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, Whamcloud, Inc.
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 "osc_internal.h"
62 #include "osc_cl_internal.h"
64 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
65 static int brw_interpret(const struct lu_env *env,
66 struct ptlrpc_request *req, void *data, int rc);
67 int osc_cleanup(struct obd_device *obd);
69 /* Pack OSC object metadata for disk storage (LE byte order). */
70 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
71 struct lov_stripe_md *lsm)
76 lmm_size = sizeof(**lmmp);
81 OBD_FREE(*lmmp, lmm_size);
87 OBD_ALLOC(*lmmp, lmm_size);
93 LASSERT(lsm->lsm_object_id);
94 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
95 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
96 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
102 /* Unpack OSC object metadata from disk storage (LE byte order). */
103 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
104 struct lov_mds_md *lmm, int lmm_bytes)
107 struct obd_import *imp = class_exp2cliimp(exp);
111 if (lmm_bytes < sizeof (*lmm)) {
112 CERROR("lov_mds_md too small: %d, need %d\n",
113 lmm_bytes, (int)sizeof(*lmm));
116 /* XXX LOV_MAGIC etc check? */
118 if (lmm->lmm_object_id == 0) {
119 CERROR("lov_mds_md: zero lmm_object_id\n");
124 lsm_size = lov_stripe_md_size(1);
128 if (*lsmp != NULL && lmm == NULL) {
129 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
130 OBD_FREE(*lsmp, lsm_size);
136 OBD_ALLOC(*lsmp, lsm_size);
139 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
140 if ((*lsmp)->lsm_oinfo[0] == NULL) {
141 OBD_FREE(*lsmp, lsm_size);
144 loi_init((*lsmp)->lsm_oinfo[0]);
148 /* XXX zero *lsmp? */
149 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
150 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
151 LASSERT((*lsmp)->lsm_object_id);
152 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
156 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
157 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
159 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165 struct ost_body *body, void *capa)
167 struct obd_capa *oc = (struct obd_capa *)capa;
168 struct lustre_capa *c;
173 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
176 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177 DEBUG_CAPA(D_SEC, c, "pack");
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181 struct obd_info *oinfo)
183 struct ost_body *body;
185 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
188 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(const struct lu_env *env,
204 struct ptlrpc_request *req,
205 struct osc_async_args *aa, int rc)
207 struct ost_body *body;
213 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
215 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
218 /* This should really be sent by the OST */
219 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222 CDEBUG(D_INFO, "can't unpack ost_body\n");
224 aa->aa_oi->oi_oa->o_valid = 0;
227 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232 struct ptlrpc_request_set *set)
234 struct ptlrpc_request *req;
235 struct osc_async_args *aa;
239 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
243 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246 ptlrpc_request_free(req);
250 osc_pack_req_body(req, oinfo);
252 ptlrpc_request_set_replen(req);
253 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
255 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256 aa = ptlrpc_req_async_args(req);
259 ptlrpc_set_add_req(set, req);
263 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
264 struct obd_info *oinfo)
266 struct ptlrpc_request *req;
267 struct ost_body *body;
271 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
275 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
278 ptlrpc_request_free(req);
282 osc_pack_req_body(req, oinfo);
284 ptlrpc_request_set_replen(req);
286 rc = ptlrpc_queue_wait(req);
290 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
292 GOTO(out, rc = -EPROTO);
294 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
295 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
297 /* This should really be sent by the OST */
298 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
299 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
303 ptlrpc_req_finished(req);
307 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
308 struct obd_info *oinfo, struct obd_trans_info *oti)
310 struct ptlrpc_request *req;
311 struct ost_body *body;
315 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
317 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
321 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
322 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
324 ptlrpc_request_free(req);
328 osc_pack_req_body(req, oinfo);
330 ptlrpc_request_set_replen(req);
332 rc = ptlrpc_queue_wait(req);
336 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
338 GOTO(out, rc = -EPROTO);
340 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
344 ptlrpc_req_finished(req);
348 static int osc_setattr_interpret(const struct lu_env *env,
349 struct ptlrpc_request *req,
350 struct osc_setattr_args *sa, int rc)
352 struct ost_body *body;
358 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
360 GOTO(out, rc = -EPROTO);
362 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
364 rc = sa->sa_upcall(sa->sa_cookie, rc);
368 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
369 struct obd_trans_info *oti,
370 obd_enqueue_update_f upcall, void *cookie,
371 struct ptlrpc_request_set *rqset)
373 struct ptlrpc_request *req;
374 struct osc_setattr_args *sa;
378 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
382 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
383 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
385 ptlrpc_request_free(req);
389 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
390 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
392 osc_pack_req_body(req, oinfo);
394 ptlrpc_request_set_replen(req);
396 /* do mds to ost setattr asynchronously */
398 /* Do not wait for response. */
399 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
401 req->rq_interpret_reply =
402 (ptlrpc_interpterer_t)osc_setattr_interpret;
404 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
405 sa = ptlrpc_req_async_args(req);
406 sa->sa_oa = oinfo->oi_oa;
407 sa->sa_upcall = upcall;
408 sa->sa_cookie = cookie;
410 if (rqset == PTLRPCD_SET)
411 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
413 ptlrpc_set_add_req(rqset, req);
419 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
420 struct obd_trans_info *oti,
421 struct ptlrpc_request_set *rqset)
423 return osc_setattr_async_base(exp, oinfo, oti,
424 oinfo->oi_cb_up, oinfo, rqset);
427 int osc_real_create(struct obd_export *exp, struct obdo *oa,
428 struct lov_stripe_md **ea, struct obd_trans_info *oti)
430 struct ptlrpc_request *req;
431 struct ost_body *body;
432 struct lov_stripe_md *lsm;
441 rc = obd_alloc_memmd(exp, &lsm);
446 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
448 GOTO(out, rc = -ENOMEM);
450 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
452 ptlrpc_request_free(req);
456 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
458 lustre_set_wire_obdo(&body->oa, oa);
460 ptlrpc_request_set_replen(req);
462 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
463 oa->o_flags == OBD_FL_DELORPHAN) {
465 "delorphan from OST integration");
466 /* Don't resend the delorphan req */
467 req->rq_no_resend = req->rq_no_delay = 1;
470 rc = ptlrpc_queue_wait(req);
474 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
476 GOTO(out_req, rc = -EPROTO);
478 lustre_get_wire_obdo(oa, &body->oa);
480 /* This should really be sent by the OST */
481 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
482 oa->o_valid |= OBD_MD_FLBLKSZ;
484 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
485 * have valid lsm_oinfo data structs, so don't go touching that.
486 * This needs to be fixed in a big way.
488 lsm->lsm_object_id = oa->o_id;
489 lsm->lsm_object_seq = oa->o_seq;
493 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
495 if (oa->o_valid & OBD_MD_FLCOOKIE) {
496 if (!oti->oti_logcookies)
497 oti_alloc_cookies(oti, 1);
498 *oti->oti_logcookies = oa->o_lcookie;
502 CDEBUG(D_HA, "transno: "LPD64"\n",
503 lustre_msg_get_transno(req->rq_repmsg));
505 ptlrpc_req_finished(req);
508 obd_free_memmd(exp, &lsm);
512 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
513 obd_enqueue_update_f upcall, void *cookie,
514 struct ptlrpc_request_set *rqset)
516 struct ptlrpc_request *req;
517 struct osc_setattr_args *sa;
518 struct ost_body *body;
522 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
526 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
527 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
529 ptlrpc_request_free(req);
532 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
533 ptlrpc_at_set_req_timeout(req);
535 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
537 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
538 osc_pack_capa(req, body, oinfo->oi_capa);
540 ptlrpc_request_set_replen(req);
542 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
543 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
544 sa = ptlrpc_req_async_args(req);
545 sa->sa_oa = oinfo->oi_oa;
546 sa->sa_upcall = upcall;
547 sa->sa_cookie = cookie;
548 if (rqset == PTLRPCD_SET)
549 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
551 ptlrpc_set_add_req(rqset, req);
556 static int osc_punch(const struct lu_env *env, struct obd_export *exp,
557 struct obd_info *oinfo, struct obd_trans_info *oti,
558 struct ptlrpc_request_set *rqset)
560 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
561 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
562 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
563 return osc_punch_base(exp, oinfo,
564 oinfo->oi_cb_up, oinfo, rqset);
567 static int osc_sync_interpret(const struct lu_env *env,
568 struct ptlrpc_request *req,
571 struct osc_fsync_args *fa = arg;
572 struct ost_body *body;
578 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
580 CERROR ("can't unpack ost_body\n");
581 GOTO(out, rc = -EPROTO);
584 *fa->fa_oi->oi_oa = body->oa;
586 rc = fa->fa_upcall(fa->fa_cookie, rc);
590 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
591 obd_enqueue_update_f upcall, void *cookie,
592 struct ptlrpc_request_set *rqset)
594 struct ptlrpc_request *req;
595 struct ost_body *body;
596 struct osc_fsync_args *fa;
600 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
604 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
605 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
607 ptlrpc_request_free(req);
611 /* overload the size and blocks fields in the oa with start/end */
612 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
614 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
615 osc_pack_capa(req, body, oinfo->oi_capa);
617 ptlrpc_request_set_replen(req);
618 req->rq_interpret_reply = osc_sync_interpret;
620 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
621 fa = ptlrpc_req_async_args(req);
623 fa->fa_upcall = upcall;
624 fa->fa_cookie = cookie;
626 if (rqset == PTLRPCD_SET)
627 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
629 ptlrpc_set_add_req(rqset, req);
634 static int osc_sync(const struct lu_env *env, struct obd_export *exp,
635 struct obd_info *oinfo, obd_size start, obd_size end,
636 struct ptlrpc_request_set *set)
641 CDEBUG(D_INFO, "oa NULL\n");
645 oinfo->oi_oa->o_size = start;
646 oinfo->oi_oa->o_blocks = end;
647 oinfo->oi_oa->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
649 RETURN(osc_sync_base(exp, oinfo, oinfo->oi_cb_up, oinfo, set));
652 /* Find and cancel locally locks matched by @mode in the resource found by
653 * @objid. Found locks are added into @cancel list. Returns the amount of
654 * locks added to @cancels list. */
655 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
657 ldlm_mode_t mode, int lock_flags)
659 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
660 struct ldlm_res_id res_id;
661 struct ldlm_resource *res;
665 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
666 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
670 LDLM_RESOURCE_ADDREF(res);
671 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
672 lock_flags, 0, NULL);
673 LDLM_RESOURCE_DELREF(res);
674 ldlm_resource_putref(res);
678 static int osc_destroy_interpret(const struct lu_env *env,
679 struct ptlrpc_request *req, void *data,
682 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
684 cfs_atomic_dec(&cli->cl_destroy_in_flight);
685 cfs_waitq_signal(&cli->cl_destroy_waitq);
689 static int osc_can_send_destroy(struct client_obd *cli)
691 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
692 cli->cl_max_rpcs_in_flight) {
693 /* The destroy request can be sent */
696 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
697 cli->cl_max_rpcs_in_flight) {
699 * The counter has been modified between the two atomic
702 cfs_waitq_signal(&cli->cl_destroy_waitq);
707 /* Destroy requests can be async always on the client, and we don't even really
708 * care about the return code since the client cannot do anything at all about
710 * When the MDS is unlinking a filename, it saves the file objects into a
711 * recovery llog, and these object records are cancelled when the OST reports
712 * they were destroyed and sync'd to disk (i.e. transaction committed).
713 * If the client dies, or the OST is down when the object should be destroyed,
714 * the records are not cancelled, and when the OST reconnects to the MDS next,
715 * it will retrieve the llog unlink logs and then sends the log cancellation
716 * cookies to the MDS after committing destroy transactions. */
717 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
718 struct obdo *oa, struct lov_stripe_md *ea,
719 struct obd_trans_info *oti, struct obd_export *md_export,
722 struct client_obd *cli = &exp->exp_obd->u.cli;
723 struct ptlrpc_request *req;
724 struct ost_body *body;
725 CFS_LIST_HEAD(cancels);
730 CDEBUG(D_INFO, "oa NULL\n");
734 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
735 LDLM_FL_DISCARD_DATA);
737 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
739 ldlm_lock_list_put(&cancels, l_bl_ast, count);
743 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
744 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
747 ptlrpc_request_free(req);
751 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
752 ptlrpc_at_set_req_timeout(req);
754 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
755 oa->o_lcookie = *oti->oti_logcookies;
756 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
758 lustre_set_wire_obdo(&body->oa, oa);
760 osc_pack_capa(req, body, (struct obd_capa *)capa);
761 ptlrpc_request_set_replen(req);
763 /* If osc_destory is for destroying the unlink orphan,
764 * sent from MDT to OST, which should not be blocked here,
765 * because the process might be triggered by ptlrpcd, and
766 * it is not good to block ptlrpcd thread (b=16006)*/
767 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
768 req->rq_interpret_reply = osc_destroy_interpret;
769 if (!osc_can_send_destroy(cli)) {
770 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
774 * Wait until the number of on-going destroy RPCs drops
775 * under max_rpc_in_flight
777 l_wait_event_exclusive(cli->cl_destroy_waitq,
778 osc_can_send_destroy(cli), &lwi);
782 /* Do not wait for response */
783 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
787 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
790 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
792 LASSERT(!(oa->o_valid & bits));
795 client_obd_list_lock(&cli->cl_loi_list_lock);
796 oa->o_dirty = cli->cl_dirty;
797 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
798 CERROR("dirty %lu - %lu > dirty_max %lu\n",
799 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
801 } else if (cfs_atomic_read(&obd_dirty_pages) -
802 cfs_atomic_read(&obd_dirty_transit_pages) >
803 obd_max_dirty_pages + 1){
804 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
805 * not covered by a lock thus they may safely race and trip
806 * this CERROR() unless we add in a small fudge factor (+1). */
807 CERROR("dirty %d - %d > system dirty_max %d\n",
808 cfs_atomic_read(&obd_dirty_pages),
809 cfs_atomic_read(&obd_dirty_transit_pages),
810 obd_max_dirty_pages);
812 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
813 CERROR("dirty %lu - dirty_max %lu too big???\n",
814 cli->cl_dirty, cli->cl_dirty_max);
817 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
818 (cli->cl_max_rpcs_in_flight + 1);
819 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
821 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
822 oa->o_dropped = cli->cl_lost_grant;
823 cli->cl_lost_grant = 0;
824 client_obd_list_unlock(&cli->cl_loi_list_lock);
825 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
826 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
830 void osc_update_next_shrink(struct client_obd *cli)
832 cli->cl_next_shrink_grant =
833 cfs_time_shift(cli->cl_grant_shrink_interval);
834 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
835 cli->cl_next_shrink_grant);
838 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
840 client_obd_list_lock(&cli->cl_loi_list_lock);
841 cli->cl_avail_grant += grant;
842 client_obd_list_unlock(&cli->cl_loi_list_lock);
845 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
847 if (body->oa.o_valid & OBD_MD_FLGRANT) {
848 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
849 __osc_update_grant(cli, body->oa.o_grant);
853 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
854 obd_count keylen, void *key, obd_count vallen,
855 void *val, struct ptlrpc_request_set *set);
857 static int osc_shrink_grant_interpret(const struct lu_env *env,
858 struct ptlrpc_request *req,
861 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
862 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
863 struct ost_body *body;
866 __osc_update_grant(cli, oa->o_grant);
870 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
872 osc_update_grant(cli, body);
878 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
880 client_obd_list_lock(&cli->cl_loi_list_lock);
881 oa->o_grant = cli->cl_avail_grant / 4;
882 cli->cl_avail_grant -= oa->o_grant;
883 client_obd_list_unlock(&cli->cl_loi_list_lock);
884 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
885 oa->o_valid |= OBD_MD_FLFLAGS;
888 oa->o_flags |= OBD_FL_SHRINK_GRANT;
889 osc_update_next_shrink(cli);
892 /* Shrink the current grant, either from some large amount to enough for a
893 * full set of in-flight RPCs, or if we have already shrunk to that limit
894 * then to enough for a single RPC. This avoids keeping more grant than
895 * needed, and avoids shrinking the grant piecemeal. */
896 static int osc_shrink_grant(struct client_obd *cli)
898 long target = (cli->cl_max_rpcs_in_flight + 1) *
899 cli->cl_max_pages_per_rpc;
901 client_obd_list_lock(&cli->cl_loi_list_lock);
902 if (cli->cl_avail_grant <= target)
903 target = cli->cl_max_pages_per_rpc;
904 client_obd_list_unlock(&cli->cl_loi_list_lock);
906 return osc_shrink_grant_to_target(cli, target);
909 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
912 struct ost_body *body;
915 client_obd_list_lock(&cli->cl_loi_list_lock);
916 /* Don't shrink if we are already above or below the desired limit
917 * We don't want to shrink below a single RPC, as that will negatively
918 * impact block allocation and long-term performance. */
919 if (target < cli->cl_max_pages_per_rpc)
920 target = cli->cl_max_pages_per_rpc;
922 if (target >= cli->cl_avail_grant) {
923 client_obd_list_unlock(&cli->cl_loi_list_lock);
926 client_obd_list_unlock(&cli->cl_loi_list_lock);
932 osc_announce_cached(cli, &body->oa, 0);
934 client_obd_list_lock(&cli->cl_loi_list_lock);
935 body->oa.o_grant = cli->cl_avail_grant - target;
936 cli->cl_avail_grant = target;
937 client_obd_list_unlock(&cli->cl_loi_list_lock);
938 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
939 body->oa.o_valid |= OBD_MD_FLFLAGS;
940 body->oa.o_flags = 0;
942 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
943 osc_update_next_shrink(cli);
945 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
946 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
947 sizeof(*body), body, NULL);
949 __osc_update_grant(cli, body->oa.o_grant);
954 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
955 static int osc_should_shrink_grant(struct client_obd *client)
957 cfs_time_t time = cfs_time_current();
958 cfs_time_t next_shrink = client->cl_next_shrink_grant;
960 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
961 OBD_CONNECT_GRANT_SHRINK) == 0)
964 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
965 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
966 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
969 osc_update_next_shrink(client);
974 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
976 struct client_obd *client;
978 cfs_list_for_each_entry(client, &item->ti_obd_list,
979 cl_grant_shrink_list) {
980 if (osc_should_shrink_grant(client))
981 osc_shrink_grant(client);
986 static int osc_add_shrink_grant(struct client_obd *client)
990 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
992 osc_grant_shrink_grant_cb, NULL,
993 &client->cl_grant_shrink_list);
995 CERROR("add grant client %s error %d\n",
996 client->cl_import->imp_obd->obd_name, rc);
999 CDEBUG(D_CACHE, "add grant client %s \n",
1000 client->cl_import->imp_obd->obd_name);
1001 osc_update_next_shrink(client);
1005 static int osc_del_shrink_grant(struct client_obd *client)
1007 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1011 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1014 * ocd_grant is the total grant amount we're expect to hold: if we've
1015 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1016 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1018 * race is tolerable here: if we're evicted, but imp_state already
1019 * left EVICTED state, then cl_dirty must be 0 already.
1021 client_obd_list_lock(&cli->cl_loi_list_lock);
1022 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1023 cli->cl_avail_grant = ocd->ocd_grant;
1025 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1027 if (cli->cl_avail_grant < 0) {
1028 CWARN("%s: available grant < 0, the OSS is probably not running"
1029 " with patch from bug20278 (%ld) \n",
1030 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1031 /* workaround for 1.6 servers which do not have
1032 * the patch from bug20278 */
1033 cli->cl_avail_grant = ocd->ocd_grant;
1036 /* determine the appropriate chunk size used by osc_extent. */
1037 cli->cl_chunkbits = max_t(int, CFS_PAGE_SHIFT, ocd->ocd_blocksize);
1038 client_obd_list_unlock(&cli->cl_loi_list_lock);
1040 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1041 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1042 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1044 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1045 cfs_list_empty(&cli->cl_grant_shrink_list))
1046 osc_add_shrink_grant(cli);
1049 /* We assume that the reason this OSC got a short read is because it read
1050 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1051 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1052 * this stripe never got written at or beyond this stripe offset yet. */
1053 static void handle_short_read(int nob_read, obd_count page_count,
1054 struct brw_page **pga)
1059 /* skip bytes read OK */
1060 while (nob_read > 0) {
1061 LASSERT (page_count > 0);
1063 if (pga[i]->count > nob_read) {
1064 /* EOF inside this page */
1065 ptr = cfs_kmap(pga[i]->pg) +
1066 (pga[i]->off & ~CFS_PAGE_MASK);
1067 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1068 cfs_kunmap(pga[i]->pg);
1074 nob_read -= pga[i]->count;
1079 /* zero remaining pages */
1080 while (page_count-- > 0) {
1081 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1082 memset(ptr, 0, pga[i]->count);
1083 cfs_kunmap(pga[i]->pg);
1088 static int check_write_rcs(struct ptlrpc_request *req,
1089 int requested_nob, int niocount,
1090 obd_count page_count, struct brw_page **pga)
1095 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1096 sizeof(*remote_rcs) *
1098 if (remote_rcs == NULL) {
1099 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1103 /* return error if any niobuf was in error */
1104 for (i = 0; i < niocount; i++) {
1105 if ((int)remote_rcs[i] < 0)
1106 return(remote_rcs[i]);
1108 if (remote_rcs[i] != 0) {
1109 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1110 i, remote_rcs[i], req);
1115 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1116 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1117 req->rq_bulk->bd_nob_transferred, requested_nob);
1124 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1126 if (p1->flag != p2->flag) {
1127 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1128 OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1130 /* warn if we try to combine flags that we don't know to be
1131 * safe to combine */
1132 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1133 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1134 "report this at http://bugs.whamcloud.com/\n",
1135 p1->flag, p2->flag);
1140 return (p1->off + p1->count == p2->off);
1143 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1144 struct brw_page **pga, int opc,
1145 cksum_type_t cksum_type)
1149 struct cfs_crypto_hash_desc *hdesc;
1150 unsigned int bufsize;
1152 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1154 LASSERT(pg_count > 0);
1156 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1157 if (IS_ERR(hdesc)) {
1158 CERROR("Unable to initialize checksum hash %s\n",
1159 cfs_crypto_hash_name(cfs_alg));
1160 return PTR_ERR(hdesc);
1163 while (nob > 0 && pg_count > 0) {
1164 int count = pga[i]->count > nob ? nob : pga[i]->count;
1166 /* corrupt the data before we compute the checksum, to
1167 * simulate an OST->client data error */
1168 if (i == 0 && opc == OST_READ &&
1169 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1170 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1171 int off = pga[i]->off & ~CFS_PAGE_MASK;
1172 memcpy(ptr + off, "bad1", min(4, nob));
1173 cfs_kunmap(pga[i]->pg);
1175 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1176 pga[i]->off & ~CFS_PAGE_MASK,
1178 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1179 (int)(pga[i]->off & ~CFS_PAGE_MASK), cksum);
1181 nob -= pga[i]->count;
1187 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1190 cfs_crypto_hash_final(hdesc, NULL, NULL);
1192 /* For sending we only compute the wrong checksum instead
1193 * of corrupting the data so it is still correct on a redo */
1194 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1200 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1201 struct lov_stripe_md *lsm, obd_count page_count,
1202 struct brw_page **pga,
1203 struct ptlrpc_request **reqp,
1204 struct obd_capa *ocapa, int reserve,
1207 struct ptlrpc_request *req;
1208 struct ptlrpc_bulk_desc *desc;
1209 struct ost_body *body;
1210 struct obd_ioobj *ioobj;
1211 struct niobuf_remote *niobuf;
1212 int niocount, i, requested_nob, opc, rc;
1213 struct osc_brw_async_args *aa;
1214 struct req_capsule *pill;
1215 struct brw_page *pg_prev;
1218 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1219 RETURN(-ENOMEM); /* Recoverable */
1220 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1221 RETURN(-EINVAL); /* Fatal */
1223 if ((cmd & OBD_BRW_WRITE) != 0) {
1225 req = ptlrpc_request_alloc_pool(cli->cl_import,
1226 cli->cl_import->imp_rq_pool,
1227 &RQF_OST_BRW_WRITE);
1230 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1235 for (niocount = i = 1; i < page_count; i++) {
1236 if (!can_merge_pages(pga[i - 1], pga[i]))
1240 pill = &req->rq_pill;
1241 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1243 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1244 niocount * sizeof(*niobuf));
1245 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1247 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1249 ptlrpc_request_free(req);
1252 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1253 ptlrpc_at_set_req_timeout(req);
1254 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1256 req->rq_no_retry_einprogress = 1;
1258 if (opc == OST_WRITE)
1259 desc = ptlrpc_prep_bulk_imp(req, page_count,
1260 BULK_GET_SOURCE, OST_BULK_PORTAL);
1262 desc = ptlrpc_prep_bulk_imp(req, page_count,
1263 BULK_PUT_SINK, OST_BULK_PORTAL);
1266 GOTO(out, rc = -ENOMEM);
1267 /* NB request now owns desc and will free it when it gets freed */
1269 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1270 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1271 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1272 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1274 lustre_set_wire_obdo(&body->oa, oa);
1276 obdo_to_ioobj(oa, ioobj);
1277 ioobj->ioo_bufcnt = niocount;
1278 osc_pack_capa(req, body, ocapa);
1279 LASSERT (page_count > 0);
1281 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1282 struct brw_page *pg = pga[i];
1283 int poff = pg->off & ~CFS_PAGE_MASK;
1285 LASSERT(pg->count > 0);
1286 /* make sure there is no gap in the middle of page array */
1287 LASSERTF(page_count == 1 ||
1288 (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1289 ergo(i > 0 && i < page_count - 1,
1290 poff == 0 && pg->count == CFS_PAGE_SIZE) &&
1291 ergo(i == page_count - 1, poff == 0)),
1292 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1293 i, page_count, pg, pg->off, pg->count);
1295 LASSERTF(i == 0 || pg->off > pg_prev->off,
1296 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1297 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1299 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1300 pg_prev->pg, page_private(pg_prev->pg),
1301 pg_prev->pg->index, pg_prev->off);
1303 LASSERTF(i == 0 || pg->off > pg_prev->off,
1304 "i %d p_c %u\n", i, page_count);
1306 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1307 (pg->flag & OBD_BRW_SRVLOCK));
1309 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1310 requested_nob += pg->count;
1312 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1314 niobuf->len += pg->count;
1316 niobuf->offset = pg->off;
1317 niobuf->len = pg->count;
1318 niobuf->flags = pg->flag;
1323 LASSERTF((void *)(niobuf - niocount) ==
1324 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1325 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1326 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1328 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1330 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1331 body->oa.o_valid |= OBD_MD_FLFLAGS;
1332 body->oa.o_flags = 0;
1334 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1337 if (osc_should_shrink_grant(cli))
1338 osc_shrink_grant_local(cli, &body->oa);
1340 /* size[REQ_REC_OFF] still sizeof (*body) */
1341 if (opc == OST_WRITE) {
1342 if (cli->cl_checksum &&
1343 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1344 /* store cl_cksum_type in a local variable since
1345 * it can be changed via lprocfs */
1346 cksum_type_t cksum_type = cli->cl_cksum_type;
1348 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1349 oa->o_flags &= OBD_FL_LOCAL_MASK;
1350 body->oa.o_flags = 0;
1352 body->oa.o_flags |= cksum_type_pack(cksum_type);
1353 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1354 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1358 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1360 /* save this in 'oa', too, for later checking */
1361 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1362 oa->o_flags |= cksum_type_pack(cksum_type);
1364 /* clear out the checksum flag, in case this is a
1365 * resend but cl_checksum is no longer set. b=11238 */
1366 oa->o_valid &= ~OBD_MD_FLCKSUM;
1368 oa->o_cksum = body->oa.o_cksum;
1369 /* 1 RC per niobuf */
1370 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1371 sizeof(__u32) * niocount);
1373 if (cli->cl_checksum &&
1374 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1375 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1376 body->oa.o_flags = 0;
1377 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1378 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1381 ptlrpc_request_set_replen(req);
1383 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1384 aa = ptlrpc_req_async_args(req);
1386 aa->aa_requested_nob = requested_nob;
1387 aa->aa_nio_count = niocount;
1388 aa->aa_page_count = page_count;
1392 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1393 if (ocapa && reserve)
1394 aa->aa_ocapa = capa_get(ocapa);
1400 ptlrpc_req_finished(req);
1404 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1405 __u32 client_cksum, __u32 server_cksum, int nob,
1406 obd_count page_count, struct brw_page **pga,
1407 cksum_type_t client_cksum_type)
1411 cksum_type_t cksum_type;
1413 if (server_cksum == client_cksum) {
1414 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1418 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1420 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1423 if (cksum_type != client_cksum_type)
1424 msg = "the server did not use the checksum type specified in "
1425 "the original request - likely a protocol problem";
1426 else if (new_cksum == server_cksum)
1427 msg = "changed on the client after we checksummed it - "
1428 "likely false positive due to mmap IO (bug 11742)";
1429 else if (new_cksum == client_cksum)
1430 msg = "changed in transit before arrival at OST";
1432 msg = "changed in transit AND doesn't match the original - "
1433 "likely false positive due to mmap IO (bug 11742)";
1435 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1436 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1437 msg, libcfs_nid2str(peer->nid),
1438 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1439 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1440 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1442 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1444 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1445 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1446 "client csum now %x\n", client_cksum, client_cksum_type,
1447 server_cksum, cksum_type, new_cksum);
1451 /* Note rc enters this function as number of bytes transferred */
1452 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1454 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1455 const lnet_process_id_t *peer =
1456 &req->rq_import->imp_connection->c_peer;
1457 struct client_obd *cli = aa->aa_cli;
1458 struct ost_body *body;
1459 __u32 client_cksum = 0;
1462 if (rc < 0 && rc != -EDQUOT) {
1463 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1467 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1468 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1470 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1474 /* set/clear over quota flag for a uid/gid */
1475 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1476 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1477 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1479 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1480 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1482 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1485 osc_update_grant(cli, body);
1490 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1491 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1493 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1495 CERROR("Unexpected +ve rc %d\n", rc);
1498 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1500 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1503 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1504 check_write_checksum(&body->oa, peer, client_cksum,
1505 body->oa.o_cksum, aa->aa_requested_nob,
1506 aa->aa_page_count, aa->aa_ppga,
1507 cksum_type_unpack(aa->aa_oa->o_flags)))
1510 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1511 aa->aa_page_count, aa->aa_ppga);
1515 /* The rest of this function executes only for OST_READs */
1517 /* if unwrap_bulk failed, return -EAGAIN to retry */
1518 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1520 GOTO(out, rc = -EAGAIN);
1522 if (rc > aa->aa_requested_nob) {
1523 CERROR("Unexpected rc %d (%d requested)\n", rc,
1524 aa->aa_requested_nob);
1528 if (rc != req->rq_bulk->bd_nob_transferred) {
1529 CERROR ("Unexpected rc %d (%d transferred)\n",
1530 rc, req->rq_bulk->bd_nob_transferred);
1534 if (rc < aa->aa_requested_nob)
1535 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1537 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1538 static int cksum_counter;
1539 __u32 server_cksum = body->oa.o_cksum;
1542 cksum_type_t cksum_type;
1544 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1545 body->oa.o_flags : 0);
1546 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1547 aa->aa_ppga, OST_READ,
1550 if (peer->nid == req->rq_bulk->bd_sender) {
1554 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1557 if (server_cksum == ~0 && rc > 0) {
1558 CERROR("Protocol error: server %s set the 'checksum' "
1559 "bit, but didn't send a checksum. Not fatal, "
1560 "but please notify on http://bugs.whamcloud.com/\n",
1561 libcfs_nid2str(peer->nid));
1562 } else if (server_cksum != client_cksum) {
1563 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1564 "%s%s%s inode "DFID" object "
1565 LPU64"/"LPU64" extent "
1566 "["LPU64"-"LPU64"]\n",
1567 req->rq_import->imp_obd->obd_name,
1568 libcfs_nid2str(peer->nid),
1570 body->oa.o_valid & OBD_MD_FLFID ?
1571 body->oa.o_parent_seq : (__u64)0,
1572 body->oa.o_valid & OBD_MD_FLFID ?
1573 body->oa.o_parent_oid : 0,
1574 body->oa.o_valid & OBD_MD_FLFID ?
1575 body->oa.o_parent_ver : 0,
1577 body->oa.o_valid & OBD_MD_FLGROUP ?
1578 body->oa.o_seq : (__u64)0,
1579 aa->aa_ppga[0]->off,
1580 aa->aa_ppga[aa->aa_page_count-1]->off +
1581 aa->aa_ppga[aa->aa_page_count-1]->count -
1583 CERROR("client %x, server %x, cksum_type %x\n",
1584 client_cksum, server_cksum, cksum_type);
1586 aa->aa_oa->o_cksum = client_cksum;
1590 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1593 } else if (unlikely(client_cksum)) {
1594 static int cksum_missed;
1597 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1598 CERROR("Checksum %u requested from %s but not sent\n",
1599 cksum_missed, libcfs_nid2str(peer->nid));
1605 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1610 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1611 struct lov_stripe_md *lsm,
1612 obd_count page_count, struct brw_page **pga,
1613 struct obd_capa *ocapa)
1615 struct ptlrpc_request *req;
1618 int generation, resends = 0;
1619 struct l_wait_info lwi;
1623 cfs_waitq_init(&waitq);
1624 generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1627 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1628 page_count, pga, &req, ocapa, 0, resends);
1633 req->rq_generation_set = 1;
1634 req->rq_import_generation = generation;
1635 req->rq_sent = cfs_time_current_sec() + resends;
1638 rc = ptlrpc_queue_wait(req);
1640 if (rc == -ETIMEDOUT && req->rq_resend) {
1641 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1642 ptlrpc_req_finished(req);
1646 rc = osc_brw_fini_request(req, rc);
1648 ptlrpc_req_finished(req);
1649 /* When server return -EINPROGRESS, client should always retry
1650 * regardless of the number of times the bulk was resent already.*/
1651 if (osc_recoverable_error(rc)) {
1653 if (rc != -EINPROGRESS &&
1654 !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1655 CERROR("%s: too many resend retries for object: "
1656 ""LPU64":"LPU64", rc = %d.\n",
1657 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1661 exp->exp_obd->u.cli.cl_import->imp_generation) {
1662 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1663 ""LPU64":"LPU64", rc = %d.\n",
1664 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1668 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1670 l_wait_event(waitq, 0, &lwi);
1675 if (rc == -EAGAIN || rc == -EINPROGRESS)
1680 static int osc_brw_redo_request(struct ptlrpc_request *request,
1681 struct osc_brw_async_args *aa, int rc)
1683 struct ptlrpc_request *new_req;
1684 struct osc_brw_async_args *new_aa;
1685 struct osc_async_page *oap;
1688 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1689 "redo for recoverable error %d", rc);
1691 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1692 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1693 aa->aa_cli, aa->aa_oa,
1694 NULL /* lsm unused by osc currently */,
1695 aa->aa_page_count, aa->aa_ppga,
1696 &new_req, aa->aa_ocapa, 0, 1);
1700 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1701 if (oap->oap_request != NULL) {
1702 LASSERTF(request == oap->oap_request,
1703 "request %p != oap_request %p\n",
1704 request, oap->oap_request);
1705 if (oap->oap_interrupted) {
1706 ptlrpc_req_finished(new_req);
1711 /* New request takes over pga and oaps from old request.
1712 * Note that copying a list_head doesn't work, need to move it... */
1714 new_req->rq_interpret_reply = request->rq_interpret_reply;
1715 new_req->rq_async_args = request->rq_async_args;
1716 /* cap resend delay to the current request timeout, this is similar to
1717 * what ptlrpc does (see after_reply()) */
1718 if (aa->aa_resends > new_req->rq_timeout)
1719 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1721 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1722 new_req->rq_generation_set = 1;
1723 new_req->rq_import_generation = request->rq_import_generation;
1725 new_aa = ptlrpc_req_async_args(new_req);
1727 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1728 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1729 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1730 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1732 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1733 if (oap->oap_request) {
1734 ptlrpc_req_finished(oap->oap_request);
1735 oap->oap_request = ptlrpc_request_addref(new_req);
1739 new_aa->aa_ocapa = aa->aa_ocapa;
1740 aa->aa_ocapa = NULL;
1742 /* XXX: This code will run into problem if we're going to support
1743 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1744 * and wait for all of them to be finished. We should inherit request
1745 * set from old request. */
1746 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1748 DEBUG_REQ(D_INFO, new_req, "new request");
1753 * ugh, we want disk allocation on the target to happen in offset order. we'll
1754 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1755 * fine for our small page arrays and doesn't require allocation. its an
1756 * insertion sort that swaps elements that are strides apart, shrinking the
1757 * stride down until its '1' and the array is sorted.
1759 static void sort_brw_pages(struct brw_page **array, int num)
1762 struct brw_page *tmp;
1766 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1771 for (i = stride ; i < num ; i++) {
1774 while (j >= stride && array[j - stride]->off > tmp->off) {
1775 array[j] = array[j - stride];
1780 } while (stride > 1);
1783 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1789 LASSERT (pages > 0);
1790 offset = pg[i]->off & ~CFS_PAGE_MASK;
1794 if (pages == 0) /* that's all */
1797 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1798 return count; /* doesn't end on page boundary */
1801 offset = pg[i]->off & ~CFS_PAGE_MASK;
1802 if (offset != 0) /* doesn't start on page boundary */
1809 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1811 struct brw_page **ppga;
1814 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1818 for (i = 0; i < count; i++)
1823 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1825 LASSERT(ppga != NULL);
1826 OBD_FREE(ppga, sizeof(*ppga) * count);
1829 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1830 obd_count page_count, struct brw_page *pga,
1831 struct obd_trans_info *oti)
1833 struct obdo *saved_oa = NULL;
1834 struct brw_page **ppga, **orig;
1835 struct obd_import *imp = class_exp2cliimp(exp);
1836 struct client_obd *cli;
1837 int rc, page_count_orig;
1840 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1841 cli = &imp->imp_obd->u.cli;
1843 if (cmd & OBD_BRW_CHECK) {
1844 /* The caller just wants to know if there's a chance that this
1845 * I/O can succeed */
1847 if (imp->imp_invalid)
1852 /* test_brw with a failed create can trip this, maybe others. */
1853 LASSERT(cli->cl_max_pages_per_rpc);
1857 orig = ppga = osc_build_ppga(pga, page_count);
1860 page_count_orig = page_count;
1862 sort_brw_pages(ppga, page_count);
1863 while (page_count) {
1864 obd_count pages_per_brw;
1866 if (page_count > cli->cl_max_pages_per_rpc)
1867 pages_per_brw = cli->cl_max_pages_per_rpc;
1869 pages_per_brw = page_count;
1871 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1873 if (saved_oa != NULL) {
1874 /* restore previously saved oa */
1875 *oinfo->oi_oa = *saved_oa;
1876 } else if (page_count > pages_per_brw) {
1877 /* save a copy of oa (brw will clobber it) */
1878 OBDO_ALLOC(saved_oa);
1879 if (saved_oa == NULL)
1880 GOTO(out, rc = -ENOMEM);
1881 *saved_oa = *oinfo->oi_oa;
1884 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1885 pages_per_brw, ppga, oinfo->oi_capa);
1890 page_count -= pages_per_brw;
1891 ppga += pages_per_brw;
1895 osc_release_ppga(orig, page_count_orig);
1897 if (saved_oa != NULL)
1898 OBDO_FREE(saved_oa);
1903 static int brw_interpret(const struct lu_env *env,
1904 struct ptlrpc_request *req, void *data, int rc)
1906 struct osc_brw_async_args *aa = data;
1907 struct osc_extent *ext;
1908 struct osc_extent *tmp;
1909 struct cl_object *obj = NULL;
1910 struct client_obd *cli = aa->aa_cli;
1913 rc = osc_brw_fini_request(req, rc);
1914 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1915 /* When server return -EINPROGRESS, client should always retry
1916 * regardless of the number of times the bulk was resent already. */
1917 if (osc_recoverable_error(rc)) {
1918 if (req->rq_import_generation !=
1919 req->rq_import->imp_generation) {
1920 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1921 ""LPU64":"LPU64", rc = %d.\n",
1922 req->rq_import->imp_obd->obd_name,
1923 aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
1924 } else if (rc == -EINPROGRESS ||
1925 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1926 rc = osc_brw_redo_request(req, aa, rc);
1928 CERROR("%s: too many resent retries for object: "
1929 ""LPU64":"LPU64", rc = %d.\n",
1930 req->rq_import->imp_obd->obd_name,
1931 aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
1936 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1941 capa_put(aa->aa_ocapa);
1942 aa->aa_ocapa = NULL;
1945 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1946 if (obj == NULL && rc == 0) {
1947 obj = osc2cl(ext->oe_obj);
1951 cfs_list_del_init(&ext->oe_link);
1952 osc_extent_finish(env, ext, 1, rc);
1954 LASSERT(cfs_list_empty(&aa->aa_exts));
1955 LASSERT(cfs_list_empty(&aa->aa_oaps));
1958 struct obdo *oa = aa->aa_oa;
1959 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1960 unsigned long valid = 0;
1963 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1964 attr->cat_blocks = oa->o_blocks;
1965 valid |= CAT_BLOCKS;
1967 if (oa->o_valid & OBD_MD_FLMTIME) {
1968 attr->cat_mtime = oa->o_mtime;
1971 if (oa->o_valid & OBD_MD_FLATIME) {
1972 attr->cat_atime = oa->o_atime;
1975 if (oa->o_valid & OBD_MD_FLCTIME) {
1976 attr->cat_ctime = oa->o_ctime;
1980 cl_object_attr_lock(obj);
1981 cl_object_attr_set(env, obj, attr, valid);
1982 cl_object_attr_unlock(obj);
1984 cl_object_put(env, obj);
1986 OBDO_FREE(aa->aa_oa);
1988 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1989 req->rq_bulk->bd_nob_transferred);
1990 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1991 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1993 client_obd_list_lock(&cli->cl_loi_list_lock);
1994 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1995 * is called so we know whether to go to sync BRWs or wait for more
1996 * RPCs to complete */
1997 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1998 cli->cl_w_in_flight--;
2000 cli->cl_r_in_flight--;
2001 osc_wake_cache_waiters(cli);
2002 client_obd_list_unlock(&cli->cl_loi_list_lock);
2004 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2009 * Build an RPC by the list of extent @ext_list. The caller must ensure
2010 * that the total pages in this list are NOT over max pages per RPC.
2011 * Extents in the list must be in OES_RPC state.
2013 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2014 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2016 struct ptlrpc_request *req = NULL;
2017 struct osc_extent *ext;
2018 CFS_LIST_HEAD(rpc_list);
2019 struct brw_page **pga = NULL;
2020 struct osc_brw_async_args *aa = NULL;
2021 struct obdo *oa = NULL;
2022 struct osc_async_page *oap;
2023 struct osc_async_page *tmp;
2024 struct cl_req *clerq = NULL;
2025 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2026 struct ldlm_lock *lock = NULL;
2027 struct cl_req_attr crattr;
2028 obd_off starting_offset = OBD_OBJECT_EOF;
2029 obd_off ending_offset = 0;
2030 int i, rc, mpflag = 0, mem_tight = 0, page_count = 0;
2033 LASSERT(!cfs_list_empty(ext_list));
2035 /* add pages into rpc_list to build BRW rpc */
2036 cfs_list_for_each_entry(ext, ext_list, oe_link) {
2037 LASSERT(ext->oe_state == OES_RPC);
2038 mem_tight |= ext->oe_memalloc;
2039 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2041 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2042 if (starting_offset > oap->oap_obj_off)
2043 starting_offset = oap->oap_obj_off;
2045 LASSERT(oap->oap_page_off == 0);
2046 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2047 ending_offset = oap->oap_obj_off +
2050 LASSERT(oap->oap_page_off + oap->oap_count ==
2056 mpflag = cfs_memory_pressure_get_and_set();
2058 memset(&crattr, 0, sizeof crattr);
2059 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2061 GOTO(out, rc = -ENOMEM);
2065 GOTO(out, rc = -ENOMEM);
2068 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2069 struct cl_page *page = oap2cl_page(oap);
2070 if (clerq == NULL) {
2071 clerq = cl_req_alloc(env, page, crt,
2072 1 /* only 1-object rpcs for
2075 GOTO(out, rc = PTR_ERR(clerq));
2076 lock = oap->oap_ldlm_lock;
2079 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2080 pga[i] = &oap->oap_brw_page;
2081 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2082 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2083 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2085 cl_req_page_add(env, clerq, page);
2088 /* always get the data for the obdo for the rpc */
2089 LASSERT(clerq != NULL);
2091 crattr.cra_capa = NULL;
2092 memset(crattr.cra_jobid, 0, JOBSTATS_JOBID_SIZE);
2093 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2095 oa->o_handle = lock->l_remote_handle;
2096 oa->o_valid |= OBD_MD_FLHANDLE;
2099 rc = cl_req_prep(env, clerq);
2101 CERROR("cl_req_prep failed: %d\n", rc);
2105 sort_brw_pages(pga, page_count);
2106 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2107 pga, &req, crattr.cra_capa, 1, 0);
2109 CERROR("prep_req failed: %d\n", rc);
2113 req->rq_interpret_reply = brw_interpret;
2115 req->rq_memalloc = 1;
2117 /* Need to update the timestamps after the request is built in case
2118 * we race with setattr (locally or in queue at OST). If OST gets
2119 * later setattr before earlier BRW (as determined by the request xid),
2120 * the OST will not use BRW timestamps. Sadly, there is no obvious
2121 * way to do this in a single call. bug 10150 */
2122 cl_req_attr_set(env, clerq, &crattr,
2123 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2125 lustre_msg_set_jobid(req->rq_reqmsg, crattr.cra_jobid);
2127 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2128 aa = ptlrpc_req_async_args(req);
2129 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2130 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2131 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2132 cfs_list_splice_init(ext_list, &aa->aa_exts);
2133 aa->aa_clerq = clerq;
2135 /* queued sync pages can be torn down while the pages
2136 * were between the pending list and the rpc */
2138 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2139 /* only one oap gets a request reference */
2142 if (oap->oap_interrupted && !req->rq_intr) {
2143 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2145 ptlrpc_mark_interrupted(req);
2149 tmp->oap_request = ptlrpc_request_addref(req);
2151 client_obd_list_lock(&cli->cl_loi_list_lock);
2152 starting_offset >>= CFS_PAGE_SHIFT;
2153 if (cmd == OBD_BRW_READ) {
2154 cli->cl_r_in_flight++;
2155 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2156 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2157 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2158 starting_offset + 1);
2160 cli->cl_w_in_flight++;
2161 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2162 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2163 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2164 starting_offset + 1);
2166 client_obd_list_unlock(&cli->cl_loi_list_lock);
2168 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2169 page_count, aa, cli->cl_r_in_flight,
2170 cli->cl_w_in_flight);
2172 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2173 * see which CPU/NUMA node the majority of pages were allocated
2174 * on, and try to assign the async RPC to the CPU core
2175 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2177 * But on the other hand, we expect that multiple ptlrpcd
2178 * threads and the initial write sponsor can run in parallel,
2179 * especially when data checksum is enabled, which is CPU-bound
2180 * operation and single ptlrpcd thread cannot process in time.
2181 * So more ptlrpcd threads sharing BRW load
2182 * (with PDL_POLICY_ROUND) seems better.
2184 ptlrpcd_add_req(req, pol, -1);
2190 cfs_memory_pressure_restore(mpflag);
2192 capa_put(crattr.cra_capa);
2194 LASSERT(req == NULL);
2199 OBD_FREE(pga, sizeof(*pga) * page_count);
2200 /* this should happen rarely and is pretty bad, it makes the
2201 * pending list not follow the dirty order */
2202 while (!cfs_list_empty(ext_list)) {
2203 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2205 cfs_list_del_init(&ext->oe_link);
2206 osc_extent_finish(env, ext, 0, rc);
2208 if (clerq && !IS_ERR(clerq))
2209 cl_req_completion(env, clerq, rc);
2214 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2215 struct ldlm_enqueue_info *einfo)
2217 void *data = einfo->ei_cbdata;
2220 LASSERT(lock != NULL);
2221 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2222 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2223 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2224 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2226 lock_res_and_lock(lock);
2227 cfs_spin_lock(&osc_ast_guard);
2229 if (lock->l_ast_data == NULL)
2230 lock->l_ast_data = data;
2231 if (lock->l_ast_data == data)
2234 cfs_spin_unlock(&osc_ast_guard);
2235 unlock_res_and_lock(lock);
2240 static int osc_set_data_with_check(struct lustre_handle *lockh,
2241 struct ldlm_enqueue_info *einfo)
2243 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2247 set = osc_set_lock_data_with_check(lock, einfo);
2248 LDLM_LOCK_PUT(lock);
2250 CERROR("lockh %p, data %p - client evicted?\n",
2251 lockh, einfo->ei_cbdata);
2255 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2256 ldlm_iterator_t replace, void *data)
2258 struct ldlm_res_id res_id;
2259 struct obd_device *obd = class_exp2obd(exp);
2261 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
2262 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2266 /* find any ldlm lock of the inode in osc
2270 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2271 ldlm_iterator_t replace, void *data)
2273 struct ldlm_res_id res_id;
2274 struct obd_device *obd = class_exp2obd(exp);
2277 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
2278 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2279 if (rc == LDLM_ITER_STOP)
2281 if (rc == LDLM_ITER_CONTINUE)
2286 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2287 obd_enqueue_update_f upcall, void *cookie,
2288 int *flags, int agl, int rc)
2290 int intent = *flags & LDLM_FL_HAS_INTENT;
2294 /* The request was created before ldlm_cli_enqueue call. */
2295 if (rc == ELDLM_LOCK_ABORTED) {
2296 struct ldlm_reply *rep;
2297 rep = req_capsule_server_get(&req->rq_pill,
2300 LASSERT(rep != NULL);
2301 if (rep->lock_policy_res1)
2302 rc = rep->lock_policy_res1;
2306 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2308 *flags |= LDLM_FL_LVB_READY;
2309 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2310 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2313 /* Call the update callback. */
2314 rc = (*upcall)(cookie, rc);
2318 static int osc_enqueue_interpret(const struct lu_env *env,
2319 struct ptlrpc_request *req,
2320 struct osc_enqueue_args *aa, int rc)
2322 struct ldlm_lock *lock;
2323 struct lustre_handle handle;
2325 struct ost_lvb *lvb;
2327 int *flags = aa->oa_flags;
2329 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2330 * might be freed anytime after lock upcall has been called. */
2331 lustre_handle_copy(&handle, aa->oa_lockh);
2332 mode = aa->oa_ei->ei_mode;
2334 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2336 lock = ldlm_handle2lock(&handle);
2338 /* Take an additional reference so that a blocking AST that
2339 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2340 * to arrive after an upcall has been executed by
2341 * osc_enqueue_fini(). */
2342 ldlm_lock_addref(&handle, mode);
2344 /* Let CP AST to grant the lock first. */
2345 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2347 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2352 lvb_len = sizeof(*aa->oa_lvb);
2355 /* Complete obtaining the lock procedure. */
2356 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2357 mode, flags, lvb, lvb_len, &handle, rc);
2358 /* Complete osc stuff. */
2359 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2360 flags, aa->oa_agl, rc);
2362 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2364 /* Release the lock for async request. */
2365 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2367 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2368 * not already released by
2369 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2371 ldlm_lock_decref(&handle, mode);
2373 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2374 aa->oa_lockh, req, aa);
2375 ldlm_lock_decref(&handle, mode);
2376 LDLM_LOCK_PUT(lock);
2380 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2381 struct lov_oinfo *loi, int flags,
2382 struct ost_lvb *lvb, __u32 mode, int rc)
2384 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2386 if (rc == ELDLM_OK) {
2389 LASSERT(lock != NULL);
2390 loi->loi_lvb = *lvb;
2391 tmp = loi->loi_lvb.lvb_size;
2392 /* Extend KMS up to the end of this lock and no further
2393 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2394 if (tmp > lock->l_policy_data.l_extent.end)
2395 tmp = lock->l_policy_data.l_extent.end + 1;
2396 if (tmp >= loi->loi_kms) {
2397 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2398 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2399 loi_kms_set(loi, tmp);
2401 LDLM_DEBUG(lock, "lock acquired, setting rss="
2402 LPU64"; leaving kms="LPU64", end="LPU64,
2403 loi->loi_lvb.lvb_size, loi->loi_kms,
2404 lock->l_policy_data.l_extent.end);
2406 ldlm_lock_allow_match(lock);
2407 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2408 LASSERT(lock != NULL);
2409 loi->loi_lvb = *lvb;
2410 ldlm_lock_allow_match(lock);
2411 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2412 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2418 ldlm_lock_fail_match(lock);
2420 LDLM_LOCK_PUT(lock);
2423 EXPORT_SYMBOL(osc_update_enqueue);
2425 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2427 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2428 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2429 * other synchronous requests, however keeping some locks and trying to obtain
2430 * others may take a considerable amount of time in a case of ost failure; and
2431 * when other sync requests do not get released lock from a client, the client
2432 * is excluded from the cluster -- such scenarious make the life difficult, so
2433 * release locks just after they are obtained. */
2434 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2435 int *flags, ldlm_policy_data_t *policy,
2436 struct ost_lvb *lvb, int kms_valid,
2437 obd_enqueue_update_f upcall, void *cookie,
2438 struct ldlm_enqueue_info *einfo,
2439 struct lustre_handle *lockh,
2440 struct ptlrpc_request_set *rqset, int async, int agl)
2442 struct obd_device *obd = exp->exp_obd;
2443 struct ptlrpc_request *req = NULL;
2444 int intent = *flags & LDLM_FL_HAS_INTENT;
2445 int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2450 /* Filesystem lock extents are extended to page boundaries so that
2451 * dealing with the page cache is a little smoother. */
2452 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2453 policy->l_extent.end |= ~CFS_PAGE_MASK;
2456 * kms is not valid when either object is completely fresh (so that no
2457 * locks are cached), or object was evicted. In the latter case cached
2458 * lock cannot be used, because it would prime inode state with
2459 * potentially stale LVB.
2464 /* Next, search for already existing extent locks that will cover us */
2465 /* If we're trying to read, we also search for an existing PW lock. The
2466 * VFS and page cache already protect us locally, so lots of readers/
2467 * writers can share a single PW lock.
2469 * There are problems with conversion deadlocks, so instead of
2470 * converting a read lock to a write lock, we'll just enqueue a new
2473 * At some point we should cancel the read lock instead of making them
2474 * send us a blocking callback, but there are problems with canceling
2475 * locks out from other users right now, too. */
2476 mode = einfo->ei_mode;
2477 if (einfo->ei_mode == LCK_PR)
2479 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2480 einfo->ei_type, policy, mode, lockh, 0);
2482 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2484 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2485 /* For AGL, if enqueue RPC is sent but the lock is not
2486 * granted, then skip to process this strpe.
2487 * Return -ECANCELED to tell the caller. */
2488 ldlm_lock_decref(lockh, mode);
2489 LDLM_LOCK_PUT(matched);
2491 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2492 *flags |= LDLM_FL_LVB_READY;
2493 /* addref the lock only if not async requests and PW
2494 * lock is matched whereas we asked for PR. */
2495 if (!rqset && einfo->ei_mode != mode)
2496 ldlm_lock_addref(lockh, LCK_PR);
2498 /* I would like to be able to ASSERT here that
2499 * rss <= kms, but I can't, for reasons which
2500 * are explained in lov_enqueue() */
2503 /* We already have a lock, and it's referenced.
2505 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2506 * AGL upcall may change it to CLS_HELD directly. */
2507 (*upcall)(cookie, ELDLM_OK);
2509 if (einfo->ei_mode != mode)
2510 ldlm_lock_decref(lockh, LCK_PW);
2512 /* For async requests, decref the lock. */
2513 ldlm_lock_decref(lockh, einfo->ei_mode);
2514 LDLM_LOCK_PUT(matched);
2517 ldlm_lock_decref(lockh, mode);
2518 LDLM_LOCK_PUT(matched);
2524 CFS_LIST_HEAD(cancels);
2525 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2526 &RQF_LDLM_ENQUEUE_LVB);
2530 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2532 ptlrpc_request_free(req);
2536 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2538 ptlrpc_request_set_replen(req);
2541 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2542 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2544 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2545 sizeof(*lvb), lockh, async);
2548 struct osc_enqueue_args *aa;
2549 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2550 aa = ptlrpc_req_async_args(req);
2553 aa->oa_flags = flags;
2554 aa->oa_upcall = upcall;
2555 aa->oa_cookie = cookie;
2557 aa->oa_lockh = lockh;
2560 req->rq_interpret_reply =
2561 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2562 if (rqset == PTLRPCD_SET)
2563 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2565 ptlrpc_set_add_req(rqset, req);
2566 } else if (intent) {
2567 ptlrpc_req_finished(req);
2572 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2574 ptlrpc_req_finished(req);
2579 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2580 struct ldlm_enqueue_info *einfo,
2581 struct ptlrpc_request_set *rqset)
2583 struct ldlm_res_id res_id;
2587 osc_build_res_name(oinfo->oi_md->lsm_object_id,
2588 oinfo->oi_md->lsm_object_seq, &res_id);
2590 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2591 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2592 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2593 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2594 rqset, rqset != NULL, 0);
2598 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2599 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2600 int *flags, void *data, struct lustre_handle *lockh,
2603 struct obd_device *obd = exp->exp_obd;
2604 int lflags = *flags;
2608 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2611 /* Filesystem lock extents are extended to page boundaries so that
2612 * dealing with the page cache is a little smoother */
2613 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2614 policy->l_extent.end |= ~CFS_PAGE_MASK;
2616 /* Next, search for already existing extent locks that will cover us */
2617 /* If we're trying to read, we also search for an existing PW lock. The
2618 * VFS and page cache already protect us locally, so lots of readers/
2619 * writers can share a single PW lock. */
2623 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2624 res_id, type, policy, rc, lockh, unref);
2627 if (!osc_set_data_with_check(lockh, data)) {
2628 if (!(lflags & LDLM_FL_TEST_LOCK))
2629 ldlm_lock_decref(lockh, rc);
2633 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2634 ldlm_lock_addref(lockh, LCK_PR);
2635 ldlm_lock_decref(lockh, LCK_PW);
2642 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2646 if (unlikely(mode == LCK_GROUP))
2647 ldlm_lock_decref_and_cancel(lockh, mode);
2649 ldlm_lock_decref(lockh, mode);
2654 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2655 __u32 mode, struct lustre_handle *lockh)
2658 RETURN(osc_cancel_base(lockh, mode));
2661 static int osc_cancel_unused(struct obd_export *exp,
2662 struct lov_stripe_md *lsm,
2663 ldlm_cancel_flags_t flags,
2666 struct obd_device *obd = class_exp2obd(exp);
2667 struct ldlm_res_id res_id, *resp = NULL;
2670 resp = osc_build_res_name(lsm->lsm_object_id,
2671 lsm->lsm_object_seq, &res_id);
2674 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2677 static int osc_statfs_interpret(const struct lu_env *env,
2678 struct ptlrpc_request *req,
2679 struct osc_async_args *aa, int rc)
2681 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
2682 struct obd_statfs *msfs;
2687 /* The request has in fact never been sent
2688 * due to issues at a higher level (LOV).
2689 * Exit immediately since the caller is
2690 * aware of the problem and takes care
2691 * of the clean up */
2694 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2695 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2701 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2703 GOTO(out, rc = -EPROTO);
2706 /* Reinitialize the RDONLY and DEGRADED flags at the client
2707 * on each statfs, so they don't stay set permanently. */
2708 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
2710 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
2711 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
2712 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
2713 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
2715 if (unlikely(msfs->os_state & OS_STATE_READONLY))
2716 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
2717 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
2718 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
2720 /* Add a bit of hysteresis so this flag isn't continually flapping,
2721 * and ensure that new files don't get extremely fragmented due to
2722 * only a small amount of available space in the filesystem.
2723 * We want to set the NOSPC flag when there is less than ~0.1% free
2724 * and clear it when there is at least ~0.2% free space, so:
2725 * avail < ~0.1% max max = avail + used
2726 * 1025 * avail < avail + used used = blocks - free
2727 * 1024 * avail < used
2728 * 1024 * avail < blocks - free
2729 * avail < ((blocks - free) >> 10)
2731 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
2732 * lose that amount of space so in those cases we report no space left
2733 * if their is less than 1 GB left. */
2734 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
2735 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
2736 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
2737 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
2738 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
2739 (msfs->os_ffree > 64) &&
2740 (msfs->os_bavail > (used << 1)))) {
2741 cli->cl_oscc.oscc_flags &= ~(OSCC_FLAG_NOSPC |
2742 OSCC_FLAG_NOSPC_BLK);
2745 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
2746 (msfs->os_bavail < used)))
2747 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC_BLK;
2749 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
2751 *aa->aa_oi->oi_osfs = *msfs;
2753 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2757 static int osc_statfs_async(struct obd_export *exp,
2758 struct obd_info *oinfo, __u64 max_age,
2759 struct ptlrpc_request_set *rqset)
2761 struct obd_device *obd = class_exp2obd(exp);
2762 struct ptlrpc_request *req;
2763 struct osc_async_args *aa;
2767 /* We could possibly pass max_age in the request (as an absolute
2768 * timestamp or a "seconds.usec ago") so the target can avoid doing
2769 * extra calls into the filesystem if that isn't necessary (e.g.
2770 * during mount that would help a bit). Having relative timestamps
2771 * is not so great if request processing is slow, while absolute
2772 * timestamps are not ideal because they need time synchronization. */
2773 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2777 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2779 ptlrpc_request_free(req);
2782 ptlrpc_request_set_replen(req);
2783 req->rq_request_portal = OST_CREATE_PORTAL;
2784 ptlrpc_at_set_req_timeout(req);
2786 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2787 /* procfs requests not want stat in wait for avoid deadlock */
2788 req->rq_no_resend = 1;
2789 req->rq_no_delay = 1;
2792 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2793 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2794 aa = ptlrpc_req_async_args(req);
2797 ptlrpc_set_add_req(rqset, req);
2801 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2802 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2804 struct obd_device *obd = class_exp2obd(exp);
2805 struct obd_statfs *msfs;
2806 struct ptlrpc_request *req;
2807 struct obd_import *imp = NULL;
2811 /*Since the request might also come from lprocfs, so we need
2812 *sync this with client_disconnect_export Bug15684*/
2813 cfs_down_read(&obd->u.cli.cl_sem);
2814 if (obd->u.cli.cl_import)
2815 imp = class_import_get(obd->u.cli.cl_import);
2816 cfs_up_read(&obd->u.cli.cl_sem);
2820 /* We could possibly pass max_age in the request (as an absolute
2821 * timestamp or a "seconds.usec ago") so the target can avoid doing
2822 * extra calls into the filesystem if that isn't necessary (e.g.
2823 * during mount that would help a bit). Having relative timestamps
2824 * is not so great if request processing is slow, while absolute
2825 * timestamps are not ideal because they need time synchronization. */
2826 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2828 class_import_put(imp);
2833 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2835 ptlrpc_request_free(req);
2838 ptlrpc_request_set_replen(req);
2839 req->rq_request_portal = OST_CREATE_PORTAL;
2840 ptlrpc_at_set_req_timeout(req);
2842 if (flags & OBD_STATFS_NODELAY) {
2843 /* procfs requests not want stat in wait for avoid deadlock */
2844 req->rq_no_resend = 1;
2845 req->rq_no_delay = 1;
2848 rc = ptlrpc_queue_wait(req);
2852 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2854 GOTO(out, rc = -EPROTO);
2861 ptlrpc_req_finished(req);
2865 /* Retrieve object striping information.
2867 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2868 * the maximum number of OST indices which will fit in the user buffer.
2869 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2871 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2873 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2874 struct lov_user_md_v3 lum, *lumk;
2875 struct lov_user_ost_data_v1 *lmm_objects;
2876 int rc = 0, lum_size;
2882 /* we only need the header part from user space to get lmm_magic and
2883 * lmm_stripe_count, (the header part is common to v1 and v3) */
2884 lum_size = sizeof(struct lov_user_md_v1);
2885 if (cfs_copy_from_user(&lum, lump, lum_size))
2888 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2889 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2892 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2893 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2894 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2895 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2897 /* we can use lov_mds_md_size() to compute lum_size
2898 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2899 if (lum.lmm_stripe_count > 0) {
2900 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2901 OBD_ALLOC(lumk, lum_size);
2905 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2906 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2908 lmm_objects = &(lumk->lmm_objects[0]);
2909 lmm_objects->l_object_id = lsm->lsm_object_id;
2911 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2915 lumk->lmm_object_id = lsm->lsm_object_id;
2916 lumk->lmm_object_seq = lsm->lsm_object_seq;
2917 lumk->lmm_stripe_count = 1;
2919 if (cfs_copy_to_user(lump, lumk, lum_size))
2923 OBD_FREE(lumk, lum_size);
2929 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2930 void *karg, void *uarg)
2932 struct obd_device *obd = exp->exp_obd;
2933 struct obd_ioctl_data *data = karg;
2937 if (!cfs_try_module_get(THIS_MODULE)) {
2938 CERROR("Can't get module. Is it alive?");
2942 case OBD_IOC_LOV_GET_CONFIG: {
2944 struct lov_desc *desc;
2945 struct obd_uuid uuid;
2949 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2950 GOTO(out, err = -EINVAL);
2952 data = (struct obd_ioctl_data *)buf;
2954 if (sizeof(*desc) > data->ioc_inllen1) {
2955 obd_ioctl_freedata(buf, len);
2956 GOTO(out, err = -EINVAL);
2959 if (data->ioc_inllen2 < sizeof(uuid)) {
2960 obd_ioctl_freedata(buf, len);
2961 GOTO(out, err = -EINVAL);
2964 desc = (struct lov_desc *)data->ioc_inlbuf1;
2965 desc->ld_tgt_count = 1;
2966 desc->ld_active_tgt_count = 1;
2967 desc->ld_default_stripe_count = 1;
2968 desc->ld_default_stripe_size = 0;
2969 desc->ld_default_stripe_offset = 0;
2970 desc->ld_pattern = 0;
2971 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2973 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2975 err = cfs_copy_to_user((void *)uarg, buf, len);
2978 obd_ioctl_freedata(buf, len);
2981 case LL_IOC_LOV_SETSTRIPE:
2982 err = obd_alloc_memmd(exp, karg);
2986 case LL_IOC_LOV_GETSTRIPE:
2987 err = osc_getstripe(karg, uarg);
2989 case OBD_IOC_CLIENT_RECOVER:
2990 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2991 data->ioc_inlbuf1, 0);
2995 case IOC_OSC_SET_ACTIVE:
2996 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2999 case OBD_IOC_POLL_QUOTACHECK:
3000 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
3002 case OBD_IOC_PING_TARGET:
3003 err = ptlrpc_obd_ping(obd);
3006 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3007 cmd, cfs_curproc_comm());
3008 GOTO(out, err = -ENOTTY);
3011 cfs_module_put(THIS_MODULE);
3015 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3016 obd_count keylen, void *key, __u32 *vallen, void *val,
3017 struct lov_stripe_md *lsm)
3020 if (!vallen || !val)
3023 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3024 __u32 *stripe = val;
3025 *vallen = sizeof(*stripe);
3028 } else if (KEY_IS(KEY_LAST_ID)) {
3029 struct ptlrpc_request *req;
3034 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3035 &RQF_OST_GET_INFO_LAST_ID);
3039 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3040 RCL_CLIENT, keylen);
3041 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3043 ptlrpc_request_free(req);
3047 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3048 memcpy(tmp, key, keylen);
3050 req->rq_no_delay = req->rq_no_resend = 1;
3051 ptlrpc_request_set_replen(req);
3052 rc = ptlrpc_queue_wait(req);
3056 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3058 GOTO(out, rc = -EPROTO);
3060 *((obd_id *)val) = *reply;
3062 ptlrpc_req_finished(req);
3064 } else if (KEY_IS(KEY_FIEMAP)) {
3065 struct ptlrpc_request *req;
3066 struct ll_user_fiemap *reply;
3070 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3071 &RQF_OST_GET_INFO_FIEMAP);
3075 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3076 RCL_CLIENT, keylen);
3077 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3078 RCL_CLIENT, *vallen);
3079 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3080 RCL_SERVER, *vallen);
3082 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3084 ptlrpc_request_free(req);
3088 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3089 memcpy(tmp, key, keylen);
3090 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3091 memcpy(tmp, val, *vallen);
3093 ptlrpc_request_set_replen(req);
3094 rc = ptlrpc_queue_wait(req);
3098 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3100 GOTO(out1, rc = -EPROTO);
3102 memcpy(val, reply, *vallen);
3104 ptlrpc_req_finished(req);
3112 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3114 struct llog_ctxt *ctxt;
3118 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3120 rc = llog_initiator_connect(ctxt);
3121 llog_ctxt_put(ctxt);
3123 /* XXX return an error? skip setting below flags? */
3126 cfs_spin_lock(&imp->imp_lock);
3127 imp->imp_server_timeout = 1;
3128 imp->imp_pingable = 1;
3129 cfs_spin_unlock(&imp->imp_lock);
3130 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3135 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3136 struct ptlrpc_request *req,
3143 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3146 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3147 obd_count keylen, void *key, obd_count vallen,
3148 void *val, struct ptlrpc_request_set *set)
3150 struct ptlrpc_request *req;
3151 struct obd_device *obd = exp->exp_obd;
3152 struct obd_import *imp = class_exp2cliimp(exp);
3157 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3159 if (KEY_IS(KEY_NEXT_ID)) {
3161 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3163 if (vallen != sizeof(obd_id))
3168 if (vallen != sizeof(obd_id))
3171 /* avoid race between allocate new object and set next id
3172 * from ll_sync thread */
3173 cfs_spin_lock(&oscc->oscc_lock);
3174 new_val = *((obd_id*)val) + 1;
3175 if (new_val > oscc->oscc_next_id)
3176 oscc->oscc_next_id = new_val;
3177 cfs_spin_unlock(&oscc->oscc_lock);
3178 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3179 exp->exp_obd->obd_name,
3180 obd->u.cli.cl_oscc.oscc_next_id);
3185 if (KEY_IS(KEY_CHECKSUM)) {
3186 if (vallen != sizeof(int))
3188 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3192 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3193 sptlrpc_conf_client_adapt(obd);
3197 if (KEY_IS(KEY_FLUSH_CTX)) {
3198 sptlrpc_import_flush_my_ctx(imp);
3202 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3205 /* We pass all other commands directly to OST. Since nobody calls osc
3206 methods directly and everybody is supposed to go through LOV, we
3207 assume lov checked invalid values for us.
3208 The only recognised values so far are evict_by_nid and mds_conn.
3209 Even if something bad goes through, we'd get a -EINVAL from OST
3212 if (KEY_IS(KEY_GRANT_SHRINK))
3213 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
3215 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
3220 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3221 RCL_CLIENT, keylen);
3222 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3223 RCL_CLIENT, vallen);
3224 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3226 ptlrpc_request_free(req);
3230 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3231 memcpy(tmp, key, keylen);
3232 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3233 memcpy(tmp, val, vallen);
3235 if (KEY_IS(KEY_MDS_CONN)) {
3236 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3238 oscc->oscc_oa.o_seq = (*(__u32 *)val);
3239 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3240 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
3241 req->rq_no_delay = req->rq_no_resend = 1;
3242 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3243 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
3244 struct osc_grant_args *aa;
3247 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3248 aa = ptlrpc_req_async_args(req);
3251 ptlrpc_req_finished(req);
3254 *oa = ((struct ost_body *)val)->oa;
3256 req->rq_interpret_reply = osc_shrink_grant_interpret;
3259 ptlrpc_request_set_replen(req);
3260 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3261 LASSERT(set != NULL);
3262 ptlrpc_set_add_req(set, req);
3263 ptlrpc_check_set(NULL, set);
3265 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3271 static struct llog_operations osc_size_repl_logops = {
3272 lop_cancel: llog_obd_repl_cancel
3275 static struct llog_operations osc_mds_ost_orig_logops;
3277 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3278 struct obd_device *tgt, struct llog_catid *catid)
3280 struct llog_ctxt *ctxt = NULL;
3281 struct llog_handle *lgh;
3286 osc_mds_ost_orig_logops = llog_lvfs_ops;
3287 osc_mds_ost_orig_logops.lop_obd_add = llog_obd_origin_add;
3288 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3289 rc = llog_setup(NULL, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt,
3290 &osc_mds_ost_orig_logops);
3294 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3297 /* context might be initialized already */
3298 if (ctxt->loc_handle != NULL) {
3299 /* sanity check for valid loc_handle */
3300 LASSERT(ctxt->loc_handle->lgh_ctxt == ctxt);
3304 /* first try to open existent llog by ID */
3305 if (likely(catid->lci_logid.lgl_oid != 0)) {
3306 rc = llog_open(NULL, ctxt, &lgh, &catid->lci_logid, NULL,
3308 /* re-create llog if it is missing */
3310 catid->lci_logid.lgl_oid = 0;
3312 GOTO(out_cleanup, rc);
3314 /* create new llog if llog ID is not specified or llog is missed */
3315 if (unlikely(catid->lci_logid.lgl_oid == 0)) {
3316 rc = llog_open_create(NULL, ctxt, &lgh, NULL, NULL);
3318 GOTO(out_cleanup, rc);
3319 catid->lci_logid = lgh->lgh_id;
3322 ctxt->loc_handle = lgh;
3324 rc = llog_cat_init_and_process(NULL, lgh);
3326 GOTO(out_close, rc);
3328 rc = llog_setup(NULL, obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt,
3329 &osc_size_repl_logops);
3331 GOTO(out_close, rc);
3333 llog_ctxt_put(ctxt);
3336 llog_cat_close(NULL, lgh);
3338 llog_cleanup(NULL, ctxt);
3339 CERROR("%s: fail to init llog #"LPX64"#"LPX64"#%08x tgt '%s': "
3340 "rc = %d\n", obd->obd_name, catid->lci_logid.lgl_oid,
3341 catid->lci_logid.lgl_oseq, catid->lci_logid.lgl_ogen,
3346 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3347 struct obd_device *disk_obd, int *index)
3349 struct llog_catid catid;
3350 static char name[32] = CATLIST;
3354 LASSERT(olg == &obd->obd_olg);
3356 cfs_mutex_lock(&olg->olg_cat_processing);
3357 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
3359 CERROR("rc: %d\n", rc);
3363 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
3364 obd->obd_name, *index, catid.lci_logid.lgl_oid,
3365 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
3367 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
3369 CERROR("rc: %d\n", rc);
3373 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
3375 CERROR("rc: %d\n", rc);
3380 cfs_mutex_unlock(&olg->olg_cat_processing);
3385 static int osc_llog_finish(struct obd_device *obd, int count)
3387 struct llog_ctxt *ctxt;
3391 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3393 llog_cat_close(NULL, ctxt->loc_handle);
3394 llog_cleanup(NULL, ctxt);
3397 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3399 llog_cleanup(NULL, ctxt);
3403 static int osc_reconnect(const struct lu_env *env,
3404 struct obd_export *exp, struct obd_device *obd,
3405 struct obd_uuid *cluuid,
3406 struct obd_connect_data *data,
3409 struct client_obd *cli = &obd->u.cli;
3411 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3414 client_obd_list_lock(&cli->cl_loi_list_lock);
3415 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3416 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3417 lost_grant = cli->cl_lost_grant;
3418 cli->cl_lost_grant = 0;
3419 client_obd_list_unlock(&cli->cl_loi_list_lock);
3421 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3422 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3423 data->ocd_version, data->ocd_grant, lost_grant);
3429 static int osc_disconnect(struct obd_export *exp)
3431 struct obd_device *obd = class_exp2obd(exp);
3432 struct llog_ctxt *ctxt;
3435 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3437 if (obd->u.cli.cl_conn_count == 1) {
3438 /* Flush any remaining cancel messages out to the
3440 llog_sync(ctxt, exp, 0);
3442 llog_ctxt_put(ctxt);
3444 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3448 rc = client_disconnect_export(exp);
3450 * Initially we put del_shrink_grant before disconnect_export, but it
3451 * causes the following problem if setup (connect) and cleanup
3452 * (disconnect) are tangled together.
3453 * connect p1 disconnect p2
3454 * ptlrpc_connect_import
3455 * ............... class_manual_cleanup
3458 * ptlrpc_connect_interrupt
3460 * add this client to shrink list
3462 * Bang! pinger trigger the shrink.
3463 * So the osc should be disconnected from the shrink list, after we
3464 * are sure the import has been destroyed. BUG18662
3466 if (obd->u.cli.cl_import == NULL)
3467 osc_del_shrink_grant(&obd->u.cli);
3471 static int osc_import_event(struct obd_device *obd,
3472 struct obd_import *imp,
3473 enum obd_import_event event)
3475 struct client_obd *cli;
3479 LASSERT(imp->imp_obd == obd);
3482 case IMP_EVENT_DISCON: {
3483 /* Only do this on the MDS OSC's */
3484 if (imp->imp_server_timeout) {
3485 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3487 cfs_spin_lock(&oscc->oscc_lock);
3488 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3489 cfs_spin_unlock(&oscc->oscc_lock);
3492 client_obd_list_lock(&cli->cl_loi_list_lock);
3493 cli->cl_avail_grant = 0;
3494 cli->cl_lost_grant = 0;
3495 client_obd_list_unlock(&cli->cl_loi_list_lock);
3498 case IMP_EVENT_INACTIVE: {
3499 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3502 case IMP_EVENT_INVALIDATE: {
3503 struct ldlm_namespace *ns = obd->obd_namespace;
3507 env = cl_env_get(&refcheck);
3511 /* all pages go to failing rpcs due to the invalid
3513 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3515 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3516 cl_env_put(env, &refcheck);
3521 case IMP_EVENT_ACTIVE: {
3522 /* Only do this on the MDS OSC's */
3523 if (imp->imp_server_timeout) {
3524 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3526 cfs_spin_lock(&oscc->oscc_lock);
3527 oscc->oscc_flags &= ~(OSCC_FLAG_NOSPC |
3528 OSCC_FLAG_NOSPC_BLK);
3529 cfs_spin_unlock(&oscc->oscc_lock);
3531 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3534 case IMP_EVENT_OCD: {
3535 struct obd_connect_data *ocd = &imp->imp_connect_data;
3537 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3538 osc_init_grant(&obd->u.cli, ocd);
3541 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3542 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3544 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3547 case IMP_EVENT_DEACTIVATE: {
3548 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3551 case IMP_EVENT_ACTIVATE: {
3552 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3556 CERROR("Unknown import event %d\n", event);
3563 * Determine whether the lock can be canceled before replaying the lock
3564 * during recovery, see bug16774 for detailed information.
3566 * \retval zero the lock can't be canceled
3567 * \retval other ok to cancel
3569 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3571 check_res_locked(lock->l_resource);
3574 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3576 * XXX as a future improvement, we can also cancel unused write lock
3577 * if it doesn't have dirty data and active mmaps.
3579 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3580 (lock->l_granted_mode == LCK_PR ||
3581 lock->l_granted_mode == LCK_CR) &&
3582 (osc_dlm_lock_pageref(lock) == 0))
3588 static int brw_queue_work(const struct lu_env *env, void *data)
3590 struct client_obd *cli = data;
3592 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3594 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3598 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3600 struct lprocfs_static_vars lvars = { 0 };
3601 struct client_obd *cli = &obd->u.cli;
3606 rc = ptlrpcd_addref();
3610 rc = client_obd_setup(obd, lcfg);
3612 GOTO(out_ptlrpcd, rc);
3614 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3615 if (IS_ERR(handler))
3616 GOTO(out_client_setup, rc = PTR_ERR(handler));
3617 cli->cl_writeback_work = handler;
3619 rc = osc_quota_setup(obd);
3621 GOTO(out_ptlrpcd_work, rc);
3623 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3624 lprocfs_osc_init_vars(&lvars);
3625 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3626 lproc_osc_attach_seqstat(obd);
3627 sptlrpc_lprocfs_cliobd_attach(obd);
3628 ptlrpc_lprocfs_register_obd(obd);
3632 /* We need to allocate a few requests more, because
3633 * brw_interpret tries to create new requests before freeing
3634 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3635 * reserved, but I'm afraid that might be too much wasted RAM
3636 * in fact, so 2 is just my guess and still should work. */
3637 cli->cl_import->imp_rq_pool =
3638 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3640 ptlrpc_add_rqs_to_pool);
3642 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3643 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3647 ptlrpcd_destroy_work(handler);
3649 client_obd_cleanup(obd);
3655 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3661 case OBD_CLEANUP_EARLY: {
3662 struct obd_import *imp;
3663 imp = obd->u.cli.cl_import;
3664 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3665 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3666 ptlrpc_deactivate_import(imp);
3667 cfs_spin_lock(&imp->imp_lock);
3668 imp->imp_pingable = 0;
3669 cfs_spin_unlock(&imp->imp_lock);
3672 case OBD_CLEANUP_EXPORTS: {
3673 struct client_obd *cli = &obd->u.cli;
3675 * for echo client, export may be on zombie list, wait for
3676 * zombie thread to cull it, because cli.cl_import will be
3677 * cleared in client_disconnect_export():
3678 * class_export_destroy() -> obd_cleanup() ->
3679 * echo_device_free() -> echo_client_cleanup() ->
3680 * obd_disconnect() -> osc_disconnect() ->
3681 * client_disconnect_export()
3683 obd_zombie_barrier();
3684 if (cli->cl_writeback_work) {
3685 ptlrpcd_destroy_work(cli->cl_writeback_work);
3686 cli->cl_writeback_work = NULL;
3688 obd_cleanup_client_import(obd);
3689 ptlrpc_lprocfs_unregister_obd(obd);
3690 lprocfs_obd_cleanup(obd);
3691 rc = obd_llog_finish(obd, 0);
3693 CERROR("failed to cleanup llogging subsystems\n");
3700 int osc_cleanup(struct obd_device *obd)
3706 /* free memory of osc quota cache */
3707 osc_quota_cleanup(obd);
3709 rc = client_obd_cleanup(obd);
3715 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3717 struct lprocfs_static_vars lvars = { 0 };
3720 lprocfs_osc_init_vars(&lvars);
3722 switch (lcfg->lcfg_command) {
3724 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3734 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3736 return osc_process_config_base(obd, buf);
3739 struct obd_ops osc_obd_ops = {
3740 .o_owner = THIS_MODULE,
3741 .o_setup = osc_setup,
3742 .o_precleanup = osc_precleanup,
3743 .o_cleanup = osc_cleanup,
3744 .o_add_conn = client_import_add_conn,
3745 .o_del_conn = client_import_del_conn,
3746 .o_connect = client_connect_import,
3747 .o_reconnect = osc_reconnect,
3748 .o_disconnect = osc_disconnect,
3749 .o_statfs = osc_statfs,
3750 .o_statfs_async = osc_statfs_async,
3751 .o_packmd = osc_packmd,
3752 .o_unpackmd = osc_unpackmd,
3753 .o_precreate = osc_precreate,
3754 .o_create = osc_create,
3755 .o_create_async = osc_create_async,
3756 .o_destroy = osc_destroy,
3757 .o_getattr = osc_getattr,
3758 .o_getattr_async = osc_getattr_async,
3759 .o_setattr = osc_setattr,
3760 .o_setattr_async = osc_setattr_async,
3762 .o_punch = osc_punch,
3764 .o_enqueue = osc_enqueue,
3765 .o_change_cbdata = osc_change_cbdata,
3766 .o_find_cbdata = osc_find_cbdata,
3767 .o_cancel = osc_cancel,
3768 .o_cancel_unused = osc_cancel_unused,
3769 .o_iocontrol = osc_iocontrol,
3770 .o_get_info = osc_get_info,
3771 .o_set_info_async = osc_set_info_async,
3772 .o_import_event = osc_import_event,
3773 .o_llog_init = osc_llog_init,
3774 .o_llog_finish = osc_llog_finish,
3775 .o_process_config = osc_process_config,
3776 .o_quotactl = osc_quotactl,
3777 .o_quotacheck = osc_quotacheck,
3780 extern struct lu_kmem_descr osc_caches[];
3781 extern cfs_spinlock_t osc_ast_guard;
3782 extern cfs_lock_class_key_t osc_ast_guard_class;
3784 int __init osc_init(void)
3786 struct lprocfs_static_vars lvars = { 0 };
3790 /* print an address of _any_ initialized kernel symbol from this
3791 * module, to allow debugging with gdb that doesn't support data
3792 * symbols from modules.*/
3793 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3795 rc = lu_kmem_init(osc_caches);
3797 lprocfs_osc_init_vars(&lvars);
3799 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3800 LUSTRE_OSC_NAME, &osc_device_type);
3802 lu_kmem_fini(osc_caches);
3806 cfs_spin_lock_init(&osc_ast_guard);
3807 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3813 static void /*__exit*/ osc_exit(void)
3815 class_unregister_type(LUSTRE_OSC_NAME);
3816 lu_kmem_fini(osc_caches);
3819 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3820 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3821 MODULE_LICENSE("GPL");
3823 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);