4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include <libcfs/libcfs.h>
42 # include <liblustre.h>
45 #include <lustre_dlm.h>
46 #include <lustre_net.h>
47 #include <lustre/lustre_user.h>
48 #include <obd_cksum.h>
54 #include <lustre_ha.h>
55 #include <lprocfs_status.h>
56 #include <lustre_ioctl.h>
57 #include <lustre_log.h>
58 #include <lustre_debug.h>
59 #include <lustre_param.h>
60 #include <lustre_fid.h>
61 #include "osc_internal.h"
62 #include "osc_cl_internal.h"
64 struct osc_brw_async_args {
68 obd_count aa_page_count;
70 struct brw_page **aa_ppga;
71 struct client_obd *aa_cli;
72 struct list_head aa_oaps;
73 struct list_head aa_exts;
74 struct obd_capa *aa_ocapa;
75 struct cl_req *aa_clerq;
78 #define osc_grant_args osc_brw_async_args
80 struct osc_async_args {
81 struct obd_info *aa_oi;
84 struct osc_setattr_args {
86 obd_enqueue_update_f sa_upcall;
90 struct osc_fsync_args {
91 struct obd_info *fa_oi;
92 obd_enqueue_update_f fa_upcall;
96 struct osc_enqueue_args {
97 struct obd_export *oa_exp;
99 obd_enqueue_update_f oa_upcall;
101 struct ost_lvb *oa_lvb;
102 struct lustre_handle *oa_lockh;
103 struct ldlm_enqueue_info *oa_ei;
104 unsigned int oa_agl:1;
107 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
108 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
111 /* Unpack OSC object metadata from disk storage (LE byte order). */
112 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
113 struct lov_mds_md *lmm, int lmm_bytes)
116 struct obd_import *imp = class_exp2cliimp(exp);
120 if (lmm_bytes < sizeof(*lmm)) {
121 CERROR("%s: lov_mds_md too small: %d, need %d\n",
122 exp->exp_obd->obd_name, lmm_bytes,
126 /* XXX LOV_MAGIC etc check? */
128 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
129 CERROR("%s: zero lmm_object_id: rc = %d\n",
130 exp->exp_obd->obd_name, -EINVAL);
135 lsm_size = lov_stripe_md_size(1);
139 if (*lsmp != NULL && lmm == NULL) {
140 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
141 OBD_FREE(*lsmp, lsm_size);
147 OBD_ALLOC(*lsmp, lsm_size);
148 if (unlikely(*lsmp == NULL))
150 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
151 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
152 OBD_FREE(*lsmp, lsm_size);
155 loi_init((*lsmp)->lsm_oinfo[0]);
156 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
161 /* XXX zero *lsmp? */
162 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
165 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
166 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
168 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
173 static inline void osc_pack_capa(struct ptlrpc_request *req,
174 struct ost_body *body, void *capa)
176 struct obd_capa *oc = (struct obd_capa *)capa;
177 struct lustre_capa *c;
182 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
185 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
186 DEBUG_CAPA(D_SEC, c, "pack");
189 static inline void osc_pack_req_body(struct ptlrpc_request *req,
190 struct obd_info *oinfo)
192 struct ost_body *body;
194 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
197 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
199 osc_pack_capa(req, body, oinfo->oi_capa);
202 static inline void osc_set_capa_size(struct ptlrpc_request *req,
203 const struct req_msg_field *field,
207 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
209 /* it is already calculated as sizeof struct obd_capa */
213 static int osc_getattr_interpret(const struct lu_env *env,
214 struct ptlrpc_request *req,
215 struct osc_async_args *aa, int rc)
217 struct ost_body *body;
223 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
225 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
226 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
227 aa->aa_oi->oi_oa, &body->oa);
229 /* This should really be sent by the OST */
230 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
231 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
233 CDEBUG(D_INFO, "can't unpack ost_body\n");
235 aa->aa_oi->oi_oa->o_valid = 0;
238 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
242 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
243 struct ptlrpc_request_set *set)
245 struct ptlrpc_request *req;
246 struct osc_async_args *aa;
250 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
254 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
255 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
257 ptlrpc_request_free(req);
261 osc_pack_req_body(req, oinfo);
263 ptlrpc_request_set_replen(req);
264 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
266 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
267 aa = ptlrpc_req_async_args(req);
270 ptlrpc_set_add_req(set, req);
274 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
275 struct obd_info *oinfo)
277 struct ptlrpc_request *req;
278 struct ost_body *body;
282 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
286 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
287 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
289 ptlrpc_request_free(req);
293 osc_pack_req_body(req, oinfo);
295 ptlrpc_request_set_replen(req);
297 rc = ptlrpc_queue_wait(req);
301 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
303 GOTO(out, rc = -EPROTO);
305 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
306 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
309 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
310 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
314 ptlrpc_req_finished(req);
318 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
319 struct obd_info *oinfo, struct obd_trans_info *oti)
321 struct ptlrpc_request *req;
322 struct ost_body *body;
326 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
328 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
332 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
333 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
335 ptlrpc_request_free(req);
339 osc_pack_req_body(req, oinfo);
341 ptlrpc_request_set_replen(req);
343 rc = ptlrpc_queue_wait(req);
347 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
349 GOTO(out, rc = -EPROTO);
351 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
356 ptlrpc_req_finished(req);
360 static int osc_setattr_interpret(const struct lu_env *env,
361 struct ptlrpc_request *req,
362 struct osc_setattr_args *sa, int rc)
364 struct ost_body *body;
370 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
372 GOTO(out, rc = -EPROTO);
374 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
377 rc = sa->sa_upcall(sa->sa_cookie, rc);
381 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
382 struct obd_trans_info *oti,
383 obd_enqueue_update_f upcall, void *cookie,
384 struct ptlrpc_request_set *rqset)
386 struct ptlrpc_request *req;
387 struct osc_setattr_args *sa;
391 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
395 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
396 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
398 ptlrpc_request_free(req);
402 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
403 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
405 osc_pack_req_body(req, oinfo);
407 ptlrpc_request_set_replen(req);
409 /* do mds to ost setattr asynchronously */
411 /* Do not wait for response. */
412 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
414 req->rq_interpret_reply =
415 (ptlrpc_interpterer_t)osc_setattr_interpret;
417 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
418 sa = ptlrpc_req_async_args(req);
419 sa->sa_oa = oinfo->oi_oa;
420 sa->sa_upcall = upcall;
421 sa->sa_cookie = cookie;
423 if (rqset == PTLRPCD_SET)
424 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
426 ptlrpc_set_add_req(rqset, req);
432 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
433 struct obd_trans_info *oti,
434 struct ptlrpc_request_set *rqset)
436 return osc_setattr_async_base(exp, oinfo, oti,
437 oinfo->oi_cb_up, oinfo, rqset);
440 int osc_real_create(struct obd_export *exp, struct obdo *oa,
441 struct lov_stripe_md **ea, struct obd_trans_info *oti)
443 struct ptlrpc_request *req;
444 struct ost_body *body;
445 struct lov_stripe_md *lsm;
454 rc = obd_alloc_memmd(exp, &lsm);
459 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
461 GOTO(out, rc = -ENOMEM);
463 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
465 ptlrpc_request_free(req);
469 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
472 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
474 ptlrpc_request_set_replen(req);
476 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
477 oa->o_flags == OBD_FL_DELORPHAN) {
479 "delorphan from OST integration");
480 /* Don't resend the delorphan req */
481 req->rq_no_resend = req->rq_no_delay = 1;
484 rc = ptlrpc_queue_wait(req);
488 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
490 GOTO(out_req, rc = -EPROTO);
492 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
493 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
495 oa->o_blksize = cli_brw_size(exp->exp_obd);
496 oa->o_valid |= OBD_MD_FLBLKSZ;
498 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
499 * have valid lsm_oinfo data structs, so don't go touching that.
500 * This needs to be fixed in a big way.
502 lsm->lsm_oi = oa->o_oi;
506 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
508 if (oa->o_valid & OBD_MD_FLCOOKIE) {
509 if (!oti->oti_logcookies)
510 oti_alloc_cookies(oti, 1);
511 *oti->oti_logcookies = oa->o_lcookie;
515 CDEBUG(D_HA, "transno: "LPD64"\n",
516 lustre_msg_get_transno(req->rq_repmsg));
518 ptlrpc_req_finished(req);
521 obd_free_memmd(exp, &lsm);
525 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
526 obd_enqueue_update_f upcall, void *cookie,
527 struct ptlrpc_request_set *rqset)
529 struct ptlrpc_request *req;
530 struct osc_setattr_args *sa;
531 struct ost_body *body;
535 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
539 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
540 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
542 ptlrpc_request_free(req);
545 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
546 ptlrpc_at_set_req_timeout(req);
548 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
550 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
552 osc_pack_capa(req, body, oinfo->oi_capa);
554 ptlrpc_request_set_replen(req);
556 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
557 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
558 sa = ptlrpc_req_async_args(req);
559 sa->sa_oa = oinfo->oi_oa;
560 sa->sa_upcall = upcall;
561 sa->sa_cookie = cookie;
562 if (rqset == PTLRPCD_SET)
563 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
565 ptlrpc_set_add_req(rqset, req);
570 static int osc_sync_interpret(const struct lu_env *env,
571 struct ptlrpc_request *req,
574 struct osc_fsync_args *fa = arg;
575 struct ost_body *body;
581 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
583 CERROR ("can't unpack ost_body\n");
584 GOTO(out, rc = -EPROTO);
587 *fa->fa_oi->oi_oa = body->oa;
589 rc = fa->fa_upcall(fa->fa_cookie, rc);
593 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
594 obd_enqueue_update_f upcall, void *cookie,
595 struct ptlrpc_request_set *rqset)
597 struct ptlrpc_request *req;
598 struct ost_body *body;
599 struct osc_fsync_args *fa;
603 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
607 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
608 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
610 ptlrpc_request_free(req);
614 /* overload the size and blocks fields in the oa with start/end */
615 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
617 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
619 osc_pack_capa(req, body, oinfo->oi_capa);
621 ptlrpc_request_set_replen(req);
622 req->rq_interpret_reply = osc_sync_interpret;
624 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
625 fa = ptlrpc_req_async_args(req);
627 fa->fa_upcall = upcall;
628 fa->fa_cookie = cookie;
630 if (rqset == PTLRPCD_SET)
631 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
633 ptlrpc_set_add_req(rqset, req);
638 /* Find and cancel locally locks matched by @mode in the resource found by
639 * @objid. Found locks are added into @cancel list. Returns the amount of
640 * locks added to @cancels list. */
641 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
642 struct list_head *cancels,
643 ldlm_mode_t mode, __u64 lock_flags)
645 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
646 struct ldlm_res_id res_id;
647 struct ldlm_resource *res;
651 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
652 * export) but disabled through procfs (flag in NS).
654 * This distinguishes from a case when ELC is not supported originally,
655 * when we still want to cancel locks in advance and just cancel them
656 * locally, without sending any RPC. */
657 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
660 ostid_build_res_name(&oa->o_oi, &res_id);
661 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
665 LDLM_RESOURCE_ADDREF(res);
666 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
667 lock_flags, 0, NULL);
668 LDLM_RESOURCE_DELREF(res);
669 ldlm_resource_putref(res);
673 static int osc_destroy_interpret(const struct lu_env *env,
674 struct ptlrpc_request *req, void *data,
677 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
679 atomic_dec(&cli->cl_destroy_in_flight);
680 wake_up(&cli->cl_destroy_waitq);
684 static int osc_can_send_destroy(struct client_obd *cli)
686 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
687 cli->cl_max_rpcs_in_flight) {
688 /* The destroy request can be sent */
691 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
692 cli->cl_max_rpcs_in_flight) {
694 * The counter has been modified between the two atomic
697 wake_up(&cli->cl_destroy_waitq);
702 int osc_create(const struct lu_env *env, struct obd_export *exp,
703 struct obdo *oa, struct lov_stripe_md **ea,
704 struct obd_trans_info *oti)
711 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
713 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
714 oa->o_flags == OBD_FL_RECREATE_OBJS) {
715 RETURN(osc_real_create(exp, oa, ea, oti));
718 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
719 RETURN(osc_real_create(exp, oa, ea, oti));
721 /* we should not get here anymore */
727 /* Destroy requests can be async always on the client, and we don't even really
728 * care about the return code since the client cannot do anything at all about
730 * When the MDS is unlinking a filename, it saves the file objects into a
731 * recovery llog, and these object records are cancelled when the OST reports
732 * they were destroyed and sync'd to disk (i.e. transaction committed).
733 * If the client dies, or the OST is down when the object should be destroyed,
734 * the records are not cancelled, and when the OST reconnects to the MDS next,
735 * it will retrieve the llog unlink logs and then sends the log cancellation
736 * cookies to the MDS after committing destroy transactions. */
737 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
738 struct obdo *oa, struct lov_stripe_md *ea,
739 struct obd_trans_info *oti, struct obd_export *md_export,
742 struct client_obd *cli = &exp->exp_obd->u.cli;
743 struct ptlrpc_request *req;
744 struct ost_body *body;
745 struct list_head cancels = LIST_HEAD_INIT(cancels);
750 CDEBUG(D_INFO, "oa NULL\n");
754 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
755 LDLM_FL_DISCARD_DATA);
757 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
759 ldlm_lock_list_put(&cancels, l_bl_ast, count);
763 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
764 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
767 ptlrpc_request_free(req);
771 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
772 ptlrpc_at_set_req_timeout(req);
774 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
775 oa->o_lcookie = *oti->oti_logcookies;
776 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
778 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
780 osc_pack_capa(req, body, (struct obd_capa *)capa);
781 ptlrpc_request_set_replen(req);
783 /* If osc_destory is for destroying the unlink orphan,
784 * sent from MDT to OST, which should not be blocked here,
785 * because the process might be triggered by ptlrpcd, and
786 * it is not good to block ptlrpcd thread (b=16006)*/
787 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
788 req->rq_interpret_reply = osc_destroy_interpret;
789 if (!osc_can_send_destroy(cli)) {
790 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
794 * Wait until the number of on-going destroy RPCs drops
795 * under max_rpc_in_flight
797 l_wait_event_exclusive(cli->cl_destroy_waitq,
798 osc_can_send_destroy(cli), &lwi);
802 /* Do not wait for response */
803 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
807 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
810 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
812 LASSERT(!(oa->o_valid & bits));
815 client_obd_list_lock(&cli->cl_loi_list_lock);
816 oa->o_dirty = cli->cl_dirty;
817 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
818 cli->cl_dirty_max)) {
819 CERROR("dirty %lu - %lu > dirty_max %lu\n",
820 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
822 } else if (unlikely(atomic_read(&obd_unstable_pages) +
823 atomic_read(&obd_dirty_pages) -
824 atomic_read(&obd_dirty_transit_pages) >
825 (long)(obd_max_dirty_pages + 1))) {
826 /* The atomic_read() allowing the atomic_inc() are
827 * not covered by a lock thus they may safely race and trip
828 * this CERROR() unless we add in a small fudge factor (+1). */
829 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
830 cli->cl_import->imp_obd->obd_name,
831 atomic_read(&obd_unstable_pages),
832 atomic_read(&obd_dirty_pages),
833 atomic_read(&obd_dirty_transit_pages),
834 obd_max_dirty_pages);
836 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
837 CERROR("dirty %lu - dirty_max %lu too big???\n",
838 cli->cl_dirty, cli->cl_dirty_max);
841 long max_in_flight = (cli->cl_max_pages_per_rpc <<
843 (cli->cl_max_rpcs_in_flight + 1);
844 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
846 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
847 oa->o_dropped = cli->cl_lost_grant;
848 cli->cl_lost_grant = 0;
849 client_obd_list_unlock(&cli->cl_loi_list_lock);
850 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
851 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
855 void osc_update_next_shrink(struct client_obd *cli)
857 cli->cl_next_shrink_grant =
858 cfs_time_shift(cli->cl_grant_shrink_interval);
859 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
860 cli->cl_next_shrink_grant);
863 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
865 client_obd_list_lock(&cli->cl_loi_list_lock);
866 cli->cl_avail_grant += grant;
867 client_obd_list_unlock(&cli->cl_loi_list_lock);
870 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
872 if (body->oa.o_valid & OBD_MD_FLGRANT) {
873 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
874 __osc_update_grant(cli, body->oa.o_grant);
878 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
879 obd_count keylen, void *key, obd_count vallen,
880 void *val, struct ptlrpc_request_set *set);
882 static int osc_shrink_grant_interpret(const struct lu_env *env,
883 struct ptlrpc_request *req,
886 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
887 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
888 struct ost_body *body;
891 __osc_update_grant(cli, oa->o_grant);
895 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
897 osc_update_grant(cli, body);
903 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
905 client_obd_list_lock(&cli->cl_loi_list_lock);
906 oa->o_grant = cli->cl_avail_grant / 4;
907 cli->cl_avail_grant -= oa->o_grant;
908 client_obd_list_unlock(&cli->cl_loi_list_lock);
909 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
910 oa->o_valid |= OBD_MD_FLFLAGS;
913 oa->o_flags |= OBD_FL_SHRINK_GRANT;
914 osc_update_next_shrink(cli);
917 /* Shrink the current grant, either from some large amount to enough for a
918 * full set of in-flight RPCs, or if we have already shrunk to that limit
919 * then to enough for a single RPC. This avoids keeping more grant than
920 * needed, and avoids shrinking the grant piecemeal. */
921 static int osc_shrink_grant(struct client_obd *cli)
923 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
924 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
926 client_obd_list_lock(&cli->cl_loi_list_lock);
927 if (cli->cl_avail_grant <= target_bytes)
928 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
929 client_obd_list_unlock(&cli->cl_loi_list_lock);
931 return osc_shrink_grant_to_target(cli, target_bytes);
934 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
937 struct ost_body *body;
940 client_obd_list_lock(&cli->cl_loi_list_lock);
941 /* Don't shrink if we are already above or below the desired limit
942 * We don't want to shrink below a single RPC, as that will negatively
943 * impact block allocation and long-term performance. */
944 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
945 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
947 if (target_bytes >= cli->cl_avail_grant) {
948 client_obd_list_unlock(&cli->cl_loi_list_lock);
951 client_obd_list_unlock(&cli->cl_loi_list_lock);
957 osc_announce_cached(cli, &body->oa, 0);
959 client_obd_list_lock(&cli->cl_loi_list_lock);
960 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
961 cli->cl_avail_grant = target_bytes;
962 client_obd_list_unlock(&cli->cl_loi_list_lock);
963 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
964 body->oa.o_valid |= OBD_MD_FLFLAGS;
965 body->oa.o_flags = 0;
967 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
968 osc_update_next_shrink(cli);
970 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
971 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
972 sizeof(*body), body, NULL);
974 __osc_update_grant(cli, body->oa.o_grant);
979 static int osc_should_shrink_grant(struct client_obd *client)
981 cfs_time_t time = cfs_time_current();
982 cfs_time_t next_shrink = client->cl_next_shrink_grant;
984 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
985 OBD_CONNECT_GRANT_SHRINK) == 0)
988 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
989 /* Get the current RPC size directly, instead of going via:
990 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
991 * Keep comment here so that it can be found by searching. */
992 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
994 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
995 client->cl_avail_grant > brw_size)
998 osc_update_next_shrink(client);
1003 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1005 struct client_obd *client;
1007 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1008 if (osc_should_shrink_grant(client))
1009 osc_shrink_grant(client);
1014 static int osc_add_shrink_grant(struct client_obd *client)
1018 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1020 osc_grant_shrink_grant_cb, NULL,
1021 &client->cl_grant_shrink_list);
1023 CERROR("add grant client %s error %d\n",
1024 client->cl_import->imp_obd->obd_name, rc);
1027 CDEBUG(D_CACHE, "add grant client %s \n",
1028 client->cl_import->imp_obd->obd_name);
1029 osc_update_next_shrink(client);
1033 static int osc_del_shrink_grant(struct client_obd *client)
1035 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1039 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1042 * ocd_grant is the total grant amount we're expect to hold: if we've
1043 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1044 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1046 * race is tolerable here: if we're evicted, but imp_state already
1047 * left EVICTED state, then cl_dirty must be 0 already.
1049 client_obd_list_lock(&cli->cl_loi_list_lock);
1050 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1051 cli->cl_avail_grant = ocd->ocd_grant;
1053 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1055 if (cli->cl_avail_grant < 0) {
1056 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1057 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1058 ocd->ocd_grant, cli->cl_dirty);
1059 /* workaround for servers which do not have the patch from
1061 cli->cl_avail_grant = ocd->ocd_grant;
1064 /* determine the appropriate chunk size used by osc_extent. */
1065 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1066 client_obd_list_unlock(&cli->cl_loi_list_lock);
1068 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1069 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1070 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1072 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1073 list_empty(&cli->cl_grant_shrink_list))
1074 osc_add_shrink_grant(cli);
1077 /* We assume that the reason this OSC got a short read is because it read
1078 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1079 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1080 * this stripe never got written at or beyond this stripe offset yet. */
1081 static void handle_short_read(int nob_read, obd_count page_count,
1082 struct brw_page **pga)
1087 /* skip bytes read OK */
1088 while (nob_read > 0) {
1089 LASSERT (page_count > 0);
1091 if (pga[i]->count > nob_read) {
1092 /* EOF inside this page */
1093 ptr = kmap(pga[i]->pg) +
1094 (pga[i]->off & ~CFS_PAGE_MASK);
1095 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1102 nob_read -= pga[i]->count;
1107 /* zero remaining pages */
1108 while (page_count-- > 0) {
1109 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1110 memset(ptr, 0, pga[i]->count);
1116 static int check_write_rcs(struct ptlrpc_request *req,
1117 int requested_nob, int niocount,
1118 obd_count page_count, struct brw_page **pga)
1123 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1124 sizeof(*remote_rcs) *
1126 if (remote_rcs == NULL) {
1127 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1131 /* return error if any niobuf was in error */
1132 for (i = 0; i < niocount; i++) {
1133 if ((int)remote_rcs[i] < 0)
1134 return(remote_rcs[i]);
1136 if (remote_rcs[i] != 0) {
1137 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1138 i, remote_rcs[i], req);
1143 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1144 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1145 req->rq_bulk->bd_nob_transferred, requested_nob);
1152 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1154 if (p1->flag != p2->flag) {
1155 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1156 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1157 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1159 /* warn if we try to combine flags that we don't know to be
1160 * safe to combine */
1161 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1162 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1163 "report this at http://bugs.whamcloud.com/\n",
1164 p1->flag, p2->flag);
1169 return (p1->off + p1->count == p2->off);
1172 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1173 struct brw_page **pga, int opc,
1174 cksum_type_t cksum_type)
1178 struct cfs_crypto_hash_desc *hdesc;
1179 unsigned int bufsize;
1181 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1183 LASSERT(pg_count > 0);
1185 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1186 if (IS_ERR(hdesc)) {
1187 CERROR("Unable to initialize checksum hash %s\n",
1188 cfs_crypto_hash_name(cfs_alg));
1189 return PTR_ERR(hdesc);
1192 while (nob > 0 && pg_count > 0) {
1193 int count = pga[i]->count > nob ? nob : pga[i]->count;
1195 /* corrupt the data before we compute the checksum, to
1196 * simulate an OST->client data error */
1197 if (i == 0 && opc == OST_READ &&
1198 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1199 unsigned char *ptr = kmap(pga[i]->pg);
1200 int off = pga[i]->off & ~CFS_PAGE_MASK;
1201 memcpy(ptr + off, "bad1", min(4, nob));
1204 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1205 pga[i]->off & ~CFS_PAGE_MASK,
1207 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1208 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1210 nob -= pga[i]->count;
1216 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1219 cfs_crypto_hash_final(hdesc, NULL, NULL);
1221 /* For sending we only compute the wrong checksum instead
1222 * of corrupting the data so it is still correct on a redo */
1223 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1229 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1230 struct lov_stripe_md *lsm, obd_count page_count,
1231 struct brw_page **pga,
1232 struct ptlrpc_request **reqp,
1233 struct obd_capa *ocapa, int reserve,
1236 struct ptlrpc_request *req;
1237 struct ptlrpc_bulk_desc *desc;
1238 struct ost_body *body;
1239 struct obd_ioobj *ioobj;
1240 struct niobuf_remote *niobuf;
1241 int niocount, i, requested_nob, opc, rc;
1242 struct osc_brw_async_args *aa;
1243 struct req_capsule *pill;
1244 struct brw_page *pg_prev;
1247 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1248 RETURN(-ENOMEM); /* Recoverable */
1249 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1250 RETURN(-EINVAL); /* Fatal */
1252 if ((cmd & OBD_BRW_WRITE) != 0) {
1254 req = ptlrpc_request_alloc_pool(cli->cl_import,
1255 cli->cl_import->imp_rq_pool,
1256 &RQF_OST_BRW_WRITE);
1259 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1264 for (niocount = i = 1; i < page_count; i++) {
1265 if (!can_merge_pages(pga[i - 1], pga[i]))
1269 pill = &req->rq_pill;
1270 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1272 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1273 niocount * sizeof(*niobuf));
1274 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1278 ptlrpc_request_free(req);
1281 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1282 ptlrpc_at_set_req_timeout(req);
1283 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1285 req->rq_no_retry_einprogress = 1;
1287 desc = ptlrpc_prep_bulk_imp(req, page_count,
1288 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1289 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1293 GOTO(out, rc = -ENOMEM);
1294 /* NB request now owns desc and will free it when it gets freed */
1296 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1297 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1298 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1299 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1301 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1303 obdo_to_ioobj(oa, ioobj);
1304 ioobj->ioo_bufcnt = niocount;
1305 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1306 * that might be send for this request. The actual number is decided
1307 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1308 * "max - 1" for old client compatibility sending "0", and also so the
1309 * the actual maximum is a power-of-two number, not one less. LU-1431 */
1310 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1311 osc_pack_capa(req, body, ocapa);
1312 LASSERT(page_count > 0);
1314 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1315 struct brw_page *pg = pga[i];
1316 int poff = pg->off & ~CFS_PAGE_MASK;
1318 LASSERT(pg->count > 0);
1319 /* make sure there is no gap in the middle of page array */
1320 LASSERTF(page_count == 1 ||
1321 (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1322 ergo(i > 0 && i < page_count - 1,
1323 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1324 ergo(i == page_count - 1, poff == 0)),
1325 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1326 i, page_count, pg, pg->off, pg->count);
1328 LASSERTF(i == 0 || pg->off > pg_prev->off,
1329 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1330 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1332 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1333 pg_prev->pg, page_private(pg_prev->pg),
1334 pg_prev->pg->index, pg_prev->off);
1336 LASSERTF(i == 0 || pg->off > pg_prev->off,
1337 "i %d p_c %u\n", i, page_count);
1339 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1340 (pg->flag & OBD_BRW_SRVLOCK));
1342 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1343 requested_nob += pg->count;
1345 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1347 niobuf->len += pg->count;
1349 niobuf->offset = pg->off;
1350 niobuf->len = pg->count;
1351 niobuf->flags = pg->flag;
1356 LASSERTF((void *)(niobuf - niocount) ==
1357 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1358 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1359 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1361 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1363 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1364 body->oa.o_valid |= OBD_MD_FLFLAGS;
1365 body->oa.o_flags = 0;
1367 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1370 if (osc_should_shrink_grant(cli))
1371 osc_shrink_grant_local(cli, &body->oa);
1373 /* size[REQ_REC_OFF] still sizeof (*body) */
1374 if (opc == OST_WRITE) {
1375 if (cli->cl_checksum &&
1376 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1377 /* store cl_cksum_type in a local variable since
1378 * it can be changed via lprocfs */
1379 cksum_type_t cksum_type = cli->cl_cksum_type;
1381 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1382 oa->o_flags &= OBD_FL_LOCAL_MASK;
1383 body->oa.o_flags = 0;
1385 body->oa.o_flags |= cksum_type_pack(cksum_type);
1386 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1387 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1391 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1393 /* save this in 'oa', too, for later checking */
1394 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1395 oa->o_flags |= cksum_type_pack(cksum_type);
1397 /* clear out the checksum flag, in case this is a
1398 * resend but cl_checksum is no longer set. b=11238 */
1399 oa->o_valid &= ~OBD_MD_FLCKSUM;
1401 oa->o_cksum = body->oa.o_cksum;
1402 /* 1 RC per niobuf */
1403 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1404 sizeof(__u32) * niocount);
1406 if (cli->cl_checksum &&
1407 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1408 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1409 body->oa.o_flags = 0;
1410 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1411 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1414 ptlrpc_request_set_replen(req);
1416 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1417 aa = ptlrpc_req_async_args(req);
1419 aa->aa_requested_nob = requested_nob;
1420 aa->aa_nio_count = niocount;
1421 aa->aa_page_count = page_count;
1425 INIT_LIST_HEAD(&aa->aa_oaps);
1426 if (ocapa && reserve)
1427 aa->aa_ocapa = capa_get(ocapa);
1433 ptlrpc_req_finished(req);
1437 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1438 __u32 client_cksum, __u32 server_cksum, int nob,
1439 obd_count page_count, struct brw_page **pga,
1440 cksum_type_t client_cksum_type)
1444 cksum_type_t cksum_type;
1446 if (server_cksum == client_cksum) {
1447 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1451 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1453 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1456 if (cksum_type != client_cksum_type)
1457 msg = "the server did not use the checksum type specified in "
1458 "the original request - likely a protocol problem";
1459 else if (new_cksum == server_cksum)
1460 msg = "changed on the client after we checksummed it - "
1461 "likely false positive due to mmap IO (bug 11742)";
1462 else if (new_cksum == client_cksum)
1463 msg = "changed in transit before arrival at OST";
1465 msg = "changed in transit AND doesn't match the original - "
1466 "likely false positive due to mmap IO (bug 11742)";
1468 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1469 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1470 msg, libcfs_nid2str(peer->nid),
1471 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1472 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1473 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1474 POSTID(&oa->o_oi), pga[0]->off,
1475 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1476 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1477 "client csum now %x\n", client_cksum, client_cksum_type,
1478 server_cksum, cksum_type, new_cksum);
1482 /* Note rc enters this function as number of bytes transferred */
1483 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1485 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1486 const lnet_process_id_t *peer =
1487 &req->rq_import->imp_connection->c_peer;
1488 struct client_obd *cli = aa->aa_cli;
1489 struct ost_body *body;
1490 __u32 client_cksum = 0;
1493 if (rc < 0 && rc != -EDQUOT) {
1494 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1498 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1499 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1501 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1505 /* set/clear over quota flag for a uid/gid */
1506 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1507 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1508 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1510 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1511 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1513 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1516 osc_update_grant(cli, body);
1521 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1522 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1524 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1526 CERROR("Unexpected +ve rc %d\n", rc);
1529 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1531 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1534 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1535 check_write_checksum(&body->oa, peer, client_cksum,
1536 body->oa.o_cksum, aa->aa_requested_nob,
1537 aa->aa_page_count, aa->aa_ppga,
1538 cksum_type_unpack(aa->aa_oa->o_flags)))
1541 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1542 aa->aa_page_count, aa->aa_ppga);
1546 /* The rest of this function executes only for OST_READs */
1548 /* if unwrap_bulk failed, return -EAGAIN to retry */
1549 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1551 GOTO(out, rc = -EAGAIN);
1553 if (rc > aa->aa_requested_nob) {
1554 CERROR("Unexpected rc %d (%d requested)\n", rc,
1555 aa->aa_requested_nob);
1559 if (rc != req->rq_bulk->bd_nob_transferred) {
1560 CERROR ("Unexpected rc %d (%d transferred)\n",
1561 rc, req->rq_bulk->bd_nob_transferred);
1565 if (rc < aa->aa_requested_nob)
1566 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1568 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1569 static int cksum_counter;
1570 __u32 server_cksum = body->oa.o_cksum;
1573 cksum_type_t cksum_type;
1575 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1576 body->oa.o_flags : 0);
1577 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1578 aa->aa_ppga, OST_READ,
1581 if (peer->nid == req->rq_bulk->bd_sender) {
1585 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1588 if (server_cksum != client_cksum) {
1589 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1590 "%s%s%s inode "DFID" object "DOSTID
1591 " extent ["LPU64"-"LPU64"]\n",
1592 req->rq_import->imp_obd->obd_name,
1593 libcfs_nid2str(peer->nid),
1595 body->oa.o_valid & OBD_MD_FLFID ?
1596 body->oa.o_parent_seq : (__u64)0,
1597 body->oa.o_valid & OBD_MD_FLFID ?
1598 body->oa.o_parent_oid : 0,
1599 body->oa.o_valid & OBD_MD_FLFID ?
1600 body->oa.o_parent_ver : 0,
1601 POSTID(&body->oa.o_oi),
1602 aa->aa_ppga[0]->off,
1603 aa->aa_ppga[aa->aa_page_count-1]->off +
1604 aa->aa_ppga[aa->aa_page_count-1]->count -
1606 CERROR("client %x, server %x, cksum_type %x\n",
1607 client_cksum, server_cksum, cksum_type);
1609 aa->aa_oa->o_cksum = client_cksum;
1613 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1616 } else if (unlikely(client_cksum)) {
1617 static int cksum_missed;
1620 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1621 CERROR("Checksum %u requested from %s but not sent\n",
1622 cksum_missed, libcfs_nid2str(peer->nid));
1628 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1629 aa->aa_oa, &body->oa);
1634 static int osc_brw_redo_request(struct ptlrpc_request *request,
1635 struct osc_brw_async_args *aa, int rc)
1637 struct ptlrpc_request *new_req;
1638 struct osc_brw_async_args *new_aa;
1639 struct osc_async_page *oap;
1642 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1643 "redo for recoverable error %d", rc);
1645 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1646 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1647 aa->aa_cli, aa->aa_oa,
1648 NULL /* lsm unused by osc currently */,
1649 aa->aa_page_count, aa->aa_ppga,
1650 &new_req, aa->aa_ocapa, 0, 1);
1654 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1655 if (oap->oap_request != NULL) {
1656 LASSERTF(request == oap->oap_request,
1657 "request %p != oap_request %p\n",
1658 request, oap->oap_request);
1659 if (oap->oap_interrupted) {
1660 ptlrpc_req_finished(new_req);
1665 /* New request takes over pga and oaps from old request.
1666 * Note that copying a list_head doesn't work, need to move it... */
1668 new_req->rq_interpret_reply = request->rq_interpret_reply;
1669 new_req->rq_async_args = request->rq_async_args;
1670 new_req->rq_commit_cb = request->rq_commit_cb;
1671 /* cap resend delay to the current request timeout, this is similar to
1672 * what ptlrpc does (see after_reply()) */
1673 if (aa->aa_resends > new_req->rq_timeout)
1674 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1676 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1677 new_req->rq_generation_set = 1;
1678 new_req->rq_import_generation = request->rq_import_generation;
1680 new_aa = ptlrpc_req_async_args(new_req);
1682 INIT_LIST_HEAD(&new_aa->aa_oaps);
1683 list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1684 INIT_LIST_HEAD(&new_aa->aa_exts);
1685 list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1686 new_aa->aa_resends = aa->aa_resends;
1688 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1689 if (oap->oap_request) {
1690 ptlrpc_req_finished(oap->oap_request);
1691 oap->oap_request = ptlrpc_request_addref(new_req);
1695 new_aa->aa_ocapa = aa->aa_ocapa;
1696 aa->aa_ocapa = NULL;
1698 /* XXX: This code will run into problem if we're going to support
1699 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1700 * and wait for all of them to be finished. We should inherit request
1701 * set from old request. */
1702 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1704 DEBUG_REQ(D_INFO, new_req, "new request");
1709 * ugh, we want disk allocation on the target to happen in offset order. we'll
1710 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1711 * fine for our small page arrays and doesn't require allocation. its an
1712 * insertion sort that swaps elements that are strides apart, shrinking the
1713 * stride down until its '1' and the array is sorted.
1715 static void sort_brw_pages(struct brw_page **array, int num)
1718 struct brw_page *tmp;
1722 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1727 for (i = stride ; i < num ; i++) {
1730 while (j >= stride && array[j - stride]->off > tmp->off) {
1731 array[j] = array[j - stride];
1736 } while (stride > 1);
1739 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1741 LASSERT(ppga != NULL);
1742 OBD_FREE(ppga, sizeof(*ppga) * count);
1745 static int brw_interpret(const struct lu_env *env,
1746 struct ptlrpc_request *req, void *data, int rc)
1748 struct osc_brw_async_args *aa = data;
1749 struct osc_extent *ext;
1750 struct osc_extent *tmp;
1751 struct client_obd *cli = aa->aa_cli;
1754 rc = osc_brw_fini_request(req, rc);
1755 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1756 /* When server return -EINPROGRESS, client should always retry
1757 * regardless of the number of times the bulk was resent already. */
1758 if (osc_recoverable_error(rc)) {
1759 if (req->rq_import_generation !=
1760 req->rq_import->imp_generation) {
1761 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1762 ""DOSTID", rc = %d.\n",
1763 req->rq_import->imp_obd->obd_name,
1764 POSTID(&aa->aa_oa->o_oi), rc);
1765 } else if (rc == -EINPROGRESS ||
1766 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1767 rc = osc_brw_redo_request(req, aa, rc);
1769 CERROR("%s: too many resent retries for object: "
1770 ""LPU64":"LPU64", rc = %d.\n",
1771 req->rq_import->imp_obd->obd_name,
1772 POSTID(&aa->aa_oa->o_oi), rc);
1777 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1782 capa_put(aa->aa_ocapa);
1783 aa->aa_ocapa = NULL;
1787 struct obdo *oa = aa->aa_oa;
1788 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1789 unsigned long valid = 0;
1790 struct cl_object *obj;
1791 struct osc_async_page *last;
1793 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1794 obj = osc2cl(last->oap_obj);
1796 cl_object_attr_lock(obj);
1797 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1798 attr->cat_blocks = oa->o_blocks;
1799 valid |= CAT_BLOCKS;
1801 if (oa->o_valid & OBD_MD_FLMTIME) {
1802 attr->cat_mtime = oa->o_mtime;
1805 if (oa->o_valid & OBD_MD_FLATIME) {
1806 attr->cat_atime = oa->o_atime;
1809 if (oa->o_valid & OBD_MD_FLCTIME) {
1810 attr->cat_ctime = oa->o_ctime;
1814 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1815 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1816 loff_t last_off = last->oap_count + last->oap_obj_off;
1818 /* Change file size if this is an out of quota or
1819 * direct IO write and it extends the file size */
1820 if (loi->loi_lvb.lvb_size < last_off) {
1821 attr->cat_size = last_off;
1824 /* Extend KMS if it's not a lockless write */
1825 if (loi->loi_kms < last_off &&
1826 oap2osc_page(last)->ops_srvlock == 0) {
1827 attr->cat_kms = last_off;
1833 cl_object_attr_set(env, obj, attr, valid);
1834 cl_object_attr_unlock(obj);
1836 OBDO_FREE(aa->aa_oa);
1838 list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1839 list_del_init(&ext->oe_link);
1840 osc_extent_finish(env, ext, 1, rc);
1842 LASSERT(list_empty(&aa->aa_exts));
1843 LASSERT(list_empty(&aa->aa_oaps));
1845 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1846 req->rq_bulk->bd_nob_transferred);
1847 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1848 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1850 client_obd_list_lock(&cli->cl_loi_list_lock);
1851 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1852 * is called so we know whether to go to sync BRWs or wait for more
1853 * RPCs to complete */
1854 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1855 cli->cl_w_in_flight--;
1857 cli->cl_r_in_flight--;
1858 osc_wake_cache_waiters(cli);
1859 client_obd_list_unlock(&cli->cl_loi_list_lock);
1861 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
1865 static void brw_commit(struct ptlrpc_request *req)
1867 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1868 * this called via the rq_commit_cb, I need to ensure
1869 * osc_dec_unstable_pages is still called. Otherwise unstable
1870 * pages may be leaked. */
1871 spin_lock(&req->rq_lock);
1872 if (likely(req->rq_unstable)) {
1873 req->rq_unstable = 0;
1874 spin_unlock(&req->rq_lock);
1876 osc_dec_unstable_pages(req);
1878 req->rq_committed = 1;
1879 spin_unlock(&req->rq_lock);
1884 * Build an RPC by the list of extent @ext_list. The caller must ensure
1885 * that the total pages in this list are NOT over max pages per RPC.
1886 * Extents in the list must be in OES_RPC state.
1888 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1889 struct list_head *ext_list, int cmd, pdl_policy_t pol)
1891 struct ptlrpc_request *req = NULL;
1892 struct osc_extent *ext;
1893 struct brw_page **pga = NULL;
1894 struct osc_brw_async_args *aa = NULL;
1895 struct obdo *oa = NULL;
1896 struct osc_async_page *oap;
1897 struct osc_async_page *tmp;
1898 struct cl_req *clerq = NULL;
1899 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
1901 struct ldlm_lock *lock = NULL;
1902 struct cl_req_attr *crattr = NULL;
1903 obd_off starting_offset = OBD_OBJECT_EOF;
1904 obd_off ending_offset = 0;
1910 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1913 LASSERT(!list_empty(ext_list));
1915 /* add pages into rpc_list to build BRW rpc */
1916 list_for_each_entry(ext, ext_list, oe_link) {
1917 LASSERT(ext->oe_state == OES_RPC);
1918 mem_tight |= ext->oe_memalloc;
1919 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1921 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1922 if (starting_offset > oap->oap_obj_off)
1923 starting_offset = oap->oap_obj_off;
1925 LASSERT(oap->oap_page_off == 0);
1926 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1927 ending_offset = oap->oap_obj_off +
1930 LASSERT(oap->oap_page_off + oap->oap_count ==
1936 mpflag = cfs_memory_pressure_get_and_set();
1938 OBD_ALLOC(crattr, sizeof(*crattr));
1940 GOTO(out, rc = -ENOMEM);
1942 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1944 GOTO(out, rc = -ENOMEM);
1948 GOTO(out, rc = -ENOMEM);
1951 list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1952 struct cl_page *page = oap2cl_page(oap);
1953 if (clerq == NULL) {
1954 clerq = cl_req_alloc(env, page, crt,
1955 1 /* only 1-object rpcs for now */);
1957 GOTO(out, rc = PTR_ERR(clerq));
1958 lock = oap->oap_ldlm_lock;
1961 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1962 pga[i] = &oap->oap_brw_page;
1963 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1964 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1965 pga[i]->pg, page_index(oap->oap_page), oap,
1968 cl_req_page_add(env, clerq, page);
1971 /* always get the data for the obdo for the rpc */
1972 LASSERT(clerq != NULL);
1973 crattr->cra_oa = oa;
1974 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1976 oa->o_handle = lock->l_remote_handle;
1977 oa->o_valid |= OBD_MD_FLHANDLE;
1980 rc = cl_req_prep(env, clerq);
1982 CERROR("cl_req_prep failed: %d\n", rc);
1986 sort_brw_pages(pga, page_count);
1987 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1988 pga, &req, crattr->cra_capa, 1, 0);
1990 CERROR("prep_req failed: %d\n", rc);
1994 req->rq_commit_cb = brw_commit;
1995 req->rq_interpret_reply = brw_interpret;
1998 req->rq_memalloc = 1;
2000 /* Need to update the timestamps after the request is built in case
2001 * we race with setattr (locally or in queue at OST). If OST gets
2002 * later setattr before earlier BRW (as determined by the request xid),
2003 * the OST will not use BRW timestamps. Sadly, there is no obvious
2004 * way to do this in a single call. bug 10150 */
2005 cl_req_attr_set(env, clerq, crattr,
2006 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2008 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2010 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2011 aa = ptlrpc_req_async_args(req);
2012 INIT_LIST_HEAD(&aa->aa_oaps);
2013 list_splice_init(&rpc_list, &aa->aa_oaps);
2014 INIT_LIST_HEAD(&aa->aa_exts);
2015 list_splice_init(ext_list, &aa->aa_exts);
2016 aa->aa_clerq = clerq;
2018 /* queued sync pages can be torn down while the pages
2019 * were between the pending list and the rpc */
2021 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2022 /* only one oap gets a request reference */
2025 if (oap->oap_interrupted && !req->rq_intr) {
2026 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2028 ptlrpc_mark_interrupted(req);
2032 tmp->oap_request = ptlrpc_request_addref(req);
2034 client_obd_list_lock(&cli->cl_loi_list_lock);
2035 starting_offset >>= PAGE_CACHE_SHIFT;
2036 if (cmd == OBD_BRW_READ) {
2037 cli->cl_r_in_flight++;
2038 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2039 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2040 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2041 starting_offset + 1);
2043 cli->cl_w_in_flight++;
2044 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2045 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2046 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2047 starting_offset + 1);
2049 client_obd_list_unlock(&cli->cl_loi_list_lock);
2051 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2052 page_count, aa, cli->cl_r_in_flight,
2053 cli->cl_w_in_flight);
2055 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2056 * see which CPU/NUMA node the majority of pages were allocated
2057 * on, and try to assign the async RPC to the CPU core
2058 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2060 * But on the other hand, we expect that multiple ptlrpcd
2061 * threads and the initial write sponsor can run in parallel,
2062 * especially when data checksum is enabled, which is CPU-bound
2063 * operation and single ptlrpcd thread cannot process in time.
2064 * So more ptlrpcd threads sharing BRW load
2065 * (with PDL_POLICY_ROUND) seems better.
2067 ptlrpcd_add_req(req, pol, -1);
2073 cfs_memory_pressure_restore(mpflag);
2075 if (crattr != NULL) {
2076 capa_put(crattr->cra_capa);
2077 OBD_FREE(crattr, sizeof(*crattr));
2081 LASSERT(req == NULL);
2086 OBD_FREE(pga, sizeof(*pga) * page_count);
2087 /* this should happen rarely and is pretty bad, it makes the
2088 * pending list not follow the dirty order */
2089 while (!list_empty(ext_list)) {
2090 ext = list_entry(ext_list->next, struct osc_extent,
2092 list_del_init(&ext->oe_link);
2093 osc_extent_finish(env, ext, 0, rc);
2095 if (clerq && !IS_ERR(clerq))
2096 cl_req_completion(env, clerq, rc);
2101 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2102 struct ldlm_enqueue_info *einfo)
2104 void *data = einfo->ei_cbdata;
2107 LASSERT(lock != NULL);
2108 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2109 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2110 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2111 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2113 lock_res_and_lock(lock);
2114 spin_lock(&osc_ast_guard);
2116 if (lock->l_ast_data == NULL)
2117 lock->l_ast_data = data;
2118 if (lock->l_ast_data == data)
2121 spin_unlock(&osc_ast_guard);
2122 unlock_res_and_lock(lock);
2127 static int osc_set_data_with_check(struct lustre_handle *lockh,
2128 struct ldlm_enqueue_info *einfo)
2130 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2134 set = osc_set_lock_data_with_check(lock, einfo);
2135 LDLM_LOCK_PUT(lock);
2137 CERROR("lockh %p, data %p - client evicted?\n",
2138 lockh, einfo->ei_cbdata);
2142 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2143 ldlm_iterator_t replace, void *data)
2145 struct ldlm_res_id res_id;
2146 struct obd_device *obd = class_exp2obd(exp);
2148 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2149 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2153 /* find any ldlm lock of the inode in osc
2157 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2158 ldlm_iterator_t replace, void *data)
2160 struct ldlm_res_id res_id;
2161 struct obd_device *obd = class_exp2obd(exp);
2164 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2165 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2166 if (rc == LDLM_ITER_STOP)
2168 if (rc == LDLM_ITER_CONTINUE)
2173 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2174 obd_enqueue_update_f upcall, void *cookie,
2175 __u64 *flags, int agl, int rc)
2177 int intent = *flags & LDLM_FL_HAS_INTENT;
2181 /* The request was created before ldlm_cli_enqueue call. */
2182 if (rc == ELDLM_LOCK_ABORTED) {
2183 struct ldlm_reply *rep;
2184 rep = req_capsule_server_get(&req->rq_pill,
2187 LASSERT(rep != NULL);
2188 rep->lock_policy_res1 =
2189 ptlrpc_status_ntoh(rep->lock_policy_res1);
2190 if (rep->lock_policy_res1)
2191 rc = rep->lock_policy_res1;
2195 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2197 *flags |= LDLM_FL_LVB_READY;
2198 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2199 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2202 /* Call the update callback. */
2203 rc = (*upcall)(cookie, rc);
2207 static int osc_enqueue_interpret(const struct lu_env *env,
2208 struct ptlrpc_request *req,
2209 struct osc_enqueue_args *aa, int rc)
2211 struct ldlm_lock *lock;
2212 struct lustre_handle handle;
2214 struct ost_lvb *lvb;
2216 __u64 *flags = aa->oa_flags;
2218 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2219 * might be freed anytime after lock upcall has been called. */
2220 lustre_handle_copy(&handle, aa->oa_lockh);
2221 mode = aa->oa_ei->ei_mode;
2223 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2225 lock = ldlm_handle2lock(&handle);
2227 /* Take an additional reference so that a blocking AST that
2228 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2229 * to arrive after an upcall has been executed by
2230 * osc_enqueue_fini(). */
2231 ldlm_lock_addref(&handle, mode);
2233 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2234 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2236 /* Let CP AST to grant the lock first. */
2237 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2239 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2244 lvb_len = sizeof(*aa->oa_lvb);
2247 /* Complete obtaining the lock procedure. */
2248 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2249 mode, flags, lvb, lvb_len, &handle, rc);
2250 /* Complete osc stuff. */
2251 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2252 flags, aa->oa_agl, rc);
2254 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2256 /* Release the lock for async request. */
2257 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2259 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2260 * not already released by
2261 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2263 ldlm_lock_decref(&handle, mode);
2265 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2266 aa->oa_lockh, req, aa);
2267 ldlm_lock_decref(&handle, mode);
2268 LDLM_LOCK_PUT(lock);
2272 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2274 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2275 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2276 * other synchronous requests, however keeping some locks and trying to obtain
2277 * others may take a considerable amount of time in a case of ost failure; and
2278 * when other sync requests do not get released lock from a client, the client
2279 * is excluded from the cluster -- such scenarious make the life difficult, so
2280 * release locks just after they are obtained. */
2281 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2282 __u64 *flags, ldlm_policy_data_t *policy,
2283 struct ost_lvb *lvb, int kms_valid,
2284 obd_enqueue_update_f upcall, void *cookie,
2285 struct ldlm_enqueue_info *einfo,
2286 struct lustre_handle *lockh,
2287 struct ptlrpc_request_set *rqset, int async, int agl)
2289 struct obd_device *obd = exp->exp_obd;
2290 struct ptlrpc_request *req = NULL;
2291 int intent = *flags & LDLM_FL_HAS_INTENT;
2292 __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2297 /* Filesystem lock extents are extended to page boundaries so that
2298 * dealing with the page cache is a little smoother. */
2299 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2300 policy->l_extent.end |= ~CFS_PAGE_MASK;
2303 * kms is not valid when either object is completely fresh (so that no
2304 * locks are cached), or object was evicted. In the latter case cached
2305 * lock cannot be used, because it would prime inode state with
2306 * potentially stale LVB.
2311 /* Next, search for already existing extent locks that will cover us */
2312 /* If we're trying to read, we also search for an existing PW lock. The
2313 * VFS and page cache already protect us locally, so lots of readers/
2314 * writers can share a single PW lock.
2316 * There are problems with conversion deadlocks, so instead of
2317 * converting a read lock to a write lock, we'll just enqueue a new
2320 * At some point we should cancel the read lock instead of making them
2321 * send us a blocking callback, but there are problems with canceling
2322 * locks out from other users right now, too. */
2323 mode = einfo->ei_mode;
2324 if (einfo->ei_mode == LCK_PR)
2326 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2327 einfo->ei_type, policy, mode, lockh, 0);
2329 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2331 if ((agl != 0) && !ldlm_is_lvb_ready(matched)) {
2332 /* For AGL, if enqueue RPC is sent but the lock is not
2333 * granted, then skip to process this strpe.
2334 * Return -ECANCELED to tell the caller. */
2335 ldlm_lock_decref(lockh, mode);
2336 LDLM_LOCK_PUT(matched);
2338 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2339 *flags |= LDLM_FL_LVB_READY;
2340 /* addref the lock only if not async requests and PW
2341 * lock is matched whereas we asked for PR. */
2342 if (!rqset && einfo->ei_mode != mode)
2343 ldlm_lock_addref(lockh, LCK_PR);
2345 /* I would like to be able to ASSERT here that
2346 * rss <= kms, but I can't, for reasons which
2347 * are explained in lov_enqueue() */
2350 /* We already have a lock, and it's referenced.
2352 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2353 * AGL upcall may change it to CLS_HELD directly. */
2354 (*upcall)(cookie, ELDLM_OK);
2356 if (einfo->ei_mode != mode)
2357 ldlm_lock_decref(lockh, LCK_PW);
2359 /* For async requests, decref the lock. */
2360 ldlm_lock_decref(lockh, einfo->ei_mode);
2361 LDLM_LOCK_PUT(matched);
2364 ldlm_lock_decref(lockh, mode);
2365 LDLM_LOCK_PUT(matched);
2371 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2372 &RQF_LDLM_ENQUEUE_LVB);
2376 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2378 ptlrpc_request_free(req);
2382 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2384 ptlrpc_request_set_replen(req);
2387 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2388 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2390 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2391 sizeof(*lvb), LVB_T_OST, lockh, async);
2394 struct osc_enqueue_args *aa;
2395 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2396 aa = ptlrpc_req_async_args(req);
2399 aa->oa_flags = flags;
2400 aa->oa_upcall = upcall;
2401 aa->oa_cookie = cookie;
2403 aa->oa_lockh = lockh;
2406 req->rq_interpret_reply =
2407 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2408 if (rqset == PTLRPCD_SET)
2409 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2411 ptlrpc_set_add_req(rqset, req);
2412 } else if (intent) {
2413 ptlrpc_req_finished(req);
2418 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2420 ptlrpc_req_finished(req);
2425 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2426 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2427 __u64 *flags, void *data, struct lustre_handle *lockh,
2430 struct obd_device *obd = exp->exp_obd;
2431 __u64 lflags = *flags;
2435 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2438 /* Filesystem lock extents are extended to page boundaries so that
2439 * dealing with the page cache is a little smoother */
2440 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2441 policy->l_extent.end |= ~CFS_PAGE_MASK;
2443 /* Next, search for already existing extent locks that will cover us */
2444 /* If we're trying to read, we also search for an existing PW lock. The
2445 * VFS and page cache already protect us locally, so lots of readers/
2446 * writers can share a single PW lock. */
2450 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2451 res_id, type, policy, rc, lockh, unref);
2454 if (!osc_set_data_with_check(lockh, data)) {
2455 if (!(lflags & LDLM_FL_TEST_LOCK))
2456 ldlm_lock_decref(lockh, rc);
2460 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2461 ldlm_lock_addref(lockh, LCK_PR);
2462 ldlm_lock_decref(lockh, LCK_PW);
2469 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2473 if (unlikely(mode == LCK_GROUP))
2474 ldlm_lock_decref_and_cancel(lockh, mode);
2476 ldlm_lock_decref(lockh, mode);
2481 static int osc_statfs_interpret(const struct lu_env *env,
2482 struct ptlrpc_request *req,
2483 struct osc_async_args *aa, int rc)
2485 struct obd_statfs *msfs;
2489 /* The request has in fact never been sent
2490 * due to issues at a higher level (LOV).
2491 * Exit immediately since the caller is
2492 * aware of the problem and takes care
2493 * of the clean up */
2496 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2497 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2503 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2505 GOTO(out, rc = -EPROTO);
2508 *aa->aa_oi->oi_osfs = *msfs;
2510 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2514 static int osc_statfs_async(struct obd_export *exp,
2515 struct obd_info *oinfo, __u64 max_age,
2516 struct ptlrpc_request_set *rqset)
2518 struct obd_device *obd = class_exp2obd(exp);
2519 struct ptlrpc_request *req;
2520 struct osc_async_args *aa;
2524 /* We could possibly pass max_age in the request (as an absolute
2525 * timestamp or a "seconds.usec ago") so the target can avoid doing
2526 * extra calls into the filesystem if that isn't necessary (e.g.
2527 * during mount that would help a bit). Having relative timestamps
2528 * is not so great if request processing is slow, while absolute
2529 * timestamps are not ideal because they need time synchronization. */
2530 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2534 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2536 ptlrpc_request_free(req);
2539 ptlrpc_request_set_replen(req);
2540 req->rq_request_portal = OST_CREATE_PORTAL;
2541 ptlrpc_at_set_req_timeout(req);
2543 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2544 /* procfs requests not want stat in wait for avoid deadlock */
2545 req->rq_no_resend = 1;
2546 req->rq_no_delay = 1;
2549 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2550 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2551 aa = ptlrpc_req_async_args(req);
2554 ptlrpc_set_add_req(rqset, req);
2558 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2559 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2561 struct obd_device *obd = class_exp2obd(exp);
2562 struct obd_statfs *msfs;
2563 struct ptlrpc_request *req;
2564 struct obd_import *imp = NULL;
2568 /*Since the request might also come from lprocfs, so we need
2569 *sync this with client_disconnect_export Bug15684*/
2570 down_read(&obd->u.cli.cl_sem);
2571 if (obd->u.cli.cl_import)
2572 imp = class_import_get(obd->u.cli.cl_import);
2573 up_read(&obd->u.cli.cl_sem);
2577 /* We could possibly pass max_age in the request (as an absolute
2578 * timestamp or a "seconds.usec ago") so the target can avoid doing
2579 * extra calls into the filesystem if that isn't necessary (e.g.
2580 * during mount that would help a bit). Having relative timestamps
2581 * is not so great if request processing is slow, while absolute
2582 * timestamps are not ideal because they need time synchronization. */
2583 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2585 class_import_put(imp);
2590 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2592 ptlrpc_request_free(req);
2595 ptlrpc_request_set_replen(req);
2596 req->rq_request_portal = OST_CREATE_PORTAL;
2597 ptlrpc_at_set_req_timeout(req);
2599 if (flags & OBD_STATFS_NODELAY) {
2600 /* procfs requests not want stat in wait for avoid deadlock */
2601 req->rq_no_resend = 1;
2602 req->rq_no_delay = 1;
2605 rc = ptlrpc_queue_wait(req);
2609 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2611 GOTO(out, rc = -EPROTO);
2618 ptlrpc_req_finished(req);
2622 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2623 void *karg, void *uarg)
2625 struct obd_device *obd = exp->exp_obd;
2626 struct obd_ioctl_data *data = karg;
2630 if (!try_module_get(THIS_MODULE)) {
2631 CERROR("Can't get module. Is it alive?");
2635 case OBD_IOC_CLIENT_RECOVER:
2636 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2637 data->ioc_inlbuf1, 0);
2641 case IOC_OSC_SET_ACTIVE:
2642 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2645 case OBD_IOC_POLL_QUOTACHECK:
2646 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2648 case OBD_IOC_PING_TARGET:
2649 err = ptlrpc_obd_ping(obd);
2652 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2653 cmd, current_comm());
2654 GOTO(out, err = -ENOTTY);
2657 module_put(THIS_MODULE);
2661 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2662 obd_count keylen, void *key, __u32 *vallen, void *val,
2663 struct lov_stripe_md *lsm)
2666 if (!vallen || !val)
2669 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2670 __u32 *stripe = val;
2671 *vallen = sizeof(*stripe);
2674 } else if (KEY_IS(KEY_LAST_ID)) {
2675 struct ptlrpc_request *req;
2680 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2681 &RQF_OST_GET_INFO_LAST_ID);
2685 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2686 RCL_CLIENT, keylen);
2687 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2689 ptlrpc_request_free(req);
2693 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2694 memcpy(tmp, key, keylen);
2696 req->rq_no_delay = req->rq_no_resend = 1;
2697 ptlrpc_request_set_replen(req);
2698 rc = ptlrpc_queue_wait(req);
2702 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2704 GOTO(out, rc = -EPROTO);
2706 *((obd_id *)val) = *reply;
2708 ptlrpc_req_finished(req);
2710 } else if (KEY_IS(KEY_FIEMAP)) {
2711 struct ll_fiemap_info_key *fm_key =
2712 (struct ll_fiemap_info_key *)key;
2713 struct ldlm_res_id res_id;
2714 ldlm_policy_data_t policy;
2715 struct lustre_handle lockh;
2716 ldlm_mode_t mode = 0;
2717 struct ptlrpc_request *req;
2718 struct ll_user_fiemap *reply;
2722 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2725 policy.l_extent.start = fm_key->fiemap.fm_start &
2728 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2729 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
2730 policy.l_extent.end = OBD_OBJECT_EOF;
2732 policy.l_extent.end = (fm_key->fiemap.fm_start +
2733 fm_key->fiemap.fm_length +
2734 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
2736 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2737 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2738 LDLM_FL_BLOCK_GRANTED |
2740 &res_id, LDLM_EXTENT, &policy,
2741 LCK_PR | LCK_PW, &lockh, 0);
2742 if (mode) { /* lock is cached on client */
2743 if (mode != LCK_PR) {
2744 ldlm_lock_addref(&lockh, LCK_PR);
2745 ldlm_lock_decref(&lockh, LCK_PW);
2747 } else { /* no cached lock, needs acquire lock on server side */
2748 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2749 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2753 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2754 &RQF_OST_GET_INFO_FIEMAP);
2756 GOTO(drop_lock, rc = -ENOMEM);
2758 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2759 RCL_CLIENT, keylen);
2760 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2761 RCL_CLIENT, *vallen);
2762 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2763 RCL_SERVER, *vallen);
2765 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2767 ptlrpc_request_free(req);
2768 GOTO(drop_lock, rc);
2771 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2772 memcpy(tmp, key, keylen);
2773 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2774 memcpy(tmp, val, *vallen);
2776 ptlrpc_request_set_replen(req);
2777 rc = ptlrpc_queue_wait(req);
2781 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2783 GOTO(fini_req, rc = -EPROTO);
2785 memcpy(val, reply, *vallen);
2787 ptlrpc_req_finished(req);
2790 ldlm_lock_decref(&lockh, LCK_PR);
2797 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2798 obd_count keylen, void *key, obd_count vallen,
2799 void *val, struct ptlrpc_request_set *set)
2801 struct ptlrpc_request *req;
2802 struct obd_device *obd = exp->exp_obd;
2803 struct obd_import *imp = class_exp2cliimp(exp);
2808 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2810 if (KEY_IS(KEY_CHECKSUM)) {
2811 if (vallen != sizeof(int))
2813 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2817 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2818 sptlrpc_conf_client_adapt(obd);
2822 if (KEY_IS(KEY_FLUSH_CTX)) {
2823 sptlrpc_import_flush_my_ctx(imp);
2827 if (KEY_IS(KEY_CACHE_SET)) {
2828 struct client_obd *cli = &obd->u.cli;
2830 LASSERT(cli->cl_cache == NULL); /* only once */
2831 cli->cl_cache = (struct cl_client_cache *)val;
2832 atomic_inc(&cli->cl_cache->ccc_users);
2833 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2835 /* add this osc into entity list */
2836 LASSERT(list_empty(&cli->cl_lru_osc));
2837 spin_lock(&cli->cl_cache->ccc_lru_lock);
2838 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2839 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2844 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2845 struct client_obd *cli = &obd->u.cli;
2846 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2847 int target = *(int *)val;
2849 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2854 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2857 /* We pass all other commands directly to OST. Since nobody calls osc
2858 methods directly and everybody is supposed to go through LOV, we
2859 assume lov checked invalid values for us.
2860 The only recognised values so far are evict_by_nid and mds_conn.
2861 Even if something bad goes through, we'd get a -EINVAL from OST
2864 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2865 &RQF_OST_SET_GRANT_INFO :
2870 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2871 RCL_CLIENT, keylen);
2872 if (!KEY_IS(KEY_GRANT_SHRINK))
2873 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2874 RCL_CLIENT, vallen);
2875 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2877 ptlrpc_request_free(req);
2881 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2882 memcpy(tmp, key, keylen);
2883 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2886 memcpy(tmp, val, vallen);
2888 if (KEY_IS(KEY_GRANT_SHRINK)) {
2889 struct osc_grant_args *aa;
2892 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2893 aa = ptlrpc_req_async_args(req);
2896 ptlrpc_req_finished(req);
2899 *oa = ((struct ost_body *)val)->oa;
2901 req->rq_interpret_reply = osc_shrink_grant_interpret;
2904 ptlrpc_request_set_replen(req);
2905 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2906 LASSERT(set != NULL);
2907 ptlrpc_set_add_req(set, req);
2908 ptlrpc_check_set(NULL, set);
2910 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2915 static int osc_reconnect(const struct lu_env *env,
2916 struct obd_export *exp, struct obd_device *obd,
2917 struct obd_uuid *cluuid,
2918 struct obd_connect_data *data,
2921 struct client_obd *cli = &obd->u.cli;
2923 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2926 client_obd_list_lock(&cli->cl_loi_list_lock);
2927 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2928 2 * cli_brw_size(obd);
2929 lost_grant = cli->cl_lost_grant;
2930 cli->cl_lost_grant = 0;
2931 client_obd_list_unlock(&cli->cl_loi_list_lock);
2933 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
2934 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2935 data->ocd_version, data->ocd_grant, lost_grant);
2941 static int osc_disconnect(struct obd_export *exp)
2943 struct obd_device *obd = class_exp2obd(exp);
2944 struct llog_ctxt *ctxt;
2947 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2949 if (obd->u.cli.cl_conn_count == 1) {
2950 /* Flush any remaining cancel messages out to the
2952 llog_sync(ctxt, exp, 0);
2954 llog_ctxt_put(ctxt);
2956 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
2960 rc = client_disconnect_export(exp);
2962 * Initially we put del_shrink_grant before disconnect_export, but it
2963 * causes the following problem if setup (connect) and cleanup
2964 * (disconnect) are tangled together.
2965 * connect p1 disconnect p2
2966 * ptlrpc_connect_import
2967 * ............... class_manual_cleanup
2970 * ptlrpc_connect_interrupt
2972 * add this client to shrink list
2974 * Bang! pinger trigger the shrink.
2975 * So the osc should be disconnected from the shrink list, after we
2976 * are sure the import has been destroyed. BUG18662
2978 if (obd->u.cli.cl_import == NULL)
2979 osc_del_shrink_grant(&obd->u.cli);
2983 static int osc_import_event(struct obd_device *obd,
2984 struct obd_import *imp,
2985 enum obd_import_event event)
2987 struct client_obd *cli;
2991 LASSERT(imp->imp_obd == obd);
2994 case IMP_EVENT_DISCON: {
2996 client_obd_list_lock(&cli->cl_loi_list_lock);
2997 cli->cl_avail_grant = 0;
2998 cli->cl_lost_grant = 0;
2999 client_obd_list_unlock(&cli->cl_loi_list_lock);
3002 case IMP_EVENT_INACTIVE: {
3003 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3006 case IMP_EVENT_INVALIDATE: {
3007 struct ldlm_namespace *ns = obd->obd_namespace;
3011 env = cl_env_get(&refcheck);
3015 /* all pages go to failing rpcs due to the invalid
3017 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3019 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3020 cl_env_put(env, &refcheck);
3025 case IMP_EVENT_ACTIVE: {
3026 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3029 case IMP_EVENT_OCD: {
3030 struct obd_connect_data *ocd = &imp->imp_connect_data;
3032 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3033 osc_init_grant(&obd->u.cli, ocd);
3036 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3037 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3039 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3042 case IMP_EVENT_DEACTIVATE: {
3043 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3046 case IMP_EVENT_ACTIVATE: {
3047 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3051 CERROR("Unknown import event %d\n", event);
3058 * Determine whether the lock can be canceled before replaying the lock
3059 * during recovery, see bug16774 for detailed information.
3061 * \retval zero the lock can't be canceled
3062 * \retval other ok to cancel
3064 static int osc_cancel_weight(struct ldlm_lock *lock)
3067 * Cancel all unused and granted extent lock.
3069 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3070 lock->l_granted_mode == lock->l_req_mode &&
3071 osc_ldlm_weigh_ast(lock) == 0)
3077 static int brw_queue_work(const struct lu_env *env, void *data)
3079 struct client_obd *cli = data;
3081 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3083 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3087 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3089 struct client_obd *cli = &obd->u.cli;
3090 struct obd_type *type;
3095 rc = ptlrpcd_addref();
3099 rc = client_obd_setup(obd, lcfg);
3101 GOTO(out_ptlrpcd, rc);
3103 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3104 if (IS_ERR(handler))
3105 GOTO(out_client_setup, rc = PTR_ERR(handler));
3106 cli->cl_writeback_work = handler;
3108 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3109 if (IS_ERR(handler))
3110 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3111 cli->cl_lru_work = handler;
3113 rc = osc_quota_setup(obd);
3115 GOTO(out_ptlrpcd_work, rc);
3117 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3120 obd->obd_vars = lprocfs_osc_obd_vars;
3122 /* If this is true then both client (osc) and server (osp) are on the
3123 * same node. The osp layer if loaded first will register the osc proc
3124 * directory. In that case this obd_device will be attached its proc
3125 * tree to type->typ_procsym instead of obd->obd_type->typ_procroot. */
3126 type = class_search_type(LUSTRE_OSP_NAME);
3127 if (type && type->typ_procsym) {
3128 obd->obd_proc_entry = lprocfs_seq_register(obd->obd_name,
3130 obd->obd_vars, obd);
3131 if (IS_ERR(obd->obd_proc_entry)) {
3132 rc = PTR_ERR(obd->obd_proc_entry);
3133 CERROR("error %d setting up lprocfs for %s\n", rc,
3135 obd->obd_proc_entry = NULL;
3138 rc = lprocfs_seq_obd_setup(obd);
3141 /* If the basic OSC proc tree construction succeeded then
3142 * lets do the rest. */
3144 lproc_osc_attach_seqstat(obd);
3145 sptlrpc_lprocfs_cliobd_attach(obd);
3146 ptlrpc_lprocfs_register_obd(obd);
3149 /* We need to allocate a few requests more, because
3150 * brw_interpret tries to create new requests before freeing
3151 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3152 * reserved, but I'm afraid that might be too much wasted RAM
3153 * in fact, so 2 is just my guess and still should work. */
3154 cli->cl_import->imp_rq_pool =
3155 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3157 ptlrpc_add_rqs_to_pool);
3159 INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3160 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3164 if (cli->cl_writeback_work != NULL) {
3165 ptlrpcd_destroy_work(cli->cl_writeback_work);
3166 cli->cl_writeback_work = NULL;
3168 if (cli->cl_lru_work != NULL) {
3169 ptlrpcd_destroy_work(cli->cl_lru_work);
3170 cli->cl_lru_work = NULL;
3173 client_obd_cleanup(obd);
3179 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3185 case OBD_CLEANUP_EARLY: {
3186 struct obd_import *imp;
3187 imp = obd->u.cli.cl_import;
3188 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3189 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3190 ptlrpc_deactivate_import(imp);
3191 spin_lock(&imp->imp_lock);
3192 imp->imp_pingable = 0;
3193 spin_unlock(&imp->imp_lock);
3196 case OBD_CLEANUP_EXPORTS: {
3197 struct client_obd *cli = &obd->u.cli;
3199 * for echo client, export may be on zombie list, wait for
3200 * zombie thread to cull it, because cli.cl_import will be
3201 * cleared in client_disconnect_export():
3202 * class_export_destroy() -> obd_cleanup() ->
3203 * echo_device_free() -> echo_client_cleanup() ->
3204 * obd_disconnect() -> osc_disconnect() ->
3205 * client_disconnect_export()
3207 obd_zombie_barrier();
3208 if (cli->cl_writeback_work) {
3209 ptlrpcd_destroy_work(cli->cl_writeback_work);
3210 cli->cl_writeback_work = NULL;
3212 if (cli->cl_lru_work) {
3213 ptlrpcd_destroy_work(cli->cl_lru_work);
3214 cli->cl_lru_work = NULL;
3216 obd_cleanup_client_import(obd);
3217 ptlrpc_lprocfs_unregister_obd(obd);
3218 lprocfs_obd_cleanup(obd);
3219 rc = obd_llog_finish(obd, 0);
3221 CERROR("failed to cleanup llogging subsystems\n");
3228 int osc_cleanup(struct obd_device *obd)
3230 struct client_obd *cli = &obd->u.cli;
3236 if (cli->cl_cache != NULL) {
3237 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3238 spin_lock(&cli->cl_cache->ccc_lru_lock);
3239 list_del_init(&cli->cl_lru_osc);
3240 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3241 cli->cl_lru_left = NULL;
3242 atomic_dec(&cli->cl_cache->ccc_users);
3243 cli->cl_cache = NULL;
3246 /* free memory of osc quota cache */
3247 osc_quota_cleanup(obd);
3249 rc = client_obd_cleanup(obd);
3255 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3257 int rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
3259 return rc > 0 ? 0: rc;
3262 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3264 return osc_process_config_base(obd, buf);
3267 struct obd_ops osc_obd_ops = {
3268 .o_owner = THIS_MODULE,
3269 .o_setup = osc_setup,
3270 .o_precleanup = osc_precleanup,
3271 .o_cleanup = osc_cleanup,
3272 .o_add_conn = client_import_add_conn,
3273 .o_del_conn = client_import_del_conn,
3274 .o_connect = client_connect_import,
3275 .o_reconnect = osc_reconnect,
3276 .o_disconnect = osc_disconnect,
3277 .o_statfs = osc_statfs,
3278 .o_statfs_async = osc_statfs_async,
3279 .o_unpackmd = osc_unpackmd,
3280 .o_create = osc_create,
3281 .o_destroy = osc_destroy,
3282 .o_getattr = osc_getattr,
3283 .o_getattr_async = osc_getattr_async,
3284 .o_setattr = osc_setattr,
3285 .o_setattr_async = osc_setattr_async,
3286 .o_change_cbdata = osc_change_cbdata,
3287 .o_find_cbdata = osc_find_cbdata,
3288 .o_iocontrol = osc_iocontrol,
3289 .o_get_info = osc_get_info,
3290 .o_set_info_async = osc_set_info_async,
3291 .o_import_event = osc_import_event,
3292 .o_process_config = osc_process_config,
3293 .o_quotactl = osc_quotactl,
3294 .o_quotacheck = osc_quotacheck,
3297 extern struct lu_kmem_descr osc_caches[];
3298 extern spinlock_t osc_ast_guard;
3299 extern struct lock_class_key osc_ast_guard_class;
3301 int __init osc_init(void)
3303 bool enable_proc = true;
3304 struct obd_type *type;
3308 /* print an address of _any_ initialized kernel symbol from this
3309 * module, to allow debugging with gdb that doesn't support data
3310 * symbols from modules.*/
3311 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3313 rc = lu_kmem_init(osc_caches);
3317 type = class_search_type(LUSTRE_OSP_NAME);
3318 if (type != NULL && type->typ_procsym != NULL)
3319 enable_proc = false;
3321 rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
3322 #ifndef HAVE_ONLY_PROCFS_SEQ
3325 LUSTRE_OSC_NAME, &osc_device_type);
3327 lu_kmem_fini(osc_caches);
3331 spin_lock_init(&osc_ast_guard);
3332 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3338 static void /*__exit*/ osc_exit(void)
3340 class_unregister_type(LUSTRE_OSC_NAME);
3341 lu_kmem_fini(osc_caches);
3344 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3345 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3346 MODULE_LICENSE("GPL");
3348 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);