354 .o_precleanup = osc_precleanup,
3355 .o_cleanup = osc_cleanup_common,
3356 .o_add_conn = client_import_add_conn,
3357 .o_del_conn = client_import_del_conn,
3358 .o_connect = client_connect_import,
3359 .o_reconnect = osc_reconnect,
3360 .o_disconnect = osc_disconnect,
3361 .o_statfs = osc_statfs,
3362 .o_statfs_async = osc_statfs_async,
3363 .o_create = osc_create,
3364 .o_destroy = osc_destroy,
3365 .o_getattr = osc_getattr,
3366 .o_setattr = osc_setattr,
3367 .o_iocontrol = osc_iocontrol,
3368 .o_set_info_async = osc_set_info_async,
3369 .o_import_event = osc_import_event,
3370 .o_process_config = osc_process_config,
3371 .o_quotactl = osc_quotactl,
3374 static struct shrinker *osc_cache_shrinker;
3375 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
3376 DEFINE_SPINLOCK(osc_shrink_lock);
3378 #ifndef HAVE_SHRINKER_COUNT
3379 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3381 struct shrink_control scv = {
3382 .nr_to_scan = shrink_param(sc, nr_to_scan),
3383 .gfp_mask = shrink_param(sc, gfp_mask)
3385 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
3386 struct shrinker *shrinker = NULL;
3389 (void)osc_cache_shrink_scan(shrinker, &scv);
3391 return osc_cache_shrink_count(shrinker, &scv);
3395 static int __init osc_init(void)
3397 bool enable_proc = true;
3398 struct obd_type *type;
3399 unsigned int reqpool_size;
3400 unsigned int reqsize;
3402 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3403 osc_cache_shrink_count, osc_cache_shrink_scan);
3406 /* print an address of _any_ initialized kernel symbol from this
3407 * module, to allow debugging with gdb that doesn't support data
3408 * symbols from modules.*/
3409 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3411 rc = lu_kmem_init(osc_caches);
3415 type = class_search_type(LUSTRE_OSP_NAME);
3416 if (type != NULL && type->typ_procsym != NULL)
3417 enable_proc = false;
3419 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
3420 LUSTRE_OSC_NAME, &osc_device_type);
3424 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3426 /* This is obviously too much memory, only prevent overflow here */
3427 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3428 GOTO(out_type, rc = -EINVAL);
3430 reqpool_size = osc_reqpool_mem_max << 20;
3433 while (reqsize < OST_IO_MAXREQSIZE)
3434 reqsize = reqsize << 1;
3437 * We don't enlarge the request count in OSC pool according to
3438 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3439 * tried after normal allocation failed. So a small OSC pool won't
3440 * cause much performance degression in most of cases.
3442 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3444 atomic_set(&osc_pool_req_count, 0);
3445 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3446 ptlrpc_add_rqs_to_pool);
3448 if (osc_rq_pool == NULL)
3449 GOTO(out_type, rc = -ENOMEM);
3451 rc = osc_start_grant_work();
3453 GOTO(out_req_pool, rc);
3458 ptlrpc_free_rq_pool(osc_rq_pool);
3460 class_unregister_type(LUSTRE_OSC_NAME);
3462 lu_kmem_fini(osc_caches);
3467 static void __exit osc_exit(void)
3469 osc_stop_grant_work();
3470 remove_shrinker(osc_cache_shrinker);
3471 class_unregister_type(LUSTRE_OSC_NAME);
3472 lu_kmem_fini(osc_caches);
3473 ptlrpc_free_rq_pool(osc_rq_pool);
3476 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3477 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3478 MODULE_VERSION(LUSTRE_VERSION_STRING);
3479 MODULE_LICENSE("GPL");
3481 module_init(osc_init);
3482 module_exit(osc_exit);