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);
1441 /* len is forced to PAGE_SIZE, and poff to 0
1442 * so store the old, clear text info
1445 pg->bp_count_diff = PAGE_SIZE - pg->count;
1446 pg->count = PAGE_SIZE;
1447 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1448 pg->off = pg->off & PAGE_MASK;
1452 for (niocount = i = 1; i < page_count; i++) {
1453 if (!can_merge_pages(pga[i - 1], pga[i]))
1457 pill = &req->rq_pill;
1458 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1460 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1461 niocount * sizeof(*niobuf));
1463 for (i = 0; i < page_count; i++)
1464 short_io_size += pga[i]->count;
1466 /* Check if read/write is small enough to be a short io. */
1467 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1468 !imp_connect_shortio(cli->cl_import))
1471 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1472 opc == OST_READ ? 0 : short_io_size);
1473 if (opc == OST_READ)
1474 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1477 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1479 ptlrpc_request_free(req);
1482 osc_set_io_portal(req);
1484 ptlrpc_at_set_req_timeout(req);
1485 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1487 req->rq_no_retry_einprogress = 1;
1489 if (short_io_size != 0) {
1491 short_io_buf = NULL;
1495 desc = ptlrpc_prep_bulk_imp(req, page_count,
1496 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1497 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1498 PTLRPC_BULK_PUT_SINK),
1500 &ptlrpc_bulk_kiov_pin_ops);
1503 GOTO(out, rc = -ENOMEM);
1504 /* NB request now owns desc and will free it when it gets freed */
1506 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1507 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1508 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1509 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1511 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1513 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1514 * and from_kgid(), because they are asynchronous. Fortunately, variable
1515 * oa contains valid o_uid and o_gid in these two operations.
1516 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1517 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1518 * other process logic */
1519 body->oa.o_uid = oa->o_uid;
1520 body->oa.o_gid = oa->o_gid;
1522 obdo_to_ioobj(oa, ioobj);
1523 ioobj->ioo_bufcnt = niocount;
1524 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1525 * that might be send for this request. The actual number is decided
1526 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1527 * "max - 1" for old client compatibility sending "0", and also so the
1528 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1530 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1532 ioobj_max_brw_set(ioobj, 0);
1534 if (short_io_size != 0) {
1535 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1536 body->oa.o_valid |= OBD_MD_FLFLAGS;
1537 body->oa.o_flags = 0;
1539 body->oa.o_flags |= OBD_FL_SHORT_IO;
1540 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1542 if (opc == OST_WRITE) {
1543 short_io_buf = req_capsule_client_get(pill,
1545 LASSERT(short_io_buf != NULL);
1549 LASSERT(page_count > 0);
1551 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1552 struct brw_page *pg = pga[i];
1553 int poff = pg->off & ~PAGE_MASK;
1555 LASSERT(pg->count > 0);
1556 /* make sure there is no gap in the middle of page array */
1557 LASSERTF(page_count == 1 ||
1558 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1559 ergo(i > 0 && i < page_count - 1,
1560 poff == 0 && pg->count == PAGE_SIZE) &&
1561 ergo(i == page_count - 1, poff == 0)),
1562 "i: %d/%d pg: %p off: %llu, count: %u\n",
1563 i, page_count, pg, pg->off, pg->count);
1564 LASSERTF(i == 0 || pg->off > pg_prev->off,
1565 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1566 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1568 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1569 pg_prev->pg, page_private(pg_prev->pg),
1570 pg_prev->pg->index, pg_prev->off);
1571 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1572 (pg->flag & OBD_BRW_SRVLOCK));
1573 if (short_io_size != 0 && opc == OST_WRITE) {
1574 unsigned char *ptr = kmap_atomic(pg->pg);
1576 LASSERT(short_io_size >= requested_nob + pg->count);
1577 memcpy(short_io_buf + requested_nob,
1581 } else if (short_io_size == 0) {
1582 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1585 requested_nob += pg->count;
1587 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1589 niobuf->rnb_len += pg->count;
1591 niobuf->rnb_offset = pg->off;
1592 niobuf->rnb_len = pg->count;
1593 niobuf->rnb_flags = pg->flag;
1598 LASSERTF((void *)(niobuf - niocount) ==
1599 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1600 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1601 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1603 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1605 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1606 body->oa.o_valid |= OBD_MD_FLFLAGS;
1607 body->oa.o_flags = 0;
1609 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1612 if (osc_should_shrink_grant(cli))
1613 osc_shrink_grant_local(cli, &body->oa);
1615 /* size[REQ_REC_OFF] still sizeof (*body) */
1616 if (opc == OST_WRITE) {
1617 if (cli->cl_checksum &&
1618 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1619 /* store cl_cksum_type in a local variable since
1620 * it can be changed via lprocfs */
1621 enum cksum_types cksum_type = cli->cl_cksum_type;
1623 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1624 body->oa.o_flags = 0;
1626 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1628 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1630 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1631 requested_nob, page_count,
1635 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1639 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1642 /* save this in 'oa', too, for later checking */
1643 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1644 oa->o_flags |= obd_cksum_type_pack(obd_name,
1647 /* clear out the checksum flag, in case this is a
1648 * resend but cl_checksum is no longer set. b=11238 */
1649 oa->o_valid &= ~OBD_MD_FLCKSUM;
1651 oa->o_cksum = body->oa.o_cksum;
1652 /* 1 RC per niobuf */
1653 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1654 sizeof(__u32) * niocount);
1656 if (cli->cl_checksum &&
1657 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1658 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1659 body->oa.o_flags = 0;
1660 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1661 cli->cl_cksum_type);
1662 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1665 /* Client cksum has been already copied to wire obdo in previous
1666 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1667 * resent due to cksum error, this will allow Server to
1668 * check+dump pages on its side */
1670 ptlrpc_request_set_replen(req);
1672 aa = ptlrpc_req_async_args(aa, req);
1674 aa->aa_requested_nob = requested_nob;
1675 aa->aa_nio_count = niocount;
1676 aa->aa_page_count = page_count;
1680 INIT_LIST_HEAD(&aa->aa_oaps);
1683 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1684 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1685 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1686 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1690 ptlrpc_req_finished(req);
1694 char dbgcksum_file_name[PATH_MAX];
1696 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1697 struct brw_page **pga, __u32 server_cksum,
1705 /* will only keep dump of pages on first error for the same range in
1706 * file/fid, not during the resends/retries. */
1707 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1708 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1709 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1710 libcfs_debug_file_path_arr :
1711 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1712 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1713 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1714 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1716 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1717 client_cksum, server_cksum);
1718 filp = filp_open(dbgcksum_file_name,
1719 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1723 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1724 "checksum error: rc = %d\n", dbgcksum_file_name,
1727 CERROR("%s: can't open to dump pages with checksum "
1728 "error: rc = %d\n", dbgcksum_file_name, rc);
1732 for (i = 0; i < page_count; i++) {
1733 len = pga[i]->count;
1734 buf = kmap(pga[i]->pg);
1736 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1738 CERROR("%s: wanted to write %u but got %d "
1739 "error\n", dbgcksum_file_name, len, rc);
1744 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1745 dbgcksum_file_name, rc);
1750 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1752 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1753 filp_close(filp, NULL);
1757 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1758 __u32 client_cksum, __u32 server_cksum,
1759 struct osc_brw_async_args *aa)
1761 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1762 enum cksum_types cksum_type;
1763 obd_dif_csum_fn *fn = NULL;
1764 int sector_size = 0;
1769 if (server_cksum == client_cksum) {
1770 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1774 if (aa->aa_cli->cl_checksum_dump)
1775 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1776 server_cksum, client_cksum);
1778 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1781 switch (cksum_type) {
1782 case OBD_CKSUM_T10IP512:
1786 case OBD_CKSUM_T10IP4K:
1790 case OBD_CKSUM_T10CRC512:
1791 fn = obd_dif_crc_fn;
1794 case OBD_CKSUM_T10CRC4K:
1795 fn = obd_dif_crc_fn;
1803 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1804 aa->aa_page_count, aa->aa_ppga,
1805 OST_WRITE, fn, sector_size,
1808 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1809 aa->aa_ppga, OST_WRITE, cksum_type,
1813 msg = "failed to calculate the client write checksum";
1814 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1815 msg = "the server did not use the checksum type specified in "
1816 "the original request - likely a protocol problem";
1817 else if (new_cksum == server_cksum)
1818 msg = "changed on the client after we checksummed it - "
1819 "likely false positive due to mmap IO (bug 11742)";
1820 else if (new_cksum == client_cksum)
1821 msg = "changed in transit before arrival at OST";
1823 msg = "changed in transit AND doesn't match the original - "
1824 "likely false positive due to mmap IO (bug 11742)";
1826 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1827 DFID " object "DOSTID" extent [%llu-%llu], original "
1828 "client csum %x (type %x), server csum %x (type %x),"
1829 " client csum now %x\n",
1830 obd_name, msg, libcfs_nid2str(peer->nid),
1831 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1832 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1833 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1834 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1835 aa->aa_ppga[aa->aa_page_count - 1]->off +
1836 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1838 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1839 server_cksum, cksum_type, new_cksum);
1843 /* Note rc enters this function as number of bytes transferred */
1844 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1846 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1847 struct client_obd *cli = aa->aa_cli;
1848 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1849 const struct lnet_process_id *peer =
1850 &req->rq_import->imp_connection->c_peer;
1851 struct ost_body *body;
1852 u32 client_cksum = 0;
1856 if (rc < 0 && rc != -EDQUOT) {
1857 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1861 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1862 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1864 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1868 /* set/clear over quota flag for a uid/gid/projid */
1869 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1870 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1871 unsigned qid[LL_MAXQUOTAS] = {
1872 body->oa.o_uid, body->oa.o_gid,
1873 body->oa.o_projid };
1875 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1876 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1877 body->oa.o_valid, body->oa.o_flags);
1878 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1882 osc_update_grant(cli, body);
1887 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1888 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1890 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1892 CERROR("%s: unexpected positive size %d\n",
1897 if (req->rq_bulk != NULL &&
1898 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1901 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1902 check_write_checksum(&body->oa, peer, client_cksum,
1903 body->oa.o_cksum, aa))
1906 rc = check_write_rcs(req, aa->aa_requested_nob,
1907 aa->aa_nio_count, aa->aa_page_count,
1912 /* The rest of this function executes only for OST_READs */
1914 if (req->rq_bulk == NULL) {
1915 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1917 LASSERT(rc == req->rq_status);
1919 /* if unwrap_bulk failed, return -EAGAIN to retry */
1920 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1923 GOTO(out, rc = -EAGAIN);
1925 if (rc > aa->aa_requested_nob) {
1926 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
1927 rc, aa->aa_requested_nob);
1931 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1932 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
1933 rc, req->rq_bulk->bd_nob_transferred);
1937 if (req->rq_bulk == NULL) {
1939 int nob, pg_count, i = 0;
1942 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1943 pg_count = aa->aa_page_count;
1944 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1947 while (nob > 0 && pg_count > 0) {
1949 int count = aa->aa_ppga[i]->count > nob ?
1950 nob : aa->aa_ppga[i]->count;
1952 CDEBUG(D_CACHE, "page %p count %d\n",
1953 aa->aa_ppga[i]->pg, count);
1954 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
1955 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1957 kunmap_atomic((void *) ptr);
1966 if (rc < aa->aa_requested_nob)
1967 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1969 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1970 static int cksum_counter;
1971 u32 server_cksum = body->oa.o_cksum;
1974 enum cksum_types cksum_type;
1975 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1976 body->oa.o_flags : 0;
1978 cksum_type = obd_cksum_type_unpack(o_flags);
1979 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
1980 aa->aa_page_count, aa->aa_ppga,
1981 OST_READ, &client_cksum);
1985 if (req->rq_bulk != NULL &&
1986 peer->nid != req->rq_bulk->bd_sender) {
1988 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1991 if (server_cksum != client_cksum) {
1992 struct ost_body *clbody;
1993 u32 page_count = aa->aa_page_count;
1995 clbody = req_capsule_client_get(&req->rq_pill,
1997 if (cli->cl_checksum_dump)
1998 dump_all_bulk_pages(&clbody->oa, page_count,
1999 aa->aa_ppga, server_cksum,
2002 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2003 "%s%s%s inode "DFID" object "DOSTID
2004 " extent [%llu-%llu], client %x, "
2005 "server %x, cksum_type %x\n",
2007 libcfs_nid2str(peer->nid),
2009 clbody->oa.o_valid & OBD_MD_FLFID ?
2010 clbody->oa.o_parent_seq : 0ULL,
2011 clbody->oa.o_valid & OBD_MD_FLFID ?
2012 clbody->oa.o_parent_oid : 0,
2013 clbody->oa.o_valid & OBD_MD_FLFID ?
2014 clbody->oa.o_parent_ver : 0,
2015 POSTID(&body->oa.o_oi),
2016 aa->aa_ppga[0]->off,
2017 aa->aa_ppga[page_count-1]->off +
2018 aa->aa_ppga[page_count-1]->count - 1,
2019 client_cksum, server_cksum,
2022 aa->aa_oa->o_cksum = client_cksum;
2026 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2029 } else if (unlikely(client_cksum)) {
2030 static int cksum_missed;
2033 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2034 CERROR("%s: checksum %u requested from %s but not sent\n",
2035 obd_name, cksum_missed,
2036 libcfs_nid2str(peer->nid));
2042 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2043 aa->aa_oa, &body->oa);
2048 static int osc_brw_redo_request(struct ptlrpc_request *request,
2049 struct osc_brw_async_args *aa, int rc)
2051 struct ptlrpc_request *new_req;
2052 struct osc_brw_async_args *new_aa;
2053 struct osc_async_page *oap;
2056 /* The below message is checked in replay-ost-single.sh test_8ae*/
2057 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2058 "redo for recoverable error %d", rc);
2060 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2061 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2062 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2063 aa->aa_ppga, &new_req, 1);
2067 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2068 if (oap->oap_request != NULL) {
2069 LASSERTF(request == oap->oap_request,
2070 "request %p != oap_request %p\n",
2071 request, oap->oap_request);
2075 * New request takes over pga and oaps from old request.
2076 * Note that copying a list_head doesn't work, need to move it...
2079 new_req->rq_interpret_reply = request->rq_interpret_reply;
2080 new_req->rq_async_args = request->rq_async_args;
2081 new_req->rq_commit_cb = request->rq_commit_cb;
2082 /* cap resend delay to the current request timeout, this is similar to
2083 * what ptlrpc does (see after_reply()) */
2084 if (aa->aa_resends > new_req->rq_timeout)
2085 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2087 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2088 new_req->rq_generation_set = 1;
2089 new_req->rq_import_generation = request->rq_import_generation;
2091 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2093 INIT_LIST_HEAD(&new_aa->aa_oaps);
2094 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2095 INIT_LIST_HEAD(&new_aa->aa_exts);
2096 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2097 new_aa->aa_resends = aa->aa_resends;
2099 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2100 if (oap->oap_request) {
2101 ptlrpc_req_finished(oap->oap_request);
2102 oap->oap_request = ptlrpc_request_addref(new_req);
2106 /* XXX: This code will run into problem if we're going to support
2107 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2108 * and wait for all of them to be finished. We should inherit request
2109 * set from old request. */
2110 ptlrpcd_add_req(new_req);
2112 DEBUG_REQ(D_INFO, new_req, "new request");
2117 * ugh, we want disk allocation on the target to happen in offset order. we'll
2118 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2119 * fine for our small page arrays and doesn't require allocation. its an
2120 * insertion sort that swaps elements that are strides apart, shrinking the
2121 * stride down until its '1' and the array is sorted.
2123 static void sort_brw_pages(struct brw_page **array, int num)
2126 struct brw_page *tmp;
2130 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2135 for (i = stride ; i < num ; i++) {
2138 while (j >= stride && array[j - stride]->off > tmp->off) {
2139 array[j] = array[j - stride];
2144 } while (stride > 1);
2147 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2149 LASSERT(ppga != NULL);
2150 OBD_FREE_PTR_ARRAY(ppga, count);
2153 static int brw_interpret(const struct lu_env *env,
2154 struct ptlrpc_request *req, void *args, int rc)
2156 struct osc_brw_async_args *aa = args;
2157 struct osc_extent *ext;
2158 struct osc_extent *tmp;
2159 struct client_obd *cli = aa->aa_cli;
2160 unsigned long transferred = 0;
2164 rc = osc_brw_fini_request(req, rc);
2165 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2167 /* restore clear text pages */
2168 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2171 * When server returns -EINPROGRESS, client should always retry
2172 * regardless of the number of times the bulk was resent already.
2174 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2175 if (req->rq_import_generation !=
2176 req->rq_import->imp_generation) {
2177 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2178 ""DOSTID", rc = %d.\n",
2179 req->rq_import->imp_obd->obd_name,
2180 POSTID(&aa->aa_oa->o_oi), rc);
2181 } else if (rc == -EINPROGRESS ||
2182 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2183 rc = osc_brw_redo_request(req, aa, rc);
2185 CERROR("%s: too many resent retries for object: "
2186 "%llu:%llu, rc = %d.\n",
2187 req->rq_import->imp_obd->obd_name,
2188 POSTID(&aa->aa_oa->o_oi), rc);
2193 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2198 struct obdo *oa = aa->aa_oa;
2199 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2200 unsigned long valid = 0;
2201 struct cl_object *obj;
2202 struct osc_async_page *last;
2204 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2205 obj = osc2cl(last->oap_obj);
2207 cl_object_attr_lock(obj);
2208 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2209 attr->cat_blocks = oa->o_blocks;
2210 valid |= CAT_BLOCKS;
2212 if (oa->o_valid & OBD_MD_FLMTIME) {
2213 attr->cat_mtime = oa->o_mtime;
2216 if (oa->o_valid & OBD_MD_FLATIME) {
2217 attr->cat_atime = oa->o_atime;
2220 if (oa->o_valid & OBD_MD_FLCTIME) {
2221 attr->cat_ctime = oa->o_ctime;
2225 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2226 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2227 loff_t last_off = last->oap_count + last->oap_obj_off +
2230 /* Change file size if this is an out of quota or
2231 * direct IO write and it extends the file size */
2232 if (loi->loi_lvb.lvb_size < last_off) {
2233 attr->cat_size = last_off;
2236 /* Extend KMS if it's not a lockless write */
2237 if (loi->loi_kms < last_off &&
2238 oap2osc_page(last)->ops_srvlock == 0) {
2239 attr->cat_kms = last_off;
2245 cl_object_attr_update(env, obj, attr, valid);
2246 cl_object_attr_unlock(obj);
2248 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2251 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2252 osc_inc_unstable_pages(req);
2254 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2255 list_del_init(&ext->oe_link);
2256 osc_extent_finish(env, ext, 1,
2257 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2259 LASSERT(list_empty(&aa->aa_exts));
2260 LASSERT(list_empty(&aa->aa_oaps));
2262 transferred = (req->rq_bulk == NULL ? /* short io */
2263 aa->aa_requested_nob :
2264 req->rq_bulk->bd_nob_transferred);
2266 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2267 ptlrpc_lprocfs_brw(req, transferred);
2269 spin_lock(&cli->cl_loi_list_lock);
2270 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2271 * is called so we know whether to go to sync BRWs or wait for more
2272 * RPCs to complete */
2273 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2274 cli->cl_w_in_flight--;
2276 cli->cl_r_in_flight--;
2277 osc_wake_cache_waiters(cli);
2278 spin_unlock(&cli->cl_loi_list_lock);
2280 osc_io_unplug(env, cli, NULL);
2284 static void brw_commit(struct ptlrpc_request *req)
2286 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2287 * this called via the rq_commit_cb, I need to ensure
2288 * osc_dec_unstable_pages is still called. Otherwise unstable
2289 * pages may be leaked. */
2290 spin_lock(&req->rq_lock);
2291 if (likely(req->rq_unstable)) {
2292 req->rq_unstable = 0;
2293 spin_unlock(&req->rq_lock);
2295 osc_dec_unstable_pages(req);
2297 req->rq_committed = 1;
2298 spin_unlock(&req->rq_lock);
2303 * Build an RPC by the list of extent @ext_list. The caller must ensure
2304 * that the total pages in this list are NOT over max pages per RPC.
2305 * Extents in the list must be in OES_RPC state.
2307 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2308 struct list_head *ext_list, int cmd)
2310 struct ptlrpc_request *req = NULL;
2311 struct osc_extent *ext;
2312 struct brw_page **pga = NULL;
2313 struct osc_brw_async_args *aa = NULL;
2314 struct obdo *oa = NULL;
2315 struct osc_async_page *oap;
2316 struct osc_object *obj = NULL;
2317 struct cl_req_attr *crattr = NULL;
2318 loff_t starting_offset = OBD_OBJECT_EOF;
2319 loff_t ending_offset = 0;
2320 /* '1' for consistency with code that checks !mpflag to restore */
2324 bool soft_sync = false;
2325 bool ndelay = false;
2329 __u32 layout_version = 0;
2330 LIST_HEAD(rpc_list);
2331 struct ost_body *body;
2333 LASSERT(!list_empty(ext_list));
2335 /* add pages into rpc_list to build BRW rpc */
2336 list_for_each_entry(ext, ext_list, oe_link) {
2337 LASSERT(ext->oe_state == OES_RPC);
2338 mem_tight |= ext->oe_memalloc;
2339 grant += ext->oe_grants;
2340 page_count += ext->oe_nr_pages;
2341 layout_version = max(layout_version, ext->oe_layout_version);
2346 soft_sync = osc_over_unstable_soft_limit(cli);
2348 mpflag = memalloc_noreclaim_save();
2350 OBD_ALLOC_PTR_ARRAY(pga, page_count);
2352 GOTO(out, rc = -ENOMEM);
2354 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2356 GOTO(out, rc = -ENOMEM);
2359 list_for_each_entry(ext, ext_list, oe_link) {
2360 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2362 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2364 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2365 pga[i] = &oap->oap_brw_page;
2366 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2369 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2370 if (starting_offset == OBD_OBJECT_EOF ||
2371 starting_offset > oap->oap_obj_off)
2372 starting_offset = oap->oap_obj_off;
2374 LASSERT(oap->oap_page_off == 0);
2375 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2376 ending_offset = oap->oap_obj_off +
2379 LASSERT(oap->oap_page_off + oap->oap_count ==
2386 /* first page in the list */
2387 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2389 crattr = &osc_env_info(env)->oti_req_attr;
2390 memset(crattr, 0, sizeof(*crattr));
2391 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2392 crattr->cra_flags = ~0ULL;
2393 crattr->cra_page = oap2cl_page(oap);
2394 crattr->cra_oa = oa;
2395 cl_req_attr_set(env, osc2cl(obj), crattr);
2397 if (cmd == OBD_BRW_WRITE) {
2398 oa->o_grant_used = grant;
2399 if (layout_version > 0) {
2400 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2401 PFID(&oa->o_oi.oi_fid), layout_version);
2403 oa->o_layout_version = layout_version;
2404 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2408 sort_brw_pages(pga, page_count);
2409 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2411 CERROR("prep_req failed: %d\n", rc);
2415 req->rq_commit_cb = brw_commit;
2416 req->rq_interpret_reply = brw_interpret;
2417 req->rq_memalloc = mem_tight != 0;
2418 oap->oap_request = ptlrpc_request_addref(req);
2420 req->rq_no_resend = req->rq_no_delay = 1;
2421 /* probably set a shorter timeout value.
2422 * to handle ETIMEDOUT in brw_interpret() correctly. */
2423 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2426 /* Need to update the timestamps after the request is built in case
2427 * we race with setattr (locally or in queue at OST). If OST gets
2428 * later setattr before earlier BRW (as determined by the request xid),
2429 * the OST will not use BRW timestamps. Sadly, there is no obvious
2430 * way to do this in a single call. bug 10150 */
2431 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2432 crattr->cra_oa = &body->oa;
2433 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2434 cl_req_attr_set(env, osc2cl(obj), crattr);
2435 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2437 aa = ptlrpc_req_async_args(aa, req);
2438 INIT_LIST_HEAD(&aa->aa_oaps);
2439 list_splice_init(&rpc_list, &aa->aa_oaps);
2440 INIT_LIST_HEAD(&aa->aa_exts);
2441 list_splice_init(ext_list, &aa->aa_exts);
2443 spin_lock(&cli->cl_loi_list_lock);
2444 starting_offset >>= PAGE_SHIFT;
2445 if (cmd == OBD_BRW_READ) {
2446 cli->cl_r_in_flight++;
2447 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2448 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2449 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2450 starting_offset + 1);
2452 cli->cl_w_in_flight++;
2453 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2454 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2455 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2456 starting_offset + 1);
2458 spin_unlock(&cli->cl_loi_list_lock);
2460 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2461 page_count, aa, cli->cl_r_in_flight,
2462 cli->cl_w_in_flight);
2463 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2465 ptlrpcd_add_req(req);
2471 memalloc_noreclaim_restore(mpflag);
2474 LASSERT(req == NULL);
2477 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2479 osc_release_bounce_pages(pga, page_count);
2480 osc_release_ppga(pga, page_count);
2482 /* this should happen rarely and is pretty bad, it makes the
2483 * pending list not follow the dirty order */
2484 while (!list_empty(ext_list)) {
2485 ext = list_entry(ext_list->next, struct osc_extent,
2487 list_del_init(&ext->oe_link);
2488 osc_extent_finish(env, ext, 0, rc);
2494 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2498 LASSERT(lock != NULL);
2500 lock_res_and_lock(lock);
2502 if (lock->l_ast_data == NULL)
2503 lock->l_ast_data = data;
2504 if (lock->l_ast_data == data)
2507 unlock_res_and_lock(lock);
2512 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2513 void *cookie, struct lustre_handle *lockh,
2514 enum ldlm_mode mode, __u64 *flags, bool speculative,
2517 bool intent = *flags & LDLM_FL_HAS_INTENT;
2521 /* The request was created before ldlm_cli_enqueue call. */
2522 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2523 struct ldlm_reply *rep;
2525 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2526 LASSERT(rep != NULL);
2528 rep->lock_policy_res1 =
2529 ptlrpc_status_ntoh(rep->lock_policy_res1);
2530 if (rep->lock_policy_res1)
2531 errcode = rep->lock_policy_res1;
2533 *flags |= LDLM_FL_LVB_READY;
2534 } else if (errcode == ELDLM_OK) {
2535 *flags |= LDLM_FL_LVB_READY;
2538 /* Call the update callback. */
2539 rc = (*upcall)(cookie, lockh, errcode);
2541 /* release the reference taken in ldlm_cli_enqueue() */
2542 if (errcode == ELDLM_LOCK_MATCHED)
2544 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2545 ldlm_lock_decref(lockh, mode);
2550 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2553 struct osc_enqueue_args *aa = args;
2554 struct ldlm_lock *lock;
2555 struct lustre_handle *lockh = &aa->oa_lockh;
2556 enum ldlm_mode mode = aa->oa_mode;
2557 struct ost_lvb *lvb = aa->oa_lvb;
2558 __u32 lvb_len = sizeof(*lvb);
2563 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2565 lock = ldlm_handle2lock(lockh);
2566 LASSERTF(lock != NULL,
2567 "lockh %#llx, req %p, aa %p - client evicted?\n",
2568 lockh->cookie, req, aa);
2570 /* Take an additional reference so that a blocking AST that
2571 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2572 * to arrive after an upcall has been executed by
2573 * osc_enqueue_fini(). */
2574 ldlm_lock_addref(lockh, mode);
2576 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2577 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2579 /* Let CP AST to grant the lock first. */
2580 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2582 if (aa->oa_speculative) {
2583 LASSERT(aa->oa_lvb == NULL);
2584 LASSERT(aa->oa_flags == NULL);
2585 aa->oa_flags = &flags;
2588 /* Complete obtaining the lock procedure. */
2589 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2590 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2592 /* Complete osc stuff. */
2593 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2594 aa->oa_flags, aa->oa_speculative, rc);
2596 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2598 ldlm_lock_decref(lockh, mode);
2599 LDLM_LOCK_PUT(lock);
2603 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2604 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2605 * other synchronous requests, however keeping some locks and trying to obtain
2606 * others may take a considerable amount of time in a case of ost failure; and
2607 * when other sync requests do not get released lock from a client, the client
2608 * is evicted from the cluster -- such scenarious make the life difficult, so
2609 * release locks just after they are obtained. */
2610 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2611 __u64 *flags, union ldlm_policy_data *policy,
2612 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2613 void *cookie, struct ldlm_enqueue_info *einfo,
2614 struct ptlrpc_request_set *rqset, int async,
2617 struct obd_device *obd = exp->exp_obd;
2618 struct lustre_handle lockh = { 0 };
2619 struct ptlrpc_request *req = NULL;
2620 int intent = *flags & LDLM_FL_HAS_INTENT;
2621 __u64 match_flags = *flags;
2622 enum ldlm_mode mode;
2626 /* Filesystem lock extents are extended to page boundaries so that
2627 * dealing with the page cache is a little smoother. */
2628 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2629 policy->l_extent.end |= ~PAGE_MASK;
2631 /* Next, search for already existing extent locks that will cover us */
2632 /* If we're trying to read, we also search for an existing PW lock. The
2633 * VFS and page cache already protect us locally, so lots of readers/
2634 * writers can share a single PW lock.
2636 * There are problems with conversion deadlocks, so instead of
2637 * converting a read lock to a write lock, we'll just enqueue a new
2640 * At some point we should cancel the read lock instead of making them
2641 * send us a blocking callback, but there are problems with canceling
2642 * locks out from other users right now, too. */
2643 mode = einfo->ei_mode;
2644 if (einfo->ei_mode == LCK_PR)
2646 /* Normal lock requests must wait for the LVB to be ready before
2647 * matching a lock; speculative lock requests do not need to,
2648 * because they will not actually use the lock. */
2650 match_flags |= LDLM_FL_LVB_READY;
2652 match_flags |= LDLM_FL_BLOCK_GRANTED;
2653 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2654 einfo->ei_type, policy, mode, &lockh, 0);
2656 struct ldlm_lock *matched;
2658 if (*flags & LDLM_FL_TEST_LOCK)
2661 matched = ldlm_handle2lock(&lockh);
2663 /* This DLM lock request is speculative, and does not
2664 * have an associated IO request. Therefore if there
2665 * is already a DLM lock, it wll just inform the
2666 * caller to cancel the request for this stripe.*/
2667 lock_res_and_lock(matched);
2668 if (ldlm_extent_equal(&policy->l_extent,
2669 &matched->l_policy_data.l_extent))
2673 unlock_res_and_lock(matched);
2675 ldlm_lock_decref(&lockh, mode);
2676 LDLM_LOCK_PUT(matched);
2678 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2679 *flags |= LDLM_FL_LVB_READY;
2681 /* We already have a lock, and it's referenced. */
2682 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2684 ldlm_lock_decref(&lockh, mode);
2685 LDLM_LOCK_PUT(matched);
2688 ldlm_lock_decref(&lockh, mode);
2689 LDLM_LOCK_PUT(matched);
2693 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2697 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2698 &RQF_LDLM_ENQUEUE_LVB);
2702 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2704 ptlrpc_request_free(req);
2708 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2710 ptlrpc_request_set_replen(req);
2713 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2714 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2716 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2717 sizeof(*lvb), LVB_T_OST, &lockh, async);
2720 struct osc_enqueue_args *aa;
2721 aa = ptlrpc_req_async_args(aa, req);
2723 aa->oa_mode = einfo->ei_mode;
2724 aa->oa_type = einfo->ei_type;
2725 lustre_handle_copy(&aa->oa_lockh, &lockh);
2726 aa->oa_upcall = upcall;
2727 aa->oa_cookie = cookie;
2728 aa->oa_speculative = speculative;
2730 aa->oa_flags = flags;
2733 /* speculative locks are essentially to enqueue
2734 * a DLM lock in advance, so we don't care
2735 * about the result of the enqueue. */
2737 aa->oa_flags = NULL;
2740 req->rq_interpret_reply = osc_enqueue_interpret;
2741 ptlrpc_set_add_req(rqset, req);
2742 } else if (intent) {
2743 ptlrpc_req_finished(req);
2748 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2749 flags, speculative, rc);
2751 ptlrpc_req_finished(req);
2756 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2757 struct ldlm_res_id *res_id, enum ldlm_type type,
2758 union ldlm_policy_data *policy, enum ldlm_mode mode,
2759 __u64 *flags, struct osc_object *obj,
2760 struct lustre_handle *lockh, int unref)
2762 struct obd_device *obd = exp->exp_obd;
2763 __u64 lflags = *flags;
2767 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2770 /* Filesystem lock extents are extended to page boundaries so that
2771 * dealing with the page cache is a little smoother */
2772 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2773 policy->l_extent.end |= ~PAGE_MASK;
2775 /* Next, search for already existing extent locks that will cover us */
2776 /* If we're trying to read, we also search for an existing PW lock. The
2777 * VFS and page cache already protect us locally, so lots of readers/
2778 * writers can share a single PW lock. */
2782 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2783 res_id, type, policy, rc, lockh, unref);
2784 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2788 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2790 LASSERT(lock != NULL);
2791 if (osc_set_lock_data(lock, obj)) {
2792 lock_res_and_lock(lock);
2793 if (!ldlm_is_lvb_cached(lock)) {
2794 LASSERT(lock->l_ast_data == obj);
2795 osc_lock_lvb_update(env, obj, lock, NULL);
2796 ldlm_set_lvb_cached(lock);
2798 unlock_res_and_lock(lock);
2800 ldlm_lock_decref(lockh, rc);
2803 LDLM_LOCK_PUT(lock);
2808 static int osc_statfs_interpret(const struct lu_env *env,
2809 struct ptlrpc_request *req, void *args, int rc)
2811 struct osc_async_args *aa = args;
2812 struct obd_statfs *msfs;
2817 * The request has in fact never been sent due to issues at
2818 * a higher level (LOV). Exit immediately since the caller
2819 * is aware of the problem and takes care of the clean up.
2823 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2824 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2830 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2832 GOTO(out, rc = -EPROTO);
2834 *aa->aa_oi->oi_osfs = *msfs;
2836 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2841 static int osc_statfs_async(struct obd_export *exp,
2842 struct obd_info *oinfo, time64_t max_age,
2843 struct ptlrpc_request_set *rqset)
2845 struct obd_device *obd = class_exp2obd(exp);
2846 struct ptlrpc_request *req;
2847 struct osc_async_args *aa;
2851 if (obd->obd_osfs_age >= max_age) {
2853 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2854 obd->obd_name, &obd->obd_osfs,
2855 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2856 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2857 spin_lock(&obd->obd_osfs_lock);
2858 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2859 spin_unlock(&obd->obd_osfs_lock);
2860 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2861 if (oinfo->oi_cb_up)
2862 oinfo->oi_cb_up(oinfo, 0);
2867 /* We could possibly pass max_age in the request (as an absolute
2868 * timestamp or a "seconds.usec ago") so the target can avoid doing
2869 * extra calls into the filesystem if that isn't necessary (e.g.
2870 * during mount that would help a bit). Having relative timestamps
2871 * is not so great if request processing is slow, while absolute
2872 * timestamps are not ideal because they need time synchronization. */
2873 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2877 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2879 ptlrpc_request_free(req);
2882 ptlrpc_request_set_replen(req);
2883 req->rq_request_portal = OST_CREATE_PORTAL;
2884 ptlrpc_at_set_req_timeout(req);
2886 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2887 /* procfs requests not want stat in wait for avoid deadlock */
2888 req->rq_no_resend = 1;
2889 req->rq_no_delay = 1;
2892 req->rq_interpret_reply = osc_statfs_interpret;
2893 aa = ptlrpc_req_async_args(aa, req);
2896 ptlrpc_set_add_req(rqset, req);
2900 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2901 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2903 struct obd_device *obd = class_exp2obd(exp);
2904 struct obd_statfs *msfs;
2905 struct ptlrpc_request *req;
2906 struct obd_import *imp = NULL;
2911 /*Since the request might also come from lprocfs, so we need
2912 *sync this with client_disconnect_export Bug15684*/
2913 down_read(&obd->u.cli.cl_sem);
2914 if (obd->u.cli.cl_import)
2915 imp = class_import_get(obd->u.cli.cl_import);
2916 up_read(&obd->u.cli.cl_sem);
2920 /* We could possibly pass max_age in the request (as an absolute
2921 * timestamp or a "seconds.usec ago") so the target can avoid doing
2922 * extra calls into the filesystem if that isn't necessary (e.g.
2923 * during mount that would help a bit). Having relative timestamps
2924 * is not so great if request processing is slow, while absolute
2925 * timestamps are not ideal because they need time synchronization. */
2926 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2928 class_import_put(imp);
2933 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2935 ptlrpc_request_free(req);
2938 ptlrpc_request_set_replen(req);
2939 req->rq_request_portal = OST_CREATE_PORTAL;
2940 ptlrpc_at_set_req_timeout(req);
2942 if (flags & OBD_STATFS_NODELAY) {
2943 /* procfs requests not want stat in wait for avoid deadlock */
2944 req->rq_no_resend = 1;
2945 req->rq_no_delay = 1;
2948 rc = ptlrpc_queue_wait(req);
2952 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2954 GOTO(out, rc = -EPROTO);
2960 ptlrpc_req_finished(req);
2964 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2965 void *karg, void __user *uarg)
2967 struct obd_device *obd = exp->exp_obd;
2968 struct obd_ioctl_data *data = karg;
2972 if (!try_module_get(THIS_MODULE)) {
2973 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2974 module_name(THIS_MODULE));
2978 case OBD_IOC_CLIENT_RECOVER:
2979 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
2980 data->ioc_inlbuf1, 0);
2984 case IOC_OSC_SET_ACTIVE:
2985 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
2990 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
2991 obd->obd_name, cmd, current->comm, rc);
2995 module_put(THIS_MODULE);
2999 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3000 u32 keylen, void *key, u32 vallen, void *val,
3001 struct ptlrpc_request_set *set)
3003 struct ptlrpc_request *req;
3004 struct obd_device *obd = exp->exp_obd;
3005 struct obd_import *imp = class_exp2cliimp(exp);
3010 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3012 if (KEY_IS(KEY_CHECKSUM)) {
3013 if (vallen != sizeof(int))
3015 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3019 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3020 sptlrpc_conf_client_adapt(obd);
3024 if (KEY_IS(KEY_FLUSH_CTX)) {
3025 sptlrpc_import_flush_my_ctx(imp);
3029 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3030 struct client_obd *cli = &obd->u.cli;
3031 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3032 long target = *(long *)val;
3034 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3039 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3042 /* We pass all other commands directly to OST. Since nobody calls osc
3043 methods directly and everybody is supposed to go through LOV, we
3044 assume lov checked invalid values for us.
3045 The only recognised values so far are evict_by_nid and mds_conn.
3046 Even if something bad goes through, we'd get a -EINVAL from OST
3049 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3050 &RQF_OST_SET_GRANT_INFO :
3055 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3056 RCL_CLIENT, keylen);
3057 if (!KEY_IS(KEY_GRANT_SHRINK))
3058 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3059 RCL_CLIENT, vallen);
3060 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3062 ptlrpc_request_free(req);
3066 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3067 memcpy(tmp, key, keylen);
3068 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3071 memcpy(tmp, val, vallen);
3073 if (KEY_IS(KEY_GRANT_SHRINK)) {
3074 struct osc_grant_args *aa;
3077 aa = ptlrpc_req_async_args(aa, req);
3078 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3080 ptlrpc_req_finished(req);
3083 *oa = ((struct ost_body *)val)->oa;
3085 req->rq_interpret_reply = osc_shrink_grant_interpret;
3088 ptlrpc_request_set_replen(req);
3089 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3090 LASSERT(set != NULL);
3091 ptlrpc_set_add_req(set, req);
3092 ptlrpc_check_set(NULL, set);
3094 ptlrpcd_add_req(req);
3099 EXPORT_SYMBOL(osc_set_info_async);
3101 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3102 struct obd_device *obd, struct obd_uuid *cluuid,
3103 struct obd_connect_data *data, void *localdata)
3105 struct client_obd *cli = &obd->u.cli;
3107 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3111 spin_lock(&cli->cl_loi_list_lock);
3112 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3113 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3114 /* restore ocd_grant_blkbits as client page bits */
3115 data->ocd_grant_blkbits = PAGE_SHIFT;
3116 grant += cli->cl_dirty_grant;
3118 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3120 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3121 lost_grant = cli->cl_lost_grant;
3122 cli->cl_lost_grant = 0;
3123 spin_unlock(&cli->cl_loi_list_lock);
3125 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3126 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3127 data->ocd_version, data->ocd_grant, lost_grant);
3132 EXPORT_SYMBOL(osc_reconnect);
3134 int osc_disconnect(struct obd_export *exp)
3136 struct obd_device *obd = class_exp2obd(exp);
3139 rc = client_disconnect_export(exp);
3141 * Initially we put del_shrink_grant before disconnect_export, but it
3142 * causes the following problem if setup (connect) and cleanup
3143 * (disconnect) are tangled together.
3144 * connect p1 disconnect p2
3145 * ptlrpc_connect_import
3146 * ............... class_manual_cleanup
3149 * ptlrpc_connect_interrupt
3151 * add this client to shrink list
3153 * Bang! grant shrink thread trigger the shrink. BUG18662
3155 osc_del_grant_list(&obd->u.cli);
3158 EXPORT_SYMBOL(osc_disconnect);
3160 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3161 struct hlist_node *hnode, void *arg)
3163 struct lu_env *env = arg;
3164 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3165 struct ldlm_lock *lock;
3166 struct osc_object *osc = NULL;
3170 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3171 if (lock->l_ast_data != NULL && osc == NULL) {
3172 osc = lock->l_ast_data;
3173 cl_object_get(osc2cl(osc));
3176 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3177 * by the 2nd round of ldlm_namespace_clean() call in
3178 * osc_import_event(). */
3179 ldlm_clear_cleaned(lock);
3184 osc_object_invalidate(env, osc);
3185 cl_object_put(env, osc2cl(osc));
3190 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3192 static int osc_import_event(struct obd_device *obd,
3193 struct obd_import *imp,
3194 enum obd_import_event event)
3196 struct client_obd *cli;
3200 LASSERT(imp->imp_obd == obd);
3203 case IMP_EVENT_DISCON: {
3205 spin_lock(&cli->cl_loi_list_lock);
3206 cli->cl_avail_grant = 0;
3207 cli->cl_lost_grant = 0;
3208 spin_unlock(&cli->cl_loi_list_lock);
3211 case IMP_EVENT_INACTIVE: {
3212 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3215 case IMP_EVENT_INVALIDATE: {
3216 struct ldlm_namespace *ns = obd->obd_namespace;
3220 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3222 env = cl_env_get(&refcheck);
3224 osc_io_unplug(env, &obd->u.cli, NULL);
3226 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3227 osc_ldlm_resource_invalidate,
3229 cl_env_put(env, &refcheck);
3231 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3236 case IMP_EVENT_ACTIVE: {
3237 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3240 case IMP_EVENT_OCD: {
3241 struct obd_connect_data *ocd = &imp->imp_connect_data;
3243 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3244 osc_init_grant(&obd->u.cli, ocd);
3247 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3248 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3250 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3253 case IMP_EVENT_DEACTIVATE: {
3254 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3257 case IMP_EVENT_ACTIVATE: {
3258 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3262 CERROR("Unknown import event %d\n", event);
3269 * Determine whether the lock can be canceled before replaying the lock
3270 * during recovery, see bug16774 for detailed information.
3272 * \retval zero the lock can't be canceled
3273 * \retval other ok to cancel
3275 static int osc_cancel_weight(struct ldlm_lock *lock)
3278 * Cancel all unused and granted extent lock.
3280 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3281 ldlm_is_granted(lock) &&
3282 osc_ldlm_weigh_ast(lock) == 0)
3288 static int brw_queue_work(const struct lu_env *env, void *data)
3290 struct client_obd *cli = data;
3292 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3294 osc_io_unplug(env, cli, NULL);
3298 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3300 struct client_obd *cli = &obd->u.cli;
3306 rc = ptlrpcd_addref();
3310 rc = client_obd_setup(obd, lcfg);
3312 GOTO(out_ptlrpcd, rc);
3315 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3316 if (IS_ERR(handler))
3317 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3318 cli->cl_writeback_work = handler;
3320 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3321 if (IS_ERR(handler))
3322 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3323 cli->cl_lru_work = handler;
3325 rc = osc_quota_setup(obd);
3327 GOTO(out_ptlrpcd_work, rc);
3329 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3330 osc_update_next_shrink(cli);
3335 if (cli->cl_writeback_work != NULL) {
3336 ptlrpcd_destroy_work(cli->cl_writeback_work);
3337 cli->cl_writeback_work = NULL;
3339 if (cli->cl_lru_work != NULL) {
3340 ptlrpcd_destroy_work(cli->cl_lru_work);
3341 cli->cl_lru_work = NULL;
3343 client_obd_cleanup(obd);
3348 EXPORT_SYMBOL(osc_setup_common);
3350 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3352 struct client_obd *cli = &obd->u.cli;
3360 rc = osc_setup_common(obd, lcfg);
3364 rc = osc_tunables_init(obd);
3369 * We try to control the total number of requests with a upper limit
3370 * osc_reqpool_maxreqcount. There might be some race which will cause
3371 * over-limit allocation, but it is fine.
3373 req_count = atomic_read(&osc_pool_req_count);
3374 if (req_count < osc_reqpool_maxreqcount) {
3375 adding = cli->cl_max_rpcs_in_flight + 2;
3376 if (req_count + adding > osc_reqpool_maxreqcount)
3377 adding = osc_reqpool_maxreqcount - req_count;
3379 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3380 atomic_add(added, &osc_pool_req_count);
3383 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3385 spin_lock(&osc_shrink_lock);
3386 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3387 spin_unlock(&osc_shrink_lock);
3388 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3389 cli->cl_import->imp_idle_debug = D_HA;
3394 int osc_precleanup_common(struct obd_device *obd)
3396 struct client_obd *cli = &obd->u.cli;
3400 * for echo client, export may be on zombie list, wait for
3401 * zombie thread to cull it, because cli.cl_import will be
3402 * cleared in client_disconnect_export():
3403 * class_export_destroy() -> obd_cleanup() ->
3404 * echo_device_free() -> echo_client_cleanup() ->
3405 * obd_disconnect() -> osc_disconnect() ->
3406 * client_disconnect_export()
3408 obd_zombie_barrier();
3409 if (cli->cl_writeback_work) {
3410 ptlrpcd_destroy_work(cli->cl_writeback_work);
3411 cli->cl_writeback_work = NULL;
3414 if (cli->cl_lru_work) {
3415 ptlrpcd_destroy_work(cli->cl_lru_work);
3416 cli->cl_lru_work = NULL;
3419 obd_cleanup_client_import(obd);
3422 EXPORT_SYMBOL(osc_precleanup_common);
3424 static int osc_precleanup(struct obd_device *obd)
3428 osc_precleanup_common(obd);
3430 ptlrpc_lprocfs_unregister_obd(obd);
3434 int osc_cleanup_common(struct obd_device *obd)
3436 struct client_obd *cli = &obd->u.cli;
3441 spin_lock(&osc_shrink_lock);
3442 list_del(&cli->cl_shrink_list);
3443 spin_unlock(&osc_shrink_lock);
3446 if (cli->cl_cache != NULL) {
3447 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3448 spin_lock(&cli->cl_cache->ccc_lru_lock);
3449 list_del_init(&cli->cl_lru_osc);
3450 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3451 cli->cl_lru_left = NULL;
3452 cl_cache_decref(cli->cl_cache);
3453 cli->cl_cache = NULL;
3456 /* free memory of osc quota cache */
3457 osc_quota_cleanup(obd);
3459 rc = client_obd_cleanup(obd);
3464 EXPORT_SYMBOL(osc_cleanup_common);
3466 static const struct obd_ops osc_obd_ops = {
3467 .o_owner = THIS_MODULE,
3468 .o_setup = osc_setup,
3469 .o_precleanup = osc_precleanup,
3470 .o_cleanup = osc_cleanup_common,
3471 .o_add_conn = client_import_add_conn,
3472 .o_del_conn = client_import_del_conn,
3473 .o_connect = client_connect_import,
3474 .o_reconnect = osc_reconnect,
3475 .o_disconnect = osc_disconnect,
3476 .o_statfs = osc_statfs,
3477 .o_statfs_async = osc_statfs_async,
3478 .o_create = osc_create,
3479 .o_destroy = osc_destroy,
3480 .o_getattr = osc_getattr,
3481 .o_setattr = osc_setattr,
3482 .o_iocontrol = osc_iocontrol,
3483 .o_set_info_async = osc_set_info_async,
3484 .o_import_event = osc_import_event,
3485 .o_quotactl = osc_quotactl,
3488 static struct shrinker *osc_cache_shrinker;
3489 LIST_HEAD(osc_shrink_list);
3490 DEFINE_SPINLOCK(osc_shrink_lock);
3492 #ifndef HAVE_SHRINKER_COUNT
3493 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3495 struct shrink_control scv = {
3496 .nr_to_scan = shrink_param(sc, nr_to_scan),
3497 .gfp_mask = shrink_param(sc, gfp_mask)
3499 (void)osc_cache_shrink_scan(shrinker, &scv);
3501 return osc_cache_shrink_count(shrinker, &scv);
3505 static int __init osc_init(void)
3507 unsigned int reqpool_size;
3508 unsigned int reqsize;
3510 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3511 osc_cache_shrink_count, osc_cache_shrink_scan);
3514 /* print an address of _any_ initialized kernel symbol from this
3515 * module, to allow debugging with gdb that doesn't support data
3516 * symbols from modules.*/
3517 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3519 rc = lu_kmem_init(osc_caches);
3523 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3524 LUSTRE_OSC_NAME, &osc_device_type);
3528 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3530 /* This is obviously too much memory, only prevent overflow here */
3531 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3532 GOTO(out_type, rc = -EINVAL);
3534 reqpool_size = osc_reqpool_mem_max << 20;
3537 while (reqsize < OST_IO_MAXREQSIZE)
3538 reqsize = reqsize << 1;
3541 * We don't enlarge the request count in OSC pool according to
3542 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3543 * tried after normal allocation failed. So a small OSC pool won't
3544 * cause much performance degression in most of cases.
3546 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3548 atomic_set(&osc_pool_req_count, 0);
3549 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3550 ptlrpc_add_rqs_to_pool);
3552 if (osc_rq_pool == NULL)
3553 GOTO(out_type, rc = -ENOMEM);
3555 rc = osc_start_grant_work();
3557 GOTO(out_req_pool, rc);
3562 ptlrpc_free_rq_pool(osc_rq_pool);
3564 class_unregister_type(LUSTRE_OSC_NAME);
3566 lu_kmem_fini(osc_caches);
3571 static void __exit osc_exit(void)
3573 osc_stop_grant_work();
3574 remove_shrinker(osc_cache_shrinker);
3575 class_unregister_type(LUSTRE_OSC_NAME);
3576 lu_kmem_fini(osc_caches);
3577 ptlrpc_free_rq_pool(osc_rq_pool);
3580 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3581 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3582 MODULE_VERSION(LUSTRE_VERSION_STRING);
3583 MODULE_LICENSE("GPL");
3585 module_init(osc_init);
3586 module_exit(osc_exit);