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, 2013, 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 <libcfs/libcfs.h>
42 # include <liblustre.h>
45 #include <lustre_dlm.h>
46 #include <lustre_net.h>
47 #include <lustre/lustre_user.h>
48 #include <obd_cksum.h>
54 #include <lustre_ha.h>
55 #include <lprocfs_status.h>
56 #include <lustre_ioctl.h>
57 #include <lustre_log.h>
58 #include <lustre_debug.h>
59 #include <lustre_param.h>
60 #include <lustre_fid.h>
61 #include "osc_internal.h"
62 #include "osc_cl_internal.h"
64 struct osc_brw_async_args {
68 obd_count aa_page_count;
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 obd_capa *aa_ocapa;
75 struct cl_req *aa_clerq;
78 #define osc_grant_args osc_brw_async_args
80 struct osc_async_args {
81 struct obd_info *aa_oi;
84 struct osc_setattr_args {
86 obd_enqueue_update_f sa_upcall;
90 struct osc_fsync_args {
91 struct obd_info *fa_oi;
92 obd_enqueue_update_f fa_upcall;
96 struct osc_enqueue_args {
97 struct obd_export *oa_exp;
99 obd_enqueue_update_f oa_upcall;
101 struct ost_lvb *oa_lvb;
102 struct lustre_handle *oa_lockh;
103 struct ldlm_enqueue_info *oa_ei;
104 unsigned int oa_agl:1;
107 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
108 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
111 /* Unpack OSC object metadata from disk storage (LE byte order). */
112 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
113 struct lov_mds_md *lmm, int lmm_bytes)
116 struct obd_import *imp = class_exp2cliimp(exp);
120 if (lmm_bytes < sizeof(*lmm)) {
121 CERROR("%s: lov_mds_md too small: %d, need %d\n",
122 exp->exp_obd->obd_name, lmm_bytes,
126 /* XXX LOV_MAGIC etc check? */
128 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
129 CERROR("%s: zero lmm_object_id: rc = %d\n",
130 exp->exp_obd->obd_name, -EINVAL);
135 lsm_size = lov_stripe_md_size(1);
139 if (*lsmp != NULL && lmm == NULL) {
140 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
141 OBD_FREE(*lsmp, lsm_size);
147 OBD_ALLOC(*lsmp, lsm_size);
148 if (unlikely(*lsmp == NULL))
150 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
151 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
152 OBD_FREE(*lsmp, lsm_size);
155 loi_init((*lsmp)->lsm_oinfo[0]);
156 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
161 /* XXX zero *lsmp? */
162 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
165 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
166 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
168 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
173 static inline void osc_pack_capa(struct ptlrpc_request *req,
174 struct ost_body *body, void *capa)
176 struct obd_capa *oc = (struct obd_capa *)capa;
177 struct lustre_capa *c;
182 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
185 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
186 DEBUG_CAPA(D_SEC, c, "pack");
189 static inline void osc_pack_req_body(struct ptlrpc_request *req,
190 struct obd_info *oinfo)
192 struct ost_body *body;
194 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
197 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
199 osc_pack_capa(req, body, oinfo->oi_capa);
202 static inline void osc_set_capa_size(struct ptlrpc_request *req,
203 const struct req_msg_field *field,
207 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
209 /* it is already calculated as sizeof struct obd_capa */
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;
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;
250 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
254 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
255 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
257 ptlrpc_request_free(req);
261 osc_pack_req_body(req, oinfo);
263 ptlrpc_request_set_replen(req);
264 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
266 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
267 aa = ptlrpc_req_async_args(req);
270 ptlrpc_set_add_req(set, req);
274 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
275 struct obd_info *oinfo)
277 struct ptlrpc_request *req;
278 struct ost_body *body;
282 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
286 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
287 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
289 ptlrpc_request_free(req);
293 osc_pack_req_body(req, oinfo);
295 ptlrpc_request_set_replen(req);
297 rc = ptlrpc_queue_wait(req);
301 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
303 GOTO(out, rc = -EPROTO);
305 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
306 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
309 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
310 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
314 ptlrpc_req_finished(req);
318 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
319 struct obd_info *oinfo, struct obd_trans_info *oti)
321 struct ptlrpc_request *req;
322 struct ost_body *body;
326 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
328 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
332 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
333 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
335 ptlrpc_request_free(req);
339 osc_pack_req_body(req, oinfo);
341 ptlrpc_request_set_replen(req);
343 rc = ptlrpc_queue_wait(req);
347 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
349 GOTO(out, rc = -EPROTO);
351 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
356 ptlrpc_req_finished(req);
360 static int osc_setattr_interpret(const struct lu_env *env,
361 struct ptlrpc_request *req,
362 struct osc_setattr_args *sa, int rc)
364 struct ost_body *body;
370 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
372 GOTO(out, rc = -EPROTO);
374 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
377 rc = sa->sa_upcall(sa->sa_cookie, rc);
381 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
382 struct obd_trans_info *oti,
383 obd_enqueue_update_f upcall, void *cookie,
384 struct ptlrpc_request_set *rqset)
386 struct ptlrpc_request *req;
387 struct osc_setattr_args *sa;
391 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
395 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
396 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
398 ptlrpc_request_free(req);
402 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
403 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
405 osc_pack_req_body(req, oinfo);
407 ptlrpc_request_set_replen(req);
409 /* do mds to ost setattr asynchronously */
411 /* Do not wait for response. */
412 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
414 req->rq_interpret_reply =
415 (ptlrpc_interpterer_t)osc_setattr_interpret;
417 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
418 sa = ptlrpc_req_async_args(req);
419 sa->sa_oa = oinfo->oi_oa;
420 sa->sa_upcall = upcall;
421 sa->sa_cookie = cookie;
423 if (rqset == PTLRPCD_SET)
424 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
426 ptlrpc_set_add_req(rqset, req);
432 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
433 struct obd_trans_info *oti,
434 struct ptlrpc_request_set *rqset)
436 return osc_setattr_async_base(exp, oinfo, oti,
437 oinfo->oi_cb_up, oinfo, rqset);
440 int osc_real_create(struct obd_export *exp, struct obdo *oa,
441 struct lov_stripe_md **ea, struct obd_trans_info *oti)
443 struct ptlrpc_request *req;
444 struct ost_body *body;
445 struct lov_stripe_md *lsm;
454 rc = obd_alloc_memmd(exp, &lsm);
459 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
461 GOTO(out, rc = -ENOMEM);
463 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
465 ptlrpc_request_free(req);
469 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
472 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
474 ptlrpc_request_set_replen(req);
476 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
477 oa->o_flags == OBD_FL_DELORPHAN) {
479 "delorphan from OST integration");
480 /* Don't resend the delorphan req */
481 req->rq_no_resend = req->rq_no_delay = 1;
484 rc = ptlrpc_queue_wait(req);
488 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
490 GOTO(out_req, rc = -EPROTO);
492 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
493 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
495 oa->o_blksize = cli_brw_size(exp->exp_obd);
496 oa->o_valid |= OBD_MD_FLBLKSZ;
498 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
499 * have valid lsm_oinfo data structs, so don't go touching that.
500 * This needs to be fixed in a big way.
502 lsm->lsm_oi = oa->o_oi;
506 if (oa->o_valid & OBD_MD_FLCOOKIE) {
507 if (oti->oti_logcookies == NULL)
508 oti->oti_logcookies = &oti->oti_onecookie;
510 *oti->oti_logcookies = oa->o_lcookie;
514 CDEBUG(D_HA, "transno: "LPD64"\n",
515 lustre_msg_get_transno(req->rq_repmsg));
517 ptlrpc_req_finished(req);
520 obd_free_memmd(exp, &lsm);
524 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
525 obd_enqueue_update_f upcall, void *cookie,
526 struct ptlrpc_request_set *rqset)
528 struct ptlrpc_request *req;
529 struct osc_setattr_args *sa;
530 struct ost_body *body;
534 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
538 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
539 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
541 ptlrpc_request_free(req);
544 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
545 ptlrpc_at_set_req_timeout(req);
547 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
549 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
551 osc_pack_capa(req, body, oinfo->oi_capa);
553 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
556 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
557 sa = ptlrpc_req_async_args(req);
558 sa->sa_oa = oinfo->oi_oa;
559 sa->sa_upcall = upcall;
560 sa->sa_cookie = cookie;
561 if (rqset == PTLRPCD_SET)
562 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
564 ptlrpc_set_add_req(rqset, req);
569 static int osc_sync_interpret(const struct lu_env *env,
570 struct ptlrpc_request *req,
573 struct osc_fsync_args *fa = arg;
574 struct ost_body *body;
580 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
582 CERROR ("can't unpack ost_body\n");
583 GOTO(out, rc = -EPROTO);
586 *fa->fa_oi->oi_oa = body->oa;
588 rc = fa->fa_upcall(fa->fa_cookie, rc);
592 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
593 obd_enqueue_update_f upcall, void *cookie,
594 struct ptlrpc_request_set *rqset)
596 struct ptlrpc_request *req;
597 struct ost_body *body;
598 struct osc_fsync_args *fa;
602 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
606 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
607 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
609 ptlrpc_request_free(req);
613 /* overload the size and blocks fields in the oa with start/end */
614 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
616 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
618 osc_pack_capa(req, body, oinfo->oi_capa);
620 ptlrpc_request_set_replen(req);
621 req->rq_interpret_reply = osc_sync_interpret;
623 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
624 fa = ptlrpc_req_async_args(req);
626 fa->fa_upcall = upcall;
627 fa->fa_cookie = cookie;
629 if (rqset == PTLRPCD_SET)
630 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
632 ptlrpc_set_add_req(rqset, req);
637 /* Find and cancel locally locks matched by @mode in the resource found by
638 * @objid. Found locks are added into @cancel list. Returns the amount of
639 * locks added to @cancels list. */
640 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
641 struct list_head *cancels,
642 ldlm_mode_t mode, __u64 lock_flags)
644 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
645 struct ldlm_res_id res_id;
646 struct ldlm_resource *res;
650 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
651 * export) but disabled through procfs (flag in NS).
653 * This distinguishes from a case when ELC is not supported originally,
654 * when we still want to cancel locks in advance and just cancel them
655 * locally, without sending any RPC. */
656 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
659 ostid_build_res_name(&oa->o_oi, &res_id);
660 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
664 LDLM_RESOURCE_ADDREF(res);
665 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
666 lock_flags, 0, NULL);
667 LDLM_RESOURCE_DELREF(res);
668 ldlm_resource_putref(res);
672 static int osc_destroy_interpret(const struct lu_env *env,
673 struct ptlrpc_request *req, void *data,
676 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
678 atomic_dec(&cli->cl_destroy_in_flight);
679 wake_up(&cli->cl_destroy_waitq);
683 static int osc_can_send_destroy(struct client_obd *cli)
685 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
686 cli->cl_max_rpcs_in_flight) {
687 /* The destroy request can be sent */
690 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
691 cli->cl_max_rpcs_in_flight) {
693 * The counter has been modified between the two atomic
696 wake_up(&cli->cl_destroy_waitq);
701 int osc_create(const struct lu_env *env, struct obd_export *exp,
702 struct obdo *oa, struct lov_stripe_md **ea,
703 struct obd_trans_info *oti)
710 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
712 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
713 oa->o_flags == OBD_FL_RECREATE_OBJS) {
714 RETURN(osc_real_create(exp, oa, ea, oti));
717 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
718 RETURN(osc_real_create(exp, oa, ea, oti));
720 /* we should not get here anymore */
726 /* Destroy requests can be async always on the client, and we don't even really
727 * care about the return code since the client cannot do anything at all about
729 * When the MDS is unlinking a filename, it saves the file objects into a
730 * recovery llog, and these object records are cancelled when the OST reports
731 * they were destroyed and sync'd to disk (i.e. transaction committed).
732 * If the client dies, or the OST is down when the object should be destroyed,
733 * the records are not cancelled, and when the OST reconnects to the MDS next,
734 * it will retrieve the llog unlink logs and then sends the log cancellation
735 * cookies to the MDS after committing destroy transactions. */
736 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
737 struct obdo *oa, struct lov_stripe_md *ea,
738 struct obd_trans_info *oti, struct obd_export *md_export,
741 struct client_obd *cli = &exp->exp_obd->u.cli;
742 struct ptlrpc_request *req;
743 struct ost_body *body;
744 struct list_head cancels = LIST_HEAD_INIT(cancels);
749 CDEBUG(D_INFO, "oa NULL\n");
753 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
754 LDLM_FL_DISCARD_DATA);
756 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
758 ldlm_lock_list_put(&cancels, l_bl_ast, count);
762 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
763 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
766 ptlrpc_request_free(req);
770 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
771 ptlrpc_at_set_req_timeout(req);
773 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
774 oa->o_lcookie = *oti->oti_logcookies;
775 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
777 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
779 osc_pack_capa(req, body, (struct obd_capa *)capa);
780 ptlrpc_request_set_replen(req);
782 /* If osc_destory is for destroying the unlink orphan,
783 * sent from MDT to OST, which should not be blocked here,
784 * because the process might be triggered by ptlrpcd, and
785 * it is not good to block ptlrpcd thread (b=16006)*/
786 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
787 req->rq_interpret_reply = osc_destroy_interpret;
788 if (!osc_can_send_destroy(cli)) {
789 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
793 * Wait until the number of on-going destroy RPCs drops
794 * under max_rpc_in_flight
796 l_wait_event_exclusive(cli->cl_destroy_waitq,
797 osc_can_send_destroy(cli), &lwi);
801 /* Do not wait for response */
802 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
806 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
809 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
811 LASSERT(!(oa->o_valid & bits));
814 client_obd_list_lock(&cli->cl_loi_list_lock);
815 oa->o_dirty = cli->cl_dirty;
816 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
817 cli->cl_dirty_max)) {
818 CERROR("dirty %lu - %lu > dirty_max %lu\n",
819 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
821 } else if (unlikely(atomic_read(&obd_unstable_pages) +
822 atomic_read(&obd_dirty_pages) -
823 atomic_read(&obd_dirty_transit_pages) >
824 (long)(obd_max_dirty_pages + 1))) {
825 /* The atomic_read() allowing the atomic_inc() are
826 * not covered by a lock thus they may safely race and trip
827 * this CERROR() unless we add in a small fudge factor (+1). */
828 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
829 cli->cl_import->imp_obd->obd_name,
830 atomic_read(&obd_unstable_pages),
831 atomic_read(&obd_dirty_pages),
832 atomic_read(&obd_dirty_transit_pages),
833 obd_max_dirty_pages);
835 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
836 CERROR("dirty %lu - dirty_max %lu too big???\n",
837 cli->cl_dirty, cli->cl_dirty_max);
840 long max_in_flight = (cli->cl_max_pages_per_rpc <<
842 (cli->cl_max_rpcs_in_flight + 1);
843 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
845 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
846 oa->o_dropped = cli->cl_lost_grant;
847 cli->cl_lost_grant = 0;
848 client_obd_list_unlock(&cli->cl_loi_list_lock);
849 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
850 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
854 void osc_update_next_shrink(struct client_obd *cli)
856 cli->cl_next_shrink_grant =
857 cfs_time_shift(cli->cl_grant_shrink_interval);
858 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
859 cli->cl_next_shrink_grant);
862 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
864 client_obd_list_lock(&cli->cl_loi_list_lock);
865 cli->cl_avail_grant += grant;
866 client_obd_list_unlock(&cli->cl_loi_list_lock);
869 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
871 if (body->oa.o_valid & OBD_MD_FLGRANT) {
872 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
873 __osc_update_grant(cli, body->oa.o_grant);
877 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
878 obd_count keylen, void *key, obd_count vallen,
879 void *val, struct ptlrpc_request_set *set);
881 static int osc_shrink_grant_interpret(const struct lu_env *env,
882 struct ptlrpc_request *req,
885 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
886 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
887 struct ost_body *body;
890 __osc_update_grant(cli, oa->o_grant);
894 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
896 osc_update_grant(cli, body);
902 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
904 client_obd_list_lock(&cli->cl_loi_list_lock);
905 oa->o_grant = cli->cl_avail_grant / 4;
906 cli->cl_avail_grant -= oa->o_grant;
907 client_obd_list_unlock(&cli->cl_loi_list_lock);
908 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
909 oa->o_valid |= OBD_MD_FLFLAGS;
912 oa->o_flags |= OBD_FL_SHRINK_GRANT;
913 osc_update_next_shrink(cli);
916 /* Shrink the current grant, either from some large amount to enough for a
917 * full set of in-flight RPCs, or if we have already shrunk to that limit
918 * then to enough for a single RPC. This avoids keeping more grant than
919 * needed, and avoids shrinking the grant piecemeal. */
920 static int osc_shrink_grant(struct client_obd *cli)
922 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
923 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
925 client_obd_list_lock(&cli->cl_loi_list_lock);
926 if (cli->cl_avail_grant <= target_bytes)
927 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
928 client_obd_list_unlock(&cli->cl_loi_list_lock);
930 return osc_shrink_grant_to_target(cli, target_bytes);
933 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
936 struct ost_body *body;
939 client_obd_list_lock(&cli->cl_loi_list_lock);
940 /* Don't shrink if we are already above or below the desired limit
941 * We don't want to shrink below a single RPC, as that will negatively
942 * impact block allocation and long-term performance. */
943 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
944 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
946 if (target_bytes >= cli->cl_avail_grant) {
947 client_obd_list_unlock(&cli->cl_loi_list_lock);
950 client_obd_list_unlock(&cli->cl_loi_list_lock);
956 osc_announce_cached(cli, &body->oa, 0);
958 client_obd_list_lock(&cli->cl_loi_list_lock);
959 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
960 cli->cl_avail_grant = target_bytes;
961 client_obd_list_unlock(&cli->cl_loi_list_lock);
962 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
963 body->oa.o_valid |= OBD_MD_FLFLAGS;
964 body->oa.o_flags = 0;
966 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
967 osc_update_next_shrink(cli);
969 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
970 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
971 sizeof(*body), body, NULL);
973 __osc_update_grant(cli, body->oa.o_grant);
978 static int osc_should_shrink_grant(struct client_obd *client)
980 cfs_time_t time = cfs_time_current();
981 cfs_time_t next_shrink = client->cl_next_shrink_grant;
983 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
984 OBD_CONNECT_GRANT_SHRINK) == 0)
987 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
988 /* Get the current RPC size directly, instead of going via:
989 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
990 * Keep comment here so that it can be found by searching. */
991 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
993 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
994 client->cl_avail_grant > brw_size)
997 osc_update_next_shrink(client);
1002 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1004 struct client_obd *client;
1006 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1007 if (osc_should_shrink_grant(client))
1008 osc_shrink_grant(client);
1013 static int osc_add_shrink_grant(struct client_obd *client)
1017 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1019 osc_grant_shrink_grant_cb, NULL,
1020 &client->cl_grant_shrink_list);
1022 CERROR("add grant client %s error %d\n",
1023 client->cl_import->imp_obd->obd_name, rc);
1026 CDEBUG(D_CACHE, "add grant client %s \n",
1027 client->cl_import->imp_obd->obd_name);
1028 osc_update_next_shrink(client);
1032 static int osc_del_shrink_grant(struct client_obd *client)
1034 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1038 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1041 * ocd_grant is the total grant amount we're expect to hold: if we've
1042 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1043 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1045 * race is tolerable here: if we're evicted, but imp_state already
1046 * left EVICTED state, then cl_dirty must be 0 already.
1048 client_obd_list_lock(&cli->cl_loi_list_lock);
1049 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1050 cli->cl_avail_grant = ocd->ocd_grant;
1052 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1054 if (cli->cl_avail_grant < 0) {
1055 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1056 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1057 ocd->ocd_grant, cli->cl_dirty);
1058 /* workaround for servers which do not have the patch from
1060 cli->cl_avail_grant = ocd->ocd_grant;
1063 /* determine the appropriate chunk size used by osc_extent. */
1064 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1065 client_obd_list_unlock(&cli->cl_loi_list_lock);
1067 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1068 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1069 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1071 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1072 list_empty(&cli->cl_grant_shrink_list))
1073 osc_add_shrink_grant(cli);
1076 /* We assume that the reason this OSC got a short read is because it read
1077 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1078 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1079 * this stripe never got written at or beyond this stripe offset yet. */
1080 static void handle_short_read(int nob_read, obd_count page_count,
1081 struct brw_page **pga)
1086 /* skip bytes read OK */
1087 while (nob_read > 0) {
1088 LASSERT (page_count > 0);
1090 if (pga[i]->count > nob_read) {
1091 /* EOF inside this page */
1092 ptr = kmap(pga[i]->pg) +
1093 (pga[i]->off & ~CFS_PAGE_MASK);
1094 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1101 nob_read -= pga[i]->count;
1106 /* zero remaining pages */
1107 while (page_count-- > 0) {
1108 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1109 memset(ptr, 0, pga[i]->count);
1115 static int check_write_rcs(struct ptlrpc_request *req,
1116 int requested_nob, int niocount,
1117 obd_count page_count, struct brw_page **pga)
1122 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1123 sizeof(*remote_rcs) *
1125 if (remote_rcs == NULL) {
1126 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1130 /* return error if any niobuf was in error */
1131 for (i = 0; i < niocount; i++) {
1132 if ((int)remote_rcs[i] < 0)
1133 return(remote_rcs[i]);
1135 if (remote_rcs[i] != 0) {
1136 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1137 i, remote_rcs[i], req);
1142 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1143 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1144 req->rq_bulk->bd_nob_transferred, requested_nob);
1151 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1153 if (p1->flag != p2->flag) {
1154 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1155 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1156 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1158 /* warn if we try to combine flags that we don't know to be
1159 * safe to combine */
1160 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1161 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1162 "report this at http://bugs.whamcloud.com/\n",
1163 p1->flag, p2->flag);
1168 return (p1->off + p1->count == p2->off);
1171 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1172 struct brw_page **pga, int opc,
1173 cksum_type_t cksum_type)
1177 struct cfs_crypto_hash_desc *hdesc;
1178 unsigned int bufsize;
1180 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1182 LASSERT(pg_count > 0);
1184 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1185 if (IS_ERR(hdesc)) {
1186 CERROR("Unable to initialize checksum hash %s\n",
1187 cfs_crypto_hash_name(cfs_alg));
1188 return PTR_ERR(hdesc);
1191 while (nob > 0 && pg_count > 0) {
1192 int count = pga[i]->count > nob ? nob : pga[i]->count;
1194 /* corrupt the data before we compute the checksum, to
1195 * simulate an OST->client data error */
1196 if (i == 0 && opc == OST_READ &&
1197 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1198 unsigned char *ptr = kmap(pga[i]->pg);
1199 int off = pga[i]->off & ~CFS_PAGE_MASK;
1200 memcpy(ptr + off, "bad1", min(4, nob));
1203 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1204 pga[i]->off & ~CFS_PAGE_MASK,
1206 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1207 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1209 nob -= pga[i]->count;
1215 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1218 cfs_crypto_hash_final(hdesc, NULL, NULL);
1220 /* For sending we only compute the wrong checksum instead
1221 * of corrupting the data so it is still correct on a redo */
1222 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1228 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1229 struct lov_stripe_md *lsm, obd_count page_count,
1230 struct brw_page **pga,
1231 struct ptlrpc_request **reqp,
1232 struct obd_capa *ocapa, int reserve,
1235 struct ptlrpc_request *req;
1236 struct ptlrpc_bulk_desc *desc;
1237 struct ost_body *body;
1238 struct obd_ioobj *ioobj;
1239 struct niobuf_remote *niobuf;
1240 int niocount, i, requested_nob, opc, rc;
1241 struct osc_brw_async_args *aa;
1242 struct req_capsule *pill;
1243 struct brw_page *pg_prev;
1246 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1247 RETURN(-ENOMEM); /* Recoverable */
1248 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1249 RETURN(-EINVAL); /* Fatal */
1251 if ((cmd & OBD_BRW_WRITE) != 0) {
1253 req = ptlrpc_request_alloc_pool(cli->cl_import,
1254 cli->cl_import->imp_rq_pool,
1255 &RQF_OST_BRW_WRITE);
1258 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1263 for (niocount = i = 1; i < page_count; i++) {
1264 if (!can_merge_pages(pga[i - 1], pga[i]))
1268 pill = &req->rq_pill;
1269 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1271 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1272 niocount * sizeof(*niobuf));
1273 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1277 ptlrpc_request_free(req);
1280 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1281 ptlrpc_at_set_req_timeout(req);
1282 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1284 req->rq_no_retry_einprogress = 1;
1286 desc = ptlrpc_prep_bulk_imp(req, page_count,
1287 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1288 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1292 GOTO(out, rc = -ENOMEM);
1293 /* NB request now owns desc and will free it when it gets freed */
1295 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1296 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1297 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1298 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1300 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1302 obdo_to_ioobj(oa, ioobj);
1303 ioobj->ioo_bufcnt = niocount;
1304 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1305 * that might be send for this request. The actual number is decided
1306 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1307 * "max - 1" for old client compatibility sending "0", and also so the
1308 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1309 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1310 osc_pack_capa(req, body, ocapa);
1311 LASSERT(page_count > 0);
1313 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1314 struct brw_page *pg = pga[i];
1315 int poff = pg->off & ~CFS_PAGE_MASK;
1317 LASSERT(pg->count > 0);
1318 /* make sure there is no gap in the middle of page array */
1319 LASSERTF(page_count == 1 ||
1320 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1321 ergo(i > 0 && i < page_count - 1,
1322 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1323 ergo(i == page_count - 1, poff == 0)),
1324 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1325 i, page_count, pg, pg->off, pg->count);
1327 LASSERTF(i == 0 || pg->off > pg_prev->off,
1328 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1329 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1331 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1332 pg_prev->pg, page_private(pg_prev->pg),
1333 pg_prev->pg->index, pg_prev->off);
1335 LASSERTF(i == 0 || pg->off > pg_prev->off,
1336 "i %d p_c %u\n", i, page_count);
1338 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1339 (pg->flag & OBD_BRW_SRVLOCK));
1341 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1342 requested_nob += pg->count;
1344 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1346 niobuf->len += pg->count;
1348 niobuf->offset = pg->off;
1349 niobuf->len = pg->count;
1350 niobuf->flags = pg->flag;
1355 LASSERTF((void *)(niobuf - niocount) ==
1356 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1357 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1358 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1360 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1362 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1363 body->oa.o_valid |= OBD_MD_FLFLAGS;
1364 body->oa.o_flags = 0;
1366 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1369 if (osc_should_shrink_grant(cli))
1370 osc_shrink_grant_local(cli, &body->oa);
1372 /* size[REQ_REC_OFF] still sizeof (*body) */
1373 if (opc == OST_WRITE) {
1374 if (cli->cl_checksum &&
1375 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1376 /* store cl_cksum_type in a local variable since
1377 * it can be changed via lprocfs */
1378 cksum_type_t cksum_type = cli->cl_cksum_type;
1380 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1381 oa->o_flags &= OBD_FL_LOCAL_MASK;
1382 body->oa.o_flags = 0;
1384 body->oa.o_flags |= cksum_type_pack(cksum_type);
1385 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1386 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1390 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1392 /* save this in 'oa', too, for later checking */
1393 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1394 oa->o_flags |= cksum_type_pack(cksum_type);
1396 /* clear out the checksum flag, in case this is a
1397 * resend but cl_checksum is no longer set. b=11238 */
1398 oa->o_valid &= ~OBD_MD_FLCKSUM;
1400 oa->o_cksum = body->oa.o_cksum;
1401 /* 1 RC per niobuf */
1402 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1403 sizeof(__u32) * niocount);
1405 if (cli->cl_checksum &&
1406 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1407 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1408 body->oa.o_flags = 0;
1409 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1410 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1413 ptlrpc_request_set_replen(req);
1415 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1416 aa = ptlrpc_req_async_args(req);
1418 aa->aa_requested_nob = requested_nob;
1419 aa->aa_nio_count = niocount;
1420 aa->aa_page_count = page_count;
1424 INIT_LIST_HEAD(&aa->aa_oaps);
1425 if (ocapa && reserve)
1426 aa->aa_ocapa = capa_get(ocapa);
1432 ptlrpc_req_finished(req);
1436 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1437 __u32 client_cksum, __u32 server_cksum, int nob,
1438 obd_count page_count, struct brw_page **pga,
1439 cksum_type_t client_cksum_type)
1443 cksum_type_t cksum_type;
1445 if (server_cksum == client_cksum) {
1446 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1450 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1452 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1455 if (cksum_type != client_cksum_type)
1456 msg = "the server did not use the checksum type specified in "
1457 "the original request - likely a protocol problem";
1458 else if (new_cksum == server_cksum)
1459 msg = "changed on the client after we checksummed it - "
1460 "likely false positive due to mmap IO (bug 11742)";
1461 else if (new_cksum == client_cksum)
1462 msg = "changed in transit before arrival at OST";
1464 msg = "changed in transit AND doesn't match the original - "
1465 "likely false positive due to mmap IO (bug 11742)";
1467 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1468 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1469 msg, libcfs_nid2str(peer->nid),
1470 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1471 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1472 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1473 POSTID(&oa->o_oi), pga[0]->off,
1474 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1475 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1476 "client csum now %x\n", client_cksum, client_cksum_type,
1477 server_cksum, cksum_type, new_cksum);
1481 /* Note rc enters this function as number of bytes transferred */
1482 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1484 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1485 const lnet_process_id_t *peer =
1486 &req->rq_import->imp_connection->c_peer;
1487 struct client_obd *cli = aa->aa_cli;
1488 struct ost_body *body;
1489 __u32 client_cksum = 0;
1492 if (rc < 0 && rc != -EDQUOT) {
1493 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1497 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1498 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1500 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1504 /* set/clear over quota flag for a uid/gid */
1505 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1506 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1507 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1509 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1510 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1512 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1515 osc_update_grant(cli, body);
1520 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1521 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1523 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1525 CERROR("Unexpected +ve rc %d\n", rc);
1528 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1530 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1533 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1534 check_write_checksum(&body->oa, peer, client_cksum,
1535 body->oa.o_cksum, aa->aa_requested_nob,
1536 aa->aa_page_count, aa->aa_ppga,
1537 cksum_type_unpack(aa->aa_oa->o_flags)))
1540 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1541 aa->aa_page_count, aa->aa_ppga);
1545 /* The rest of this function executes only for OST_READs */
1547 /* if unwrap_bulk failed, return -EAGAIN to retry */
1548 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1550 GOTO(out, rc = -EAGAIN);
1552 if (rc > aa->aa_requested_nob) {
1553 CERROR("Unexpected rc %d (%d requested)\n", rc,
1554 aa->aa_requested_nob);
1558 if (rc != req->rq_bulk->bd_nob_transferred) {
1559 CERROR ("Unexpected rc %d (%d transferred)\n",
1560 rc, req->rq_bulk->bd_nob_transferred);
1564 if (rc < aa->aa_requested_nob)
1565 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1567 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1568 static int cksum_counter;
1569 __u32 server_cksum = body->oa.o_cksum;
1572 cksum_type_t cksum_type;
1574 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1575 body->oa.o_flags : 0);
1576 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1577 aa->aa_ppga, OST_READ,
1580 if (peer->nid == req->rq_bulk->bd_sender) {
1584 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1587 if (server_cksum != client_cksum) {
1588 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1589 "%s%s%s inode "DFID" object "DOSTID
1590 " extent ["LPU64"-"LPU64"]\n",
1591 req->rq_import->imp_obd->obd_name,
1592 libcfs_nid2str(peer->nid),
1594 body->oa.o_valid & OBD_MD_FLFID ?
1595 body->oa.o_parent_seq : (__u64)0,
1596 body->oa.o_valid & OBD_MD_FLFID ?
1597 body->oa.o_parent_oid : 0,
1598 body->oa.o_valid & OBD_MD_FLFID ?
1599 body->oa.o_parent_ver : 0,
1600 POSTID(&body->oa.o_oi),
1601 aa->aa_ppga[0]->off,
1602 aa->aa_ppga[aa->aa_page_count-1]->off +
1603 aa->aa_ppga[aa->aa_page_count-1]->count -
1605 CERROR("client %x, server %x, cksum_type %x\n",
1606 client_cksum, server_cksum, cksum_type);
1608 aa->aa_oa->o_cksum = client_cksum;
1612 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1615 } else if (unlikely(client_cksum)) {
1616 static int cksum_missed;
1619 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1620 CERROR("Checksum %u requested from %s but not sent\n",
1621 cksum_missed, libcfs_nid2str(peer->nid));
1627 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1628 aa->aa_oa, &body->oa);
1633 static int osc_brw_redo_request(struct ptlrpc_request *request,
1634 struct osc_brw_async_args *aa, int rc)
1636 struct ptlrpc_request *new_req;
1637 struct osc_brw_async_args *new_aa;
1638 struct osc_async_page *oap;
1641 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1642 "redo for recoverable error %d", rc);
1644 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1645 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1646 aa->aa_cli, aa->aa_oa,
1647 NULL /* lsm unused by osc currently */,
1648 aa->aa_page_count, aa->aa_ppga,
1649 &new_req, aa->aa_ocapa, 0, 1);
1653 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1654 if (oap->oap_request != NULL) {
1655 LASSERTF(request == oap->oap_request,
1656 "request %p != oap_request %p\n",
1657 request, oap->oap_request);
1658 if (oap->oap_interrupted) {
1659 ptlrpc_req_finished(new_req);
1664 /* New request takes over pga and oaps from old request.
1665 * Note that copying a list_head doesn't work, need to move it... */
1667 new_req->rq_interpret_reply = request->rq_interpret_reply;
1668 new_req->rq_async_args = request->rq_async_args;
1669 new_req->rq_commit_cb = request->rq_commit_cb;
1670 /* cap resend delay to the current request timeout, this is similar to
1671 * what ptlrpc does (see after_reply()) */
1672 if (aa->aa_resends > new_req->rq_timeout)
1673 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1675 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1676 new_req->rq_generation_set = 1;
1677 new_req->rq_import_generation = request->rq_import_generation;
1679 new_aa = ptlrpc_req_async_args(new_req);
1681 INIT_LIST_HEAD(&new_aa->aa_oaps);
1682 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1683 INIT_LIST_HEAD(&new_aa->aa_exts);
1684 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1685 new_aa->aa_resends = aa->aa_resends;
1687 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1688 if (oap->oap_request) {
1689 ptlrpc_req_finished(oap->oap_request);
1690 oap->oap_request = ptlrpc_request_addref(new_req);
1694 new_aa->aa_ocapa = aa->aa_ocapa;
1695 aa->aa_ocapa = NULL;
1697 /* XXX: This code will run into problem if we're going to support
1698 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1699 * and wait for all of them to be finished. We should inherit request
1700 * set from old request. */
1701 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1703 DEBUG_REQ(D_INFO, new_req, "new request");
1708 * ugh, we want disk allocation on the target to happen in offset order. we'll
1709 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1710 * fine for our small page arrays and doesn't require allocation. its an
1711 * insertion sort that swaps elements that are strides apart, shrinking the
1712 * stride down until its '1' and the array is sorted.
1714 static void sort_brw_pages(struct brw_page **array, int num)
1717 struct brw_page *tmp;
1721 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1726 for (i = stride ; i < num ; i++) {
1729 while (j >= stride && array[j - stride]->off > tmp->off) {
1730 array[j] = array[j - stride];
1735 } while (stride > 1);
1738 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1740 LASSERT(ppga != NULL);
1741 OBD_FREE(ppga, sizeof(*ppga) * count);
1744 static int brw_interpret(const struct lu_env *env,
1745 struct ptlrpc_request *req, void *data, int rc)
1747 struct osc_brw_async_args *aa = data;
1748 struct osc_extent *ext;
1749 struct osc_extent *tmp;
1750 struct client_obd *cli = aa->aa_cli;
1753 rc = osc_brw_fini_request(req, rc);
1754 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1755 /* When server return -EINPROGRESS, client should always retry
1756 * regardless of the number of times the bulk was resent already. */
1757 if (osc_recoverable_error(rc)) {
1758 if (req->rq_import_generation !=
1759 req->rq_import->imp_generation) {
1760 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1761 ""DOSTID", rc = %d.\n",
1762 req->rq_import->imp_obd->obd_name,
1763 POSTID(&aa->aa_oa->o_oi), rc);
1764 } else if (rc == -EINPROGRESS ||
1765 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1766 rc = osc_brw_redo_request(req, aa, rc);
1768 CERROR("%s: too many resent retries for object: "
1769 ""LPU64":"LPU64", rc = %d.\n",
1770 req->rq_import->imp_obd->obd_name,
1771 POSTID(&aa->aa_oa->o_oi), rc);
1776 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1781 capa_put(aa->aa_ocapa);
1782 aa->aa_ocapa = NULL;
1786 struct obdo *oa = aa->aa_oa;
1787 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1788 unsigned long valid = 0;
1789 struct cl_object *obj;
1790 struct osc_async_page *last;
1792 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1793 obj = osc2cl(last->oap_obj);
1795 cl_object_attr_lock(obj);
1796 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1797 attr->cat_blocks = oa->o_blocks;
1798 valid |= CAT_BLOCKS;
1800 if (oa->o_valid & OBD_MD_FLMTIME) {
1801 attr->cat_mtime = oa->o_mtime;
1804 if (oa->o_valid & OBD_MD_FLATIME) {
1805 attr->cat_atime = oa->o_atime;
1808 if (oa->o_valid & OBD_MD_FLCTIME) {
1809 attr->cat_ctime = oa->o_ctime;
1813 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1814 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1815 loff_t last_off = last->oap_count + last->oap_obj_off;
1817 /* Change file size if this is an out of quota or
1818 * direct IO write and it extends the file size */
1819 if (loi->loi_lvb.lvb_size < last_off) {
1820 attr->cat_size = last_off;
1823 /* Extend KMS if it's not a lockless write */
1824 if (loi->loi_kms < last_off &&
1825 oap2osc_page(last)->ops_srvlock == 0) {
1826 attr->cat_kms = last_off;
1832 cl_object_attr_set(env, obj, attr, valid);
1833 cl_object_attr_unlock(obj);
1835 OBDO_FREE(aa->aa_oa);
1837 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1838 list_del_init(&ext->oe_link);
1839 osc_extent_finish(env, ext, 1, rc);
1841 LASSERT(list_empty(&aa->aa_exts));
1842 LASSERT(list_empty(&aa->aa_oaps));
1844 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1845 req->rq_bulk->bd_nob_transferred);
1846 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1847 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1849 client_obd_list_lock(&cli->cl_loi_list_lock);
1850 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1851 * is called so we know whether to go to sync BRWs or wait for more
1852 * RPCs to complete */
1853 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1854 cli->cl_w_in_flight--;
1856 cli->cl_r_in_flight--;
1857 osc_wake_cache_waiters(cli);
1858 client_obd_list_unlock(&cli->cl_loi_list_lock);
1860 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
1864 static void brw_commit(struct ptlrpc_request *req)
1866 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1867 * this called via the rq_commit_cb, I need to ensure
1868 * osc_dec_unstable_pages is still called. Otherwise unstable
1869 * pages may be leaked. */
1870 spin_lock(&req->rq_lock);
1871 if (likely(req->rq_unstable)) {
1872 req->rq_unstable = 0;
1873 spin_unlock(&req->rq_lock);
1875 osc_dec_unstable_pages(req);
1877 req->rq_committed = 1;
1878 spin_unlock(&req->rq_lock);
1883 * Build an RPC by the list of extent @ext_list. The caller must ensure
1884 * that the total pages in this list are NOT over max pages per RPC.
1885 * Extents in the list must be in OES_RPC state.
1887 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1888 struct list_head *ext_list, int cmd, pdl_policy_t pol)
1890 struct ptlrpc_request *req = NULL;
1891 struct osc_extent *ext;
1892 struct brw_page **pga = NULL;
1893 struct osc_brw_async_args *aa = NULL;
1894 struct obdo *oa = NULL;
1895 struct osc_async_page *oap;
1896 struct osc_async_page *tmp;
1897 struct cl_req *clerq = NULL;
1898 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
1900 struct ldlm_lock *lock = NULL;
1901 struct cl_req_attr *crattr = NULL;
1902 obd_off starting_offset = OBD_OBJECT_EOF;
1903 obd_off ending_offset = 0;
1909 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1912 LASSERT(!list_empty(ext_list));
1914 /* add pages into rpc_list to build BRW rpc */
1915 list_for_each_entry(ext, ext_list, oe_link) {
1916 LASSERT(ext->oe_state == OES_RPC);
1917 mem_tight |= ext->oe_memalloc;
1918 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1920 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1921 if (starting_offset > oap->oap_obj_off)
1922 starting_offset = oap->oap_obj_off;
1924 LASSERT(oap->oap_page_off == 0);
1925 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1926 ending_offset = oap->oap_obj_off +
1929 LASSERT(oap->oap_page_off + oap->oap_count ==
1935 mpflag = cfs_memory_pressure_get_and_set();
1937 OBD_ALLOC(crattr, sizeof(*crattr));
1939 GOTO(out, rc = -ENOMEM);
1941 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1943 GOTO(out, rc = -ENOMEM);
1947 GOTO(out, rc = -ENOMEM);
1950 list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1951 struct cl_page *page = oap2cl_page(oap);
1952 if (clerq == NULL) {
1953 clerq = cl_req_alloc(env, page, crt,
1954 1 /* only 1-object rpcs for now */);
1956 GOTO(out, rc = PTR_ERR(clerq));
1957 lock = oap->oap_ldlm_lock;
1960 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1961 pga[i] = &oap->oap_brw_page;
1962 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1963 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1964 pga[i]->pg, page_index(oap->oap_page), oap,
1967 cl_req_page_add(env, clerq, page);
1970 /* always get the data for the obdo for the rpc */
1971 LASSERT(clerq != NULL);
1972 crattr->cra_oa = oa;
1973 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1975 oa->o_handle = lock->l_remote_handle;
1976 oa->o_valid |= OBD_MD_FLHANDLE;
1979 rc = cl_req_prep(env, clerq);
1981 CERROR("cl_req_prep failed: %d\n", rc);
1985 sort_brw_pages(pga, page_count);
1986 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1987 pga, &req, crattr->cra_capa, 1, 0);
1989 CERROR("prep_req failed: %d\n", rc);
1993 req->rq_commit_cb = brw_commit;
1994 req->rq_interpret_reply = brw_interpret;
1997 req->rq_memalloc = 1;
1999 /* Need to update the timestamps after the request is built in case
2000 * we race with setattr (locally or in queue at OST). If OST gets
2001 * later setattr before earlier BRW (as determined by the request xid),
2002 * the OST will not use BRW timestamps. Sadly, there is no obvious
2003 * way to do this in a single call. bug 10150 */
2004 cl_req_attr_set(env, clerq, crattr,
2005 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2007 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2009 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2010 aa = ptlrpc_req_async_args(req);
2011 INIT_LIST_HEAD(&aa->aa_oaps);
2012 list_splice_init(&rpc_list, &aa->aa_oaps);
2013 INIT_LIST_HEAD(&aa->aa_exts);
2014 list_splice_init(ext_list, &aa->aa_exts);
2015 aa->aa_clerq = clerq;
2017 /* queued sync pages can be torn down while the pages
2018 * were between the pending list and the rpc */
2020 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2021 /* only one oap gets a request reference */
2024 if (oap->oap_interrupted && !req->rq_intr) {
2025 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2027 ptlrpc_mark_interrupted(req);
2031 tmp->oap_request = ptlrpc_request_addref(req);
2033 client_obd_list_lock(&cli->cl_loi_list_lock);
2034 starting_offset >>= PAGE_CACHE_SHIFT;
2035 if (cmd == OBD_BRW_READ) {
2036 cli->cl_r_in_flight++;
2037 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2038 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2039 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2040 starting_offset + 1);
2042 cli->cl_w_in_flight++;
2043 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2044 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2045 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2046 starting_offset + 1);
2048 client_obd_list_unlock(&cli->cl_loi_list_lock);
2050 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2051 page_count, aa, cli->cl_r_in_flight,
2052 cli->cl_w_in_flight);
2054 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2055 * see which CPU/NUMA node the majority of pages were allocated
2056 * on, and try to assign the async RPC to the CPU core
2057 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2059 * But on the other hand, we expect that multiple ptlrpcd
2060 * threads and the initial write sponsor can run in parallel,
2061 * especially when data checksum is enabled, which is CPU-bound
2062 * operation and single ptlrpcd thread cannot process in time.
2063 * So more ptlrpcd threads sharing BRW load
2064 * (with PDL_POLICY_ROUND) seems better.
2066 ptlrpcd_add_req(req, pol, -1);
2072 cfs_memory_pressure_restore(mpflag);
2074 if (crattr != NULL) {
2075 capa_put(crattr->cra_capa);
2076 OBD_FREE(crattr, sizeof(*crattr));
2080 LASSERT(req == NULL);
2085 OBD_FREE(pga, sizeof(*pga) * page_count);
2086 /* this should happen rarely and is pretty bad, it makes the
2087 * pending list not follow the dirty order */
2088 while (!list_empty(ext_list)) {
2089 ext = list_entry(ext_list->next, struct osc_extent,
2091 list_del_init(&ext->oe_link);
2092 osc_extent_finish(env, ext, 0, rc);
2094 if (clerq && !IS_ERR(clerq))
2095 cl_req_completion(env, clerq, rc);
2100 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2101 struct ldlm_enqueue_info *einfo)
2103 void *data = einfo->ei_cbdata;
2106 LASSERT(lock != NULL);
2107 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2108 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2109 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2110 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2112 lock_res_and_lock(lock);
2113 spin_lock(&osc_ast_guard);
2115 if (lock->l_ast_data == NULL)
2116 lock->l_ast_data = data;
2117 if (lock->l_ast_data == data)
2120 spin_unlock(&osc_ast_guard);
2121 unlock_res_and_lock(lock);
2126 static int osc_set_data_with_check(struct lustre_handle *lockh,
2127 struct ldlm_enqueue_info *einfo)
2129 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2133 set = osc_set_lock_data_with_check(lock, einfo);
2134 LDLM_LOCK_PUT(lock);
2136 CERROR("lockh %p, data %p - client evicted?\n",
2137 lockh, einfo->ei_cbdata);
2141 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2142 ldlm_iterator_t replace, void *data)
2144 struct ldlm_res_id res_id;
2145 struct obd_device *obd = class_exp2obd(exp);
2147 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2148 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2152 /* find any ldlm lock of the inode in osc
2156 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2157 ldlm_iterator_t replace, void *data)
2159 struct ldlm_res_id res_id;
2160 struct obd_device *obd = class_exp2obd(exp);
2163 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2164 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2165 if (rc == LDLM_ITER_STOP)
2167 if (rc == LDLM_ITER_CONTINUE)
2172 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2173 obd_enqueue_update_f upcall, void *cookie,
2174 __u64 *flags, int agl, int rc)
2176 int intent = *flags & LDLM_FL_HAS_INTENT;
2180 /* The request was created before ldlm_cli_enqueue call. */
2181 if (rc == ELDLM_LOCK_ABORTED) {
2182 struct ldlm_reply *rep;
2183 rep = req_capsule_server_get(&req->rq_pill,
2186 LASSERT(rep != NULL);
2187 rep->lock_policy_res1 =
2188 ptlrpc_status_ntoh(rep->lock_policy_res1);
2189 if (rep->lock_policy_res1)
2190 rc = rep->lock_policy_res1;
2194 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2196 *flags |= LDLM_FL_LVB_READY;
2197 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2198 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2201 /* Call the update callback. */
2202 rc = (*upcall)(cookie, rc);
2206 static int osc_enqueue_interpret(const struct lu_env *env,
2207 struct ptlrpc_request *req,
2208 struct osc_enqueue_args *aa, int rc)
2210 struct ldlm_lock *lock;
2211 struct lustre_handle handle;
2213 struct ost_lvb *lvb;
2215 __u64 *flags = aa->oa_flags;
2217 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2218 * might be freed anytime after lock upcall has been called. */
2219 lustre_handle_copy(&handle, aa->oa_lockh);
2220 mode = aa->oa_ei->ei_mode;
2222 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2224 lock = ldlm_handle2lock(&handle);
2226 /* Take an additional reference so that a blocking AST that
2227 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2228 * to arrive after an upcall has been executed by
2229 * osc_enqueue_fini(). */
2230 ldlm_lock_addref(&handle, mode);
2232 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2233 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2235 /* Let CP AST to grant the lock first. */
2236 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2238 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2243 lvb_len = sizeof(*aa->oa_lvb);
2246 /* Complete obtaining the lock procedure. */
2247 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2248 mode, flags, lvb, lvb_len, &handle, rc);
2249 /* Complete osc stuff. */
2250 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2251 flags, aa->oa_agl, rc);
2253 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2255 /* Release the lock for async request. */
2256 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2258 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2259 * not already released by
2260 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2262 ldlm_lock_decref(&handle, mode);
2264 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2265 aa->oa_lockh, req, aa);
2266 ldlm_lock_decref(&handle, mode);
2267 LDLM_LOCK_PUT(lock);
2271 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2273 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2274 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2275 * other synchronous requests, however keeping some locks and trying to obtain
2276 * others may take a considerable amount of time in a case of ost failure; and
2277 * when other sync requests do not get released lock from a client, the client
2278 * is excluded from the cluster -- such scenarious make the life difficult, so
2279 * release locks just after they are obtained. */
2280 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2281 __u64 *flags, ldlm_policy_data_t *policy,
2282 struct ost_lvb *lvb, int kms_valid,
2283 obd_enqueue_update_f upcall, void *cookie,
2284 struct ldlm_enqueue_info *einfo,
2285 struct lustre_handle *lockh,
2286 struct ptlrpc_request_set *rqset, int async, int agl)
2288 struct obd_device *obd = exp->exp_obd;
2289 struct ptlrpc_request *req = NULL;
2290 int intent = *flags & LDLM_FL_HAS_INTENT;
2291 __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2296 /* Filesystem lock extents are extended to page boundaries so that
2297 * dealing with the page cache is a little smoother. */
2298 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2299 policy->l_extent.end |= ~CFS_PAGE_MASK;
2302 * kms is not valid when either object is completely fresh (so that no
2303 * locks are cached), or object was evicted. In the latter case cached
2304 * lock cannot be used, because it would prime inode state with
2305 * potentially stale LVB.
2310 /* Next, search for already existing extent locks that will cover us */
2311 /* If we're trying to read, we also search for an existing PW lock. The
2312 * VFS and page cache already protect us locally, so lots of readers/
2313 * writers can share a single PW lock.
2315 * There are problems with conversion deadlocks, so instead of
2316 * converting a read lock to a write lock, we'll just enqueue a new
2319 * At some point we should cancel the read lock instead of making them
2320 * send us a blocking callback, but there are problems with canceling
2321 * locks out from other users right now, too. */
2322 mode = einfo->ei_mode;
2323 if (einfo->ei_mode == LCK_PR)
2325 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2326 einfo->ei_type, policy, mode, lockh, 0);
2328 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2330 if ((agl != 0) && !ldlm_is_lvb_ready(matched)) {
2331 /* For AGL, if enqueue RPC is sent but the lock is not
2332 * granted, then skip to process this strpe.
2333 * Return -ECANCELED to tell the caller. */
2334 ldlm_lock_decref(lockh, mode);
2335 LDLM_LOCK_PUT(matched);
2337 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2338 *flags |= LDLM_FL_LVB_READY;
2339 /* addref the lock only if not async requests and PW
2340 * lock is matched whereas we asked for PR. */
2341 if (!rqset && einfo->ei_mode != mode)
2342 ldlm_lock_addref(lockh, LCK_PR);
2344 /* I would like to be able to ASSERT here that
2345 * rss <= kms, but I can't, for reasons which
2346 * are explained in lov_enqueue() */
2349 /* We already have a lock, and it's referenced.
2351 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2352 * AGL upcall may change it to CLS_HELD directly. */
2353 (*upcall)(cookie, ELDLM_OK);
2355 if (einfo->ei_mode != mode)
2356 ldlm_lock_decref(lockh, LCK_PW);
2358 /* For async requests, decref the lock. */
2359 ldlm_lock_decref(lockh, einfo->ei_mode);
2360 LDLM_LOCK_PUT(matched);
2363 ldlm_lock_decref(lockh, mode);
2364 LDLM_LOCK_PUT(matched);
2370 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2371 &RQF_LDLM_ENQUEUE_LVB);
2375 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2377 ptlrpc_request_free(req);
2381 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2383 ptlrpc_request_set_replen(req);
2386 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2387 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2389 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2390 sizeof(*lvb), LVB_T_OST, lockh, async);
2393 struct osc_enqueue_args *aa;
2394 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2395 aa = ptlrpc_req_async_args(req);
2398 aa->oa_flags = flags;
2399 aa->oa_upcall = upcall;
2400 aa->oa_cookie = cookie;
2402 aa->oa_lockh = lockh;
2405 req->rq_interpret_reply =
2406 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2407 if (rqset == PTLRPCD_SET)
2408 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2410 ptlrpc_set_add_req(rqset, req);
2411 } else if (intent) {
2412 ptlrpc_req_finished(req);
2417 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2419 ptlrpc_req_finished(req);
2424 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2425 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2426 __u64 *flags, void *data, struct lustre_handle *lockh,
2429 struct obd_device *obd = exp->exp_obd;
2430 __u64 lflags = *flags;
2434 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2437 /* Filesystem lock extents are extended to page boundaries so that
2438 * dealing with the page cache is a little smoother */
2439 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2440 policy->l_extent.end |= ~CFS_PAGE_MASK;
2442 /* Next, search for already existing extent locks that will cover us */
2443 /* If we're trying to read, we also search for an existing PW lock. The
2444 * VFS and page cache already protect us locally, so lots of readers/
2445 * writers can share a single PW lock. */
2449 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2450 res_id, type, policy, rc, lockh, unref);
2453 if (!osc_set_data_with_check(lockh, data)) {
2454 if (!(lflags & LDLM_FL_TEST_LOCK))
2455 ldlm_lock_decref(lockh, rc);
2459 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2460 ldlm_lock_addref(lockh, LCK_PR);
2461 ldlm_lock_decref(lockh, LCK_PW);
2468 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2472 if (unlikely(mode == LCK_GROUP))
2473 ldlm_lock_decref_and_cancel(lockh, mode);
2475 ldlm_lock_decref(lockh, mode);
2480 static int osc_statfs_interpret(const struct lu_env *env,
2481 struct ptlrpc_request *req,
2482 struct osc_async_args *aa, int rc)
2484 struct obd_statfs *msfs;
2488 /* The request has in fact never been sent
2489 * due to issues at a higher level (LOV).
2490 * Exit immediately since the caller is
2491 * aware of the problem and takes care
2492 * of the clean up */
2495 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2496 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2502 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2504 GOTO(out, rc = -EPROTO);
2507 *aa->aa_oi->oi_osfs = *msfs;
2509 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2513 static int osc_statfs_async(struct obd_export *exp,
2514 struct obd_info *oinfo, __u64 max_age,
2515 struct ptlrpc_request_set *rqset)
2517 struct obd_device *obd = class_exp2obd(exp);
2518 struct ptlrpc_request *req;
2519 struct osc_async_args *aa;
2523 /* We could possibly pass max_age in the request (as an absolute
2524 * timestamp or a "seconds.usec ago") so the target can avoid doing
2525 * extra calls into the filesystem if that isn't necessary (e.g.
2526 * during mount that would help a bit). Having relative timestamps
2527 * is not so great if request processing is slow, while absolute
2528 * timestamps are not ideal because they need time synchronization. */
2529 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2533 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2535 ptlrpc_request_free(req);
2538 ptlrpc_request_set_replen(req);
2539 req->rq_request_portal = OST_CREATE_PORTAL;
2540 ptlrpc_at_set_req_timeout(req);
2542 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2543 /* procfs requests not want stat in wait for avoid deadlock */
2544 req->rq_no_resend = 1;
2545 req->rq_no_delay = 1;
2548 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2549 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2550 aa = ptlrpc_req_async_args(req);
2553 ptlrpc_set_add_req(rqset, req);
2557 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2558 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2560 struct obd_device *obd = class_exp2obd(exp);
2561 struct obd_statfs *msfs;
2562 struct ptlrpc_request *req;
2563 struct obd_import *imp = NULL;
2567 /*Since the request might also come from lprocfs, so we need
2568 *sync this with client_disconnect_export Bug15684*/
2569 down_read(&obd->u.cli.cl_sem);
2570 if (obd->u.cli.cl_import)
2571 imp = class_import_get(obd->u.cli.cl_import);
2572 up_read(&obd->u.cli.cl_sem);
2576 /* We could possibly pass max_age in the request (as an absolute
2577 * timestamp or a "seconds.usec ago") so the target can avoid doing
2578 * extra calls into the filesystem if that isn't necessary (e.g.
2579 * during mount that would help a bit). Having relative timestamps
2580 * is not so great if request processing is slow, while absolute
2581 * timestamps are not ideal because they need time synchronization. */
2582 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2584 class_import_put(imp);
2589 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2591 ptlrpc_request_free(req);
2594 ptlrpc_request_set_replen(req);
2595 req->rq_request_portal = OST_CREATE_PORTAL;
2596 ptlrpc_at_set_req_timeout(req);
2598 if (flags & OBD_STATFS_NODELAY) {
2599 /* procfs requests not want stat in wait for avoid deadlock */
2600 req->rq_no_resend = 1;
2601 req->rq_no_delay = 1;
2604 rc = ptlrpc_queue_wait(req);
2608 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2610 GOTO(out, rc = -EPROTO);
2617 ptlrpc_req_finished(req);
2621 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2622 void *karg, void *uarg)
2624 struct obd_device *obd = exp->exp_obd;
2625 struct obd_ioctl_data *data = karg;
2629 if (!try_module_get(THIS_MODULE)) {
2630 CERROR("Can't get module. Is it alive?");
2634 case OBD_IOC_CLIENT_RECOVER:
2635 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2636 data->ioc_inlbuf1, 0);
2640 case IOC_OSC_SET_ACTIVE:
2641 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2644 case OBD_IOC_POLL_QUOTACHECK:
2645 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2647 case OBD_IOC_PING_TARGET:
2648 err = ptlrpc_obd_ping(obd);
2651 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2652 cmd, current_comm());
2653 GOTO(out, err = -ENOTTY);
2656 module_put(THIS_MODULE);
2660 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2661 obd_count keylen, void *key, __u32 *vallen, void *val,
2662 struct lov_stripe_md *lsm)
2665 if (!vallen || !val)
2668 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2669 __u32 *stripe = val;
2670 *vallen = sizeof(*stripe);
2673 } else if (KEY_IS(KEY_LAST_ID)) {
2674 struct ptlrpc_request *req;
2679 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2680 &RQF_OST_GET_INFO_LAST_ID);
2684 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2685 RCL_CLIENT, keylen);
2686 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2688 ptlrpc_request_free(req);
2692 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2693 memcpy(tmp, key, keylen);
2695 req->rq_no_delay = req->rq_no_resend = 1;
2696 ptlrpc_request_set_replen(req);
2697 rc = ptlrpc_queue_wait(req);
2701 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2703 GOTO(out, rc = -EPROTO);
2705 *((obd_id *)val) = *reply;
2707 ptlrpc_req_finished(req);
2709 } else if (KEY_IS(KEY_FIEMAP)) {
2710 struct ll_fiemap_info_key *fm_key =
2711 (struct ll_fiemap_info_key *)key;
2712 struct ldlm_res_id res_id;
2713 ldlm_policy_data_t policy;
2714 struct lustre_handle lockh;
2715 ldlm_mode_t mode = 0;
2716 struct ptlrpc_request *req;
2717 struct ll_user_fiemap *reply;
2721 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2724 policy.l_extent.start = fm_key->fiemap.fm_start &
2727 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2728 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
2729 policy.l_extent.end = OBD_OBJECT_EOF;
2731 policy.l_extent.end = (fm_key->fiemap.fm_start +
2732 fm_key->fiemap.fm_length +
2733 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
2735 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2736 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2737 LDLM_FL_BLOCK_GRANTED |
2739 &res_id, LDLM_EXTENT, &policy,
2740 LCK_PR | LCK_PW, &lockh, 0);
2741 if (mode) { /* lock is cached on client */
2742 if (mode != LCK_PR) {
2743 ldlm_lock_addref(&lockh, LCK_PR);
2744 ldlm_lock_decref(&lockh, LCK_PW);
2746 } else { /* no cached lock, needs acquire lock on server side */
2747 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2748 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2752 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2753 &RQF_OST_GET_INFO_FIEMAP);
2755 GOTO(drop_lock, rc = -ENOMEM);
2757 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2758 RCL_CLIENT, keylen);
2759 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2760 RCL_CLIENT, *vallen);
2761 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2762 RCL_SERVER, *vallen);
2764 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2766 ptlrpc_request_free(req);
2767 GOTO(drop_lock, rc);
2770 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2771 memcpy(tmp, key, keylen);
2772 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2773 memcpy(tmp, val, *vallen);
2775 ptlrpc_request_set_replen(req);
2776 rc = ptlrpc_queue_wait(req);
2780 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2782 GOTO(fini_req, rc = -EPROTO);
2784 memcpy(val, reply, *vallen);
2786 ptlrpc_req_finished(req);
2789 ldlm_lock_decref(&lockh, LCK_PR);
2796 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2797 obd_count keylen, void *key, obd_count vallen,
2798 void *val, struct ptlrpc_request_set *set)
2800 struct ptlrpc_request *req;
2801 struct obd_device *obd = exp->exp_obd;
2802 struct obd_import *imp = class_exp2cliimp(exp);
2807 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2809 if (KEY_IS(KEY_CHECKSUM)) {
2810 if (vallen != sizeof(int))
2812 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2816 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2817 sptlrpc_conf_client_adapt(obd);
2821 if (KEY_IS(KEY_FLUSH_CTX)) {
2822 sptlrpc_import_flush_my_ctx(imp);
2826 if (KEY_IS(KEY_CACHE_SET)) {
2827 struct client_obd *cli = &obd->u.cli;
2829 LASSERT(cli->cl_cache == NULL); /* only once */
2830 cli->cl_cache = (struct cl_client_cache *)val;
2831 atomic_inc(&cli->cl_cache->ccc_users);
2832 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2834 /* add this osc into entity list */
2835 LASSERT(list_empty(&cli->cl_lru_osc));
2836 spin_lock(&cli->cl_cache->ccc_lru_lock);
2837 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2838 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2843 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2844 struct client_obd *cli = &obd->u.cli;
2845 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2846 int target = *(int *)val;
2848 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2853 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2856 /* We pass all other commands directly to OST. Since nobody calls osc
2857 methods directly and everybody is supposed to go through LOV, we
2858 assume lov checked invalid values for us.
2859 The only recognised values so far are evict_by_nid and mds_conn.
2860 Even if something bad goes through, we'd get a -EINVAL from OST
2863 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2864 &RQF_OST_SET_GRANT_INFO :
2869 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2870 RCL_CLIENT, keylen);
2871 if (!KEY_IS(KEY_GRANT_SHRINK))
2872 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2873 RCL_CLIENT, vallen);
2874 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2876 ptlrpc_request_free(req);
2880 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2881 memcpy(tmp, key, keylen);
2882 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2885 memcpy(tmp, val, vallen);
2887 if (KEY_IS(KEY_GRANT_SHRINK)) {
2888 struct osc_grant_args *aa;
2891 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2892 aa = ptlrpc_req_async_args(req);
2895 ptlrpc_req_finished(req);
2898 *oa = ((struct ost_body *)val)->oa;
2900 req->rq_interpret_reply = osc_shrink_grant_interpret;
2903 ptlrpc_request_set_replen(req);
2904 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2905 LASSERT(set != NULL);
2906 ptlrpc_set_add_req(set, req);
2907 ptlrpc_check_set(NULL, set);
2909 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2914 static int osc_reconnect(const struct lu_env *env,
2915 struct obd_export *exp, struct obd_device *obd,
2916 struct obd_uuid *cluuid,
2917 struct obd_connect_data *data,
2920 struct client_obd *cli = &obd->u.cli;
2922 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2925 client_obd_list_lock(&cli->cl_loi_list_lock);
2926 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2927 2 * cli_brw_size(obd);
2928 lost_grant = cli->cl_lost_grant;
2929 cli->cl_lost_grant = 0;
2930 client_obd_list_unlock(&cli->cl_loi_list_lock);
2932 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
2933 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2934 data->ocd_version, data->ocd_grant, lost_grant);
2940 static int osc_disconnect(struct obd_export *exp)
2942 struct obd_device *obd = class_exp2obd(exp);
2943 struct llog_ctxt *ctxt;
2946 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2948 if (obd->u.cli.cl_conn_count == 1) {
2949 /* Flush any remaining cancel messages out to the
2951 llog_sync(ctxt, exp, 0);
2953 llog_ctxt_put(ctxt);
2955 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
2959 rc = client_disconnect_export(exp);
2961 * Initially we put del_shrink_grant before disconnect_export, but it
2962 * causes the following problem if setup (connect) and cleanup
2963 * (disconnect) are tangled together.
2964 * connect p1 disconnect p2
2965 * ptlrpc_connect_import
2966 * ............... class_manual_cleanup
2969 * ptlrpc_connect_interrupt
2971 * add this client to shrink list
2973 * Bang! pinger trigger the shrink.
2974 * So the osc should be disconnected from the shrink list, after we
2975 * are sure the import has been destroyed. BUG18662
2977 if (obd->u.cli.cl_import == NULL)
2978 osc_del_shrink_grant(&obd->u.cli);
2982 static int osc_import_event(struct obd_device *obd,
2983 struct obd_import *imp,
2984 enum obd_import_event event)
2986 struct client_obd *cli;
2990 LASSERT(imp->imp_obd == obd);
2993 case IMP_EVENT_DISCON: {
2995 client_obd_list_lock(&cli->cl_loi_list_lock);
2996 cli->cl_avail_grant = 0;
2997 cli->cl_lost_grant = 0;
2998 client_obd_list_unlock(&cli->cl_loi_list_lock);
3001 case IMP_EVENT_INACTIVE: {
3002 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3005 case IMP_EVENT_INVALIDATE: {
3006 struct ldlm_namespace *ns = obd->obd_namespace;
3010 env = cl_env_get(&refcheck);
3014 /* all pages go to failing rpcs due to the invalid
3016 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3018 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3019 cl_env_put(env, &refcheck);
3024 case IMP_EVENT_ACTIVE: {
3025 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3028 case IMP_EVENT_OCD: {
3029 struct obd_connect_data *ocd = &imp->imp_connect_data;
3031 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3032 osc_init_grant(&obd->u.cli, ocd);
3035 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3036 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3038 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3041 case IMP_EVENT_DEACTIVATE: {
3042 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3045 case IMP_EVENT_ACTIVATE: {
3046 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3050 CERROR("Unknown import event %d\n", event);
3057 * Determine whether the lock can be canceled before replaying the lock
3058 * during recovery, see bug16774 for detailed information.
3060 * \retval zero the lock can't be canceled
3061 * \retval other ok to cancel
3063 static int osc_cancel_weight(struct ldlm_lock *lock)
3066 * Cancel all unused and granted extent lock.
3068 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3069 lock->l_granted_mode == lock->l_req_mode &&
3070 osc_ldlm_weigh_ast(lock) == 0)
3076 static int brw_queue_work(const struct lu_env *env, void *data)
3078 struct client_obd *cli = data;
3080 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3082 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3086 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3088 struct client_obd *cli = &obd->u.cli;
3089 struct obd_type *type;
3094 rc = ptlrpcd_addref();
3098 rc = client_obd_setup(obd, lcfg);
3100 GOTO(out_ptlrpcd, rc);
3102 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3103 if (IS_ERR(handler))
3104 GOTO(out_client_setup, rc = PTR_ERR(handler));
3105 cli->cl_writeback_work = handler;
3107 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3108 if (IS_ERR(handler))
3109 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3110 cli->cl_lru_work = handler;
3112 rc = osc_quota_setup(obd);
3114 GOTO(out_ptlrpcd_work, rc);
3116 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3119 obd->obd_vars = lprocfs_osc_obd_vars;
3121 /* If this is true then both client (osc) and server (osp) are on the
3122 * same node. The osp layer if loaded first will register the osc proc
3123 * directory. In that case this obd_device will be attached its proc
3124 * tree to type->typ_procsym instead of obd->obd_type->typ_procroot. */
3125 type = class_search_type(LUSTRE_OSP_NAME);
3126 if (type && type->typ_procsym) {
3127 obd->obd_proc_entry = lprocfs_seq_register(obd->obd_name,
3129 obd->obd_vars, obd);
3130 if (IS_ERR(obd->obd_proc_entry)) {
3131 rc = PTR_ERR(obd->obd_proc_entry);
3132 CERROR("error %d setting up lprocfs for %s\n", rc,
3134 obd->obd_proc_entry = NULL;
3137 rc = lprocfs_seq_obd_setup(obd);
3140 /* If the basic OSC proc tree construction succeeded then
3141 * lets do the rest. */
3143 lproc_osc_attach_seqstat(obd);
3144 sptlrpc_lprocfs_cliobd_attach(obd);
3145 ptlrpc_lprocfs_register_obd(obd);
3148 /* We need to allocate a few requests more, because
3149 * brw_interpret tries to create new requests before freeing
3150 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3151 * reserved, but I'm afraid that might be too much wasted RAM
3152 * in fact, so 2 is just my guess and still should work. */
3153 cli->cl_import->imp_rq_pool =
3154 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3156 ptlrpc_add_rqs_to_pool);
3158 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3159 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3163 if (cli->cl_writeback_work != NULL) {
3164 ptlrpcd_destroy_work(cli->cl_writeback_work);
3165 cli->cl_writeback_work = NULL;
3167 if (cli->cl_lru_work != NULL) {
3168 ptlrpcd_destroy_work(cli->cl_lru_work);
3169 cli->cl_lru_work = NULL;
3172 client_obd_cleanup(obd);
3178 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3184 case OBD_CLEANUP_EARLY: {
3185 struct obd_import *imp;
3186 imp = obd->u.cli.cl_import;
3187 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3188 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3189 ptlrpc_deactivate_import(imp);
3190 spin_lock(&imp->imp_lock);
3191 imp->imp_pingable = 0;
3192 spin_unlock(&imp->imp_lock);
3195 case OBD_CLEANUP_EXPORTS: {
3196 struct client_obd *cli = &obd->u.cli;
3198 * for echo client, export may be on zombie list, wait for
3199 * zombie thread to cull it, because cli.cl_import will be
3200 * cleared in client_disconnect_export():
3201 * class_export_destroy() -> obd_cleanup() ->
3202 * echo_device_free() -> echo_client_cleanup() ->
3203 * obd_disconnect() -> osc_disconnect() ->
3204 * client_disconnect_export()
3206 obd_zombie_barrier();
3207 if (cli->cl_writeback_work) {
3208 ptlrpcd_destroy_work(cli->cl_writeback_work);
3209 cli->cl_writeback_work = NULL;
3211 if (cli->cl_lru_work) {
3212 ptlrpcd_destroy_work(cli->cl_lru_work);
3213 cli->cl_lru_work = NULL;
3215 obd_cleanup_client_import(obd);
3216 ptlrpc_lprocfs_unregister_obd(obd);
3217 lprocfs_obd_cleanup(obd);
3218 rc = obd_llog_finish(obd, 0);
3220 CERROR("failed to cleanup llogging subsystems\n");
3227 int osc_cleanup(struct obd_device *obd)
3229 struct client_obd *cli = &obd->u.cli;
3235 if (cli->cl_cache != NULL) {
3236 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3237 spin_lock(&cli->cl_cache->ccc_lru_lock);
3238 list_del_init(&cli->cl_lru_osc);
3239 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3240 cli->cl_lru_left = NULL;
3241 atomic_dec(&cli->cl_cache->ccc_users);
3242 cli->cl_cache = NULL;
3245 /* free memory of osc quota cache */
3246 osc_quota_cleanup(obd);
3248 rc = client_obd_cleanup(obd);
3254 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3256 int rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
3258 return rc > 0 ? 0: rc;
3261 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3263 return osc_process_config_base(obd, buf);
3266 struct obd_ops osc_obd_ops = {
3267 .o_owner = THIS_MODULE,
3268 .o_setup = osc_setup,
3269 .o_precleanup = osc_precleanup,
3270 .o_cleanup = osc_cleanup,
3271 .o_add_conn = client_import_add_conn,
3272 .o_del_conn = client_import_del_conn,
3273 .o_connect = client_connect_import,
3274 .o_reconnect = osc_reconnect,
3275 .o_disconnect = osc_disconnect,
3276 .o_statfs = osc_statfs,
3277 .o_statfs_async = osc_statfs_async,
3278 .o_unpackmd = osc_unpackmd,
3279 .o_create = osc_create,
3280 .o_destroy = osc_destroy,
3281 .o_getattr = osc_getattr,
3282 .o_getattr_async = osc_getattr_async,
3283 .o_setattr = osc_setattr,
3284 .o_setattr_async = osc_setattr_async,
3285 .o_change_cbdata = osc_change_cbdata,
3286 .o_find_cbdata = osc_find_cbdata,
3287 .o_iocontrol = osc_iocontrol,
3288 .o_get_info = osc_get_info,
3289 .o_set_info_async = osc_set_info_async,
3290 .o_import_event = osc_import_event,
3291 .o_process_config = osc_process_config,
3292 .o_quotactl = osc_quotactl,
3293 .o_quotacheck = osc_quotacheck,
3296 extern struct lu_kmem_descr osc_caches[];
3297 extern spinlock_t osc_ast_guard;
3298 extern struct lock_class_key osc_ast_guard_class;
3300 int __init osc_init(void)
3302 bool enable_proc = true;
3303 struct obd_type *type;
3307 /* print an address of _any_ initialized kernel symbol from this
3308 * module, to allow debugging with gdb that doesn't support data
3309 * symbols from modules.*/
3310 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3312 rc = lu_kmem_init(osc_caches);
3316 type = class_search_type(LUSTRE_OSP_NAME);
3317 if (type != NULL && type->typ_procsym != NULL)
3318 enable_proc = false;
3320 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
3321 #ifndef HAVE_ONLY_PROCFS_SEQ
3324 LUSTRE_OSC_NAME, &osc_device_type);
3326 lu_kmem_fini(osc_caches);
3330 spin_lock_init(&osc_ast_guard);
3331 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3337 static void /*__exit*/ osc_exit(void)
3339 class_unregister_type(LUSTRE_OSC_NAME);
3340 lu_kmem_fini(osc_caches);
3343 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3344 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3345 MODULE_LICENSE("GPL");
3347 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);