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, __u64 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 wake_up(&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 wake_up(&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_unstable_pages) +
842 cfs_atomic_read(&obd_dirty_pages) -
843 cfs_atomic_read(&obd_dirty_transit_pages) >
844 (long)(obd_max_dirty_pages + 1))) {
845 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
846 * not covered by a lock thus they may safely race and trip
847 * this CERROR() unless we add in a small fudge factor (+1). */
848 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
849 cli->cl_import->imp_obd->obd_name,
850 cfs_atomic_read(&obd_unstable_pages),
851 cfs_atomic_read(&obd_dirty_pages),
852 cfs_atomic_read(&obd_dirty_transit_pages),
853 obd_max_dirty_pages);
855 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
856 CERROR("dirty %lu - dirty_max %lu too big???\n",
857 cli->cl_dirty, cli->cl_dirty_max);
860 long max_in_flight = (cli->cl_max_pages_per_rpc <<
862 (cli->cl_max_rpcs_in_flight + 1);
863 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
865 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
866 oa->o_dropped = cli->cl_lost_grant;
867 cli->cl_lost_grant = 0;
868 client_obd_list_unlock(&cli->cl_loi_list_lock);
869 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
870 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
874 void osc_update_next_shrink(struct client_obd *cli)
876 cli->cl_next_shrink_grant =
877 cfs_time_shift(cli->cl_grant_shrink_interval);
878 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
879 cli->cl_next_shrink_grant);
882 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
884 client_obd_list_lock(&cli->cl_loi_list_lock);
885 cli->cl_avail_grant += grant;
886 client_obd_list_unlock(&cli->cl_loi_list_lock);
889 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
891 if (body->oa.o_valid & OBD_MD_FLGRANT) {
892 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
893 __osc_update_grant(cli, body->oa.o_grant);
897 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
898 obd_count keylen, void *key, obd_count vallen,
899 void *val, struct ptlrpc_request_set *set);
901 static int osc_shrink_grant_interpret(const struct lu_env *env,
902 struct ptlrpc_request *req,
905 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
906 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
907 struct ost_body *body;
910 __osc_update_grant(cli, oa->o_grant);
914 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
916 osc_update_grant(cli, body);
922 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
924 client_obd_list_lock(&cli->cl_loi_list_lock);
925 oa->o_grant = cli->cl_avail_grant / 4;
926 cli->cl_avail_grant -= oa->o_grant;
927 client_obd_list_unlock(&cli->cl_loi_list_lock);
928 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
929 oa->o_valid |= OBD_MD_FLFLAGS;
932 oa->o_flags |= OBD_FL_SHRINK_GRANT;
933 osc_update_next_shrink(cli);
936 /* Shrink the current grant, either from some large amount to enough for a
937 * full set of in-flight RPCs, or if we have already shrunk to that limit
938 * then to enough for a single RPC. This avoids keeping more grant than
939 * needed, and avoids shrinking the grant piecemeal. */
940 static int osc_shrink_grant(struct client_obd *cli)
942 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
943 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
945 client_obd_list_lock(&cli->cl_loi_list_lock);
946 if (cli->cl_avail_grant <= target_bytes)
947 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
948 client_obd_list_unlock(&cli->cl_loi_list_lock);
950 return osc_shrink_grant_to_target(cli, target_bytes);
953 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
956 struct ost_body *body;
959 client_obd_list_lock(&cli->cl_loi_list_lock);
960 /* Don't shrink if we are already above or below the desired limit
961 * We don't want to shrink below a single RPC, as that will negatively
962 * impact block allocation and long-term performance. */
963 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
964 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
966 if (target_bytes >= cli->cl_avail_grant) {
967 client_obd_list_unlock(&cli->cl_loi_list_lock);
970 client_obd_list_unlock(&cli->cl_loi_list_lock);
976 osc_announce_cached(cli, &body->oa, 0);
978 client_obd_list_lock(&cli->cl_loi_list_lock);
979 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
980 cli->cl_avail_grant = target_bytes;
981 client_obd_list_unlock(&cli->cl_loi_list_lock);
982 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
983 body->oa.o_valid |= OBD_MD_FLFLAGS;
984 body->oa.o_flags = 0;
986 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
987 osc_update_next_shrink(cli);
989 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
990 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
991 sizeof(*body), body, NULL);
993 __osc_update_grant(cli, body->oa.o_grant);
998 static int osc_should_shrink_grant(struct client_obd *client)
1000 cfs_time_t time = cfs_time_current();
1001 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1003 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1004 OBD_CONNECT_GRANT_SHRINK) == 0)
1007 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1008 /* Get the current RPC size directly, instead of going via:
1009 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
1010 * Keep comment here so that it can be found by searching. */
1011 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
1013 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1014 client->cl_avail_grant > brw_size)
1017 osc_update_next_shrink(client);
1022 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1024 struct client_obd *client;
1026 cfs_list_for_each_entry(client, &item->ti_obd_list,
1027 cl_grant_shrink_list) {
1028 if (osc_should_shrink_grant(client))
1029 osc_shrink_grant(client);
1034 static int osc_add_shrink_grant(struct client_obd *client)
1038 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1040 osc_grant_shrink_grant_cb, NULL,
1041 &client->cl_grant_shrink_list);
1043 CERROR("add grant client %s error %d\n",
1044 client->cl_import->imp_obd->obd_name, rc);
1047 CDEBUG(D_CACHE, "add grant client %s \n",
1048 client->cl_import->imp_obd->obd_name);
1049 osc_update_next_shrink(client);
1053 static int osc_del_shrink_grant(struct client_obd *client)
1055 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1059 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1062 * ocd_grant is the total grant amount we're expect to hold: if we've
1063 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1064 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1066 * race is tolerable here: if we're evicted, but imp_state already
1067 * left EVICTED state, then cl_dirty must be 0 already.
1069 client_obd_list_lock(&cli->cl_loi_list_lock);
1070 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1071 cli->cl_avail_grant = ocd->ocd_grant;
1073 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1075 if (cli->cl_avail_grant < 0) {
1076 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1077 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1078 ocd->ocd_grant, cli->cl_dirty);
1079 /* workaround for servers which do not have the patch from
1081 cli->cl_avail_grant = ocd->ocd_grant;
1084 /* determine the appropriate chunk size used by osc_extent. */
1085 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1086 client_obd_list_unlock(&cli->cl_loi_list_lock);
1088 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1089 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1090 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1092 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1093 cfs_list_empty(&cli->cl_grant_shrink_list))
1094 osc_add_shrink_grant(cli);
1097 /* We assume that the reason this OSC got a short read is because it read
1098 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1099 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1100 * this stripe never got written at or beyond this stripe offset yet. */
1101 static void handle_short_read(int nob_read, obd_count page_count,
1102 struct brw_page **pga)
1107 /* skip bytes read OK */
1108 while (nob_read > 0) {
1109 LASSERT (page_count > 0);
1111 if (pga[i]->count > nob_read) {
1112 /* EOF inside this page */
1113 ptr = kmap(pga[i]->pg) +
1114 (pga[i]->off & ~CFS_PAGE_MASK);
1115 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1122 nob_read -= pga[i]->count;
1127 /* zero remaining pages */
1128 while (page_count-- > 0) {
1129 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1130 memset(ptr, 0, pga[i]->count);
1136 static int check_write_rcs(struct ptlrpc_request *req,
1137 int requested_nob, int niocount,
1138 obd_count page_count, struct brw_page **pga)
1143 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1144 sizeof(*remote_rcs) *
1146 if (remote_rcs == NULL) {
1147 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1151 /* return error if any niobuf was in error */
1152 for (i = 0; i < niocount; i++) {
1153 if ((int)remote_rcs[i] < 0)
1154 return(remote_rcs[i]);
1156 if (remote_rcs[i] != 0) {
1157 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1158 i, remote_rcs[i], req);
1163 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1164 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1165 req->rq_bulk->bd_nob_transferred, requested_nob);
1172 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1174 if (p1->flag != p2->flag) {
1175 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1176 OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1178 /* warn if we try to combine flags that we don't know to be
1179 * safe to combine */
1180 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1181 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1182 "report this at http://bugs.whamcloud.com/\n",
1183 p1->flag, p2->flag);
1188 return (p1->off + p1->count == p2->off);
1191 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1192 struct brw_page **pga, int opc,
1193 cksum_type_t cksum_type)
1197 struct cfs_crypto_hash_desc *hdesc;
1198 unsigned int bufsize;
1200 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1202 LASSERT(pg_count > 0);
1204 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1205 if (IS_ERR(hdesc)) {
1206 CERROR("Unable to initialize checksum hash %s\n",
1207 cfs_crypto_hash_name(cfs_alg));
1208 return PTR_ERR(hdesc);
1211 while (nob > 0 && pg_count > 0) {
1212 int count = pga[i]->count > nob ? nob : pga[i]->count;
1214 /* corrupt the data before we compute the checksum, to
1215 * simulate an OST->client data error */
1216 if (i == 0 && opc == OST_READ &&
1217 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1218 unsigned char *ptr = kmap(pga[i]->pg);
1219 int off = pga[i]->off & ~CFS_PAGE_MASK;
1220 memcpy(ptr + off, "bad1", min(4, nob));
1223 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1224 pga[i]->off & ~CFS_PAGE_MASK,
1226 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1227 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1229 nob -= pga[i]->count;
1235 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1238 cfs_crypto_hash_final(hdesc, NULL, NULL);
1240 /* For sending we only compute the wrong checksum instead
1241 * of corrupting the data so it is still correct on a redo */
1242 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1248 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1249 struct lov_stripe_md *lsm, obd_count page_count,
1250 struct brw_page **pga,
1251 struct ptlrpc_request **reqp,
1252 struct obd_capa *ocapa, int reserve,
1255 struct ptlrpc_request *req;
1256 struct ptlrpc_bulk_desc *desc;
1257 struct ost_body *body;
1258 struct obd_ioobj *ioobj;
1259 struct niobuf_remote *niobuf;
1260 int niocount, i, requested_nob, opc, rc;
1261 struct osc_brw_async_args *aa;
1262 struct req_capsule *pill;
1263 struct brw_page *pg_prev;
1266 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1267 RETURN(-ENOMEM); /* Recoverable */
1268 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1269 RETURN(-EINVAL); /* Fatal */
1271 if ((cmd & OBD_BRW_WRITE) != 0) {
1273 req = ptlrpc_request_alloc_pool(cli->cl_import,
1274 cli->cl_import->imp_rq_pool,
1275 &RQF_OST_BRW_WRITE);
1278 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1283 for (niocount = i = 1; i < page_count; i++) {
1284 if (!can_merge_pages(pga[i - 1], pga[i]))
1288 pill = &req->rq_pill;
1289 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1291 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1292 niocount * sizeof(*niobuf));
1293 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1295 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1297 ptlrpc_request_free(req);
1300 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1301 ptlrpc_at_set_req_timeout(req);
1302 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1304 req->rq_no_retry_einprogress = 1;
1306 desc = ptlrpc_prep_bulk_imp(req, page_count,
1307 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1308 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1312 GOTO(out, rc = -ENOMEM);
1313 /* NB request now owns desc and will free it when it gets freed */
1315 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1316 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1317 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1318 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1320 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1322 obdo_to_ioobj(oa, ioobj);
1323 ioobj->ioo_bufcnt = niocount;
1324 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1325 * that might be send for this request. The actual number is decided
1326 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1327 * "max - 1" for old client compatibility sending "0", and also so the
1328 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1329 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1330 osc_pack_capa(req, body, ocapa);
1331 LASSERT(page_count > 0);
1333 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1334 struct brw_page *pg = pga[i];
1335 int poff = pg->off & ~CFS_PAGE_MASK;
1337 LASSERT(pg->count > 0);
1338 /* make sure there is no gap in the middle of page array */
1339 LASSERTF(page_count == 1 ||
1340 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1341 ergo(i > 0 && i < page_count - 1,
1342 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1343 ergo(i == page_count - 1, poff == 0)),
1344 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1345 i, page_count, pg, pg->off, pg->count);
1347 LASSERTF(i == 0 || pg->off > pg_prev->off,
1348 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1349 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1351 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1352 pg_prev->pg, page_private(pg_prev->pg),
1353 pg_prev->pg->index, pg_prev->off);
1355 LASSERTF(i == 0 || pg->off > pg_prev->off,
1356 "i %d p_c %u\n", i, page_count);
1358 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1359 (pg->flag & OBD_BRW_SRVLOCK));
1361 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1362 requested_nob += pg->count;
1364 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1366 niobuf->len += pg->count;
1368 niobuf->offset = pg->off;
1369 niobuf->len = pg->count;
1370 niobuf->flags = pg->flag;
1375 LASSERTF((void *)(niobuf - niocount) ==
1376 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1377 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1378 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1380 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1382 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1383 body->oa.o_valid |= OBD_MD_FLFLAGS;
1384 body->oa.o_flags = 0;
1386 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1389 if (osc_should_shrink_grant(cli))
1390 osc_shrink_grant_local(cli, &body->oa);
1392 /* size[REQ_REC_OFF] still sizeof (*body) */
1393 if (opc == OST_WRITE) {
1394 if (cli->cl_checksum &&
1395 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1396 /* store cl_cksum_type in a local variable since
1397 * it can be changed via lprocfs */
1398 cksum_type_t cksum_type = cli->cl_cksum_type;
1400 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1401 oa->o_flags &= OBD_FL_LOCAL_MASK;
1402 body->oa.o_flags = 0;
1404 body->oa.o_flags |= cksum_type_pack(cksum_type);
1405 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1406 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1410 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1412 /* save this in 'oa', too, for later checking */
1413 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1414 oa->o_flags |= cksum_type_pack(cksum_type);
1416 /* clear out the checksum flag, in case this is a
1417 * resend but cl_checksum is no longer set. b=11238 */
1418 oa->o_valid &= ~OBD_MD_FLCKSUM;
1420 oa->o_cksum = body->oa.o_cksum;
1421 /* 1 RC per niobuf */
1422 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1423 sizeof(__u32) * niocount);
1425 if (cli->cl_checksum &&
1426 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1427 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1428 body->oa.o_flags = 0;
1429 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1430 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1433 ptlrpc_request_set_replen(req);
1435 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1436 aa = ptlrpc_req_async_args(req);
1438 aa->aa_requested_nob = requested_nob;
1439 aa->aa_nio_count = niocount;
1440 aa->aa_page_count = page_count;
1444 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1445 if (ocapa && reserve)
1446 aa->aa_ocapa = capa_get(ocapa);
1452 ptlrpc_req_finished(req);
1456 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1457 __u32 client_cksum, __u32 server_cksum, int nob,
1458 obd_count page_count, struct brw_page **pga,
1459 cksum_type_t client_cksum_type)
1463 cksum_type_t cksum_type;
1465 if (server_cksum == client_cksum) {
1466 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1470 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1472 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1475 if (cksum_type != client_cksum_type)
1476 msg = "the server did not use the checksum type specified in "
1477 "the original request - likely a protocol problem";
1478 else if (new_cksum == server_cksum)
1479 msg = "changed on the client after we checksummed it - "
1480 "likely false positive due to mmap IO (bug 11742)";
1481 else if (new_cksum == client_cksum)
1482 msg = "changed in transit before arrival at OST";
1484 msg = "changed in transit AND doesn't match the original - "
1485 "likely false positive due to mmap IO (bug 11742)";
1487 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1488 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1489 msg, libcfs_nid2str(peer->nid),
1490 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1491 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1492 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1493 POSTID(&oa->o_oi), pga[0]->off,
1494 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1495 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1496 "client csum now %x\n", client_cksum, client_cksum_type,
1497 server_cksum, cksum_type, new_cksum);
1501 /* Note rc enters this function as number of bytes transferred */
1502 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1504 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1505 const lnet_process_id_t *peer =
1506 &req->rq_import->imp_connection->c_peer;
1507 struct client_obd *cli = aa->aa_cli;
1508 struct ost_body *body;
1509 __u32 client_cksum = 0;
1512 if (rc < 0 && rc != -EDQUOT) {
1513 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1517 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1518 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1520 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1524 /* set/clear over quota flag for a uid/gid */
1525 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1526 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1527 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1529 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1530 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1532 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1535 osc_update_grant(cli, body);
1540 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1541 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1543 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1545 CERROR("Unexpected +ve rc %d\n", rc);
1548 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1550 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1553 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1554 check_write_checksum(&body->oa, peer, client_cksum,
1555 body->oa.o_cksum, aa->aa_requested_nob,
1556 aa->aa_page_count, aa->aa_ppga,
1557 cksum_type_unpack(aa->aa_oa->o_flags)))
1560 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1561 aa->aa_page_count, aa->aa_ppga);
1565 /* The rest of this function executes only for OST_READs */
1567 /* if unwrap_bulk failed, return -EAGAIN to retry */
1568 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1570 GOTO(out, rc = -EAGAIN);
1572 if (rc > aa->aa_requested_nob) {
1573 CERROR("Unexpected rc %d (%d requested)\n", rc,
1574 aa->aa_requested_nob);
1578 if (rc != req->rq_bulk->bd_nob_transferred) {
1579 CERROR ("Unexpected rc %d (%d transferred)\n",
1580 rc, req->rq_bulk->bd_nob_transferred);
1584 if (rc < aa->aa_requested_nob)
1585 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1587 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1588 static int cksum_counter;
1589 __u32 server_cksum = body->oa.o_cksum;
1592 cksum_type_t cksum_type;
1594 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1595 body->oa.o_flags : 0);
1596 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1597 aa->aa_ppga, OST_READ,
1600 if (peer->nid == req->rq_bulk->bd_sender) {
1604 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1607 if (server_cksum == ~0 && rc > 0) {
1608 CERROR("Protocol error: server %s set the 'checksum' "
1609 "bit, but didn't send a checksum. Not fatal, "
1610 "but please notify on http://bugs.whamcloud.com/\n",
1611 libcfs_nid2str(peer->nid));
1612 } else if (server_cksum != client_cksum) {
1613 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1614 "%s%s%s inode "DFID" object "DOSTID
1615 " extent ["LPU64"-"LPU64"]\n",
1616 req->rq_import->imp_obd->obd_name,
1617 libcfs_nid2str(peer->nid),
1619 body->oa.o_valid & OBD_MD_FLFID ?
1620 body->oa.o_parent_seq : (__u64)0,
1621 body->oa.o_valid & OBD_MD_FLFID ?
1622 body->oa.o_parent_oid : 0,
1623 body->oa.o_valid & OBD_MD_FLFID ?
1624 body->oa.o_parent_ver : 0,
1625 POSTID(&body->oa.o_oi),
1626 aa->aa_ppga[0]->off,
1627 aa->aa_ppga[aa->aa_page_count-1]->off +
1628 aa->aa_ppga[aa->aa_page_count-1]->count -
1630 CERROR("client %x, server %x, cksum_type %x\n",
1631 client_cksum, server_cksum, cksum_type);
1633 aa->aa_oa->o_cksum = client_cksum;
1637 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1640 } else if (unlikely(client_cksum)) {
1641 static int cksum_missed;
1644 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1645 CERROR("Checksum %u requested from %s but not sent\n",
1646 cksum_missed, libcfs_nid2str(peer->nid));
1652 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1653 aa->aa_oa, &body->oa);
1658 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1659 struct lov_stripe_md *lsm,
1660 obd_count page_count, struct brw_page **pga,
1661 struct obd_capa *ocapa)
1663 struct ptlrpc_request *req;
1665 wait_queue_head_t waitq;
1666 int generation, resends = 0;
1667 struct l_wait_info lwi;
1671 init_waitqueue_head(&waitq);
1672 generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1675 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1676 page_count, pga, &req, ocapa, 0, resends);
1681 req->rq_generation_set = 1;
1682 req->rq_import_generation = generation;
1683 req->rq_sent = cfs_time_current_sec() + resends;
1686 rc = ptlrpc_queue_wait(req);
1688 if (rc == -ETIMEDOUT && req->rq_resend) {
1689 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1690 ptlrpc_req_finished(req);
1694 rc = osc_brw_fini_request(req, rc);
1696 ptlrpc_req_finished(req);
1697 /* When server return -EINPROGRESS, client should always retry
1698 * regardless of the number of times the bulk was resent already.*/
1699 if (osc_recoverable_error(rc)) {
1701 if (rc != -EINPROGRESS &&
1702 !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1703 CERROR("%s: too many resend retries for object: "
1704 ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1705 POSTID(&oa->o_oi), rc);
1709 exp->exp_obd->u.cli.cl_import->imp_generation) {
1710 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1711 ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1712 POSTID(&oa->o_oi), rc);
1716 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1718 l_wait_event(waitq, 0, &lwi);
1723 if (rc == -EAGAIN || rc == -EINPROGRESS)
1728 static int osc_brw_redo_request(struct ptlrpc_request *request,
1729 struct osc_brw_async_args *aa, int rc)
1731 struct ptlrpc_request *new_req;
1732 struct osc_brw_async_args *new_aa;
1733 struct osc_async_page *oap;
1736 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1737 "redo for recoverable error %d", rc);
1739 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1740 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1741 aa->aa_cli, aa->aa_oa,
1742 NULL /* lsm unused by osc currently */,
1743 aa->aa_page_count, aa->aa_ppga,
1744 &new_req, aa->aa_ocapa, 0, 1);
1748 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1749 if (oap->oap_request != NULL) {
1750 LASSERTF(request == oap->oap_request,
1751 "request %p != oap_request %p\n",
1752 request, oap->oap_request);
1753 if (oap->oap_interrupted) {
1754 ptlrpc_req_finished(new_req);
1759 /* New request takes over pga and oaps from old request.
1760 * Note that copying a list_head doesn't work, need to move it... */
1762 new_req->rq_interpret_reply = request->rq_interpret_reply;
1763 new_req->rq_async_args = request->rq_async_args;
1764 new_req->rq_commit_cb = request->rq_commit_cb;
1765 /* cap resend delay to the current request timeout, this is similar to
1766 * what ptlrpc does (see after_reply()) */
1767 if (aa->aa_resends > new_req->rq_timeout)
1768 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1770 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1771 new_req->rq_generation_set = 1;
1772 new_req->rq_import_generation = request->rq_import_generation;
1774 new_aa = ptlrpc_req_async_args(new_req);
1776 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1777 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1778 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1779 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1780 new_aa->aa_resends = aa->aa_resends;
1782 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1783 if (oap->oap_request) {
1784 ptlrpc_req_finished(oap->oap_request);
1785 oap->oap_request = ptlrpc_request_addref(new_req);
1789 new_aa->aa_ocapa = aa->aa_ocapa;
1790 aa->aa_ocapa = NULL;
1792 /* XXX: This code will run into problem if we're going to support
1793 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1794 * and wait for all of them to be finished. We should inherit request
1795 * set from old request. */
1796 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1798 DEBUG_REQ(D_INFO, new_req, "new request");
1803 * ugh, we want disk allocation on the target to happen in offset order. we'll
1804 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1805 * fine for our small page arrays and doesn't require allocation. its an
1806 * insertion sort that swaps elements that are strides apart, shrinking the
1807 * stride down until its '1' and the array is sorted.
1809 static void sort_brw_pages(struct brw_page **array, int num)
1812 struct brw_page *tmp;
1816 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1821 for (i = stride ; i < num ; i++) {
1824 while (j >= stride && array[j - stride]->off > tmp->off) {
1825 array[j] = array[j - stride];
1830 } while (stride > 1);
1833 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1839 LASSERT (pages > 0);
1840 offset = pg[i]->off & ~CFS_PAGE_MASK;
1844 if (pages == 0) /* that's all */
1847 if (offset + pg[i]->count < PAGE_CACHE_SIZE)
1848 return count; /* doesn't end on page boundary */
1851 offset = pg[i]->off & ~CFS_PAGE_MASK;
1852 if (offset != 0) /* doesn't start on page boundary */
1859 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1861 struct brw_page **ppga;
1864 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1868 for (i = 0; i < count; i++)
1873 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1875 LASSERT(ppga != NULL);
1876 OBD_FREE(ppga, sizeof(*ppga) * count);
1879 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1880 obd_count page_count, struct brw_page *pga,
1881 struct obd_trans_info *oti)
1883 struct obdo *saved_oa = NULL;
1884 struct brw_page **ppga, **orig;
1885 struct obd_import *imp = class_exp2cliimp(exp);
1886 struct client_obd *cli;
1887 int rc, page_count_orig;
1890 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1891 cli = &imp->imp_obd->u.cli;
1893 if (cmd & OBD_BRW_CHECK) {
1894 /* The caller just wants to know if there's a chance that this
1895 * I/O can succeed */
1897 if (imp->imp_invalid)
1902 /* test_brw with a failed create can trip this, maybe others. */
1903 LASSERT(cli->cl_max_pages_per_rpc);
1907 orig = ppga = osc_build_ppga(pga, page_count);
1910 page_count_orig = page_count;
1912 sort_brw_pages(ppga, page_count);
1913 while (page_count) {
1914 obd_count pages_per_brw;
1916 if (page_count > cli->cl_max_pages_per_rpc)
1917 pages_per_brw = cli->cl_max_pages_per_rpc;
1919 pages_per_brw = page_count;
1921 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1923 if (saved_oa != NULL) {
1924 /* restore previously saved oa */
1925 *oinfo->oi_oa = *saved_oa;
1926 } else if (page_count > pages_per_brw) {
1927 /* save a copy of oa (brw will clobber it) */
1928 OBDO_ALLOC(saved_oa);
1929 if (saved_oa == NULL)
1930 GOTO(out, rc = -ENOMEM);
1931 *saved_oa = *oinfo->oi_oa;
1934 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1935 pages_per_brw, ppga, oinfo->oi_capa);
1940 page_count -= pages_per_brw;
1941 ppga += pages_per_brw;
1945 osc_release_ppga(orig, page_count_orig);
1947 if (saved_oa != NULL)
1948 OBDO_FREE(saved_oa);
1953 static int brw_interpret(const struct lu_env *env,
1954 struct ptlrpc_request *req, void *data, int rc)
1956 struct osc_brw_async_args *aa = data;
1957 struct osc_extent *ext;
1958 struct osc_extent *tmp;
1959 struct cl_object *obj = NULL;
1960 struct client_obd *cli = aa->aa_cli;
1963 rc = osc_brw_fini_request(req, rc);
1964 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1965 /* When server return -EINPROGRESS, client should always retry
1966 * regardless of the number of times the bulk was resent already. */
1967 if (osc_recoverable_error(rc)) {
1968 if (req->rq_import_generation !=
1969 req->rq_import->imp_generation) {
1970 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1971 ""DOSTID", rc = %d.\n",
1972 req->rq_import->imp_obd->obd_name,
1973 POSTID(&aa->aa_oa->o_oi), rc);
1974 } else if (rc == -EINPROGRESS ||
1975 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1976 rc = osc_brw_redo_request(req, aa, rc);
1978 CERROR("%s: too many resent retries for object: "
1979 ""LPU64":"LPU64", rc = %d.\n",
1980 req->rq_import->imp_obd->obd_name,
1981 POSTID(&aa->aa_oa->o_oi), rc);
1986 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1991 capa_put(aa->aa_ocapa);
1992 aa->aa_ocapa = NULL;
1995 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1996 if (obj == NULL && rc == 0) {
1997 obj = osc2cl(ext->oe_obj);
2001 cfs_list_del_init(&ext->oe_link);
2002 osc_extent_finish(env, ext, 1, rc);
2004 LASSERT(cfs_list_empty(&aa->aa_exts));
2005 LASSERT(cfs_list_empty(&aa->aa_oaps));
2008 struct obdo *oa = aa->aa_oa;
2009 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2010 unsigned long valid = 0;
2013 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2014 attr->cat_blocks = oa->o_blocks;
2015 valid |= CAT_BLOCKS;
2017 if (oa->o_valid & OBD_MD_FLMTIME) {
2018 attr->cat_mtime = oa->o_mtime;
2021 if (oa->o_valid & OBD_MD_FLATIME) {
2022 attr->cat_atime = oa->o_atime;
2025 if (oa->o_valid & OBD_MD_FLCTIME) {
2026 attr->cat_ctime = oa->o_ctime;
2030 cl_object_attr_lock(obj);
2031 cl_object_attr_set(env, obj, attr, valid);
2032 cl_object_attr_unlock(obj);
2034 cl_object_put(env, obj);
2036 OBDO_FREE(aa->aa_oa);
2038 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2039 req->rq_bulk->bd_nob_transferred);
2040 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2041 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2043 client_obd_list_lock(&cli->cl_loi_list_lock);
2044 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2045 * is called so we know whether to go to sync BRWs or wait for more
2046 * RPCs to complete */
2047 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2048 cli->cl_w_in_flight--;
2050 cli->cl_r_in_flight--;
2051 osc_wake_cache_waiters(cli);
2052 client_obd_list_unlock(&cli->cl_loi_list_lock);
2054 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2058 static void brw_commit(struct ptlrpc_request *req)
2060 spin_lock(&req->rq_lock);
2061 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2062 * this called via the rq_commit_cb, I need to ensure
2063 * osc_dec_unstable_pages is still called. Otherwise unstable
2064 * pages may be leaked. */
2065 if (req->rq_unstable) {
2066 spin_unlock(&req->rq_lock);
2067 osc_dec_unstable_pages(req);
2068 spin_lock(&req->rq_lock);
2070 req->rq_committed = 1;
2072 spin_unlock(&req->rq_lock);
2076 * Build an RPC by the list of extent @ext_list. The caller must ensure
2077 * that the total pages in this list are NOT over max pages per RPC.
2078 * Extents in the list must be in OES_RPC state.
2080 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2081 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2083 struct ptlrpc_request *req = NULL;
2084 struct osc_extent *ext;
2085 struct brw_page **pga = NULL;
2086 struct osc_brw_async_args *aa = NULL;
2087 struct obdo *oa = NULL;
2088 struct osc_async_page *oap;
2089 struct osc_async_page *tmp;
2090 struct cl_req *clerq = NULL;
2091 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
2093 struct ldlm_lock *lock = NULL;
2094 struct cl_req_attr *crattr = NULL;
2095 obd_off starting_offset = OBD_OBJECT_EOF;
2096 obd_off ending_offset = 0;
2102 CFS_LIST_HEAD(rpc_list);
2105 LASSERT(!cfs_list_empty(ext_list));
2107 /* add pages into rpc_list to build BRW rpc */
2108 cfs_list_for_each_entry(ext, ext_list, oe_link) {
2109 LASSERT(ext->oe_state == OES_RPC);
2110 mem_tight |= ext->oe_memalloc;
2111 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2113 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2114 if (starting_offset > oap->oap_obj_off)
2115 starting_offset = oap->oap_obj_off;
2117 LASSERT(oap->oap_page_off == 0);
2118 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2119 ending_offset = oap->oap_obj_off +
2122 LASSERT(oap->oap_page_off + oap->oap_count ==
2128 mpflag = cfs_memory_pressure_get_and_set();
2130 OBD_ALLOC(crattr, sizeof(*crattr));
2132 GOTO(out, rc = -ENOMEM);
2134 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2136 GOTO(out, rc = -ENOMEM);
2140 GOTO(out, rc = -ENOMEM);
2143 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2144 struct cl_page *page = oap2cl_page(oap);
2145 if (clerq == NULL) {
2146 clerq = cl_req_alloc(env, page, crt,
2147 1 /* only 1-object rpcs for now */);
2149 GOTO(out, rc = PTR_ERR(clerq));
2150 lock = oap->oap_ldlm_lock;
2153 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2154 pga[i] = &oap->oap_brw_page;
2155 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2156 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2157 pga[i]->pg, page_index(oap->oap_page), oap,
2160 cl_req_page_add(env, clerq, page);
2163 /* always get the data for the obdo for the rpc */
2164 LASSERT(clerq != NULL);
2165 crattr->cra_oa = oa;
2166 cl_req_attr_set(env, clerq, crattr, ~0ULL);
2168 oa->o_handle = lock->l_remote_handle;
2169 oa->o_valid |= OBD_MD_FLHANDLE;
2172 rc = cl_req_prep(env, clerq);
2174 CERROR("cl_req_prep failed: %d\n", rc);
2178 sort_brw_pages(pga, page_count);
2179 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2180 pga, &req, crattr->cra_capa, 1, 0);
2182 CERROR("prep_req failed: %d\n", rc);
2186 req->rq_commit_cb = brw_commit;
2187 req->rq_interpret_reply = brw_interpret;
2190 req->rq_memalloc = 1;
2192 /* Need to update the timestamps after the request is built in case
2193 * we race with setattr (locally or in queue at OST). If OST gets
2194 * later setattr before earlier BRW (as determined by the request xid),
2195 * the OST will not use BRW timestamps. Sadly, there is no obvious
2196 * way to do this in a single call. bug 10150 */
2197 cl_req_attr_set(env, clerq, crattr,
2198 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2200 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2202 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2203 aa = ptlrpc_req_async_args(req);
2204 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2205 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2206 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2207 cfs_list_splice_init(ext_list, &aa->aa_exts);
2208 aa->aa_clerq = clerq;
2210 /* queued sync pages can be torn down while the pages
2211 * were between the pending list and the rpc */
2213 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2214 /* only one oap gets a request reference */
2217 if (oap->oap_interrupted && !req->rq_intr) {
2218 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2220 ptlrpc_mark_interrupted(req);
2224 tmp->oap_request = ptlrpc_request_addref(req);
2226 client_obd_list_lock(&cli->cl_loi_list_lock);
2227 starting_offset >>= PAGE_CACHE_SHIFT;
2228 if (cmd == OBD_BRW_READ) {
2229 cli->cl_r_in_flight++;
2230 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2231 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2232 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2233 starting_offset + 1);
2235 cli->cl_w_in_flight++;
2236 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2237 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2238 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2239 starting_offset + 1);
2241 client_obd_list_unlock(&cli->cl_loi_list_lock);
2243 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2244 page_count, aa, cli->cl_r_in_flight,
2245 cli->cl_w_in_flight);
2247 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2248 * see which CPU/NUMA node the majority of pages were allocated
2249 * on, and try to assign the async RPC to the CPU core
2250 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2252 * But on the other hand, we expect that multiple ptlrpcd
2253 * threads and the initial write sponsor can run in parallel,
2254 * especially when data checksum is enabled, which is CPU-bound
2255 * operation and single ptlrpcd thread cannot process in time.
2256 * So more ptlrpcd threads sharing BRW load
2257 * (with PDL_POLICY_ROUND) seems better.
2259 ptlrpcd_add_req(req, pol, -1);
2265 cfs_memory_pressure_restore(mpflag);
2267 if (crattr != NULL) {
2268 capa_put(crattr->cra_capa);
2269 OBD_FREE(crattr, sizeof(*crattr));
2273 LASSERT(req == NULL);
2278 OBD_FREE(pga, sizeof(*pga) * page_count);
2279 /* this should happen rarely and is pretty bad, it makes the
2280 * pending list not follow the dirty order */
2281 while (!cfs_list_empty(ext_list)) {
2282 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2284 cfs_list_del_init(&ext->oe_link);
2285 osc_extent_finish(env, ext, 0, rc);
2287 if (clerq && !IS_ERR(clerq))
2288 cl_req_completion(env, clerq, rc);
2293 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2294 struct ldlm_enqueue_info *einfo)
2296 void *data = einfo->ei_cbdata;
2299 LASSERT(lock != NULL);
2300 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2301 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2302 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2303 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2305 lock_res_and_lock(lock);
2306 spin_lock(&osc_ast_guard);
2308 if (lock->l_ast_data == NULL)
2309 lock->l_ast_data = data;
2310 if (lock->l_ast_data == data)
2313 spin_unlock(&osc_ast_guard);
2314 unlock_res_and_lock(lock);
2319 static int osc_set_data_with_check(struct lustre_handle *lockh,
2320 struct ldlm_enqueue_info *einfo)
2322 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2326 set = osc_set_lock_data_with_check(lock, einfo);
2327 LDLM_LOCK_PUT(lock);
2329 CERROR("lockh %p, data %p - client evicted?\n",
2330 lockh, einfo->ei_cbdata);
2334 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2335 ldlm_iterator_t replace, void *data)
2337 struct ldlm_res_id res_id;
2338 struct obd_device *obd = class_exp2obd(exp);
2340 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2341 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2345 /* find any ldlm lock of the inode in osc
2349 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2350 ldlm_iterator_t replace, void *data)
2352 struct ldlm_res_id res_id;
2353 struct obd_device *obd = class_exp2obd(exp);
2356 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2357 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2358 if (rc == LDLM_ITER_STOP)
2360 if (rc == LDLM_ITER_CONTINUE)
2365 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2366 obd_enqueue_update_f upcall, void *cookie,
2367 __u64 *flags, int agl, int rc)
2369 int intent = *flags & LDLM_FL_HAS_INTENT;
2373 /* The request was created before ldlm_cli_enqueue call. */
2374 if (rc == ELDLM_LOCK_ABORTED) {
2375 struct ldlm_reply *rep;
2376 rep = req_capsule_server_get(&req->rq_pill,
2379 LASSERT(rep != NULL);
2380 rep->lock_policy_res1 =
2381 ptlrpc_status_ntoh(rep->lock_policy_res1);
2382 if (rep->lock_policy_res1)
2383 rc = rep->lock_policy_res1;
2387 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2389 *flags |= LDLM_FL_LVB_READY;
2390 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2391 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2394 /* Call the update callback. */
2395 rc = (*upcall)(cookie, rc);
2399 static int osc_enqueue_interpret(const struct lu_env *env,
2400 struct ptlrpc_request *req,
2401 struct osc_enqueue_args *aa, int rc)
2403 struct ldlm_lock *lock;
2404 struct lustre_handle handle;
2406 struct ost_lvb *lvb;
2408 __u64 *flags = aa->oa_flags;
2410 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2411 * might be freed anytime after lock upcall has been called. */
2412 lustre_handle_copy(&handle, aa->oa_lockh);
2413 mode = aa->oa_ei->ei_mode;
2415 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2417 lock = ldlm_handle2lock(&handle);
2419 /* Take an additional reference so that a blocking AST that
2420 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2421 * to arrive after an upcall has been executed by
2422 * osc_enqueue_fini(). */
2423 ldlm_lock_addref(&handle, mode);
2425 /* Let CP AST to grant the lock first. */
2426 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2428 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2433 lvb_len = sizeof(*aa->oa_lvb);
2436 /* Complete obtaining the lock procedure. */
2437 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2438 mode, flags, lvb, lvb_len, &handle, rc);
2439 /* Complete osc stuff. */
2440 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2441 flags, aa->oa_agl, rc);
2443 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2445 /* Release the lock for async request. */
2446 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2448 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2449 * not already released by
2450 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2452 ldlm_lock_decref(&handle, mode);
2454 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2455 aa->oa_lockh, req, aa);
2456 ldlm_lock_decref(&handle, mode);
2457 LDLM_LOCK_PUT(lock);
2461 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2462 struct lov_oinfo *loi, __u64 flags,
2463 struct ost_lvb *lvb, __u32 mode, int rc)
2465 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2467 if (rc == ELDLM_OK) {
2470 LASSERT(lock != NULL);
2471 loi->loi_lvb = *lvb;
2472 tmp = loi->loi_lvb.lvb_size;
2473 /* Extend KMS up to the end of this lock and no further
2474 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2475 if (tmp > lock->l_policy_data.l_extent.end)
2476 tmp = lock->l_policy_data.l_extent.end + 1;
2477 if (tmp >= loi->loi_kms) {
2478 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2479 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2480 loi_kms_set(loi, tmp);
2482 LDLM_DEBUG(lock, "lock acquired, setting rss="
2483 LPU64"; leaving kms="LPU64", end="LPU64,
2484 loi->loi_lvb.lvb_size, loi->loi_kms,
2485 lock->l_policy_data.l_extent.end);
2487 ldlm_lock_allow_match(lock);
2488 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2489 LASSERT(lock != NULL);
2490 loi->loi_lvb = *lvb;
2491 ldlm_lock_allow_match(lock);
2492 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2493 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2499 ldlm_lock_fail_match(lock);
2501 LDLM_LOCK_PUT(lock);
2504 EXPORT_SYMBOL(osc_update_enqueue);
2506 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2508 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2509 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2510 * other synchronous requests, however keeping some locks and trying to obtain
2511 * others may take a considerable amount of time in a case of ost failure; and
2512 * when other sync requests do not get released lock from a client, the client
2513 * is excluded from the cluster -- such scenarious make the life difficult, so
2514 * release locks just after they are obtained. */
2515 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2516 __u64 *flags, ldlm_policy_data_t *policy,
2517 struct ost_lvb *lvb, int kms_valid,
2518 obd_enqueue_update_f upcall, void *cookie,
2519 struct ldlm_enqueue_info *einfo,
2520 struct lustre_handle *lockh,
2521 struct ptlrpc_request_set *rqset, int async, int agl)
2523 struct obd_device *obd = exp->exp_obd;
2524 struct ptlrpc_request *req = NULL;
2525 int intent = *flags & LDLM_FL_HAS_INTENT;
2526 __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2531 /* Filesystem lock extents are extended to page boundaries so that
2532 * dealing with the page cache is a little smoother. */
2533 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2534 policy->l_extent.end |= ~CFS_PAGE_MASK;
2537 * kms is not valid when either object is completely fresh (so that no
2538 * locks are cached), or object was evicted. In the latter case cached
2539 * lock cannot be used, because it would prime inode state with
2540 * potentially stale LVB.
2545 /* Next, search for already existing extent locks that will cover us */
2546 /* If we're trying to read, we also search for an existing PW lock. The
2547 * VFS and page cache already protect us locally, so lots of readers/
2548 * writers can share a single PW lock.
2550 * There are problems with conversion deadlocks, so instead of
2551 * converting a read lock to a write lock, we'll just enqueue a new
2554 * At some point we should cancel the read lock instead of making them
2555 * send us a blocking callback, but there are problems with canceling
2556 * locks out from other users right now, too. */
2557 mode = einfo->ei_mode;
2558 if (einfo->ei_mode == LCK_PR)
2560 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2561 einfo->ei_type, policy, mode, lockh, 0);
2563 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2565 if ((agl != 0) && !ldlm_is_lvb_ready(matched)) {
2566 /* For AGL, if enqueue RPC is sent but the lock is not
2567 * granted, then skip to process this strpe.
2568 * Return -ECANCELED to tell the caller. */
2569 ldlm_lock_decref(lockh, mode);
2570 LDLM_LOCK_PUT(matched);
2572 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2573 *flags |= LDLM_FL_LVB_READY;
2574 /* addref the lock only if not async requests and PW
2575 * lock is matched whereas we asked for PR. */
2576 if (!rqset && einfo->ei_mode != mode)
2577 ldlm_lock_addref(lockh, LCK_PR);
2579 /* I would like to be able to ASSERT here that
2580 * rss <= kms, but I can't, for reasons which
2581 * are explained in lov_enqueue() */
2584 /* We already have a lock, and it's referenced.
2586 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2587 * AGL upcall may change it to CLS_HELD directly. */
2588 (*upcall)(cookie, ELDLM_OK);
2590 if (einfo->ei_mode != mode)
2591 ldlm_lock_decref(lockh, LCK_PW);
2593 /* For async requests, decref the lock. */
2594 ldlm_lock_decref(lockh, einfo->ei_mode);
2595 LDLM_LOCK_PUT(matched);
2598 ldlm_lock_decref(lockh, mode);
2599 LDLM_LOCK_PUT(matched);
2605 CFS_LIST_HEAD(cancels);
2606 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2607 &RQF_LDLM_ENQUEUE_LVB);
2611 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2613 ptlrpc_request_free(req);
2617 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2619 ptlrpc_request_set_replen(req);
2622 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2623 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2625 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2626 sizeof(*lvb), LVB_T_OST, lockh, async);
2629 struct osc_enqueue_args *aa;
2630 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2631 aa = ptlrpc_req_async_args(req);
2634 aa->oa_flags = flags;
2635 aa->oa_upcall = upcall;
2636 aa->oa_cookie = cookie;
2638 aa->oa_lockh = lockh;
2641 req->rq_interpret_reply =
2642 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2643 if (rqset == PTLRPCD_SET)
2644 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2646 ptlrpc_set_add_req(rqset, req);
2647 } else if (intent) {
2648 ptlrpc_req_finished(req);
2653 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2655 ptlrpc_req_finished(req);
2660 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2661 struct ldlm_enqueue_info *einfo,
2662 struct ptlrpc_request_set *rqset)
2664 struct ldlm_res_id res_id;
2668 ostid_build_res_name(&oinfo->oi_md->lsm_oi, &res_id);
2669 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2670 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2671 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2672 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2673 rqset, rqset != NULL, 0);
2677 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2678 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2679 __u64 *flags, void *data, struct lustre_handle *lockh,
2682 struct obd_device *obd = exp->exp_obd;
2683 __u64 lflags = *flags;
2687 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2690 /* Filesystem lock extents are extended to page boundaries so that
2691 * dealing with the page cache is a little smoother */
2692 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2693 policy->l_extent.end |= ~CFS_PAGE_MASK;
2695 /* Next, search for already existing extent locks that will cover us */
2696 /* If we're trying to read, we also search for an existing PW lock. The
2697 * VFS and page cache already protect us locally, so lots of readers/
2698 * writers can share a single PW lock. */
2702 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2703 res_id, type, policy, rc, lockh, unref);
2706 if (!osc_set_data_with_check(lockh, data)) {
2707 if (!(lflags & LDLM_FL_TEST_LOCK))
2708 ldlm_lock_decref(lockh, rc);
2712 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2713 ldlm_lock_addref(lockh, LCK_PR);
2714 ldlm_lock_decref(lockh, LCK_PW);
2721 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2725 if (unlikely(mode == LCK_GROUP))
2726 ldlm_lock_decref_and_cancel(lockh, mode);
2728 ldlm_lock_decref(lockh, mode);
2733 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2734 __u32 mode, struct lustre_handle *lockh)
2737 RETURN(osc_cancel_base(lockh, mode));
2740 static int osc_cancel_unused(struct obd_export *exp,
2741 struct lov_stripe_md *lsm,
2742 ldlm_cancel_flags_t flags,
2745 struct obd_device *obd = class_exp2obd(exp);
2746 struct ldlm_res_id res_id, *resp = NULL;
2749 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2753 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2756 static int osc_statfs_interpret(const struct lu_env *env,
2757 struct ptlrpc_request *req,
2758 struct osc_async_args *aa, int rc)
2760 struct obd_statfs *msfs;
2764 /* The request has in fact never been sent
2765 * due to issues at a higher level (LOV).
2766 * Exit immediately since the caller is
2767 * aware of the problem and takes care
2768 * of the clean up */
2771 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2772 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2778 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2780 GOTO(out, rc = -EPROTO);
2783 *aa->aa_oi->oi_osfs = *msfs;
2785 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2789 static int osc_statfs_async(struct obd_export *exp,
2790 struct obd_info *oinfo, __u64 max_age,
2791 struct ptlrpc_request_set *rqset)
2793 struct obd_device *obd = class_exp2obd(exp);
2794 struct ptlrpc_request *req;
2795 struct osc_async_args *aa;
2799 /* We could possibly pass max_age in the request (as an absolute
2800 * timestamp or a "seconds.usec ago") so the target can avoid doing
2801 * extra calls into the filesystem if that isn't necessary (e.g.
2802 * during mount that would help a bit). Having relative timestamps
2803 * is not so great if request processing is slow, while absolute
2804 * timestamps are not ideal because they need time synchronization. */
2805 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2809 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2811 ptlrpc_request_free(req);
2814 ptlrpc_request_set_replen(req);
2815 req->rq_request_portal = OST_CREATE_PORTAL;
2816 ptlrpc_at_set_req_timeout(req);
2818 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2819 /* procfs requests not want stat in wait for avoid deadlock */
2820 req->rq_no_resend = 1;
2821 req->rq_no_delay = 1;
2824 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2825 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2826 aa = ptlrpc_req_async_args(req);
2829 ptlrpc_set_add_req(rqset, req);
2833 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2834 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2836 struct obd_device *obd = class_exp2obd(exp);
2837 struct obd_statfs *msfs;
2838 struct ptlrpc_request *req;
2839 struct obd_import *imp = NULL;
2843 /*Since the request might also come from lprocfs, so we need
2844 *sync this with client_disconnect_export Bug15684*/
2845 down_read(&obd->u.cli.cl_sem);
2846 if (obd->u.cli.cl_import)
2847 imp = class_import_get(obd->u.cli.cl_import);
2848 up_read(&obd->u.cli.cl_sem);
2852 /* We could possibly pass max_age in the request (as an absolute
2853 * timestamp or a "seconds.usec ago") so the target can avoid doing
2854 * extra calls into the filesystem if that isn't necessary (e.g.
2855 * during mount that would help a bit). Having relative timestamps
2856 * is not so great if request processing is slow, while absolute
2857 * timestamps are not ideal because they need time synchronization. */
2858 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2860 class_import_put(imp);
2865 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2867 ptlrpc_request_free(req);
2870 ptlrpc_request_set_replen(req);
2871 req->rq_request_portal = OST_CREATE_PORTAL;
2872 ptlrpc_at_set_req_timeout(req);
2874 if (flags & OBD_STATFS_NODELAY) {
2875 /* procfs requests not want stat in wait for avoid deadlock */
2876 req->rq_no_resend = 1;
2877 req->rq_no_delay = 1;
2880 rc = ptlrpc_queue_wait(req);
2884 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2886 GOTO(out, rc = -EPROTO);
2893 ptlrpc_req_finished(req);
2897 /* Retrieve object striping information.
2899 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2900 * the maximum number of OST indices which will fit in the user buffer.
2901 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2903 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2905 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2906 struct lov_user_md_v3 lum, *lumk;
2907 struct lov_user_ost_data_v1 *lmm_objects;
2908 int rc = 0, lum_size;
2914 /* we only need the header part from user space to get lmm_magic and
2915 * lmm_stripe_count, (the header part is common to v1 and v3) */
2916 lum_size = sizeof(struct lov_user_md_v1);
2917 if (copy_from_user(&lum, lump, lum_size))
2920 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2921 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2924 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2925 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2926 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2927 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2929 /* we can use lov_mds_md_size() to compute lum_size
2930 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2931 if (lum.lmm_stripe_count > 0) {
2932 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2933 OBD_ALLOC(lumk, lum_size);
2937 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2939 &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2941 lmm_objects = &(lumk->lmm_objects[0]);
2942 lmm_objects->l_ost_oi = lsm->lsm_oi;
2944 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2948 lumk->lmm_oi = lsm->lsm_oi;
2949 lumk->lmm_stripe_count = 1;
2951 if (copy_to_user(lump, lumk, lum_size))
2955 OBD_FREE(lumk, lum_size);
2961 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2962 void *karg, void *uarg)
2964 struct obd_device *obd = exp->exp_obd;
2965 struct obd_ioctl_data *data = karg;
2969 if (!try_module_get(THIS_MODULE)) {
2970 CERROR("Can't get module. Is it alive?");
2974 case OBD_IOC_LOV_GET_CONFIG: {
2976 struct lov_desc *desc;
2977 struct obd_uuid uuid;
2981 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2982 GOTO(out, err = -EINVAL);
2984 data = (struct obd_ioctl_data *)buf;
2986 if (sizeof(*desc) > data->ioc_inllen1) {
2987 obd_ioctl_freedata(buf, len);
2988 GOTO(out, err = -EINVAL);
2991 if (data->ioc_inllen2 < sizeof(uuid)) {
2992 obd_ioctl_freedata(buf, len);
2993 GOTO(out, err = -EINVAL);
2996 desc = (struct lov_desc *)data->ioc_inlbuf1;
2997 desc->ld_tgt_count = 1;
2998 desc->ld_active_tgt_count = 1;
2999 desc->ld_default_stripe_count = 1;
3000 desc->ld_default_stripe_size = 0;
3001 desc->ld_default_stripe_offset = 0;
3002 desc->ld_pattern = 0;
3003 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3005 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3007 err = copy_to_user((void *)uarg, buf, len);
3010 obd_ioctl_freedata(buf, len);
3013 case LL_IOC_LOV_SETSTRIPE:
3014 err = obd_alloc_memmd(exp, karg);
3018 case LL_IOC_LOV_GETSTRIPE:
3019 err = osc_getstripe(karg, uarg);
3021 case OBD_IOC_CLIENT_RECOVER:
3022 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3023 data->ioc_inlbuf1, 0);
3027 case IOC_OSC_SET_ACTIVE:
3028 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3031 case OBD_IOC_POLL_QUOTACHECK:
3032 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
3034 case OBD_IOC_PING_TARGET:
3035 err = ptlrpc_obd_ping(obd);
3038 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3039 cmd, current_comm());
3040 GOTO(out, err = -ENOTTY);
3043 module_put(THIS_MODULE);
3047 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3048 obd_count keylen, void *key, __u32 *vallen, void *val,
3049 struct lov_stripe_md *lsm)
3052 if (!vallen || !val)
3055 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3056 __u32 *stripe = val;
3057 *vallen = sizeof(*stripe);
3060 } else if (KEY_IS(KEY_LAST_ID)) {
3061 struct ptlrpc_request *req;
3066 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3067 &RQF_OST_GET_INFO_LAST_ID);
3071 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3072 RCL_CLIENT, keylen);
3073 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3075 ptlrpc_request_free(req);
3079 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3080 memcpy(tmp, key, keylen);
3082 req->rq_no_delay = req->rq_no_resend = 1;
3083 ptlrpc_request_set_replen(req);
3084 rc = ptlrpc_queue_wait(req);
3088 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3090 GOTO(out, rc = -EPROTO);
3092 *((obd_id *)val) = *reply;
3094 ptlrpc_req_finished(req);
3096 } else if (KEY_IS(KEY_FIEMAP)) {
3097 struct ll_fiemap_info_key *fm_key =
3098 (struct ll_fiemap_info_key *)key;
3099 struct ldlm_res_id res_id;
3100 ldlm_policy_data_t policy;
3101 struct lustre_handle lockh;
3102 ldlm_mode_t mode = 0;
3103 struct ptlrpc_request *req;
3104 struct ll_user_fiemap *reply;
3108 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
3111 policy.l_extent.start = fm_key->fiemap.fm_start &
3114 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
3115 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
3116 policy.l_extent.end = OBD_OBJECT_EOF;
3118 policy.l_extent.end = (fm_key->fiemap.fm_start +
3119 fm_key->fiemap.fm_length +
3120 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
3122 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
3123 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
3124 LDLM_FL_BLOCK_GRANTED |
3126 &res_id, LDLM_EXTENT, &policy,
3127 LCK_PR | LCK_PW, &lockh, 0);
3128 if (mode) { /* lock is cached on client */
3129 if (mode != LCK_PR) {
3130 ldlm_lock_addref(&lockh, LCK_PR);
3131 ldlm_lock_decref(&lockh, LCK_PW);
3133 } else { /* no cached lock, needs acquire lock on server side */
3134 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
3135 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
3139 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3140 &RQF_OST_GET_INFO_FIEMAP);
3142 GOTO(drop_lock, rc = -ENOMEM);
3144 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3145 RCL_CLIENT, keylen);
3146 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3147 RCL_CLIENT, *vallen);
3148 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3149 RCL_SERVER, *vallen);
3151 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3153 ptlrpc_request_free(req);
3154 GOTO(drop_lock, rc);
3157 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3158 memcpy(tmp, key, keylen);
3159 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3160 memcpy(tmp, val, *vallen);
3162 ptlrpc_request_set_replen(req);
3163 rc = ptlrpc_queue_wait(req);
3167 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3169 GOTO(fini_req, rc = -EPROTO);
3171 memcpy(val, reply, *vallen);
3173 ptlrpc_req_finished(req);
3176 ldlm_lock_decref(&lockh, LCK_PR);
3183 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3184 obd_count keylen, void *key, obd_count vallen,
3185 void *val, struct ptlrpc_request_set *set)
3187 struct ptlrpc_request *req;
3188 struct obd_device *obd = exp->exp_obd;
3189 struct obd_import *imp = class_exp2cliimp(exp);
3194 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3196 if (KEY_IS(KEY_CHECKSUM)) {
3197 if (vallen != sizeof(int))
3199 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3203 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3204 sptlrpc_conf_client_adapt(obd);
3208 if (KEY_IS(KEY_FLUSH_CTX)) {
3209 sptlrpc_import_flush_my_ctx(imp);
3213 if (KEY_IS(KEY_CACHE_SET)) {
3214 struct client_obd *cli = &obd->u.cli;
3216 LASSERT(cli->cl_cache == NULL); /* only once */
3217 cli->cl_cache = (struct cl_client_cache *)val;
3218 cfs_atomic_inc(&cli->cl_cache->ccc_users);
3219 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
3221 /* add this osc into entity list */
3222 LASSERT(cfs_list_empty(&cli->cl_lru_osc));
3223 spin_lock(&cli->cl_cache->ccc_lru_lock);
3224 cfs_list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
3225 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3230 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3231 struct client_obd *cli = &obd->u.cli;
3232 int nr = cfs_atomic_read(&cli->cl_lru_in_list) >> 1;
3233 int target = *(int *)val;
3235 nr = osc_lru_shrink(cli, min(nr, target));
3240 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3243 /* We pass all other commands directly to OST. Since nobody calls osc
3244 methods directly and everybody is supposed to go through LOV, we
3245 assume lov checked invalid values for us.
3246 The only recognised values so far are evict_by_nid and mds_conn.
3247 Even if something bad goes through, we'd get a -EINVAL from OST
3250 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3251 &RQF_OST_SET_GRANT_INFO :
3256 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3257 RCL_CLIENT, keylen);
3258 if (!KEY_IS(KEY_GRANT_SHRINK))
3259 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3260 RCL_CLIENT, vallen);
3261 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3263 ptlrpc_request_free(req);
3267 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3268 memcpy(tmp, key, keylen);
3269 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3272 memcpy(tmp, val, vallen);
3274 if (KEY_IS(KEY_GRANT_SHRINK)) {
3275 struct osc_grant_args *aa;
3278 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3279 aa = ptlrpc_req_async_args(req);
3282 ptlrpc_req_finished(req);
3285 *oa = ((struct ost_body *)val)->oa;
3287 req->rq_interpret_reply = osc_shrink_grant_interpret;
3290 ptlrpc_request_set_replen(req);
3291 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3292 LASSERT(set != NULL);
3293 ptlrpc_set_add_req(set, req);
3294 ptlrpc_check_set(NULL, set);
3296 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3302 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3303 struct obd_device *disk_obd, int *index)
3305 /* this code is not supposed to be used with LOD/OSP
3306 * to be removed soon */
3311 static int osc_llog_finish(struct obd_device *obd, int count)
3313 struct llog_ctxt *ctxt;
3317 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3319 llog_cat_close(NULL, ctxt->loc_handle);
3320 llog_cleanup(NULL, ctxt);
3323 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3325 llog_cleanup(NULL, ctxt);
3329 static int osc_reconnect(const struct lu_env *env,
3330 struct obd_export *exp, struct obd_device *obd,
3331 struct obd_uuid *cluuid,
3332 struct obd_connect_data *data,
3335 struct client_obd *cli = &obd->u.cli;
3337 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3340 client_obd_list_lock(&cli->cl_loi_list_lock);
3341 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3342 2 * cli_brw_size(obd);
3343 lost_grant = cli->cl_lost_grant;
3344 cli->cl_lost_grant = 0;
3345 client_obd_list_unlock(&cli->cl_loi_list_lock);
3347 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3348 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3349 data->ocd_version, data->ocd_grant, lost_grant);
3355 static int osc_disconnect(struct obd_export *exp)
3357 struct obd_device *obd = class_exp2obd(exp);
3358 struct llog_ctxt *ctxt;
3361 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3363 if (obd->u.cli.cl_conn_count == 1) {
3364 /* Flush any remaining cancel messages out to the
3366 llog_sync(ctxt, exp, 0);
3368 llog_ctxt_put(ctxt);
3370 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3374 rc = client_disconnect_export(exp);
3376 * Initially we put del_shrink_grant before disconnect_export, but it
3377 * causes the following problem if setup (connect) and cleanup
3378 * (disconnect) are tangled together.
3379 * connect p1 disconnect p2
3380 * ptlrpc_connect_import
3381 * ............... class_manual_cleanup
3384 * ptlrpc_connect_interrupt
3386 * add this client to shrink list
3388 * Bang! pinger trigger the shrink.
3389 * So the osc should be disconnected from the shrink list, after we
3390 * are sure the import has been destroyed. BUG18662
3392 if (obd->u.cli.cl_import == NULL)
3393 osc_del_shrink_grant(&obd->u.cli);
3397 static int osc_import_event(struct obd_device *obd,
3398 struct obd_import *imp,
3399 enum obd_import_event event)
3401 struct client_obd *cli;
3405 LASSERT(imp->imp_obd == obd);
3408 case IMP_EVENT_DISCON: {
3410 client_obd_list_lock(&cli->cl_loi_list_lock);
3411 cli->cl_avail_grant = 0;
3412 cli->cl_lost_grant = 0;
3413 client_obd_list_unlock(&cli->cl_loi_list_lock);
3416 case IMP_EVENT_INACTIVE: {
3417 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3420 case IMP_EVENT_INVALIDATE: {
3421 struct ldlm_namespace *ns = obd->obd_namespace;
3425 env = cl_env_get(&refcheck);
3429 /* all pages go to failing rpcs due to the invalid
3431 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3433 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3434 cl_env_put(env, &refcheck);
3439 case IMP_EVENT_ACTIVE: {
3440 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3443 case IMP_EVENT_OCD: {
3444 struct obd_connect_data *ocd = &imp->imp_connect_data;
3446 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3447 osc_init_grant(&obd->u.cli, ocd);
3450 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3451 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3453 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3456 case IMP_EVENT_DEACTIVATE: {
3457 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3460 case IMP_EVENT_ACTIVATE: {
3461 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3465 CERROR("Unknown import event %d\n", event);
3472 * Determine whether the lock can be canceled before replaying the lock
3473 * during recovery, see bug16774 for detailed information.
3475 * \retval zero the lock can't be canceled
3476 * \retval other ok to cancel
3478 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3480 check_res_locked(lock->l_resource);
3483 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3485 * XXX as a future improvement, we can also cancel unused write lock
3486 * if it doesn't have dirty data and active mmaps.
3488 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3489 (lock->l_granted_mode == LCK_PR ||
3490 lock->l_granted_mode == LCK_CR) &&
3491 (osc_dlm_lock_pageref(lock) == 0))
3497 static int brw_queue_work(const struct lu_env *env, void *data)
3499 struct client_obd *cli = data;
3501 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3503 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3507 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3509 struct lprocfs_static_vars lvars = { 0 };
3510 struct client_obd *cli = &obd->u.cli;
3515 rc = ptlrpcd_addref();
3519 rc = client_obd_setup(obd, lcfg);
3521 GOTO(out_ptlrpcd, rc);
3523 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3524 if (IS_ERR(handler))
3525 GOTO(out_client_setup, rc = PTR_ERR(handler));
3526 cli->cl_writeback_work = handler;
3528 rc = osc_quota_setup(obd);
3530 GOTO(out_ptlrpcd_work, rc);
3532 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3533 lprocfs_osc_init_vars(&lvars);
3534 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3535 lproc_osc_attach_seqstat(obd);
3536 sptlrpc_lprocfs_cliobd_attach(obd);
3537 ptlrpc_lprocfs_register_obd(obd);
3540 /* We need to allocate a few requests more, because
3541 * brw_interpret tries to create new requests before freeing
3542 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3543 * reserved, but I'm afraid that might be too much wasted RAM
3544 * in fact, so 2 is just my guess and still should work. */
3545 cli->cl_import->imp_rq_pool =
3546 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3548 ptlrpc_add_rqs_to_pool);
3550 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3551 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3555 ptlrpcd_destroy_work(handler);
3557 client_obd_cleanup(obd);
3563 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3569 case OBD_CLEANUP_EARLY: {
3570 struct obd_import *imp;
3571 imp = obd->u.cli.cl_import;
3572 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3573 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3574 ptlrpc_deactivate_import(imp);
3575 spin_lock(&imp->imp_lock);
3576 imp->imp_pingable = 0;
3577 spin_unlock(&imp->imp_lock);
3580 case OBD_CLEANUP_EXPORTS: {
3581 struct client_obd *cli = &obd->u.cli;
3583 * for echo client, export may be on zombie list, wait for
3584 * zombie thread to cull it, because cli.cl_import will be
3585 * cleared in client_disconnect_export():
3586 * class_export_destroy() -> obd_cleanup() ->
3587 * echo_device_free() -> echo_client_cleanup() ->
3588 * obd_disconnect() -> osc_disconnect() ->
3589 * client_disconnect_export()
3591 obd_zombie_barrier();
3592 if (cli->cl_writeback_work) {
3593 ptlrpcd_destroy_work(cli->cl_writeback_work);
3594 cli->cl_writeback_work = NULL;
3596 obd_cleanup_client_import(obd);
3597 ptlrpc_lprocfs_unregister_obd(obd);
3598 lprocfs_obd_cleanup(obd);
3599 rc = obd_llog_finish(obd, 0);
3601 CERROR("failed to cleanup llogging subsystems\n");
3608 int osc_cleanup(struct obd_device *obd)
3610 struct client_obd *cli = &obd->u.cli;
3616 if (cli->cl_cache != NULL) {
3617 LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_users) > 0);
3618 spin_lock(&cli->cl_cache->ccc_lru_lock);
3619 cfs_list_del_init(&cli->cl_lru_osc);
3620 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3621 cli->cl_lru_left = NULL;
3622 cfs_atomic_dec(&cli->cl_cache->ccc_users);
3623 cli->cl_cache = NULL;
3626 /* free memory of osc quota cache */
3627 osc_quota_cleanup(obd);
3629 rc = client_obd_cleanup(obd);
3635 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3637 struct lprocfs_static_vars lvars = { 0 };
3640 lprocfs_osc_init_vars(&lvars);
3642 switch (lcfg->lcfg_command) {
3644 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3654 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3656 return osc_process_config_base(obd, buf);
3659 struct obd_ops osc_obd_ops = {
3660 .o_owner = THIS_MODULE,
3661 .o_setup = osc_setup,
3662 .o_precleanup = osc_precleanup,
3663 .o_cleanup = osc_cleanup,
3664 .o_add_conn = client_import_add_conn,
3665 .o_del_conn = client_import_del_conn,
3666 .o_connect = client_connect_import,
3667 .o_reconnect = osc_reconnect,
3668 .o_disconnect = osc_disconnect,
3669 .o_statfs = osc_statfs,
3670 .o_statfs_async = osc_statfs_async,
3671 .o_packmd = osc_packmd,
3672 .o_unpackmd = osc_unpackmd,
3673 .o_create = osc_create,
3674 .o_destroy = osc_destroy,
3675 .o_getattr = osc_getattr,
3676 .o_getattr_async = osc_getattr_async,
3677 .o_setattr = osc_setattr,
3678 .o_setattr_async = osc_setattr_async,
3680 .o_punch = osc_punch,
3682 .o_enqueue = osc_enqueue,
3683 .o_change_cbdata = osc_change_cbdata,
3684 .o_find_cbdata = osc_find_cbdata,
3685 .o_cancel = osc_cancel,
3686 .o_cancel_unused = osc_cancel_unused,
3687 .o_iocontrol = osc_iocontrol,
3688 .o_get_info = osc_get_info,
3689 .o_set_info_async = osc_set_info_async,
3690 .o_import_event = osc_import_event,
3691 .o_llog_init = osc_llog_init,
3692 .o_llog_finish = osc_llog_finish,
3693 .o_process_config = osc_process_config,
3694 .o_quotactl = osc_quotactl,
3695 .o_quotacheck = osc_quotacheck,
3698 extern struct lu_kmem_descr osc_caches[];
3699 extern spinlock_t osc_ast_guard;
3700 extern struct lock_class_key osc_ast_guard_class;
3702 int __init osc_init(void)
3704 struct lprocfs_static_vars lvars = { 0 };
3708 /* print an address of _any_ initialized kernel symbol from this
3709 * module, to allow debugging with gdb that doesn't support data
3710 * symbols from modules.*/
3711 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3713 rc = lu_kmem_init(osc_caches);
3717 lprocfs_osc_init_vars(&lvars);
3719 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3720 LUSTRE_OSC_NAME, &osc_device_type);
3722 lu_kmem_fini(osc_caches);
3726 spin_lock_init(&osc_ast_guard);
3727 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3733 static void /*__exit*/ osc_exit(void)
3735 class_unregister_type(LUSTRE_OSC_NAME);
3736 lu_kmem_fini(osc_caches);
3739 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3740 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3741 MODULE_LICENSE("GPL");
3743 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);