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, 0,
1160 guard_start + used_number,
1161 guard_number - used_number,
1167 used_number += used;
1168 if (used_number == guard_number) {
1169 cfs_crypto_hash_update_page(req, __page, 0,
1170 used_number * sizeof(*guard_start));
1174 nob -= pga[i]->count;
1182 if (used_number != 0)
1183 cfs_crypto_hash_update_page(req, __page, 0,
1184 used_number * sizeof(*guard_start));
1186 bufsize = sizeof(cksum);
1187 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1189 /* For sending we only compute the wrong checksum instead
1190 * of corrupting the data so it is still correct on a redo */
1191 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1196 __free_page(__page);
1200 static int osc_checksum_bulk(int nob, size_t pg_count,
1201 struct brw_page **pga, int opc,
1202 enum cksum_types cksum_type,
1206 struct ahash_request *req;
1207 unsigned int bufsize;
1208 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1210 LASSERT(pg_count > 0);
1212 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1214 CERROR("Unable to initialize checksum hash %s\n",
1215 cfs_crypto_hash_name(cfs_alg));
1216 return PTR_ERR(req);
1219 while (nob > 0 && pg_count > 0) {
1220 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1222 /* corrupt the data before we compute the checksum, to
1223 * simulate an OST->client data error */
1224 if (i == 0 && opc == OST_READ &&
1225 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1226 unsigned char *ptr = kmap(pga[i]->pg);
1227 int off = pga[i]->off & ~PAGE_MASK;
1229 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1232 cfs_crypto_hash_update_page(req, pga[i]->pg,
1233 pga[i]->off & ~PAGE_MASK,
1235 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1236 (int)(pga[i]->off & ~PAGE_MASK));
1238 nob -= pga[i]->count;
1243 bufsize = sizeof(*cksum);
1244 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1246 /* For sending we only compute the wrong checksum instead
1247 * of corrupting the data so it is still correct on a redo */
1248 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1254 static int osc_checksum_bulk_rw(const char *obd_name,
1255 enum cksum_types cksum_type,
1256 int nob, size_t pg_count,
1257 struct brw_page **pga, int opc,
1260 obd_dif_csum_fn *fn = NULL;
1261 int sector_size = 0;
1265 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1268 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1269 opc, fn, sector_size, check_sum);
1271 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1278 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1279 u32 page_count, struct brw_page **pga,
1280 struct ptlrpc_request **reqp, int resend)
1282 struct ptlrpc_request *req;
1283 struct ptlrpc_bulk_desc *desc;
1284 struct ost_body *body;
1285 struct obd_ioobj *ioobj;
1286 struct niobuf_remote *niobuf;
1287 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1288 struct osc_brw_async_args *aa;
1289 struct req_capsule *pill;
1290 struct brw_page *pg_prev;
1292 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1295 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1296 RETURN(-ENOMEM); /* Recoverable */
1297 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1298 RETURN(-EINVAL); /* Fatal */
1300 if ((cmd & OBD_BRW_WRITE) != 0) {
1302 req = ptlrpc_request_alloc_pool(cli->cl_import,
1304 &RQF_OST_BRW_WRITE);
1307 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1312 for (niocount = i = 1; i < page_count; i++) {
1313 if (!can_merge_pages(pga[i - 1], pga[i]))
1317 pill = &req->rq_pill;
1318 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1320 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1321 niocount * sizeof(*niobuf));
1323 for (i = 0; i < page_count; i++)
1324 short_io_size += pga[i]->count;
1326 /* Check if read/write is small enough to be a short io. */
1327 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1328 !imp_connect_shortio(cli->cl_import))
1331 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1332 opc == OST_READ ? 0 : short_io_size);
1333 if (opc == OST_READ)
1334 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1337 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1339 ptlrpc_request_free(req);
1342 osc_set_io_portal(req);
1344 ptlrpc_at_set_req_timeout(req);
1345 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1347 req->rq_no_retry_einprogress = 1;
1349 if (short_io_size != 0) {
1351 short_io_buf = NULL;
1355 desc = ptlrpc_prep_bulk_imp(req, page_count,
1356 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1357 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1358 PTLRPC_BULK_PUT_SINK) |
1359 PTLRPC_BULK_BUF_KIOV,
1361 &ptlrpc_bulk_kiov_pin_ops);
1364 GOTO(out, rc = -ENOMEM);
1365 /* NB request now owns desc and will free it when it gets freed */
1367 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1368 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1369 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1370 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1372 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1374 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1375 * and from_kgid(), because they are asynchronous. Fortunately, variable
1376 * oa contains valid o_uid and o_gid in these two operations.
1377 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1378 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1379 * other process logic */
1380 body->oa.o_uid = oa->o_uid;
1381 body->oa.o_gid = oa->o_gid;
1383 obdo_to_ioobj(oa, ioobj);
1384 ioobj->ioo_bufcnt = niocount;
1385 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1386 * that might be send for this request. The actual number is decided
1387 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1388 * "max - 1" for old client compatibility sending "0", and also so the
1389 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1391 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1393 ioobj_max_brw_set(ioobj, 0);
1395 if (short_io_size != 0) {
1396 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1397 body->oa.o_valid |= OBD_MD_FLFLAGS;
1398 body->oa.o_flags = 0;
1400 body->oa.o_flags |= OBD_FL_SHORT_IO;
1401 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1403 if (opc == OST_WRITE) {
1404 short_io_buf = req_capsule_client_get(pill,
1406 LASSERT(short_io_buf != NULL);
1410 LASSERT(page_count > 0);
1412 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1413 struct brw_page *pg = pga[i];
1414 int poff = pg->off & ~PAGE_MASK;
1416 LASSERT(pg->count > 0);
1417 /* make sure there is no gap in the middle of page array */
1418 LASSERTF(page_count == 1 ||
1419 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1420 ergo(i > 0 && i < page_count - 1,
1421 poff == 0 && pg->count == PAGE_SIZE) &&
1422 ergo(i == page_count - 1, poff == 0)),
1423 "i: %d/%d pg: %p off: %llu, count: %u\n",
1424 i, page_count, pg, pg->off, pg->count);
1425 LASSERTF(i == 0 || pg->off > pg_prev->off,
1426 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1427 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1429 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1430 pg_prev->pg, page_private(pg_prev->pg),
1431 pg_prev->pg->index, pg_prev->off);
1432 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1433 (pg->flag & OBD_BRW_SRVLOCK));
1434 if (short_io_size != 0 && opc == OST_WRITE) {
1435 unsigned char *ptr = ll_kmap_atomic(pg->pg, KM_USER0);
1437 LASSERT(short_io_size >= requested_nob + pg->count);
1438 memcpy(short_io_buf + requested_nob,
1441 ll_kunmap_atomic(ptr, KM_USER0);
1442 } else if (short_io_size == 0) {
1443 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1446 requested_nob += pg->count;
1448 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1450 niobuf->rnb_len += pg->count;
1452 niobuf->rnb_offset = pg->off;
1453 niobuf->rnb_len = pg->count;
1454 niobuf->rnb_flags = pg->flag;
1459 LASSERTF((void *)(niobuf - niocount) ==
1460 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1461 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1462 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1464 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1466 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1467 body->oa.o_valid |= OBD_MD_FLFLAGS;
1468 body->oa.o_flags = 0;
1470 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1473 if (osc_should_shrink_grant(cli))
1474 osc_shrink_grant_local(cli, &body->oa);
1476 /* size[REQ_REC_OFF] still sizeof (*body) */
1477 if (opc == OST_WRITE) {
1478 if (cli->cl_checksum &&
1479 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1480 /* store cl_cksum_type in a local variable since
1481 * it can be changed via lprocfs */
1482 enum cksum_types cksum_type = cli->cl_cksum_type;
1484 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1485 body->oa.o_flags = 0;
1487 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1489 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1491 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1492 requested_nob, page_count,
1496 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1500 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1503 /* save this in 'oa', too, for later checking */
1504 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1505 oa->o_flags |= obd_cksum_type_pack(obd_name,
1508 /* clear out the checksum flag, in case this is a
1509 * resend but cl_checksum is no longer set. b=11238 */
1510 oa->o_valid &= ~OBD_MD_FLCKSUM;
1512 oa->o_cksum = body->oa.o_cksum;
1513 /* 1 RC per niobuf */
1514 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1515 sizeof(__u32) * niocount);
1517 if (cli->cl_checksum &&
1518 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1519 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1520 body->oa.o_flags = 0;
1521 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1522 cli->cl_cksum_type);
1523 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1526 /* Client cksum has been already copied to wire obdo in previous
1527 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1528 * resent due to cksum error, this will allow Server to
1529 * check+dump pages on its side */
1531 ptlrpc_request_set_replen(req);
1533 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1534 aa = ptlrpc_req_async_args(req);
1536 aa->aa_requested_nob = requested_nob;
1537 aa->aa_nio_count = niocount;
1538 aa->aa_page_count = page_count;
1542 INIT_LIST_HEAD(&aa->aa_oaps);
1545 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1546 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1547 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1548 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1552 ptlrpc_req_finished(req);
1556 char dbgcksum_file_name[PATH_MAX];
1558 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1559 struct brw_page **pga, __u32 server_cksum,
1567 /* will only keep dump of pages on first error for the same range in
1568 * file/fid, not during the resends/retries. */
1569 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1570 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1571 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1572 libcfs_debug_file_path_arr :
1573 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1574 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1575 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1576 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1578 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1579 client_cksum, server_cksum);
1580 filp = filp_open(dbgcksum_file_name,
1581 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1585 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1586 "checksum error: rc = %d\n", dbgcksum_file_name,
1589 CERROR("%s: can't open to dump pages with checksum "
1590 "error: rc = %d\n", dbgcksum_file_name, rc);
1594 for (i = 0; i < page_count; i++) {
1595 len = pga[i]->count;
1596 buf = kmap(pga[i]->pg);
1598 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1600 CERROR("%s: wanted to write %u but got %d "
1601 "error\n", dbgcksum_file_name, len, rc);
1606 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1607 dbgcksum_file_name, rc);
1612 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1614 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1615 filp_close(filp, NULL);
1620 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1621 __u32 client_cksum, __u32 server_cksum,
1622 struct osc_brw_async_args *aa)
1624 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1625 enum cksum_types cksum_type;
1626 obd_dif_csum_fn *fn = NULL;
1627 int sector_size = 0;
1633 if (server_cksum == client_cksum) {
1634 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1638 if (aa->aa_cli->cl_checksum_dump)
1639 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1640 server_cksum, client_cksum);
1642 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1645 switch (cksum_type) {
1646 case OBD_CKSUM_T10IP512:
1651 case OBD_CKSUM_T10IP4K:
1656 case OBD_CKSUM_T10CRC512:
1658 fn = obd_dif_crc_fn;
1661 case OBD_CKSUM_T10CRC4K:
1663 fn = obd_dif_crc_fn;
1671 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1679 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1680 aa->aa_ppga, OST_WRITE, cksum_type,
1684 msg = "failed to calculate the client write checksum";
1685 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1686 msg = "the server did not use the checksum type specified in "
1687 "the original request - likely a protocol problem";
1688 else if (new_cksum == server_cksum)
1689 msg = "changed on the client after we checksummed it - "
1690 "likely false positive due to mmap IO (bug 11742)";
1691 else if (new_cksum == client_cksum)
1692 msg = "changed in transit before arrival at OST";
1694 msg = "changed in transit AND doesn't match the original - "
1695 "likely false positive due to mmap IO (bug 11742)";
1697 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1698 DFID " object "DOSTID" extent [%llu-%llu], original "
1699 "client csum %x (type %x), server csum %x (type %x),"
1700 " client csum now %x\n",
1701 obd_name, msg, libcfs_nid2str(peer->nid),
1702 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1703 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1704 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1705 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1706 aa->aa_ppga[aa->aa_page_count - 1]->off +
1707 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1709 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1710 server_cksum, cksum_type, new_cksum);
1714 /* Note rc enters this function as number of bytes transferred */
1715 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1717 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1718 struct client_obd *cli = aa->aa_cli;
1719 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1720 const struct lnet_process_id *peer =
1721 &req->rq_import->imp_connection->c_peer;
1722 struct ost_body *body;
1723 u32 client_cksum = 0;
1726 if (rc < 0 && rc != -EDQUOT) {
1727 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1731 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1732 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1734 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1738 /* set/clear over quota flag for a uid/gid/projid */
1739 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1740 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1741 unsigned qid[LL_MAXQUOTAS] = {
1742 body->oa.o_uid, body->oa.o_gid,
1743 body->oa.o_projid };
1744 CDEBUG(D_QUOTA, "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1745 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1746 body->oa.o_valid, body->oa.o_flags);
1747 osc_quota_setdq(cli, qid, body->oa.o_valid,
1751 osc_update_grant(cli, body);
1756 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1757 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1759 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1761 CERROR("Unexpected +ve rc %d\n", rc);
1765 if (req->rq_bulk != NULL &&
1766 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1769 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1770 check_write_checksum(&body->oa, peer, client_cksum,
1771 body->oa.o_cksum, aa))
1774 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1775 aa->aa_page_count, aa->aa_ppga);
1779 /* The rest of this function executes only for OST_READs */
1781 if (req->rq_bulk == NULL) {
1782 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1784 LASSERT(rc == req->rq_status);
1786 /* if unwrap_bulk failed, return -EAGAIN to retry */
1787 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1790 GOTO(out, rc = -EAGAIN);
1792 if (rc > aa->aa_requested_nob) {
1793 CERROR("Unexpected rc %d (%d requested)\n", rc,
1794 aa->aa_requested_nob);
1798 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1799 CERROR ("Unexpected rc %d (%d transferred)\n",
1800 rc, req->rq_bulk->bd_nob_transferred);
1804 if (req->rq_bulk == NULL) {
1806 int nob, pg_count, i = 0;
1809 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1810 pg_count = aa->aa_page_count;
1811 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1814 while (nob > 0 && pg_count > 0) {
1816 int count = aa->aa_ppga[i]->count > nob ?
1817 nob : aa->aa_ppga[i]->count;
1819 CDEBUG(D_CACHE, "page %p count %d\n",
1820 aa->aa_ppga[i]->pg, count);
1821 ptr = ll_kmap_atomic(aa->aa_ppga[i]->pg, KM_USER0);
1822 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1824 ll_kunmap_atomic((void *) ptr, KM_USER0);
1833 if (rc < aa->aa_requested_nob)
1834 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1836 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1837 static int cksum_counter;
1838 u32 server_cksum = body->oa.o_cksum;
1841 enum cksum_types cksum_type;
1842 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1843 body->oa.o_flags : 0;
1845 cksum_type = obd_cksum_type_unpack(o_flags);
1846 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1847 aa->aa_page_count, aa->aa_ppga,
1848 OST_READ, &client_cksum);
1852 if (req->rq_bulk != NULL &&
1853 peer->nid != req->rq_bulk->bd_sender) {
1855 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1858 if (server_cksum != client_cksum) {
1859 struct ost_body *clbody;
1860 u32 page_count = aa->aa_page_count;
1862 clbody = req_capsule_client_get(&req->rq_pill,
1864 if (cli->cl_checksum_dump)
1865 dump_all_bulk_pages(&clbody->oa, page_count,
1866 aa->aa_ppga, server_cksum,
1869 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1870 "%s%s%s inode "DFID" object "DOSTID
1871 " extent [%llu-%llu], client %x, "
1872 "server %x, cksum_type %x\n",
1874 libcfs_nid2str(peer->nid),
1876 clbody->oa.o_valid & OBD_MD_FLFID ?
1877 clbody->oa.o_parent_seq : 0ULL,
1878 clbody->oa.o_valid & OBD_MD_FLFID ?
1879 clbody->oa.o_parent_oid : 0,
1880 clbody->oa.o_valid & OBD_MD_FLFID ?
1881 clbody->oa.o_parent_ver : 0,
1882 POSTID(&body->oa.o_oi),
1883 aa->aa_ppga[0]->off,
1884 aa->aa_ppga[page_count-1]->off +
1885 aa->aa_ppga[page_count-1]->count - 1,
1886 client_cksum, server_cksum,
1889 aa->aa_oa->o_cksum = client_cksum;
1893 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1896 } else if (unlikely(client_cksum)) {
1897 static int cksum_missed;
1900 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1901 CERROR("Checksum %u requested from %s but not sent\n",
1902 cksum_missed, libcfs_nid2str(peer->nid));
1908 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1909 aa->aa_oa, &body->oa);
1914 static int osc_brw_redo_request(struct ptlrpc_request *request,
1915 struct osc_brw_async_args *aa, int rc)
1917 struct ptlrpc_request *new_req;
1918 struct osc_brw_async_args *new_aa;
1919 struct osc_async_page *oap;
1922 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1923 "redo for recoverable error %d", rc);
1925 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1926 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1927 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1928 aa->aa_ppga, &new_req, 1);
1932 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1933 if (oap->oap_request != NULL) {
1934 LASSERTF(request == oap->oap_request,
1935 "request %p != oap_request %p\n",
1936 request, oap->oap_request);
1937 if (oap->oap_interrupted) {
1938 ptlrpc_req_finished(new_req);
1943 /* New request takes over pga and oaps from old request.
1944 * Note that copying a list_head doesn't work, need to move it... */
1946 new_req->rq_interpret_reply = request->rq_interpret_reply;
1947 new_req->rq_async_args = request->rq_async_args;
1948 new_req->rq_commit_cb = request->rq_commit_cb;
1949 /* cap resend delay to the current request timeout, this is similar to
1950 * what ptlrpc does (see after_reply()) */
1951 if (aa->aa_resends > new_req->rq_timeout)
1952 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1954 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1955 new_req->rq_generation_set = 1;
1956 new_req->rq_import_generation = request->rq_import_generation;
1958 new_aa = ptlrpc_req_async_args(new_req);
1960 INIT_LIST_HEAD(&new_aa->aa_oaps);
1961 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1962 INIT_LIST_HEAD(&new_aa->aa_exts);
1963 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1964 new_aa->aa_resends = aa->aa_resends;
1966 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1967 if (oap->oap_request) {
1968 ptlrpc_req_finished(oap->oap_request);
1969 oap->oap_request = ptlrpc_request_addref(new_req);
1973 /* XXX: This code will run into problem if we're going to support
1974 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1975 * and wait for all of them to be finished. We should inherit request
1976 * set from old request. */
1977 ptlrpcd_add_req(new_req);
1979 DEBUG_REQ(D_INFO, new_req, "new request");
1984 * ugh, we want disk allocation on the target to happen in offset order. we'll
1985 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1986 * fine for our small page arrays and doesn't require allocation. its an
1987 * insertion sort that swaps elements that are strides apart, shrinking the
1988 * stride down until its '1' and the array is sorted.
1990 static void sort_brw_pages(struct brw_page **array, int num)
1993 struct brw_page *tmp;
1997 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2002 for (i = stride ; i < num ; i++) {
2005 while (j >= stride && array[j - stride]->off > tmp->off) {
2006 array[j] = array[j - stride];
2011 } while (stride > 1);
2014 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2016 LASSERT(ppga != NULL);
2017 OBD_FREE(ppga, sizeof(*ppga) * count);
2020 static int brw_interpret(const struct lu_env *env,
2021 struct ptlrpc_request *req, void *data, int rc)
2023 struct osc_brw_async_args *aa = data;
2024 struct osc_extent *ext;
2025 struct osc_extent *tmp;
2026 struct client_obd *cli = aa->aa_cli;
2027 unsigned long transferred = 0;
2030 rc = osc_brw_fini_request(req, rc);
2031 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2032 /* When server return -EINPROGRESS, client should always retry
2033 * regardless of the number of times the bulk was resent already. */
2034 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2035 if (req->rq_import_generation !=
2036 req->rq_import->imp_generation) {
2037 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2038 ""DOSTID", rc = %d.\n",
2039 req->rq_import->imp_obd->obd_name,
2040 POSTID(&aa->aa_oa->o_oi), rc);
2041 } else if (rc == -EINPROGRESS ||
2042 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2043 rc = osc_brw_redo_request(req, aa, rc);
2045 CERROR("%s: too many resent retries for object: "
2046 "%llu:%llu, rc = %d.\n",
2047 req->rq_import->imp_obd->obd_name,
2048 POSTID(&aa->aa_oa->o_oi), rc);
2053 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2058 struct obdo *oa = aa->aa_oa;
2059 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2060 unsigned long valid = 0;
2061 struct cl_object *obj;
2062 struct osc_async_page *last;
2064 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2065 obj = osc2cl(last->oap_obj);
2067 cl_object_attr_lock(obj);
2068 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2069 attr->cat_blocks = oa->o_blocks;
2070 valid |= CAT_BLOCKS;
2072 if (oa->o_valid & OBD_MD_FLMTIME) {
2073 attr->cat_mtime = oa->o_mtime;
2076 if (oa->o_valid & OBD_MD_FLATIME) {
2077 attr->cat_atime = oa->o_atime;
2080 if (oa->o_valid & OBD_MD_FLCTIME) {
2081 attr->cat_ctime = oa->o_ctime;
2085 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2086 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2087 loff_t last_off = last->oap_count + last->oap_obj_off +
2090 /* Change file size if this is an out of quota or
2091 * direct IO write and it extends the file size */
2092 if (loi->loi_lvb.lvb_size < last_off) {
2093 attr->cat_size = last_off;
2096 /* Extend KMS if it's not a lockless write */
2097 if (loi->loi_kms < last_off &&
2098 oap2osc_page(last)->ops_srvlock == 0) {
2099 attr->cat_kms = last_off;
2105 cl_object_attr_update(env, obj, attr, valid);
2106 cl_object_attr_unlock(obj);
2108 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2110 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2111 osc_inc_unstable_pages(req);
2113 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2114 list_del_init(&ext->oe_link);
2115 osc_extent_finish(env, ext, 1,
2116 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2118 LASSERT(list_empty(&aa->aa_exts));
2119 LASSERT(list_empty(&aa->aa_oaps));
2121 transferred = (req->rq_bulk == NULL ? /* short io */
2122 aa->aa_requested_nob :
2123 req->rq_bulk->bd_nob_transferred);
2125 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2126 ptlrpc_lprocfs_brw(req, transferred);
2128 spin_lock(&cli->cl_loi_list_lock);
2129 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2130 * is called so we know whether to go to sync BRWs or wait for more
2131 * RPCs to complete */
2132 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2133 cli->cl_w_in_flight--;
2135 cli->cl_r_in_flight--;
2136 osc_wake_cache_waiters(cli);
2137 spin_unlock(&cli->cl_loi_list_lock);
2139 osc_io_unplug(env, cli, NULL);
2143 static void brw_commit(struct ptlrpc_request *req)
2145 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2146 * this called via the rq_commit_cb, I need to ensure
2147 * osc_dec_unstable_pages is still called. Otherwise unstable
2148 * pages may be leaked. */
2149 spin_lock(&req->rq_lock);
2150 if (likely(req->rq_unstable)) {
2151 req->rq_unstable = 0;
2152 spin_unlock(&req->rq_lock);
2154 osc_dec_unstable_pages(req);
2156 req->rq_committed = 1;
2157 spin_unlock(&req->rq_lock);
2162 * Build an RPC by the list of extent @ext_list. The caller must ensure
2163 * that the total pages in this list are NOT over max pages per RPC.
2164 * Extents in the list must be in OES_RPC state.
2166 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2167 struct list_head *ext_list, int cmd)
2169 struct ptlrpc_request *req = NULL;
2170 struct osc_extent *ext;
2171 struct brw_page **pga = NULL;
2172 struct osc_brw_async_args *aa = NULL;
2173 struct obdo *oa = NULL;
2174 struct osc_async_page *oap;
2175 struct osc_object *obj = NULL;
2176 struct cl_req_attr *crattr = NULL;
2177 loff_t starting_offset = OBD_OBJECT_EOF;
2178 loff_t ending_offset = 0;
2182 bool soft_sync = false;
2183 bool interrupted = false;
2184 bool ndelay = false;
2188 __u32 layout_version = 0;
2189 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
2190 struct ost_body *body;
2192 LASSERT(!list_empty(ext_list));
2194 /* add pages into rpc_list to build BRW rpc */
2195 list_for_each_entry(ext, ext_list, oe_link) {
2196 LASSERT(ext->oe_state == OES_RPC);
2197 mem_tight |= ext->oe_memalloc;
2198 grant += ext->oe_grants;
2199 page_count += ext->oe_nr_pages;
2200 layout_version = MAX(layout_version, ext->oe_layout_version);
2205 soft_sync = osc_over_unstable_soft_limit(cli);
2207 mpflag = cfs_memory_pressure_get_and_set();
2209 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2211 GOTO(out, rc = -ENOMEM);
2213 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2215 GOTO(out, rc = -ENOMEM);
2218 list_for_each_entry(ext, ext_list, oe_link) {
2219 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2221 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2223 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2224 pga[i] = &oap->oap_brw_page;
2225 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2228 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2229 if (starting_offset == OBD_OBJECT_EOF ||
2230 starting_offset > oap->oap_obj_off)
2231 starting_offset = oap->oap_obj_off;
2233 LASSERT(oap->oap_page_off == 0);
2234 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2235 ending_offset = oap->oap_obj_off +
2238 LASSERT(oap->oap_page_off + oap->oap_count ==
2240 if (oap->oap_interrupted)
2247 /* first page in the list */
2248 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2250 crattr = &osc_env_info(env)->oti_req_attr;
2251 memset(crattr, 0, sizeof(*crattr));
2252 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2253 crattr->cra_flags = ~0ULL;
2254 crattr->cra_page = oap2cl_page(oap);
2255 crattr->cra_oa = oa;
2256 cl_req_attr_set(env, osc2cl(obj), crattr);
2258 if (cmd == OBD_BRW_WRITE) {
2259 oa->o_grant_used = grant;
2260 if (layout_version > 0) {
2261 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2262 PFID(&oa->o_oi.oi_fid), layout_version);
2264 oa->o_layout_version = layout_version;
2265 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2269 sort_brw_pages(pga, page_count);
2270 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2272 CERROR("prep_req failed: %d\n", rc);
2276 req->rq_commit_cb = brw_commit;
2277 req->rq_interpret_reply = brw_interpret;
2278 req->rq_memalloc = mem_tight != 0;
2279 oap->oap_request = ptlrpc_request_addref(req);
2280 if (interrupted && !req->rq_intr)
2281 ptlrpc_mark_interrupted(req);
2283 req->rq_no_resend = req->rq_no_delay = 1;
2284 /* probably set a shorter timeout value.
2285 * to handle ETIMEDOUT in brw_interpret() correctly. */
2286 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2289 /* Need to update the timestamps after the request is built in case
2290 * we race with setattr (locally or in queue at OST). If OST gets
2291 * later setattr before earlier BRW (as determined by the request xid),
2292 * the OST will not use BRW timestamps. Sadly, there is no obvious
2293 * way to do this in a single call. bug 10150 */
2294 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2295 crattr->cra_oa = &body->oa;
2296 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2297 cl_req_attr_set(env, osc2cl(obj), crattr);
2298 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2300 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2301 aa = ptlrpc_req_async_args(req);
2302 INIT_LIST_HEAD(&aa->aa_oaps);
2303 list_splice_init(&rpc_list, &aa->aa_oaps);
2304 INIT_LIST_HEAD(&aa->aa_exts);
2305 list_splice_init(ext_list, &aa->aa_exts);
2307 spin_lock(&cli->cl_loi_list_lock);
2308 starting_offset >>= PAGE_SHIFT;
2309 if (cmd == OBD_BRW_READ) {
2310 cli->cl_r_in_flight++;
2311 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2312 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2313 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2314 starting_offset + 1);
2316 cli->cl_w_in_flight++;
2317 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2318 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2319 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2320 starting_offset + 1);
2322 spin_unlock(&cli->cl_loi_list_lock);
2324 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
2325 page_count, aa, cli->cl_r_in_flight,
2326 cli->cl_w_in_flight);
2327 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2329 ptlrpcd_add_req(req);
2335 cfs_memory_pressure_restore(mpflag);
2338 LASSERT(req == NULL);
2341 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2343 OBD_FREE(pga, sizeof(*pga) * page_count);
2344 /* this should happen rarely and is pretty bad, it makes the
2345 * pending list not follow the dirty order */
2346 while (!list_empty(ext_list)) {
2347 ext = list_entry(ext_list->next, struct osc_extent,
2349 list_del_init(&ext->oe_link);
2350 osc_extent_finish(env, ext, 0, rc);
2356 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2360 LASSERT(lock != NULL);
2362 lock_res_and_lock(lock);
2364 if (lock->l_ast_data == NULL)
2365 lock->l_ast_data = data;
2366 if (lock->l_ast_data == data)
2369 unlock_res_and_lock(lock);
2374 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2375 void *cookie, struct lustre_handle *lockh,
2376 enum ldlm_mode mode, __u64 *flags, bool speculative,
2379 bool intent = *flags & LDLM_FL_HAS_INTENT;
2383 /* The request was created before ldlm_cli_enqueue call. */
2384 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2385 struct ldlm_reply *rep;
2387 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2388 LASSERT(rep != NULL);
2390 rep->lock_policy_res1 =
2391 ptlrpc_status_ntoh(rep->lock_policy_res1);
2392 if (rep->lock_policy_res1)
2393 errcode = rep->lock_policy_res1;
2395 *flags |= LDLM_FL_LVB_READY;
2396 } else if (errcode == ELDLM_OK) {
2397 *flags |= LDLM_FL_LVB_READY;
2400 /* Call the update callback. */
2401 rc = (*upcall)(cookie, lockh, errcode);
2403 /* release the reference taken in ldlm_cli_enqueue() */
2404 if (errcode == ELDLM_LOCK_MATCHED)
2406 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2407 ldlm_lock_decref(lockh, mode);
2412 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2413 struct osc_enqueue_args *aa, int rc)
2415 struct ldlm_lock *lock;
2416 struct lustre_handle *lockh = &aa->oa_lockh;
2417 enum ldlm_mode mode = aa->oa_mode;
2418 struct ost_lvb *lvb = aa->oa_lvb;
2419 __u32 lvb_len = sizeof(*lvb);
2424 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2426 lock = ldlm_handle2lock(lockh);
2427 LASSERTF(lock != NULL,
2428 "lockh %#llx, req %p, aa %p - client evicted?\n",
2429 lockh->cookie, req, aa);
2431 /* Take an additional reference so that a blocking AST that
2432 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2433 * to arrive after an upcall has been executed by
2434 * osc_enqueue_fini(). */
2435 ldlm_lock_addref(lockh, mode);
2437 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2438 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2440 /* Let CP AST to grant the lock first. */
2441 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2443 if (aa->oa_speculative) {
2444 LASSERT(aa->oa_lvb == NULL);
2445 LASSERT(aa->oa_flags == NULL);
2446 aa->oa_flags = &flags;
2449 /* Complete obtaining the lock procedure. */
2450 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2451 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2453 /* Complete osc stuff. */
2454 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2455 aa->oa_flags, aa->oa_speculative, rc);
2457 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2459 ldlm_lock_decref(lockh, mode);
2460 LDLM_LOCK_PUT(lock);
2464 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2466 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2467 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2468 * other synchronous requests, however keeping some locks and trying to obtain
2469 * others may take a considerable amount of time in a case of ost failure; and
2470 * when other sync requests do not get released lock from a client, the client
2471 * is evicted from the cluster -- such scenarious make the life difficult, so
2472 * release locks just after they are obtained. */
2473 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2474 __u64 *flags, union ldlm_policy_data *policy,
2475 struct ost_lvb *lvb, int kms_valid,
2476 osc_enqueue_upcall_f upcall, void *cookie,
2477 struct ldlm_enqueue_info *einfo,
2478 struct ptlrpc_request_set *rqset, int async,
2481 struct obd_device *obd = exp->exp_obd;
2482 struct lustre_handle lockh = { 0 };
2483 struct ptlrpc_request *req = NULL;
2484 int intent = *flags & LDLM_FL_HAS_INTENT;
2485 __u64 match_flags = *flags;
2486 enum ldlm_mode mode;
2490 /* Filesystem lock extents are extended to page boundaries so that
2491 * dealing with the page cache is a little smoother. */
2492 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2493 policy->l_extent.end |= ~PAGE_MASK;
2496 * kms is not valid when either object is completely fresh (so that no
2497 * locks are cached), or object was evicted. In the latter case cached
2498 * lock cannot be used, because it would prime inode state with
2499 * potentially stale LVB.
2504 /* Next, search for already existing extent locks that will cover us */
2505 /* If we're trying to read, we also search for an existing PW lock. The
2506 * VFS and page cache already protect us locally, so lots of readers/
2507 * writers can share a single PW lock.
2509 * There are problems with conversion deadlocks, so instead of
2510 * converting a read lock to a write lock, we'll just enqueue a new
2513 * At some point we should cancel the read lock instead of making them
2514 * send us a blocking callback, but there are problems with canceling
2515 * locks out from other users right now, too. */
2516 mode = einfo->ei_mode;
2517 if (einfo->ei_mode == LCK_PR)
2519 /* Normal lock requests must wait for the LVB to be ready before
2520 * matching a lock; speculative lock requests do not need to,
2521 * because they will not actually use the lock. */
2523 match_flags |= LDLM_FL_LVB_READY;
2525 match_flags |= LDLM_FL_BLOCK_GRANTED;
2526 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2527 einfo->ei_type, policy, mode, &lockh, 0);
2529 struct ldlm_lock *matched;
2531 if (*flags & LDLM_FL_TEST_LOCK)
2534 matched = ldlm_handle2lock(&lockh);
2536 /* This DLM lock request is speculative, and does not
2537 * have an associated IO request. Therefore if there
2538 * is already a DLM lock, it wll just inform the
2539 * caller to cancel the request for this stripe.*/
2540 lock_res_and_lock(matched);
2541 if (ldlm_extent_equal(&policy->l_extent,
2542 &matched->l_policy_data.l_extent))
2546 unlock_res_and_lock(matched);
2548 ldlm_lock_decref(&lockh, mode);
2549 LDLM_LOCK_PUT(matched);
2551 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2552 *flags |= LDLM_FL_LVB_READY;
2554 /* We already have a lock, and it's referenced. */
2555 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2557 ldlm_lock_decref(&lockh, mode);
2558 LDLM_LOCK_PUT(matched);
2561 ldlm_lock_decref(&lockh, mode);
2562 LDLM_LOCK_PUT(matched);
2567 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2571 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2572 &RQF_LDLM_ENQUEUE_LVB);
2576 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2578 ptlrpc_request_free(req);
2582 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2584 ptlrpc_request_set_replen(req);
2587 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2588 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2590 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2591 sizeof(*lvb), LVB_T_OST, &lockh, async);
2594 struct osc_enqueue_args *aa;
2595 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2596 aa = ptlrpc_req_async_args(req);
2598 aa->oa_mode = einfo->ei_mode;
2599 aa->oa_type = einfo->ei_type;
2600 lustre_handle_copy(&aa->oa_lockh, &lockh);
2601 aa->oa_upcall = upcall;
2602 aa->oa_cookie = cookie;
2603 aa->oa_speculative = speculative;
2605 aa->oa_flags = flags;
2608 /* speculative locks are essentially to enqueue
2609 * a DLM lock in advance, so we don't care
2610 * about the result of the enqueue. */
2612 aa->oa_flags = NULL;
2615 req->rq_interpret_reply =
2616 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2617 if (rqset == PTLRPCD_SET)
2618 ptlrpcd_add_req(req);
2620 ptlrpc_set_add_req(rqset, req);
2621 } else if (intent) {
2622 ptlrpc_req_finished(req);
2627 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2628 flags, speculative, rc);
2630 ptlrpc_req_finished(req);
2635 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2636 enum ldlm_type type, union ldlm_policy_data *policy,
2637 enum ldlm_mode mode, __u64 *flags, void *data,
2638 struct lustre_handle *lockh, int unref)
2640 struct obd_device *obd = exp->exp_obd;
2641 __u64 lflags = *flags;
2645 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2648 /* Filesystem lock extents are extended to page boundaries so that
2649 * dealing with the page cache is a little smoother */
2650 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2651 policy->l_extent.end |= ~PAGE_MASK;
2653 /* Next, search for already existing extent locks that will cover us */
2654 /* If we're trying to read, we also search for an existing PW lock. The
2655 * VFS and page cache already protect us locally, so lots of readers/
2656 * writers can share a single PW lock. */
2660 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2661 res_id, type, policy, rc, lockh, unref);
2662 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2666 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2668 LASSERT(lock != NULL);
2669 if (!osc_set_lock_data(lock, data)) {
2670 ldlm_lock_decref(lockh, rc);
2673 LDLM_LOCK_PUT(lock);
2678 static int osc_statfs_interpret(const struct lu_env *env,
2679 struct ptlrpc_request *req,
2680 struct osc_async_args *aa, int rc)
2682 struct obd_statfs *msfs;
2686 /* The request has in fact never been sent
2687 * due to issues at a higher level (LOV).
2688 * Exit immediately since the caller is
2689 * aware of the problem and takes care
2690 * of the clean up */
2693 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2694 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2700 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2702 GOTO(out, rc = -EPROTO);
2705 *aa->aa_oi->oi_osfs = *msfs;
2707 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2711 static int osc_statfs_async(struct obd_export *exp,
2712 struct obd_info *oinfo, time64_t max_age,
2713 struct ptlrpc_request_set *rqset)
2715 struct obd_device *obd = class_exp2obd(exp);
2716 struct ptlrpc_request *req;
2717 struct osc_async_args *aa;
2721 /* We could possibly pass max_age in the request (as an absolute
2722 * timestamp or a "seconds.usec ago") so the target can avoid doing
2723 * extra calls into the filesystem if that isn't necessary (e.g.
2724 * during mount that would help a bit). Having relative timestamps
2725 * is not so great if request processing is slow, while absolute
2726 * timestamps are not ideal because they need time synchronization. */
2727 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2731 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2733 ptlrpc_request_free(req);
2736 ptlrpc_request_set_replen(req);
2737 req->rq_request_portal = OST_CREATE_PORTAL;
2738 ptlrpc_at_set_req_timeout(req);
2740 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2741 /* procfs requests not want stat in wait for avoid deadlock */
2742 req->rq_no_resend = 1;
2743 req->rq_no_delay = 1;
2746 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2747 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2748 aa = ptlrpc_req_async_args(req);
2751 ptlrpc_set_add_req(rqset, req);
2755 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2756 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2758 struct obd_device *obd = class_exp2obd(exp);
2759 struct obd_statfs *msfs;
2760 struct ptlrpc_request *req;
2761 struct obd_import *imp = NULL;
2766 /*Since the request might also come from lprocfs, so we need
2767 *sync this with client_disconnect_export Bug15684*/
2768 down_read(&obd->u.cli.cl_sem);
2769 if (obd->u.cli.cl_import)
2770 imp = class_import_get(obd->u.cli.cl_import);
2771 up_read(&obd->u.cli.cl_sem);
2775 /* We could possibly pass max_age in the request (as an absolute
2776 * timestamp or a "seconds.usec ago") so the target can avoid doing
2777 * extra calls into the filesystem if that isn't necessary (e.g.
2778 * during mount that would help a bit). Having relative timestamps
2779 * is not so great if request processing is slow, while absolute
2780 * timestamps are not ideal because they need time synchronization. */
2781 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2783 class_import_put(imp);
2788 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2790 ptlrpc_request_free(req);
2793 ptlrpc_request_set_replen(req);
2794 req->rq_request_portal = OST_CREATE_PORTAL;
2795 ptlrpc_at_set_req_timeout(req);
2797 if (flags & OBD_STATFS_NODELAY) {
2798 /* procfs requests not want stat in wait for avoid deadlock */
2799 req->rq_no_resend = 1;
2800 req->rq_no_delay = 1;
2803 rc = ptlrpc_queue_wait(req);
2807 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2809 GOTO(out, rc = -EPROTO);
2815 ptlrpc_req_finished(req);
2819 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2820 void *karg, void __user *uarg)
2822 struct obd_device *obd = exp->exp_obd;
2823 struct obd_ioctl_data *data = karg;
2827 if (!try_module_get(THIS_MODULE)) {
2828 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2829 module_name(THIS_MODULE));
2833 case OBD_IOC_CLIENT_RECOVER:
2834 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2835 data->ioc_inlbuf1, 0);
2839 case IOC_OSC_SET_ACTIVE:
2840 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2843 case OBD_IOC_PING_TARGET:
2844 err = ptlrpc_obd_ping(obd);
2847 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2848 cmd, current_comm());
2849 GOTO(out, err = -ENOTTY);
2852 module_put(THIS_MODULE);
2856 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2857 u32 keylen, void *key, u32 vallen, void *val,
2858 struct ptlrpc_request_set *set)
2860 struct ptlrpc_request *req;
2861 struct obd_device *obd = exp->exp_obd;
2862 struct obd_import *imp = class_exp2cliimp(exp);
2867 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2869 if (KEY_IS(KEY_CHECKSUM)) {
2870 if (vallen != sizeof(int))
2872 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2876 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2877 sptlrpc_conf_client_adapt(obd);
2881 if (KEY_IS(KEY_FLUSH_CTX)) {
2882 sptlrpc_import_flush_my_ctx(imp);
2886 if (KEY_IS(KEY_CACHE_SET)) {
2887 struct client_obd *cli = &obd->u.cli;
2889 LASSERT(cli->cl_cache == NULL); /* only once */
2890 cli->cl_cache = (struct cl_client_cache *)val;
2891 cl_cache_incref(cli->cl_cache);
2892 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2894 /* add this osc into entity list */
2895 LASSERT(list_empty(&cli->cl_lru_osc));
2896 spin_lock(&cli->cl_cache->ccc_lru_lock);
2897 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2898 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2903 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2904 struct client_obd *cli = &obd->u.cli;
2905 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2906 long target = *(long *)val;
2908 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2913 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2916 /* We pass all other commands directly to OST. Since nobody calls osc
2917 methods directly and everybody is supposed to go through LOV, we
2918 assume lov checked invalid values for us.
2919 The only recognised values so far are evict_by_nid and mds_conn.
2920 Even if something bad goes through, we'd get a -EINVAL from OST
2923 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2924 &RQF_OST_SET_GRANT_INFO :
2929 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2930 RCL_CLIENT, keylen);
2931 if (!KEY_IS(KEY_GRANT_SHRINK))
2932 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2933 RCL_CLIENT, vallen);
2934 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2936 ptlrpc_request_free(req);
2940 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2941 memcpy(tmp, key, keylen);
2942 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2945 memcpy(tmp, val, vallen);
2947 if (KEY_IS(KEY_GRANT_SHRINK)) {
2948 struct osc_grant_args *aa;
2951 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2952 aa = ptlrpc_req_async_args(req);
2953 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2955 ptlrpc_req_finished(req);
2958 *oa = ((struct ost_body *)val)->oa;
2960 req->rq_interpret_reply = osc_shrink_grant_interpret;
2963 ptlrpc_request_set_replen(req);
2964 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2965 LASSERT(set != NULL);
2966 ptlrpc_set_add_req(set, req);
2967 ptlrpc_check_set(NULL, set);
2969 ptlrpcd_add_req(req);
2974 EXPORT_SYMBOL(osc_set_info_async);
2976 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
2977 struct obd_device *obd, struct obd_uuid *cluuid,
2978 struct obd_connect_data *data, void *localdata)
2980 struct client_obd *cli = &obd->u.cli;
2982 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2986 spin_lock(&cli->cl_loi_list_lock);
2987 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2988 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM)
2989 grant += cli->cl_dirty_grant;
2991 grant += cli->cl_dirty_pages << PAGE_SHIFT;
2992 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
2993 lost_grant = cli->cl_lost_grant;
2994 cli->cl_lost_grant = 0;
2995 spin_unlock(&cli->cl_loi_list_lock);
2997 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
2998 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2999 data->ocd_version, data->ocd_grant, lost_grant);
3004 EXPORT_SYMBOL(osc_reconnect);
3006 int osc_disconnect(struct obd_export *exp)
3008 struct obd_device *obd = class_exp2obd(exp);
3011 rc = client_disconnect_export(exp);
3013 * Initially we put del_shrink_grant before disconnect_export, but it
3014 * causes the following problem if setup (connect) and cleanup
3015 * (disconnect) are tangled together.
3016 * connect p1 disconnect p2
3017 * ptlrpc_connect_import
3018 * ............... class_manual_cleanup
3021 * ptlrpc_connect_interrupt
3023 * add this client to shrink list
3025 * Bang! grant shrink thread trigger the shrink. BUG18662
3027 osc_del_grant_list(&obd->u.cli);
3030 EXPORT_SYMBOL(osc_disconnect);
3032 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3033 struct hlist_node *hnode, void *arg)
3035 struct lu_env *env = arg;
3036 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3037 struct ldlm_lock *lock;
3038 struct osc_object *osc = NULL;
3042 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3043 if (lock->l_ast_data != NULL && osc == NULL) {
3044 osc = lock->l_ast_data;
3045 cl_object_get(osc2cl(osc));
3048 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3049 * by the 2nd round of ldlm_namespace_clean() call in
3050 * osc_import_event(). */
3051 ldlm_clear_cleaned(lock);
3056 osc_object_invalidate(env, osc);
3057 cl_object_put(env, osc2cl(osc));
3062 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3064 static int osc_import_event(struct obd_device *obd,
3065 struct obd_import *imp,
3066 enum obd_import_event event)
3068 struct client_obd *cli;
3072 LASSERT(imp->imp_obd == obd);
3075 case IMP_EVENT_DISCON: {
3077 spin_lock(&cli->cl_loi_list_lock);
3078 cli->cl_avail_grant = 0;
3079 cli->cl_lost_grant = 0;
3080 spin_unlock(&cli->cl_loi_list_lock);
3083 case IMP_EVENT_INACTIVE: {
3084 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3087 case IMP_EVENT_INVALIDATE: {
3088 struct ldlm_namespace *ns = obd->obd_namespace;
3092 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3094 env = cl_env_get(&refcheck);
3096 osc_io_unplug(env, &obd->u.cli, NULL);
3098 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3099 osc_ldlm_resource_invalidate,
3101 cl_env_put(env, &refcheck);
3103 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3108 case IMP_EVENT_ACTIVE: {
3109 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3112 case IMP_EVENT_OCD: {
3113 struct obd_connect_data *ocd = &imp->imp_connect_data;
3115 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3116 osc_init_grant(&obd->u.cli, ocd);
3119 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3120 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3122 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3125 case IMP_EVENT_DEACTIVATE: {
3126 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3129 case IMP_EVENT_ACTIVATE: {
3130 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3134 CERROR("Unknown import event %d\n", event);
3141 * Determine whether the lock can be canceled before replaying the lock
3142 * during recovery, see bug16774 for detailed information.
3144 * \retval zero the lock can't be canceled
3145 * \retval other ok to cancel
3147 static int osc_cancel_weight(struct ldlm_lock *lock)
3150 * Cancel all unused and granted extent lock.
3152 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3153 lock->l_granted_mode == lock->l_req_mode &&
3154 osc_ldlm_weigh_ast(lock) == 0)
3160 static int brw_queue_work(const struct lu_env *env, void *data)
3162 struct client_obd *cli = data;
3164 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3166 osc_io_unplug(env, cli, NULL);
3170 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3172 struct client_obd *cli = &obd->u.cli;
3178 rc = ptlrpcd_addref();
3182 rc = client_obd_setup(obd, lcfg);
3184 GOTO(out_ptlrpcd, rc);
3187 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3188 if (IS_ERR(handler))
3189 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3190 cli->cl_writeback_work = handler;
3192 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3193 if (IS_ERR(handler))
3194 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3195 cli->cl_lru_work = handler;
3197 rc = osc_quota_setup(obd);
3199 GOTO(out_ptlrpcd_work, rc);
3201 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3202 osc_update_next_shrink(cli);
3207 if (cli->cl_writeback_work != NULL) {
3208 ptlrpcd_destroy_work(cli->cl_writeback_work);
3209 cli->cl_writeback_work = NULL;
3211 if (cli->cl_lru_work != NULL) {
3212 ptlrpcd_destroy_work(cli->cl_lru_work);
3213 cli->cl_lru_work = NULL;
3215 client_obd_cleanup(obd);
3220 EXPORT_SYMBOL(osc_setup_common);
3222 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3224 struct client_obd *cli = &obd->u.cli;
3232 rc = osc_setup_common(obd, lcfg);
3236 rc = osc_tunables_init(obd);
3241 * We try to control the total number of requests with a upper limit
3242 * osc_reqpool_maxreqcount. There might be some race which will cause
3243 * over-limit allocation, but it is fine.
3245 req_count = atomic_read(&osc_pool_req_count);
3246 if (req_count < osc_reqpool_maxreqcount) {
3247 adding = cli->cl_max_rpcs_in_flight + 2;
3248 if (req_count + adding > osc_reqpool_maxreqcount)
3249 adding = osc_reqpool_maxreqcount - req_count;
3251 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3252 atomic_add(added, &osc_pool_req_count);
3255 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3257 spin_lock(&osc_shrink_lock);
3258 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3259 spin_unlock(&osc_shrink_lock);
3260 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3261 cli->cl_import->imp_idle_debug = D_HA;
3266 int osc_precleanup_common(struct obd_device *obd)
3268 struct client_obd *cli = &obd->u.cli;
3272 * for echo client, export may be on zombie list, wait for
3273 * zombie thread to cull it, because cli.cl_import will be
3274 * cleared in client_disconnect_export():
3275 * class_export_destroy() -> obd_cleanup() ->
3276 * echo_device_free() -> echo_client_cleanup() ->
3277 * obd_disconnect() -> osc_disconnect() ->
3278 * client_disconnect_export()
3280 obd_zombie_barrier();
3281 if (cli->cl_writeback_work) {
3282 ptlrpcd_destroy_work(cli->cl_writeback_work);
3283 cli->cl_writeback_work = NULL;
3286 if (cli->cl_lru_work) {
3287 ptlrpcd_destroy_work(cli->cl_lru_work);
3288 cli->cl_lru_work = NULL;
3291 obd_cleanup_client_import(obd);
3294 EXPORT_SYMBOL(osc_precleanup_common);
3296 static int osc_precleanup(struct obd_device *obd)
3300 osc_precleanup_common(obd);
3302 ptlrpc_lprocfs_unregister_obd(obd);
3306 int osc_cleanup_common(struct obd_device *obd)
3308 struct client_obd *cli = &obd->u.cli;
3313 spin_lock(&osc_shrink_lock);
3314 list_del(&cli->cl_shrink_list);
3315 spin_unlock(&osc_shrink_lock);
3318 if (cli->cl_cache != NULL) {
3319 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3320 spin_lock(&cli->cl_cache->ccc_lru_lock);
3321 list_del_init(&cli->cl_lru_osc);
3322 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3323 cli->cl_lru_left = NULL;
3324 cl_cache_decref(cli->cl_cache);
3325 cli->cl_cache = NULL;
3328 /* free memory of osc quota cache */
3329 osc_quota_cleanup(obd);
3331 rc = client_obd_cleanup(obd);
3336 EXPORT_SYMBOL(osc_cleanup_common);
3338 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3340 ssize_t count = class_modify_config(lcfg, PARAM_OSC,
3341 &obd->obd_kset.kobj);
3342 return count > 0 ? 0 : count;
3345 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
3347 return osc_process_config_base(obd, buf);
3350 static struct obd_ops osc_obd_ops = {
3351 .o_owner = THIS_MODULE,
3352 .o_setup = osc_setup,
3353 .o_precleanup = osc_precleanup,
3354 .o_cleanup = osc_cleanup_common,
3355 .o_add_conn = client_import_add_conn,
3356 .o_del_conn = client_import_del_conn,
3357 .o_connect = client_connect_import,
3358 .o_reconnect = osc_reconnect,
3359 .o_disconnect = osc_disconnect,
3360 .o_statfs = osc_statfs,
3361 .o_statfs_async = osc_statfs_async,
3362 .o_create = osc_create,
3363 .o_destroy = osc_destroy,
3364 .o_getattr = osc_getattr,
3365 .o_setattr = osc_setattr,
3366 .o_iocontrol = osc_iocontrol,
3367 .o_set_info_async = osc_set_info_async,
3368 .o_import_event = osc_import_event,
3369 .o_process_config = osc_process_config,
3370 .o_quotactl = osc_quotactl,
3373 static struct shrinker *osc_cache_shrinker;
3374 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
3375 DEFINE_SPINLOCK(osc_shrink_lock);
3377 #ifndef HAVE_SHRINKER_COUNT
3378 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3380 struct shrink_control scv = {
3381 .nr_to_scan = shrink_param(sc, nr_to_scan),
3382 .gfp_mask = shrink_param(sc, gfp_mask)
3384 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
3385 struct shrinker *shrinker = NULL;
3388 (void)osc_cache_shrink_scan(shrinker, &scv);
3390 return osc_cache_shrink_count(shrinker, &scv);
3394 static int __init osc_init(void)
3396 bool enable_proc = true;
3397 struct obd_type *type;
3398 unsigned int reqpool_size;
3399 unsigned int reqsize;
3401 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3402 osc_cache_shrink_count, osc_cache_shrink_scan);
3405 /* print an address of _any_ initialized kernel symbol from this
3406 * module, to allow debugging with gdb that doesn't support data
3407 * symbols from modules.*/
3408 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3410 rc = lu_kmem_init(osc_caches);
3414 type = class_search_type(LUSTRE_OSP_NAME);
3415 if (type != NULL && type->typ_procsym != NULL)
3416 enable_proc = false;
3418 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
3419 LUSTRE_OSC_NAME, &osc_device_type);
3423 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3425 /* This is obviously too much memory, only prevent overflow here */
3426 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3427 GOTO(out_type, rc = -EINVAL);
3429 reqpool_size = osc_reqpool_mem_max << 20;
3432 while (reqsize < OST_IO_MAXREQSIZE)
3433 reqsize = reqsize << 1;
3436 * We don't enlarge the request count in OSC pool according to
3437 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3438 * tried after normal allocation failed. So a small OSC pool won't
3439 * cause much performance degression in most of cases.
3441 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3443 atomic_set(&osc_pool_req_count, 0);
3444 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3445 ptlrpc_add_rqs_to_pool);
3447 if (osc_rq_pool == NULL)
3448 GOTO(out_type, rc = -ENOMEM);
3450 rc = osc_start_grant_work();
3452 GOTO(out_req_pool, rc);
3457 ptlrpc_free_rq_pool(osc_rq_pool);
3459 class_unregister_type(LUSTRE_OSC_NAME);
3461 lu_kmem_fini(osc_caches);
3466 static void __exit osc_exit(void)
3468 osc_stop_grant_work();
3469 remove_shrinker(osc_cache_shrinker);
3470 class_unregister_type(LUSTRE_OSC_NAME);
3471 lu_kmem_fini(osc_caches);
3472 ptlrpc_free_rq_pool(osc_rq_pool);
3475 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3476 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3477 MODULE_VERSION(LUSTRE_VERSION_STRING);
3478 MODULE_LICENSE("GPL");
3480 module_init(osc_init);
3481 module_exit(osc_exit);