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>
56 #include <lustre_ha.h>
57 #include <lprocfs_status.h>
58 #include <lustre_log.h>
59 #include <lustre_debug.h>
60 #include <lustre_param.h>
61 #include <lustre_fid.h>
62 #include "osc_internal.h"
63 #include "osc_cl_internal.h"
65 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
66 static int brw_interpret(const struct lu_env *env,
67 struct ptlrpc_request *req, void *data, int rc);
68 int osc_cleanup(struct obd_device *obd);
70 /* Pack OSC object metadata for disk storage (LE byte order). */
71 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
72 struct lov_stripe_md *lsm)
77 lmm_size = sizeof(**lmmp);
81 if (*lmmp != NULL && lsm == NULL) {
82 OBD_FREE(*lmmp, lmm_size);
85 } else if (unlikely(lsm != NULL && ostid_id(&lsm->lsm_oi) == 0)) {
90 OBD_ALLOC(*lmmp, lmm_size);
96 ostid_cpu_to_le(&lsm->lsm_oi, &(*lmmp)->lmm_oi);
101 /* Unpack OSC object metadata from disk storage (LE byte order). */
102 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
103 struct lov_mds_md *lmm, int lmm_bytes)
106 struct obd_import *imp = class_exp2cliimp(exp);
110 if (lmm_bytes < sizeof(*lmm)) {
111 CERROR("%s: lov_mds_md too small: %d, need %d\n",
112 exp->exp_obd->obd_name, lmm_bytes,
116 /* XXX LOV_MAGIC etc check? */
118 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
119 CERROR("%s: zero lmm_object_id: rc = %d\n",
120 exp->exp_obd->obd_name, -EINVAL);
125 lsm_size = lov_stripe_md_size(1);
129 if (*lsmp != NULL && lmm == NULL) {
130 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
131 OBD_FREE(*lsmp, lsm_size);
137 OBD_ALLOC(*lsmp, lsm_size);
138 if (unlikely(*lsmp == NULL))
140 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
141 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
142 OBD_FREE(*lsmp, lsm_size);
145 loi_init((*lsmp)->lsm_oinfo[0]);
146 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
151 /* XXX zero *lsmp? */
152 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
155 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
156 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
158 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
163 static inline void osc_pack_capa(struct ptlrpc_request *req,
164 struct ost_body *body, void *capa)
166 struct obd_capa *oc = (struct obd_capa *)capa;
167 struct lustre_capa *c;
172 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
175 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
176 DEBUG_CAPA(D_SEC, c, "pack");
179 static inline void osc_pack_req_body(struct ptlrpc_request *req,
180 struct obd_info *oinfo)
182 struct ost_body *body;
184 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
187 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(const struct lu_env *env,
204 struct ptlrpc_request *req,
205 struct osc_async_args *aa, int rc)
207 struct ost_body *body;
213 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
215 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
217 aa->aa_oi->oi_oa, &body->oa);
219 /* This should really be sent by the OST */
220 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
221 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
223 CDEBUG(D_INFO, "can't unpack ost_body\n");
225 aa->aa_oi->oi_oa->o_valid = 0;
228 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
232 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
233 struct ptlrpc_request_set *set)
235 struct ptlrpc_request *req;
236 struct osc_async_args *aa;
240 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
244 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
245 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
247 ptlrpc_request_free(req);
251 osc_pack_req_body(req, oinfo);
253 ptlrpc_request_set_replen(req);
254 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
256 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
257 aa = ptlrpc_req_async_args(req);
260 ptlrpc_set_add_req(set, req);
264 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
265 struct obd_info *oinfo)
267 struct ptlrpc_request *req;
268 struct ost_body *body;
272 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
276 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
277 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
279 ptlrpc_request_free(req);
283 osc_pack_req_body(req, oinfo);
285 ptlrpc_request_set_replen(req);
287 rc = ptlrpc_queue_wait(req);
291 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
293 GOTO(out, rc = -EPROTO);
295 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
296 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
299 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
300 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
304 ptlrpc_req_finished(req);
308 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
309 struct obd_info *oinfo, struct obd_trans_info *oti)
311 struct ptlrpc_request *req;
312 struct ost_body *body;
316 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
318 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
322 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
323 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
325 ptlrpc_request_free(req);
329 osc_pack_req_body(req, oinfo);
331 ptlrpc_request_set_replen(req);
333 rc = ptlrpc_queue_wait(req);
337 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
339 GOTO(out, rc = -EPROTO);
341 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
346 ptlrpc_req_finished(req);
350 static int osc_setattr_interpret(const struct lu_env *env,
351 struct ptlrpc_request *req,
352 struct osc_setattr_args *sa, int rc)
354 struct ost_body *body;
360 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
362 GOTO(out, rc = -EPROTO);
364 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
367 rc = sa->sa_upcall(sa->sa_cookie, rc);
371 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
372 struct obd_trans_info *oti,
373 obd_enqueue_update_f upcall, void *cookie,
374 struct ptlrpc_request_set *rqset)
376 struct ptlrpc_request *req;
377 struct osc_setattr_args *sa;
381 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
385 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
386 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
388 ptlrpc_request_free(req);
392 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
393 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
395 osc_pack_req_body(req, oinfo);
397 ptlrpc_request_set_replen(req);
399 /* do mds to ost setattr asynchronously */
401 /* Do not wait for response. */
402 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
404 req->rq_interpret_reply =
405 (ptlrpc_interpterer_t)osc_setattr_interpret;
407 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
408 sa = ptlrpc_req_async_args(req);
409 sa->sa_oa = oinfo->oi_oa;
410 sa->sa_upcall = upcall;
411 sa->sa_cookie = cookie;
413 if (rqset == PTLRPCD_SET)
414 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
416 ptlrpc_set_add_req(rqset, req);
422 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
423 struct obd_trans_info *oti,
424 struct ptlrpc_request_set *rqset)
426 return osc_setattr_async_base(exp, oinfo, oti,
427 oinfo->oi_cb_up, oinfo, rqset);
430 int osc_real_create(struct obd_export *exp, struct obdo *oa,
431 struct lov_stripe_md **ea, struct obd_trans_info *oti)
433 struct ptlrpc_request *req;
434 struct ost_body *body;
435 struct lov_stripe_md *lsm;
444 rc = obd_alloc_memmd(exp, &lsm);
449 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
451 GOTO(out, rc = -ENOMEM);
453 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
455 ptlrpc_request_free(req);
459 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
462 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
464 ptlrpc_request_set_replen(req);
466 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
467 oa->o_flags == OBD_FL_DELORPHAN) {
469 "delorphan from OST integration");
470 /* Don't resend the delorphan req */
471 req->rq_no_resend = req->rq_no_delay = 1;
474 rc = ptlrpc_queue_wait(req);
478 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
480 GOTO(out_req, rc = -EPROTO);
482 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
483 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
485 oa->o_blksize = cli_brw_size(exp->exp_obd);
486 oa->o_valid |= OBD_MD_FLBLKSZ;
488 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
489 * have valid lsm_oinfo data structs, so don't go touching that.
490 * This needs to be fixed in a big way.
492 lsm->lsm_oi = oa->o_oi;
496 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
498 if (oa->o_valid & OBD_MD_FLCOOKIE) {
499 if (!oti->oti_logcookies)
500 oti_alloc_cookies(oti, 1);
501 *oti->oti_logcookies = oa->o_lcookie;
505 CDEBUG(D_HA, "transno: "LPD64"\n",
506 lustre_msg_get_transno(req->rq_repmsg));
508 ptlrpc_req_finished(req);
511 obd_free_memmd(exp, &lsm);
515 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
516 obd_enqueue_update_f upcall, void *cookie,
517 struct ptlrpc_request_set *rqset)
519 struct ptlrpc_request *req;
520 struct osc_setattr_args *sa;
521 struct ost_body *body;
525 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
529 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
530 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
532 ptlrpc_request_free(req);
535 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
536 ptlrpc_at_set_req_timeout(req);
538 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
540 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
542 osc_pack_capa(req, body, oinfo->oi_capa);
544 ptlrpc_request_set_replen(req);
546 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
547 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
548 sa = ptlrpc_req_async_args(req);
549 sa->sa_oa = oinfo->oi_oa;
550 sa->sa_upcall = upcall;
551 sa->sa_cookie = cookie;
552 if (rqset == PTLRPCD_SET)
553 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
555 ptlrpc_set_add_req(rqset, req);
560 static int osc_punch(const struct lu_env *env, struct obd_export *exp,
561 struct obd_info *oinfo, struct obd_trans_info *oti,
562 struct ptlrpc_request_set *rqset)
564 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
565 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
566 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
567 return osc_punch_base(exp, oinfo,
568 oinfo->oi_cb_up, oinfo, rqset);
571 static int osc_sync_interpret(const struct lu_env *env,
572 struct ptlrpc_request *req,
575 struct osc_fsync_args *fa = arg;
576 struct ost_body *body;
582 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
584 CERROR ("can't unpack ost_body\n");
585 GOTO(out, rc = -EPROTO);
588 *fa->fa_oi->oi_oa = body->oa;
590 rc = fa->fa_upcall(fa->fa_cookie, rc);
594 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
595 obd_enqueue_update_f upcall, void *cookie,
596 struct ptlrpc_request_set *rqset)
598 struct ptlrpc_request *req;
599 struct ost_body *body;
600 struct osc_fsync_args *fa;
604 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
608 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
609 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
611 ptlrpc_request_free(req);
615 /* overload the size and blocks fields in the oa with start/end */
616 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
618 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
620 osc_pack_capa(req, body, oinfo->oi_capa);
622 ptlrpc_request_set_replen(req);
623 req->rq_interpret_reply = osc_sync_interpret;
625 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
626 fa = ptlrpc_req_async_args(req);
628 fa->fa_upcall = upcall;
629 fa->fa_cookie = cookie;
631 if (rqset == PTLRPCD_SET)
632 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
634 ptlrpc_set_add_req(rqset, req);
639 static int osc_sync(const struct lu_env *env, struct obd_export *exp,
640 struct obd_info *oinfo, obd_size start, obd_size end,
641 struct ptlrpc_request_set *set)
646 CDEBUG(D_INFO, "oa NULL\n");
650 oinfo->oi_oa->o_size = start;
651 oinfo->oi_oa->o_blocks = end;
652 oinfo->oi_oa->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
654 RETURN(osc_sync_base(exp, oinfo, oinfo->oi_cb_up, oinfo, set));
657 /* Find and cancel locally locks matched by @mode in the resource found by
658 * @objid. Found locks are added into @cancel list. Returns the amount of
659 * locks added to @cancels list. */
660 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
662 ldlm_mode_t mode, int lock_flags)
664 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
665 struct ldlm_res_id res_id;
666 struct ldlm_resource *res;
670 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
671 * export) but disabled through procfs (flag in NS).
673 * This distinguishes from a case when ELC is not supported originally,
674 * when we still want to cancel locks in advance and just cancel them
675 * locally, without sending any RPC. */
676 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
679 ostid_build_res_name(&oa->o_oi, &res_id);
680 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
684 LDLM_RESOURCE_ADDREF(res);
685 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
686 lock_flags, 0, NULL);
687 LDLM_RESOURCE_DELREF(res);
688 ldlm_resource_putref(res);
692 static int osc_destroy_interpret(const struct lu_env *env,
693 struct ptlrpc_request *req, void *data,
696 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
698 cfs_atomic_dec(&cli->cl_destroy_in_flight);
699 cfs_waitq_signal(&cli->cl_destroy_waitq);
703 static int osc_can_send_destroy(struct client_obd *cli)
705 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
706 cli->cl_max_rpcs_in_flight) {
707 /* The destroy request can be sent */
710 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
711 cli->cl_max_rpcs_in_flight) {
713 * The counter has been modified between the two atomic
716 cfs_waitq_signal(&cli->cl_destroy_waitq);
721 int osc_create(const struct lu_env *env, struct obd_export *exp,
722 struct obdo *oa, struct lov_stripe_md **ea,
723 struct obd_trans_info *oti)
730 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
732 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
733 oa->o_flags == OBD_FL_RECREATE_OBJS) {
734 RETURN(osc_real_create(exp, oa, ea, oti));
737 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
738 RETURN(osc_real_create(exp, oa, ea, oti));
740 /* we should not get here anymore */
746 /* Destroy requests can be async always on the client, and we don't even really
747 * care about the return code since the client cannot do anything at all about
749 * When the MDS is unlinking a filename, it saves the file objects into a
750 * recovery llog, and these object records are cancelled when the OST reports
751 * they were destroyed and sync'd to disk (i.e. transaction committed).
752 * If the client dies, or the OST is down when the object should be destroyed,
753 * the records are not cancelled, and when the OST reconnects to the MDS next,
754 * it will retrieve the llog unlink logs and then sends the log cancellation
755 * cookies to the MDS after committing destroy transactions. */
756 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
757 struct obdo *oa, struct lov_stripe_md *ea,
758 struct obd_trans_info *oti, struct obd_export *md_export,
761 struct client_obd *cli = &exp->exp_obd->u.cli;
762 struct ptlrpc_request *req;
763 struct ost_body *body;
764 CFS_LIST_HEAD(cancels);
769 CDEBUG(D_INFO, "oa NULL\n");
773 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
774 LDLM_FL_DISCARD_DATA);
776 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
778 ldlm_lock_list_put(&cancels, l_bl_ast, count);
782 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
783 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
786 ptlrpc_request_free(req);
790 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
791 ptlrpc_at_set_req_timeout(req);
793 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
794 oa->o_lcookie = *oti->oti_logcookies;
795 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
797 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
799 osc_pack_capa(req, body, (struct obd_capa *)capa);
800 ptlrpc_request_set_replen(req);
802 /* If osc_destory is for destroying the unlink orphan,
803 * sent from MDT to OST, which should not be blocked here,
804 * because the process might be triggered by ptlrpcd, and
805 * it is not good to block ptlrpcd thread (b=16006)*/
806 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
807 req->rq_interpret_reply = osc_destroy_interpret;
808 if (!osc_can_send_destroy(cli)) {
809 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
813 * Wait until the number of on-going destroy RPCs drops
814 * under max_rpc_in_flight
816 l_wait_event_exclusive(cli->cl_destroy_waitq,
817 osc_can_send_destroy(cli), &lwi);
821 /* Do not wait for response */
822 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
826 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
829 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
831 LASSERT(!(oa->o_valid & bits));
834 client_obd_list_lock(&cli->cl_loi_list_lock);
835 oa->o_dirty = cli->cl_dirty;
836 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
837 cli->cl_dirty_max)) {
838 CERROR("dirty %lu - %lu > dirty_max %lu\n",
839 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
841 } else if (unlikely(cfs_atomic_read(&obd_dirty_pages) -
842 cfs_atomic_read(&obd_dirty_transit_pages) >
843 (long)(obd_max_dirty_pages + 1))) {
844 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
845 * not covered by a lock thus they may safely race and trip
846 * this CERROR() unless we add in a small fudge factor (+1). */
847 CERROR("dirty %d - %d > system dirty_max %d\n",
848 cfs_atomic_read(&obd_dirty_pages),
849 cfs_atomic_read(&obd_dirty_transit_pages),
850 obd_max_dirty_pages);
852 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
853 CERROR("dirty %lu - dirty_max %lu too big???\n",
854 cli->cl_dirty, cli->cl_dirty_max);
857 long max_in_flight = (cli->cl_max_pages_per_rpc <<
859 (cli->cl_max_rpcs_in_flight + 1);
860 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
862 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
863 oa->o_dropped = cli->cl_lost_grant;
864 cli->cl_lost_grant = 0;
865 client_obd_list_unlock(&cli->cl_loi_list_lock);
866 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
867 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
871 void osc_update_next_shrink(struct client_obd *cli)
873 cli->cl_next_shrink_grant =
874 cfs_time_shift(cli->cl_grant_shrink_interval);
875 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
876 cli->cl_next_shrink_grant);
879 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
881 client_obd_list_lock(&cli->cl_loi_list_lock);
882 cli->cl_avail_grant += grant;
883 client_obd_list_unlock(&cli->cl_loi_list_lock);
886 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
888 if (body->oa.o_valid & OBD_MD_FLGRANT) {
889 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
890 __osc_update_grant(cli, body->oa.o_grant);
894 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
895 obd_count keylen, void *key, obd_count vallen,
896 void *val, struct ptlrpc_request_set *set);
898 static int osc_shrink_grant_interpret(const struct lu_env *env,
899 struct ptlrpc_request *req,
902 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
903 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
904 struct ost_body *body;
907 __osc_update_grant(cli, oa->o_grant);
911 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
913 osc_update_grant(cli, body);
919 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
921 client_obd_list_lock(&cli->cl_loi_list_lock);
922 oa->o_grant = cli->cl_avail_grant / 4;
923 cli->cl_avail_grant -= oa->o_grant;
924 client_obd_list_unlock(&cli->cl_loi_list_lock);
925 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
926 oa->o_valid |= OBD_MD_FLFLAGS;
929 oa->o_flags |= OBD_FL_SHRINK_GRANT;
930 osc_update_next_shrink(cli);
933 /* Shrink the current grant, either from some large amount to enough for a
934 * full set of in-flight RPCs, or if we have already shrunk to that limit
935 * then to enough for a single RPC. This avoids keeping more grant than
936 * needed, and avoids shrinking the grant piecemeal. */
937 static int osc_shrink_grant(struct client_obd *cli)
939 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
940 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
942 client_obd_list_lock(&cli->cl_loi_list_lock);
943 if (cli->cl_avail_grant <= target_bytes)
944 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
945 client_obd_list_unlock(&cli->cl_loi_list_lock);
947 return osc_shrink_grant_to_target(cli, target_bytes);
950 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
953 struct ost_body *body;
956 client_obd_list_lock(&cli->cl_loi_list_lock);
957 /* Don't shrink if we are already above or below the desired limit
958 * We don't want to shrink below a single RPC, as that will negatively
959 * impact block allocation and long-term performance. */
960 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
961 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
963 if (target_bytes >= cli->cl_avail_grant) {
964 client_obd_list_unlock(&cli->cl_loi_list_lock);
967 client_obd_list_unlock(&cli->cl_loi_list_lock);
973 osc_announce_cached(cli, &body->oa, 0);
975 client_obd_list_lock(&cli->cl_loi_list_lock);
976 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
977 cli->cl_avail_grant = target_bytes;
978 client_obd_list_unlock(&cli->cl_loi_list_lock);
979 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
980 body->oa.o_valid |= OBD_MD_FLFLAGS;
981 body->oa.o_flags = 0;
983 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
984 osc_update_next_shrink(cli);
986 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
987 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
988 sizeof(*body), body, NULL);
990 __osc_update_grant(cli, body->oa.o_grant);
995 static int osc_should_shrink_grant(struct client_obd *client)
997 cfs_time_t time = cfs_time_current();
998 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1000 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1001 OBD_CONNECT_GRANT_SHRINK) == 0)
1004 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1005 /* Get the current RPC size directly, instead of going via:
1006 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
1007 * Keep comment here so that it can be found by searching. */
1008 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
1010 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1011 client->cl_avail_grant > brw_size)
1014 osc_update_next_shrink(client);
1019 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1021 struct client_obd *client;
1023 cfs_list_for_each_entry(client, &item->ti_obd_list,
1024 cl_grant_shrink_list) {
1025 if (osc_should_shrink_grant(client))
1026 osc_shrink_grant(client);
1031 static int osc_add_shrink_grant(struct client_obd *client)
1035 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1037 osc_grant_shrink_grant_cb, NULL,
1038 &client->cl_grant_shrink_list);
1040 CERROR("add grant client %s error %d\n",
1041 client->cl_import->imp_obd->obd_name, rc);
1044 CDEBUG(D_CACHE, "add grant client %s \n",
1045 client->cl_import->imp_obd->obd_name);
1046 osc_update_next_shrink(client);
1050 static int osc_del_shrink_grant(struct client_obd *client)
1052 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1056 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1059 * ocd_grant is the total grant amount we're expect to hold: if we've
1060 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1061 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1063 * race is tolerable here: if we're evicted, but imp_state already
1064 * left EVICTED state, then cl_dirty must be 0 already.
1066 client_obd_list_lock(&cli->cl_loi_list_lock);
1067 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1068 cli->cl_avail_grant = ocd->ocd_grant;
1070 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1072 if (cli->cl_avail_grant < 0) {
1073 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1074 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1075 ocd->ocd_grant, cli->cl_dirty);
1076 /* workaround for servers which do not have the patch from
1078 cli->cl_avail_grant = ocd->ocd_grant;
1081 /* determine the appropriate chunk size used by osc_extent. */
1082 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1083 client_obd_list_unlock(&cli->cl_loi_list_lock);
1085 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1086 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1087 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1089 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1090 cfs_list_empty(&cli->cl_grant_shrink_list))
1091 osc_add_shrink_grant(cli);
1094 /* We assume that the reason this OSC got a short read is because it read
1095 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1096 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1097 * this stripe never got written at or beyond this stripe offset yet. */
1098 static void handle_short_read(int nob_read, obd_count page_count,
1099 struct brw_page **pga)
1104 /* skip bytes read OK */
1105 while (nob_read > 0) {
1106 LASSERT (page_count > 0);
1108 if (pga[i]->count > nob_read) {
1109 /* EOF inside this page */
1110 ptr = kmap(pga[i]->pg) +
1111 (pga[i]->off & ~CFS_PAGE_MASK);
1112 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1119 nob_read -= pga[i]->count;
1124 /* zero remaining pages */
1125 while (page_count-- > 0) {
1126 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1127 memset(ptr, 0, pga[i]->count);
1133 static int check_write_rcs(struct ptlrpc_request *req,
1134 int requested_nob, int niocount,
1135 obd_count page_count, struct brw_page **pga)
1140 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1141 sizeof(*remote_rcs) *
1143 if (remote_rcs == NULL) {
1144 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1148 /* return error if any niobuf was in error */
1149 for (i = 0; i < niocount; i++) {
1150 if ((int)remote_rcs[i] < 0)
1151 return(remote_rcs[i]);
1153 if (remote_rcs[i] != 0) {
1154 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1155 i, remote_rcs[i], req);
1160 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1161 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1162 req->rq_bulk->bd_nob_transferred, requested_nob);
1169 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1171 if (p1->flag != p2->flag) {
1172 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1173 OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1175 /* warn if we try to combine flags that we don't know to be
1176 * safe to combine */
1177 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1178 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1179 "report this at http://bugs.whamcloud.com/\n",
1180 p1->flag, p2->flag);
1185 return (p1->off + p1->count == p2->off);
1188 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1189 struct brw_page **pga, int opc,
1190 cksum_type_t cksum_type)
1194 struct cfs_crypto_hash_desc *hdesc;
1195 unsigned int bufsize;
1197 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1199 LASSERT(pg_count > 0);
1201 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1202 if (IS_ERR(hdesc)) {
1203 CERROR("Unable to initialize checksum hash %s\n",
1204 cfs_crypto_hash_name(cfs_alg));
1205 return PTR_ERR(hdesc);
1208 while (nob > 0 && pg_count > 0) {
1209 int count = pga[i]->count > nob ? nob : pga[i]->count;
1211 /* corrupt the data before we compute the checksum, to
1212 * simulate an OST->client data error */
1213 if (i == 0 && opc == OST_READ &&
1214 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1215 unsigned char *ptr = kmap(pga[i]->pg);
1216 int off = pga[i]->off & ~CFS_PAGE_MASK;
1217 memcpy(ptr + off, "bad1", min(4, nob));
1220 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1221 pga[i]->off & ~CFS_PAGE_MASK,
1223 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1224 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1226 nob -= pga[i]->count;
1232 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1235 cfs_crypto_hash_final(hdesc, NULL, NULL);
1237 /* For sending we only compute the wrong checksum instead
1238 * of corrupting the data so it is still correct on a redo */
1239 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1245 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1246 struct lov_stripe_md *lsm, obd_count page_count,
1247 struct brw_page **pga,
1248 struct ptlrpc_request **reqp,
1249 struct obd_capa *ocapa, int reserve,
1252 struct ptlrpc_request *req;
1253 struct ptlrpc_bulk_desc *desc;
1254 struct ost_body *body;
1255 struct obd_ioobj *ioobj;
1256 struct niobuf_remote *niobuf;
1257 int niocount, i, requested_nob, opc, rc;
1258 struct osc_brw_async_args *aa;
1259 struct req_capsule *pill;
1260 struct brw_page *pg_prev;
1263 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1264 RETURN(-ENOMEM); /* Recoverable */
1265 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1266 RETURN(-EINVAL); /* Fatal */
1268 if ((cmd & OBD_BRW_WRITE) != 0) {
1270 req = ptlrpc_request_alloc_pool(cli->cl_import,
1271 cli->cl_import->imp_rq_pool,
1272 &RQF_OST_BRW_WRITE);
1275 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1280 for (niocount = i = 1; i < page_count; i++) {
1281 if (!can_merge_pages(pga[i - 1], pga[i]))
1285 pill = &req->rq_pill;
1286 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1288 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1289 niocount * sizeof(*niobuf));
1290 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1292 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1294 ptlrpc_request_free(req);
1297 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1298 ptlrpc_at_set_req_timeout(req);
1299 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1301 req->rq_no_retry_einprogress = 1;
1303 desc = ptlrpc_prep_bulk_imp(req, page_count,
1304 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1305 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1309 GOTO(out, rc = -ENOMEM);
1310 /* NB request now owns desc and will free it when it gets freed */
1312 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1313 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1314 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1315 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1317 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1319 obdo_to_ioobj(oa, ioobj);
1320 ioobj->ioo_bufcnt = niocount;
1321 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1322 * that might be send for this request. The actual number is decided
1323 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1324 * "max - 1" for old client compatibility sending "0", and also so the
1325 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1326 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1327 osc_pack_capa(req, body, ocapa);
1328 LASSERT(page_count > 0);
1330 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1331 struct brw_page *pg = pga[i];
1332 int poff = pg->off & ~CFS_PAGE_MASK;
1334 LASSERT(pg->count > 0);
1335 /* make sure there is no gap in the middle of page array */
1336 LASSERTF(page_count == 1 ||
1337 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1338 ergo(i > 0 && i < page_count - 1,
1339 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1340 ergo(i == page_count - 1, poff == 0)),
1341 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1342 i, page_count, pg, pg->off, pg->count);
1344 LASSERTF(i == 0 || pg->off > pg_prev->off,
1345 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1346 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1348 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1349 pg_prev->pg, page_private(pg_prev->pg),
1350 pg_prev->pg->index, pg_prev->off);
1352 LASSERTF(i == 0 || pg->off > pg_prev->off,
1353 "i %d p_c %u\n", i, page_count);
1355 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1356 (pg->flag & OBD_BRW_SRVLOCK));
1358 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1359 requested_nob += pg->count;
1361 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1363 niobuf->len += pg->count;
1365 niobuf->offset = pg->off;
1366 niobuf->len = pg->count;
1367 niobuf->flags = pg->flag;
1372 LASSERTF((void *)(niobuf - niocount) ==
1373 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1374 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1375 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1377 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1379 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1380 body->oa.o_valid |= OBD_MD_FLFLAGS;
1381 body->oa.o_flags = 0;
1383 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1386 if (osc_should_shrink_grant(cli))
1387 osc_shrink_grant_local(cli, &body->oa);
1389 /* size[REQ_REC_OFF] still sizeof (*body) */
1390 if (opc == OST_WRITE) {
1391 if (cli->cl_checksum &&
1392 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1393 /* store cl_cksum_type in a local variable since
1394 * it can be changed via lprocfs */
1395 cksum_type_t cksum_type = cli->cl_cksum_type;
1397 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1398 oa->o_flags &= OBD_FL_LOCAL_MASK;
1399 body->oa.o_flags = 0;
1401 body->oa.o_flags |= cksum_type_pack(cksum_type);
1402 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1403 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1407 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1409 /* save this in 'oa', too, for later checking */
1410 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1411 oa->o_flags |= cksum_type_pack(cksum_type);
1413 /* clear out the checksum flag, in case this is a
1414 * resend but cl_checksum is no longer set. b=11238 */
1415 oa->o_valid &= ~OBD_MD_FLCKSUM;
1417 oa->o_cksum = body->oa.o_cksum;
1418 /* 1 RC per niobuf */
1419 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1420 sizeof(__u32) * niocount);
1422 if (cli->cl_checksum &&
1423 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1424 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1425 body->oa.o_flags = 0;
1426 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1427 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1430 ptlrpc_request_set_replen(req);
1432 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1433 aa = ptlrpc_req_async_args(req);
1435 aa->aa_requested_nob = requested_nob;
1436 aa->aa_nio_count = niocount;
1437 aa->aa_page_count = page_count;
1441 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1442 if (ocapa && reserve)
1443 aa->aa_ocapa = capa_get(ocapa);
1449 ptlrpc_req_finished(req);
1453 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1454 __u32 client_cksum, __u32 server_cksum, int nob,
1455 obd_count page_count, struct brw_page **pga,
1456 cksum_type_t client_cksum_type)
1460 cksum_type_t cksum_type;
1462 if (server_cksum == client_cksum) {
1463 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1467 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1469 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1472 if (cksum_type != client_cksum_type)
1473 msg = "the server did not use the checksum type specified in "
1474 "the original request - likely a protocol problem";
1475 else if (new_cksum == server_cksum)
1476 msg = "changed on the client after we checksummed it - "
1477 "likely false positive due to mmap IO (bug 11742)";
1478 else if (new_cksum == client_cksum)
1479 msg = "changed in transit before arrival at OST";
1481 msg = "changed in transit AND doesn't match the original - "
1482 "likely false positive due to mmap IO (bug 11742)";
1484 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1485 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1486 msg, libcfs_nid2str(peer->nid),
1487 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1488 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1489 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1490 POSTID(&oa->o_oi), pga[0]->off,
1491 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1492 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1493 "client csum now %x\n", client_cksum, client_cksum_type,
1494 server_cksum, cksum_type, new_cksum);
1498 /* Note rc enters this function as number of bytes transferred */
1499 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1501 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1502 const lnet_process_id_t *peer =
1503 &req->rq_import->imp_connection->c_peer;
1504 struct client_obd *cli = aa->aa_cli;
1505 struct ost_body *body;
1506 __u32 client_cksum = 0;
1509 if (rc < 0 && rc != -EDQUOT) {
1510 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1514 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1515 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1517 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1521 /* set/clear over quota flag for a uid/gid */
1522 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1523 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1524 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1526 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1527 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1529 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1532 osc_update_grant(cli, body);
1537 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1538 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1540 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1542 CERROR("Unexpected +ve rc %d\n", rc);
1545 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1547 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1550 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1551 check_write_checksum(&body->oa, peer, client_cksum,
1552 body->oa.o_cksum, aa->aa_requested_nob,
1553 aa->aa_page_count, aa->aa_ppga,
1554 cksum_type_unpack(aa->aa_oa->o_flags)))
1557 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1558 aa->aa_page_count, aa->aa_ppga);
1562 /* The rest of this function executes only for OST_READs */
1564 /* if unwrap_bulk failed, return -EAGAIN to retry */
1565 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1567 GOTO(out, rc = -EAGAIN);
1569 if (rc > aa->aa_requested_nob) {
1570 CERROR("Unexpected rc %d (%d requested)\n", rc,
1571 aa->aa_requested_nob);
1575 if (rc != req->rq_bulk->bd_nob_transferred) {
1576 CERROR ("Unexpected rc %d (%d transferred)\n",
1577 rc, req->rq_bulk->bd_nob_transferred);
1581 if (rc < aa->aa_requested_nob)
1582 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1584 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1585 static int cksum_counter;
1586 __u32 server_cksum = body->oa.o_cksum;
1589 cksum_type_t cksum_type;
1591 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1592 body->oa.o_flags : 0);
1593 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1594 aa->aa_ppga, OST_READ,
1597 if (peer->nid == req->rq_bulk->bd_sender) {
1601 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1604 if (server_cksum == ~0 && rc > 0) {
1605 CERROR("Protocol error: server %s set the 'checksum' "
1606 "bit, but didn't send a checksum. Not fatal, "
1607 "but please notify on http://bugs.whamcloud.com/\n",
1608 libcfs_nid2str(peer->nid));
1609 } else if (server_cksum != client_cksum) {
1610 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1611 "%s%s%s inode "DFID" object "DOSTID
1612 " extent ["LPU64"-"LPU64"]\n",
1613 req->rq_import->imp_obd->obd_name,
1614 libcfs_nid2str(peer->nid),
1616 body->oa.o_valid & OBD_MD_FLFID ?
1617 body->oa.o_parent_seq : (__u64)0,
1618 body->oa.o_valid & OBD_MD_FLFID ?
1619 body->oa.o_parent_oid : 0,
1620 body->oa.o_valid & OBD_MD_FLFID ?
1621 body->oa.o_parent_ver : 0,
1622 POSTID(&body->oa.o_oi),
1623 aa->aa_ppga[0]->off,
1624 aa->aa_ppga[aa->aa_page_count-1]->off +
1625 aa->aa_ppga[aa->aa_page_count-1]->count -
1627 CERROR("client %x, server %x, cksum_type %x\n",
1628 client_cksum, server_cksum, cksum_type);
1630 aa->aa_oa->o_cksum = client_cksum;
1634 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1637 } else if (unlikely(client_cksum)) {
1638 static int cksum_missed;
1641 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1642 CERROR("Checksum %u requested from %s but not sent\n",
1643 cksum_missed, libcfs_nid2str(peer->nid));
1649 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1650 aa->aa_oa, &body->oa);
1655 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1656 struct lov_stripe_md *lsm,
1657 obd_count page_count, struct brw_page **pga,
1658 struct obd_capa *ocapa)
1660 struct ptlrpc_request *req;
1663 int generation, resends = 0;
1664 struct l_wait_info lwi;
1668 cfs_waitq_init(&waitq);
1669 generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1672 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1673 page_count, pga, &req, ocapa, 0, resends);
1678 req->rq_generation_set = 1;
1679 req->rq_import_generation = generation;
1680 req->rq_sent = cfs_time_current_sec() + resends;
1683 rc = ptlrpc_queue_wait(req);
1685 if (rc == -ETIMEDOUT && req->rq_resend) {
1686 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1687 ptlrpc_req_finished(req);
1691 rc = osc_brw_fini_request(req, rc);
1693 ptlrpc_req_finished(req);
1694 /* When server return -EINPROGRESS, client should always retry
1695 * regardless of the number of times the bulk was resent already.*/
1696 if (osc_recoverable_error(rc)) {
1698 if (rc != -EINPROGRESS &&
1699 !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1700 CERROR("%s: too many resend retries for object: "
1701 ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1702 POSTID(&oa->o_oi), rc);
1706 exp->exp_obd->u.cli.cl_import->imp_generation) {
1707 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1708 ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1709 POSTID(&oa->o_oi), rc);
1713 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1715 l_wait_event(waitq, 0, &lwi);
1720 if (rc == -EAGAIN || rc == -EINPROGRESS)
1725 static int osc_brw_redo_request(struct ptlrpc_request *request,
1726 struct osc_brw_async_args *aa, int rc)
1728 struct ptlrpc_request *new_req;
1729 struct osc_brw_async_args *new_aa;
1730 struct osc_async_page *oap;
1733 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1734 "redo for recoverable error %d", rc);
1736 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1737 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1738 aa->aa_cli, aa->aa_oa,
1739 NULL /* lsm unused by osc currently */,
1740 aa->aa_page_count, aa->aa_ppga,
1741 &new_req, aa->aa_ocapa, 0, 1);
1745 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1746 if (oap->oap_request != NULL) {
1747 LASSERTF(request == oap->oap_request,
1748 "request %p != oap_request %p\n",
1749 request, oap->oap_request);
1750 if (oap->oap_interrupted) {
1751 ptlrpc_req_finished(new_req);
1756 /* New request takes over pga and oaps from old request.
1757 * Note that copying a list_head doesn't work, need to move it... */
1759 new_req->rq_interpret_reply = request->rq_interpret_reply;
1760 new_req->rq_async_args = request->rq_async_args;
1761 /* cap resend delay to the current request timeout, this is similar to
1762 * what ptlrpc does (see after_reply()) */
1763 if (aa->aa_resends > new_req->rq_timeout)
1764 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1766 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1767 new_req->rq_generation_set = 1;
1768 new_req->rq_import_generation = request->rq_import_generation;
1770 new_aa = ptlrpc_req_async_args(new_req);
1772 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1773 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1774 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1775 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1776 new_aa->aa_resends = aa->aa_resends;
1778 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1779 if (oap->oap_request) {
1780 ptlrpc_req_finished(oap->oap_request);
1781 oap->oap_request = ptlrpc_request_addref(new_req);
1785 new_aa->aa_ocapa = aa->aa_ocapa;
1786 aa->aa_ocapa = NULL;
1788 /* XXX: This code will run into problem if we're going to support
1789 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1790 * and wait for all of them to be finished. We should inherit request
1791 * set from old request. */
1792 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1794 DEBUG_REQ(D_INFO, new_req, "new request");
1799 * ugh, we want disk allocation on the target to happen in offset order. we'll
1800 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1801 * fine for our small page arrays and doesn't require allocation. its an
1802 * insertion sort that swaps elements that are strides apart, shrinking the
1803 * stride down until its '1' and the array is sorted.
1805 static void sort_brw_pages(struct brw_page **array, int num)
1808 struct brw_page *tmp;
1812 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1817 for (i = stride ; i < num ; i++) {
1820 while (j >= stride && array[j - stride]->off > tmp->off) {
1821 array[j] = array[j - stride];
1826 } while (stride > 1);
1829 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1835 LASSERT (pages > 0);
1836 offset = pg[i]->off & ~CFS_PAGE_MASK;
1840 if (pages == 0) /* that's all */
1843 if (offset + pg[i]->count < PAGE_CACHE_SIZE)
1844 return count; /* doesn't end on page boundary */
1847 offset = pg[i]->off & ~CFS_PAGE_MASK;
1848 if (offset != 0) /* doesn't start on page boundary */
1855 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1857 struct brw_page **ppga;
1860 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1864 for (i = 0; i < count; i++)
1869 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1871 LASSERT(ppga != NULL);
1872 OBD_FREE(ppga, sizeof(*ppga) * count);
1875 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1876 obd_count page_count, struct brw_page *pga,
1877 struct obd_trans_info *oti)
1879 struct obdo *saved_oa = NULL;
1880 struct brw_page **ppga, **orig;
1881 struct obd_import *imp = class_exp2cliimp(exp);
1882 struct client_obd *cli;
1883 int rc, page_count_orig;
1886 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1887 cli = &imp->imp_obd->u.cli;
1889 if (cmd & OBD_BRW_CHECK) {
1890 /* The caller just wants to know if there's a chance that this
1891 * I/O can succeed */
1893 if (imp->imp_invalid)
1898 /* test_brw with a failed create can trip this, maybe others. */
1899 LASSERT(cli->cl_max_pages_per_rpc);
1903 orig = ppga = osc_build_ppga(pga, page_count);
1906 page_count_orig = page_count;
1908 sort_brw_pages(ppga, page_count);
1909 while (page_count) {
1910 obd_count pages_per_brw;
1912 if (page_count > cli->cl_max_pages_per_rpc)
1913 pages_per_brw = cli->cl_max_pages_per_rpc;
1915 pages_per_brw = page_count;
1917 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1919 if (saved_oa != NULL) {
1920 /* restore previously saved oa */
1921 *oinfo->oi_oa = *saved_oa;
1922 } else if (page_count > pages_per_brw) {
1923 /* save a copy of oa (brw will clobber it) */
1924 OBDO_ALLOC(saved_oa);
1925 if (saved_oa == NULL)
1926 GOTO(out, rc = -ENOMEM);
1927 *saved_oa = *oinfo->oi_oa;
1930 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1931 pages_per_brw, ppga, oinfo->oi_capa);
1936 page_count -= pages_per_brw;
1937 ppga += pages_per_brw;
1941 osc_release_ppga(orig, page_count_orig);
1943 if (saved_oa != NULL)
1944 OBDO_FREE(saved_oa);
1949 static int brw_interpret(const struct lu_env *env,
1950 struct ptlrpc_request *req, void *data, int rc)
1952 struct osc_brw_async_args *aa = data;
1953 struct osc_extent *ext;
1954 struct osc_extent *tmp;
1955 struct cl_object *obj = NULL;
1956 struct client_obd *cli = aa->aa_cli;
1959 rc = osc_brw_fini_request(req, rc);
1960 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1961 /* When server return -EINPROGRESS, client should always retry
1962 * regardless of the number of times the bulk was resent already. */
1963 if (osc_recoverable_error(rc)) {
1964 if (req->rq_import_generation !=
1965 req->rq_import->imp_generation) {
1966 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1967 ""DOSTID", rc = %d.\n",
1968 req->rq_import->imp_obd->obd_name,
1969 POSTID(&aa->aa_oa->o_oi), rc);
1970 } else if (rc == -EINPROGRESS ||
1971 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1972 rc = osc_brw_redo_request(req, aa, rc);
1974 CERROR("%s: too many resent retries for object: "
1975 ""LPU64":"LPU64", rc = %d.\n",
1976 req->rq_import->imp_obd->obd_name,
1977 POSTID(&aa->aa_oa->o_oi), rc);
1982 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1987 capa_put(aa->aa_ocapa);
1988 aa->aa_ocapa = NULL;
1991 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1992 if (obj == NULL && rc == 0) {
1993 obj = osc2cl(ext->oe_obj);
1997 cfs_list_del_init(&ext->oe_link);
1998 osc_extent_finish(env, ext, 1, rc);
2000 LASSERT(cfs_list_empty(&aa->aa_exts));
2001 LASSERT(cfs_list_empty(&aa->aa_oaps));
2004 struct obdo *oa = aa->aa_oa;
2005 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2006 unsigned long valid = 0;
2009 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2010 attr->cat_blocks = oa->o_blocks;
2011 valid |= CAT_BLOCKS;
2013 if (oa->o_valid & OBD_MD_FLMTIME) {
2014 attr->cat_mtime = oa->o_mtime;
2017 if (oa->o_valid & OBD_MD_FLATIME) {
2018 attr->cat_atime = oa->o_atime;
2021 if (oa->o_valid & OBD_MD_FLCTIME) {
2022 attr->cat_ctime = oa->o_ctime;
2026 cl_object_attr_lock(obj);
2027 cl_object_attr_set(env, obj, attr, valid);
2028 cl_object_attr_unlock(obj);
2030 cl_object_put(env, obj);
2032 OBDO_FREE(aa->aa_oa);
2034 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2035 req->rq_bulk->bd_nob_transferred);
2036 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2037 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2039 client_obd_list_lock(&cli->cl_loi_list_lock);
2040 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2041 * is called so we know whether to go to sync BRWs or wait for more
2042 * RPCs to complete */
2043 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2044 cli->cl_w_in_flight--;
2046 cli->cl_r_in_flight--;
2047 osc_wake_cache_waiters(cli);
2048 client_obd_list_unlock(&cli->cl_loi_list_lock);
2050 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2055 * Build an RPC by the list of extent @ext_list. The caller must ensure
2056 * that the total pages in this list are NOT over max pages per RPC.
2057 * Extents in the list must be in OES_RPC state.
2059 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2060 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2062 struct ptlrpc_request *req = NULL;
2063 struct osc_extent *ext;
2064 struct brw_page **pga = NULL;
2065 struct osc_brw_async_args *aa = NULL;
2066 struct obdo *oa = NULL;
2067 struct osc_async_page *oap;
2068 struct osc_async_page *tmp;
2069 struct cl_req *clerq = NULL;
2070 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
2072 struct ldlm_lock *lock = NULL;
2073 struct cl_req_attr *crattr = NULL;
2074 obd_off starting_offset = OBD_OBJECT_EOF;
2075 obd_off ending_offset = 0;
2081 CFS_LIST_HEAD(rpc_list);
2084 LASSERT(!cfs_list_empty(ext_list));
2086 /* add pages into rpc_list to build BRW rpc */
2087 cfs_list_for_each_entry(ext, ext_list, oe_link) {
2088 LASSERT(ext->oe_state == OES_RPC);
2089 mem_tight |= ext->oe_memalloc;
2090 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2092 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2093 if (starting_offset > oap->oap_obj_off)
2094 starting_offset = oap->oap_obj_off;
2096 LASSERT(oap->oap_page_off == 0);
2097 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2098 ending_offset = oap->oap_obj_off +
2101 LASSERT(oap->oap_page_off + oap->oap_count ==
2107 mpflag = cfs_memory_pressure_get_and_set();
2109 OBD_ALLOC(crattr, sizeof(*crattr));
2111 GOTO(out, rc = -ENOMEM);
2113 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2115 GOTO(out, rc = -ENOMEM);
2119 GOTO(out, rc = -ENOMEM);
2122 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2123 struct cl_page *page = oap2cl_page(oap);
2124 if (clerq == NULL) {
2125 clerq = cl_req_alloc(env, page, crt,
2126 1 /* only 1-object rpcs for now */);
2128 GOTO(out, rc = PTR_ERR(clerq));
2129 lock = oap->oap_ldlm_lock;
2132 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2133 pga[i] = &oap->oap_brw_page;
2134 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2135 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2136 pga[i]->pg, page_index(oap->oap_page), oap,
2139 cl_req_page_add(env, clerq, page);
2142 /* always get the data for the obdo for the rpc */
2143 LASSERT(clerq != NULL);
2144 crattr->cra_oa = oa;
2145 cl_req_attr_set(env, clerq, crattr, ~0ULL);
2147 oa->o_handle = lock->l_remote_handle;
2148 oa->o_valid |= OBD_MD_FLHANDLE;
2151 rc = cl_req_prep(env, clerq);
2153 CERROR("cl_req_prep failed: %d\n", rc);
2157 sort_brw_pages(pga, page_count);
2158 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2159 pga, &req, crattr->cra_capa, 1, 0);
2161 CERROR("prep_req failed: %d\n", rc);
2165 req->rq_interpret_reply = brw_interpret;
2167 req->rq_memalloc = 1;
2169 /* Need to update the timestamps after the request is built in case
2170 * we race with setattr (locally or in queue at OST). If OST gets
2171 * later setattr before earlier BRW (as determined by the request xid),
2172 * the OST will not use BRW timestamps. Sadly, there is no obvious
2173 * way to do this in a single call. bug 10150 */
2174 cl_req_attr_set(env, clerq, crattr,
2175 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2177 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2179 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2180 aa = ptlrpc_req_async_args(req);
2181 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2182 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2183 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2184 cfs_list_splice_init(ext_list, &aa->aa_exts);
2185 aa->aa_clerq = clerq;
2187 /* queued sync pages can be torn down while the pages
2188 * were between the pending list and the rpc */
2190 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2191 /* only one oap gets a request reference */
2194 if (oap->oap_interrupted && !req->rq_intr) {
2195 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2197 ptlrpc_mark_interrupted(req);
2201 tmp->oap_request = ptlrpc_request_addref(req);
2203 client_obd_list_lock(&cli->cl_loi_list_lock);
2204 starting_offset >>= PAGE_CACHE_SHIFT;
2205 if (cmd == OBD_BRW_READ) {
2206 cli->cl_r_in_flight++;
2207 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2208 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2209 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2210 starting_offset + 1);
2212 cli->cl_w_in_flight++;
2213 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2214 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2215 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2216 starting_offset + 1);
2218 client_obd_list_unlock(&cli->cl_loi_list_lock);
2220 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2221 page_count, aa, cli->cl_r_in_flight,
2222 cli->cl_w_in_flight);
2224 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2225 * see which CPU/NUMA node the majority of pages were allocated
2226 * on, and try to assign the async RPC to the CPU core
2227 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2229 * But on the other hand, we expect that multiple ptlrpcd
2230 * threads and the initial write sponsor can run in parallel,
2231 * especially when data checksum is enabled, which is CPU-bound
2232 * operation and single ptlrpcd thread cannot process in time.
2233 * So more ptlrpcd threads sharing BRW load
2234 * (with PDL_POLICY_ROUND) seems better.
2236 ptlrpcd_add_req(req, pol, -1);
2242 cfs_memory_pressure_restore(mpflag);
2244 if (crattr != NULL) {
2245 capa_put(crattr->cra_capa);
2246 OBD_FREE(crattr, sizeof(*crattr));
2250 LASSERT(req == NULL);
2255 OBD_FREE(pga, sizeof(*pga) * page_count);
2256 /* this should happen rarely and is pretty bad, it makes the
2257 * pending list not follow the dirty order */
2258 while (!cfs_list_empty(ext_list)) {
2259 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2261 cfs_list_del_init(&ext->oe_link);
2262 osc_extent_finish(env, ext, 0, rc);
2264 if (clerq && !IS_ERR(clerq))
2265 cl_req_completion(env, clerq, rc);
2270 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2271 struct ldlm_enqueue_info *einfo)
2273 void *data = einfo->ei_cbdata;
2276 LASSERT(lock != NULL);
2277 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2278 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2279 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2280 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2282 lock_res_and_lock(lock);
2283 spin_lock(&osc_ast_guard);
2285 if (lock->l_ast_data == NULL)
2286 lock->l_ast_data = data;
2287 if (lock->l_ast_data == data)
2290 spin_unlock(&osc_ast_guard);
2291 unlock_res_and_lock(lock);
2296 static int osc_set_data_with_check(struct lustre_handle *lockh,
2297 struct ldlm_enqueue_info *einfo)
2299 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2303 set = osc_set_lock_data_with_check(lock, einfo);
2304 LDLM_LOCK_PUT(lock);
2306 CERROR("lockh %p, data %p - client evicted?\n",
2307 lockh, einfo->ei_cbdata);
2311 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2312 ldlm_iterator_t replace, void *data)
2314 struct ldlm_res_id res_id;
2315 struct obd_device *obd = class_exp2obd(exp);
2317 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2318 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2322 /* find any ldlm lock of the inode in osc
2326 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2327 ldlm_iterator_t replace, void *data)
2329 struct ldlm_res_id res_id;
2330 struct obd_device *obd = class_exp2obd(exp);
2333 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2334 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2335 if (rc == LDLM_ITER_STOP)
2337 if (rc == LDLM_ITER_CONTINUE)
2342 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2343 obd_enqueue_update_f upcall, void *cookie,
2344 __u64 *flags, int agl, int rc)
2346 int intent = *flags & LDLM_FL_HAS_INTENT;
2350 /* The request was created before ldlm_cli_enqueue call. */
2351 if (rc == ELDLM_LOCK_ABORTED) {
2352 struct ldlm_reply *rep;
2353 rep = req_capsule_server_get(&req->rq_pill,
2356 LASSERT(rep != NULL);
2357 rep->lock_policy_res1 =
2358 ptlrpc_status_ntoh(rep->lock_policy_res1);
2359 if (rep->lock_policy_res1)
2360 rc = rep->lock_policy_res1;
2364 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2366 *flags |= LDLM_FL_LVB_READY;
2367 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2368 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2371 /* Call the update callback. */
2372 rc = (*upcall)(cookie, rc);
2376 static int osc_enqueue_interpret(const struct lu_env *env,
2377 struct ptlrpc_request *req,
2378 struct osc_enqueue_args *aa, int rc)
2380 struct ldlm_lock *lock;
2381 struct lustre_handle handle;
2383 struct ost_lvb *lvb;
2385 __u64 *flags = aa->oa_flags;
2387 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2388 * might be freed anytime after lock upcall has been called. */
2389 lustre_handle_copy(&handle, aa->oa_lockh);
2390 mode = aa->oa_ei->ei_mode;
2392 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2394 lock = ldlm_handle2lock(&handle);
2396 /* Take an additional reference so that a blocking AST that
2397 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2398 * to arrive after an upcall has been executed by
2399 * osc_enqueue_fini(). */
2400 ldlm_lock_addref(&handle, mode);
2402 /* Let CP AST to grant the lock first. */
2403 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2405 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2410 lvb_len = sizeof(*aa->oa_lvb);
2413 /* Complete obtaining the lock procedure. */
2414 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2415 mode, flags, lvb, lvb_len, &handle, rc);
2416 /* Complete osc stuff. */
2417 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2418 flags, aa->oa_agl, rc);
2420 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2422 /* Release the lock for async request. */
2423 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2425 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2426 * not already released by
2427 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2429 ldlm_lock_decref(&handle, mode);
2431 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2432 aa->oa_lockh, req, aa);
2433 ldlm_lock_decref(&handle, mode);
2434 LDLM_LOCK_PUT(lock);
2438 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2439 struct lov_oinfo *loi, int flags,
2440 struct ost_lvb *lvb, __u32 mode, int rc)
2442 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2444 if (rc == ELDLM_OK) {
2447 LASSERT(lock != NULL);
2448 loi->loi_lvb = *lvb;
2449 tmp = loi->loi_lvb.lvb_size;
2450 /* Extend KMS up to the end of this lock and no further
2451 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2452 if (tmp > lock->l_policy_data.l_extent.end)
2453 tmp = lock->l_policy_data.l_extent.end + 1;
2454 if (tmp >= loi->loi_kms) {
2455 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2456 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2457 loi_kms_set(loi, tmp);
2459 LDLM_DEBUG(lock, "lock acquired, setting rss="
2460 LPU64"; leaving kms="LPU64", end="LPU64,
2461 loi->loi_lvb.lvb_size, loi->loi_kms,
2462 lock->l_policy_data.l_extent.end);
2464 ldlm_lock_allow_match(lock);
2465 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2466 LASSERT(lock != NULL);
2467 loi->loi_lvb = *lvb;
2468 ldlm_lock_allow_match(lock);
2469 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2470 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2476 ldlm_lock_fail_match(lock);
2478 LDLM_LOCK_PUT(lock);
2481 EXPORT_SYMBOL(osc_update_enqueue);
2483 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2485 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2486 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2487 * other synchronous requests, however keeping some locks and trying to obtain
2488 * others may take a considerable amount of time in a case of ost failure; and
2489 * when other sync requests do not get released lock from a client, the client
2490 * is excluded from the cluster -- such scenarious make the life difficult, so
2491 * release locks just after they are obtained. */
2492 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2493 __u64 *flags, ldlm_policy_data_t *policy,
2494 struct ost_lvb *lvb, int kms_valid,
2495 obd_enqueue_update_f upcall, void *cookie,
2496 struct ldlm_enqueue_info *einfo,
2497 struct lustre_handle *lockh,
2498 struct ptlrpc_request_set *rqset, int async, int agl)
2500 struct obd_device *obd = exp->exp_obd;
2501 struct ptlrpc_request *req = NULL;
2502 int intent = *flags & LDLM_FL_HAS_INTENT;
2503 int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2508 /* Filesystem lock extents are extended to page boundaries so that
2509 * dealing with the page cache is a little smoother. */
2510 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2511 policy->l_extent.end |= ~CFS_PAGE_MASK;
2514 * kms is not valid when either object is completely fresh (so that no
2515 * locks are cached), or object was evicted. In the latter case cached
2516 * lock cannot be used, because it would prime inode state with
2517 * potentially stale LVB.
2522 /* Next, search for already existing extent locks that will cover us */
2523 /* If we're trying to read, we also search for an existing PW lock. The
2524 * VFS and page cache already protect us locally, so lots of readers/
2525 * writers can share a single PW lock.
2527 * There are problems with conversion deadlocks, so instead of
2528 * converting a read lock to a write lock, we'll just enqueue a new
2531 * At some point we should cancel the read lock instead of making them
2532 * send us a blocking callback, but there are problems with canceling
2533 * locks out from other users right now, too. */
2534 mode = einfo->ei_mode;
2535 if (einfo->ei_mode == LCK_PR)
2537 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2538 einfo->ei_type, policy, mode, lockh, 0);
2540 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2542 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2543 /* For AGL, if enqueue RPC is sent but the lock is not
2544 * granted, then skip to process this strpe.
2545 * Return -ECANCELED to tell the caller. */
2546 ldlm_lock_decref(lockh, mode);
2547 LDLM_LOCK_PUT(matched);
2549 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2550 *flags |= LDLM_FL_LVB_READY;
2551 /* addref the lock only if not async requests and PW
2552 * lock is matched whereas we asked for PR. */
2553 if (!rqset && einfo->ei_mode != mode)
2554 ldlm_lock_addref(lockh, LCK_PR);
2556 /* I would like to be able to ASSERT here that
2557 * rss <= kms, but I can't, for reasons which
2558 * are explained in lov_enqueue() */
2561 /* We already have a lock, and it's referenced.
2563 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2564 * AGL upcall may change it to CLS_HELD directly. */
2565 (*upcall)(cookie, ELDLM_OK);
2567 if (einfo->ei_mode != mode)
2568 ldlm_lock_decref(lockh, LCK_PW);
2570 /* For async requests, decref the lock. */
2571 ldlm_lock_decref(lockh, einfo->ei_mode);
2572 LDLM_LOCK_PUT(matched);
2575 ldlm_lock_decref(lockh, mode);
2576 LDLM_LOCK_PUT(matched);
2582 CFS_LIST_HEAD(cancels);
2583 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2584 &RQF_LDLM_ENQUEUE_LVB);
2588 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2590 ptlrpc_request_free(req);
2594 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2596 ptlrpc_request_set_replen(req);
2599 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2600 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2602 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2603 sizeof(*lvb), LVB_T_OST, lockh, async);
2606 struct osc_enqueue_args *aa;
2607 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2608 aa = ptlrpc_req_async_args(req);
2611 aa->oa_flags = flags;
2612 aa->oa_upcall = upcall;
2613 aa->oa_cookie = cookie;
2615 aa->oa_lockh = lockh;
2618 req->rq_interpret_reply =
2619 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2620 if (rqset == PTLRPCD_SET)
2621 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2623 ptlrpc_set_add_req(rqset, req);
2624 } else if (intent) {
2625 ptlrpc_req_finished(req);
2630 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2632 ptlrpc_req_finished(req);
2637 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2638 struct ldlm_enqueue_info *einfo,
2639 struct ptlrpc_request_set *rqset)
2641 struct ldlm_res_id res_id;
2645 ostid_build_res_name(&oinfo->oi_md->lsm_oi, &res_id);
2646 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2647 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2648 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2649 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2650 rqset, rqset != NULL, 0);
2654 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2655 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2656 int *flags, void *data, struct lustre_handle *lockh,
2659 struct obd_device *obd = exp->exp_obd;
2660 int lflags = *flags;
2664 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2667 /* Filesystem lock extents are extended to page boundaries so that
2668 * dealing with the page cache is a little smoother */
2669 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2670 policy->l_extent.end |= ~CFS_PAGE_MASK;
2672 /* Next, search for already existing extent locks that will cover us */
2673 /* If we're trying to read, we also search for an existing PW lock. The
2674 * VFS and page cache already protect us locally, so lots of readers/
2675 * writers can share a single PW lock. */
2679 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2680 res_id, type, policy, rc, lockh, unref);
2683 if (!osc_set_data_with_check(lockh, data)) {
2684 if (!(lflags & LDLM_FL_TEST_LOCK))
2685 ldlm_lock_decref(lockh, rc);
2689 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2690 ldlm_lock_addref(lockh, LCK_PR);
2691 ldlm_lock_decref(lockh, LCK_PW);
2698 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2702 if (unlikely(mode == LCK_GROUP))
2703 ldlm_lock_decref_and_cancel(lockh, mode);
2705 ldlm_lock_decref(lockh, mode);
2710 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2711 __u32 mode, struct lustre_handle *lockh)
2714 RETURN(osc_cancel_base(lockh, mode));
2717 static int osc_cancel_unused(struct obd_export *exp,
2718 struct lov_stripe_md *lsm,
2719 ldlm_cancel_flags_t flags,
2722 struct obd_device *obd = class_exp2obd(exp);
2723 struct ldlm_res_id res_id, *resp = NULL;
2726 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2730 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2733 static int osc_statfs_interpret(const struct lu_env *env,
2734 struct ptlrpc_request *req,
2735 struct osc_async_args *aa, int rc)
2737 struct obd_statfs *msfs;
2741 /* The request has in fact never been sent
2742 * due to issues at a higher level (LOV).
2743 * Exit immediately since the caller is
2744 * aware of the problem and takes care
2745 * of the clean up */
2748 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2749 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2755 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2757 GOTO(out, rc = -EPROTO);
2760 *aa->aa_oi->oi_osfs = *msfs;
2762 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2766 static int osc_statfs_async(struct obd_export *exp,
2767 struct obd_info *oinfo, __u64 max_age,
2768 struct ptlrpc_request_set *rqset)
2770 struct obd_device *obd = class_exp2obd(exp);
2771 struct ptlrpc_request *req;
2772 struct osc_async_args *aa;
2776 /* We could possibly pass max_age in the request (as an absolute
2777 * timestamp or a "seconds.usec ago") so the target can avoid doing
2778 * extra calls into the filesystem if that isn't necessary (e.g.
2779 * during mount that would help a bit). Having relative timestamps
2780 * is not so great if request processing is slow, while absolute
2781 * timestamps are not ideal because they need time synchronization. */
2782 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2786 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2788 ptlrpc_request_free(req);
2791 ptlrpc_request_set_replen(req);
2792 req->rq_request_portal = OST_CREATE_PORTAL;
2793 ptlrpc_at_set_req_timeout(req);
2795 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2796 /* procfs requests not want stat in wait for avoid deadlock */
2797 req->rq_no_resend = 1;
2798 req->rq_no_delay = 1;
2801 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2802 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2803 aa = ptlrpc_req_async_args(req);
2806 ptlrpc_set_add_req(rqset, req);
2810 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2811 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2813 struct obd_device *obd = class_exp2obd(exp);
2814 struct obd_statfs *msfs;
2815 struct ptlrpc_request *req;
2816 struct obd_import *imp = NULL;
2820 /*Since the request might also come from lprocfs, so we need
2821 *sync this with client_disconnect_export Bug15684*/
2822 down_read(&obd->u.cli.cl_sem);
2823 if (obd->u.cli.cl_import)
2824 imp = class_import_get(obd->u.cli.cl_import);
2825 up_read(&obd->u.cli.cl_sem);
2829 /* We could possibly pass max_age in the request (as an absolute
2830 * timestamp or a "seconds.usec ago") so the target can avoid doing
2831 * extra calls into the filesystem if that isn't necessary (e.g.
2832 * during mount that would help a bit). Having relative timestamps
2833 * is not so great if request processing is slow, while absolute
2834 * timestamps are not ideal because they need time synchronization. */
2835 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2837 class_import_put(imp);
2842 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2844 ptlrpc_request_free(req);
2847 ptlrpc_request_set_replen(req);
2848 req->rq_request_portal = OST_CREATE_PORTAL;
2849 ptlrpc_at_set_req_timeout(req);
2851 if (flags & OBD_STATFS_NODELAY) {
2852 /* procfs requests not want stat in wait for avoid deadlock */
2853 req->rq_no_resend = 1;
2854 req->rq_no_delay = 1;
2857 rc = ptlrpc_queue_wait(req);
2861 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2863 GOTO(out, rc = -EPROTO);
2870 ptlrpc_req_finished(req);
2874 /* Retrieve object striping information.
2876 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2877 * the maximum number of OST indices which will fit in the user buffer.
2878 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2880 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2882 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2883 struct lov_user_md_v3 lum, *lumk;
2884 struct lov_user_ost_data_v1 *lmm_objects;
2885 int rc = 0, lum_size;
2891 /* we only need the header part from user space to get lmm_magic and
2892 * lmm_stripe_count, (the header part is common to v1 and v3) */
2893 lum_size = sizeof(struct lov_user_md_v1);
2894 if (copy_from_user(&lum, lump, lum_size))
2897 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2898 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2901 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2902 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2903 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2904 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2906 /* we can use lov_mds_md_size() to compute lum_size
2907 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2908 if (lum.lmm_stripe_count > 0) {
2909 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2910 OBD_ALLOC(lumk, lum_size);
2914 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2916 &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2918 lmm_objects = &(lumk->lmm_objects[0]);
2919 lmm_objects->l_ost_oi = lsm->lsm_oi;
2921 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2925 lumk->lmm_oi = lsm->lsm_oi;
2926 lumk->lmm_stripe_count = 1;
2928 if (copy_to_user(lump, lumk, lum_size))
2932 OBD_FREE(lumk, lum_size);
2938 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2939 void *karg, void *uarg)
2941 struct obd_device *obd = exp->exp_obd;
2942 struct obd_ioctl_data *data = karg;
2946 if (!cfs_try_module_get(THIS_MODULE)) {
2947 CERROR("Can't get module. Is it alive?");
2951 case OBD_IOC_LOV_GET_CONFIG: {
2953 struct lov_desc *desc;
2954 struct obd_uuid uuid;
2958 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2959 GOTO(out, err = -EINVAL);
2961 data = (struct obd_ioctl_data *)buf;
2963 if (sizeof(*desc) > data->ioc_inllen1) {
2964 obd_ioctl_freedata(buf, len);
2965 GOTO(out, err = -EINVAL);
2968 if (data->ioc_inllen2 < sizeof(uuid)) {
2969 obd_ioctl_freedata(buf, len);
2970 GOTO(out, err = -EINVAL);
2973 desc = (struct lov_desc *)data->ioc_inlbuf1;
2974 desc->ld_tgt_count = 1;
2975 desc->ld_active_tgt_count = 1;
2976 desc->ld_default_stripe_count = 1;
2977 desc->ld_default_stripe_size = 0;
2978 desc->ld_default_stripe_offset = 0;
2979 desc->ld_pattern = 0;
2980 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2982 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2984 err = copy_to_user((void *)uarg, buf, len);
2987 obd_ioctl_freedata(buf, len);
2990 case LL_IOC_LOV_SETSTRIPE:
2991 err = obd_alloc_memmd(exp, karg);
2995 case LL_IOC_LOV_GETSTRIPE:
2996 err = osc_getstripe(karg, uarg);
2998 case OBD_IOC_CLIENT_RECOVER:
2999 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3000 data->ioc_inlbuf1, 0);
3004 case IOC_OSC_SET_ACTIVE:
3005 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3008 case OBD_IOC_POLL_QUOTACHECK:
3009 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
3011 case OBD_IOC_PING_TARGET:
3012 err = ptlrpc_obd_ping(obd);
3015 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3016 cmd, cfs_curproc_comm());
3017 GOTO(out, err = -ENOTTY);
3020 cfs_module_put(THIS_MODULE);
3024 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3025 obd_count keylen, void *key, __u32 *vallen, void *val,
3026 struct lov_stripe_md *lsm)
3029 if (!vallen || !val)
3032 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3033 __u32 *stripe = val;
3034 *vallen = sizeof(*stripe);
3037 } else if (KEY_IS(KEY_LAST_ID)) {
3038 struct ptlrpc_request *req;
3043 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3044 &RQF_OST_GET_INFO_LAST_ID);
3048 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3049 RCL_CLIENT, keylen);
3050 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3052 ptlrpc_request_free(req);
3056 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3057 memcpy(tmp, key, keylen);
3059 req->rq_no_delay = req->rq_no_resend = 1;
3060 ptlrpc_request_set_replen(req);
3061 rc = ptlrpc_queue_wait(req);
3065 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3067 GOTO(out, rc = -EPROTO);
3069 *((obd_id *)val) = *reply;
3071 ptlrpc_req_finished(req);
3073 } else if (KEY_IS(KEY_FIEMAP)) {
3074 struct ll_fiemap_info_key *fm_key =
3075 (struct ll_fiemap_info_key *)key;
3076 struct ldlm_res_id res_id;
3077 ldlm_policy_data_t policy;
3078 struct lustre_handle lockh;
3079 ldlm_mode_t mode = 0;
3080 struct ptlrpc_request *req;
3081 struct ll_user_fiemap *reply;
3085 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
3088 policy.l_extent.start = fm_key->fiemap.fm_start &
3091 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
3092 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
3093 policy.l_extent.end = OBD_OBJECT_EOF;
3095 policy.l_extent.end = (fm_key->fiemap.fm_start +
3096 fm_key->fiemap.fm_length +
3097 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
3099 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
3100 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
3101 LDLM_FL_BLOCK_GRANTED |
3103 &res_id, LDLM_EXTENT, &policy,
3104 LCK_PR | LCK_PW, &lockh, 0);
3105 if (mode) { /* lock is cached on client */
3106 if (mode != LCK_PR) {
3107 ldlm_lock_addref(&lockh, LCK_PR);
3108 ldlm_lock_decref(&lockh, LCK_PW);
3110 } else { /* no cached lock, needs acquire lock on server side */
3111 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
3112 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
3116 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3117 &RQF_OST_GET_INFO_FIEMAP);
3119 GOTO(drop_lock, rc = -ENOMEM);
3121 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3122 RCL_CLIENT, keylen);
3123 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3124 RCL_CLIENT, *vallen);
3125 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3126 RCL_SERVER, *vallen);
3128 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3130 ptlrpc_request_free(req);
3131 GOTO(drop_lock, rc);
3134 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3135 memcpy(tmp, key, keylen);
3136 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3137 memcpy(tmp, val, *vallen);
3139 ptlrpc_request_set_replen(req);
3140 rc = ptlrpc_queue_wait(req);
3144 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3146 GOTO(fini_req, rc = -EPROTO);
3148 memcpy(val, reply, *vallen);
3150 ptlrpc_req_finished(req);
3153 ldlm_lock_decref(&lockh, LCK_PR);
3160 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3161 obd_count keylen, void *key, obd_count vallen,
3162 void *val, struct ptlrpc_request_set *set)
3164 struct ptlrpc_request *req;
3165 struct obd_device *obd = exp->exp_obd;
3166 struct obd_import *imp = class_exp2cliimp(exp);
3171 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3173 if (KEY_IS(KEY_CHECKSUM)) {
3174 if (vallen != sizeof(int))
3176 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3180 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3181 sptlrpc_conf_client_adapt(obd);
3185 if (KEY_IS(KEY_FLUSH_CTX)) {
3186 sptlrpc_import_flush_my_ctx(imp);
3190 if (KEY_IS(KEY_CACHE_SET)) {
3191 struct client_obd *cli = &obd->u.cli;
3193 LASSERT(cli->cl_cache == NULL); /* only once */
3194 cli->cl_cache = (struct cl_client_cache *)val;
3195 cfs_atomic_inc(&cli->cl_cache->ccc_users);
3196 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
3198 /* add this osc into entity list */
3199 LASSERT(cfs_list_empty(&cli->cl_lru_osc));
3200 spin_lock(&cli->cl_cache->ccc_lru_lock);
3201 cfs_list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
3202 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3207 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3208 struct client_obd *cli = &obd->u.cli;
3209 int nr = cfs_atomic_read(&cli->cl_lru_in_list) >> 1;
3210 int target = *(int *)val;
3212 nr = osc_lru_shrink(cli, min(nr, target));
3217 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3220 /* We pass all other commands directly to OST. Since nobody calls osc
3221 methods directly and everybody is supposed to go through LOV, we
3222 assume lov checked invalid values for us.
3223 The only recognised values so far are evict_by_nid and mds_conn.
3224 Even if something bad goes through, we'd get a -EINVAL from OST
3227 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3228 &RQF_OST_SET_GRANT_INFO :
3233 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3234 RCL_CLIENT, keylen);
3235 if (!KEY_IS(KEY_GRANT_SHRINK))
3236 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3237 RCL_CLIENT, vallen);
3238 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3240 ptlrpc_request_free(req);
3244 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3245 memcpy(tmp, key, keylen);
3246 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3249 memcpy(tmp, val, vallen);
3251 if (KEY_IS(KEY_GRANT_SHRINK)) {
3252 struct osc_grant_args *aa;
3255 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3256 aa = ptlrpc_req_async_args(req);
3259 ptlrpc_req_finished(req);
3262 *oa = ((struct ost_body *)val)->oa;
3264 req->rq_interpret_reply = osc_shrink_grant_interpret;
3267 ptlrpc_request_set_replen(req);
3268 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3269 LASSERT(set != NULL);
3270 ptlrpc_set_add_req(set, req);
3271 ptlrpc_check_set(NULL, set);
3273 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3279 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3280 struct obd_device *disk_obd, int *index)
3282 /* this code is not supposed to be used with LOD/OSP
3283 * to be removed soon */
3288 static int osc_llog_finish(struct obd_device *obd, int count)
3290 struct llog_ctxt *ctxt;
3294 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3296 llog_cat_close(NULL, ctxt->loc_handle);
3297 llog_cleanup(NULL, ctxt);
3300 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3302 llog_cleanup(NULL, ctxt);
3306 static int osc_reconnect(const struct lu_env *env,
3307 struct obd_export *exp, struct obd_device *obd,
3308 struct obd_uuid *cluuid,
3309 struct obd_connect_data *data,
3312 struct client_obd *cli = &obd->u.cli;
3314 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3317 client_obd_list_lock(&cli->cl_loi_list_lock);
3318 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3319 2 * cli_brw_size(obd);
3320 lost_grant = cli->cl_lost_grant;
3321 cli->cl_lost_grant = 0;
3322 client_obd_list_unlock(&cli->cl_loi_list_lock);
3324 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3325 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3326 data->ocd_version, data->ocd_grant, lost_grant);
3332 static int osc_disconnect(struct obd_export *exp)
3334 struct obd_device *obd = class_exp2obd(exp);
3335 struct llog_ctxt *ctxt;
3338 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3340 if (obd->u.cli.cl_conn_count == 1) {
3341 /* Flush any remaining cancel messages out to the
3343 llog_sync(ctxt, exp, 0);
3345 llog_ctxt_put(ctxt);
3347 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3351 rc = client_disconnect_export(exp);
3353 * Initially we put del_shrink_grant before disconnect_export, but it
3354 * causes the following problem if setup (connect) and cleanup
3355 * (disconnect) are tangled together.
3356 * connect p1 disconnect p2
3357 * ptlrpc_connect_import
3358 * ............... class_manual_cleanup
3361 * ptlrpc_connect_interrupt
3363 * add this client to shrink list
3365 * Bang! pinger trigger the shrink.
3366 * So the osc should be disconnected from the shrink list, after we
3367 * are sure the import has been destroyed. BUG18662
3369 if (obd->u.cli.cl_import == NULL)
3370 osc_del_shrink_grant(&obd->u.cli);
3374 static int osc_import_event(struct obd_device *obd,
3375 struct obd_import *imp,
3376 enum obd_import_event event)
3378 struct client_obd *cli;
3382 LASSERT(imp->imp_obd == obd);
3385 case IMP_EVENT_DISCON: {
3387 client_obd_list_lock(&cli->cl_loi_list_lock);
3388 cli->cl_avail_grant = 0;
3389 cli->cl_lost_grant = 0;
3390 client_obd_list_unlock(&cli->cl_loi_list_lock);
3393 case IMP_EVENT_INACTIVE: {
3394 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3397 case IMP_EVENT_INVALIDATE: {
3398 struct ldlm_namespace *ns = obd->obd_namespace;
3402 env = cl_env_get(&refcheck);
3406 /* all pages go to failing rpcs due to the invalid
3408 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3410 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3411 cl_env_put(env, &refcheck);
3416 case IMP_EVENT_ACTIVE: {
3417 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3420 case IMP_EVENT_OCD: {
3421 struct obd_connect_data *ocd = &imp->imp_connect_data;
3423 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3424 osc_init_grant(&obd->u.cli, ocd);
3427 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3428 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3430 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3433 case IMP_EVENT_DEACTIVATE: {
3434 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3437 case IMP_EVENT_ACTIVATE: {
3438 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3442 CERROR("Unknown import event %d\n", event);
3449 * Determine whether the lock can be canceled before replaying the lock
3450 * during recovery, see bug16774 for detailed information.
3452 * \retval zero the lock can't be canceled
3453 * \retval other ok to cancel
3455 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3457 check_res_locked(lock->l_resource);
3460 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3462 * XXX as a future improvement, we can also cancel unused write lock
3463 * if it doesn't have dirty data and active mmaps.
3465 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3466 (lock->l_granted_mode == LCK_PR ||
3467 lock->l_granted_mode == LCK_CR) &&
3468 (osc_dlm_lock_pageref(lock) == 0))
3474 static int brw_queue_work(const struct lu_env *env, void *data)
3476 struct client_obd *cli = data;
3478 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3480 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3484 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3486 struct lprocfs_static_vars lvars = { 0 };
3487 struct client_obd *cli = &obd->u.cli;
3492 rc = ptlrpcd_addref();
3496 rc = client_obd_setup(obd, lcfg);
3498 GOTO(out_ptlrpcd, rc);
3500 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3501 if (IS_ERR(handler))
3502 GOTO(out_client_setup, rc = PTR_ERR(handler));
3503 cli->cl_writeback_work = handler;
3505 rc = osc_quota_setup(obd);
3507 GOTO(out_ptlrpcd_work, rc);
3509 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3510 lprocfs_osc_init_vars(&lvars);
3511 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3512 lproc_osc_attach_seqstat(obd);
3513 sptlrpc_lprocfs_cliobd_attach(obd);
3514 ptlrpc_lprocfs_register_obd(obd);
3517 /* We need to allocate a few requests more, because
3518 * brw_interpret tries to create new requests before freeing
3519 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3520 * reserved, but I'm afraid that might be too much wasted RAM
3521 * in fact, so 2 is just my guess and still should work. */
3522 cli->cl_import->imp_rq_pool =
3523 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3525 ptlrpc_add_rqs_to_pool);
3527 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3528 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3532 ptlrpcd_destroy_work(handler);
3534 client_obd_cleanup(obd);
3540 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3546 case OBD_CLEANUP_EARLY: {
3547 struct obd_import *imp;
3548 imp = obd->u.cli.cl_import;
3549 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3550 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3551 ptlrpc_deactivate_import(imp);
3552 spin_lock(&imp->imp_lock);
3553 imp->imp_pingable = 0;
3554 spin_unlock(&imp->imp_lock);
3557 case OBD_CLEANUP_EXPORTS: {
3558 struct client_obd *cli = &obd->u.cli;
3560 * for echo client, export may be on zombie list, wait for
3561 * zombie thread to cull it, because cli.cl_import will be
3562 * cleared in client_disconnect_export():
3563 * class_export_destroy() -> obd_cleanup() ->
3564 * echo_device_free() -> echo_client_cleanup() ->
3565 * obd_disconnect() -> osc_disconnect() ->
3566 * client_disconnect_export()
3568 obd_zombie_barrier();
3569 if (cli->cl_writeback_work) {
3570 ptlrpcd_destroy_work(cli->cl_writeback_work);
3571 cli->cl_writeback_work = NULL;
3573 obd_cleanup_client_import(obd);
3574 ptlrpc_lprocfs_unregister_obd(obd);
3575 lprocfs_obd_cleanup(obd);
3576 rc = obd_llog_finish(obd, 0);
3578 CERROR("failed to cleanup llogging subsystems\n");
3585 int osc_cleanup(struct obd_device *obd)
3587 struct client_obd *cli = &obd->u.cli;
3593 if (cli->cl_cache != NULL) {
3594 LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_users) > 0);
3595 spin_lock(&cli->cl_cache->ccc_lru_lock);
3596 cfs_list_del_init(&cli->cl_lru_osc);
3597 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3598 cli->cl_lru_left = NULL;
3599 cfs_atomic_dec(&cli->cl_cache->ccc_users);
3600 cli->cl_cache = NULL;
3603 /* free memory of osc quota cache */
3604 osc_quota_cleanup(obd);
3606 rc = client_obd_cleanup(obd);
3612 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3614 struct lprocfs_static_vars lvars = { 0 };
3617 lprocfs_osc_init_vars(&lvars);
3619 switch (lcfg->lcfg_command) {
3621 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3631 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3633 return osc_process_config_base(obd, buf);
3636 struct obd_ops osc_obd_ops = {
3637 .o_owner = THIS_MODULE,
3638 .o_setup = osc_setup,
3639 .o_precleanup = osc_precleanup,
3640 .o_cleanup = osc_cleanup,
3641 .o_add_conn = client_import_add_conn,
3642 .o_del_conn = client_import_del_conn,
3643 .o_connect = client_connect_import,
3644 .o_reconnect = osc_reconnect,
3645 .o_disconnect = osc_disconnect,
3646 .o_statfs = osc_statfs,
3647 .o_statfs_async = osc_statfs_async,
3648 .o_packmd = osc_packmd,
3649 .o_unpackmd = osc_unpackmd,
3650 .o_create = osc_create,
3651 .o_destroy = osc_destroy,
3652 .o_getattr = osc_getattr,
3653 .o_getattr_async = osc_getattr_async,
3654 .o_setattr = osc_setattr,
3655 .o_setattr_async = osc_setattr_async,
3657 .o_punch = osc_punch,
3659 .o_enqueue = osc_enqueue,
3660 .o_change_cbdata = osc_change_cbdata,
3661 .o_find_cbdata = osc_find_cbdata,
3662 .o_cancel = osc_cancel,
3663 .o_cancel_unused = osc_cancel_unused,
3664 .o_iocontrol = osc_iocontrol,
3665 .o_get_info = osc_get_info,
3666 .o_set_info_async = osc_set_info_async,
3667 .o_import_event = osc_import_event,
3668 .o_llog_init = osc_llog_init,
3669 .o_llog_finish = osc_llog_finish,
3670 .o_process_config = osc_process_config,
3671 .o_quotactl = osc_quotactl,
3672 .o_quotacheck = osc_quotacheck,
3675 extern struct lu_kmem_descr osc_caches[];
3676 extern spinlock_t osc_ast_guard;
3677 extern struct lock_class_key osc_ast_guard_class;
3679 int __init osc_init(void)
3681 struct lprocfs_static_vars lvars = { 0 };
3685 /* print an address of _any_ initialized kernel symbol from this
3686 * module, to allow debugging with gdb that doesn't support data
3687 * symbols from modules.*/
3688 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3690 rc = lu_kmem_init(osc_caches);
3694 lprocfs_osc_init_vars(&lvars);
3696 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3697 LUSTRE_OSC_NAME, &osc_device_type);
3699 lu_kmem_fini(osc_caches);
3703 spin_lock_init(&osc_ast_guard);
3704 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3710 static void /*__exit*/ osc_exit(void)
3712 class_unregister_type(LUSTRE_OSC_NAME);
3713 lu_kmem_fini(osc_caches);
3716 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3717 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3718 MODULE_LICENSE("GPL");
3720 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);