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 */
1359 llcrypt_finalize_bounce_page(&pga[i]->pg);
1360 pga[i]->count -= pga[i]->bp_count_diff;
1361 pga[i]->off += pga[i]->bp_off_diff;
1367 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1368 u32 page_count, struct brw_page **pga,
1369 struct ptlrpc_request **reqp, int resend)
1371 struct ptlrpc_request *req;
1372 struct ptlrpc_bulk_desc *desc;
1373 struct ost_body *body;
1374 struct obd_ioobj *ioobj;
1375 struct niobuf_remote *niobuf;
1376 int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1377 struct osc_brw_async_args *aa;
1378 struct req_capsule *pill;
1379 struct brw_page *pg_prev;
1381 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1382 struct inode *inode;
1385 inode = page2inode(pga[0]->pg);
1386 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1387 RETURN(-ENOMEM); /* Recoverable */
1388 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1389 RETURN(-EINVAL); /* Fatal */
1391 if ((cmd & OBD_BRW_WRITE) != 0) {
1393 req = ptlrpc_request_alloc_pool(cli->cl_import,
1395 &RQF_OST_BRW_WRITE);
1398 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1403 if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1404 for (i = 0; i < page_count; i++) {
1405 struct brw_page *pg = pga[i];
1406 struct page *data_page = NULL;
1407 bool retried = false;
1408 bool lockedbymyself;
1411 /* The page can already be locked when we arrive here.
1412 * This is possible when cl_page_assume/vvp_page_assume
1413 * is stuck on wait_on_page_writeback with page lock
1414 * held. In this case there is no risk for the lock to
1415 * be released while we are doing our encryption
1416 * processing, because writeback against that page will
1417 * end in vvp_page_completion_write/cl_page_completion,
1418 * which means only once the page is fully processed.
1420 lockedbymyself = trylock_page(pg->pg);
1422 llcrypt_encrypt_pagecache_blocks(pg->pg,
1426 unlock_page(pg->pg);
1427 if (IS_ERR(data_page)) {
1428 rc = PTR_ERR(data_page);
1429 if (rc == -ENOMEM && !retried) {
1434 ptlrpc_request_free(req);
1438 /* there should be no gap in the middle of page array */
1439 if (i == page_count - 1) {
1440 struct osc_async_page *oap = brw_page2oap(pg);
1442 oa->o_size = oap->oap_count +
1443 oap->oap_obj_off + oap->oap_page_off;
1445 /* len is forced to PAGE_SIZE, and poff to 0
1446 * so store the old, clear text info
1448 pg->bp_count_diff = PAGE_SIZE - pg->count;
1449 pg->count = PAGE_SIZE;
1450 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1451 pg->off = pg->off & PAGE_MASK;
1453 } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode)) {
1454 for (i = 0; i < page_count; i++) {
1455 struct brw_page *pg = pga[i];
1457 /* count/off are forced to cover the whole page so that
1458 * all encrypted data is stored on the OST, so adjust
1459 * bp_{count,off}_diff for the size of the clear text.
1461 pg->bp_count_diff = PAGE_SIZE - pg->count;
1462 pg->count = PAGE_SIZE;
1463 pg->bp_off_diff = pg->off & ~PAGE_MASK;
1464 pg->off = pg->off & PAGE_MASK;
1468 for (niocount = i = 1; i < page_count; i++) {
1469 if (!can_merge_pages(pga[i - 1], pga[i]))
1473 pill = &req->rq_pill;
1474 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1476 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1477 niocount * sizeof(*niobuf));
1479 for (i = 0; i < page_count; i++) {
1480 short_io_size += pga[i]->count;
1481 if (!inode || !IS_ENCRYPTED(inode)) {
1482 pga[i]->bp_count_diff = 0;
1483 pga[i]->bp_off_diff = 0;
1487 /* Check if read/write is small enough to be a short io. */
1488 if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1489 !imp_connect_shortio(cli->cl_import))
1492 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1493 opc == OST_READ ? 0 : short_io_size);
1494 if (opc == OST_READ)
1495 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1498 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1500 ptlrpc_request_free(req);
1503 osc_set_io_portal(req);
1505 ptlrpc_at_set_req_timeout(req);
1506 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1508 req->rq_no_retry_einprogress = 1;
1510 if (short_io_size != 0) {
1512 short_io_buf = NULL;
1516 desc = ptlrpc_prep_bulk_imp(req, page_count,
1517 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1518 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1519 PTLRPC_BULK_PUT_SINK),
1521 &ptlrpc_bulk_kiov_pin_ops);
1524 GOTO(out, rc = -ENOMEM);
1525 /* NB request now owns desc and will free it when it gets freed */
1527 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1528 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1529 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1530 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1532 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1534 /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1535 * and from_kgid(), because they are asynchronous. Fortunately, variable
1536 * oa contains valid o_uid and o_gid in these two operations.
1537 * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1538 * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1539 * other process logic */
1540 body->oa.o_uid = oa->o_uid;
1541 body->oa.o_gid = oa->o_gid;
1543 obdo_to_ioobj(oa, ioobj);
1544 ioobj->ioo_bufcnt = niocount;
1545 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1546 * that might be send for this request. The actual number is decided
1547 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1548 * "max - 1" for old client compatibility sending "0", and also so the
1549 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1551 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1553 ioobj_max_brw_set(ioobj, 0);
1555 if (short_io_size != 0) {
1556 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1557 body->oa.o_valid |= OBD_MD_FLFLAGS;
1558 body->oa.o_flags = 0;
1560 body->oa.o_flags |= OBD_FL_SHORT_IO;
1561 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1563 if (opc == OST_WRITE) {
1564 short_io_buf = req_capsule_client_get(pill,
1566 LASSERT(short_io_buf != NULL);
1570 LASSERT(page_count > 0);
1572 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1573 struct brw_page *pg = pga[i];
1574 int poff = pg->off & ~PAGE_MASK;
1576 LASSERT(pg->count > 0);
1577 /* make sure there is no gap in the middle of page array */
1578 LASSERTF(page_count == 1 ||
1579 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1580 ergo(i > 0 && i < page_count - 1,
1581 poff == 0 && pg->count == PAGE_SIZE) &&
1582 ergo(i == page_count - 1, poff == 0)),
1583 "i: %d/%d pg: %p off: %llu, count: %u\n",
1584 i, page_count, pg, pg->off, pg->count);
1585 LASSERTF(i == 0 || pg->off > pg_prev->off,
1586 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1587 " prev_pg %p [pri %lu ind %lu] off %llu\n",
1589 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1590 pg_prev->pg, page_private(pg_prev->pg),
1591 pg_prev->pg->index, pg_prev->off);
1592 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1593 (pg->flag & OBD_BRW_SRVLOCK));
1594 if (short_io_size != 0 && opc == OST_WRITE) {
1595 unsigned char *ptr = kmap_atomic(pg->pg);
1597 LASSERT(short_io_size >= requested_nob + pg->count);
1598 memcpy(short_io_buf + requested_nob,
1602 } else if (short_io_size == 0) {
1603 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1606 requested_nob += pg->count;
1608 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1610 niobuf->rnb_len += pg->count;
1612 niobuf->rnb_offset = pg->off;
1613 niobuf->rnb_len = pg->count;
1614 niobuf->rnb_flags = pg->flag;
1619 LASSERTF((void *)(niobuf - niocount) ==
1620 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1621 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1622 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1624 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1626 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1627 body->oa.o_valid |= OBD_MD_FLFLAGS;
1628 body->oa.o_flags = 0;
1630 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1633 if (osc_should_shrink_grant(cli))
1634 osc_shrink_grant_local(cli, &body->oa);
1636 /* size[REQ_REC_OFF] still sizeof (*body) */
1637 if (opc == OST_WRITE) {
1638 if (cli->cl_checksum &&
1639 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1640 /* store cl_cksum_type in a local variable since
1641 * it can be changed via lprocfs */
1642 enum cksum_types cksum_type = cli->cl_cksum_type;
1644 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1645 body->oa.o_flags = 0;
1647 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1649 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1651 rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1652 requested_nob, page_count,
1656 CDEBUG(D_PAGE, "failed to checksum, rc = %d\n",
1660 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1663 /* save this in 'oa', too, for later checking */
1664 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1665 oa->o_flags |= obd_cksum_type_pack(obd_name,
1668 /* clear out the checksum flag, in case this is a
1669 * resend but cl_checksum is no longer set. b=11238 */
1670 oa->o_valid &= ~OBD_MD_FLCKSUM;
1672 oa->o_cksum = body->oa.o_cksum;
1673 /* 1 RC per niobuf */
1674 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1675 sizeof(__u32) * niocount);
1677 if (cli->cl_checksum &&
1678 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1679 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1680 body->oa.o_flags = 0;
1681 body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1682 cli->cl_cksum_type);
1683 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1686 /* Client cksum has been already copied to wire obdo in previous
1687 * lustre_set_wire_obdo(), and in the case a bulk-read is being
1688 * resent due to cksum error, this will allow Server to
1689 * check+dump pages on its side */
1691 ptlrpc_request_set_replen(req);
1693 aa = ptlrpc_req_async_args(aa, req);
1695 aa->aa_requested_nob = requested_nob;
1696 aa->aa_nio_count = niocount;
1697 aa->aa_page_count = page_count;
1701 INIT_LIST_HEAD(&aa->aa_oaps);
1704 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1705 CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1706 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1707 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1711 ptlrpc_req_finished(req);
1715 char dbgcksum_file_name[PATH_MAX];
1717 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1718 struct brw_page **pga, __u32 server_cksum,
1726 /* will only keep dump of pages on first error for the same range in
1727 * file/fid, not during the resends/retries. */
1728 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1729 "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1730 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1731 libcfs_debug_file_path_arr :
1732 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1733 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1734 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1735 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1737 pga[page_count-1]->off + pga[page_count-1]->count - 1,
1738 client_cksum, server_cksum);
1739 filp = filp_open(dbgcksum_file_name,
1740 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1744 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1745 "checksum error: rc = %d\n", dbgcksum_file_name,
1748 CERROR("%s: can't open to dump pages with checksum "
1749 "error: rc = %d\n", dbgcksum_file_name, rc);
1753 for (i = 0; i < page_count; i++) {
1754 len = pga[i]->count;
1755 buf = kmap(pga[i]->pg);
1757 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1759 CERROR("%s: wanted to write %u but got %d "
1760 "error\n", dbgcksum_file_name, len, rc);
1765 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1766 dbgcksum_file_name, rc);
1771 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1773 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1774 filp_close(filp, NULL);
1778 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1779 __u32 client_cksum, __u32 server_cksum,
1780 struct osc_brw_async_args *aa)
1782 const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1783 enum cksum_types cksum_type;
1784 obd_dif_csum_fn *fn = NULL;
1785 int sector_size = 0;
1790 if (server_cksum == client_cksum) {
1791 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1795 if (aa->aa_cli->cl_checksum_dump)
1796 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1797 server_cksum, client_cksum);
1799 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1802 switch (cksum_type) {
1803 case OBD_CKSUM_T10IP512:
1807 case OBD_CKSUM_T10IP4K:
1811 case OBD_CKSUM_T10CRC512:
1812 fn = obd_dif_crc_fn;
1815 case OBD_CKSUM_T10CRC4K:
1816 fn = obd_dif_crc_fn;
1824 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1825 aa->aa_page_count, aa->aa_ppga,
1826 OST_WRITE, fn, sector_size,
1829 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1830 aa->aa_ppga, OST_WRITE, cksum_type,
1834 msg = "failed to calculate the client write checksum";
1835 else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1836 msg = "the server did not use the checksum type specified in "
1837 "the original request - likely a protocol problem";
1838 else if (new_cksum == server_cksum)
1839 msg = "changed on the client after we checksummed it - "
1840 "likely false positive due to mmap IO (bug 11742)";
1841 else if (new_cksum == client_cksum)
1842 msg = "changed in transit before arrival at OST";
1844 msg = "changed in transit AND doesn't match the original - "
1845 "likely false positive due to mmap IO (bug 11742)";
1847 LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1848 DFID " object "DOSTID" extent [%llu-%llu], original "
1849 "client csum %x (type %x), server csum %x (type %x),"
1850 " client csum now %x\n",
1851 obd_name, msg, libcfs_nid2str(peer->nid),
1852 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1853 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1854 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1855 POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1856 aa->aa_ppga[aa->aa_page_count - 1]->off +
1857 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1859 obd_cksum_type_unpack(aa->aa_oa->o_flags),
1860 server_cksum, cksum_type, new_cksum);
1864 /* Note rc enters this function as number of bytes transferred */
1865 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1867 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1868 struct client_obd *cli = aa->aa_cli;
1869 const char *obd_name = cli->cl_import->imp_obd->obd_name;
1870 const struct lnet_process_id *peer =
1871 &req->rq_import->imp_connection->c_peer;
1872 struct ost_body *body;
1873 u32 client_cksum = 0;
1874 struct inode *inode;
1878 if (rc < 0 && rc != -EDQUOT) {
1879 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1883 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1884 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1886 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1890 /* set/clear over quota flag for a uid/gid/projid */
1891 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1892 body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1893 unsigned qid[LL_MAXQUOTAS] = {
1894 body->oa.o_uid, body->oa.o_gid,
1895 body->oa.o_projid };
1897 "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1898 body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1899 body->oa.o_valid, body->oa.o_flags);
1900 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1904 osc_update_grant(cli, body);
1909 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1910 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1912 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1914 CERROR("%s: unexpected positive size %d\n",
1919 if (req->rq_bulk != NULL &&
1920 sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1923 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1924 check_write_checksum(&body->oa, peer, client_cksum,
1925 body->oa.o_cksum, aa))
1928 rc = check_write_rcs(req, aa->aa_requested_nob,
1929 aa->aa_nio_count, aa->aa_page_count,
1934 /* The rest of this function executes only for OST_READs */
1936 if (req->rq_bulk == NULL) {
1937 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
1939 LASSERT(rc == req->rq_status);
1941 /* if unwrap_bulk failed, return -EAGAIN to retry */
1942 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1945 GOTO(out, rc = -EAGAIN);
1947 if (rc > aa->aa_requested_nob) {
1948 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
1949 rc, aa->aa_requested_nob);
1953 if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
1954 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
1955 rc, req->rq_bulk->bd_nob_transferred);
1959 if (req->rq_bulk == NULL) {
1961 int nob, pg_count, i = 0;
1964 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
1965 pg_count = aa->aa_page_count;
1966 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
1969 while (nob > 0 && pg_count > 0) {
1971 int count = aa->aa_ppga[i]->count > nob ?
1972 nob : aa->aa_ppga[i]->count;
1974 CDEBUG(D_CACHE, "page %p count %d\n",
1975 aa->aa_ppga[i]->pg, count);
1976 ptr = kmap_atomic(aa->aa_ppga[i]->pg);
1977 memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
1979 kunmap_atomic((void *) ptr);
1988 if (rc < aa->aa_requested_nob)
1989 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1991 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1992 static int cksum_counter;
1993 u32 server_cksum = body->oa.o_cksum;
1996 enum cksum_types cksum_type;
1997 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
1998 body->oa.o_flags : 0;
2000 cksum_type = obd_cksum_type_unpack(o_flags);
2001 rc = osc_checksum_bulk_rw(obd_name, cksum_type, rc,
2002 aa->aa_page_count, aa->aa_ppga,
2003 OST_READ, &client_cksum);
2007 if (req->rq_bulk != NULL &&
2008 peer->nid != req->rq_bulk->bd_sender) {
2010 router = libcfs_nid2str(req->rq_bulk->bd_sender);
2013 if (server_cksum != client_cksum) {
2014 struct ost_body *clbody;
2015 u32 page_count = aa->aa_page_count;
2017 clbody = req_capsule_client_get(&req->rq_pill,
2019 if (cli->cl_checksum_dump)
2020 dump_all_bulk_pages(&clbody->oa, page_count,
2021 aa->aa_ppga, server_cksum,
2024 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2025 "%s%s%s inode "DFID" object "DOSTID
2026 " extent [%llu-%llu], client %x, "
2027 "server %x, cksum_type %x\n",
2029 libcfs_nid2str(peer->nid),
2031 clbody->oa.o_valid & OBD_MD_FLFID ?
2032 clbody->oa.o_parent_seq : 0ULL,
2033 clbody->oa.o_valid & OBD_MD_FLFID ?
2034 clbody->oa.o_parent_oid : 0,
2035 clbody->oa.o_valid & OBD_MD_FLFID ?
2036 clbody->oa.o_parent_ver : 0,
2037 POSTID(&body->oa.o_oi),
2038 aa->aa_ppga[0]->off,
2039 aa->aa_ppga[page_count-1]->off +
2040 aa->aa_ppga[page_count-1]->count - 1,
2041 client_cksum, server_cksum,
2044 aa->aa_oa->o_cksum = client_cksum;
2048 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2051 } else if (unlikely(client_cksum)) {
2052 static int cksum_missed;
2055 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2056 CERROR("%s: checksum %u requested from %s but not sent\n",
2057 obd_name, cksum_missed,
2058 libcfs_nid2str(peer->nid));
2063 inode = page2inode(aa->aa_ppga[0]->pg);
2064 if (inode && IS_ENCRYPTED(inode)) {
2067 if (!llcrypt_has_encryption_key(inode)) {
2068 CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2071 for (idx = 0; idx < aa->aa_page_count; idx++) {
2072 struct brw_page *pg = aa->aa_ppga[idx];
2074 /* do not decrypt if page is all 0s */
2075 if (memchr_inv(page_address(pg->pg), 0,
2076 PAGE_SIZE) == NULL) {
2077 /* if page is empty forward info to upper layers
2078 * (ll_io_zero_page) by clearing PagePrivate2
2080 ClearPagePrivate2(pg->pg);
2084 /* The page is already locked when we arrive here,
2085 * except when we deal with a twisted page for
2086 * specific Direct IO support, in which case
2087 * PageChecked flag is set on page.
2089 if (PageChecked(pg->pg))
2091 rc = llcrypt_decrypt_pagecache_blocks(pg->pg,
2093 if (PageChecked(pg->pg))
2094 unlock_page(pg->pg);
2102 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2103 aa->aa_oa, &body->oa);
2108 static int osc_brw_redo_request(struct ptlrpc_request *request,
2109 struct osc_brw_async_args *aa, int rc)
2111 struct ptlrpc_request *new_req;
2112 struct osc_brw_async_args *new_aa;
2113 struct osc_async_page *oap;
2116 /* The below message is checked in replay-ost-single.sh test_8ae*/
2117 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2118 "redo for recoverable error %d", rc);
2120 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2121 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2122 aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2123 aa->aa_ppga, &new_req, 1);
2127 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2128 if (oap->oap_request != NULL) {
2129 LASSERTF(request == oap->oap_request,
2130 "request %p != oap_request %p\n",
2131 request, oap->oap_request);
2135 * New request takes over pga and oaps from old request.
2136 * Note that copying a list_head doesn't work, need to move it...
2139 new_req->rq_interpret_reply = request->rq_interpret_reply;
2140 new_req->rq_async_args = request->rq_async_args;
2141 new_req->rq_commit_cb = request->rq_commit_cb;
2142 /* cap resend delay to the current request timeout, this is similar to
2143 * what ptlrpc does (see after_reply()) */
2144 if (aa->aa_resends > new_req->rq_timeout)
2145 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2147 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2148 new_req->rq_generation_set = 1;
2149 new_req->rq_import_generation = request->rq_import_generation;
2151 new_aa = ptlrpc_req_async_args(new_aa, new_req);
2153 INIT_LIST_HEAD(&new_aa->aa_oaps);
2154 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2155 INIT_LIST_HEAD(&new_aa->aa_exts);
2156 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2157 new_aa->aa_resends = aa->aa_resends;
2159 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2160 if (oap->oap_request) {
2161 ptlrpc_req_finished(oap->oap_request);
2162 oap->oap_request = ptlrpc_request_addref(new_req);
2166 /* XXX: This code will run into problem if we're going to support
2167 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2168 * and wait for all of them to be finished. We should inherit request
2169 * set from old request. */
2170 ptlrpcd_add_req(new_req);
2172 DEBUG_REQ(D_INFO, new_req, "new request");
2177 * ugh, we want disk allocation on the target to happen in offset order. we'll
2178 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2179 * fine for our small page arrays and doesn't require allocation. its an
2180 * insertion sort that swaps elements that are strides apart, shrinking the
2181 * stride down until its '1' and the array is sorted.
2183 static void sort_brw_pages(struct brw_page **array, int num)
2186 struct brw_page *tmp;
2190 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2195 for (i = stride ; i < num ; i++) {
2198 while (j >= stride && array[j - stride]->off > tmp->off) {
2199 array[j] = array[j - stride];
2204 } while (stride > 1);
2207 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2209 LASSERT(ppga != NULL);
2210 OBD_FREE_PTR_ARRAY(ppga, count);
2213 static int brw_interpret(const struct lu_env *env,
2214 struct ptlrpc_request *req, void *args, int rc)
2216 struct osc_brw_async_args *aa = args;
2217 struct osc_extent *ext;
2218 struct osc_extent *tmp;
2219 struct client_obd *cli = aa->aa_cli;
2220 unsigned long transferred = 0;
2224 rc = osc_brw_fini_request(req, rc);
2225 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2227 /* restore clear text pages */
2228 osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2231 * When server returns -EINPROGRESS, client should always retry
2232 * regardless of the number of times the bulk was resent already.
2234 if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2235 if (req->rq_import_generation !=
2236 req->rq_import->imp_generation) {
2237 CDEBUG(D_HA, "%s: resend cross eviction for object: "
2238 ""DOSTID", rc = %d.\n",
2239 req->rq_import->imp_obd->obd_name,
2240 POSTID(&aa->aa_oa->o_oi), rc);
2241 } else if (rc == -EINPROGRESS ||
2242 client_should_resend(aa->aa_resends, aa->aa_cli)) {
2243 rc = osc_brw_redo_request(req, aa, rc);
2245 CERROR("%s: too many resent retries for object: "
2246 "%llu:%llu, rc = %d.\n",
2247 req->rq_import->imp_obd->obd_name,
2248 POSTID(&aa->aa_oa->o_oi), rc);
2253 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2258 struct obdo *oa = aa->aa_oa;
2259 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2260 unsigned long valid = 0;
2261 struct cl_object *obj;
2262 struct osc_async_page *last;
2264 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2265 obj = osc2cl(last->oap_obj);
2267 cl_object_attr_lock(obj);
2268 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2269 attr->cat_blocks = oa->o_blocks;
2270 valid |= CAT_BLOCKS;
2272 if (oa->o_valid & OBD_MD_FLMTIME) {
2273 attr->cat_mtime = oa->o_mtime;
2276 if (oa->o_valid & OBD_MD_FLATIME) {
2277 attr->cat_atime = oa->o_atime;
2280 if (oa->o_valid & OBD_MD_FLCTIME) {
2281 attr->cat_ctime = oa->o_ctime;
2285 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2286 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2287 loff_t last_off = last->oap_count + last->oap_obj_off +
2290 /* Change file size if this is an out of quota or
2291 * direct IO write and it extends the file size */
2292 if (loi->loi_lvb.lvb_size < last_off) {
2293 attr->cat_size = last_off;
2296 /* Extend KMS if it's not a lockless write */
2297 if (loi->loi_kms < last_off &&
2298 oap2osc_page(last)->ops_srvlock == 0) {
2299 attr->cat_kms = last_off;
2305 cl_object_attr_update(env, obj, attr, valid);
2306 cl_object_attr_unlock(obj);
2308 OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2311 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2312 osc_inc_unstable_pages(req);
2314 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2315 list_del_init(&ext->oe_link);
2316 osc_extent_finish(env, ext, 1,
2317 rc && req->rq_no_delay ? -EWOULDBLOCK : rc);
2319 LASSERT(list_empty(&aa->aa_exts));
2320 LASSERT(list_empty(&aa->aa_oaps));
2322 transferred = (req->rq_bulk == NULL ? /* short io */
2323 aa->aa_requested_nob :
2324 req->rq_bulk->bd_nob_transferred);
2326 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2327 ptlrpc_lprocfs_brw(req, transferred);
2329 spin_lock(&cli->cl_loi_list_lock);
2330 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2331 * is called so we know whether to go to sync BRWs or wait for more
2332 * RPCs to complete */
2333 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2334 cli->cl_w_in_flight--;
2336 cli->cl_r_in_flight--;
2337 osc_wake_cache_waiters(cli);
2338 spin_unlock(&cli->cl_loi_list_lock);
2340 osc_io_unplug(env, cli, NULL);
2344 static void brw_commit(struct ptlrpc_request *req)
2346 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2347 * this called via the rq_commit_cb, I need to ensure
2348 * osc_dec_unstable_pages is still called. Otherwise unstable
2349 * pages may be leaked. */
2350 spin_lock(&req->rq_lock);
2351 if (likely(req->rq_unstable)) {
2352 req->rq_unstable = 0;
2353 spin_unlock(&req->rq_lock);
2355 osc_dec_unstable_pages(req);
2357 req->rq_committed = 1;
2358 spin_unlock(&req->rq_lock);
2363 * Build an RPC by the list of extent @ext_list. The caller must ensure
2364 * that the total pages in this list are NOT over max pages per RPC.
2365 * Extents in the list must be in OES_RPC state.
2367 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2368 struct list_head *ext_list, int cmd)
2370 struct ptlrpc_request *req = NULL;
2371 struct osc_extent *ext;
2372 struct brw_page **pga = NULL;
2373 struct osc_brw_async_args *aa = NULL;
2374 struct obdo *oa = NULL;
2375 struct osc_async_page *oap;
2376 struct osc_object *obj = NULL;
2377 struct cl_req_attr *crattr = NULL;
2378 loff_t starting_offset = OBD_OBJECT_EOF;
2379 loff_t ending_offset = 0;
2380 /* '1' for consistency with code that checks !mpflag to restore */
2384 bool soft_sync = false;
2385 bool ndelay = false;
2389 __u32 layout_version = 0;
2390 LIST_HEAD(rpc_list);
2391 struct ost_body *body;
2393 LASSERT(!list_empty(ext_list));
2395 /* add pages into rpc_list to build BRW rpc */
2396 list_for_each_entry(ext, ext_list, oe_link) {
2397 LASSERT(ext->oe_state == OES_RPC);
2398 mem_tight |= ext->oe_memalloc;
2399 grant += ext->oe_grants;
2400 page_count += ext->oe_nr_pages;
2401 layout_version = max(layout_version, ext->oe_layout_version);
2406 soft_sync = osc_over_unstable_soft_limit(cli);
2408 mpflag = memalloc_noreclaim_save();
2410 OBD_ALLOC_PTR_ARRAY(pga, page_count);
2412 GOTO(out, rc = -ENOMEM);
2414 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2416 GOTO(out, rc = -ENOMEM);
2419 list_for_each_entry(ext, ext_list, oe_link) {
2420 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2422 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2424 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2425 pga[i] = &oap->oap_brw_page;
2426 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2429 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2430 if (starting_offset == OBD_OBJECT_EOF ||
2431 starting_offset > oap->oap_obj_off)
2432 starting_offset = oap->oap_obj_off;
2434 LASSERT(oap->oap_page_off == 0);
2435 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2436 ending_offset = oap->oap_obj_off +
2439 LASSERT(oap->oap_page_off + oap->oap_count ==
2446 /* first page in the list */
2447 oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
2449 crattr = &osc_env_info(env)->oti_req_attr;
2450 memset(crattr, 0, sizeof(*crattr));
2451 crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2452 crattr->cra_flags = ~0ULL;
2453 crattr->cra_page = oap2cl_page(oap);
2454 crattr->cra_oa = oa;
2455 cl_req_attr_set(env, osc2cl(obj), crattr);
2457 if (cmd == OBD_BRW_WRITE) {
2458 oa->o_grant_used = grant;
2459 if (layout_version > 0) {
2460 CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2461 PFID(&oa->o_oi.oi_fid), layout_version);
2463 oa->o_layout_version = layout_version;
2464 oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2468 sort_brw_pages(pga, page_count);
2469 rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2471 CERROR("prep_req failed: %d\n", rc);
2475 req->rq_commit_cb = brw_commit;
2476 req->rq_interpret_reply = brw_interpret;
2477 req->rq_memalloc = mem_tight != 0;
2478 oap->oap_request = ptlrpc_request_addref(req);
2480 req->rq_no_resend = req->rq_no_delay = 1;
2481 /* probably set a shorter timeout value.
2482 * to handle ETIMEDOUT in brw_interpret() correctly. */
2483 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2486 /* Need to update the timestamps after the request is built in case
2487 * we race with setattr (locally or in queue at OST). If OST gets
2488 * later setattr before earlier BRW (as determined by the request xid),
2489 * the OST will not use BRW timestamps. Sadly, there is no obvious
2490 * way to do this in a single call. bug 10150 */
2491 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2492 crattr->cra_oa = &body->oa;
2493 crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2494 cl_req_attr_set(env, osc2cl(obj), crattr);
2495 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2497 aa = ptlrpc_req_async_args(aa, req);
2498 INIT_LIST_HEAD(&aa->aa_oaps);
2499 list_splice_init(&rpc_list, &aa->aa_oaps);
2500 INIT_LIST_HEAD(&aa->aa_exts);
2501 list_splice_init(ext_list, &aa->aa_exts);
2503 spin_lock(&cli->cl_loi_list_lock);
2504 starting_offset >>= PAGE_SHIFT;
2505 if (cmd == OBD_BRW_READ) {
2506 cli->cl_r_in_flight++;
2507 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2508 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2509 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2510 starting_offset + 1);
2512 cli->cl_w_in_flight++;
2513 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2514 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2515 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2516 starting_offset + 1);
2518 spin_unlock(&cli->cl_loi_list_lock);
2520 DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2521 page_count, aa, cli->cl_r_in_flight,
2522 cli->cl_w_in_flight);
2523 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2525 ptlrpcd_add_req(req);
2531 memalloc_noreclaim_restore(mpflag);
2534 LASSERT(req == NULL);
2537 OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2539 osc_release_bounce_pages(pga, page_count);
2540 osc_release_ppga(pga, page_count);
2542 /* this should happen rarely and is pretty bad, it makes the
2543 * pending list not follow the dirty order */
2544 while (!list_empty(ext_list)) {
2545 ext = list_entry(ext_list->next, struct osc_extent,
2547 list_del_init(&ext->oe_link);
2548 osc_extent_finish(env, ext, 0, rc);
2554 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2558 LASSERT(lock != NULL);
2560 lock_res_and_lock(lock);
2562 if (lock->l_ast_data == NULL)
2563 lock->l_ast_data = data;
2564 if (lock->l_ast_data == data)
2567 unlock_res_and_lock(lock);
2572 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2573 void *cookie, struct lustre_handle *lockh,
2574 enum ldlm_mode mode, __u64 *flags, bool speculative,
2577 bool intent = *flags & LDLM_FL_HAS_INTENT;
2581 /* The request was created before ldlm_cli_enqueue call. */
2582 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2583 struct ldlm_reply *rep;
2585 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2586 LASSERT(rep != NULL);
2588 rep->lock_policy_res1 =
2589 ptlrpc_status_ntoh(rep->lock_policy_res1);
2590 if (rep->lock_policy_res1)
2591 errcode = rep->lock_policy_res1;
2593 *flags |= LDLM_FL_LVB_READY;
2594 } else if (errcode == ELDLM_OK) {
2595 *flags |= LDLM_FL_LVB_READY;
2598 /* Call the update callback. */
2599 rc = (*upcall)(cookie, lockh, errcode);
2601 /* release the reference taken in ldlm_cli_enqueue() */
2602 if (errcode == ELDLM_LOCK_MATCHED)
2604 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2605 ldlm_lock_decref(lockh, mode);
2610 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2613 struct osc_enqueue_args *aa = args;
2614 struct ldlm_lock *lock;
2615 struct lustre_handle *lockh = &aa->oa_lockh;
2616 enum ldlm_mode mode = aa->oa_mode;
2617 struct ost_lvb *lvb = aa->oa_lvb;
2618 __u32 lvb_len = sizeof(*lvb);
2623 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2625 lock = ldlm_handle2lock(lockh);
2626 LASSERTF(lock != NULL,
2627 "lockh %#llx, req %p, aa %p - client evicted?\n",
2628 lockh->cookie, req, aa);
2630 /* Take an additional reference so that a blocking AST that
2631 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2632 * to arrive after an upcall has been executed by
2633 * osc_enqueue_fini(). */
2634 ldlm_lock_addref(lockh, mode);
2636 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2637 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2639 /* Let CP AST to grant the lock first. */
2640 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2642 if (aa->oa_speculative) {
2643 LASSERT(aa->oa_lvb == NULL);
2644 LASSERT(aa->oa_flags == NULL);
2645 aa->oa_flags = &flags;
2648 /* Complete obtaining the lock procedure. */
2649 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2650 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2652 /* Complete osc stuff. */
2653 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2654 aa->oa_flags, aa->oa_speculative, rc);
2656 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2658 ldlm_lock_decref(lockh, mode);
2659 LDLM_LOCK_PUT(lock);
2663 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2664 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2665 * other synchronous requests, however keeping some locks and trying to obtain
2666 * others may take a considerable amount of time in a case of ost failure; and
2667 * when other sync requests do not get released lock from a client, the client
2668 * is evicted from the cluster -- such scenarious make the life difficult, so
2669 * release locks just after they are obtained. */
2670 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2671 __u64 *flags, union ldlm_policy_data *policy,
2672 struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2673 void *cookie, struct ldlm_enqueue_info *einfo,
2674 struct ptlrpc_request_set *rqset, int async,
2677 struct obd_device *obd = exp->exp_obd;
2678 struct lustre_handle lockh = { 0 };
2679 struct ptlrpc_request *req = NULL;
2680 int intent = *flags & LDLM_FL_HAS_INTENT;
2681 __u64 match_flags = *flags;
2682 enum ldlm_mode mode;
2686 /* Filesystem lock extents are extended to page boundaries so that
2687 * dealing with the page cache is a little smoother. */
2688 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2689 policy->l_extent.end |= ~PAGE_MASK;
2691 /* Next, search for already existing extent locks that will cover us */
2692 /* If we're trying to read, we also search for an existing PW lock. The
2693 * VFS and page cache already protect us locally, so lots of readers/
2694 * writers can share a single PW lock.
2696 * There are problems with conversion deadlocks, so instead of
2697 * converting a read lock to a write lock, we'll just enqueue a new
2700 * At some point we should cancel the read lock instead of making them
2701 * send us a blocking callback, but there are problems with canceling
2702 * locks out from other users right now, too. */
2703 mode = einfo->ei_mode;
2704 if (einfo->ei_mode == LCK_PR)
2706 /* Normal lock requests must wait for the LVB to be ready before
2707 * matching a lock; speculative lock requests do not need to,
2708 * because they will not actually use the lock. */
2710 match_flags |= LDLM_FL_LVB_READY;
2712 match_flags |= LDLM_FL_BLOCK_GRANTED;
2713 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2714 einfo->ei_type, policy, mode, &lockh, 0);
2716 struct ldlm_lock *matched;
2718 if (*flags & LDLM_FL_TEST_LOCK)
2721 matched = ldlm_handle2lock(&lockh);
2723 /* This DLM lock request is speculative, and does not
2724 * have an associated IO request. Therefore if there
2725 * is already a DLM lock, it wll just inform the
2726 * caller to cancel the request for this stripe.*/
2727 lock_res_and_lock(matched);
2728 if (ldlm_extent_equal(&policy->l_extent,
2729 &matched->l_policy_data.l_extent))
2733 unlock_res_and_lock(matched);
2735 ldlm_lock_decref(&lockh, mode);
2736 LDLM_LOCK_PUT(matched);
2738 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2739 *flags |= LDLM_FL_LVB_READY;
2741 /* We already have a lock, and it's referenced. */
2742 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2744 ldlm_lock_decref(&lockh, mode);
2745 LDLM_LOCK_PUT(matched);
2748 ldlm_lock_decref(&lockh, mode);
2749 LDLM_LOCK_PUT(matched);
2753 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2757 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2758 &RQF_LDLM_ENQUEUE_LVB);
2762 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2764 ptlrpc_request_free(req);
2768 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2770 ptlrpc_request_set_replen(req);
2773 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2774 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2776 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2777 sizeof(*lvb), LVB_T_OST, &lockh, async);
2780 struct osc_enqueue_args *aa;
2781 aa = ptlrpc_req_async_args(aa, req);
2783 aa->oa_mode = einfo->ei_mode;
2784 aa->oa_type = einfo->ei_type;
2785 lustre_handle_copy(&aa->oa_lockh, &lockh);
2786 aa->oa_upcall = upcall;
2787 aa->oa_cookie = cookie;
2788 aa->oa_speculative = speculative;
2790 aa->oa_flags = flags;
2793 /* speculative locks are essentially to enqueue
2794 * a DLM lock in advance, so we don't care
2795 * about the result of the enqueue. */
2797 aa->oa_flags = NULL;
2800 req->rq_interpret_reply = osc_enqueue_interpret;
2801 ptlrpc_set_add_req(rqset, req);
2802 } else if (intent) {
2803 ptlrpc_req_finished(req);
2808 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2809 flags, speculative, rc);
2811 ptlrpc_req_finished(req);
2816 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2817 struct ldlm_res_id *res_id, enum ldlm_type type,
2818 union ldlm_policy_data *policy, enum ldlm_mode mode,
2819 __u64 *flags, struct osc_object *obj,
2820 struct lustre_handle *lockh, int unref)
2822 struct obd_device *obd = exp->exp_obd;
2823 __u64 lflags = *flags;
2827 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2830 /* Filesystem lock extents are extended to page boundaries so that
2831 * dealing with the page cache is a little smoother */
2832 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2833 policy->l_extent.end |= ~PAGE_MASK;
2835 /* Next, search for already existing extent locks that will cover us */
2836 /* If we're trying to read, we also search for an existing PW lock. The
2837 * VFS and page cache already protect us locally, so lots of readers/
2838 * writers can share a single PW lock. */
2842 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2843 res_id, type, policy, rc, lockh, unref);
2844 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2848 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2850 LASSERT(lock != NULL);
2851 if (osc_set_lock_data(lock, obj)) {
2852 lock_res_and_lock(lock);
2853 if (!ldlm_is_lvb_cached(lock)) {
2854 LASSERT(lock->l_ast_data == obj);
2855 osc_lock_lvb_update(env, obj, lock, NULL);
2856 ldlm_set_lvb_cached(lock);
2858 unlock_res_and_lock(lock);
2860 ldlm_lock_decref(lockh, rc);
2863 LDLM_LOCK_PUT(lock);
2868 static int osc_statfs_interpret(const struct lu_env *env,
2869 struct ptlrpc_request *req, void *args, int rc)
2871 struct osc_async_args *aa = args;
2872 struct obd_statfs *msfs;
2877 * The request has in fact never been sent due to issues at
2878 * a higher level (LOV). Exit immediately since the caller
2879 * is aware of the problem and takes care of the clean up.
2883 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2884 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2890 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2892 GOTO(out, rc = -EPROTO);
2894 *aa->aa_oi->oi_osfs = *msfs;
2896 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2901 static int osc_statfs_async(struct obd_export *exp,
2902 struct obd_info *oinfo, time64_t max_age,
2903 struct ptlrpc_request_set *rqset)
2905 struct obd_device *obd = class_exp2obd(exp);
2906 struct ptlrpc_request *req;
2907 struct osc_async_args *aa;
2911 if (obd->obd_osfs_age >= max_age) {
2913 "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
2914 obd->obd_name, &obd->obd_osfs,
2915 obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
2916 obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
2917 spin_lock(&obd->obd_osfs_lock);
2918 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
2919 spin_unlock(&obd->obd_osfs_lock);
2920 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
2921 if (oinfo->oi_cb_up)
2922 oinfo->oi_cb_up(oinfo, 0);
2927 /* We could possibly pass max_age in the request (as an absolute
2928 * timestamp or a "seconds.usec ago") so the target can avoid doing
2929 * extra calls into the filesystem if that isn't necessary (e.g.
2930 * during mount that would help a bit). Having relative timestamps
2931 * is not so great if request processing is slow, while absolute
2932 * timestamps are not ideal because they need time synchronization. */
2933 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2937 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2939 ptlrpc_request_free(req);
2942 ptlrpc_request_set_replen(req);
2943 req->rq_request_portal = OST_CREATE_PORTAL;
2944 ptlrpc_at_set_req_timeout(req);
2946 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2947 /* procfs requests not want stat in wait for avoid deadlock */
2948 req->rq_no_resend = 1;
2949 req->rq_no_delay = 1;
2952 req->rq_interpret_reply = osc_statfs_interpret;
2953 aa = ptlrpc_req_async_args(aa, req);
2956 ptlrpc_set_add_req(rqset, req);
2960 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2961 struct obd_statfs *osfs, time64_t max_age, __u32 flags)
2963 struct obd_device *obd = class_exp2obd(exp);
2964 struct obd_statfs *msfs;
2965 struct ptlrpc_request *req;
2966 struct obd_import *imp = NULL;
2971 /*Since the request might also come from lprocfs, so we need
2972 *sync this with client_disconnect_export Bug15684*/
2973 down_read(&obd->u.cli.cl_sem);
2974 if (obd->u.cli.cl_import)
2975 imp = class_import_get(obd->u.cli.cl_import);
2976 up_read(&obd->u.cli.cl_sem);
2980 /* We could possibly pass max_age in the request (as an absolute
2981 * timestamp or a "seconds.usec ago") so the target can avoid doing
2982 * extra calls into the filesystem if that isn't necessary (e.g.
2983 * during mount that would help a bit). Having relative timestamps
2984 * is not so great if request processing is slow, while absolute
2985 * timestamps are not ideal because they need time synchronization. */
2986 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2988 class_import_put(imp);
2993 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2995 ptlrpc_request_free(req);
2998 ptlrpc_request_set_replen(req);
2999 req->rq_request_portal = OST_CREATE_PORTAL;
3000 ptlrpc_at_set_req_timeout(req);
3002 if (flags & OBD_STATFS_NODELAY) {
3003 /* procfs requests not want stat in wait for avoid deadlock */
3004 req->rq_no_resend = 1;
3005 req->rq_no_delay = 1;
3008 rc = ptlrpc_queue_wait(req);
3012 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3014 GOTO(out, rc = -EPROTO);
3020 ptlrpc_req_finished(req);
3024 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3025 void *karg, void __user *uarg)
3027 struct obd_device *obd = exp->exp_obd;
3028 struct obd_ioctl_data *data = karg;
3032 if (!try_module_get(THIS_MODULE)) {
3033 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3034 module_name(THIS_MODULE));
3038 case OBD_IOC_CLIENT_RECOVER:
3039 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3040 data->ioc_inlbuf1, 0);
3044 case IOC_OSC_SET_ACTIVE:
3045 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3050 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
3051 obd->obd_name, cmd, current->comm, rc);
3055 module_put(THIS_MODULE);
3059 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3060 u32 keylen, void *key, u32 vallen, void *val,
3061 struct ptlrpc_request_set *set)
3063 struct ptlrpc_request *req;
3064 struct obd_device *obd = exp->exp_obd;
3065 struct obd_import *imp = class_exp2cliimp(exp);
3070 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3072 if (KEY_IS(KEY_CHECKSUM)) {
3073 if (vallen != sizeof(int))
3075 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3079 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3080 sptlrpc_conf_client_adapt(obd);
3084 if (KEY_IS(KEY_FLUSH_CTX)) {
3085 sptlrpc_import_flush_my_ctx(imp);
3089 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3090 struct client_obd *cli = &obd->u.cli;
3091 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3092 long target = *(long *)val;
3094 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3099 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3102 /* We pass all other commands directly to OST. Since nobody calls osc
3103 methods directly and everybody is supposed to go through LOV, we
3104 assume lov checked invalid values for us.
3105 The only recognised values so far are evict_by_nid and mds_conn.
3106 Even if something bad goes through, we'd get a -EINVAL from OST
3109 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3110 &RQF_OST_SET_GRANT_INFO :
3115 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3116 RCL_CLIENT, keylen);
3117 if (!KEY_IS(KEY_GRANT_SHRINK))
3118 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3119 RCL_CLIENT, vallen);
3120 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3122 ptlrpc_request_free(req);
3126 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3127 memcpy(tmp, key, keylen);
3128 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3131 memcpy(tmp, val, vallen);
3133 if (KEY_IS(KEY_GRANT_SHRINK)) {
3134 struct osc_grant_args *aa;
3137 aa = ptlrpc_req_async_args(aa, req);
3138 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3140 ptlrpc_req_finished(req);
3143 *oa = ((struct ost_body *)val)->oa;
3145 req->rq_interpret_reply = osc_shrink_grant_interpret;
3148 ptlrpc_request_set_replen(req);
3149 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3150 LASSERT(set != NULL);
3151 ptlrpc_set_add_req(set, req);
3152 ptlrpc_check_set(NULL, set);
3154 ptlrpcd_add_req(req);
3159 EXPORT_SYMBOL(osc_set_info_async);
3161 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3162 struct obd_device *obd, struct obd_uuid *cluuid,
3163 struct obd_connect_data *data, void *localdata)
3165 struct client_obd *cli = &obd->u.cli;
3167 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3171 spin_lock(&cli->cl_loi_list_lock);
3172 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3173 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3174 /* restore ocd_grant_blkbits as client page bits */
3175 data->ocd_grant_blkbits = PAGE_SHIFT;
3176 grant += cli->cl_dirty_grant;
3178 grant += cli->cl_dirty_pages << PAGE_SHIFT;
3180 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3181 lost_grant = cli->cl_lost_grant;
3182 cli->cl_lost_grant = 0;
3183 spin_unlock(&cli->cl_loi_list_lock);
3185 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3186 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3187 data->ocd_version, data->ocd_grant, lost_grant);
3192 EXPORT_SYMBOL(osc_reconnect);
3194 int osc_disconnect(struct obd_export *exp)
3196 struct obd_device *obd = class_exp2obd(exp);
3199 rc = client_disconnect_export(exp);
3201 * Initially we put del_shrink_grant before disconnect_export, but it
3202 * causes the following problem if setup (connect) and cleanup
3203 * (disconnect) are tangled together.
3204 * connect p1 disconnect p2
3205 * ptlrpc_connect_import
3206 * ............... class_manual_cleanup
3209 * ptlrpc_connect_interrupt
3211 * add this client to shrink list
3213 * Bang! grant shrink thread trigger the shrink. BUG18662
3215 osc_del_grant_list(&obd->u.cli);
3218 EXPORT_SYMBOL(osc_disconnect);
3220 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3221 struct hlist_node *hnode, void *arg)
3223 struct lu_env *env = arg;
3224 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3225 struct ldlm_lock *lock;
3226 struct osc_object *osc = NULL;
3230 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3231 if (lock->l_ast_data != NULL && osc == NULL) {
3232 osc = lock->l_ast_data;
3233 cl_object_get(osc2cl(osc));
3236 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3237 * by the 2nd round of ldlm_namespace_clean() call in
3238 * osc_import_event(). */
3239 ldlm_clear_cleaned(lock);
3244 osc_object_invalidate(env, osc);
3245 cl_object_put(env, osc2cl(osc));
3250 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3252 static int osc_import_event(struct obd_device *obd,
3253 struct obd_import *imp,
3254 enum obd_import_event event)
3256 struct client_obd *cli;
3260 LASSERT(imp->imp_obd == obd);
3263 case IMP_EVENT_DISCON: {
3265 spin_lock(&cli->cl_loi_list_lock);
3266 cli->cl_avail_grant = 0;
3267 cli->cl_lost_grant = 0;
3268 spin_unlock(&cli->cl_loi_list_lock);
3271 case IMP_EVENT_INACTIVE: {
3272 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3275 case IMP_EVENT_INVALIDATE: {
3276 struct ldlm_namespace *ns = obd->obd_namespace;
3280 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3282 env = cl_env_get(&refcheck);
3284 osc_io_unplug(env, &obd->u.cli, NULL);
3286 cfs_hash_for_each_nolock(ns->ns_rs_hash,
3287 osc_ldlm_resource_invalidate,
3289 cl_env_put(env, &refcheck);
3291 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3296 case IMP_EVENT_ACTIVE: {
3297 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3300 case IMP_EVENT_OCD: {
3301 struct obd_connect_data *ocd = &imp->imp_connect_data;
3303 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3304 osc_init_grant(&obd->u.cli, ocd);
3307 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3308 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3310 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3313 case IMP_EVENT_DEACTIVATE: {
3314 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3317 case IMP_EVENT_ACTIVATE: {
3318 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3322 CERROR("Unknown import event %d\n", event);
3329 * Determine whether the lock can be canceled before replaying the lock
3330 * during recovery, see bug16774 for detailed information.
3332 * \retval zero the lock can't be canceled
3333 * \retval other ok to cancel
3335 static int osc_cancel_weight(struct ldlm_lock *lock)
3338 * Cancel all unused and granted extent lock.
3340 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3341 ldlm_is_granted(lock) &&
3342 osc_ldlm_weigh_ast(lock) == 0)
3348 static int brw_queue_work(const struct lu_env *env, void *data)
3350 struct client_obd *cli = data;
3352 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3354 osc_io_unplug(env, cli, NULL);
3358 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3360 struct client_obd *cli = &obd->u.cli;
3366 rc = ptlrpcd_addref();
3370 rc = client_obd_setup(obd, lcfg);
3372 GOTO(out_ptlrpcd, rc);
3375 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3376 if (IS_ERR(handler))
3377 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3378 cli->cl_writeback_work = handler;
3380 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3381 if (IS_ERR(handler))
3382 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3383 cli->cl_lru_work = handler;
3385 rc = osc_quota_setup(obd);
3387 GOTO(out_ptlrpcd_work, rc);
3389 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3390 osc_update_next_shrink(cli);
3395 if (cli->cl_writeback_work != NULL) {
3396 ptlrpcd_destroy_work(cli->cl_writeback_work);
3397 cli->cl_writeback_work = NULL;
3399 if (cli->cl_lru_work != NULL) {
3400 ptlrpcd_destroy_work(cli->cl_lru_work);
3401 cli->cl_lru_work = NULL;
3403 client_obd_cleanup(obd);
3408 EXPORT_SYMBOL(osc_setup_common);
3410 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3412 struct client_obd *cli = &obd->u.cli;
3420 rc = osc_setup_common(obd, lcfg);
3424 rc = osc_tunables_init(obd);
3429 * We try to control the total number of requests with a upper limit
3430 * osc_reqpool_maxreqcount. There might be some race which will cause
3431 * over-limit allocation, but it is fine.
3433 req_count = atomic_read(&osc_pool_req_count);
3434 if (req_count < osc_reqpool_maxreqcount) {
3435 adding = cli->cl_max_rpcs_in_flight + 2;
3436 if (req_count + adding > osc_reqpool_maxreqcount)
3437 adding = osc_reqpool_maxreqcount - req_count;
3439 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3440 atomic_add(added, &osc_pool_req_count);
3443 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3445 spin_lock(&osc_shrink_lock);
3446 list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3447 spin_unlock(&osc_shrink_lock);
3448 cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3449 cli->cl_import->imp_idle_debug = D_HA;
3454 int osc_precleanup_common(struct obd_device *obd)
3456 struct client_obd *cli = &obd->u.cli;
3460 * for echo client, export may be on zombie list, wait for
3461 * zombie thread to cull it, because cli.cl_import will be
3462 * cleared in client_disconnect_export():
3463 * class_export_destroy() -> obd_cleanup() ->
3464 * echo_device_free() -> echo_client_cleanup() ->
3465 * obd_disconnect() -> osc_disconnect() ->
3466 * client_disconnect_export()
3468 obd_zombie_barrier();
3469 if (cli->cl_writeback_work) {
3470 ptlrpcd_destroy_work(cli->cl_writeback_work);
3471 cli->cl_writeback_work = NULL;
3474 if (cli->cl_lru_work) {
3475 ptlrpcd_destroy_work(cli->cl_lru_work);
3476 cli->cl_lru_work = NULL;
3479 obd_cleanup_client_import(obd);
3482 EXPORT_SYMBOL(osc_precleanup_common);
3484 static int osc_precleanup(struct obd_device *obd)
3488 osc_precleanup_common(obd);
3490 ptlrpc_lprocfs_unregister_obd(obd);
3494 int osc_cleanup_common(struct obd_device *obd)
3496 struct client_obd *cli = &obd->u.cli;
3501 spin_lock(&osc_shrink_lock);
3502 list_del(&cli->cl_shrink_list);
3503 spin_unlock(&osc_shrink_lock);
3506 if (cli->cl_cache != NULL) {
3507 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3508 spin_lock(&cli->cl_cache->ccc_lru_lock);
3509 list_del_init(&cli->cl_lru_osc);
3510 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3511 cli->cl_lru_left = NULL;
3512 cl_cache_decref(cli->cl_cache);
3513 cli->cl_cache = NULL;
3516 /* free memory of osc quota cache */
3517 osc_quota_cleanup(obd);
3519 rc = client_obd_cleanup(obd);
3524 EXPORT_SYMBOL(osc_cleanup_common);
3526 static const struct obd_ops osc_obd_ops = {
3527 .o_owner = THIS_MODULE,
3528 .o_setup = osc_setup,
3529 .o_precleanup = osc_precleanup,
3530 .o_cleanup = osc_cleanup_common,
3531 .o_add_conn = client_import_add_conn,
3532 .o_del_conn = client_import_del_conn,
3533 .o_connect = client_connect_import,
3534 .o_reconnect = osc_reconnect,
3535 .o_disconnect = osc_disconnect,
3536 .o_statfs = osc_statfs,
3537 .o_statfs_async = osc_statfs_async,
3538 .o_create = osc_create,
3539 .o_destroy = osc_destroy,
3540 .o_getattr = osc_getattr,
3541 .o_setattr = osc_setattr,
3542 .o_iocontrol = osc_iocontrol,
3543 .o_set_info_async = osc_set_info_async,
3544 .o_import_event = osc_import_event,
3545 .o_quotactl = osc_quotactl,
3548 static struct shrinker *osc_cache_shrinker;
3549 LIST_HEAD(osc_shrink_list);
3550 DEFINE_SPINLOCK(osc_shrink_lock);
3552 #ifndef HAVE_SHRINKER_COUNT
3553 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
3555 struct shrink_control scv = {
3556 .nr_to_scan = shrink_param(sc, nr_to_scan),
3557 .gfp_mask = shrink_param(sc, gfp_mask)
3559 (void)osc_cache_shrink_scan(shrinker, &scv);
3561 return osc_cache_shrink_count(shrinker, &scv);
3565 static int __init osc_init(void)
3567 unsigned int reqpool_size;
3568 unsigned int reqsize;
3570 DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
3571 osc_cache_shrink_count, osc_cache_shrink_scan);
3574 /* print an address of _any_ initialized kernel symbol from this
3575 * module, to allow debugging with gdb that doesn't support data
3576 * symbols from modules.*/
3577 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3579 rc = lu_kmem_init(osc_caches);
3583 rc = class_register_type(&osc_obd_ops, NULL, true, NULL,
3584 LUSTRE_OSC_NAME, &osc_device_type);
3588 osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
3590 /* This is obviously too much memory, only prevent overflow here */
3591 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3592 GOTO(out_type, rc = -EINVAL);
3594 reqpool_size = osc_reqpool_mem_max << 20;
3597 while (reqsize < OST_IO_MAXREQSIZE)
3598 reqsize = reqsize << 1;
3601 * We don't enlarge the request count in OSC pool according to
3602 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3603 * tried after normal allocation failed. So a small OSC pool won't
3604 * cause much performance degression in most of cases.
3606 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3608 atomic_set(&osc_pool_req_count, 0);
3609 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3610 ptlrpc_add_rqs_to_pool);
3612 if (osc_rq_pool == NULL)
3613 GOTO(out_type, rc = -ENOMEM);
3615 rc = osc_start_grant_work();
3617 GOTO(out_req_pool, rc);
3622 ptlrpc_free_rq_pool(osc_rq_pool);
3624 class_unregister_type(LUSTRE_OSC_NAME);
3626 lu_kmem_fini(osc_caches);
3631 static void __exit osc_exit(void)
3633 osc_stop_grant_work();
3634 remove_shrinker(osc_cache_shrinker);
3635 class_unregister_type(LUSTRE_OSC_NAME);
3636 lu_kmem_fini(osc_caches);
3637 ptlrpc_free_rq_pool(osc_rq_pool);
3640 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3641 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3642 MODULE_VERSION(LUSTRE_VERSION_STRING);
3643 MODULE_LICENSE("GPL");
3645 module_init(osc_init);
3646 module_exit(osc_exit);