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 <libcfs/libcfs.h>
37 #include <linux/falloc.h>
38 #include <lprocfs_status.h>
39 #include <lustre_debug.h>
40 #include <lustre_dlm.h>
41 #include <lustre_fid.h>
42 #include <lustre_ha.h>
43 #include <uapi/linux/lustre/lustre_ioctl.h>
44 #include <lustre_net.h>
45 #include <lustre_obdo.h>
47 #include <obd_cksum.h>
48 #include <obd_class.h>
49 #include <lustre_osc.h>
50 #include <linux/falloc.h>
52 #include "osc_internal.h"
54 atomic_t osc_pool_req_count;
55 unsigned int osc_reqpool_maxreqcount;
56 struct ptlrpc_request_pool *osc_rq_pool;
58 /* max memory used for request pool, unit is MB */
59 static unsigned int osc_reqpool_mem_max = 5;
60 module_param(osc_reqpool_mem_max, uint, 0444);
62 static int osc_idle_timeout = 20;
63 module_param(osc_idle_timeout, uint, 0644);
65 #define osc_grant_args osc_brw_async_args
67 struct osc_setattr_args {
69 obd_enqueue_update_f sa_upcall;
73 struct osc_fsync_args {
74 struct osc_object *fa_obj;
76 obd_enqueue_update_f fa_upcall;
80 struct osc_ladvise_args {
82 obd_enqueue_update_f la_upcall;
86 static void osc_release_ppga(struct brw_page **ppga, size_t count);
87 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
90 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
92 struct ost_body *body;
94 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
97 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
100 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
103 struct ptlrpc_request *req;
104 struct ost_body *body;
108 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
112 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
114 ptlrpc_request_free(req);
118 osc_pack_req_body(req, oa);
120 ptlrpc_request_set_replen(req);
122 rc = ptlrpc_queue_wait(req);
126 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
128 GOTO(out, rc = -EPROTO);
130 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
131 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
133 oa->o_blksize = cli_brw_size(exp->exp_obd);
134 oa->o_valid |= OBD_MD_FLBLKSZ;
138 ptlrpc_req_finished(req);
143 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
146 struct ptlrpc_request *req;
147 struct ost_body *body;
151 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
153 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
157 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
159 ptlrpc_request_free(req);
163 osc_pack_req_body(req, oa);
165 ptlrpc_request_set_replen(req);
167 rc = ptlrpc_queue_wait(req);
171 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
173 GOTO(out, rc = -EPROTO);
175 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
179 ptlrpc_req_finished(req);
184 static int osc_setattr_interpret(const struct lu_env *env,
185 struct ptlrpc_request *req, void *args, int rc)
187 struct osc_setattr_args *sa = args;
188 struct ost_body *body;
195 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
197 GOTO(out, rc = -EPROTO);
199 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
202 rc = sa->sa_upcall(sa->sa_cookie, rc);
206 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
207 obd_enqueue_update_f upcall, void *cookie,
208 struct ptlrpc_request_set *rqset)
210 struct ptlrpc_request *req;
211 struct osc_setattr_args *sa;
216 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
220 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
222 ptlrpc_request_free(req);
226 osc_pack_req_body(req, oa);
228 ptlrpc_request_set_replen(req);
230 /* do mds to ost setattr asynchronously */
232 /* Do not wait for response. */
233 ptlrpcd_add_req(req);
235 req->rq_interpret_reply = osc_setattr_interpret;
237 sa = ptlrpc_req_async_args(sa, req);
239 sa->sa_upcall = upcall;
240 sa->sa_cookie = cookie;
242 ptlrpc_set_add_req(rqset, req);
248 static int osc_ladvise_interpret(const struct lu_env *env,
249 struct ptlrpc_request *req,
252 struct osc_ladvise_args *la = arg;
253 struct ost_body *body;
259 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
261 GOTO(out, rc = -EPROTO);
263 *la->la_oa = body->oa;
265 rc = la->la_upcall(la->la_cookie, rc);
270 * If rqset is NULL, do not wait for response. Upcall and cookie could also
271 * be NULL in this case
273 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
274 struct ladvise_hdr *ladvise_hdr,
275 obd_enqueue_update_f upcall, void *cookie,
276 struct ptlrpc_request_set *rqset)
278 struct ptlrpc_request *req;
279 struct ost_body *body;
280 struct osc_ladvise_args *la;
282 struct lu_ladvise *req_ladvise;
283 struct lu_ladvise *ladvise = ladvise_hdr->lah_advise;
284 int num_advise = ladvise_hdr->lah_count;
285 struct ladvise_hdr *req_ladvise_hdr;
288 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
292 req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
293 num_advise * sizeof(*ladvise));
294 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
296 ptlrpc_request_free(req);
299 req->rq_request_portal = OST_IO_PORTAL;
300 ptlrpc_at_set_req_timeout(req);
302 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
304 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
307 req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
308 &RMF_OST_LADVISE_HDR);
309 memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
311 req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
312 memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
313 ptlrpc_request_set_replen(req);
316 /* Do not wait for response. */
317 ptlrpcd_add_req(req);
321 req->rq_interpret_reply = osc_ladvise_interpret;
322 la = ptlrpc_req_async_args(la, req);
324 la->la_upcall = upcall;
325 la->la_cookie = cookie;
327 ptlrpc_set_add_req(rqset, req);
332 static int osc_create(const struct lu_env *env, struct obd_export *exp,
335 struct ptlrpc_request *req;
336 struct ost_body *body;
341 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
342 LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
344 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
346 GOTO(out, rc = -ENOMEM);
348 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
350 ptlrpc_request_free(req);
354 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
357 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
359 ptlrpc_request_set_replen(req);
361 rc = ptlrpc_queue_wait(req);
365 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
367 GOTO(out_req, rc = -EPROTO);
369 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
370 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
372 oa->o_blksize = cli_brw_size(exp->exp_obd);
373 oa->o_valid |= OBD_MD_FLBLKSZ;
375 CDEBUG(D_HA, "transno: %lld\n",
376 lustre_msg_get_transno(req->rq_repmsg));
378 ptlrpc_req_finished(req);
383 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
384 obd_enqueue_update_f upcall, void *cookie)
386 struct ptlrpc_request *req;
387 struct osc_setattr_args *sa;
388 struct obd_import *imp = class_exp2cliimp(exp);
389 struct ost_body *body;
394 req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
398 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
400 ptlrpc_request_free(req);
404 osc_set_io_portal(req);
406 ptlrpc_at_set_req_timeout(req);
408 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
410 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
412 ptlrpc_request_set_replen(req);
414 req->rq_interpret_reply = osc_setattr_interpret;
415 sa = ptlrpc_req_async_args(sa, req);
417 sa->sa_upcall = upcall;
418 sa->sa_cookie = cookie;
420 ptlrpcd_add_req(req);
424 EXPORT_SYMBOL(osc_punch_send);
427 * osc_fallocate_base() - Handles fallocate request.
429 * @exp: Export structure
430 * @oa: Attributes passed to OSS from client (obdo structure)
431 * @upcall: Primary & supplementary group information
432 * @cookie: Exclusive identifier
433 * @rqset: Request list.
434 * @mode: Operation done on given range.
436 * osc_fallocate_base() - Handles fallocate requests only. Only block
437 * allocation or standard preallocate operation is supported currently.
438 * Other mode flags is not supported yet. ftruncate(2) or truncate(2)
439 * is supported via SETATTR request.
441 * Return: Non-zero on failure and O on success.
443 int osc_fallocate_base(struct obd_export *exp, struct obdo *oa,
444 obd_enqueue_update_f upcall, void *cookie, int mode)
446 struct ptlrpc_request *req;
447 struct osc_setattr_args *sa;
448 struct ost_body *body;
449 struct obd_import *imp = class_exp2cliimp(exp);
454 * Only mode == 0 (which is standard prealloc) is supported now.
455 * Punch is not supported yet.
457 if (mode & ~FALLOC_FL_KEEP_SIZE)
459 oa->o_falloc_mode = mode;
461 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
466 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_FALLOCATE);
468 ptlrpc_request_free(req);
472 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
475 lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
477 ptlrpc_request_set_replen(req);
479 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
480 BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
481 sa = ptlrpc_req_async_args(sa, req);
483 sa->sa_upcall = upcall;
484 sa->sa_cookie = cookie;
486 ptlrpcd_add_req(req);
491 static int osc_sync_interpret(const struct lu_env *env,
492 struct ptlrpc_request *req, void *args, int rc)
494 struct osc_fsync_args *fa = args;
495 struct ost_body *body;
496 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
497 unsigned long valid = 0;
498 struct cl_object *obj;
504 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
506 CERROR("can't unpack ost_body\n");
507 GOTO(out, rc = -EPROTO);
510 *fa->fa_oa = body->oa;
511 obj = osc2cl(fa->fa_obj);
513 /* Update osc object's blocks attribute */
514 cl_object_attr_lock(obj);
515 if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
516 attr->cat_blocks = body->oa.o_blocks;
521 cl_object_attr_update(env, obj, attr, valid);
522 cl_object_attr_unlock(obj);
525 rc = fa->fa_upcall(fa->fa_cookie, rc);
529 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
530 obd_enqueue_update_f upcall, void *cookie,
531 struct ptlrpc_request_set *rqset)
533 struct obd_export *exp = osc_export(obj);
534 struct ptlrpc_request *req;
535 struct ost_body *body;
536 struct osc_fsync_args *fa;
540 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
544 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
546 ptlrpc_request_free(req);
550 /* overload the size and blocks fields in the oa with start/end */
551 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
553 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
555 ptlrpc_request_set_replen(req);
556 req->rq_interpret_reply = osc_sync_interpret;
558 fa = ptlrpc_req_async_args(fa, req);
561 fa->fa_upcall = upcall;
562 fa->fa_cookie = cookie;
564 ptlrpc_set_add_req(rqset, req);
569 /* Find and cancel locally locks matched by @mode in the resource found by
570 * @objid. Found locks are added into @cancel list. Returns the amount of
571 * locks added to @cancels list. */
572 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
573 struct list_head *cancels,
574 enum ldlm_mode mode, __u64 lock_flags)
576 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
577 struct ldlm_res_id res_id;
578 struct ldlm_resource *res;
582 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
583 * export) but disabled through procfs (flag in NS).
585 * This distinguishes from a case when ELC is not supported originally,
586 * when we still want to cancel locks in advance and just cancel them
587 * locally, without sending any RPC. */
588 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
591 ostid_build_res_name(&oa->o_oi, &res_id);
592 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
596 LDLM_RESOURCE_ADDREF(res);
597 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
598 lock_flags, 0, NULL);
599 LDLM_RESOURCE_DELREF(res);
600 ldlm_resource_putref(res);
604 static int osc_destroy_interpret(const struct lu_env *env,
605 struct ptlrpc_request *req, void *args, int rc)
607 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
609 atomic_dec(&cli->cl_destroy_in_flight);
610 wake_up(&cli->cl_destroy_waitq);
615 static int osc_can_send_destroy(struct client_obd *cli)
617 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
618 cli->cl_max_rpcs_in_flight) {
619 /* The destroy request can be sent */
622 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
623 cli->cl_max_rpcs_in_flight) {
625 * The counter has been modified between the two atomic
628 wake_up(&cli->cl_destroy_waitq);
633 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
636 struct client_obd *cli = &exp->exp_obd->u.cli;
637 struct ptlrpc_request *req;
638 struct ost_body *body;
644 CDEBUG(D_INFO, "oa NULL\n");
648 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
649 LDLM_FL_DISCARD_DATA);
651 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
653 ldlm_lock_list_put(&cancels, l_bl_ast, count);
657 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
660 ptlrpc_request_free(req);
664 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
665 ptlrpc_at_set_req_timeout(req);
667 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
669 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
671 ptlrpc_request_set_replen(req);
673 req->rq_interpret_reply = osc_destroy_interpret;
674 if (!osc_can_send_destroy(cli)) {
676 * Wait until the number of on-going destroy RPCs drops
677 * under max_rpc_in_flight
679 rc = l_wait_event_abortable_exclusive(
680 cli->cl_destroy_waitq,
681 osc_can_send_destroy(cli));
683 ptlrpc_req_finished(req);
688 /* Do not wait for response */
689 ptlrpcd_add_req(req);
693 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
696 u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
698 LASSERT(!(oa->o_valid & bits));
701 spin_lock(&cli->cl_loi_list_lock);
702 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
703 oa->o_dirty = cli->cl_dirty_grant;
705 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
706 if (unlikely(cli->cl_dirty_pages > cli->cl_dirty_max_pages)) {
707 CERROR("dirty %lu > dirty_max %lu\n",
709 cli->cl_dirty_max_pages);
711 } else if (unlikely(atomic_long_read(&obd_dirty_pages) >
712 (long)(obd_max_dirty_pages + 1))) {
713 /* The atomic_read() allowing the atomic_inc() are
714 * not covered by a lock thus they may safely race and trip
715 * this CERROR() unless we add in a small fudge factor (+1). */
716 CERROR("%s: dirty %ld > system dirty_max %ld\n",
717 cli_name(cli), atomic_long_read(&obd_dirty_pages),
718 obd_max_dirty_pages);
720 } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
722 CERROR("dirty %lu - dirty_max %lu too big???\n",
723 cli->cl_dirty_pages, cli->cl_dirty_max_pages);
726 unsigned long nrpages;
727 unsigned long undirty;
729 nrpages = cli->cl_max_pages_per_rpc;
730 nrpages *= cli->cl_max_rpcs_in_flight + 1;
731 nrpages = max(nrpages, cli->cl_dirty_max_pages);
732 undirty = nrpages << PAGE_SHIFT;
733 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
737 /* take extent tax into account when asking for more
739 nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
740 cli->cl_max_extent_pages;
741 undirty += nrextents * cli->cl_grant_extent_tax;
743 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
744 * to add extent tax, etc.
746 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
747 ~(PTLRPC_MAX_BRW_SIZE * 4UL));
749 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
750 oa->o_dropped = cli->cl_lost_grant;
751 cli->cl_lost_grant = 0;
752 spin_unlock(&cli->cl_loi_list_lock);
753 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
754 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
757 void osc_update_next_shrink(struct client_obd *cli)
759 cli->cl_next_shrink_grant = ktime_get_seconds() +
760 cli->cl_grant_shrink_interval;
762 CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
763 cli->cl_next_shrink_grant);
766 static void __osc_update_grant(struct client_obd *cli, u64 grant)
768 spin_lock(&cli->cl_loi_list_lock);
769 cli->cl_avail_grant += grant;
770 spin_unlock(&cli->cl_loi_list_lock);
773 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
775 if (body->oa.o_valid & OBD_MD_FLGRANT) {
776 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
777 __osc_update_grant(cli, body->oa.o_grant);
782 * grant thread data for shrinking space.
784 struct grant_thread_data {
785 struct list_head gtd_clients;
786 struct mutex gtd_mutex;
787 unsigned long gtd_stopped:1;
789 static struct grant_thread_data client_gtd;
791 static int osc_shrink_grant_interpret(const struct lu_env *env,
792 struct ptlrpc_request *req,
795 struct osc_grant_args *aa = args;
796 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
797 struct ost_body *body;
800 __osc_update_grant(cli, aa->aa_oa->o_grant);
804 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
806 osc_update_grant(cli, body);
808 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
814 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
816 spin_lock(&cli->cl_loi_list_lock);
817 oa->o_grant = cli->cl_avail_grant / 4;
818 cli->cl_avail_grant -= oa->o_grant;
819 spin_unlock(&cli->cl_loi_list_lock);
820 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
821 oa->o_valid |= OBD_MD_FLFLAGS;
824 oa->o_flags |= OBD_FL_SHRINK_GRANT;
825 osc_update_next_shrink(cli);
828 /* Shrink the current grant, either from some large amount to enough for a
829 * full set of in-flight RPCs, or if we have already shrunk to that limit
830 * then to enough for a single RPC. This avoids keeping more grant than
831 * needed, and avoids shrinking the grant piecemeal. */
832 static int osc_shrink_grant(struct client_obd *cli)
834 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
835 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
837 spin_lock(&cli->cl_loi_list_lock);
838 if (cli->cl_avail_grant <= target_bytes)
839 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
840 spin_unlock(&cli->cl_loi_list_lock);
842 return osc_shrink_grant_to_target(cli, target_bytes);
845 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
848 struct ost_body *body;
851 spin_lock(&cli->cl_loi_list_lock);
852 /* Don't shrink if we are already above or below the desired limit
853 * We don't want to shrink below a single RPC, as that will negatively
854 * impact block allocation and long-term performance. */
855 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
856 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
858 if (target_bytes >= cli->cl_avail_grant) {
859 spin_unlock(&cli->cl_loi_list_lock);
862 spin_unlock(&cli->cl_loi_list_lock);
868 osc_announce_cached(cli, &body->oa, 0);
870 spin_lock(&cli->cl_loi_list_lock);
871 if (target_bytes >= cli->cl_avail_grant) {
872 /* available grant has changed since target calculation */
873 spin_unlock(&cli->cl_loi_list_lock);
874 GOTO(out_free, rc = 0);
876 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
877 cli->cl_avail_grant = target_bytes;
878 spin_unlock(&cli->cl_loi_list_lock);
879 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
880 body->oa.o_valid |= OBD_MD_FLFLAGS;
881 body->oa.o_flags = 0;
883 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
884 osc_update_next_shrink(cli);
886 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
887 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
888 sizeof(*body), body, NULL);
890 __osc_update_grant(cli, body->oa.o_grant);
896 static int osc_should_shrink_grant(struct client_obd *client)
898 time64_t next_shrink = client->cl_next_shrink_grant;
900 if (client->cl_import == NULL)
903 if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
904 client->cl_import->imp_grant_shrink_disabled) {
905 osc_update_next_shrink(client);
909 if (ktime_get_seconds() >= next_shrink - 5) {
910 /* Get the current RPC size directly, instead of going via:
911 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
912 * Keep comment here so that it can be found by searching. */
913 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
915 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
916 client->cl_avail_grant > brw_size)
919 osc_update_next_shrink(client);
924 #define GRANT_SHRINK_RPC_BATCH 100
926 static struct delayed_work work;
928 static void osc_grant_work_handler(struct work_struct *data)
930 struct client_obd *cli;
932 bool init_next_shrink = true;
933 time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
936 mutex_lock(&client_gtd.gtd_mutex);
937 list_for_each_entry(cli, &client_gtd.gtd_clients,
939 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
940 osc_should_shrink_grant(cli)) {
941 osc_shrink_grant(cli);
945 if (!init_next_shrink) {
946 if (cli->cl_next_shrink_grant < next_shrink &&
947 cli->cl_next_shrink_grant > ktime_get_seconds())
948 next_shrink = cli->cl_next_shrink_grant;
950 init_next_shrink = false;
951 next_shrink = cli->cl_next_shrink_grant;
954 mutex_unlock(&client_gtd.gtd_mutex);
956 if (client_gtd.gtd_stopped == 1)
959 if (next_shrink > ktime_get_seconds()) {
960 time64_t delay = next_shrink - ktime_get_seconds();
962 schedule_delayed_work(&work, cfs_time_seconds(delay));
964 schedule_work(&work.work);
968 void osc_schedule_grant_work(void)
970 cancel_delayed_work_sync(&work);
971 schedule_work(&work.work);
975 * Start grant thread for returing grant to server for idle clients.
977 static int osc_start_grant_work(void)
979 client_gtd.gtd_stopped = 0;
980 mutex_init(&client_gtd.gtd_mutex);
981 INIT_LIST_HEAD(&client_gtd.gtd_clients);
983 INIT_DELAYED_WORK(&work, osc_grant_work_handler);
984 schedule_work(&work.work);
989 static void osc_stop_grant_work(void)
991 client_gtd.gtd_stopped = 1;
992 cancel_delayed_work_sync(&work);
995 static void osc_add_grant_list(struct client_obd *client)
997 mutex_lock(&client_gtd.gtd_mutex);
998 list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
999 mutex_unlock(&client_gtd.gtd_mutex);
1002 static void osc_del_grant_list(struct client_obd *client)
1004 if (list_empty(&client->cl_grant_chain))
1007 mutex_lock(&client_gtd.gtd_mutex);
1008 list_del_init(&client->cl_grant_chain);
1009 mutex_unlock(&client_gtd.gtd_mutex);
1012 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1015 * ocd_grant is the total grant amount we're expect to hold: if we've
1016 * been evicted, it's the new avail_grant amount, cl_dirty_pages will
1017 * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
1020 * race is tolerable here: if we're evicted, but imp_state already
1021 * left EVICTED state, then cl_dirty_pages must be 0 already.
1023 spin_lock(&cli->cl_loi_list_lock);
1024 cli->cl_avail_grant = ocd->ocd_grant;
1025 if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
1026 cli->cl_avail_grant -= cli->cl_reserved_grant;
1027 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
1028 cli->cl_avail_grant -= cli->cl_dirty_grant;
1030 cli->cl_avail_grant -=
1031 cli->cl_dirty_pages << PAGE_SHIFT;
1034 if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
1038 /* overhead for each extent insertion */
1039 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
1040 /* determine the appropriate chunk size used by osc_extent. */
1041 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
1042 ocd->ocd_grant_blkbits);
1043 /* max_pages_per_rpc must be chunk aligned */
1044 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
1045 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
1046 ~chunk_mask) & chunk_mask;
1047 /* determine maximum extent size, in #pages */
1048 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
1049 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
1050 if (cli->cl_max_extent_pages == 0)
1051 cli->cl_max_extent_pages = 1;
1053 cli->cl_grant_extent_tax = 0;
1054 cli->cl_chunkbits = PAGE_SHIFT;
1055 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1057 spin_unlock(&cli->cl_loi_list_lock);
1060 "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld. chunk bits: %d cl_max_extent_pages: %d\n",
1062 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1063 cli->cl_max_extent_pages);
1065 if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1066 osc_add_grant_list(cli);
1068 EXPORT_SYMBOL(osc_init_grant);
1070 /* We assume that the reason this OSC got a short read is because it read
1071 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1072 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1073 * this stripe never got written at or beyond this stripe offset yet. */
1074 static void handle_short_read(int nob_read, size_t page_count,
1075 struct brw_page **pga)
1080 /* skip bytes read OK */
1081 while (nob_read > 0) {
1082 LASSERT (page_count > 0);
1084 if (pga[i]->count > nob_read) {
1085 /* EOF inside this page */
1086 ptr = kmap(pga[i]->pg) +
1087 (pga[i]->off & ~PAGE_MASK);
1088 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1095 nob_read -= pga[i]->count;
1100 /* zero remaining pages */
1101 while (page_count-- > 0) {
1102 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1103 memset(ptr, 0, pga[i]->count);
1109 static int check_write_rcs(struct ptlrpc_request *req,
1110 int requested_nob, int niocount,
1111 size_t page_count, struct brw_page **pga)
1116 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1117 sizeof(*remote_rcs) *
1119 if (remote_rcs == NULL) {
1120 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1124 /* return error if any niobuf was in error */
1125 for (i = 0; i < niocount; i++) {
1126 if ((int)remote_rcs[i] < 0) {
1127 CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1128 i, remote_rcs[i], req);
1129 return remote_rcs[i];
1132 if (remote_rcs[i] != 0) {
1133 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1134 i, remote_rcs[i], req);
1138 if (req->rq_bulk != NULL &&
1139 req->rq_bulk->bd_nob_transferred != requested_nob) {
1140 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1141 req->rq_bulk->bd_nob_transferred, requested_nob);
1148 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1150 if (p1->flag != p2->flag) {
1151 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1152 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1153 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1155 /* warn if we try to combine flags that we don't know to be
1156 * safe to combine */
1157 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1158 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1159 "report this at https://jira.whamcloud.com/\n",
1160 p1->flag, p2->flag);
1165 return (p1->off + p1->count == p2->off);
1168 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1169 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1170 size_t pg_count, struct brw_page **pga,
1171 int opc, obd_dif_csum_fn *fn,
1175 struct ahash_request *req;
1176 /* Used Adler as the default checksum type on top of DIF tags */
1177 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1178 struct page *__page;
1179 unsigned char *buffer;
1181 unsigned int bufsize;
1183 int used_number = 0;
1189 LASSERT(pg_count > 0);
1191 __page = alloc_page(GFP_KERNEL);
1195 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1198 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1199 obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1203 buffer = kmap(__page);
1204 guard_start = (__u16 *)buffer;
1205 guard_number = PAGE_SIZE / sizeof(*guard_start);
1206 while (nob > 0 && pg_count > 0) {
1207 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1209 /* corrupt the data before we compute the checksum, to
1210 * simulate an OST->client data error */
1211 if (unlikely(i == 0 && opc == OST_READ &&
1212 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1213 unsigned char *ptr = kmap(pga[i]->pg);
1214 int off = pga[i]->off & ~PAGE_MASK;
1216 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1221 * The left guard number should be able to hold checksums of a
1224 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1225 pga[i]->off & ~PAGE_MASK,
1227 guard_start + used_number,
1228 guard_number - used_number,
1234 used_number += used;
1235 if (used_number == guard_number) {
1236 cfs_crypto_hash_update_page(req, __page, 0,
1237 used_number * sizeof(*guard_start));
1241 nob -= pga[i]->count;
1249 if (used_number != 0)
1250 cfs_crypto_hash_update_page(req, __page, 0,
1251 used_number * sizeof(*guard_start));
1253 bufsize = sizeof(cksum);
1254 cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1256 /* For sending we only compute the wrong checksum instead
1257 * of corrupting the data so it is still correct on a redo */
1258 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1263 __free_page(__page);
1266 #else /* !CONFIG_CRC_T10DIF */
1267 #define obd_dif_ip_fn NULL
1268 #define obd_dif_crc_fn NULL
1269 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum) \
1271 #endif /* CONFIG_CRC_T10DIF */
1273 static int osc_checksum_bulk(int nob, size_t pg_count,
1274 struct brw_page **pga, int opc,
1275 enum cksum_types cksum_type,
1279 struct ahash_request *req;
1280 unsigned int bufsize;
1281 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1283 LASSERT(pg_count > 0);
1285 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1287 CERROR("Unable to initialize checksum hash %s\n",
1288 cfs_crypto_hash_name(cfs_alg));
1289 return PTR_ERR(req);
1292 while (nob > 0 && pg_count > 0) {
1293 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1295 /* corrupt the data before we compute the checksum, to
1296 * simulate an OST->client data error */
1297 if (i == 0 && opc == OST_READ &&
1298 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1299 unsigned char *ptr = kmap(pga[i]->pg);
1300 int off = pga[i]->off & ~PAGE_MASK;
1302 memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1305 cfs_crypto_hash_update_page(req, pga[i]->pg,
1306 pga[i]->off & ~PAGE_MASK,
1308 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1309 (int)(pga[i]->off & ~PAGE_MASK));
1311 nob -= pga[i]->count;
1316 bufsize = sizeof(*cksum);
1317 cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1319 /* For sending we only compute the wrong checksum instead
1320 * of corrupting the data so it is still correct on a redo */
1321 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1327 static int osc_checksum_bulk_rw(const char *obd_name,
1328 enum cksum_types cksum_type,
1329 int nob, size_t pg_count,
1330 struct brw_page **pga, int opc,
1333 obd_dif_csum_fn *fn = NULL;
1334 int sector_size = 0;
1338 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
1341 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1342 opc, fn, sector_size, check_sum);
1344 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1351 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1352 u32 page_count, struct brw_page **pga,
1353 struct ptlrpc_request **reqp, int resend)
1355 struct ptlrpc_request *req;
1356 struct ptlrpc_bulk_desc *desc;
1357 struct ost_body *body;
1358 struct obd_ioobj *ioobj;
1359 struct niobuf_remote *niobuf;
1360 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1361 struct osc_brw_async_args *aa;
1362 struct req_capsule *pill;
1363 struct brw_page *pg_prev;
1365 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1368 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1369 RETURN(-ENOMEM); /* Recoverable */
1370 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1371 RETURN(-EINVAL); /* Fatal */
1373 if ((cmd & OBD_BRW_WRITE) != 0) {
1375 req = ptlrpc_request_alloc_pool(cli->cl_import,
1377 &RQF_OST_BRW_WRITE);
1380 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1385 for (niocount = i = 1; i < page_count; i++) {
1386 if (!can_merge_pages(pga[i - 1], pga[i]))
1390 pill = &req->rq_pill;
1391 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1393 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1394 niocount * sizeof(*niobuf));
1396 for (i = 0; i < page_count; i++)
1397 short_io_size += pga[i]->count;
1399 /* Check if read/write is small enough to be a short io. */
1400 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1401 !imp_connect_shortio(cli->cl_import))
1404 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1405 opc == OST_READ ? 0 : short_io_size);
1406 if (opc == OST_READ)
1407 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1410 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1412 ptlrpc_request_free(req);
1415 osc_set_io_portal(req);
1417 ptlrpc_at_set_req_timeout(req);
1418 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1420 req->rq_no_retry_einprogress = 1;
1422 if (short_io_size != 0) {
1424 short_io_buf = NULL;
1428 desc = ptlrpc_prep_bulk_imp(req, page_count,
1429 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1430 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1431 PTLRPC_BULK_PUT_SINK),
1433 &ptlrpc_bulk_kiov_pin_ops);
1436 GOTO(out, rc = -ENOMEM);
1437 /* NB request now owns desc and will free it when it gets freed */
1439 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1440 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1441 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1442 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1444 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1446 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1447 * and from_kgid(), because they are asynchronous. Fortunately, variable
1448 * oa contains valid o_uid and o_gid in these two operations.
1449 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1450 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1451 * other process logic */
1452 body->oa.o_uid = oa->o_uid;
1453 body->oa.o_gid = oa->o_gid;
1455 obdo_to_ioobj(oa, ioobj);
1456 ioobj->ioo_bufcnt = niocount;
1457 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1458 * that might be send for this request. The actual number is decided
1459 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1460 * "max - 1" for old client compatibility sending "0", and also so the
1461 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1463 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1465 ioobj_max_brw_set(ioobj, 0);
1467 if (short_io_size != 0) {
1468 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1469 body->oa.o_valid |= OBD_MD_FLFLAGS;
1470 body->oa.o_flags = 0;
1472 body->oa.o_flags |= OBD_FL_SHORT_IO;
1473 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1475 if (opc == OST_WRITE) {
1476 short_io_buf = req_capsule_client_get(pill,
1478 LASSERT(short_io_buf != NULL);
1482 LASSERT(page_count > 0);
1484 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1485 struct brw_page *pg = pga[i];
1486 int poff = pg->off & ~PAGE_MASK;
1488 LASSERT(pg->count > 0);
1489 /* make sure there is no gap in the middle of page array */
1490 LASSERTF(page_count == 1 ||
1491 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1492 ergo(i > 0 && i < page_count - 1,
1493 poff == 0 && pg->count == PAGE_SIZE) &&
1494 ergo(i == page_count - 1, poff == 0)),
1495 "i: %d/%d pg: %p off: %llu, count: %u\n",
1496 i, page_count, pg, pg->off, pg->count);
1497 LASSERTF(i == 0 || pg->off > pg_prev->off,
1498 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1499 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1501 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1502 pg_prev->pg, page_private(pg_prev->pg),
1503 pg_prev->pg->index, pg_prev->off);
1504 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1505 (pg->flag & OBD_BRW_SRVLOCK));
1506 if (short_io_size != 0 && opc == OST_WRITE) {
1507 unsigned char *ptr = kmap_atomic(pg->pg);
1509 LASSERT(short_io_size >= requested_nob + pg->count);
1510 memcpy(short_io_buf + requested_nob,
1514 } else if (short_io_size == 0) {
1515 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1518 requested_nob += pg->count;
1520 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1522 niobuf->rnb_len += pg->count;
1524 niobuf->rnb_offset = pg->off;
1525 niobuf->rnb_len = pg->count;
1526 niobuf->rnb_flags = pg->flag;
1531 LASSERTF((void *)(niobuf - niocount) ==
1532 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1533 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1534 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1536 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1538 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1539 body->oa.o_valid |= OBD_MD_FLFLAGS;
1540 body->oa.o_flags = 0;
1542 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1545 if (osc_should_shrink_grant(cli))
1546 osc_shrink_grant_local(cli, &body->oa);
1548 /* size[REQ_REC_OFF] still sizeof (*body) */
1549 if (opc == OST_WRITE) {
1550 if (cli->cl_checksum &&
1551 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1552 /* store cl_cksum_type in a local variable since
1553 * it can be changed via lprocfs */
1554 enum cksum_types cksum_type = cli->cl_cksum_type;
1556 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1557 body->oa.o_flags = 0;
1559 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1561 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1563 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1564 requested_nob, page_count,
1568 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1572 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1575 /* save this in 'oa', too, for later checking */
1576 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1577 oa->o_flags |= obd_cksum_type_pack(obd_name,
1580 /* clear out the checksum flag, in case this is a
1581 * resend but cl_checksum is no longer set. b=11238 */
1582 oa->o_valid &= ~OBD_MD_FLCKSUM;
1584 oa->o_cksum = body->oa.o_cksum;
1585 /* 1 RC per niobuf */
1586 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1587 sizeof(__u32) * niocount);
1589 if (cli->cl_checksum &&
1590 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1591 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1592 body->oa.o_flags = 0;
1593 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1594 cli->cl_cksum_type);
1595 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1598 /* Client cksum has been already copied to wire obdo in previous
1599 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1600 * resent due to cksum error, this will allow Server to
1601 * check+dump pages on its side */
1603 ptlrpc_request_set_replen(req);
1605 aa = ptlrpc_req_async_args(aa, req);
1607 aa->aa_requested_nob = requested_nob;
1608 aa->aa_nio_count = niocount;
1609 aa->aa_page_count = page_count;
1613 INIT_LIST_HEAD(&aa->aa_oaps);
1616 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1617 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1618 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1619 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1623 ptlrpc_req_finished(req);
1627 char dbgcksum_file_name[PATH_MAX];
1629 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1630 struct brw_page **pga, __u32 server_cksum,
1638 /* will only keep dump of pages on first error for the same range in
1639 * file/fid, not during the resends/retries. */
1640 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1641 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1642 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1643 libcfs_debug_file_path_arr :
1644 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1645 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1646 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1647 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1649 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1650 client_cksum, server_cksum);
1651 filp = filp_open(dbgcksum_file_name,
1652 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1656 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1657 "checksum error: rc = %d\n", dbgcksum_file_name,
1660 CERROR("%s: can't open to dump pages with checksum "
1661 "error: rc = %d\n", dbgcksum_file_name, rc);
1665 for (i = 0; i < page_count; i++) {
1666 len = pga[i]->count;
1667 buf = kmap(pga[i]->pg);
1669 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1671 CERROR("%s: wanted to write %u but got %d "
1672 "error\n", dbgcksum_file_name, len, rc);
1677 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1678 dbgcksum_file_name, rc);
1683 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1685 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1686 filp_close(filp, NULL);
1690 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1691 __u32 client_cksum, __u32 server_cksum,
1692 struct osc_brw_async_args *aa)
1694 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1695 enum cksum_types cksum_type;
1696 obd_dif_csum_fn *fn = NULL;
1697 int sector_size = 0;
1702 if (server_cksum == client_cksum) {
1703 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1707 if (aa->aa_cli->cl_checksum_dump)
1708 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1709 server_cksum, client_cksum);
1711 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1714 switch (cksum_type) {
1715 case OBD_CKSUM_T10IP512:
1719 case OBD_CKSUM_T10IP4K:
1723 case OBD_CKSUM_T10CRC512:
1724 fn = obd_dif_crc_fn;
1727 case OBD_CKSUM_T10CRC4K:
1728 fn = obd_dif_crc_fn;
1736 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1737 aa->aa_page_count, aa->aa_ppga,
1738 OST_WRITE, fn, sector_size,
1741 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1742 aa->aa_ppga, OST_WRITE, cksum_type,
1746 msg = "failed to calculate the client write checksum";
1747 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1748 msg = "the server did not use the checksum type specified in "
1749 "the original request - likely a protocol problem";
1750 else if (new_cksum == server_cksum)
1751 msg = "changed on the client after we checksummed it - "
1752 "likely false positive due to mmap IO (bug 11742)";
1753 else if (new_cksum == client_cksum)
1754 msg = "changed in transit before arrival at OST";
1756 msg = "changed in transit AND doesn't match the original - "
1757 "likely false positive due to mmap IO (bug 11742)";
1759 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1760 DFID " object "DOSTID" extent [%llu-%llu], original "
1761 "client csum %x (type %x), server csum %x (type %x),"
1762 " client csum now %x\n",
1763 obd_name, msg, libcfs_nid2str(peer->nid),
1764 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1765 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1766 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1767 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1768 aa->aa_ppga[aa->aa_page_count - 1]->off +
1769 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1771 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1772 server_cksum, cksum_type, new_cksum);
1776 /* Note rc enters this function as number of bytes transferred */
1777 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1779 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1780 struct client_obd *cli = aa->aa_cli;
1781 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1782 const struct lnet_process_id *peer =
1783 &req->rq_import->imp_connection->c_peer;
1784 struct ost_body *body;
1785 u32 client_cksum = 0;
1789 if (rc < 0 && rc != -EDQUOT) {
1790 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1794 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1795 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1797 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1801 /* set/clear over quota flag for a uid/gid/projid */
1802 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1803 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1804 unsigned qid[LL_MAXQUOTAS] = {
1805 body->oa.o_uid, body->oa.o_gid,
1806 body->oa.o_projid };
1808 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1809 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1810 body->oa.o_valid, body->oa.o_flags);
1811 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1815 osc_update_grant(cli, body);
1820 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1821 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1823 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1825 CERROR("%s: unexpected positive size %d\n",
1830 if (req->rq_bulk != NULL &&
1831 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1834 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1835 check_write_checksum(&body->oa, peer, client_cksum,
1836 body->oa.o_cksum, aa))
1839 rc = check_write_rcs(req, aa->aa_requested_nob,
1840 aa->aa_nio_count, aa->aa_page_count,
1845 /* The rest of this function executes only for OST_READs */
1847 if (req->rq_bulk == NULL) {
1848 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1850 LASSERT(rc == req->rq_status);
1852 /* if unwrap_bulk failed, return -EAGAIN to retry */
1853 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1856 GOTO(out, rc = -EAGAIN);
1858 if (rc > aa->aa_requested_nob) {
1859 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
1860 rc, aa->aa_requested_nob);
1864 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1865 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
1866 rc, req->rq_bulk->bd_nob_transferred);
1870 if (req->rq_bulk == NULL) {
1872 int nob, pg_count, i = 0;
1875 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1876 pg_count = aa->aa_page_count;
1877 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1880 while (nob > 0 && pg_count > 0) {
1882 int count = aa->aa_ppga[i]->count > nob ?
1883 nob : aa->aa_ppga[i]->count;
1885 CDEBUG(D_CACHE, "page %p count %d\n",
1886 aa->aa_ppga[i]->pg, count);
1887 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
1888 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1890 kunmap_atomic((void *) ptr);
1899 if (rc < aa->aa_requested_nob)
1900 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1902 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1903 static int cksum_counter;
1904 u32 server_cksum = body->oa.o_cksum;
1907 enum cksum_types cksum_type;
1908 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1909 body->oa.o_flags : 0;
1911 cksum_type = obd_cksum_type_unpack(o_flags);
1912 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1913 aa->aa_page_count, aa->aa_ppga,
1914 OST_READ, &client_cksum);
1918 if (req->rq_bulk != NULL &&
1919 peer->nid != req->rq_bulk->bd_sender) {
1921 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1924 if (server_cksum != client_cksum) {
1925 struct ost_body *clbody;
1926 u32 page_count = aa->aa_page_count;
1928 clbody = req_capsule_client_get(&req->rq_pill,
1930 if (cli->cl_checksum_dump)
1931 dump_all_bulk_pages(&clbody->oa, page_count,
1932 aa->aa_ppga, server_cksum,
1935 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1936 "%s%s%s inode "DFID" object "DOSTID
1937 " extent [%llu-%llu], client %x, "
1938 "server %x, cksum_type %x\n",
1940 libcfs_nid2str(peer->nid),
1942 clbody->oa.o_valid & OBD_MD_FLFID ?
1943 clbody->oa.o_parent_seq : 0ULL,
1944 clbody->oa.o_valid & OBD_MD_FLFID ?
1945 clbody->oa.o_parent_oid : 0,
1946 clbody->oa.o_valid & OBD_MD_FLFID ?
1947 clbody->oa.o_parent_ver : 0,
1948 POSTID(&body->oa.o_oi),
1949 aa->aa_ppga[0]->off,
1950 aa->aa_ppga[page_count-1]->off +
1951 aa->aa_ppga[page_count-1]->count - 1,
1952 client_cksum, server_cksum,
1955 aa->aa_oa->o_cksum = client_cksum;
1959 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1962 } else if (unlikely(client_cksum)) {
1963 static int cksum_missed;
1966 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1967 CERROR("%s: checksum %u requested from %s but not sent\n",
1968 obd_name, cksum_missed,
1969 libcfs_nid2str(peer->nid));
1975 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1976 aa->aa_oa, &body->oa);
1981 static int osc_brw_redo_request(struct ptlrpc_request *request,
1982 struct osc_brw_async_args *aa, int rc)
1984 struct ptlrpc_request *new_req;
1985 struct osc_brw_async_args *new_aa;
1986 struct osc_async_page *oap;
1989 /* The below message is checked in replay-ost-single.sh test_8ae*/
1990 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1991 "redo for recoverable error %d", rc);
1993 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1994 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1995 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1996 aa->aa_ppga, &new_req, 1);
2000 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2001 if (oap->oap_request != NULL) {
2002 LASSERTF(request == oap->oap_request,
2003 "request %p != oap_request %p\n",
2004 request, oap->oap_request);
2008 * New request takes over pga and oaps from old request.
2009 * Note that copying a list_head doesn't work, need to move it...
2012 new_req->rq_interpret_reply = request->rq_interpret_reply;
2013 new_req->rq_async_args = request->rq_async_args;
2014 new_req->rq_commit_cb = request->rq_commit_cb;
2015 /* cap resend delay to the current request timeout, this is similar to
2016 * what ptlrpc does (see after_reply()) */
2017 if (aa->aa_resends > new_req->rq_timeout)
2018 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2020 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2021 new_req->rq_generation_set = 1;
2022 new_req->rq_import_generation = request->rq_import_generation;
2024 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2026 INIT_LIST_HEAD(&new_aa->aa_oaps);
2027 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2028 INIT_LIST_HEAD(&new_aa->aa_exts);
2029 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2030 new_aa->aa_resends = aa->aa_resends;
2032 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2033 if (oap->oap_request) {
2034 ptlrpc_req_finished(oap->oap_request);
2035 oap->oap_request = ptlrpc_request_addref(new_req);
2039 /* XXX: This code will run into problem if we're going to support
2040 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2041 * and wait for all of them to be finished. We should inherit request
2042 * set from old request. */
2043 ptlrpcd_add_req(new_req);
2045 DEBUG_REQ(D_INFO, new_req, "new request");
2050 * ugh, we want disk allocation on the target to happen in offset order. we'll
2051 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2052 * fine for our small page arrays and doesn't require allocation. its an
2053 * insertion sort that swaps elements that are strides apart, shrinking the
2054 * stride down until its '1' and the array is sorted.
2056 static void sort_brw_pages(struct brw_page **array, int num)
2059 struct brw_page *tmp;
2063 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2068 for (i = stride ; i < num ; i++) {
2071 while (j >= stride && array[j - stride]->off > tmp->off) {
2072 array[j] = array[j - stride];
2077 } while (stride > 1);
2080 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2082 LASSERT(ppga != NULL);
2083 OBD_FREE_PTR_ARRAY(ppga, count);
2086 static int brw_interpret(const struct lu_env *env,
2087 struct ptlrpc_request *req, void *args, int rc)
2089 struct osc_brw_async_args *aa = args;
2090 struct osc_extent *ext;
2091 struct osc_extent *tmp;
2092 struct client_obd *cli = aa->aa_cli;
2093 unsigned long transferred = 0;
2097 rc = osc_brw_fini_request(req, rc);
2098 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2100 * When server returns -EINPROGRESS, client should always retry
2101 * regardless of the number of times the bulk was resent already.
2103 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2104 if (req->rq_import_generation !=
2105 req->rq_import->imp_generation) {
2106 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2107 ""DOSTID", rc = %d.\n",
2108 req->rq_import->imp_obd->obd_name,
2109 POSTID(&aa->aa_oa->o_oi), rc);
2110 } else if (rc == -EINPROGRESS ||
2111 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2112 rc = osc_brw_redo_request(req, aa, rc);
2114 CERROR("%s: too many resent retries for object: "
2115 "%llu:%llu, rc = %d.\n",
2116 req->rq_import->imp_obd->obd_name,
2117 POSTID(&aa->aa_oa->o_oi), rc);
2122 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2127 struct obdo *oa = aa->aa_oa;
2128 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2129 unsigned long valid = 0;
2130 struct cl_object *obj;
2131 struct osc_async_page *last;
2133 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2134 obj = osc2cl(last->oap_obj);
2136 cl_object_attr_lock(obj);
2137 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2138 attr->cat_blocks = oa->o_blocks;
2139 valid |= CAT_BLOCKS;
2141 if (oa->o_valid & OBD_MD_FLMTIME) {
2142 attr->cat_mtime = oa->o_mtime;
2145 if (oa->o_valid & OBD_MD_FLATIME) {
2146 attr->cat_atime = oa->o_atime;
2149 if (oa->o_valid & OBD_MD_FLCTIME) {
2150 attr->cat_ctime = oa->o_ctime;
2154 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2155 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2156 loff_t last_off = last->oap_count + last->oap_obj_off +
2159 /* Change file size if this is an out of quota or
2160 * direct IO write and it extends the file size */
2161 if (loi->loi_lvb.lvb_size < last_off) {
2162 attr->cat_size = last_off;
2165 /* Extend KMS if it's not a lockless write */
2166 if (loi->loi_kms < last_off &&
2167 oap2osc_page(last)->ops_srvlock == 0) {
2168 attr->cat_kms = last_off;
2174 cl_object_attr_update(env, obj, attr, valid);
2175 cl_object_attr_unlock(obj);
2177 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2180 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2181 osc_inc_unstable_pages(req);
2183 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2184 list_del_init(&ext->oe_link);
2185 osc_extent_finish(env, ext, 1,
2186 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2188 LASSERT(list_empty(&aa->aa_exts));
2189 LASSERT(list_empty(&aa->aa_oaps));
2191 transferred = (req->rq_bulk == NULL ? /* short io */
2192 aa->aa_requested_nob :
2193 req->rq_bulk->bd_nob_transferred);
2195 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2196 ptlrpc_lprocfs_brw(req, transferred);
2198 spin_lock(&cli->cl_loi_list_lock);
2199 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2200 * is called so we know whether to go to sync BRWs or wait for more
2201 * RPCs to complete */
2202 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2203 cli->cl_w_in_flight--;
2205 cli->cl_r_in_flight--;
2206 osc_wake_cache_waiters(cli);
2207 spin_unlock(&cli->cl_loi_list_lock);
2209 osc_io_unplug(env, cli, NULL);
2213 static void brw_commit(struct ptlrpc_request *req)
2215 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2216 * this called via the rq_commit_cb, I need to ensure
2217 * osc_dec_unstable_pages is still called. Otherwise unstable
2218 * pages may be leaked. */
2219 spin_lock(&req->rq_lock);
2220 if (likely(req->rq_unstable)) {
2221 req->rq_unstable = 0;
2222 spin_unlock(&req->rq_lock);
2224 osc_dec_unstable_pages(req);
2226 req->rq_committed = 1;
2227 spin_unlock(&req->rq_lock);
2232 * Build an RPC by the list of extent @ext_list. The caller must ensure
2233 * that the total pages in this list are NOT over max pages per RPC.
2234 * Extents in the list must be in OES_RPC state.
2236 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2237 struct list_head *ext_list, int cmd)
2239 struct ptlrpc_request *req = NULL;
2240 struct osc_extent *ext;
2241 struct brw_page **pga = NULL;
2242 struct osc_brw_async_args *aa = NULL;
2243 struct obdo *oa = NULL;
2244 struct osc_async_page *oap;
2245 struct osc_object *obj = NULL;
2246 struct cl_req_attr *crattr = NULL;
2247 loff_t starting_offset = OBD_OBJECT_EOF;
2248 loff_t ending_offset = 0;
2249 /* '1' for consistency with code that checks !mpflag to restore */
2253 bool soft_sync = false;
2254 bool ndelay = false;
2258 __u32 layout_version = 0;
2259 LIST_HEAD(rpc_list);
2260 struct ost_body *body;
2262 LASSERT(!list_empty(ext_list));
2264 /* add pages into rpc_list to build BRW rpc */
2265 list_for_each_entry(ext, ext_list, oe_link) {
2266 LASSERT(ext->oe_state == OES_RPC);
2267 mem_tight |= ext->oe_memalloc;
2268 grant += ext->oe_grants;
2269 page_count += ext->oe_nr_pages;
2270 layout_version = max(layout_version, ext->oe_layout_version);
2275 soft_sync = osc_over_unstable_soft_limit(cli);
2277 mpflag = memalloc_noreclaim_save();
2279 OBD_ALLOC_PTR_ARRAY(pga, page_count);
2281 GOTO(out, rc = -ENOMEM);
2283 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2285 GOTO(out, rc = -ENOMEM);
2288 list_for_each_entry(ext, ext_list, oe_link) {
2289 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2291 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2293 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2294 pga[i] = &oap->oap_brw_page;
2295 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2298 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2299 if (starting_offset == OBD_OBJECT_EOF ||
2300 starting_offset > oap->oap_obj_off)
2301 starting_offset = oap->oap_obj_off;
2303 LASSERT(oap->oap_page_off == 0);
2304 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2305 ending_offset = oap->oap_obj_off +
2308 LASSERT(oap->oap_page_off + oap->oap_count ==
2315 /* first page in the list */
2316 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2318 crattr = &osc_env_info(env)->oti_req_attr;
2319 memset(crattr, 0, sizeof(*crattr));
2320 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2321 crattr->cra_flags = ~0ULL;
2322 crattr->cra_page = oap2cl_page(oap);
2323 crattr->cra_oa = oa;
2324 cl_req_attr_set(env, osc2cl(obj), crattr);
2326 if (cmd == OBD_BRW_WRITE) {
2327 oa->o_grant_used = grant;
2328 if (layout_version > 0) {
2329 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2330 PFID(&oa->o_oi.oi_fid), layout_version);
2332 oa->o_layout_version = layout_version;
2333 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2337 sort_brw_pages(pga, page_count);
2338 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2340 CERROR("prep_req failed: %d\n", rc);
2344 req->rq_commit_cb = brw_commit;
2345 req->rq_interpret_reply = brw_interpret;
2346 req->rq_memalloc = mem_tight != 0;
2347 oap->oap_request = ptlrpc_request_addref(req);
2349 req->rq_no_resend = req->rq_no_delay = 1;
2350 /* probably set a shorter timeout value.
2351 * to handle ETIMEDOUT in brw_interpret() correctly. */
2352 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2355 /* Need to update the timestamps after the request is built in case
2356 * we race with setattr (locally or in queue at OST). If OST gets
2357 * later setattr before earlier BRW (as determined by the request xid),
2358 * the OST will not use BRW timestamps. Sadly, there is no obvious
2359 * way to do this in a single call. bug 10150 */
2360 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2361 crattr->cra_oa = &body->oa;
2362 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2363 cl_req_attr_set(env, osc2cl(obj), crattr);
2364 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2366 aa = ptlrpc_req_async_args(aa, req);
2367 INIT_LIST_HEAD(&aa->aa_oaps);
2368 list_splice_init(&rpc_list, &aa->aa_oaps);
2369 INIT_LIST_HEAD(&aa->aa_exts);
2370 list_splice_init(ext_list, &aa->aa_exts);
2372 spin_lock(&cli->cl_loi_list_lock);
2373 starting_offset >>= PAGE_SHIFT;
2374 if (cmd == OBD_BRW_READ) {
2375 cli->cl_r_in_flight++;
2376 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2377 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2378 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2379 starting_offset + 1);
2381 cli->cl_w_in_flight++;
2382 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2383 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2384 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2385 starting_offset + 1);
2387 spin_unlock(&cli->cl_loi_list_lock);
2389 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2390 page_count, aa, cli->cl_r_in_flight,
2391 cli->cl_w_in_flight);
2392 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2394 ptlrpcd_add_req(req);
2400 memalloc_noreclaim_restore(mpflag);
2403 LASSERT(req == NULL);
2406 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2408 OBD_FREE_PTR_ARRAY(pga, page_count);
2409 /* this should happen rarely and is pretty bad, it makes the
2410 * pending list not follow the dirty order */
2411 while (!list_empty(ext_list)) {
2412 ext = list_entry(ext_list->next, struct osc_extent,
2414 list_del_init(&ext->oe_link);
2415 osc_extent_finish(env, ext, 0, rc);
2421 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2425 LASSERT(lock != NULL);
2427 lock_res_and_lock(lock);
2429 if (lock->l_ast_data == NULL)
2430 lock->l_ast_data = data;
2431 if (lock->l_ast_data == data)
2434 unlock_res_and_lock(lock);
2439 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2440 void *cookie, struct lustre_handle *lockh,
2441 enum ldlm_mode mode, __u64 *flags, bool speculative,
2444 bool intent = *flags & LDLM_FL_HAS_INTENT;
2448 /* The request was created before ldlm_cli_enqueue call. */
2449 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2450 struct ldlm_reply *rep;
2452 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2453 LASSERT(rep != NULL);
2455 rep->lock_policy_res1 =
2456 ptlrpc_status_ntoh(rep->lock_policy_res1);
2457 if (rep->lock_policy_res1)
2458 errcode = rep->lock_policy_res1;
2460 *flags |= LDLM_FL_LVB_READY;
2461 } else if (errcode == ELDLM_OK) {
2462 *flags |= LDLM_FL_LVB_READY;
2465 /* Call the update callback. */
2466 rc = (*upcall)(cookie, lockh, errcode);
2468 /* release the reference taken in ldlm_cli_enqueue() */
2469 if (errcode == ELDLM_LOCK_MATCHED)
2471 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2472 ldlm_lock_decref(lockh, mode);
2477 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2480 struct osc_enqueue_args *aa = args;
2481 struct ldlm_lock *lock;
2482 struct lustre_handle *lockh = &aa->oa_lockh;
2483 enum ldlm_mode mode = aa->oa_mode;
2484 struct ost_lvb *lvb = aa->oa_lvb;
2485 __u32 lvb_len = sizeof(*lvb);
2490 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2492 lock = ldlm_handle2lock(lockh);
2493 LASSERTF(lock != NULL,
2494 "lockh %#llx, req %p, aa %p - client evicted?\n",
2495 lockh->cookie, req, aa);
2497 /* Take an additional reference so that a blocking AST that
2498 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2499 * to arrive after an upcall has been executed by
2500 * osc_enqueue_fini(). */
2501 ldlm_lock_addref(lockh, mode);
2503 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2504 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2506 /* Let CP AST to grant the lock first. */
2507 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2509 if (aa->oa_speculative) {
2510 LASSERT(aa->oa_lvb == NULL);
2511 LASSERT(aa->oa_flags == NULL);
2512 aa->oa_flags = &flags;
2515 /* Complete obtaining the lock procedure. */
2516 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2517 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2519 /* Complete osc stuff. */
2520 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2521 aa->oa_flags, aa->oa_speculative, rc);
2523 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2525 ldlm_lock_decref(lockh, mode);
2526 LDLM_LOCK_PUT(lock);
2530 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2531 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2532 * other synchronous requests, however keeping some locks and trying to obtain
2533 * others may take a considerable amount of time in a case of ost failure; and
2534 * when other sync requests do not get released lock from a client, the client
2535 * is evicted from the cluster -- such scenarious make the life difficult, so
2536 * release locks just after they are obtained. */
2537 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2538 __u64 *flags, union ldlm_policy_data *policy,
2539 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2540 void *cookie, struct ldlm_enqueue_info *einfo,
2541 struct ptlrpc_request_set *rqset, int async,
2544 struct obd_device *obd = exp->exp_obd;
2545 struct lustre_handle lockh = { 0 };
2546 struct ptlrpc_request *req = NULL;
2547 int intent = *flags & LDLM_FL_HAS_INTENT;
2548 __u64 match_flags = *flags;
2549 enum ldlm_mode mode;
2553 /* Filesystem lock extents are extended to page boundaries so that
2554 * dealing with the page cache is a little smoother. */
2555 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2556 policy->l_extent.end |= ~PAGE_MASK;
2558 /* Next, search for already existing extent locks that will cover us */
2559 /* If we're trying to read, we also search for an existing PW lock. The
2560 * VFS and page cache already protect us locally, so lots of readers/
2561 * writers can share a single PW lock.
2563 * There are problems with conversion deadlocks, so instead of
2564 * converting a read lock to a write lock, we'll just enqueue a new
2567 * At some point we should cancel the read lock instead of making them
2568 * send us a blocking callback, but there are problems with canceling
2569 * locks out from other users right now, too. */
2570 mode = einfo->ei_mode;
2571 if (einfo->ei_mode == LCK_PR)
2573 /* Normal lock requests must wait for the LVB to be ready before
2574 * matching a lock; speculative lock requests do not need to,
2575 * because they will not actually use the lock. */
2577 match_flags |= LDLM_FL_LVB_READY;
2579 match_flags |= LDLM_FL_BLOCK_GRANTED;
2580 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2581 einfo->ei_type, policy, mode, &lockh, 0);
2583 struct ldlm_lock *matched;
2585 if (*flags & LDLM_FL_TEST_LOCK)
2588 matched = ldlm_handle2lock(&lockh);
2590 /* This DLM lock request is speculative, and does not
2591 * have an associated IO request. Therefore if there
2592 * is already a DLM lock, it wll just inform the
2593 * caller to cancel the request for this stripe.*/
2594 lock_res_and_lock(matched);
2595 if (ldlm_extent_equal(&policy->l_extent,
2596 &matched->l_policy_data.l_extent))
2600 unlock_res_and_lock(matched);
2602 ldlm_lock_decref(&lockh, mode);
2603 LDLM_LOCK_PUT(matched);
2605 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2606 *flags |= LDLM_FL_LVB_READY;
2608 /* We already have a lock, and it's referenced. */
2609 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2611 ldlm_lock_decref(&lockh, mode);
2612 LDLM_LOCK_PUT(matched);
2615 ldlm_lock_decref(&lockh, mode);
2616 LDLM_LOCK_PUT(matched);
2620 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2624 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2625 &RQF_LDLM_ENQUEUE_LVB);
2629 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2631 ptlrpc_request_free(req);
2635 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2637 ptlrpc_request_set_replen(req);
2640 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2641 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2643 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2644 sizeof(*lvb), LVB_T_OST, &lockh, async);
2647 struct osc_enqueue_args *aa;
2648 aa = ptlrpc_req_async_args(aa, req);
2650 aa->oa_mode = einfo->ei_mode;
2651 aa->oa_type = einfo->ei_type;
2652 lustre_handle_copy(&aa->oa_lockh, &lockh);
2653 aa->oa_upcall = upcall;
2654 aa->oa_cookie = cookie;
2655 aa->oa_speculative = speculative;
2657 aa->oa_flags = flags;
2660 /* speculative locks are essentially to enqueue
2661 * a DLM lock in advance, so we don't care
2662 * about the result of the enqueue. */
2664 aa->oa_flags = NULL;
2667 req->rq_interpret_reply = osc_enqueue_interpret;
2668 ptlrpc_set_add_req(rqset, req);
2669 } else if (intent) {
2670 ptlrpc_req_finished(req);
2675 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2676 flags, speculative, rc);
2678 ptlrpc_req_finished(req);
2683 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2684 struct ldlm_res_id *res_id, enum ldlm_type type,
2685 union ldlm_policy_data *policy, enum ldlm_mode mode,
2686 __u64 *flags, struct osc_object *obj,
2687 struct lustre_handle *lockh, int unref)
2689 struct obd_device *obd = exp->exp_obd;
2690 __u64 lflags = *flags;
2694 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2697 /* Filesystem lock extents are extended to page boundaries so that
2698 * dealing with the page cache is a little smoother */
2699 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2700 policy->l_extent.end |= ~PAGE_MASK;
2702 /* Next, search for already existing extent locks that will cover us */
2703 /* If we're trying to read, we also search for an existing PW lock. The
2704 * VFS and page cache already protect us locally, so lots of readers/
2705 * writers can share a single PW lock. */
2709 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2710 res_id, type, policy, rc, lockh, unref);
2711 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2715 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2717 LASSERT(lock != NULL);
2718 if (osc_set_lock_data(lock, obj)) {
2719 lock_res_and_lock(lock);
2720 if (!ldlm_is_lvb_cached(lock)) {
2721 LASSERT(lock->l_ast_data == obj);
2722 osc_lock_lvb_update(env, obj, lock, NULL);
2723 ldlm_set_lvb_cached(lock);
2725 unlock_res_and_lock(lock);
2727 ldlm_lock_decref(lockh, rc);
2730 LDLM_LOCK_PUT(lock);
2735 static int osc_statfs_interpret(const struct lu_env *env,
2736 struct ptlrpc_request *req, void *args, int rc)
2738 struct osc_async_args *aa = args;
2739 struct obd_statfs *msfs;
2744 * The request has in fact never been sent due to issues at
2745 * a higher level (LOV). Exit immediately since the caller
2746 * is aware of the problem and takes care of the clean up.
2750 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2751 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2757 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2759 GOTO(out, rc = -EPROTO);
2761 *aa->aa_oi->oi_osfs = *msfs;
2763 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2768 static int osc_statfs_async(struct obd_export *exp,
2769 struct obd_info *oinfo, time64_t max_age,
2770 struct ptlrpc_request_set *rqset)
2772 struct obd_device *obd = class_exp2obd(exp);
2773 struct ptlrpc_request *req;
2774 struct osc_async_args *aa;
2778 if (obd->obd_osfs_age >= max_age) {
2780 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2781 obd->obd_name, &obd->obd_osfs,
2782 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2783 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2784 spin_lock(&obd->obd_osfs_lock);
2785 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2786 spin_unlock(&obd->obd_osfs_lock);
2787 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2788 if (oinfo->oi_cb_up)
2789 oinfo->oi_cb_up(oinfo, 0);
2794 /* We could possibly pass max_age in the request (as an absolute
2795 * timestamp or a "seconds.usec ago") so the target can avoid doing
2796 * extra calls into the filesystem if that isn't necessary (e.g.
2797 * during mount that would help a bit). Having relative timestamps
2798 * is not so great if request processing is slow, while absolute
2799 * timestamps are not ideal because they need time synchronization. */
2800 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2804 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2806 ptlrpc_request_free(req);
2809 ptlrpc_request_set_replen(req);
2810 req->rq_request_portal = OST_CREATE_PORTAL;
2811 ptlrpc_at_set_req_timeout(req);
2813 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2814 /* procfs requests not want stat in wait for avoid deadlock */
2815 req->rq_no_resend = 1;
2816 req->rq_no_delay = 1;
2819 req->rq_interpret_reply = osc_statfs_interpret;
2820 aa = ptlrpc_req_async_args(aa, req);
2823 ptlrpc_set_add_req(rqset, req);
2827 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2828 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2830 struct obd_device *obd = class_exp2obd(exp);
2831 struct obd_statfs *msfs;
2832 struct ptlrpc_request *req;
2833 struct obd_import *imp = NULL;
2838 /*Since the request might also come from lprocfs, so we need
2839 *sync this with client_disconnect_export Bug15684*/
2840 down_read(&obd->u.cli.cl_sem);
2841 if (obd->u.cli.cl_import)
2842 imp = class_import_get(obd->u.cli.cl_import);
2843 up_read(&obd->u.cli.cl_sem);
2847 /* We could possibly pass max_age in the request (as an absolute
2848 * timestamp or a "seconds.usec ago") so the target can avoid doing
2849 * extra calls into the filesystem if that isn't necessary (e.g.
2850 * during mount that would help a bit). Having relative timestamps
2851 * is not so great if request processing is slow, while absolute
2852 * timestamps are not ideal because they need time synchronization. */
2853 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2855 class_import_put(imp);
2860 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2862 ptlrpc_request_free(req);
2865 ptlrpc_request_set_replen(req);
2866 req->rq_request_portal = OST_CREATE_PORTAL;
2867 ptlrpc_at_set_req_timeout(req);
2869 if (flags & OBD_STATFS_NODELAY) {
2870 /* procfs requests not want stat in wait for avoid deadlock */
2871 req->rq_no_resend = 1;
2872 req->rq_no_delay = 1;
2875 rc = ptlrpc_queue_wait(req);
2879 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2881 GOTO(out, rc = -EPROTO);
2887 ptlrpc_req_finished(req);
2891 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2892 void *karg, void __user *uarg)
2894 struct obd_device *obd = exp->exp_obd;
2895 struct obd_ioctl_data *data = karg;
2899 if (!try_module_get(THIS_MODULE)) {
2900 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2901 module_name(THIS_MODULE));
2905 case OBD_IOC_CLIENT_RECOVER:
2906 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
2907 data->ioc_inlbuf1, 0);
2911 case IOC_OSC_SET_ACTIVE:
2912 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
2917 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
2918 obd->obd_name, cmd, current->comm, rc);
2922 module_put(THIS_MODULE);
2926 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2927 u32 keylen, void *key, u32 vallen, void *val,
2928 struct ptlrpc_request_set *set)
2930 struct ptlrpc_request *req;
2931 struct obd_device *obd = exp->exp_obd;
2932 struct obd_import *imp = class_exp2cliimp(exp);
2937 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2939 if (KEY_IS(KEY_CHECKSUM)) {
2940 if (vallen != sizeof(int))
2942 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2946 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2947 sptlrpc_conf_client_adapt(obd);
2951 if (KEY_IS(KEY_FLUSH_CTX)) {
2952 sptlrpc_import_flush_my_ctx(imp);
2956 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2957 struct client_obd *cli = &obd->u.cli;
2958 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2959 long target = *(long *)val;
2961 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2966 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2969 /* We pass all other commands directly to OST. Since nobody calls osc
2970 methods directly and everybody is supposed to go through LOV, we
2971 assume lov checked invalid values for us.
2972 The only recognised values so far are evict_by_nid and mds_conn.
2973 Even if something bad goes through, we'd get a -EINVAL from OST
2976 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2977 &RQF_OST_SET_GRANT_INFO :
2982 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2983 RCL_CLIENT, keylen);
2984 if (!KEY_IS(KEY_GRANT_SHRINK))
2985 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2986 RCL_CLIENT, vallen);
2987 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2989 ptlrpc_request_free(req);
2993 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2994 memcpy(tmp, key, keylen);
2995 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2998 memcpy(tmp, val, vallen);
3000 if (KEY_IS(KEY_GRANT_SHRINK)) {
3001 struct osc_grant_args *aa;
3004 aa = ptlrpc_req_async_args(aa, req);
3005 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3007 ptlrpc_req_finished(req);
3010 *oa = ((struct ost_body *)val)->oa;
3012 req->rq_interpret_reply = osc_shrink_grant_interpret;
3015 ptlrpc_request_set_replen(req);
3016 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3017 LASSERT(set != NULL);
3018 ptlrpc_set_add_req(set, req);
3019 ptlrpc_check_set(NULL, set);
3021 ptlrpcd_add_req(req);
3026 EXPORT_SYMBOL(osc_set_info_async);
3028 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3029 struct obd_device *obd, struct obd_uuid *cluuid,
3030 struct obd_connect_data *data, void *localdata)
3032 struct client_obd *cli = &obd->u.cli;
3034 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3038 spin_lock(&cli->cl_loi_list_lock);
3039 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3040 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3041 /* restore ocd_grant_blkbits as client page bits */
3042 data->ocd_grant_blkbits = PAGE_SHIFT;
3043 grant += cli->cl_dirty_grant;
3045 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3047 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3048 lost_grant = cli->cl_lost_grant;
3049 cli->cl_lost_grant = 0;
3050 spin_unlock(&cli->cl_loi_list_lock);
3052 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3053 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3054 data->ocd_version, data->ocd_grant, lost_grant);
3059 EXPORT_SYMBOL(osc_reconnect);
3061 int osc_disconnect(struct obd_export *exp)
3063 struct obd_device *obd = class_exp2obd(exp);
3066 rc = client_disconnect_export(exp);
3068 * Initially we put del_shrink_grant before disconnect_export, but it
3069 * causes the following problem if setup (connect) and cleanup
3070 * (disconnect) are tangled together.
3071 * connect p1 disconnect p2
3072 * ptlrpc_connect_import
3073 * ............... class_manual_cleanup
3076 * ptlrpc_connect_interrupt
3078 * add this client to shrink list
3080 * Bang! grant shrink thread trigger the shrink. BUG18662
3082 osc_del_grant_list(&obd->u.cli);
3085 EXPORT_SYMBOL(osc_disconnect);
3087 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3088 struct hlist_node *hnode, void *arg)
3090 struct lu_env *env = arg;
3091 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3092 struct ldlm_lock *lock;
3093 struct osc_object *osc = NULL;
3097 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3098 if (lock->l_ast_data != NULL && osc == NULL) {
3099 osc = lock->l_ast_data;
3100 cl_object_get(osc2cl(osc));
3103 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3104 * by the 2nd round of ldlm_namespace_clean() call in
3105 * osc_import_event(). */
3106 ldlm_clear_cleaned(lock);
3111 osc_object_invalidate(env, osc);
3112 cl_object_put(env, osc2cl(osc));
3117 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3119 static int osc_import_event(struct obd_device *obd,
3120 struct obd_import *imp,
3121 enum obd_import_event event)
3123 struct client_obd *cli;
3127 LASSERT(imp->imp_obd == obd);
3130 case IMP_EVENT_DISCON: {
3132 spin_lock(&cli->cl_loi_list_lock);
3133 cli->cl_avail_grant = 0;
3134 cli->cl_lost_grant = 0;
3135 spin_unlock(&cli->cl_loi_list_lock);
3138 case IMP_EVENT_INACTIVE: {
3139 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3142 case IMP_EVENT_INVALIDATE: {
3143 struct ldlm_namespace *ns = obd->obd_namespace;
3147 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3149 env = cl_env_get(&refcheck);
3151 osc_io_unplug(env, &obd->u.cli, NULL);
3153 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3154 osc_ldlm_resource_invalidate,
3156 cl_env_put(env, &refcheck);
3158 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3163 case IMP_EVENT_ACTIVE: {
3164 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3167 case IMP_EVENT_OCD: {
3168 struct obd_connect_data *ocd = &imp->imp_connect_data;
3170 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3171 osc_init_grant(&obd->u.cli, ocd);
3174 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3175 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3177 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3180 case IMP_EVENT_DEACTIVATE: {
3181 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3184 case IMP_EVENT_ACTIVATE: {
3185 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3189 CERROR("Unknown import event %d\n", event);
3196 * Determine whether the lock can be canceled before replaying the lock
3197 * during recovery, see bug16774 for detailed information.
3199 * \retval zero the lock can't be canceled
3200 * \retval other ok to cancel
3202 static int osc_cancel_weight(struct ldlm_lock *lock)
3205 * Cancel all unused and granted extent lock.
3207 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3208 ldlm_is_granted(lock) &&
3209 osc_ldlm_weigh_ast(lock) == 0)
3215 static int brw_queue_work(const struct lu_env *env, void *data)
3217 struct client_obd *cli = data;
3219 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3221 osc_io_unplug(env, cli, NULL);
3225 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3227 struct client_obd *cli = &obd->u.cli;
3233 rc = ptlrpcd_addref();
3237 rc = client_obd_setup(obd, lcfg);
3239 GOTO(out_ptlrpcd, rc);
3242 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3243 if (IS_ERR(handler))
3244 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3245 cli->cl_writeback_work = handler;
3247 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3248 if (IS_ERR(handler))
3249 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3250 cli->cl_lru_work = handler;
3252 rc = osc_quota_setup(obd);
3254 GOTO(out_ptlrpcd_work, rc);
3256 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3257 osc_update_next_shrink(cli);
3262 if (cli->cl_writeback_work != NULL) {
3263 ptlrpcd_destroy_work(cli->cl_writeback_work);
3264 cli->cl_writeback_work = NULL;
3266 if (cli->cl_lru_work != NULL) {
3267 ptlrpcd_destroy_work(cli->cl_lru_work);
3268 cli->cl_lru_work = NULL;
3270 client_obd_cleanup(obd);
3275 EXPORT_SYMBOL(osc_setup_common);
3277 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3279 struct client_obd *cli = &obd->u.cli;
3287 rc = osc_setup_common(obd, lcfg);
3291 rc = osc_tunables_init(obd);
3296 * We try to control the total number of requests with a upper limit
3297 * osc_reqpool_maxreqcount. There might be some race which will cause
3298 * over-limit allocation, but it is fine.
3300 req_count = atomic_read(&osc_pool_req_count);
3301 if (req_count < osc_reqpool_maxreqcount) {
3302 adding = cli->cl_max_rpcs_in_flight + 2;
3303 if (req_count + adding > osc_reqpool_maxreqcount)
3304 adding = osc_reqpool_maxreqcount - req_count;
3306 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3307 atomic_add(added, &osc_pool_req_count);
3310 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3312 spin_lock(&osc_shrink_lock);
3313 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3314 spin_unlock(&osc_shrink_lock);
3315 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3316 cli->cl_import->imp_idle_debug = D_HA;
3321 int osc_precleanup_common(struct obd_device *obd)
3323 struct client_obd *cli = &obd->u.cli;
3327 * for echo client, export may be on zombie list, wait for
3328 * zombie thread to cull it, because cli.cl_import will be
3329 * cleared in client_disconnect_export():
3330 * class_export_destroy() -> obd_cleanup() ->
3331 * echo_device_free() -> echo_client_cleanup() ->
3332 * obd_disconnect() -> osc_disconnect() ->
3333 * client_disconnect_export()
3335 obd_zombie_barrier();
3336 if (cli->cl_writeback_work) {
3337 ptlrpcd_destroy_work(cli->cl_writeback_work);
3338 cli->cl_writeback_work = NULL;
3341 if (cli->cl_lru_work) {
3342 ptlrpcd_destroy_work(cli->cl_lru_work);
3343 cli->cl_lru_work = NULL;
3346 obd_cleanup_client_import(obd);
3349 EXPORT_SYMBOL(osc_precleanup_common);
3351 static int osc_precleanup(struct obd_device *obd)
3355 osc_precleanup_common(obd);
3357 ptlrpc_lprocfs_unregister_obd(obd);
3361 int osc_cleanup_common(struct obd_device *obd)
3363 struct client_obd *cli = &obd->u.cli;
3368 spin_lock(&osc_shrink_lock);
3369 list_del(&cli->cl_shrink_list);
3370 spin_unlock(&osc_shrink_lock);
3373 if (cli->cl_cache != NULL) {
3374 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3375 spin_lock(&cli->cl_cache->ccc_lru_lock);
3376 list_del_init(&cli->cl_lru_osc);
3377 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3378 cli->cl_lru_left = NULL;
3379 cl_cache_decref(cli->cl_cache);
3380 cli->cl_cache = NULL;
3383 /* free memory of osc quota cache */
3384 osc_quota_cleanup(obd);
3386 rc = client_obd_cleanup(obd);
3391 EXPORT_SYMBOL(osc_cleanup_common);
3393 static const struct obd_ops osc_obd_ops = {
3394 .o_owner = THIS_MODULE,
3395 .o_setup = osc_setup,
3396 .o_precleanup = osc_precleanup,
3397 .o_cleanup = osc_cleanup_common,
3398 .o_add_conn = client_import_add_conn,
3399 .o_del_conn = client_import_del_conn,
3400 .o_connect = client_connect_import,
3401 .o_reconnect = osc_reconnect,
3402 .o_disconnect = osc_disconnect,
3403 .o_statfs = osc_statfs,
3404 .o_statfs_async = osc_statfs_async,
3405 .o_create = osc_create,
3406 .o_destroy = osc_destroy,
3407 .o_getattr = osc_getattr,
3408 .o_setattr = osc_setattr,
3409 .o_iocontrol = osc_iocontrol,
3410 .o_set_info_async = osc_set_info_async,
3411 .o_import_event = osc_import_event,
3412 .o_quotactl = osc_quotactl,
3415 static struct shrinker *osc_cache_shrinker;
3416 LIST_HEAD(osc_shrink_list);
3417 DEFINE_SPINLOCK(osc_shrink_lock);
3419 #ifndef HAVE_SHRINKER_COUNT
3420 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3422 struct shrink_control scv = {
3423 .nr_to_scan = shrink_param(sc, nr_to_scan),
3424 .gfp_mask = shrink_param(sc, gfp_mask)
3426 (void)osc_cache_shrink_scan(shrinker, &scv);
3428 return osc_cache_shrink_count(shrinker, &scv);
3432 static int __init osc_init(void)
3434 unsigned int reqpool_size;
3435 unsigned int reqsize;
3437 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3438 osc_cache_shrink_count, osc_cache_shrink_scan);
3441 /* print an address of _any_ initialized kernel symbol from this
3442 * module, to allow debugging with gdb that doesn't support data
3443 * symbols from modules.*/
3444 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3446 rc = lu_kmem_init(osc_caches);
3450 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3451 LUSTRE_OSC_NAME, &osc_device_type);
3455 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3457 /* This is obviously too much memory, only prevent overflow here */
3458 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3459 GOTO(out_type, rc = -EINVAL);
3461 reqpool_size = osc_reqpool_mem_max << 20;
3464 while (reqsize < OST_IO_MAXREQSIZE)
3465 reqsize = reqsize << 1;
3468 * We don't enlarge the request count in OSC pool according to
3469 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3470 * tried after normal allocation failed. So a small OSC pool won't
3471 * cause much performance degression in most of cases.
3473 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3475 atomic_set(&osc_pool_req_count, 0);
3476 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3477 ptlrpc_add_rqs_to_pool);
3479 if (osc_rq_pool == NULL)
3480 GOTO(out_type, rc = -ENOMEM);
3482 rc = osc_start_grant_work();
3484 GOTO(out_req_pool, rc);
3489 ptlrpc_free_rq_pool(osc_rq_pool);
3491 class_unregister_type(LUSTRE_OSC_NAME);
3493 lu_kmem_fini(osc_caches);
3498 static void __exit osc_exit(void)
3500 osc_stop_grant_work();
3501 remove_shrinker(osc_cache_shrinker);
3502 class_unregister_type(LUSTRE_OSC_NAME);
3503 lu_kmem_fini(osc_caches);
3504 ptlrpc_free_rq_pool(osc_rq_pool);
3507 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3508 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3509 MODULE_VERSION(LUSTRE_VERSION_STRING);
3510 MODULE_LICENSE("GPL");
3512 module_init(osc_init);
3513 module_exit(osc_exit);