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 2008 Sun Microsystems, Inc. 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_MDS_GROUP(lsm->lsm_object_gr);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
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_gr = le64_to_cpu (lmm->lmm_object_gr);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT_MDS_GROUP((*lsmp)->lsm_object_gr);
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 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 = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
214 lustre_swab_ost_body);
216 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
217 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
219 /* This should really be sent by the OST */
220 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
221 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
223 CDEBUG(D_INFO, "can't unpack ost_body\n");
225 aa->aa_oi->oi_oa->o_valid = 0;
228 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
232 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
233 struct ptlrpc_request_set *set)
235 struct ptlrpc_request *req;
236 struct osc_async_args *aa;
240 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
244 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
245 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
247 ptlrpc_request_free(req);
251 osc_pack_req_body(req, oinfo);
253 ptlrpc_request_set_replen(req);
254 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
256 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
257 aa = ptlrpc_req_async_args(req);
260 ptlrpc_set_add_req(set, req);
264 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
266 struct ptlrpc_request *req;
267 struct ost_body *body;
271 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
275 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
278 ptlrpc_request_free(req);
282 osc_pack_req_body(req, oinfo);
284 ptlrpc_request_set_replen(req);
286 rc = ptlrpc_queue_wait(req);
290 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
292 GOTO(out, rc = -EPROTO);
294 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
295 *oinfo->oi_oa = body->oa;
297 /* This should really be sent by the OST */
298 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
299 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
303 ptlrpc_req_finished(req);
307 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
308 struct obd_trans_info *oti)
310 struct ptlrpc_request *req;
311 struct ost_body *body;
315 LASSERTF(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
316 CHECK_MDS_GROUP(oinfo->oi_oa->o_gr),
317 "oinfo->oi_oa->o_valid="LPU64" oinfo->oi_oa->o_gr="LPU64"\n",
318 oinfo->oi_oa->o_valid, oinfo->oi_oa->o_gr);
320 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
324 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
325 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
327 ptlrpc_request_free(req);
331 osc_pack_req_body(req, oinfo);
333 ptlrpc_request_set_replen(req);
335 rc = ptlrpc_queue_wait(req);
339 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
341 GOTO(out, rc = -EPROTO);
343 *oinfo->oi_oa = body->oa;
347 ptlrpc_req_finished(req);
351 static int osc_setattr_interpret(const struct lu_env *env,
352 struct ptlrpc_request *req,
353 struct osc_async_args *aa, int rc)
355 struct ost_body *body;
361 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
363 GOTO(out, rc = -EPROTO);
365 *aa->aa_oi->oi_oa = body->oa;
367 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
371 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
372 struct obd_trans_info *oti,
373 struct ptlrpc_request_set *rqset)
375 struct ptlrpc_request *req;
376 struct osc_async_args *aa;
380 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
384 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
385 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
387 ptlrpc_request_free(req);
391 osc_pack_req_body(req, oinfo);
393 ptlrpc_request_set_replen(req);
395 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
397 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
400 /* do mds to ost setattr asynchronously */
402 /* Do not wait for response. */
403 ptlrpcd_add_req(req, PSCOPE_OTHER);
405 req->rq_interpret_reply =
406 (ptlrpc_interpterer_t)osc_setattr_interpret;
408 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
409 aa = ptlrpc_req_async_args(req);
412 ptlrpc_set_add_req(rqset, req);
418 int osc_real_create(struct obd_export *exp, struct obdo *oa,
419 struct lov_stripe_md **ea, struct obd_trans_info *oti)
421 struct ptlrpc_request *req;
422 struct ost_body *body;
423 struct lov_stripe_md *lsm;
432 rc = obd_alloc_memmd(exp, &lsm);
437 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
439 GOTO(out, rc = -ENOMEM);
441 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
443 ptlrpc_request_free(req);
447 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
451 ptlrpc_request_set_replen(req);
453 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
454 oa->o_flags == OBD_FL_DELORPHAN) {
456 "delorphan from OST integration");
457 /* Don't resend the delorphan req */
458 req->rq_no_resend = req->rq_no_delay = 1;
461 rc = ptlrpc_queue_wait(req);
465 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
467 GOTO(out_req, rc = -EPROTO);
471 /* This should really be sent by the OST */
472 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
473 oa->o_valid |= OBD_MD_FLBLKSZ;
475 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
476 * have valid lsm_oinfo data structs, so don't go touching that.
477 * This needs to be fixed in a big way.
479 lsm->lsm_object_id = oa->o_id;
480 lsm->lsm_object_gr = oa->o_gr;
484 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
486 if (oa->o_valid & OBD_MD_FLCOOKIE) {
487 if (!oti->oti_logcookies)
488 oti_alloc_cookies(oti, 1);
489 *oti->oti_logcookies = oa->o_lcookie;
493 CDEBUG(D_HA, "transno: "LPD64"\n",
494 lustre_msg_get_transno(req->rq_repmsg));
496 ptlrpc_req_finished(req);
499 obd_free_memmd(exp, &lsm);
503 static int osc_punch_interpret(const struct lu_env *env,
504 struct ptlrpc_request *req,
505 struct osc_punch_args *aa, int rc)
507 struct ost_body *body;
513 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
515 GOTO(out, rc = -EPROTO);
517 *aa->pa_oa = body->oa;
519 rc = aa->pa_upcall(aa->pa_cookie, rc);
523 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
524 struct obd_capa *capa,
525 obd_enqueue_update_f upcall, void *cookie,
526 struct ptlrpc_request_set *rqset)
528 struct ptlrpc_request *req;
529 struct osc_punch_args *aa;
530 struct ost_body *body;
534 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
538 osc_set_capa_size(req, &RMF_CAPA1, capa);
539 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
541 ptlrpc_request_free(req);
544 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
545 ptlrpc_at_set_req_timeout(req);
547 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
550 osc_pack_capa(req, body, capa);
552 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_punch_interpret;
556 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
557 aa = ptlrpc_req_async_args(req);
559 aa->pa_upcall = upcall;
560 aa->pa_cookie = cookie;
561 if (rqset == PTLRPCD_SET)
562 ptlrpcd_add_req(req, PSCOPE_OTHER);
564 ptlrpc_set_add_req(rqset, req);
569 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
570 struct obd_trans_info *oti,
571 struct ptlrpc_request_set *rqset)
573 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
574 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
575 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
576 return osc_punch_base(exp, oinfo->oi_oa, oinfo->oi_capa,
577 oinfo->oi_cb_up, oinfo, rqset);
580 static int osc_sync(struct obd_export *exp, struct obdo *oa,
581 struct lov_stripe_md *md, obd_size start, obd_size end,
584 struct ptlrpc_request *req;
585 struct ost_body *body;
590 CDEBUG(D_INFO, "oa NULL\n");
594 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
598 osc_set_capa_size(req, &RMF_CAPA1, capa);
599 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
601 ptlrpc_request_free(req);
605 /* overload the size and blocks fields in the oa with start/end */
606 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
609 body->oa.o_size = start;
610 body->oa.o_blocks = end;
611 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
612 osc_pack_capa(req, body, capa);
614 ptlrpc_request_set_replen(req);
616 rc = ptlrpc_queue_wait(req);
620 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
622 GOTO(out, rc = -EPROTO);
628 ptlrpc_req_finished(req);
632 /* Find and cancel locally locks matched by @mode in the resource found by
633 * @objid. Found locks are added into @cancel list. Returns the amount of
634 * locks added to @cancels list. */
635 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
636 struct list_head *cancels, ldlm_mode_t mode,
639 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
640 struct ldlm_res_id res_id;
641 struct ldlm_resource *res;
645 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
646 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
650 LDLM_RESOURCE_ADDREF(res);
651 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
652 lock_flags, 0, NULL);
653 LDLM_RESOURCE_DELREF(res);
654 ldlm_resource_putref(res);
658 static int osc_destroy_interpret(const struct lu_env *env,
659 struct ptlrpc_request *req, void *data,
662 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
664 atomic_dec(&cli->cl_destroy_in_flight);
665 cfs_waitq_signal(&cli->cl_destroy_waitq);
669 static int osc_can_send_destroy(struct client_obd *cli)
671 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
672 cli->cl_max_rpcs_in_flight) {
673 /* The destroy request can be sent */
676 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
677 cli->cl_max_rpcs_in_flight) {
679 * The counter has been modified between the two atomic
682 cfs_waitq_signal(&cli->cl_destroy_waitq);
687 /* Destroy requests can be async always on the client, and we don't even really
688 * care about the return code since the client cannot do anything at all about
690 * When the MDS is unlinking a filename, it saves the file objects into a
691 * recovery llog, and these object records are cancelled when the OST reports
692 * they were destroyed and sync'd to disk (i.e. transaction committed).
693 * If the client dies, or the OST is down when the object should be destroyed,
694 * the records are not cancelled, and when the OST reconnects to the MDS next,
695 * it will retrieve the llog unlink logs and then sends the log cancellation
696 * cookies to the MDS after committing destroy transactions. */
697 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
698 struct lov_stripe_md *ea, struct obd_trans_info *oti,
699 struct obd_export *md_export, void *capa)
701 struct client_obd *cli = &exp->exp_obd->u.cli;
702 struct ptlrpc_request *req;
703 struct ost_body *body;
704 CFS_LIST_HEAD(cancels);
709 CDEBUG(D_INFO, "oa NULL\n");
713 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
714 LDLM_FL_DISCARD_DATA);
716 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
718 ldlm_lock_list_put(&cancels, l_bl_ast, count);
722 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
723 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
726 ptlrpc_request_free(req);
730 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
731 ptlrpc_at_set_req_timeout(req);
733 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
734 oa->o_lcookie = *oti->oti_logcookies;
735 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
739 osc_pack_capa(req, body, (struct obd_capa *)capa);
740 ptlrpc_request_set_replen(req);
742 /* don't throttle destroy RPCs for the MDT */
743 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
744 req->rq_interpret_reply = osc_destroy_interpret;
745 if (!osc_can_send_destroy(cli)) {
746 struct l_wait_info lwi = { 0 };
749 * Wait until the number of on-going destroy RPCs drops
750 * under max_rpc_in_flight
752 l_wait_event_exclusive(cli->cl_destroy_waitq,
753 osc_can_send_destroy(cli), &lwi);
757 /* Do not wait for response */
758 ptlrpcd_add_req(req, PSCOPE_OTHER);
762 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
765 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
767 LASSERT(!(oa->o_valid & bits));
770 client_obd_list_lock(&cli->cl_loi_list_lock);
771 oa->o_dirty = cli->cl_dirty;
772 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
773 CERROR("dirty %lu - %lu > dirty_max %lu\n",
774 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
776 } else if (atomic_read(&obd_dirty_pages) -
777 atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages){
778 CERROR("dirty %d - %d > system dirty_max %d\n",
779 atomic_read(&obd_dirty_pages),
780 atomic_read(&obd_dirty_transit_pages),
781 obd_max_dirty_pages);
783 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
784 CERROR("dirty %lu - dirty_max %lu too big???\n",
785 cli->cl_dirty, cli->cl_dirty_max);
788 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
789 (cli->cl_max_rpcs_in_flight + 1);
790 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
792 oa->o_grant = cli->cl_avail_grant;
793 oa->o_dropped = cli->cl_lost_grant;
794 cli->cl_lost_grant = 0;
795 client_obd_list_unlock(&cli->cl_loi_list_lock);
796 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
797 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
800 /* caller must hold loi_list_lock */
801 static void osc_consume_write_grant(struct client_obd *cli,
802 struct brw_page *pga)
804 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
805 atomic_inc(&obd_dirty_pages);
806 cli->cl_dirty += CFS_PAGE_SIZE;
807 cli->cl_avail_grant -= CFS_PAGE_SIZE;
808 pga->flag |= OBD_BRW_FROM_GRANT;
809 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
810 CFS_PAGE_SIZE, pga, pga->pg);
811 LASSERT(cli->cl_avail_grant >= 0);
814 /* the companion to osc_consume_write_grant, called when a brw has completed.
815 * must be called with the loi lock held. */
816 static void osc_release_write_grant(struct client_obd *cli,
817 struct brw_page *pga, int sent)
819 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
822 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
827 pga->flag &= ~OBD_BRW_FROM_GRANT;
828 atomic_dec(&obd_dirty_pages);
829 cli->cl_dirty -= CFS_PAGE_SIZE;
830 if (pga->flag & OBD_BRW_NOCACHE) {
831 pga->flag &= ~OBD_BRW_NOCACHE;
832 atomic_dec(&obd_dirty_transit_pages);
833 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
836 cli->cl_lost_grant += CFS_PAGE_SIZE;
837 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
838 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
839 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
840 /* For short writes we shouldn't count parts of pages that
841 * span a whole block on the OST side, or our accounting goes
842 * wrong. Should match the code in filter_grant_check. */
843 int offset = pga->off & ~CFS_PAGE_MASK;
844 int count = pga->count + (offset & (blocksize - 1));
845 int end = (offset + pga->count) & (blocksize - 1);
847 count += blocksize - end;
849 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
850 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
851 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
852 cli->cl_avail_grant, cli->cl_dirty);
858 static unsigned long rpcs_in_flight(struct client_obd *cli)
860 return cli->cl_r_in_flight + cli->cl_w_in_flight;
863 /* caller must hold loi_list_lock */
864 void osc_wake_cache_waiters(struct client_obd *cli)
866 struct list_head *l, *tmp;
867 struct osc_cache_waiter *ocw;
870 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
871 /* if we can't dirty more, we must wait until some is written */
872 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
873 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
874 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
875 "osc max %ld, sys max %d\n", cli->cl_dirty,
876 cli->cl_dirty_max, obd_max_dirty_pages);
880 /* if still dirty cache but no grant wait for pending RPCs that
881 * may yet return us some grant before doing sync writes */
882 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
883 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
884 cli->cl_w_in_flight);
888 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
889 list_del_init(&ocw->ocw_entry);
890 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
891 /* no more RPCs in flight to return grant, do sync IO */
892 ocw->ocw_rc = -EDQUOT;
893 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
895 osc_consume_write_grant(cli,
896 &ocw->ocw_oap->oap_brw_page);
899 cfs_waitq_signal(&ocw->ocw_waitq);
905 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
907 client_obd_list_lock(&cli->cl_loi_list_lock);
908 cli->cl_avail_grant = ocd->ocd_grant;
909 client_obd_list_unlock(&cli->cl_loi_list_lock);
911 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
912 cli->cl_avail_grant, cli->cl_lost_grant);
913 LASSERT(cli->cl_avail_grant >= 0);
916 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
918 client_obd_list_lock(&cli->cl_loi_list_lock);
919 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
920 if (body->oa.o_valid & OBD_MD_FLGRANT)
921 cli->cl_avail_grant += body->oa.o_grant;
922 /* waiters are woken in brw_interpret */
923 client_obd_list_unlock(&cli->cl_loi_list_lock);
926 /* We assume that the reason this OSC got a short read is because it read
927 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
928 * via the LOV, and it _knows_ it's reading inside the file, it's just that
929 * this stripe never got written at or beyond this stripe offset yet. */
930 static void handle_short_read(int nob_read, obd_count page_count,
931 struct brw_page **pga)
936 /* skip bytes read OK */
937 while (nob_read > 0) {
938 LASSERT (page_count > 0);
940 if (pga[i]->count > nob_read) {
941 /* EOF inside this page */
942 ptr = cfs_kmap(pga[i]->pg) +
943 (pga[i]->off & ~CFS_PAGE_MASK);
944 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
945 cfs_kunmap(pga[i]->pg);
951 nob_read -= pga[i]->count;
956 /* zero remaining pages */
957 while (page_count-- > 0) {
958 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
959 memset(ptr, 0, pga[i]->count);
960 cfs_kunmap(pga[i]->pg);
965 static int check_write_rcs(struct ptlrpc_request *req,
966 int requested_nob, int niocount,
967 obd_count page_count, struct brw_page **pga)
971 /* return error if any niobuf was in error */
972 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
973 sizeof(*remote_rcs) * niocount, NULL);
974 if (remote_rcs == NULL) {
975 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
978 if (lustre_msg_swabbed(req->rq_repmsg))
979 for (i = 0; i < niocount; i++)
980 __swab32s(&remote_rcs[i]);
982 for (i = 0; i < niocount; i++) {
983 if (remote_rcs[i] < 0)
984 return(remote_rcs[i]);
986 if (remote_rcs[i] != 0) {
987 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
988 i, remote_rcs[i], req);
993 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
994 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
995 req->rq_bulk->bd_nob_transferred, requested_nob);
1002 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1004 if (p1->flag != p2->flag) {
1005 unsigned mask = ~(OBD_BRW_FROM_GRANT|OBD_BRW_NOCACHE);
1007 /* warn if we try to combine flags that we don't know to be
1008 * safe to combine */
1009 if ((p1->flag & mask) != (p2->flag & mask))
1010 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1011 "same brw?\n", p1->flag, p2->flag);
1015 return (p1->off + p1->count == p2->off);
1018 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1019 struct brw_page **pga, int opc,
1020 cksum_type_t cksum_type)
1025 LASSERT (pg_count > 0);
1026 cksum = init_checksum(cksum_type);
1027 while (nob > 0 && pg_count > 0) {
1028 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1029 int off = pga[i]->off & ~CFS_PAGE_MASK;
1030 int count = pga[i]->count > nob ? nob : pga[i]->count;
1032 /* corrupt the data before we compute the checksum, to
1033 * simulate an OST->client data error */
1034 if (i == 0 && opc == OST_READ &&
1035 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1036 memcpy(ptr + off, "bad1", min(4, nob));
1037 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1038 cfs_kunmap(pga[i]->pg);
1039 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1042 nob -= pga[i]->count;
1046 /* For sending we only compute the wrong checksum instead
1047 * of corrupting the data so it is still correct on a redo */
1048 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1054 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1055 struct lov_stripe_md *lsm, obd_count page_count,
1056 struct brw_page **pga,
1057 struct ptlrpc_request **reqp,
1058 struct obd_capa *ocapa, int reserve)
1060 struct ptlrpc_request *req;
1061 struct ptlrpc_bulk_desc *desc;
1062 struct ost_body *body;
1063 struct obd_ioobj *ioobj;
1064 struct niobuf_remote *niobuf;
1065 int niocount, i, requested_nob, opc, rc;
1066 struct osc_brw_async_args *aa;
1067 struct req_capsule *pill;
1068 struct brw_page *pg_prev;
1071 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1072 RETURN(-ENOMEM); /* Recoverable */
1073 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1074 RETURN(-EINVAL); /* Fatal */
1076 if ((cmd & OBD_BRW_WRITE) != 0) {
1078 req = ptlrpc_request_alloc_pool(cli->cl_import,
1079 cli->cl_import->imp_rq_pool,
1083 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1088 for (niocount = i = 1; i < page_count; i++) {
1089 if (!can_merge_pages(pga[i - 1], pga[i]))
1093 pill = &req->rq_pill;
1094 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1095 niocount * sizeof(*niobuf));
1096 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1098 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1100 ptlrpc_request_free(req);
1103 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1104 ptlrpc_at_set_req_timeout(req);
1106 if (opc == OST_WRITE)
1107 desc = ptlrpc_prep_bulk_imp(req, page_count,
1108 BULK_GET_SOURCE, OST_BULK_PORTAL);
1110 desc = ptlrpc_prep_bulk_imp(req, page_count,
1111 BULK_PUT_SINK, OST_BULK_PORTAL);
1114 GOTO(out, rc = -ENOMEM);
1115 /* NB request now owns desc and will free it when it gets freed */
1117 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1118 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1119 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1120 LASSERT(body && ioobj && niobuf);
1124 obdo_to_ioobj(oa, ioobj);
1125 ioobj->ioo_bufcnt = niocount;
1126 osc_pack_capa(req, body, ocapa);
1127 LASSERT (page_count > 0);
1129 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1130 struct brw_page *pg = pga[i];
1132 LASSERT(pg->count > 0);
1133 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1134 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1135 pg->off, pg->count);
1137 LASSERTF(i == 0 || pg->off > pg_prev->off,
1138 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1139 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1141 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1142 pg_prev->pg, page_private(pg_prev->pg),
1143 pg_prev->pg->index, pg_prev->off);
1145 LASSERTF(i == 0 || pg->off > pg_prev->off,
1146 "i %d p_c %u\n", i, page_count);
1148 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1149 (pg->flag & OBD_BRW_SRVLOCK));
1151 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1153 requested_nob += pg->count;
1155 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1157 niobuf->len += pg->count;
1159 niobuf->offset = pg->off;
1160 niobuf->len = pg->count;
1161 niobuf->flags = pg->flag;
1166 LASSERTF((void *)(niobuf - niocount) ==
1167 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1168 niocount * sizeof(*niobuf)),
1169 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1170 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1171 (void *)(niobuf - niocount));
1173 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1175 /* size[REQ_REC_OFF] still sizeof (*body) */
1176 if (opc == OST_WRITE) {
1177 if (unlikely(cli->cl_checksum) &&
1178 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1179 /* store cl_cksum_type in a local variable since
1180 * it can be changed via lprocfs */
1181 cksum_type_t cksum_type = cli->cl_cksum_type;
1183 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1184 oa->o_flags = body->oa.o_flags = 0;
1185 body->oa.o_flags |= cksum_type_pack(cksum_type);
1186 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1187 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1191 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1193 /* save this in 'oa', too, for later checking */
1194 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1195 oa->o_flags |= cksum_type_pack(cksum_type);
1197 /* clear out the checksum flag, in case this is a
1198 * resend but cl_checksum is no longer set. b=11238 */
1199 oa->o_valid &= ~OBD_MD_FLCKSUM;
1201 oa->o_cksum = body->oa.o_cksum;
1202 /* 1 RC per niobuf */
1203 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1204 sizeof(__u32) * niocount);
1206 if (unlikely(cli->cl_checksum) &&
1207 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1208 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1209 body->oa.o_flags = 0;
1210 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1211 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1213 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1214 /* 1 RC for the whole I/O */
1216 ptlrpc_request_set_replen(req);
1218 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1219 aa = ptlrpc_req_async_args(req);
1221 aa->aa_requested_nob = requested_nob;
1222 aa->aa_nio_count = niocount;
1223 aa->aa_page_count = page_count;
1227 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1228 if (ocapa && reserve)
1229 aa->aa_ocapa = capa_get(ocapa);
1235 ptlrpc_req_finished(req);
1239 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1240 __u32 client_cksum, __u32 server_cksum, int nob,
1241 obd_count page_count, struct brw_page **pga,
1242 cksum_type_t client_cksum_type)
1246 cksum_type_t cksum_type;
1248 if (server_cksum == client_cksum) {
1249 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1253 if (oa->o_valid & OBD_MD_FLFLAGS)
1254 cksum_type = cksum_type_unpack(oa->o_flags);
1256 cksum_type = OBD_CKSUM_CRC32;
1258 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1261 if (cksum_type != client_cksum_type)
1262 msg = "the server did not use the checksum type specified in "
1263 "the original request - likely a protocol problem";
1264 else if (new_cksum == server_cksum)
1265 msg = "changed on the client after we checksummed it - "
1266 "likely false positive due to mmap IO (bug 11742)";
1267 else if (new_cksum == client_cksum)
1268 msg = "changed in transit before arrival at OST";
1270 msg = "changed in transit AND doesn't match the original - "
1271 "likely false positive due to mmap IO (bug 11742)";
1273 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1274 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1275 "["LPU64"-"LPU64"]\n",
1276 msg, libcfs_nid2str(peer->nid),
1277 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1278 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1281 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1283 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1284 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1285 "client csum now %x\n", client_cksum, client_cksum_type,
1286 server_cksum, cksum_type, new_cksum);
1290 /* Note rc enters this function as number of bytes transferred */
1291 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1293 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1294 const lnet_process_id_t *peer =
1295 &req->rq_import->imp_connection->c_peer;
1296 struct client_obd *cli = aa->aa_cli;
1297 struct ost_body *body;
1298 __u32 client_cksum = 0;
1301 if (rc < 0 && rc != -EDQUOT)
1304 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1305 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1306 lustre_swab_ost_body);
1308 CDEBUG(D_INFO, "Can't unpack body\n");
1312 /* set/clear over quota flag for a uid/gid */
1313 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1314 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1315 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1316 body->oa.o_gid, body->oa.o_valid,
1322 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1323 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1325 osc_update_grant(cli, body);
1327 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1329 CERROR("Unexpected +ve rc %d\n", rc);
1332 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1334 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1337 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1338 check_write_checksum(&body->oa, peer, client_cksum,
1339 body->oa.o_cksum, aa->aa_requested_nob,
1340 aa->aa_page_count, aa->aa_ppga,
1341 cksum_type_unpack(aa->aa_oa->o_flags)))
1344 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1345 aa->aa_page_count, aa->aa_ppga);
1349 /* The rest of this function executes only for OST_READs */
1351 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1355 if (rc > aa->aa_requested_nob) {
1356 CERROR("Unexpected rc %d (%d requested)\n", rc,
1357 aa->aa_requested_nob);
1361 if (rc != req->rq_bulk->bd_nob_transferred) {
1362 CERROR ("Unexpected rc %d (%d transferred)\n",
1363 rc, req->rq_bulk->bd_nob_transferred);
1367 if (rc < aa->aa_requested_nob)
1368 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1370 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1371 static int cksum_counter;
1372 __u32 server_cksum = body->oa.o_cksum;
1375 cksum_type_t cksum_type;
1377 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1378 cksum_type = cksum_type_unpack(body->oa.o_flags);
1380 cksum_type = OBD_CKSUM_CRC32;
1381 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1382 aa->aa_ppga, OST_READ,
1385 if (peer->nid == req->rq_bulk->bd_sender) {
1389 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1392 if (server_cksum == ~0 && rc > 0) {
1393 CERROR("Protocol error: server %s set the 'checksum' "
1394 "bit, but didn't send a checksum. Not fatal, "
1395 "but please notify on http://bugzilla.lustre.org/\n",
1396 libcfs_nid2str(peer->nid));
1397 } else if (server_cksum != client_cksum) {
1398 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1399 "%s%s%s inum "LPU64"/"LPU64" object "
1400 LPU64"/"LPU64" extent "
1401 "["LPU64"-"LPU64"]\n",
1402 req->rq_import->imp_obd->obd_name,
1403 libcfs_nid2str(peer->nid),
1405 body->oa.o_valid & OBD_MD_FLFID ?
1406 body->oa.o_fid : (__u64)0,
1407 body->oa.o_valid & OBD_MD_FLFID ?
1408 body->oa.o_generation :(__u64)0,
1410 body->oa.o_valid & OBD_MD_FLGROUP ?
1411 body->oa.o_gr : (__u64)0,
1412 aa->aa_ppga[0]->off,
1413 aa->aa_ppga[aa->aa_page_count-1]->off +
1414 aa->aa_ppga[aa->aa_page_count-1]->count -
1416 CERROR("client %x, server %x, cksum_type %x\n",
1417 client_cksum, server_cksum, cksum_type);
1419 aa->aa_oa->o_cksum = client_cksum;
1423 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1426 } else if (unlikely(client_cksum)) {
1427 static int cksum_missed;
1430 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1431 CERROR("Checksum %u requested from %s but not sent\n",
1432 cksum_missed, libcfs_nid2str(peer->nid));
1438 *aa->aa_oa = body->oa;
1443 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1444 struct lov_stripe_md *lsm,
1445 obd_count page_count, struct brw_page **pga,
1446 struct obd_capa *ocapa)
1448 struct ptlrpc_request *req;
1452 struct l_wait_info lwi;
1456 cfs_waitq_init(&waitq);
1459 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1460 page_count, pga, &req, ocapa, 0);
1464 rc = ptlrpc_queue_wait(req);
1466 if (rc == -ETIMEDOUT && req->rq_resend) {
1467 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1468 ptlrpc_req_finished(req);
1472 rc = osc_brw_fini_request(req, rc);
1474 ptlrpc_req_finished(req);
1475 if (osc_recoverable_error(rc)) {
1477 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1478 CERROR("too many resend retries, returning error\n");
1482 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1483 l_wait_event(waitq, 0, &lwi);
1491 int osc_brw_redo_request(struct ptlrpc_request *request,
1492 struct osc_brw_async_args *aa)
1494 struct ptlrpc_request *new_req;
1495 struct ptlrpc_request_set *set = request->rq_set;
1496 struct osc_brw_async_args *new_aa;
1497 struct osc_async_page *oap;
1501 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1502 CERROR("too many resend retries, returning error\n");
1506 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1508 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1509 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1510 aa->aa_cli, aa->aa_oa,
1511 NULL /* lsm unused by osc currently */,
1512 aa->aa_page_count, aa->aa_ppga,
1513 &new_req, aa->aa_ocapa, 0);
1517 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1519 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1520 if (oap->oap_request != NULL) {
1521 LASSERTF(request == oap->oap_request,
1522 "request %p != oap_request %p\n",
1523 request, oap->oap_request);
1524 if (oap->oap_interrupted) {
1525 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1526 ptlrpc_req_finished(new_req);
1531 /* New request takes over pga and oaps from old request.
1532 * Note that copying a list_head doesn't work, need to move it... */
1534 new_req->rq_interpret_reply = request->rq_interpret_reply;
1535 new_req->rq_async_args = request->rq_async_args;
1536 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1538 new_aa = ptlrpc_req_async_args(new_req);
1540 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1541 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1542 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1544 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1545 if (oap->oap_request) {
1546 ptlrpc_req_finished(oap->oap_request);
1547 oap->oap_request = ptlrpc_request_addref(new_req);
1551 new_aa->aa_ocapa = aa->aa_ocapa;
1552 aa->aa_ocapa = NULL;
1554 /* use ptlrpc_set_add_req is safe because interpret functions work
1555 * in check_set context. only one way exist with access to request
1556 * from different thread got -EINTR - this way protected with
1557 * cl_loi_list_lock */
1558 ptlrpc_set_add_req(set, new_req);
1560 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1562 DEBUG_REQ(D_INFO, new_req, "new request");
1567 * ugh, we want disk allocation on the target to happen in offset order. we'll
1568 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1569 * fine for our small page arrays and doesn't require allocation. its an
1570 * insertion sort that swaps elements that are strides apart, shrinking the
1571 * stride down until its '1' and the array is sorted.
1573 static void sort_brw_pages(struct brw_page **array, int num)
1576 struct brw_page *tmp;
1580 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1585 for (i = stride ; i < num ; i++) {
1588 while (j >= stride && array[j - stride]->off > tmp->off) {
1589 array[j] = array[j - stride];
1594 } while (stride > 1);
1597 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1603 LASSERT (pages > 0);
1604 offset = pg[i]->off & ~CFS_PAGE_MASK;
1608 if (pages == 0) /* that's all */
1611 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1612 return count; /* doesn't end on page boundary */
1615 offset = pg[i]->off & ~CFS_PAGE_MASK;
1616 if (offset != 0) /* doesn't start on page boundary */
1623 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1625 struct brw_page **ppga;
1628 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1632 for (i = 0; i < count; i++)
1637 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1639 LASSERT(ppga != NULL);
1640 OBD_FREE(ppga, sizeof(*ppga) * count);
1643 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1644 obd_count page_count, struct brw_page *pga,
1645 struct obd_trans_info *oti)
1647 struct obdo *saved_oa = NULL;
1648 struct brw_page **ppga, **orig;
1649 struct obd_import *imp = class_exp2cliimp(exp);
1650 struct client_obd *cli = &imp->imp_obd->u.cli;
1651 int rc, page_count_orig;
1654 if (cmd & OBD_BRW_CHECK) {
1655 /* The caller just wants to know if there's a chance that this
1656 * I/O can succeed */
1658 if (imp == NULL || imp->imp_invalid)
1663 /* test_brw with a failed create can trip this, maybe others. */
1664 LASSERT(cli->cl_max_pages_per_rpc);
1668 orig = ppga = osc_build_ppga(pga, page_count);
1671 page_count_orig = page_count;
1673 sort_brw_pages(ppga, page_count);
1674 while (page_count) {
1675 obd_count pages_per_brw;
1677 if (page_count > cli->cl_max_pages_per_rpc)
1678 pages_per_brw = cli->cl_max_pages_per_rpc;
1680 pages_per_brw = page_count;
1682 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1684 if (saved_oa != NULL) {
1685 /* restore previously saved oa */
1686 *oinfo->oi_oa = *saved_oa;
1687 } else if (page_count > pages_per_brw) {
1688 /* save a copy of oa (brw will clobber it) */
1689 OBDO_ALLOC(saved_oa);
1690 if (saved_oa == NULL)
1691 GOTO(out, rc = -ENOMEM);
1692 *saved_oa = *oinfo->oi_oa;
1695 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1696 pages_per_brw, ppga, oinfo->oi_capa);
1701 page_count -= pages_per_brw;
1702 ppga += pages_per_brw;
1706 osc_release_ppga(orig, page_count_orig);
1708 if (saved_oa != NULL)
1709 OBDO_FREE(saved_oa);
1714 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1715 * the dirty accounting. Writeback completes or truncate happens before
1716 * writing starts. Must be called with the loi lock held. */
1717 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1720 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1724 /* This maintains the lists of pending pages to read/write for a given object
1725 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1726 * to quickly find objects that are ready to send an RPC. */
1727 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1733 if (lop->lop_num_pending == 0)
1736 /* if we have an invalid import we want to drain the queued pages
1737 * by forcing them through rpcs that immediately fail and complete
1738 * the pages. recovery relies on this to empty the queued pages
1739 * before canceling the locks and evicting down the llite pages */
1740 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1743 /* stream rpcs in queue order as long as as there is an urgent page
1744 * queued. this is our cheap solution for good batching in the case
1745 * where writepage marks some random page in the middle of the file
1746 * as urgent because of, say, memory pressure */
1747 if (!list_empty(&lop->lop_urgent)) {
1748 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1751 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1752 optimal = cli->cl_max_pages_per_rpc;
1753 if (cmd & OBD_BRW_WRITE) {
1754 /* trigger a write rpc stream as long as there are dirtiers
1755 * waiting for space. as they're waiting, they're not going to
1756 * create more pages to coallesce with what's waiting.. */
1757 if (!list_empty(&cli->cl_cache_waiters)) {
1758 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1761 /* +16 to avoid triggering rpcs that would want to include pages
1762 * that are being queued but which can't be made ready until
1763 * the queuer finishes with the page. this is a wart for
1764 * llite::commit_write() */
1767 if (lop->lop_num_pending >= optimal)
1773 static void on_list(struct list_head *item, struct list_head *list,
1776 if (list_empty(item) && should_be_on)
1777 list_add_tail(item, list);
1778 else if (!list_empty(item) && !should_be_on)
1779 list_del_init(item);
1782 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1783 * can find pages to build into rpcs quickly */
1784 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1786 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1787 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1788 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1790 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1791 loi->loi_write_lop.lop_num_pending);
1793 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1794 loi->loi_read_lop.lop_num_pending);
1797 static void lop_update_pending(struct client_obd *cli,
1798 struct loi_oap_pages *lop, int cmd, int delta)
1800 lop->lop_num_pending += delta;
1801 if (cmd & OBD_BRW_WRITE)
1802 cli->cl_pending_w_pages += delta;
1804 cli->cl_pending_r_pages += delta;
1808 * this is called when a sync waiter receives an interruption. Its job is to
1809 * get the caller woken as soon as possible. If its page hasn't been put in an
1810 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1811 * desiring interruption which will forcefully complete the rpc once the rpc
1814 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
1816 struct loi_oap_pages *lop;
1817 struct lov_oinfo *loi;
1821 LASSERT(!oap->oap_interrupted);
1822 oap->oap_interrupted = 1;
1824 /* ok, it's been put in an rpc. only one oap gets a request reference */
1825 if (oap->oap_request != NULL) {
1826 ptlrpc_mark_interrupted(oap->oap_request);
1827 ptlrpcd_wake(oap->oap_request);
1828 ptlrpc_req_finished(oap->oap_request);
1829 oap->oap_request = NULL;
1833 * page completion may be called only if ->cpo_prep() method was
1834 * executed by osc_io_submit(), that also adds page the to pending list
1836 if (!list_empty(&oap->oap_pending_item)) {
1837 list_del_init(&oap->oap_pending_item);
1838 list_del_init(&oap->oap_urgent_item);
1841 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1842 &loi->loi_write_lop : &loi->loi_read_lop;
1843 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1844 loi_list_maint(oap->oap_cli, oap->oap_loi);
1845 rc = oap->oap_caller_ops->ap_completion(env,
1846 oap->oap_caller_data,
1847 oap->oap_cmd, NULL, -EINTR);
1853 /* this is trying to propogate async writeback errors back up to the
1854 * application. As an async write fails we record the error code for later if
1855 * the app does an fsync. As long as errors persist we force future rpcs to be
1856 * sync so that the app can get a sync error and break the cycle of queueing
1857 * pages for which writeback will fail. */
1858 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1865 ar->ar_force_sync = 1;
1866 ar->ar_min_xid = ptlrpc_sample_next_xid();
1871 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1872 ar->ar_force_sync = 0;
1875 void osc_oap_to_pending(struct osc_async_page *oap)
1877 struct loi_oap_pages *lop;
1879 if (oap->oap_cmd & OBD_BRW_WRITE)
1880 lop = &oap->oap_loi->loi_write_lop;
1882 lop = &oap->oap_loi->loi_read_lop;
1884 if (oap->oap_async_flags & ASYNC_URGENT)
1885 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1886 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1887 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1890 /* this must be called holding the loi list lock to give coverage to exit_cache,
1891 * async_flag maintenance, and oap_request */
1892 static void osc_ap_completion(const struct lu_env *env,
1893 struct client_obd *cli, struct obdo *oa,
1894 struct osc_async_page *oap, int sent, int rc)
1899 if (oap->oap_request != NULL) {
1900 xid = ptlrpc_req_xid(oap->oap_request);
1901 ptlrpc_req_finished(oap->oap_request);
1902 oap->oap_request = NULL;
1905 oap->oap_async_flags = 0;
1906 oap->oap_interrupted = 0;
1908 if (oap->oap_cmd & OBD_BRW_WRITE) {
1909 osc_process_ar(&cli->cl_ar, xid, rc);
1910 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
1913 if (rc == 0 && oa != NULL) {
1914 if (oa->o_valid & OBD_MD_FLBLOCKS)
1915 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
1916 if (oa->o_valid & OBD_MD_FLMTIME)
1917 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
1918 if (oa->o_valid & OBD_MD_FLATIME)
1919 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
1920 if (oa->o_valid & OBD_MD_FLCTIME)
1921 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
1924 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
1925 oap->oap_cmd, oa, rc);
1927 /* ll_ap_completion (from llite) drops PG_locked. so, a new
1928 * I/O on the page could start, but OSC calls it under lock
1929 * and thus we can add oap back to pending safely */
1931 /* upper layer wants to leave the page on pending queue */
1932 osc_oap_to_pending(oap);
1934 osc_exit_cache(cli, oap, sent);
1938 static int brw_interpret(const struct lu_env *env,
1939 struct ptlrpc_request *req, void *data, int rc)
1941 struct osc_brw_async_args *aa = data;
1942 struct client_obd *cli;
1946 rc = osc_brw_fini_request(req, rc);
1947 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1948 if (osc_recoverable_error(rc)) {
1949 rc = osc_brw_redo_request(req, aa);
1955 capa_put(aa->aa_ocapa);
1956 aa->aa_ocapa = NULL;
1961 client_obd_list_lock(&cli->cl_loi_list_lock);
1963 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1964 * is called so we know whether to go to sync BRWs or wait for more
1965 * RPCs to complete */
1966 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1967 cli->cl_w_in_flight--;
1969 cli->cl_r_in_flight--;
1971 async = list_empty(&aa->aa_oaps);
1972 if (!async) { /* from osc_send_oap_rpc() */
1973 struct osc_async_page *oap, *tmp;
1974 /* the caller may re-use the oap after the completion call so
1975 * we need to clean it up a little */
1976 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
1977 list_del_init(&oap->oap_rpc_item);
1978 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
1980 OBDO_FREE(aa->aa_oa);
1981 } else { /* from async_internal() */
1983 for (i = 0; i < aa->aa_page_count; i++)
1984 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
1986 osc_wake_cache_waiters(cli);
1987 osc_check_rpcs(env, cli);
1988 client_obd_list_unlock(&cli->cl_loi_list_lock);
1990 cl_req_completion(env, aa->aa_clerq, rc);
1991 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1995 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
1996 struct client_obd *cli,
1997 struct list_head *rpc_list,
1998 int page_count, int cmd)
2000 struct ptlrpc_request *req;
2001 struct brw_page **pga = NULL;
2002 struct osc_brw_async_args *aa;
2003 struct obdo *oa = NULL;
2004 const struct obd_async_page_ops *ops = NULL;
2005 void *caller_data = NULL;
2006 struct osc_async_page *oap;
2007 struct osc_async_page *tmp;
2008 struct ost_body *body;
2009 struct cl_req *clerq = NULL;
2010 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2011 struct ldlm_lock *lock = NULL;
2012 struct cl_req_attr crattr;
2016 LASSERT(!list_empty(rpc_list));
2018 memset(&crattr, 0, sizeof crattr);
2019 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2021 GOTO(out, req = ERR_PTR(-ENOMEM));
2025 GOTO(out, req = ERR_PTR(-ENOMEM));
2028 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2029 struct cl_page *page = osc_oap2cl_page(oap);
2031 ops = oap->oap_caller_ops;
2032 caller_data = oap->oap_caller_data;
2034 clerq = cl_req_alloc(env, page, crt,
2035 1 /* only 1-object rpcs for
2038 GOTO(out, req = (void *)clerq);
2039 lock = oap->oap_ldlm_lock;
2041 pga[i] = &oap->oap_brw_page;
2042 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2043 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2044 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2046 cl_req_page_add(env, clerq, page);
2049 /* always get the data for the obdo for the rpc */
2050 LASSERT(ops != NULL);
2052 crattr.cra_capa = NULL;
2053 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2055 oa->o_handle = lock->l_remote_handle;
2056 oa->o_valid |= OBD_MD_FLHANDLE;
2059 rc = cl_req_prep(env, clerq);
2061 CERROR("cl_req_prep failed: %d\n", rc);
2062 GOTO(out, req = ERR_PTR(rc));
2065 sort_brw_pages(pga, page_count);
2066 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2067 pga, &req, crattr.cra_capa, 1);
2069 CERROR("prep_req failed: %d\n", rc);
2070 GOTO(out, req = ERR_PTR(rc));
2073 /* Need to update the timestamps after the request is built in case
2074 * we race with setattr (locally or in queue at OST). If OST gets
2075 * later setattr before earlier BRW (as determined by the request xid),
2076 * the OST will not use BRW timestamps. Sadly, there is no obvious
2077 * way to do this in a single call. bug 10150 */
2078 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2079 cl_req_attr_set(env, clerq, &crattr,
2080 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2082 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2083 aa = ptlrpc_req_async_args(req);
2084 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2085 list_splice(rpc_list, &aa->aa_oaps);
2086 CFS_INIT_LIST_HEAD(rpc_list);
2087 aa->aa_clerq = clerq;
2089 capa_put(crattr.cra_capa);
2094 OBD_FREE(pga, sizeof(*pga) * page_count);
2095 /* this should happen rarely and is pretty bad, it makes the
2096 * pending list not follow the dirty order */
2097 client_obd_list_lock(&cli->cl_loi_list_lock);
2098 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2099 list_del_init(&oap->oap_rpc_item);
2101 /* queued sync pages can be torn down while the pages
2102 * were between the pending list and the rpc */
2103 if (oap->oap_interrupted) {
2104 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2105 osc_ap_completion(env, cli, NULL, oap, 0,
2109 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2111 if (clerq && !IS_ERR(clerq))
2112 cl_req_completion(env, clerq, PTR_ERR(req));
2118 * prepare pages for ASYNC io and put pages in send queue.
2122 * \param cmd - OBD_BRW_* macroses
2123 * \param lop - pending pages
2125 * \return zero if pages successfully add to send queue.
2126 * \return not zere if error occurring.
2129 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2130 struct lov_oinfo *loi,
2131 int cmd, struct loi_oap_pages *lop)
2133 struct ptlrpc_request *req;
2134 obd_count page_count = 0;
2135 struct osc_async_page *oap = NULL, *tmp;
2136 struct osc_brw_async_args *aa;
2137 const struct obd_async_page_ops *ops;
2138 CFS_LIST_HEAD(rpc_list);
2139 unsigned int ending_offset;
2140 unsigned starting_offset = 0;
2142 struct cl_object *clob = NULL;
2145 /* first we find the pages we're allowed to work with */
2146 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2148 ops = oap->oap_caller_ops;
2150 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2151 "magic 0x%x\n", oap, oap->oap_magic);
2154 /* pin object in memory, so that completion call-backs
2155 * can be safely called under client_obd_list lock. */
2156 clob = osc_oap2cl_page(oap)->cp_obj;
2157 cl_object_get(clob);
2160 if (page_count != 0 &&
2161 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2162 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2163 " oap %p, page %p, srvlock %u\n",
2164 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2167 /* in llite being 'ready' equates to the page being locked
2168 * until completion unlocks it. commit_write submits a page
2169 * as not ready because its unlock will happen unconditionally
2170 * as the call returns. if we race with commit_write giving
2171 * us that page we dont' want to create a hole in the page
2172 * stream, so we stop and leave the rpc to be fired by
2173 * another dirtier or kupdated interval (the not ready page
2174 * will still be on the dirty list). we could call in
2175 * at the end of ll_file_write to process the queue again. */
2176 if (!(oap->oap_async_flags & ASYNC_READY)) {
2177 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2180 CDEBUG(D_INODE, "oap %p page %p returned %d "
2181 "instead of ready\n", oap,
2185 /* llite is telling us that the page is still
2186 * in commit_write and that we should try
2187 * and put it in an rpc again later. we
2188 * break out of the loop so we don't create
2189 * a hole in the sequence of pages in the rpc
2194 /* the io isn't needed.. tell the checks
2195 * below to complete the rpc with EINTR */
2196 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2197 oap->oap_count = -EINTR;
2200 oap->oap_async_flags |= ASYNC_READY;
2203 LASSERTF(0, "oap %p page %p returned %d "
2204 "from make_ready\n", oap,
2212 * Page submitted for IO has to be locked. Either by
2213 * ->ap_make_ready() or by higher layers.
2215 #if defined(__KERNEL__) && defined(__linux__)
2217 struct cl_page *page;
2219 page = osc_oap2cl_page(oap);
2221 if (page->cp_type == CPT_CACHEABLE &&
2222 !(PageLocked(oap->oap_page) &&
2223 (CheckWriteback(oap->oap_page, cmd)))) {
2224 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2226 (long)oap->oap_page->flags,
2227 oap->oap_async_flags);
2232 /* If there is a gap at the start of this page, it can't merge
2233 * with any previous page, so we'll hand the network a
2234 * "fragmented" page array that it can't transfer in 1 RDMA */
2235 if (page_count != 0 && oap->oap_page_off != 0)
2238 /* take the page out of our book-keeping */
2239 list_del_init(&oap->oap_pending_item);
2240 lop_update_pending(cli, lop, cmd, -1);
2241 list_del_init(&oap->oap_urgent_item);
2243 if (page_count == 0)
2244 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2245 (PTLRPC_MAX_BRW_SIZE - 1);
2247 /* ask the caller for the size of the io as the rpc leaves. */
2248 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2250 ops->ap_refresh_count(env, oap->oap_caller_data,
2252 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2254 if (oap->oap_count <= 0) {
2255 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2257 osc_ap_completion(env, cli, NULL,
2258 oap, 0, oap->oap_count);
2262 /* now put the page back in our accounting */
2263 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2264 if (page_count == 0)
2265 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2266 if (++page_count >= cli->cl_max_pages_per_rpc)
2269 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2270 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2271 * have the same alignment as the initial writes that allocated
2272 * extents on the server. */
2273 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2274 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2275 if (ending_offset == 0)
2278 /* If there is a gap at the end of this page, it can't merge
2279 * with any subsequent pages, so we'll hand the network a
2280 * "fragmented" page array that it can't transfer in 1 RDMA */
2281 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2285 osc_wake_cache_waiters(cli);
2287 loi_list_maint(cli, loi);
2289 client_obd_list_unlock(&cli->cl_loi_list_lock);
2292 cl_object_put(env, clob);
2294 if (page_count == 0) {
2295 client_obd_list_lock(&cli->cl_loi_list_lock);
2299 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2301 LASSERT(list_empty(&rpc_list));
2302 loi_list_maint(cli, loi);
2303 RETURN(PTR_ERR(req));
2306 aa = ptlrpc_req_async_args(req);
2308 if (cmd == OBD_BRW_READ) {
2309 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2310 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2311 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2312 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2314 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2315 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2316 cli->cl_w_in_flight);
2317 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2318 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2320 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2322 client_obd_list_lock(&cli->cl_loi_list_lock);
2324 if (cmd == OBD_BRW_READ)
2325 cli->cl_r_in_flight++;
2327 cli->cl_w_in_flight++;
2329 /* queued sync pages can be torn down while the pages
2330 * were between the pending list and the rpc */
2332 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2333 /* only one oap gets a request reference */
2336 if (oap->oap_interrupted && !req->rq_intr) {
2337 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2339 ptlrpc_mark_interrupted(req);
2343 tmp->oap_request = ptlrpc_request_addref(req);
2345 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2346 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2348 req->rq_interpret_reply = brw_interpret;
2349 ptlrpcd_add_req(req, PSCOPE_BRW);
2353 #define LOI_DEBUG(LOI, STR, args...) \
2354 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2355 !list_empty(&(LOI)->loi_cli_item), \
2356 (LOI)->loi_write_lop.lop_num_pending, \
2357 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2358 (LOI)->loi_read_lop.lop_num_pending, \
2359 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2362 /* This is called by osc_check_rpcs() to find which objects have pages that
2363 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2364 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2367 /* first return all objects which we already know to have
2368 * pages ready to be stuffed into rpcs */
2369 if (!list_empty(&cli->cl_loi_ready_list))
2370 RETURN(list_entry(cli->cl_loi_ready_list.next,
2371 struct lov_oinfo, loi_cli_item));
2373 /* then if we have cache waiters, return all objects with queued
2374 * writes. This is especially important when many small files
2375 * have filled up the cache and not been fired into rpcs because
2376 * they don't pass the nr_pending/object threshhold */
2377 if (!list_empty(&cli->cl_cache_waiters) &&
2378 !list_empty(&cli->cl_loi_write_list))
2379 RETURN(list_entry(cli->cl_loi_write_list.next,
2380 struct lov_oinfo, loi_write_item));
2382 /* then return all queued objects when we have an invalid import
2383 * so that they get flushed */
2384 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2385 if (!list_empty(&cli->cl_loi_write_list))
2386 RETURN(list_entry(cli->cl_loi_write_list.next,
2387 struct lov_oinfo, loi_write_item));
2388 if (!list_empty(&cli->cl_loi_read_list))
2389 RETURN(list_entry(cli->cl_loi_read_list.next,
2390 struct lov_oinfo, loi_read_item));
2395 /* called with the loi list lock held */
2396 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2398 struct lov_oinfo *loi;
2399 int rc = 0, race_counter = 0;
2402 while ((loi = osc_next_loi(cli)) != NULL) {
2403 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2405 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2408 /* attempt some read/write balancing by alternating between
2409 * reads and writes in an object. The makes_rpc checks here
2410 * would be redundant if we were getting read/write work items
2411 * instead of objects. we don't want send_oap_rpc to drain a
2412 * partial read pending queue when we're given this object to
2413 * do io on writes while there are cache waiters */
2414 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2415 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2416 &loi->loi_write_lop);
2424 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2425 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2426 &loi->loi_read_lop);
2435 /* attempt some inter-object balancing by issueing rpcs
2436 * for each object in turn */
2437 if (!list_empty(&loi->loi_cli_item))
2438 list_del_init(&loi->loi_cli_item);
2439 if (!list_empty(&loi->loi_write_item))
2440 list_del_init(&loi->loi_write_item);
2441 if (!list_empty(&loi->loi_read_item))
2442 list_del_init(&loi->loi_read_item);
2444 loi_list_maint(cli, loi);
2446 /* send_oap_rpc fails with 0 when make_ready tells it to
2447 * back off. llite's make_ready does this when it tries
2448 * to lock a page queued for write that is already locked.
2449 * we want to try sending rpcs from many objects, but we
2450 * don't want to spin failing with 0. */
2451 if (race_counter == 10)
2457 /* we're trying to queue a page in the osc so we're subject to the
2458 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2459 * If the osc's queued pages are already at that limit, then we want to sleep
2460 * until there is space in the osc's queue for us. We also may be waiting for
2461 * write credits from the OST if there are RPCs in flight that may return some
2462 * before we fall back to sync writes.
2464 * We need this know our allocation was granted in the presence of signals */
2465 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2469 client_obd_list_lock(&cli->cl_loi_list_lock);
2470 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2471 client_obd_list_unlock(&cli->cl_loi_list_lock);
2476 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2479 int osc_enter_cache_try(const struct lu_env *env,
2480 struct client_obd *cli, struct lov_oinfo *loi,
2481 struct osc_async_page *oap, int transient)
2485 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2487 osc_consume_write_grant(cli, &oap->oap_brw_page);
2489 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2490 atomic_inc(&obd_dirty_transit_pages);
2491 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2497 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2498 * grant or cache space. */
2499 static int osc_enter_cache(const struct lu_env *env,
2500 struct client_obd *cli, struct lov_oinfo *loi,
2501 struct osc_async_page *oap)
2503 struct osc_cache_waiter ocw;
2504 struct l_wait_info lwi = { 0 };
2508 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2509 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2510 cli->cl_dirty_max, obd_max_dirty_pages,
2511 cli->cl_lost_grant, cli->cl_avail_grant);
2513 /* force the caller to try sync io. this can jump the list
2514 * of queued writes and create a discontiguous rpc stream */
2515 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2516 loi->loi_ar.ar_force_sync)
2519 /* Hopefully normal case - cache space and write credits available */
2520 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2521 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2522 osc_enter_cache_try(env, cli, loi, oap, 0))
2525 /* Make sure that there are write rpcs in flight to wait for. This
2526 * is a little silly as this object may not have any pending but
2527 * other objects sure might. */
2528 if (cli->cl_w_in_flight) {
2529 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2530 cfs_waitq_init(&ocw.ocw_waitq);
2534 loi_list_maint(cli, loi);
2535 osc_check_rpcs(env, cli);
2536 client_obd_list_unlock(&cli->cl_loi_list_lock);
2538 CDEBUG(D_CACHE, "sleeping for cache space\n");
2539 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2541 client_obd_list_lock(&cli->cl_loi_list_lock);
2542 if (!list_empty(&ocw.ocw_entry)) {
2543 list_del(&ocw.ocw_entry);
2553 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2554 struct lov_oinfo *loi, cfs_page_t *page,
2555 obd_off offset, const struct obd_async_page_ops *ops,
2556 void *data, void **res, int nocache,
2557 struct lustre_handle *lockh)
2559 struct osc_async_page *oap;
2564 return size_round(sizeof(*oap));
2567 oap->oap_magic = OAP_MAGIC;
2568 oap->oap_cli = &exp->exp_obd->u.cli;
2571 oap->oap_caller_ops = ops;
2572 oap->oap_caller_data = data;
2574 oap->oap_page = page;
2575 oap->oap_obj_off = offset;
2576 if (!client_is_remote(exp) &&
2577 cfs_capable(CFS_CAP_SYS_RESOURCE))
2578 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2580 LASSERT(!(offset & ~CFS_PAGE_MASK));
2582 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2583 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2584 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2585 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2587 spin_lock_init(&oap->oap_lock);
2588 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2592 struct osc_async_page *oap_from_cookie(void *cookie)
2594 struct osc_async_page *oap = cookie;
2595 if (oap->oap_magic != OAP_MAGIC)
2596 return ERR_PTR(-EINVAL);
2600 int osc_queue_async_io(const struct lu_env *env,
2601 struct obd_export *exp, struct lov_stripe_md *lsm,
2602 struct lov_oinfo *loi, void *cookie,
2603 int cmd, obd_off off, int count,
2604 obd_flag brw_flags, enum async_flags async_flags)
2606 struct client_obd *cli = &exp->exp_obd->u.cli;
2607 struct osc_async_page *oap;
2611 oap = oap_from_cookie(cookie);
2613 RETURN(PTR_ERR(oap));
2615 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2618 if (!list_empty(&oap->oap_pending_item) ||
2619 !list_empty(&oap->oap_urgent_item) ||
2620 !list_empty(&oap->oap_rpc_item))
2623 /* check if the file's owner/group is over quota */
2624 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2625 struct cl_object *obj;
2626 struct cl_attr attr; /* XXX put attr into thread info */
2628 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2630 cl_object_attr_lock(obj);
2631 rc = cl_object_attr_get(env, obj, &attr);
2632 cl_object_attr_unlock(obj);
2634 if (rc == 0 && lquota_chkdq(quota_interface, cli, attr.cat_uid,
2635 attr.cat_gid) == NO_QUOTA)
2642 loi = lsm->lsm_oinfo[0];
2644 client_obd_list_lock(&cli->cl_loi_list_lock);
2646 LASSERT(off + count <= CFS_PAGE_SIZE);
2648 oap->oap_page_off = off;
2649 oap->oap_count = count;
2650 oap->oap_brw_flags = brw_flags;
2651 oap->oap_async_flags = async_flags;
2653 if (cmd & OBD_BRW_WRITE) {
2654 rc = osc_enter_cache(env, cli, loi, oap);
2656 client_obd_list_unlock(&cli->cl_loi_list_lock);
2661 osc_oap_to_pending(oap);
2662 loi_list_maint(cli, loi);
2664 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2667 osc_check_rpcs(env, cli);
2668 client_obd_list_unlock(&cli->cl_loi_list_lock);
2673 /* aka (~was & now & flag), but this is more clear :) */
2674 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2676 int osc_set_async_flags_base(struct client_obd *cli,
2677 struct lov_oinfo *loi, struct osc_async_page *oap,
2678 obd_flag async_flags)
2680 struct loi_oap_pages *lop;
2683 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2686 if (oap->oap_cmd & OBD_BRW_WRITE) {
2687 lop = &loi->loi_write_lop;
2689 lop = &loi->loi_read_lop;
2692 if (list_empty(&oap->oap_pending_item))
2695 if ((oap->oap_async_flags & async_flags) == async_flags)
2698 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2699 oap->oap_async_flags |= ASYNC_READY;
2701 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2702 if (list_empty(&oap->oap_rpc_item)) {
2703 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2704 loi_list_maint(cli, loi);
2708 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2709 oap->oap_async_flags);
2713 int osc_teardown_async_page(struct obd_export *exp,
2714 struct lov_stripe_md *lsm,
2715 struct lov_oinfo *loi, void *cookie)
2717 struct client_obd *cli = &exp->exp_obd->u.cli;
2718 struct loi_oap_pages *lop;
2719 struct osc_async_page *oap;
2723 oap = oap_from_cookie(cookie);
2725 RETURN(PTR_ERR(oap));
2728 loi = lsm->lsm_oinfo[0];
2730 if (oap->oap_cmd & OBD_BRW_WRITE) {
2731 lop = &loi->loi_write_lop;
2733 lop = &loi->loi_read_lop;
2736 client_obd_list_lock(&cli->cl_loi_list_lock);
2738 if (!list_empty(&oap->oap_rpc_item))
2739 GOTO(out, rc = -EBUSY);
2741 osc_exit_cache(cli, oap, 0);
2742 osc_wake_cache_waiters(cli);
2744 if (!list_empty(&oap->oap_urgent_item)) {
2745 list_del_init(&oap->oap_urgent_item);
2746 oap->oap_async_flags &= ~ASYNC_URGENT;
2748 if (!list_empty(&oap->oap_pending_item)) {
2749 list_del_init(&oap->oap_pending_item);
2750 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2752 loi_list_maint(cli, loi);
2753 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2755 client_obd_list_unlock(&cli->cl_loi_list_lock);
2759 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
2760 struct ldlm_enqueue_info *einfo,
2763 void *data = einfo->ei_cbdata;
2765 LASSERT(lock != NULL);
2766 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2767 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2768 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2769 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2771 lock_res_and_lock(lock);
2772 spin_lock(&osc_ast_guard);
2773 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
2774 lock->l_ast_data = data;
2775 spin_unlock(&osc_ast_guard);
2776 unlock_res_and_lock(lock);
2779 static void osc_set_data_with_check(struct lustre_handle *lockh,
2780 struct ldlm_enqueue_info *einfo,
2783 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2786 osc_set_lock_data_with_check(lock, einfo, flags);
2787 LDLM_LOCK_PUT(lock);
2789 CERROR("lockh %p, data %p - client evicted?\n",
2790 lockh, einfo->ei_cbdata);
2793 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2794 ldlm_iterator_t replace, void *data)
2796 struct ldlm_res_id res_id;
2797 struct obd_device *obd = class_exp2obd(exp);
2799 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
2800 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2804 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2805 obd_enqueue_update_f upcall, void *cookie,
2808 int intent = *flags & LDLM_FL_HAS_INTENT;
2812 /* The request was created before ldlm_cli_enqueue call. */
2813 if (rc == ELDLM_LOCK_ABORTED) {
2814 struct ldlm_reply *rep;
2815 rep = req_capsule_server_get(&req->rq_pill,
2818 LASSERT(rep != NULL);
2819 if (rep->lock_policy_res1)
2820 rc = rep->lock_policy_res1;
2824 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
2825 *flags |= LDLM_FL_LVB_READY;
2826 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2827 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2830 /* Call the update callback. */
2831 rc = (*upcall)(cookie, rc);
2835 static int osc_enqueue_interpret(const struct lu_env *env,
2836 struct ptlrpc_request *req,
2837 struct osc_enqueue_args *aa, int rc)
2839 struct ldlm_lock *lock;
2840 struct lustre_handle handle;
2843 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2844 * might be freed anytime after lock upcall has been called. */
2845 lustre_handle_copy(&handle, aa->oa_lockh);
2846 mode = aa->oa_ei->ei_mode;
2848 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2850 lock = ldlm_handle2lock(&handle);
2852 /* Take an additional reference so that a blocking AST that
2853 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2854 * to arrive after an upcall has been executed by
2855 * osc_enqueue_fini(). */
2856 ldlm_lock_addref(&handle, mode);
2858 /* Complete obtaining the lock procedure. */
2859 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2860 mode, aa->oa_flags, aa->oa_lvb,
2861 sizeof(*aa->oa_lvb), lustre_swab_ost_lvb,
2863 /* Complete osc stuff. */
2864 rc = osc_enqueue_fini(req, aa->oa_lvb,
2865 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
2866 /* Release the lock for async request. */
2867 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2869 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2870 * not already released by
2871 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2873 ldlm_lock_decref(&handle, mode);
2875 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2876 aa->oa_lockh, req, aa);
2877 ldlm_lock_decref(&handle, mode);
2878 LDLM_LOCK_PUT(lock);
2882 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2883 struct lov_oinfo *loi, int flags,
2884 struct ost_lvb *lvb, __u32 mode, int rc)
2886 if (rc == ELDLM_OK) {
2887 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2890 LASSERT(lock != NULL);
2891 loi->loi_lvb = *lvb;
2892 tmp = loi->loi_lvb.lvb_size;
2893 /* Extend KMS up to the end of this lock and no further
2894 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2895 if (tmp > lock->l_policy_data.l_extent.end)
2896 tmp = lock->l_policy_data.l_extent.end + 1;
2897 if (tmp >= loi->loi_kms) {
2898 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2899 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2900 loi_kms_set(loi, tmp);
2902 LDLM_DEBUG(lock, "lock acquired, setting rss="
2903 LPU64"; leaving kms="LPU64", end="LPU64,
2904 loi->loi_lvb.lvb_size, loi->loi_kms,
2905 lock->l_policy_data.l_extent.end);
2907 ldlm_lock_allow_match(lock);
2908 LDLM_LOCK_PUT(lock);
2909 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2910 loi->loi_lvb = *lvb;
2911 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2912 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2916 EXPORT_SYMBOL(osc_update_enqueue);
2918 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2920 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2921 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2922 * other synchronous requests, however keeping some locks and trying to obtain
2923 * others may take a considerable amount of time in a case of ost failure; and
2924 * when other sync requests do not get released lock from a client, the client
2925 * is excluded from the cluster -- such scenarious make the life difficult, so
2926 * release locks just after they are obtained. */
2927 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2928 int *flags, ldlm_policy_data_t *policy,
2929 struct ost_lvb *lvb, int kms_valid,
2930 obd_enqueue_update_f upcall, void *cookie,
2931 struct ldlm_enqueue_info *einfo,
2932 struct lustre_handle *lockh,
2933 struct ptlrpc_request_set *rqset, int async)
2935 struct obd_device *obd = exp->exp_obd;
2936 struct ptlrpc_request *req = NULL;
2937 int intent = *flags & LDLM_FL_HAS_INTENT;
2942 /* Filesystem lock extents are extended to page boundaries so that
2943 * dealing with the page cache is a little smoother. */
2944 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2945 policy->l_extent.end |= ~CFS_PAGE_MASK;
2948 * kms is not valid when either object is completely fresh (so that no
2949 * locks are cached), or object was evicted. In the latter case cached
2950 * lock cannot be used, because it would prime inode state with
2951 * potentially stale LVB.
2956 /* Next, search for already existing extent locks that will cover us */
2957 /* If we're trying to read, we also search for an existing PW lock. The
2958 * VFS and page cache already protect us locally, so lots of readers/
2959 * writers can share a single PW lock.
2961 * There are problems with conversion deadlocks, so instead of
2962 * converting a read lock to a write lock, we'll just enqueue a new
2965 * At some point we should cancel the read lock instead of making them
2966 * send us a blocking callback, but there are problems with canceling
2967 * locks out from other users right now, too. */
2968 mode = einfo->ei_mode;
2969 if (einfo->ei_mode == LCK_PR)
2971 mode = ldlm_lock_match(obd->obd_namespace,
2972 *flags | LDLM_FL_LVB_READY, res_id,
2973 einfo->ei_type, policy, mode, lockh, 0);
2975 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2977 if (matched->l_ast_data == NULL ||
2978 matched->l_ast_data == einfo->ei_cbdata) {
2979 /* addref the lock only if not async requests and PW
2980 * lock is matched whereas we asked for PR. */
2981 if (!rqset && einfo->ei_mode != mode)
2982 ldlm_lock_addref(lockh, LCK_PR);
2983 osc_set_lock_data_with_check(matched, einfo, *flags);
2985 /* I would like to be able to ASSERT here that
2986 * rss <= kms, but I can't, for reasons which
2987 * are explained in lov_enqueue() */
2990 /* We already have a lock, and it's referenced */
2991 (*upcall)(cookie, ELDLM_OK);
2993 /* For async requests, decref the lock. */
2994 if (einfo->ei_mode != mode)
2995 ldlm_lock_decref(lockh, LCK_PW);
2997 ldlm_lock_decref(lockh, einfo->ei_mode);
2998 LDLM_LOCK_PUT(matched);
3001 ldlm_lock_decref(lockh, mode);
3002 LDLM_LOCK_PUT(matched);
3007 CFS_LIST_HEAD(cancels);
3008 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3009 &RQF_LDLM_ENQUEUE_LVB);
3013 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3017 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3019 ptlrpc_request_set_replen(req);
3022 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3023 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3025 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3026 sizeof(*lvb), lustre_swab_ost_lvb, lockh, async);
3029 struct osc_enqueue_args *aa;
3030 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3031 aa = ptlrpc_req_async_args(req);
3034 aa->oa_flags = flags;
3035 aa->oa_upcall = upcall;
3036 aa->oa_cookie = cookie;
3038 aa->oa_lockh = lockh;
3040 req->rq_interpret_reply =
3041 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3042 if (rqset == PTLRPCD_SET)
3043 ptlrpcd_add_req(req, PSCOPE_OTHER);
3045 ptlrpc_set_add_req(rqset, req);
3046 } else if (intent) {
3047 ptlrpc_req_finished(req);
3052 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3054 ptlrpc_req_finished(req);
3059 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3060 struct ldlm_enqueue_info *einfo,
3061 struct ptlrpc_request_set *rqset)
3063 struct ldlm_res_id res_id;
3067 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3068 oinfo->oi_md->lsm_object_gr, &res_id);
3070 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3071 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3072 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3073 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3074 rqset, rqset != NULL);
3078 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3079 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3080 int *flags, void *data, struct lustre_handle *lockh,
3083 struct obd_device *obd = exp->exp_obd;
3084 int lflags = *flags;
3088 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3091 /* Filesystem lock extents are extended to page boundaries so that
3092 * dealing with the page cache is a little smoother */
3093 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3094 policy->l_extent.end |= ~CFS_PAGE_MASK;
3096 /* Next, search for already existing extent locks that will cover us */
3097 /* If we're trying to read, we also search for an existing PW lock. The
3098 * VFS and page cache already protect us locally, so lots of readers/
3099 * writers can share a single PW lock. */
3103 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3104 res_id, type, policy, rc, lockh, unref);
3107 osc_set_data_with_check(lockh, data, lflags);
3108 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3109 ldlm_lock_addref(lockh, LCK_PR);
3110 ldlm_lock_decref(lockh, LCK_PW);
3117 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3121 if (unlikely(mode == LCK_GROUP))
3122 ldlm_lock_decref_and_cancel(lockh, mode);
3124 ldlm_lock_decref(lockh, mode);
3129 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3130 __u32 mode, struct lustre_handle *lockh)
3133 RETURN(osc_cancel_base(lockh, mode));
3136 static int osc_cancel_unused(struct obd_export *exp,
3137 struct lov_stripe_md *lsm, int flags,
3140 struct obd_device *obd = class_exp2obd(exp);
3141 struct ldlm_res_id res_id, *resp = NULL;
3144 resp = osc_build_res_name(lsm->lsm_object_id,
3145 lsm->lsm_object_gr, &res_id);
3148 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3151 static int osc_statfs_interpret(const struct lu_env *env,
3152 struct ptlrpc_request *req,
3153 struct osc_async_args *aa, int rc)
3155 struct obd_statfs *msfs;
3161 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3163 GOTO(out, rc = -EPROTO);
3166 *aa->aa_oi->oi_osfs = *msfs;
3168 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3172 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3173 __u64 max_age, struct ptlrpc_request_set *rqset)
3175 struct ptlrpc_request *req;
3176 struct osc_async_args *aa;
3180 /* We could possibly pass max_age in the request (as an absolute
3181 * timestamp or a "seconds.usec ago") so the target can avoid doing
3182 * extra calls into the filesystem if that isn't necessary (e.g.
3183 * during mount that would help a bit). Having relative timestamps
3184 * is not so great if request processing is slow, while absolute
3185 * timestamps are not ideal because they need time synchronization. */
3186 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3190 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3192 ptlrpc_request_free(req);
3195 ptlrpc_request_set_replen(req);
3196 req->rq_request_portal = OST_CREATE_PORTAL;
3197 ptlrpc_at_set_req_timeout(req);
3199 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3200 /* procfs requests not want stat in wait for avoid deadlock */
3201 req->rq_no_resend = 1;
3202 req->rq_no_delay = 1;
3205 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3206 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3207 aa = ptlrpc_req_async_args(req);
3210 ptlrpc_set_add_req(rqset, req);
3214 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3215 __u64 max_age, __u32 flags)
3217 struct obd_statfs *msfs;
3218 struct ptlrpc_request *req;
3219 struct obd_import *imp = NULL;
3223 /*Since the request might also come from lprocfs, so we need
3224 *sync this with client_disconnect_export Bug15684*/
3225 down_read(&obd->u.cli.cl_sem);
3226 if (obd->u.cli.cl_import)
3227 imp = class_import_get(obd->u.cli.cl_import);
3228 up_read(&obd->u.cli.cl_sem);
3232 /* We could possibly pass max_age in the request (as an absolute
3233 * timestamp or a "seconds.usec ago") so the target can avoid doing
3234 * extra calls into the filesystem if that isn't necessary (e.g.
3235 * during mount that would help a bit). Having relative timestamps
3236 * is not so great if request processing is slow, while absolute
3237 * timestamps are not ideal because they need time synchronization. */
3238 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3240 class_import_put(imp);
3245 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3247 ptlrpc_request_free(req);
3250 ptlrpc_request_set_replen(req);
3251 req->rq_request_portal = OST_CREATE_PORTAL;
3252 ptlrpc_at_set_req_timeout(req);
3254 if (flags & OBD_STATFS_NODELAY) {
3255 /* procfs requests not want stat in wait for avoid deadlock */
3256 req->rq_no_resend = 1;
3257 req->rq_no_delay = 1;
3260 rc = ptlrpc_queue_wait(req);
3264 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3266 GOTO(out, rc = -EPROTO);
3273 ptlrpc_req_finished(req);
3277 /* Retrieve object striping information.
3279 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3280 * the maximum number of OST indices which will fit in the user buffer.
3281 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3283 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3285 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3286 struct lov_user_md_v3 lum, *lumk;
3287 struct lov_user_ost_data_v1 *lmm_objects;
3288 int rc = 0, lum_size;
3294 /* we only need the header part from user space to get lmm_magic and
3295 * lmm_stripe_count, (the header part is common to v1 and v3) */
3296 lum_size = sizeof(struct lov_user_md_v1);
3297 if (copy_from_user(&lum, lump, lum_size))
3300 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3301 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3304 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3305 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3306 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3307 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3309 /* we can use lov_mds_md_size() to compute lum_size
3310 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3311 if (lum.lmm_stripe_count > 0) {
3312 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3313 OBD_ALLOC(lumk, lum_size);
3317 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3318 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3320 lmm_objects = &(lumk->lmm_objects[0]);
3321 lmm_objects->l_object_id = lsm->lsm_object_id;
3323 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3327 lumk->lmm_object_id = lsm->lsm_object_id;
3328 lumk->lmm_object_gr = lsm->lsm_object_gr;
3329 lumk->lmm_stripe_count = 1;
3331 if (copy_to_user(lump, lumk, lum_size))
3335 OBD_FREE(lumk, lum_size);
3341 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3342 void *karg, void *uarg)
3344 struct obd_device *obd = exp->exp_obd;
3345 struct obd_ioctl_data *data = karg;
3349 if (!try_module_get(THIS_MODULE)) {
3350 CERROR("Can't get module. Is it alive?");
3354 case OBD_IOC_LOV_GET_CONFIG: {
3356 struct lov_desc *desc;
3357 struct obd_uuid uuid;
3361 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3362 GOTO(out, err = -EINVAL);
3364 data = (struct obd_ioctl_data *)buf;
3366 if (sizeof(*desc) > data->ioc_inllen1) {
3367 obd_ioctl_freedata(buf, len);
3368 GOTO(out, err = -EINVAL);
3371 if (data->ioc_inllen2 < sizeof(uuid)) {
3372 obd_ioctl_freedata(buf, len);
3373 GOTO(out, err = -EINVAL);
3376 desc = (struct lov_desc *)data->ioc_inlbuf1;
3377 desc->ld_tgt_count = 1;
3378 desc->ld_active_tgt_count = 1;
3379 desc->ld_default_stripe_count = 1;
3380 desc->ld_default_stripe_size = 0;
3381 desc->ld_default_stripe_offset = 0;
3382 desc->ld_pattern = 0;
3383 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3385 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3387 err = copy_to_user((void *)uarg, buf, len);
3390 obd_ioctl_freedata(buf, len);
3393 case LL_IOC_LOV_SETSTRIPE:
3394 err = obd_alloc_memmd(exp, karg);
3398 case LL_IOC_LOV_GETSTRIPE:
3399 err = osc_getstripe(karg, uarg);
3401 case OBD_IOC_CLIENT_RECOVER:
3402 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3407 case IOC_OSC_SET_ACTIVE:
3408 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3411 case OBD_IOC_POLL_QUOTACHECK:
3412 err = lquota_poll_check(quota_interface, exp,
3413 (struct if_quotacheck *)karg);
3415 case OBD_IOC_PING_TARGET:
3416 err = ptlrpc_obd_ping(obd);
3419 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3420 cmd, cfs_curproc_comm());
3421 GOTO(out, err = -ENOTTY);
3424 module_put(THIS_MODULE);
3428 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3429 void *key, __u32 *vallen, void *val,
3430 struct lov_stripe_md *lsm)
3433 if (!vallen || !val)
3436 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3437 __u32 *stripe = val;
3438 *vallen = sizeof(*stripe);
3441 } else if (KEY_IS(KEY_LAST_ID)) {
3442 struct ptlrpc_request *req;
3447 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3448 &RQF_OST_GET_INFO_LAST_ID);
3452 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3453 RCL_CLIENT, keylen);
3454 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3456 ptlrpc_request_free(req);
3460 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3461 memcpy(tmp, key, keylen);
3463 ptlrpc_request_set_replen(req);
3464 rc = ptlrpc_queue_wait(req);
3468 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3470 GOTO(out, rc = -EPROTO);
3472 *((obd_id *)val) = *reply;
3474 ptlrpc_req_finished(req);
3476 } else if (KEY_IS(KEY_FIEMAP)) {
3477 struct ptlrpc_request *req;
3478 struct ll_user_fiemap *reply;
3482 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3483 &RQF_OST_GET_INFO_FIEMAP);
3487 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3488 RCL_CLIENT, keylen);
3489 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3490 RCL_CLIENT, *vallen);
3491 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3492 RCL_SERVER, *vallen);
3494 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3496 ptlrpc_request_free(req);
3500 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3501 memcpy(tmp, key, keylen);
3502 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3503 memcpy(tmp, val, *vallen);
3505 ptlrpc_request_set_replen(req);
3506 rc = ptlrpc_queue_wait(req);
3510 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3512 GOTO(out1, rc = -EPROTO);
3514 memcpy(val, reply, *vallen);
3516 ptlrpc_req_finished(req);
3524 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3525 struct ptlrpc_request *req,
3528 struct llog_ctxt *ctxt;
3529 struct obd_import *imp = req->rq_import;
3535 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3538 rc = llog_initiator_connect(ctxt);
3540 CERROR("cannot establish connection for "
3541 "ctxt %p: %d\n", ctxt, rc);
3544 llog_ctxt_put(ctxt);
3545 spin_lock(&imp->imp_lock);
3546 imp->imp_server_timeout = 1;
3547 imp->imp_pingable = 1;
3548 spin_unlock(&imp->imp_lock);
3549 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3554 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3555 void *key, obd_count vallen, void *val,
3556 struct ptlrpc_request_set *set)
3558 struct ptlrpc_request *req;
3559 struct obd_device *obd = exp->exp_obd;
3560 struct obd_import *imp = class_exp2cliimp(exp);
3565 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3567 if (KEY_IS(KEY_NEXT_ID)) {
3568 if (vallen != sizeof(obd_id))
3572 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3573 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3574 exp->exp_obd->obd_name,
3575 obd->u.cli.cl_oscc.oscc_next_id);
3580 if (KEY_IS(KEY_UNLINKED)) {
3581 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3582 spin_lock(&oscc->oscc_lock);
3583 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3584 spin_unlock(&oscc->oscc_lock);
3588 if (KEY_IS(KEY_INIT_RECOV)) {
3589 if (vallen != sizeof(int))
3591 spin_lock(&imp->imp_lock);
3592 imp->imp_initial_recov = *(int *)val;
3593 spin_unlock(&imp->imp_lock);
3594 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3595 exp->exp_obd->obd_name,
3596 imp->imp_initial_recov);
3600 if (KEY_IS(KEY_CHECKSUM)) {
3601 if (vallen != sizeof(int))
3603 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3607 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3608 sptlrpc_conf_client_adapt(obd);
3612 if (KEY_IS(KEY_FLUSH_CTX)) {
3613 sptlrpc_import_flush_my_ctx(imp);
3620 /* We pass all other commands directly to OST. Since nobody calls osc
3621 methods directly and everybody is supposed to go through LOV, we
3622 assume lov checked invalid values for us.
3623 The only recognised values so far are evict_by_nid and mds_conn.
3624 Even if something bad goes through, we'd get a -EINVAL from OST
3628 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3632 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3633 RCL_CLIENT, keylen);
3634 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3635 RCL_CLIENT, vallen);
3636 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3638 ptlrpc_request_free(req);
3642 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3643 memcpy(tmp, key, keylen);
3644 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3645 memcpy(tmp, val, vallen);
3647 if (KEY_IS(KEY_MDS_CONN)) {
3648 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3650 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3651 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3652 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
3653 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3656 ptlrpc_request_set_replen(req);
3657 ptlrpc_set_add_req(set, req);
3658 ptlrpc_check_set(NULL, set);
3664 static struct llog_operations osc_size_repl_logops = {
3665 lop_cancel: llog_obd_repl_cancel
3668 static struct llog_operations osc_mds_ost_orig_logops;
3669 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3670 struct obd_device *tgt, int count,
3671 struct llog_catid *catid, struct obd_uuid *uuid)
3676 LASSERT(olg == &obd->obd_olg);
3677 spin_lock(&obd->obd_dev_lock);
3678 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3679 osc_mds_ost_orig_logops = llog_lvfs_ops;
3680 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3681 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3682 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3683 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3685 spin_unlock(&obd->obd_dev_lock);
3687 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3688 &catid->lci_logid, &osc_mds_ost_orig_logops);
3690 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3694 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3695 NULL, &osc_size_repl_logops);
3697 struct llog_ctxt *ctxt =
3698 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3701 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3706 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3707 obd->obd_name, tgt->obd_name, count, catid, rc);
3708 CERROR("logid "LPX64":0x%x\n",
3709 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3714 static int osc_llog_finish(struct obd_device *obd, int count)
3716 struct llog_ctxt *ctxt;
3717 int rc = 0, rc2 = 0;
3720 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3722 rc = llog_cleanup(ctxt);
3724 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3726 rc2 = llog_cleanup(ctxt);
3733 static int osc_reconnect(const struct lu_env *env,
3734 struct obd_export *exp, struct obd_device *obd,
3735 struct obd_uuid *cluuid,
3736 struct obd_connect_data *data,
3739 struct client_obd *cli = &obd->u.cli;
3741 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3744 client_obd_list_lock(&cli->cl_loi_list_lock);
3745 data->ocd_grant = cli->cl_avail_grant ?:
3746 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3747 lost_grant = cli->cl_lost_grant;
3748 cli->cl_lost_grant = 0;
3749 client_obd_list_unlock(&cli->cl_loi_list_lock);
3751 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3752 "cl_lost_grant: %ld\n", data->ocd_grant,
3753 cli->cl_avail_grant, lost_grant);
3754 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3755 " ocd_grant: %d\n", data->ocd_connect_flags,
3756 data->ocd_version, data->ocd_grant);
3762 static int osc_disconnect(struct obd_export *exp)
3764 struct obd_device *obd = class_exp2obd(exp);
3765 struct llog_ctxt *ctxt;
3768 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3770 if (obd->u.cli.cl_conn_count == 1) {
3771 /* Flush any remaining cancel messages out to the
3773 llog_sync(ctxt, exp);
3775 llog_ctxt_put(ctxt);
3777 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3781 rc = client_disconnect_export(exp);
3785 static int osc_import_event(struct obd_device *obd,
3786 struct obd_import *imp,
3787 enum obd_import_event event)
3789 struct client_obd *cli;
3793 LASSERT(imp->imp_obd == obd);
3796 case IMP_EVENT_DISCON: {
3797 /* Only do this on the MDS OSC's */
3798 if (imp->imp_server_timeout) {
3799 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3801 spin_lock(&oscc->oscc_lock);
3802 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3803 spin_unlock(&oscc->oscc_lock);
3806 client_obd_list_lock(&cli->cl_loi_list_lock);
3807 cli->cl_avail_grant = 0;
3808 cli->cl_lost_grant = 0;
3809 client_obd_list_unlock(&cli->cl_loi_list_lock);
3812 case IMP_EVENT_INACTIVE: {
3813 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3816 case IMP_EVENT_INVALIDATE: {
3817 struct ldlm_namespace *ns = obd->obd_namespace;
3821 env = cl_env_get(&refcheck);
3825 client_obd_list_lock(&cli->cl_loi_list_lock);
3826 /* all pages go to failing rpcs due to the invalid
3828 osc_check_rpcs(env, cli);
3829 client_obd_list_unlock(&cli->cl_loi_list_lock);
3831 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3832 cl_env_put(env, &refcheck);
3837 case IMP_EVENT_ACTIVE: {
3838 /* Only do this on the MDS OSC's */
3839 if (imp->imp_server_timeout) {
3840 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3842 spin_lock(&oscc->oscc_lock);
3843 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3844 spin_unlock(&oscc->oscc_lock);
3846 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3849 case IMP_EVENT_OCD: {
3850 struct obd_connect_data *ocd = &imp->imp_connect_data;
3852 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3853 osc_init_grant(&obd->u.cli, ocd);
3856 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3857 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3859 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3863 CERROR("Unknown import event %d\n", event);
3869 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3875 rc = ptlrpcd_addref();
3879 rc = client_obd_setup(obd, lcfg);
3883 struct lprocfs_static_vars lvars = { 0 };
3884 struct client_obd *cli = &obd->u.cli;
3886 lprocfs_osc_init_vars(&lvars);
3887 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3888 lproc_osc_attach_seqstat(obd);
3889 sptlrpc_lprocfs_cliobd_attach(obd);
3890 ptlrpc_lprocfs_register_obd(obd);
3894 /* We need to allocate a few requests more, because
3895 brw_interpret tries to create new requests before freeing
3896 previous ones. Ideally we want to have 2x max_rpcs_in_flight
3897 reserved, but I afraid that might be too much wasted RAM
3898 in fact, so 2 is just my guess and still should work. */
3899 cli->cl_import->imp_rq_pool =
3900 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3902 ptlrpc_add_rqs_to_pool);
3908 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3914 case OBD_CLEANUP_EARLY: {
3915 struct obd_import *imp;
3916 imp = obd->u.cli.cl_import;
3917 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3918 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3919 ptlrpc_deactivate_import(imp);
3920 spin_lock(&imp->imp_lock);
3921 imp->imp_pingable = 0;
3922 spin_unlock(&imp->imp_lock);
3925 case OBD_CLEANUP_EXPORTS: {
3926 /* If we set up but never connected, the
3927 client import will not have been cleaned. */
3928 if (obd->u.cli.cl_import) {
3929 struct obd_import *imp;
3930 down_write(&obd->u.cli.cl_sem);
3931 imp = obd->u.cli.cl_import;
3932 CDEBUG(D_CONFIG, "%s: client import never connected\n",
3934 ptlrpc_invalidate_import(imp);
3935 if (imp->imp_rq_pool) {
3936 ptlrpc_free_rq_pool(imp->imp_rq_pool);
3937 imp->imp_rq_pool = NULL;
3939 class_destroy_import(imp);
3940 up_write(&obd->u.cli.cl_sem);
3941 obd->u.cli.cl_import = NULL;
3943 rc = obd_llog_finish(obd, 0);
3945 CERROR("failed to cleanup llogging subsystems\n");
3952 int osc_cleanup(struct obd_device *obd)
3954 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3958 ptlrpc_lprocfs_unregister_obd(obd);
3959 lprocfs_obd_cleanup(obd);
3961 spin_lock(&oscc->oscc_lock);
3962 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3963 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3964 spin_unlock(&oscc->oscc_lock);
3966 /* free memory of osc quota cache */
3967 lquota_cleanup(quota_interface, obd);
3969 rc = client_obd_cleanup(obd);
3975 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3977 struct lprocfs_static_vars lvars = { 0 };
3980 lprocfs_osc_init_vars(&lvars);
3982 switch (lcfg->lcfg_command) {
3984 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3994 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3996 return osc_process_config_base(obd, buf);
3999 struct obd_ops osc_obd_ops = {
4000 .o_owner = THIS_MODULE,
4001 .o_setup = osc_setup,
4002 .o_precleanup = osc_precleanup,
4003 .o_cleanup = osc_cleanup,
4004 .o_add_conn = client_import_add_conn,
4005 .o_del_conn = client_import_del_conn,
4006 .o_connect = client_connect_import,
4007 .o_reconnect = osc_reconnect,
4008 .o_disconnect = osc_disconnect,
4009 .o_statfs = osc_statfs,
4010 .o_statfs_async = osc_statfs_async,
4011 .o_packmd = osc_packmd,
4012 .o_unpackmd = osc_unpackmd,
4013 .o_precreate = osc_precreate,
4014 .o_create = osc_create,
4015 .o_destroy = osc_destroy,
4016 .o_getattr = osc_getattr,
4017 .o_getattr_async = osc_getattr_async,
4018 .o_setattr = osc_setattr,
4019 .o_setattr_async = osc_setattr_async,
4021 .o_punch = osc_punch,
4023 .o_enqueue = osc_enqueue,
4024 .o_change_cbdata = osc_change_cbdata,
4025 .o_cancel = osc_cancel,
4026 .o_cancel_unused = osc_cancel_unused,
4027 .o_iocontrol = osc_iocontrol,
4028 .o_get_info = osc_get_info,
4029 .o_set_info_async = osc_set_info_async,
4030 .o_import_event = osc_import_event,
4031 .o_llog_init = osc_llog_init,
4032 .o_llog_finish = osc_llog_finish,
4033 .o_process_config = osc_process_config,
4036 extern struct lu_kmem_descr osc_caches[];
4037 extern spinlock_t osc_ast_guard;
4038 extern struct lock_class_key osc_ast_guard_class;
4040 int __init osc_init(void)
4042 struct lprocfs_static_vars lvars = { 0 };
4046 /* print an address of _any_ initialized kernel symbol from this
4047 * module, to allow debugging with gdb that doesn't support data
4048 * symbols from modules.*/
4049 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4051 rc = lu_kmem_init(osc_caches);
4053 lprocfs_osc_init_vars(&lvars);
4055 request_module("lquota");
4056 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4057 lquota_init(quota_interface);
4058 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4060 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4061 LUSTRE_OSC_NAME, &osc_device_type);
4063 if (quota_interface)
4064 PORTAL_SYMBOL_PUT(osc_quota_interface);
4065 lu_kmem_fini(osc_caches);
4069 spin_lock_init(&osc_ast_guard);
4070 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4076 static void /*__exit*/ osc_exit(void)
4078 lu_device_type_fini(&osc_device_type);
4080 lquota_exit(quota_interface);
4081 if (quota_interface)
4082 PORTAL_SYMBOL_PUT(osc_quota_interface);
4084 class_unregister_type(LUSTRE_OSC_NAME);
4085 lu_kmem_fini(osc_caches);
4088 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4089 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4090 MODULE_LICENSE("GPL");
4092 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);