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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include "../../include/linux/libcfs/libcfs.h"
42 #include "../include/lustre_dlm.h"
43 #include "../include/lustre_net.h"
44 #include "../include/lustre/lustre_user.h"
45 #include "../include/obd_cksum.h"
47 #include "../include/lustre_ha.h"
48 #include "../include/lprocfs_status.h"
49 #include "../include/lustre_debug.h"
50 #include "../include/lustre_param.h"
51 #include "../include/lustre_fid.h"
52 #include "../include/obd_class.h"
53 #include "../include/obd.h"
54 #include "osc_internal.h"
55 #include "osc_cl_internal.h"
57 atomic_t osc_pool_req_count;
58 unsigned int osc_reqpool_maxreqcount;
59 struct ptlrpc_request_pool *osc_rq_pool;
61 /* max memory used for request pool, unit is MB */
62 static unsigned int osc_reqpool_mem_max = 5;
63 module_param(osc_reqpool_mem_max, uint, 0444);
65 struct osc_brw_async_args {
71 struct brw_page **aa_ppga;
72 struct client_obd *aa_cli;
73 struct list_head aa_oaps;
74 struct list_head aa_exts;
75 struct cl_req *aa_clerq;
78 struct osc_async_args {
79 struct obd_info *aa_oi;
82 struct osc_setattr_args {
84 obd_enqueue_update_f sa_upcall;
88 struct osc_fsync_args {
89 struct obd_info *fa_oi;
90 obd_enqueue_update_f fa_upcall;
94 struct osc_enqueue_args {
95 struct obd_export *oa_exp;
97 obd_enqueue_update_f oa_upcall;
99 struct ost_lvb *oa_lvb;
100 struct lustre_handle *oa_lockh;
101 struct ldlm_enqueue_info *oa_ei;
102 unsigned int oa_agl:1;
105 static void osc_release_ppga(struct brw_page **ppga, u32 count);
106 static int brw_interpret(const struct lu_env *env,
107 struct ptlrpc_request *req, void *data, int rc);
108 int osc_cleanup(struct obd_device *obd);
110 /* Pack OSC object metadata for disk storage (LE byte order). */
111 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
112 struct lov_stripe_md *lsm)
116 lmm_size = sizeof(**lmmp);
120 if (*lmmp != NULL && lsm == NULL) {
124 } else if (unlikely(lsm != NULL && ostid_id(&lsm->lsm_oi) == 0)) {
129 *lmmp = kzalloc(lmm_size, GFP_NOFS);
135 ostid_cpu_to_le(&lsm->lsm_oi, &(*lmmp)->lmm_oi);
140 /* Unpack OSC object metadata from disk storage (LE byte order). */
141 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
142 struct lov_mds_md *lmm, int lmm_bytes)
145 struct obd_import *imp = class_exp2cliimp(exp);
148 if (lmm_bytes < sizeof(*lmm)) {
149 CERROR("%s: lov_mds_md too small: %d, need %d\n",
150 exp->exp_obd->obd_name, lmm_bytes,
154 /* XXX LOV_MAGIC etc check? */
156 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
157 CERROR("%s: zero lmm_object_id: rc = %d\n",
158 exp->exp_obd->obd_name, -EINVAL);
163 lsm_size = lov_stripe_md_size(1);
167 if (*lsmp != NULL && lmm == NULL) {
168 kfree((*lsmp)->lsm_oinfo[0]);
175 *lsmp = kzalloc(lsm_size, GFP_NOFS);
176 if (unlikely(*lsmp == NULL))
178 (*lsmp)->lsm_oinfo[0] = kzalloc(sizeof(struct lov_oinfo),
180 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
184 loi_init((*lsmp)->lsm_oinfo[0]);
185 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
190 /* XXX zero *lsmp? */
191 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
194 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
195 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
197 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
202 static inline void osc_pack_req_body(struct ptlrpc_request *req,
203 struct obd_info *oinfo)
205 struct ost_body *body;
207 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
210 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
214 static int osc_getattr_interpret(const struct lu_env *env,
215 struct ptlrpc_request *req,
216 struct osc_async_args *aa, int rc)
218 struct ost_body *body;
223 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
225 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
226 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
227 aa->aa_oi->oi_oa, &body->oa);
229 /* This should really be sent by the OST */
230 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
231 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
233 CDEBUG(D_INFO, "can't unpack ost_body\n");
235 aa->aa_oi->oi_oa->o_valid = 0;
238 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
242 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
243 struct ptlrpc_request_set *set)
245 struct ptlrpc_request *req;
246 struct osc_async_args *aa;
249 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
253 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
255 ptlrpc_request_free(req);
259 osc_pack_req_body(req, oinfo);
261 ptlrpc_request_set_replen(req);
262 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
264 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
265 aa = ptlrpc_req_async_args(req);
268 ptlrpc_set_add_req(set, req);
272 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
273 struct obd_info *oinfo)
275 struct ptlrpc_request *req;
276 struct ost_body *body;
279 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
283 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
285 ptlrpc_request_free(req);
289 osc_pack_req_body(req, oinfo);
291 ptlrpc_request_set_replen(req);
293 rc = ptlrpc_queue_wait(req);
297 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
303 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
304 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
307 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
308 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
311 ptlrpc_req_finished(req);
315 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
316 struct obd_info *oinfo, struct obd_trans_info *oti)
318 struct ptlrpc_request *req;
319 struct ost_body *body;
322 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
324 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
328 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
330 ptlrpc_request_free(req);
334 osc_pack_req_body(req, oinfo);
336 ptlrpc_request_set_replen(req);
338 rc = ptlrpc_queue_wait(req);
342 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
348 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
352 ptlrpc_req_finished(req);
356 static int osc_setattr_interpret(const struct lu_env *env,
357 struct ptlrpc_request *req,
358 struct osc_setattr_args *sa, int rc)
360 struct ost_body *body;
365 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
371 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
374 rc = sa->sa_upcall(sa->sa_cookie, rc);
378 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
379 struct obd_trans_info *oti,
380 obd_enqueue_update_f upcall, void *cookie,
381 struct ptlrpc_request_set *rqset)
383 struct ptlrpc_request *req;
384 struct osc_setattr_args *sa;
387 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
391 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
393 ptlrpc_request_free(req);
397 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
398 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
400 osc_pack_req_body(req, oinfo);
402 ptlrpc_request_set_replen(req);
404 /* do mds to ost setattr asynchronously */
406 /* Do not wait for response. */
407 ptlrpcd_add_req(req);
409 req->rq_interpret_reply =
410 (ptlrpc_interpterer_t)osc_setattr_interpret;
412 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
413 sa = ptlrpc_req_async_args(req);
414 sa->sa_oa = oinfo->oi_oa;
415 sa->sa_upcall = upcall;
416 sa->sa_cookie = cookie;
418 if (rqset == PTLRPCD_SET)
419 ptlrpcd_add_req(req);
421 ptlrpc_set_add_req(rqset, req);
427 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
428 struct obd_trans_info *oti,
429 struct ptlrpc_request_set *rqset)
431 return osc_setattr_async_base(exp, oinfo, oti,
432 oinfo->oi_cb_up, oinfo, rqset);
435 int osc_real_create(struct obd_export *exp, struct obdo *oa,
436 struct lov_stripe_md **ea, struct obd_trans_info *oti)
438 struct ptlrpc_request *req;
439 struct ost_body *body;
440 struct lov_stripe_md *lsm;
448 rc = obd_alloc_memmd(exp, &lsm);
453 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
459 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
461 ptlrpc_request_free(req);
465 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
468 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
470 ptlrpc_request_set_replen(req);
472 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
473 oa->o_flags == OBD_FL_DELORPHAN) {
475 "delorphan from OST integration");
476 /* Don't resend the delorphan req */
477 req->rq_no_resend = req->rq_no_delay = 1;
480 rc = ptlrpc_queue_wait(req);
484 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
490 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
491 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
493 oa->o_blksize = cli_brw_size(exp->exp_obd);
494 oa->o_valid |= OBD_MD_FLBLKSZ;
496 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
497 * have valid lsm_oinfo data structs, so don't go touching that.
498 * This needs to be fixed in a big way.
500 lsm->lsm_oi = oa->o_oi;
504 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
506 if (oa->o_valid & OBD_MD_FLCOOKIE) {
507 if (!oti->oti_logcookies)
508 oti_alloc_cookies(oti, 1);
509 *oti->oti_logcookies = oa->o_lcookie;
513 CDEBUG(D_HA, "transno: %lld\n",
514 lustre_msg_get_transno(req->rq_repmsg));
516 ptlrpc_req_finished(req);
519 obd_free_memmd(exp, &lsm);
523 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
524 obd_enqueue_update_f upcall, void *cookie,
525 struct ptlrpc_request_set *rqset)
527 struct ptlrpc_request *req;
528 struct osc_setattr_args *sa;
529 struct ost_body *body;
532 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
536 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
538 ptlrpc_request_free(req);
541 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
542 ptlrpc_at_set_req_timeout(req);
544 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
546 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
549 ptlrpc_request_set_replen(req);
551 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
552 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
553 sa = ptlrpc_req_async_args(req);
554 sa->sa_oa = oinfo->oi_oa;
555 sa->sa_upcall = upcall;
556 sa->sa_cookie = cookie;
557 if (rqset == PTLRPCD_SET)
558 ptlrpcd_add_req(req);
560 ptlrpc_set_add_req(rqset, req);
565 static int osc_sync_interpret(const struct lu_env *env,
566 struct ptlrpc_request *req,
569 struct osc_fsync_args *fa = arg;
570 struct ost_body *body;
575 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
577 CERROR("can't unpack ost_body\n");
582 *fa->fa_oi->oi_oa = body->oa;
584 rc = fa->fa_upcall(fa->fa_cookie, rc);
588 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
589 obd_enqueue_update_f upcall, void *cookie,
590 struct ptlrpc_request_set *rqset)
592 struct ptlrpc_request *req;
593 struct ost_body *body;
594 struct osc_fsync_args *fa;
597 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
601 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
603 ptlrpc_request_free(req);
607 /* overload the size and blocks fields in the oa with start/end */
608 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
610 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
613 ptlrpc_request_set_replen(req);
614 req->rq_interpret_reply = osc_sync_interpret;
616 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
617 fa = ptlrpc_req_async_args(req);
619 fa->fa_upcall = upcall;
620 fa->fa_cookie = cookie;
622 if (rqset == PTLRPCD_SET)
623 ptlrpcd_add_req(req);
625 ptlrpc_set_add_req(rqset, req);
630 /* Find and cancel locally locks matched by @mode in the resource found by
631 * @objid. Found locks are added into @cancel list. Returns the amount of
632 * locks added to @cancels list. */
633 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
634 struct list_head *cancels,
635 ldlm_mode_t mode, __u64 lock_flags)
637 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
638 struct ldlm_res_id res_id;
639 struct ldlm_resource *res;
642 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
643 * export) but disabled through procfs (flag in NS).
645 * This distinguishes from a case when ELC is not supported originally,
646 * when we still want to cancel locks in advance and just cancel them
647 * locally, without sending any RPC. */
648 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
651 ostid_build_res_name(&oa->o_oi, &res_id);
652 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
656 LDLM_RESOURCE_ADDREF(res);
657 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
658 lock_flags, 0, NULL);
659 LDLM_RESOURCE_DELREF(res);
660 ldlm_resource_putref(res);
664 static int osc_destroy_interpret(const struct lu_env *env,
665 struct ptlrpc_request *req, void *data,
668 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
670 atomic_dec(&cli->cl_destroy_in_flight);
671 wake_up(&cli->cl_destroy_waitq);
675 static int osc_can_send_destroy(struct client_obd *cli)
677 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
678 cli->cl_max_rpcs_in_flight) {
679 /* The destroy request can be sent */
682 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
683 cli->cl_max_rpcs_in_flight) {
685 * The counter has been modified between the two atomic
688 wake_up(&cli->cl_destroy_waitq);
693 int osc_create(const struct lu_env *env, struct obd_export *exp,
694 struct obdo *oa, struct lov_stripe_md **ea,
695 struct obd_trans_info *oti)
701 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
703 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
704 oa->o_flags == OBD_FL_RECREATE_OBJS) {
705 return osc_real_create(exp, oa, ea, oti);
708 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
709 return osc_real_create(exp, oa, ea, oti);
711 /* we should not get here anymore */
717 /* Destroy requests can be async always on the client, and we don't even really
718 * care about the return code since the client cannot do anything at all about
720 * When the MDS is unlinking a filename, it saves the file objects into a
721 * recovery llog, and these object records are cancelled when the OST reports
722 * they were destroyed and sync'd to disk (i.e. transaction committed).
723 * If the client dies, or the OST is down when the object should be destroyed,
724 * the records are not cancelled, and when the OST reconnects to the MDS next,
725 * it will retrieve the llog unlink logs and then sends the log cancellation
726 * cookies to the MDS after committing destroy transactions. */
727 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
728 struct obdo *oa, struct lov_stripe_md *ea,
729 struct obd_trans_info *oti, struct obd_export *md_export)
731 struct client_obd *cli = &exp->exp_obd->u.cli;
732 struct ptlrpc_request *req;
733 struct ost_body *body;
738 CDEBUG(D_INFO, "oa NULL\n");
742 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
743 LDLM_FL_DISCARD_DATA);
745 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
747 ldlm_lock_list_put(&cancels, l_bl_ast, count);
751 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
754 ptlrpc_request_free(req);
758 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
759 ptlrpc_at_set_req_timeout(req);
761 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
762 oa->o_lcookie = *oti->oti_logcookies;
763 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
765 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
767 ptlrpc_request_set_replen(req);
769 /* If osc_destroy is for destroying the unlink orphan,
770 * sent from MDT to OST, which should not be blocked here,
771 * because the process might be triggered by ptlrpcd, and
772 * it is not good to block ptlrpcd thread (b=16006)*/
773 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
774 req->rq_interpret_reply = osc_destroy_interpret;
775 if (!osc_can_send_destroy(cli)) {
776 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
780 * Wait until the number of on-going destroy RPCs drops
781 * under max_rpc_in_flight
783 l_wait_event_exclusive(cli->cl_destroy_waitq,
784 osc_can_send_destroy(cli), &lwi);
788 /* Do not wait for response */
789 ptlrpcd_add_req(req);
793 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
796 u32 bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
798 LASSERT(!(oa->o_valid & bits));
801 client_obd_list_lock(&cli->cl_loi_list_lock);
802 oa->o_dirty = cli->cl_dirty;
803 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
804 cli->cl_dirty_max)) {
805 CERROR("dirty %lu - %lu > dirty_max %lu\n",
806 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
808 } else if (unlikely(atomic_read(&obd_dirty_pages) -
809 atomic_read(&obd_dirty_transit_pages) >
810 (long)(obd_max_dirty_pages + 1))) {
811 /* The atomic_read() allowing the atomic_inc() are
812 * not covered by a lock thus they may safely race and trip
813 * this CERROR() unless we add in a small fudge factor (+1). */
814 CERROR("dirty %d - %d > system dirty_max %d\n",
815 atomic_read(&obd_dirty_pages),
816 atomic_read(&obd_dirty_transit_pages),
817 obd_max_dirty_pages);
819 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
820 CERROR("dirty %lu - dirty_max %lu too big???\n",
821 cli->cl_dirty, cli->cl_dirty_max);
824 long max_in_flight = (cli->cl_max_pages_per_rpc <<
826 (cli->cl_max_rpcs_in_flight + 1);
827 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
829 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
830 oa->o_dropped = cli->cl_lost_grant;
831 cli->cl_lost_grant = 0;
832 client_obd_list_unlock(&cli->cl_loi_list_lock);
833 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
834 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
838 void osc_update_next_shrink(struct client_obd *cli)
840 cli->cl_next_shrink_grant =
841 cfs_time_shift(cli->cl_grant_shrink_interval);
842 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
843 cli->cl_next_shrink_grant);
846 static void __osc_update_grant(struct client_obd *cli, u64 grant)
848 client_obd_list_lock(&cli->cl_loi_list_lock);
849 cli->cl_avail_grant += grant;
850 client_obd_list_unlock(&cli->cl_loi_list_lock);
853 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
855 if (body->oa.o_valid & OBD_MD_FLGRANT) {
856 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
857 __osc_update_grant(cli, body->oa.o_grant);
861 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
862 u32 keylen, void *key, u32 vallen,
863 void *val, struct ptlrpc_request_set *set);
865 static int osc_shrink_grant_interpret(const struct lu_env *env,
866 struct ptlrpc_request *req,
869 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
870 struct obdo *oa = ((struct osc_brw_async_args *)aa)->aa_oa;
871 struct ost_body *body;
874 __osc_update_grant(cli, oa->o_grant);
878 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
880 osc_update_grant(cli, body);
886 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
888 client_obd_list_lock(&cli->cl_loi_list_lock);
889 oa->o_grant = cli->cl_avail_grant / 4;
890 cli->cl_avail_grant -= oa->o_grant;
891 client_obd_list_unlock(&cli->cl_loi_list_lock);
892 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
893 oa->o_valid |= OBD_MD_FLFLAGS;
896 oa->o_flags |= OBD_FL_SHRINK_GRANT;
897 osc_update_next_shrink(cli);
900 /* Shrink the current grant, either from some large amount to enough for a
901 * full set of in-flight RPCs, or if we have already shrunk to that limit
902 * then to enough for a single RPC. This avoids keeping more grant than
903 * needed, and avoids shrinking the grant piecemeal. */
904 static int osc_shrink_grant(struct client_obd *cli)
906 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
907 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
909 client_obd_list_lock(&cli->cl_loi_list_lock);
910 if (cli->cl_avail_grant <= target_bytes)
911 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
912 client_obd_list_unlock(&cli->cl_loi_list_lock);
914 return osc_shrink_grant_to_target(cli, target_bytes);
917 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
920 struct ost_body *body;
922 client_obd_list_lock(&cli->cl_loi_list_lock);
923 /* Don't shrink if we are already above or below the desired limit
924 * We don't want to shrink below a single RPC, as that will negatively
925 * impact block allocation and long-term performance. */
926 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
927 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
929 if (target_bytes >= cli->cl_avail_grant) {
930 client_obd_list_unlock(&cli->cl_loi_list_lock);
933 client_obd_list_unlock(&cli->cl_loi_list_lock);
935 body = kzalloc(sizeof(*body), GFP_NOFS);
939 osc_announce_cached(cli, &body->oa, 0);
941 client_obd_list_lock(&cli->cl_loi_list_lock);
942 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
943 cli->cl_avail_grant = target_bytes;
944 client_obd_list_unlock(&cli->cl_loi_list_lock);
945 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
946 body->oa.o_valid |= OBD_MD_FLFLAGS;
947 body->oa.o_flags = 0;
949 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
950 osc_update_next_shrink(cli);
952 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
953 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
954 sizeof(*body), body, NULL);
956 __osc_update_grant(cli, body->oa.o_grant);
961 static int osc_should_shrink_grant(struct client_obd *client)
963 unsigned long time = cfs_time_current();
964 unsigned long next_shrink = client->cl_next_shrink_grant;
966 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
967 OBD_CONNECT_GRANT_SHRINK) == 0)
970 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
971 /* Get the current RPC size directly, instead of going via:
972 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
973 * Keep comment here so that it can be found by searching. */
974 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
976 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
977 client->cl_avail_grant > brw_size)
980 osc_update_next_shrink(client);
985 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
987 struct client_obd *client;
989 list_for_each_entry(client, &item->ti_obd_list,
990 cl_grant_shrink_list) {
991 if (osc_should_shrink_grant(client))
992 osc_shrink_grant(client);
997 static int osc_add_shrink_grant(struct client_obd *client)
1001 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1003 osc_grant_shrink_grant_cb, NULL,
1004 &client->cl_grant_shrink_list);
1006 CERROR("add grant client %s error %d\n",
1007 client->cl_import->imp_obd->obd_name, rc);
1010 CDEBUG(D_CACHE, "add grant client %s \n",
1011 client->cl_import->imp_obd->obd_name);
1012 osc_update_next_shrink(client);
1016 static int osc_del_shrink_grant(struct client_obd *client)
1018 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1022 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1025 * ocd_grant is the total grant amount we're expect to hold: if we've
1026 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1027 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1029 * race is tolerable here: if we're evicted, but imp_state already
1030 * left EVICTED state, then cl_dirty must be 0 already.
1032 client_obd_list_lock(&cli->cl_loi_list_lock);
1033 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1034 cli->cl_avail_grant = ocd->ocd_grant;
1036 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1038 if (cli->cl_avail_grant < 0) {
1039 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1040 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1041 ocd->ocd_grant, cli->cl_dirty);
1042 /* workaround for servers which do not have the patch from
1044 cli->cl_avail_grant = ocd->ocd_grant;
1047 /* determine the appropriate chunk size used by osc_extent. */
1048 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1049 client_obd_list_unlock(&cli->cl_loi_list_lock);
1051 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld chunk bits: %d\n",
1052 cli->cl_import->imp_obd->obd_name,
1053 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1055 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1056 list_empty(&cli->cl_grant_shrink_list))
1057 osc_add_shrink_grant(cli);
1060 /* We assume that the reason this OSC got a short read is because it read
1061 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1062 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1063 * this stripe never got written at or beyond this stripe offset yet. */
1064 static void handle_short_read(int nob_read, u32 page_count,
1065 struct brw_page **pga)
1070 /* skip bytes read OK */
1071 while (nob_read > 0) {
1072 LASSERT(page_count > 0);
1074 if (pga[i]->count > nob_read) {
1075 /* EOF inside this page */
1076 ptr = kmap(pga[i]->pg) +
1077 (pga[i]->off & ~CFS_PAGE_MASK);
1078 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1085 nob_read -= pga[i]->count;
1090 /* zero remaining pages */
1091 while (page_count-- > 0) {
1092 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1093 memset(ptr, 0, pga[i]->count);
1099 static int check_write_rcs(struct ptlrpc_request *req,
1100 int requested_nob, int niocount,
1101 u32 page_count, struct brw_page **pga)
1106 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1107 sizeof(*remote_rcs) *
1109 if (remote_rcs == NULL) {
1110 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1114 /* return error if any niobuf was in error */
1115 for (i = 0; i < niocount; i++) {
1116 if ((int)remote_rcs[i] < 0)
1117 return remote_rcs[i];
1119 if (remote_rcs[i] != 0) {
1120 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1121 i, remote_rcs[i], req);
1126 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1127 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1128 req->rq_bulk->bd_nob_transferred, requested_nob);
1135 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1137 if (p1->flag != p2->flag) {
1138 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1139 OBD_BRW_SYNC | OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1141 /* warn if we try to combine flags that we don't know to be
1142 * safe to combine */
1143 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1144 CWARN("Saw flags 0x%x and 0x%x in the same brw, please report this at http://bugs.whamcloud.com/\n",
1145 p1->flag, p2->flag);
1150 return (p1->off + p1->count == p2->off);
1153 static u32 osc_checksum_bulk(int nob, u32 pg_count,
1154 struct brw_page **pga, int opc,
1155 cksum_type_t cksum_type)
1159 struct cfs_crypto_hash_desc *hdesc;
1160 unsigned int bufsize;
1162 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1164 LASSERT(pg_count > 0);
1166 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1167 if (IS_ERR(hdesc)) {
1168 CERROR("Unable to initialize checksum hash %s\n",
1169 cfs_crypto_hash_name(cfs_alg));
1170 return PTR_ERR(hdesc);
1173 while (nob > 0 && pg_count > 0) {
1174 int count = pga[i]->count > nob ? nob : pga[i]->count;
1176 /* corrupt the data before we compute the checksum, to
1177 * simulate an OST->client data error */
1178 if (i == 0 && opc == OST_READ &&
1179 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1180 unsigned char *ptr = kmap(pga[i]->pg);
1181 int off = pga[i]->off & ~CFS_PAGE_MASK;
1182 memcpy(ptr + off, "bad1", min(4, nob));
1185 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1186 pga[i]->off & ~CFS_PAGE_MASK,
1189 "page %p map %p index %lu flags %lx count %u priv %0lx: off %d\n",
1190 pga[i]->pg, pga[i]->pg->mapping, pga[i]->pg->index,
1191 (long)pga[i]->pg->flags, page_count(pga[i]->pg),
1192 page_private(pga[i]->pg),
1193 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1195 nob -= pga[i]->count;
1201 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1204 cfs_crypto_hash_final(hdesc, NULL, NULL);
1206 /* For sending we only compute the wrong checksum instead
1207 * of corrupting the data so it is still correct on a redo */
1208 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1214 static int osc_brw_prep_request(int cmd, struct client_obd *cli,
1216 struct lov_stripe_md *lsm, u32 page_count,
1217 struct brw_page **pga,
1218 struct ptlrpc_request **reqp,
1222 struct ptlrpc_request *req;
1223 struct ptlrpc_bulk_desc *desc;
1224 struct ost_body *body;
1225 struct obd_ioobj *ioobj;
1226 struct niobuf_remote *niobuf;
1227 int niocount, i, requested_nob, opc, rc;
1228 struct osc_brw_async_args *aa;
1229 struct req_capsule *pill;
1230 struct brw_page *pg_prev;
1232 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1233 return -ENOMEM; /* Recoverable */
1234 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1235 return -EINVAL; /* Fatal */
1237 if ((cmd & OBD_BRW_WRITE) != 0) {
1239 req = ptlrpc_request_alloc_pool(cli->cl_import,
1241 &RQF_OST_BRW_WRITE);
1244 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1249 for (niocount = i = 1; i < page_count; i++) {
1250 if (!can_merge_pages(pga[i - 1], pga[i]))
1254 pill = &req->rq_pill;
1255 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1257 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1258 niocount * sizeof(*niobuf));
1260 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1262 ptlrpc_request_free(req);
1265 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1266 ptlrpc_at_set_req_timeout(req);
1267 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1269 req->rq_no_retry_einprogress = 1;
1271 desc = ptlrpc_prep_bulk_imp(req, page_count,
1272 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1273 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1280 /* NB request now owns desc and will free it when it gets freed */
1282 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1283 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1284 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1285 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1287 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1289 obdo_to_ioobj(oa, ioobj);
1290 ioobj->ioo_bufcnt = niocount;
1291 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1292 * that might be send for this request. The actual number is decided
1293 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1294 * "max - 1" for old client compatibility sending "0", and also so the
1295 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1296 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1297 LASSERT(page_count > 0);
1299 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1300 struct brw_page *pg = pga[i];
1301 int poff = pg->off & ~CFS_PAGE_MASK;
1303 LASSERT(pg->count > 0);
1304 /* make sure there is no gap in the middle of page array */
1305 LASSERTF(page_count == 1 ||
1306 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1307 ergo(i > 0 && i < page_count - 1,
1308 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1309 ergo(i == page_count - 1, poff == 0)),
1310 "i: %d/%d pg: %p off: %llu, count: %u\n",
1311 i, page_count, pg, pg->off, pg->count);
1312 LASSERTF(i == 0 || pg->off > pg_prev->off,
1313 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu prev_pg %p [pri %lu ind %lu] off %llu\n",
1315 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1316 pg_prev->pg, page_private(pg_prev->pg),
1317 pg_prev->pg->index, pg_prev->off);
1318 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1319 (pg->flag & OBD_BRW_SRVLOCK));
1321 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1322 requested_nob += pg->count;
1324 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1326 niobuf->len += pg->count;
1328 niobuf->offset = pg->off;
1329 niobuf->len = pg->count;
1330 niobuf->flags = pg->flag;
1335 LASSERTF((void *)(niobuf - niocount) ==
1336 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1337 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1338 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1340 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1342 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1343 body->oa.o_valid |= OBD_MD_FLFLAGS;
1344 body->oa.o_flags = 0;
1346 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1349 if (osc_should_shrink_grant(cli))
1350 osc_shrink_grant_local(cli, &body->oa);
1352 /* size[REQ_REC_OFF] still sizeof (*body) */
1353 if (opc == OST_WRITE) {
1354 if (cli->cl_checksum &&
1355 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1356 /* store cl_cksum_type in a local variable since
1357 * it can be changed via lprocfs */
1358 cksum_type_t cksum_type = cli->cl_cksum_type;
1360 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1361 oa->o_flags &= OBD_FL_LOCAL_MASK;
1362 body->oa.o_flags = 0;
1364 body->oa.o_flags |= cksum_type_pack(cksum_type);
1365 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1366 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1370 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1372 /* save this in 'oa', too, for later checking */
1373 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1374 oa->o_flags |= cksum_type_pack(cksum_type);
1376 /* clear out the checksum flag, in case this is a
1377 * resend but cl_checksum is no longer set. b=11238 */
1378 oa->o_valid &= ~OBD_MD_FLCKSUM;
1380 oa->o_cksum = body->oa.o_cksum;
1381 /* 1 RC per niobuf */
1382 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1383 sizeof(__u32) * niocount);
1385 if (cli->cl_checksum &&
1386 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1387 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1388 body->oa.o_flags = 0;
1389 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1390 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1393 ptlrpc_request_set_replen(req);
1395 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1396 aa = ptlrpc_req_async_args(req);
1398 aa->aa_requested_nob = requested_nob;
1399 aa->aa_nio_count = niocount;
1400 aa->aa_page_count = page_count;
1404 INIT_LIST_HEAD(&aa->aa_oaps);
1410 ptlrpc_req_finished(req);
1414 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1415 __u32 client_cksum, __u32 server_cksum, int nob,
1416 u32 page_count, struct brw_page **pga,
1417 cksum_type_t client_cksum_type)
1421 cksum_type_t cksum_type;
1423 if (server_cksum == client_cksum) {
1424 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1428 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1430 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1433 if (cksum_type != client_cksum_type)
1434 msg = "the server did not use the checksum type specified in the original request - likely a protocol problem"
1436 else if (new_cksum == server_cksum)
1437 msg = "changed on the client after we checksummed it - likely false positive due to mmap IO (bug 11742)"
1439 else if (new_cksum == client_cksum)
1440 msg = "changed in transit before arrival at OST";
1442 msg = "changed in transit AND doesn't match the original - likely false positive due to mmap IO (bug 11742)"
1445 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1446 " object "DOSTID" extent [%llu-%llu]\n",
1447 msg, libcfs_nid2str(peer->nid),
1448 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1449 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1450 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1451 POSTID(&oa->o_oi), pga[0]->off,
1452 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1453 CERROR("original client csum %x (type %x), server csum %x (type %x), client csum now %x\n",
1454 client_cksum, client_cksum_type,
1455 server_cksum, cksum_type, new_cksum);
1459 /* Note rc enters this function as number of bytes transferred */
1460 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1462 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1463 const lnet_process_id_t *peer =
1464 &req->rq_import->imp_connection->c_peer;
1465 struct client_obd *cli = aa->aa_cli;
1466 struct ost_body *body;
1467 __u32 client_cksum = 0;
1469 if (rc < 0 && rc != -EDQUOT) {
1470 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1474 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1475 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1477 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1481 /* set/clear over quota flag for a uid/gid */
1482 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1483 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1484 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1486 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1487 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1489 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1492 osc_update_grant(cli, body);
1497 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1498 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1500 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1502 CERROR("Unexpected +ve rc %d\n", rc);
1505 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1507 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1510 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1511 check_write_checksum(&body->oa, peer, client_cksum,
1512 body->oa.o_cksum, aa->aa_requested_nob,
1513 aa->aa_page_count, aa->aa_ppga,
1514 cksum_type_unpack(aa->aa_oa->o_flags)))
1517 rc = check_write_rcs(req, aa->aa_requested_nob,
1519 aa->aa_page_count, aa->aa_ppga);
1523 /* The rest of this function executes only for OST_READs */
1525 /* if unwrap_bulk failed, return -EAGAIN to retry */
1526 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1532 if (rc > aa->aa_requested_nob) {
1533 CERROR("Unexpected rc %d (%d requested)\n", rc,
1534 aa->aa_requested_nob);
1538 if (rc != req->rq_bulk->bd_nob_transferred) {
1539 CERROR("Unexpected rc %d (%d transferred)\n",
1540 rc, req->rq_bulk->bd_nob_transferred);
1544 if (rc < aa->aa_requested_nob)
1545 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1547 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1548 static int cksum_counter;
1549 __u32 server_cksum = body->oa.o_cksum;
1552 cksum_type_t cksum_type;
1554 cksum_type = cksum_type_unpack(body->oa.o_valid&OBD_MD_FLFLAGS ?
1555 body->oa.o_flags : 0);
1556 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1557 aa->aa_ppga, OST_READ,
1560 if (peer->nid == req->rq_bulk->bd_sender) {
1564 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1567 if (server_cksum != client_cksum) {
1568 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from %s%s%s inode " DFID " object " DOSTID " extent [%llu-%llu]\n",
1569 req->rq_import->imp_obd->obd_name,
1570 libcfs_nid2str(peer->nid),
1572 body->oa.o_valid & OBD_MD_FLFID ?
1573 body->oa.o_parent_seq : (__u64)0,
1574 body->oa.o_valid & OBD_MD_FLFID ?
1575 body->oa.o_parent_oid : 0,
1576 body->oa.o_valid & OBD_MD_FLFID ?
1577 body->oa.o_parent_ver : 0,
1578 POSTID(&body->oa.o_oi),
1579 aa->aa_ppga[0]->off,
1580 aa->aa_ppga[aa->aa_page_count-1]->off +
1581 aa->aa_ppga[aa->aa_page_count-1]->count -
1583 CERROR("client %x, server %x, cksum_type %x\n",
1584 client_cksum, server_cksum, cksum_type);
1586 aa->aa_oa->o_cksum = client_cksum;
1590 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1593 } else if (unlikely(client_cksum)) {
1594 static int cksum_missed;
1597 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1598 CERROR("Checksum %u requested from %s but not sent\n",
1599 cksum_missed, libcfs_nid2str(peer->nid));
1605 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1606 aa->aa_oa, &body->oa);
1611 static int osc_brw_redo_request(struct ptlrpc_request *request,
1612 struct osc_brw_async_args *aa, int rc)
1614 struct ptlrpc_request *new_req;
1615 struct osc_brw_async_args *new_aa;
1616 struct osc_async_page *oap;
1618 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1619 "redo for recoverable error %d", rc);
1621 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1622 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1623 aa->aa_cli, aa->aa_oa,
1624 NULL /* lsm unused by osc currently */,
1625 aa->aa_page_count, aa->aa_ppga,
1630 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1631 if (oap->oap_request != NULL) {
1632 LASSERTF(request == oap->oap_request,
1633 "request %p != oap_request %p\n",
1634 request, oap->oap_request);
1635 if (oap->oap_interrupted) {
1636 ptlrpc_req_finished(new_req);
1641 /* New request takes over pga and oaps from old request.
1642 * Note that copying a list_head doesn't work, need to move it... */
1644 new_req->rq_interpret_reply = request->rq_interpret_reply;
1645 new_req->rq_async_args = request->rq_async_args;
1646 /* cap resend delay to the current request timeout, this is similar to
1647 * what ptlrpc does (see after_reply()) */
1648 if (aa->aa_resends > new_req->rq_timeout)
1649 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1651 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1652 new_req->rq_generation_set = 1;
1653 new_req->rq_import_generation = request->rq_import_generation;
1655 new_aa = ptlrpc_req_async_args(new_req);
1657 INIT_LIST_HEAD(&new_aa->aa_oaps);
1658 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1659 INIT_LIST_HEAD(&new_aa->aa_exts);
1660 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1661 new_aa->aa_resends = aa->aa_resends;
1663 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1664 if (oap->oap_request) {
1665 ptlrpc_req_finished(oap->oap_request);
1666 oap->oap_request = ptlrpc_request_addref(new_req);
1670 /* XXX: This code will run into problem if we're going to support
1671 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1672 * and wait for all of them to be finished. We should inherit request
1673 * set from old request. */
1674 ptlrpcd_add_req(new_req);
1676 DEBUG_REQ(D_INFO, new_req, "new request");
1681 * ugh, we want disk allocation on the target to happen in offset order. we'll
1682 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1683 * fine for our small page arrays and doesn't require allocation. its an
1684 * insertion sort that swaps elements that are strides apart, shrinking the
1685 * stride down until its '1' and the array is sorted.
1687 static void sort_brw_pages(struct brw_page **array, int num)
1690 struct brw_page *tmp;
1694 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1699 for (i = stride ; i < num ; i++) {
1702 while (j >= stride && array[j - stride]->off > tmp->off) {
1703 array[j] = array[j - stride];
1708 } while (stride > 1);
1711 static void osc_release_ppga(struct brw_page **ppga, u32 count)
1713 LASSERT(ppga != NULL);
1717 static int brw_interpret(const struct lu_env *env,
1718 struct ptlrpc_request *req, void *data, int rc)
1720 struct osc_brw_async_args *aa = data;
1721 struct osc_extent *ext;
1722 struct osc_extent *tmp;
1723 struct cl_object *obj = NULL;
1724 struct client_obd *cli = aa->aa_cli;
1726 rc = osc_brw_fini_request(req, rc);
1727 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1728 /* When server return -EINPROGRESS, client should always retry
1729 * regardless of the number of times the bulk was resent already. */
1730 if (osc_recoverable_error(rc)) {
1731 if (req->rq_import_generation !=
1732 req->rq_import->imp_generation) {
1733 CDEBUG(D_HA, "%s: resend cross eviction for object: " DOSTID ", rc = %d.\n",
1734 req->rq_import->imp_obd->obd_name,
1735 POSTID(&aa->aa_oa->o_oi), rc);
1736 } else if (rc == -EINPROGRESS ||
1737 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1738 rc = osc_brw_redo_request(req, aa, rc);
1740 CERROR("%s: too many resent retries for object: %llu:%llu, rc = %d.\n",
1741 req->rq_import->imp_obd->obd_name,
1742 POSTID(&aa->aa_oa->o_oi), rc);
1747 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1751 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1752 if (obj == NULL && rc == 0) {
1753 obj = osc2cl(ext->oe_obj);
1757 list_del_init(&ext->oe_link);
1758 osc_extent_finish(env, ext, 1, rc);
1760 LASSERT(list_empty(&aa->aa_exts));
1761 LASSERT(list_empty(&aa->aa_oaps));
1764 struct obdo *oa = aa->aa_oa;
1765 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1766 unsigned long valid = 0;
1769 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1770 attr->cat_blocks = oa->o_blocks;
1771 valid |= CAT_BLOCKS;
1773 if (oa->o_valid & OBD_MD_FLMTIME) {
1774 attr->cat_mtime = oa->o_mtime;
1777 if (oa->o_valid & OBD_MD_FLATIME) {
1778 attr->cat_atime = oa->o_atime;
1781 if (oa->o_valid & OBD_MD_FLCTIME) {
1782 attr->cat_ctime = oa->o_ctime;
1786 cl_object_attr_lock(obj);
1787 cl_object_attr_set(env, obj, attr, valid);
1788 cl_object_attr_unlock(obj);
1790 cl_object_put(env, obj);
1792 OBDO_FREE(aa->aa_oa);
1794 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1795 req->rq_bulk->bd_nob_transferred);
1796 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1797 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1799 client_obd_list_lock(&cli->cl_loi_list_lock);
1800 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1801 * is called so we know whether to go to sync BRWs or wait for more
1802 * RPCs to complete */
1803 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1804 cli->cl_w_in_flight--;
1806 cli->cl_r_in_flight--;
1807 osc_wake_cache_waiters(cli);
1808 client_obd_list_unlock(&cli->cl_loi_list_lock);
1810 osc_io_unplug(env, cli, NULL);
1815 * Build an RPC by the list of extent @ext_list. The caller must ensure
1816 * that the total pages in this list are NOT over max pages per RPC.
1817 * Extents in the list must be in OES_RPC state.
1819 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1820 struct list_head *ext_list, int cmd)
1822 struct ptlrpc_request *req = NULL;
1823 struct osc_extent *ext;
1824 struct brw_page **pga = NULL;
1825 struct osc_brw_async_args *aa = NULL;
1826 struct obdo *oa = NULL;
1827 struct osc_async_page *oap;
1828 struct osc_async_page *tmp;
1829 struct cl_req *clerq = NULL;
1830 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1831 struct ldlm_lock *lock = NULL;
1832 struct cl_req_attr *crattr = NULL;
1833 u64 starting_offset = OBD_OBJECT_EOF;
1834 u64 ending_offset = 0;
1840 struct ost_body *body;
1841 LIST_HEAD(rpc_list);
1843 LASSERT(!list_empty(ext_list));
1845 /* add pages into rpc_list to build BRW rpc */
1846 list_for_each_entry(ext, ext_list, oe_link) {
1847 LASSERT(ext->oe_state == OES_RPC);
1848 mem_tight |= ext->oe_memalloc;
1849 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1851 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1852 if (starting_offset > oap->oap_obj_off)
1853 starting_offset = oap->oap_obj_off;
1855 LASSERT(oap->oap_page_off == 0);
1856 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1857 ending_offset = oap->oap_obj_off +
1860 LASSERT(oap->oap_page_off + oap->oap_count ==
1866 mpflag = cfs_memory_pressure_get_and_set();
1868 crattr = kzalloc(sizeof(*crattr), GFP_NOFS);
1874 pga = kcalloc(page_count, sizeof(*pga), GFP_NOFS);
1887 list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1888 struct cl_page *page = oap2cl_page(oap);
1889 if (clerq == NULL) {
1890 clerq = cl_req_alloc(env, page, crt,
1891 1 /* only 1-object rpcs for now */);
1892 if (IS_ERR(clerq)) {
1893 rc = PTR_ERR(clerq);
1896 lock = oap->oap_ldlm_lock;
1899 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1900 pga[i] = &oap->oap_brw_page;
1901 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1902 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1903 pga[i]->pg, page_index(oap->oap_page), oap,
1906 cl_req_page_add(env, clerq, page);
1909 /* always get the data for the obdo for the rpc */
1910 LASSERT(clerq != NULL);
1911 crattr->cra_oa = oa;
1912 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1914 oa->o_handle = lock->l_remote_handle;
1915 oa->o_valid |= OBD_MD_FLHANDLE;
1918 rc = cl_req_prep(env, clerq);
1920 CERROR("cl_req_prep failed: %d\n", rc);
1924 sort_brw_pages(pga, page_count);
1925 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1928 CERROR("prep_req failed: %d\n", rc);
1932 req->rq_interpret_reply = brw_interpret;
1935 req->rq_memalloc = 1;
1937 /* Need to update the timestamps after the request is built in case
1938 * we race with setattr (locally or in queue at OST). If OST gets
1939 * later setattr before earlier BRW (as determined by the request xid),
1940 * the OST will not use BRW timestamps. Sadly, there is no obvious
1941 * way to do this in a single call. bug 10150 */
1942 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1943 crattr->cra_oa = &body->oa;
1944 cl_req_attr_set(env, clerq, crattr,
1945 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
1947 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1949 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1950 aa = ptlrpc_req_async_args(req);
1951 INIT_LIST_HEAD(&aa->aa_oaps);
1952 list_splice_init(&rpc_list, &aa->aa_oaps);
1953 INIT_LIST_HEAD(&aa->aa_exts);
1954 list_splice_init(ext_list, &aa->aa_exts);
1955 aa->aa_clerq = clerq;
1957 /* queued sync pages can be torn down while the pages
1958 * were between the pending list and the rpc */
1960 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1961 /* only one oap gets a request reference */
1964 if (oap->oap_interrupted && !req->rq_intr) {
1965 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1967 ptlrpc_mark_interrupted(req);
1971 tmp->oap_request = ptlrpc_request_addref(req);
1973 client_obd_list_lock(&cli->cl_loi_list_lock);
1974 starting_offset >>= PAGE_CACHE_SHIFT;
1975 if (cmd == OBD_BRW_READ) {
1976 cli->cl_r_in_flight++;
1977 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1978 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1979 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
1980 starting_offset + 1);
1982 cli->cl_w_in_flight++;
1983 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1984 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
1985 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
1986 starting_offset + 1);
1988 client_obd_list_unlock(&cli->cl_loi_list_lock);
1990 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
1991 page_count, aa, cli->cl_r_in_flight,
1992 cli->cl_w_in_flight);
1994 ptlrpcd_add_req(req);
1999 cfs_memory_pressure_restore(mpflag);
2004 LASSERT(req == NULL);
2009 /* this should happen rarely and is pretty bad, it makes the
2010 * pending list not follow the dirty order */
2011 while (!list_empty(ext_list)) {
2012 ext = list_entry(ext_list->next, struct osc_extent,
2014 list_del_init(&ext->oe_link);
2015 osc_extent_finish(env, ext, 0, rc);
2017 if (clerq && !IS_ERR(clerq))
2018 cl_req_completion(env, clerq, rc);
2023 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2024 struct ldlm_enqueue_info *einfo)
2026 void *data = einfo->ei_cbdata;
2029 LASSERT(lock != NULL);
2030 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2031 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2032 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2033 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2035 lock_res_and_lock(lock);
2036 spin_lock(&osc_ast_guard);
2038 if (lock->l_ast_data == NULL)
2039 lock->l_ast_data = data;
2040 if (lock->l_ast_data == data)
2043 spin_unlock(&osc_ast_guard);
2044 unlock_res_and_lock(lock);
2049 static int osc_set_data_with_check(struct lustre_handle *lockh,
2050 struct ldlm_enqueue_info *einfo)
2052 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2056 set = osc_set_lock_data_with_check(lock, einfo);
2057 LDLM_LOCK_PUT(lock);
2059 CERROR("lockh %p, data %p - client evicted?\n",
2060 lockh, einfo->ei_cbdata);
2064 /* find any ldlm lock of the inode in osc
2068 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2069 ldlm_iterator_t replace, void *data)
2071 struct ldlm_res_id res_id;
2072 struct obd_device *obd = class_exp2obd(exp);
2075 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2076 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2077 if (rc == LDLM_ITER_STOP)
2079 if (rc == LDLM_ITER_CONTINUE)
2084 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2085 obd_enqueue_update_f upcall, void *cookie,
2086 __u64 *flags, int agl, int rc)
2088 int intent = *flags & LDLM_FL_HAS_INTENT;
2091 /* The request was created before ldlm_cli_enqueue call. */
2092 if (rc == ELDLM_LOCK_ABORTED) {
2093 struct ldlm_reply *rep;
2094 rep = req_capsule_server_get(&req->rq_pill,
2097 LASSERT(rep != NULL);
2098 rep->lock_policy_res1 =
2099 ptlrpc_status_ntoh(rep->lock_policy_res1);
2100 if (rep->lock_policy_res1)
2101 rc = rep->lock_policy_res1;
2105 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2107 *flags |= LDLM_FL_LVB_READY;
2108 CDEBUG(D_INODE, "got kms %llu blocks %llu mtime %llu\n",
2109 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2112 /* Call the update callback. */
2113 rc = (*upcall)(cookie, rc);
2117 static int osc_enqueue_interpret(const struct lu_env *env,
2118 struct ptlrpc_request *req,
2119 struct osc_enqueue_args *aa, int rc)
2121 struct ldlm_lock *lock;
2122 struct lustre_handle handle;
2124 struct ost_lvb *lvb;
2126 __u64 *flags = aa->oa_flags;
2128 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2129 * might be freed anytime after lock upcall has been called. */
2130 lustre_handle_copy(&handle, aa->oa_lockh);
2131 mode = aa->oa_ei->ei_mode;
2133 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2135 lock = ldlm_handle2lock(&handle);
2137 /* Take an additional reference so that a blocking AST that
2138 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2139 * to arrive after an upcall has been executed by
2140 * osc_enqueue_fini(). */
2141 ldlm_lock_addref(&handle, mode);
2143 /* Let CP AST to grant the lock first. */
2144 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2146 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2151 lvb_len = sizeof(*aa->oa_lvb);
2154 /* Complete obtaining the lock procedure. */
2155 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2156 mode, flags, lvb, lvb_len, &handle, rc);
2157 /* Complete osc stuff. */
2158 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2159 flags, aa->oa_agl, rc);
2161 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2163 /* Release the lock for async request. */
2164 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2166 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2167 * not already released by
2168 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2170 ldlm_lock_decref(&handle, mode);
2172 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2173 aa->oa_lockh, req, aa);
2174 ldlm_lock_decref(&handle, mode);
2175 LDLM_LOCK_PUT(lock);
2179 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2181 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2182 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2183 * other synchronous requests, however keeping some locks and trying to obtain
2184 * others may take a considerable amount of time in a case of ost failure; and
2185 * when other sync requests do not get released lock from a client, the client
2186 * is excluded from the cluster -- such scenarious make the life difficult, so
2187 * release locks just after they are obtained. */
2188 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2189 __u64 *flags, ldlm_policy_data_t *policy,
2190 struct ost_lvb *lvb, int kms_valid,
2191 obd_enqueue_update_f upcall, void *cookie,
2192 struct ldlm_enqueue_info *einfo,
2193 struct lustre_handle *lockh,
2194 struct ptlrpc_request_set *rqset, int async, int agl)
2196 struct obd_device *obd = exp->exp_obd;
2197 struct ptlrpc_request *req = NULL;
2198 int intent = *flags & LDLM_FL_HAS_INTENT;
2199 __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2203 /* Filesystem lock extents are extended to page boundaries so that
2204 * dealing with the page cache is a little smoother. */
2205 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2206 policy->l_extent.end |= ~CFS_PAGE_MASK;
2209 * kms is not valid when either object is completely fresh (so that no
2210 * locks are cached), or object was evicted. In the latter case cached
2211 * lock cannot be used, because it would prime inode state with
2212 * potentially stale LVB.
2217 /* Next, search for already existing extent locks that will cover us */
2218 /* If we're trying to read, we also search for an existing PW lock. The
2219 * VFS and page cache already protect us locally, so lots of readers/
2220 * writers can share a single PW lock.
2222 * There are problems with conversion deadlocks, so instead of
2223 * converting a read lock to a write lock, we'll just enqueue a new
2226 * At some point we should cancel the read lock instead of making them
2227 * send us a blocking callback, but there are problems with canceling
2228 * locks out from other users right now, too. */
2229 mode = einfo->ei_mode;
2230 if (einfo->ei_mode == LCK_PR)
2232 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2233 einfo->ei_type, policy, mode, lockh, 0);
2235 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2237 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2238 /* For AGL, if enqueue RPC is sent but the lock is not
2239 * granted, then skip to process this strpe.
2240 * Return -ECANCELED to tell the caller. */
2241 ldlm_lock_decref(lockh, mode);
2242 LDLM_LOCK_PUT(matched);
2246 if (osc_set_lock_data_with_check(matched, einfo)) {
2247 *flags |= LDLM_FL_LVB_READY;
2248 /* addref the lock only if not async requests and PW
2249 * lock is matched whereas we asked for PR. */
2250 if (!rqset && einfo->ei_mode != mode)
2251 ldlm_lock_addref(lockh, LCK_PR);
2253 /* I would like to be able to ASSERT here that
2254 * rss <= kms, but I can't, for reasons which
2255 * are explained in lov_enqueue() */
2258 /* We already have a lock, and it's referenced.
2260 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2261 * AGL upcall may change it to CLS_HELD directly. */
2262 (*upcall)(cookie, ELDLM_OK);
2264 if (einfo->ei_mode != mode)
2265 ldlm_lock_decref(lockh, LCK_PW);
2267 /* For async requests, decref the lock. */
2268 ldlm_lock_decref(lockh, einfo->ei_mode);
2269 LDLM_LOCK_PUT(matched);
2273 ldlm_lock_decref(lockh, mode);
2274 LDLM_LOCK_PUT(matched);
2280 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2281 &RQF_LDLM_ENQUEUE_LVB);
2285 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2287 ptlrpc_request_free(req);
2291 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2293 ptlrpc_request_set_replen(req);
2296 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2297 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2299 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2300 sizeof(*lvb), LVB_T_OST, lockh, async);
2303 struct osc_enqueue_args *aa;
2304 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2305 aa = ptlrpc_req_async_args(req);
2308 aa->oa_flags = flags;
2309 aa->oa_upcall = upcall;
2310 aa->oa_cookie = cookie;
2312 aa->oa_lockh = lockh;
2315 req->rq_interpret_reply =
2316 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2317 if (rqset == PTLRPCD_SET)
2318 ptlrpcd_add_req(req);
2320 ptlrpc_set_add_req(rqset, req);
2321 } else if (intent) {
2322 ptlrpc_req_finished(req);
2327 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2329 ptlrpc_req_finished(req);
2334 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2335 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2336 __u64 *flags, void *data, struct lustre_handle *lockh,
2339 struct obd_device *obd = exp->exp_obd;
2340 __u64 lflags = *flags;
2343 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2346 /* Filesystem lock extents are extended to page boundaries so that
2347 * dealing with the page cache is a little smoother */
2348 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2349 policy->l_extent.end |= ~CFS_PAGE_MASK;
2351 /* Next, search for already existing extent locks that will cover us */
2352 /* If we're trying to read, we also search for an existing PW lock. The
2353 * VFS and page cache already protect us locally, so lots of readers/
2354 * writers can share a single PW lock. */
2358 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2359 res_id, type, policy, rc, lockh, unref);
2362 if (!osc_set_data_with_check(lockh, data)) {
2363 if (!(lflags & LDLM_FL_TEST_LOCK))
2364 ldlm_lock_decref(lockh, rc);
2368 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2369 ldlm_lock_addref(lockh, LCK_PR);
2370 ldlm_lock_decref(lockh, LCK_PW);
2377 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2379 if (unlikely(mode == LCK_GROUP))
2380 ldlm_lock_decref_and_cancel(lockh, mode);
2382 ldlm_lock_decref(lockh, mode);
2387 static int osc_statfs_interpret(const struct lu_env *env,
2388 struct ptlrpc_request *req,
2389 struct osc_async_args *aa, int rc)
2391 struct obd_statfs *msfs;
2394 /* The request has in fact never been sent
2395 * due to issues at a higher level (LOV).
2396 * Exit immediately since the caller is
2397 * aware of the problem and takes care
2398 * of the clean up */
2401 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2402 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY)) {
2410 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2416 *aa->aa_oi->oi_osfs = *msfs;
2418 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2422 static int osc_statfs_async(struct obd_export *exp,
2423 struct obd_info *oinfo, __u64 max_age,
2424 struct ptlrpc_request_set *rqset)
2426 struct obd_device *obd = class_exp2obd(exp);
2427 struct ptlrpc_request *req;
2428 struct osc_async_args *aa;
2431 /* We could possibly pass max_age in the request (as an absolute
2432 * timestamp or a "seconds.usec ago") so the target can avoid doing
2433 * extra calls into the filesystem if that isn't necessary (e.g.
2434 * during mount that would help a bit). Having relative timestamps
2435 * is not so great if request processing is slow, while absolute
2436 * timestamps are not ideal because they need time synchronization. */
2437 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2441 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2443 ptlrpc_request_free(req);
2446 ptlrpc_request_set_replen(req);
2447 req->rq_request_portal = OST_CREATE_PORTAL;
2448 ptlrpc_at_set_req_timeout(req);
2450 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2451 /* procfs requests not want stat in wait for avoid deadlock */
2452 req->rq_no_resend = 1;
2453 req->rq_no_delay = 1;
2456 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2457 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2458 aa = ptlrpc_req_async_args(req);
2461 ptlrpc_set_add_req(rqset, req);
2465 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2466 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2468 struct obd_device *obd = class_exp2obd(exp);
2469 struct obd_statfs *msfs;
2470 struct ptlrpc_request *req;
2471 struct obd_import *imp = NULL;
2474 /*Since the request might also come from lprocfs, so we need
2475 *sync this with client_disconnect_export Bug15684*/
2476 down_read(&obd->u.cli.cl_sem);
2477 if (obd->u.cli.cl_import)
2478 imp = class_import_get(obd->u.cli.cl_import);
2479 up_read(&obd->u.cli.cl_sem);
2483 /* We could possibly pass max_age in the request (as an absolute
2484 * timestamp or a "seconds.usec ago") so the target can avoid doing
2485 * extra calls into the filesystem if that isn't necessary (e.g.
2486 * during mount that would help a bit). Having relative timestamps
2487 * is not so great if request processing is slow, while absolute
2488 * timestamps are not ideal because they need time synchronization. */
2489 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2491 class_import_put(imp);
2496 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2498 ptlrpc_request_free(req);
2501 ptlrpc_request_set_replen(req);
2502 req->rq_request_portal = OST_CREATE_PORTAL;
2503 ptlrpc_at_set_req_timeout(req);
2505 if (flags & OBD_STATFS_NODELAY) {
2506 /* procfs requests not want stat in wait for avoid deadlock */
2507 req->rq_no_resend = 1;
2508 req->rq_no_delay = 1;
2511 rc = ptlrpc_queue_wait(req);
2515 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2524 ptlrpc_req_finished(req);
2528 /* Retrieve object striping information.
2530 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2531 * the maximum number of OST indices which will fit in the user buffer.
2532 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2534 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2536 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2537 struct lov_user_md_v3 lum, *lumk;
2538 struct lov_user_ost_data_v1 *lmm_objects;
2539 int rc = 0, lum_size;
2544 /* we only need the header part from user space to get lmm_magic and
2545 * lmm_stripe_count, (the header part is common to v1 and v3) */
2546 lum_size = sizeof(struct lov_user_md_v1);
2547 if (copy_from_user(&lum, lump, lum_size))
2550 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2551 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2554 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2555 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2556 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2557 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2559 /* we can use lov_mds_md_size() to compute lum_size
2560 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2561 if (lum.lmm_stripe_count > 0) {
2562 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2563 lumk = kzalloc(lum_size, GFP_NOFS);
2567 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2569 &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2571 lmm_objects = &(lumk->lmm_objects[0]);
2572 lmm_objects->l_ost_oi = lsm->lsm_oi;
2574 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2578 lumk->lmm_oi = lsm->lsm_oi;
2579 lumk->lmm_stripe_count = 1;
2581 if (copy_to_user(lump, lumk, lum_size))
2591 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2592 void *karg, void *uarg)
2594 struct obd_device *obd = exp->exp_obd;
2595 struct obd_ioctl_data *data = karg;
2598 if (!try_module_get(THIS_MODULE)) {
2599 CERROR("Can't get module. Is it alive?");
2603 case OBD_IOC_LOV_GET_CONFIG: {
2605 struct lov_desc *desc;
2606 struct obd_uuid uuid;
2610 if (obd_ioctl_getdata(&buf, &len, uarg)) {
2615 data = (struct obd_ioctl_data *)buf;
2617 if (sizeof(*desc) > data->ioc_inllen1) {
2618 obd_ioctl_freedata(buf, len);
2623 if (data->ioc_inllen2 < sizeof(uuid)) {
2624 obd_ioctl_freedata(buf, len);
2629 desc = (struct lov_desc *)data->ioc_inlbuf1;
2630 desc->ld_tgt_count = 1;
2631 desc->ld_active_tgt_count = 1;
2632 desc->ld_default_stripe_count = 1;
2633 desc->ld_default_stripe_size = 0;
2634 desc->ld_default_stripe_offset = 0;
2635 desc->ld_pattern = 0;
2636 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2638 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2640 err = copy_to_user(uarg, buf, len);
2643 obd_ioctl_freedata(buf, len);
2646 case LL_IOC_LOV_SETSTRIPE:
2647 err = obd_alloc_memmd(exp, karg);
2651 case LL_IOC_LOV_GETSTRIPE:
2652 err = osc_getstripe(karg, uarg);
2654 case OBD_IOC_CLIENT_RECOVER:
2655 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2656 data->ioc_inlbuf1, 0);
2660 case IOC_OSC_SET_ACTIVE:
2661 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2664 case OBD_IOC_POLL_QUOTACHECK:
2665 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2667 case OBD_IOC_PING_TARGET:
2668 err = ptlrpc_obd_ping(obd);
2671 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2672 cmd, current_comm());
2677 module_put(THIS_MODULE);
2681 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2682 u32 keylen, void *key, __u32 *vallen, void *val,
2683 struct lov_stripe_md *lsm)
2685 if (!vallen || !val)
2688 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2689 __u32 *stripe = val;
2690 *vallen = sizeof(*stripe);
2693 } else if (KEY_IS(KEY_LAST_ID)) {
2694 struct ptlrpc_request *req;
2699 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2700 &RQF_OST_GET_INFO_LAST_ID);
2704 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2705 RCL_CLIENT, keylen);
2706 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2708 ptlrpc_request_free(req);
2712 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2713 memcpy(tmp, key, keylen);
2715 req->rq_no_delay = req->rq_no_resend = 1;
2716 ptlrpc_request_set_replen(req);
2717 rc = ptlrpc_queue_wait(req);
2721 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2722 if (reply == NULL) {
2727 *((u64 *)val) = *reply;
2729 ptlrpc_req_finished(req);
2731 } else if (KEY_IS(KEY_FIEMAP)) {
2732 struct ll_fiemap_info_key *fm_key =
2733 (struct ll_fiemap_info_key *)key;
2734 struct ldlm_res_id res_id;
2735 ldlm_policy_data_t policy;
2736 struct lustre_handle lockh;
2737 ldlm_mode_t mode = 0;
2738 struct ptlrpc_request *req;
2739 struct ll_user_fiemap *reply;
2743 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2746 policy.l_extent.start = fm_key->fiemap.fm_start &
2749 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2750 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
2751 policy.l_extent.end = OBD_OBJECT_EOF;
2753 policy.l_extent.end = (fm_key->fiemap.fm_start +
2754 fm_key->fiemap.fm_length +
2755 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
2757 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2758 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2759 LDLM_FL_BLOCK_GRANTED |
2761 &res_id, LDLM_EXTENT, &policy,
2762 LCK_PR | LCK_PW, &lockh, 0);
2763 if (mode) { /* lock is cached on client */
2764 if (mode != LCK_PR) {
2765 ldlm_lock_addref(&lockh, LCK_PR);
2766 ldlm_lock_decref(&lockh, LCK_PW);
2768 } else { /* no cached lock, needs acquire lock on server side */
2769 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2770 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2774 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2775 &RQF_OST_GET_INFO_FIEMAP);
2781 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2782 RCL_CLIENT, keylen);
2783 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2784 RCL_CLIENT, *vallen);
2785 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2786 RCL_SERVER, *vallen);
2788 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2790 ptlrpc_request_free(req);
2794 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2795 memcpy(tmp, key, keylen);
2796 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2797 memcpy(tmp, val, *vallen);
2799 ptlrpc_request_set_replen(req);
2800 rc = ptlrpc_queue_wait(req);
2804 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2805 if (reply == NULL) {
2810 memcpy(val, reply, *vallen);
2812 ptlrpc_req_finished(req);
2815 ldlm_lock_decref(&lockh, LCK_PR);
2822 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2823 u32 keylen, void *key, u32 vallen,
2824 void *val, struct ptlrpc_request_set *set)
2826 struct ptlrpc_request *req;
2827 struct obd_device *obd = exp->exp_obd;
2828 struct obd_import *imp = class_exp2cliimp(exp);
2832 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2834 if (KEY_IS(KEY_CHECKSUM)) {
2835 if (vallen != sizeof(int))
2837 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2841 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2842 sptlrpc_conf_client_adapt(obd);
2846 if (KEY_IS(KEY_FLUSH_CTX)) {
2847 sptlrpc_import_flush_my_ctx(imp);
2851 if (KEY_IS(KEY_CACHE_SET)) {
2852 struct client_obd *cli = &obd->u.cli;
2854 LASSERT(cli->cl_cache == NULL); /* only once */
2855 cli->cl_cache = (struct cl_client_cache *)val;
2856 atomic_inc(&cli->cl_cache->ccc_users);
2857 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2859 /* add this osc into entity list */
2860 LASSERT(list_empty(&cli->cl_lru_osc));
2861 spin_lock(&cli->cl_cache->ccc_lru_lock);
2862 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2863 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2868 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2869 struct client_obd *cli = &obd->u.cli;
2870 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2871 int target = *(int *)val;
2873 nr = osc_lru_shrink(cli, min(nr, target));
2878 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2881 /* We pass all other commands directly to OST. Since nobody calls osc
2882 methods directly and everybody is supposed to go through LOV, we
2883 assume lov checked invalid values for us.
2884 The only recognised values so far are evict_by_nid and mds_conn.
2885 Even if something bad goes through, we'd get a -EINVAL from OST
2888 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2889 &RQF_OST_SET_GRANT_INFO :
2894 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2895 RCL_CLIENT, keylen);
2896 if (!KEY_IS(KEY_GRANT_SHRINK))
2897 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2898 RCL_CLIENT, vallen);
2899 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2901 ptlrpc_request_free(req);
2905 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2906 memcpy(tmp, key, keylen);
2907 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2910 memcpy(tmp, val, vallen);
2912 if (KEY_IS(KEY_GRANT_SHRINK)) {
2913 struct osc_brw_async_args *aa;
2916 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2917 aa = ptlrpc_req_async_args(req);
2920 ptlrpc_req_finished(req);
2923 *oa = ((struct ost_body *)val)->oa;
2925 req->rq_interpret_reply = osc_shrink_grant_interpret;
2928 ptlrpc_request_set_replen(req);
2929 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2930 LASSERT(set != NULL);
2931 ptlrpc_set_add_req(set, req);
2932 ptlrpc_check_set(NULL, set);
2934 ptlrpcd_add_req(req);
2940 static int osc_reconnect(const struct lu_env *env,
2941 struct obd_export *exp, struct obd_device *obd,
2942 struct obd_uuid *cluuid,
2943 struct obd_connect_data *data,
2946 struct client_obd *cli = &obd->u.cli;
2948 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2951 client_obd_list_lock(&cli->cl_loi_list_lock);
2952 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2953 2 * cli_brw_size(obd);
2954 lost_grant = cli->cl_lost_grant;
2955 cli->cl_lost_grant = 0;
2956 client_obd_list_unlock(&cli->cl_loi_list_lock);
2958 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d ocd_grant: %d, lost: %ld.\n",
2959 data->ocd_connect_flags,
2960 data->ocd_version, data->ocd_grant, lost_grant);
2966 static int osc_disconnect(struct obd_export *exp)
2968 struct obd_device *obd = class_exp2obd(exp);
2971 rc = client_disconnect_export(exp);
2973 * Initially we put del_shrink_grant before disconnect_export, but it
2974 * causes the following problem if setup (connect) and cleanup
2975 * (disconnect) are tangled together.
2976 * connect p1 disconnect p2
2977 * ptlrpc_connect_import
2978 * ............... class_manual_cleanup
2981 * ptlrpc_connect_interrupt
2983 * add this client to shrink list
2985 * Bang! pinger trigger the shrink.
2986 * So the osc should be disconnected from the shrink list, after we
2987 * are sure the import has been destroyed. BUG18662
2989 if (obd->u.cli.cl_import == NULL)
2990 osc_del_shrink_grant(&obd->u.cli);
2994 static int osc_import_event(struct obd_device *obd,
2995 struct obd_import *imp,
2996 enum obd_import_event event)
2998 struct client_obd *cli;
3001 LASSERT(imp->imp_obd == obd);
3004 case IMP_EVENT_DISCON: {
3006 client_obd_list_lock(&cli->cl_loi_list_lock);
3007 cli->cl_avail_grant = 0;
3008 cli->cl_lost_grant = 0;
3009 client_obd_list_unlock(&cli->cl_loi_list_lock);
3012 case IMP_EVENT_INACTIVE: {
3013 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3016 case IMP_EVENT_INVALIDATE: {
3017 struct ldlm_namespace *ns = obd->obd_namespace;
3021 env = cl_env_get(&refcheck);
3025 /* all pages go to failing rpcs due to the invalid
3027 osc_io_unplug(env, cli, NULL);
3029 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3030 cl_env_put(env, &refcheck);
3035 case IMP_EVENT_ACTIVE: {
3036 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3039 case IMP_EVENT_OCD: {
3040 struct obd_connect_data *ocd = &imp->imp_connect_data;
3042 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3043 osc_init_grant(&obd->u.cli, ocd);
3046 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3047 imp->imp_client->cli_request_portal = OST_REQUEST_PORTAL;
3049 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3052 case IMP_EVENT_DEACTIVATE: {
3053 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3056 case IMP_EVENT_ACTIVATE: {
3057 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3061 CERROR("Unknown import event %d\n", event);
3068 * Determine whether the lock can be canceled before replaying the lock
3069 * during recovery, see bug16774 for detailed information.
3071 * \retval zero the lock can't be canceled
3072 * \retval other ok to cancel
3074 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3076 check_res_locked(lock->l_resource);
3079 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3081 * XXX as a future improvement, we can also cancel unused write lock
3082 * if it doesn't have dirty data and active mmaps.
3084 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3085 (lock->l_granted_mode == LCK_PR ||
3086 lock->l_granted_mode == LCK_CR) &&
3087 (osc_dlm_lock_pageref(lock) == 0))
3093 static int brw_queue_work(const struct lu_env *env, void *data)
3095 struct client_obd *cli = data;
3097 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3099 osc_io_unplug(env, cli, NULL);
3103 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3105 struct lprocfs_static_vars lvars = { NULL };
3106 struct client_obd *cli = &obd->u.cli;
3113 rc = ptlrpcd_addref();
3117 rc = client_obd_setup(obd, lcfg);
3121 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3122 if (IS_ERR(handler)) {
3123 rc = PTR_ERR(handler);
3124 goto out_client_setup;
3126 cli->cl_writeback_work = handler;
3128 rc = osc_quota_setup(obd);
3130 goto out_ptlrpcd_work;
3132 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3133 lprocfs_osc_init_vars(&lvars);
3134 if (lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars) == 0) {
3135 lproc_osc_attach_seqstat(obd);
3136 sptlrpc_lprocfs_cliobd_attach(obd);
3137 ptlrpc_lprocfs_register_obd(obd);
3141 * We try to control the total number of requests with a upper limit
3142 * osc_reqpool_maxreqcount. There might be some race which will cause
3143 * over-limit allocation, but it is fine.
3145 req_count = atomic_read(&osc_pool_req_count);
3146 if (req_count < osc_reqpool_maxreqcount) {
3147 adding = cli->cl_max_rpcs_in_flight + 2;
3148 if (req_count + adding > osc_reqpool_maxreqcount)
3149 adding = osc_reqpool_maxreqcount - req_count;
3151 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3152 atomic_add(added, &osc_pool_req_count);
3155 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3156 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3160 ptlrpcd_destroy_work(handler);
3162 client_obd_cleanup(obd);
3168 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3171 case OBD_CLEANUP_EARLY: {
3172 struct obd_import *imp;
3173 imp = obd->u.cli.cl_import;
3174 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3175 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3176 ptlrpc_deactivate_import(imp);
3177 spin_lock(&imp->imp_lock);
3178 imp->imp_pingable = 0;
3179 spin_unlock(&imp->imp_lock);
3182 case OBD_CLEANUP_EXPORTS: {
3183 struct client_obd *cli = &obd->u.cli;
3185 * for echo client, export may be on zombie list, wait for
3186 * zombie thread to cull it, because cli.cl_import will be
3187 * cleared in client_disconnect_export():
3188 * class_export_destroy() -> obd_cleanup() ->
3189 * echo_device_free() -> echo_client_cleanup() ->
3190 * obd_disconnect() -> osc_disconnect() ->
3191 * client_disconnect_export()
3193 obd_zombie_barrier();
3194 if (cli->cl_writeback_work) {
3195 ptlrpcd_destroy_work(cli->cl_writeback_work);
3196 cli->cl_writeback_work = NULL;
3198 obd_cleanup_client_import(obd);
3199 ptlrpc_lprocfs_unregister_obd(obd);
3200 lprocfs_obd_cleanup(obd);
3207 int osc_cleanup(struct obd_device *obd)
3209 struct client_obd *cli = &obd->u.cli;
3213 if (cli->cl_cache != NULL) {
3214 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3215 spin_lock(&cli->cl_cache->ccc_lru_lock);
3216 list_del_init(&cli->cl_lru_osc);
3217 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3218 cli->cl_lru_left = NULL;
3219 atomic_dec(&cli->cl_cache->ccc_users);
3220 cli->cl_cache = NULL;
3223 /* free memory of osc quota cache */
3224 osc_quota_cleanup(obd);
3226 rc = client_obd_cleanup(obd);
3232 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3234 struct lprocfs_static_vars lvars = { NULL };
3237 lprocfs_osc_init_vars(&lvars);
3239 switch (lcfg->lcfg_command) {
3241 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3251 static int osc_process_config(struct obd_device *obd, u32 len, void *buf)
3253 return osc_process_config_base(obd, buf);
3256 struct obd_ops osc_obd_ops = {
3257 .o_owner = THIS_MODULE,
3258 .o_setup = osc_setup,
3259 .o_precleanup = osc_precleanup,
3260 .o_cleanup = osc_cleanup,
3261 .o_add_conn = client_import_add_conn,
3262 .o_del_conn = client_import_del_conn,
3263 .o_connect = client_connect_import,
3264 .o_reconnect = osc_reconnect,
3265 .o_disconnect = osc_disconnect,
3266 .o_statfs = osc_statfs,
3267 .o_statfs_async = osc_statfs_async,
3268 .o_packmd = osc_packmd,
3269 .o_unpackmd = osc_unpackmd,
3270 .o_create = osc_create,
3271 .o_destroy = osc_destroy,
3272 .o_getattr = osc_getattr,
3273 .o_getattr_async = osc_getattr_async,
3274 .o_setattr = osc_setattr,
3275 .o_setattr_async = osc_setattr_async,
3276 .o_find_cbdata = osc_find_cbdata,
3277 .o_iocontrol = osc_iocontrol,
3278 .o_get_info = osc_get_info,
3279 .o_set_info_async = osc_set_info_async,
3280 .o_import_event = osc_import_event,
3281 .o_process_config = osc_process_config,
3282 .o_quotactl = osc_quotactl,
3283 .o_quotacheck = osc_quotacheck,
3286 extern struct lu_kmem_descr osc_caches[];
3287 extern spinlock_t osc_ast_guard;
3288 extern struct lock_class_key osc_ast_guard_class;
3290 static int __init osc_init(void)
3292 struct lprocfs_static_vars lvars = { NULL };
3293 unsigned int reqpool_size;
3294 unsigned int reqsize;
3297 /* print an address of _any_ initialized kernel symbol from this
3298 * module, to allow debugging with gdb that doesn't support data
3299 * symbols from modules.*/
3300 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3302 rc = lu_kmem_init(osc_caches);
3306 lprocfs_osc_init_vars(&lvars);
3308 rc = class_register_type(&osc_obd_ops, NULL,
3309 LUSTRE_OSC_NAME, &osc_device_type);
3313 spin_lock_init(&osc_ast_guard);
3314 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3316 /* This is obviously too much memory, only prevent overflow here */
3317 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0) {
3322 reqpool_size = osc_reqpool_mem_max << 20;
3325 while (reqsize < OST_MAXREQSIZE)
3326 reqsize = reqsize << 1;
3329 * We don't enlarge the request count in OSC pool according to
3330 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3331 * tried after normal allocation failed. So a small OSC pool won't
3332 * cause much performance degression in most of cases.
3334 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3336 atomic_set(&osc_pool_req_count, 0);
3337 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_MAXREQSIZE,
3338 ptlrpc_add_rqs_to_pool);
3346 class_unregister_type(LUSTRE_OSC_NAME);
3348 lu_kmem_fini(osc_caches);
3352 static void /*__exit*/ osc_exit(void)
3354 class_unregister_type(LUSTRE_OSC_NAME);
3355 lu_kmem_fini(osc_caches);
3356 ptlrpc_free_rq_pool(osc_rq_pool);
3359 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3360 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3361 MODULE_LICENSE("GPL");
3362 MODULE_VERSION(LUSTRE_VERSION_STRING);
3364 module_init(osc_init);
3365 module_exit(osc_exit);