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,
1236 struct ptlrpc_request *req;
1237 struct ptlrpc_bulk_desc *desc;
1238 struct ost_body *body;
1239 struct obd_ioobj *ioobj;
1240 struct niobuf_remote *niobuf;
1241 int niocount, i, requested_nob, opc, rc;
1242 struct osc_brw_async_args *aa;
1243 struct req_capsule *pill;
1244 struct brw_page *pg_prev;
1247 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1248 RETURN(-ENOMEM); /* Recoverable */
1249 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1250 RETURN(-EINVAL); /* Fatal */
1252 if ((cmd & OBD_BRW_WRITE) != 0) {
1254 req = ptlrpc_request_alloc_pool(cli->cl_import,
1255 cli->cl_import->imp_rq_pool,
1256 &RQF_OST_BRW_WRITE);
1259 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1264 for (niocount = i = 1; i < page_count; i++) {
1265 if (!can_merge_pages(pga[i - 1], pga[i]))
1269 pill = &req->rq_pill;
1270 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1272 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1273 niocount * sizeof(*niobuf));
1274 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1278 ptlrpc_request_free(req);
1281 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1282 ptlrpc_at_set_req_timeout(req);
1284 if (opc == OST_WRITE)
1285 desc = ptlrpc_prep_bulk_imp(req, page_count,
1286 BULK_GET_SOURCE, OST_BULK_PORTAL);
1288 desc = ptlrpc_prep_bulk_imp(req, page_count,
1289 BULK_PUT_SINK, OST_BULK_PORTAL);
1292 GOTO(out, rc = -ENOMEM);
1293 /* NB request now owns desc and will free it when it gets freed */
1295 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1296 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1297 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1298 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1300 lustre_set_wire_obdo(&body->oa, oa);
1302 obdo_to_ioobj(oa, ioobj);
1303 ioobj->ioo_bufcnt = niocount;
1304 osc_pack_capa(req, body, ocapa);
1305 LASSERT (page_count > 0);
1307 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1308 struct brw_page *pg = pga[i];
1310 LASSERT(pg->count > 0);
1311 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1312 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1313 pg->off, pg->count);
1315 LASSERTF(i == 0 || pg->off > pg_prev->off,
1316 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1317 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1319 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1320 pg_prev->pg, page_private(pg_prev->pg),
1321 pg_prev->pg->index, pg_prev->off);
1323 LASSERTF(i == 0 || pg->off > pg_prev->off,
1324 "i %d p_c %u\n", i, page_count);
1326 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1327 (pg->flag & OBD_BRW_SRVLOCK));
1329 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1331 requested_nob += pg->count;
1333 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1335 niobuf->len += pg->count;
1337 niobuf->offset = pg->off;
1338 niobuf->len = pg->count;
1339 niobuf->flags = pg->flag;
1344 LASSERTF((void *)(niobuf - niocount) ==
1345 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1346 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1347 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1349 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1351 body->oa.o_valid |= OBD_FL_RECOV_RESEND;
1353 if (osc_should_shrink_grant(cli))
1354 osc_shrink_grant_local(cli, &body->oa);
1356 /* size[REQ_REC_OFF] still sizeof (*body) */
1357 if (opc == OST_WRITE) {
1358 if (unlikely(cli->cl_checksum) &&
1359 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1360 /* store cl_cksum_type in a local variable since
1361 * it can be changed via lprocfs */
1362 cksum_type_t cksum_type = cli->cl_cksum_type;
1364 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1365 oa->o_flags &= OBD_FL_LOCAL_MASK;
1366 body->oa.o_flags = 0;
1368 body->oa.o_flags |= cksum_type_pack(cksum_type);
1369 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1370 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1374 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1376 /* save this in 'oa', too, for later checking */
1377 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1378 oa->o_flags |= cksum_type_pack(cksum_type);
1380 /* clear out the checksum flag, in case this is a
1381 * resend but cl_checksum is no longer set. b=11238 */
1382 oa->o_valid &= ~OBD_MD_FLCKSUM;
1384 oa->o_cksum = body->oa.o_cksum;
1385 /* 1 RC per niobuf */
1386 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1387 sizeof(__u32) * niocount);
1389 if (unlikely(cli->cl_checksum) &&
1390 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1391 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1392 body->oa.o_flags = 0;
1393 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1394 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1397 ptlrpc_request_set_replen(req);
1399 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1400 aa = ptlrpc_req_async_args(req);
1402 aa->aa_requested_nob = requested_nob;
1403 aa->aa_nio_count = niocount;
1404 aa->aa_page_count = page_count;
1408 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1409 if (ocapa && reserve)
1410 aa->aa_ocapa = capa_get(ocapa);
1416 ptlrpc_req_finished(req);
1420 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1421 __u32 client_cksum, __u32 server_cksum, int nob,
1422 obd_count page_count, struct brw_page **pga,
1423 cksum_type_t client_cksum_type)
1427 cksum_type_t cksum_type;
1429 if (server_cksum == client_cksum) {
1430 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
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, resends);
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, 1);
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 coallesce 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 rc = osc_brw_redo_request(req, aa);
2180 capa_put(aa->aa_ocapa);
2181 aa->aa_ocapa = NULL;
2186 client_obd_list_lock(&cli->cl_loi_list_lock);
2188 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2189 * is called so we know whether to go to sync BRWs or wait for more
2190 * RPCs to complete */
2191 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2192 cli->cl_w_in_flight--;
2194 cli->cl_r_in_flight--;
2196 async = cfs_list_empty(&aa->aa_oaps);
2197 if (!async) { /* from osc_send_oap_rpc() */
2198 struct osc_async_page *oap, *tmp;
2199 /* the caller may re-use the oap after the completion call so
2200 * we need to clean it up a little */
2201 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2203 cfs_list_del_init(&oap->oap_rpc_item);
2204 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2206 OBDO_FREE(aa->aa_oa);
2207 } else { /* from async_internal() */
2209 for (i = 0; i < aa->aa_page_count; i++)
2210 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2212 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2213 OBDO_FREE(aa->aa_oa);
2215 osc_wake_cache_waiters(cli);
2216 osc_check_rpcs(env, cli);
2217 client_obd_list_unlock(&cli->cl_loi_list_lock);
2219 cl_req_completion(env, aa->aa_clerq, rc);
2220 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2224 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2225 struct client_obd *cli,
2226 cfs_list_t *rpc_list,
2227 int page_count, int cmd)
2229 struct ptlrpc_request *req;
2230 struct brw_page **pga = NULL;
2231 struct osc_brw_async_args *aa;
2232 struct obdo *oa = NULL;
2233 const struct obd_async_page_ops *ops = NULL;
2234 void *caller_data = NULL;
2235 struct osc_async_page *oap;
2236 struct osc_async_page *tmp;
2237 struct ost_body *body;
2238 struct cl_req *clerq = NULL;
2239 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2240 struct ldlm_lock *lock = NULL;
2241 struct cl_req_attr crattr;
2245 LASSERT(!cfs_list_empty(rpc_list));
2247 memset(&crattr, 0, sizeof crattr);
2248 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2250 GOTO(out, req = ERR_PTR(-ENOMEM));
2254 GOTO(out, req = ERR_PTR(-ENOMEM));
2257 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2258 struct cl_page *page = osc_oap2cl_page(oap);
2260 ops = oap->oap_caller_ops;
2261 caller_data = oap->oap_caller_data;
2263 clerq = cl_req_alloc(env, page, crt,
2264 1 /* only 1-object rpcs for
2267 GOTO(out, req = (void *)clerq);
2268 lock = oap->oap_ldlm_lock;
2270 pga[i] = &oap->oap_brw_page;
2271 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2272 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2273 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2275 cl_req_page_add(env, clerq, page);
2278 /* always get the data for the obdo for the rpc */
2279 LASSERT(ops != NULL);
2281 crattr.cra_capa = NULL;
2282 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2284 oa->o_handle = lock->l_remote_handle;
2285 oa->o_valid |= OBD_MD_FLHANDLE;
2288 rc = cl_req_prep(env, clerq);
2290 CERROR("cl_req_prep failed: %d\n", rc);
2291 GOTO(out, req = ERR_PTR(rc));
2294 sort_brw_pages(pga, page_count);
2295 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2296 pga, &req, crattr.cra_capa, 1, 0);
2298 CERROR("prep_req failed: %d\n", rc);
2299 GOTO(out, req = ERR_PTR(rc));
2302 /* Need to update the timestamps after the request is built in case
2303 * we race with setattr (locally or in queue at OST). If OST gets
2304 * later setattr before earlier BRW (as determined by the request xid),
2305 * the OST will not use BRW timestamps. Sadly, there is no obvious
2306 * way to do this in a single call. bug 10150 */
2307 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2308 cl_req_attr_set(env, clerq, &crattr,
2309 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2311 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2312 aa = ptlrpc_req_async_args(req);
2313 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2314 cfs_list_splice(rpc_list, &aa->aa_oaps);
2315 CFS_INIT_LIST_HEAD(rpc_list);
2316 aa->aa_clerq = clerq;
2318 capa_put(crattr.cra_capa);
2323 OBD_FREE(pga, sizeof(*pga) * page_count);
2324 /* this should happen rarely and is pretty bad, it makes the
2325 * pending list not follow the dirty order */
2326 client_obd_list_lock(&cli->cl_loi_list_lock);
2327 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2328 cfs_list_del_init(&oap->oap_rpc_item);
2330 /* queued sync pages can be torn down while the pages
2331 * were between the pending list and the rpc */
2332 if (oap->oap_interrupted) {
2333 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2334 osc_ap_completion(env, cli, NULL, oap, 0,
2338 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2340 if (clerq && !IS_ERR(clerq))
2341 cl_req_completion(env, clerq, PTR_ERR(req));
2347 * prepare pages for ASYNC io and put pages in send queue.
2349 * \param cmd OBD_BRW_* macroses
2350 * \param lop pending pages
2352 * \return zero if pages successfully add to send queue.
2353 * \return not zere if error occurring.
2356 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2357 struct lov_oinfo *loi,
2358 int cmd, struct loi_oap_pages *lop)
2360 struct ptlrpc_request *req;
2361 obd_count page_count = 0;
2362 struct osc_async_page *oap = NULL, *tmp;
2363 struct osc_brw_async_args *aa;
2364 const struct obd_async_page_ops *ops;
2365 CFS_LIST_HEAD(rpc_list);
2366 CFS_LIST_HEAD(tmp_list);
2367 unsigned int ending_offset;
2368 unsigned starting_offset = 0;
2370 struct cl_object *clob = NULL;
2373 /* ASYNC_HP pages first. At present, when the lock the pages is
2374 * to be canceled, the pages covered by the lock will be sent out
2375 * with ASYNC_HP. We have to send out them as soon as possible. */
2376 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2377 if (oap->oap_async_flags & ASYNC_HP)
2378 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2380 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2381 if (++page_count >= cli->cl_max_pages_per_rpc)
2385 cfs_list_splice(&tmp_list, &lop->lop_pending);
2388 /* first we find the pages we're allowed to work with */
2389 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2391 ops = oap->oap_caller_ops;
2393 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2394 "magic 0x%x\n", oap, oap->oap_magic);
2397 /* pin object in memory, so that completion call-backs
2398 * can be safely called under client_obd_list lock. */
2399 clob = osc_oap2cl_page(oap)->cp_obj;
2400 cl_object_get(clob);
2403 if (page_count != 0 &&
2404 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2405 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2406 " oap %p, page %p, srvlock %u\n",
2407 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2411 /* If there is a gap at the start of this page, it can't merge
2412 * with any previous page, so we'll hand the network a
2413 * "fragmented" page array that it can't transfer in 1 RDMA */
2414 if (page_count != 0 && oap->oap_page_off != 0)
2417 /* in llite being 'ready' equates to the page being locked
2418 * until completion unlocks it. commit_write submits a page
2419 * as not ready because its unlock will happen unconditionally
2420 * as the call returns. if we race with commit_write giving
2421 * us that page we dont' want to create a hole in the page
2422 * stream, so we stop and leave the rpc to be fired by
2423 * another dirtier or kupdated interval (the not ready page
2424 * will still be on the dirty list). we could call in
2425 * at the end of ll_file_write to process the queue again. */
2426 if (!(oap->oap_async_flags & ASYNC_READY)) {
2427 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2430 CDEBUG(D_INODE, "oap %p page %p returned %d "
2431 "instead of ready\n", oap,
2435 /* llite is telling us that the page is still
2436 * in commit_write and that we should try
2437 * and put it in an rpc again later. we
2438 * break out of the loop so we don't create
2439 * a hole in the sequence of pages in the rpc
2444 /* the io isn't needed.. tell the checks
2445 * below to complete the rpc with EINTR */
2446 cfs_spin_lock(&oap->oap_lock);
2447 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2448 cfs_spin_unlock(&oap->oap_lock);
2449 oap->oap_count = -EINTR;
2452 cfs_spin_lock(&oap->oap_lock);
2453 oap->oap_async_flags |= ASYNC_READY;
2454 cfs_spin_unlock(&oap->oap_lock);
2457 LASSERTF(0, "oap %p page %p returned %d "
2458 "from make_ready\n", oap,
2466 * Page submitted for IO has to be locked. Either by
2467 * ->ap_make_ready() or by higher layers.
2469 #if defined(__KERNEL__) && defined(__linux__)
2471 struct cl_page *page;
2473 page = osc_oap2cl_page(oap);
2475 if (page->cp_type == CPT_CACHEABLE &&
2476 !(PageLocked(oap->oap_page) &&
2477 (CheckWriteback(oap->oap_page, cmd)))) {
2478 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2480 (long)oap->oap_page->flags,
2481 oap->oap_async_flags);
2487 /* take the page out of our book-keeping */
2488 cfs_list_del_init(&oap->oap_pending_item);
2489 lop_update_pending(cli, lop, cmd, -1);
2490 cfs_list_del_init(&oap->oap_urgent_item);
2492 if (page_count == 0)
2493 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2494 (PTLRPC_MAX_BRW_SIZE - 1);
2496 /* ask the caller for the size of the io as the rpc leaves. */
2497 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2499 ops->ap_refresh_count(env, oap->oap_caller_data,
2501 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2503 if (oap->oap_count <= 0) {
2504 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2506 osc_ap_completion(env, cli, NULL,
2507 oap, 0, oap->oap_count);
2511 /* now put the page back in our accounting */
2512 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2513 if (page_count == 0)
2514 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2515 if (++page_count >= cli->cl_max_pages_per_rpc)
2518 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2519 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2520 * have the same alignment as the initial writes that allocated
2521 * extents on the server. */
2522 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2523 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2524 if (ending_offset == 0)
2527 /* If there is a gap at the end of this page, it can't merge
2528 * with any subsequent pages, so we'll hand the network a
2529 * "fragmented" page array that it can't transfer in 1 RDMA */
2530 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2534 osc_wake_cache_waiters(cli);
2536 loi_list_maint(cli, loi);
2538 client_obd_list_unlock(&cli->cl_loi_list_lock);
2541 cl_object_put(env, clob);
2543 if (page_count == 0) {
2544 client_obd_list_lock(&cli->cl_loi_list_lock);
2548 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2550 LASSERT(cfs_list_empty(&rpc_list));
2551 loi_list_maint(cli, loi);
2552 RETURN(PTR_ERR(req));
2555 aa = ptlrpc_req_async_args(req);
2557 if (cmd == OBD_BRW_READ) {
2558 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2559 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2560 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2561 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2563 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2564 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2565 cli->cl_w_in_flight);
2566 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2567 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2569 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2571 client_obd_list_lock(&cli->cl_loi_list_lock);
2573 if (cmd == OBD_BRW_READ)
2574 cli->cl_r_in_flight++;
2576 cli->cl_w_in_flight++;
2578 /* queued sync pages can be torn down while the pages
2579 * were between the pending list and the rpc */
2581 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2582 /* only one oap gets a request reference */
2585 if (oap->oap_interrupted && !req->rq_intr) {
2586 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2588 ptlrpc_mark_interrupted(req);
2592 tmp->oap_request = ptlrpc_request_addref(req);
2594 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2595 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2597 req->rq_interpret_reply = brw_interpret;
2598 ptlrpcd_add_req(req, PSCOPE_BRW);
2602 #define LOI_DEBUG(LOI, STR, args...) \
2603 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2604 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2605 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2606 (LOI)->loi_write_lop.lop_num_pending, \
2607 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2608 (LOI)->loi_read_lop.lop_num_pending, \
2609 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2612 /* This is called by osc_check_rpcs() to find which objects have pages that
2613 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2614 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2618 /* First return objects that have blocked locks so that they
2619 * will be flushed quickly and other clients can get the lock,
2620 * then objects which have pages ready to be stuffed into RPCs */
2621 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2622 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2623 struct lov_oinfo, loi_hp_ready_item));
2624 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2625 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2626 struct lov_oinfo, loi_ready_item));
2628 /* then if we have cache waiters, return all objects with queued
2629 * writes. This is especially important when many small files
2630 * have filled up the cache and not been fired into rpcs because
2631 * they don't pass the nr_pending/object threshhold */
2632 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2633 !cfs_list_empty(&cli->cl_loi_write_list))
2634 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2635 struct lov_oinfo, loi_write_item));
2637 /* then return all queued objects when we have an invalid import
2638 * so that they get flushed */
2639 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2640 if (!cfs_list_empty(&cli->cl_loi_write_list))
2641 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2644 if (!cfs_list_empty(&cli->cl_loi_read_list))
2645 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2646 struct lov_oinfo, loi_read_item));
2651 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2653 struct osc_async_page *oap;
2656 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2657 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2658 struct osc_async_page, oap_urgent_item);
2659 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2662 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2663 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2664 struct osc_async_page, oap_urgent_item);
2665 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2668 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2671 /* called with the loi list lock held */
2672 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2674 struct lov_oinfo *loi;
2675 int rc = 0, race_counter = 0;
2678 while ((loi = osc_next_loi(cli)) != NULL) {
2679 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2681 if (osc_max_rpc_in_flight(cli, loi))
2684 /* attempt some read/write balancing by alternating between
2685 * reads and writes in an object. The makes_rpc checks here
2686 * would be redundant if we were getting read/write work items
2687 * instead of objects. we don't want send_oap_rpc to drain a
2688 * partial read pending queue when we're given this object to
2689 * do io on writes while there are cache waiters */
2690 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2691 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2692 &loi->loi_write_lop);
2694 CERROR("Write request failed with %d\n", rc);
2696 /* osc_send_oap_rpc failed, mostly because of
2699 * It can't break here, because if:
2700 * - a page was submitted by osc_io_submit, so
2702 * - no request in flight
2703 * - no subsequent request
2704 * The system will be in live-lock state,
2705 * because there is no chance to call
2706 * osc_io_unplug() and osc_check_rpcs() any
2707 * more. pdflush can't help in this case,
2708 * because it might be blocked at grabbing
2709 * the page lock as we mentioned.
2711 * Anyway, continue to drain pages. */
2720 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2721 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2722 &loi->loi_read_lop);
2724 CERROR("Read request failed with %d\n", rc);
2732 /* attempt some inter-object balancing by issueing rpcs
2733 * for each object in turn */
2734 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2735 cfs_list_del_init(&loi->loi_hp_ready_item);
2736 if (!cfs_list_empty(&loi->loi_ready_item))
2737 cfs_list_del_init(&loi->loi_ready_item);
2738 if (!cfs_list_empty(&loi->loi_write_item))
2739 cfs_list_del_init(&loi->loi_write_item);
2740 if (!cfs_list_empty(&loi->loi_read_item))
2741 cfs_list_del_init(&loi->loi_read_item);
2743 loi_list_maint(cli, loi);
2745 /* send_oap_rpc fails with 0 when make_ready tells it to
2746 * back off. llite's make_ready does this when it tries
2747 * to lock a page queued for write that is already locked.
2748 * we want to try sending rpcs from many objects, but we
2749 * don't want to spin failing with 0. */
2750 if (race_counter == 10)
2756 /* we're trying to queue a page in the osc so we're subject to the
2757 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2758 * If the osc's queued pages are already at that limit, then we want to sleep
2759 * until there is space in the osc's queue for us. We also may be waiting for
2760 * write credits from the OST if there are RPCs in flight that may return some
2761 * before we fall back to sync writes.
2763 * We need this know our allocation was granted in the presence of signals */
2764 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2768 client_obd_list_lock(&cli->cl_loi_list_lock);
2769 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2770 client_obd_list_unlock(&cli->cl_loi_list_lock);
2775 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2778 int osc_enter_cache_try(const struct lu_env *env,
2779 struct client_obd *cli, struct lov_oinfo *loi,
2780 struct osc_async_page *oap, int transient)
2784 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2786 osc_consume_write_grant(cli, &oap->oap_brw_page);
2788 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2789 cfs_atomic_inc(&obd_dirty_transit_pages);
2790 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2796 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2797 * grant or cache space. */
2798 static int osc_enter_cache(const struct lu_env *env,
2799 struct client_obd *cli, struct lov_oinfo *loi,
2800 struct osc_async_page *oap)
2802 struct osc_cache_waiter ocw;
2803 struct l_wait_info lwi = { 0 };
2807 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2808 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2809 cli->cl_dirty_max, obd_max_dirty_pages,
2810 cli->cl_lost_grant, cli->cl_avail_grant);
2812 /* force the caller to try sync io. this can jump the list
2813 * of queued writes and create a discontiguous rpc stream */
2814 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2815 loi->loi_ar.ar_force_sync)
2818 /* Hopefully normal case - cache space and write credits available */
2819 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2820 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2821 osc_enter_cache_try(env, cli, loi, oap, 0))
2824 /* It is safe to block as a cache waiter as long as there is grant
2825 * space available or the hope of additional grant being returned
2826 * when an in flight write completes. Using the write back cache
2827 * if possible is preferable to sending the data synchronously
2828 * because write pages can then be merged in to large requests.
2829 * The addition of this cache waiter will causing pending write
2830 * pages to be sent immediately. */
2831 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2832 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2833 cfs_waitq_init(&ocw.ocw_waitq);
2837 loi_list_maint(cli, loi);
2838 osc_check_rpcs(env, cli);
2839 client_obd_list_unlock(&cli->cl_loi_list_lock);
2841 CDEBUG(D_CACHE, "sleeping for cache space\n");
2842 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2844 client_obd_list_lock(&cli->cl_loi_list_lock);
2845 if (!cfs_list_empty(&ocw.ocw_entry)) {
2846 cfs_list_del(&ocw.ocw_entry);
2856 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2857 struct lov_oinfo *loi, cfs_page_t *page,
2858 obd_off offset, const struct obd_async_page_ops *ops,
2859 void *data, void **res, int nocache,
2860 struct lustre_handle *lockh)
2862 struct osc_async_page *oap;
2867 return cfs_size_round(sizeof(*oap));
2870 oap->oap_magic = OAP_MAGIC;
2871 oap->oap_cli = &exp->exp_obd->u.cli;
2874 oap->oap_caller_ops = ops;
2875 oap->oap_caller_data = data;
2877 oap->oap_page = page;
2878 oap->oap_obj_off = offset;
2879 if (!client_is_remote(exp) &&
2880 cfs_capable(CFS_CAP_SYS_RESOURCE))
2881 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2883 LASSERT(!(offset & ~CFS_PAGE_MASK));
2885 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2886 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2887 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2888 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2890 cfs_spin_lock_init(&oap->oap_lock);
2891 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2895 struct osc_async_page *oap_from_cookie(void *cookie)
2897 struct osc_async_page *oap = cookie;
2898 if (oap->oap_magic != OAP_MAGIC)
2899 return ERR_PTR(-EINVAL);
2903 int osc_queue_async_io(const struct lu_env *env,
2904 struct obd_export *exp, struct lov_stripe_md *lsm,
2905 struct lov_oinfo *loi, void *cookie,
2906 int cmd, obd_off off, int count,
2907 obd_flag brw_flags, enum async_flags async_flags)
2909 struct client_obd *cli = &exp->exp_obd->u.cli;
2910 struct osc_async_page *oap;
2914 oap = oap_from_cookie(cookie);
2916 RETURN(PTR_ERR(oap));
2918 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2921 if (!cfs_list_empty(&oap->oap_pending_item) ||
2922 !cfs_list_empty(&oap->oap_urgent_item) ||
2923 !cfs_list_empty(&oap->oap_rpc_item))
2926 /* check if the file's owner/group is over quota */
2927 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2928 struct cl_object *obj;
2929 struct cl_attr attr; /* XXX put attr into thread info */
2930 unsigned int qid[MAXQUOTAS];
2932 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2934 cl_object_attr_lock(obj);
2935 rc = cl_object_attr_get(env, obj, &attr);
2936 cl_object_attr_unlock(obj);
2938 qid[USRQUOTA] = attr.cat_uid;
2939 qid[GRPQUOTA] = attr.cat_gid;
2941 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2948 loi = lsm->lsm_oinfo[0];
2950 client_obd_list_lock(&cli->cl_loi_list_lock);
2952 LASSERT(off + count <= CFS_PAGE_SIZE);
2954 oap->oap_page_off = off;
2955 oap->oap_count = count;
2956 oap->oap_brw_flags = brw_flags;
2957 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2958 if (libcfs_memory_pressure_get())
2959 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2960 cfs_spin_lock(&oap->oap_lock);
2961 oap->oap_async_flags = async_flags;
2962 cfs_spin_unlock(&oap->oap_lock);
2964 if (cmd & OBD_BRW_WRITE) {
2965 rc = osc_enter_cache(env, cli, loi, oap);
2967 client_obd_list_unlock(&cli->cl_loi_list_lock);
2972 osc_oap_to_pending(oap);
2973 loi_list_maint(cli, loi);
2975 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2978 osc_check_rpcs(env, cli);
2979 client_obd_list_unlock(&cli->cl_loi_list_lock);
2984 /* aka (~was & now & flag), but this is more clear :) */
2985 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2987 int osc_set_async_flags_base(struct client_obd *cli,
2988 struct lov_oinfo *loi, struct osc_async_page *oap,
2989 obd_flag async_flags)
2991 struct loi_oap_pages *lop;
2995 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
2997 if (oap->oap_cmd & OBD_BRW_WRITE) {
2998 lop = &loi->loi_write_lop;
3000 lop = &loi->loi_read_lop;
3003 if ((oap->oap_async_flags & async_flags) == async_flags)
3006 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3007 flags |= ASYNC_READY;
3009 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3010 cfs_list_empty(&oap->oap_rpc_item)) {
3011 if (oap->oap_async_flags & ASYNC_HP)
3012 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3014 cfs_list_add_tail(&oap->oap_urgent_item,
3016 flags |= ASYNC_URGENT;
3017 loi_list_maint(cli, loi);
3019 cfs_spin_lock(&oap->oap_lock);
3020 oap->oap_async_flags |= flags;
3021 cfs_spin_unlock(&oap->oap_lock);
3023 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3024 oap->oap_async_flags);
3028 int osc_teardown_async_page(struct obd_export *exp,
3029 struct lov_stripe_md *lsm,
3030 struct lov_oinfo *loi, void *cookie)
3032 struct client_obd *cli = &exp->exp_obd->u.cli;
3033 struct loi_oap_pages *lop;
3034 struct osc_async_page *oap;
3038 oap = oap_from_cookie(cookie);
3040 RETURN(PTR_ERR(oap));
3043 loi = lsm->lsm_oinfo[0];
3045 if (oap->oap_cmd & OBD_BRW_WRITE) {
3046 lop = &loi->loi_write_lop;
3048 lop = &loi->loi_read_lop;
3051 client_obd_list_lock(&cli->cl_loi_list_lock);
3053 if (!cfs_list_empty(&oap->oap_rpc_item))
3054 GOTO(out, rc = -EBUSY);
3056 osc_exit_cache(cli, oap, 0);
3057 osc_wake_cache_waiters(cli);
3059 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3060 cfs_list_del_init(&oap->oap_urgent_item);
3061 cfs_spin_lock(&oap->oap_lock);
3062 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3063 cfs_spin_unlock(&oap->oap_lock);
3065 if (!cfs_list_empty(&oap->oap_pending_item)) {
3066 cfs_list_del_init(&oap->oap_pending_item);
3067 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3069 loi_list_maint(cli, loi);
3070 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3072 client_obd_list_unlock(&cli->cl_loi_list_lock);
3076 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3077 struct ldlm_enqueue_info *einfo,
3080 void *data = einfo->ei_cbdata;
3082 LASSERT(lock != NULL);
3083 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3084 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3085 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3086 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3088 lock_res_and_lock(lock);
3089 cfs_spin_lock(&osc_ast_guard);
3090 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3091 lock->l_ast_data = data;
3092 cfs_spin_unlock(&osc_ast_guard);
3093 unlock_res_and_lock(lock);
3096 static void osc_set_data_with_check(struct lustre_handle *lockh,
3097 struct ldlm_enqueue_info *einfo,
3100 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3103 osc_set_lock_data_with_check(lock, einfo, flags);
3104 LDLM_LOCK_PUT(lock);
3106 CERROR("lockh %p, data %p - client evicted?\n",
3107 lockh, einfo->ei_cbdata);
3110 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3111 ldlm_iterator_t replace, void *data)
3113 struct ldlm_res_id res_id;
3114 struct obd_device *obd = class_exp2obd(exp);
3116 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3117 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3121 /* find any ldlm lock of the inode in osc
3125 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3126 ldlm_iterator_t replace, void *data)
3128 struct ldlm_res_id res_id;
3129 struct obd_device *obd = class_exp2obd(exp);
3132 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3133 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3134 if (rc == LDLM_ITER_STOP)
3136 if (rc == LDLM_ITER_CONTINUE)
3141 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3142 obd_enqueue_update_f upcall, void *cookie,
3145 int intent = *flags & LDLM_FL_HAS_INTENT;
3149 /* The request was created before ldlm_cli_enqueue call. */
3150 if (rc == ELDLM_LOCK_ABORTED) {
3151 struct ldlm_reply *rep;
3152 rep = req_capsule_server_get(&req->rq_pill,
3155 LASSERT(rep != NULL);
3156 if (rep->lock_policy_res1)
3157 rc = rep->lock_policy_res1;
3161 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3162 *flags |= LDLM_FL_LVB_READY;
3163 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3164 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3167 /* Call the update callback. */
3168 rc = (*upcall)(cookie, rc);
3172 static int osc_enqueue_interpret(const struct lu_env *env,
3173 struct ptlrpc_request *req,
3174 struct osc_enqueue_args *aa, int rc)
3176 struct ldlm_lock *lock;
3177 struct lustre_handle handle;
3180 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3181 * might be freed anytime after lock upcall has been called. */
3182 lustre_handle_copy(&handle, aa->oa_lockh);
3183 mode = aa->oa_ei->ei_mode;
3185 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3187 lock = ldlm_handle2lock(&handle);
3189 /* Take an additional reference so that a blocking AST that
3190 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3191 * to arrive after an upcall has been executed by
3192 * osc_enqueue_fini(). */
3193 ldlm_lock_addref(&handle, mode);
3195 /* Complete obtaining the lock procedure. */
3196 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3197 mode, aa->oa_flags, aa->oa_lvb,
3198 sizeof(*aa->oa_lvb), &handle, rc);
3199 /* Complete osc stuff. */
3200 rc = osc_enqueue_fini(req, aa->oa_lvb,
3201 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3203 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3205 /* Release the lock for async request. */
3206 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3208 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3209 * not already released by
3210 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3212 ldlm_lock_decref(&handle, mode);
3214 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3215 aa->oa_lockh, req, aa);
3216 ldlm_lock_decref(&handle, mode);
3217 LDLM_LOCK_PUT(lock);
3221 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3222 struct lov_oinfo *loi, int flags,
3223 struct ost_lvb *lvb, __u32 mode, int rc)
3225 if (rc == ELDLM_OK) {
3226 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3229 LASSERT(lock != NULL);
3230 loi->loi_lvb = *lvb;
3231 tmp = loi->loi_lvb.lvb_size;
3232 /* Extend KMS up to the end of this lock and no further
3233 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3234 if (tmp > lock->l_policy_data.l_extent.end)
3235 tmp = lock->l_policy_data.l_extent.end + 1;
3236 if (tmp >= loi->loi_kms) {
3237 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3238 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3239 loi_kms_set(loi, tmp);
3241 LDLM_DEBUG(lock, "lock acquired, setting rss="
3242 LPU64"; leaving kms="LPU64", end="LPU64,
3243 loi->loi_lvb.lvb_size, loi->loi_kms,
3244 lock->l_policy_data.l_extent.end);
3246 ldlm_lock_allow_match(lock);
3247 LDLM_LOCK_PUT(lock);
3248 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3249 loi->loi_lvb = *lvb;
3250 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3251 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3255 EXPORT_SYMBOL(osc_update_enqueue);
3257 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3259 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3260 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3261 * other synchronous requests, however keeping some locks and trying to obtain
3262 * others may take a considerable amount of time in a case of ost failure; and
3263 * when other sync requests do not get released lock from a client, the client
3264 * is excluded from the cluster -- such scenarious make the life difficult, so
3265 * release locks just after they are obtained. */
3266 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3267 int *flags, ldlm_policy_data_t *policy,
3268 struct ost_lvb *lvb, int kms_valid,
3269 obd_enqueue_update_f upcall, void *cookie,
3270 struct ldlm_enqueue_info *einfo,
3271 struct lustre_handle *lockh,
3272 struct ptlrpc_request_set *rqset, int async)
3274 struct obd_device *obd = exp->exp_obd;
3275 struct ptlrpc_request *req = NULL;
3276 int intent = *flags & LDLM_FL_HAS_INTENT;
3281 /* Filesystem lock extents are extended to page boundaries so that
3282 * dealing with the page cache is a little smoother. */
3283 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3284 policy->l_extent.end |= ~CFS_PAGE_MASK;
3287 * kms is not valid when either object is completely fresh (so that no
3288 * locks are cached), or object was evicted. In the latter case cached
3289 * lock cannot be used, because it would prime inode state with
3290 * potentially stale LVB.
3295 /* Next, search for already existing extent locks that will cover us */
3296 /* If we're trying to read, we also search for an existing PW lock. The
3297 * VFS and page cache already protect us locally, so lots of readers/
3298 * writers can share a single PW lock.
3300 * There are problems with conversion deadlocks, so instead of
3301 * converting a read lock to a write lock, we'll just enqueue a new
3304 * At some point we should cancel the read lock instead of making them
3305 * send us a blocking callback, but there are problems with canceling
3306 * locks out from other users right now, too. */
3307 mode = einfo->ei_mode;
3308 if (einfo->ei_mode == LCK_PR)
3310 mode = ldlm_lock_match(obd->obd_namespace,
3311 *flags | LDLM_FL_LVB_READY, res_id,
3312 einfo->ei_type, policy, mode, lockh, 0);
3314 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3316 if (matched->l_ast_data == NULL ||
3317 matched->l_ast_data == einfo->ei_cbdata) {
3318 /* addref the lock only if not async requests and PW
3319 * lock is matched whereas we asked for PR. */
3320 if (!rqset && einfo->ei_mode != mode)
3321 ldlm_lock_addref(lockh, LCK_PR);
3322 osc_set_lock_data_with_check(matched, einfo, *flags);
3324 /* I would like to be able to ASSERT here that
3325 * rss <= kms, but I can't, for reasons which
3326 * are explained in lov_enqueue() */
3329 /* We already have a lock, and it's referenced */
3330 (*upcall)(cookie, ELDLM_OK);
3332 /* For async requests, decref the lock. */
3333 if (einfo->ei_mode != mode)
3334 ldlm_lock_decref(lockh, LCK_PW);
3336 ldlm_lock_decref(lockh, einfo->ei_mode);
3337 LDLM_LOCK_PUT(matched);
3340 ldlm_lock_decref(lockh, mode);
3341 LDLM_LOCK_PUT(matched);
3346 CFS_LIST_HEAD(cancels);
3347 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3348 &RQF_LDLM_ENQUEUE_LVB);
3352 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3356 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3358 ptlrpc_request_set_replen(req);
3361 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3362 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3364 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3365 sizeof(*lvb), lockh, async);
3368 struct osc_enqueue_args *aa;
3369 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3370 aa = ptlrpc_req_async_args(req);
3373 aa->oa_flags = flags;
3374 aa->oa_upcall = upcall;
3375 aa->oa_cookie = cookie;
3377 aa->oa_lockh = lockh;
3379 req->rq_interpret_reply =
3380 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3381 if (rqset == PTLRPCD_SET)
3382 ptlrpcd_add_req(req, PSCOPE_OTHER);
3384 ptlrpc_set_add_req(rqset, req);
3385 } else if (intent) {
3386 ptlrpc_req_finished(req);
3391 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3393 ptlrpc_req_finished(req);
3398 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3399 struct ldlm_enqueue_info *einfo,
3400 struct ptlrpc_request_set *rqset)
3402 struct ldlm_res_id res_id;
3406 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3407 oinfo->oi_md->lsm_object_seq, &res_id);
3409 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3410 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3411 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3412 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3413 rqset, rqset != NULL);
3417 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3418 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3419 int *flags, void *data, struct lustre_handle *lockh,
3422 struct obd_device *obd = exp->exp_obd;
3423 int lflags = *flags;
3427 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3430 /* Filesystem lock extents are extended to page boundaries so that
3431 * dealing with the page cache is a little smoother */
3432 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3433 policy->l_extent.end |= ~CFS_PAGE_MASK;
3435 /* Next, search for already existing extent locks that will cover us */
3436 /* If we're trying to read, we also search for an existing PW lock. The
3437 * VFS and page cache already protect us locally, so lots of readers/
3438 * writers can share a single PW lock. */
3442 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3443 res_id, type, policy, rc, lockh, unref);
3446 osc_set_data_with_check(lockh, data, lflags);
3447 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3448 ldlm_lock_addref(lockh, LCK_PR);
3449 ldlm_lock_decref(lockh, LCK_PW);
3456 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3460 if (unlikely(mode == LCK_GROUP))
3461 ldlm_lock_decref_and_cancel(lockh, mode);
3463 ldlm_lock_decref(lockh, mode);
3468 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3469 __u32 mode, struct lustre_handle *lockh)
3472 RETURN(osc_cancel_base(lockh, mode));
3475 static int osc_cancel_unused(struct obd_export *exp,
3476 struct lov_stripe_md *lsm,
3477 ldlm_cancel_flags_t flags,
3480 struct obd_device *obd = class_exp2obd(exp);
3481 struct ldlm_res_id res_id, *resp = NULL;
3484 resp = osc_build_res_name(lsm->lsm_object_id,
3485 lsm->lsm_object_seq, &res_id);
3488 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3491 static int osc_statfs_interpret(const struct lu_env *env,
3492 struct ptlrpc_request *req,
3493 struct osc_async_args *aa, int rc)
3495 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3496 struct obd_statfs *msfs;
3501 /* The request has in fact never been sent
3502 * due to issues at a higher level (LOV).
3503 * Exit immediately since the caller is
3504 * aware of the problem and takes care
3505 * of the clean up */
3508 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3509 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3515 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3517 GOTO(out, rc = -EPROTO);
3520 /* Reinitialize the RDONLY and DEGRADED flags at the client
3521 * on each statfs, so they don't stay set permanently. */
3522 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3524 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3525 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3526 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3527 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3529 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3530 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3531 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3532 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3534 /* Add a bit of hysteresis so this flag isn't continually flapping,
3535 * and ensure that new files don't get extremely fragmented due to
3536 * only a small amount of available space in the filesystem.
3537 * We want to set the NOSPC flag when there is less than ~0.1% free
3538 * and clear it when there is at least ~0.2% free space, so:
3539 * avail < ~0.1% max max = avail + used
3540 * 1025 * avail < avail + used used = blocks - free
3541 * 1024 * avail < used
3542 * 1024 * avail < blocks - free
3543 * avail < ((blocks - free) >> 10)
3545 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3546 * lose that amount of space so in those cases we report no space left
3547 * if their is less than 1 GB left. */
3548 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3549 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3550 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3551 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3552 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3553 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3554 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3556 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3558 *aa->aa_oi->oi_osfs = *msfs;
3560 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3564 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3565 __u64 max_age, struct ptlrpc_request_set *rqset)
3567 struct ptlrpc_request *req;
3568 struct osc_async_args *aa;
3572 /* We could possibly pass max_age in the request (as an absolute
3573 * timestamp or a "seconds.usec ago") so the target can avoid doing
3574 * extra calls into the filesystem if that isn't necessary (e.g.
3575 * during mount that would help a bit). Having relative timestamps
3576 * is not so great if request processing is slow, while absolute
3577 * timestamps are not ideal because they need time synchronization. */
3578 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3582 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3584 ptlrpc_request_free(req);
3587 ptlrpc_request_set_replen(req);
3588 req->rq_request_portal = OST_CREATE_PORTAL;
3589 ptlrpc_at_set_req_timeout(req);
3591 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3592 /* procfs requests not want stat in wait for avoid deadlock */
3593 req->rq_no_resend = 1;
3594 req->rq_no_delay = 1;
3597 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3598 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3599 aa = ptlrpc_req_async_args(req);
3602 ptlrpc_set_add_req(rqset, req);
3606 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3607 __u64 max_age, __u32 flags)
3609 struct obd_statfs *msfs;
3610 struct ptlrpc_request *req;
3611 struct obd_import *imp = NULL;
3615 /*Since the request might also come from lprocfs, so we need
3616 *sync this with client_disconnect_export Bug15684*/
3617 cfs_down_read(&obd->u.cli.cl_sem);
3618 if (obd->u.cli.cl_import)
3619 imp = class_import_get(obd->u.cli.cl_import);
3620 cfs_up_read(&obd->u.cli.cl_sem);
3624 /* We could possibly pass max_age in the request (as an absolute
3625 * timestamp or a "seconds.usec ago") so the target can avoid doing
3626 * extra calls into the filesystem if that isn't necessary (e.g.
3627 * during mount that would help a bit). Having relative timestamps
3628 * is not so great if request processing is slow, while absolute
3629 * timestamps are not ideal because they need time synchronization. */
3630 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3632 class_import_put(imp);
3637 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3639 ptlrpc_request_free(req);
3642 ptlrpc_request_set_replen(req);
3643 req->rq_request_portal = OST_CREATE_PORTAL;
3644 ptlrpc_at_set_req_timeout(req);
3646 if (flags & OBD_STATFS_NODELAY) {
3647 /* procfs requests not want stat in wait for avoid deadlock */
3648 req->rq_no_resend = 1;
3649 req->rq_no_delay = 1;
3652 rc = ptlrpc_queue_wait(req);
3656 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3658 GOTO(out, rc = -EPROTO);
3665 ptlrpc_req_finished(req);
3669 /* Retrieve object striping information.
3671 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3672 * the maximum number of OST indices which will fit in the user buffer.
3673 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3675 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3677 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3678 struct lov_user_md_v3 lum, *lumk;
3679 struct lov_user_ost_data_v1 *lmm_objects;
3680 int rc = 0, lum_size;
3686 /* we only need the header part from user space to get lmm_magic and
3687 * lmm_stripe_count, (the header part is common to v1 and v3) */
3688 lum_size = sizeof(struct lov_user_md_v1);
3689 if (cfs_copy_from_user(&lum, lump, lum_size))
3692 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3693 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3696 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3697 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3698 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3699 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3701 /* we can use lov_mds_md_size() to compute lum_size
3702 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3703 if (lum.lmm_stripe_count > 0) {
3704 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3705 OBD_ALLOC(lumk, lum_size);
3709 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3710 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3712 lmm_objects = &(lumk->lmm_objects[0]);
3713 lmm_objects->l_object_id = lsm->lsm_object_id;
3715 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3719 lumk->lmm_object_id = lsm->lsm_object_id;
3720 lumk->lmm_object_seq = lsm->lsm_object_seq;
3721 lumk->lmm_stripe_count = 1;
3723 if (cfs_copy_to_user(lump, lumk, lum_size))
3727 OBD_FREE(lumk, lum_size);
3733 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3734 void *karg, void *uarg)
3736 struct obd_device *obd = exp->exp_obd;
3737 struct obd_ioctl_data *data = karg;
3741 if (!cfs_try_module_get(THIS_MODULE)) {
3742 CERROR("Can't get module. Is it alive?");
3746 case OBD_IOC_LOV_GET_CONFIG: {
3748 struct lov_desc *desc;
3749 struct obd_uuid uuid;
3753 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3754 GOTO(out, err = -EINVAL);
3756 data = (struct obd_ioctl_data *)buf;
3758 if (sizeof(*desc) > data->ioc_inllen1) {
3759 obd_ioctl_freedata(buf, len);
3760 GOTO(out, err = -EINVAL);
3763 if (data->ioc_inllen2 < sizeof(uuid)) {
3764 obd_ioctl_freedata(buf, len);
3765 GOTO(out, err = -EINVAL);
3768 desc = (struct lov_desc *)data->ioc_inlbuf1;
3769 desc->ld_tgt_count = 1;
3770 desc->ld_active_tgt_count = 1;
3771 desc->ld_default_stripe_count = 1;
3772 desc->ld_default_stripe_size = 0;
3773 desc->ld_default_stripe_offset = 0;
3774 desc->ld_pattern = 0;
3775 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3777 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3779 err = cfs_copy_to_user((void *)uarg, buf, len);
3782 obd_ioctl_freedata(buf, len);
3785 case LL_IOC_LOV_SETSTRIPE:
3786 err = obd_alloc_memmd(exp, karg);
3790 case LL_IOC_LOV_GETSTRIPE:
3791 err = osc_getstripe(karg, uarg);
3793 case OBD_IOC_CLIENT_RECOVER:
3794 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3799 case IOC_OSC_SET_ACTIVE:
3800 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3803 case OBD_IOC_POLL_QUOTACHECK:
3804 err = lquota_poll_check(quota_interface, exp,
3805 (struct if_quotacheck *)karg);
3807 case OBD_IOC_PING_TARGET:
3808 err = ptlrpc_obd_ping(obd);
3811 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3812 cmd, cfs_curproc_comm());
3813 GOTO(out, err = -ENOTTY);
3816 cfs_module_put(THIS_MODULE);
3820 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3821 void *key, __u32 *vallen, void *val,
3822 struct lov_stripe_md *lsm)
3825 if (!vallen || !val)
3828 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3829 __u32 *stripe = val;
3830 *vallen = sizeof(*stripe);
3833 } else if (KEY_IS(KEY_LAST_ID)) {
3834 struct ptlrpc_request *req;
3839 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3840 &RQF_OST_GET_INFO_LAST_ID);
3844 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3845 RCL_CLIENT, keylen);
3846 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3848 ptlrpc_request_free(req);
3852 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3853 memcpy(tmp, key, keylen);
3855 req->rq_no_delay = req->rq_no_resend = 1;
3856 ptlrpc_request_set_replen(req);
3857 rc = ptlrpc_queue_wait(req);
3861 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3863 GOTO(out, rc = -EPROTO);
3865 *((obd_id *)val) = *reply;
3867 ptlrpc_req_finished(req);
3869 } else if (KEY_IS(KEY_FIEMAP)) {
3870 struct ptlrpc_request *req;
3871 struct ll_user_fiemap *reply;
3875 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3876 &RQF_OST_GET_INFO_FIEMAP);
3880 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3881 RCL_CLIENT, keylen);
3882 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3883 RCL_CLIENT, *vallen);
3884 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3885 RCL_SERVER, *vallen);
3887 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3889 ptlrpc_request_free(req);
3893 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3894 memcpy(tmp, key, keylen);
3895 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3896 memcpy(tmp, val, *vallen);
3898 ptlrpc_request_set_replen(req);
3899 rc = ptlrpc_queue_wait(req);
3903 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3905 GOTO(out1, rc = -EPROTO);
3907 memcpy(val, reply, *vallen);
3909 ptlrpc_req_finished(req);
3917 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3919 struct llog_ctxt *ctxt;
3923 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3925 rc = llog_initiator_connect(ctxt);
3926 llog_ctxt_put(ctxt);
3928 /* XXX return an error? skip setting below flags? */
3931 cfs_spin_lock(&imp->imp_lock);
3932 imp->imp_server_timeout = 1;
3933 imp->imp_pingable = 1;
3934 cfs_spin_unlock(&imp->imp_lock);
3935 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3940 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3941 struct ptlrpc_request *req,
3948 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3951 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3952 void *key, obd_count vallen, void *val,
3953 struct ptlrpc_request_set *set)
3955 struct ptlrpc_request *req;
3956 struct obd_device *obd = exp->exp_obd;
3957 struct obd_import *imp = class_exp2cliimp(exp);
3962 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3964 if (KEY_IS(KEY_NEXT_ID)) {
3966 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3968 if (vallen != sizeof(obd_id))
3973 if (vallen != sizeof(obd_id))
3976 /* avoid race between allocate new object and set next id
3977 * from ll_sync thread */
3978 cfs_spin_lock(&oscc->oscc_lock);
3979 new_val = *((obd_id*)val) + 1;
3980 if (new_val > oscc->oscc_next_id)
3981 oscc->oscc_next_id = new_val;
3982 cfs_spin_unlock(&oscc->oscc_lock);
3983 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3984 exp->exp_obd->obd_name,
3985 obd->u.cli.cl_oscc.oscc_next_id);
3990 if (KEY_IS(KEY_CHECKSUM)) {
3991 if (vallen != sizeof(int))
3993 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3997 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3998 sptlrpc_conf_client_adapt(obd);
4002 if (KEY_IS(KEY_FLUSH_CTX)) {
4003 sptlrpc_import_flush_my_ctx(imp);
4007 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4010 /* We pass all other commands directly to OST. Since nobody calls osc
4011 methods directly and everybody is supposed to go through LOV, we
4012 assume lov checked invalid values for us.
4013 The only recognised values so far are evict_by_nid and mds_conn.
4014 Even if something bad goes through, we'd get a -EINVAL from OST
4017 if (KEY_IS(KEY_GRANT_SHRINK))
4018 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4020 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4025 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4026 RCL_CLIENT, keylen);
4027 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4028 RCL_CLIENT, vallen);
4029 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4031 ptlrpc_request_free(req);
4035 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4036 memcpy(tmp, key, keylen);
4037 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4038 memcpy(tmp, val, vallen);
4040 if (KEY_IS(KEY_MDS_CONN)) {
4041 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4043 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4044 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4045 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4046 req->rq_no_delay = req->rq_no_resend = 1;
4047 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4048 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4049 struct osc_grant_args *aa;
4052 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4053 aa = ptlrpc_req_async_args(req);
4056 ptlrpc_req_finished(req);
4059 *oa = ((struct ost_body *)val)->oa;
4061 req->rq_interpret_reply = osc_shrink_grant_interpret;
4064 ptlrpc_request_set_replen(req);
4065 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4066 LASSERT(set != NULL);
4067 ptlrpc_set_add_req(set, req);
4068 ptlrpc_check_set(NULL, set);
4070 ptlrpcd_add_req(req, PSCOPE_OTHER);
4076 static struct llog_operations osc_size_repl_logops = {
4077 lop_cancel: llog_obd_repl_cancel
4080 static struct llog_operations osc_mds_ost_orig_logops;
4082 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4083 struct obd_device *tgt, struct llog_catid *catid)
4088 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4089 &catid->lci_logid, &osc_mds_ost_orig_logops);
4091 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4095 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4096 NULL, &osc_size_repl_logops);
4098 struct llog_ctxt *ctxt =
4099 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4102 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4107 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4108 obd->obd_name, tgt->obd_name, catid, rc);
4109 CERROR("logid "LPX64":0x%x\n",
4110 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4115 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4116 struct obd_device *disk_obd, int *index)
4118 struct llog_catid catid;
4119 static char name[32] = CATLIST;
4123 LASSERT(olg == &obd->obd_olg);
4125 cfs_mutex_down(&olg->olg_cat_processing);
4126 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4128 CERROR("rc: %d\n", rc);
4132 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4133 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4134 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4136 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4138 CERROR("rc: %d\n", rc);
4142 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4144 CERROR("rc: %d\n", rc);
4149 cfs_mutex_up(&olg->olg_cat_processing);
4154 static int osc_llog_finish(struct obd_device *obd, int count)
4156 struct llog_ctxt *ctxt;
4157 int rc = 0, rc2 = 0;
4160 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4162 rc = llog_cleanup(ctxt);
4164 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4166 rc2 = llog_cleanup(ctxt);
4173 static int osc_reconnect(const struct lu_env *env,
4174 struct obd_export *exp, struct obd_device *obd,
4175 struct obd_uuid *cluuid,
4176 struct obd_connect_data *data,
4179 struct client_obd *cli = &obd->u.cli;
4181 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4184 client_obd_list_lock(&cli->cl_loi_list_lock);
4185 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4186 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4187 lost_grant = cli->cl_lost_grant;
4188 cli->cl_lost_grant = 0;
4189 client_obd_list_unlock(&cli->cl_loi_list_lock);
4191 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4192 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4193 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4194 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4195 " ocd_grant: %d\n", data->ocd_connect_flags,
4196 data->ocd_version, data->ocd_grant);
4202 static int osc_disconnect(struct obd_export *exp)
4204 struct obd_device *obd = class_exp2obd(exp);
4205 struct llog_ctxt *ctxt;
4208 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4210 if (obd->u.cli.cl_conn_count == 1) {
4211 /* Flush any remaining cancel messages out to the
4213 llog_sync(ctxt, exp);
4215 llog_ctxt_put(ctxt);
4217 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4221 rc = client_disconnect_export(exp);
4223 * Initially we put del_shrink_grant before disconnect_export, but it
4224 * causes the following problem if setup (connect) and cleanup
4225 * (disconnect) are tangled together.
4226 * connect p1 disconnect p2
4227 * ptlrpc_connect_import
4228 * ............... class_manual_cleanup
4231 * ptlrpc_connect_interrupt
4233 * add this client to shrink list
4235 * Bang! pinger trigger the shrink.
4236 * So the osc should be disconnected from the shrink list, after we
4237 * are sure the import has been destroyed. BUG18662
4239 if (obd->u.cli.cl_import == NULL)
4240 osc_del_shrink_grant(&obd->u.cli);
4244 static int osc_import_event(struct obd_device *obd,
4245 struct obd_import *imp,
4246 enum obd_import_event event)
4248 struct client_obd *cli;
4252 LASSERT(imp->imp_obd == obd);
4255 case IMP_EVENT_DISCON: {
4256 /* Only do this on the MDS OSC's */
4257 if (imp->imp_server_timeout) {
4258 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4260 cfs_spin_lock(&oscc->oscc_lock);
4261 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4262 cfs_spin_unlock(&oscc->oscc_lock);
4265 client_obd_list_lock(&cli->cl_loi_list_lock);
4266 cli->cl_avail_grant = 0;
4267 cli->cl_lost_grant = 0;
4268 client_obd_list_unlock(&cli->cl_loi_list_lock);
4271 case IMP_EVENT_INACTIVE: {
4272 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4275 case IMP_EVENT_INVALIDATE: {
4276 struct ldlm_namespace *ns = obd->obd_namespace;
4280 env = cl_env_get(&refcheck);
4284 client_obd_list_lock(&cli->cl_loi_list_lock);
4285 /* all pages go to failing rpcs due to the invalid
4287 osc_check_rpcs(env, cli);
4288 client_obd_list_unlock(&cli->cl_loi_list_lock);
4290 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4291 cl_env_put(env, &refcheck);
4296 case IMP_EVENT_ACTIVE: {
4297 /* Only do this on the MDS OSC's */
4298 if (imp->imp_server_timeout) {
4299 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4301 cfs_spin_lock(&oscc->oscc_lock);
4302 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4303 cfs_spin_unlock(&oscc->oscc_lock);
4305 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4308 case IMP_EVENT_OCD: {
4309 struct obd_connect_data *ocd = &imp->imp_connect_data;
4311 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4312 osc_init_grant(&obd->u.cli, ocd);
4315 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4316 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4318 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4322 CERROR("Unknown import event %d\n", event);
4328 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4334 rc = ptlrpcd_addref();
4338 rc = client_obd_setup(obd, lcfg);
4342 struct lprocfs_static_vars lvars = { 0 };
4343 struct client_obd *cli = &obd->u.cli;
4345 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4346 lprocfs_osc_init_vars(&lvars);
4347 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4348 lproc_osc_attach_seqstat(obd);
4349 sptlrpc_lprocfs_cliobd_attach(obd);
4350 ptlrpc_lprocfs_register_obd(obd);
4354 /* We need to allocate a few requests more, because
4355 brw_interpret tries to create new requests before freeing
4356 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4357 reserved, but I afraid that might be too much wasted RAM
4358 in fact, so 2 is just my guess and still should work. */
4359 cli->cl_import->imp_rq_pool =
4360 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4362 ptlrpc_add_rqs_to_pool);
4364 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4365 cfs_sema_init(&cli->cl_grant_sem, 1);
4371 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4377 case OBD_CLEANUP_EARLY: {
4378 struct obd_import *imp;
4379 imp = obd->u.cli.cl_import;
4380 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4381 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4382 ptlrpc_deactivate_import(imp);
4383 cfs_spin_lock(&imp->imp_lock);
4384 imp->imp_pingable = 0;
4385 cfs_spin_unlock(&imp->imp_lock);
4388 case OBD_CLEANUP_EXPORTS: {
4389 /* If we set up but never connected, the
4390 client import will not have been cleaned. */
4391 if (obd->u.cli.cl_import) {
4392 struct obd_import *imp;
4393 cfs_down_write(&obd->u.cli.cl_sem);
4394 imp = obd->u.cli.cl_import;
4395 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4397 ptlrpc_invalidate_import(imp);
4398 if (imp->imp_rq_pool) {
4399 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4400 imp->imp_rq_pool = NULL;
4402 class_destroy_import(imp);
4403 cfs_up_write(&obd->u.cli.cl_sem);
4404 obd->u.cli.cl_import = NULL;
4406 rc = obd_llog_finish(obd, 0);
4408 CERROR("failed to cleanup llogging subsystems\n");
4415 int osc_cleanup(struct obd_device *obd)
4420 ptlrpc_lprocfs_unregister_obd(obd);
4421 lprocfs_obd_cleanup(obd);
4423 /* free memory of osc quota cache */
4424 lquota_cleanup(quota_interface, obd);
4426 rc = client_obd_cleanup(obd);
4432 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4434 struct lprocfs_static_vars lvars = { 0 };
4437 lprocfs_osc_init_vars(&lvars);
4439 switch (lcfg->lcfg_command) {
4441 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4451 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4453 return osc_process_config_base(obd, buf);
4456 struct obd_ops osc_obd_ops = {
4457 .o_owner = THIS_MODULE,
4458 .o_setup = osc_setup,
4459 .o_precleanup = osc_precleanup,
4460 .o_cleanup = osc_cleanup,
4461 .o_add_conn = client_import_add_conn,
4462 .o_del_conn = client_import_del_conn,
4463 .o_connect = client_connect_import,
4464 .o_reconnect = osc_reconnect,
4465 .o_disconnect = osc_disconnect,
4466 .o_statfs = osc_statfs,
4467 .o_statfs_async = osc_statfs_async,
4468 .o_packmd = osc_packmd,
4469 .o_unpackmd = osc_unpackmd,
4470 .o_precreate = osc_precreate,
4471 .o_create = osc_create,
4472 .o_create_async = osc_create_async,
4473 .o_destroy = osc_destroy,
4474 .o_getattr = osc_getattr,
4475 .o_getattr_async = osc_getattr_async,
4476 .o_setattr = osc_setattr,
4477 .o_setattr_async = osc_setattr_async,
4479 .o_punch = osc_punch,
4481 .o_enqueue = osc_enqueue,
4482 .o_change_cbdata = osc_change_cbdata,
4483 .o_find_cbdata = osc_find_cbdata,
4484 .o_cancel = osc_cancel,
4485 .o_cancel_unused = osc_cancel_unused,
4486 .o_iocontrol = osc_iocontrol,
4487 .o_get_info = osc_get_info,
4488 .o_set_info_async = osc_set_info_async,
4489 .o_import_event = osc_import_event,
4490 .o_llog_init = osc_llog_init,
4491 .o_llog_finish = osc_llog_finish,
4492 .o_process_config = osc_process_config,
4495 extern struct lu_kmem_descr osc_caches[];
4496 extern cfs_spinlock_t osc_ast_guard;
4497 extern cfs_lock_class_key_t osc_ast_guard_class;
4499 int __init osc_init(void)
4501 struct lprocfs_static_vars lvars = { 0 };
4505 /* print an address of _any_ initialized kernel symbol from this
4506 * module, to allow debugging with gdb that doesn't support data
4507 * symbols from modules.*/
4508 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4510 rc = lu_kmem_init(osc_caches);
4512 lprocfs_osc_init_vars(&lvars);
4514 cfs_request_module("lquota");
4515 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4516 lquota_init(quota_interface);
4517 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4519 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4520 LUSTRE_OSC_NAME, &osc_device_type);
4522 if (quota_interface)
4523 PORTAL_SYMBOL_PUT(osc_quota_interface);
4524 lu_kmem_fini(osc_caches);
4528 cfs_spin_lock_init(&osc_ast_guard);
4529 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4531 osc_mds_ost_orig_logops = llog_lvfs_ops;
4532 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4533 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4534 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4535 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4541 static void /*__exit*/ osc_exit(void)
4543 lu_device_type_fini(&osc_device_type);
4545 lquota_exit(quota_interface);
4546 if (quota_interface)
4547 PORTAL_SYMBOL_PUT(osc_quota_interface);
4549 class_unregister_type(LUSTRE_OSC_NAME);
4550 lu_kmem_fini(osc_caches);
4553 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4554 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4555 MODULE_LICENSE("GPL");
4557 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);