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 this is mmaped file - it can be changed at any time */
1431 if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1434 if (oa->o_valid & OBD_MD_FLFLAGS)
1435 cksum_type = cksum_type_unpack(oa->o_flags);
1437 cksum_type = OBD_CKSUM_CRC32;
1439 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1442 if (cksum_type != client_cksum_type)
1443 msg = "the server did not use the checksum type specified in "
1444 "the original request - likely a protocol problem";
1445 else if (new_cksum == server_cksum)
1446 msg = "changed on the client after we checksummed it - "
1447 "likely false positive due to mmap IO (bug 11742)";
1448 else if (new_cksum == client_cksum)
1449 msg = "changed in transit before arrival at OST";
1451 msg = "changed in transit AND doesn't match the original - "
1452 "likely false positive due to mmap IO (bug 11742)";
1454 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1455 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1456 msg, libcfs_nid2str(peer->nid),
1457 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1458 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1459 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1461 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1463 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1464 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1465 "client csum now %x\n", client_cksum, client_cksum_type,
1466 server_cksum, cksum_type, new_cksum);
1470 /* Note rc enters this function as number of bytes transferred */
1471 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1473 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1474 const lnet_process_id_t *peer =
1475 &req->rq_import->imp_connection->c_peer;
1476 struct client_obd *cli = aa->aa_cli;
1477 struct ost_body *body;
1478 __u32 client_cksum = 0;
1481 if (rc < 0 && rc != -EDQUOT) {
1482 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1486 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1487 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1489 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1493 #ifdef HAVE_QUOTA_SUPPORT
1494 /* set/clear over quota flag for a uid/gid */
1495 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1496 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1497 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1499 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1500 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1502 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1507 osc_update_grant(cli, body);
1512 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1513 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1515 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1517 CERROR("Unexpected +ve rc %d\n", rc);
1520 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1522 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1525 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1526 check_write_checksum(&body->oa, peer, client_cksum,
1527 body->oa.o_cksum, aa->aa_requested_nob,
1528 aa->aa_page_count, aa->aa_ppga,
1529 cksum_type_unpack(aa->aa_oa->o_flags)))
1532 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1533 aa->aa_page_count, aa->aa_ppga);
1537 /* The rest of this function executes only for OST_READs */
1539 /* if unwrap_bulk failed, return -EAGAIN to retry */
1540 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1542 GOTO(out, rc = -EAGAIN);
1544 if (rc > aa->aa_requested_nob) {
1545 CERROR("Unexpected rc %d (%d requested)\n", rc,
1546 aa->aa_requested_nob);
1550 if (rc != req->rq_bulk->bd_nob_transferred) {
1551 CERROR ("Unexpected rc %d (%d transferred)\n",
1552 rc, req->rq_bulk->bd_nob_transferred);
1556 if (rc < aa->aa_requested_nob)
1557 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1559 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1560 static int cksum_counter;
1561 __u32 server_cksum = body->oa.o_cksum;
1564 cksum_type_t cksum_type;
1566 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1567 cksum_type = cksum_type_unpack(body->oa.o_flags);
1569 cksum_type = OBD_CKSUM_CRC32;
1570 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1571 aa->aa_ppga, OST_READ,
1574 if (peer->nid == req->rq_bulk->bd_sender) {
1578 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1581 if (server_cksum == ~0 && rc > 0) {
1582 CERROR("Protocol error: server %s set the 'checksum' "
1583 "bit, but didn't send a checksum. Not fatal, "
1584 "but please notify on http://bugzilla.lustre.org/\n",
1585 libcfs_nid2str(peer->nid));
1586 } else if (server_cksum != client_cksum) {
1587 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1588 "%s%s%s inode "DFID" object "
1589 LPU64"/"LPU64" extent "
1590 "["LPU64"-"LPU64"]\n",
1591 req->rq_import->imp_obd->obd_name,
1592 libcfs_nid2str(peer->nid),
1594 body->oa.o_valid & OBD_MD_FLFID ?
1595 body->oa.o_parent_seq : (__u64)0,
1596 body->oa.o_valid & OBD_MD_FLFID ?
1597 body->oa.o_parent_oid : 0,
1598 body->oa.o_valid & OBD_MD_FLFID ?
1599 body->oa.o_parent_ver : 0,
1601 body->oa.o_valid & OBD_MD_FLGROUP ?
1602 body->oa.o_seq : (__u64)0,
1603 aa->aa_ppga[0]->off,
1604 aa->aa_ppga[aa->aa_page_count-1]->off +
1605 aa->aa_ppga[aa->aa_page_count-1]->count -
1607 CERROR("client %x, server %x, cksum_type %x\n",
1608 client_cksum, server_cksum, cksum_type);
1610 aa->aa_oa->o_cksum = client_cksum;
1614 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1617 } else if (unlikely(client_cksum)) {
1618 static int cksum_missed;
1621 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1622 CERROR("Checksum %u requested from %s but not sent\n",
1623 cksum_missed, libcfs_nid2str(peer->nid));
1629 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1634 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1635 struct lov_stripe_md *lsm,
1636 obd_count page_count, struct brw_page **pga,
1637 struct obd_capa *ocapa)
1639 struct ptlrpc_request *req;
1643 struct l_wait_info lwi;
1647 cfs_waitq_init(&waitq);
1650 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1651 page_count, pga, &req, ocapa, 0);
1655 rc = ptlrpc_queue_wait(req);
1657 if (rc == -ETIMEDOUT && req->rq_resend) {
1658 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1659 ptlrpc_req_finished(req);
1663 rc = osc_brw_fini_request(req, rc);
1665 ptlrpc_req_finished(req);
1666 if (osc_recoverable_error(rc)) {
1668 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1669 CERROR("too many resend retries, returning error\n");
1673 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1674 l_wait_event(waitq, 0, &lwi);
1682 int osc_brw_redo_request(struct ptlrpc_request *request,
1683 struct osc_brw_async_args *aa)
1685 struct ptlrpc_request *new_req;
1686 struct ptlrpc_request_set *set = request->rq_set;
1687 struct osc_brw_async_args *new_aa;
1688 struct osc_async_page *oap;
1692 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1693 CERROR("too many resent retries, returning error\n");
1697 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1699 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1700 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1701 aa->aa_cli, aa->aa_oa,
1702 NULL /* lsm unused by osc currently */,
1703 aa->aa_page_count, aa->aa_ppga,
1704 &new_req, aa->aa_ocapa, 0);
1708 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1710 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1711 if (oap->oap_request != NULL) {
1712 LASSERTF(request == oap->oap_request,
1713 "request %p != oap_request %p\n",
1714 request, oap->oap_request);
1715 if (oap->oap_interrupted) {
1716 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1717 ptlrpc_req_finished(new_req);
1722 /* New request takes over pga and oaps from old request.
1723 * Note that copying a list_head doesn't work, need to move it... */
1725 new_req->rq_interpret_reply = request->rq_interpret_reply;
1726 new_req->rq_async_args = request->rq_async_args;
1727 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1729 new_aa = ptlrpc_req_async_args(new_req);
1731 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1732 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1733 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1735 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1736 if (oap->oap_request) {
1737 ptlrpc_req_finished(oap->oap_request);
1738 oap->oap_request = ptlrpc_request_addref(new_req);
1742 new_aa->aa_ocapa = aa->aa_ocapa;
1743 aa->aa_ocapa = NULL;
1745 /* use ptlrpc_set_add_req is safe because interpret functions work
1746 * in check_set context. only one way exist with access to request
1747 * from different thread got -EINTR - this way protected with
1748 * cl_loi_list_lock */
1749 ptlrpc_set_add_req(set, new_req);
1751 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1753 DEBUG_REQ(D_INFO, new_req, "new request");
1758 * ugh, we want disk allocation on the target to happen in offset order. we'll
1759 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1760 * fine for our small page arrays and doesn't require allocation. its an
1761 * insertion sort that swaps elements that are strides apart, shrinking the
1762 * stride down until its '1' and the array is sorted.
1764 static void sort_brw_pages(struct brw_page **array, int num)
1767 struct brw_page *tmp;
1771 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1776 for (i = stride ; i < num ; i++) {
1779 while (j >= stride && array[j - stride]->off > tmp->off) {
1780 array[j] = array[j - stride];
1785 } while (stride > 1);
1788 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1794 LASSERT (pages > 0);
1795 offset = pg[i]->off & ~CFS_PAGE_MASK;
1799 if (pages == 0) /* that's all */
1802 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1803 return count; /* doesn't end on page boundary */
1806 offset = pg[i]->off & ~CFS_PAGE_MASK;
1807 if (offset != 0) /* doesn't start on page boundary */
1814 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1816 struct brw_page **ppga;
1819 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1823 for (i = 0; i < count; i++)
1828 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1830 LASSERT(ppga != NULL);
1831 OBD_FREE(ppga, sizeof(*ppga) * count);
1834 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1835 obd_count page_count, struct brw_page *pga,
1836 struct obd_trans_info *oti)
1838 struct obdo *saved_oa = NULL;
1839 struct brw_page **ppga, **orig;
1840 struct obd_import *imp = class_exp2cliimp(exp);
1841 struct client_obd *cli;
1842 int rc, page_count_orig;
1845 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1846 cli = &imp->imp_obd->u.cli;
1848 if (cmd & OBD_BRW_CHECK) {
1849 /* The caller just wants to know if there's a chance that this
1850 * I/O can succeed */
1852 if (imp->imp_invalid)
1857 /* test_brw with a failed create can trip this, maybe others. */
1858 LASSERT(cli->cl_max_pages_per_rpc);
1862 orig = ppga = osc_build_ppga(pga, page_count);
1865 page_count_orig = page_count;
1867 sort_brw_pages(ppga, page_count);
1868 while (page_count) {
1869 obd_count pages_per_brw;
1871 if (page_count > cli->cl_max_pages_per_rpc)
1872 pages_per_brw = cli->cl_max_pages_per_rpc;
1874 pages_per_brw = page_count;
1876 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1878 if (saved_oa != NULL) {
1879 /* restore previously saved oa */
1880 *oinfo->oi_oa = *saved_oa;
1881 } else if (page_count > pages_per_brw) {
1882 /* save a copy of oa (brw will clobber it) */
1883 OBDO_ALLOC(saved_oa);
1884 if (saved_oa == NULL)
1885 GOTO(out, rc = -ENOMEM);
1886 *saved_oa = *oinfo->oi_oa;
1889 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1890 pages_per_brw, ppga, oinfo->oi_capa);
1895 page_count -= pages_per_brw;
1896 ppga += pages_per_brw;
1900 osc_release_ppga(orig, page_count_orig);
1902 if (saved_oa != NULL)
1903 OBDO_FREE(saved_oa);
1908 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1909 * the dirty accounting. Writeback completes or truncate happens before
1910 * writing starts. Must be called with the loi lock held. */
1911 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1914 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1918 /* This maintains the lists of pending pages to read/write for a given object
1919 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1920 * to quickly find objects that are ready to send an RPC. */
1921 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1927 if (lop->lop_num_pending == 0)
1930 /* if we have an invalid import we want to drain the queued pages
1931 * by forcing them through rpcs that immediately fail and complete
1932 * the pages. recovery relies on this to empty the queued pages
1933 * before canceling the locks and evicting down the llite pages */
1934 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1937 /* stream rpcs in queue order as long as as there is an urgent page
1938 * queued. this is our cheap solution for good batching in the case
1939 * where writepage marks some random page in the middle of the file
1940 * as urgent because of, say, memory pressure */
1941 if (!cfs_list_empty(&lop->lop_urgent)) {
1942 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1945 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1946 optimal = cli->cl_max_pages_per_rpc;
1947 if (cmd & OBD_BRW_WRITE) {
1948 /* trigger a write rpc stream as long as there are dirtiers
1949 * waiting for space. as they're waiting, they're not going to
1950 * create more pages to coalesce with what's waiting.. */
1951 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1952 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1955 /* +16 to avoid triggering rpcs that would want to include pages
1956 * that are being queued but which can't be made ready until
1957 * the queuer finishes with the page. this is a wart for
1958 * llite::commit_write() */
1961 if (lop->lop_num_pending >= optimal)
1967 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1969 struct osc_async_page *oap;
1972 if (cfs_list_empty(&lop->lop_urgent))
1975 oap = cfs_list_entry(lop->lop_urgent.next,
1976 struct osc_async_page, oap_urgent_item);
1978 if (oap->oap_async_flags & ASYNC_HP) {
1979 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1986 static void on_list(cfs_list_t *item, cfs_list_t *list,
1989 if (cfs_list_empty(item) && should_be_on)
1990 cfs_list_add_tail(item, list);
1991 else if (!cfs_list_empty(item) && !should_be_on)
1992 cfs_list_del_init(item);
1995 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1996 * can find pages to build into rpcs quickly */
1997 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1999 if (lop_makes_hprpc(&loi->loi_write_lop) ||
2000 lop_makes_hprpc(&loi->loi_read_lop)) {
2002 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2003 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2005 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2006 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2007 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2008 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2011 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2012 loi->loi_write_lop.lop_num_pending);
2014 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2015 loi->loi_read_lop.lop_num_pending);
2018 static void lop_update_pending(struct client_obd *cli,
2019 struct loi_oap_pages *lop, int cmd, int delta)
2021 lop->lop_num_pending += delta;
2022 if (cmd & OBD_BRW_WRITE)
2023 cli->cl_pending_w_pages += delta;
2025 cli->cl_pending_r_pages += delta;
2029 * this is called when a sync waiter receives an interruption. Its job is to
2030 * get the caller woken as soon as possible. If its page hasn't been put in an
2031 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2032 * desiring interruption which will forcefully complete the rpc once the rpc
2035 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2037 struct loi_oap_pages *lop;
2038 struct lov_oinfo *loi;
2042 LASSERT(!oap->oap_interrupted);
2043 oap->oap_interrupted = 1;
2045 /* ok, it's been put in an rpc. only one oap gets a request reference */
2046 if (oap->oap_request != NULL) {
2047 ptlrpc_mark_interrupted(oap->oap_request);
2048 ptlrpcd_wake(oap->oap_request);
2049 ptlrpc_req_finished(oap->oap_request);
2050 oap->oap_request = NULL;
2054 * page completion may be called only if ->cpo_prep() method was
2055 * executed by osc_io_submit(), that also adds page the to pending list
2057 if (!cfs_list_empty(&oap->oap_pending_item)) {
2058 cfs_list_del_init(&oap->oap_pending_item);
2059 cfs_list_del_init(&oap->oap_urgent_item);
2062 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2063 &loi->loi_write_lop : &loi->loi_read_lop;
2064 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2065 loi_list_maint(oap->oap_cli, oap->oap_loi);
2066 rc = oap->oap_caller_ops->ap_completion(env,
2067 oap->oap_caller_data,
2068 oap->oap_cmd, NULL, -EINTR);
2074 /* this is trying to propogate async writeback errors back up to the
2075 * application. As an async write fails we record the error code for later if
2076 * the app does an fsync. As long as errors persist we force future rpcs to be
2077 * sync so that the app can get a sync error and break the cycle of queueing
2078 * pages for which writeback will fail. */
2079 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2086 ar->ar_force_sync = 1;
2087 ar->ar_min_xid = ptlrpc_sample_next_xid();
2092 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2093 ar->ar_force_sync = 0;
2096 void osc_oap_to_pending(struct osc_async_page *oap)
2098 struct loi_oap_pages *lop;
2100 if (oap->oap_cmd & OBD_BRW_WRITE)
2101 lop = &oap->oap_loi->loi_write_lop;
2103 lop = &oap->oap_loi->loi_read_lop;
2105 if (oap->oap_async_flags & ASYNC_HP)
2106 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2107 else if (oap->oap_async_flags & ASYNC_URGENT)
2108 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2109 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2110 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2113 /* this must be called holding the loi list lock to give coverage to exit_cache,
2114 * async_flag maintenance, and oap_request */
2115 static void osc_ap_completion(const struct lu_env *env,
2116 struct client_obd *cli, struct obdo *oa,
2117 struct osc_async_page *oap, int sent, int rc)
2122 if (oap->oap_request != NULL) {
2123 xid = ptlrpc_req_xid(oap->oap_request);
2124 ptlrpc_req_finished(oap->oap_request);
2125 oap->oap_request = NULL;
2128 cfs_spin_lock(&oap->oap_lock);
2129 oap->oap_async_flags = 0;
2130 cfs_spin_unlock(&oap->oap_lock);
2131 oap->oap_interrupted = 0;
2133 if (oap->oap_cmd & OBD_BRW_WRITE) {
2134 osc_process_ar(&cli->cl_ar, xid, rc);
2135 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2138 if (rc == 0 && oa != NULL) {
2139 if (oa->o_valid & OBD_MD_FLBLOCKS)
2140 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2141 if (oa->o_valid & OBD_MD_FLMTIME)
2142 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2143 if (oa->o_valid & OBD_MD_FLATIME)
2144 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2145 if (oa->o_valid & OBD_MD_FLCTIME)
2146 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2149 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2150 oap->oap_cmd, oa, rc);
2152 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2153 * I/O on the page could start, but OSC calls it under lock
2154 * and thus we can add oap back to pending safely */
2156 /* upper layer wants to leave the page on pending queue */
2157 osc_oap_to_pending(oap);
2159 osc_exit_cache(cli, oap, sent);
2163 static int brw_interpret(const struct lu_env *env,
2164 struct ptlrpc_request *req, void *data, int rc)
2166 struct osc_brw_async_args *aa = data;
2167 struct client_obd *cli;
2171 rc = osc_brw_fini_request(req, rc);
2172 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2173 if (osc_recoverable_error(rc)) {
2174 /* Only retry once for mmaped files since the mmaped page
2175 * might be modified at anytime. We have to retry at least
2176 * once in case there WAS really a corruption of the page
2177 * on the network, that was not caused by mmap() modifying
2178 * the page. Bug11742 */
2179 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2180 aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2181 aa->aa_oa->o_flags & OBD_FL_MMAP) {
2184 rc = osc_brw_redo_request(req, aa);
2191 capa_put(aa->aa_ocapa);
2192 aa->aa_ocapa = NULL;
2197 client_obd_list_lock(&cli->cl_loi_list_lock);
2199 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2200 * is called so we know whether to go to sync BRWs or wait for more
2201 * RPCs to complete */
2202 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2203 cli->cl_w_in_flight--;
2205 cli->cl_r_in_flight--;
2207 async = cfs_list_empty(&aa->aa_oaps);
2208 if (!async) { /* from osc_send_oap_rpc() */
2209 struct osc_async_page *oap, *tmp;
2210 /* the caller may re-use the oap after the completion call so
2211 * we need to clean it up a little */
2212 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2214 cfs_list_del_init(&oap->oap_rpc_item);
2215 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2217 OBDO_FREE(aa->aa_oa);
2218 } else { /* from async_internal() */
2220 for (i = 0; i < aa->aa_page_count; i++)
2221 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2223 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2224 OBDO_FREE(aa->aa_oa);
2226 osc_wake_cache_waiters(cli);
2227 osc_check_rpcs(env, cli);
2228 client_obd_list_unlock(&cli->cl_loi_list_lock);
2230 cl_req_completion(env, aa->aa_clerq, rc);
2231 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2235 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2236 struct client_obd *cli,
2237 cfs_list_t *rpc_list,
2238 int page_count, int cmd)
2240 struct ptlrpc_request *req;
2241 struct brw_page **pga = NULL;
2242 struct osc_brw_async_args *aa;
2243 struct obdo *oa = NULL;
2244 const struct obd_async_page_ops *ops = NULL;
2245 void *caller_data = NULL;
2246 struct osc_async_page *oap;
2247 struct osc_async_page *tmp;
2248 struct ost_body *body;
2249 struct cl_req *clerq = NULL;
2250 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2251 struct ldlm_lock *lock = NULL;
2252 struct cl_req_attr crattr;
2253 int i, rc, mpflag = 0;
2256 LASSERT(!cfs_list_empty(rpc_list));
2258 if (cmd & OBD_BRW_MEMALLOC)
2259 mpflag = cfs_memory_pressure_get_and_set();
2261 memset(&crattr, 0, sizeof crattr);
2262 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2264 GOTO(out, req = ERR_PTR(-ENOMEM));
2268 GOTO(out, req = ERR_PTR(-ENOMEM));
2271 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2272 struct cl_page *page = osc_oap2cl_page(oap);
2274 ops = oap->oap_caller_ops;
2275 caller_data = oap->oap_caller_data;
2277 clerq = cl_req_alloc(env, page, crt,
2278 1 /* only 1-object rpcs for
2281 GOTO(out, req = (void *)clerq);
2282 lock = oap->oap_ldlm_lock;
2284 pga[i] = &oap->oap_brw_page;
2285 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2286 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2287 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2289 cl_req_page_add(env, clerq, page);
2292 /* always get the data for the obdo for the rpc */
2293 LASSERT(ops != NULL);
2295 crattr.cra_capa = NULL;
2296 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2298 oa->o_handle = lock->l_remote_handle;
2299 oa->o_valid |= OBD_MD_FLHANDLE;
2302 rc = cl_req_prep(env, clerq);
2304 CERROR("cl_req_prep failed: %d\n", rc);
2305 GOTO(out, req = ERR_PTR(rc));
2308 sort_brw_pages(pga, page_count);
2309 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2310 pga, &req, crattr.cra_capa, 1);
2312 CERROR("prep_req failed: %d\n", rc);
2313 GOTO(out, req = ERR_PTR(rc));
2316 if (cmd & OBD_BRW_MEMALLOC)
2317 req->rq_memalloc = 1;
2319 /* Need to update the timestamps after the request is built in case
2320 * we race with setattr (locally or in queue at OST). If OST gets
2321 * later setattr before earlier BRW (as determined by the request xid),
2322 * the OST will not use BRW timestamps. Sadly, there is no obvious
2323 * way to do this in a single call. bug 10150 */
2324 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2325 cl_req_attr_set(env, clerq, &crattr,
2326 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2328 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2329 aa = ptlrpc_req_async_args(req);
2330 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2331 cfs_list_splice(rpc_list, &aa->aa_oaps);
2332 CFS_INIT_LIST_HEAD(rpc_list);
2333 aa->aa_clerq = clerq;
2335 if (cmd & OBD_BRW_MEMALLOC)
2336 cfs_memory_pressure_restore(mpflag);
2338 capa_put(crattr.cra_capa);
2343 OBD_FREE(pga, sizeof(*pga) * page_count);
2344 /* this should happen rarely and is pretty bad, it makes the
2345 * pending list not follow the dirty order */
2346 client_obd_list_lock(&cli->cl_loi_list_lock);
2347 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2348 cfs_list_del_init(&oap->oap_rpc_item);
2350 /* queued sync pages can be torn down while the pages
2351 * were between the pending list and the rpc */
2352 if (oap->oap_interrupted) {
2353 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2354 osc_ap_completion(env, cli, NULL, oap, 0,
2358 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2360 if (clerq && !IS_ERR(clerq))
2361 cl_req_completion(env, clerq, PTR_ERR(req));
2367 * prepare pages for ASYNC io and put pages in send queue.
2369 * \param cmd OBD_BRW_* macroses
2370 * \param lop pending pages
2372 * \return zero if no page added to send queue.
2373 * \return 1 if pages successfully added to send queue.
2374 * \return negative on errors.
2377 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2378 struct lov_oinfo *loi,
2379 int cmd, struct loi_oap_pages *lop)
2381 struct ptlrpc_request *req;
2382 obd_count page_count = 0;
2383 struct osc_async_page *oap = NULL, *tmp;
2384 struct osc_brw_async_args *aa;
2385 const struct obd_async_page_ops *ops;
2386 CFS_LIST_HEAD(rpc_list);
2387 CFS_LIST_HEAD(tmp_list);
2388 unsigned int ending_offset;
2389 unsigned starting_offset = 0;
2390 int srvlock = 0, mem_tight = 0;
2391 struct cl_object *clob = NULL;
2394 /* ASYNC_HP pages first. At present, when the lock the pages is
2395 * to be canceled, the pages covered by the lock will be sent out
2396 * with ASYNC_HP. We have to send out them as soon as possible. */
2397 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2398 if (oap->oap_async_flags & ASYNC_HP)
2399 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2401 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2402 if (++page_count >= cli->cl_max_pages_per_rpc)
2406 cfs_list_splice(&tmp_list, &lop->lop_pending);
2409 /* first we find the pages we're allowed to work with */
2410 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2412 ops = oap->oap_caller_ops;
2414 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2415 "magic 0x%x\n", oap, oap->oap_magic);
2418 /* pin object in memory, so that completion call-backs
2419 * can be safely called under client_obd_list lock. */
2420 clob = osc_oap2cl_page(oap)->cp_obj;
2421 cl_object_get(clob);
2424 if (page_count != 0 &&
2425 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2426 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2427 " oap %p, page %p, srvlock %u\n",
2428 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2432 /* If there is a gap at the start of this page, it can't merge
2433 * with any previous page, so we'll hand the network a
2434 * "fragmented" page array that it can't transfer in 1 RDMA */
2435 if (page_count != 0 && oap->oap_page_off != 0)
2438 /* in llite being 'ready' equates to the page being locked
2439 * until completion unlocks it. commit_write submits a page
2440 * as not ready because its unlock will happen unconditionally
2441 * as the call returns. if we race with commit_write giving
2442 * us that page we don't want to create a hole in the page
2443 * stream, so we stop and leave the rpc to be fired by
2444 * another dirtier or kupdated interval (the not ready page
2445 * will still be on the dirty list). we could call in
2446 * at the end of ll_file_write to process the queue again. */
2447 if (!(oap->oap_async_flags & ASYNC_READY)) {
2448 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2451 CDEBUG(D_INODE, "oap %p page %p returned %d "
2452 "instead of ready\n", oap,
2456 /* llite is telling us that the page is still
2457 * in commit_write and that we should try
2458 * and put it in an rpc again later. we
2459 * break out of the loop so we don't create
2460 * a hole in the sequence of pages in the rpc
2465 /* the io isn't needed.. tell the checks
2466 * below to complete the rpc with EINTR */
2467 cfs_spin_lock(&oap->oap_lock);
2468 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2469 cfs_spin_unlock(&oap->oap_lock);
2470 oap->oap_count = -EINTR;
2473 cfs_spin_lock(&oap->oap_lock);
2474 oap->oap_async_flags |= ASYNC_READY;
2475 cfs_spin_unlock(&oap->oap_lock);
2478 LASSERTF(0, "oap %p page %p returned %d "
2479 "from make_ready\n", oap,
2487 * Page submitted for IO has to be locked. Either by
2488 * ->ap_make_ready() or by higher layers.
2490 #if defined(__KERNEL__) && defined(__linux__)
2492 struct cl_page *page;
2494 page = osc_oap2cl_page(oap);
2496 if (page->cp_type == CPT_CACHEABLE &&
2497 !(PageLocked(oap->oap_page) &&
2498 (CheckWriteback(oap->oap_page, cmd)))) {
2499 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2501 (long)oap->oap_page->flags,
2502 oap->oap_async_flags);
2508 /* take the page out of our book-keeping */
2509 cfs_list_del_init(&oap->oap_pending_item);
2510 lop_update_pending(cli, lop, cmd, -1);
2511 cfs_list_del_init(&oap->oap_urgent_item);
2513 if (page_count == 0)
2514 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2515 (PTLRPC_MAX_BRW_SIZE - 1);
2517 /* ask the caller for the size of the io as the rpc leaves. */
2518 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2520 ops->ap_refresh_count(env, oap->oap_caller_data,
2522 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2524 if (oap->oap_count <= 0) {
2525 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2527 osc_ap_completion(env, cli, NULL,
2528 oap, 0, oap->oap_count);
2532 /* now put the page back in our accounting */
2533 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2534 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2536 if (page_count == 0)
2537 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2538 if (++page_count >= cli->cl_max_pages_per_rpc)
2541 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2542 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2543 * have the same alignment as the initial writes that allocated
2544 * extents on the server. */
2545 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2546 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2547 if (ending_offset == 0)
2550 /* If there is a gap at the end of this page, it can't merge
2551 * with any subsequent pages, so we'll hand the network a
2552 * "fragmented" page array that it can't transfer in 1 RDMA */
2553 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2557 osc_wake_cache_waiters(cli);
2559 loi_list_maint(cli, loi);
2561 client_obd_list_unlock(&cli->cl_loi_list_lock);
2564 cl_object_put(env, clob);
2566 if (page_count == 0) {
2567 client_obd_list_lock(&cli->cl_loi_list_lock);
2571 req = osc_build_req(env, cli, &rpc_list, page_count,
2572 mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2574 LASSERT(cfs_list_empty(&rpc_list));
2575 loi_list_maint(cli, loi);
2576 RETURN(PTR_ERR(req));
2579 aa = ptlrpc_req_async_args(req);
2581 if (cmd == OBD_BRW_READ) {
2582 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2583 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2584 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2585 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2587 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2588 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2589 cli->cl_w_in_flight);
2590 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2591 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2593 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2595 client_obd_list_lock(&cli->cl_loi_list_lock);
2597 if (cmd == OBD_BRW_READ)
2598 cli->cl_r_in_flight++;
2600 cli->cl_w_in_flight++;
2602 /* queued sync pages can be torn down while the pages
2603 * were between the pending list and the rpc */
2605 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2606 /* only one oap gets a request reference */
2609 if (oap->oap_interrupted && !req->rq_intr) {
2610 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2612 ptlrpc_mark_interrupted(req);
2616 tmp->oap_request = ptlrpc_request_addref(req);
2618 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2619 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2621 req->rq_interpret_reply = brw_interpret;
2622 ptlrpcd_add_req(req, PSCOPE_BRW);
2626 #define LOI_DEBUG(LOI, STR, args...) \
2627 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2628 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2629 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2630 (LOI)->loi_write_lop.lop_num_pending, \
2631 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2632 (LOI)->loi_read_lop.lop_num_pending, \
2633 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2636 /* This is called by osc_check_rpcs() to find which objects have pages that
2637 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2638 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2642 /* First return objects that have blocked locks so that they
2643 * will be flushed quickly and other clients can get the lock,
2644 * then objects which have pages ready to be stuffed into RPCs */
2645 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2646 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2647 struct lov_oinfo, loi_hp_ready_item));
2648 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2649 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2650 struct lov_oinfo, loi_ready_item));
2652 /* then if we have cache waiters, return all objects with queued
2653 * writes. This is especially important when many small files
2654 * have filled up the cache and not been fired into rpcs because
2655 * they don't pass the nr_pending/object threshhold */
2656 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2657 !cfs_list_empty(&cli->cl_loi_write_list))
2658 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2659 struct lov_oinfo, loi_write_item));
2661 /* then return all queued objects when we have an invalid import
2662 * so that they get flushed */
2663 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2664 if (!cfs_list_empty(&cli->cl_loi_write_list))
2665 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2668 if (!cfs_list_empty(&cli->cl_loi_read_list))
2669 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2670 struct lov_oinfo, loi_read_item));
2675 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2677 struct osc_async_page *oap;
2680 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2681 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2682 struct osc_async_page, oap_urgent_item);
2683 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2686 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2687 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2688 struct osc_async_page, oap_urgent_item);
2689 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2692 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2695 /* called with the loi list lock held */
2696 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2698 struct lov_oinfo *loi;
2699 int rc = 0, race_counter = 0;
2702 while ((loi = osc_next_loi(cli)) != NULL) {
2703 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2705 if (osc_max_rpc_in_flight(cli, loi))
2708 /* attempt some read/write balancing by alternating between
2709 * reads and writes in an object. The makes_rpc checks here
2710 * would be redundant if we were getting read/write work items
2711 * instead of objects. we don't want send_oap_rpc to drain a
2712 * partial read pending queue when we're given this object to
2713 * do io on writes while there are cache waiters */
2714 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2715 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2716 &loi->loi_write_lop);
2718 CERROR("Write request failed with %d\n", rc);
2720 /* osc_send_oap_rpc failed, mostly because of
2723 * It can't break here, because if:
2724 * - a page was submitted by osc_io_submit, so
2726 * - no request in flight
2727 * - no subsequent request
2728 * The system will be in live-lock state,
2729 * because there is no chance to call
2730 * osc_io_unplug() and osc_check_rpcs() any
2731 * more. pdflush can't help in this case,
2732 * because it might be blocked at grabbing
2733 * the page lock as we mentioned.
2735 * Anyway, continue to drain pages. */
2744 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2745 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2746 &loi->loi_read_lop);
2748 CERROR("Read request failed with %d\n", rc);
2756 /* attempt some inter-object balancing by issuing rpcs
2757 * for each object in turn */
2758 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2759 cfs_list_del_init(&loi->loi_hp_ready_item);
2760 if (!cfs_list_empty(&loi->loi_ready_item))
2761 cfs_list_del_init(&loi->loi_ready_item);
2762 if (!cfs_list_empty(&loi->loi_write_item))
2763 cfs_list_del_init(&loi->loi_write_item);
2764 if (!cfs_list_empty(&loi->loi_read_item))
2765 cfs_list_del_init(&loi->loi_read_item);
2767 loi_list_maint(cli, loi);
2769 /* send_oap_rpc fails with 0 when make_ready tells it to
2770 * back off. llite's make_ready does this when it tries
2771 * to lock a page queued for write that is already locked.
2772 * we want to try sending rpcs from many objects, but we
2773 * don't want to spin failing with 0. */
2774 if (race_counter == 10)
2780 /* we're trying to queue a page in the osc so we're subject to the
2781 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2782 * If the osc's queued pages are already at that limit, then we want to sleep
2783 * until there is space in the osc's queue for us. We also may be waiting for
2784 * write credits from the OST if there are RPCs in flight that may return some
2785 * before we fall back to sync writes.
2787 * We need this know our allocation was granted in the presence of signals */
2788 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2792 client_obd_list_lock(&cli->cl_loi_list_lock);
2793 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2794 client_obd_list_unlock(&cli->cl_loi_list_lock);
2799 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2802 int osc_enter_cache_try(const struct lu_env *env,
2803 struct client_obd *cli, struct lov_oinfo *loi,
2804 struct osc_async_page *oap, int transient)
2808 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2810 osc_consume_write_grant(cli, &oap->oap_brw_page);
2812 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2813 cfs_atomic_inc(&obd_dirty_transit_pages);
2814 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2820 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2821 * grant or cache space. */
2822 static int osc_enter_cache(const struct lu_env *env,
2823 struct client_obd *cli, struct lov_oinfo *loi,
2824 struct osc_async_page *oap)
2826 struct osc_cache_waiter ocw;
2827 struct l_wait_info lwi = { 0 };
2831 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2832 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2833 cli->cl_dirty_max, obd_max_dirty_pages,
2834 cli->cl_lost_grant, cli->cl_avail_grant);
2836 /* force the caller to try sync io. this can jump the list
2837 * of queued writes and create a discontiguous rpc stream */
2838 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2839 loi->loi_ar.ar_force_sync)
2842 /* Hopefully normal case - cache space and write credits available */
2843 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2844 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2845 osc_enter_cache_try(env, cli, loi, oap, 0))
2848 /* It is safe to block as a cache waiter as long as there is grant
2849 * space available or the hope of additional grant being returned
2850 * when an in flight write completes. Using the write back cache
2851 * if possible is preferable to sending the data synchronously
2852 * because write pages can then be merged in to large requests.
2853 * The addition of this cache waiter will causing pending write
2854 * pages to be sent immediately. */
2855 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2856 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2857 cfs_waitq_init(&ocw.ocw_waitq);
2861 loi_list_maint(cli, loi);
2862 osc_check_rpcs(env, cli);
2863 client_obd_list_unlock(&cli->cl_loi_list_lock);
2865 CDEBUG(D_CACHE, "sleeping for cache space\n");
2866 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2868 client_obd_list_lock(&cli->cl_loi_list_lock);
2869 if (!cfs_list_empty(&ocw.ocw_entry)) {
2870 cfs_list_del(&ocw.ocw_entry);
2880 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2881 struct lov_oinfo *loi, cfs_page_t *page,
2882 obd_off offset, const struct obd_async_page_ops *ops,
2883 void *data, void **res, int nocache,
2884 struct lustre_handle *lockh)
2886 struct osc_async_page *oap;
2891 return cfs_size_round(sizeof(*oap));
2894 oap->oap_magic = OAP_MAGIC;
2895 oap->oap_cli = &exp->exp_obd->u.cli;
2898 oap->oap_caller_ops = ops;
2899 oap->oap_caller_data = data;
2901 oap->oap_page = page;
2902 oap->oap_obj_off = offset;
2903 if (!client_is_remote(exp) &&
2904 cfs_capable(CFS_CAP_SYS_RESOURCE))
2905 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2907 LASSERT(!(offset & ~CFS_PAGE_MASK));
2909 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2910 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2911 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2912 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2914 cfs_spin_lock_init(&oap->oap_lock);
2915 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2919 struct osc_async_page *oap_from_cookie(void *cookie)
2921 struct osc_async_page *oap = cookie;
2922 if (oap->oap_magic != OAP_MAGIC)
2923 return ERR_PTR(-EINVAL);
2927 int osc_queue_async_io(const struct lu_env *env,
2928 struct obd_export *exp, struct lov_stripe_md *lsm,
2929 struct lov_oinfo *loi, void *cookie,
2930 int cmd, obd_off off, int count,
2931 obd_flag brw_flags, enum async_flags async_flags)
2933 struct client_obd *cli = &exp->exp_obd->u.cli;
2934 struct osc_async_page *oap;
2938 oap = oap_from_cookie(cookie);
2940 RETURN(PTR_ERR(oap));
2942 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2945 if (!cfs_list_empty(&oap->oap_pending_item) ||
2946 !cfs_list_empty(&oap->oap_urgent_item) ||
2947 !cfs_list_empty(&oap->oap_rpc_item))
2950 /* check if the file's owner/group is over quota */
2951 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2952 struct cl_object *obj;
2953 struct cl_attr attr; /* XXX put attr into thread info */
2954 unsigned int qid[MAXQUOTAS];
2956 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2958 cl_object_attr_lock(obj);
2959 rc = cl_object_attr_get(env, obj, &attr);
2960 cl_object_attr_unlock(obj);
2962 qid[USRQUOTA] = attr.cat_uid;
2963 qid[GRPQUOTA] = attr.cat_gid;
2965 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2972 loi = lsm->lsm_oinfo[0];
2974 client_obd_list_lock(&cli->cl_loi_list_lock);
2976 LASSERT(off + count <= CFS_PAGE_SIZE);
2978 oap->oap_page_off = off;
2979 oap->oap_count = count;
2980 oap->oap_brw_flags = brw_flags;
2981 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2982 if (cfs_memory_pressure_get())
2983 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2984 cfs_spin_lock(&oap->oap_lock);
2985 oap->oap_async_flags = async_flags;
2986 cfs_spin_unlock(&oap->oap_lock);
2988 if (cmd & OBD_BRW_WRITE) {
2989 rc = osc_enter_cache(env, cli, loi, oap);
2991 client_obd_list_unlock(&cli->cl_loi_list_lock);
2996 osc_oap_to_pending(oap);
2997 loi_list_maint(cli, loi);
2999 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3002 osc_check_rpcs(env, cli);
3003 client_obd_list_unlock(&cli->cl_loi_list_lock);
3008 /* aka (~was & now & flag), but this is more clear :) */
3009 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3011 int osc_set_async_flags_base(struct client_obd *cli,
3012 struct lov_oinfo *loi, struct osc_async_page *oap,
3013 obd_flag async_flags)
3015 struct loi_oap_pages *lop;
3019 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3021 if (oap->oap_cmd & OBD_BRW_WRITE) {
3022 lop = &loi->loi_write_lop;
3024 lop = &loi->loi_read_lop;
3027 if ((oap->oap_async_flags & async_flags) == async_flags)
3030 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3031 flags |= ASYNC_READY;
3033 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3034 cfs_list_empty(&oap->oap_rpc_item)) {
3035 if (oap->oap_async_flags & ASYNC_HP)
3036 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3038 cfs_list_add_tail(&oap->oap_urgent_item,
3040 flags |= ASYNC_URGENT;
3041 loi_list_maint(cli, loi);
3043 cfs_spin_lock(&oap->oap_lock);
3044 oap->oap_async_flags |= flags;
3045 cfs_spin_unlock(&oap->oap_lock);
3047 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3048 oap->oap_async_flags);
3052 int osc_teardown_async_page(struct obd_export *exp,
3053 struct lov_stripe_md *lsm,
3054 struct lov_oinfo *loi, void *cookie)
3056 struct client_obd *cli = &exp->exp_obd->u.cli;
3057 struct loi_oap_pages *lop;
3058 struct osc_async_page *oap;
3062 oap = oap_from_cookie(cookie);
3064 RETURN(PTR_ERR(oap));
3067 loi = lsm->lsm_oinfo[0];
3069 if (oap->oap_cmd & OBD_BRW_WRITE) {
3070 lop = &loi->loi_write_lop;
3072 lop = &loi->loi_read_lop;
3075 client_obd_list_lock(&cli->cl_loi_list_lock);
3077 if (!cfs_list_empty(&oap->oap_rpc_item))
3078 GOTO(out, rc = -EBUSY);
3080 osc_exit_cache(cli, oap, 0);
3081 osc_wake_cache_waiters(cli);
3083 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3084 cfs_list_del_init(&oap->oap_urgent_item);
3085 cfs_spin_lock(&oap->oap_lock);
3086 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3087 cfs_spin_unlock(&oap->oap_lock);
3089 if (!cfs_list_empty(&oap->oap_pending_item)) {
3090 cfs_list_del_init(&oap->oap_pending_item);
3091 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3093 loi_list_maint(cli, loi);
3094 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3096 client_obd_list_unlock(&cli->cl_loi_list_lock);
3100 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3101 struct ldlm_enqueue_info *einfo,
3104 void *data = einfo->ei_cbdata;
3106 LASSERT(lock != NULL);
3107 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3108 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3109 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3110 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3112 lock_res_and_lock(lock);
3113 cfs_spin_lock(&osc_ast_guard);
3114 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3115 lock->l_ast_data = data;
3116 cfs_spin_unlock(&osc_ast_guard);
3117 unlock_res_and_lock(lock);
3120 static void osc_set_data_with_check(struct lustre_handle *lockh,
3121 struct ldlm_enqueue_info *einfo,
3124 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3127 osc_set_lock_data_with_check(lock, einfo, flags);
3128 LDLM_LOCK_PUT(lock);
3130 CERROR("lockh %p, data %p - client evicted?\n",
3131 lockh, einfo->ei_cbdata);
3134 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3135 ldlm_iterator_t replace, void *data)
3137 struct ldlm_res_id res_id;
3138 struct obd_device *obd = class_exp2obd(exp);
3140 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3141 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3145 /* find any ldlm lock of the inode in osc
3149 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3150 ldlm_iterator_t replace, void *data)
3152 struct ldlm_res_id res_id;
3153 struct obd_device *obd = class_exp2obd(exp);
3156 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3157 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3158 if (rc == LDLM_ITER_STOP)
3160 if (rc == LDLM_ITER_CONTINUE)
3165 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3166 obd_enqueue_update_f upcall, void *cookie,
3169 int intent = *flags & LDLM_FL_HAS_INTENT;
3173 /* The request was created before ldlm_cli_enqueue call. */
3174 if (rc == ELDLM_LOCK_ABORTED) {
3175 struct ldlm_reply *rep;
3176 rep = req_capsule_server_get(&req->rq_pill,
3179 LASSERT(rep != NULL);
3180 if (rep->lock_policy_res1)
3181 rc = rep->lock_policy_res1;
3185 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3186 *flags |= LDLM_FL_LVB_READY;
3187 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3188 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3191 /* Call the update callback. */
3192 rc = (*upcall)(cookie, rc);
3196 static int osc_enqueue_interpret(const struct lu_env *env,
3197 struct ptlrpc_request *req,
3198 struct osc_enqueue_args *aa, int rc)
3200 struct ldlm_lock *lock;
3201 struct lustre_handle handle;
3204 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3205 * might be freed anytime after lock upcall has been called. */
3206 lustre_handle_copy(&handle, aa->oa_lockh);
3207 mode = aa->oa_ei->ei_mode;
3209 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3211 lock = ldlm_handle2lock(&handle);
3213 /* Take an additional reference so that a blocking AST that
3214 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3215 * to arrive after an upcall has been executed by
3216 * osc_enqueue_fini(). */
3217 ldlm_lock_addref(&handle, mode);
3219 /* Complete obtaining the lock procedure. */
3220 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3221 mode, aa->oa_flags, aa->oa_lvb,
3222 sizeof(*aa->oa_lvb), &handle, rc);
3223 /* Complete osc stuff. */
3224 rc = osc_enqueue_fini(req, aa->oa_lvb,
3225 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3227 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3229 /* Release the lock for async request. */
3230 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3232 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3233 * not already released by
3234 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3236 ldlm_lock_decref(&handle, mode);
3238 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3239 aa->oa_lockh, req, aa);
3240 ldlm_lock_decref(&handle, mode);
3241 LDLM_LOCK_PUT(lock);
3245 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3246 struct lov_oinfo *loi, int flags,
3247 struct ost_lvb *lvb, __u32 mode, int rc)
3249 if (rc == ELDLM_OK) {
3250 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3253 LASSERT(lock != NULL);
3254 loi->loi_lvb = *lvb;
3255 tmp = loi->loi_lvb.lvb_size;
3256 /* Extend KMS up to the end of this lock and no further
3257 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3258 if (tmp > lock->l_policy_data.l_extent.end)
3259 tmp = lock->l_policy_data.l_extent.end + 1;
3260 if (tmp >= loi->loi_kms) {
3261 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3262 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3263 loi_kms_set(loi, tmp);
3265 LDLM_DEBUG(lock, "lock acquired, setting rss="
3266 LPU64"; leaving kms="LPU64", end="LPU64,
3267 loi->loi_lvb.lvb_size, loi->loi_kms,
3268 lock->l_policy_data.l_extent.end);
3270 ldlm_lock_allow_match(lock);
3271 LDLM_LOCK_PUT(lock);
3272 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3273 loi->loi_lvb = *lvb;
3274 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3275 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3279 EXPORT_SYMBOL(osc_update_enqueue);
3281 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3283 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3284 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3285 * other synchronous requests, however keeping some locks and trying to obtain
3286 * others may take a considerable amount of time in a case of ost failure; and
3287 * when other sync requests do not get released lock from a client, the client
3288 * is excluded from the cluster -- such scenarious make the life difficult, so
3289 * release locks just after they are obtained. */
3290 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3291 int *flags, ldlm_policy_data_t *policy,
3292 struct ost_lvb *lvb, int kms_valid,
3293 obd_enqueue_update_f upcall, void *cookie,
3294 struct ldlm_enqueue_info *einfo,
3295 struct lustre_handle *lockh,
3296 struct ptlrpc_request_set *rqset, int async)
3298 struct obd_device *obd = exp->exp_obd;
3299 struct ptlrpc_request *req = NULL;
3300 int intent = *flags & LDLM_FL_HAS_INTENT;
3305 /* Filesystem lock extents are extended to page boundaries so that
3306 * dealing with the page cache is a little smoother. */
3307 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3308 policy->l_extent.end |= ~CFS_PAGE_MASK;
3311 * kms is not valid when either object is completely fresh (so that no
3312 * locks are cached), or object was evicted. In the latter case cached
3313 * lock cannot be used, because it would prime inode state with
3314 * potentially stale LVB.
3319 /* Next, search for already existing extent locks that will cover us */
3320 /* If we're trying to read, we also search for an existing PW lock. The
3321 * VFS and page cache already protect us locally, so lots of readers/
3322 * writers can share a single PW lock.
3324 * There are problems with conversion deadlocks, so instead of
3325 * converting a read lock to a write lock, we'll just enqueue a new
3328 * At some point we should cancel the read lock instead of making them
3329 * send us a blocking callback, but there are problems with canceling
3330 * locks out from other users right now, too. */
3331 mode = einfo->ei_mode;
3332 if (einfo->ei_mode == LCK_PR)
3334 mode = ldlm_lock_match(obd->obd_namespace,
3335 *flags | LDLM_FL_LVB_READY, res_id,
3336 einfo->ei_type, policy, mode, lockh, 0);
3338 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3340 if (matched->l_ast_data == NULL ||
3341 matched->l_ast_data == einfo->ei_cbdata) {
3342 /* addref the lock only if not async requests and PW
3343 * lock is matched whereas we asked for PR. */
3344 if (!rqset && einfo->ei_mode != mode)
3345 ldlm_lock_addref(lockh, LCK_PR);
3346 osc_set_lock_data_with_check(matched, einfo, *flags);
3348 /* I would like to be able to ASSERT here that
3349 * rss <= kms, but I can't, for reasons which
3350 * are explained in lov_enqueue() */
3353 /* We already have a lock, and it's referenced */
3354 (*upcall)(cookie, ELDLM_OK);
3356 /* For async requests, decref the lock. */
3357 if (einfo->ei_mode != mode)
3358 ldlm_lock_decref(lockh, LCK_PW);
3360 ldlm_lock_decref(lockh, einfo->ei_mode);
3361 LDLM_LOCK_PUT(matched);
3364 ldlm_lock_decref(lockh, mode);
3365 LDLM_LOCK_PUT(matched);
3370 CFS_LIST_HEAD(cancels);
3371 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3372 &RQF_LDLM_ENQUEUE_LVB);
3376 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3380 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3382 ptlrpc_request_set_replen(req);
3385 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3386 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3388 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3389 sizeof(*lvb), lockh, async);
3392 struct osc_enqueue_args *aa;
3393 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3394 aa = ptlrpc_req_async_args(req);
3397 aa->oa_flags = flags;
3398 aa->oa_upcall = upcall;
3399 aa->oa_cookie = cookie;
3401 aa->oa_lockh = lockh;
3403 req->rq_interpret_reply =
3404 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3405 if (rqset == PTLRPCD_SET)
3406 ptlrpcd_add_req(req, PSCOPE_OTHER);
3408 ptlrpc_set_add_req(rqset, req);
3409 } else if (intent) {
3410 ptlrpc_req_finished(req);
3415 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3417 ptlrpc_req_finished(req);
3422 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3423 struct ldlm_enqueue_info *einfo,
3424 struct ptlrpc_request_set *rqset)
3426 struct ldlm_res_id res_id;
3430 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3431 oinfo->oi_md->lsm_object_seq, &res_id);
3433 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3434 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3435 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3436 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3437 rqset, rqset != NULL);
3441 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3442 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3443 int *flags, void *data, struct lustre_handle *lockh,
3446 struct obd_device *obd = exp->exp_obd;
3447 int lflags = *flags;
3451 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3454 /* Filesystem lock extents are extended to page boundaries so that
3455 * dealing with the page cache is a little smoother */
3456 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3457 policy->l_extent.end |= ~CFS_PAGE_MASK;
3459 /* Next, search for already existing extent locks that will cover us */
3460 /* If we're trying to read, we also search for an existing PW lock. The
3461 * VFS and page cache already protect us locally, so lots of readers/
3462 * writers can share a single PW lock. */
3466 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3467 res_id, type, policy, rc, lockh, unref);
3470 osc_set_data_with_check(lockh, data, lflags);
3471 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3472 ldlm_lock_addref(lockh, LCK_PR);
3473 ldlm_lock_decref(lockh, LCK_PW);
3480 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3484 if (unlikely(mode == LCK_GROUP))
3485 ldlm_lock_decref_and_cancel(lockh, mode);
3487 ldlm_lock_decref(lockh, mode);
3492 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3493 __u32 mode, struct lustre_handle *lockh)
3496 RETURN(osc_cancel_base(lockh, mode));
3499 static int osc_cancel_unused(struct obd_export *exp,
3500 struct lov_stripe_md *lsm,
3501 ldlm_cancel_flags_t flags,
3504 struct obd_device *obd = class_exp2obd(exp);
3505 struct ldlm_res_id res_id, *resp = NULL;
3508 resp = osc_build_res_name(lsm->lsm_object_id,
3509 lsm->lsm_object_seq, &res_id);
3512 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3515 static int osc_statfs_interpret(const struct lu_env *env,
3516 struct ptlrpc_request *req,
3517 struct osc_async_args *aa, int rc)
3519 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3520 struct obd_statfs *msfs;
3525 /* The request has in fact never been sent
3526 * due to issues at a higher level (LOV).
3527 * Exit immediately since the caller is
3528 * aware of the problem and takes care
3529 * of the clean up */
3532 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3533 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3539 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3541 GOTO(out, rc = -EPROTO);
3544 /* Reinitialize the RDONLY and DEGRADED flags at the client
3545 * on each statfs, so they don't stay set permanently. */
3546 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3548 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3549 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3550 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3551 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3553 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3554 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3555 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3556 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3558 /* Add a bit of hysteresis so this flag isn't continually flapping,
3559 * and ensure that new files don't get extremely fragmented due to
3560 * only a small amount of available space in the filesystem.
3561 * We want to set the NOSPC flag when there is less than ~0.1% free
3562 * and clear it when there is at least ~0.2% free space, so:
3563 * avail < ~0.1% max max = avail + used
3564 * 1025 * avail < avail + used used = blocks - free
3565 * 1024 * avail < used
3566 * 1024 * avail < blocks - free
3567 * avail < ((blocks - free) >> 10)
3569 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3570 * lose that amount of space so in those cases we report no space left
3571 * if their is less than 1 GB left. */
3572 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3573 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3574 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3575 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3576 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3577 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3578 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3580 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3582 *aa->aa_oi->oi_osfs = *msfs;
3584 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3588 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3589 __u64 max_age, struct ptlrpc_request_set *rqset)
3591 struct ptlrpc_request *req;
3592 struct osc_async_args *aa;
3596 /* We could possibly pass max_age in the request (as an absolute
3597 * timestamp or a "seconds.usec ago") so the target can avoid doing
3598 * extra calls into the filesystem if that isn't necessary (e.g.
3599 * during mount that would help a bit). Having relative timestamps
3600 * is not so great if request processing is slow, while absolute
3601 * timestamps are not ideal because they need time synchronization. */
3602 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3606 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3608 ptlrpc_request_free(req);
3611 ptlrpc_request_set_replen(req);
3612 req->rq_request_portal = OST_CREATE_PORTAL;
3613 ptlrpc_at_set_req_timeout(req);
3615 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3616 /* procfs requests not want stat in wait for avoid deadlock */
3617 req->rq_no_resend = 1;
3618 req->rq_no_delay = 1;
3621 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3622 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3623 aa = ptlrpc_req_async_args(req);
3626 ptlrpc_set_add_req(rqset, req);
3630 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3631 __u64 max_age, __u32 flags)
3633 struct obd_statfs *msfs;
3634 struct ptlrpc_request *req;
3635 struct obd_import *imp = NULL;
3639 /*Since the request might also come from lprocfs, so we need
3640 *sync this with client_disconnect_export Bug15684*/
3641 cfs_down_read(&obd->u.cli.cl_sem);
3642 if (obd->u.cli.cl_import)
3643 imp = class_import_get(obd->u.cli.cl_import);
3644 cfs_up_read(&obd->u.cli.cl_sem);
3648 /* We could possibly pass max_age in the request (as an absolute
3649 * timestamp or a "seconds.usec ago") so the target can avoid doing
3650 * extra calls into the filesystem if that isn't necessary (e.g.
3651 * during mount that would help a bit). Having relative timestamps
3652 * is not so great if request processing is slow, while absolute
3653 * timestamps are not ideal because they need time synchronization. */
3654 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3656 class_import_put(imp);
3661 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3663 ptlrpc_request_free(req);
3666 ptlrpc_request_set_replen(req);
3667 req->rq_request_portal = OST_CREATE_PORTAL;
3668 ptlrpc_at_set_req_timeout(req);
3670 if (flags & OBD_STATFS_NODELAY) {
3671 /* procfs requests not want stat in wait for avoid deadlock */
3672 req->rq_no_resend = 1;
3673 req->rq_no_delay = 1;
3676 rc = ptlrpc_queue_wait(req);
3680 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3682 GOTO(out, rc = -EPROTO);
3689 ptlrpc_req_finished(req);
3693 /* Retrieve object striping information.
3695 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3696 * the maximum number of OST indices which will fit in the user buffer.
3697 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3699 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3701 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3702 struct lov_user_md_v3 lum, *lumk;
3703 struct lov_user_ost_data_v1 *lmm_objects;
3704 int rc = 0, lum_size;
3710 /* we only need the header part from user space to get lmm_magic and
3711 * lmm_stripe_count, (the header part is common to v1 and v3) */
3712 lum_size = sizeof(struct lov_user_md_v1);
3713 if (cfs_copy_from_user(&lum, lump, lum_size))
3716 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3717 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3720 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3721 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3722 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3723 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3725 /* we can use lov_mds_md_size() to compute lum_size
3726 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3727 if (lum.lmm_stripe_count > 0) {
3728 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3729 OBD_ALLOC(lumk, lum_size);
3733 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3734 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3736 lmm_objects = &(lumk->lmm_objects[0]);
3737 lmm_objects->l_object_id = lsm->lsm_object_id;
3739 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3743 lumk->lmm_object_id = lsm->lsm_object_id;
3744 lumk->lmm_object_seq = lsm->lsm_object_seq;
3745 lumk->lmm_stripe_count = 1;
3747 if (cfs_copy_to_user(lump, lumk, lum_size))
3751 OBD_FREE(lumk, lum_size);
3757 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3758 void *karg, void *uarg)
3760 struct obd_device *obd = exp->exp_obd;
3761 struct obd_ioctl_data *data = karg;
3765 if (!cfs_try_module_get(THIS_MODULE)) {
3766 CERROR("Can't get module. Is it alive?");
3770 case OBD_IOC_LOV_GET_CONFIG: {
3772 struct lov_desc *desc;
3773 struct obd_uuid uuid;
3777 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3778 GOTO(out, err = -EINVAL);
3780 data = (struct obd_ioctl_data *)buf;
3782 if (sizeof(*desc) > data->ioc_inllen1) {
3783 obd_ioctl_freedata(buf, len);
3784 GOTO(out, err = -EINVAL);
3787 if (data->ioc_inllen2 < sizeof(uuid)) {
3788 obd_ioctl_freedata(buf, len);
3789 GOTO(out, err = -EINVAL);
3792 desc = (struct lov_desc *)data->ioc_inlbuf1;
3793 desc->ld_tgt_count = 1;
3794 desc->ld_active_tgt_count = 1;
3795 desc->ld_default_stripe_count = 1;
3796 desc->ld_default_stripe_size = 0;
3797 desc->ld_default_stripe_offset = 0;
3798 desc->ld_pattern = 0;
3799 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3801 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3803 err = cfs_copy_to_user((void *)uarg, buf, len);
3806 obd_ioctl_freedata(buf, len);
3809 case LL_IOC_LOV_SETSTRIPE:
3810 err = obd_alloc_memmd(exp, karg);
3814 case LL_IOC_LOV_GETSTRIPE:
3815 err = osc_getstripe(karg, uarg);
3817 case OBD_IOC_CLIENT_RECOVER:
3818 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3823 case IOC_OSC_SET_ACTIVE:
3824 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3827 case OBD_IOC_POLL_QUOTACHECK:
3828 err = lquota_poll_check(quota_interface, exp,
3829 (struct if_quotacheck *)karg);
3831 case OBD_IOC_PING_TARGET:
3832 err = ptlrpc_obd_ping(obd);
3835 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3836 cmd, cfs_curproc_comm());
3837 GOTO(out, err = -ENOTTY);
3840 cfs_module_put(THIS_MODULE);
3844 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3845 void *key, __u32 *vallen, void *val,
3846 struct lov_stripe_md *lsm)
3849 if (!vallen || !val)
3852 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3853 __u32 *stripe = val;
3854 *vallen = sizeof(*stripe);
3857 } else if (KEY_IS(KEY_LAST_ID)) {
3858 struct ptlrpc_request *req;
3863 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3864 &RQF_OST_GET_INFO_LAST_ID);
3868 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3869 RCL_CLIENT, keylen);
3870 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3872 ptlrpc_request_free(req);
3876 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3877 memcpy(tmp, key, keylen);
3879 req->rq_no_delay = req->rq_no_resend = 1;
3880 ptlrpc_request_set_replen(req);
3881 rc = ptlrpc_queue_wait(req);
3885 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3887 GOTO(out, rc = -EPROTO);
3889 *((obd_id *)val) = *reply;
3891 ptlrpc_req_finished(req);
3893 } else if (KEY_IS(KEY_FIEMAP)) {
3894 struct ptlrpc_request *req;
3895 struct ll_user_fiemap *reply;
3899 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3900 &RQF_OST_GET_INFO_FIEMAP);
3904 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3905 RCL_CLIENT, keylen);
3906 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3907 RCL_CLIENT, *vallen);
3908 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3909 RCL_SERVER, *vallen);
3911 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3913 ptlrpc_request_free(req);
3917 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3918 memcpy(tmp, key, keylen);
3919 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3920 memcpy(tmp, val, *vallen);
3922 ptlrpc_request_set_replen(req);
3923 rc = ptlrpc_queue_wait(req);
3927 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3929 GOTO(out1, rc = -EPROTO);
3931 memcpy(val, reply, *vallen);
3933 ptlrpc_req_finished(req);
3941 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3943 struct llog_ctxt *ctxt;
3947 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3949 rc = llog_initiator_connect(ctxt);
3950 llog_ctxt_put(ctxt);
3952 /* XXX return an error? skip setting below flags? */
3955 cfs_spin_lock(&imp->imp_lock);
3956 imp->imp_server_timeout = 1;
3957 imp->imp_pingable = 1;
3958 cfs_spin_unlock(&imp->imp_lock);
3959 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3964 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3965 struct ptlrpc_request *req,
3972 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3975 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3976 void *key, obd_count vallen, void *val,
3977 struct ptlrpc_request_set *set)
3979 struct ptlrpc_request *req;
3980 struct obd_device *obd = exp->exp_obd;
3981 struct obd_import *imp = class_exp2cliimp(exp);
3986 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3988 if (KEY_IS(KEY_NEXT_ID)) {
3990 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3992 if (vallen != sizeof(obd_id))
3997 if (vallen != sizeof(obd_id))
4000 /* avoid race between allocate new object and set next id
4001 * from ll_sync thread */
4002 cfs_spin_lock(&oscc->oscc_lock);
4003 new_val = *((obd_id*)val) + 1;
4004 if (new_val > oscc->oscc_next_id)
4005 oscc->oscc_next_id = new_val;
4006 cfs_spin_unlock(&oscc->oscc_lock);
4007 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
4008 exp->exp_obd->obd_name,
4009 obd->u.cli.cl_oscc.oscc_next_id);
4014 if (KEY_IS(KEY_CHECKSUM)) {
4015 if (vallen != sizeof(int))
4017 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
4021 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4022 sptlrpc_conf_client_adapt(obd);
4026 if (KEY_IS(KEY_FLUSH_CTX)) {
4027 sptlrpc_import_flush_my_ctx(imp);
4031 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4034 /* We pass all other commands directly to OST. Since nobody calls osc
4035 methods directly and everybody is supposed to go through LOV, we
4036 assume lov checked invalid values for us.
4037 The only recognised values so far are evict_by_nid and mds_conn.
4038 Even if something bad goes through, we'd get a -EINVAL from OST
4041 if (KEY_IS(KEY_GRANT_SHRINK))
4042 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4044 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4049 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4050 RCL_CLIENT, keylen);
4051 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4052 RCL_CLIENT, vallen);
4053 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4055 ptlrpc_request_free(req);
4059 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4060 memcpy(tmp, key, keylen);
4061 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4062 memcpy(tmp, val, vallen);
4064 if (KEY_IS(KEY_MDS_CONN)) {
4065 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4067 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4068 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4069 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4070 req->rq_no_delay = req->rq_no_resend = 1;
4071 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4072 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4073 struct osc_grant_args *aa;
4076 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4077 aa = ptlrpc_req_async_args(req);
4080 ptlrpc_req_finished(req);
4083 *oa = ((struct ost_body *)val)->oa;
4085 req->rq_interpret_reply = osc_shrink_grant_interpret;
4088 ptlrpc_request_set_replen(req);
4089 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4090 LASSERT(set != NULL);
4091 ptlrpc_set_add_req(set, req);
4092 ptlrpc_check_set(NULL, set);
4094 ptlrpcd_add_req(req, PSCOPE_OTHER);
4100 static struct llog_operations osc_size_repl_logops = {
4101 lop_cancel: llog_obd_repl_cancel
4104 static struct llog_operations osc_mds_ost_orig_logops;
4106 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4107 struct obd_device *tgt, struct llog_catid *catid)
4112 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4113 &catid->lci_logid, &osc_mds_ost_orig_logops);
4115 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4119 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4120 NULL, &osc_size_repl_logops);
4122 struct llog_ctxt *ctxt =
4123 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4126 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4131 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4132 obd->obd_name, tgt->obd_name, catid, rc);
4133 CERROR("logid "LPX64":0x%x\n",
4134 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4139 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4140 struct obd_device *disk_obd, int *index)
4142 struct llog_catid catid;
4143 static char name[32] = CATLIST;
4147 LASSERT(olg == &obd->obd_olg);
4149 cfs_mutex_down(&olg->olg_cat_processing);
4150 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4152 CERROR("rc: %d\n", rc);
4156 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4157 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4158 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4160 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4162 CERROR("rc: %d\n", rc);
4166 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4168 CERROR("rc: %d\n", rc);
4173 cfs_mutex_up(&olg->olg_cat_processing);
4178 static int osc_llog_finish(struct obd_device *obd, int count)
4180 struct llog_ctxt *ctxt;
4181 int rc = 0, rc2 = 0;
4184 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4186 rc = llog_cleanup(ctxt);
4188 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4190 rc2 = llog_cleanup(ctxt);
4197 static int osc_reconnect(const struct lu_env *env,
4198 struct obd_export *exp, struct obd_device *obd,
4199 struct obd_uuid *cluuid,
4200 struct obd_connect_data *data,
4203 struct client_obd *cli = &obd->u.cli;
4205 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4208 client_obd_list_lock(&cli->cl_loi_list_lock);
4209 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4210 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4211 lost_grant = cli->cl_lost_grant;
4212 cli->cl_lost_grant = 0;
4213 client_obd_list_unlock(&cli->cl_loi_list_lock);
4215 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4216 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4217 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4218 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4219 " ocd_grant: %d\n", data->ocd_connect_flags,
4220 data->ocd_version, data->ocd_grant);
4226 static int osc_disconnect(struct obd_export *exp)
4228 struct obd_device *obd = class_exp2obd(exp);
4229 struct llog_ctxt *ctxt;
4232 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4234 if (obd->u.cli.cl_conn_count == 1) {
4235 /* Flush any remaining cancel messages out to the
4237 llog_sync(ctxt, exp);
4239 llog_ctxt_put(ctxt);
4241 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4245 rc = client_disconnect_export(exp);
4247 * Initially we put del_shrink_grant before disconnect_export, but it
4248 * causes the following problem if setup (connect) and cleanup
4249 * (disconnect) are tangled together.
4250 * connect p1 disconnect p2
4251 * ptlrpc_connect_import
4252 * ............... class_manual_cleanup
4255 * ptlrpc_connect_interrupt
4257 * add this client to shrink list
4259 * Bang! pinger trigger the shrink.
4260 * So the osc should be disconnected from the shrink list, after we
4261 * are sure the import has been destroyed. BUG18662
4263 if (obd->u.cli.cl_import == NULL)
4264 osc_del_shrink_grant(&obd->u.cli);
4268 static int osc_import_event(struct obd_device *obd,
4269 struct obd_import *imp,
4270 enum obd_import_event event)
4272 struct client_obd *cli;
4276 LASSERT(imp->imp_obd == obd);
4279 case IMP_EVENT_DISCON: {
4280 /* Only do this on the MDS OSC's */
4281 if (imp->imp_server_timeout) {
4282 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4284 cfs_spin_lock(&oscc->oscc_lock);
4285 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4286 cfs_spin_unlock(&oscc->oscc_lock);
4289 client_obd_list_lock(&cli->cl_loi_list_lock);
4290 cli->cl_avail_grant = 0;
4291 cli->cl_lost_grant = 0;
4292 client_obd_list_unlock(&cli->cl_loi_list_lock);
4295 case IMP_EVENT_INACTIVE: {
4296 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4299 case IMP_EVENT_INVALIDATE: {
4300 struct ldlm_namespace *ns = obd->obd_namespace;
4304 env = cl_env_get(&refcheck);
4308 client_obd_list_lock(&cli->cl_loi_list_lock);
4309 /* all pages go to failing rpcs due to the invalid
4311 osc_check_rpcs(env, cli);
4312 client_obd_list_unlock(&cli->cl_loi_list_lock);
4314 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4315 cl_env_put(env, &refcheck);
4320 case IMP_EVENT_ACTIVE: {
4321 /* Only do this on the MDS OSC's */
4322 if (imp->imp_server_timeout) {
4323 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4325 cfs_spin_lock(&oscc->oscc_lock);
4326 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4327 cfs_spin_unlock(&oscc->oscc_lock);
4329 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4332 case IMP_EVENT_OCD: {
4333 struct obd_connect_data *ocd = &imp->imp_connect_data;
4335 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4336 osc_init_grant(&obd->u.cli, ocd);
4339 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4340 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4342 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4346 CERROR("Unknown import event %d\n", event);
4352 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4358 rc = ptlrpcd_addref();
4362 rc = client_obd_setup(obd, lcfg);
4366 struct lprocfs_static_vars lvars = { 0 };
4367 struct client_obd *cli = &obd->u.cli;
4369 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4370 lprocfs_osc_init_vars(&lvars);
4371 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4372 lproc_osc_attach_seqstat(obd);
4373 sptlrpc_lprocfs_cliobd_attach(obd);
4374 ptlrpc_lprocfs_register_obd(obd);
4378 /* We need to allocate a few requests more, because
4379 brw_interpret tries to create new requests before freeing
4380 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4381 reserved, but I afraid that might be too much wasted RAM
4382 in fact, so 2 is just my guess and still should work. */
4383 cli->cl_import->imp_rq_pool =
4384 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4386 ptlrpc_add_rqs_to_pool);
4388 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4389 cfs_sema_init(&cli->cl_grant_sem, 1);
4395 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4401 case OBD_CLEANUP_EARLY: {
4402 struct obd_import *imp;
4403 imp = obd->u.cli.cl_import;
4404 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4405 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4406 ptlrpc_deactivate_import(imp);
4407 cfs_spin_lock(&imp->imp_lock);
4408 imp->imp_pingable = 0;
4409 cfs_spin_unlock(&imp->imp_lock);
4412 case OBD_CLEANUP_EXPORTS: {
4413 /* If we set up but never connected, the
4414 client import will not have been cleaned. */
4415 if (obd->u.cli.cl_import) {
4416 struct obd_import *imp;
4417 cfs_down_write(&obd->u.cli.cl_sem);
4418 imp = obd->u.cli.cl_import;
4419 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4421 ptlrpc_invalidate_import(imp);
4422 if (imp->imp_rq_pool) {
4423 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4424 imp->imp_rq_pool = NULL;
4426 class_destroy_import(imp);
4427 cfs_up_write(&obd->u.cli.cl_sem);
4428 obd->u.cli.cl_import = NULL;
4430 rc = obd_llog_finish(obd, 0);
4432 CERROR("failed to cleanup llogging subsystems\n");
4439 int osc_cleanup(struct obd_device *obd)
4444 ptlrpc_lprocfs_unregister_obd(obd);
4445 lprocfs_obd_cleanup(obd);
4447 /* free memory of osc quota cache */
4448 lquota_cleanup(quota_interface, obd);
4450 rc = client_obd_cleanup(obd);
4456 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4458 struct lprocfs_static_vars lvars = { 0 };
4461 lprocfs_osc_init_vars(&lvars);
4463 switch (lcfg->lcfg_command) {
4465 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4475 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4477 return osc_process_config_base(obd, buf);
4480 struct obd_ops osc_obd_ops = {
4481 .o_owner = THIS_MODULE,
4482 .o_setup = osc_setup,
4483 .o_precleanup = osc_precleanup,
4484 .o_cleanup = osc_cleanup,
4485 .o_add_conn = client_import_add_conn,
4486 .o_del_conn = client_import_del_conn,
4487 .o_connect = client_connect_import,
4488 .o_reconnect = osc_reconnect,
4489 .o_disconnect = osc_disconnect,
4490 .o_statfs = osc_statfs,
4491 .o_statfs_async = osc_statfs_async,
4492 .o_packmd = osc_packmd,
4493 .o_unpackmd = osc_unpackmd,
4494 .o_precreate = osc_precreate,
4495 .o_create = osc_create,
4496 .o_create_async = osc_create_async,
4497 .o_destroy = osc_destroy,
4498 .o_getattr = osc_getattr,
4499 .o_getattr_async = osc_getattr_async,
4500 .o_setattr = osc_setattr,
4501 .o_setattr_async = osc_setattr_async,
4503 .o_punch = osc_punch,
4505 .o_enqueue = osc_enqueue,
4506 .o_change_cbdata = osc_change_cbdata,
4507 .o_find_cbdata = osc_find_cbdata,
4508 .o_cancel = osc_cancel,
4509 .o_cancel_unused = osc_cancel_unused,
4510 .o_iocontrol = osc_iocontrol,
4511 .o_get_info = osc_get_info,
4512 .o_set_info_async = osc_set_info_async,
4513 .o_import_event = osc_import_event,
4514 .o_llog_init = osc_llog_init,
4515 .o_llog_finish = osc_llog_finish,
4516 .o_process_config = osc_process_config,
4519 extern struct lu_kmem_descr osc_caches[];
4520 extern cfs_spinlock_t osc_ast_guard;
4521 extern cfs_lock_class_key_t osc_ast_guard_class;
4523 int __init osc_init(void)
4525 struct lprocfs_static_vars lvars = { 0 };
4529 /* print an address of _any_ initialized kernel symbol from this
4530 * module, to allow debugging with gdb that doesn't support data
4531 * symbols from modules.*/
4532 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4534 rc = lu_kmem_init(osc_caches);
4536 lprocfs_osc_init_vars(&lvars);
4538 cfs_request_module("lquota");
4539 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4540 lquota_init(quota_interface);
4541 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4543 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4544 LUSTRE_OSC_NAME, &osc_device_type);
4546 if (quota_interface)
4547 PORTAL_SYMBOL_PUT(osc_quota_interface);
4548 lu_kmem_fini(osc_caches);
4552 cfs_spin_lock_init(&osc_ast_guard);
4553 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4555 osc_mds_ost_orig_logops = llog_lvfs_ops;
4556 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4557 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4558 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4559 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4565 static void /*__exit*/ osc_exit(void)
4567 lu_device_type_fini(&osc_device_type);
4569 lquota_exit(quota_interface);
4570 if (quota_interface)
4571 PORTAL_SYMBOL_PUT(osc_quota_interface);
4573 class_unregister_type(LUSTRE_OSC_NAME);
4574 lu_kmem_fini(osc_caches);
4577 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4578 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4579 MODULE_LICENSE("GPL");
4581 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);