1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_OSC
42 #include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71 struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76 struct lov_stripe_md *lsm)
81 lmm_size = sizeof(**lmmp);
86 OBD_FREE(*lmmp, lmm_size);
92 OBD_ALLOC(*lmmp, lmm_size);
98 LASSERT(lsm->lsm_object_id);
99 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109 struct lov_mds_md *lmm, int lmm_bytes)
115 if (lmm_bytes < sizeof (*lmm)) {
116 CERROR("lov_mds_md too small: %d, need %d\n",
117 lmm_bytes, (int)sizeof(*lmm));
120 /* XXX LOV_MAGIC etc check? */
122 if (lmm->lmm_object_id == 0) {
123 CERROR("lov_mds_md: zero lmm_object_id\n");
128 lsm_size = lov_stripe_md_size(1);
132 if (*lsmp != NULL && lmm == NULL) {
133 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134 OBD_FREE(*lsmp, lsm_size);
140 OBD_ALLOC(*lsmp, lsm_size);
143 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145 OBD_FREE(*lsmp, lsm_size);
148 loi_init((*lsmp)->lsm_oinfo[0]);
152 /* XXX zero *lsmp? */
153 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
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(struct obd_export *exp, struct obd_info *oinfo)
265 struct ptlrpc_request *req;
266 struct ost_body *body;
270 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
274 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
277 ptlrpc_request_free(req);
281 osc_pack_req_body(req, oinfo);
283 ptlrpc_request_set_replen(req);
285 rc = ptlrpc_queue_wait(req);
289 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
291 GOTO(out, rc = -EPROTO);
293 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
296 /* This should really be sent by the OST */
297 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
302 ptlrpc_req_finished(req);
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307 struct obd_trans_info *oti)
309 struct ptlrpc_request *req;
310 struct ost_body *body;
314 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
316 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
320 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
321 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
323 ptlrpc_request_free(req);
327 osc_pack_req_body(req, oinfo);
329 ptlrpc_request_set_replen(req);
331 rc = ptlrpc_queue_wait(req);
335 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
337 GOTO(out, rc = -EPROTO);
339 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
343 ptlrpc_req_finished(req);
347 static int osc_setattr_interpret(const struct lu_env *env,
348 struct ptlrpc_request *req,
349 struct osc_setattr_args *sa, int rc)
351 struct ost_body *body;
357 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
359 GOTO(out, rc = -EPROTO);
361 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
363 rc = sa->sa_upcall(sa->sa_cookie, rc);
367 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
368 struct obd_trans_info *oti,
369 obd_enqueue_update_f upcall, void *cookie,
370 struct ptlrpc_request_set *rqset)
372 struct ptlrpc_request *req;
373 struct osc_setattr_args *sa;
377 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
381 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
382 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
384 ptlrpc_request_free(req);
388 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
389 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
391 osc_pack_req_body(req, oinfo);
393 ptlrpc_request_set_replen(req);
395 /* do mds to ost setattr asynchronously */
397 /* Do not wait for response. */
398 ptlrpcd_add_req(req, PSCOPE_OTHER);
400 req->rq_interpret_reply =
401 (ptlrpc_interpterer_t)osc_setattr_interpret;
403 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
404 sa = ptlrpc_req_async_args(req);
405 sa->sa_oa = oinfo->oi_oa;
406 sa->sa_upcall = upcall;
407 sa->sa_cookie = cookie;
409 if (rqset == PTLRPCD_SET)
410 ptlrpcd_add_req(req, PSCOPE_OTHER);
412 ptlrpc_set_add_req(rqset, req);
418 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
419 struct obd_trans_info *oti,
420 struct ptlrpc_request_set *rqset)
422 return osc_setattr_async_base(exp, oinfo, oti,
423 oinfo->oi_cb_up, oinfo, rqset);
426 int osc_real_create(struct obd_export *exp, struct obdo *oa,
427 struct lov_stripe_md **ea, struct obd_trans_info *oti)
429 struct ptlrpc_request *req;
430 struct ost_body *body;
431 struct lov_stripe_md *lsm;
440 rc = obd_alloc_memmd(exp, &lsm);
445 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
447 GOTO(out, rc = -ENOMEM);
449 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
451 ptlrpc_request_free(req);
455 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
457 lustre_set_wire_obdo(&body->oa, oa);
459 ptlrpc_request_set_replen(req);
461 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462 oa->o_flags == OBD_FL_DELORPHAN) {
464 "delorphan from OST integration");
465 /* Don't resend the delorphan req */
466 req->rq_no_resend = req->rq_no_delay = 1;
469 rc = ptlrpc_queue_wait(req);
473 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
475 GOTO(out_req, rc = -EPROTO);
477 lustre_get_wire_obdo(oa, &body->oa);
479 /* This should really be sent by the OST */
480 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
481 oa->o_valid |= OBD_MD_FLBLKSZ;
483 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
484 * have valid lsm_oinfo data structs, so don't go touching that.
485 * This needs to be fixed in a big way.
487 lsm->lsm_object_id = oa->o_id;
488 lsm->lsm_object_seq = oa->o_seq;
492 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
494 if (oa->o_valid & OBD_MD_FLCOOKIE) {
495 if (!oti->oti_logcookies)
496 oti_alloc_cookies(oti, 1);
497 *oti->oti_logcookies = oa->o_lcookie;
501 CDEBUG(D_HA, "transno: "LPD64"\n",
502 lustre_msg_get_transno(req->rq_repmsg));
504 ptlrpc_req_finished(req);
507 obd_free_memmd(exp, &lsm);
511 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
512 obd_enqueue_update_f upcall, void *cookie,
513 struct ptlrpc_request_set *rqset)
515 struct ptlrpc_request *req;
516 struct osc_setattr_args *sa;
517 struct ost_body *body;
521 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
525 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
526 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
528 ptlrpc_request_free(req);
531 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
532 ptlrpc_at_set_req_timeout(req);
534 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
536 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
537 osc_pack_capa(req, body, oinfo->oi_capa);
539 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, PSCOPE_OTHER);
551 ptlrpc_set_add_req(rqset, req);
556 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
557 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(struct obd_export *exp, struct obdo *oa,
568 struct lov_stripe_md *md, obd_size start, obd_size end,
571 struct ptlrpc_request *req;
572 struct ost_body *body;
577 CDEBUG(D_INFO, "oa NULL\n");
581 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
585 osc_set_capa_size(req, &RMF_CAPA1, capa);
586 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
588 ptlrpc_request_free(req);
592 /* overload the size and blocks fields in the oa with start/end */
593 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
595 lustre_set_wire_obdo(&body->oa, oa);
596 body->oa.o_size = start;
597 body->oa.o_blocks = end;
598 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
599 osc_pack_capa(req, body, capa);
601 ptlrpc_request_set_replen(req);
603 rc = ptlrpc_queue_wait(req);
607 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
609 GOTO(out, rc = -EPROTO);
611 lustre_get_wire_obdo(oa, &body->oa);
615 ptlrpc_req_finished(req);
619 /* Find and cancel locally locks matched by @mode in the resource found by
620 * @objid. Found locks are added into @cancel list. Returns the amount of
621 * locks added to @cancels list. */
622 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
624 ldlm_mode_t mode, int lock_flags)
626 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
627 struct ldlm_res_id res_id;
628 struct ldlm_resource *res;
632 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
633 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
637 LDLM_RESOURCE_ADDREF(res);
638 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
639 lock_flags, 0, NULL);
640 LDLM_RESOURCE_DELREF(res);
641 ldlm_resource_putref(res);
645 static int osc_destroy_interpret(const struct lu_env *env,
646 struct ptlrpc_request *req, void *data,
649 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
651 cfs_atomic_dec(&cli->cl_destroy_in_flight);
652 cfs_waitq_signal(&cli->cl_destroy_waitq);
656 static int osc_can_send_destroy(struct client_obd *cli)
658 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
659 cli->cl_max_rpcs_in_flight) {
660 /* The destroy request can be sent */
663 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
664 cli->cl_max_rpcs_in_flight) {
666 * The counter has been modified between the two atomic
669 cfs_waitq_signal(&cli->cl_destroy_waitq);
674 /* Destroy requests can be async always on the client, and we don't even really
675 * care about the return code since the client cannot do anything at all about
677 * When the MDS is unlinking a filename, it saves the file objects into a
678 * recovery llog, and these object records are cancelled when the OST reports
679 * they were destroyed and sync'd to disk (i.e. transaction committed).
680 * If the client dies, or the OST is down when the object should be destroyed,
681 * the records are not cancelled, and when the OST reconnects to the MDS next,
682 * it will retrieve the llog unlink logs and then sends the log cancellation
683 * cookies to the MDS after committing destroy transactions. */
684 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
685 struct lov_stripe_md *ea, struct obd_trans_info *oti,
686 struct obd_export *md_export, void *capa)
688 struct client_obd *cli = &exp->exp_obd->u.cli;
689 struct ptlrpc_request *req;
690 struct ost_body *body;
691 CFS_LIST_HEAD(cancels);
696 CDEBUG(D_INFO, "oa NULL\n");
700 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
701 LDLM_FL_DISCARD_DATA);
703 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
705 ldlm_lock_list_put(&cancels, l_bl_ast, count);
709 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
710 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
713 ptlrpc_request_free(req);
717 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
718 ptlrpc_at_set_req_timeout(req);
720 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
721 oa->o_lcookie = *oti->oti_logcookies;
722 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
724 lustre_set_wire_obdo(&body->oa, oa);
726 osc_pack_capa(req, body, (struct obd_capa *)capa);
727 ptlrpc_request_set_replen(req);
729 /* don't throttle destroy RPCs for the MDT */
730 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
731 req->rq_interpret_reply = osc_destroy_interpret;
732 if (!osc_can_send_destroy(cli)) {
733 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
737 * Wait until the number of on-going destroy RPCs drops
738 * under max_rpc_in_flight
740 l_wait_event_exclusive(cli->cl_destroy_waitq,
741 osc_can_send_destroy(cli), &lwi);
745 /* Do not wait for response */
746 ptlrpcd_add_req(req, PSCOPE_OTHER);
750 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
753 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
755 LASSERT(!(oa->o_valid & bits));
758 client_obd_list_lock(&cli->cl_loi_list_lock);
759 oa->o_dirty = cli->cl_dirty;
760 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
761 CERROR("dirty %lu - %lu > dirty_max %lu\n",
762 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
764 } else if (cfs_atomic_read(&obd_dirty_pages) -
765 cfs_atomic_read(&obd_dirty_transit_pages) >
766 obd_max_dirty_pages + 1){
767 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
768 * not covered by a lock thus they may safely race and trip
769 * this CERROR() unless we add in a small fudge factor (+1). */
770 CERROR("dirty %d - %d > system dirty_max %d\n",
771 cfs_atomic_read(&obd_dirty_pages),
772 cfs_atomic_read(&obd_dirty_transit_pages),
773 obd_max_dirty_pages);
775 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
776 CERROR("dirty %lu - dirty_max %lu too big???\n",
777 cli->cl_dirty, cli->cl_dirty_max);
780 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
781 (cli->cl_max_rpcs_in_flight + 1);
782 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
784 oa->o_grant = cli->cl_avail_grant;
785 oa->o_dropped = cli->cl_lost_grant;
786 cli->cl_lost_grant = 0;
787 client_obd_list_unlock(&cli->cl_loi_list_lock);
788 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
789 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
793 static void osc_update_next_shrink(struct client_obd *cli)
795 cli->cl_next_shrink_grant =
796 cfs_time_shift(cli->cl_grant_shrink_interval);
797 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
798 cli->cl_next_shrink_grant);
801 /* caller must hold loi_list_lock */
802 static void osc_consume_write_grant(struct client_obd *cli,
803 struct brw_page *pga)
805 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
806 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
807 cfs_atomic_inc(&obd_dirty_pages);
808 cli->cl_dirty += CFS_PAGE_SIZE;
809 cli->cl_avail_grant -= CFS_PAGE_SIZE;
810 pga->flag |= OBD_BRW_FROM_GRANT;
811 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
812 CFS_PAGE_SIZE, pga, pga->pg);
813 LASSERT(cli->cl_avail_grant >= 0);
814 osc_update_next_shrink(cli);
817 /* the companion to osc_consume_write_grant, called when a brw has completed.
818 * must be called with the loi lock held. */
819 static void osc_release_write_grant(struct client_obd *cli,
820 struct brw_page *pga, int sent)
822 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
825 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
826 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
831 pga->flag &= ~OBD_BRW_FROM_GRANT;
832 cfs_atomic_dec(&obd_dirty_pages);
833 cli->cl_dirty -= CFS_PAGE_SIZE;
834 if (pga->flag & OBD_BRW_NOCACHE) {
835 pga->flag &= ~OBD_BRW_NOCACHE;
836 cfs_atomic_dec(&obd_dirty_transit_pages);
837 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
840 cli->cl_lost_grant += CFS_PAGE_SIZE;
841 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
842 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
843 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
844 /* For short writes we shouldn't count parts of pages that
845 * span a whole block on the OST side, or our accounting goes
846 * wrong. Should match the code in filter_grant_check. */
847 int offset = pga->off & ~CFS_PAGE_MASK;
848 int count = pga->count + (offset & (blocksize - 1));
849 int end = (offset + pga->count) & (blocksize - 1);
851 count += blocksize - end;
853 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
854 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
855 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
856 cli->cl_avail_grant, cli->cl_dirty);
862 static unsigned long rpcs_in_flight(struct client_obd *cli)
864 return cli->cl_r_in_flight + cli->cl_w_in_flight;
867 /* caller must hold loi_list_lock */
868 void osc_wake_cache_waiters(struct client_obd *cli)
871 struct osc_cache_waiter *ocw;
874 cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
875 /* if we can't dirty more, we must wait until some is written */
876 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
877 (cfs_atomic_read(&obd_dirty_pages) + 1 >
878 obd_max_dirty_pages)) {
879 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
880 "osc max %ld, sys max %d\n", cli->cl_dirty,
881 cli->cl_dirty_max, obd_max_dirty_pages);
885 /* if still dirty cache but no grant wait for pending RPCs that
886 * may yet return us some grant before doing sync writes */
887 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
888 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
889 cli->cl_w_in_flight);
893 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
894 cfs_list_del_init(&ocw->ocw_entry);
895 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
896 /* no more RPCs in flight to return grant, do sync IO */
897 ocw->ocw_rc = -EDQUOT;
898 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
900 osc_consume_write_grant(cli,
901 &ocw->ocw_oap->oap_brw_page);
904 cfs_waitq_signal(&ocw->ocw_waitq);
910 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
912 client_obd_list_lock(&cli->cl_loi_list_lock);
913 cli->cl_avail_grant += grant;
914 client_obd_list_unlock(&cli->cl_loi_list_lock);
917 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
919 if (body->oa.o_valid & OBD_MD_FLGRANT) {
920 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
921 __osc_update_grant(cli, body->oa.o_grant);
925 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
926 void *key, obd_count vallen, void *val,
927 struct ptlrpc_request_set *set);
929 static int osc_shrink_grant_interpret(const struct lu_env *env,
930 struct ptlrpc_request *req,
933 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
934 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
935 struct ost_body *body;
938 __osc_update_grant(cli, oa->o_grant);
942 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
944 osc_update_grant(cli, body);
950 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
952 client_obd_list_lock(&cli->cl_loi_list_lock);
953 oa->o_grant = cli->cl_avail_grant / 4;
954 cli->cl_avail_grant -= oa->o_grant;
955 client_obd_list_unlock(&cli->cl_loi_list_lock);
956 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
957 oa->o_valid |= OBD_MD_FLFLAGS;
960 oa->o_flags |= OBD_FL_SHRINK_GRANT;
961 osc_update_next_shrink(cli);
964 /* Shrink the current grant, either from some large amount to enough for a
965 * full set of in-flight RPCs, or if we have already shrunk to that limit
966 * then to enough for a single RPC. This avoids keeping more grant than
967 * needed, and avoids shrinking the grant piecemeal. */
968 static int osc_shrink_grant(struct client_obd *cli)
970 long target = (cli->cl_max_rpcs_in_flight + 1) *
971 cli->cl_max_pages_per_rpc;
973 client_obd_list_lock(&cli->cl_loi_list_lock);
974 if (cli->cl_avail_grant <= target)
975 target = cli->cl_max_pages_per_rpc;
976 client_obd_list_unlock(&cli->cl_loi_list_lock);
978 return osc_shrink_grant_to_target(cli, target);
981 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
984 struct ost_body *body;
987 client_obd_list_lock(&cli->cl_loi_list_lock);
988 /* Don't shrink if we are already above or below the desired limit
989 * We don't want to shrink below a single RPC, as that will negatively
990 * impact block allocation and long-term performance. */
991 if (target < cli->cl_max_pages_per_rpc)
992 target = cli->cl_max_pages_per_rpc;
994 if (target >= cli->cl_avail_grant) {
995 client_obd_list_unlock(&cli->cl_loi_list_lock);
998 client_obd_list_unlock(&cli->cl_loi_list_lock);
1000 OBD_ALLOC_PTR(body);
1004 osc_announce_cached(cli, &body->oa, 0);
1006 client_obd_list_lock(&cli->cl_loi_list_lock);
1007 body->oa.o_grant = cli->cl_avail_grant - target;
1008 cli->cl_avail_grant = target;
1009 client_obd_list_unlock(&cli->cl_loi_list_lock);
1010 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1011 body->oa.o_valid |= OBD_MD_FLFLAGS;
1012 body->oa.o_flags = 0;
1014 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1015 osc_update_next_shrink(cli);
1017 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1018 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1019 sizeof(*body), body, NULL);
1021 __osc_update_grant(cli, body->oa.o_grant);
1026 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1027 static int osc_should_shrink_grant(struct client_obd *client)
1029 cfs_time_t time = cfs_time_current();
1030 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1032 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1033 OBD_CONNECT_GRANT_SHRINK) == 0)
1036 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1037 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1038 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1041 osc_update_next_shrink(client);
1046 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1048 struct client_obd *client;
1050 cfs_list_for_each_entry(client, &item->ti_obd_list,
1051 cl_grant_shrink_list) {
1052 if (osc_should_shrink_grant(client))
1053 osc_shrink_grant(client);
1058 static int osc_add_shrink_grant(struct client_obd *client)
1062 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1064 osc_grant_shrink_grant_cb, NULL,
1065 &client->cl_grant_shrink_list);
1067 CERROR("add grant client %s error %d\n",
1068 client->cl_import->imp_obd->obd_name, rc);
1071 CDEBUG(D_CACHE, "add grant client %s \n",
1072 client->cl_import->imp_obd->obd_name);
1073 osc_update_next_shrink(client);
1077 static int osc_del_shrink_grant(struct client_obd *client)
1079 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1083 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1086 * ocd_grant is the total grant amount we're expect to hold: if we've
1087 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1088 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1090 * race is tolerable here: if we're evicted, but imp_state already
1091 * left EVICTED state, then cl_dirty must be 0 already.
1093 client_obd_list_lock(&cli->cl_loi_list_lock);
1094 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1095 cli->cl_avail_grant = ocd->ocd_grant;
1097 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1099 if (cli->cl_avail_grant < 0) {
1100 CWARN("%s: available grant < 0, the OSS is probably not running"
1101 " with patch from bug20278 (%ld) \n",
1102 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1103 /* workaround for 1.6 servers which do not have
1104 * the patch from bug20278 */
1105 cli->cl_avail_grant = ocd->ocd_grant;
1108 client_obd_list_unlock(&cli->cl_loi_list_lock);
1110 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1111 cli->cl_import->imp_obd->obd_name,
1112 cli->cl_avail_grant, cli->cl_lost_grant);
1114 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1115 cfs_list_empty(&cli->cl_grant_shrink_list))
1116 osc_add_shrink_grant(cli);
1119 /* We assume that the reason this OSC got a short read is because it read
1120 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1121 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1122 * this stripe never got written at or beyond this stripe offset yet. */
1123 static void handle_short_read(int nob_read, obd_count page_count,
1124 struct brw_page **pga)
1129 /* skip bytes read OK */
1130 while (nob_read > 0) {
1131 LASSERT (page_count > 0);
1133 if (pga[i]->count > nob_read) {
1134 /* EOF inside this page */
1135 ptr = cfs_kmap(pga[i]->pg) +
1136 (pga[i]->off & ~CFS_PAGE_MASK);
1137 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1138 cfs_kunmap(pga[i]->pg);
1144 nob_read -= pga[i]->count;
1149 /* zero remaining pages */
1150 while (page_count-- > 0) {
1151 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1152 memset(ptr, 0, pga[i]->count);
1153 cfs_kunmap(pga[i]->pg);
1158 static int check_write_rcs(struct ptlrpc_request *req,
1159 int requested_nob, int niocount,
1160 obd_count page_count, struct brw_page **pga)
1165 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1166 sizeof(*remote_rcs) *
1168 if (remote_rcs == NULL) {
1169 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1173 /* return error if any niobuf was in error */
1174 for (i = 0; i < niocount; i++) {
1175 if (remote_rcs[i] < 0)
1176 return(remote_rcs[i]);
1178 if (remote_rcs[i] != 0) {
1179 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1180 i, remote_rcs[i], req);
1185 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1186 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1187 req->rq_bulk->bd_nob_transferred, requested_nob);
1194 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1196 if (p1->flag != p2->flag) {
1197 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1198 OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
1200 /* warn if we try to combine flags that we don't know to be
1201 * safe to combine */
1202 if ((p1->flag & mask) != (p2->flag & mask))
1203 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1204 "same brw?\n", p1->flag, p2->flag);
1208 return (p1->off + p1->count == p2->off);
1211 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1212 struct brw_page **pga, int opc,
1213 cksum_type_t cksum_type)
1218 LASSERT (pg_count > 0);
1219 cksum = init_checksum(cksum_type);
1220 while (nob > 0 && pg_count > 0) {
1221 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1222 int off = pga[i]->off & ~CFS_PAGE_MASK;
1223 int count = pga[i]->count > nob ? nob : pga[i]->count;
1225 /* corrupt the data before we compute the checksum, to
1226 * simulate an OST->client data error */
1227 if (i == 0 && opc == OST_READ &&
1228 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1229 memcpy(ptr + off, "bad1", min(4, nob));
1230 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1231 cfs_kunmap(pga[i]->pg);
1232 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1235 nob -= pga[i]->count;
1239 /* For sending we only compute the wrong checksum instead
1240 * of corrupting the data so it is still correct on a redo */
1241 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1247 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1248 struct lov_stripe_md *lsm, obd_count page_count,
1249 struct brw_page **pga,
1250 struct ptlrpc_request **reqp,
1251 struct obd_capa *ocapa, int reserve,
1254 struct ptlrpc_request *req;
1255 struct ptlrpc_bulk_desc *desc;
1256 struct ost_body *body;
1257 struct obd_ioobj *ioobj;
1258 struct niobuf_remote *niobuf;
1259 int niocount, i, requested_nob, opc, rc;
1260 struct osc_brw_async_args *aa;
1261 struct req_capsule *pill;
1262 struct brw_page *pg_prev;
1265 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1266 RETURN(-ENOMEM); /* Recoverable */
1267 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1268 RETURN(-EINVAL); /* Fatal */
1270 if ((cmd & OBD_BRW_WRITE) != 0) {
1272 req = ptlrpc_request_alloc_pool(cli->cl_import,
1273 cli->cl_import->imp_rq_pool,
1274 &RQF_OST_BRW_WRITE);
1277 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1282 for (niocount = i = 1; i < page_count; i++) {
1283 if (!can_merge_pages(pga[i - 1], pga[i]))
1287 pill = &req->rq_pill;
1288 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1290 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1291 niocount * sizeof(*niobuf));
1292 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1294 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1296 ptlrpc_request_free(req);
1299 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1300 ptlrpc_at_set_req_timeout(req);
1302 if (opc == OST_WRITE)
1303 desc = ptlrpc_prep_bulk_imp(req, page_count,
1304 BULK_GET_SOURCE, OST_BULK_PORTAL);
1306 desc = ptlrpc_prep_bulk_imp(req, page_count,
1307 BULK_PUT_SINK, OST_BULK_PORTAL);
1310 GOTO(out, rc = -ENOMEM);
1311 /* NB request now owns desc and will free it when it gets freed */
1313 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1314 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1315 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1316 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1318 lustre_set_wire_obdo(&body->oa, oa);
1320 obdo_to_ioobj(oa, ioobj);
1321 ioobj->ioo_bufcnt = niocount;
1322 osc_pack_capa(req, body, ocapa);
1323 LASSERT (page_count > 0);
1325 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1326 struct brw_page *pg = pga[i];
1328 LASSERT(pg->count > 0);
1329 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1330 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1331 pg->off, pg->count);
1333 LASSERTF(i == 0 || pg->off > pg_prev->off,
1334 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1335 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1337 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1338 pg_prev->pg, page_private(pg_prev->pg),
1339 pg_prev->pg->index, pg_prev->off);
1341 LASSERTF(i == 0 || pg->off > pg_prev->off,
1342 "i %d p_c %u\n", i, page_count);
1344 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1345 (pg->flag & OBD_BRW_SRVLOCK));
1347 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1349 requested_nob += pg->count;
1351 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1353 niobuf->len += pg->count;
1355 niobuf->offset = pg->off;
1356 niobuf->len = pg->count;
1357 niobuf->flags = pg->flag;
1362 LASSERTF((void *)(niobuf - niocount) ==
1363 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1364 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1365 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1367 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1369 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1370 body->oa.o_valid |= OBD_MD_FLFLAGS;
1371 body->oa.o_flags = 0;
1373 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1376 if (osc_should_shrink_grant(cli))
1377 osc_shrink_grant_local(cli, &body->oa);
1379 /* size[REQ_REC_OFF] still sizeof (*body) */
1380 if (opc == OST_WRITE) {
1381 if (unlikely(cli->cl_checksum) &&
1382 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1383 /* store cl_cksum_type in a local variable since
1384 * it can be changed via lprocfs */
1385 cksum_type_t cksum_type = cli->cl_cksum_type;
1387 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1388 oa->o_flags &= OBD_FL_LOCAL_MASK;
1389 body->oa.o_flags = 0;
1391 body->oa.o_flags |= cksum_type_pack(cksum_type);
1392 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1393 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1397 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1399 /* save this in 'oa', too, for later checking */
1400 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1401 oa->o_flags |= cksum_type_pack(cksum_type);
1403 /* clear out the checksum flag, in case this is a
1404 * resend but cl_checksum is no longer set. b=11238 */
1405 oa->o_valid &= ~OBD_MD_FLCKSUM;
1407 oa->o_cksum = body->oa.o_cksum;
1408 /* 1 RC per niobuf */
1409 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1410 sizeof(__u32) * niocount);
1412 if (unlikely(cli->cl_checksum) &&
1413 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1414 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1415 body->oa.o_flags = 0;
1416 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1417 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1420 ptlrpc_request_set_replen(req);
1422 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1423 aa = ptlrpc_req_async_args(req);
1425 aa->aa_requested_nob = requested_nob;
1426 aa->aa_nio_count = niocount;
1427 aa->aa_page_count = page_count;
1431 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1432 if (ocapa && reserve)
1433 aa->aa_ocapa = capa_get(ocapa);
1439 ptlrpc_req_finished(req);
1443 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1444 __u32 client_cksum, __u32 server_cksum, int nob,
1445 obd_count page_count, struct brw_page **pga,
1446 cksum_type_t client_cksum_type)
1450 cksum_type_t cksum_type;
1452 if (server_cksum == client_cksum) {
1453 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1457 /* If this is mmaped file - it can be changed at any time */
1458 if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1461 if (oa->o_valid & OBD_MD_FLFLAGS)
1462 cksum_type = cksum_type_unpack(oa->o_flags);
1464 cksum_type = OBD_CKSUM_CRC32;
1466 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1469 if (cksum_type != client_cksum_type)
1470 msg = "the server did not use the checksum type specified in "
1471 "the original request - likely a protocol problem";
1472 else if (new_cksum == server_cksum)
1473 msg = "changed on the client after we checksummed it - "
1474 "likely false positive due to mmap IO (bug 11742)";
1475 else if (new_cksum == client_cksum)
1476 msg = "changed in transit before arrival at OST";
1478 msg = "changed in transit AND doesn't match the original - "
1479 "likely false positive due to mmap IO (bug 11742)";
1481 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1482 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1483 msg, libcfs_nid2str(peer->nid),
1484 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1485 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1486 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1488 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1490 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1491 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1492 "client csum now %x\n", client_cksum, client_cksum_type,
1493 server_cksum, cksum_type, new_cksum);
1497 /* Note rc enters this function as number of bytes transferred */
1498 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1500 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1501 const lnet_process_id_t *peer =
1502 &req->rq_import->imp_connection->c_peer;
1503 struct client_obd *cli = aa->aa_cli;
1504 struct ost_body *body;
1505 __u32 client_cksum = 0;
1508 if (rc < 0 && rc != -EDQUOT) {
1509 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1513 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1514 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1516 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1520 #ifdef HAVE_QUOTA_SUPPORT
1521 /* set/clear over quota flag for a uid/gid */
1522 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1523 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1524 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1526 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1527 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1529 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1534 osc_update_grant(cli, body);
1539 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1540 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1542 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1544 CERROR("Unexpected +ve rc %d\n", rc);
1547 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1549 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1552 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1553 check_write_checksum(&body->oa, peer, client_cksum,
1554 body->oa.o_cksum, aa->aa_requested_nob,
1555 aa->aa_page_count, aa->aa_ppga,
1556 cksum_type_unpack(aa->aa_oa->o_flags)))
1559 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1560 aa->aa_page_count, aa->aa_ppga);
1564 /* The rest of this function executes only for OST_READs */
1566 /* if unwrap_bulk failed, return -EAGAIN to retry */
1567 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1569 GOTO(out, rc = -EAGAIN);
1571 if (rc > aa->aa_requested_nob) {
1572 CERROR("Unexpected rc %d (%d requested)\n", rc,
1573 aa->aa_requested_nob);
1577 if (rc != req->rq_bulk->bd_nob_transferred) {
1578 CERROR ("Unexpected rc %d (%d transferred)\n",
1579 rc, req->rq_bulk->bd_nob_transferred);
1583 if (rc < aa->aa_requested_nob)
1584 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1586 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1587 static int cksum_counter;
1588 __u32 server_cksum = body->oa.o_cksum;
1591 cksum_type_t cksum_type;
1593 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1594 cksum_type = cksum_type_unpack(body->oa.o_flags);
1596 cksum_type = OBD_CKSUM_CRC32;
1597 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1598 aa->aa_ppga, OST_READ,
1601 if (peer->nid == req->rq_bulk->bd_sender) {
1605 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1608 if (server_cksum == ~0 && rc > 0) {
1609 CERROR("Protocol error: server %s set the 'checksum' "
1610 "bit, but didn't send a checksum. Not fatal, "
1611 "but please notify on http://bugzilla.lustre.org/\n",
1612 libcfs_nid2str(peer->nid));
1613 } else if (server_cksum != client_cksum) {
1614 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1615 "%s%s%s inode "DFID" object "
1616 LPU64"/"LPU64" extent "
1617 "["LPU64"-"LPU64"]\n",
1618 req->rq_import->imp_obd->obd_name,
1619 libcfs_nid2str(peer->nid),
1621 body->oa.o_valid & OBD_MD_FLFID ?
1622 body->oa.o_parent_seq : (__u64)0,
1623 body->oa.o_valid & OBD_MD_FLFID ?
1624 body->oa.o_parent_oid : 0,
1625 body->oa.o_valid & OBD_MD_FLFID ?
1626 body->oa.o_parent_ver : 0,
1628 body->oa.o_valid & OBD_MD_FLGROUP ?
1629 body->oa.o_seq : (__u64)0,
1630 aa->aa_ppga[0]->off,
1631 aa->aa_ppga[aa->aa_page_count-1]->off +
1632 aa->aa_ppga[aa->aa_page_count-1]->count -
1634 CERROR("client %x, server %x, cksum_type %x\n",
1635 client_cksum, server_cksum, cksum_type);
1637 aa->aa_oa->o_cksum = client_cksum;
1641 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1644 } else if (unlikely(client_cksum)) {
1645 static int cksum_missed;
1648 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1649 CERROR("Checksum %u requested from %s but not sent\n",
1650 cksum_missed, libcfs_nid2str(peer->nid));
1656 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1661 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1662 struct lov_stripe_md *lsm,
1663 obd_count page_count, struct brw_page **pga,
1664 struct obd_capa *ocapa)
1666 struct ptlrpc_request *req;
1670 struct l_wait_info lwi;
1674 cfs_waitq_init(&waitq);
1677 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1678 page_count, pga, &req, ocapa, 0, resends);
1682 rc = ptlrpc_queue_wait(req);
1684 if (rc == -ETIMEDOUT && req->rq_resend) {
1685 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1686 ptlrpc_req_finished(req);
1690 rc = osc_brw_fini_request(req, rc);
1692 ptlrpc_req_finished(req);
1693 if (osc_recoverable_error(rc)) {
1695 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1696 CERROR("too many resend retries, returning error\n");
1700 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1701 l_wait_event(waitq, 0, &lwi);
1709 int osc_brw_redo_request(struct ptlrpc_request *request,
1710 struct osc_brw_async_args *aa)
1712 struct ptlrpc_request *new_req;
1713 struct ptlrpc_request_set *set = request->rq_set;
1714 struct osc_brw_async_args *new_aa;
1715 struct osc_async_page *oap;
1719 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1720 CERROR("too many resent retries, returning error\n");
1724 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1726 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1727 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1728 aa->aa_cli, aa->aa_oa,
1729 NULL /* lsm unused by osc currently */,
1730 aa->aa_page_count, aa->aa_ppga,
1731 &new_req, aa->aa_ocapa, 0, 1);
1735 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1737 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1738 if (oap->oap_request != NULL) {
1739 LASSERTF(request == oap->oap_request,
1740 "request %p != oap_request %p\n",
1741 request, oap->oap_request);
1742 if (oap->oap_interrupted) {
1743 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1744 ptlrpc_req_finished(new_req);
1749 /* New request takes over pga and oaps from old request.
1750 * Note that copying a list_head doesn't work, need to move it... */
1752 new_req->rq_interpret_reply = request->rq_interpret_reply;
1753 new_req->rq_async_args = request->rq_async_args;
1754 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1756 new_aa = ptlrpc_req_async_args(new_req);
1758 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1759 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1760 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1762 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1763 if (oap->oap_request) {
1764 ptlrpc_req_finished(oap->oap_request);
1765 oap->oap_request = ptlrpc_request_addref(new_req);
1769 new_aa->aa_ocapa = aa->aa_ocapa;
1770 aa->aa_ocapa = NULL;
1772 /* use ptlrpc_set_add_req is safe because interpret functions work
1773 * in check_set context. only one way exist with access to request
1774 * from different thread got -EINTR - this way protected with
1775 * cl_loi_list_lock */
1776 ptlrpc_set_add_req(set, new_req);
1778 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1780 DEBUG_REQ(D_INFO, new_req, "new request");
1785 * ugh, we want disk allocation on the target to happen in offset order. we'll
1786 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1787 * fine for our small page arrays and doesn't require allocation. its an
1788 * insertion sort that swaps elements that are strides apart, shrinking the
1789 * stride down until its '1' and the array is sorted.
1791 static void sort_brw_pages(struct brw_page **array, int num)
1794 struct brw_page *tmp;
1798 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1803 for (i = stride ; i < num ; i++) {
1806 while (j >= stride && array[j - stride]->off > tmp->off) {
1807 array[j] = array[j - stride];
1812 } while (stride > 1);
1815 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1821 LASSERT (pages > 0);
1822 offset = pg[i]->off & ~CFS_PAGE_MASK;
1826 if (pages == 0) /* that's all */
1829 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1830 return count; /* doesn't end on page boundary */
1833 offset = pg[i]->off & ~CFS_PAGE_MASK;
1834 if (offset != 0) /* doesn't start on page boundary */
1841 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1843 struct brw_page **ppga;
1846 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1850 for (i = 0; i < count; i++)
1855 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1857 LASSERT(ppga != NULL);
1858 OBD_FREE(ppga, sizeof(*ppga) * count);
1861 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1862 obd_count page_count, struct brw_page *pga,
1863 struct obd_trans_info *oti)
1865 struct obdo *saved_oa = NULL;
1866 struct brw_page **ppga, **orig;
1867 struct obd_import *imp = class_exp2cliimp(exp);
1868 struct client_obd *cli;
1869 int rc, page_count_orig;
1872 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1873 cli = &imp->imp_obd->u.cli;
1875 if (cmd & OBD_BRW_CHECK) {
1876 /* The caller just wants to know if there's a chance that this
1877 * I/O can succeed */
1879 if (imp->imp_invalid)
1884 /* test_brw with a failed create can trip this, maybe others. */
1885 LASSERT(cli->cl_max_pages_per_rpc);
1889 orig = ppga = osc_build_ppga(pga, page_count);
1892 page_count_orig = page_count;
1894 sort_brw_pages(ppga, page_count);
1895 while (page_count) {
1896 obd_count pages_per_brw;
1898 if (page_count > cli->cl_max_pages_per_rpc)
1899 pages_per_brw = cli->cl_max_pages_per_rpc;
1901 pages_per_brw = page_count;
1903 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1905 if (saved_oa != NULL) {
1906 /* restore previously saved oa */
1907 *oinfo->oi_oa = *saved_oa;
1908 } else if (page_count > pages_per_brw) {
1909 /* save a copy of oa (brw will clobber it) */
1910 OBDO_ALLOC(saved_oa);
1911 if (saved_oa == NULL)
1912 GOTO(out, rc = -ENOMEM);
1913 *saved_oa = *oinfo->oi_oa;
1916 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1917 pages_per_brw, ppga, oinfo->oi_capa);
1922 page_count -= pages_per_brw;
1923 ppga += pages_per_brw;
1927 osc_release_ppga(orig, page_count_orig);
1929 if (saved_oa != NULL)
1930 OBDO_FREE(saved_oa);
1935 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1936 * the dirty accounting. Writeback completes or truncate happens before
1937 * writing starts. Must be called with the loi lock held. */
1938 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1941 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1945 /* This maintains the lists of pending pages to read/write for a given object
1946 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1947 * to quickly find objects that are ready to send an RPC. */
1948 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1954 if (lop->lop_num_pending == 0)
1957 /* if we have an invalid import we want to drain the queued pages
1958 * by forcing them through rpcs that immediately fail and complete
1959 * the pages. recovery relies on this to empty the queued pages
1960 * before canceling the locks and evicting down the llite pages */
1961 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1964 /* stream rpcs in queue order as long as as there is an urgent page
1965 * queued. this is our cheap solution for good batching in the case
1966 * where writepage marks some random page in the middle of the file
1967 * as urgent because of, say, memory pressure */
1968 if (!cfs_list_empty(&lop->lop_urgent)) {
1969 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1972 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1973 optimal = cli->cl_max_pages_per_rpc;
1974 if (cmd & OBD_BRW_WRITE) {
1975 /* trigger a write rpc stream as long as there are dirtiers
1976 * waiting for space. as they're waiting, they're not going to
1977 * create more pages to coalesce with what's waiting.. */
1978 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1979 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1982 /* +16 to avoid triggering rpcs that would want to include pages
1983 * that are being queued but which can't be made ready until
1984 * the queuer finishes with the page. this is a wart for
1985 * llite::commit_write() */
1988 if (lop->lop_num_pending >= optimal)
1994 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1996 struct osc_async_page *oap;
1999 if (cfs_list_empty(&lop->lop_urgent))
2002 oap = cfs_list_entry(lop->lop_urgent.next,
2003 struct osc_async_page, oap_urgent_item);
2005 if (oap->oap_async_flags & ASYNC_HP) {
2006 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2013 static void on_list(cfs_list_t *item, cfs_list_t *list,
2016 if (cfs_list_empty(item) && should_be_on)
2017 cfs_list_add_tail(item, list);
2018 else if (!cfs_list_empty(item) && !should_be_on)
2019 cfs_list_del_init(item);
2022 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2023 * can find pages to build into rpcs quickly */
2024 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2026 if (lop_makes_hprpc(&loi->loi_write_lop) ||
2027 lop_makes_hprpc(&loi->loi_read_lop)) {
2029 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2030 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2032 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2033 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2034 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2035 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2038 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2039 loi->loi_write_lop.lop_num_pending);
2041 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2042 loi->loi_read_lop.lop_num_pending);
2045 static void lop_update_pending(struct client_obd *cli,
2046 struct loi_oap_pages *lop, int cmd, int delta)
2048 lop->lop_num_pending += delta;
2049 if (cmd & OBD_BRW_WRITE)
2050 cli->cl_pending_w_pages += delta;
2052 cli->cl_pending_r_pages += delta;
2056 * this is called when a sync waiter receives an interruption. Its job is to
2057 * get the caller woken as soon as possible. If its page hasn't been put in an
2058 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2059 * desiring interruption which will forcefully complete the rpc once the rpc
2062 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2064 struct loi_oap_pages *lop;
2065 struct lov_oinfo *loi;
2069 LASSERT(!oap->oap_interrupted);
2070 oap->oap_interrupted = 1;
2072 /* ok, it's been put in an rpc. only one oap gets a request reference */
2073 if (oap->oap_request != NULL) {
2074 ptlrpc_mark_interrupted(oap->oap_request);
2075 ptlrpcd_wake(oap->oap_request);
2076 ptlrpc_req_finished(oap->oap_request);
2077 oap->oap_request = NULL;
2081 * page completion may be called only if ->cpo_prep() method was
2082 * executed by osc_io_submit(), that also adds page the to pending list
2084 if (!cfs_list_empty(&oap->oap_pending_item)) {
2085 cfs_list_del_init(&oap->oap_pending_item);
2086 cfs_list_del_init(&oap->oap_urgent_item);
2089 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2090 &loi->loi_write_lop : &loi->loi_read_lop;
2091 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2092 loi_list_maint(oap->oap_cli, oap->oap_loi);
2093 rc = oap->oap_caller_ops->ap_completion(env,
2094 oap->oap_caller_data,
2095 oap->oap_cmd, NULL, -EINTR);
2101 /* this is trying to propogate async writeback errors back up to the
2102 * application. As an async write fails we record the error code for later if
2103 * the app does an fsync. As long as errors persist we force future rpcs to be
2104 * sync so that the app can get a sync error and break the cycle of queueing
2105 * pages for which writeback will fail. */
2106 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2113 ar->ar_force_sync = 1;
2114 ar->ar_min_xid = ptlrpc_sample_next_xid();
2119 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2120 ar->ar_force_sync = 0;
2123 void osc_oap_to_pending(struct osc_async_page *oap)
2125 struct loi_oap_pages *lop;
2127 if (oap->oap_cmd & OBD_BRW_WRITE)
2128 lop = &oap->oap_loi->loi_write_lop;
2130 lop = &oap->oap_loi->loi_read_lop;
2132 if (oap->oap_async_flags & ASYNC_HP)
2133 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2134 else if (oap->oap_async_flags & ASYNC_URGENT)
2135 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2136 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2137 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2140 /* this must be called holding the loi list lock to give coverage to exit_cache,
2141 * async_flag maintenance, and oap_request */
2142 static void osc_ap_completion(const struct lu_env *env,
2143 struct client_obd *cli, struct obdo *oa,
2144 struct osc_async_page *oap, int sent, int rc)
2149 if (oap->oap_request != NULL) {
2150 xid = ptlrpc_req_xid(oap->oap_request);
2151 ptlrpc_req_finished(oap->oap_request);
2152 oap->oap_request = NULL;
2155 cfs_spin_lock(&oap->oap_lock);
2156 oap->oap_async_flags = 0;
2157 cfs_spin_unlock(&oap->oap_lock);
2158 oap->oap_interrupted = 0;
2160 if (oap->oap_cmd & OBD_BRW_WRITE) {
2161 osc_process_ar(&cli->cl_ar, xid, rc);
2162 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2165 if (rc == 0 && oa != NULL) {
2166 if (oa->o_valid & OBD_MD_FLBLOCKS)
2167 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2168 if (oa->o_valid & OBD_MD_FLMTIME)
2169 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2170 if (oa->o_valid & OBD_MD_FLATIME)
2171 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2172 if (oa->o_valid & OBD_MD_FLCTIME)
2173 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2176 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2177 oap->oap_cmd, oa, rc);
2179 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2180 * I/O on the page could start, but OSC calls it under lock
2181 * and thus we can add oap back to pending safely */
2183 /* upper layer wants to leave the page on pending queue */
2184 osc_oap_to_pending(oap);
2186 osc_exit_cache(cli, oap, sent);
2190 static int brw_interpret(const struct lu_env *env,
2191 struct ptlrpc_request *req, void *data, int rc)
2193 struct osc_brw_async_args *aa = data;
2194 struct client_obd *cli;
2198 rc = osc_brw_fini_request(req, rc);
2199 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2200 if (osc_recoverable_error(rc)) {
2201 /* Only retry once for mmaped files since the mmaped page
2202 * might be modified at anytime. We have to retry at least
2203 * once in case there WAS really a corruption of the page
2204 * on the network, that was not caused by mmap() modifying
2205 * the page. Bug11742 */
2206 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2207 aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2208 aa->aa_oa->o_flags & OBD_FL_MMAP) {
2211 rc = osc_brw_redo_request(req, aa);
2218 capa_put(aa->aa_ocapa);
2219 aa->aa_ocapa = NULL;
2224 client_obd_list_lock(&cli->cl_loi_list_lock);
2226 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2227 * is called so we know whether to go to sync BRWs or wait for more
2228 * RPCs to complete */
2229 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2230 cli->cl_w_in_flight--;
2232 cli->cl_r_in_flight--;
2234 async = cfs_list_empty(&aa->aa_oaps);
2235 if (!async) { /* from osc_send_oap_rpc() */
2236 struct osc_async_page *oap, *tmp;
2237 /* the caller may re-use the oap after the completion call so
2238 * we need to clean it up a little */
2239 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2241 cfs_list_del_init(&oap->oap_rpc_item);
2242 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2244 OBDO_FREE(aa->aa_oa);
2245 } else { /* from async_internal() */
2247 for (i = 0; i < aa->aa_page_count; i++)
2248 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2250 osc_wake_cache_waiters(cli);
2251 osc_check_rpcs(env, cli);
2252 client_obd_list_unlock(&cli->cl_loi_list_lock);
2254 cl_req_completion(env, aa->aa_clerq, rc);
2255 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2260 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2261 struct client_obd *cli,
2262 cfs_list_t *rpc_list,
2263 int page_count, int cmd)
2265 struct ptlrpc_request *req;
2266 struct brw_page **pga = NULL;
2267 struct osc_brw_async_args *aa;
2268 struct obdo *oa = NULL;
2269 const struct obd_async_page_ops *ops = NULL;
2270 void *caller_data = NULL;
2271 struct osc_async_page *oap;
2272 struct osc_async_page *tmp;
2273 struct ost_body *body;
2274 struct cl_req *clerq = NULL;
2275 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2276 struct ldlm_lock *lock = NULL;
2277 struct cl_req_attr crattr;
2278 int i, rc, mpflag = 0;
2281 LASSERT(!cfs_list_empty(rpc_list));
2283 if (cmd & OBD_BRW_MEMALLOC)
2284 mpflag = cfs_memory_pressure_get_and_set();
2286 memset(&crattr, 0, sizeof crattr);
2287 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2289 GOTO(out, req = ERR_PTR(-ENOMEM));
2293 GOTO(out, req = ERR_PTR(-ENOMEM));
2296 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2297 struct cl_page *page = osc_oap2cl_page(oap);
2299 ops = oap->oap_caller_ops;
2300 caller_data = oap->oap_caller_data;
2302 clerq = cl_req_alloc(env, page, crt,
2303 1 /* only 1-object rpcs for
2306 GOTO(out, req = (void *)clerq);
2307 lock = oap->oap_ldlm_lock;
2309 pga[i] = &oap->oap_brw_page;
2310 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2311 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2312 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2314 cl_req_page_add(env, clerq, page);
2317 /* always get the data for the obdo for the rpc */
2318 LASSERT(ops != NULL);
2320 crattr.cra_capa = NULL;
2321 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2323 oa->o_handle = lock->l_remote_handle;
2324 oa->o_valid |= OBD_MD_FLHANDLE;
2327 rc = cl_req_prep(env, clerq);
2329 CERROR("cl_req_prep failed: %d\n", rc);
2330 GOTO(out, req = ERR_PTR(rc));
2333 sort_brw_pages(pga, page_count);
2334 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2335 pga, &req, crattr.cra_capa, 1, 0);
2337 CERROR("prep_req failed: %d\n", rc);
2338 GOTO(out, req = ERR_PTR(rc));
2341 if (cmd & OBD_BRW_MEMALLOC)
2342 req->rq_memalloc = 1;
2344 /* Need to update the timestamps after the request is built in case
2345 * we race with setattr (locally or in queue at OST). If OST gets
2346 * later setattr before earlier BRW (as determined by the request xid),
2347 * the OST will not use BRW timestamps. Sadly, there is no obvious
2348 * way to do this in a single call. bug 10150 */
2349 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2350 cl_req_attr_set(env, clerq, &crattr,
2351 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2353 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2354 aa = ptlrpc_req_async_args(req);
2355 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2356 cfs_list_splice(rpc_list, &aa->aa_oaps);
2357 CFS_INIT_LIST_HEAD(rpc_list);
2358 aa->aa_clerq = clerq;
2360 if (cmd & OBD_BRW_MEMALLOC)
2361 cfs_memory_pressure_restore(mpflag);
2363 capa_put(crattr.cra_capa);
2368 OBD_FREE(pga, sizeof(*pga) * page_count);
2369 /* this should happen rarely and is pretty bad, it makes the
2370 * pending list not follow the dirty order */
2371 client_obd_list_lock(&cli->cl_loi_list_lock);
2372 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2373 cfs_list_del_init(&oap->oap_rpc_item);
2375 /* queued sync pages can be torn down while the pages
2376 * were between the pending list and the rpc */
2377 if (oap->oap_interrupted) {
2378 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2379 osc_ap_completion(env, cli, NULL, oap, 0,
2383 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2385 if (clerq && !IS_ERR(clerq))
2386 cl_req_completion(env, clerq, PTR_ERR(req));
2392 * prepare pages for ASYNC io and put pages in send queue.
2394 * \param cmd OBD_BRW_* macroses
2395 * \param lop pending pages
2397 * \return zero if no page added to send queue.
2398 * \return 1 if pages successfully added to send queue.
2399 * \return negative on errors.
2402 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2403 struct lov_oinfo *loi,
2404 int cmd, struct loi_oap_pages *lop)
2406 struct ptlrpc_request *req;
2407 obd_count page_count = 0;
2408 struct osc_async_page *oap = NULL, *tmp;
2409 struct osc_brw_async_args *aa;
2410 const struct obd_async_page_ops *ops;
2411 CFS_LIST_HEAD(rpc_list);
2412 CFS_LIST_HEAD(tmp_list);
2413 unsigned int ending_offset;
2414 unsigned starting_offset = 0;
2415 int srvlock = 0, mem_tight = 0;
2416 struct cl_object *clob = NULL;
2419 /* ASYNC_HP pages first. At present, when the lock the pages is
2420 * to be canceled, the pages covered by the lock will be sent out
2421 * with ASYNC_HP. We have to send out them as soon as possible. */
2422 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2423 if (oap->oap_async_flags & ASYNC_HP)
2424 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2426 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2427 if (++page_count >= cli->cl_max_pages_per_rpc)
2431 cfs_list_splice(&tmp_list, &lop->lop_pending);
2434 /* first we find the pages we're allowed to work with */
2435 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2437 ops = oap->oap_caller_ops;
2439 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2440 "magic 0x%x\n", oap, oap->oap_magic);
2443 /* pin object in memory, so that completion call-backs
2444 * can be safely called under client_obd_list lock. */
2445 clob = osc_oap2cl_page(oap)->cp_obj;
2446 cl_object_get(clob);
2449 if (page_count != 0 &&
2450 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2451 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2452 " oap %p, page %p, srvlock %u\n",
2453 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2457 /* If there is a gap at the start of this page, it can't merge
2458 * with any previous page, so we'll hand the network a
2459 * "fragmented" page array that it can't transfer in 1 RDMA */
2460 if (page_count != 0 && oap->oap_page_off != 0)
2463 /* in llite being 'ready' equates to the page being locked
2464 * until completion unlocks it. commit_write submits a page
2465 * as not ready because its unlock will happen unconditionally
2466 * as the call returns. if we race with commit_write giving
2467 * us that page we don't want to create a hole in the page
2468 * stream, so we stop and leave the rpc to be fired by
2469 * another dirtier or kupdated interval (the not ready page
2470 * will still be on the dirty list). we could call in
2471 * at the end of ll_file_write to process the queue again. */
2472 if (!(oap->oap_async_flags & ASYNC_READY)) {
2473 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2476 CDEBUG(D_INODE, "oap %p page %p returned %d "
2477 "instead of ready\n", oap,
2481 /* llite is telling us that the page is still
2482 * in commit_write and that we should try
2483 * and put it in an rpc again later. we
2484 * break out of the loop so we don't create
2485 * a hole in the sequence of pages in the rpc
2490 /* the io isn't needed.. tell the checks
2491 * below to complete the rpc with EINTR */
2492 cfs_spin_lock(&oap->oap_lock);
2493 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2494 cfs_spin_unlock(&oap->oap_lock);
2495 oap->oap_count = -EINTR;
2498 cfs_spin_lock(&oap->oap_lock);
2499 oap->oap_async_flags |= ASYNC_READY;
2500 cfs_spin_unlock(&oap->oap_lock);
2503 LASSERTF(0, "oap %p page %p returned %d "
2504 "from make_ready\n", oap,
2512 * Page submitted for IO has to be locked. Either by
2513 * ->ap_make_ready() or by higher layers.
2515 #if defined(__KERNEL__) && defined(__linux__)
2517 struct cl_page *page;
2519 page = osc_oap2cl_page(oap);
2521 if (page->cp_type == CPT_CACHEABLE &&
2522 !(PageLocked(oap->oap_page) &&
2523 (CheckWriteback(oap->oap_page, cmd)))) {
2524 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2526 (long)oap->oap_page->flags,
2527 oap->oap_async_flags);
2533 /* take the page out of our book-keeping */
2534 cfs_list_del_init(&oap->oap_pending_item);
2535 lop_update_pending(cli, lop, cmd, -1);
2536 cfs_list_del_init(&oap->oap_urgent_item);
2538 if (page_count == 0)
2539 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2540 (PTLRPC_MAX_BRW_SIZE - 1);
2542 /* ask the caller for the size of the io as the rpc leaves. */
2543 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2545 ops->ap_refresh_count(env, oap->oap_caller_data,
2547 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2549 if (oap->oap_count <= 0) {
2550 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2552 osc_ap_completion(env, cli, NULL,
2553 oap, 0, oap->oap_count);
2557 /* now put the page back in our accounting */
2558 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2559 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2561 if (page_count == 0)
2562 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2563 if (++page_count >= cli->cl_max_pages_per_rpc)
2566 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2567 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2568 * have the same alignment as the initial writes that allocated
2569 * extents on the server. */
2570 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2571 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2572 if (ending_offset == 0)
2575 /* If there is a gap at the end of this page, it can't merge
2576 * with any subsequent pages, so we'll hand the network a
2577 * "fragmented" page array that it can't transfer in 1 RDMA */
2578 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2582 osc_wake_cache_waiters(cli);
2584 loi_list_maint(cli, loi);
2586 client_obd_list_unlock(&cli->cl_loi_list_lock);
2589 cl_object_put(env, clob);
2591 if (page_count == 0) {
2592 client_obd_list_lock(&cli->cl_loi_list_lock);
2596 req = osc_build_req(env, cli, &rpc_list, page_count,
2597 mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2599 LASSERT(cfs_list_empty(&rpc_list));
2600 loi_list_maint(cli, loi);
2601 RETURN(PTR_ERR(req));
2604 aa = ptlrpc_req_async_args(req);
2606 if (cmd == OBD_BRW_READ) {
2607 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2608 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2609 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2610 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2612 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2613 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2614 cli->cl_w_in_flight);
2615 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2616 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2618 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2620 client_obd_list_lock(&cli->cl_loi_list_lock);
2622 if (cmd == OBD_BRW_READ)
2623 cli->cl_r_in_flight++;
2625 cli->cl_w_in_flight++;
2627 /* queued sync pages can be torn down while the pages
2628 * were between the pending list and the rpc */
2630 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2631 /* only one oap gets a request reference */
2634 if (oap->oap_interrupted && !req->rq_intr) {
2635 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2637 ptlrpc_mark_interrupted(req);
2641 tmp->oap_request = ptlrpc_request_addref(req);
2643 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2644 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2646 req->rq_interpret_reply = brw_interpret;
2647 ptlrpcd_add_req(req, PSCOPE_BRW);
2651 #define LOI_DEBUG(LOI, STR, args...) \
2652 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2653 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2654 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2655 (LOI)->loi_write_lop.lop_num_pending, \
2656 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2657 (LOI)->loi_read_lop.lop_num_pending, \
2658 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2661 /* This is called by osc_check_rpcs() to find which objects have pages that
2662 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2663 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2667 /* First return objects that have blocked locks so that they
2668 * will be flushed quickly and other clients can get the lock,
2669 * then objects which have pages ready to be stuffed into RPCs */
2670 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2671 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2672 struct lov_oinfo, loi_hp_ready_item));
2673 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2674 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2675 struct lov_oinfo, loi_ready_item));
2677 /* then if we have cache waiters, return all objects with queued
2678 * writes. This is especially important when many small files
2679 * have filled up the cache and not been fired into rpcs because
2680 * they don't pass the nr_pending/object threshhold */
2681 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2682 !cfs_list_empty(&cli->cl_loi_write_list))
2683 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2684 struct lov_oinfo, loi_write_item));
2686 /* then return all queued objects when we have an invalid import
2687 * so that they get flushed */
2688 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2689 if (!cfs_list_empty(&cli->cl_loi_write_list))
2690 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2693 if (!cfs_list_empty(&cli->cl_loi_read_list))
2694 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2695 struct lov_oinfo, loi_read_item));
2700 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2702 struct osc_async_page *oap;
2705 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2706 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2707 struct osc_async_page, oap_urgent_item);
2708 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2711 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2712 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2713 struct osc_async_page, oap_urgent_item);
2714 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2717 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2720 /* called with the loi list lock held */
2721 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2723 struct lov_oinfo *loi;
2724 int rc = 0, race_counter = 0;
2727 while ((loi = osc_next_loi(cli)) != NULL) {
2728 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2730 if (osc_max_rpc_in_flight(cli, loi))
2733 /* attempt some read/write balancing by alternating between
2734 * reads and writes in an object. The makes_rpc checks here
2735 * would be redundant if we were getting read/write work items
2736 * instead of objects. we don't want send_oap_rpc to drain a
2737 * partial read pending queue when we're given this object to
2738 * do io on writes while there are cache waiters */
2739 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2740 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2741 &loi->loi_write_lop);
2743 CERROR("Write request failed with %d\n", rc);
2745 /* osc_send_oap_rpc failed, mostly because of
2748 * It can't break here, because if:
2749 * - a page was submitted by osc_io_submit, so
2751 * - no request in flight
2752 * - no subsequent request
2753 * The system will be in live-lock state,
2754 * because there is no chance to call
2755 * osc_io_unplug() and osc_check_rpcs() any
2756 * more. pdflush can't help in this case,
2757 * because it might be blocked at grabbing
2758 * the page lock as we mentioned.
2760 * Anyway, continue to drain pages. */
2769 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2770 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2771 &loi->loi_read_lop);
2773 CERROR("Read request failed with %d\n", rc);
2781 /* attempt some inter-object balancing by issuing rpcs
2782 * for each object in turn */
2783 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2784 cfs_list_del_init(&loi->loi_hp_ready_item);
2785 if (!cfs_list_empty(&loi->loi_ready_item))
2786 cfs_list_del_init(&loi->loi_ready_item);
2787 if (!cfs_list_empty(&loi->loi_write_item))
2788 cfs_list_del_init(&loi->loi_write_item);
2789 if (!cfs_list_empty(&loi->loi_read_item))
2790 cfs_list_del_init(&loi->loi_read_item);
2792 loi_list_maint(cli, loi);
2794 /* send_oap_rpc fails with 0 when make_ready tells it to
2795 * back off. llite's make_ready does this when it tries
2796 * to lock a page queued for write that is already locked.
2797 * we want to try sending rpcs from many objects, but we
2798 * don't want to spin failing with 0. */
2799 if (race_counter == 10)
2805 /* we're trying to queue a page in the osc so we're subject to the
2806 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2807 * If the osc's queued pages are already at that limit, then we want to sleep
2808 * until there is space in the osc's queue for us. We also may be waiting for
2809 * write credits from the OST if there are RPCs in flight that may return some
2810 * before we fall back to sync writes.
2812 * We need this know our allocation was granted in the presence of signals */
2813 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2817 client_obd_list_lock(&cli->cl_loi_list_lock);
2818 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2819 client_obd_list_unlock(&cli->cl_loi_list_lock);
2824 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2827 int osc_enter_cache_try(const struct lu_env *env,
2828 struct client_obd *cli, struct lov_oinfo *loi,
2829 struct osc_async_page *oap, int transient)
2833 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2835 osc_consume_write_grant(cli, &oap->oap_brw_page);
2837 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2838 cfs_atomic_inc(&obd_dirty_transit_pages);
2839 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2845 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2846 * grant or cache space. */
2847 static int osc_enter_cache(const struct lu_env *env,
2848 struct client_obd *cli, struct lov_oinfo *loi,
2849 struct osc_async_page *oap)
2851 struct osc_cache_waiter ocw;
2852 struct l_wait_info lwi = { 0 };
2856 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2857 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2858 cli->cl_dirty_max, obd_max_dirty_pages,
2859 cli->cl_lost_grant, cli->cl_avail_grant);
2861 /* force the caller to try sync io. this can jump the list
2862 * of queued writes and create a discontiguous rpc stream */
2863 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2864 loi->loi_ar.ar_force_sync)
2867 /* Hopefully normal case - cache space and write credits available */
2868 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2869 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2870 osc_enter_cache_try(env, cli, loi, oap, 0))
2873 /* It is safe to block as a cache waiter as long as there is grant
2874 * space available or the hope of additional grant being returned
2875 * when an in flight write completes. Using the write back cache
2876 * if possible is preferable to sending the data synchronously
2877 * because write pages can then be merged in to large requests.
2878 * The addition of this cache waiter will causing pending write
2879 * pages to be sent immediately. */
2880 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2881 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2882 cfs_waitq_init(&ocw.ocw_waitq);
2886 loi_list_maint(cli, loi);
2887 osc_check_rpcs(env, cli);
2888 client_obd_list_unlock(&cli->cl_loi_list_lock);
2890 CDEBUG(D_CACHE, "sleeping for cache space\n");
2891 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2893 client_obd_list_lock(&cli->cl_loi_list_lock);
2894 if (!cfs_list_empty(&ocw.ocw_entry)) {
2895 cfs_list_del(&ocw.ocw_entry);
2905 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2906 struct lov_oinfo *loi, cfs_page_t *page,
2907 obd_off offset, const struct obd_async_page_ops *ops,
2908 void *data, void **res, int nocache,
2909 struct lustre_handle *lockh)
2911 struct osc_async_page *oap;
2916 return cfs_size_round(sizeof(*oap));
2919 oap->oap_magic = OAP_MAGIC;
2920 oap->oap_cli = &exp->exp_obd->u.cli;
2923 oap->oap_caller_ops = ops;
2924 oap->oap_caller_data = data;
2926 oap->oap_page = page;
2927 oap->oap_obj_off = offset;
2928 if (!client_is_remote(exp) &&
2929 cfs_capable(CFS_CAP_SYS_RESOURCE))
2930 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2932 LASSERT(!(offset & ~CFS_PAGE_MASK));
2934 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2935 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2936 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2937 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2939 cfs_spin_lock_init(&oap->oap_lock);
2940 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2944 struct osc_async_page *oap_from_cookie(void *cookie)
2946 struct osc_async_page *oap = cookie;
2947 if (oap->oap_magic != OAP_MAGIC)
2948 return ERR_PTR(-EINVAL);
2952 int osc_queue_async_io(const struct lu_env *env,
2953 struct obd_export *exp, struct lov_stripe_md *lsm,
2954 struct lov_oinfo *loi, void *cookie,
2955 int cmd, obd_off off, int count,
2956 obd_flag brw_flags, enum async_flags async_flags)
2958 struct client_obd *cli = &exp->exp_obd->u.cli;
2959 struct osc_async_page *oap;
2963 oap = oap_from_cookie(cookie);
2965 RETURN(PTR_ERR(oap));
2967 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2970 if (!cfs_list_empty(&oap->oap_pending_item) ||
2971 !cfs_list_empty(&oap->oap_urgent_item) ||
2972 !cfs_list_empty(&oap->oap_rpc_item))
2975 /* check if the file's owner/group is over quota */
2976 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2977 struct cl_object *obj;
2978 struct cl_attr attr; /* XXX put attr into thread info */
2979 unsigned int qid[MAXQUOTAS];
2981 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2983 cl_object_attr_lock(obj);
2984 rc = cl_object_attr_get(env, obj, &attr);
2985 cl_object_attr_unlock(obj);
2987 qid[USRQUOTA] = attr.cat_uid;
2988 qid[GRPQUOTA] = attr.cat_gid;
2990 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2997 loi = lsm->lsm_oinfo[0];
2999 client_obd_list_lock(&cli->cl_loi_list_lock);
3001 LASSERT(off + count <= CFS_PAGE_SIZE);
3003 oap->oap_page_off = off;
3004 oap->oap_count = count;
3005 oap->oap_brw_flags = brw_flags;
3006 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3007 if (cfs_memory_pressure_get())
3008 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3009 cfs_spin_lock(&oap->oap_lock);
3010 oap->oap_async_flags = async_flags;
3011 cfs_spin_unlock(&oap->oap_lock);
3013 if (cmd & OBD_BRW_WRITE) {
3014 rc = osc_enter_cache(env, cli, loi, oap);
3016 client_obd_list_unlock(&cli->cl_loi_list_lock);
3021 osc_oap_to_pending(oap);
3022 loi_list_maint(cli, loi);
3024 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3027 osc_check_rpcs(env, cli);
3028 client_obd_list_unlock(&cli->cl_loi_list_lock);
3033 /* aka (~was & now & flag), but this is more clear :) */
3034 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3036 int osc_set_async_flags_base(struct client_obd *cli,
3037 struct lov_oinfo *loi, struct osc_async_page *oap,
3038 obd_flag async_flags)
3040 struct loi_oap_pages *lop;
3044 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3046 if (oap->oap_cmd & OBD_BRW_WRITE) {
3047 lop = &loi->loi_write_lop;
3049 lop = &loi->loi_read_lop;
3052 if ((oap->oap_async_flags & async_flags) == async_flags)
3055 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3056 flags |= ASYNC_READY;
3058 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3059 cfs_list_empty(&oap->oap_rpc_item)) {
3060 if (oap->oap_async_flags & ASYNC_HP)
3061 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3063 cfs_list_add_tail(&oap->oap_urgent_item,
3065 flags |= ASYNC_URGENT;
3066 loi_list_maint(cli, loi);
3068 cfs_spin_lock(&oap->oap_lock);
3069 oap->oap_async_flags |= flags;
3070 cfs_spin_unlock(&oap->oap_lock);
3072 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3073 oap->oap_async_flags);
3077 int osc_teardown_async_page(struct obd_export *exp,
3078 struct lov_stripe_md *lsm,
3079 struct lov_oinfo *loi, void *cookie)
3081 struct client_obd *cli = &exp->exp_obd->u.cli;
3082 struct loi_oap_pages *lop;
3083 struct osc_async_page *oap;
3087 oap = oap_from_cookie(cookie);
3089 RETURN(PTR_ERR(oap));
3092 loi = lsm->lsm_oinfo[0];
3094 if (oap->oap_cmd & OBD_BRW_WRITE) {
3095 lop = &loi->loi_write_lop;
3097 lop = &loi->loi_read_lop;
3100 client_obd_list_lock(&cli->cl_loi_list_lock);
3102 if (!cfs_list_empty(&oap->oap_rpc_item))
3103 GOTO(out, rc = -EBUSY);
3105 osc_exit_cache(cli, oap, 0);
3106 osc_wake_cache_waiters(cli);
3108 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3109 cfs_list_del_init(&oap->oap_urgent_item);
3110 cfs_spin_lock(&oap->oap_lock);
3111 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3112 cfs_spin_unlock(&oap->oap_lock);
3114 if (!cfs_list_empty(&oap->oap_pending_item)) {
3115 cfs_list_del_init(&oap->oap_pending_item);
3116 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3118 loi_list_maint(cli, loi);
3119 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3121 client_obd_list_unlock(&cli->cl_loi_list_lock);
3125 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3126 struct ldlm_enqueue_info *einfo,
3129 void *data = einfo->ei_cbdata;
3131 LASSERT(lock != NULL);
3132 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3133 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3134 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3135 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3137 lock_res_and_lock(lock);
3138 cfs_spin_lock(&osc_ast_guard);
3139 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3140 lock->l_ast_data = data;
3141 cfs_spin_unlock(&osc_ast_guard);
3142 unlock_res_and_lock(lock);
3145 static void osc_set_data_with_check(struct lustre_handle *lockh,
3146 struct ldlm_enqueue_info *einfo,
3149 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3152 osc_set_lock_data_with_check(lock, einfo, flags);
3153 LDLM_LOCK_PUT(lock);
3155 CERROR("lockh %p, data %p - client evicted?\n",
3156 lockh, einfo->ei_cbdata);
3159 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3160 ldlm_iterator_t replace, void *data)
3162 struct ldlm_res_id res_id;
3163 struct obd_device *obd = class_exp2obd(exp);
3165 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3166 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3170 /* find any ldlm lock of the inode in osc
3174 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3175 ldlm_iterator_t replace, void *data)
3177 struct ldlm_res_id res_id;
3178 struct obd_device *obd = class_exp2obd(exp);
3181 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3182 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3183 if (rc == LDLM_ITER_STOP)
3185 if (rc == LDLM_ITER_CONTINUE)
3190 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3191 obd_enqueue_update_f upcall, void *cookie,
3194 int intent = *flags & LDLM_FL_HAS_INTENT;
3198 /* The request was created before ldlm_cli_enqueue call. */
3199 if (rc == ELDLM_LOCK_ABORTED) {
3200 struct ldlm_reply *rep;
3201 rep = req_capsule_server_get(&req->rq_pill,
3204 LASSERT(rep != NULL);
3205 if (rep->lock_policy_res1)
3206 rc = rep->lock_policy_res1;
3210 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3211 *flags |= LDLM_FL_LVB_READY;
3212 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3213 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3216 /* Call the update callback. */
3217 rc = (*upcall)(cookie, rc);
3221 static int osc_enqueue_interpret(const struct lu_env *env,
3222 struct ptlrpc_request *req,
3223 struct osc_enqueue_args *aa, int rc)
3225 struct ldlm_lock *lock;
3226 struct lustre_handle handle;
3229 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3230 * might be freed anytime after lock upcall has been called. */
3231 lustre_handle_copy(&handle, aa->oa_lockh);
3232 mode = aa->oa_ei->ei_mode;
3234 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3236 lock = ldlm_handle2lock(&handle);
3238 /* Take an additional reference so that a blocking AST that
3239 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3240 * to arrive after an upcall has been executed by
3241 * osc_enqueue_fini(). */
3242 ldlm_lock_addref(&handle, mode);
3244 /* Let CP AST to grant the lock first. */
3245 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3247 /* Complete obtaining the lock procedure. */
3248 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3249 mode, aa->oa_flags, aa->oa_lvb,
3250 sizeof(*aa->oa_lvb), &handle, rc);
3251 /* Complete osc stuff. */
3252 rc = osc_enqueue_fini(req, aa->oa_lvb,
3253 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3255 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3257 /* Release the lock for async request. */
3258 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3260 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3261 * not already released by
3262 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3264 ldlm_lock_decref(&handle, mode);
3266 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3267 aa->oa_lockh, req, aa);
3268 ldlm_lock_decref(&handle, mode);
3269 LDLM_LOCK_PUT(lock);
3273 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3274 struct lov_oinfo *loi, int flags,
3275 struct ost_lvb *lvb, __u32 mode, int rc)
3277 if (rc == ELDLM_OK) {
3278 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3281 LASSERT(lock != NULL);
3282 loi->loi_lvb = *lvb;
3283 tmp = loi->loi_lvb.lvb_size;
3284 /* Extend KMS up to the end of this lock and no further
3285 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3286 if (tmp > lock->l_policy_data.l_extent.end)
3287 tmp = lock->l_policy_data.l_extent.end + 1;
3288 if (tmp >= loi->loi_kms) {
3289 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3290 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3291 loi_kms_set(loi, tmp);
3293 LDLM_DEBUG(lock, "lock acquired, setting rss="
3294 LPU64"; leaving kms="LPU64", end="LPU64,
3295 loi->loi_lvb.lvb_size, loi->loi_kms,
3296 lock->l_policy_data.l_extent.end);
3298 ldlm_lock_allow_match(lock);
3299 LDLM_LOCK_PUT(lock);
3300 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3301 loi->loi_lvb = *lvb;
3302 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3303 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3307 EXPORT_SYMBOL(osc_update_enqueue);
3309 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3311 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3312 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3313 * other synchronous requests, however keeping some locks and trying to obtain
3314 * others may take a considerable amount of time in a case of ost failure; and
3315 * when other sync requests do not get released lock from a client, the client
3316 * is excluded from the cluster -- such scenarious make the life difficult, so
3317 * release locks just after they are obtained. */
3318 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3319 int *flags, ldlm_policy_data_t *policy,
3320 struct ost_lvb *lvb, int kms_valid,
3321 obd_enqueue_update_f upcall, void *cookie,
3322 struct ldlm_enqueue_info *einfo,
3323 struct lustre_handle *lockh,
3324 struct ptlrpc_request_set *rqset, int async)
3326 struct obd_device *obd = exp->exp_obd;
3327 struct ptlrpc_request *req = NULL;
3328 int intent = *flags & LDLM_FL_HAS_INTENT;
3333 /* Filesystem lock extents are extended to page boundaries so that
3334 * dealing with the page cache is a little smoother. */
3335 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3336 policy->l_extent.end |= ~CFS_PAGE_MASK;
3339 * kms is not valid when either object is completely fresh (so that no
3340 * locks are cached), or object was evicted. In the latter case cached
3341 * lock cannot be used, because it would prime inode state with
3342 * potentially stale LVB.
3347 /* Next, search for already existing extent locks that will cover us */
3348 /* If we're trying to read, we also search for an existing PW lock. The
3349 * VFS and page cache already protect us locally, so lots of readers/
3350 * writers can share a single PW lock.
3352 * There are problems with conversion deadlocks, so instead of
3353 * converting a read lock to a write lock, we'll just enqueue a new
3356 * At some point we should cancel the read lock instead of making them
3357 * send us a blocking callback, but there are problems with canceling
3358 * locks out from other users right now, too. */
3359 mode = einfo->ei_mode;
3360 if (einfo->ei_mode == LCK_PR)
3362 mode = ldlm_lock_match(obd->obd_namespace,
3363 *flags | LDLM_FL_LVB_READY, res_id,
3364 einfo->ei_type, policy, mode, lockh, 0);
3366 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3368 if (matched->l_ast_data == NULL ||
3369 matched->l_ast_data == einfo->ei_cbdata) {
3370 /* addref the lock only if not async requests and PW
3371 * lock is matched whereas we asked for PR. */
3372 if (!rqset && einfo->ei_mode != mode)
3373 ldlm_lock_addref(lockh, LCK_PR);
3374 osc_set_lock_data_with_check(matched, einfo, *flags);
3376 /* I would like to be able to ASSERT here that
3377 * rss <= kms, but I can't, for reasons which
3378 * are explained in lov_enqueue() */
3381 /* We already have a lock, and it's referenced */
3382 (*upcall)(cookie, ELDLM_OK);
3384 /* For async requests, decref the lock. */
3385 if (einfo->ei_mode != mode)
3386 ldlm_lock_decref(lockh, LCK_PW);
3388 ldlm_lock_decref(lockh, einfo->ei_mode);
3389 LDLM_LOCK_PUT(matched);
3392 ldlm_lock_decref(lockh, mode);
3393 LDLM_LOCK_PUT(matched);
3398 CFS_LIST_HEAD(cancels);
3399 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3400 &RQF_LDLM_ENQUEUE_LVB);
3404 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3406 ptlrpc_request_free(req);
3410 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3412 ptlrpc_request_set_replen(req);
3415 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3416 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3418 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3419 sizeof(*lvb), lockh, async);
3422 struct osc_enqueue_args *aa;
3423 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3424 aa = ptlrpc_req_async_args(req);
3427 aa->oa_flags = flags;
3428 aa->oa_upcall = upcall;
3429 aa->oa_cookie = cookie;
3431 aa->oa_lockh = lockh;
3433 req->rq_interpret_reply =
3434 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3435 if (rqset == PTLRPCD_SET)
3436 ptlrpcd_add_req(req, PSCOPE_OTHER);
3438 ptlrpc_set_add_req(rqset, req);
3439 } else if (intent) {
3440 ptlrpc_req_finished(req);
3445 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3447 ptlrpc_req_finished(req);
3452 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3453 struct ldlm_enqueue_info *einfo,
3454 struct ptlrpc_request_set *rqset)
3456 struct ldlm_res_id res_id;
3460 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3461 oinfo->oi_md->lsm_object_seq, &res_id);
3463 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3464 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3465 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3466 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3467 rqset, rqset != NULL);
3471 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3472 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3473 int *flags, void *data, struct lustre_handle *lockh,
3476 struct obd_device *obd = exp->exp_obd;
3477 int lflags = *flags;
3481 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3484 /* Filesystem lock extents are extended to page boundaries so that
3485 * dealing with the page cache is a little smoother */
3486 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3487 policy->l_extent.end |= ~CFS_PAGE_MASK;
3489 /* Next, search for already existing extent locks that will cover us */
3490 /* If we're trying to read, we also search for an existing PW lock. The
3491 * VFS and page cache already protect us locally, so lots of readers/
3492 * writers can share a single PW lock. */
3496 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3497 res_id, type, policy, rc, lockh, unref);
3500 osc_set_data_with_check(lockh, data, lflags);
3501 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3502 ldlm_lock_addref(lockh, LCK_PR);
3503 ldlm_lock_decref(lockh, LCK_PW);
3510 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3514 if (unlikely(mode == LCK_GROUP))
3515 ldlm_lock_decref_and_cancel(lockh, mode);
3517 ldlm_lock_decref(lockh, mode);
3522 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3523 __u32 mode, struct lustre_handle *lockh)
3526 RETURN(osc_cancel_base(lockh, mode));
3529 static int osc_cancel_unused(struct obd_export *exp,
3530 struct lov_stripe_md *lsm,
3531 ldlm_cancel_flags_t flags,
3534 struct obd_device *obd = class_exp2obd(exp);
3535 struct ldlm_res_id res_id, *resp = NULL;
3538 resp = osc_build_res_name(lsm->lsm_object_id,
3539 lsm->lsm_object_seq, &res_id);
3542 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3545 static int osc_statfs_interpret(const struct lu_env *env,
3546 struct ptlrpc_request *req,
3547 struct osc_async_args *aa, int rc)
3549 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3550 struct obd_statfs *msfs;
3555 /* The request has in fact never been sent
3556 * due to issues at a higher level (LOV).
3557 * Exit immediately since the caller is
3558 * aware of the problem and takes care
3559 * of the clean up */
3562 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3563 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3569 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3571 GOTO(out, rc = -EPROTO);
3574 /* Reinitialize the RDONLY and DEGRADED flags at the client
3575 * on each statfs, so they don't stay set permanently. */
3576 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3578 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3579 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3580 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3581 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3583 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3584 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3585 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3586 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3588 /* Add a bit of hysteresis so this flag isn't continually flapping,
3589 * and ensure that new files don't get extremely fragmented due to
3590 * only a small amount of available space in the filesystem.
3591 * We want to set the NOSPC flag when there is less than ~0.1% free
3592 * and clear it when there is at least ~0.2% free space, so:
3593 * avail < ~0.1% max max = avail + used
3594 * 1025 * avail < avail + used used = blocks - free
3595 * 1024 * avail < used
3596 * 1024 * avail < blocks - free
3597 * avail < ((blocks - free) >> 10)
3599 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3600 * lose that amount of space so in those cases we report no space left
3601 * if their is less than 1 GB left. */
3602 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3603 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3604 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3605 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3606 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3607 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3608 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3610 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3612 *aa->aa_oi->oi_osfs = *msfs;
3614 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3618 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3619 __u64 max_age, struct ptlrpc_request_set *rqset)
3621 struct ptlrpc_request *req;
3622 struct osc_async_args *aa;
3626 /* We could possibly pass max_age in the request (as an absolute
3627 * timestamp or a "seconds.usec ago") so the target can avoid doing
3628 * extra calls into the filesystem if that isn't necessary (e.g.
3629 * during mount that would help a bit). Having relative timestamps
3630 * is not so great if request processing is slow, while absolute
3631 * timestamps are not ideal because they need time synchronization. */
3632 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3636 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3638 ptlrpc_request_free(req);
3641 ptlrpc_request_set_replen(req);
3642 req->rq_request_portal = OST_CREATE_PORTAL;
3643 ptlrpc_at_set_req_timeout(req);
3645 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3646 /* procfs requests not want stat in wait for avoid deadlock */
3647 req->rq_no_resend = 1;
3648 req->rq_no_delay = 1;
3651 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3652 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3653 aa = ptlrpc_req_async_args(req);
3656 ptlrpc_set_add_req(rqset, req);
3660 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3661 __u64 max_age, __u32 flags)
3663 struct obd_statfs *msfs;
3664 struct ptlrpc_request *req;
3665 struct obd_import *imp = NULL;
3669 /*Since the request might also come from lprocfs, so we need
3670 *sync this with client_disconnect_export Bug15684*/
3671 cfs_down_read(&obd->u.cli.cl_sem);
3672 if (obd->u.cli.cl_import)
3673 imp = class_import_get(obd->u.cli.cl_import);
3674 cfs_up_read(&obd->u.cli.cl_sem);
3678 /* We could possibly pass max_age in the request (as an absolute
3679 * timestamp or a "seconds.usec ago") so the target can avoid doing
3680 * extra calls into the filesystem if that isn't necessary (e.g.
3681 * during mount that would help a bit). Having relative timestamps
3682 * is not so great if request processing is slow, while absolute
3683 * timestamps are not ideal because they need time synchronization. */
3684 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3686 class_import_put(imp);
3691 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3693 ptlrpc_request_free(req);
3696 ptlrpc_request_set_replen(req);
3697 req->rq_request_portal = OST_CREATE_PORTAL;
3698 ptlrpc_at_set_req_timeout(req);
3700 if (flags & OBD_STATFS_NODELAY) {
3701 /* procfs requests not want stat in wait for avoid deadlock */
3702 req->rq_no_resend = 1;
3703 req->rq_no_delay = 1;
3706 rc = ptlrpc_queue_wait(req);
3710 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3712 GOTO(out, rc = -EPROTO);
3719 ptlrpc_req_finished(req);
3723 /* Retrieve object striping information.
3725 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3726 * the maximum number of OST indices which will fit in the user buffer.
3727 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3729 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3731 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3732 struct lov_user_md_v3 lum, *lumk;
3733 struct lov_user_ost_data_v1 *lmm_objects;
3734 int rc = 0, lum_size;
3740 /* we only need the header part from user space to get lmm_magic and
3741 * lmm_stripe_count, (the header part is common to v1 and v3) */
3742 lum_size = sizeof(struct lov_user_md_v1);
3743 if (cfs_copy_from_user(&lum, lump, lum_size))
3746 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3747 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3750 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3751 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3752 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3753 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3755 /* we can use lov_mds_md_size() to compute lum_size
3756 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3757 if (lum.lmm_stripe_count > 0) {
3758 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3759 OBD_ALLOC(lumk, lum_size);
3763 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3764 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3766 lmm_objects = &(lumk->lmm_objects[0]);
3767 lmm_objects->l_object_id = lsm->lsm_object_id;
3769 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3773 lumk->lmm_object_id = lsm->lsm_object_id;
3774 lumk->lmm_object_seq = lsm->lsm_object_seq;
3775 lumk->lmm_stripe_count = 1;
3777 if (cfs_copy_to_user(lump, lumk, lum_size))
3781 OBD_FREE(lumk, lum_size);
3787 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3788 void *karg, void *uarg)
3790 struct obd_device *obd = exp->exp_obd;
3791 struct obd_ioctl_data *data = karg;
3795 if (!cfs_try_module_get(THIS_MODULE)) {
3796 CERROR("Can't get module. Is it alive?");
3800 case OBD_IOC_LOV_GET_CONFIG: {
3802 struct lov_desc *desc;
3803 struct obd_uuid uuid;
3807 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3808 GOTO(out, err = -EINVAL);
3810 data = (struct obd_ioctl_data *)buf;
3812 if (sizeof(*desc) > data->ioc_inllen1) {
3813 obd_ioctl_freedata(buf, len);
3814 GOTO(out, err = -EINVAL);
3817 if (data->ioc_inllen2 < sizeof(uuid)) {
3818 obd_ioctl_freedata(buf, len);
3819 GOTO(out, err = -EINVAL);
3822 desc = (struct lov_desc *)data->ioc_inlbuf1;
3823 desc->ld_tgt_count = 1;
3824 desc->ld_active_tgt_count = 1;
3825 desc->ld_default_stripe_count = 1;
3826 desc->ld_default_stripe_size = 0;
3827 desc->ld_default_stripe_offset = 0;
3828 desc->ld_pattern = 0;
3829 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3831 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3833 err = cfs_copy_to_user((void *)uarg, buf, len);
3836 obd_ioctl_freedata(buf, len);
3839 case LL_IOC_LOV_SETSTRIPE:
3840 err = obd_alloc_memmd(exp, karg);
3844 case LL_IOC_LOV_GETSTRIPE:
3845 err = osc_getstripe(karg, uarg);
3847 case OBD_IOC_CLIENT_RECOVER:
3848 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3853 case IOC_OSC_SET_ACTIVE:
3854 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3857 case OBD_IOC_POLL_QUOTACHECK:
3858 err = lquota_poll_check(quota_interface, exp,
3859 (struct if_quotacheck *)karg);
3861 case OBD_IOC_PING_TARGET:
3862 err = ptlrpc_obd_ping(obd);
3865 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3866 cmd, cfs_curproc_comm());
3867 GOTO(out, err = -ENOTTY);
3870 cfs_module_put(THIS_MODULE);
3874 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3875 void *key, __u32 *vallen, void *val,
3876 struct lov_stripe_md *lsm)
3879 if (!vallen || !val)
3882 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3883 __u32 *stripe = val;
3884 *vallen = sizeof(*stripe);
3887 } else if (KEY_IS(KEY_LAST_ID)) {
3888 struct ptlrpc_request *req;
3893 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3894 &RQF_OST_GET_INFO_LAST_ID);
3898 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3899 RCL_CLIENT, keylen);
3900 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3902 ptlrpc_request_free(req);
3906 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3907 memcpy(tmp, key, keylen);
3909 req->rq_no_delay = req->rq_no_resend = 1;
3910 ptlrpc_request_set_replen(req);
3911 rc = ptlrpc_queue_wait(req);
3915 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3917 GOTO(out, rc = -EPROTO);
3919 *((obd_id *)val) = *reply;
3921 ptlrpc_req_finished(req);
3923 } else if (KEY_IS(KEY_FIEMAP)) {
3924 struct ptlrpc_request *req;
3925 struct ll_user_fiemap *reply;
3929 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3930 &RQF_OST_GET_INFO_FIEMAP);
3934 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3935 RCL_CLIENT, keylen);
3936 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3937 RCL_CLIENT, *vallen);
3938 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3939 RCL_SERVER, *vallen);
3941 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3943 ptlrpc_request_free(req);
3947 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3948 memcpy(tmp, key, keylen);
3949 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3950 memcpy(tmp, val, *vallen);
3952 ptlrpc_request_set_replen(req);
3953 rc = ptlrpc_queue_wait(req);
3957 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3959 GOTO(out1, rc = -EPROTO);
3961 memcpy(val, reply, *vallen);
3963 ptlrpc_req_finished(req);
3971 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3973 struct llog_ctxt *ctxt;
3977 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3979 rc = llog_initiator_connect(ctxt);
3980 llog_ctxt_put(ctxt);
3982 /* XXX return an error? skip setting below flags? */
3985 cfs_spin_lock(&imp->imp_lock);
3986 imp->imp_server_timeout = 1;
3987 imp->imp_pingable = 1;
3988 cfs_spin_unlock(&imp->imp_lock);
3989 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3994 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3995 struct ptlrpc_request *req,
4002 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
4005 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
4006 void *key, obd_count vallen, void *val,
4007 struct ptlrpc_request_set *set)
4009 struct ptlrpc_request *req;
4010 struct obd_device *obd = exp->exp_obd;
4011 struct obd_import *imp = class_exp2cliimp(exp);
4016 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
4018 if (KEY_IS(KEY_NEXT_ID)) {
4020 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4022 if (vallen != sizeof(obd_id))
4027 if (vallen != sizeof(obd_id))
4030 /* avoid race between allocate new object and set next id
4031 * from ll_sync thread */
4032 cfs_spin_lock(&oscc->oscc_lock);
4033 new_val = *((obd_id*)val) + 1;
4034 if (new_val > oscc->oscc_next_id)
4035 oscc->oscc_next_id = new_val;
4036 cfs_spin_unlock(&oscc->oscc_lock);
4037 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
4038 exp->exp_obd->obd_name,
4039 obd->u.cli.cl_oscc.oscc_next_id);
4044 if (KEY_IS(KEY_CHECKSUM)) {
4045 if (vallen != sizeof(int))
4047 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
4051 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4052 sptlrpc_conf_client_adapt(obd);
4056 if (KEY_IS(KEY_FLUSH_CTX)) {
4057 sptlrpc_import_flush_my_ctx(imp);
4061 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4064 /* We pass all other commands directly to OST. Since nobody calls osc
4065 methods directly and everybody is supposed to go through LOV, we
4066 assume lov checked invalid values for us.
4067 The only recognised values so far are evict_by_nid and mds_conn.
4068 Even if something bad goes through, we'd get a -EINVAL from OST
4071 if (KEY_IS(KEY_GRANT_SHRINK))
4072 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4074 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4079 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4080 RCL_CLIENT, keylen);
4081 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4082 RCL_CLIENT, vallen);
4083 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4085 ptlrpc_request_free(req);
4089 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4090 memcpy(tmp, key, keylen);
4091 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4092 memcpy(tmp, val, vallen);
4094 if (KEY_IS(KEY_MDS_CONN)) {
4095 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4097 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4098 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4099 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4100 req->rq_no_delay = req->rq_no_resend = 1;
4101 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4102 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4103 struct osc_grant_args *aa;
4106 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4107 aa = ptlrpc_req_async_args(req);
4110 ptlrpc_req_finished(req);
4113 *oa = ((struct ost_body *)val)->oa;
4115 req->rq_interpret_reply = osc_shrink_grant_interpret;
4118 ptlrpc_request_set_replen(req);
4119 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4120 LASSERT(set != NULL);
4121 ptlrpc_set_add_req(set, req);
4122 ptlrpc_check_set(NULL, set);
4124 ptlrpcd_add_req(req, PSCOPE_OTHER);
4130 static struct llog_operations osc_size_repl_logops = {
4131 lop_cancel: llog_obd_repl_cancel
4134 static struct llog_operations osc_mds_ost_orig_logops;
4136 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4137 struct obd_device *tgt, struct llog_catid *catid)
4142 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4143 &catid->lci_logid, &osc_mds_ost_orig_logops);
4145 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4149 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4150 NULL, &osc_size_repl_logops);
4152 struct llog_ctxt *ctxt =
4153 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4156 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4161 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4162 obd->obd_name, tgt->obd_name, catid, rc);
4163 CERROR("logid "LPX64":0x%x\n",
4164 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4169 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4170 struct obd_device *disk_obd, int *index)
4172 struct llog_catid catid;
4173 static char name[32] = CATLIST;
4177 LASSERT(olg == &obd->obd_olg);
4179 cfs_mutex_down(&olg->olg_cat_processing);
4180 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4182 CERROR("rc: %d\n", rc);
4186 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4187 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4188 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4190 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4192 CERROR("rc: %d\n", rc);
4196 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4198 CERROR("rc: %d\n", rc);
4203 cfs_mutex_up(&olg->olg_cat_processing);
4208 static int osc_llog_finish(struct obd_device *obd, int count)
4210 struct llog_ctxt *ctxt;
4211 int rc = 0, rc2 = 0;
4214 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4216 rc = llog_cleanup(ctxt);
4218 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4220 rc2 = llog_cleanup(ctxt);
4227 static int osc_reconnect(const struct lu_env *env,
4228 struct obd_export *exp, struct obd_device *obd,
4229 struct obd_uuid *cluuid,
4230 struct obd_connect_data *data,
4233 struct client_obd *cli = &obd->u.cli;
4235 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4238 client_obd_list_lock(&cli->cl_loi_list_lock);
4239 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4240 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4241 lost_grant = cli->cl_lost_grant;
4242 cli->cl_lost_grant = 0;
4243 client_obd_list_unlock(&cli->cl_loi_list_lock);
4245 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4246 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4247 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4248 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4249 " ocd_grant: %d\n", data->ocd_connect_flags,
4250 data->ocd_version, data->ocd_grant);
4256 static int osc_disconnect(struct obd_export *exp)
4258 struct obd_device *obd = class_exp2obd(exp);
4259 struct llog_ctxt *ctxt;
4262 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4264 if (obd->u.cli.cl_conn_count == 1) {
4265 /* Flush any remaining cancel messages out to the
4267 llog_sync(ctxt, exp);
4269 llog_ctxt_put(ctxt);
4271 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4275 rc = client_disconnect_export(exp);
4277 * Initially we put del_shrink_grant before disconnect_export, but it
4278 * causes the following problem if setup (connect) and cleanup
4279 * (disconnect) are tangled together.
4280 * connect p1 disconnect p2
4281 * ptlrpc_connect_import
4282 * ............... class_manual_cleanup
4285 * ptlrpc_connect_interrupt
4287 * add this client to shrink list
4289 * Bang! pinger trigger the shrink.
4290 * So the osc should be disconnected from the shrink list, after we
4291 * are sure the import has been destroyed. BUG18662
4293 if (obd->u.cli.cl_import == NULL)
4294 osc_del_shrink_grant(&obd->u.cli);
4298 static int osc_import_event(struct obd_device *obd,
4299 struct obd_import *imp,
4300 enum obd_import_event event)
4302 struct client_obd *cli;
4306 LASSERT(imp->imp_obd == obd);
4309 case IMP_EVENT_DISCON: {
4310 /* Only do this on the MDS OSC's */
4311 if (imp->imp_server_timeout) {
4312 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4314 cfs_spin_lock(&oscc->oscc_lock);
4315 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4316 cfs_spin_unlock(&oscc->oscc_lock);
4319 client_obd_list_lock(&cli->cl_loi_list_lock);
4320 cli->cl_avail_grant = 0;
4321 cli->cl_lost_grant = 0;
4322 client_obd_list_unlock(&cli->cl_loi_list_lock);
4325 case IMP_EVENT_INACTIVE: {
4326 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4329 case IMP_EVENT_INVALIDATE: {
4330 struct ldlm_namespace *ns = obd->obd_namespace;
4334 env = cl_env_get(&refcheck);
4338 client_obd_list_lock(&cli->cl_loi_list_lock);
4339 /* all pages go to failing rpcs due to the invalid
4341 osc_check_rpcs(env, cli);
4342 client_obd_list_unlock(&cli->cl_loi_list_lock);
4344 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4345 cl_env_put(env, &refcheck);
4350 case IMP_EVENT_ACTIVE: {
4351 /* Only do this on the MDS OSC's */
4352 if (imp->imp_server_timeout) {
4353 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4355 cfs_spin_lock(&oscc->oscc_lock);
4356 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4357 cfs_spin_unlock(&oscc->oscc_lock);
4359 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4362 case IMP_EVENT_OCD: {
4363 struct obd_connect_data *ocd = &imp->imp_connect_data;
4365 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4366 osc_init_grant(&obd->u.cli, ocd);
4369 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4370 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4372 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4375 case IMP_EVENT_DEACTIVATE: {
4376 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
4379 case IMP_EVENT_ACTIVATE: {
4380 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
4384 CERROR("Unknown import event %d\n", event);
4391 * Determine whether the lock can be canceled before replaying the lock
4392 * during recovery, see bug16774 for detailed information.
4394 * \retval zero the lock can't be canceled
4395 * \retval other ok to cancel
4397 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
4399 check_res_locked(lock->l_resource);
4402 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
4404 * XXX as a future improvement, we can also cancel unused write lock
4405 * if it doesn't have dirty data and active mmaps.
4407 if (lock->l_resource->lr_type == LDLM_EXTENT &&
4408 (lock->l_granted_mode == LCK_PR ||
4409 lock->l_granted_mode == LCK_CR) &&
4410 (osc_dlm_lock_pageref(lock) == 0))
4416 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4422 rc = ptlrpcd_addref();
4426 rc = client_obd_setup(obd, lcfg);
4430 struct lprocfs_static_vars lvars = { 0 };
4431 struct client_obd *cli = &obd->u.cli;
4433 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4434 lprocfs_osc_init_vars(&lvars);
4435 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4436 lproc_osc_attach_seqstat(obd);
4437 sptlrpc_lprocfs_cliobd_attach(obd);
4438 ptlrpc_lprocfs_register_obd(obd);
4442 /* We need to allocate a few requests more, because
4443 brw_interpret tries to create new requests before freeing
4444 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4445 reserved, but I afraid that might be too much wasted RAM
4446 in fact, so 2 is just my guess and still should work. */
4447 cli->cl_import->imp_rq_pool =
4448 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4450 ptlrpc_add_rqs_to_pool);
4452 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4453 cfs_sema_init(&cli->cl_grant_sem, 1);
4455 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
4461 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4467 case OBD_CLEANUP_EARLY: {
4468 struct obd_import *imp;
4469 imp = obd->u.cli.cl_import;
4470 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4471 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4472 ptlrpc_deactivate_import(imp);
4473 cfs_spin_lock(&imp->imp_lock);
4474 imp->imp_pingable = 0;
4475 cfs_spin_unlock(&imp->imp_lock);
4478 case OBD_CLEANUP_EXPORTS: {
4479 /* If we set up but never connected, the
4480 client import will not have been cleaned. */
4481 if (obd->u.cli.cl_import) {
4482 struct obd_import *imp;
4483 cfs_down_write(&obd->u.cli.cl_sem);
4484 imp = obd->u.cli.cl_import;
4485 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4487 ptlrpc_invalidate_import(imp);
4488 if (imp->imp_rq_pool) {
4489 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4490 imp->imp_rq_pool = NULL;
4492 class_destroy_import(imp);
4493 cfs_up_write(&obd->u.cli.cl_sem);
4494 obd->u.cli.cl_import = NULL;
4496 rc = obd_llog_finish(obd, 0);
4498 CERROR("failed to cleanup llogging subsystems\n");
4505 int osc_cleanup(struct obd_device *obd)
4510 ptlrpc_lprocfs_unregister_obd(obd);
4511 lprocfs_obd_cleanup(obd);
4513 /* free memory of osc quota cache */
4514 lquota_cleanup(quota_interface, obd);
4516 rc = client_obd_cleanup(obd);
4522 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4524 struct lprocfs_static_vars lvars = { 0 };
4527 lprocfs_osc_init_vars(&lvars);
4529 switch (lcfg->lcfg_command) {
4531 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4541 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4543 return osc_process_config_base(obd, buf);
4546 struct obd_ops osc_obd_ops = {
4547 .o_owner = THIS_MODULE,
4548 .o_setup = osc_setup,
4549 .o_precleanup = osc_precleanup,
4550 .o_cleanup = osc_cleanup,
4551 .o_add_conn = client_import_add_conn,
4552 .o_del_conn = client_import_del_conn,
4553 .o_connect = client_connect_import,
4554 .o_reconnect = osc_reconnect,
4555 .o_disconnect = osc_disconnect,
4556 .o_statfs = osc_statfs,
4557 .o_statfs_async = osc_statfs_async,
4558 .o_packmd = osc_packmd,
4559 .o_unpackmd = osc_unpackmd,
4560 .o_precreate = osc_precreate,
4561 .o_create = osc_create,
4562 .o_create_async = osc_create_async,
4563 .o_destroy = osc_destroy,
4564 .o_getattr = osc_getattr,
4565 .o_getattr_async = osc_getattr_async,
4566 .o_setattr = osc_setattr,
4567 .o_setattr_async = osc_setattr_async,
4569 .o_punch = osc_punch,
4571 .o_enqueue = osc_enqueue,
4572 .o_change_cbdata = osc_change_cbdata,
4573 .o_find_cbdata = osc_find_cbdata,
4574 .o_cancel = osc_cancel,
4575 .o_cancel_unused = osc_cancel_unused,
4576 .o_iocontrol = osc_iocontrol,
4577 .o_get_info = osc_get_info,
4578 .o_set_info_async = osc_set_info_async,
4579 .o_import_event = osc_import_event,
4580 .o_llog_init = osc_llog_init,
4581 .o_llog_finish = osc_llog_finish,
4582 .o_process_config = osc_process_config,
4585 extern struct lu_kmem_descr osc_caches[];
4586 extern cfs_spinlock_t osc_ast_guard;
4587 extern cfs_lock_class_key_t osc_ast_guard_class;
4589 int __init osc_init(void)
4591 struct lprocfs_static_vars lvars = { 0 };
4595 /* print an address of _any_ initialized kernel symbol from this
4596 * module, to allow debugging with gdb that doesn't support data
4597 * symbols from modules.*/
4598 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4600 rc = lu_kmem_init(osc_caches);
4602 lprocfs_osc_init_vars(&lvars);
4604 cfs_request_module("lquota");
4605 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4606 lquota_init(quota_interface);
4607 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4609 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4610 LUSTRE_OSC_NAME, &osc_device_type);
4612 if (quota_interface)
4613 PORTAL_SYMBOL_PUT(osc_quota_interface);
4614 lu_kmem_fini(osc_caches);
4618 cfs_spin_lock_init(&osc_ast_guard);
4619 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4621 osc_mds_ost_orig_logops = llog_lvfs_ops;
4622 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4623 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4624 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4625 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4631 static void /*__exit*/ osc_exit(void)
4633 lu_device_type_fini(&osc_device_type);
4635 lquota_exit(quota_interface);
4636 if (quota_interface)
4637 PORTAL_SYMBOL_PUT(osc_quota_interface);
4639 class_unregister_type(LUSTRE_OSC_NAME);
4640 lu_kmem_fini(osc_caches);
4643 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4644 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4645 MODULE_LICENSE("GPL");
4647 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);