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,
1350 static inline void osc_release_bounce_pages(struct brw_page **pga,
1353 #ifdef HAVE_LUSTRE_CRYPTO
1356 for (i = 0; i < page_count; i++) {
1357 if (pga[i]->pg->mapping)
1358 /* bounce pages are unmapped */
1360 if (pga[i]->flag & OBD_BRW_SYNC)
1361 /* sync transfer cannot have encrypted pages */
1363 llcrypt_finalize_bounce_page(&pga[i]->pg);
1364 pga[i]->count -= pga[i]->bp_count_diff;
1365 pga[i]->off += pga[i]->bp_off_diff;
1371 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1372 u32 page_count, struct brw_page **pga,
1373 struct ptlrpc_request **reqp, int resend)
1375 struct ptlrpc_request *req;
1376 struct ptlrpc_bulk_desc *desc;
1377 struct ost_body *body;
1378 struct obd_ioobj *ioobj;
1379 struct niobuf_remote *niobuf;
1380 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1381 struct osc_brw_async_args *aa;
1382 struct req_capsule *pill;
1383 struct brw_page *pg_prev;
1385 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1386 struct inode *inode;
1389 inode = page2inode(pga[0]->pg);
1390 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1391 RETURN(-ENOMEM); /* Recoverable */
1392 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1393 RETURN(-EINVAL); /* Fatal */
1395 if ((cmd & OBD_BRW_WRITE) != 0) {
1397 req = ptlrpc_request_alloc_pool(cli->cl_import,
1399 &RQF_OST_BRW_WRITE);
1402 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1407 if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1408 for (i = 0; i < page_count; i++) {
1409 struct brw_page *pg = pga[i];
1410 struct page *data_page = NULL;
1411 bool retried = false;
1412 bool lockedbymyself;
1415 /* The page can already be locked when we arrive here.
1416 * This is possible when cl_page_assume/vvp_page_assume
1417 * is stuck on wait_on_page_writeback with page lock
1418 * held. In this case there is no risk for the lock to
1419 * be released while we are doing our encryption
1420 * processing, because writeback against that page will
1421 * end in vvp_page_completion_write/cl_page_completion,
1422 * which means only once the page is fully processed.
1424 lockedbymyself = trylock_page(pg->pg);
1426 llcrypt_encrypt_pagecache_blocks(pg->pg,
1430 unlock_page(pg->pg);
1431 if (IS_ERR(data_page)) {
1432 rc = PTR_ERR(data_page);
1433 if (rc == -ENOMEM && !retried) {
1438 ptlrpc_request_free(req);
1442 /* there should be no gap in the middle of page array */
1443 if (i == page_count - 1) {
1444 struct osc_async_page *oap = brw_page2oap(pg);
1446 oa->o_size = oap->oap_count +
1447 oap->oap_obj_off + oap->oap_page_off;
1449 /* len is forced to PAGE_SIZE, and poff to 0
1450 * so store the old, clear text info
1452 pg->bp_count_diff = PAGE_SIZE - pg->count;
1453 pg->count = PAGE_SIZE;
1454 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1455 pg->off = pg->off & PAGE_MASK;
1459 for (niocount = i = 1; i < page_count; i++) {
1460 if (!can_merge_pages(pga[i - 1], pga[i]))
1464 pill = &req->rq_pill;
1465 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1467 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1468 niocount * sizeof(*niobuf));
1470 for (i = 0; i < page_count; i++)
1471 short_io_size += pga[i]->count;
1473 /* Check if read/write is small enough to be a short io. */
1474 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1475 !imp_connect_shortio(cli->cl_import))
1478 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1479 opc == OST_READ ? 0 : short_io_size);
1480 if (opc == OST_READ)
1481 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1484 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1486 ptlrpc_request_free(req);
1489 osc_set_io_portal(req);
1491 ptlrpc_at_set_req_timeout(req);
1492 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1494 req->rq_no_retry_einprogress = 1;
1496 if (short_io_size != 0) {
1498 short_io_buf = NULL;
1502 desc = ptlrpc_prep_bulk_imp(req, page_count,
1503 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1504 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1505 PTLRPC_BULK_PUT_SINK),
1507 &ptlrpc_bulk_kiov_pin_ops);
1510 GOTO(out, rc = -ENOMEM);
1511 /* NB request now owns desc and will free it when it gets freed */
1513 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1514 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1515 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1516 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1518 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1520 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1521 * and from_kgid(), because they are asynchronous. Fortunately, variable
1522 * oa contains valid o_uid and o_gid in these two operations.
1523 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1524 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1525 * other process logic */
1526 body->oa.o_uid = oa->o_uid;
1527 body->oa.o_gid = oa->o_gid;
1529 obdo_to_ioobj(oa, ioobj);
1530 ioobj->ioo_bufcnt = niocount;
1531 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1532 * that might be send for this request. The actual number is decided
1533 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1534 * "max - 1" for old client compatibility sending "0", and also so the
1535 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1537 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1539 ioobj_max_brw_set(ioobj, 0);
1541 if (short_io_size != 0) {
1542 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1543 body->oa.o_valid |= OBD_MD_FLFLAGS;
1544 body->oa.o_flags = 0;
1546 body->oa.o_flags |= OBD_FL_SHORT_IO;
1547 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1549 if (opc == OST_WRITE) {
1550 short_io_buf = req_capsule_client_get(pill,
1552 LASSERT(short_io_buf != NULL);
1556 LASSERT(page_count > 0);
1558 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1559 struct brw_page *pg = pga[i];
1560 int poff = pg->off & ~PAGE_MASK;
1562 LASSERT(pg->count > 0);
1563 /* make sure there is no gap in the middle of page array */
1564 LASSERTF(page_count == 1 ||
1565 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1566 ergo(i > 0 && i < page_count - 1,
1567 poff == 0 && pg->count == PAGE_SIZE) &&
1568 ergo(i == page_count - 1, poff == 0)),
1569 "i: %d/%d pg: %p off: %llu, count: %u\n",
1570 i, page_count, pg, pg->off, pg->count);
1571 LASSERTF(i == 0 || pg->off > pg_prev->off,
1572 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1573 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1575 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1576 pg_prev->pg, page_private(pg_prev->pg),
1577 pg_prev->pg->index, pg_prev->off);
1578 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1579 (pg->flag & OBD_BRW_SRVLOCK));
1580 if (short_io_size != 0 && opc == OST_WRITE) {
1581 unsigned char *ptr = kmap_atomic(pg->pg);
1583 LASSERT(short_io_size >= requested_nob + pg->count);
1584 memcpy(short_io_buf + requested_nob,
1588 } else if (short_io_size == 0) {
1589 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1592 requested_nob += pg->count;
1594 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1596 niobuf->rnb_len += pg->count;
1598 niobuf->rnb_offset = pg->off;
1599 niobuf->rnb_len = pg->count;
1600 niobuf->rnb_flags = pg->flag;
1605 LASSERTF((void *)(niobuf - niocount) ==
1606 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1607 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1608 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1610 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1612 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1613 body->oa.o_valid |= OBD_MD_FLFLAGS;
1614 body->oa.o_flags = 0;
1616 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1619 if (osc_should_shrink_grant(cli))
1620 osc_shrink_grant_local(cli, &body->oa);
1622 /* size[REQ_REC_OFF] still sizeof (*body) */
1623 if (opc == OST_WRITE) {
1624 if (cli->cl_checksum &&
1625 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1626 /* store cl_cksum_type in a local variable since
1627 * it can be changed via lprocfs */
1628 enum cksum_types cksum_type = cli->cl_cksum_type;
1630 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1631 body->oa.o_flags = 0;
1633 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1635 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1637 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1638 requested_nob, page_count,
1642 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1646 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1649 /* save this in 'oa', too, for later checking */
1650 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1651 oa->o_flags |= obd_cksum_type_pack(obd_name,
1654 /* clear out the checksum flag, in case this is a
1655 * resend but cl_checksum is no longer set. b=11238 */
1656 oa->o_valid &= ~OBD_MD_FLCKSUM;
1658 oa->o_cksum = body->oa.o_cksum;
1659 /* 1 RC per niobuf */
1660 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1661 sizeof(__u32) * niocount);
1663 if (cli->cl_checksum &&
1664 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1665 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1666 body->oa.o_flags = 0;
1667 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1668 cli->cl_cksum_type);
1669 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1672 /* Client cksum has been already copied to wire obdo in previous
1673 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1674 * resent due to cksum error, this will allow Server to
1675 * check+dump pages on its side */
1677 ptlrpc_request_set_replen(req);
1679 aa = ptlrpc_req_async_args(aa, req);
1681 aa->aa_requested_nob = requested_nob;
1682 aa->aa_nio_count = niocount;
1683 aa->aa_page_count = page_count;
1687 INIT_LIST_HEAD(&aa->aa_oaps);
1690 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1691 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1692 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1693 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1697 ptlrpc_req_finished(req);
1701 char dbgcksum_file_name[PATH_MAX];
1703 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1704 struct brw_page **pga, __u32 server_cksum,
1712 /* will only keep dump of pages on first error for the same range in
1713 * file/fid, not during the resends/retries. */
1714 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1715 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1716 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1717 libcfs_debug_file_path_arr :
1718 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1719 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1720 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1721 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1723 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1724 client_cksum, server_cksum);
1725 filp = filp_open(dbgcksum_file_name,
1726 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1730 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1731 "checksum error: rc = %d\n", dbgcksum_file_name,
1734 CERROR("%s: can't open to dump pages with checksum "
1735 "error: rc = %d\n", dbgcksum_file_name, rc);
1739 for (i = 0; i < page_count; i++) {
1740 len = pga[i]->count;
1741 buf = kmap(pga[i]->pg);
1743 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1745 CERROR("%s: wanted to write %u but got %d "
1746 "error\n", dbgcksum_file_name, len, rc);
1751 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1752 dbgcksum_file_name, rc);
1757 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1759 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1760 filp_close(filp, NULL);
1764 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1765 __u32 client_cksum, __u32 server_cksum,
1766 struct osc_brw_async_args *aa)
1768 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1769 enum cksum_types cksum_type;
1770 obd_dif_csum_fn *fn = NULL;
1771 int sector_size = 0;
1776 if (server_cksum == client_cksum) {
1777 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1781 if (aa->aa_cli->cl_checksum_dump)
1782 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1783 server_cksum, client_cksum);
1785 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1788 switch (cksum_type) {
1789 case OBD_CKSUM_T10IP512:
1793 case OBD_CKSUM_T10IP4K:
1797 case OBD_CKSUM_T10CRC512:
1798 fn = obd_dif_crc_fn;
1801 case OBD_CKSUM_T10CRC4K:
1802 fn = obd_dif_crc_fn;
1810 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1811 aa->aa_page_count, aa->aa_ppga,
1812 OST_WRITE, fn, sector_size,
1815 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1816 aa->aa_ppga, OST_WRITE, cksum_type,
1820 msg = "failed to calculate the client write checksum";
1821 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1822 msg = "the server did not use the checksum type specified in "
1823 "the original request - likely a protocol problem";
1824 else if (new_cksum == server_cksum)
1825 msg = "changed on the client after we checksummed it - "
1826 "likely false positive due to mmap IO (bug 11742)";
1827 else if (new_cksum == client_cksum)
1828 msg = "changed in transit before arrival at OST";
1830 msg = "changed in transit AND doesn't match the original - "
1831 "likely false positive due to mmap IO (bug 11742)";
1833 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1834 DFID " object "DOSTID" extent [%llu-%llu], original "
1835 "client csum %x (type %x), server csum %x (type %x),"
1836 " client csum now %x\n",
1837 obd_name, msg, libcfs_nid2str(peer->nid),
1838 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1839 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1840 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1841 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1842 aa->aa_ppga[aa->aa_page_count - 1]->off +
1843 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1845 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1846 server_cksum, cksum_type, new_cksum);
1850 /* Note rc enters this function as number of bytes transferred */
1851 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1853 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1854 struct client_obd *cli = aa->aa_cli;
1855 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1856 const struct lnet_process_id *peer =
1857 &req->rq_import->imp_connection->c_peer;
1858 struct ost_body *body;
1859 u32 client_cksum = 0;
1860 struct inode *inode;
1864 if (rc < 0 && rc != -EDQUOT) {
1865 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1869 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1870 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1872 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1876 /* set/clear over quota flag for a uid/gid/projid */
1877 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1878 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1879 unsigned qid[LL_MAXQUOTAS] = {
1880 body->oa.o_uid, body->oa.o_gid,
1881 body->oa.o_projid };
1883 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1884 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1885 body->oa.o_valid, body->oa.o_flags);
1886 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1890 osc_update_grant(cli, body);
1895 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1896 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1898 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1900 CERROR("%s: unexpected positive size %d\n",
1905 if (req->rq_bulk != NULL &&
1906 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1909 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1910 check_write_checksum(&body->oa, peer, client_cksum,
1911 body->oa.o_cksum, aa))
1914 rc = check_write_rcs(req, aa->aa_requested_nob,
1915 aa->aa_nio_count, aa->aa_page_count,
1920 /* The rest of this function executes only for OST_READs */
1922 if (req->rq_bulk == NULL) {
1923 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1925 LASSERT(rc == req->rq_status);
1927 /* if unwrap_bulk failed, return -EAGAIN to retry */
1928 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1931 GOTO(out, rc = -EAGAIN);
1933 if (rc > aa->aa_requested_nob) {
1934 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
1935 rc, aa->aa_requested_nob);
1939 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1940 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
1941 rc, req->rq_bulk->bd_nob_transferred);
1945 if (req->rq_bulk == NULL) {
1947 int nob, pg_count, i = 0;
1950 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1951 pg_count = aa->aa_page_count;
1952 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1955 while (nob > 0 && pg_count > 0) {
1957 int count = aa->aa_ppga[i]->count > nob ?
1958 nob : aa->aa_ppga[i]->count;
1960 CDEBUG(D_CACHE, "page %p count %d\n",
1961 aa->aa_ppga[i]->pg, count);
1962 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
1963 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1965 kunmap_atomic((void *) ptr);
1974 if (rc < aa->aa_requested_nob)
1975 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1977 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1978 static int cksum_counter;
1979 u32 server_cksum = body->oa.o_cksum;
1982 enum cksum_types cksum_type;
1983 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1984 body->oa.o_flags : 0;
1986 cksum_type = obd_cksum_type_unpack(o_flags);
1987 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1988 aa->aa_page_count, aa->aa_ppga,
1989 OST_READ, &client_cksum);
1993 if (req->rq_bulk != NULL &&
1994 peer->nid != req->rq_bulk->bd_sender) {
1996 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1999 if (server_cksum != client_cksum) {
2000 struct ost_body *clbody;
2001 u32 page_count = aa->aa_page_count;
2003 clbody = req_capsule_client_get(&req->rq_pill,
2005 if (cli->cl_checksum_dump)
2006 dump_all_bulk_pages(&clbody->oa, page_count,
2007 aa->aa_ppga, server_cksum,
2010 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2011 "%s%s%s inode "DFID" object "DOSTID
2012 " extent [%llu-%llu], client %x, "
2013 "server %x, cksum_type %x\n",
2015 libcfs_nid2str(peer->nid),
2017 clbody->oa.o_valid & OBD_MD_FLFID ?
2018 clbody->oa.o_parent_seq : 0ULL,
2019 clbody->oa.o_valid & OBD_MD_FLFID ?
2020 clbody->oa.o_parent_oid : 0,
2021 clbody->oa.o_valid & OBD_MD_FLFID ?
2022 clbody->oa.o_parent_ver : 0,
2023 POSTID(&body->oa.o_oi),
2024 aa->aa_ppga[0]->off,
2025 aa->aa_ppga[page_count-1]->off +
2026 aa->aa_ppga[page_count-1]->count - 1,
2027 client_cksum, server_cksum,
2030 aa->aa_oa->o_cksum = client_cksum;
2034 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2037 } else if (unlikely(client_cksum)) {
2038 static int cksum_missed;
2041 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2042 CERROR("%s: checksum %u requested from %s but not sent\n",
2043 obd_name, cksum_missed,
2044 libcfs_nid2str(peer->nid));
2049 inode = page2inode(aa->aa_ppga[0]->pg);
2050 if (inode && IS_ENCRYPTED(inode)) {
2053 if (!llcrypt_has_encryption_key(inode)) {
2054 CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2057 for (idx = 0; idx < aa->aa_page_count; idx++) {
2058 struct brw_page *pg = aa->aa_ppga[idx];
2061 /* do not decrypt if page is all 0s */
2062 p = q = page_address(pg->pg);
2063 while (p - q < PAGE_SIZE / sizeof(*p)) {
2068 if (p - q == PAGE_SIZE / sizeof(*p))
2071 rc = llcrypt_decrypt_pagecache_blocks(pg->pg,
2080 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2081 aa->aa_oa, &body->oa);
2086 static int osc_brw_redo_request(struct ptlrpc_request *request,
2087 struct osc_brw_async_args *aa, int rc)
2089 struct ptlrpc_request *new_req;
2090 struct osc_brw_async_args *new_aa;
2091 struct osc_async_page *oap;
2094 /* The below message is checked in replay-ost-single.sh test_8ae*/
2095 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2096 "redo for recoverable error %d", rc);
2098 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2099 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2100 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2101 aa->aa_ppga, &new_req, 1);
2105 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2106 if (oap->oap_request != NULL) {
2107 LASSERTF(request == oap->oap_request,
2108 "request %p != oap_request %p\n",
2109 request, oap->oap_request);
2113 * New request takes over pga and oaps from old request.
2114 * Note that copying a list_head doesn't work, need to move it...
2117 new_req->rq_interpret_reply = request->rq_interpret_reply;
2118 new_req->rq_async_args = request->rq_async_args;
2119 new_req->rq_commit_cb = request->rq_commit_cb;
2120 /* cap resend delay to the current request timeout, this is similar to
2121 * what ptlrpc does (see after_reply()) */
2122 if (aa->aa_resends > new_req->rq_timeout)
2123 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2125 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2126 new_req->rq_generation_set = 1;
2127 new_req->rq_import_generation = request->rq_import_generation;
2129 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2131 INIT_LIST_HEAD(&new_aa->aa_oaps);
2132 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2133 INIT_LIST_HEAD(&new_aa->aa_exts);
2134 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2135 new_aa->aa_resends = aa->aa_resends;
2137 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2138 if (oap->oap_request) {
2139 ptlrpc_req_finished(oap->oap_request);
2140 oap->oap_request = ptlrpc_request_addref(new_req);
2144 /* XXX: This code will run into problem if we're going to support
2145 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2146 * and wait for all of them to be finished. We should inherit request
2147 * set from old request. */
2148 ptlrpcd_add_req(new_req);
2150 DEBUG_REQ(D_INFO, new_req, "new request");
2155 * ugh, we want disk allocation on the target to happen in offset order. we'll
2156 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2157 * fine for our small page arrays and doesn't require allocation. its an
2158 * insertion sort that swaps elements that are strides apart, shrinking the
2159 * stride down until its '1' and the array is sorted.
2161 static void sort_brw_pages(struct brw_page **array, int num)
2164 struct brw_page *tmp;
2168 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2173 for (i = stride ; i < num ; i++) {
2176 while (j >= stride && array[j - stride]->off > tmp->off) {
2177 array[j] = array[j - stride];
2182 } while (stride > 1);
2185 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2187 LASSERT(ppga != NULL);
2188 OBD_FREE_PTR_ARRAY(ppga, count);
2191 static int brw_interpret(const struct lu_env *env,
2192 struct ptlrpc_request *req, void *args, int rc)
2194 struct osc_brw_async_args *aa = args;
2195 struct osc_extent *ext;
2196 struct osc_extent *tmp;
2197 struct client_obd *cli = aa->aa_cli;
2198 unsigned long transferred = 0;
2202 rc = osc_brw_fini_request(req, rc);
2203 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2205 /* restore clear text pages */
2206 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2209 * When server returns -EINPROGRESS, client should always retry
2210 * regardless of the number of times the bulk was resent already.
2212 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2213 if (req->rq_import_generation !=
2214 req->rq_import->imp_generation) {
2215 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2216 ""DOSTID", rc = %d.\n",
2217 req->rq_import->imp_obd->obd_name,
2218 POSTID(&aa->aa_oa->o_oi), rc);
2219 } else if (rc == -EINPROGRESS ||
2220 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2221 rc = osc_brw_redo_request(req, aa, rc);
2223 CERROR("%s: too many resent retries for object: "
2224 "%llu:%llu, rc = %d.\n",
2225 req->rq_import->imp_obd->obd_name,
2226 POSTID(&aa->aa_oa->o_oi), rc);
2231 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2236 struct obdo *oa = aa->aa_oa;
2237 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2238 unsigned long valid = 0;
2239 struct cl_object *obj;
2240 struct osc_async_page *last;
2242 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2243 obj = osc2cl(last->oap_obj);
2245 cl_object_attr_lock(obj);
2246 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2247 attr->cat_blocks = oa->o_blocks;
2248 valid |= CAT_BLOCKS;
2250 if (oa->o_valid & OBD_MD_FLMTIME) {
2251 attr->cat_mtime = oa->o_mtime;
2254 if (oa->o_valid & OBD_MD_FLATIME) {
2255 attr->cat_atime = oa->o_atime;
2258 if (oa->o_valid & OBD_MD_FLCTIME) {
2259 attr->cat_ctime = oa->o_ctime;
2263 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2264 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2265 loff_t last_off = last->oap_count + last->oap_obj_off +
2268 /* Change file size if this is an out of quota or
2269 * direct IO write and it extends the file size */
2270 if (loi->loi_lvb.lvb_size < last_off) {
2271 attr->cat_size = last_off;
2274 /* Extend KMS if it's not a lockless write */
2275 if (loi->loi_kms < last_off &&
2276 oap2osc_page(last)->ops_srvlock == 0) {
2277 attr->cat_kms = last_off;
2283 cl_object_attr_update(env, obj, attr, valid);
2284 cl_object_attr_unlock(obj);
2286 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2289 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2290 osc_inc_unstable_pages(req);
2292 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2293 list_del_init(&ext->oe_link);
2294 osc_extent_finish(env, ext, 1,
2295 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2297 LASSERT(list_empty(&aa->aa_exts));
2298 LASSERT(list_empty(&aa->aa_oaps));
2300 transferred = (req->rq_bulk == NULL ? /* short io */
2301 aa->aa_requested_nob :
2302 req->rq_bulk->bd_nob_transferred);
2304 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2305 ptlrpc_lprocfs_brw(req, transferred);
2307 spin_lock(&cli->cl_loi_list_lock);
2308 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2309 * is called so we know whether to go to sync BRWs or wait for more
2310 * RPCs to complete */
2311 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2312 cli->cl_w_in_flight--;
2314 cli->cl_r_in_flight--;
2315 osc_wake_cache_waiters(cli);
2316 spin_unlock(&cli->cl_loi_list_lock);
2318 osc_io_unplug(env, cli, NULL);
2322 static void brw_commit(struct ptlrpc_request *req)
2324 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2325 * this called via the rq_commit_cb, I need to ensure
2326 * osc_dec_unstable_pages is still called. Otherwise unstable
2327 * pages may be leaked. */
2328 spin_lock(&req->rq_lock);
2329 if (likely(req->rq_unstable)) {
2330 req->rq_unstable = 0;
2331 spin_unlock(&req->rq_lock);
2333 osc_dec_unstable_pages(req);
2335 req->rq_committed = 1;
2336 spin_unlock(&req->rq_lock);
2341 * Build an RPC by the list of extent @ext_list. The caller must ensure
2342 * that the total pages in this list are NOT over max pages per RPC.
2343 * Extents in the list must be in OES_RPC state.
2345 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2346 struct list_head *ext_list, int cmd)
2348 struct ptlrpc_request *req = NULL;
2349 struct osc_extent *ext;
2350 struct brw_page **pga = NULL;
2351 struct osc_brw_async_args *aa = NULL;
2352 struct obdo *oa = NULL;
2353 struct osc_async_page *oap;
2354 struct osc_object *obj = NULL;
2355 struct cl_req_attr *crattr = NULL;
2356 loff_t starting_offset = OBD_OBJECT_EOF;
2357 loff_t ending_offset = 0;
2358 /* '1' for consistency with code that checks !mpflag to restore */
2362 bool soft_sync = false;
2363 bool ndelay = false;
2367 __u32 layout_version = 0;
2368 LIST_HEAD(rpc_list);
2369 struct ost_body *body;
2371 LASSERT(!list_empty(ext_list));
2373 /* add pages into rpc_list to build BRW rpc */
2374 list_for_each_entry(ext, ext_list, oe_link) {
2375 LASSERT(ext->oe_state == OES_RPC);
2376 mem_tight |= ext->oe_memalloc;
2377 grant += ext->oe_grants;
2378 page_count += ext->oe_nr_pages;
2379 layout_version = max(layout_version, ext->oe_layout_version);
2384 soft_sync = osc_over_unstable_soft_limit(cli);
2386 mpflag = memalloc_noreclaim_save();
2388 OBD_ALLOC_PTR_ARRAY(pga, page_count);
2390 GOTO(out, rc = -ENOMEM);
2392 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2394 GOTO(out, rc = -ENOMEM);
2397 list_for_each_entry(ext, ext_list, oe_link) {
2398 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2400 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2402 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2403 pga[i] = &oap->oap_brw_page;
2404 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2407 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2408 if (starting_offset == OBD_OBJECT_EOF ||
2409 starting_offset > oap->oap_obj_off)
2410 starting_offset = oap->oap_obj_off;
2412 LASSERT(oap->oap_page_off == 0);
2413 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2414 ending_offset = oap->oap_obj_off +
2417 LASSERT(oap->oap_page_off + oap->oap_count ==
2424 /* first page in the list */
2425 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2427 crattr = &osc_env_info(env)->oti_req_attr;
2428 memset(crattr, 0, sizeof(*crattr));
2429 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2430 crattr->cra_flags = ~0ULL;
2431 crattr->cra_page = oap2cl_page(oap);
2432 crattr->cra_oa = oa;
2433 cl_req_attr_set(env, osc2cl(obj), crattr);
2435 if (cmd == OBD_BRW_WRITE) {
2436 oa->o_grant_used = grant;
2437 if (layout_version > 0) {
2438 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2439 PFID(&oa->o_oi.oi_fid), layout_version);
2441 oa->o_layout_version = layout_version;
2442 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2446 sort_brw_pages(pga, page_count);
2447 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2449 CERROR("prep_req failed: %d\n", rc);
2453 req->rq_commit_cb = brw_commit;
2454 req->rq_interpret_reply = brw_interpret;
2455 req->rq_memalloc = mem_tight != 0;
2456 oap->oap_request = ptlrpc_request_addref(req);
2458 req->rq_no_resend = req->rq_no_delay = 1;
2459 /* probably set a shorter timeout value.
2460 * to handle ETIMEDOUT in brw_interpret() correctly. */
2461 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2464 /* Need to update the timestamps after the request is built in case
2465 * we race with setattr (locally or in queue at OST). If OST gets
2466 * later setattr before earlier BRW (as determined by the request xid),
2467 * the OST will not use BRW timestamps. Sadly, there is no obvious
2468 * way to do this in a single call. bug 10150 */
2469 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2470 crattr->cra_oa = &body->oa;
2471 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2472 cl_req_attr_set(env, osc2cl(obj), crattr);
2473 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2475 aa = ptlrpc_req_async_args(aa, req);
2476 INIT_LIST_HEAD(&aa->aa_oaps);
2477 list_splice_init(&rpc_list, &aa->aa_oaps);
2478 INIT_LIST_HEAD(&aa->aa_exts);
2479 list_splice_init(ext_list, &aa->aa_exts);
2481 spin_lock(&cli->cl_loi_list_lock);
2482 starting_offset >>= PAGE_SHIFT;
2483 if (cmd == OBD_BRW_READ) {
2484 cli->cl_r_in_flight++;
2485 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2486 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2487 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2488 starting_offset + 1);
2490 cli->cl_w_in_flight++;
2491 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2492 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2493 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2494 starting_offset + 1);
2496 spin_unlock(&cli->cl_loi_list_lock);
2498 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2499 page_count, aa, cli->cl_r_in_flight,
2500 cli->cl_w_in_flight);
2501 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2503 ptlrpcd_add_req(req);
2509 memalloc_noreclaim_restore(mpflag);
2512 LASSERT(req == NULL);
2515 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2517 osc_release_bounce_pages(pga, page_count);
2518 osc_release_ppga(pga, page_count);
2520 /* this should happen rarely and is pretty bad, it makes the
2521 * pending list not follow the dirty order */
2522 while (!list_empty(ext_list)) {
2523 ext = list_entry(ext_list->next, struct osc_extent,
2525 list_del_init(&ext->oe_link);
2526 osc_extent_finish(env, ext, 0, rc);
2532 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2536 LASSERT(lock != NULL);
2538 lock_res_and_lock(lock);
2540 if (lock->l_ast_data == NULL)
2541 lock->l_ast_data = data;
2542 if (lock->l_ast_data == data)
2545 unlock_res_and_lock(lock);
2550 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2551 void *cookie, struct lustre_handle *lockh,
2552 enum ldlm_mode mode, __u64 *flags, bool speculative,
2555 bool intent = *flags & LDLM_FL_HAS_INTENT;
2559 /* The request was created before ldlm_cli_enqueue call. */
2560 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2561 struct ldlm_reply *rep;
2563 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2564 LASSERT(rep != NULL);
2566 rep->lock_policy_res1 =
2567 ptlrpc_status_ntoh(rep->lock_policy_res1);
2568 if (rep->lock_policy_res1)
2569 errcode = rep->lock_policy_res1;
2571 *flags |= LDLM_FL_LVB_READY;
2572 } else if (errcode == ELDLM_OK) {
2573 *flags |= LDLM_FL_LVB_READY;
2576 /* Call the update callback. */
2577 rc = (*upcall)(cookie, lockh, errcode);
2579 /* release the reference taken in ldlm_cli_enqueue() */
2580 if (errcode == ELDLM_LOCK_MATCHED)
2582 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2583 ldlm_lock_decref(lockh, mode);
2588 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2591 struct osc_enqueue_args *aa = args;
2592 struct ldlm_lock *lock;
2593 struct lustre_handle *lockh = &aa->oa_lockh;
2594 enum ldlm_mode mode = aa->oa_mode;
2595 struct ost_lvb *lvb = aa->oa_lvb;
2596 __u32 lvb_len = sizeof(*lvb);
2601 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2603 lock = ldlm_handle2lock(lockh);
2604 LASSERTF(lock != NULL,
2605 "lockh %#llx, req %p, aa %p - client evicted?\n",
2606 lockh->cookie, req, aa);
2608 /* Take an additional reference so that a blocking AST that
2609 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2610 * to arrive after an upcall has been executed by
2611 * osc_enqueue_fini(). */
2612 ldlm_lock_addref(lockh, mode);
2614 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2615 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2617 /* Let CP AST to grant the lock first. */
2618 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2620 if (aa->oa_speculative) {
2621 LASSERT(aa->oa_lvb == NULL);
2622 LASSERT(aa->oa_flags == NULL);
2623 aa->oa_flags = &flags;
2626 /* Complete obtaining the lock procedure. */
2627 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2628 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2630 /* Complete osc stuff. */
2631 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2632 aa->oa_flags, aa->oa_speculative, rc);
2634 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2636 ldlm_lock_decref(lockh, mode);
2637 LDLM_LOCK_PUT(lock);
2641 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2642 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2643 * other synchronous requests, however keeping some locks and trying to obtain
2644 * others may take a considerable amount of time in a case of ost failure; and
2645 * when other sync requests do not get released lock from a client, the client
2646 * is evicted from the cluster -- such scenarious make the life difficult, so
2647 * release locks just after they are obtained. */
2648 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2649 __u64 *flags, union ldlm_policy_data *policy,
2650 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2651 void *cookie, struct ldlm_enqueue_info *einfo,
2652 struct ptlrpc_request_set *rqset, int async,
2655 struct obd_device *obd = exp->exp_obd;
2656 struct lustre_handle lockh = { 0 };
2657 struct ptlrpc_request *req = NULL;
2658 int intent = *flags & LDLM_FL_HAS_INTENT;
2659 __u64 match_flags = *flags;
2660 enum ldlm_mode mode;
2664 /* Filesystem lock extents are extended to page boundaries so that
2665 * dealing with the page cache is a little smoother. */
2666 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2667 policy->l_extent.end |= ~PAGE_MASK;
2669 /* Next, search for already existing extent locks that will cover us */
2670 /* If we're trying to read, we also search for an existing PW lock. The
2671 * VFS and page cache already protect us locally, so lots of readers/
2672 * writers can share a single PW lock.
2674 * There are problems with conversion deadlocks, so instead of
2675 * converting a read lock to a write lock, we'll just enqueue a new
2678 * At some point we should cancel the read lock instead of making them
2679 * send us a blocking callback, but there are problems with canceling
2680 * locks out from other users right now, too. */
2681 mode = einfo->ei_mode;
2682 if (einfo->ei_mode == LCK_PR)
2684 /* Normal lock requests must wait for the LVB to be ready before
2685 * matching a lock; speculative lock requests do not need to,
2686 * because they will not actually use the lock. */
2688 match_flags |= LDLM_FL_LVB_READY;
2690 match_flags |= LDLM_FL_BLOCK_GRANTED;
2691 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2692 einfo->ei_type, policy, mode, &lockh, 0);
2694 struct ldlm_lock *matched;
2696 if (*flags & LDLM_FL_TEST_LOCK)
2699 matched = ldlm_handle2lock(&lockh);
2701 /* This DLM lock request is speculative, and does not
2702 * have an associated IO request. Therefore if there
2703 * is already a DLM lock, it wll just inform the
2704 * caller to cancel the request for this stripe.*/
2705 lock_res_and_lock(matched);
2706 if (ldlm_extent_equal(&policy->l_extent,
2707 &matched->l_policy_data.l_extent))
2711 unlock_res_and_lock(matched);
2713 ldlm_lock_decref(&lockh, mode);
2714 LDLM_LOCK_PUT(matched);
2716 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2717 *flags |= LDLM_FL_LVB_READY;
2719 /* We already have a lock, and it's referenced. */
2720 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2722 ldlm_lock_decref(&lockh, mode);
2723 LDLM_LOCK_PUT(matched);
2726 ldlm_lock_decref(&lockh, mode);
2727 LDLM_LOCK_PUT(matched);
2731 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2735 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2736 &RQF_LDLM_ENQUEUE_LVB);
2740 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2742 ptlrpc_request_free(req);
2746 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2748 ptlrpc_request_set_replen(req);
2751 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2752 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2754 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2755 sizeof(*lvb), LVB_T_OST, &lockh, async);
2758 struct osc_enqueue_args *aa;
2759 aa = ptlrpc_req_async_args(aa, req);
2761 aa->oa_mode = einfo->ei_mode;
2762 aa->oa_type = einfo->ei_type;
2763 lustre_handle_copy(&aa->oa_lockh, &lockh);
2764 aa->oa_upcall = upcall;
2765 aa->oa_cookie = cookie;
2766 aa->oa_speculative = speculative;
2768 aa->oa_flags = flags;
2771 /* speculative locks are essentially to enqueue
2772 * a DLM lock in advance, so we don't care
2773 * about the result of the enqueue. */
2775 aa->oa_flags = NULL;
2778 req->rq_interpret_reply = osc_enqueue_interpret;
2779 ptlrpc_set_add_req(rqset, req);
2780 } else if (intent) {
2781 ptlrpc_req_finished(req);
2786 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2787 flags, speculative, rc);
2789 ptlrpc_req_finished(req);
2794 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2795 struct ldlm_res_id *res_id, enum ldlm_type type,
2796 union ldlm_policy_data *policy, enum ldlm_mode mode,
2797 __u64 *flags, struct osc_object *obj,
2798 struct lustre_handle *lockh, int unref)
2800 struct obd_device *obd = exp->exp_obd;
2801 __u64 lflags = *flags;
2805 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2808 /* Filesystem lock extents are extended to page boundaries so that
2809 * dealing with the page cache is a little smoother */
2810 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2811 policy->l_extent.end |= ~PAGE_MASK;
2813 /* Next, search for already existing extent locks that will cover us */
2814 /* If we're trying to read, we also search for an existing PW lock. The
2815 * VFS and page cache already protect us locally, so lots of readers/
2816 * writers can share a single PW lock. */
2820 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2821 res_id, type, policy, rc, lockh, unref);
2822 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2826 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2828 LASSERT(lock != NULL);
2829 if (osc_set_lock_data(lock, obj)) {
2830 lock_res_and_lock(lock);
2831 if (!ldlm_is_lvb_cached(lock)) {
2832 LASSERT(lock->l_ast_data == obj);
2833 osc_lock_lvb_update(env, obj, lock, NULL);
2834 ldlm_set_lvb_cached(lock);
2836 unlock_res_and_lock(lock);
2838 ldlm_lock_decref(lockh, rc);
2841 LDLM_LOCK_PUT(lock);
2846 static int osc_statfs_interpret(const struct lu_env *env,
2847 struct ptlrpc_request *req, void *args, int rc)
2849 struct osc_async_args *aa = args;
2850 struct obd_statfs *msfs;
2855 * The request has in fact never been sent due to issues at
2856 * a higher level (LOV). Exit immediately since the caller
2857 * is aware of the problem and takes care of the clean up.
2861 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2862 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2868 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2870 GOTO(out, rc = -EPROTO);
2872 *aa->aa_oi->oi_osfs = *msfs;
2874 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2879 static int osc_statfs_async(struct obd_export *exp,
2880 struct obd_info *oinfo, time64_t max_age,
2881 struct ptlrpc_request_set *rqset)
2883 struct obd_device *obd = class_exp2obd(exp);
2884 struct ptlrpc_request *req;
2885 struct osc_async_args *aa;
2889 if (obd->obd_osfs_age >= max_age) {
2891 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2892 obd->obd_name, &obd->obd_osfs,
2893 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2894 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2895 spin_lock(&obd->obd_osfs_lock);
2896 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2897 spin_unlock(&obd->obd_osfs_lock);
2898 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2899 if (oinfo->oi_cb_up)
2900 oinfo->oi_cb_up(oinfo, 0);
2905 /* We could possibly pass max_age in the request (as an absolute
2906 * timestamp or a "seconds.usec ago") so the target can avoid doing
2907 * extra calls into the filesystem if that isn't necessary (e.g.
2908 * during mount that would help a bit). Having relative timestamps
2909 * is not so great if request processing is slow, while absolute
2910 * timestamps are not ideal because they need time synchronization. */
2911 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2915 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2917 ptlrpc_request_free(req);
2920 ptlrpc_request_set_replen(req);
2921 req->rq_request_portal = OST_CREATE_PORTAL;
2922 ptlrpc_at_set_req_timeout(req);
2924 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2925 /* procfs requests not want stat in wait for avoid deadlock */
2926 req->rq_no_resend = 1;
2927 req->rq_no_delay = 1;
2930 req->rq_interpret_reply = osc_statfs_interpret;
2931 aa = ptlrpc_req_async_args(aa, req);
2934 ptlrpc_set_add_req(rqset, req);
2938 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2939 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2941 struct obd_device *obd = class_exp2obd(exp);
2942 struct obd_statfs *msfs;
2943 struct ptlrpc_request *req;
2944 struct obd_import *imp = NULL;
2949 /*Since the request might also come from lprocfs, so we need
2950 *sync this with client_disconnect_export Bug15684*/
2951 down_read(&obd->u.cli.cl_sem);
2952 if (obd->u.cli.cl_import)
2953 imp = class_import_get(obd->u.cli.cl_import);
2954 up_read(&obd->u.cli.cl_sem);
2958 /* We could possibly pass max_age in the request (as an absolute
2959 * timestamp or a "seconds.usec ago") so the target can avoid doing
2960 * extra calls into the filesystem if that isn't necessary (e.g.
2961 * during mount that would help a bit). Having relative timestamps
2962 * is not so great if request processing is slow, while absolute
2963 * timestamps are not ideal because they need time synchronization. */
2964 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2966 class_import_put(imp);
2971 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2973 ptlrpc_request_free(req);
2976 ptlrpc_request_set_replen(req);
2977 req->rq_request_portal = OST_CREATE_PORTAL;
2978 ptlrpc_at_set_req_timeout(req);
2980 if (flags & OBD_STATFS_NODELAY) {
2981 /* procfs requests not want stat in wait for avoid deadlock */
2982 req->rq_no_resend = 1;
2983 req->rq_no_delay = 1;
2986 rc = ptlrpc_queue_wait(req);
2990 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2992 GOTO(out, rc = -EPROTO);
2998 ptlrpc_req_finished(req);
3002 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3003 void *karg, void __user *uarg)
3005 struct obd_device *obd = exp->exp_obd;
3006 struct obd_ioctl_data *data = karg;
3010 if (!try_module_get(THIS_MODULE)) {
3011 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3012 module_name(THIS_MODULE));
3016 case OBD_IOC_CLIENT_RECOVER:
3017 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3018 data->ioc_inlbuf1, 0);
3022 case IOC_OSC_SET_ACTIVE:
3023 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3028 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
3029 obd->obd_name, cmd, current->comm, rc);
3033 module_put(THIS_MODULE);
3037 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3038 u32 keylen, void *key, u32 vallen, void *val,
3039 struct ptlrpc_request_set *set)
3041 struct ptlrpc_request *req;
3042 struct obd_device *obd = exp->exp_obd;
3043 struct obd_import *imp = class_exp2cliimp(exp);
3048 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3050 if (KEY_IS(KEY_CHECKSUM)) {
3051 if (vallen != sizeof(int))
3053 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3057 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3058 sptlrpc_conf_client_adapt(obd);
3062 if (KEY_IS(KEY_FLUSH_CTX)) {
3063 sptlrpc_import_flush_my_ctx(imp);
3067 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3068 struct client_obd *cli = &obd->u.cli;
3069 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3070 long target = *(long *)val;
3072 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3077 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3080 /* We pass all other commands directly to OST. Since nobody calls osc
3081 methods directly and everybody is supposed to go through LOV, we
3082 assume lov checked invalid values for us.
3083 The only recognised values so far are evict_by_nid and mds_conn.
3084 Even if something bad goes through, we'd get a -EINVAL from OST
3087 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3088 &RQF_OST_SET_GRANT_INFO :
3093 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3094 RCL_CLIENT, keylen);
3095 if (!KEY_IS(KEY_GRANT_SHRINK))
3096 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3097 RCL_CLIENT, vallen);
3098 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3100 ptlrpc_request_free(req);
3104 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3105 memcpy(tmp, key, keylen);
3106 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3109 memcpy(tmp, val, vallen);
3111 if (KEY_IS(KEY_GRANT_SHRINK)) {
3112 struct osc_grant_args *aa;
3115 aa = ptlrpc_req_async_args(aa, req);
3116 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3118 ptlrpc_req_finished(req);
3121 *oa = ((struct ost_body *)val)->oa;
3123 req->rq_interpret_reply = osc_shrink_grant_interpret;
3126 ptlrpc_request_set_replen(req);
3127 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3128 LASSERT(set != NULL);
3129 ptlrpc_set_add_req(set, req);
3130 ptlrpc_check_set(NULL, set);
3132 ptlrpcd_add_req(req);
3137 EXPORT_SYMBOL(osc_set_info_async);
3139 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3140 struct obd_device *obd, struct obd_uuid *cluuid,
3141 struct obd_connect_data *data, void *localdata)
3143 struct client_obd *cli = &obd->u.cli;
3145 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3149 spin_lock(&cli->cl_loi_list_lock);
3150 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3151 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3152 /* restore ocd_grant_blkbits as client page bits */
3153 data->ocd_grant_blkbits = PAGE_SHIFT;
3154 grant += cli->cl_dirty_grant;
3156 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3158 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3159 lost_grant = cli->cl_lost_grant;
3160 cli->cl_lost_grant = 0;
3161 spin_unlock(&cli->cl_loi_list_lock);
3163 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3164 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3165 data->ocd_version, data->ocd_grant, lost_grant);
3170 EXPORT_SYMBOL(osc_reconnect);
3172 int osc_disconnect(struct obd_export *exp)
3174 struct obd_device *obd = class_exp2obd(exp);
3177 rc = client_disconnect_export(exp);
3179 * Initially we put del_shrink_grant before disconnect_export, but it
3180 * causes the following problem if setup (connect) and cleanup
3181 * (disconnect) are tangled together.
3182 * connect p1 disconnect p2
3183 * ptlrpc_connect_import
3184 * ............... class_manual_cleanup
3187 * ptlrpc_connect_interrupt
3189 * add this client to shrink list
3191 * Bang! grant shrink thread trigger the shrink. BUG18662
3193 osc_del_grant_list(&obd->u.cli);
3196 EXPORT_SYMBOL(osc_disconnect);
3198 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3199 struct hlist_node *hnode, void *arg)
3201 struct lu_env *env = arg;
3202 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3203 struct ldlm_lock *lock;
3204 struct osc_object *osc = NULL;
3208 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3209 if (lock->l_ast_data != NULL && osc == NULL) {
3210 osc = lock->l_ast_data;
3211 cl_object_get(osc2cl(osc));
3214 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3215 * by the 2nd round of ldlm_namespace_clean() call in
3216 * osc_import_event(). */
3217 ldlm_clear_cleaned(lock);
3222 osc_object_invalidate(env, osc);
3223 cl_object_put(env, osc2cl(osc));
3228 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3230 static int osc_import_event(struct obd_device *obd,
3231 struct obd_import *imp,
3232 enum obd_import_event event)
3234 struct client_obd *cli;
3238 LASSERT(imp->imp_obd == obd);
3241 case IMP_EVENT_DISCON: {
3243 spin_lock(&cli->cl_loi_list_lock);
3244 cli->cl_avail_grant = 0;
3245 cli->cl_lost_grant = 0;
3246 spin_unlock(&cli->cl_loi_list_lock);
3249 case IMP_EVENT_INACTIVE: {
3250 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3253 case IMP_EVENT_INVALIDATE: {
3254 struct ldlm_namespace *ns = obd->obd_namespace;
3258 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3260 env = cl_env_get(&refcheck);
3262 osc_io_unplug(env, &obd->u.cli, NULL);
3264 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3265 osc_ldlm_resource_invalidate,
3267 cl_env_put(env, &refcheck);
3269 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3274 case IMP_EVENT_ACTIVE: {
3275 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3278 case IMP_EVENT_OCD: {
3279 struct obd_connect_data *ocd = &imp->imp_connect_data;
3281 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3282 osc_init_grant(&obd->u.cli, ocd);
3285 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3286 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3288 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3291 case IMP_EVENT_DEACTIVATE: {
3292 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3295 case IMP_EVENT_ACTIVATE: {
3296 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3300 CERROR("Unknown import event %d\n", event);
3307 * Determine whether the lock can be canceled before replaying the lock
3308 * during recovery, see bug16774 for detailed information.
3310 * \retval zero the lock can't be canceled
3311 * \retval other ok to cancel
3313 static int osc_cancel_weight(struct ldlm_lock *lock)
3316 * Cancel all unused and granted extent lock.
3318 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3319 ldlm_is_granted(lock) &&
3320 osc_ldlm_weigh_ast(lock) == 0)
3326 static int brw_queue_work(const struct lu_env *env, void *data)
3328 struct client_obd *cli = data;
3330 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3332 osc_io_unplug(env, cli, NULL);
3336 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3338 struct client_obd *cli = &obd->u.cli;
3344 rc = ptlrpcd_addref();
3348 rc = client_obd_setup(obd, lcfg);
3350 GOTO(out_ptlrpcd, rc);
3353 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3354 if (IS_ERR(handler))
3355 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3356 cli->cl_writeback_work = handler;
3358 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3359 if (IS_ERR(handler))
3360 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3361 cli->cl_lru_work = handler;
3363 rc = osc_quota_setup(obd);
3365 GOTO(out_ptlrpcd_work, rc);
3367 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3368 osc_update_next_shrink(cli);
3373 if (cli->cl_writeback_work != NULL) {
3374 ptlrpcd_destroy_work(cli->cl_writeback_work);
3375 cli->cl_writeback_work = NULL;
3377 if (cli->cl_lru_work != NULL) {
3378 ptlrpcd_destroy_work(cli->cl_lru_work);
3379 cli->cl_lru_work = NULL;
3381 client_obd_cleanup(obd);
3386 EXPORT_SYMBOL(osc_setup_common);
3388 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3390 struct client_obd *cli = &obd->u.cli;
3398 rc = osc_setup_common(obd, lcfg);
3402 rc = osc_tunables_init(obd);
3407 * We try to control the total number of requests with a upper limit
3408 * osc_reqpool_maxreqcount. There might be some race which will cause
3409 * over-limit allocation, but it is fine.
3411 req_count = atomic_read(&osc_pool_req_count);
3412 if (req_count < osc_reqpool_maxreqcount) {
3413 adding = cli->cl_max_rpcs_in_flight + 2;
3414 if (req_count + adding > osc_reqpool_maxreqcount)
3415 adding = osc_reqpool_maxreqcount - req_count;
3417 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3418 atomic_add(added, &osc_pool_req_count);
3421 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3423 spin_lock(&osc_shrink_lock);
3424 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3425 spin_unlock(&osc_shrink_lock);
3426 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3427 cli->cl_import->imp_idle_debug = D_HA;
3432 int osc_precleanup_common(struct obd_device *obd)
3434 struct client_obd *cli = &obd->u.cli;
3438 * for echo client, export may be on zombie list, wait for
3439 * zombie thread to cull it, because cli.cl_import will be
3440 * cleared in client_disconnect_export():
3441 * class_export_destroy() -> obd_cleanup() ->
3442 * echo_device_free() -> echo_client_cleanup() ->
3443 * obd_disconnect() -> osc_disconnect() ->
3444 * client_disconnect_export()
3446 obd_zombie_barrier();
3447 if (cli->cl_writeback_work) {
3448 ptlrpcd_destroy_work(cli->cl_writeback_work);
3449 cli->cl_writeback_work = NULL;
3452 if (cli->cl_lru_work) {
3453 ptlrpcd_destroy_work(cli->cl_lru_work);
3454 cli->cl_lru_work = NULL;
3457 obd_cleanup_client_import(obd);
3460 EXPORT_SYMBOL(osc_precleanup_common);
3462 static int osc_precleanup(struct obd_device *obd)
3466 osc_precleanup_common(obd);
3468 ptlrpc_lprocfs_unregister_obd(obd);
3472 int osc_cleanup_common(struct obd_device *obd)
3474 struct client_obd *cli = &obd->u.cli;
3479 spin_lock(&osc_shrink_lock);
3480 list_del(&cli->cl_shrink_list);
3481 spin_unlock(&osc_shrink_lock);
3484 if (cli->cl_cache != NULL) {
3485 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3486 spin_lock(&cli->cl_cache->ccc_lru_lock);
3487 list_del_init(&cli->cl_lru_osc);
3488 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3489 cli->cl_lru_left = NULL;
3490 cl_cache_decref(cli->cl_cache);
3491 cli->cl_cache = NULL;
3494 /* free memory of osc quota cache */
3495 osc_quota_cleanup(obd);
3497 rc = client_obd_cleanup(obd);
3502 EXPORT_SYMBOL(osc_cleanup_common);
3504 static const struct obd_ops osc_obd_ops = {
3505 .o_owner = THIS_MODULE,
3506 .o_setup = osc_setup,
3507 .o_precleanup = osc_precleanup,
3508 .o_cleanup = osc_cleanup_common,
3509 .o_add_conn = client_import_add_conn,
3510 .o_del_conn = client_import_del_conn,
3511 .o_connect = client_connect_import,
3512 .o_reconnect = osc_reconnect,
3513 .o_disconnect = osc_disconnect,
3514 .o_statfs = osc_statfs,
3515 .o_statfs_async = osc_statfs_async,
3516 .o_create = osc_create,
3517 .o_destroy = osc_destroy,
3518 .o_getattr = osc_getattr,
3519 .o_setattr = osc_setattr,
3520 .o_iocontrol = osc_iocontrol,
3521 .o_set_info_async = osc_set_info_async,
3522 .o_import_event = osc_import_event,
3523 .o_quotactl = osc_quotactl,
3526 static struct shrinker *osc_cache_shrinker;
3527 LIST_HEAD(osc_shrink_list);
3528 DEFINE_SPINLOCK(osc_shrink_lock);
3530 #ifndef HAVE_SHRINKER_COUNT
3531 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3533 struct shrink_control scv = {
3534 .nr_to_scan = shrink_param(sc, nr_to_scan),
3535 .gfp_mask = shrink_param(sc, gfp_mask)
3537 (void)osc_cache_shrink_scan(shrinker, &scv);
3539 return osc_cache_shrink_count(shrinker, &scv);
3543 static int __init osc_init(void)
3545 unsigned int reqpool_size;
3546 unsigned int reqsize;
3548 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3549 osc_cache_shrink_count, osc_cache_shrink_scan);
3552 /* print an address of _any_ initialized kernel symbol from this
3553 * module, to allow debugging with gdb that doesn't support data
3554 * symbols from modules.*/
3555 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3557 rc = lu_kmem_init(osc_caches);
3561 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3562 LUSTRE_OSC_NAME, &osc_device_type);
3566 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3568 /* This is obviously too much memory, only prevent overflow here */
3569 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3570 GOTO(out_type, rc = -EINVAL);
3572 reqpool_size = osc_reqpool_mem_max << 20;
3575 while (reqsize < OST_IO_MAXREQSIZE)
3576 reqsize = reqsize << 1;
3579 * We don't enlarge the request count in OSC pool according to
3580 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3581 * tried after normal allocation failed. So a small OSC pool won't
3582 * cause much performance degression in most of cases.
3584 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3586 atomic_set(&osc_pool_req_count, 0);
3587 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3588 ptlrpc_add_rqs_to_pool);
3590 if (osc_rq_pool == NULL)
3591 GOTO(out_type, rc = -ENOMEM);
3593 rc = osc_start_grant_work();
3595 GOTO(out_req_pool, rc);
3600 ptlrpc_free_rq_pool(osc_rq_pool);
3602 class_unregister_type(LUSTRE_OSC_NAME);
3604 lu_kmem_fini(osc_caches);
3609 static void __exit osc_exit(void)
3611 osc_stop_grant_work();
3612 remove_shrinker(osc_cache_shrinker);
3613 class_unregister_type(LUSTRE_OSC_NAME);
3614 lu_kmem_fini(osc_caches);
3615 ptlrpc_free_rq_pool(osc_rq_pool);
3618 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3619 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3620 MODULE_VERSION(LUSTRE_VERSION_STRING);
3621 MODULE_LICENSE("GPL");
3623 module_init(osc_init);
3624 module_exit(osc_exit);