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 oa->o_flags |= OBD_FL_SHRINK_GRANT;
957 osc_update_next_shrink(cli);
960 /* Shrink the current grant, either from some large amount to enough for a
961 * full set of in-flight RPCs, or if we have already shrunk to that limit
962 * then to enough for a single RPC. This avoids keeping more grant than
963 * needed, and avoids shrinking the grant piecemeal. */
964 static int osc_shrink_grant(struct client_obd *cli)
966 long target = (cli->cl_max_rpcs_in_flight + 1) *
967 cli->cl_max_pages_per_rpc;
969 client_obd_list_lock(&cli->cl_loi_list_lock);
970 if (cli->cl_avail_grant <= target)
971 target = cli->cl_max_pages_per_rpc;
972 client_obd_list_unlock(&cli->cl_loi_list_lock);
974 return osc_shrink_grant_to_target(cli, target);
977 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
980 struct ost_body *body;
983 client_obd_list_lock(&cli->cl_loi_list_lock);
984 /* Don't shrink if we are already above or below the desired limit
985 * We don't want to shrink below a single RPC, as that will negatively
986 * impact block allocation and long-term performance. */
987 if (target < cli->cl_max_pages_per_rpc)
988 target = cli->cl_max_pages_per_rpc;
990 if (target >= cli->cl_avail_grant) {
991 client_obd_list_unlock(&cli->cl_loi_list_lock);
994 client_obd_list_unlock(&cli->cl_loi_list_lock);
1000 osc_announce_cached(cli, &body->oa, 0);
1002 client_obd_list_lock(&cli->cl_loi_list_lock);
1003 body->oa.o_grant = cli->cl_avail_grant - target;
1004 cli->cl_avail_grant = target;
1005 client_obd_list_unlock(&cli->cl_loi_list_lock);
1006 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1007 osc_update_next_shrink(cli);
1009 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1010 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1011 sizeof(*body), body, NULL);
1013 __osc_update_grant(cli, body->oa.o_grant);
1018 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1019 static int osc_should_shrink_grant(struct client_obd *client)
1021 cfs_time_t time = cfs_time_current();
1022 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1024 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1025 OBD_CONNECT_GRANT_SHRINK) == 0)
1028 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1029 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1030 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1033 osc_update_next_shrink(client);
1038 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1040 struct client_obd *client;
1042 cfs_list_for_each_entry(client, &item->ti_obd_list,
1043 cl_grant_shrink_list) {
1044 if (osc_should_shrink_grant(client))
1045 osc_shrink_grant(client);
1050 static int osc_add_shrink_grant(struct client_obd *client)
1054 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1056 osc_grant_shrink_grant_cb, NULL,
1057 &client->cl_grant_shrink_list);
1059 CERROR("add grant client %s error %d\n",
1060 client->cl_import->imp_obd->obd_name, rc);
1063 CDEBUG(D_CACHE, "add grant client %s \n",
1064 client->cl_import->imp_obd->obd_name);
1065 osc_update_next_shrink(client);
1069 static int osc_del_shrink_grant(struct client_obd *client)
1071 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1075 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1078 * ocd_grant is the total grant amount we're expect to hold: if we've
1079 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1080 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1082 * race is tolerable here: if we're evicted, but imp_state already
1083 * left EVICTED state, then cl_dirty must be 0 already.
1085 client_obd_list_lock(&cli->cl_loi_list_lock);
1086 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1087 cli->cl_avail_grant = ocd->ocd_grant;
1089 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1090 client_obd_list_unlock(&cli->cl_loi_list_lock);
1092 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1093 cli->cl_avail_grant, cli->cl_lost_grant);
1094 LASSERT(cli->cl_avail_grant >= 0);
1096 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1097 cfs_list_empty(&cli->cl_grant_shrink_list))
1098 osc_add_shrink_grant(cli);
1101 /* We assume that the reason this OSC got a short read is because it read
1102 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1103 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1104 * this stripe never got written at or beyond this stripe offset yet. */
1105 static void handle_short_read(int nob_read, obd_count page_count,
1106 struct brw_page **pga)
1111 /* skip bytes read OK */
1112 while (nob_read > 0) {
1113 LASSERT (page_count > 0);
1115 if (pga[i]->count > nob_read) {
1116 /* EOF inside this page */
1117 ptr = cfs_kmap(pga[i]->pg) +
1118 (pga[i]->off & ~CFS_PAGE_MASK);
1119 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1120 cfs_kunmap(pga[i]->pg);
1126 nob_read -= pga[i]->count;
1131 /* zero remaining pages */
1132 while (page_count-- > 0) {
1133 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1134 memset(ptr, 0, pga[i]->count);
1135 cfs_kunmap(pga[i]->pg);
1140 static int check_write_rcs(struct ptlrpc_request *req,
1141 int requested_nob, int niocount,
1142 obd_count page_count, struct brw_page **pga)
1147 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1148 sizeof(*remote_rcs) *
1150 if (remote_rcs == NULL) {
1151 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1155 /* return error if any niobuf was in error */
1156 for (i = 0; i < niocount; i++) {
1157 if (remote_rcs[i] < 0)
1158 return(remote_rcs[i]);
1160 if (remote_rcs[i] != 0) {
1161 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1162 i, remote_rcs[i], req);
1167 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1168 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1169 req->rq_bulk->bd_nob_transferred, requested_nob);
1176 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1178 if (p1->flag != p2->flag) {
1179 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1180 OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1182 /* warn if we try to combine flags that we don't know to be
1183 * safe to combine */
1184 if ((p1->flag & mask) != (p2->flag & mask))
1185 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1186 "same brw?\n", p1->flag, p2->flag);
1190 return (p1->off + p1->count == p2->off);
1193 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1194 struct brw_page **pga, int opc,
1195 cksum_type_t cksum_type)
1200 LASSERT (pg_count > 0);
1201 cksum = init_checksum(cksum_type);
1202 while (nob > 0 && pg_count > 0) {
1203 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1204 int off = pga[i]->off & ~CFS_PAGE_MASK;
1205 int count = pga[i]->count > nob ? nob : pga[i]->count;
1207 /* corrupt the data before we compute the checksum, to
1208 * simulate an OST->client data error */
1209 if (i == 0 && opc == OST_READ &&
1210 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1211 memcpy(ptr + off, "bad1", min(4, nob));
1212 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1213 cfs_kunmap(pga[i]->pg);
1214 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1217 nob -= pga[i]->count;
1221 /* For sending we only compute the wrong checksum instead
1222 * of corrupting the data so it is still correct on a redo */
1223 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1229 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1230 struct lov_stripe_md *lsm, obd_count page_count,
1231 struct brw_page **pga,
1232 struct ptlrpc_request **reqp,
1233 struct obd_capa *ocapa, int reserve)
1235 struct ptlrpc_request *req;
1236 struct ptlrpc_bulk_desc *desc;
1237 struct ost_body *body;
1238 struct obd_ioobj *ioobj;
1239 struct niobuf_remote *niobuf;
1240 int niocount, i, requested_nob, opc, rc;
1241 struct osc_brw_async_args *aa;
1242 struct req_capsule *pill;
1243 struct brw_page *pg_prev;
1246 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1247 RETURN(-ENOMEM); /* Recoverable */
1248 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1249 RETURN(-EINVAL); /* Fatal */
1251 if ((cmd & OBD_BRW_WRITE) != 0) {
1253 req = ptlrpc_request_alloc_pool(cli->cl_import,
1254 cli->cl_import->imp_rq_pool,
1255 &RQF_OST_BRW_WRITE);
1258 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1263 for (niocount = i = 1; i < page_count; i++) {
1264 if (!can_merge_pages(pga[i - 1], pga[i]))
1268 pill = &req->rq_pill;
1269 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1271 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1272 niocount * sizeof(*niobuf));
1273 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1277 ptlrpc_request_free(req);
1280 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1281 ptlrpc_at_set_req_timeout(req);
1283 if (opc == OST_WRITE)
1284 desc = ptlrpc_prep_bulk_imp(req, page_count,
1285 BULK_GET_SOURCE, OST_BULK_PORTAL);
1287 desc = ptlrpc_prep_bulk_imp(req, page_count,
1288 BULK_PUT_SINK, OST_BULK_PORTAL);
1291 GOTO(out, rc = -ENOMEM);
1292 /* NB request now owns desc and will free it when it gets freed */
1294 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1295 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1296 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1297 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1299 lustre_set_wire_obdo(&body->oa, oa);
1301 obdo_to_ioobj(oa, ioobj);
1302 ioobj->ioo_bufcnt = niocount;
1303 osc_pack_capa(req, body, ocapa);
1304 LASSERT (page_count > 0);
1306 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1307 struct brw_page *pg = pga[i];
1309 LASSERT(pg->count > 0);
1310 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1311 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1312 pg->off, pg->count);
1314 LASSERTF(i == 0 || pg->off > pg_prev->off,
1315 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1316 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1318 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1319 pg_prev->pg, page_private(pg_prev->pg),
1320 pg_prev->pg->index, pg_prev->off);
1322 LASSERTF(i == 0 || pg->off > pg_prev->off,
1323 "i %d p_c %u\n", i, page_count);
1325 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1326 (pg->flag & OBD_BRW_SRVLOCK));
1328 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1330 requested_nob += pg->count;
1332 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1334 niobuf->len += pg->count;
1336 niobuf->offset = pg->off;
1337 niobuf->len = pg->count;
1338 niobuf->flags = pg->flag;
1343 LASSERTF((void *)(niobuf - niocount) ==
1344 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1345 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1346 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1348 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1349 if (osc_should_shrink_grant(cli))
1350 osc_shrink_grant_local(cli, &body->oa);
1352 /* size[REQ_REC_OFF] still sizeof (*body) */
1353 if (opc == OST_WRITE) {
1354 if (unlikely(cli->cl_checksum) &&
1355 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1356 /* store cl_cksum_type in a local variable since
1357 * it can be changed via lprocfs */
1358 cksum_type_t cksum_type = cli->cl_cksum_type;
1360 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1361 oa->o_flags &= OBD_FL_LOCAL_MASK;
1362 body->oa.o_flags = 0;
1364 body->oa.o_flags |= cksum_type_pack(cksum_type);
1365 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1366 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1370 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1372 /* save this in 'oa', too, for later checking */
1373 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1374 oa->o_flags |= cksum_type_pack(cksum_type);
1376 /* clear out the checksum flag, in case this is a
1377 * resend but cl_checksum is no longer set. b=11238 */
1378 oa->o_valid &= ~OBD_MD_FLCKSUM;
1380 oa->o_cksum = body->oa.o_cksum;
1381 /* 1 RC per niobuf */
1382 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1383 sizeof(__u32) * niocount);
1385 if (unlikely(cli->cl_checksum) &&
1386 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1387 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1388 body->oa.o_flags = 0;
1389 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1390 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1393 ptlrpc_request_set_replen(req);
1395 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1396 aa = ptlrpc_req_async_args(req);
1398 aa->aa_requested_nob = requested_nob;
1399 aa->aa_nio_count = niocount;
1400 aa->aa_page_count = page_count;
1404 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1405 if (ocapa && reserve)
1406 aa->aa_ocapa = capa_get(ocapa);
1412 ptlrpc_req_finished(req);
1416 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1417 __u32 client_cksum, __u32 server_cksum, int nob,
1418 obd_count page_count, struct brw_page **pga,
1419 cksum_type_t client_cksum_type)
1423 cksum_type_t cksum_type;
1425 if (server_cksum == client_cksum) {
1426 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1430 if (oa->o_valid & OBD_MD_FLFLAGS)
1431 cksum_type = cksum_type_unpack(oa->o_flags);
1433 cksum_type = OBD_CKSUM_CRC32;
1435 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1438 if (cksum_type != client_cksum_type)
1439 msg = "the server did not use the checksum type specified in "
1440 "the original request - likely a protocol problem";
1441 else if (new_cksum == server_cksum)
1442 msg = "changed on the client after we checksummed it - "
1443 "likely false positive due to mmap IO (bug 11742)";
1444 else if (new_cksum == client_cksum)
1445 msg = "changed in transit before arrival at OST";
1447 msg = "changed in transit AND doesn't match the original - "
1448 "likely false positive due to mmap IO (bug 11742)";
1450 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1451 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1452 msg, libcfs_nid2str(peer->nid),
1453 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1454 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1455 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1457 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1459 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1460 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1461 "client csum now %x\n", client_cksum, client_cksum_type,
1462 server_cksum, cksum_type, new_cksum);
1466 /* Note rc enters this function as number of bytes transferred */
1467 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1469 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1470 const lnet_process_id_t *peer =
1471 &req->rq_import->imp_connection->c_peer;
1472 struct client_obd *cli = aa->aa_cli;
1473 struct ost_body *body;
1474 __u32 client_cksum = 0;
1477 if (rc < 0 && rc != -EDQUOT) {
1478 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1482 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1483 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1485 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1489 #ifdef HAVE_QUOTA_SUPPORT
1490 /* set/clear over quota flag for a uid/gid */
1491 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1492 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1493 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1495 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1496 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1498 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1503 osc_update_grant(cli, body);
1508 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1509 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1511 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1513 CERROR("Unexpected +ve rc %d\n", rc);
1516 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1518 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1521 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1522 check_write_checksum(&body->oa, peer, client_cksum,
1523 body->oa.o_cksum, aa->aa_requested_nob,
1524 aa->aa_page_count, aa->aa_ppga,
1525 cksum_type_unpack(aa->aa_oa->o_flags)))
1528 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1529 aa->aa_page_count, aa->aa_ppga);
1533 /* The rest of this function executes only for OST_READs */
1535 /* if unwrap_bulk failed, return -EAGAIN to retry */
1536 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1538 GOTO(out, rc = -EAGAIN);
1540 if (rc > aa->aa_requested_nob) {
1541 CERROR("Unexpected rc %d (%d requested)\n", rc,
1542 aa->aa_requested_nob);
1546 if (rc != req->rq_bulk->bd_nob_transferred) {
1547 CERROR ("Unexpected rc %d (%d transferred)\n",
1548 rc, req->rq_bulk->bd_nob_transferred);
1552 if (rc < aa->aa_requested_nob)
1553 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1555 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1556 static int cksum_counter;
1557 __u32 server_cksum = body->oa.o_cksum;
1560 cksum_type_t cksum_type;
1562 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1563 cksum_type = cksum_type_unpack(body->oa.o_flags);
1565 cksum_type = OBD_CKSUM_CRC32;
1566 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1567 aa->aa_ppga, OST_READ,
1570 if (peer->nid == req->rq_bulk->bd_sender) {
1574 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1577 if (server_cksum == ~0 && rc > 0) {
1578 CERROR("Protocol error: server %s set the 'checksum' "
1579 "bit, but didn't send a checksum. Not fatal, "
1580 "but please notify on http://bugzilla.lustre.org/\n",
1581 libcfs_nid2str(peer->nid));
1582 } else if (server_cksum != client_cksum) {
1583 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1584 "%s%s%s inode "DFID" object "
1585 LPU64"/"LPU64" extent "
1586 "["LPU64"-"LPU64"]\n",
1587 req->rq_import->imp_obd->obd_name,
1588 libcfs_nid2str(peer->nid),
1590 body->oa.o_valid & OBD_MD_FLFID ?
1591 body->oa.o_parent_seq : (__u64)0,
1592 body->oa.o_valid & OBD_MD_FLFID ?
1593 body->oa.o_parent_oid : 0,
1594 body->oa.o_valid & OBD_MD_FLFID ?
1595 body->oa.o_parent_ver : 0,
1597 body->oa.o_valid & OBD_MD_FLGROUP ?
1598 body->oa.o_seq : (__u64)0,
1599 aa->aa_ppga[0]->off,
1600 aa->aa_ppga[aa->aa_page_count-1]->off +
1601 aa->aa_ppga[aa->aa_page_count-1]->count -
1603 CERROR("client %x, server %x, cksum_type %x\n",
1604 client_cksum, server_cksum, cksum_type);
1606 aa->aa_oa->o_cksum = client_cksum;
1610 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1613 } else if (unlikely(client_cksum)) {
1614 static int cksum_missed;
1617 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1618 CERROR("Checksum %u requested from %s but not sent\n",
1619 cksum_missed, libcfs_nid2str(peer->nid));
1625 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1630 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1631 struct lov_stripe_md *lsm,
1632 obd_count page_count, struct brw_page **pga,
1633 struct obd_capa *ocapa)
1635 struct ptlrpc_request *req;
1639 struct l_wait_info lwi;
1643 cfs_waitq_init(&waitq);
1646 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1647 page_count, pga, &req, ocapa, 0);
1651 rc = ptlrpc_queue_wait(req);
1653 if (rc == -ETIMEDOUT && req->rq_resend) {
1654 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1655 ptlrpc_req_finished(req);
1659 rc = osc_brw_fini_request(req, rc);
1661 ptlrpc_req_finished(req);
1662 if (osc_recoverable_error(rc)) {
1664 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1665 CERROR("too many resend retries, returning error\n");
1669 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1670 l_wait_event(waitq, 0, &lwi);
1678 int osc_brw_redo_request(struct ptlrpc_request *request,
1679 struct osc_brw_async_args *aa)
1681 struct ptlrpc_request *new_req;
1682 struct ptlrpc_request_set *set = request->rq_set;
1683 struct osc_brw_async_args *new_aa;
1684 struct osc_async_page *oap;
1688 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1689 CERROR("too many resent retries, returning error\n");
1693 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1695 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1696 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1697 aa->aa_cli, aa->aa_oa,
1698 NULL /* lsm unused by osc currently */,
1699 aa->aa_page_count, aa->aa_ppga,
1700 &new_req, aa->aa_ocapa, 0);
1704 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1706 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1707 if (oap->oap_request != NULL) {
1708 LASSERTF(request == oap->oap_request,
1709 "request %p != oap_request %p\n",
1710 request, oap->oap_request);
1711 if (oap->oap_interrupted) {
1712 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1713 ptlrpc_req_finished(new_req);
1718 /* New request takes over pga and oaps from old request.
1719 * Note that copying a list_head doesn't work, need to move it... */
1721 new_req->rq_interpret_reply = request->rq_interpret_reply;
1722 new_req->rq_async_args = request->rq_async_args;
1723 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1725 new_aa = ptlrpc_req_async_args(new_req);
1727 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1728 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1729 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1731 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1732 if (oap->oap_request) {
1733 ptlrpc_req_finished(oap->oap_request);
1734 oap->oap_request = ptlrpc_request_addref(new_req);
1738 new_aa->aa_ocapa = aa->aa_ocapa;
1739 aa->aa_ocapa = NULL;
1741 /* use ptlrpc_set_add_req is safe because interpret functions work
1742 * in check_set context. only one way exist with access to request
1743 * from different thread got -EINTR - this way protected with
1744 * cl_loi_list_lock */
1745 ptlrpc_set_add_req(set, new_req);
1747 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1749 DEBUG_REQ(D_INFO, new_req, "new request");
1754 * ugh, we want disk allocation on the target to happen in offset order. we'll
1755 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1756 * fine for our small page arrays and doesn't require allocation. its an
1757 * insertion sort that swaps elements that are strides apart, shrinking the
1758 * stride down until its '1' and the array is sorted.
1760 static void sort_brw_pages(struct brw_page **array, int num)
1763 struct brw_page *tmp;
1767 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1772 for (i = stride ; i < num ; i++) {
1775 while (j >= stride && array[j - stride]->off > tmp->off) {
1776 array[j] = array[j - stride];
1781 } while (stride > 1);
1784 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1790 LASSERT (pages > 0);
1791 offset = pg[i]->off & ~CFS_PAGE_MASK;
1795 if (pages == 0) /* that's all */
1798 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1799 return count; /* doesn't end on page boundary */
1802 offset = pg[i]->off & ~CFS_PAGE_MASK;
1803 if (offset != 0) /* doesn't start on page boundary */
1810 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1812 struct brw_page **ppga;
1815 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1819 for (i = 0; i < count; i++)
1824 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1826 LASSERT(ppga != NULL);
1827 OBD_FREE(ppga, sizeof(*ppga) * count);
1830 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1831 obd_count page_count, struct brw_page *pga,
1832 struct obd_trans_info *oti)
1834 struct obdo *saved_oa = NULL;
1835 struct brw_page **ppga, **orig;
1836 struct obd_import *imp = class_exp2cliimp(exp);
1837 struct client_obd *cli;
1838 int rc, page_count_orig;
1841 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1842 cli = &imp->imp_obd->u.cli;
1844 if (cmd & OBD_BRW_CHECK) {
1845 /* The caller just wants to know if there's a chance that this
1846 * I/O can succeed */
1848 if (imp->imp_invalid)
1853 /* test_brw with a failed create can trip this, maybe others. */
1854 LASSERT(cli->cl_max_pages_per_rpc);
1858 orig = ppga = osc_build_ppga(pga, page_count);
1861 page_count_orig = page_count;
1863 sort_brw_pages(ppga, page_count);
1864 while (page_count) {
1865 obd_count pages_per_brw;
1867 if (page_count > cli->cl_max_pages_per_rpc)
1868 pages_per_brw = cli->cl_max_pages_per_rpc;
1870 pages_per_brw = page_count;
1872 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1874 if (saved_oa != NULL) {
1875 /* restore previously saved oa */
1876 *oinfo->oi_oa = *saved_oa;
1877 } else if (page_count > pages_per_brw) {
1878 /* save a copy of oa (brw will clobber it) */
1879 OBDO_ALLOC(saved_oa);
1880 if (saved_oa == NULL)
1881 GOTO(out, rc = -ENOMEM);
1882 *saved_oa = *oinfo->oi_oa;
1885 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1886 pages_per_brw, ppga, oinfo->oi_capa);
1891 page_count -= pages_per_brw;
1892 ppga += pages_per_brw;
1896 osc_release_ppga(orig, page_count_orig);
1898 if (saved_oa != NULL)
1899 OBDO_FREE(saved_oa);
1904 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1905 * the dirty accounting. Writeback completes or truncate happens before
1906 * writing starts. Must be called with the loi lock held. */
1907 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1910 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1914 /* This maintains the lists of pending pages to read/write for a given object
1915 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1916 * to quickly find objects that are ready to send an RPC. */
1917 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1923 if (lop->lop_num_pending == 0)
1926 /* if we have an invalid import we want to drain the queued pages
1927 * by forcing them through rpcs that immediately fail and complete
1928 * the pages. recovery relies on this to empty the queued pages
1929 * before canceling the locks and evicting down the llite pages */
1930 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1933 /* stream rpcs in queue order as long as as there is an urgent page
1934 * queued. this is our cheap solution for good batching in the case
1935 * where writepage marks some random page in the middle of the file
1936 * as urgent because of, say, memory pressure */
1937 if (!cfs_list_empty(&lop->lop_urgent)) {
1938 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1941 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1942 optimal = cli->cl_max_pages_per_rpc;
1943 if (cmd & OBD_BRW_WRITE) {
1944 /* trigger a write rpc stream as long as there are dirtiers
1945 * waiting for space. as they're waiting, they're not going to
1946 * create more pages to coallesce with what's waiting.. */
1947 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1948 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1951 /* +16 to avoid triggering rpcs that would want to include pages
1952 * that are being queued but which can't be made ready until
1953 * the queuer finishes with the page. this is a wart for
1954 * llite::commit_write() */
1957 if (lop->lop_num_pending >= optimal)
1963 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1965 struct osc_async_page *oap;
1968 if (cfs_list_empty(&lop->lop_urgent))
1971 oap = cfs_list_entry(lop->lop_urgent.next,
1972 struct osc_async_page, oap_urgent_item);
1974 if (oap->oap_async_flags & ASYNC_HP) {
1975 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1982 static void on_list(cfs_list_t *item, cfs_list_t *list,
1985 if (cfs_list_empty(item) && should_be_on)
1986 cfs_list_add_tail(item, list);
1987 else if (!cfs_list_empty(item) && !should_be_on)
1988 cfs_list_del_init(item);
1991 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1992 * can find pages to build into rpcs quickly */
1993 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1995 if (lop_makes_hprpc(&loi->loi_write_lop) ||
1996 lop_makes_hprpc(&loi->loi_read_lop)) {
1998 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1999 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2001 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2002 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2003 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2004 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2007 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2008 loi->loi_write_lop.lop_num_pending);
2010 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2011 loi->loi_read_lop.lop_num_pending);
2014 static void lop_update_pending(struct client_obd *cli,
2015 struct loi_oap_pages *lop, int cmd, int delta)
2017 lop->lop_num_pending += delta;
2018 if (cmd & OBD_BRW_WRITE)
2019 cli->cl_pending_w_pages += delta;
2021 cli->cl_pending_r_pages += delta;
2025 * this is called when a sync waiter receives an interruption. Its job is to
2026 * get the caller woken as soon as possible. If its page hasn't been put in an
2027 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2028 * desiring interruption which will forcefully complete the rpc once the rpc
2031 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2033 struct loi_oap_pages *lop;
2034 struct lov_oinfo *loi;
2038 LASSERT(!oap->oap_interrupted);
2039 oap->oap_interrupted = 1;
2041 /* ok, it's been put in an rpc. only one oap gets a request reference */
2042 if (oap->oap_request != NULL) {
2043 ptlrpc_mark_interrupted(oap->oap_request);
2044 ptlrpcd_wake(oap->oap_request);
2045 ptlrpc_req_finished(oap->oap_request);
2046 oap->oap_request = NULL;
2050 * page completion may be called only if ->cpo_prep() method was
2051 * executed by osc_io_submit(), that also adds page the to pending list
2053 if (!cfs_list_empty(&oap->oap_pending_item)) {
2054 cfs_list_del_init(&oap->oap_pending_item);
2055 cfs_list_del_init(&oap->oap_urgent_item);
2058 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2059 &loi->loi_write_lop : &loi->loi_read_lop;
2060 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2061 loi_list_maint(oap->oap_cli, oap->oap_loi);
2062 rc = oap->oap_caller_ops->ap_completion(env,
2063 oap->oap_caller_data,
2064 oap->oap_cmd, NULL, -EINTR);
2070 /* this is trying to propogate async writeback errors back up to the
2071 * application. As an async write fails we record the error code for later if
2072 * the app does an fsync. As long as errors persist we force future rpcs to be
2073 * sync so that the app can get a sync error and break the cycle of queueing
2074 * pages for which writeback will fail. */
2075 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2082 ar->ar_force_sync = 1;
2083 ar->ar_min_xid = ptlrpc_sample_next_xid();
2088 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2089 ar->ar_force_sync = 0;
2092 void osc_oap_to_pending(struct osc_async_page *oap)
2094 struct loi_oap_pages *lop;
2096 if (oap->oap_cmd & OBD_BRW_WRITE)
2097 lop = &oap->oap_loi->loi_write_lop;
2099 lop = &oap->oap_loi->loi_read_lop;
2101 if (oap->oap_async_flags & ASYNC_HP)
2102 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2103 else if (oap->oap_async_flags & ASYNC_URGENT)
2104 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2105 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2106 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2109 /* this must be called holding the loi list lock to give coverage to exit_cache,
2110 * async_flag maintenance, and oap_request */
2111 static void osc_ap_completion(const struct lu_env *env,
2112 struct client_obd *cli, struct obdo *oa,
2113 struct osc_async_page *oap, int sent, int rc)
2118 if (oap->oap_request != NULL) {
2119 xid = ptlrpc_req_xid(oap->oap_request);
2120 ptlrpc_req_finished(oap->oap_request);
2121 oap->oap_request = NULL;
2124 cfs_spin_lock(&oap->oap_lock);
2125 oap->oap_async_flags = 0;
2126 cfs_spin_unlock(&oap->oap_lock);
2127 oap->oap_interrupted = 0;
2129 if (oap->oap_cmd & OBD_BRW_WRITE) {
2130 osc_process_ar(&cli->cl_ar, xid, rc);
2131 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2134 if (rc == 0 && oa != NULL) {
2135 if (oa->o_valid & OBD_MD_FLBLOCKS)
2136 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2137 if (oa->o_valid & OBD_MD_FLMTIME)
2138 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2139 if (oa->o_valid & OBD_MD_FLATIME)
2140 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2141 if (oa->o_valid & OBD_MD_FLCTIME)
2142 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2145 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2146 oap->oap_cmd, oa, rc);
2148 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2149 * I/O on the page could start, but OSC calls it under lock
2150 * and thus we can add oap back to pending safely */
2152 /* upper layer wants to leave the page on pending queue */
2153 osc_oap_to_pending(oap);
2155 osc_exit_cache(cli, oap, sent);
2159 static int brw_interpret(const struct lu_env *env,
2160 struct ptlrpc_request *req, void *data, int rc)
2162 struct osc_brw_async_args *aa = data;
2163 struct client_obd *cli;
2167 rc = osc_brw_fini_request(req, rc);
2168 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2169 if (osc_recoverable_error(rc)) {
2170 rc = osc_brw_redo_request(req, aa);
2176 capa_put(aa->aa_ocapa);
2177 aa->aa_ocapa = NULL;
2182 client_obd_list_lock(&cli->cl_loi_list_lock);
2184 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2185 * is called so we know whether to go to sync BRWs or wait for more
2186 * RPCs to complete */
2187 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2188 cli->cl_w_in_flight--;
2190 cli->cl_r_in_flight--;
2192 async = cfs_list_empty(&aa->aa_oaps);
2193 if (!async) { /* from osc_send_oap_rpc() */
2194 struct osc_async_page *oap, *tmp;
2195 /* the caller may re-use the oap after the completion call so
2196 * we need to clean it up a little */
2197 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2199 cfs_list_del_init(&oap->oap_rpc_item);
2200 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2202 OBDO_FREE(aa->aa_oa);
2203 } else { /* from async_internal() */
2205 for (i = 0; i < aa->aa_page_count; i++)
2206 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2208 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2209 OBDO_FREE(aa->aa_oa);
2211 osc_wake_cache_waiters(cli);
2212 osc_check_rpcs(env, cli);
2213 client_obd_list_unlock(&cli->cl_loi_list_lock);
2215 cl_req_completion(env, aa->aa_clerq, rc);
2216 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2220 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2221 struct client_obd *cli,
2222 cfs_list_t *rpc_list,
2223 int page_count, int cmd)
2225 struct ptlrpc_request *req;
2226 struct brw_page **pga = NULL;
2227 struct osc_brw_async_args *aa;
2228 struct obdo *oa = NULL;
2229 const struct obd_async_page_ops *ops = NULL;
2230 void *caller_data = NULL;
2231 struct osc_async_page *oap;
2232 struct osc_async_page *tmp;
2233 struct ost_body *body;
2234 struct cl_req *clerq = NULL;
2235 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2236 struct ldlm_lock *lock = NULL;
2237 struct cl_req_attr crattr;
2241 LASSERT(!cfs_list_empty(rpc_list));
2243 memset(&crattr, 0, sizeof crattr);
2244 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2246 GOTO(out, req = ERR_PTR(-ENOMEM));
2250 GOTO(out, req = ERR_PTR(-ENOMEM));
2253 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2254 struct cl_page *page = osc_oap2cl_page(oap);
2256 ops = oap->oap_caller_ops;
2257 caller_data = oap->oap_caller_data;
2259 clerq = cl_req_alloc(env, page, crt,
2260 1 /* only 1-object rpcs for
2263 GOTO(out, req = (void *)clerq);
2264 lock = oap->oap_ldlm_lock;
2266 pga[i] = &oap->oap_brw_page;
2267 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2268 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2269 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2271 cl_req_page_add(env, clerq, page);
2274 /* always get the data for the obdo for the rpc */
2275 LASSERT(ops != NULL);
2277 crattr.cra_capa = NULL;
2278 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2280 oa->o_handle = lock->l_remote_handle;
2281 oa->o_valid |= OBD_MD_FLHANDLE;
2284 rc = cl_req_prep(env, clerq);
2286 CERROR("cl_req_prep failed: %d\n", rc);
2287 GOTO(out, req = ERR_PTR(rc));
2290 sort_brw_pages(pga, page_count);
2291 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2292 pga, &req, crattr.cra_capa, 1);
2294 CERROR("prep_req failed: %d\n", rc);
2295 GOTO(out, req = ERR_PTR(rc));
2298 /* Need to update the timestamps after the request is built in case
2299 * we race with setattr (locally or in queue at OST). If OST gets
2300 * later setattr before earlier BRW (as determined by the request xid),
2301 * the OST will not use BRW timestamps. Sadly, there is no obvious
2302 * way to do this in a single call. bug 10150 */
2303 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2304 cl_req_attr_set(env, clerq, &crattr,
2305 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2307 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2308 aa = ptlrpc_req_async_args(req);
2309 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2310 cfs_list_splice(rpc_list, &aa->aa_oaps);
2311 CFS_INIT_LIST_HEAD(rpc_list);
2312 aa->aa_clerq = clerq;
2314 capa_put(crattr.cra_capa);
2319 OBD_FREE(pga, sizeof(*pga) * page_count);
2320 /* this should happen rarely and is pretty bad, it makes the
2321 * pending list not follow the dirty order */
2322 client_obd_list_lock(&cli->cl_loi_list_lock);
2323 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2324 cfs_list_del_init(&oap->oap_rpc_item);
2326 /* queued sync pages can be torn down while the pages
2327 * were between the pending list and the rpc */
2328 if (oap->oap_interrupted) {
2329 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2330 osc_ap_completion(env, cli, NULL, oap, 0,
2334 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2336 if (clerq && !IS_ERR(clerq))
2337 cl_req_completion(env, clerq, PTR_ERR(req));
2343 * prepare pages for ASYNC io and put pages in send queue.
2345 * \param cmd OBD_BRW_* macroses
2346 * \param lop pending pages
2348 * \return zero if pages successfully add to send queue.
2349 * \return not zere if error occurring.
2352 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2353 struct lov_oinfo *loi,
2354 int cmd, struct loi_oap_pages *lop)
2356 struct ptlrpc_request *req;
2357 obd_count page_count = 0;
2358 struct osc_async_page *oap = NULL, *tmp;
2359 struct osc_brw_async_args *aa;
2360 const struct obd_async_page_ops *ops;
2361 CFS_LIST_HEAD(rpc_list);
2362 CFS_LIST_HEAD(tmp_list);
2363 unsigned int ending_offset;
2364 unsigned starting_offset = 0;
2366 struct cl_object *clob = NULL;
2369 /* ASYNC_HP pages first. At present, when the lock the pages is
2370 * to be canceled, the pages covered by the lock will be sent out
2371 * with ASYNC_HP. We have to send out them as soon as possible. */
2372 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2373 if (oap->oap_async_flags & ASYNC_HP)
2374 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2376 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2377 if (++page_count >= cli->cl_max_pages_per_rpc)
2381 cfs_list_splice(&tmp_list, &lop->lop_pending);
2384 /* first we find the pages we're allowed to work with */
2385 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2387 ops = oap->oap_caller_ops;
2389 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2390 "magic 0x%x\n", oap, oap->oap_magic);
2393 /* pin object in memory, so that completion call-backs
2394 * can be safely called under client_obd_list lock. */
2395 clob = osc_oap2cl_page(oap)->cp_obj;
2396 cl_object_get(clob);
2399 if (page_count != 0 &&
2400 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2401 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2402 " oap %p, page %p, srvlock %u\n",
2403 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2407 /* If there is a gap at the start of this page, it can't merge
2408 * with any previous page, so we'll hand the network a
2409 * "fragmented" page array that it can't transfer in 1 RDMA */
2410 if (page_count != 0 && oap->oap_page_off != 0)
2413 /* in llite being 'ready' equates to the page being locked
2414 * until completion unlocks it. commit_write submits a page
2415 * as not ready because its unlock will happen unconditionally
2416 * as the call returns. if we race with commit_write giving
2417 * us that page we dont' want to create a hole in the page
2418 * stream, so we stop and leave the rpc to be fired by
2419 * another dirtier or kupdated interval (the not ready page
2420 * will still be on the dirty list). we could call in
2421 * at the end of ll_file_write to process the queue again. */
2422 if (!(oap->oap_async_flags & ASYNC_READY)) {
2423 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2426 CDEBUG(D_INODE, "oap %p page %p returned %d "
2427 "instead of ready\n", oap,
2431 /* llite is telling us that the page is still
2432 * in commit_write and that we should try
2433 * and put it in an rpc again later. we
2434 * break out of the loop so we don't create
2435 * a hole in the sequence of pages in the rpc
2440 /* the io isn't needed.. tell the checks
2441 * below to complete the rpc with EINTR */
2442 cfs_spin_lock(&oap->oap_lock);
2443 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2444 cfs_spin_unlock(&oap->oap_lock);
2445 oap->oap_count = -EINTR;
2448 cfs_spin_lock(&oap->oap_lock);
2449 oap->oap_async_flags |= ASYNC_READY;
2450 cfs_spin_unlock(&oap->oap_lock);
2453 LASSERTF(0, "oap %p page %p returned %d "
2454 "from make_ready\n", oap,
2462 * Page submitted for IO has to be locked. Either by
2463 * ->ap_make_ready() or by higher layers.
2465 #if defined(__KERNEL__) && defined(__linux__)
2467 struct cl_page *page;
2469 page = osc_oap2cl_page(oap);
2471 if (page->cp_type == CPT_CACHEABLE &&
2472 !(PageLocked(oap->oap_page) &&
2473 (CheckWriteback(oap->oap_page, cmd)))) {
2474 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2476 (long)oap->oap_page->flags,
2477 oap->oap_async_flags);
2483 /* take the page out of our book-keeping */
2484 cfs_list_del_init(&oap->oap_pending_item);
2485 lop_update_pending(cli, lop, cmd, -1);
2486 cfs_list_del_init(&oap->oap_urgent_item);
2488 if (page_count == 0)
2489 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2490 (PTLRPC_MAX_BRW_SIZE - 1);
2492 /* ask the caller for the size of the io as the rpc leaves. */
2493 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2495 ops->ap_refresh_count(env, oap->oap_caller_data,
2497 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2499 if (oap->oap_count <= 0) {
2500 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2502 osc_ap_completion(env, cli, NULL,
2503 oap, 0, oap->oap_count);
2507 /* now put the page back in our accounting */
2508 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2509 if (page_count == 0)
2510 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2511 if (++page_count >= cli->cl_max_pages_per_rpc)
2514 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2515 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2516 * have the same alignment as the initial writes that allocated
2517 * extents on the server. */
2518 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2519 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2520 if (ending_offset == 0)
2523 /* If there is a gap at the end of this page, it can't merge
2524 * with any subsequent pages, so we'll hand the network a
2525 * "fragmented" page array that it can't transfer in 1 RDMA */
2526 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2530 osc_wake_cache_waiters(cli);
2532 loi_list_maint(cli, loi);
2534 client_obd_list_unlock(&cli->cl_loi_list_lock);
2537 cl_object_put(env, clob);
2539 if (page_count == 0) {
2540 client_obd_list_lock(&cli->cl_loi_list_lock);
2544 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2546 LASSERT(cfs_list_empty(&rpc_list));
2547 loi_list_maint(cli, loi);
2548 RETURN(PTR_ERR(req));
2551 aa = ptlrpc_req_async_args(req);
2553 if (cmd == OBD_BRW_READ) {
2554 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2555 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2556 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2557 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2559 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2560 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2561 cli->cl_w_in_flight);
2562 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2563 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2565 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2567 client_obd_list_lock(&cli->cl_loi_list_lock);
2569 if (cmd == OBD_BRW_READ)
2570 cli->cl_r_in_flight++;
2572 cli->cl_w_in_flight++;
2574 /* queued sync pages can be torn down while the pages
2575 * were between the pending list and the rpc */
2577 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2578 /* only one oap gets a request reference */
2581 if (oap->oap_interrupted && !req->rq_intr) {
2582 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2584 ptlrpc_mark_interrupted(req);
2588 tmp->oap_request = ptlrpc_request_addref(req);
2590 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2591 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2593 req->rq_interpret_reply = brw_interpret;
2594 ptlrpcd_add_req(req, PSCOPE_BRW);
2598 #define LOI_DEBUG(LOI, STR, args...) \
2599 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2600 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2601 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2602 (LOI)->loi_write_lop.lop_num_pending, \
2603 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2604 (LOI)->loi_read_lop.lop_num_pending, \
2605 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2608 /* This is called by osc_check_rpcs() to find which objects have pages that
2609 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2610 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2614 /* First return objects that have blocked locks so that they
2615 * will be flushed quickly and other clients can get the lock,
2616 * then objects which have pages ready to be stuffed into RPCs */
2617 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2618 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2619 struct lov_oinfo, loi_hp_ready_item));
2620 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2621 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2622 struct lov_oinfo, loi_ready_item));
2624 /* then if we have cache waiters, return all objects with queued
2625 * writes. This is especially important when many small files
2626 * have filled up the cache and not been fired into rpcs because
2627 * they don't pass the nr_pending/object threshhold */
2628 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2629 !cfs_list_empty(&cli->cl_loi_write_list))
2630 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2631 struct lov_oinfo, loi_write_item));
2633 /* then return all queued objects when we have an invalid import
2634 * so that they get flushed */
2635 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2636 if (!cfs_list_empty(&cli->cl_loi_write_list))
2637 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2640 if (!cfs_list_empty(&cli->cl_loi_read_list))
2641 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2642 struct lov_oinfo, loi_read_item));
2647 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2649 struct osc_async_page *oap;
2652 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2653 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2654 struct osc_async_page, oap_urgent_item);
2655 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2658 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2659 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2660 struct osc_async_page, oap_urgent_item);
2661 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2664 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2667 /* called with the loi list lock held */
2668 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2670 struct lov_oinfo *loi;
2671 int rc = 0, race_counter = 0;
2674 while ((loi = osc_next_loi(cli)) != NULL) {
2675 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2677 if (osc_max_rpc_in_flight(cli, loi))
2680 /* attempt some read/write balancing by alternating between
2681 * reads and writes in an object. The makes_rpc checks here
2682 * would be redundant if we were getting read/write work items
2683 * instead of objects. we don't want send_oap_rpc to drain a
2684 * partial read pending queue when we're given this object to
2685 * do io on writes while there are cache waiters */
2686 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2687 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2688 &loi->loi_write_lop);
2690 CERROR("Write request failed with %d\n", rc);
2692 /* osc_send_oap_rpc failed, mostly because of
2695 * It can't break here, because if:
2696 * - a page was submitted by osc_io_submit, so
2698 * - no request in flight
2699 * - no subsequent request
2700 * The system will be in live-lock state,
2701 * because there is no chance to call
2702 * osc_io_unplug() and osc_check_rpcs() any
2703 * more. pdflush can't help in this case,
2704 * because it might be blocked at grabbing
2705 * the page lock as we mentioned.
2707 * Anyway, continue to drain pages. */
2716 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2717 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2718 &loi->loi_read_lop);
2720 CERROR("Read request failed with %d\n", rc);
2728 /* attempt some inter-object balancing by issueing rpcs
2729 * for each object in turn */
2730 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2731 cfs_list_del_init(&loi->loi_hp_ready_item);
2732 if (!cfs_list_empty(&loi->loi_ready_item))
2733 cfs_list_del_init(&loi->loi_ready_item);
2734 if (!cfs_list_empty(&loi->loi_write_item))
2735 cfs_list_del_init(&loi->loi_write_item);
2736 if (!cfs_list_empty(&loi->loi_read_item))
2737 cfs_list_del_init(&loi->loi_read_item);
2739 loi_list_maint(cli, loi);
2741 /* send_oap_rpc fails with 0 when make_ready tells it to
2742 * back off. llite's make_ready does this when it tries
2743 * to lock a page queued for write that is already locked.
2744 * we want to try sending rpcs from many objects, but we
2745 * don't want to spin failing with 0. */
2746 if (race_counter == 10)
2752 /* we're trying to queue a page in the osc so we're subject to the
2753 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2754 * If the osc's queued pages are already at that limit, then we want to sleep
2755 * until there is space in the osc's queue for us. We also may be waiting for
2756 * write credits from the OST if there are RPCs in flight that may return some
2757 * before we fall back to sync writes.
2759 * We need this know our allocation was granted in the presence of signals */
2760 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2764 client_obd_list_lock(&cli->cl_loi_list_lock);
2765 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2766 client_obd_list_unlock(&cli->cl_loi_list_lock);
2771 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2774 int osc_enter_cache_try(const struct lu_env *env,
2775 struct client_obd *cli, struct lov_oinfo *loi,
2776 struct osc_async_page *oap, int transient)
2780 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2782 osc_consume_write_grant(cli, &oap->oap_brw_page);
2784 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2785 cfs_atomic_inc(&obd_dirty_transit_pages);
2786 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2792 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2793 * grant or cache space. */
2794 static int osc_enter_cache(const struct lu_env *env,
2795 struct client_obd *cli, struct lov_oinfo *loi,
2796 struct osc_async_page *oap)
2798 struct osc_cache_waiter ocw;
2799 struct l_wait_info lwi = { 0 };
2803 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2804 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2805 cli->cl_dirty_max, obd_max_dirty_pages,
2806 cli->cl_lost_grant, cli->cl_avail_grant);
2808 /* force the caller to try sync io. this can jump the list
2809 * of queued writes and create a discontiguous rpc stream */
2810 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2811 loi->loi_ar.ar_force_sync)
2814 /* Hopefully normal case - cache space and write credits available */
2815 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2816 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2817 osc_enter_cache_try(env, cli, loi, oap, 0))
2820 /* It is safe to block as a cache waiter as long as there is grant
2821 * space available or the hope of additional grant being returned
2822 * when an in flight write completes. Using the write back cache
2823 * if possible is preferable to sending the data synchronously
2824 * because write pages can then be merged in to large requests.
2825 * The addition of this cache waiter will causing pending write
2826 * pages to be sent immediately. */
2827 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2828 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2829 cfs_waitq_init(&ocw.ocw_waitq);
2833 loi_list_maint(cli, loi);
2834 osc_check_rpcs(env, cli);
2835 client_obd_list_unlock(&cli->cl_loi_list_lock);
2837 CDEBUG(D_CACHE, "sleeping for cache space\n");
2838 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2840 client_obd_list_lock(&cli->cl_loi_list_lock);
2841 if (!cfs_list_empty(&ocw.ocw_entry)) {
2842 cfs_list_del(&ocw.ocw_entry);
2852 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2853 struct lov_oinfo *loi, cfs_page_t *page,
2854 obd_off offset, const struct obd_async_page_ops *ops,
2855 void *data, void **res, int nocache,
2856 struct lustre_handle *lockh)
2858 struct osc_async_page *oap;
2863 return cfs_size_round(sizeof(*oap));
2866 oap->oap_magic = OAP_MAGIC;
2867 oap->oap_cli = &exp->exp_obd->u.cli;
2870 oap->oap_caller_ops = ops;
2871 oap->oap_caller_data = data;
2873 oap->oap_page = page;
2874 oap->oap_obj_off = offset;
2875 if (!client_is_remote(exp) &&
2876 cfs_capable(CFS_CAP_SYS_RESOURCE))
2877 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2879 LASSERT(!(offset & ~CFS_PAGE_MASK));
2881 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2882 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2883 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2884 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2886 cfs_spin_lock_init(&oap->oap_lock);
2887 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2891 struct osc_async_page *oap_from_cookie(void *cookie)
2893 struct osc_async_page *oap = cookie;
2894 if (oap->oap_magic != OAP_MAGIC)
2895 return ERR_PTR(-EINVAL);
2899 int osc_queue_async_io(const struct lu_env *env,
2900 struct obd_export *exp, struct lov_stripe_md *lsm,
2901 struct lov_oinfo *loi, void *cookie,
2902 int cmd, obd_off off, int count,
2903 obd_flag brw_flags, enum async_flags async_flags)
2905 struct client_obd *cli = &exp->exp_obd->u.cli;
2906 struct osc_async_page *oap;
2910 oap = oap_from_cookie(cookie);
2912 RETURN(PTR_ERR(oap));
2914 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2917 if (!cfs_list_empty(&oap->oap_pending_item) ||
2918 !cfs_list_empty(&oap->oap_urgent_item) ||
2919 !cfs_list_empty(&oap->oap_rpc_item))
2922 /* check if the file's owner/group is over quota */
2923 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2924 struct cl_object *obj;
2925 struct cl_attr attr; /* XXX put attr into thread info */
2926 unsigned int qid[MAXQUOTAS];
2928 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2930 cl_object_attr_lock(obj);
2931 rc = cl_object_attr_get(env, obj, &attr);
2932 cl_object_attr_unlock(obj);
2934 qid[USRQUOTA] = attr.cat_uid;
2935 qid[GRPQUOTA] = attr.cat_gid;
2937 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2944 loi = lsm->lsm_oinfo[0];
2946 client_obd_list_lock(&cli->cl_loi_list_lock);
2948 LASSERT(off + count <= CFS_PAGE_SIZE);
2950 oap->oap_page_off = off;
2951 oap->oap_count = count;
2952 oap->oap_brw_flags = brw_flags;
2953 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2954 if (libcfs_memory_pressure_get())
2955 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2956 cfs_spin_lock(&oap->oap_lock);
2957 oap->oap_async_flags = async_flags;
2958 cfs_spin_unlock(&oap->oap_lock);
2960 if (cmd & OBD_BRW_WRITE) {
2961 rc = osc_enter_cache(env, cli, loi, oap);
2963 client_obd_list_unlock(&cli->cl_loi_list_lock);
2968 osc_oap_to_pending(oap);
2969 loi_list_maint(cli, loi);
2971 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2974 osc_check_rpcs(env, cli);
2975 client_obd_list_unlock(&cli->cl_loi_list_lock);
2980 /* aka (~was & now & flag), but this is more clear :) */
2981 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2983 int osc_set_async_flags_base(struct client_obd *cli,
2984 struct lov_oinfo *loi, struct osc_async_page *oap,
2985 obd_flag async_flags)
2987 struct loi_oap_pages *lop;
2991 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
2993 if (oap->oap_cmd & OBD_BRW_WRITE) {
2994 lop = &loi->loi_write_lop;
2996 lop = &loi->loi_read_lop;
2999 if ((oap->oap_async_flags & async_flags) == async_flags)
3002 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3003 flags |= ASYNC_READY;
3005 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3006 cfs_list_empty(&oap->oap_rpc_item)) {
3007 if (oap->oap_async_flags & ASYNC_HP)
3008 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3010 cfs_list_add_tail(&oap->oap_urgent_item,
3012 flags |= ASYNC_URGENT;
3013 loi_list_maint(cli, loi);
3015 cfs_spin_lock(&oap->oap_lock);
3016 oap->oap_async_flags |= flags;
3017 cfs_spin_unlock(&oap->oap_lock);
3019 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3020 oap->oap_async_flags);
3024 int osc_teardown_async_page(struct obd_export *exp,
3025 struct lov_stripe_md *lsm,
3026 struct lov_oinfo *loi, void *cookie)
3028 struct client_obd *cli = &exp->exp_obd->u.cli;
3029 struct loi_oap_pages *lop;
3030 struct osc_async_page *oap;
3034 oap = oap_from_cookie(cookie);
3036 RETURN(PTR_ERR(oap));
3039 loi = lsm->lsm_oinfo[0];
3041 if (oap->oap_cmd & OBD_BRW_WRITE) {
3042 lop = &loi->loi_write_lop;
3044 lop = &loi->loi_read_lop;
3047 client_obd_list_lock(&cli->cl_loi_list_lock);
3049 if (!cfs_list_empty(&oap->oap_rpc_item))
3050 GOTO(out, rc = -EBUSY);
3052 osc_exit_cache(cli, oap, 0);
3053 osc_wake_cache_waiters(cli);
3055 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3056 cfs_list_del_init(&oap->oap_urgent_item);
3057 cfs_spin_lock(&oap->oap_lock);
3058 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3059 cfs_spin_unlock(&oap->oap_lock);
3061 if (!cfs_list_empty(&oap->oap_pending_item)) {
3062 cfs_list_del_init(&oap->oap_pending_item);
3063 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3065 loi_list_maint(cli, loi);
3066 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3068 client_obd_list_unlock(&cli->cl_loi_list_lock);
3072 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3073 struct ldlm_enqueue_info *einfo,
3076 void *data = einfo->ei_cbdata;
3078 LASSERT(lock != NULL);
3079 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3080 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3081 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3082 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3084 lock_res_and_lock(lock);
3085 cfs_spin_lock(&osc_ast_guard);
3086 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3087 lock->l_ast_data = data;
3088 cfs_spin_unlock(&osc_ast_guard);
3089 unlock_res_and_lock(lock);
3092 static void osc_set_data_with_check(struct lustre_handle *lockh,
3093 struct ldlm_enqueue_info *einfo,
3096 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3099 osc_set_lock_data_with_check(lock, einfo, flags);
3100 LDLM_LOCK_PUT(lock);
3102 CERROR("lockh %p, data %p - client evicted?\n",
3103 lockh, einfo->ei_cbdata);
3106 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3107 ldlm_iterator_t replace, void *data)
3109 struct ldlm_res_id res_id;
3110 struct obd_device *obd = class_exp2obd(exp);
3112 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3113 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3117 /* find any ldlm lock of the inode in osc
3121 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3122 ldlm_iterator_t replace, void *data)
3124 struct ldlm_res_id res_id;
3125 struct obd_device *obd = class_exp2obd(exp);
3128 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3129 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3130 if (rc == LDLM_ITER_STOP)
3132 if (rc == LDLM_ITER_CONTINUE)
3137 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3138 obd_enqueue_update_f upcall, void *cookie,
3141 int intent = *flags & LDLM_FL_HAS_INTENT;
3145 /* The request was created before ldlm_cli_enqueue call. */
3146 if (rc == ELDLM_LOCK_ABORTED) {
3147 struct ldlm_reply *rep;
3148 rep = req_capsule_server_get(&req->rq_pill,
3151 LASSERT(rep != NULL);
3152 if (rep->lock_policy_res1)
3153 rc = rep->lock_policy_res1;
3157 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3158 *flags |= LDLM_FL_LVB_READY;
3159 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3160 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3163 /* Call the update callback. */
3164 rc = (*upcall)(cookie, rc);
3168 static int osc_enqueue_interpret(const struct lu_env *env,
3169 struct ptlrpc_request *req,
3170 struct osc_enqueue_args *aa, int rc)
3172 struct ldlm_lock *lock;
3173 struct lustre_handle handle;
3176 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3177 * might be freed anytime after lock upcall has been called. */
3178 lustre_handle_copy(&handle, aa->oa_lockh);
3179 mode = aa->oa_ei->ei_mode;
3181 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3183 lock = ldlm_handle2lock(&handle);
3185 /* Take an additional reference so that a blocking AST that
3186 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3187 * to arrive after an upcall has been executed by
3188 * osc_enqueue_fini(). */
3189 ldlm_lock_addref(&handle, mode);
3191 /* Complete obtaining the lock procedure. */
3192 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3193 mode, aa->oa_flags, aa->oa_lvb,
3194 sizeof(*aa->oa_lvb), &handle, rc);
3195 /* Complete osc stuff. */
3196 rc = osc_enqueue_fini(req, aa->oa_lvb,
3197 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3199 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3201 /* Release the lock for async request. */
3202 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3204 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3205 * not already released by
3206 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3208 ldlm_lock_decref(&handle, mode);
3210 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3211 aa->oa_lockh, req, aa);
3212 ldlm_lock_decref(&handle, mode);
3213 LDLM_LOCK_PUT(lock);
3217 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3218 struct lov_oinfo *loi, int flags,
3219 struct ost_lvb *lvb, __u32 mode, int rc)
3221 if (rc == ELDLM_OK) {
3222 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3225 LASSERT(lock != NULL);
3226 loi->loi_lvb = *lvb;
3227 tmp = loi->loi_lvb.lvb_size;
3228 /* Extend KMS up to the end of this lock and no further
3229 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3230 if (tmp > lock->l_policy_data.l_extent.end)
3231 tmp = lock->l_policy_data.l_extent.end + 1;
3232 if (tmp >= loi->loi_kms) {
3233 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3234 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3235 loi_kms_set(loi, tmp);
3237 LDLM_DEBUG(lock, "lock acquired, setting rss="
3238 LPU64"; leaving kms="LPU64", end="LPU64,
3239 loi->loi_lvb.lvb_size, loi->loi_kms,
3240 lock->l_policy_data.l_extent.end);
3242 ldlm_lock_allow_match(lock);
3243 LDLM_LOCK_PUT(lock);
3244 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3245 loi->loi_lvb = *lvb;
3246 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3247 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3251 EXPORT_SYMBOL(osc_update_enqueue);
3253 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3255 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3256 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3257 * other synchronous requests, however keeping some locks and trying to obtain
3258 * others may take a considerable amount of time in a case of ost failure; and
3259 * when other sync requests do not get released lock from a client, the client
3260 * is excluded from the cluster -- such scenarious make the life difficult, so
3261 * release locks just after they are obtained. */
3262 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3263 int *flags, ldlm_policy_data_t *policy,
3264 struct ost_lvb *lvb, int kms_valid,
3265 obd_enqueue_update_f upcall, void *cookie,
3266 struct ldlm_enqueue_info *einfo,
3267 struct lustre_handle *lockh,
3268 struct ptlrpc_request_set *rqset, int async)
3270 struct obd_device *obd = exp->exp_obd;
3271 struct ptlrpc_request *req = NULL;
3272 int intent = *flags & LDLM_FL_HAS_INTENT;
3277 /* Filesystem lock extents are extended to page boundaries so that
3278 * dealing with the page cache is a little smoother. */
3279 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3280 policy->l_extent.end |= ~CFS_PAGE_MASK;
3283 * kms is not valid when either object is completely fresh (so that no
3284 * locks are cached), or object was evicted. In the latter case cached
3285 * lock cannot be used, because it would prime inode state with
3286 * potentially stale LVB.
3291 /* Next, search for already existing extent locks that will cover us */
3292 /* If we're trying to read, we also search for an existing PW lock. The
3293 * VFS and page cache already protect us locally, so lots of readers/
3294 * writers can share a single PW lock.
3296 * There are problems with conversion deadlocks, so instead of
3297 * converting a read lock to a write lock, we'll just enqueue a new
3300 * At some point we should cancel the read lock instead of making them
3301 * send us a blocking callback, but there are problems with canceling
3302 * locks out from other users right now, too. */
3303 mode = einfo->ei_mode;
3304 if (einfo->ei_mode == LCK_PR)
3306 mode = ldlm_lock_match(obd->obd_namespace,
3307 *flags | LDLM_FL_LVB_READY, res_id,
3308 einfo->ei_type, policy, mode, lockh, 0);
3310 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3312 if (matched->l_ast_data == NULL ||
3313 matched->l_ast_data == einfo->ei_cbdata) {
3314 /* addref the lock only if not async requests and PW
3315 * lock is matched whereas we asked for PR. */
3316 if (!rqset && einfo->ei_mode != mode)
3317 ldlm_lock_addref(lockh, LCK_PR);
3318 osc_set_lock_data_with_check(matched, einfo, *flags);
3320 /* I would like to be able to ASSERT here that
3321 * rss <= kms, but I can't, for reasons which
3322 * are explained in lov_enqueue() */
3325 /* We already have a lock, and it's referenced */
3326 (*upcall)(cookie, ELDLM_OK);
3328 /* For async requests, decref the lock. */
3329 if (einfo->ei_mode != mode)
3330 ldlm_lock_decref(lockh, LCK_PW);
3332 ldlm_lock_decref(lockh, einfo->ei_mode);
3333 LDLM_LOCK_PUT(matched);
3336 ldlm_lock_decref(lockh, mode);
3337 LDLM_LOCK_PUT(matched);
3342 CFS_LIST_HEAD(cancels);
3343 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3344 &RQF_LDLM_ENQUEUE_LVB);
3348 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3352 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3354 ptlrpc_request_set_replen(req);
3357 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3358 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3360 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3361 sizeof(*lvb), lockh, async);
3364 struct osc_enqueue_args *aa;
3365 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3366 aa = ptlrpc_req_async_args(req);
3369 aa->oa_flags = flags;
3370 aa->oa_upcall = upcall;
3371 aa->oa_cookie = cookie;
3373 aa->oa_lockh = lockh;
3375 req->rq_interpret_reply =
3376 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3377 if (rqset == PTLRPCD_SET)
3378 ptlrpcd_add_req(req, PSCOPE_OTHER);
3380 ptlrpc_set_add_req(rqset, req);
3381 } else if (intent) {
3382 ptlrpc_req_finished(req);
3387 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3389 ptlrpc_req_finished(req);
3394 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3395 struct ldlm_enqueue_info *einfo,
3396 struct ptlrpc_request_set *rqset)
3398 struct ldlm_res_id res_id;
3402 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3403 oinfo->oi_md->lsm_object_seq, &res_id);
3405 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3406 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3407 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3408 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3409 rqset, rqset != NULL);
3413 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3414 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3415 int *flags, void *data, struct lustre_handle *lockh,
3418 struct obd_device *obd = exp->exp_obd;
3419 int lflags = *flags;
3423 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3426 /* Filesystem lock extents are extended to page boundaries so that
3427 * dealing with the page cache is a little smoother */
3428 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3429 policy->l_extent.end |= ~CFS_PAGE_MASK;
3431 /* Next, search for already existing extent locks that will cover us */
3432 /* If we're trying to read, we also search for an existing PW lock. The
3433 * VFS and page cache already protect us locally, so lots of readers/
3434 * writers can share a single PW lock. */
3438 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3439 res_id, type, policy, rc, lockh, unref);
3442 osc_set_data_with_check(lockh, data, lflags);
3443 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3444 ldlm_lock_addref(lockh, LCK_PR);
3445 ldlm_lock_decref(lockh, LCK_PW);
3452 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3456 if (unlikely(mode == LCK_GROUP))
3457 ldlm_lock_decref_and_cancel(lockh, mode);
3459 ldlm_lock_decref(lockh, mode);
3464 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3465 __u32 mode, struct lustre_handle *lockh)
3468 RETURN(osc_cancel_base(lockh, mode));
3471 static int osc_cancel_unused(struct obd_export *exp,
3472 struct lov_stripe_md *lsm,
3473 ldlm_cancel_flags_t flags,
3476 struct obd_device *obd = class_exp2obd(exp);
3477 struct ldlm_res_id res_id, *resp = NULL;
3480 resp = osc_build_res_name(lsm->lsm_object_id,
3481 lsm->lsm_object_seq, &res_id);
3484 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3487 static int osc_statfs_interpret(const struct lu_env *env,
3488 struct ptlrpc_request *req,
3489 struct osc_async_args *aa, int rc)
3491 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3492 struct obd_statfs *msfs;
3497 /* The request has in fact never been sent
3498 * due to issues at a higher level (LOV).
3499 * Exit immediately since the caller is
3500 * aware of the problem and takes care
3501 * of the clean up */
3504 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3505 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3511 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3513 GOTO(out, rc = -EPROTO);
3516 /* Reinitialize the RDONLY and DEGRADED flags at the client
3517 * on each statfs, so they don't stay set permanently. */
3518 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3520 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3521 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3522 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3523 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3525 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3526 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3527 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3528 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3530 /* Add a bit of hysteresis so this flag isn't continually flapping,
3531 * and ensure that new files don't get extremely fragmented due to
3532 * only a small amount of available space in the filesystem.
3533 * We want to set the NOSPC flag when there is less than ~0.1% free
3534 * and clear it when there is at least ~0.2% free space, so:
3535 * avail < ~0.1% max max = avail + used
3536 * 1025 * avail < avail + used used = blocks - free
3537 * 1024 * avail < used
3538 * 1024 * avail < blocks - free
3539 * avail < ((blocks - free) >> 10)
3541 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3542 * lose that amount of space so in those cases we report no space left
3543 * if their is less than 1 GB left. */
3544 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3545 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3546 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3547 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3548 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3549 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3550 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3552 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3554 *aa->aa_oi->oi_osfs = *msfs;
3556 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3560 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3561 __u64 max_age, struct ptlrpc_request_set *rqset)
3563 struct ptlrpc_request *req;
3564 struct osc_async_args *aa;
3568 /* We could possibly pass max_age in the request (as an absolute
3569 * timestamp or a "seconds.usec ago") so the target can avoid doing
3570 * extra calls into the filesystem if that isn't necessary (e.g.
3571 * during mount that would help a bit). Having relative timestamps
3572 * is not so great if request processing is slow, while absolute
3573 * timestamps are not ideal because they need time synchronization. */
3574 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3578 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3580 ptlrpc_request_free(req);
3583 ptlrpc_request_set_replen(req);
3584 req->rq_request_portal = OST_CREATE_PORTAL;
3585 ptlrpc_at_set_req_timeout(req);
3587 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3588 /* procfs requests not want stat in wait for avoid deadlock */
3589 req->rq_no_resend = 1;
3590 req->rq_no_delay = 1;
3593 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3594 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3595 aa = ptlrpc_req_async_args(req);
3598 ptlrpc_set_add_req(rqset, req);
3602 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3603 __u64 max_age, __u32 flags)
3605 struct obd_statfs *msfs;
3606 struct ptlrpc_request *req;
3607 struct obd_import *imp = NULL;
3611 /*Since the request might also come from lprocfs, so we need
3612 *sync this with client_disconnect_export Bug15684*/
3613 cfs_down_read(&obd->u.cli.cl_sem);
3614 if (obd->u.cli.cl_import)
3615 imp = class_import_get(obd->u.cli.cl_import);
3616 cfs_up_read(&obd->u.cli.cl_sem);
3620 /* We could possibly pass max_age in the request (as an absolute
3621 * timestamp or a "seconds.usec ago") so the target can avoid doing
3622 * extra calls into the filesystem if that isn't necessary (e.g.
3623 * during mount that would help a bit). Having relative timestamps
3624 * is not so great if request processing is slow, while absolute
3625 * timestamps are not ideal because they need time synchronization. */
3626 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3628 class_import_put(imp);
3633 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3635 ptlrpc_request_free(req);
3638 ptlrpc_request_set_replen(req);
3639 req->rq_request_portal = OST_CREATE_PORTAL;
3640 ptlrpc_at_set_req_timeout(req);
3642 if (flags & OBD_STATFS_NODELAY) {
3643 /* procfs requests not want stat in wait for avoid deadlock */
3644 req->rq_no_resend = 1;
3645 req->rq_no_delay = 1;
3648 rc = ptlrpc_queue_wait(req);
3652 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3654 GOTO(out, rc = -EPROTO);
3661 ptlrpc_req_finished(req);
3665 /* Retrieve object striping information.
3667 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3668 * the maximum number of OST indices which will fit in the user buffer.
3669 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3671 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3673 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3674 struct lov_user_md_v3 lum, *lumk;
3675 struct lov_user_ost_data_v1 *lmm_objects;
3676 int rc = 0, lum_size;
3682 /* we only need the header part from user space to get lmm_magic and
3683 * lmm_stripe_count, (the header part is common to v1 and v3) */
3684 lum_size = sizeof(struct lov_user_md_v1);
3685 if (cfs_copy_from_user(&lum, lump, lum_size))
3688 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3689 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3692 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3693 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3694 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3695 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3697 /* we can use lov_mds_md_size() to compute lum_size
3698 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3699 if (lum.lmm_stripe_count > 0) {
3700 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3701 OBD_ALLOC(lumk, lum_size);
3705 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3706 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3708 lmm_objects = &(lumk->lmm_objects[0]);
3709 lmm_objects->l_object_id = lsm->lsm_object_id;
3711 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3715 lumk->lmm_object_id = lsm->lsm_object_id;
3716 lumk->lmm_object_seq = lsm->lsm_object_seq;
3717 lumk->lmm_stripe_count = 1;
3719 if (cfs_copy_to_user(lump, lumk, lum_size))
3723 OBD_FREE(lumk, lum_size);
3729 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3730 void *karg, void *uarg)
3732 struct obd_device *obd = exp->exp_obd;
3733 struct obd_ioctl_data *data = karg;
3737 if (!cfs_try_module_get(THIS_MODULE)) {
3738 CERROR("Can't get module. Is it alive?");
3742 case OBD_IOC_LOV_GET_CONFIG: {
3744 struct lov_desc *desc;
3745 struct obd_uuid uuid;
3749 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3750 GOTO(out, err = -EINVAL);
3752 data = (struct obd_ioctl_data *)buf;
3754 if (sizeof(*desc) > data->ioc_inllen1) {
3755 obd_ioctl_freedata(buf, len);
3756 GOTO(out, err = -EINVAL);
3759 if (data->ioc_inllen2 < sizeof(uuid)) {
3760 obd_ioctl_freedata(buf, len);
3761 GOTO(out, err = -EINVAL);
3764 desc = (struct lov_desc *)data->ioc_inlbuf1;
3765 desc->ld_tgt_count = 1;
3766 desc->ld_active_tgt_count = 1;
3767 desc->ld_default_stripe_count = 1;
3768 desc->ld_default_stripe_size = 0;
3769 desc->ld_default_stripe_offset = 0;
3770 desc->ld_pattern = 0;
3771 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3773 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3775 err = cfs_copy_to_user((void *)uarg, buf, len);
3778 obd_ioctl_freedata(buf, len);
3781 case LL_IOC_LOV_SETSTRIPE:
3782 err = obd_alloc_memmd(exp, karg);
3786 case LL_IOC_LOV_GETSTRIPE:
3787 err = osc_getstripe(karg, uarg);
3789 case OBD_IOC_CLIENT_RECOVER:
3790 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3795 case IOC_OSC_SET_ACTIVE:
3796 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3799 case OBD_IOC_POLL_QUOTACHECK:
3800 err = lquota_poll_check(quota_interface, exp,
3801 (struct if_quotacheck *)karg);
3803 case OBD_IOC_PING_TARGET:
3804 err = ptlrpc_obd_ping(obd);
3807 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3808 cmd, cfs_curproc_comm());
3809 GOTO(out, err = -ENOTTY);
3812 cfs_module_put(THIS_MODULE);
3816 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3817 void *key, __u32 *vallen, void *val,
3818 struct lov_stripe_md *lsm)
3821 if (!vallen || !val)
3824 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3825 __u32 *stripe = val;
3826 *vallen = sizeof(*stripe);
3829 } else if (KEY_IS(KEY_LAST_ID)) {
3830 struct ptlrpc_request *req;
3835 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3836 &RQF_OST_GET_INFO_LAST_ID);
3840 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3841 RCL_CLIENT, keylen);
3842 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3844 ptlrpc_request_free(req);
3848 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3849 memcpy(tmp, key, keylen);
3851 req->rq_no_delay = req->rq_no_resend = 1;
3852 ptlrpc_request_set_replen(req);
3853 rc = ptlrpc_queue_wait(req);
3857 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3859 GOTO(out, rc = -EPROTO);
3861 *((obd_id *)val) = *reply;
3863 ptlrpc_req_finished(req);
3865 } else if (KEY_IS(KEY_FIEMAP)) {
3866 struct ptlrpc_request *req;
3867 struct ll_user_fiemap *reply;
3871 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3872 &RQF_OST_GET_INFO_FIEMAP);
3876 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3877 RCL_CLIENT, keylen);
3878 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3879 RCL_CLIENT, *vallen);
3880 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3881 RCL_SERVER, *vallen);
3883 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3885 ptlrpc_request_free(req);
3889 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3890 memcpy(tmp, key, keylen);
3891 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3892 memcpy(tmp, val, *vallen);
3894 ptlrpc_request_set_replen(req);
3895 rc = ptlrpc_queue_wait(req);
3899 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3901 GOTO(out1, rc = -EPROTO);
3903 memcpy(val, reply, *vallen);
3905 ptlrpc_req_finished(req);
3913 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3915 struct llog_ctxt *ctxt;
3919 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3921 rc = llog_initiator_connect(ctxt);
3922 llog_ctxt_put(ctxt);
3924 /* XXX return an error? skip setting below flags? */
3927 cfs_spin_lock(&imp->imp_lock);
3928 imp->imp_server_timeout = 1;
3929 imp->imp_pingable = 1;
3930 cfs_spin_unlock(&imp->imp_lock);
3931 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3936 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3937 struct ptlrpc_request *req,
3944 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3947 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3948 void *key, obd_count vallen, void *val,
3949 struct ptlrpc_request_set *set)
3951 struct ptlrpc_request *req;
3952 struct obd_device *obd = exp->exp_obd;
3953 struct obd_import *imp = class_exp2cliimp(exp);
3958 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3960 if (KEY_IS(KEY_NEXT_ID)) {
3962 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3964 if (vallen != sizeof(obd_id))
3969 if (vallen != sizeof(obd_id))
3972 /* avoid race between allocate new object and set next id
3973 * from ll_sync thread */
3974 cfs_spin_lock(&oscc->oscc_lock);
3975 new_val = *((obd_id*)val) + 1;
3976 if (new_val > oscc->oscc_next_id)
3977 oscc->oscc_next_id = new_val;
3978 cfs_spin_unlock(&oscc->oscc_lock);
3979 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3980 exp->exp_obd->obd_name,
3981 obd->u.cli.cl_oscc.oscc_next_id);
3986 if (KEY_IS(KEY_CHECKSUM)) {
3987 if (vallen != sizeof(int))
3989 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3993 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3994 sptlrpc_conf_client_adapt(obd);
3998 if (KEY_IS(KEY_FLUSH_CTX)) {
3999 sptlrpc_import_flush_my_ctx(imp);
4003 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4006 /* We pass all other commands directly to OST. Since nobody calls osc
4007 methods directly and everybody is supposed to go through LOV, we
4008 assume lov checked invalid values for us.
4009 The only recognised values so far are evict_by_nid and mds_conn.
4010 Even if something bad goes through, we'd get a -EINVAL from OST
4013 if (KEY_IS(KEY_GRANT_SHRINK))
4014 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4016 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4021 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4022 RCL_CLIENT, keylen);
4023 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4024 RCL_CLIENT, vallen);
4025 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4027 ptlrpc_request_free(req);
4031 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4032 memcpy(tmp, key, keylen);
4033 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4034 memcpy(tmp, val, vallen);
4036 if (KEY_IS(KEY_MDS_CONN)) {
4037 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4039 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4040 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4041 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4042 req->rq_no_delay = req->rq_no_resend = 1;
4043 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4044 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4045 struct osc_grant_args *aa;
4048 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4049 aa = ptlrpc_req_async_args(req);
4052 ptlrpc_req_finished(req);
4055 *oa = ((struct ost_body *)val)->oa;
4057 req->rq_interpret_reply = osc_shrink_grant_interpret;
4060 ptlrpc_request_set_replen(req);
4061 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4062 LASSERT(set != NULL);
4063 ptlrpc_set_add_req(set, req);
4064 ptlrpc_check_set(NULL, set);
4066 ptlrpcd_add_req(req, PSCOPE_OTHER);
4072 static struct llog_operations osc_size_repl_logops = {
4073 lop_cancel: llog_obd_repl_cancel
4076 static struct llog_operations osc_mds_ost_orig_logops;
4078 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4079 struct obd_device *tgt, struct llog_catid *catid)
4084 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4085 &catid->lci_logid, &osc_mds_ost_orig_logops);
4087 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4091 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4092 NULL, &osc_size_repl_logops);
4094 struct llog_ctxt *ctxt =
4095 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4098 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4103 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4104 obd->obd_name, tgt->obd_name, catid, rc);
4105 CERROR("logid "LPX64":0x%x\n",
4106 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4111 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4112 struct obd_device *disk_obd, int *index)
4114 struct llog_catid catid;
4115 static char name[32] = CATLIST;
4119 LASSERT(olg == &obd->obd_olg);
4121 cfs_mutex_down(&olg->olg_cat_processing);
4122 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4124 CERROR("rc: %d\n", rc);
4128 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4129 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4130 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4132 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4134 CERROR("rc: %d\n", rc);
4138 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4140 CERROR("rc: %d\n", rc);
4145 cfs_mutex_up(&olg->olg_cat_processing);
4150 static int osc_llog_finish(struct obd_device *obd, int count)
4152 struct llog_ctxt *ctxt;
4153 int rc = 0, rc2 = 0;
4156 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4158 rc = llog_cleanup(ctxt);
4160 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4162 rc2 = llog_cleanup(ctxt);
4169 static int osc_reconnect(const struct lu_env *env,
4170 struct obd_export *exp, struct obd_device *obd,
4171 struct obd_uuid *cluuid,
4172 struct obd_connect_data *data,
4175 struct client_obd *cli = &obd->u.cli;
4177 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4180 client_obd_list_lock(&cli->cl_loi_list_lock);
4181 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4182 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4183 lost_grant = cli->cl_lost_grant;
4184 cli->cl_lost_grant = 0;
4185 client_obd_list_unlock(&cli->cl_loi_list_lock);
4187 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4188 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4189 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4190 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4191 " ocd_grant: %d\n", data->ocd_connect_flags,
4192 data->ocd_version, data->ocd_grant);
4198 static int osc_disconnect(struct obd_export *exp)
4200 struct obd_device *obd = class_exp2obd(exp);
4201 struct llog_ctxt *ctxt;
4204 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4206 if (obd->u.cli.cl_conn_count == 1) {
4207 /* Flush any remaining cancel messages out to the
4209 llog_sync(ctxt, exp);
4211 llog_ctxt_put(ctxt);
4213 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4217 rc = client_disconnect_export(exp);
4219 * Initially we put del_shrink_grant before disconnect_export, but it
4220 * causes the following problem if setup (connect) and cleanup
4221 * (disconnect) are tangled together.
4222 * connect p1 disconnect p2
4223 * ptlrpc_connect_import
4224 * ............... class_manual_cleanup
4227 * ptlrpc_connect_interrupt
4229 * add this client to shrink list
4231 * Bang! pinger trigger the shrink.
4232 * So the osc should be disconnected from the shrink list, after we
4233 * are sure the import has been destroyed. BUG18662
4235 if (obd->u.cli.cl_import == NULL)
4236 osc_del_shrink_grant(&obd->u.cli);
4240 static int osc_import_event(struct obd_device *obd,
4241 struct obd_import *imp,
4242 enum obd_import_event event)
4244 struct client_obd *cli;
4248 LASSERT(imp->imp_obd == obd);
4251 case IMP_EVENT_DISCON: {
4252 /* Only do this on the MDS OSC's */
4253 if (imp->imp_server_timeout) {
4254 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4256 cfs_spin_lock(&oscc->oscc_lock);
4257 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4258 cfs_spin_unlock(&oscc->oscc_lock);
4261 client_obd_list_lock(&cli->cl_loi_list_lock);
4262 cli->cl_avail_grant = 0;
4263 cli->cl_lost_grant = 0;
4264 client_obd_list_unlock(&cli->cl_loi_list_lock);
4267 case IMP_EVENT_INACTIVE: {
4268 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4271 case IMP_EVENT_INVALIDATE: {
4272 struct ldlm_namespace *ns = obd->obd_namespace;
4276 env = cl_env_get(&refcheck);
4280 client_obd_list_lock(&cli->cl_loi_list_lock);
4281 /* all pages go to failing rpcs due to the invalid
4283 osc_check_rpcs(env, cli);
4284 client_obd_list_unlock(&cli->cl_loi_list_lock);
4286 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4287 cl_env_put(env, &refcheck);
4292 case IMP_EVENT_ACTIVE: {
4293 /* Only do this on the MDS OSC's */
4294 if (imp->imp_server_timeout) {
4295 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4297 cfs_spin_lock(&oscc->oscc_lock);
4298 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4299 cfs_spin_unlock(&oscc->oscc_lock);
4301 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4304 case IMP_EVENT_OCD: {
4305 struct obd_connect_data *ocd = &imp->imp_connect_data;
4307 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4308 osc_init_grant(&obd->u.cli, ocd);
4311 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4312 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4314 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4318 CERROR("Unknown import event %d\n", event);
4324 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4330 rc = ptlrpcd_addref();
4334 rc = client_obd_setup(obd, lcfg);
4338 struct lprocfs_static_vars lvars = { 0 };
4339 struct client_obd *cli = &obd->u.cli;
4341 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4342 lprocfs_osc_init_vars(&lvars);
4343 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4344 lproc_osc_attach_seqstat(obd);
4345 sptlrpc_lprocfs_cliobd_attach(obd);
4346 ptlrpc_lprocfs_register_obd(obd);
4350 /* We need to allocate a few requests more, because
4351 brw_interpret tries to create new requests before freeing
4352 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4353 reserved, but I afraid that might be too much wasted RAM
4354 in fact, so 2 is just my guess and still should work. */
4355 cli->cl_import->imp_rq_pool =
4356 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4358 ptlrpc_add_rqs_to_pool);
4360 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4361 cfs_sema_init(&cli->cl_grant_sem, 1);
4367 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4373 case OBD_CLEANUP_EARLY: {
4374 struct obd_import *imp;
4375 imp = obd->u.cli.cl_import;
4376 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4377 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4378 ptlrpc_deactivate_import(imp);
4379 cfs_spin_lock(&imp->imp_lock);
4380 imp->imp_pingable = 0;
4381 cfs_spin_unlock(&imp->imp_lock);
4384 case OBD_CLEANUP_EXPORTS: {
4385 /* If we set up but never connected, the
4386 client import will not have been cleaned. */
4387 if (obd->u.cli.cl_import) {
4388 struct obd_import *imp;
4389 cfs_down_write(&obd->u.cli.cl_sem);
4390 imp = obd->u.cli.cl_import;
4391 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4393 ptlrpc_invalidate_import(imp);
4394 if (imp->imp_rq_pool) {
4395 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4396 imp->imp_rq_pool = NULL;
4398 class_destroy_import(imp);
4399 cfs_up_write(&obd->u.cli.cl_sem);
4400 obd->u.cli.cl_import = NULL;
4402 rc = obd_llog_finish(obd, 0);
4404 CERROR("failed to cleanup llogging subsystems\n");
4411 int osc_cleanup(struct obd_device *obd)
4416 ptlrpc_lprocfs_unregister_obd(obd);
4417 lprocfs_obd_cleanup(obd);
4419 /* free memory of osc quota cache */
4420 lquota_cleanup(quota_interface, obd);
4422 rc = client_obd_cleanup(obd);
4428 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4430 struct lprocfs_static_vars lvars = { 0 };
4433 lprocfs_osc_init_vars(&lvars);
4435 switch (lcfg->lcfg_command) {
4437 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4447 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4449 return osc_process_config_base(obd, buf);
4452 struct obd_ops osc_obd_ops = {
4453 .o_owner = THIS_MODULE,
4454 .o_setup = osc_setup,
4455 .o_precleanup = osc_precleanup,
4456 .o_cleanup = osc_cleanup,
4457 .o_add_conn = client_import_add_conn,
4458 .o_del_conn = client_import_del_conn,
4459 .o_connect = client_connect_import,
4460 .o_reconnect = osc_reconnect,
4461 .o_disconnect = osc_disconnect,
4462 .o_statfs = osc_statfs,
4463 .o_statfs_async = osc_statfs_async,
4464 .o_packmd = osc_packmd,
4465 .o_unpackmd = osc_unpackmd,
4466 .o_precreate = osc_precreate,
4467 .o_create = osc_create,
4468 .o_create_async = osc_create_async,
4469 .o_destroy = osc_destroy,
4470 .o_getattr = osc_getattr,
4471 .o_getattr_async = osc_getattr_async,
4472 .o_setattr = osc_setattr,
4473 .o_setattr_async = osc_setattr_async,
4475 .o_punch = osc_punch,
4477 .o_enqueue = osc_enqueue,
4478 .o_change_cbdata = osc_change_cbdata,
4479 .o_find_cbdata = osc_find_cbdata,
4480 .o_cancel = osc_cancel,
4481 .o_cancel_unused = osc_cancel_unused,
4482 .o_iocontrol = osc_iocontrol,
4483 .o_get_info = osc_get_info,
4484 .o_set_info_async = osc_set_info_async,
4485 .o_import_event = osc_import_event,
4486 .o_llog_init = osc_llog_init,
4487 .o_llog_finish = osc_llog_finish,
4488 .o_process_config = osc_process_config,
4491 extern struct lu_kmem_descr osc_caches[];
4492 extern cfs_spinlock_t osc_ast_guard;
4493 extern cfs_lock_class_key_t osc_ast_guard_class;
4495 int __init osc_init(void)
4497 struct lprocfs_static_vars lvars = { 0 };
4501 /* print an address of _any_ initialized kernel symbol from this
4502 * module, to allow debugging with gdb that doesn't support data
4503 * symbols from modules.*/
4504 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4506 rc = lu_kmem_init(osc_caches);
4508 lprocfs_osc_init_vars(&lvars);
4510 cfs_request_module("lquota");
4511 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4512 lquota_init(quota_interface);
4513 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4515 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4516 LUSTRE_OSC_NAME, &osc_device_type);
4518 if (quota_interface)
4519 PORTAL_SYMBOL_PUT(osc_quota_interface);
4520 lu_kmem_fini(osc_caches);
4524 cfs_spin_lock_init(&osc_ast_guard);
4525 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4527 osc_mds_ost_orig_logops = llog_lvfs_ops;
4528 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4529 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4530 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4531 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4537 static void /*__exit*/ osc_exit(void)
4539 lu_device_type_fini(&osc_device_type);
4541 lquota_exit(quota_interface);
4542 if (quota_interface)
4543 PORTAL_SYMBOL_PUT(osc_quota_interface);
4545 class_unregister_type(LUSTRE_OSC_NAME);
4546 lu_kmem_fini(osc_caches);
4549 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4550 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4551 MODULE_LICENSE("GPL");
4553 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);