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);
1255 if (opc == OST_WRITE)
1256 desc = ptlrpc_prep_bulk_imp(req, page_count,
1257 BULK_GET_SOURCE, OST_BULK_PORTAL);
1259 desc = ptlrpc_prep_bulk_imp(req, page_count,
1260 BULK_PUT_SINK, OST_BULK_PORTAL);
1263 GOTO(out, rc = -ENOMEM);
1264 /* NB request now owns desc and will free it when it gets freed */
1266 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1267 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1268 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1269 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1271 lustre_set_wire_obdo(&body->oa, oa);
1273 obdo_to_ioobj(oa, ioobj);
1274 ioobj->ioo_bufcnt = niocount;
1275 osc_pack_capa(req, body, ocapa);
1276 LASSERT (page_count > 0);
1278 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1279 struct brw_page *pg = pga[i];
1280 int poff = pg->off & ~CFS_PAGE_MASK;
1282 LASSERT(pg->count > 0);
1283 /* make sure there is no gap in the middle of page array */
1284 LASSERTF(page_count == 1 ||
1285 (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1286 ergo(i > 0 && i < page_count - 1,
1287 poff == 0 && pg->count == CFS_PAGE_SIZE) &&
1288 ergo(i == page_count - 1, poff == 0)),
1289 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1290 i, page_count, pg, pg->off, pg->count);
1292 LASSERTF(i == 0 || pg->off > pg_prev->off,
1293 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1294 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1296 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1297 pg_prev->pg, page_private(pg_prev->pg),
1298 pg_prev->pg->index, pg_prev->off);
1300 LASSERTF(i == 0 || pg->off > pg_prev->off,
1301 "i %d p_c %u\n", i, page_count);
1303 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1304 (pg->flag & OBD_BRW_SRVLOCK));
1306 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1307 requested_nob += pg->count;
1309 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1311 niobuf->len += pg->count;
1313 niobuf->offset = pg->off;
1314 niobuf->len = pg->count;
1315 niobuf->flags = pg->flag;
1320 LASSERTF((void *)(niobuf - niocount) ==
1321 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1322 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1323 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1325 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1327 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1328 body->oa.o_valid |= OBD_MD_FLFLAGS;
1329 body->oa.o_flags = 0;
1331 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1334 if (osc_should_shrink_grant(cli))
1335 osc_shrink_grant_local(cli, &body->oa);
1337 /* size[REQ_REC_OFF] still sizeof (*body) */
1338 if (opc == OST_WRITE) {
1339 if (cli->cl_checksum &&
1340 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1341 /* store cl_cksum_type in a local variable since
1342 * it can be changed via lprocfs */
1343 cksum_type_t cksum_type = cli->cl_cksum_type;
1345 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1346 oa->o_flags &= OBD_FL_LOCAL_MASK;
1347 body->oa.o_flags = 0;
1349 body->oa.o_flags |= cksum_type_pack(cksum_type);
1350 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1351 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1355 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1357 /* save this in 'oa', too, for later checking */
1358 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1359 oa->o_flags |= cksum_type_pack(cksum_type);
1361 /* clear out the checksum flag, in case this is a
1362 * resend but cl_checksum is no longer set. b=11238 */
1363 oa->o_valid &= ~OBD_MD_FLCKSUM;
1365 oa->o_cksum = body->oa.o_cksum;
1366 /* 1 RC per niobuf */
1367 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1368 sizeof(__u32) * niocount);
1370 if (cli->cl_checksum &&
1371 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1372 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1373 body->oa.o_flags = 0;
1374 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1375 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1378 ptlrpc_request_set_replen(req);
1380 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1381 aa = ptlrpc_req_async_args(req);
1383 aa->aa_requested_nob = requested_nob;
1384 aa->aa_nio_count = niocount;
1385 aa->aa_page_count = page_count;
1389 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1390 if (ocapa && reserve)
1391 aa->aa_ocapa = capa_get(ocapa);
1397 ptlrpc_req_finished(req);
1401 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1402 __u32 client_cksum, __u32 server_cksum, int nob,
1403 obd_count page_count, struct brw_page **pga,
1404 cksum_type_t client_cksum_type)
1408 cksum_type_t cksum_type;
1410 if (server_cksum == client_cksum) {
1411 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1415 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1417 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1420 if (cksum_type != client_cksum_type)
1421 msg = "the server did not use the checksum type specified in "
1422 "the original request - likely a protocol problem";
1423 else if (new_cksum == server_cksum)
1424 msg = "changed on the client after we checksummed it - "
1425 "likely false positive due to mmap IO (bug 11742)";
1426 else if (new_cksum == client_cksum)
1427 msg = "changed in transit before arrival at OST";
1429 msg = "changed in transit AND doesn't match the original - "
1430 "likely false positive due to mmap IO (bug 11742)";
1432 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1433 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1434 msg, libcfs_nid2str(peer->nid),
1435 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1436 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1437 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1439 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1441 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1442 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1443 "client csum now %x\n", client_cksum, client_cksum_type,
1444 server_cksum, cksum_type, new_cksum);
1448 /* Note rc enters this function as number of bytes transferred */
1449 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1451 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1452 const lnet_process_id_t *peer =
1453 &req->rq_import->imp_connection->c_peer;
1454 struct client_obd *cli = aa->aa_cli;
1455 struct ost_body *body;
1456 __u32 client_cksum = 0;
1459 if (rc < 0 && rc != -EDQUOT) {
1460 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1464 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1465 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1467 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1471 /* set/clear over quota flag for a uid/gid */
1472 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1473 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1474 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1476 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1477 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1479 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1482 osc_update_grant(cli, body);
1487 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1488 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1490 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1492 CERROR("Unexpected +ve rc %d\n", rc);
1495 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1497 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1500 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1501 check_write_checksum(&body->oa, peer, client_cksum,
1502 body->oa.o_cksum, aa->aa_requested_nob,
1503 aa->aa_page_count, aa->aa_ppga,
1504 cksum_type_unpack(aa->aa_oa->o_flags)))
1507 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1508 aa->aa_page_count, aa->aa_ppga);
1512 /* The rest of this function executes only for OST_READs */
1514 /* if unwrap_bulk failed, return -EAGAIN to retry */
1515 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1517 GOTO(out, rc = -EAGAIN);
1519 if (rc > aa->aa_requested_nob) {
1520 CERROR("Unexpected rc %d (%d requested)\n", rc,
1521 aa->aa_requested_nob);
1525 if (rc != req->rq_bulk->bd_nob_transferred) {
1526 CERROR ("Unexpected rc %d (%d transferred)\n",
1527 rc, req->rq_bulk->bd_nob_transferred);
1531 if (rc < aa->aa_requested_nob)
1532 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1534 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1535 static int cksum_counter;
1536 __u32 server_cksum = body->oa.o_cksum;
1539 cksum_type_t cksum_type;
1541 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1542 body->oa.o_flags : 0);
1543 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1544 aa->aa_ppga, OST_READ,
1547 if (peer->nid == req->rq_bulk->bd_sender) {
1551 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1554 if (server_cksum == ~0 && rc > 0) {
1555 CERROR("Protocol error: server %s set the 'checksum' "
1556 "bit, but didn't send a checksum. Not fatal, "
1557 "but please notify on http://bugs.whamcloud.com/\n",
1558 libcfs_nid2str(peer->nid));
1559 } else if (server_cksum != client_cksum) {
1560 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1561 "%s%s%s inode "DFID" object "
1562 LPU64"/"LPU64" extent "
1563 "["LPU64"-"LPU64"]\n",
1564 req->rq_import->imp_obd->obd_name,
1565 libcfs_nid2str(peer->nid),
1567 body->oa.o_valid & OBD_MD_FLFID ?
1568 body->oa.o_parent_seq : (__u64)0,
1569 body->oa.o_valid & OBD_MD_FLFID ?
1570 body->oa.o_parent_oid : 0,
1571 body->oa.o_valid & OBD_MD_FLFID ?
1572 body->oa.o_parent_ver : 0,
1574 body->oa.o_valid & OBD_MD_FLGROUP ?
1575 body->oa.o_seq : (__u64)0,
1576 aa->aa_ppga[0]->off,
1577 aa->aa_ppga[aa->aa_page_count-1]->off +
1578 aa->aa_ppga[aa->aa_page_count-1]->count -
1580 CERROR("client %x, server %x, cksum_type %x\n",
1581 client_cksum, server_cksum, cksum_type);
1583 aa->aa_oa->o_cksum = client_cksum;
1587 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1590 } else if (unlikely(client_cksum)) {
1591 static int cksum_missed;
1594 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1595 CERROR("Checksum %u requested from %s but not sent\n",
1596 cksum_missed, libcfs_nid2str(peer->nid));
1602 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1607 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1608 struct lov_stripe_md *lsm,
1609 obd_count page_count, struct brw_page **pga,
1610 struct obd_capa *ocapa)
1612 struct ptlrpc_request *req;
1615 int generation, resends = 0;
1616 struct l_wait_info lwi;
1620 cfs_waitq_init(&waitq);
1621 generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1624 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1625 page_count, pga, &req, ocapa, 0, resends);
1630 req->rq_generation_set = 1;
1631 req->rq_import_generation = generation;
1632 req->rq_sent = cfs_time_current_sec() + resends;
1635 rc = ptlrpc_queue_wait(req);
1637 if (rc == -ETIMEDOUT && req->rq_resend) {
1638 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1639 ptlrpc_req_finished(req);
1643 rc = osc_brw_fini_request(req, rc);
1645 ptlrpc_req_finished(req);
1646 /* When server return -EINPROGRESS, client should always retry
1647 * regardless of the number of times the bulk was resent already.*/
1648 if (osc_recoverable_error(rc)) {
1650 if (rc != -EINPROGRESS &&
1651 !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1652 CERROR("%s: too many resend retries for object: "
1653 ""LPU64":"LPU64", rc = %d.\n",
1654 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1658 exp->exp_obd->u.cli.cl_import->imp_generation) {
1659 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1660 ""LPU64":"LPU64", rc = %d.\n",
1661 exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1665 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1667 l_wait_event(waitq, 0, &lwi);
1672 if (rc == -EAGAIN || rc == -EINPROGRESS)
1677 int osc_brw_redo_request(struct ptlrpc_request *request,
1678 struct osc_brw_async_args *aa)
1680 struct ptlrpc_request *new_req;
1681 struct osc_brw_async_args *new_aa;
1682 struct osc_async_page *oap;
1686 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1688 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1689 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1690 aa->aa_cli, aa->aa_oa,
1691 NULL /* lsm unused by osc currently */,
1692 aa->aa_page_count, aa->aa_ppga,
1693 &new_req, aa->aa_ocapa, 0, 1);
1697 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1698 if (oap->oap_request != NULL) {
1699 LASSERTF(request == oap->oap_request,
1700 "request %p != oap_request %p\n",
1701 request, oap->oap_request);
1702 if (oap->oap_interrupted) {
1703 ptlrpc_req_finished(new_req);
1708 /* New request takes over pga and oaps from old request.
1709 * Note that copying a list_head doesn't work, need to move it... */
1711 new_req->rq_interpret_reply = request->rq_interpret_reply;
1712 new_req->rq_async_args = request->rq_async_args;
1713 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1714 new_req->rq_generation_set = 1;
1715 new_req->rq_import_generation = request->rq_import_generation;
1717 new_aa = ptlrpc_req_async_args(new_req);
1719 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1720 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1721 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1722 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1724 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1725 if (oap->oap_request) {
1726 ptlrpc_req_finished(oap->oap_request);
1727 oap->oap_request = ptlrpc_request_addref(new_req);
1731 new_aa->aa_ocapa = aa->aa_ocapa;
1732 aa->aa_ocapa = NULL;
1734 /* XXX: This code will run into problem if we're going to support
1735 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1736 * and wait for all of them to be finished. We should inherit request
1737 * set from old request. */
1738 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1740 DEBUG_REQ(D_INFO, new_req, "new request");
1745 * ugh, we want disk allocation on the target to happen in offset order. we'll
1746 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1747 * fine for our small page arrays and doesn't require allocation. its an
1748 * insertion sort that swaps elements that are strides apart, shrinking the
1749 * stride down until its '1' and the array is sorted.
1751 static void sort_brw_pages(struct brw_page **array, int num)
1754 struct brw_page *tmp;
1758 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1763 for (i = stride ; i < num ; i++) {
1766 while (j >= stride && array[j - stride]->off > tmp->off) {
1767 array[j] = array[j - stride];
1772 } while (stride > 1);
1775 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1781 LASSERT (pages > 0);
1782 offset = pg[i]->off & ~CFS_PAGE_MASK;
1786 if (pages == 0) /* that's all */
1789 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1790 return count; /* doesn't end on page boundary */
1793 offset = pg[i]->off & ~CFS_PAGE_MASK;
1794 if (offset != 0) /* doesn't start on page boundary */
1801 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1803 struct brw_page **ppga;
1806 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1810 for (i = 0; i < count; i++)
1815 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1817 LASSERT(ppga != NULL);
1818 OBD_FREE(ppga, sizeof(*ppga) * count);
1821 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1822 obd_count page_count, struct brw_page *pga,
1823 struct obd_trans_info *oti)
1825 struct obdo *saved_oa = NULL;
1826 struct brw_page **ppga, **orig;
1827 struct obd_import *imp = class_exp2cliimp(exp);
1828 struct client_obd *cli;
1829 int rc, page_count_orig;
1832 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1833 cli = &imp->imp_obd->u.cli;
1835 if (cmd & OBD_BRW_CHECK) {
1836 /* The caller just wants to know if there's a chance that this
1837 * I/O can succeed */
1839 if (imp->imp_invalid)
1844 /* test_brw with a failed create can trip this, maybe others. */
1845 LASSERT(cli->cl_max_pages_per_rpc);
1849 orig = ppga = osc_build_ppga(pga, page_count);
1852 page_count_orig = page_count;
1854 sort_brw_pages(ppga, page_count);
1855 while (page_count) {
1856 obd_count pages_per_brw;
1858 if (page_count > cli->cl_max_pages_per_rpc)
1859 pages_per_brw = cli->cl_max_pages_per_rpc;
1861 pages_per_brw = page_count;
1863 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1865 if (saved_oa != NULL) {
1866 /* restore previously saved oa */
1867 *oinfo->oi_oa = *saved_oa;
1868 } else if (page_count > pages_per_brw) {
1869 /* save a copy of oa (brw will clobber it) */
1870 OBDO_ALLOC(saved_oa);
1871 if (saved_oa == NULL)
1872 GOTO(out, rc = -ENOMEM);
1873 *saved_oa = *oinfo->oi_oa;
1876 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1877 pages_per_brw, ppga, oinfo->oi_capa);
1882 page_count -= pages_per_brw;
1883 ppga += pages_per_brw;
1887 osc_release_ppga(orig, page_count_orig);
1889 if (saved_oa != NULL)
1890 OBDO_FREE(saved_oa);
1895 static int brw_interpret(const struct lu_env *env,
1896 struct ptlrpc_request *req, void *data, int rc)
1898 struct osc_brw_async_args *aa = data;
1899 struct osc_extent *ext;
1900 struct osc_extent *tmp;
1901 struct cl_object *obj = NULL;
1902 struct client_obd *cli = aa->aa_cli;
1905 rc = osc_brw_fini_request(req, rc);
1906 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1907 /* When server return -EINPROGRESS, client should always retry
1908 * regardless of the number of times the bulk was resent already. */
1909 if (osc_recoverable_error(rc)) {
1910 if (req->rq_import_generation !=
1911 req->rq_import->imp_generation) {
1912 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1913 ""LPU64":"LPU64", rc = %d.\n",
1914 req->rq_import->imp_obd->obd_name,
1915 aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
1916 } else if (rc == -EINPROGRESS ||
1917 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1918 rc = osc_brw_redo_request(req, aa);
1920 CERROR("%s: too many resent retries 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);
1928 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1933 capa_put(aa->aa_ocapa);
1934 aa->aa_ocapa = NULL;
1937 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1938 if (obj == NULL && rc == 0) {
1939 obj = osc2cl(ext->oe_obj);
1943 cfs_list_del_init(&ext->oe_link);
1944 osc_extent_finish(env, ext, 1, rc);
1946 LASSERT(cfs_list_empty(&aa->aa_exts));
1947 LASSERT(cfs_list_empty(&aa->aa_oaps));
1950 struct obdo *oa = aa->aa_oa;
1951 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1952 unsigned long valid = 0;
1955 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1956 attr->cat_blocks = oa->o_blocks;
1957 valid |= CAT_BLOCKS;
1959 if (oa->o_valid & OBD_MD_FLMTIME) {
1960 attr->cat_mtime = oa->o_mtime;
1963 if (oa->o_valid & OBD_MD_FLATIME) {
1964 attr->cat_atime = oa->o_atime;
1967 if (oa->o_valid & OBD_MD_FLCTIME) {
1968 attr->cat_ctime = oa->o_ctime;
1972 cl_object_attr_lock(obj);
1973 cl_object_attr_set(env, obj, attr, valid);
1974 cl_object_attr_unlock(obj);
1976 cl_object_put(env, obj);
1978 OBDO_FREE(aa->aa_oa);
1980 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1981 req->rq_bulk->bd_nob_transferred);
1982 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1983 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1985 client_obd_list_lock(&cli->cl_loi_list_lock);
1986 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1987 * is called so we know whether to go to sync BRWs or wait for more
1988 * RPCs to complete */
1989 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1990 cli->cl_w_in_flight--;
1992 cli->cl_r_in_flight--;
1993 osc_wake_cache_waiters(cli);
1994 client_obd_list_unlock(&cli->cl_loi_list_lock);
1996 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2001 * Build an RPC by the list of extent @ext_list. The caller must ensure
2002 * that the total pages in this list are NOT over max pages per RPC.
2003 * Extents in the list must be in OES_RPC state.
2005 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2006 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2008 struct ptlrpc_request *req = NULL;
2009 struct osc_extent *ext;
2010 CFS_LIST_HEAD(rpc_list);
2011 struct brw_page **pga = NULL;
2012 struct osc_brw_async_args *aa = NULL;
2013 struct obdo *oa = NULL;
2014 struct osc_async_page *oap;
2015 struct osc_async_page *tmp;
2016 struct cl_req *clerq = NULL;
2017 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2018 struct ldlm_lock *lock = NULL;
2019 struct cl_req_attr crattr;
2020 obd_off starting_offset = OBD_OBJECT_EOF;
2021 obd_off ending_offset = 0;
2022 int i, rc, mpflag = 0, mem_tight = 0, page_count = 0;
2025 LASSERT(!cfs_list_empty(ext_list));
2027 /* add pages into rpc_list to build BRW rpc */
2028 cfs_list_for_each_entry(ext, ext_list, oe_link) {
2029 LASSERT(ext->oe_state == OES_RPC);
2030 mem_tight |= ext->oe_memalloc;
2031 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2033 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2034 if (starting_offset > oap->oap_obj_off)
2035 starting_offset = oap->oap_obj_off;
2037 LASSERT(oap->oap_page_off == 0);
2038 if (ending_offset < oap->oap_obj_off + oap->oap_count)
2039 ending_offset = oap->oap_obj_off +
2042 LASSERT(oap->oap_page_off + oap->oap_count ==
2048 mpflag = cfs_memory_pressure_get_and_set();
2050 memset(&crattr, 0, sizeof crattr);
2051 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2053 GOTO(out, rc = -ENOMEM);
2057 GOTO(out, rc = -ENOMEM);
2060 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2061 struct cl_page *page = oap2cl_page(oap);
2062 if (clerq == NULL) {
2063 clerq = cl_req_alloc(env, page, crt,
2064 1 /* only 1-object rpcs for
2067 GOTO(out, rc = PTR_ERR(clerq));
2068 lock = oap->oap_ldlm_lock;
2071 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2072 pga[i] = &oap->oap_brw_page;
2073 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2074 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2075 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2077 cl_req_page_add(env, clerq, page);
2080 /* always get the data for the obdo for the rpc */
2081 LASSERT(clerq != NULL);
2083 crattr.cra_capa = NULL;
2084 memset(crattr.cra_jobid, 0, JOBSTATS_JOBID_SIZE);
2085 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2087 oa->o_handle = lock->l_remote_handle;
2088 oa->o_valid |= OBD_MD_FLHANDLE;
2091 rc = cl_req_prep(env, clerq);
2093 CERROR("cl_req_prep failed: %d\n", rc);
2097 sort_brw_pages(pga, page_count);
2098 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2099 pga, &req, crattr.cra_capa, 1, 0);
2101 CERROR("prep_req failed: %d\n", rc);
2105 req->rq_interpret_reply = brw_interpret;
2107 req->rq_memalloc = 1;
2109 /* Need to update the timestamps after the request is built in case
2110 * we race with setattr (locally or in queue at OST). If OST gets
2111 * later setattr before earlier BRW (as determined by the request xid),
2112 * the OST will not use BRW timestamps. Sadly, there is no obvious
2113 * way to do this in a single call. bug 10150 */
2114 cl_req_attr_set(env, clerq, &crattr,
2115 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2117 lustre_msg_set_jobid(req->rq_reqmsg, crattr.cra_jobid);
2119 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2120 aa = ptlrpc_req_async_args(req);
2121 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2122 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2123 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2124 cfs_list_splice_init(ext_list, &aa->aa_exts);
2125 aa->aa_clerq = clerq;
2127 /* queued sync pages can be torn down while the pages
2128 * were between the pending list and the rpc */
2130 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2131 /* only one oap gets a request reference */
2134 if (oap->oap_interrupted && !req->rq_intr) {
2135 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2137 ptlrpc_mark_interrupted(req);
2141 tmp->oap_request = ptlrpc_request_addref(req);
2143 client_obd_list_lock(&cli->cl_loi_list_lock);
2144 starting_offset >>= CFS_PAGE_SHIFT;
2145 if (cmd == OBD_BRW_READ) {
2146 cli->cl_r_in_flight++;
2147 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2148 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2149 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2150 starting_offset + 1);
2152 cli->cl_w_in_flight++;
2153 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2154 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2155 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2156 starting_offset + 1);
2158 client_obd_list_unlock(&cli->cl_loi_list_lock);
2160 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2161 page_count, aa, cli->cl_r_in_flight,
2162 cli->cl_w_in_flight);
2164 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2165 * see which CPU/NUMA node the majority of pages were allocated
2166 * on, and try to assign the async RPC to the CPU core
2167 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2169 * But on the other hand, we expect that multiple ptlrpcd
2170 * threads and the initial write sponsor can run in parallel,
2171 * especially when data checksum is enabled, which is CPU-bound
2172 * operation and single ptlrpcd thread cannot process in time.
2173 * So more ptlrpcd threads sharing BRW load
2174 * (with PDL_POLICY_ROUND) seems better.
2176 ptlrpcd_add_req(req, pol, -1);
2182 cfs_memory_pressure_restore(mpflag);
2184 capa_put(crattr.cra_capa);
2186 LASSERT(req == NULL);
2191 OBD_FREE(pga, sizeof(*pga) * page_count);
2192 /* this should happen rarely and is pretty bad, it makes the
2193 * pending list not follow the dirty order */
2194 while (!cfs_list_empty(ext_list)) {
2195 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2197 cfs_list_del_init(&ext->oe_link);
2198 osc_extent_finish(env, ext, 0, rc);
2200 if (clerq && !IS_ERR(clerq))
2201 cl_req_completion(env, clerq, rc);
2206 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2207 struct ldlm_enqueue_info *einfo)
2209 void *data = einfo->ei_cbdata;
2212 LASSERT(lock != NULL);
2213 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2214 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2215 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2216 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2218 lock_res_and_lock(lock);
2219 cfs_spin_lock(&osc_ast_guard);
2221 if (lock->l_ast_data == NULL)
2222 lock->l_ast_data = data;
2223 if (lock->l_ast_data == data)
2226 cfs_spin_unlock(&osc_ast_guard);
2227 unlock_res_and_lock(lock);
2232 static int osc_set_data_with_check(struct lustre_handle *lockh,
2233 struct ldlm_enqueue_info *einfo)
2235 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2239 set = osc_set_lock_data_with_check(lock, einfo);
2240 LDLM_LOCK_PUT(lock);
2242 CERROR("lockh %p, data %p - client evicted?\n",
2243 lockh, einfo->ei_cbdata);
2247 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2248 ldlm_iterator_t replace, void *data)
2250 struct ldlm_res_id res_id;
2251 struct obd_device *obd = class_exp2obd(exp);
2253 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
2254 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2258 /* find any ldlm lock of the inode in osc
2262 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2263 ldlm_iterator_t replace, void *data)
2265 struct ldlm_res_id res_id;
2266 struct obd_device *obd = class_exp2obd(exp);
2269 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
2270 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2271 if (rc == LDLM_ITER_STOP)
2273 if (rc == LDLM_ITER_CONTINUE)
2278 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2279 obd_enqueue_update_f upcall, void *cookie,
2280 int *flags, int agl, int rc)
2282 int intent = *flags & LDLM_FL_HAS_INTENT;
2286 /* The request was created before ldlm_cli_enqueue call. */
2287 if (rc == ELDLM_LOCK_ABORTED) {
2288 struct ldlm_reply *rep;
2289 rep = req_capsule_server_get(&req->rq_pill,
2292 LASSERT(rep != NULL);
2293 if (rep->lock_policy_res1)
2294 rc = rep->lock_policy_res1;
2298 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2300 *flags |= LDLM_FL_LVB_READY;
2301 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2302 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2305 /* Call the update callback. */
2306 rc = (*upcall)(cookie, rc);
2310 static int osc_enqueue_interpret(const struct lu_env *env,
2311 struct ptlrpc_request *req,
2312 struct osc_enqueue_args *aa, int rc)
2314 struct ldlm_lock *lock;
2315 struct lustre_handle handle;
2317 struct ost_lvb *lvb;
2319 int *flags = aa->oa_flags;
2321 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2322 * might be freed anytime after lock upcall has been called. */
2323 lustre_handle_copy(&handle, aa->oa_lockh);
2324 mode = aa->oa_ei->ei_mode;
2326 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2328 lock = ldlm_handle2lock(&handle);
2330 /* Take an additional reference so that a blocking AST that
2331 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2332 * to arrive after an upcall has been executed by
2333 * osc_enqueue_fini(). */
2334 ldlm_lock_addref(&handle, mode);
2336 /* Let CP AST to grant the lock first. */
2337 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2339 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2344 lvb_len = sizeof(*aa->oa_lvb);
2347 /* Complete obtaining the lock procedure. */
2348 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2349 mode, flags, lvb, lvb_len, &handle, rc);
2350 /* Complete osc stuff. */
2351 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2352 flags, aa->oa_agl, rc);
2354 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2356 /* Release the lock for async request. */
2357 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2359 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2360 * not already released by
2361 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2363 ldlm_lock_decref(&handle, mode);
2365 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2366 aa->oa_lockh, req, aa);
2367 ldlm_lock_decref(&handle, mode);
2368 LDLM_LOCK_PUT(lock);
2372 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2373 struct lov_oinfo *loi, int flags,
2374 struct ost_lvb *lvb, __u32 mode, int rc)
2376 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2378 if (rc == ELDLM_OK) {
2381 LASSERT(lock != NULL);
2382 loi->loi_lvb = *lvb;
2383 tmp = loi->loi_lvb.lvb_size;
2384 /* Extend KMS up to the end of this lock and no further
2385 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2386 if (tmp > lock->l_policy_data.l_extent.end)
2387 tmp = lock->l_policy_data.l_extent.end + 1;
2388 if (tmp >= loi->loi_kms) {
2389 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2390 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2391 loi_kms_set(loi, tmp);
2393 LDLM_DEBUG(lock, "lock acquired, setting rss="
2394 LPU64"; leaving kms="LPU64", end="LPU64,
2395 loi->loi_lvb.lvb_size, loi->loi_kms,
2396 lock->l_policy_data.l_extent.end);
2398 ldlm_lock_allow_match(lock);
2399 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2400 LASSERT(lock != NULL);
2401 loi->loi_lvb = *lvb;
2402 ldlm_lock_allow_match(lock);
2403 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2404 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2410 ldlm_lock_fail_match(lock);
2412 LDLM_LOCK_PUT(lock);
2415 EXPORT_SYMBOL(osc_update_enqueue);
2417 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2419 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2420 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2421 * other synchronous requests, however keeping some locks and trying to obtain
2422 * others may take a considerable amount of time in a case of ost failure; and
2423 * when other sync requests do not get released lock from a client, the client
2424 * is excluded from the cluster -- such scenarious make the life difficult, so
2425 * release locks just after they are obtained. */
2426 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2427 int *flags, ldlm_policy_data_t *policy,
2428 struct ost_lvb *lvb, int kms_valid,
2429 obd_enqueue_update_f upcall, void *cookie,
2430 struct ldlm_enqueue_info *einfo,
2431 struct lustre_handle *lockh,
2432 struct ptlrpc_request_set *rqset, int async, int agl)
2434 struct obd_device *obd = exp->exp_obd;
2435 struct ptlrpc_request *req = NULL;
2436 int intent = *flags & LDLM_FL_HAS_INTENT;
2437 int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2442 /* Filesystem lock extents are extended to page boundaries so that
2443 * dealing with the page cache is a little smoother. */
2444 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2445 policy->l_extent.end |= ~CFS_PAGE_MASK;
2448 * kms is not valid when either object is completely fresh (so that no
2449 * locks are cached), or object was evicted. In the latter case cached
2450 * lock cannot be used, because it would prime inode state with
2451 * potentially stale LVB.
2456 /* Next, search for already existing extent locks that will cover us */
2457 /* If we're trying to read, we also search for an existing PW lock. The
2458 * VFS and page cache already protect us locally, so lots of readers/
2459 * writers can share a single PW lock.
2461 * There are problems with conversion deadlocks, so instead of
2462 * converting a read lock to a write lock, we'll just enqueue a new
2465 * At some point we should cancel the read lock instead of making them
2466 * send us a blocking callback, but there are problems with canceling
2467 * locks out from other users right now, too. */
2468 mode = einfo->ei_mode;
2469 if (einfo->ei_mode == LCK_PR)
2471 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2472 einfo->ei_type, policy, mode, lockh, 0);
2474 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2476 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2477 /* For AGL, if enqueue RPC is sent but the lock is not
2478 * granted, then skip to process this strpe.
2479 * Return -ECANCELED to tell the caller. */
2480 ldlm_lock_decref(lockh, mode);
2481 LDLM_LOCK_PUT(matched);
2483 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2484 *flags |= LDLM_FL_LVB_READY;
2485 /* addref the lock only if not async requests and PW
2486 * lock is matched whereas we asked for PR. */
2487 if (!rqset && einfo->ei_mode != mode)
2488 ldlm_lock_addref(lockh, LCK_PR);
2490 /* I would like to be able to ASSERT here that
2491 * rss <= kms, but I can't, for reasons which
2492 * are explained in lov_enqueue() */
2495 /* We already have a lock, and it's referenced */
2496 (*upcall)(cookie, ELDLM_OK);
2498 if (einfo->ei_mode != mode)
2499 ldlm_lock_decref(lockh, LCK_PW);
2501 /* For async requests, decref the lock. */
2502 ldlm_lock_decref(lockh, einfo->ei_mode);
2503 LDLM_LOCK_PUT(matched);
2506 ldlm_lock_decref(lockh, mode);
2507 LDLM_LOCK_PUT(matched);
2513 CFS_LIST_HEAD(cancels);
2514 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2515 &RQF_LDLM_ENQUEUE_LVB);
2519 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2521 ptlrpc_request_free(req);
2525 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2527 ptlrpc_request_set_replen(req);
2530 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2531 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2533 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2534 sizeof(*lvb), lockh, async);
2537 struct osc_enqueue_args *aa;
2538 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2539 aa = ptlrpc_req_async_args(req);
2542 aa->oa_flags = flags;
2543 aa->oa_upcall = upcall;
2544 aa->oa_cookie = cookie;
2546 aa->oa_lockh = lockh;
2549 req->rq_interpret_reply =
2550 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2551 if (rqset == PTLRPCD_SET)
2552 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2554 ptlrpc_set_add_req(rqset, req);
2555 } else if (intent) {
2556 ptlrpc_req_finished(req);
2561 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2563 ptlrpc_req_finished(req);
2568 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2569 struct ldlm_enqueue_info *einfo,
2570 struct ptlrpc_request_set *rqset)
2572 struct ldlm_res_id res_id;
2576 osc_build_res_name(oinfo->oi_md->lsm_object_id,
2577 oinfo->oi_md->lsm_object_seq, &res_id);
2579 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2580 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2581 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2582 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2583 rqset, rqset != NULL, 0);
2587 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2588 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2589 int *flags, void *data, struct lustre_handle *lockh,
2592 struct obd_device *obd = exp->exp_obd;
2593 int lflags = *flags;
2597 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2600 /* Filesystem lock extents are extended to page boundaries so that
2601 * dealing with the page cache is a little smoother */
2602 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2603 policy->l_extent.end |= ~CFS_PAGE_MASK;
2605 /* Next, search for already existing extent locks that will cover us */
2606 /* If we're trying to read, we also search for an existing PW lock. The
2607 * VFS and page cache already protect us locally, so lots of readers/
2608 * writers can share a single PW lock. */
2612 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2613 res_id, type, policy, rc, lockh, unref);
2616 if (!osc_set_data_with_check(lockh, data)) {
2617 if (!(lflags & LDLM_FL_TEST_LOCK))
2618 ldlm_lock_decref(lockh, rc);
2622 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2623 ldlm_lock_addref(lockh, LCK_PR);
2624 ldlm_lock_decref(lockh, LCK_PW);
2631 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2635 if (unlikely(mode == LCK_GROUP))
2636 ldlm_lock_decref_and_cancel(lockh, mode);
2638 ldlm_lock_decref(lockh, mode);
2643 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2644 __u32 mode, struct lustre_handle *lockh)
2647 RETURN(osc_cancel_base(lockh, mode));
2650 static int osc_cancel_unused(struct obd_export *exp,
2651 struct lov_stripe_md *lsm,
2652 ldlm_cancel_flags_t flags,
2655 struct obd_device *obd = class_exp2obd(exp);
2656 struct ldlm_res_id res_id, *resp = NULL;
2659 resp = osc_build_res_name(lsm->lsm_object_id,
2660 lsm->lsm_object_seq, &res_id);
2663 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2666 static int osc_statfs_interpret(const struct lu_env *env,
2667 struct ptlrpc_request *req,
2668 struct osc_async_args *aa, int rc)
2670 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
2671 struct obd_statfs *msfs;
2676 /* The request has in fact never been sent
2677 * due to issues at a higher level (LOV).
2678 * Exit immediately since the caller is
2679 * aware of the problem and takes care
2680 * of the clean up */
2683 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2684 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2690 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2692 GOTO(out, rc = -EPROTO);
2695 /* Reinitialize the RDONLY and DEGRADED flags at the client
2696 * on each statfs, so they don't stay set permanently. */
2697 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
2699 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
2700 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
2701 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
2702 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
2704 if (unlikely(msfs->os_state & OS_STATE_READONLY))
2705 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
2706 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
2707 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
2709 /* Add a bit of hysteresis so this flag isn't continually flapping,
2710 * and ensure that new files don't get extremely fragmented due to
2711 * only a small amount of available space in the filesystem.
2712 * We want to set the NOSPC flag when there is less than ~0.1% free
2713 * and clear it when there is at least ~0.2% free space, so:
2714 * avail < ~0.1% max max = avail + used
2715 * 1025 * avail < avail + used used = blocks - free
2716 * 1024 * avail < used
2717 * 1024 * avail < blocks - free
2718 * avail < ((blocks - free) >> 10)
2720 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
2721 * lose that amount of space so in those cases we report no space left
2722 * if their is less than 1 GB left. */
2723 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
2724 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
2725 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
2726 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
2727 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
2728 (msfs->os_ffree > 64) &&
2729 (msfs->os_bavail > (used << 1)))) {
2730 cli->cl_oscc.oscc_flags &= ~(OSCC_FLAG_NOSPC |
2731 OSCC_FLAG_NOSPC_BLK);
2734 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
2735 (msfs->os_bavail < used)))
2736 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC_BLK;
2738 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
2740 *aa->aa_oi->oi_osfs = *msfs;
2742 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2746 static int osc_statfs_async(struct obd_export *exp,
2747 struct obd_info *oinfo, __u64 max_age,
2748 struct ptlrpc_request_set *rqset)
2750 struct obd_device *obd = class_exp2obd(exp);
2751 struct ptlrpc_request *req;
2752 struct osc_async_args *aa;
2756 /* We could possibly pass max_age in the request (as an absolute
2757 * timestamp or a "seconds.usec ago") so the target can avoid doing
2758 * extra calls into the filesystem if that isn't necessary (e.g.
2759 * during mount that would help a bit). Having relative timestamps
2760 * is not so great if request processing is slow, while absolute
2761 * timestamps are not ideal because they need time synchronization. */
2762 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2766 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2768 ptlrpc_request_free(req);
2771 ptlrpc_request_set_replen(req);
2772 req->rq_request_portal = OST_CREATE_PORTAL;
2773 ptlrpc_at_set_req_timeout(req);
2775 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2776 /* procfs requests not want stat in wait for avoid deadlock */
2777 req->rq_no_resend = 1;
2778 req->rq_no_delay = 1;
2781 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2782 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2783 aa = ptlrpc_req_async_args(req);
2786 ptlrpc_set_add_req(rqset, req);
2790 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2791 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2793 struct obd_device *obd = class_exp2obd(exp);
2794 struct obd_statfs *msfs;
2795 struct ptlrpc_request *req;
2796 struct obd_import *imp = NULL;
2800 /*Since the request might also come from lprocfs, so we need
2801 *sync this with client_disconnect_export Bug15684*/
2802 cfs_down_read(&obd->u.cli.cl_sem);
2803 if (obd->u.cli.cl_import)
2804 imp = class_import_get(obd->u.cli.cl_import);
2805 cfs_up_read(&obd->u.cli.cl_sem);
2809 /* We could possibly pass max_age in the request (as an absolute
2810 * timestamp or a "seconds.usec ago") so the target can avoid doing
2811 * extra calls into the filesystem if that isn't necessary (e.g.
2812 * during mount that would help a bit). Having relative timestamps
2813 * is not so great if request processing is slow, while absolute
2814 * timestamps are not ideal because they need time synchronization. */
2815 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2817 class_import_put(imp);
2822 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2824 ptlrpc_request_free(req);
2827 ptlrpc_request_set_replen(req);
2828 req->rq_request_portal = OST_CREATE_PORTAL;
2829 ptlrpc_at_set_req_timeout(req);
2831 if (flags & OBD_STATFS_NODELAY) {
2832 /* procfs requests not want stat in wait for avoid deadlock */
2833 req->rq_no_resend = 1;
2834 req->rq_no_delay = 1;
2837 rc = ptlrpc_queue_wait(req);
2841 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2843 GOTO(out, rc = -EPROTO);
2850 ptlrpc_req_finished(req);
2854 /* Retrieve object striping information.
2856 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2857 * the maximum number of OST indices which will fit in the user buffer.
2858 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2860 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2862 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2863 struct lov_user_md_v3 lum, *lumk;
2864 struct lov_user_ost_data_v1 *lmm_objects;
2865 int rc = 0, lum_size;
2871 /* we only need the header part from user space to get lmm_magic and
2872 * lmm_stripe_count, (the header part is common to v1 and v3) */
2873 lum_size = sizeof(struct lov_user_md_v1);
2874 if (cfs_copy_from_user(&lum, lump, lum_size))
2877 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2878 (lum.lmm_magic != LOV_USER_MAGIC_V3))
2881 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2882 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2883 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2884 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2886 /* we can use lov_mds_md_size() to compute lum_size
2887 * because lov_user_md_vX and lov_mds_md_vX have the same size */
2888 if (lum.lmm_stripe_count > 0) {
2889 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2890 OBD_ALLOC(lumk, lum_size);
2894 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2895 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2897 lmm_objects = &(lumk->lmm_objects[0]);
2898 lmm_objects->l_object_id = lsm->lsm_object_id;
2900 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2904 lumk->lmm_object_id = lsm->lsm_object_id;
2905 lumk->lmm_object_seq = lsm->lsm_object_seq;
2906 lumk->lmm_stripe_count = 1;
2908 if (cfs_copy_to_user(lump, lumk, lum_size))
2912 OBD_FREE(lumk, lum_size);
2918 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2919 void *karg, void *uarg)
2921 struct obd_device *obd = exp->exp_obd;
2922 struct obd_ioctl_data *data = karg;
2926 if (!cfs_try_module_get(THIS_MODULE)) {
2927 CERROR("Can't get module. Is it alive?");
2931 case OBD_IOC_LOV_GET_CONFIG: {
2933 struct lov_desc *desc;
2934 struct obd_uuid uuid;
2938 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2939 GOTO(out, err = -EINVAL);
2941 data = (struct obd_ioctl_data *)buf;
2943 if (sizeof(*desc) > data->ioc_inllen1) {
2944 obd_ioctl_freedata(buf, len);
2945 GOTO(out, err = -EINVAL);
2948 if (data->ioc_inllen2 < sizeof(uuid)) {
2949 obd_ioctl_freedata(buf, len);
2950 GOTO(out, err = -EINVAL);
2953 desc = (struct lov_desc *)data->ioc_inlbuf1;
2954 desc->ld_tgt_count = 1;
2955 desc->ld_active_tgt_count = 1;
2956 desc->ld_default_stripe_count = 1;
2957 desc->ld_default_stripe_size = 0;
2958 desc->ld_default_stripe_offset = 0;
2959 desc->ld_pattern = 0;
2960 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2962 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2964 err = cfs_copy_to_user((void *)uarg, buf, len);
2967 obd_ioctl_freedata(buf, len);
2970 case LL_IOC_LOV_SETSTRIPE:
2971 err = obd_alloc_memmd(exp, karg);
2975 case LL_IOC_LOV_GETSTRIPE:
2976 err = osc_getstripe(karg, uarg);
2978 case OBD_IOC_CLIENT_RECOVER:
2979 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2980 data->ioc_inlbuf1, 0);
2984 case IOC_OSC_SET_ACTIVE:
2985 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2988 case OBD_IOC_POLL_QUOTACHECK:
2989 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2991 case OBD_IOC_PING_TARGET:
2992 err = ptlrpc_obd_ping(obd);
2995 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2996 cmd, cfs_curproc_comm());
2997 GOTO(out, err = -ENOTTY);
3000 cfs_module_put(THIS_MODULE);
3004 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3005 obd_count keylen, void *key, __u32 *vallen, void *val,
3006 struct lov_stripe_md *lsm)
3009 if (!vallen || !val)
3012 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3013 __u32 *stripe = val;
3014 *vallen = sizeof(*stripe);
3017 } else if (KEY_IS(KEY_LAST_ID)) {
3018 struct ptlrpc_request *req;
3023 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3024 &RQF_OST_GET_INFO_LAST_ID);
3028 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3029 RCL_CLIENT, keylen);
3030 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3032 ptlrpc_request_free(req);
3036 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3037 memcpy(tmp, key, keylen);
3039 req->rq_no_delay = req->rq_no_resend = 1;
3040 ptlrpc_request_set_replen(req);
3041 rc = ptlrpc_queue_wait(req);
3045 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3047 GOTO(out, rc = -EPROTO);
3049 *((obd_id *)val) = *reply;
3051 ptlrpc_req_finished(req);
3053 } else if (KEY_IS(KEY_FIEMAP)) {
3054 struct ptlrpc_request *req;
3055 struct ll_user_fiemap *reply;
3059 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3060 &RQF_OST_GET_INFO_FIEMAP);
3064 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3065 RCL_CLIENT, keylen);
3066 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3067 RCL_CLIENT, *vallen);
3068 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3069 RCL_SERVER, *vallen);
3071 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3073 ptlrpc_request_free(req);
3077 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3078 memcpy(tmp, key, keylen);
3079 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3080 memcpy(tmp, val, *vallen);
3082 ptlrpc_request_set_replen(req);
3083 rc = ptlrpc_queue_wait(req);
3087 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3089 GOTO(out1, rc = -EPROTO);
3091 memcpy(val, reply, *vallen);
3093 ptlrpc_req_finished(req);
3101 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3103 struct llog_ctxt *ctxt;
3107 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3109 rc = llog_initiator_connect(ctxt);
3110 llog_ctxt_put(ctxt);
3112 /* XXX return an error? skip setting below flags? */
3115 cfs_spin_lock(&imp->imp_lock);
3116 imp->imp_server_timeout = 1;
3117 imp->imp_pingable = 1;
3118 cfs_spin_unlock(&imp->imp_lock);
3119 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3124 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3125 struct ptlrpc_request *req,
3132 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3135 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3136 obd_count keylen, void *key, obd_count vallen,
3137 void *val, struct ptlrpc_request_set *set)
3139 struct ptlrpc_request *req;
3140 struct obd_device *obd = exp->exp_obd;
3141 struct obd_import *imp = class_exp2cliimp(exp);
3146 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3148 if (KEY_IS(KEY_NEXT_ID)) {
3150 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3152 if (vallen != sizeof(obd_id))
3157 if (vallen != sizeof(obd_id))
3160 /* avoid race between allocate new object and set next id
3161 * from ll_sync thread */
3162 cfs_spin_lock(&oscc->oscc_lock);
3163 new_val = *((obd_id*)val) + 1;
3164 if (new_val > oscc->oscc_next_id)
3165 oscc->oscc_next_id = new_val;
3166 cfs_spin_unlock(&oscc->oscc_lock);
3167 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3168 exp->exp_obd->obd_name,
3169 obd->u.cli.cl_oscc.oscc_next_id);
3174 if (KEY_IS(KEY_CHECKSUM)) {
3175 if (vallen != sizeof(int))
3177 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3181 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3182 sptlrpc_conf_client_adapt(obd);
3186 if (KEY_IS(KEY_FLUSH_CTX)) {
3187 sptlrpc_import_flush_my_ctx(imp);
3191 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3194 /* We pass all other commands directly to OST. Since nobody calls osc
3195 methods directly and everybody is supposed to go through LOV, we
3196 assume lov checked invalid values for us.
3197 The only recognised values so far are evict_by_nid and mds_conn.
3198 Even if something bad goes through, we'd get a -EINVAL from OST
3201 if (KEY_IS(KEY_GRANT_SHRINK))
3202 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
3204 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
3209 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3210 RCL_CLIENT, keylen);
3211 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3212 RCL_CLIENT, vallen);
3213 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3215 ptlrpc_request_free(req);
3219 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3220 memcpy(tmp, key, keylen);
3221 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3222 memcpy(tmp, val, vallen);
3224 if (KEY_IS(KEY_MDS_CONN)) {
3225 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3227 oscc->oscc_oa.o_seq = (*(__u32 *)val);
3228 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3229 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
3230 req->rq_no_delay = req->rq_no_resend = 1;
3231 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3232 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
3233 struct osc_grant_args *aa;
3236 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3237 aa = ptlrpc_req_async_args(req);
3240 ptlrpc_req_finished(req);
3243 *oa = ((struct ost_body *)val)->oa;
3245 req->rq_interpret_reply = osc_shrink_grant_interpret;
3248 ptlrpc_request_set_replen(req);
3249 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3250 LASSERT(set != NULL);
3251 ptlrpc_set_add_req(set, req);
3252 ptlrpc_check_set(NULL, set);
3254 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3260 static struct llog_operations osc_size_repl_logops = {
3261 lop_cancel: llog_obd_repl_cancel
3264 static struct llog_operations osc_mds_ost_orig_logops;
3266 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3267 struct obd_device *tgt, struct llog_catid *catid)
3272 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
3273 &catid->lci_logid, &osc_mds_ost_orig_logops);
3275 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3279 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
3280 NULL, &osc_size_repl_logops);
3282 struct llog_ctxt *ctxt =
3283 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3286 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3291 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
3292 obd->obd_name, tgt->obd_name, catid, rc);
3293 CERROR("logid "LPX64":0x%x\n",
3294 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3299 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3300 struct obd_device *disk_obd, int *index)
3302 struct llog_catid catid;
3303 static char name[32] = CATLIST;
3307 LASSERT(olg == &obd->obd_olg);
3309 cfs_mutex_lock(&olg->olg_cat_processing);
3310 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
3312 CERROR("rc: %d\n", rc);
3316 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
3317 obd->obd_name, *index, catid.lci_logid.lgl_oid,
3318 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
3320 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
3322 CERROR("rc: %d\n", rc);
3326 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
3328 CERROR("rc: %d\n", rc);
3333 cfs_mutex_unlock(&olg->olg_cat_processing);
3338 static int osc_llog_finish(struct obd_device *obd, int count)
3340 struct llog_ctxt *ctxt;
3341 int rc = 0, rc2 = 0;
3344 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3346 rc = llog_cleanup(ctxt);
3348 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3350 rc2 = llog_cleanup(ctxt);
3357 static int osc_reconnect(const struct lu_env *env,
3358 struct obd_export *exp, struct obd_device *obd,
3359 struct obd_uuid *cluuid,
3360 struct obd_connect_data *data,
3363 struct client_obd *cli = &obd->u.cli;
3365 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3368 client_obd_list_lock(&cli->cl_loi_list_lock);
3369 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3370 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3371 lost_grant = cli->cl_lost_grant;
3372 cli->cl_lost_grant = 0;
3373 client_obd_list_unlock(&cli->cl_loi_list_lock);
3375 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3376 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3377 data->ocd_version, data->ocd_grant, lost_grant);
3383 static int osc_disconnect(struct obd_export *exp)
3385 struct obd_device *obd = class_exp2obd(exp);
3386 struct llog_ctxt *ctxt;
3389 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3391 if (obd->u.cli.cl_conn_count == 1) {
3392 /* Flush any remaining cancel messages out to the
3394 llog_sync(ctxt, exp, 0);
3396 llog_ctxt_put(ctxt);
3398 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3402 rc = client_disconnect_export(exp);
3404 * Initially we put del_shrink_grant before disconnect_export, but it
3405 * causes the following problem if setup (connect) and cleanup
3406 * (disconnect) are tangled together.
3407 * connect p1 disconnect p2
3408 * ptlrpc_connect_import
3409 * ............... class_manual_cleanup
3412 * ptlrpc_connect_interrupt
3414 * add this client to shrink list
3416 * Bang! pinger trigger the shrink.
3417 * So the osc should be disconnected from the shrink list, after we
3418 * are sure the import has been destroyed. BUG18662
3420 if (obd->u.cli.cl_import == NULL)
3421 osc_del_shrink_grant(&obd->u.cli);
3425 static int osc_import_event(struct obd_device *obd,
3426 struct obd_import *imp,
3427 enum obd_import_event event)
3429 struct client_obd *cli;
3433 LASSERT(imp->imp_obd == obd);
3436 case IMP_EVENT_DISCON: {
3437 /* Only do this on the MDS OSC's */
3438 if (imp->imp_server_timeout) {
3439 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3441 cfs_spin_lock(&oscc->oscc_lock);
3442 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3443 cfs_spin_unlock(&oscc->oscc_lock);
3446 client_obd_list_lock(&cli->cl_loi_list_lock);
3447 cli->cl_avail_grant = 0;
3448 cli->cl_lost_grant = 0;
3449 client_obd_list_unlock(&cli->cl_loi_list_lock);
3452 case IMP_EVENT_INACTIVE: {
3453 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3456 case IMP_EVENT_INVALIDATE: {
3457 struct ldlm_namespace *ns = obd->obd_namespace;
3461 env = cl_env_get(&refcheck);
3465 /* all pages go to failing rpcs due to the invalid
3467 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3469 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3470 cl_env_put(env, &refcheck);
3475 case IMP_EVENT_ACTIVE: {
3476 /* Only do this on the MDS OSC's */
3477 if (imp->imp_server_timeout) {
3478 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3480 cfs_spin_lock(&oscc->oscc_lock);
3481 oscc->oscc_flags &= ~(OSCC_FLAG_NOSPC |
3482 OSCC_FLAG_NOSPC_BLK);
3483 cfs_spin_unlock(&oscc->oscc_lock);
3485 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3488 case IMP_EVENT_OCD: {
3489 struct obd_connect_data *ocd = &imp->imp_connect_data;
3491 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3492 osc_init_grant(&obd->u.cli, ocd);
3495 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3496 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3498 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3501 case IMP_EVENT_DEACTIVATE: {
3502 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3505 case IMP_EVENT_ACTIVATE: {
3506 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3510 CERROR("Unknown import event %d\n", event);
3517 * Determine whether the lock can be canceled before replaying the lock
3518 * during recovery, see bug16774 for detailed information.
3520 * \retval zero the lock can't be canceled
3521 * \retval other ok to cancel
3523 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3525 check_res_locked(lock->l_resource);
3528 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3530 * XXX as a future improvement, we can also cancel unused write lock
3531 * if it doesn't have dirty data and active mmaps.
3533 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3534 (lock->l_granted_mode == LCK_PR ||
3535 lock->l_granted_mode == LCK_CR) &&
3536 (osc_dlm_lock_pageref(lock) == 0))
3542 static int brw_queue_work(const struct lu_env *env, void *data)
3544 struct client_obd *cli = data;
3546 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3548 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3552 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3554 struct client_obd *cli = &obd->u.cli;
3559 rc = ptlrpcd_addref();
3563 rc = client_obd_setup(obd, lcfg);
3566 handler = ptlrpcd_alloc_work(cli->cl_import,
3567 brw_queue_work, cli);
3568 if (!IS_ERR(handler))
3569 cli->cl_writeback_work = handler;
3571 rc = PTR_ERR(handler);
3575 struct lprocfs_static_vars lvars = { 0 };
3577 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3578 lprocfs_osc_init_vars(&lvars);
3579 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3580 lproc_osc_attach_seqstat(obd);
3581 sptlrpc_lprocfs_cliobd_attach(obd);
3582 ptlrpc_lprocfs_register_obd(obd);
3586 /* We need to allocate a few requests more, because
3587 brw_interpret tries to create new requests before freeing
3588 previous ones. Ideally we want to have 2x max_rpcs_in_flight
3589 reserved, but I afraid that might be too much wasted RAM
3590 in fact, so 2 is just my guess and still should work. */
3591 cli->cl_import->imp_rq_pool =
3592 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3594 ptlrpc_add_rqs_to_pool);
3596 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3598 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3606 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3612 case OBD_CLEANUP_EARLY: {
3613 struct obd_import *imp;
3614 imp = obd->u.cli.cl_import;
3615 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3616 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3617 ptlrpc_deactivate_import(imp);
3618 cfs_spin_lock(&imp->imp_lock);
3619 imp->imp_pingable = 0;
3620 cfs_spin_unlock(&imp->imp_lock);
3623 case OBD_CLEANUP_EXPORTS: {
3624 struct client_obd *cli = &obd->u.cli;
3626 * for echo client, export may be on zombie list, wait for
3627 * zombie thread to cull it, because cli.cl_import will be
3628 * cleared in client_disconnect_export():
3629 * class_export_destroy() -> obd_cleanup() ->
3630 * echo_device_free() -> echo_client_cleanup() ->
3631 * obd_disconnect() -> osc_disconnect() ->
3632 * client_disconnect_export()
3634 obd_zombie_barrier();
3635 if (cli->cl_writeback_work) {
3636 ptlrpcd_destroy_work(cli->cl_writeback_work);
3637 cli->cl_writeback_work = NULL;
3639 obd_cleanup_client_import(obd);
3640 ptlrpc_lprocfs_unregister_obd(obd);
3641 lprocfs_obd_cleanup(obd);
3642 rc = obd_llog_finish(obd, 0);
3644 CERROR("failed to cleanup llogging subsystems\n");
3651 int osc_cleanup(struct obd_device *obd)
3657 /* free memory of osc quota cache */
3658 osc_quota_cleanup(obd);
3660 rc = client_obd_cleanup(obd);
3666 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3668 struct lprocfs_static_vars lvars = { 0 };
3671 lprocfs_osc_init_vars(&lvars);
3673 switch (lcfg->lcfg_command) {
3675 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3685 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3687 return osc_process_config_base(obd, buf);
3690 struct obd_ops osc_obd_ops = {
3691 .o_owner = THIS_MODULE,
3692 .o_setup = osc_setup,
3693 .o_precleanup = osc_precleanup,
3694 .o_cleanup = osc_cleanup,
3695 .o_add_conn = client_import_add_conn,
3696 .o_del_conn = client_import_del_conn,
3697 .o_connect = client_connect_import,
3698 .o_reconnect = osc_reconnect,
3699 .o_disconnect = osc_disconnect,
3700 .o_statfs = osc_statfs,
3701 .o_statfs_async = osc_statfs_async,
3702 .o_packmd = osc_packmd,
3703 .o_unpackmd = osc_unpackmd,
3704 .o_precreate = osc_precreate,
3705 .o_create = osc_create,
3706 .o_create_async = osc_create_async,
3707 .o_destroy = osc_destroy,
3708 .o_getattr = osc_getattr,
3709 .o_getattr_async = osc_getattr_async,
3710 .o_setattr = osc_setattr,
3711 .o_setattr_async = osc_setattr_async,
3713 .o_punch = osc_punch,
3715 .o_enqueue = osc_enqueue,
3716 .o_change_cbdata = osc_change_cbdata,
3717 .o_find_cbdata = osc_find_cbdata,
3718 .o_cancel = osc_cancel,
3719 .o_cancel_unused = osc_cancel_unused,
3720 .o_iocontrol = osc_iocontrol,
3721 .o_get_info = osc_get_info,
3722 .o_set_info_async = osc_set_info_async,
3723 .o_import_event = osc_import_event,
3724 .o_llog_init = osc_llog_init,
3725 .o_llog_finish = osc_llog_finish,
3726 .o_process_config = osc_process_config,
3727 .o_quotactl = osc_quotactl,
3728 .o_quotacheck = osc_quotacheck,
3729 .o_quota_adjust_qunit = osc_quota_adjust_qunit,
3732 extern struct lu_kmem_descr osc_caches[];
3733 extern cfs_spinlock_t osc_ast_guard;
3734 extern cfs_lock_class_key_t osc_ast_guard_class;
3736 int __init osc_init(void)
3738 struct lprocfs_static_vars lvars = { 0 };
3742 /* print an address of _any_ initialized kernel symbol from this
3743 * module, to allow debugging with gdb that doesn't support data
3744 * symbols from modules.*/
3745 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3747 rc = lu_kmem_init(osc_caches);
3749 lprocfs_osc_init_vars(&lvars);
3752 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3753 LUSTRE_OSC_NAME, &osc_device_type);
3755 lu_kmem_fini(osc_caches);
3759 cfs_spin_lock_init(&osc_ast_guard);
3760 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3762 osc_mds_ost_orig_logops = llog_lvfs_ops;
3763 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3764 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3765 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3766 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3772 static void /*__exit*/ osc_exit(void)
3775 class_unregister_type(LUSTRE_OSC_NAME);
3776 lu_kmem_fini(osc_caches);
3779 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3780 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3781 MODULE_LICENSE("GPL");
3783 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);