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, 2015, 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"
41 #include "../include/lustre_dlm.h"
42 #include "../include/lustre_net.h"
43 #include "../include/lustre/lustre_user.h"
44 #include "../include/obd_cksum.h"
46 #include "../include/lustre_ha.h"
47 #include "../include/lprocfs_status.h"
48 #include "../include/lustre_debug.h"
49 #include "../include/lustre_param.h"
50 #include "../include/lustre_fid.h"
51 #include "../include/obd_class.h"
52 #include "../include/obd.h"
53 #include "osc_internal.h"
54 #include "osc_cl_internal.h"
56 atomic_t osc_pool_req_count;
57 unsigned int osc_reqpool_maxreqcount;
58 struct ptlrpc_request_pool *osc_rq_pool;
60 /* max memory used for request pool, unit is MB */
61 static unsigned int osc_reqpool_mem_max = 5;
62 module_param(osc_reqpool_mem_max, uint, 0444);
64 struct osc_brw_async_args {
70 struct brw_page **aa_ppga;
71 struct client_obd *aa_cli;
72 struct list_head aa_oaps;
73 struct list_head aa_exts;
74 struct cl_req *aa_clerq;
77 struct osc_async_args {
78 struct obd_info *aa_oi;
81 struct osc_setattr_args {
83 obd_enqueue_update_f sa_upcall;
87 struct osc_fsync_args {
88 struct obd_info *fa_oi;
89 obd_enqueue_update_f fa_upcall;
93 struct osc_enqueue_args {
94 struct obd_export *oa_exp;
95 enum ldlm_type oa_type;
96 enum ldlm_mode oa_mode;
98 osc_enqueue_upcall_f oa_upcall;
100 struct ost_lvb *oa_lvb;
101 struct lustre_handle oa_lockh;
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);
109 /* Pack OSC object metadata for disk storage (LE byte order). */
110 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
111 struct lov_stripe_md *lsm)
115 lmm_size = sizeof(**lmmp);
123 } else if (unlikely(lsm && ostid_id(&lsm->lsm_oi) == 0)) {
128 *lmmp = kzalloc(lmm_size, GFP_NOFS);
134 ostid_cpu_to_le(&lsm->lsm_oi, &(*lmmp)->lmm_oi);
139 /* Unpack OSC object metadata from disk storage (LE byte order). */
140 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
141 struct lov_mds_md *lmm, int lmm_bytes)
144 struct obd_import *imp = class_exp2cliimp(exp);
147 if (lmm_bytes < sizeof(*lmm)) {
148 CERROR("%s: lov_mds_md too small: %d, need %d\n",
149 exp->exp_obd->obd_name, lmm_bytes,
153 /* XXX LOV_MAGIC etc check? */
155 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
156 CERROR("%s: zero lmm_object_id: rc = %d\n",
157 exp->exp_obd->obd_name, -EINVAL);
162 lsm_size = lov_stripe_md_size(1);
167 kfree((*lsmp)->lsm_oinfo[0]);
174 *lsmp = kzalloc(lsm_size, GFP_NOFS);
175 if (unlikely(!*lsmp))
177 (*lsmp)->lsm_oinfo[0] = kzalloc(sizeof(struct lov_oinfo),
179 if (unlikely(!(*lsmp)->lsm_oinfo[0])) {
183 loi_init((*lsmp)->lsm_oinfo[0]);
184 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
189 /* XXX zero *lsmp? */
190 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
193 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
194 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
196 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
201 static inline void osc_pack_req_body(struct ptlrpc_request *req,
202 struct obd_info *oinfo)
204 struct ost_body *body;
206 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
209 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
213 static int osc_getattr_interpret(const struct lu_env *env,
214 struct ptlrpc_request *req,
215 struct osc_async_args *aa, int rc)
217 struct ost_body *body;
222 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
224 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
225 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
226 aa->aa_oi->oi_oa, &body->oa);
228 /* This should really be sent by the OST */
229 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
230 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
232 CDEBUG(D_INFO, "can't unpack ost_body\n");
234 aa->aa_oi->oi_oa->o_valid = 0;
237 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
241 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
242 struct ptlrpc_request_set *set)
244 struct ptlrpc_request *req;
245 struct osc_async_args *aa;
248 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
252 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
254 ptlrpc_request_free(req);
258 osc_pack_req_body(req, oinfo);
260 ptlrpc_request_set_replen(req);
261 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
263 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
264 aa = ptlrpc_req_async_args(req);
267 ptlrpc_set_add_req(set, req);
271 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
272 struct obd_info *oinfo)
274 struct ptlrpc_request *req;
275 struct ost_body *body;
278 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
282 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
284 ptlrpc_request_free(req);
288 osc_pack_req_body(req, oinfo);
290 ptlrpc_request_set_replen(req);
292 rc = ptlrpc_queue_wait(req);
296 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
302 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
303 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
306 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
307 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
310 ptlrpc_req_finished(req);
314 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
315 struct obd_info *oinfo, struct obd_trans_info *oti)
317 struct ptlrpc_request *req;
318 struct ost_body *body;
321 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
323 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
327 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
329 ptlrpc_request_free(req);
333 osc_pack_req_body(req, oinfo);
335 ptlrpc_request_set_replen(req);
337 rc = ptlrpc_queue_wait(req);
341 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
347 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
351 ptlrpc_req_finished(req);
355 static int osc_setattr_interpret(const struct lu_env *env,
356 struct ptlrpc_request *req,
357 struct osc_setattr_args *sa, int rc)
359 struct ost_body *body;
364 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
370 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
373 rc = sa->sa_upcall(sa->sa_cookie, rc);
377 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
378 struct obd_trans_info *oti,
379 obd_enqueue_update_f upcall, void *cookie,
380 struct ptlrpc_request_set *rqset)
382 struct ptlrpc_request *req;
383 struct osc_setattr_args *sa;
386 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
390 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
392 ptlrpc_request_free(req);
396 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
397 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
399 osc_pack_req_body(req, oinfo);
401 ptlrpc_request_set_replen(req);
403 /* do mds to ost setattr asynchronously */
405 /* Do not wait for response. */
406 ptlrpcd_add_req(req);
408 req->rq_interpret_reply =
409 (ptlrpc_interpterer_t)osc_setattr_interpret;
411 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
412 sa = ptlrpc_req_async_args(req);
413 sa->sa_oa = oinfo->oi_oa;
414 sa->sa_upcall = upcall;
415 sa->sa_cookie = cookie;
417 if (rqset == PTLRPCD_SET)
418 ptlrpcd_add_req(req);
420 ptlrpc_set_add_req(rqset, req);
426 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
427 struct obd_trans_info *oti,
428 struct ptlrpc_request_set *rqset)
430 return osc_setattr_async_base(exp, oinfo, oti,
431 oinfo->oi_cb_up, oinfo, rqset);
434 static int osc_real_create(struct obd_export *exp, struct obdo *oa,
435 struct lov_stripe_md **ea,
436 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.
634 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
635 struct list_head *cancels,
636 enum ldlm_mode mode, __u64 lock_flags)
638 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
639 struct ldlm_res_id res_id;
640 struct ldlm_resource *res;
643 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
644 * export) but disabled through procfs (flag in NS).
646 * This distinguishes from a case when ELC is not supported originally,
647 * when we still want to cancel locks in advance and just cancel them
648 * locally, without sending any RPC.
650 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
653 ostid_build_res_name(&oa->o_oi, &res_id);
654 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
658 LDLM_RESOURCE_ADDREF(res);
659 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
660 lock_flags, 0, NULL);
661 LDLM_RESOURCE_DELREF(res);
662 ldlm_resource_putref(res);
666 static int osc_destroy_interpret(const struct lu_env *env,
667 struct ptlrpc_request *req, void *data,
670 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
672 atomic_dec(&cli->cl_destroy_in_flight);
673 wake_up(&cli->cl_destroy_waitq);
677 static int osc_can_send_destroy(struct client_obd *cli)
679 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
680 cli->cl_max_rpcs_in_flight) {
681 /* The destroy request can be sent */
684 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
685 cli->cl_max_rpcs_in_flight) {
687 * The counter has been modified between the two atomic
690 wake_up(&cli->cl_destroy_waitq);
695 static int osc_create(const struct lu_env *env, struct obd_export *exp,
696 struct obdo *oa, struct lov_stripe_md **ea,
697 struct obd_trans_info *oti)
703 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
705 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
706 oa->o_flags == OBD_FL_RECREATE_OBJS) {
707 return osc_real_create(exp, oa, ea, oti);
710 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
711 return osc_real_create(exp, oa, ea, oti);
713 /* we should not get here anymore */
719 /* Destroy requests can be async always on the client, and we don't even really
720 * care about the return code since the client cannot do anything at all about
722 * When the MDS is unlinking a filename, it saves the file objects into a
723 * recovery llog, and these object records are cancelled when the OST reports
724 * they were destroyed and sync'd to disk (i.e. transaction committed).
725 * If the client dies, or the OST is down when the object should be destroyed,
726 * the records are not cancelled, and when the OST reconnects to the MDS next,
727 * it will retrieve the llog unlink logs and then sends the log cancellation
728 * cookies to the MDS after committing destroy transactions.
730 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
731 struct obdo *oa, struct lov_stripe_md *ea,
732 struct obd_trans_info *oti, struct obd_export *md_export)
734 struct client_obd *cli = &exp->exp_obd->u.cli;
735 struct ptlrpc_request *req;
736 struct ost_body *body;
741 CDEBUG(D_INFO, "oa NULL\n");
745 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
746 LDLM_FL_DISCARD_DATA);
748 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
750 ldlm_lock_list_put(&cancels, l_bl_ast, count);
754 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
757 ptlrpc_request_free(req);
761 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
762 ptlrpc_at_set_req_timeout(req);
764 if (oti && oa->o_valid & OBD_MD_FLCOOKIE)
765 oa->o_lcookie = *oti->oti_logcookies;
766 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
768 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
770 ptlrpc_request_set_replen(req);
772 /* If osc_destroy is for destroying the unlink orphan,
773 * sent from MDT to OST, which should not be blocked here,
774 * because the process might be triggered by ptlrpcd, and
775 * it is not good to block ptlrpcd thread (b=16006
777 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
778 req->rq_interpret_reply = osc_destroy_interpret;
779 if (!osc_can_send_destroy(cli)) {
780 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
784 * Wait until the number of on-going destroy RPCs drops
785 * under max_rpc_in_flight
787 l_wait_event_exclusive(cli->cl_destroy_waitq,
788 osc_can_send_destroy(cli), &lwi);
792 /* Do not wait for response */
793 ptlrpcd_add_req(req);
797 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
800 u32 bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
802 LASSERT(!(oa->o_valid & bits));
805 spin_lock(&cli->cl_loi_list_lock);
806 oa->o_dirty = cli->cl_dirty;
807 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
808 cli->cl_dirty_max)) {
809 CERROR("dirty %lu - %lu > dirty_max %lu\n",
810 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
812 } else if (unlikely(atomic_read(&obd_dirty_pages) -
813 atomic_read(&obd_dirty_transit_pages) >
814 (long)(obd_max_dirty_pages + 1))) {
815 /* The atomic_read() allowing the atomic_inc() are
816 * not covered by a lock thus they may safely race and trip
817 * this CERROR() unless we add in a small fudge factor (+1).
819 CERROR("dirty %d - %d > system dirty_max %d\n",
820 atomic_read(&obd_dirty_pages),
821 atomic_read(&obd_dirty_transit_pages),
822 obd_max_dirty_pages);
824 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
825 CERROR("dirty %lu - dirty_max %lu too big???\n",
826 cli->cl_dirty, cli->cl_dirty_max);
829 long max_in_flight = (cli->cl_max_pages_per_rpc <<
831 (cli->cl_max_rpcs_in_flight + 1);
832 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
834 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
835 oa->o_dropped = cli->cl_lost_grant;
836 cli->cl_lost_grant = 0;
837 spin_unlock(&cli->cl_loi_list_lock);
838 CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
839 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
842 void osc_update_next_shrink(struct client_obd *cli)
844 cli->cl_next_shrink_grant =
845 cfs_time_shift(cli->cl_grant_shrink_interval);
846 CDEBUG(D_CACHE, "next time %ld to shrink grant\n",
847 cli->cl_next_shrink_grant);
850 static void __osc_update_grant(struct client_obd *cli, u64 grant)
852 spin_lock(&cli->cl_loi_list_lock);
853 cli->cl_avail_grant += grant;
854 spin_unlock(&cli->cl_loi_list_lock);
857 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
859 if (body->oa.o_valid & OBD_MD_FLGRANT) {
860 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
861 __osc_update_grant(cli, body->oa.o_grant);
865 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
866 u32 keylen, void *key, u32 vallen,
867 void *val, struct ptlrpc_request_set *set);
869 static int osc_shrink_grant_interpret(const struct lu_env *env,
870 struct ptlrpc_request *req,
873 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
874 struct obdo *oa = ((struct osc_brw_async_args *)aa)->aa_oa;
875 struct ost_body *body;
878 __osc_update_grant(cli, oa->o_grant);
882 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
884 osc_update_grant(cli, body);
886 kmem_cache_free(obdo_cachep, oa);
890 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
892 spin_lock(&cli->cl_loi_list_lock);
893 oa->o_grant = cli->cl_avail_grant / 4;
894 cli->cl_avail_grant -= oa->o_grant;
895 spin_unlock(&cli->cl_loi_list_lock);
896 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
897 oa->o_valid |= OBD_MD_FLFLAGS;
900 oa->o_flags |= OBD_FL_SHRINK_GRANT;
901 osc_update_next_shrink(cli);
904 /* Shrink the current grant, either from some large amount to enough for a
905 * full set of in-flight RPCs, or if we have already shrunk to that limit
906 * then to enough for a single RPC. This avoids keeping more grant than
907 * needed, and avoids shrinking the grant piecemeal.
909 static int osc_shrink_grant(struct client_obd *cli)
911 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
912 (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
914 spin_lock(&cli->cl_loi_list_lock);
915 if (cli->cl_avail_grant <= target_bytes)
916 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
917 spin_unlock(&cli->cl_loi_list_lock);
919 return osc_shrink_grant_to_target(cli, target_bytes);
922 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
925 struct ost_body *body;
927 spin_lock(&cli->cl_loi_list_lock);
928 /* Don't shrink if we are already above or below the desired limit
929 * We don't want to shrink below a single RPC, as that will negatively
930 * impact block allocation and long-term performance.
932 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
933 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
935 if (target_bytes >= cli->cl_avail_grant) {
936 spin_unlock(&cli->cl_loi_list_lock);
939 spin_unlock(&cli->cl_loi_list_lock);
941 body = kzalloc(sizeof(*body), GFP_NOFS);
945 osc_announce_cached(cli, &body->oa, 0);
947 spin_lock(&cli->cl_loi_list_lock);
948 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
949 cli->cl_avail_grant = target_bytes;
950 spin_unlock(&cli->cl_loi_list_lock);
951 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
952 body->oa.o_valid |= OBD_MD_FLFLAGS;
953 body->oa.o_flags = 0;
955 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
956 osc_update_next_shrink(cli);
958 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
959 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
960 sizeof(*body), body, NULL);
962 __osc_update_grant(cli, body->oa.o_grant);
967 static int osc_should_shrink_grant(struct client_obd *client)
969 unsigned long time = cfs_time_current();
970 unsigned long next_shrink = client->cl_next_shrink_grant;
972 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
973 OBD_CONNECT_GRANT_SHRINK) == 0)
976 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
977 /* Get the current RPC size directly, instead of going via:
978 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
979 * Keep comment here so that it can be found by searching.
981 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
983 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
984 client->cl_avail_grant > brw_size)
987 osc_update_next_shrink(client);
992 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
994 struct client_obd *client;
996 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
997 if (osc_should_shrink_grant(client))
998 osc_shrink_grant(client);
1003 static int osc_add_shrink_grant(struct client_obd *client)
1007 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1009 osc_grant_shrink_grant_cb, NULL,
1010 &client->cl_grant_shrink_list);
1012 CERROR("add grant client %s error %d\n",
1013 client->cl_import->imp_obd->obd_name, rc);
1016 CDEBUG(D_CACHE, "add grant client %s\n",
1017 client->cl_import->imp_obd->obd_name);
1018 osc_update_next_shrink(client);
1022 static int osc_del_shrink_grant(struct client_obd *client)
1024 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1028 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1031 * ocd_grant is the total grant amount we're expect to hold: if we've
1032 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1033 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1035 * race is tolerable here: if we're evicted, but imp_state already
1036 * left EVICTED state, then cl_dirty must be 0 already.
1038 spin_lock(&cli->cl_loi_list_lock);
1039 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1040 cli->cl_avail_grant = ocd->ocd_grant;
1042 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1044 if (cli->cl_avail_grant < 0) {
1045 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1046 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1047 ocd->ocd_grant, cli->cl_dirty);
1048 /* workaround for servers which do not have the patch from
1051 cli->cl_avail_grant = ocd->ocd_grant;
1054 /* determine the appropriate chunk size used by osc_extent. */
1055 cli->cl_chunkbits = max_t(int, PAGE_SHIFT, ocd->ocd_blocksize);
1056 spin_unlock(&cli->cl_loi_list_lock);
1058 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld chunk bits: %d\n",
1059 cli->cl_import->imp_obd->obd_name,
1060 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1062 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1063 list_empty(&cli->cl_grant_shrink_list))
1064 osc_add_shrink_grant(cli);
1067 /* We assume that the reason this OSC got a short read is because it read
1068 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1069 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1070 * this stripe never got written at or beyond this stripe offset yet.
1072 static void handle_short_read(int nob_read, u32 page_count,
1073 struct brw_page **pga)
1078 /* skip bytes read OK */
1079 while (nob_read > 0) {
1080 LASSERT(page_count > 0);
1082 if (pga[i]->count > nob_read) {
1083 /* EOF inside this page */
1084 ptr = kmap(pga[i]->pg) +
1085 (pga[i]->off & ~PAGE_MASK);
1086 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1093 nob_read -= pga[i]->count;
1098 /* zero remaining pages */
1099 while (page_count-- > 0) {
1100 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1101 memset(ptr, 0, pga[i]->count);
1107 static int check_write_rcs(struct ptlrpc_request *req,
1108 int requested_nob, int niocount,
1109 u32 page_count, struct brw_page **pga)
1114 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1115 sizeof(*remote_rcs) *
1118 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1122 /* return error if any niobuf was in error */
1123 for (i = 0; i < niocount; i++) {
1124 if ((int)remote_rcs[i] < 0)
1125 return remote_rcs[i];
1127 if (remote_rcs[i] != 0) {
1128 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1129 i, remote_rcs[i], req);
1134 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1135 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1136 req->rq_bulk->bd_nob_transferred, requested_nob);
1143 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1145 if (p1->flag != p2->flag) {
1146 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1147 OBD_BRW_SYNC | OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1149 /* warn if we try to combine flags that we don't know to be
1152 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1153 CWARN("Saw flags 0x%x and 0x%x in the same brw, please report this at http://bugs.whamcloud.com/\n",
1154 p1->flag, p2->flag);
1159 return (p1->off + p1->count == p2->off);
1162 static u32 osc_checksum_bulk(int nob, u32 pg_count,
1163 struct brw_page **pga, int opc,
1164 enum cksum_type cksum_type)
1168 struct cfs_crypto_hash_desc *hdesc;
1169 unsigned int bufsize;
1171 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1173 LASSERT(pg_count > 0);
1175 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1176 if (IS_ERR(hdesc)) {
1177 CERROR("Unable to initialize checksum hash %s\n",
1178 cfs_crypto_hash_name(cfs_alg));
1179 return PTR_ERR(hdesc);
1182 while (nob > 0 && pg_count > 0) {
1183 int count = pga[i]->count > nob ? nob : pga[i]->count;
1185 /* corrupt the data before we compute the checksum, to
1186 * simulate an OST->client data error
1188 if (i == 0 && opc == OST_READ &&
1189 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1190 unsigned char *ptr = kmap(pga[i]->pg);
1191 int off = pga[i]->off & ~PAGE_MASK;
1193 memcpy(ptr + off, "bad1", min(4, nob));
1196 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1197 pga[i]->off & ~PAGE_MASK,
1200 "page %p map %p index %lu flags %lx count %u priv %0lx: off %d\n",
1201 pga[i]->pg, pga[i]->pg->mapping, pga[i]->pg->index,
1202 (long)pga[i]->pg->flags, page_count(pga[i]->pg),
1203 page_private(pga[i]->pg),
1204 (int)(pga[i]->off & ~PAGE_MASK));
1206 nob -= pga[i]->count;
1211 bufsize = sizeof(cksum);
1212 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1214 /* For sending we only compute the wrong checksum instead
1215 * of corrupting the data so it is still correct on a redo
1217 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1223 static int osc_brw_prep_request(int cmd, struct client_obd *cli,
1225 struct lov_stripe_md *lsm, u32 page_count,
1226 struct brw_page **pga,
1227 struct ptlrpc_request **reqp,
1231 struct ptlrpc_request *req;
1232 struct ptlrpc_bulk_desc *desc;
1233 struct ost_body *body;
1234 struct obd_ioobj *ioobj;
1235 struct niobuf_remote *niobuf;
1236 int niocount, i, requested_nob, opc, rc;
1237 struct osc_brw_async_args *aa;
1238 struct req_capsule *pill;
1239 struct brw_page *pg_prev;
1241 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1242 return -ENOMEM; /* Recoverable */
1243 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1244 return -EINVAL; /* Fatal */
1246 if ((cmd & OBD_BRW_WRITE) != 0) {
1248 req = ptlrpc_request_alloc_pool(cli->cl_import,
1250 &RQF_OST_BRW_WRITE);
1253 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1258 for (niocount = i = 1; i < page_count; i++) {
1259 if (!can_merge_pages(pga[i - 1], pga[i]))
1263 pill = &req->rq_pill;
1264 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1266 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1267 niocount * sizeof(*niobuf));
1269 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1271 ptlrpc_request_free(req);
1274 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1275 ptlrpc_at_set_req_timeout(req);
1276 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1279 req->rq_no_retry_einprogress = 1;
1281 desc = ptlrpc_prep_bulk_imp(req, page_count,
1282 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1283 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1290 /* NB request now owns desc and will free it when it gets freed */
1292 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1293 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1294 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1295 LASSERT(body && ioobj && niobuf);
1297 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1299 obdo_to_ioobj(oa, ioobj);
1300 ioobj->ioo_bufcnt = niocount;
1301 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1302 * that might be send for this request. The actual number is decided
1303 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1304 * "max - 1" for old client compatibility sending "0", and also so the
1305 * the actual maximum is a power-of-two number, not one less. LU-1431
1307 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1308 LASSERT(page_count > 0);
1310 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1311 struct brw_page *pg = pga[i];
1312 int poff = pg->off & ~PAGE_MASK;
1314 LASSERT(pg->count > 0);
1315 /* make sure there is no gap in the middle of page array */
1316 LASSERTF(page_count == 1 ||
1317 (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1318 ergo(i > 0 && i < page_count - 1,
1319 poff == 0 && pg->count == PAGE_SIZE) &&
1320 ergo(i == page_count - 1, poff == 0)),
1321 "i: %d/%d pg: %p off: %llu, count: %u\n",
1322 i, page_count, pg, pg->off, pg->count);
1323 LASSERTF(i == 0 || pg->off > pg_prev->off,
1324 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu prev_pg %p [pri %lu ind %lu] off %llu\n",
1326 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1327 pg_prev->pg, page_private(pg_prev->pg),
1328 pg_prev->pg->index, pg_prev->off);
1329 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1330 (pg->flag & OBD_BRW_SRVLOCK));
1332 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1333 requested_nob += pg->count;
1335 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1337 niobuf->len += pg->count;
1339 niobuf->offset = pg->off;
1340 niobuf->len = pg->count;
1341 niobuf->flags = pg->flag;
1346 LASSERTF((void *)(niobuf - niocount) ==
1347 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1348 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1349 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1351 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1353 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1354 body->oa.o_valid |= OBD_MD_FLFLAGS;
1355 body->oa.o_flags = 0;
1357 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1360 if (osc_should_shrink_grant(cli))
1361 osc_shrink_grant_local(cli, &body->oa);
1363 /* size[REQ_REC_OFF] still sizeof (*body) */
1364 if (opc == OST_WRITE) {
1365 if (cli->cl_checksum &&
1366 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1367 /* store cl_cksum_type in a local variable since
1368 * it can be changed via lprocfs
1370 enum cksum_type cksum_type = cli->cl_cksum_type;
1372 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1373 oa->o_flags &= OBD_FL_LOCAL_MASK;
1374 body->oa.o_flags = 0;
1376 body->oa.o_flags |= cksum_type_pack(cksum_type);
1377 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1378 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1382 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1384 /* save this in 'oa', too, for later checking */
1385 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1386 oa->o_flags |= cksum_type_pack(cksum_type);
1388 /* clear out the checksum flag, in case this is a
1389 * resend but cl_checksum is no longer set. b=11238
1391 oa->o_valid &= ~OBD_MD_FLCKSUM;
1393 oa->o_cksum = body->oa.o_cksum;
1394 /* 1 RC per niobuf */
1395 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1396 sizeof(__u32) * niocount);
1398 if (cli->cl_checksum &&
1399 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1400 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1401 body->oa.o_flags = 0;
1402 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1403 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1406 ptlrpc_request_set_replen(req);
1408 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1409 aa = ptlrpc_req_async_args(req);
1411 aa->aa_requested_nob = requested_nob;
1412 aa->aa_nio_count = niocount;
1413 aa->aa_page_count = page_count;
1417 INIT_LIST_HEAD(&aa->aa_oaps);
1423 ptlrpc_req_finished(req);
1427 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1428 __u32 client_cksum, __u32 server_cksum, int nob,
1429 u32 page_count, struct brw_page **pga,
1430 enum cksum_type client_cksum_type)
1434 enum cksum_type cksum_type;
1436 if (server_cksum == client_cksum) {
1437 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1441 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1443 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1446 if (cksum_type != client_cksum_type)
1447 msg = "the server did not use the checksum type specified in the original request - likely a protocol problem"
1449 else if (new_cksum == server_cksum)
1450 msg = "changed on the client after we checksummed it - likely false positive due to mmap IO (bug 11742)"
1452 else if (new_cksum == client_cksum)
1453 msg = "changed in transit before arrival at OST";
1455 msg = "changed in transit AND doesn't match the original - likely false positive due to mmap IO (bug 11742)"
1458 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1459 " object "DOSTID" extent [%llu-%llu]\n",
1460 msg, libcfs_nid2str(peer->nid),
1461 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1462 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1463 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1464 POSTID(&oa->o_oi), pga[0]->off,
1465 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1466 CERROR("original client csum %x (type %x), server csum %x (type %x), client csum now %x\n",
1467 client_cksum, client_cksum_type,
1468 server_cksum, cksum_type, new_cksum);
1472 /* Note rc enters this function as number of bytes transferred */
1473 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1475 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1476 const lnet_process_id_t *peer =
1477 &req->rq_import->imp_connection->c_peer;
1478 struct client_obd *cli = aa->aa_cli;
1479 struct ost_body *body;
1480 __u32 client_cksum = 0;
1482 if (rc < 0 && rc != -EDQUOT) {
1483 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1487 LASSERTF(req->rq_repmsg, "rc = %d\n", rc);
1488 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1490 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1494 /* set/clear over quota flag for a uid/gid */
1495 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1496 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1497 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1499 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1500 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1502 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1505 osc_update_grant(cli, body);
1510 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1511 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1513 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1515 CERROR("Unexpected +ve rc %d\n", rc);
1518 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1520 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1523 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1524 check_write_checksum(&body->oa, peer, client_cksum,
1525 body->oa.o_cksum, aa->aa_requested_nob,
1526 aa->aa_page_count, aa->aa_ppga,
1527 cksum_type_unpack(aa->aa_oa->o_flags)))
1530 rc = check_write_rcs(req, aa->aa_requested_nob,
1532 aa->aa_page_count, aa->aa_ppga);
1536 /* The rest of this function executes only for OST_READs */
1538 /* if unwrap_bulk failed, return -EAGAIN to retry */
1539 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1545 if (rc > aa->aa_requested_nob) {
1546 CERROR("Unexpected rc %d (%d requested)\n", rc,
1547 aa->aa_requested_nob);
1551 if (rc != req->rq_bulk->bd_nob_transferred) {
1552 CERROR("Unexpected rc %d (%d transferred)\n",
1553 rc, req->rq_bulk->bd_nob_transferred);
1557 if (rc < aa->aa_requested_nob)
1558 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1560 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1561 static int cksum_counter;
1562 __u32 server_cksum = body->oa.o_cksum;
1565 enum cksum_type cksum_type;
1567 cksum_type = cksum_type_unpack(body->oa.o_valid&OBD_MD_FLFLAGS ?
1568 body->oa.o_flags : 0);
1569 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1570 aa->aa_ppga, OST_READ,
1573 if (peer->nid != req->rq_bulk->bd_sender) {
1575 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1578 if (server_cksum != client_cksum) {
1579 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from %s%s%s inode " DFID " object " DOSTID " extent [%llu-%llu]\n",
1580 req->rq_import->imp_obd->obd_name,
1581 libcfs_nid2str(peer->nid),
1583 body->oa.o_valid & OBD_MD_FLFID ?
1584 body->oa.o_parent_seq : (__u64)0,
1585 body->oa.o_valid & OBD_MD_FLFID ?
1586 body->oa.o_parent_oid : 0,
1587 body->oa.o_valid & OBD_MD_FLFID ?
1588 body->oa.o_parent_ver : 0,
1589 POSTID(&body->oa.o_oi),
1590 aa->aa_ppga[0]->off,
1591 aa->aa_ppga[aa->aa_page_count-1]->off +
1592 aa->aa_ppga[aa->aa_page_count-1]->count -
1594 CERROR("client %x, server %x, cksum_type %x\n",
1595 client_cksum, server_cksum, cksum_type);
1597 aa->aa_oa->o_cksum = client_cksum;
1601 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1604 } else if (unlikely(client_cksum)) {
1605 static int cksum_missed;
1608 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1609 CERROR("Checksum %u requested from %s but not sent\n",
1610 cksum_missed, libcfs_nid2str(peer->nid));
1616 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1617 aa->aa_oa, &body->oa);
1622 static int osc_brw_redo_request(struct ptlrpc_request *request,
1623 struct osc_brw_async_args *aa, int rc)
1625 struct ptlrpc_request *new_req;
1626 struct osc_brw_async_args *new_aa;
1627 struct osc_async_page *oap;
1629 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1630 "redo for recoverable error %d", rc);
1632 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1633 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1634 aa->aa_cli, aa->aa_oa,
1635 NULL /* lsm unused by osc currently */,
1636 aa->aa_page_count, aa->aa_ppga,
1641 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1642 if (oap->oap_request) {
1643 LASSERTF(request == oap->oap_request,
1644 "request %p != oap_request %p\n",
1645 request, oap->oap_request);
1646 if (oap->oap_interrupted) {
1647 ptlrpc_req_finished(new_req);
1652 /* New request takes over pga and oaps from old request.
1653 * Note that copying a list_head doesn't work, need to move it...
1656 new_req->rq_interpret_reply = request->rq_interpret_reply;
1657 new_req->rq_async_args = request->rq_async_args;
1658 /* cap resend delay to the current request timeout, this is similar to
1659 * what ptlrpc does (see after_reply())
1661 if (aa->aa_resends > new_req->rq_timeout)
1662 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1664 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1665 new_req->rq_generation_set = 1;
1666 new_req->rq_import_generation = request->rq_import_generation;
1668 new_aa = ptlrpc_req_async_args(new_req);
1670 INIT_LIST_HEAD(&new_aa->aa_oaps);
1671 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1672 INIT_LIST_HEAD(&new_aa->aa_exts);
1673 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1674 new_aa->aa_resends = aa->aa_resends;
1676 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1677 if (oap->oap_request) {
1678 ptlrpc_req_finished(oap->oap_request);
1679 oap->oap_request = ptlrpc_request_addref(new_req);
1683 /* XXX: This code will run into problem if we're going to support
1684 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1685 * and wait for all of them to be finished. We should inherit request
1686 * set from old request.
1688 ptlrpcd_add_req(new_req);
1690 DEBUG_REQ(D_INFO, new_req, "new request");
1695 * ugh, we want disk allocation on the target to happen in offset order. we'll
1696 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1697 * fine for our small page arrays and doesn't require allocation. its an
1698 * insertion sort that swaps elements that are strides apart, shrinking the
1699 * stride down until its '1' and the array is sorted.
1701 static void sort_brw_pages(struct brw_page **array, int num)
1704 struct brw_page *tmp;
1708 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1713 for (i = stride ; i < num ; i++) {
1716 while (j >= stride && array[j - stride]->off > tmp->off) {
1717 array[j] = array[j - stride];
1722 } while (stride > 1);
1725 static void osc_release_ppga(struct brw_page **ppga, u32 count)
1731 static int brw_interpret(const struct lu_env *env,
1732 struct ptlrpc_request *req, void *data, int rc)
1734 struct osc_brw_async_args *aa = data;
1735 struct osc_extent *ext;
1736 struct osc_extent *tmp;
1737 struct client_obd *cli = aa->aa_cli;
1739 rc = osc_brw_fini_request(req, rc);
1740 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1741 /* When server return -EINPROGRESS, client should always retry
1742 * regardless of the number of times the bulk was resent already.
1744 if (osc_recoverable_error(rc)) {
1745 if (req->rq_import_generation !=
1746 req->rq_import->imp_generation) {
1747 CDEBUG(D_HA, "%s: resend cross eviction for object: " DOSTID ", rc = %d.\n",
1748 req->rq_import->imp_obd->obd_name,
1749 POSTID(&aa->aa_oa->o_oi), rc);
1750 } else if (rc == -EINPROGRESS ||
1751 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1752 rc = osc_brw_redo_request(req, aa, rc);
1754 CERROR("%s: too many resent retries for object: %llu:%llu, rc = %d.\n",
1755 req->rq_import->imp_obd->obd_name,
1756 POSTID(&aa->aa_oa->o_oi), rc);
1761 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1766 struct obdo *oa = aa->aa_oa;
1767 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1768 unsigned long valid = 0;
1769 struct cl_object *obj;
1770 struct osc_async_page *last;
1772 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1773 obj = osc2cl(last->oap_obj);
1775 cl_object_attr_lock(obj);
1776 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1777 attr->cat_blocks = oa->o_blocks;
1778 valid |= CAT_BLOCKS;
1780 if (oa->o_valid & OBD_MD_FLMTIME) {
1781 attr->cat_mtime = oa->o_mtime;
1784 if (oa->o_valid & OBD_MD_FLATIME) {
1785 attr->cat_atime = oa->o_atime;
1788 if (oa->o_valid & OBD_MD_FLCTIME) {
1789 attr->cat_ctime = oa->o_ctime;
1793 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1794 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1795 loff_t last_off = last->oap_count + last->oap_obj_off;
1797 /* Change file size if this is an out of quota or
1798 * direct IO write and it extends the file size
1800 if (loi->loi_lvb.lvb_size < last_off) {
1801 attr->cat_size = last_off;
1804 /* Extend KMS if it's not a lockless write */
1805 if (loi->loi_kms < last_off &&
1806 oap2osc_page(last)->ops_srvlock == 0) {
1807 attr->cat_kms = last_off;
1813 cl_object_attr_set(env, obj, attr, valid);
1814 cl_object_attr_unlock(obj);
1816 kmem_cache_free(obdo_cachep, aa->aa_oa);
1818 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1819 list_del_init(&ext->oe_link);
1820 osc_extent_finish(env, ext, 1, rc);
1822 LASSERT(list_empty(&aa->aa_exts));
1823 LASSERT(list_empty(&aa->aa_oaps));
1825 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1826 req->rq_bulk->bd_nob_transferred);
1827 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1828 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1830 spin_lock(&cli->cl_loi_list_lock);
1831 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1832 * is called so we know whether to go to sync BRWs or wait for more
1835 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1836 cli->cl_w_in_flight--;
1838 cli->cl_r_in_flight--;
1839 osc_wake_cache_waiters(cli);
1840 spin_unlock(&cli->cl_loi_list_lock);
1842 osc_io_unplug(env, cli, NULL);
1847 * Build an RPC by the list of extent @ext_list. The caller must ensure
1848 * that the total pages in this list are NOT over max pages per RPC.
1849 * Extents in the list must be in OES_RPC state.
1851 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1852 struct list_head *ext_list, int cmd)
1854 struct ptlrpc_request *req = NULL;
1855 struct osc_extent *ext;
1856 struct brw_page **pga = NULL;
1857 struct osc_brw_async_args *aa = NULL;
1858 struct obdo *oa = NULL;
1859 struct osc_async_page *oap;
1860 struct osc_async_page *tmp;
1861 struct cl_req *clerq = NULL;
1862 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1863 struct ldlm_lock *lock = NULL;
1864 struct cl_req_attr *crattr = NULL;
1865 u64 starting_offset = OBD_OBJECT_EOF;
1866 u64 ending_offset = 0;
1872 struct ost_body *body;
1873 LIST_HEAD(rpc_list);
1875 LASSERT(!list_empty(ext_list));
1877 /* add pages into rpc_list to build BRW rpc */
1878 list_for_each_entry(ext, ext_list, oe_link) {
1879 LASSERT(ext->oe_state == OES_RPC);
1880 mem_tight |= ext->oe_memalloc;
1881 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1883 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1884 if (starting_offset > oap->oap_obj_off)
1885 starting_offset = oap->oap_obj_off;
1887 LASSERT(oap->oap_page_off == 0);
1888 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1889 ending_offset = oap->oap_obj_off +
1892 LASSERT(oap->oap_page_off + oap->oap_count ==
1898 mpflag = cfs_memory_pressure_get_and_set();
1900 crattr = kzalloc(sizeof(*crattr), GFP_NOFS);
1906 pga = kcalloc(page_count, sizeof(*pga), GFP_NOFS);
1912 oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
1919 list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1920 struct cl_page *page = oap2cl_page(oap);
1923 clerq = cl_req_alloc(env, page, crt,
1924 1 /* only 1-object rpcs for now */);
1925 if (IS_ERR(clerq)) {
1926 rc = PTR_ERR(clerq);
1929 lock = oap->oap_ldlm_lock;
1932 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1933 pga[i] = &oap->oap_brw_page;
1934 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1935 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1936 pga[i]->pg, oap->oap_page->index, oap,
1939 cl_req_page_add(env, clerq, page);
1942 /* always get the data for the obdo for the rpc */
1944 crattr->cra_oa = oa;
1945 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1947 oa->o_handle = lock->l_remote_handle;
1948 oa->o_valid |= OBD_MD_FLHANDLE;
1951 rc = cl_req_prep(env, clerq);
1953 CERROR("cl_req_prep failed: %d\n", rc);
1957 sort_brw_pages(pga, page_count);
1958 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1961 CERROR("prep_req failed: %d\n", rc);
1965 req->rq_interpret_reply = brw_interpret;
1968 req->rq_memalloc = 1;
1970 /* Need to update the timestamps after the request is built in case
1971 * we race with setattr (locally or in queue at OST). If OST gets
1972 * later setattr before earlier BRW (as determined by the request xid),
1973 * the OST will not use BRW timestamps. Sadly, there is no obvious
1974 * way to do this in a single call. bug 10150
1976 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1977 crattr->cra_oa = &body->oa;
1978 cl_req_attr_set(env, clerq, crattr,
1979 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
1981 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1983 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1984 aa = ptlrpc_req_async_args(req);
1985 INIT_LIST_HEAD(&aa->aa_oaps);
1986 list_splice_init(&rpc_list, &aa->aa_oaps);
1987 INIT_LIST_HEAD(&aa->aa_exts);
1988 list_splice_init(ext_list, &aa->aa_exts);
1989 aa->aa_clerq = clerq;
1991 /* queued sync pages can be torn down while the pages
1992 * were between the pending list and the rpc
1995 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1996 /* only one oap gets a request reference */
1999 if (oap->oap_interrupted && !req->rq_intr) {
2000 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2002 ptlrpc_mark_interrupted(req);
2006 tmp->oap_request = ptlrpc_request_addref(req);
2008 spin_lock(&cli->cl_loi_list_lock);
2009 starting_offset >>= PAGE_SHIFT;
2010 if (cmd == OBD_BRW_READ) {
2011 cli->cl_r_in_flight++;
2012 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2013 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2014 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2015 starting_offset + 1);
2017 cli->cl_w_in_flight++;
2018 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2019 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2020 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2021 starting_offset + 1);
2023 spin_unlock(&cli->cl_loi_list_lock);
2025 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2026 page_count, aa, cli->cl_r_in_flight,
2027 cli->cl_w_in_flight);
2029 ptlrpcd_add_req(req);
2034 cfs_memory_pressure_restore(mpflag);
2042 kmem_cache_free(obdo_cachep, oa);
2044 /* this should happen rarely and is pretty bad, it makes the
2045 * pending list not follow the dirty order
2047 while (!list_empty(ext_list)) {
2048 ext = list_entry(ext_list->next, struct osc_extent,
2050 list_del_init(&ext->oe_link);
2051 osc_extent_finish(env, ext, 0, rc);
2053 if (clerq && !IS_ERR(clerq))
2054 cl_req_completion(env, clerq, rc);
2059 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2060 struct ldlm_enqueue_info *einfo)
2062 void *data = einfo->ei_cbdata;
2065 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2066 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2067 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2068 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2070 lock_res_and_lock(lock);
2072 if (!lock->l_ast_data)
2073 lock->l_ast_data = data;
2074 if (lock->l_ast_data == data)
2077 unlock_res_and_lock(lock);
2082 static int osc_set_data_with_check(struct lustre_handle *lockh,
2083 struct ldlm_enqueue_info *einfo)
2085 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2089 set = osc_set_lock_data_with_check(lock, einfo);
2090 LDLM_LOCK_PUT(lock);
2092 CERROR("lockh %p, data %p - client evicted?\n",
2093 lockh, einfo->ei_cbdata);
2097 /* find any ldlm lock of the inode in osc
2102 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2103 ldlm_iterator_t replace, void *data)
2105 struct ldlm_res_id res_id;
2106 struct obd_device *obd = class_exp2obd(exp);
2109 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2110 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2111 if (rc == LDLM_ITER_STOP)
2113 if (rc == LDLM_ITER_CONTINUE)
2118 static int osc_enqueue_fini(struct ptlrpc_request *req,
2119 osc_enqueue_upcall_f upcall, void *cookie,
2120 struct lustre_handle *lockh, enum ldlm_mode mode,
2121 __u64 *flags, int agl, int errcode)
2123 bool intent = *flags & LDLM_FL_HAS_INTENT;
2126 /* The request was created before ldlm_cli_enqueue call. */
2127 if (intent && errcode == ELDLM_LOCK_ABORTED) {
2128 struct ldlm_reply *rep;
2130 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2132 rep->lock_policy_res1 =
2133 ptlrpc_status_ntoh(rep->lock_policy_res1);
2134 if (rep->lock_policy_res1)
2135 errcode = rep->lock_policy_res1;
2137 *flags |= LDLM_FL_LVB_READY;
2138 } else if (errcode == ELDLM_OK) {
2139 *flags |= LDLM_FL_LVB_READY;
2142 /* Call the update callback. */
2143 rc = (*upcall)(cookie, lockh, errcode);
2144 /* release the reference taken in ldlm_cli_enqueue() */
2145 if (errcode == ELDLM_LOCK_MATCHED)
2147 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2148 ldlm_lock_decref(lockh, mode);
2153 static int osc_enqueue_interpret(const struct lu_env *env,
2154 struct ptlrpc_request *req,
2155 struct osc_enqueue_args *aa, int rc)
2157 struct ldlm_lock *lock;
2158 struct lustre_handle *lockh = &aa->oa_lockh;
2159 enum ldlm_mode mode = aa->oa_mode;
2160 struct ost_lvb *lvb = aa->oa_lvb;
2161 __u32 lvb_len = sizeof(*lvb);
2165 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2168 lock = ldlm_handle2lock(lockh);
2169 LASSERTF(lock, "lockh %llx, req %p, aa %p - client evicted?\n",
2170 lockh->cookie, req, aa);
2172 /* Take an additional reference so that a blocking AST that
2173 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2174 * to arrive after an upcall has been executed by
2175 * osc_enqueue_fini().
2177 ldlm_lock_addref(lockh, mode);
2179 /* Let CP AST to grant the lock first. */
2180 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2183 LASSERT(!aa->oa_lvb);
2184 LASSERT(!aa->oa_flags);
2185 aa->oa_flags = &flags;
2188 /* Complete obtaining the lock procedure. */
2189 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2190 aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2192 /* Complete osc stuff. */
2193 rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2194 aa->oa_flags, aa->oa_agl, rc);
2196 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2198 ldlm_lock_decref(lockh, mode);
2199 LDLM_LOCK_PUT(lock);
2203 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2205 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2206 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2207 * other synchronous requests, however keeping some locks and trying to obtain
2208 * others may take a considerable amount of time in a case of ost failure; and
2209 * when other sync requests do not get released lock from a client, the client
2210 * is evicted from the cluster -- such scenaries make the life difficult, so
2211 * release locks just after they are obtained.
2213 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2214 __u64 *flags, ldlm_policy_data_t *policy,
2215 struct ost_lvb *lvb, int kms_valid,
2216 osc_enqueue_upcall_f upcall, void *cookie,
2217 struct ldlm_enqueue_info *einfo,
2218 struct ptlrpc_request_set *rqset, int async, int agl)
2220 struct obd_device *obd = exp->exp_obd;
2221 struct lustre_handle lockh = { 0 };
2222 struct ptlrpc_request *req = NULL;
2223 int intent = *flags & LDLM_FL_HAS_INTENT;
2224 __u64 match_lvb = agl ? 0 : LDLM_FL_LVB_READY;
2225 enum ldlm_mode mode;
2228 /* Filesystem lock extents are extended to page boundaries so that
2229 * dealing with the page cache is a little smoother.
2231 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2232 policy->l_extent.end |= ~PAGE_MASK;
2235 * kms is not valid when either object is completely fresh (so that no
2236 * locks are cached), or object was evicted. In the latter case cached
2237 * lock cannot be used, because it would prime inode state with
2238 * potentially stale LVB.
2243 /* Next, search for already existing extent locks that will cover us */
2244 /* If we're trying to read, we also search for an existing PW lock. The
2245 * VFS and page cache already protect us locally, so lots of readers/
2246 * writers can share a single PW lock.
2248 * There are problems with conversion deadlocks, so instead of
2249 * converting a read lock to a write lock, we'll just enqueue a new
2252 * At some point we should cancel the read lock instead of making them
2253 * send us a blocking callback, but there are problems with canceling
2254 * locks out from other users right now, too.
2256 mode = einfo->ei_mode;
2257 if (einfo->ei_mode == LCK_PR)
2259 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2260 einfo->ei_type, policy, mode, &lockh, 0);
2262 struct ldlm_lock *matched;
2264 if (*flags & LDLM_FL_TEST_LOCK)
2267 matched = ldlm_handle2lock(&lockh);
2269 /* AGL enqueues DLM locks speculatively. Therefore if
2270 * it already exists a DLM lock, it wll just inform the
2271 * caller to cancel the AGL process for this stripe.
2273 ldlm_lock_decref(&lockh, mode);
2274 LDLM_LOCK_PUT(matched);
2276 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2277 *flags |= LDLM_FL_LVB_READY;
2278 /* We already have a lock, and it's referenced. */
2279 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2281 ldlm_lock_decref(&lockh, mode);
2282 LDLM_LOCK_PUT(matched);
2285 ldlm_lock_decref(&lockh, mode);
2286 LDLM_LOCK_PUT(matched);
2291 if (*flags & LDLM_FL_TEST_LOCK)
2294 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2295 &RQF_LDLM_ENQUEUE_LVB);
2299 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2301 ptlrpc_request_free(req);
2305 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2307 ptlrpc_request_set_replen(req);
2310 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2311 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2313 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2314 sizeof(*lvb), LVB_T_OST, &lockh, async);
2317 struct osc_enqueue_args *aa;
2319 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2320 aa = ptlrpc_req_async_args(req);
2322 aa->oa_mode = einfo->ei_mode;
2323 aa->oa_type = einfo->ei_type;
2324 lustre_handle_copy(&aa->oa_lockh, &lockh);
2325 aa->oa_upcall = upcall;
2326 aa->oa_cookie = cookie;
2329 aa->oa_flags = flags;
2332 /* AGL is essentially to enqueue an DLM lock
2333 * in advance, so we don't care about the
2334 * result of AGL enqueue.
2337 aa->oa_flags = NULL;
2340 req->rq_interpret_reply =
2341 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2342 if (rqset == PTLRPCD_SET)
2343 ptlrpcd_add_req(req);
2345 ptlrpc_set_add_req(rqset, req);
2346 } else if (intent) {
2347 ptlrpc_req_finished(req);
2352 rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2355 ptlrpc_req_finished(req);
2360 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2361 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2362 __u64 *flags, void *data, struct lustre_handle *lockh,
2365 struct obd_device *obd = exp->exp_obd;
2366 __u64 lflags = *flags;
2369 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2372 /* Filesystem lock extents are extended to page boundaries so that
2373 * dealing with the page cache is a little smoother
2375 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2376 policy->l_extent.end |= ~PAGE_MASK;
2378 /* Next, search for already existing extent locks that will cover us */
2379 /* If we're trying to read, we also search for an existing PW lock. The
2380 * VFS and page cache already protect us locally, so lots of readers/
2381 * writers can share a single PW lock.
2386 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2387 res_id, type, policy, rc, lockh, unref);
2390 if (!osc_set_data_with_check(lockh, data)) {
2391 if (!(lflags & LDLM_FL_TEST_LOCK))
2392 ldlm_lock_decref(lockh, rc);
2396 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2397 ldlm_lock_addref(lockh, LCK_PR);
2398 ldlm_lock_decref(lockh, LCK_PW);
2405 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2407 if (unlikely(mode == LCK_GROUP))
2408 ldlm_lock_decref_and_cancel(lockh, mode);
2410 ldlm_lock_decref(lockh, mode);
2415 static int osc_statfs_interpret(const struct lu_env *env,
2416 struct ptlrpc_request *req,
2417 struct osc_async_args *aa, int rc)
2419 struct obd_statfs *msfs;
2422 /* The request has in fact never been sent
2423 * due to issues at a higher level (LOV).
2424 * Exit immediately since the caller is
2425 * aware of the problem and takes care
2430 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2431 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY)) {
2439 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2445 *aa->aa_oi->oi_osfs = *msfs;
2447 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2451 static int osc_statfs_async(struct obd_export *exp,
2452 struct obd_info *oinfo, __u64 max_age,
2453 struct ptlrpc_request_set *rqset)
2455 struct obd_device *obd = class_exp2obd(exp);
2456 struct ptlrpc_request *req;
2457 struct osc_async_args *aa;
2460 /* We could possibly pass max_age in the request (as an absolute
2461 * timestamp or a "seconds.usec ago") so the target can avoid doing
2462 * extra calls into the filesystem if that isn't necessary (e.g.
2463 * during mount that would help a bit). Having relative timestamps
2464 * is not so great if request processing is slow, while absolute
2465 * timestamps are not ideal because they need time synchronization.
2467 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2471 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2473 ptlrpc_request_free(req);
2476 ptlrpc_request_set_replen(req);
2477 req->rq_request_portal = OST_CREATE_PORTAL;
2478 ptlrpc_at_set_req_timeout(req);
2480 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2481 /* procfs requests not want stat in wait for avoid deadlock */
2482 req->rq_no_resend = 1;
2483 req->rq_no_delay = 1;
2486 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2487 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2488 aa = ptlrpc_req_async_args(req);
2491 ptlrpc_set_add_req(rqset, req);
2495 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2496 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2498 struct obd_device *obd = class_exp2obd(exp);
2499 struct obd_statfs *msfs;
2500 struct ptlrpc_request *req;
2501 struct obd_import *imp = NULL;
2504 /* Since the request might also come from lprocfs, so we need
2505 * sync this with client_disconnect_export Bug15684
2507 down_read(&obd->u.cli.cl_sem);
2508 if (obd->u.cli.cl_import)
2509 imp = class_import_get(obd->u.cli.cl_import);
2510 up_read(&obd->u.cli.cl_sem);
2514 /* We could possibly pass max_age in the request (as an absolute
2515 * timestamp or a "seconds.usec ago") so the target can avoid doing
2516 * extra calls into the filesystem if that isn't necessary (e.g.
2517 * during mount that would help a bit). Having relative timestamps
2518 * is not so great if request processing is slow, while absolute
2519 * timestamps are not ideal because they need time synchronization.
2521 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2523 class_import_put(imp);
2528 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2530 ptlrpc_request_free(req);
2533 ptlrpc_request_set_replen(req);
2534 req->rq_request_portal = OST_CREATE_PORTAL;
2535 ptlrpc_at_set_req_timeout(req);
2537 if (flags & OBD_STATFS_NODELAY) {
2538 /* procfs requests not want stat in wait for avoid deadlock */
2539 req->rq_no_resend = 1;
2540 req->rq_no_delay = 1;
2543 rc = ptlrpc_queue_wait(req);
2547 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2556 ptlrpc_req_finished(req);
2560 /* Retrieve object striping information.
2562 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2563 * the maximum number of OST indices which will fit in the user buffer.
2564 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2566 static int osc_getstripe(struct lov_stripe_md *lsm,
2567 struct lov_user_md __user *lump)
2569 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2570 struct lov_user_md_v3 lum, *lumk;
2571 struct lov_user_ost_data_v1 *lmm_objects;
2572 int rc = 0, lum_size;
2577 /* we only need the header part from user space to get lmm_magic and
2578 * lmm_stripe_count, (the header part is common to v1 and v3)
2580 lum_size = sizeof(struct lov_user_md_v1);
2581 if (copy_from_user(&lum, lump, lum_size))
2584 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2585 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2588 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2589 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2590 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2591 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2593 /* we can use lov_mds_md_size() to compute lum_size
2594 * because lov_user_md_vX and lov_mds_md_vX have the same size
2596 if (lum.lmm_stripe_count > 0) {
2597 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2598 lumk = kzalloc(lum_size, GFP_NOFS);
2602 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2604 &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2606 lmm_objects = &(lumk->lmm_objects[0]);
2607 lmm_objects->l_ost_oi = lsm->lsm_oi;
2609 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2613 lumk->lmm_oi = lsm->lsm_oi;
2614 lumk->lmm_stripe_count = 1;
2616 if (copy_to_user(lump, lumk, lum_size))
2625 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2626 void *karg, void __user *uarg)
2628 struct obd_device *obd = exp->exp_obd;
2629 struct obd_ioctl_data *data = karg;
2632 if (!try_module_get(THIS_MODULE)) {
2633 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2634 module_name(THIS_MODULE));
2638 case OBD_IOC_LOV_GET_CONFIG: {
2640 struct lov_desc *desc;
2641 struct obd_uuid uuid;
2645 if (obd_ioctl_getdata(&buf, &len, uarg)) {
2650 data = (struct obd_ioctl_data *)buf;
2652 if (sizeof(*desc) > data->ioc_inllen1) {
2653 obd_ioctl_freedata(buf, len);
2658 if (data->ioc_inllen2 < sizeof(uuid)) {
2659 obd_ioctl_freedata(buf, len);
2664 desc = (struct lov_desc *)data->ioc_inlbuf1;
2665 desc->ld_tgt_count = 1;
2666 desc->ld_active_tgt_count = 1;
2667 desc->ld_default_stripe_count = 1;
2668 desc->ld_default_stripe_size = 0;
2669 desc->ld_default_stripe_offset = 0;
2670 desc->ld_pattern = 0;
2671 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2673 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2675 err = copy_to_user(uarg, buf, len);
2678 obd_ioctl_freedata(buf, len);
2681 case LL_IOC_LOV_SETSTRIPE:
2682 err = obd_alloc_memmd(exp, karg);
2686 case LL_IOC_LOV_GETSTRIPE:
2687 err = osc_getstripe(karg, uarg);
2689 case OBD_IOC_CLIENT_RECOVER:
2690 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2691 data->ioc_inlbuf1, 0);
2695 case IOC_OSC_SET_ACTIVE:
2696 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2699 case OBD_IOC_POLL_QUOTACHECK:
2700 err = osc_quota_poll_check(exp, karg);
2702 case OBD_IOC_PING_TARGET:
2703 err = ptlrpc_obd_ping(obd);
2706 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2707 cmd, current_comm());
2712 module_put(THIS_MODULE);
2716 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2717 u32 keylen, void *key, __u32 *vallen, void *val,
2718 struct lov_stripe_md *lsm)
2720 if (!vallen || !val)
2723 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2724 __u32 *stripe = val;
2725 *vallen = sizeof(*stripe);
2728 } else if (KEY_IS(KEY_LAST_ID)) {
2729 struct ptlrpc_request *req;
2734 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2735 &RQF_OST_GET_INFO_LAST_ID);
2739 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2740 RCL_CLIENT, keylen);
2741 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2743 ptlrpc_request_free(req);
2747 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2748 memcpy(tmp, key, keylen);
2750 req->rq_no_delay = req->rq_no_resend = 1;
2751 ptlrpc_request_set_replen(req);
2752 rc = ptlrpc_queue_wait(req);
2756 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2762 *((u64 *)val) = *reply;
2764 ptlrpc_req_finished(req);
2766 } else if (KEY_IS(KEY_FIEMAP)) {
2767 struct ll_fiemap_info_key *fm_key = key;
2768 struct ldlm_res_id res_id;
2769 ldlm_policy_data_t policy;
2770 struct lustre_handle lockh;
2771 enum ldlm_mode mode = 0;
2772 struct ptlrpc_request *req;
2773 struct ll_user_fiemap *reply;
2777 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2780 policy.l_extent.start = fm_key->fiemap.fm_start &
2783 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2784 fm_key->fiemap.fm_start + PAGE_SIZE - 1)
2785 policy.l_extent.end = OBD_OBJECT_EOF;
2787 policy.l_extent.end = (fm_key->fiemap.fm_start +
2788 fm_key->fiemap.fm_length +
2789 PAGE_SIZE - 1) & PAGE_MASK;
2791 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2792 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2793 LDLM_FL_BLOCK_GRANTED |
2795 &res_id, LDLM_EXTENT, &policy,
2796 LCK_PR | LCK_PW, &lockh, 0);
2797 if (mode) { /* lock is cached on client */
2798 if (mode != LCK_PR) {
2799 ldlm_lock_addref(&lockh, LCK_PR);
2800 ldlm_lock_decref(&lockh, LCK_PW);
2802 } else { /* no cached lock, needs acquire lock on server side */
2803 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2804 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2808 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2809 &RQF_OST_GET_INFO_FIEMAP);
2815 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2816 RCL_CLIENT, keylen);
2817 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2818 RCL_CLIENT, *vallen);
2819 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2820 RCL_SERVER, *vallen);
2822 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2824 ptlrpc_request_free(req);
2828 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2829 memcpy(tmp, key, keylen);
2830 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2831 memcpy(tmp, val, *vallen);
2833 ptlrpc_request_set_replen(req);
2834 rc = ptlrpc_queue_wait(req);
2838 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2844 memcpy(val, reply, *vallen);
2846 ptlrpc_req_finished(req);
2849 ldlm_lock_decref(&lockh, LCK_PR);
2856 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2857 u32 keylen, void *key, u32 vallen,
2858 void *val, struct ptlrpc_request_set *set)
2860 struct ptlrpc_request *req;
2861 struct obd_device *obd = exp->exp_obd;
2862 struct obd_import *imp = class_exp2cliimp(exp);
2866 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2868 if (KEY_IS(KEY_CHECKSUM)) {
2869 if (vallen != sizeof(int))
2871 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2875 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2876 sptlrpc_conf_client_adapt(obd);
2880 if (KEY_IS(KEY_FLUSH_CTX)) {
2881 sptlrpc_import_flush_my_ctx(imp);
2885 if (KEY_IS(KEY_CACHE_SET)) {
2886 struct client_obd *cli = &obd->u.cli;
2888 LASSERT(!cli->cl_cache); /* only once */
2889 cli->cl_cache = val;
2890 atomic_inc(&cli->cl_cache->ccc_users);
2891 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2893 /* add this osc into entity list */
2894 LASSERT(list_empty(&cli->cl_lru_osc));
2895 spin_lock(&cli->cl_cache->ccc_lru_lock);
2896 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2897 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2902 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2903 struct client_obd *cli = &obd->u.cli;
2904 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2905 int target = *(int *)val;
2907 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2912 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2915 /* We pass all other commands directly to OST. Since nobody calls osc
2916 * methods directly and everybody is supposed to go through LOV, we
2917 * assume lov checked invalid values for us.
2918 * The only recognised values so far are evict_by_nid and mds_conn.
2919 * Even if something bad goes through, we'd get a -EINVAL from OST
2923 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2924 &RQF_OST_SET_GRANT_INFO :
2929 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2930 RCL_CLIENT, keylen);
2931 if (!KEY_IS(KEY_GRANT_SHRINK))
2932 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2933 RCL_CLIENT, vallen);
2934 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2936 ptlrpc_request_free(req);
2940 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2941 memcpy(tmp, key, keylen);
2942 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2945 memcpy(tmp, val, vallen);
2947 if (KEY_IS(KEY_GRANT_SHRINK)) {
2948 struct osc_brw_async_args *aa;
2951 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2952 aa = ptlrpc_req_async_args(req);
2953 oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
2955 ptlrpc_req_finished(req);
2958 *oa = ((struct ost_body *)val)->oa;
2960 req->rq_interpret_reply = osc_shrink_grant_interpret;
2963 ptlrpc_request_set_replen(req);
2964 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2966 ptlrpc_set_add_req(set, req);
2967 ptlrpc_check_set(NULL, set);
2969 ptlrpcd_add_req(req);
2975 static int osc_reconnect(const struct lu_env *env,
2976 struct obd_export *exp, struct obd_device *obd,
2977 struct obd_uuid *cluuid,
2978 struct obd_connect_data *data,
2981 struct client_obd *cli = &obd->u.cli;
2983 if (data && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2986 spin_lock(&cli->cl_loi_list_lock);
2987 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2988 2 * cli_brw_size(obd);
2989 lost_grant = cli->cl_lost_grant;
2990 cli->cl_lost_grant = 0;
2991 spin_unlock(&cli->cl_loi_list_lock);
2993 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d ocd_grant: %d, lost: %ld.\n",
2994 data->ocd_connect_flags,
2995 data->ocd_version, data->ocd_grant, lost_grant);
3001 static int osc_disconnect(struct obd_export *exp)
3003 struct obd_device *obd = class_exp2obd(exp);
3006 rc = client_disconnect_export(exp);
3008 * Initially we put del_shrink_grant before disconnect_export, but it
3009 * causes the following problem if setup (connect) and cleanup
3010 * (disconnect) are tangled together.
3011 * connect p1 disconnect p2
3012 * ptlrpc_connect_import
3013 * ............... class_manual_cleanup
3016 * ptlrpc_connect_interrupt
3018 * add this client to shrink list
3020 * Bang! pinger trigger the shrink.
3021 * So the osc should be disconnected from the shrink list, after we
3022 * are sure the import has been destroyed. BUG18662
3024 if (!obd->u.cli.cl_import)
3025 osc_del_shrink_grant(&obd->u.cli);
3029 static int osc_import_event(struct obd_device *obd,
3030 struct obd_import *imp,
3031 enum obd_import_event event)
3033 struct client_obd *cli;
3036 LASSERT(imp->imp_obd == obd);
3039 case IMP_EVENT_DISCON: {
3041 spin_lock(&cli->cl_loi_list_lock);
3042 cli->cl_avail_grant = 0;
3043 cli->cl_lost_grant = 0;
3044 spin_unlock(&cli->cl_loi_list_lock);
3047 case IMP_EVENT_INACTIVE: {
3048 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3051 case IMP_EVENT_INVALIDATE: {
3052 struct ldlm_namespace *ns = obd->obd_namespace;
3056 env = cl_env_get(&refcheck);
3060 /* all pages go to failing rpcs due to the invalid
3063 osc_io_unplug(env, cli, NULL);
3065 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3066 cl_env_put(env, &refcheck);
3072 case IMP_EVENT_ACTIVE: {
3073 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3076 case IMP_EVENT_OCD: {
3077 struct obd_connect_data *ocd = &imp->imp_connect_data;
3079 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3080 osc_init_grant(&obd->u.cli, ocd);
3083 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3084 imp->imp_client->cli_request_portal = OST_REQUEST_PORTAL;
3086 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3089 case IMP_EVENT_DEACTIVATE: {
3090 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3093 case IMP_EVENT_ACTIVATE: {
3094 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3098 CERROR("Unknown import event %d\n", event);
3105 * Determine whether the lock can be canceled before replaying the lock
3106 * during recovery, see bug16774 for detailed information.
3108 * \retval zero the lock can't be canceled
3109 * \retval other ok to cancel
3111 static int osc_cancel_weight(struct ldlm_lock *lock)
3114 * Cancel all unused and granted extent lock.
3116 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3117 lock->l_granted_mode == lock->l_req_mode &&
3118 osc_ldlm_weigh_ast(lock) == 0)
3124 static int brw_queue_work(const struct lu_env *env, void *data)
3126 struct client_obd *cli = data;
3128 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3130 osc_io_unplug(env, cli, NULL);
3134 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3136 struct lprocfs_static_vars lvars = { NULL };
3137 struct client_obd *cli = &obd->u.cli;
3144 rc = ptlrpcd_addref();
3148 rc = client_obd_setup(obd, lcfg);
3152 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3153 if (IS_ERR(handler)) {
3154 rc = PTR_ERR(handler);
3155 goto out_client_setup;
3157 cli->cl_writeback_work = handler;
3159 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3160 if (IS_ERR(handler)) {
3161 rc = PTR_ERR(handler);
3162 goto out_ptlrpcd_work;
3165 cli->cl_lru_work = handler;
3167 rc = osc_quota_setup(obd);
3169 goto out_ptlrpcd_work;
3171 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3172 lprocfs_osc_init_vars(&lvars);
3173 if (lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars) == 0) {
3174 lproc_osc_attach_seqstat(obd);
3175 sptlrpc_lprocfs_cliobd_attach(obd);
3176 ptlrpc_lprocfs_register_obd(obd);
3180 * We try to control the total number of requests with a upper limit
3181 * osc_reqpool_maxreqcount. There might be some race which will cause
3182 * over-limit allocation, but it is fine.
3184 req_count = atomic_read(&osc_pool_req_count);
3185 if (req_count < osc_reqpool_maxreqcount) {
3186 adding = cli->cl_max_rpcs_in_flight + 2;
3187 if (req_count + adding > osc_reqpool_maxreqcount)
3188 adding = osc_reqpool_maxreqcount - req_count;
3190 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3191 atomic_add(added, &osc_pool_req_count);
3194 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3195 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3199 if (cli->cl_writeback_work) {
3200 ptlrpcd_destroy_work(cli->cl_writeback_work);
3201 cli->cl_writeback_work = NULL;
3203 if (cli->cl_lru_work) {
3204 ptlrpcd_destroy_work(cli->cl_lru_work);
3205 cli->cl_lru_work = NULL;
3208 client_obd_cleanup(obd);
3214 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3217 case OBD_CLEANUP_EARLY: {
3218 struct obd_import *imp;
3220 imp = obd->u.cli.cl_import;
3221 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3222 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3223 ptlrpc_deactivate_import(imp);
3224 spin_lock(&imp->imp_lock);
3225 imp->imp_pingable = 0;
3226 spin_unlock(&imp->imp_lock);
3229 case OBD_CLEANUP_EXPORTS: {
3230 struct client_obd *cli = &obd->u.cli;
3232 * for echo client, export may be on zombie list, wait for
3233 * zombie thread to cull it, because cli.cl_import will be
3234 * cleared in client_disconnect_export():
3235 * class_export_destroy() -> obd_cleanup() ->
3236 * echo_device_free() -> echo_client_cleanup() ->
3237 * obd_disconnect() -> osc_disconnect() ->
3238 * client_disconnect_export()
3240 obd_zombie_barrier();
3241 if (cli->cl_writeback_work) {
3242 ptlrpcd_destroy_work(cli->cl_writeback_work);
3243 cli->cl_writeback_work = NULL;
3245 if (cli->cl_lru_work) {
3246 ptlrpcd_destroy_work(cli->cl_lru_work);
3247 cli->cl_lru_work = NULL;
3249 obd_cleanup_client_import(obd);
3250 ptlrpc_lprocfs_unregister_obd(obd);
3251 lprocfs_obd_cleanup(obd);
3258 static int osc_cleanup(struct obd_device *obd)
3260 struct client_obd *cli = &obd->u.cli;
3264 if (cli->cl_cache) {
3265 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3266 spin_lock(&cli->cl_cache->ccc_lru_lock);
3267 list_del_init(&cli->cl_lru_osc);
3268 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3269 cli->cl_lru_left = NULL;
3270 atomic_dec(&cli->cl_cache->ccc_users);
3271 cli->cl_cache = NULL;
3274 /* free memory of osc quota cache */
3275 osc_quota_cleanup(obd);
3277 rc = client_obd_cleanup(obd);
3283 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3285 struct lprocfs_static_vars lvars = { NULL };
3288 lprocfs_osc_init_vars(&lvars);
3290 switch (lcfg->lcfg_command) {
3292 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3302 static int osc_process_config(struct obd_device *obd, u32 len, void *buf)
3304 return osc_process_config_base(obd, buf);
3307 static struct obd_ops osc_obd_ops = {
3308 .owner = THIS_MODULE,
3310 .precleanup = osc_precleanup,
3311 .cleanup = osc_cleanup,
3312 .add_conn = client_import_add_conn,
3313 .del_conn = client_import_del_conn,
3314 .connect = client_connect_import,
3315 .reconnect = osc_reconnect,
3316 .disconnect = osc_disconnect,
3317 .statfs = osc_statfs,
3318 .statfs_async = osc_statfs_async,
3319 .packmd = osc_packmd,
3320 .unpackmd = osc_unpackmd,
3321 .create = osc_create,
3322 .destroy = osc_destroy,
3323 .getattr = osc_getattr,
3324 .getattr_async = osc_getattr_async,
3325 .setattr = osc_setattr,
3326 .setattr_async = osc_setattr_async,
3327 .find_cbdata = osc_find_cbdata,
3328 .iocontrol = osc_iocontrol,
3329 .get_info = osc_get_info,
3330 .set_info_async = osc_set_info_async,
3331 .import_event = osc_import_event,
3332 .process_config = osc_process_config,
3333 .quotactl = osc_quotactl,
3334 .quotacheck = osc_quotacheck,
3337 extern struct lu_kmem_descr osc_caches[];
3338 extern struct lock_class_key osc_ast_guard_class;
3340 static int __init osc_init(void)
3342 struct lprocfs_static_vars lvars = { NULL };
3343 unsigned int reqpool_size;
3344 unsigned int reqsize;
3347 /* print an address of _any_ initialized kernel symbol from this
3348 * module, to allow debugging with gdb that doesn't support data
3349 * symbols from modules.
3351 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3353 rc = lu_kmem_init(osc_caches);
3357 lprocfs_osc_init_vars(&lvars);
3359 rc = class_register_type(&osc_obd_ops, NULL,
3360 LUSTRE_OSC_NAME, &osc_device_type);
3364 /* This is obviously too much memory, only prevent overflow here */
3365 if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0) {
3370 reqpool_size = osc_reqpool_mem_max << 20;
3373 while (reqsize < OST_MAXREQSIZE)
3374 reqsize = reqsize << 1;
3377 * We don't enlarge the request count in OSC pool according to
3378 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3379 * tried after normal allocation failed. So a small OSC pool won't
3380 * cause much performance degression in most of cases.
3382 osc_reqpool_maxreqcount = reqpool_size / reqsize;
3384 atomic_set(&osc_pool_req_count, 0);
3385 osc_rq_pool = ptlrpc_init_rq_pool(0, OST_MAXREQSIZE,
3386 ptlrpc_add_rqs_to_pool);
3394 class_unregister_type(LUSTRE_OSC_NAME);
3396 lu_kmem_fini(osc_caches);
3400 static void /*__exit*/ osc_exit(void)
3402 class_unregister_type(LUSTRE_OSC_NAME);
3403 lu_kmem_fini(osc_caches);
3404 ptlrpc_free_rq_pool(osc_rq_pool);
3407 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3408 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3409 MODULE_LICENSE("GPL");
3410 MODULE_VERSION(LUSTRE_VERSION_STRING);
3412 module_init(osc_init);
3413 module_exit(osc_exit);