4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <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);
82 OBD_FREE(*lmmp, lmm_size);
88 OBD_ALLOC(*lmmp, lmm_size);
94 LASSERT(lsm->lsm_object_id);
95 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
96 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
97 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
103 /* Unpack OSC object metadata from disk storage (LE byte order). */
104 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
105 struct lov_mds_md *lmm, int lmm_bytes)
108 struct obd_import *imp = class_exp2cliimp(exp);
112 if (lmm_bytes < sizeof (*lmm)) {
113 CERROR("lov_mds_md too small: %d, need %d\n",
114 lmm_bytes, (int)sizeof(*lmm));
117 /* XXX LOV_MAGIC etc check? */
119 if (lmm->lmm_object_id == 0) {
120 CERROR("lov_mds_md: zero lmm_object_id\n");
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);
140 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
141 if ((*lsmp)->lsm_oinfo[0] == NULL) {
142 OBD_FREE(*lsmp, lsm_size);
145 loi_init((*lsmp)->lsm_oinfo[0]);
149 /* XXX zero *lsmp? */
150 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
151 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
152 LASSERT((*lsmp)->lsm_object_id);
153 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
157 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
158 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
160 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
165 static inline void osc_pack_capa(struct ptlrpc_request *req,
166 struct ost_body *body, void *capa)
168 struct obd_capa *oc = (struct obd_capa *)capa;
169 struct lustre_capa *c;
174 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
177 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
178 DEBUG_CAPA(D_SEC, c, "pack");
181 static inline void osc_pack_req_body(struct ptlrpc_request *req,
182 struct obd_info *oinfo)
184 struct ost_body *body;
186 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
189 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
190 osc_pack_capa(req, body, oinfo->oi_capa);
193 static inline void osc_set_capa_size(struct ptlrpc_request *req,
194 const struct req_msg_field *field,
198 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
200 /* it is already calculated as sizeof struct obd_capa */
204 static int osc_getattr_interpret(const struct lu_env *env,
205 struct ptlrpc_request *req,
206 struct osc_async_args *aa, int rc)
208 struct ost_body *body;
214 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
216 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
217 lustre_get_wire_obdo(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(oinfo->oi_oa, &body->oa);
298 /* This should really be sent by the OST */
299 oinfo->oi_oa->o_blksize = exp_brw_size(exp);
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(oinfo->oi_oa, &body->oa);
345 ptlrpc_req_finished(req);
349 static int osc_setattr_interpret(const struct lu_env *env,
350 struct ptlrpc_request *req,
351 struct osc_setattr_args *sa, int rc)
353 struct ost_body *body;
359 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
361 GOTO(out, rc = -EPROTO);
363 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
365 rc = sa->sa_upcall(sa->sa_cookie, rc);
369 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
370 struct obd_trans_info *oti,
371 obd_enqueue_update_f upcall, void *cookie,
372 struct ptlrpc_request_set *rqset)
374 struct ptlrpc_request *req;
375 struct osc_setattr_args *sa;
379 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
383 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
384 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
386 ptlrpc_request_free(req);
390 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
391 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
393 osc_pack_req_body(req, oinfo);
395 ptlrpc_request_set_replen(req);
397 /* do mds to ost setattr asynchronously */
399 /* Do not wait for response. */
400 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
402 req->rq_interpret_reply =
403 (ptlrpc_interpterer_t)osc_setattr_interpret;
405 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
406 sa = ptlrpc_req_async_args(req);
407 sa->sa_oa = oinfo->oi_oa;
408 sa->sa_upcall = upcall;
409 sa->sa_cookie = cookie;
411 if (rqset == PTLRPCD_SET)
412 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
414 ptlrpc_set_add_req(rqset, req);
420 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
421 struct obd_trans_info *oti,
422 struct ptlrpc_request_set *rqset)
424 return osc_setattr_async_base(exp, oinfo, oti,
425 oinfo->oi_cb_up, oinfo, rqset);
428 int osc_real_create(struct obd_export *exp, struct obdo *oa,
429 struct lov_stripe_md **ea, struct obd_trans_info *oti)
431 struct ptlrpc_request *req;
432 struct ost_body *body;
433 struct lov_stripe_md *lsm;
442 rc = obd_alloc_memmd(exp, &lsm);
447 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
449 GOTO(out, rc = -ENOMEM);
451 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
453 ptlrpc_request_free(req);
457 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
459 lustre_set_wire_obdo(&body->oa, oa);
461 ptlrpc_request_set_replen(req);
463 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
464 oa->o_flags == OBD_FL_DELORPHAN) {
466 "delorphan from OST integration");
467 /* Don't resend the delorphan req */
468 req->rq_no_resend = req->rq_no_delay = 1;
471 rc = ptlrpc_queue_wait(req);
475 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
477 GOTO(out_req, rc = -EPROTO);
479 lustre_get_wire_obdo(oa, &body->oa);
481 /* This should really be sent by the OST */
482 oa->o_blksize = exp_brw_size(exp);
483 oa->o_valid |= OBD_MD_FLBLKSZ;
485 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
486 * have valid lsm_oinfo data structs, so don't go touching that.
487 * This needs to be fixed in a big way.
489 lsm->lsm_object_id = oa->o_id;
490 lsm->lsm_object_seq = oa->o_seq;
494 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
496 if (oa->o_valid & OBD_MD_FLCOOKIE) {
497 if (!oti->oti_logcookies)
498 oti_alloc_cookies(oti, 1);
499 *oti->oti_logcookies = oa->o_lcookie;
503 CDEBUG(D_HA, "transno: "LPD64"\n",
504 lustre_msg_get_transno(req->rq_repmsg));
506 ptlrpc_req_finished(req);
509 obd_free_memmd(exp, &lsm);
513 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
514 obd_enqueue_update_f upcall, void *cookie,
515 struct ptlrpc_request_set *rqset)
517 struct ptlrpc_request *req;
518 struct osc_setattr_args *sa;
519 struct ost_body *body;
523 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
527 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
528 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
530 ptlrpc_request_free(req);
533 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
534 ptlrpc_at_set_req_timeout(req);
536 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
538 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
539 osc_pack_capa(req, body, oinfo->oi_capa);
541 ptlrpc_request_set_replen(req);
543 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
544 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
545 sa = ptlrpc_req_async_args(req);
546 sa->sa_oa = oinfo->oi_oa;
547 sa->sa_upcall = upcall;
548 sa->sa_cookie = cookie;
549 if (rqset == PTLRPCD_SET)
550 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
552 ptlrpc_set_add_req(rqset, req);
557 static int osc_punch(const struct lu_env *env, struct obd_export *exp,
558 struct obd_info *oinfo, struct obd_trans_info *oti,
559 struct ptlrpc_request_set *rqset)
561 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
562 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
563 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
564 return osc_punch_base(exp, oinfo,
565 oinfo->oi_cb_up, oinfo, rqset);
568 static int osc_sync_interpret(const struct lu_env *env,
569 struct ptlrpc_request *req,
572 struct osc_fsync_args *fa = arg;
573 struct ost_body *body;
579 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
581 CERROR ("can't unpack ost_body\n");
582 GOTO(out, rc = -EPROTO);
585 *fa->fa_oi->oi_oa = body->oa;
587 rc = fa->fa_upcall(fa->fa_cookie, rc);
591 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
592 obd_enqueue_update_f upcall, void *cookie,
593 struct ptlrpc_request_set *rqset)
595 struct ptlrpc_request *req;
596 struct ost_body *body;
597 struct osc_fsync_args *fa;
601 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
605 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
606 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
608 ptlrpc_request_free(req);
612 /* overload the size and blocks fields in the oa with start/end */
613 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
615 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
616 osc_pack_capa(req, body, oinfo->oi_capa);
618 ptlrpc_request_set_replen(req);
619 req->rq_interpret_reply = osc_sync_interpret;
621 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
622 fa = ptlrpc_req_async_args(req);
624 fa->fa_upcall = upcall;
625 fa->fa_cookie = cookie;
627 if (rqset == PTLRPCD_SET)
628 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
630 ptlrpc_set_add_req(rqset, req);
635 static int osc_sync(const struct lu_env *env, struct obd_export *exp,
636 struct obd_info *oinfo, obd_size start, obd_size end,
637 struct ptlrpc_request_set *set)
642 CDEBUG(D_INFO, "oa NULL\n");
646 oinfo->oi_oa->o_size = start;
647 oinfo->oi_oa->o_blocks = end;
648 oinfo->oi_oa->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
650 RETURN(osc_sync_base(exp, oinfo, oinfo->oi_cb_up, oinfo, set));
653 /* Find and cancel locally locks matched by @mode in the resource found by
654 * @objid. Found locks are added into @cancel list. Returns the amount of
655 * locks added to @cancels list. */
656 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
658 ldlm_mode_t mode, int lock_flags)
660 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
661 struct ldlm_res_id res_id;
662 struct ldlm_resource *res;
666 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
667 * export) but disabled through procfs (flag in NS).
669 * This distinguishes from a case when ELC is not supported originally,
670 * when we still want to cancel locks in advance and just cancel them
671 * locally, without sending any RPC. */
672 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
675 ostid_build_res_name(&oa->o_oi, &res_id);
676 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
680 LDLM_RESOURCE_ADDREF(res);
681 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
682 lock_flags, 0, NULL);
683 LDLM_RESOURCE_DELREF(res);
684 ldlm_resource_putref(res);
688 static int osc_destroy_interpret(const struct lu_env *env,
689 struct ptlrpc_request *req, void *data,
692 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
694 cfs_atomic_dec(&cli->cl_destroy_in_flight);
695 cfs_waitq_signal(&cli->cl_destroy_waitq);
699 static int osc_can_send_destroy(struct client_obd *cli)
701 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
702 cli->cl_max_rpcs_in_flight) {
703 /* The destroy request can be sent */
706 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
707 cli->cl_max_rpcs_in_flight) {
709 * The counter has been modified between the two atomic
712 cfs_waitq_signal(&cli->cl_destroy_waitq);
717 int osc_create(const struct lu_env *env, struct obd_export *exp,
718 struct obdo *oa, struct lov_stripe_md **ea,
719 struct obd_trans_info *oti)
726 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
728 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
729 oa->o_flags == OBD_FL_RECREATE_OBJS) {
730 RETURN(osc_real_create(exp, oa, ea, oti));
733 if (!fid_seq_is_mdt(oa->o_seq))
734 RETURN(osc_real_create(exp, oa, ea, oti));
736 /* we should not get here anymore */
742 /* Destroy requests can be async always on the client, and we don't even really
743 * care about the return code since the client cannot do anything at all about
745 * When the MDS is unlinking a filename, it saves the file objects into a
746 * recovery llog, and these object records are cancelled when the OST reports
747 * they were destroyed and sync'd to disk (i.e. transaction committed).
748 * If the client dies, or the OST is down when the object should be destroyed,
749 * the records are not cancelled, and when the OST reconnects to the MDS next,
750 * it will retrieve the llog unlink logs and then sends the log cancellation
751 * cookies to the MDS after committing destroy transactions. */
752 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
753 struct obdo *oa, struct lov_stripe_md *ea,
754 struct obd_trans_info *oti, struct obd_export *md_export,
757 struct client_obd *cli = &exp->exp_obd->u.cli;
758 struct ptlrpc_request *req;
759 struct ost_body *body;
760 CFS_LIST_HEAD(cancels);
765 CDEBUG(D_INFO, "oa NULL\n");
769 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
770 LDLM_FL_DISCARD_DATA);
772 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
774 ldlm_lock_list_put(&cancels, l_bl_ast, count);
778 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
779 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
782 ptlrpc_request_free(req);
786 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
787 ptlrpc_at_set_req_timeout(req);
789 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
790 oa->o_lcookie = *oti->oti_logcookies;
791 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
793 lustre_set_wire_obdo(&body->oa, oa);
795 osc_pack_capa(req, body, (struct obd_capa *)capa);
796 ptlrpc_request_set_replen(req);
798 /* If osc_destory is for destroying the unlink orphan,
799 * sent from MDT to OST, which should not be blocked here,
800 * because the process might be triggered by ptlrpcd, and
801 * it is not good to block ptlrpcd thread (b=16006)*/
802 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
803 req->rq_interpret_reply = osc_destroy_interpret;
804 if (!osc_can_send_destroy(cli)) {
805 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
809 * Wait until the number of on-going destroy RPCs drops
810 * under max_rpc_in_flight
812 l_wait_event_exclusive(cli->cl_destroy_waitq,
813 osc_can_send_destroy(cli), &lwi);
817 /* Do not wait for response */
818 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
822 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
825 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
827 LASSERT(!(oa->o_valid & bits));
830 client_obd_list_lock(&cli->cl_loi_list_lock);
831 oa->o_dirty = cli->cl_dirty;
832 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
833 cli->cl_dirty_max)) {
834 CERROR("dirty %lu - %lu > dirty_max %lu\n",
835 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
837 } else if (unlikely(cfs_atomic_read(&obd_dirty_pages) -
838 cfs_atomic_read(&obd_dirty_transit_pages) >
839 (long)(obd_max_dirty_pages + 1))) {
840 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
841 * not covered by a lock thus they may safely race and trip
842 * this CERROR() unless we add in a small fudge factor (+1). */
843 CERROR("dirty %d - %d > system dirty_max %d\n",
844 cfs_atomic_read(&obd_dirty_pages),
845 cfs_atomic_read(&obd_dirty_transit_pages),
846 obd_max_dirty_pages);
848 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
849 CERROR("dirty %lu - dirty_max %lu too big???\n",
850 cli->cl_dirty, cli->cl_dirty_max);
853 long max_in_flight = (cli->cl_max_pages_per_rpc <<
855 (cli->cl_max_rpcs_in_flight + 1);
856 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
858 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
859 oa->o_dropped = cli->cl_lost_grant;
860 cli->cl_lost_grant = 0;
861 client_obd_list_unlock(&cli->cl_loi_list_lock);
862 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
863 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
867 void osc_update_next_shrink(struct client_obd *cli)
869 cli->cl_next_shrink_grant =
870 cfs_time_shift(cli->cl_grant_shrink_interval);
871 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
872 cli->cl_next_shrink_grant);
875 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
877 client_obd_list_lock(&cli->cl_loi_list_lock);
878 cli->cl_avail_grant += grant;
879 client_obd_list_unlock(&cli->cl_loi_list_lock);
882 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
884 if (body->oa.o_valid & OBD_MD_FLGRANT) {
885 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
886 __osc_update_grant(cli, body->oa.o_grant);
890 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
891 obd_count keylen, void *key, obd_count vallen,
892 void *val, struct ptlrpc_request_set *set);
894 static int osc_shrink_grant_interpret(const struct lu_env *env,
895 struct ptlrpc_request *req,
898 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
899 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
900 struct ost_body *body;
903 __osc_update_grant(cli, oa->o_grant);
907 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
909 osc_update_grant(cli, body);
915 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
917 client_obd_list_lock(&cli->cl_loi_list_lock);
918 oa->o_grant = cli->cl_avail_grant / 4;
919 cli->cl_avail_grant -= oa->o_grant;
920 client_obd_list_unlock(&cli->cl_loi_list_lock);
921 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
922 oa->o_valid |= OBD_MD_FLFLAGS;
925 oa->o_flags |= OBD_FL_SHRINK_GRANT;
926 osc_update_next_shrink(cli);
929 /* Shrink the current grant, either from some large amount to enough for a
930 * full set of in-flight RPCs, or if we have already shrunk to that limit
931 * then to enough for a single RPC. This avoids keeping more grant than
932 * needed, and avoids shrinking the grant piecemeal. */
933 static int osc_shrink_grant(struct client_obd *cli)
935 long target = (cli->cl_max_rpcs_in_flight + 1) *
936 cli->cl_max_pages_per_rpc;
938 client_obd_list_lock(&cli->cl_loi_list_lock);
939 if (cli->cl_avail_grant <= target)
940 target = cli->cl_max_pages_per_rpc;
941 client_obd_list_unlock(&cli->cl_loi_list_lock);
943 return osc_shrink_grant_to_target(cli, target);
946 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
949 struct ost_body *body;
952 client_obd_list_lock(&cli->cl_loi_list_lock);
953 /* Don't shrink if we are already above or below the desired limit
954 * We don't want to shrink below a single RPC, as that will negatively
955 * impact block allocation and long-term performance. */
956 if (target < cli->cl_max_pages_per_rpc)
957 target = cli->cl_max_pages_per_rpc;
959 if (target >= cli->cl_avail_grant) {
960 client_obd_list_unlock(&cli->cl_loi_list_lock);
963 client_obd_list_unlock(&cli->cl_loi_list_lock);
969 osc_announce_cached(cli, &body->oa, 0);
971 client_obd_list_lock(&cli->cl_loi_list_lock);
972 body->oa.o_grant = cli->cl_avail_grant - target;
973 cli->cl_avail_grant = target;
974 client_obd_list_unlock(&cli->cl_loi_list_lock);
975 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
976 body->oa.o_valid |= OBD_MD_FLFLAGS;
977 body->oa.o_flags = 0;
979 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
980 osc_update_next_shrink(cli);
982 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
983 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
984 sizeof(*body), body, NULL);
986 __osc_update_grant(cli, body->oa.o_grant);
991 static int osc_should_shrink_grant(struct client_obd *client)
993 cfs_time_t time = cfs_time_current();
994 cfs_time_t next_shrink = client->cl_next_shrink_grant;
996 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
997 OBD_CONNECT_GRANT_SHRINK) == 0)
1000 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1001 int brw_size = exp_brw_size(
1002 client->cl_import->imp_obd->obd_self_export);
1004 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1005 client->cl_avail_grant > brw_size)
1008 osc_update_next_shrink(client);
1013 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1015 struct client_obd *client;
1017 cfs_list_for_each_entry(client, &item->ti_obd_list,
1018 cl_grant_shrink_list) {
1019 if (osc_should_shrink_grant(client))
1020 osc_shrink_grant(client);
1025 static int osc_add_shrink_grant(struct client_obd *client)
1029 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1031 osc_grant_shrink_grant_cb, NULL,
1032 &client->cl_grant_shrink_list);
1034 CERROR("add grant client %s error %d\n",
1035 client->cl_import->imp_obd->obd_name, rc);
1038 CDEBUG(D_CACHE, "add grant client %s \n",
1039 client->cl_import->imp_obd->obd_name);
1040 osc_update_next_shrink(client);
1044 static int osc_del_shrink_grant(struct client_obd *client)
1046 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1050 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1053 * ocd_grant is the total grant amount we're expect to hold: if we've
1054 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1055 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1057 * race is tolerable here: if we're evicted, but imp_state already
1058 * left EVICTED state, then cl_dirty must be 0 already.
1060 client_obd_list_lock(&cli->cl_loi_list_lock);
1061 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1062 cli->cl_avail_grant = ocd->ocd_grant;
1064 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1066 if (cli->cl_avail_grant < 0) {
1067 CWARN("%s: available grant < 0, the OSS is probably not running"
1068 " with patch from bug20278 (%ld) \n",
1069 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1070 /* workaround for 1.6 servers which do not have
1071 * the patch from bug20278 */
1072 cli->cl_avail_grant = ocd->ocd_grant;
1075 /* determine the appropriate chunk size used by osc_extent. */
1076 cli->cl_chunkbits = max_t(int, CFS_PAGE_SHIFT, ocd->ocd_blocksize);
1077 client_obd_list_unlock(&cli->cl_loi_list_lock);
1079 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1080 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1081 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1083 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1084 cfs_list_empty(&cli->cl_grant_shrink_list))
1085 osc_add_shrink_grant(cli);
1088 /* We assume that the reason this OSC got a short read is because it read
1089 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1090 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1091 * this stripe never got written at or beyond this stripe offset yet. */
1092 static void handle_short_read(int nob_read, obd_count page_count,
1093 struct brw_page **pga)
1098 /* skip bytes read OK */
1099 while (nob_read > 0) {
1100 LASSERT (page_count > 0);
1102 if (pga[i]->count > nob_read) {
1103 /* EOF inside this page */
1104 ptr = cfs_kmap(pga[i]->pg) +
1105 (pga[i]->off & ~CFS_PAGE_MASK);
1106 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1107 cfs_kunmap(pga[i]->pg);
1113 nob_read -= pga[i]->count;
1118 /* zero remaining pages */
1119 while (page_count-- > 0) {
1120 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1121 memset(ptr, 0, pga[i]->count);
1122 cfs_kunmap(pga[i]->pg);
1127 static int check_write_rcs(struct ptlrpc_request *req,
1128 int requested_nob, int niocount,
1129 obd_count page_count, struct brw_page **pga)
1134 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1135 sizeof(*remote_rcs) *
1137 if (remote_rcs == NULL) {
1138 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1142 /* return error if any niobuf was in error */
1143 for (i = 0; i < niocount; i++) {
1144 if ((int)remote_rcs[i] < 0)
1145 return(remote_rcs[i]);
1147 if (remote_rcs[i] != 0) {
1148 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1149 i, remote_rcs[i], req);
1154 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1155 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1156 req->rq_bulk->bd_nob_transferred, requested_nob);
1163 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1165 if (p1->flag != p2->flag) {
1166 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1167 OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1169 /* warn if we try to combine flags that we don't know to be
1170 * safe to combine */
1171 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1172 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1173 "report this at http://bugs.whamcloud.com/\n",
1174 p1->flag, p2->flag);
1179 return (p1->off + p1->count == p2->off);
1182 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1183 struct brw_page **pga, int opc,
1184 cksum_type_t cksum_type)
1188 struct cfs_crypto_hash_desc *hdesc;
1189 unsigned int bufsize;
1191 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1193 LASSERT(pg_count > 0);
1195 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1196 if (IS_ERR(hdesc)) {
1197 CERROR("Unable to initialize checksum hash %s\n",
1198 cfs_crypto_hash_name(cfs_alg));
1199 return PTR_ERR(hdesc);
1202 while (nob > 0 && pg_count > 0) {
1203 int count = pga[i]->count > nob ? nob : pga[i]->count;
1205 /* corrupt the data before we compute the checksum, to
1206 * simulate an OST->client data error */
1207 if (i == 0 && opc == OST_READ &&
1208 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1209 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1210 int off = pga[i]->off & ~CFS_PAGE_MASK;
1211 memcpy(ptr + off, "bad1", min(4, nob));
1212 cfs_kunmap(pga[i]->pg);
1214 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1215 pga[i]->off & ~CFS_PAGE_MASK,
1217 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1218 (int)(pga[i]->off & ~CFS_PAGE_MASK), cksum);
1220 nob -= pga[i]->count;
1226 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1229 cfs_crypto_hash_final(hdesc, NULL, NULL);
1231 /* For sending we only compute the wrong checksum instead
1232 * of corrupting the data so it is still correct on a redo */
1233 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1239 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1240 struct lov_stripe_md *lsm, obd_count page_count,
1241 struct brw_page **pga,
1242 struct ptlrpc_request **reqp,
1243 struct obd_capa *ocapa, int reserve,
1246 struct ptlrpc_request *req;
1247 struct ptlrpc_bulk_desc *desc;
1248 struct ost_body *body;
1249 struct obd_ioobj *ioobj;
1250 struct niobuf_remote *niobuf;
1251 int niocount, i, requested_nob, opc, rc;
1252 struct osc_brw_async_args *aa;
1253 struct req_capsule *pill;
1254 struct brw_page *pg_prev;
1257 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1258 RETURN(-ENOMEM); /* Recoverable */
1259 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1260 RETURN(-EINVAL); /* Fatal */
1262 if ((cmd & OBD_BRW_WRITE) != 0) {
1264 req = ptlrpc_request_alloc_pool(cli->cl_import,
1265 cli->cl_import->imp_rq_pool,
1266 &RQF_OST_BRW_WRITE);
1269 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1274 for (niocount = i = 1; i < page_count; i++) {
1275 if (!can_merge_pages(pga[i - 1], pga[i]))
1279 pill = &req->rq_pill;
1280 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1282 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1283 niocount * sizeof(*niobuf));
1284 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1286 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1288 ptlrpc_request_free(req);
1291 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1292 ptlrpc_at_set_req_timeout(req);
1293 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1295 req->rq_no_retry_einprogress = 1;
1297 if (opc == OST_WRITE)
1298 desc = ptlrpc_prep_bulk_imp(req, page_count,
1299 BULK_GET_SOURCE, OST_BULK_PORTAL);
1301 desc = ptlrpc_prep_bulk_imp(req, page_count,
1302 BULK_PUT_SINK, OST_BULK_PORTAL);
1305 GOTO(out, rc = -ENOMEM);
1306 /* NB request now owns desc and will free it when it gets freed */
1308 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1309 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1310 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1311 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1313 lustre_set_wire_obdo(&body->oa, oa);
1315 obdo_to_ioobj(oa, ioobj);
1316 ioobj->ioo_bufcnt = niocount;
1317 osc_pack_capa(req, body, ocapa);
1318 LASSERT (page_count > 0);
1320 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1321 struct brw_page *pg = pga[i];
1322 int poff = pg->off & ~CFS_PAGE_MASK;
1324 LASSERT(pg->count > 0);
1325 /* make sure there is no gap in the middle of page array */
1326 LASSERTF(page_count == 1 ||
1327 (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1328 ergo(i > 0 && i < page_count - 1,
1329 poff == 0 && pg->count == CFS_PAGE_SIZE) &&
1330 ergo(i == page_count - 1, poff == 0)),
1331 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1332 i, page_count, pg, pg->off, pg->count);
1334 LASSERTF(i == 0 || pg->off > pg_prev->off,
1335 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1336 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1338 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1339 pg_prev->pg, page_private(pg_prev->pg),
1340 pg_prev->pg->index, pg_prev->off);
1342 LASSERTF(i == 0 || pg->off > pg_prev->off,
1343 "i %d p_c %u\n", i, page_count);
1345 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1346 (pg->flag & OBD_BRW_SRVLOCK));
1348 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1349 requested_nob += pg->count;
1351 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1353 niobuf->len += pg->count;
1355 niobuf->offset = pg->off;
1356 niobuf->len = pg->count;
1357 niobuf->flags = pg->flag;
1362 LASSERTF((void *)(niobuf - niocount) ==
1363 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1364 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1365 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1367 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1369 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1370 body->oa.o_valid |= OBD_MD_FLFLAGS;
1371 body->oa.o_flags = 0;
1373 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1376 if (osc_should_shrink_grant(cli))
1377 osc_shrink_grant_local(cli, &body->oa);
1379 /* size[REQ_REC_OFF] still sizeof (*body) */
1380 if (opc == OST_WRITE) {
1381 if (cli->cl_checksum &&
1382 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1383 /* store cl_cksum_type in a local variable since
1384 * it can be changed via lprocfs */
1385 cksum_type_t cksum_type = cli->cl_cksum_type;
1387 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1388 oa->o_flags &= OBD_FL_LOCAL_MASK;
1389 body->oa.o_flags = 0;
1391 body->oa.o_flags |= cksum_type_pack(cksum_type);
1392 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1393 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1397 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1399 /* save this in 'oa', too, for later checking */
1400 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1401 oa->o_flags |= cksum_type_pack(cksum_type);
1403 /* clear out the checksum flag, in case this is a
1404 * resend but cl_checksum is no longer set. b=11238 */
1405 oa->o_valid &= ~OBD_MD_FLCKSUM;
1407 oa->o_cksum = body->oa.o_cksum;
1408 /* 1 RC per niobuf */
1409 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1410 sizeof(__u32) * niocount);
1412 if (cli->cl_checksum &&
1413 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1414 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1415 body->oa.o_flags = 0;
1416 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1417 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1420 ptlrpc_request_set_replen(req);
1422 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1423 aa = ptlrpc_req_async_args(req);
1425 aa->aa_requested_nob = requested_nob;
1426 aa->aa_nio_count = niocount;
1427 aa->aa_page_count = page_count;
1431 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1432 if (ocapa && reserve)
1433 aa->aa_ocapa = capa_get(ocapa);
1439 ptlrpc_req_finished(req);
1443 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1444 __u32 client_cksum, __u32 server_cksum, int nob,
1445 obd_count page_count, struct brw_page **pga,
1446 cksum_type_t client_cksum_type)
1450 cksum_type_t cksum_type;
1452 if (server_cksum == client_cksum) {
1453 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1457 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1459 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1462 if (cksum_type != client_cksum_type)
1463 msg = "the server did not use the checksum type specified in "
1464 "the original request - likely a protocol problem";
1465 else if (new_cksum == server_cksum)
1466 msg = "changed on the client after we checksummed it - "
1467 "likely false positive due to mmap IO (bug 11742)";
1468 else if (new_cksum == client_cksum)
1469 msg = "changed in transit before arrival at OST";
1471 msg = "changed in transit AND doesn't match the original - "
1472 "likely false positive due to mmap IO (bug 11742)";
1474 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1475 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1476 msg, libcfs_nid2str(peer->nid),
1477 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1478 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1479 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1481 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1483 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1484 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1485 "client csum now %x\n", client_cksum, client_cksum_type,
1486 server_cksum, cksum_type, new_cksum);
1490 /* Note rc enters this function as number of bytes transferred */
1491 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1493 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1494 const lnet_process_id_t *peer =
1495 &req->rq_import->imp_connection->c_peer;
1496 struct client_obd *cli = aa->aa_cli;
1497 struct ost_body *body;
1498 __u32 client_cksum = 0;
1501 if (rc < 0 && rc != -EDQUOT) {
1502 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1506 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1507 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1509 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1513 /* set/clear over quota flag for a uid/gid */
1514 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1515 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1516 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1518 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1519 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1521 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1524 osc_update_grant(cli, body);
1529 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1530 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1532 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1534 CERROR("Unexpected +ve rc %d\n", rc);
1537 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1539 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1542 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1543 check_write_checksum(&body->oa, peer, client_cksum,
1544 body->oa.o_cksum, aa->aa_requested_nob,
1545 aa->aa_page_count, aa->aa_ppga,
1546 cksum_type_unpack(aa->aa_oa->o_flags)))
1549 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1550 aa->aa_page_count, aa->aa_ppga);
1554 /* The rest of this function executes only for OST_READs */
1556 /* if unwrap_bulk failed, return -EAGAIN to retry */
1557 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1559 GOTO(out, rc = -EAGAIN);
1561 if (rc > aa->aa_requested_nob) {
1562 CERROR("Unexpected rc %d (%d requested)\n", rc,
1563 aa->aa_requested_nob);
1567 if (rc != req->rq_bulk->bd_nob_transferred) {
1568 CERROR ("Unexpected rc %d (%d transferred)\n",
1569 rc, req->rq_bulk->bd_nob_transferred);
1573 if (rc < aa->aa_requested_nob)
1574 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1576 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1577 static int cksum_counter;
1578 __u32 server_cksum = body->oa.o_cksum;
1581 cksum_type_t cksum_type;
1583 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1584 body->oa.o_flags : 0);
1585 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1586 aa->aa_ppga, OST_READ,
1589 if (peer->nid == req->rq_bulk->bd_sender) {
1593 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1596 if (server_cksum == ~0 && rc > 0) {
1597 CERROR("Protocol error: server %s set the 'checksum' "
1598 "bit, but didn't send a checksum. Not fatal, "
1599 "but please notify on http://bugs.whamcloud.com/\n",
1600 libcfs_nid2str(peer->nid));
1601 } else if (server_cksum != client_cksum) {
1602 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1603 "%s%s%s inode "DFID" object "
1604 LPU64"/"LPU64" extent "
1605 "["LPU64"-"LPU64"]\n",
1606 req->rq_import->imp_obd->obd_name,
1607 libcfs_nid2str(peer->nid),
1609 body->oa.o_valid & OBD_MD_FLFID ?
1610 body->oa.o_parent_seq : (__u64)0,
1611 body->oa.o_valid & OBD_MD_FLFID ?
1612 body->oa.o_parent_oid : 0,
1613 body->oa.o_valid & OBD_MD_FLFID ?
1614 body->oa.o_parent_ver : 0,
1616 body->oa.o_valid & OBD_MD_FLGROUP ?
1617 body->oa.o_seq : (__u64)0,
1618 aa->aa_ppga[0]->off,
1619 aa->aa_ppga[aa->aa_page_count-1]->off +
1620 aa->aa_ppga[aa->aa_page_count-1]->count -
1622 CERROR("client %x, server %x, cksum_type %x\n",
1623 client_cksum, server_cksum, cksum_type);
1625 aa->aa_oa->o_cksum = client_cksum;
1629 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1632 } else if (unlikely(client_cksum)) {
1633 static int cksum_missed;
1636 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1637 CERROR("Checksum %u requested from %s but not sent\n",
1638 cksum_missed, libcfs_nid2str(peer->nid));
1644 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1649 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1650 struct lov_stripe_md *lsm,
1651 obd_count page_count, struct brw_page **pga,
1652 struct obd_capa *ocapa)
1654 struct ptlrpc_request *req;
1657 int generation, resends = 0;
1658 struct l_wait_info lwi;
1662 cfs_waitq_init(&waitq);
1663 generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1666 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1667 page_count, pga, &req, ocapa, 0, resends);
1672 req->rq_generation_set = 1;
1673 req->rq_import_generation = generation;
1674 req->rq_sent = cfs_time_current_sec() + resends;
1677 rc = ptlrpc_queue_wait(req);
1679 if (rc == -ETIMEDOUT && req->rq_resend) {
1680 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1681 ptlrpc_req_finished(req);
1685 rc = osc_brw_fini_request(req, rc);
1687 ptlrpc_req_finished(req);
1688 /* When server return -EINPROGRESS, client should always retry
1689 * regardless of the number of times the bulk was resent already.*/
1690 if (osc_recoverable_error(rc)) {
1692 if (rc != -EINPROGRESS &&
1693 !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1694 CERROR("%s: too many resend retries for object: "
1695 ""LPU64":"LPU64", rc = %d.\n",
1696 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1700 exp->exp_obd->u.cli.cl_import->imp_generation) {
1701 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1702 ""LPU64":"LPU64", rc = %d.\n",
1703 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1707 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1709 l_wait_event(waitq, 0, &lwi);
1714 if (rc == -EAGAIN || rc == -EINPROGRESS)
1719 static int osc_brw_redo_request(struct ptlrpc_request *request,
1720 struct osc_brw_async_args *aa, int rc)
1722 struct ptlrpc_request *new_req;
1723 struct osc_brw_async_args *new_aa;
1724 struct osc_async_page *oap;
1727 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1728 "redo for recoverable error %d", rc);
1730 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1731 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1732 aa->aa_cli, aa->aa_oa,
1733 NULL /* lsm unused by osc currently */,
1734 aa->aa_page_count, aa->aa_ppga,
1735 &new_req, aa->aa_ocapa, 0, 1);
1739 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1740 if (oap->oap_request != NULL) {
1741 LASSERTF(request == oap->oap_request,
1742 "request %p != oap_request %p\n",
1743 request, oap->oap_request);
1744 if (oap->oap_interrupted) {
1745 ptlrpc_req_finished(new_req);
1750 /* New request takes over pga and oaps from old request.
1751 * Note that copying a list_head doesn't work, need to move it... */
1753 new_req->rq_interpret_reply = request->rq_interpret_reply;
1754 new_req->rq_async_args = request->rq_async_args;
1755 /* cap resend delay to the current request timeout, this is similar to
1756 * what ptlrpc does (see after_reply()) */
1757 if (aa->aa_resends > new_req->rq_timeout)
1758 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1760 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1761 new_req->rq_generation_set = 1;
1762 new_req->rq_import_generation = request->rq_import_generation;
1764 new_aa = ptlrpc_req_async_args(new_req);
1766 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1767 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1768 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1769 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1770 new_aa->aa_resends = aa->aa_resends;
1772 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1773 if (oap->oap_request) {
1774 ptlrpc_req_finished(oap->oap_request);
1775 oap->oap_request = ptlrpc_request_addref(new_req);
1779 new_aa->aa_ocapa = aa->aa_ocapa;
1780 aa->aa_ocapa = NULL;
1782 /* XXX: This code will run into problem if we're going to support
1783 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1784 * and wait for all of them to be finished. We should inherit request
1785 * set from old request. */
1786 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1788 DEBUG_REQ(D_INFO, new_req, "new request");
1793 * ugh, we want disk allocation on the target to happen in offset order. we'll
1794 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1795 * fine for our small page arrays and doesn't require allocation. its an
1796 * insertion sort that swaps elements that are strides apart, shrinking the
1797 * stride down until its '1' and the array is sorted.
1799 static void sort_brw_pages(struct brw_page **array, int num)
1802 struct brw_page *tmp;
1806 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1811 for (i = stride ; i < num ; i++) {
1814 while (j >= stride && array[j - stride]->off > tmp->off) {
1815 array[j] = array[j - stride];
1820 } while (stride > 1);
1823 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1829 LASSERT (pages > 0);
1830 offset = pg[i]->off & ~CFS_PAGE_MASK;
1834 if (pages == 0) /* that's all */
1837 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1838 return count; /* doesn't end on page boundary */
1841 offset = pg[i]->off & ~CFS_PAGE_MASK;
1842 if (offset != 0) /* doesn't start on page boundary */
1849 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1851 struct brw_page **ppga;
1854 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1858 for (i = 0; i < count; i++)
1863 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1865 LASSERT(ppga != NULL);
1866 OBD_FREE(ppga, sizeof(*ppga) * count);
1869 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1870 obd_count page_count, struct brw_page *pga,
1871 struct obd_trans_info *oti)
1873 struct obdo *saved_oa = NULL;
1874 struct brw_page **ppga, **orig;
1875 struct obd_import *imp = class_exp2cliimp(exp);
1876 struct client_obd *cli;
1877 int rc, page_count_orig;
1880 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1881 cli = &imp->imp_obd->u.cli;
1883 if (cmd & OBD_BRW_CHECK) {
1884 /* The caller just wants to know if there's a chance that this
1885 * I/O can succeed */
1887 if (imp->imp_invalid)
1892 /* test_brw with a failed create can trip this, maybe others. */
1893 LASSERT(cli->cl_max_pages_per_rpc);
1897 orig = ppga = osc_build_ppga(pga, page_count);
1900 page_count_orig = page_count;
1902 sort_brw_pages(ppga, page_count);
1903 while (page_count) {
1904 obd_count pages_per_brw;
1906 if (page_count > cli->cl_max_pages_per_rpc)
1907 pages_per_brw = cli->cl_max_pages_per_rpc;
1909 pages_per_brw = page_count;
1911 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1913 if (saved_oa != NULL) {
1914 /* restore previously saved oa */
1915 *oinfo->oi_oa = *saved_oa;
1916 } else if (page_count > pages_per_brw) {
1917 /* save a copy of oa (brw will clobber it) */
1918 OBDO_ALLOC(saved_oa);
1919 if (saved_oa == NULL)
1920 GOTO(out, rc = -ENOMEM);
1921 *saved_oa = *oinfo->oi_oa;
1924 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1925 pages_per_brw, ppga, oinfo->oi_capa);
1930 page_count -= pages_per_brw;
1931 ppga += pages_per_brw;
1935 osc_release_ppga(orig, page_count_orig);
1937 if (saved_oa != NULL)
1938 OBDO_FREE(saved_oa);
1943 static int brw_interpret(const struct lu_env *env,
1944 struct ptlrpc_request *req, void *data, int rc)
1946 struct osc_brw_async_args *aa = data;
1947 struct osc_extent *ext;
1948 struct osc_extent *tmp;
1949 struct cl_object *obj = NULL;
1950 struct client_obd *cli = aa->aa_cli;
1953 rc = osc_brw_fini_request(req, rc);
1954 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1955 /* When server return -EINPROGRESS, client should always retry
1956 * regardless of the number of times the bulk was resent already. */
1957 if (osc_recoverable_error(rc)) {
1958 if (req->rq_import_generation !=
1959 req->rq_import->imp_generation) {
1960 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1961 ""LPU64":"LPU64", rc = %d.\n",
1962 req->rq_import->imp_obd->obd_name,
1963 aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
1964 } else if (rc == -EINPROGRESS ||
1965 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1966 rc = osc_brw_redo_request(req, aa, rc);
1968 CERROR("%s: too many resent retries for object: "
1969 ""LPU64":"LPU64", rc = %d.\n",
1970 req->rq_import->imp_obd->obd_name,
1971 aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
1976 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1981 capa_put(aa->aa_ocapa);
1982 aa->aa_ocapa = NULL;
1985 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1986 if (obj == NULL && rc == 0) {
1987 obj = osc2cl(ext->oe_obj);
1991 cfs_list_del_init(&ext->oe_link);
1992 osc_extent_finish(env, ext, 1, rc);
1994 LASSERT(cfs_list_empty(&aa->aa_exts));
1995 LASSERT(cfs_list_empty(&aa->aa_oaps));
1998 struct obdo *oa = aa->aa_oa;
1999 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2000 unsigned long valid = 0;
2003 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2004 attr->cat_blocks = oa->o_blocks;
2005 valid |= CAT_BLOCKS;
2007 if (oa->o_valid & OBD_MD_FLMTIME) {
2008 attr->cat_mtime = oa->o_mtime;
2011 if (oa->o_valid & OBD_MD_FLATIME) {
2012 attr->cat_atime = oa->o_atime;
2015 if (oa->o_valid & OBD_MD_FLCTIME) {
2016 attr->cat_ctime = oa->o_ctime;
2020 cl_object_attr_lock(obj);
2021 cl_object_attr_set(env, obj, attr, valid);
2022 cl_object_attr_unlock(obj);
2024 cl_object_put(env, obj);
2026 OBDO_FREE(aa->aa_oa);
2028 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2029 req->rq_bulk->bd_nob_transferred);
2030 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2031 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2033 client_obd_list_lock(&cli->cl_loi_list_lock);
2034 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2035 * is called so we know whether to go to sync BRWs or wait for more
2036 * RPCs to complete */
2037 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2038 cli->cl_w_in_flight--;
2040 cli->cl_r_in_flight--;
2041 osc_wake_cache_waiters(cli);
2042 client_obd_list_unlock(&cli->cl_loi_list_lock);
2044 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2049 * Build an RPC by the list of extent @ext_list. The caller must ensure
2050 * that the total pages in this list are NOT over max pages per RPC.
2051 * Extents in the list must be in OES_RPC state.
2053 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2054 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2056 struct ptlrpc_request *req = NULL;
2057 struct osc_extent *ext;
2058 CFS_LIST_HEAD(rpc_list);
2059 struct brw_page **pga = NULL;
2060 struct osc_brw_async_args *aa = NULL;
2061 struct obdo *oa = NULL;
2062 struct osc_async_page *oap;
2063 struct osc_async_page *tmp;
2064 struct cl_req *clerq = NULL;
2065 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2066 struct ldlm_lock *lock = NULL;
2067 struct cl_req_attr crattr;
2068 obd_off starting_offset = OBD_OBJECT_EOF;
2069 obd_off ending_offset = 0;
2070 int i, rc, mpflag = 0, mem_tight = 0, page_count = 0;
2073 LASSERT(!cfs_list_empty(ext_list));
2075 /* add pages into rpc_list to build BRW rpc */
2076 cfs_list_for_each_entry(ext, ext_list, oe_link) {
2077 LASSERT(ext->oe_state == OES_RPC);
2078 mem_tight |= ext->oe_memalloc;
2079 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2081 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2082 if (starting_offset > oap->oap_obj_off)
2083 starting_offset = oap->oap_obj_off;
2085 LASSERT(oap->oap_page_off == 0);
2086 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2087 ending_offset = oap->oap_obj_off +
2090 LASSERT(oap->oap_page_off + oap->oap_count ==
2096 mpflag = cfs_memory_pressure_get_and_set();
2098 memset(&crattr, 0, sizeof crattr);
2099 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2101 GOTO(out, rc = -ENOMEM);
2105 GOTO(out, rc = -ENOMEM);
2108 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2109 struct cl_page *page = oap2cl_page(oap);
2110 if (clerq == NULL) {
2111 clerq = cl_req_alloc(env, page, crt,
2112 1 /* only 1-object rpcs for
2115 GOTO(out, rc = PTR_ERR(clerq));
2116 lock = oap->oap_ldlm_lock;
2119 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2120 pga[i] = &oap->oap_brw_page;
2121 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2122 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2123 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2125 cl_req_page_add(env, clerq, page);
2128 /* always get the data for the obdo for the rpc */
2129 LASSERT(clerq != NULL);
2131 crattr.cra_capa = NULL;
2132 memset(crattr.cra_jobid, 0, JOBSTATS_JOBID_SIZE);
2133 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2135 oa->o_handle = lock->l_remote_handle;
2136 oa->o_valid |= OBD_MD_FLHANDLE;
2139 rc = cl_req_prep(env, clerq);
2141 CERROR("cl_req_prep failed: %d\n", rc);
2145 sort_brw_pages(pga, page_count);
2146 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2147 pga, &req, crattr.cra_capa, 1, 0);
2149 CERROR("prep_req failed: %d\n", rc);
2153 req->rq_interpret_reply = brw_interpret;
2155 req->rq_memalloc = 1;
2157 /* Need to update the timestamps after the request is built in case
2158 * we race with setattr (locally or in queue at OST). If OST gets
2159 * later setattr before earlier BRW (as determined by the request xid),
2160 * the OST will not use BRW timestamps. Sadly, there is no obvious
2161 * way to do this in a single call. bug 10150 */
2162 cl_req_attr_set(env, clerq, &crattr,
2163 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2165 lustre_msg_set_jobid(req->rq_reqmsg, crattr.cra_jobid);
2167 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2168 aa = ptlrpc_req_async_args(req);
2169 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2170 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2171 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2172 cfs_list_splice_init(ext_list, &aa->aa_exts);
2173 aa->aa_clerq = clerq;
2175 /* queued sync pages can be torn down while the pages
2176 * were between the pending list and the rpc */
2178 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2179 /* only one oap gets a request reference */
2182 if (oap->oap_interrupted && !req->rq_intr) {
2183 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2185 ptlrpc_mark_interrupted(req);
2189 tmp->oap_request = ptlrpc_request_addref(req);
2191 client_obd_list_lock(&cli->cl_loi_list_lock);
2192 starting_offset >>= CFS_PAGE_SHIFT;
2193 if (cmd == OBD_BRW_READ) {
2194 cli->cl_r_in_flight++;
2195 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2196 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2197 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2198 starting_offset + 1);
2200 cli->cl_w_in_flight++;
2201 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2202 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2203 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2204 starting_offset + 1);
2206 client_obd_list_unlock(&cli->cl_loi_list_lock);
2208 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2209 page_count, aa, cli->cl_r_in_flight,
2210 cli->cl_w_in_flight);
2212 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2213 * see which CPU/NUMA node the majority of pages were allocated
2214 * on, and try to assign the async RPC to the CPU core
2215 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2217 * But on the other hand, we expect that multiple ptlrpcd
2218 * threads and the initial write sponsor can run in parallel,
2219 * especially when data checksum is enabled, which is CPU-bound
2220 * operation and single ptlrpcd thread cannot process in time.
2221 * So more ptlrpcd threads sharing BRW load
2222 * (with PDL_POLICY_ROUND) seems better.
2224 ptlrpcd_add_req(req, pol, -1);
2230 cfs_memory_pressure_restore(mpflag);
2232 capa_put(crattr.cra_capa);
2234 LASSERT(req == NULL);
2239 OBD_FREE(pga, sizeof(*pga) * page_count);
2240 /* this should happen rarely and is pretty bad, it makes the
2241 * pending list not follow the dirty order */
2242 while (!cfs_list_empty(ext_list)) {
2243 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2245 cfs_list_del_init(&ext->oe_link);
2246 osc_extent_finish(env, ext, 0, rc);
2248 if (clerq && !IS_ERR(clerq))
2249 cl_req_completion(env, clerq, rc);
2254 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2255 struct ldlm_enqueue_info *einfo)
2257 void *data = einfo->ei_cbdata;
2260 LASSERT(lock != NULL);
2261 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2262 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2263 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2264 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2266 lock_res_and_lock(lock);
2267 spin_lock(&osc_ast_guard);
2269 if (lock->l_ast_data == NULL)
2270 lock->l_ast_data = data;
2271 if (lock->l_ast_data == data)
2274 spin_unlock(&osc_ast_guard);
2275 unlock_res_and_lock(lock);
2280 static int osc_set_data_with_check(struct lustre_handle *lockh,
2281 struct ldlm_enqueue_info *einfo)
2283 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2287 set = osc_set_lock_data_with_check(lock, einfo);
2288 LDLM_LOCK_PUT(lock);
2290 CERROR("lockh %p, data %p - client evicted?\n",
2291 lockh, einfo->ei_cbdata);
2295 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2296 ldlm_iterator_t replace, void *data)
2298 struct ldlm_res_id res_id;
2299 struct obd_device *obd = class_exp2obd(exp);
2301 ostid_build_res_name(&lsm->lsm_object_oid, &res_id);
2302 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2306 /* find any ldlm lock of the inode in osc
2310 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2311 ldlm_iterator_t replace, void *data)
2313 struct ldlm_res_id res_id;
2314 struct obd_device *obd = class_exp2obd(exp);
2317 ostid_build_res_name(&lsm->lsm_object_oid, &res_id);
2318 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2319 if (rc == LDLM_ITER_STOP)
2321 if (rc == LDLM_ITER_CONTINUE)
2326 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2327 obd_enqueue_update_f upcall, void *cookie,
2328 __u64 *flags, int agl, int rc)
2330 int intent = *flags & LDLM_FL_HAS_INTENT;
2334 /* The request was created before ldlm_cli_enqueue call. */
2335 if (rc == ELDLM_LOCK_ABORTED) {
2336 struct ldlm_reply *rep;
2337 rep = req_capsule_server_get(&req->rq_pill,
2340 LASSERT(rep != NULL);
2341 if (rep->lock_policy_res1)
2342 rc = rep->lock_policy_res1;
2346 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2348 *flags |= LDLM_FL_LVB_READY;
2349 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2350 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2353 /* Call the update callback. */
2354 rc = (*upcall)(cookie, rc);
2358 static int osc_enqueue_interpret(const struct lu_env *env,
2359 struct ptlrpc_request *req,
2360 struct osc_enqueue_args *aa, int rc)
2362 struct ldlm_lock *lock;
2363 struct lustre_handle handle;
2365 struct ost_lvb *lvb;
2367 __u64 *flags = aa->oa_flags;
2369 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2370 * might be freed anytime after lock upcall has been called. */
2371 lustre_handle_copy(&handle, aa->oa_lockh);
2372 mode = aa->oa_ei->ei_mode;
2374 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2376 lock = ldlm_handle2lock(&handle);
2378 /* Take an additional reference so that a blocking AST that
2379 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2380 * to arrive after an upcall has been executed by
2381 * osc_enqueue_fini(). */
2382 ldlm_lock_addref(&handle, mode);
2384 /* Let CP AST to grant the lock first. */
2385 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2387 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2392 lvb_len = sizeof(*aa->oa_lvb);
2395 /* Complete obtaining the lock procedure. */
2396 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2397 mode, flags, lvb, lvb_len, &handle, rc);
2398 /* Complete osc stuff. */
2399 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2400 flags, aa->oa_agl, rc);
2402 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2404 /* Release the lock for async request. */
2405 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2407 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2408 * not already released by
2409 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2411 ldlm_lock_decref(&handle, mode);
2413 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2414 aa->oa_lockh, req, aa);
2415 ldlm_lock_decref(&handle, mode);
2416 LDLM_LOCK_PUT(lock);
2420 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2421 struct lov_oinfo *loi, int flags,
2422 struct ost_lvb *lvb, __u32 mode, int rc)
2424 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2426 if (rc == ELDLM_OK) {
2429 LASSERT(lock != NULL);
2430 loi->loi_lvb = *lvb;
2431 tmp = loi->loi_lvb.lvb_size;
2432 /* Extend KMS up to the end of this lock and no further
2433 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2434 if (tmp > lock->l_policy_data.l_extent.end)
2435 tmp = lock->l_policy_data.l_extent.end + 1;
2436 if (tmp >= loi->loi_kms) {
2437 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2438 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2439 loi_kms_set(loi, tmp);
2441 LDLM_DEBUG(lock, "lock acquired, setting rss="
2442 LPU64"; leaving kms="LPU64", end="LPU64,
2443 loi->loi_lvb.lvb_size, loi->loi_kms,
2444 lock->l_policy_data.l_extent.end);
2446 ldlm_lock_allow_match(lock);
2447 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2448 LASSERT(lock != NULL);
2449 loi->loi_lvb = *lvb;
2450 ldlm_lock_allow_match(lock);
2451 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2452 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2458 ldlm_lock_fail_match(lock);
2460 LDLM_LOCK_PUT(lock);
2463 EXPORT_SYMBOL(osc_update_enqueue);
2465 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2467 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2468 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2469 * other synchronous requests, however keeping some locks and trying to obtain
2470 * others may take a considerable amount of time in a case of ost failure; and
2471 * when other sync requests do not get released lock from a client, the client
2472 * is excluded from the cluster -- such scenarious make the life difficult, so
2473 * release locks just after they are obtained. */
2474 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2475 __u64 *flags, ldlm_policy_data_t *policy,
2476 struct ost_lvb *lvb, int kms_valid,
2477 obd_enqueue_update_f upcall, void *cookie,
2478 struct ldlm_enqueue_info *einfo,
2479 struct lustre_handle *lockh,
2480 struct ptlrpc_request_set *rqset, int async, int agl)
2482 struct obd_device *obd = exp->exp_obd;
2483 struct ptlrpc_request *req = NULL;
2484 int intent = *flags & LDLM_FL_HAS_INTENT;
2485 int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2490 /* Filesystem lock extents are extended to page boundaries so that
2491 * dealing with the page cache is a little smoother. */
2492 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2493 policy->l_extent.end |= ~CFS_PAGE_MASK;
2496 * kms is not valid when either object is completely fresh (so that no
2497 * locks are cached), or object was evicted. In the latter case cached
2498 * lock cannot be used, because it would prime inode state with
2499 * potentially stale LVB.
2504 /* Next, search for already existing extent locks that will cover us */
2505 /* If we're trying to read, we also search for an existing PW lock. The
2506 * VFS and page cache already protect us locally, so lots of readers/
2507 * writers can share a single PW lock.
2509 * There are problems with conversion deadlocks, so instead of
2510 * converting a read lock to a write lock, we'll just enqueue a new
2513 * At some point we should cancel the read lock instead of making them
2514 * send us a blocking callback, but there are problems with canceling
2515 * locks out from other users right now, too. */
2516 mode = einfo->ei_mode;
2517 if (einfo->ei_mode == LCK_PR)
2519 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2520 einfo->ei_type, policy, mode, lockh, 0);
2522 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2524 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2525 /* For AGL, if enqueue RPC is sent but the lock is not
2526 * granted, then skip to process this strpe.
2527 * Return -ECANCELED to tell the caller. */
2528 ldlm_lock_decref(lockh, mode);
2529 LDLM_LOCK_PUT(matched);
2531 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2532 *flags |= LDLM_FL_LVB_READY;
2533 /* addref the lock only if not async requests and PW
2534 * lock is matched whereas we asked for PR. */
2535 if (!rqset && einfo->ei_mode != mode)
2536 ldlm_lock_addref(lockh, LCK_PR);
2538 /* I would like to be able to ASSERT here that
2539 * rss <= kms, but I can't, for reasons which
2540 * are explained in lov_enqueue() */
2543 /* We already have a lock, and it's referenced.
2545 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2546 * AGL upcall may change it to CLS_HELD directly. */
2547 (*upcall)(cookie, ELDLM_OK);
2549 if (einfo->ei_mode != mode)
2550 ldlm_lock_decref(lockh, LCK_PW);
2552 /* For async requests, decref the lock. */
2553 ldlm_lock_decref(lockh, einfo->ei_mode);
2554 LDLM_LOCK_PUT(matched);
2557 ldlm_lock_decref(lockh, mode);
2558 LDLM_LOCK_PUT(matched);
2564 CFS_LIST_HEAD(cancels);
2565 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2566 &RQF_LDLM_ENQUEUE_LVB);
2570 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2572 ptlrpc_request_free(req);
2576 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2578 ptlrpc_request_set_replen(req);
2581 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2582 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2584 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2585 sizeof(*lvb), LVB_T_OST, lockh, async);
2588 struct osc_enqueue_args *aa;
2589 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2590 aa = ptlrpc_req_async_args(req);
2593 aa->oa_flags = flags;
2594 aa->oa_upcall = upcall;
2595 aa->oa_cookie = cookie;
2597 aa->oa_lockh = lockh;
2600 req->rq_interpret_reply =
2601 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2602 if (rqset == PTLRPCD_SET)
2603 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2605 ptlrpc_set_add_req(rqset, req);
2606 } else if (intent) {
2607 ptlrpc_req_finished(req);
2612 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2614 ptlrpc_req_finished(req);
2619 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2620 struct ldlm_enqueue_info *einfo,
2621 struct ptlrpc_request_set *rqset)
2623 struct ldlm_res_id res_id;
2627 ostid_build_res_name(&oinfo->oi_md->lsm_object_oid, &res_id);
2628 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2629 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2630 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2631 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2632 rqset, rqset != NULL, 0);
2636 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2637 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2638 int *flags, void *data, struct lustre_handle *lockh,
2641 struct obd_device *obd = exp->exp_obd;
2642 int lflags = *flags;
2646 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2649 /* Filesystem lock extents are extended to page boundaries so that
2650 * dealing with the page cache is a little smoother */
2651 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2652 policy->l_extent.end |= ~CFS_PAGE_MASK;
2654 /* Next, search for already existing extent locks that will cover us */
2655 /* If we're trying to read, we also search for an existing PW lock. The
2656 * VFS and page cache already protect us locally, so lots of readers/
2657 * writers can share a single PW lock. */
2661 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2662 res_id, type, policy, rc, lockh, unref);
2665 if (!osc_set_data_with_check(lockh, data)) {
2666 if (!(lflags & LDLM_FL_TEST_LOCK))
2667 ldlm_lock_decref(lockh, rc);
2671 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2672 ldlm_lock_addref(lockh, LCK_PR);
2673 ldlm_lock_decref(lockh, LCK_PW);
2680 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2684 if (unlikely(mode == LCK_GROUP))
2685 ldlm_lock_decref_and_cancel(lockh, mode);
2687 ldlm_lock_decref(lockh, mode);
2692 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2693 __u32 mode, struct lustre_handle *lockh)
2696 RETURN(osc_cancel_base(lockh, mode));
2699 static int osc_cancel_unused(struct obd_export *exp,
2700 struct lov_stripe_md *lsm,
2701 ldlm_cancel_flags_t flags,
2704 struct obd_device *obd = class_exp2obd(exp);
2705 struct ldlm_res_id res_id, *resp = NULL;
2708 ostid_build_res_name(&lsm->lsm_object_oid, &res_id);
2712 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2715 static int osc_statfs_interpret(const struct lu_env *env,
2716 struct ptlrpc_request *req,
2717 struct osc_async_args *aa, int rc)
2719 struct obd_statfs *msfs;
2723 /* The request has in fact never been sent
2724 * due to issues at a higher level (LOV).
2725 * Exit immediately since the caller is
2726 * aware of the problem and takes care
2727 * of the clean up */
2730 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2731 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2737 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2739 GOTO(out, rc = -EPROTO);
2742 *aa->aa_oi->oi_osfs = *msfs;
2744 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2748 static int osc_statfs_async(struct obd_export *exp,
2749 struct obd_info *oinfo, __u64 max_age,
2750 struct ptlrpc_request_set *rqset)
2752 struct obd_device *obd = class_exp2obd(exp);
2753 struct ptlrpc_request *req;
2754 struct osc_async_args *aa;
2758 /* We could possibly pass max_age in the request (as an absolute
2759 * timestamp or a "seconds.usec ago") so the target can avoid doing
2760 * extra calls into the filesystem if that isn't necessary (e.g.
2761 * during mount that would help a bit). Having relative timestamps
2762 * is not so great if request processing is slow, while absolute
2763 * timestamps are not ideal because they need time synchronization. */
2764 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2768 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2770 ptlrpc_request_free(req);
2773 ptlrpc_request_set_replen(req);
2774 req->rq_request_portal = OST_CREATE_PORTAL;
2775 ptlrpc_at_set_req_timeout(req);
2777 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2778 /* procfs requests not want stat in wait for avoid deadlock */
2779 req->rq_no_resend = 1;
2780 req->rq_no_delay = 1;
2783 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2784 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2785 aa = ptlrpc_req_async_args(req);
2788 ptlrpc_set_add_req(rqset, req);
2792 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2793 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2795 struct obd_device *obd = class_exp2obd(exp);
2796 struct obd_statfs *msfs;
2797 struct ptlrpc_request *req;
2798 struct obd_import *imp = NULL;
2802 /*Since the request might also come from lprocfs, so we need
2803 *sync this with client_disconnect_export Bug15684*/
2804 down_read(&obd->u.cli.cl_sem);
2805 if (obd->u.cli.cl_import)
2806 imp = class_import_get(obd->u.cli.cl_import);
2807 up_read(&obd->u.cli.cl_sem);
2811 /* We could possibly pass max_age in the request (as an absolute
2812 * timestamp or a "seconds.usec ago") so the target can avoid doing
2813 * extra calls into the filesystem if that isn't necessary (e.g.
2814 * during mount that would help a bit). Having relative timestamps
2815 * is not so great if request processing is slow, while absolute
2816 * timestamps are not ideal because they need time synchronization. */
2817 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2819 class_import_put(imp);
2824 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2826 ptlrpc_request_free(req);
2829 ptlrpc_request_set_replen(req);
2830 req->rq_request_portal = OST_CREATE_PORTAL;
2831 ptlrpc_at_set_req_timeout(req);
2833 if (flags & OBD_STATFS_NODELAY) {
2834 /* procfs requests not want stat in wait for avoid deadlock */
2835 req->rq_no_resend = 1;
2836 req->rq_no_delay = 1;
2839 rc = ptlrpc_queue_wait(req);
2843 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2845 GOTO(out, rc = -EPROTO);
2852 ptlrpc_req_finished(req);
2856 /* Retrieve object striping information.
2858 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2859 * the maximum number of OST indices which will fit in the user buffer.
2860 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2862 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2864 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2865 struct lov_user_md_v3 lum, *lumk;
2866 struct lov_user_ost_data_v1 *lmm_objects;
2867 int rc = 0, lum_size;
2873 /* we only need the header part from user space to get lmm_magic and
2874 * lmm_stripe_count, (the header part is common to v1 and v3) */
2875 lum_size = sizeof(struct lov_user_md_v1);
2876 if (cfs_copy_from_user(&lum, lump, lum_size))
2879 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2880 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2883 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2884 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2885 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2886 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2888 /* we can use lov_mds_md_size() to compute lum_size
2889 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2890 if (lum.lmm_stripe_count > 0) {
2891 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2892 OBD_ALLOC(lumk, lum_size);
2896 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2897 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2899 lmm_objects = &(lumk->lmm_objects[0]);
2900 lmm_objects->l_object_id = lsm->lsm_object_id;
2902 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2906 lumk->lmm_object_id = lsm->lsm_object_id;
2907 lumk->lmm_object_seq = lsm->lsm_object_seq;
2908 lumk->lmm_stripe_count = 1;
2910 if (cfs_copy_to_user(lump, lumk, lum_size))
2914 OBD_FREE(lumk, lum_size);
2920 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2921 void *karg, void *uarg)
2923 struct obd_device *obd = exp->exp_obd;
2924 struct obd_ioctl_data *data = karg;
2928 if (!cfs_try_module_get(THIS_MODULE)) {
2929 CERROR("Can't get module. Is it alive?");
2933 case OBD_IOC_LOV_GET_CONFIG: {
2935 struct lov_desc *desc;
2936 struct obd_uuid uuid;
2940 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2941 GOTO(out, err = -EINVAL);
2943 data = (struct obd_ioctl_data *)buf;
2945 if (sizeof(*desc) > data->ioc_inllen1) {
2946 obd_ioctl_freedata(buf, len);
2947 GOTO(out, err = -EINVAL);
2950 if (data->ioc_inllen2 < sizeof(uuid)) {
2951 obd_ioctl_freedata(buf, len);
2952 GOTO(out, err = -EINVAL);
2955 desc = (struct lov_desc *)data->ioc_inlbuf1;
2956 desc->ld_tgt_count = 1;
2957 desc->ld_active_tgt_count = 1;
2958 desc->ld_default_stripe_count = 1;
2959 desc->ld_default_stripe_size = 0;
2960 desc->ld_default_stripe_offset = 0;
2961 desc->ld_pattern = 0;
2962 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2964 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2966 err = cfs_copy_to_user((void *)uarg, buf, len);
2969 obd_ioctl_freedata(buf, len);
2972 case LL_IOC_LOV_SETSTRIPE:
2973 err = obd_alloc_memmd(exp, karg);
2977 case LL_IOC_LOV_GETSTRIPE:
2978 err = osc_getstripe(karg, uarg);
2980 case OBD_IOC_CLIENT_RECOVER:
2981 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2982 data->ioc_inlbuf1, 0);
2986 case IOC_OSC_SET_ACTIVE:
2987 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2990 case OBD_IOC_POLL_QUOTACHECK:
2991 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2993 case OBD_IOC_PING_TARGET:
2994 err = ptlrpc_obd_ping(obd);
2997 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2998 cmd, cfs_curproc_comm());
2999 GOTO(out, err = -ENOTTY);
3002 cfs_module_put(THIS_MODULE);
3006 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3007 obd_count keylen, void *key, __u32 *vallen, void *val,
3008 struct lov_stripe_md *lsm)
3011 if (!vallen || !val)
3014 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3015 __u32 *stripe = val;
3016 *vallen = sizeof(*stripe);
3019 } else if (KEY_IS(KEY_LAST_ID)) {
3020 struct ptlrpc_request *req;
3025 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3026 &RQF_OST_GET_INFO_LAST_ID);
3030 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3031 RCL_CLIENT, keylen);
3032 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3034 ptlrpc_request_free(req);
3038 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3039 memcpy(tmp, key, keylen);
3041 req->rq_no_delay = req->rq_no_resend = 1;
3042 ptlrpc_request_set_replen(req);
3043 rc = ptlrpc_queue_wait(req);
3047 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3049 GOTO(out, rc = -EPROTO);
3051 *((obd_id *)val) = *reply;
3053 ptlrpc_req_finished(req);
3055 } else if (KEY_IS(KEY_FIEMAP)) {
3056 struct ptlrpc_request *req;
3057 struct ll_user_fiemap *reply;
3061 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3062 &RQF_OST_GET_INFO_FIEMAP);
3066 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3067 RCL_CLIENT, keylen);
3068 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3069 RCL_CLIENT, *vallen);
3070 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3071 RCL_SERVER, *vallen);
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_FIEMAP_KEY);
3080 memcpy(tmp, key, keylen);
3081 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3082 memcpy(tmp, val, *vallen);
3084 ptlrpc_request_set_replen(req);
3085 rc = ptlrpc_queue_wait(req);
3089 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3091 GOTO(out1, rc = -EPROTO);
3093 memcpy(val, reply, *vallen);
3095 ptlrpc_req_finished(req);
3103 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3104 obd_count keylen, void *key, obd_count vallen,
3105 void *val, struct ptlrpc_request_set *set)
3107 struct ptlrpc_request *req;
3108 struct obd_device *obd = exp->exp_obd;
3109 struct obd_import *imp = class_exp2cliimp(exp);
3114 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3116 if (KEY_IS(KEY_CHECKSUM)) {
3117 if (vallen != sizeof(int))
3119 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3123 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3124 sptlrpc_conf_client_adapt(obd);
3128 if (KEY_IS(KEY_FLUSH_CTX)) {
3129 sptlrpc_import_flush_my_ctx(imp);
3133 if (KEY_IS(KEY_CACHE_SET)) {
3134 struct client_obd *cli = &obd->u.cli;
3136 LASSERT(cli->cl_cache == NULL); /* only once */
3137 cli->cl_cache = (struct cl_client_cache *)val;
3138 cfs_atomic_inc(&cli->cl_cache->ccc_users);
3139 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
3141 /* add this osc into entity list */
3142 LASSERT(cfs_list_empty(&cli->cl_lru_osc));
3143 spin_lock(&cli->cl_cache->ccc_lru_lock);
3144 cfs_list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
3145 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3150 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3151 struct client_obd *cli = &obd->u.cli;
3152 int nr = cfs_atomic_read(&cli->cl_lru_in_list) >> 1;
3153 int target = *(int *)val;
3155 nr = osc_lru_shrink(cli, min(nr, target));
3160 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3163 /* We pass all other commands directly to OST. Since nobody calls osc
3164 methods directly and everybody is supposed to go through LOV, we
3165 assume lov checked invalid values for us.
3166 The only recognised values so far are evict_by_nid and mds_conn.
3167 Even if something bad goes through, we'd get a -EINVAL from OST
3170 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3171 &RQF_OST_SET_GRANT_INFO :
3176 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3177 RCL_CLIENT, keylen);
3178 if (!KEY_IS(KEY_GRANT_SHRINK))
3179 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3180 RCL_CLIENT, vallen);
3181 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3183 ptlrpc_request_free(req);
3187 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3188 memcpy(tmp, key, keylen);
3189 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3192 memcpy(tmp, val, vallen);
3194 if (KEY_IS(KEY_GRANT_SHRINK)) {
3195 struct osc_grant_args *aa;
3198 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3199 aa = ptlrpc_req_async_args(req);
3202 ptlrpc_req_finished(req);
3205 *oa = ((struct ost_body *)val)->oa;
3207 req->rq_interpret_reply = osc_shrink_grant_interpret;
3210 ptlrpc_request_set_replen(req);
3211 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3212 LASSERT(set != NULL);
3213 ptlrpc_set_add_req(set, req);
3214 ptlrpc_check_set(NULL, set);
3216 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3222 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3223 struct obd_device *disk_obd, int *index)
3225 /* this code is not supposed to be used with LOD/OSP
3226 * to be removed soon */
3231 static int osc_llog_finish(struct obd_device *obd, int count)
3233 struct llog_ctxt *ctxt;
3237 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3239 llog_cat_close(NULL, ctxt->loc_handle);
3240 llog_cleanup(NULL, ctxt);
3243 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3245 llog_cleanup(NULL, ctxt);
3249 static int osc_reconnect(const struct lu_env *env,
3250 struct obd_export *exp, struct obd_device *obd,
3251 struct obd_uuid *cluuid,
3252 struct obd_connect_data *data,
3255 struct client_obd *cli = &obd->u.cli;
3257 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3260 client_obd_list_lock(&cli->cl_loi_list_lock);
3261 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3262 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3263 lost_grant = cli->cl_lost_grant;
3264 cli->cl_lost_grant = 0;
3265 client_obd_list_unlock(&cli->cl_loi_list_lock);
3267 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3268 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3269 data->ocd_version, data->ocd_grant, lost_grant);
3275 static int osc_disconnect(struct obd_export *exp)
3277 struct obd_device *obd = class_exp2obd(exp);
3278 struct llog_ctxt *ctxt;
3281 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3283 if (obd->u.cli.cl_conn_count == 1) {
3284 /* Flush any remaining cancel messages out to the
3286 llog_sync(ctxt, exp, 0);
3288 llog_ctxt_put(ctxt);
3290 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3294 rc = client_disconnect_export(exp);
3296 * Initially we put del_shrink_grant before disconnect_export, but it
3297 * causes the following problem if setup (connect) and cleanup
3298 * (disconnect) are tangled together.
3299 * connect p1 disconnect p2
3300 * ptlrpc_connect_import
3301 * ............... class_manual_cleanup
3304 * ptlrpc_connect_interrupt
3306 * add this client to shrink list
3308 * Bang! pinger trigger the shrink.
3309 * So the osc should be disconnected from the shrink list, after we
3310 * are sure the import has been destroyed. BUG18662
3312 if (obd->u.cli.cl_import == NULL)
3313 osc_del_shrink_grant(&obd->u.cli);
3317 static int osc_import_event(struct obd_device *obd,
3318 struct obd_import *imp,
3319 enum obd_import_event event)
3321 struct client_obd *cli;
3325 LASSERT(imp->imp_obd == obd);
3328 case IMP_EVENT_DISCON: {
3330 client_obd_list_lock(&cli->cl_loi_list_lock);
3331 cli->cl_avail_grant = 0;
3332 cli->cl_lost_grant = 0;
3333 client_obd_list_unlock(&cli->cl_loi_list_lock);
3336 case IMP_EVENT_INACTIVE: {
3337 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3340 case IMP_EVENT_INVALIDATE: {
3341 struct ldlm_namespace *ns = obd->obd_namespace;
3345 env = cl_env_get(&refcheck);
3349 /* all pages go to failing rpcs due to the invalid
3351 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3353 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3354 cl_env_put(env, &refcheck);
3359 case IMP_EVENT_ACTIVE: {
3360 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3363 case IMP_EVENT_OCD: {
3364 struct obd_connect_data *ocd = &imp->imp_connect_data;
3366 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3367 osc_init_grant(&obd->u.cli, ocd);
3370 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3371 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3373 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3376 case IMP_EVENT_DEACTIVATE: {
3377 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3380 case IMP_EVENT_ACTIVATE: {
3381 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3385 CERROR("Unknown import event %d\n", event);
3392 * Determine whether the lock can be canceled before replaying the lock
3393 * during recovery, see bug16774 for detailed information.
3395 * \retval zero the lock can't be canceled
3396 * \retval other ok to cancel
3398 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3400 check_res_locked(lock->l_resource);
3403 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3405 * XXX as a future improvement, we can also cancel unused write lock
3406 * if it doesn't have dirty data and active mmaps.
3408 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3409 (lock->l_granted_mode == LCK_PR ||
3410 lock->l_granted_mode == LCK_CR) &&
3411 (osc_dlm_lock_pageref(lock) == 0))
3417 static int brw_queue_work(const struct lu_env *env, void *data)
3419 struct client_obd *cli = data;
3421 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3423 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3427 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3429 struct lprocfs_static_vars lvars = { 0 };
3430 struct client_obd *cli = &obd->u.cli;
3435 rc = ptlrpcd_addref();
3439 rc = client_obd_setup(obd, lcfg);
3441 GOTO(out_ptlrpcd, rc);
3443 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3444 if (IS_ERR(handler))
3445 GOTO(out_client_setup, rc = PTR_ERR(handler));
3446 cli->cl_writeback_work = handler;
3448 rc = osc_quota_setup(obd);
3450 GOTO(out_ptlrpcd_work, rc);
3452 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3453 lprocfs_osc_init_vars(&lvars);
3454 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3455 lproc_osc_attach_seqstat(obd);
3456 sptlrpc_lprocfs_cliobd_attach(obd);
3457 ptlrpc_lprocfs_register_obd(obd);
3460 /* We need to allocate a few requests more, because
3461 * brw_interpret tries to create new requests before freeing
3462 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3463 * reserved, but I'm afraid that might be too much wasted RAM
3464 * in fact, so 2 is just my guess and still should work. */
3465 cli->cl_import->imp_rq_pool =
3466 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3468 ptlrpc_add_rqs_to_pool);
3470 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3471 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3475 ptlrpcd_destroy_work(handler);
3477 client_obd_cleanup(obd);
3483 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3489 case OBD_CLEANUP_EARLY: {
3490 struct obd_import *imp;
3491 imp = obd->u.cli.cl_import;
3492 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3493 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3494 ptlrpc_deactivate_import(imp);
3495 spin_lock(&imp->imp_lock);
3496 imp->imp_pingable = 0;
3497 spin_unlock(&imp->imp_lock);
3500 case OBD_CLEANUP_EXPORTS: {
3501 struct client_obd *cli = &obd->u.cli;
3503 * for echo client, export may be on zombie list, wait for
3504 * zombie thread to cull it, because cli.cl_import will be
3505 * cleared in client_disconnect_export():
3506 * class_export_destroy() -> obd_cleanup() ->
3507 * echo_device_free() -> echo_client_cleanup() ->
3508 * obd_disconnect() -> osc_disconnect() ->
3509 * client_disconnect_export()
3511 obd_zombie_barrier();
3512 if (cli->cl_writeback_work) {
3513 ptlrpcd_destroy_work(cli->cl_writeback_work);
3514 cli->cl_writeback_work = NULL;
3516 obd_cleanup_client_import(obd);
3517 ptlrpc_lprocfs_unregister_obd(obd);
3518 lprocfs_obd_cleanup(obd);
3519 rc = obd_llog_finish(obd, 0);
3521 CERROR("failed to cleanup llogging subsystems\n");
3528 int osc_cleanup(struct obd_device *obd)
3530 struct client_obd *cli = &obd->u.cli;
3536 if (cli->cl_cache != NULL) {
3537 LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_users) > 0);
3538 spin_lock(&cli->cl_cache->ccc_lru_lock);
3539 cfs_list_del_init(&cli->cl_lru_osc);
3540 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3541 cli->cl_lru_left = NULL;
3542 cfs_atomic_dec(&cli->cl_cache->ccc_users);
3543 cli->cl_cache = NULL;
3546 /* free memory of osc quota cache */
3547 osc_quota_cleanup(obd);
3549 rc = client_obd_cleanup(obd);
3555 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3557 struct lprocfs_static_vars lvars = { 0 };
3560 lprocfs_osc_init_vars(&lvars);
3562 switch (lcfg->lcfg_command) {
3564 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3574 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3576 return osc_process_config_base(obd, buf);
3579 struct obd_ops osc_obd_ops = {
3580 .o_owner = THIS_MODULE,
3581 .o_setup = osc_setup,
3582 .o_precleanup = osc_precleanup,
3583 .o_cleanup = osc_cleanup,
3584 .o_add_conn = client_import_add_conn,
3585 .o_del_conn = client_import_del_conn,
3586 .o_connect = client_connect_import,
3587 .o_reconnect = osc_reconnect,
3588 .o_disconnect = osc_disconnect,
3589 .o_statfs = osc_statfs,
3590 .o_statfs_async = osc_statfs_async,
3591 .o_packmd = osc_packmd,
3592 .o_unpackmd = osc_unpackmd,
3593 .o_create = osc_create,
3594 .o_destroy = osc_destroy,
3595 .o_getattr = osc_getattr,
3596 .o_getattr_async = osc_getattr_async,
3597 .o_setattr = osc_setattr,
3598 .o_setattr_async = osc_setattr_async,
3600 .o_punch = osc_punch,
3602 .o_enqueue = osc_enqueue,
3603 .o_change_cbdata = osc_change_cbdata,
3604 .o_find_cbdata = osc_find_cbdata,
3605 .o_cancel = osc_cancel,
3606 .o_cancel_unused = osc_cancel_unused,
3607 .o_iocontrol = osc_iocontrol,
3608 .o_get_info = osc_get_info,
3609 .o_set_info_async = osc_set_info_async,
3610 .o_import_event = osc_import_event,
3611 .o_llog_init = osc_llog_init,
3612 .o_llog_finish = osc_llog_finish,
3613 .o_process_config = osc_process_config,
3614 .o_quotactl = osc_quotactl,
3615 .o_quotacheck = osc_quotacheck,
3618 extern struct lu_kmem_descr osc_caches[];
3619 extern spinlock_t osc_ast_guard;
3620 extern struct lock_class_key osc_ast_guard_class;
3622 int __init osc_init(void)
3624 struct lprocfs_static_vars lvars = { 0 };
3628 /* print an address of _any_ initialized kernel symbol from this
3629 * module, to allow debugging with gdb that doesn't support data
3630 * symbols from modules.*/
3631 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3633 rc = lu_kmem_init(osc_caches);
3635 lprocfs_osc_init_vars(&lvars);
3637 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3638 LUSTRE_OSC_NAME, &osc_device_type);
3640 lu_kmem_fini(osc_caches);
3644 spin_lock_init(&osc_ast_guard);
3645 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3651 static void /*__exit*/ osc_exit(void)
3653 class_unregister_type(LUSTRE_OSC_NAME);
3654 lu_kmem_fini(osc_caches);
3657 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3658 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3659 MODULE_LICENSE("GPL");
3661 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);