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 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
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 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
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 ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1335 check_write_checksum(&body->oa, peer, client_cksum,
1336 body->oa.o_cksum, aa->aa_requested_nob,
1337 aa->aa_page_count, aa->aa_ppga,
1338 cksum_type_unpack(aa->aa_oa->o_flags)))
1341 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
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 */
1350 if (rc > aa->aa_requested_nob) {
1351 CERROR("Unexpected rc %d (%d requested)\n", rc,
1352 aa->aa_requested_nob);
1356 if (rc != req->rq_bulk->bd_nob_transferred) {
1357 CERROR ("Unexpected rc %d (%d transferred)\n",
1358 rc, req->rq_bulk->bd_nob_transferred);
1362 if (rc < aa->aa_requested_nob)
1363 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1365 if (sptlrpc_cli_unwrap_bulk_read(req, rc, aa->aa_page_count,
1367 GOTO(out, rc = -EAGAIN);
1369 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1370 static int cksum_counter;
1371 __u32 server_cksum = body->oa.o_cksum;
1374 cksum_type_t cksum_type;
1376 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1377 cksum_type = cksum_type_unpack(body->oa.o_flags);
1379 cksum_type = OBD_CKSUM_CRC32;
1380 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1381 aa->aa_ppga, OST_READ,
1384 if (peer->nid == req->rq_bulk->bd_sender) {
1388 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1391 if (server_cksum == ~0 && rc > 0) {
1392 CERROR("Protocol error: server %s set the 'checksum' "
1393 "bit, but didn't send a checksum. Not fatal, "
1394 "but please notify on http://bugzilla.lustre.org/\n",
1395 libcfs_nid2str(peer->nid));
1396 } else if (server_cksum != client_cksum) {
1397 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1398 "%s%s%s inum "LPU64"/"LPU64" object "
1399 LPU64"/"LPU64" extent "
1400 "["LPU64"-"LPU64"]\n",
1401 req->rq_import->imp_obd->obd_name,
1402 libcfs_nid2str(peer->nid),
1404 body->oa.o_valid & OBD_MD_FLFID ?
1405 body->oa.o_fid : (__u64)0,
1406 body->oa.o_valid & OBD_MD_FLFID ?
1407 body->oa.o_generation :(__u64)0,
1409 body->oa.o_valid & OBD_MD_FLGROUP ?
1410 body->oa.o_gr : (__u64)0,
1411 aa->aa_ppga[0]->off,
1412 aa->aa_ppga[aa->aa_page_count-1]->off +
1413 aa->aa_ppga[aa->aa_page_count-1]->count -
1415 CERROR("client %x, server %x, cksum_type %x\n",
1416 client_cksum, server_cksum, cksum_type);
1418 aa->aa_oa->o_cksum = client_cksum;
1422 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1425 } else if (unlikely(client_cksum)) {
1426 static int cksum_missed;
1429 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1430 CERROR("Checksum %u requested from %s but not sent\n",
1431 cksum_missed, libcfs_nid2str(peer->nid));
1437 *aa->aa_oa = body->oa;
1442 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1443 struct lov_stripe_md *lsm,
1444 obd_count page_count, struct brw_page **pga,
1445 struct obd_capa *ocapa)
1447 struct ptlrpc_request *req;
1451 struct l_wait_info lwi;
1455 cfs_waitq_init(&waitq);
1458 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1459 page_count, pga, &req, ocapa, 0);
1463 rc = ptlrpc_queue_wait(req);
1465 if (rc == -ETIMEDOUT && req->rq_resend) {
1466 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1467 ptlrpc_req_finished(req);
1471 rc = osc_brw_fini_request(req, rc);
1473 ptlrpc_req_finished(req);
1474 if (osc_recoverable_error(rc)) {
1476 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1477 CERROR("too many resend retries, returning error\n");
1481 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1482 l_wait_event(waitq, 0, &lwi);
1490 int osc_brw_redo_request(struct ptlrpc_request *request,
1491 struct osc_brw_async_args *aa)
1493 struct ptlrpc_request *new_req;
1494 struct ptlrpc_request_set *set = request->rq_set;
1495 struct osc_brw_async_args *new_aa;
1496 struct osc_async_page *oap;
1500 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1501 CERROR("too many resend retries, returning error\n");
1505 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1507 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1508 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1509 aa->aa_cli, aa->aa_oa,
1510 NULL /* lsm unused by osc currently */,
1511 aa->aa_page_count, aa->aa_ppga,
1512 &new_req, aa->aa_ocapa, 0);
1516 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1518 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1519 if (oap->oap_request != NULL) {
1520 LASSERTF(request == oap->oap_request,
1521 "request %p != oap_request %p\n",
1522 request, oap->oap_request);
1523 if (oap->oap_interrupted) {
1524 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1525 ptlrpc_req_finished(new_req);
1530 /* New request takes over pga and oaps from old request.
1531 * Note that copying a list_head doesn't work, need to move it... */
1533 new_req->rq_interpret_reply = request->rq_interpret_reply;
1534 new_req->rq_async_args = request->rq_async_args;
1535 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1537 new_aa = ptlrpc_req_async_args(new_req);
1539 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1540 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1541 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1543 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1544 if (oap->oap_request) {
1545 ptlrpc_req_finished(oap->oap_request);
1546 oap->oap_request = ptlrpc_request_addref(new_req);
1550 new_aa->aa_ocapa = aa->aa_ocapa;
1551 aa->aa_ocapa = NULL;
1553 /* use ptlrpc_set_add_req is safe because interpret functions work
1554 * in check_set context. only one way exist with access to request
1555 * from different thread got -EINTR - this way protected with
1556 * cl_loi_list_lock */
1557 ptlrpc_set_add_req(set, new_req);
1559 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1561 DEBUG_REQ(D_INFO, new_req, "new request");
1566 * ugh, we want disk allocation on the target to happen in offset order. we'll
1567 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1568 * fine for our small page arrays and doesn't require allocation. its an
1569 * insertion sort that swaps elements that are strides apart, shrinking the
1570 * stride down until its '1' and the array is sorted.
1572 static void sort_brw_pages(struct brw_page **array, int num)
1575 struct brw_page *tmp;
1579 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1584 for (i = stride ; i < num ; i++) {
1587 while (j >= stride && array[j - stride]->off > tmp->off) {
1588 array[j] = array[j - stride];
1593 } while (stride > 1);
1596 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1602 LASSERT (pages > 0);
1603 offset = pg[i]->off & ~CFS_PAGE_MASK;
1607 if (pages == 0) /* that's all */
1610 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1611 return count; /* doesn't end on page boundary */
1614 offset = pg[i]->off & ~CFS_PAGE_MASK;
1615 if (offset != 0) /* doesn't start on page boundary */
1622 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1624 struct brw_page **ppga;
1627 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1631 for (i = 0; i < count; i++)
1636 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1638 LASSERT(ppga != NULL);
1639 OBD_FREE(ppga, sizeof(*ppga) * count);
1642 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1643 obd_count page_count, struct brw_page *pga,
1644 struct obd_trans_info *oti)
1646 struct obdo *saved_oa = NULL;
1647 struct brw_page **ppga, **orig;
1648 struct obd_import *imp = class_exp2cliimp(exp);
1649 struct client_obd *cli = &imp->imp_obd->u.cli;
1650 int rc, page_count_orig;
1653 if (cmd & OBD_BRW_CHECK) {
1654 /* The caller just wants to know if there's a chance that this
1655 * I/O can succeed */
1657 if (imp == NULL || imp->imp_invalid)
1662 /* test_brw with a failed create can trip this, maybe others. */
1663 LASSERT(cli->cl_max_pages_per_rpc);
1667 orig = ppga = osc_build_ppga(pga, page_count);
1670 page_count_orig = page_count;
1672 sort_brw_pages(ppga, page_count);
1673 while (page_count) {
1674 obd_count pages_per_brw;
1676 if (page_count > cli->cl_max_pages_per_rpc)
1677 pages_per_brw = cli->cl_max_pages_per_rpc;
1679 pages_per_brw = page_count;
1681 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1683 if (saved_oa != NULL) {
1684 /* restore previously saved oa */
1685 *oinfo->oi_oa = *saved_oa;
1686 } else if (page_count > pages_per_brw) {
1687 /* save a copy of oa (brw will clobber it) */
1688 OBDO_ALLOC(saved_oa);
1689 if (saved_oa == NULL)
1690 GOTO(out, rc = -ENOMEM);
1691 *saved_oa = *oinfo->oi_oa;
1694 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1695 pages_per_brw, ppga, oinfo->oi_capa);
1700 page_count -= pages_per_brw;
1701 ppga += pages_per_brw;
1705 osc_release_ppga(orig, page_count_orig);
1707 if (saved_oa != NULL)
1708 OBDO_FREE(saved_oa);
1713 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1714 * the dirty accounting. Writeback completes or truncate happens before
1715 * writing starts. Must be called with the loi lock held. */
1716 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1719 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1723 /* This maintains the lists of pending pages to read/write for a given object
1724 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1725 * to quickly find objects that are ready to send an RPC. */
1726 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1732 if (lop->lop_num_pending == 0)
1735 /* if we have an invalid import we want to drain the queued pages
1736 * by forcing them through rpcs that immediately fail and complete
1737 * the pages. recovery relies on this to empty the queued pages
1738 * before canceling the locks and evicting down the llite pages */
1739 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1742 /* stream rpcs in queue order as long as as there is an urgent page
1743 * queued. this is our cheap solution for good batching in the case
1744 * where writepage marks some random page in the middle of the file
1745 * as urgent because of, say, memory pressure */
1746 if (!list_empty(&lop->lop_urgent)) {
1747 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1750 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1751 optimal = cli->cl_max_pages_per_rpc;
1752 if (cmd & OBD_BRW_WRITE) {
1753 /* trigger a write rpc stream as long as there are dirtiers
1754 * waiting for space. as they're waiting, they're not going to
1755 * create more pages to coallesce with what's waiting.. */
1756 if (!list_empty(&cli->cl_cache_waiters)) {
1757 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1760 /* +16 to avoid triggering rpcs that would want to include pages
1761 * that are being queued but which can't be made ready until
1762 * the queuer finishes with the page. this is a wart for
1763 * llite::commit_write() */
1766 if (lop->lop_num_pending >= optimal)
1772 static void on_list(struct list_head *item, struct list_head *list,
1775 if (list_empty(item) && should_be_on)
1776 list_add_tail(item, list);
1777 else if (!list_empty(item) && !should_be_on)
1778 list_del_init(item);
1781 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1782 * can find pages to build into rpcs quickly */
1783 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1785 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1786 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1787 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1789 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1790 loi->loi_write_lop.lop_num_pending);
1792 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1793 loi->loi_read_lop.lop_num_pending);
1796 static void lop_update_pending(struct client_obd *cli,
1797 struct loi_oap_pages *lop, int cmd, int delta)
1799 lop->lop_num_pending += delta;
1800 if (cmd & OBD_BRW_WRITE)
1801 cli->cl_pending_w_pages += delta;
1803 cli->cl_pending_r_pages += delta;
1807 * this is called when a sync waiter receives an interruption. Its job is to
1808 * get the caller woken as soon as possible. If its page hasn't been put in an
1809 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1810 * desiring interruption which will forcefully complete the rpc once the rpc
1813 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
1815 struct loi_oap_pages *lop;
1816 struct lov_oinfo *loi;
1820 LASSERT(!oap->oap_interrupted);
1821 oap->oap_interrupted = 1;
1823 /* ok, it's been put in an rpc. only one oap gets a request reference */
1824 if (oap->oap_request != NULL) {
1825 ptlrpc_mark_interrupted(oap->oap_request);
1826 ptlrpcd_wake(oap->oap_request);
1827 ptlrpc_req_finished(oap->oap_request);
1828 oap->oap_request = NULL;
1832 * page completion may be called only if ->cpo_prep() method was
1833 * executed by osc_io_submit(), that also adds page the to pending list
1835 if (!list_empty(&oap->oap_pending_item)) {
1836 list_del_init(&oap->oap_pending_item);
1837 list_del_init(&oap->oap_urgent_item);
1840 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1841 &loi->loi_write_lop : &loi->loi_read_lop;
1842 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1843 loi_list_maint(oap->oap_cli, oap->oap_loi);
1844 rc = oap->oap_caller_ops->ap_completion(env,
1845 oap->oap_caller_data,
1846 oap->oap_cmd, NULL, -EINTR);
1852 /* this is trying to propogate async writeback errors back up to the
1853 * application. As an async write fails we record the error code for later if
1854 * the app does an fsync. As long as errors persist we force future rpcs to be
1855 * sync so that the app can get a sync error and break the cycle of queueing
1856 * pages for which writeback will fail. */
1857 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1864 ar->ar_force_sync = 1;
1865 ar->ar_min_xid = ptlrpc_sample_next_xid();
1870 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1871 ar->ar_force_sync = 0;
1874 void osc_oap_to_pending(struct osc_async_page *oap)
1876 struct loi_oap_pages *lop;
1878 if (oap->oap_cmd & OBD_BRW_WRITE)
1879 lop = &oap->oap_loi->loi_write_lop;
1881 lop = &oap->oap_loi->loi_read_lop;
1883 if (oap->oap_async_flags & ASYNC_URGENT)
1884 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1885 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1886 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1889 /* this must be called holding the loi list lock to give coverage to exit_cache,
1890 * async_flag maintenance, and oap_request */
1891 static void osc_ap_completion(const struct lu_env *env,
1892 struct client_obd *cli, struct obdo *oa,
1893 struct osc_async_page *oap, int sent, int rc)
1898 if (oap->oap_request != NULL) {
1899 xid = ptlrpc_req_xid(oap->oap_request);
1900 ptlrpc_req_finished(oap->oap_request);
1901 oap->oap_request = NULL;
1904 oap->oap_async_flags = 0;
1905 oap->oap_interrupted = 0;
1907 if (oap->oap_cmd & OBD_BRW_WRITE) {
1908 osc_process_ar(&cli->cl_ar, xid, rc);
1909 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
1912 if (rc == 0 && oa != NULL) {
1913 if (oa->o_valid & OBD_MD_FLBLOCKS)
1914 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
1915 if (oa->o_valid & OBD_MD_FLMTIME)
1916 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
1917 if (oa->o_valid & OBD_MD_FLATIME)
1918 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
1919 if (oa->o_valid & OBD_MD_FLCTIME)
1920 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
1923 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
1924 oap->oap_cmd, oa, rc);
1926 /* ll_ap_completion (from llite) drops PG_locked. so, a new
1927 * I/O on the page could start, but OSC calls it under lock
1928 * and thus we can add oap back to pending safely */
1930 /* upper layer wants to leave the page on pending queue */
1931 osc_oap_to_pending(oap);
1933 osc_exit_cache(cli, oap, sent);
1937 static int brw_interpret(const struct lu_env *env,
1938 struct ptlrpc_request *req, void *data, int rc)
1940 struct osc_brw_async_args *aa = data;
1941 struct client_obd *cli;
1945 rc = osc_brw_fini_request(req, rc);
1946 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1947 if (osc_recoverable_error(rc)) {
1948 rc = osc_brw_redo_request(req, aa);
1954 capa_put(aa->aa_ocapa);
1955 aa->aa_ocapa = NULL;
1960 client_obd_list_lock(&cli->cl_loi_list_lock);
1962 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1963 * is called so we know whether to go to sync BRWs or wait for more
1964 * RPCs to complete */
1965 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1966 cli->cl_w_in_flight--;
1968 cli->cl_r_in_flight--;
1970 async = list_empty(&aa->aa_oaps);
1971 if (!async) { /* from osc_send_oap_rpc() */
1972 struct osc_async_page *oap, *tmp;
1973 /* the caller may re-use the oap after the completion call so
1974 * we need to clean it up a little */
1975 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
1976 list_del_init(&oap->oap_rpc_item);
1977 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
1979 OBDO_FREE(aa->aa_oa);
1980 } else { /* from async_internal() */
1982 for (i = 0; i < aa->aa_page_count; i++)
1983 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
1985 osc_wake_cache_waiters(cli);
1986 osc_check_rpcs(env, cli);
1987 client_obd_list_unlock(&cli->cl_loi_list_lock);
1989 cl_req_completion(env, aa->aa_clerq, rc);
1990 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1994 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
1995 struct client_obd *cli,
1996 struct list_head *rpc_list,
1997 int page_count, int cmd)
1999 struct ptlrpc_request *req;
2000 struct brw_page **pga = NULL;
2001 struct osc_brw_async_args *aa;
2002 struct obdo *oa = NULL;
2003 const struct obd_async_page_ops *ops = NULL;
2004 void *caller_data = NULL;
2005 struct osc_async_page *oap;
2006 struct osc_async_page *tmp;
2007 struct ost_body *body;
2008 struct cl_req *clerq = NULL;
2009 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2010 struct ldlm_lock *lock = NULL;
2011 struct cl_req_attr crattr;
2015 LASSERT(!list_empty(rpc_list));
2017 memset(&crattr, 0, sizeof crattr);
2018 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2020 GOTO(out, req = ERR_PTR(-ENOMEM));
2024 GOTO(out, req = ERR_PTR(-ENOMEM));
2027 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2028 struct cl_page *page = osc_oap2cl_page(oap);
2030 ops = oap->oap_caller_ops;
2031 caller_data = oap->oap_caller_data;
2033 clerq = cl_req_alloc(env, page, crt,
2034 1 /* only 1-object rpcs for
2037 GOTO(out, req = (void *)clerq);
2038 lock = oap->oap_ldlm_lock;
2040 pga[i] = &oap->oap_brw_page;
2041 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2042 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2043 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2045 cl_req_page_add(env, clerq, page);
2048 /* always get the data for the obdo for the rpc */
2049 LASSERT(ops != NULL);
2051 crattr.cra_capa = NULL;
2052 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2054 oa->o_handle = lock->l_remote_handle;
2055 oa->o_valid |= OBD_MD_FLHANDLE;
2058 rc = cl_req_prep(env, clerq);
2060 CERROR("cl_req_prep failed: %d\n", rc);
2061 GOTO(out, req = ERR_PTR(rc));
2064 sort_brw_pages(pga, page_count);
2065 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2066 pga, &req, crattr.cra_capa, 1);
2068 CERROR("prep_req failed: %d\n", rc);
2069 GOTO(out, req = ERR_PTR(rc));
2072 /* Need to update the timestamps after the request is built in case
2073 * we race with setattr (locally or in queue at OST). If OST gets
2074 * later setattr before earlier BRW (as determined by the request xid),
2075 * the OST will not use BRW timestamps. Sadly, there is no obvious
2076 * way to do this in a single call. bug 10150 */
2077 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2078 cl_req_attr_set(env, clerq, &crattr,
2079 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2081 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2082 aa = ptlrpc_req_async_args(req);
2083 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2084 list_splice(rpc_list, &aa->aa_oaps);
2085 CFS_INIT_LIST_HEAD(rpc_list);
2086 aa->aa_clerq = clerq;
2088 capa_put(crattr.cra_capa);
2093 OBD_FREE(pga, sizeof(*pga) * page_count);
2094 /* this should happen rarely and is pretty bad, it makes the
2095 * pending list not follow the dirty order */
2096 client_obd_list_lock(&cli->cl_loi_list_lock);
2097 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2098 list_del_init(&oap->oap_rpc_item);
2100 /* queued sync pages can be torn down while the pages
2101 * were between the pending list and the rpc */
2102 if (oap->oap_interrupted) {
2103 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2104 osc_ap_completion(env, cli, NULL, oap, 0,
2108 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2110 if (clerq && !IS_ERR(clerq))
2111 cl_req_completion(env, clerq, PTR_ERR(req));
2117 * prepare pages for ASYNC io and put pages in send queue.
2121 * \param cmd - OBD_BRW_* macroses
2122 * \param lop - pending pages
2124 * \return zero if pages successfully add to send queue.
2125 * \return not zere if error occurring.
2128 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2129 struct lov_oinfo *loi,
2130 int cmd, struct loi_oap_pages *lop)
2132 struct ptlrpc_request *req;
2133 obd_count page_count = 0;
2134 struct osc_async_page *oap = NULL, *tmp;
2135 struct osc_brw_async_args *aa;
2136 const struct obd_async_page_ops *ops;
2137 CFS_LIST_HEAD(rpc_list);
2138 unsigned int ending_offset;
2139 unsigned starting_offset = 0;
2141 struct cl_object *clob = NULL;
2144 /* first we find the pages we're allowed to work with */
2145 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2147 ops = oap->oap_caller_ops;
2149 LASSERT(oap->oap_magic == OAP_MAGIC);
2152 /* pin object in memory, so that completion call-backs
2153 * can be safely called under client_obd_list lock. */
2154 clob = osc_oap2cl_page(oap)->cp_obj;
2155 cl_object_get(clob);
2158 if (page_count != 0 &&
2159 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2160 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2161 " oap %p, page %p, srvlock %u\n",
2162 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2165 /* in llite being 'ready' equates to the page being locked
2166 * until completion unlocks it. commit_write submits a page
2167 * as not ready because its unlock will happen unconditionally
2168 * as the call returns. if we race with commit_write giving
2169 * us that page we dont' want to create a hole in the page
2170 * stream, so we stop and leave the rpc to be fired by
2171 * another dirtier or kupdated interval (the not ready page
2172 * will still be on the dirty list). we could call in
2173 * at the end of ll_file_write to process the queue again. */
2174 if (!(oap->oap_async_flags & ASYNC_READY)) {
2175 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2178 CDEBUG(D_INODE, "oap %p page %p returned %d "
2179 "instead of ready\n", oap,
2183 /* llite is telling us that the page is still
2184 * in commit_write and that we should try
2185 * and put it in an rpc again later. we
2186 * break out of the loop so we don't create
2187 * a hole in the sequence of pages in the rpc
2192 /* the io isn't needed.. tell the checks
2193 * below to complete the rpc with EINTR */
2194 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2195 oap->oap_count = -EINTR;
2198 oap->oap_async_flags |= ASYNC_READY;
2201 LASSERTF(0, "oap %p page %p returned %d "
2202 "from make_ready\n", oap,
2210 * Page submitted for IO has to be locked. Either by
2211 * ->ap_make_ready() or by higher layers.
2213 #if defined(__KERNEL__) && defined(__linux__)
2215 struct cl_page *page;
2217 page = osc_oap2cl_page(oap);
2219 if (page->cp_type == CPT_CACHEABLE &&
2220 !(PageLocked(oap->oap_page) &&
2221 (CheckWriteback(oap->oap_page, cmd)))) {
2222 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2224 (long)oap->oap_page->flags,
2225 oap->oap_async_flags);
2230 /* If there is a gap at the start of this page, it can't merge
2231 * with any previous page, so we'll hand the network a
2232 * "fragmented" page array that it can't transfer in 1 RDMA */
2233 if (page_count != 0 && oap->oap_page_off != 0)
2236 /* take the page out of our book-keeping */
2237 list_del_init(&oap->oap_pending_item);
2238 lop_update_pending(cli, lop, cmd, -1);
2239 list_del_init(&oap->oap_urgent_item);
2241 if (page_count == 0)
2242 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2243 (PTLRPC_MAX_BRW_SIZE - 1);
2245 /* ask the caller for the size of the io as the rpc leaves. */
2246 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2248 ops->ap_refresh_count(env, oap->oap_caller_data,
2250 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2252 if (oap->oap_count <= 0) {
2253 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2255 osc_ap_completion(env, cli, NULL,
2256 oap, 0, oap->oap_count);
2260 /* now put the page back in our accounting */
2261 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2262 if (page_count == 0)
2263 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2264 if (++page_count >= cli->cl_max_pages_per_rpc)
2267 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2268 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2269 * have the same alignment as the initial writes that allocated
2270 * extents on the server. */
2271 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2272 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2273 if (ending_offset == 0)
2276 /* If there is a gap at the end of this page, it can't merge
2277 * with any subsequent pages, so we'll hand the network a
2278 * "fragmented" page array that it can't transfer in 1 RDMA */
2279 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2283 osc_wake_cache_waiters(cli);
2285 loi_list_maint(cli, loi);
2287 client_obd_list_unlock(&cli->cl_loi_list_lock);
2290 cl_object_put(env, clob);
2292 if (page_count == 0) {
2293 client_obd_list_lock(&cli->cl_loi_list_lock);
2297 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2299 LASSERT(list_empty(&rpc_list));
2300 loi_list_maint(cli, loi);
2301 RETURN(PTR_ERR(req));
2304 aa = ptlrpc_req_async_args(req);
2306 if (cmd == OBD_BRW_READ) {
2307 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2308 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2309 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2310 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2312 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2313 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2314 cli->cl_w_in_flight);
2315 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2316 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2318 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2320 client_obd_list_lock(&cli->cl_loi_list_lock);
2322 if (cmd == OBD_BRW_READ)
2323 cli->cl_r_in_flight++;
2325 cli->cl_w_in_flight++;
2327 /* queued sync pages can be torn down while the pages
2328 * were between the pending list and the rpc */
2330 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2331 /* only one oap gets a request reference */
2334 if (oap->oap_interrupted && !req->rq_intr) {
2335 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2337 ptlrpc_mark_interrupted(req);
2341 tmp->oap_request = ptlrpc_request_addref(req);
2343 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2344 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2346 req->rq_interpret_reply = brw_interpret;
2347 ptlrpcd_add_req(req, PSCOPE_BRW);
2351 #define LOI_DEBUG(LOI, STR, args...) \
2352 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2353 !list_empty(&(LOI)->loi_cli_item), \
2354 (LOI)->loi_write_lop.lop_num_pending, \
2355 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2356 (LOI)->loi_read_lop.lop_num_pending, \
2357 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2360 /* This is called by osc_check_rpcs() to find which objects have pages that
2361 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2362 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2365 /* first return all objects which we already know to have
2366 * pages ready to be stuffed into rpcs */
2367 if (!list_empty(&cli->cl_loi_ready_list))
2368 RETURN(list_entry(cli->cl_loi_ready_list.next,
2369 struct lov_oinfo, loi_cli_item));
2371 /* then if we have cache waiters, return all objects with queued
2372 * writes. This is especially important when many small files
2373 * have filled up the cache and not been fired into rpcs because
2374 * they don't pass the nr_pending/object threshhold */
2375 if (!list_empty(&cli->cl_cache_waiters) &&
2376 !list_empty(&cli->cl_loi_write_list))
2377 RETURN(list_entry(cli->cl_loi_write_list.next,
2378 struct lov_oinfo, loi_write_item));
2380 /* then return all queued objects when we have an invalid import
2381 * so that they get flushed */
2382 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2383 if (!list_empty(&cli->cl_loi_write_list))
2384 RETURN(list_entry(cli->cl_loi_write_list.next,
2385 struct lov_oinfo, loi_write_item));
2386 if (!list_empty(&cli->cl_loi_read_list))
2387 RETURN(list_entry(cli->cl_loi_read_list.next,
2388 struct lov_oinfo, loi_read_item));
2393 /* called with the loi list lock held */
2394 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2396 struct lov_oinfo *loi;
2397 int rc = 0, race_counter = 0;
2400 while ((loi = osc_next_loi(cli)) != NULL) {
2401 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2403 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2406 /* attempt some read/write balancing by alternating between
2407 * reads and writes in an object. The makes_rpc checks here
2408 * would be redundant if we were getting read/write work items
2409 * instead of objects. we don't want send_oap_rpc to drain a
2410 * partial read pending queue when we're given this object to
2411 * do io on writes while there are cache waiters */
2412 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2413 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2414 &loi->loi_write_lop);
2422 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2423 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2424 &loi->loi_read_lop);
2433 /* attempt some inter-object balancing by issueing rpcs
2434 * for each object in turn */
2435 if (!list_empty(&loi->loi_cli_item))
2436 list_del_init(&loi->loi_cli_item);
2437 if (!list_empty(&loi->loi_write_item))
2438 list_del_init(&loi->loi_write_item);
2439 if (!list_empty(&loi->loi_read_item))
2440 list_del_init(&loi->loi_read_item);
2442 loi_list_maint(cli, loi);
2444 /* send_oap_rpc fails with 0 when make_ready tells it to
2445 * back off. llite's make_ready does this when it tries
2446 * to lock a page queued for write that is already locked.
2447 * we want to try sending rpcs from many objects, but we
2448 * don't want to spin failing with 0. */
2449 if (race_counter == 10)
2455 /* we're trying to queue a page in the osc so we're subject to the
2456 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2457 * If the osc's queued pages are already at that limit, then we want to sleep
2458 * until there is space in the osc's queue for us. We also may be waiting for
2459 * write credits from the OST if there are RPCs in flight that may return some
2460 * before we fall back to sync writes.
2462 * We need this know our allocation was granted in the presence of signals */
2463 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2467 client_obd_list_lock(&cli->cl_loi_list_lock);
2468 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2469 client_obd_list_unlock(&cli->cl_loi_list_lock);
2474 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2477 int osc_enter_cache_try(const struct lu_env *env,
2478 struct client_obd *cli, struct lov_oinfo *loi,
2479 struct osc_async_page *oap, int transient)
2483 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2485 osc_consume_write_grant(cli, &oap->oap_brw_page);
2487 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2488 atomic_inc(&obd_dirty_transit_pages);
2489 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2495 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2496 * grant or cache space. */
2497 static int osc_enter_cache(const struct lu_env *env,
2498 struct client_obd *cli, struct lov_oinfo *loi,
2499 struct osc_async_page *oap)
2501 struct osc_cache_waiter ocw;
2502 struct l_wait_info lwi = { 0 };
2506 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2507 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2508 cli->cl_dirty_max, obd_max_dirty_pages,
2509 cli->cl_lost_grant, cli->cl_avail_grant);
2511 /* force the caller to try sync io. this can jump the list
2512 * of queued writes and create a discontiguous rpc stream */
2513 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2514 loi->loi_ar.ar_force_sync)
2517 /* Hopefully normal case - cache space and write credits available */
2518 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2519 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2520 osc_enter_cache_try(env, cli, loi, oap, 0))
2523 /* Make sure that there are write rpcs in flight to wait for. This
2524 * is a little silly as this object may not have any pending but
2525 * other objects sure might. */
2526 if (cli->cl_w_in_flight) {
2527 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2528 cfs_waitq_init(&ocw.ocw_waitq);
2532 loi_list_maint(cli, loi);
2533 osc_check_rpcs(env, cli);
2534 client_obd_list_unlock(&cli->cl_loi_list_lock);
2536 CDEBUG(D_CACHE, "sleeping for cache space\n");
2537 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2539 client_obd_list_lock(&cli->cl_loi_list_lock);
2540 if (!list_empty(&ocw.ocw_entry)) {
2541 list_del(&ocw.ocw_entry);
2551 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2552 struct lov_oinfo *loi, cfs_page_t *page,
2553 obd_off offset, const struct obd_async_page_ops *ops,
2554 void *data, void **res, int nocache,
2555 struct lustre_handle *lockh)
2557 struct osc_async_page *oap;
2562 return size_round(sizeof(*oap));
2565 oap->oap_magic = OAP_MAGIC;
2566 oap->oap_cli = &exp->exp_obd->u.cli;
2569 oap->oap_caller_ops = ops;
2570 oap->oap_caller_data = data;
2572 oap->oap_page = page;
2573 oap->oap_obj_off = offset;
2574 if (!client_is_remote(exp) &&
2575 cfs_capable(CFS_CAP_SYS_RESOURCE))
2576 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2578 LASSERT(!(offset & ~CFS_PAGE_MASK));
2580 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2581 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2582 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2583 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2585 spin_lock_init(&oap->oap_lock);
2586 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2590 struct osc_async_page *oap_from_cookie(void *cookie)
2592 struct osc_async_page *oap = cookie;
2593 if (oap->oap_magic != OAP_MAGIC)
2594 return ERR_PTR(-EINVAL);
2598 int osc_queue_async_io(const struct lu_env *env,
2599 struct obd_export *exp, struct lov_stripe_md *lsm,
2600 struct lov_oinfo *loi, void *cookie,
2601 int cmd, obd_off off, int count,
2602 obd_flag brw_flags, enum async_flags async_flags)
2604 struct client_obd *cli = &exp->exp_obd->u.cli;
2605 struct osc_async_page *oap;
2609 oap = oap_from_cookie(cookie);
2611 RETURN(PTR_ERR(oap));
2613 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2616 if (!list_empty(&oap->oap_pending_item) ||
2617 !list_empty(&oap->oap_urgent_item) ||
2618 !list_empty(&oap->oap_rpc_item))
2621 /* check if the file's owner/group is over quota */
2622 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2623 struct cl_object *obj;
2624 struct cl_attr attr; /* XXX put attr into thread info */
2626 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2628 cl_object_attr_lock(obj);
2629 rc = cl_object_attr_get(env, obj, &attr);
2630 cl_object_attr_unlock(obj);
2632 if (rc == 0 && lquota_chkdq(quota_interface, cli, attr.cat_uid,
2633 attr.cat_gid) == NO_QUOTA)
2640 loi = lsm->lsm_oinfo[0];
2642 client_obd_list_lock(&cli->cl_loi_list_lock);
2644 LASSERT(off + count <= CFS_PAGE_SIZE);
2646 oap->oap_page_off = off;
2647 oap->oap_count = count;
2648 oap->oap_brw_flags = brw_flags;
2649 oap->oap_async_flags = async_flags;
2651 if (cmd & OBD_BRW_WRITE) {
2652 rc = osc_enter_cache(env, cli, loi, oap);
2654 client_obd_list_unlock(&cli->cl_loi_list_lock);
2659 osc_oap_to_pending(oap);
2660 loi_list_maint(cli, loi);
2662 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2665 osc_check_rpcs(env, cli);
2666 client_obd_list_unlock(&cli->cl_loi_list_lock);
2671 /* aka (~was & now & flag), but this is more clear :) */
2672 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2674 int osc_set_async_flags_base(struct client_obd *cli,
2675 struct lov_oinfo *loi, struct osc_async_page *oap,
2676 obd_flag async_flags)
2678 struct loi_oap_pages *lop;
2681 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2684 if (oap->oap_cmd & OBD_BRW_WRITE) {
2685 lop = &loi->loi_write_lop;
2687 lop = &loi->loi_read_lop;
2690 if (list_empty(&oap->oap_pending_item))
2693 if ((oap->oap_async_flags & async_flags) == async_flags)
2696 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2697 oap->oap_async_flags |= ASYNC_READY;
2699 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2700 if (list_empty(&oap->oap_rpc_item)) {
2701 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2702 loi_list_maint(cli, loi);
2706 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2707 oap->oap_async_flags);
2711 int osc_teardown_async_page(struct obd_export *exp,
2712 struct lov_stripe_md *lsm,
2713 struct lov_oinfo *loi, void *cookie)
2715 struct client_obd *cli = &exp->exp_obd->u.cli;
2716 struct loi_oap_pages *lop;
2717 struct osc_async_page *oap;
2721 oap = oap_from_cookie(cookie);
2723 RETURN(PTR_ERR(oap));
2726 loi = lsm->lsm_oinfo[0];
2728 if (oap->oap_cmd & OBD_BRW_WRITE) {
2729 lop = &loi->loi_write_lop;
2731 lop = &loi->loi_read_lop;
2734 client_obd_list_lock(&cli->cl_loi_list_lock);
2736 if (!list_empty(&oap->oap_rpc_item))
2737 GOTO(out, rc = -EBUSY);
2739 osc_exit_cache(cli, oap, 0);
2740 osc_wake_cache_waiters(cli);
2742 if (!list_empty(&oap->oap_urgent_item)) {
2743 list_del_init(&oap->oap_urgent_item);
2744 oap->oap_async_flags &= ~ASYNC_URGENT;
2746 if (!list_empty(&oap->oap_pending_item)) {
2747 list_del_init(&oap->oap_pending_item);
2748 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2750 loi_list_maint(cli, loi);
2751 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2753 client_obd_list_unlock(&cli->cl_loi_list_lock);
2757 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
2758 struct ldlm_enqueue_info *einfo,
2761 void *data = einfo->ei_cbdata;
2763 LASSERT(lock != NULL);
2764 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2765 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2766 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2767 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2769 lock_res_and_lock(lock);
2770 spin_lock(&osc_ast_guard);
2771 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
2772 lock->l_ast_data = data;
2773 spin_unlock(&osc_ast_guard);
2774 unlock_res_and_lock(lock);
2777 static void osc_set_data_with_check(struct lustre_handle *lockh,
2778 struct ldlm_enqueue_info *einfo,
2781 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2784 osc_set_lock_data_with_check(lock, einfo, flags);
2785 LDLM_LOCK_PUT(lock);
2787 CERROR("lockh %p, data %p - client evicted?\n",
2788 lockh, einfo->ei_cbdata);
2791 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2792 ldlm_iterator_t replace, void *data)
2794 struct ldlm_res_id res_id;
2795 struct obd_device *obd = class_exp2obd(exp);
2797 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
2798 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2802 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2803 obd_enqueue_update_f upcall, void *cookie,
2806 int intent = *flags & LDLM_FL_HAS_INTENT;
2810 /* The request was created before ldlm_cli_enqueue call. */
2811 if (rc == ELDLM_LOCK_ABORTED) {
2812 struct ldlm_reply *rep;
2813 rep = req_capsule_server_get(&req->rq_pill,
2816 LASSERT(rep != NULL);
2817 if (rep->lock_policy_res1)
2818 rc = rep->lock_policy_res1;
2822 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
2823 *flags |= LDLM_FL_LVB_READY;
2824 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2825 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2828 /* Call the update callback. */
2829 rc = (*upcall)(cookie, rc);
2833 static int osc_enqueue_interpret(const struct lu_env *env,
2834 struct ptlrpc_request *req,
2835 struct osc_enqueue_args *aa, int rc)
2837 struct ldlm_lock *lock;
2838 struct lustre_handle handle;
2841 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2842 * might be freed anytime after lock upcall has been called. */
2843 lustre_handle_copy(&handle, aa->oa_lockh);
2844 mode = aa->oa_ei->ei_mode;
2846 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2848 lock = ldlm_handle2lock(&handle);
2850 /* Take an additional reference so that a blocking AST that
2851 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2852 * to arrive after an upcall has been executed by
2853 * osc_enqueue_fini(). */
2854 ldlm_lock_addref(&handle, mode);
2856 /* Complete obtaining the lock procedure. */
2857 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2858 mode, aa->oa_flags, aa->oa_lvb,
2859 sizeof(*aa->oa_lvb), lustre_swab_ost_lvb,
2861 /* Complete osc stuff. */
2862 rc = osc_enqueue_fini(req, aa->oa_lvb,
2863 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
2864 /* Release the lock for async request. */
2865 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2867 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2868 * not already released by
2869 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2871 ldlm_lock_decref(&handle, mode);
2873 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2874 aa->oa_lockh, req, aa);
2875 ldlm_lock_decref(&handle, mode);
2876 LDLM_LOCK_PUT(lock);
2880 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2881 struct lov_oinfo *loi, int flags,
2882 struct ost_lvb *lvb, __u32 mode, int rc)
2884 if (rc == ELDLM_OK) {
2885 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2888 LASSERT(lock != NULL);
2889 loi->loi_lvb = *lvb;
2890 tmp = loi->loi_lvb.lvb_size;
2891 /* Extend KMS up to the end of this lock and no further
2892 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2893 if (tmp > lock->l_policy_data.l_extent.end)
2894 tmp = lock->l_policy_data.l_extent.end + 1;
2895 if (tmp >= loi->loi_kms) {
2896 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2897 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2898 loi_kms_set(loi, tmp);
2900 LDLM_DEBUG(lock, "lock acquired, setting rss="
2901 LPU64"; leaving kms="LPU64", end="LPU64,
2902 loi->loi_lvb.lvb_size, loi->loi_kms,
2903 lock->l_policy_data.l_extent.end);
2905 ldlm_lock_allow_match(lock);
2906 LDLM_LOCK_PUT(lock);
2907 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2908 loi->loi_lvb = *lvb;
2909 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2910 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2914 EXPORT_SYMBOL(osc_update_enqueue);
2916 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2918 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2919 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2920 * other synchronous requests, however keeping some locks and trying to obtain
2921 * others may take a considerable amount of time in a case of ost failure; and
2922 * when other sync requests do not get released lock from a client, the client
2923 * is excluded from the cluster -- such scenarious make the life difficult, so
2924 * release locks just after they are obtained. */
2925 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2926 int *flags, ldlm_policy_data_t *policy,
2927 struct ost_lvb *lvb, int kms_valid,
2928 obd_enqueue_update_f upcall, void *cookie,
2929 struct ldlm_enqueue_info *einfo,
2930 struct lustre_handle *lockh,
2931 struct ptlrpc_request_set *rqset, int async)
2933 struct obd_device *obd = exp->exp_obd;
2934 struct ptlrpc_request *req = NULL;
2935 int intent = *flags & LDLM_FL_HAS_INTENT;
2940 /* Filesystem lock extents are extended to page boundaries so that
2941 * dealing with the page cache is a little smoother. */
2942 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2943 policy->l_extent.end |= ~CFS_PAGE_MASK;
2946 * kms is not valid when either object is completely fresh (so that no
2947 * locks are cached), or object was evicted. In the latter case cached
2948 * lock cannot be used, because it would prime inode state with
2949 * potentially stale LVB.
2954 /* Next, search for already existing extent locks that will cover us */
2955 /* If we're trying to read, we also search for an existing PW lock. The
2956 * VFS and page cache already protect us locally, so lots of readers/
2957 * writers can share a single PW lock.
2959 * There are problems with conversion deadlocks, so instead of
2960 * converting a read lock to a write lock, we'll just enqueue a new
2963 * At some point we should cancel the read lock instead of making them
2964 * send us a blocking callback, but there are problems with canceling
2965 * locks out from other users right now, too. */
2966 mode = einfo->ei_mode;
2967 if (einfo->ei_mode == LCK_PR)
2969 mode = ldlm_lock_match(obd->obd_namespace,
2970 *flags | LDLM_FL_LVB_READY, res_id,
2971 einfo->ei_type, policy, mode, lockh, 0);
2973 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2975 if (matched->l_ast_data == NULL ||
2976 matched->l_ast_data == einfo->ei_cbdata) {
2977 /* addref the lock only if not async requests and PW
2978 * lock is matched whereas we asked for PR. */
2979 if (!rqset && einfo->ei_mode != mode)
2980 ldlm_lock_addref(lockh, LCK_PR);
2981 osc_set_lock_data_with_check(matched, einfo, *flags);
2983 /* I would like to be able to ASSERT here that
2984 * rss <= kms, but I can't, for reasons which
2985 * are explained in lov_enqueue() */
2988 /* We already have a lock, and it's referenced */
2989 (*upcall)(cookie, ELDLM_OK);
2991 /* For async requests, decref the lock. */
2992 if (einfo->ei_mode != mode)
2993 ldlm_lock_decref(lockh, LCK_PW);
2995 ldlm_lock_decref(lockh, einfo->ei_mode);
2996 LDLM_LOCK_PUT(matched);
2999 ldlm_lock_decref(lockh, mode);
3000 LDLM_LOCK_PUT(matched);
3005 CFS_LIST_HEAD(cancels);
3006 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3007 &RQF_LDLM_ENQUEUE_LVB);
3011 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3015 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3017 ptlrpc_request_set_replen(req);
3020 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3021 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3023 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3024 sizeof(*lvb), lustre_swab_ost_lvb, lockh, async);
3027 struct osc_enqueue_args *aa;
3028 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3029 aa = ptlrpc_req_async_args(req);
3032 aa->oa_flags = flags;
3033 aa->oa_upcall = upcall;
3034 aa->oa_cookie = cookie;
3036 aa->oa_lockh = lockh;
3038 req->rq_interpret_reply =
3039 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3040 if (rqset == PTLRPCD_SET)
3041 ptlrpcd_add_req(req, PSCOPE_OTHER);
3043 ptlrpc_set_add_req(rqset, req);
3044 } else if (intent) {
3045 ptlrpc_req_finished(req);
3050 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3052 ptlrpc_req_finished(req);
3057 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3058 struct ldlm_enqueue_info *einfo,
3059 struct ptlrpc_request_set *rqset)
3061 struct ldlm_res_id res_id;
3065 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3066 oinfo->oi_md->lsm_object_gr, &res_id);
3068 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3069 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3070 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3071 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3072 rqset, rqset != NULL);
3076 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3077 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3078 int *flags, void *data, struct lustre_handle *lockh,
3081 struct obd_device *obd = exp->exp_obd;
3082 int lflags = *flags;
3086 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3089 /* Filesystem lock extents are extended to page boundaries so that
3090 * dealing with the page cache is a little smoother */
3091 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3092 policy->l_extent.end |= ~CFS_PAGE_MASK;
3094 /* Next, search for already existing extent locks that will cover us */
3095 /* If we're trying to read, we also search for an existing PW lock. The
3096 * VFS and page cache already protect us locally, so lots of readers/
3097 * writers can share a single PW lock. */
3101 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3102 res_id, type, policy, rc, lockh, unref);
3105 osc_set_data_with_check(lockh, data, lflags);
3106 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3107 ldlm_lock_addref(lockh, LCK_PR);
3108 ldlm_lock_decref(lockh, LCK_PW);
3115 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3119 if (unlikely(mode == LCK_GROUP))
3120 ldlm_lock_decref_and_cancel(lockh, mode);
3122 ldlm_lock_decref(lockh, mode);
3127 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3128 __u32 mode, struct lustre_handle *lockh)
3131 RETURN(osc_cancel_base(lockh, mode));
3134 static int osc_cancel_unused(struct obd_export *exp,
3135 struct lov_stripe_md *lsm, int flags,
3138 struct obd_device *obd = class_exp2obd(exp);
3139 struct ldlm_res_id res_id, *resp = NULL;
3142 resp = osc_build_res_name(lsm->lsm_object_id,
3143 lsm->lsm_object_gr, &res_id);
3146 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3149 static int osc_statfs_interpret(const struct lu_env *env,
3150 struct ptlrpc_request *req,
3151 struct osc_async_args *aa, int rc)
3153 struct obd_statfs *msfs;
3159 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3161 GOTO(out, rc = -EPROTO);
3164 *aa->aa_oi->oi_osfs = *msfs;
3166 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3170 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3171 __u64 max_age, struct ptlrpc_request_set *rqset)
3173 struct ptlrpc_request *req;
3174 struct osc_async_args *aa;
3178 /* We could possibly pass max_age in the request (as an absolute
3179 * timestamp or a "seconds.usec ago") so the target can avoid doing
3180 * extra calls into the filesystem if that isn't necessary (e.g.
3181 * during mount that would help a bit). Having relative timestamps
3182 * is not so great if request processing is slow, while absolute
3183 * timestamps are not ideal because they need time synchronization. */
3184 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3188 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3190 ptlrpc_request_free(req);
3193 ptlrpc_request_set_replen(req);
3194 req->rq_request_portal = OST_CREATE_PORTAL;
3195 ptlrpc_at_set_req_timeout(req);
3197 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3198 /* procfs requests not want stat in wait for avoid deadlock */
3199 req->rq_no_resend = 1;
3200 req->rq_no_delay = 1;
3203 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3204 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3205 aa = ptlrpc_req_async_args(req);
3208 ptlrpc_set_add_req(rqset, req);
3212 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3213 __u64 max_age, __u32 flags)
3215 struct obd_statfs *msfs;
3216 struct ptlrpc_request *req;
3217 struct obd_import *imp = NULL;
3221 /*Since the request might also come from lprocfs, so we need
3222 *sync this with client_disconnect_export Bug15684*/
3223 down_read(&obd->u.cli.cl_sem);
3224 if (obd->u.cli.cl_import)
3225 imp = class_import_get(obd->u.cli.cl_import);
3226 up_read(&obd->u.cli.cl_sem);
3230 /* We could possibly pass max_age in the request (as an absolute
3231 * timestamp or a "seconds.usec ago") so the target can avoid doing
3232 * extra calls into the filesystem if that isn't necessary (e.g.
3233 * during mount that would help a bit). Having relative timestamps
3234 * is not so great if request processing is slow, while absolute
3235 * timestamps are not ideal because they need time synchronization. */
3236 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3238 class_import_put(imp);
3243 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3245 ptlrpc_request_free(req);
3248 ptlrpc_request_set_replen(req);
3249 req->rq_request_portal = OST_CREATE_PORTAL;
3250 ptlrpc_at_set_req_timeout(req);
3252 if (flags & OBD_STATFS_NODELAY) {
3253 /* procfs requests not want stat in wait for avoid deadlock */
3254 req->rq_no_resend = 1;
3255 req->rq_no_delay = 1;
3258 rc = ptlrpc_queue_wait(req);
3262 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3264 GOTO(out, rc = -EPROTO);
3271 ptlrpc_req_finished(req);
3275 /* Retrieve object striping information.
3277 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3278 * the maximum number of OST indices which will fit in the user buffer.
3279 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3281 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3283 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3284 struct lov_user_md_v3 lum, *lumk;
3285 struct lov_user_ost_data_v1 *lmm_objects;
3286 int rc = 0, lum_size;
3292 /* we only need the header part from user space to get lmm_magic and
3293 * lmm_stripe_count, (the header part is common to v1 and v3) */
3294 lum_size = sizeof(struct lov_user_md_v1);
3295 if (copy_from_user(&lum, lump, lum_size))
3298 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3299 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3302 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3303 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3304 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3305 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3307 /* we can use lov_mds_md_size() to compute lum_size
3308 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3309 if (lum.lmm_stripe_count > 0) {
3310 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3311 OBD_ALLOC(lumk, lum_size);
3315 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3316 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3318 lmm_objects = &(lumk->lmm_objects[0]);
3319 lmm_objects->l_object_id = lsm->lsm_object_id;
3321 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3325 lumk->lmm_object_id = lsm->lsm_object_id;
3326 lumk->lmm_object_gr = lsm->lsm_object_gr;
3327 lumk->lmm_stripe_count = 1;
3329 if (copy_to_user(lump, lumk, lum_size))
3333 OBD_FREE(lumk, lum_size);
3339 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3340 void *karg, void *uarg)
3342 struct obd_device *obd = exp->exp_obd;
3343 struct obd_ioctl_data *data = karg;
3347 if (!try_module_get(THIS_MODULE)) {
3348 CERROR("Can't get module. Is it alive?");
3352 case OBD_IOC_LOV_GET_CONFIG: {
3354 struct lov_desc *desc;
3355 struct obd_uuid uuid;
3359 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3360 GOTO(out, err = -EINVAL);
3362 data = (struct obd_ioctl_data *)buf;
3364 if (sizeof(*desc) > data->ioc_inllen1) {
3365 obd_ioctl_freedata(buf, len);
3366 GOTO(out, err = -EINVAL);
3369 if (data->ioc_inllen2 < sizeof(uuid)) {
3370 obd_ioctl_freedata(buf, len);
3371 GOTO(out, err = -EINVAL);
3374 desc = (struct lov_desc *)data->ioc_inlbuf1;
3375 desc->ld_tgt_count = 1;
3376 desc->ld_active_tgt_count = 1;
3377 desc->ld_default_stripe_count = 1;
3378 desc->ld_default_stripe_size = 0;
3379 desc->ld_default_stripe_offset = 0;
3380 desc->ld_pattern = 0;
3381 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3383 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3385 err = copy_to_user((void *)uarg, buf, len);
3388 obd_ioctl_freedata(buf, len);
3391 case LL_IOC_LOV_SETSTRIPE:
3392 err = obd_alloc_memmd(exp, karg);
3396 case LL_IOC_LOV_GETSTRIPE:
3397 err = osc_getstripe(karg, uarg);
3399 case OBD_IOC_CLIENT_RECOVER:
3400 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3405 case IOC_OSC_SET_ACTIVE:
3406 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3409 case OBD_IOC_POLL_QUOTACHECK:
3410 err = lquota_poll_check(quota_interface, exp,
3411 (struct if_quotacheck *)karg);
3413 case OBD_IOC_PING_TARGET:
3414 err = ptlrpc_obd_ping(obd);
3417 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3418 cmd, cfs_curproc_comm());
3419 GOTO(out, err = -ENOTTY);
3422 module_put(THIS_MODULE);
3426 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3427 void *key, __u32 *vallen, void *val,
3428 struct lov_stripe_md *lsm)
3431 if (!vallen || !val)
3434 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3435 __u32 *stripe = val;
3436 *vallen = sizeof(*stripe);
3439 } else if (KEY_IS(KEY_LAST_ID)) {
3440 struct ptlrpc_request *req;
3445 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3446 &RQF_OST_GET_INFO_LAST_ID);
3450 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3451 RCL_CLIENT, keylen);
3452 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3454 ptlrpc_request_free(req);
3458 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3459 memcpy(tmp, key, keylen);
3461 ptlrpc_request_set_replen(req);
3462 rc = ptlrpc_queue_wait(req);
3466 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3468 GOTO(out, rc = -EPROTO);
3470 *((obd_id *)val) = *reply;
3472 ptlrpc_req_finished(req);
3474 } else if (KEY_IS(KEY_FIEMAP)) {
3475 struct ptlrpc_request *req;
3476 struct ll_user_fiemap *reply;
3480 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3481 &RQF_OST_GET_INFO_FIEMAP);
3485 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3486 RCL_CLIENT, keylen);
3487 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3488 RCL_CLIENT, *vallen);
3489 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3490 RCL_SERVER, *vallen);
3492 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3494 ptlrpc_request_free(req);
3498 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3499 memcpy(tmp, key, keylen);
3500 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3501 memcpy(tmp, val, *vallen);
3503 ptlrpc_request_set_replen(req);
3504 rc = ptlrpc_queue_wait(req);
3508 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3510 GOTO(out1, rc = -EPROTO);
3512 memcpy(val, reply, *vallen);
3514 ptlrpc_req_finished(req);
3522 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3523 struct ptlrpc_request *req,
3526 struct llog_ctxt *ctxt;
3527 struct obd_import *imp = req->rq_import;
3533 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3536 rc = llog_initiator_connect(ctxt);
3538 CERROR("cannot establish connection for "
3539 "ctxt %p: %d\n", ctxt, rc);
3542 llog_ctxt_put(ctxt);
3543 spin_lock(&imp->imp_lock);
3544 imp->imp_server_timeout = 1;
3545 imp->imp_pingable = 1;
3546 spin_unlock(&imp->imp_lock);
3547 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3552 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3553 void *key, obd_count vallen, void *val,
3554 struct ptlrpc_request_set *set)
3556 struct ptlrpc_request *req;
3557 struct obd_device *obd = exp->exp_obd;
3558 struct obd_import *imp = class_exp2cliimp(exp);
3563 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3565 if (KEY_IS(KEY_NEXT_ID)) {
3566 if (vallen != sizeof(obd_id))
3570 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3571 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3572 exp->exp_obd->obd_name,
3573 obd->u.cli.cl_oscc.oscc_next_id);
3578 if (KEY_IS(KEY_UNLINKED)) {
3579 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3580 spin_lock(&oscc->oscc_lock);
3581 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3582 spin_unlock(&oscc->oscc_lock);
3586 if (KEY_IS(KEY_INIT_RECOV)) {
3587 if (vallen != sizeof(int))
3589 spin_lock(&imp->imp_lock);
3590 imp->imp_initial_recov = *(int *)val;
3591 spin_unlock(&imp->imp_lock);
3592 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3593 exp->exp_obd->obd_name,
3594 imp->imp_initial_recov);
3598 if (KEY_IS(KEY_CHECKSUM)) {
3599 if (vallen != sizeof(int))
3601 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3605 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3606 sptlrpc_conf_client_adapt(obd);
3610 if (KEY_IS(KEY_FLUSH_CTX)) {
3611 sptlrpc_import_flush_my_ctx(imp);
3618 /* We pass all other commands directly to OST. Since nobody calls osc
3619 methods directly and everybody is supposed to go through LOV, we
3620 assume lov checked invalid values for us.
3621 The only recognised values so far are evict_by_nid and mds_conn.
3622 Even if something bad goes through, we'd get a -EINVAL from OST
3626 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3630 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3631 RCL_CLIENT, keylen);
3632 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3633 RCL_CLIENT, vallen);
3634 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3636 ptlrpc_request_free(req);
3640 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3641 memcpy(tmp, key, keylen);
3642 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3643 memcpy(tmp, val, vallen);
3645 if (KEY_IS(KEY_MDS_CONN)) {
3646 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3648 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3649 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3650 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
3651 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3654 ptlrpc_request_set_replen(req);
3655 ptlrpc_set_add_req(set, req);
3656 ptlrpc_check_set(NULL, set);
3662 static struct llog_operations osc_size_repl_logops = {
3663 lop_cancel: llog_obd_repl_cancel
3666 static struct llog_operations osc_mds_ost_orig_logops;
3667 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3668 struct obd_device *tgt, int count,
3669 struct llog_catid *catid, struct obd_uuid *uuid)
3674 LASSERT(olg == &obd->obd_olg);
3675 spin_lock(&obd->obd_dev_lock);
3676 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3677 osc_mds_ost_orig_logops = llog_lvfs_ops;
3678 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3679 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3680 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3681 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3683 spin_unlock(&obd->obd_dev_lock);
3685 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3686 &catid->lci_logid, &osc_mds_ost_orig_logops);
3688 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3692 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3693 NULL, &osc_size_repl_logops);
3695 struct llog_ctxt *ctxt =
3696 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3699 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3704 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3705 obd->obd_name, tgt->obd_name, count, catid, rc);
3706 CERROR("logid "LPX64":0x%x\n",
3707 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3712 static int osc_llog_finish(struct obd_device *obd, int count)
3714 struct llog_ctxt *ctxt;
3715 int rc = 0, rc2 = 0;
3718 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3720 rc = llog_cleanup(ctxt);
3722 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3724 rc2 = llog_cleanup(ctxt);
3731 static int osc_reconnect(const struct lu_env *env,
3732 struct obd_export *exp, struct obd_device *obd,
3733 struct obd_uuid *cluuid,
3734 struct obd_connect_data *data,
3737 struct client_obd *cli = &obd->u.cli;
3739 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3742 client_obd_list_lock(&cli->cl_loi_list_lock);
3743 data->ocd_grant = cli->cl_avail_grant ?:
3744 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3745 lost_grant = cli->cl_lost_grant;
3746 cli->cl_lost_grant = 0;
3747 client_obd_list_unlock(&cli->cl_loi_list_lock);
3749 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3750 "cl_lost_grant: %ld\n", data->ocd_grant,
3751 cli->cl_avail_grant, lost_grant);
3752 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3753 " ocd_grant: %d\n", data->ocd_connect_flags,
3754 data->ocd_version, data->ocd_grant);
3760 static int osc_disconnect(struct obd_export *exp)
3762 struct obd_device *obd = class_exp2obd(exp);
3763 struct llog_ctxt *ctxt;
3766 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3768 if (obd->u.cli.cl_conn_count == 1) {
3769 /* Flush any remaining cancel messages out to the
3771 llog_sync(ctxt, exp);
3773 llog_ctxt_put(ctxt);
3775 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3779 rc = client_disconnect_export(exp);
3783 static int osc_import_event(struct obd_device *obd,
3784 struct obd_import *imp,
3785 enum obd_import_event event)
3787 struct client_obd *cli;
3791 LASSERT(imp->imp_obd == obd);
3794 case IMP_EVENT_DISCON: {
3795 /* Only do this on the MDS OSC's */
3796 if (imp->imp_server_timeout) {
3797 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3799 spin_lock(&oscc->oscc_lock);
3800 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3801 spin_unlock(&oscc->oscc_lock);
3804 client_obd_list_lock(&cli->cl_loi_list_lock);
3805 cli->cl_avail_grant = 0;
3806 cli->cl_lost_grant = 0;
3807 client_obd_list_unlock(&cli->cl_loi_list_lock);
3810 case IMP_EVENT_INACTIVE: {
3811 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3814 case IMP_EVENT_INVALIDATE: {
3815 struct ldlm_namespace *ns = obd->obd_namespace;
3819 env = cl_env_get(&refcheck);
3823 client_obd_list_lock(&cli->cl_loi_list_lock);
3824 /* all pages go to failing rpcs due to the invalid
3826 osc_check_rpcs(env, cli);
3827 client_obd_list_unlock(&cli->cl_loi_list_lock);
3829 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3830 cl_env_put(env, &refcheck);
3835 case IMP_EVENT_ACTIVE: {
3836 /* Only do this on the MDS OSC's */
3837 if (imp->imp_server_timeout) {
3838 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3840 spin_lock(&oscc->oscc_lock);
3841 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3842 spin_unlock(&oscc->oscc_lock);
3844 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3847 case IMP_EVENT_OCD: {
3848 struct obd_connect_data *ocd = &imp->imp_connect_data;
3850 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3851 osc_init_grant(&obd->u.cli, ocd);
3854 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3855 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3857 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3861 CERROR("Unknown import event %d\n", event);
3867 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3873 rc = ptlrpcd_addref();
3877 rc = client_obd_setup(obd, lcfg);
3881 struct lprocfs_static_vars lvars = { 0 };
3882 struct client_obd *cli = &obd->u.cli;
3884 lprocfs_osc_init_vars(&lvars);
3885 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3886 lproc_osc_attach_seqstat(obd);
3887 sptlrpc_lprocfs_cliobd_attach(obd);
3888 ptlrpc_lprocfs_register_obd(obd);
3892 /* We need to allocate a few requests more, because
3893 brw_interpret tries to create new requests before freeing
3894 previous ones. Ideally we want to have 2x max_rpcs_in_flight
3895 reserved, but I afraid that might be too much wasted RAM
3896 in fact, so 2 is just my guess and still should work. */
3897 cli->cl_import->imp_rq_pool =
3898 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3900 ptlrpc_add_rqs_to_pool);
3906 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3912 case OBD_CLEANUP_EARLY: {
3913 struct obd_import *imp;
3914 imp = obd->u.cli.cl_import;
3915 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3916 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3917 ptlrpc_deactivate_import(imp);
3918 spin_lock(&imp->imp_lock);
3919 imp->imp_pingable = 0;
3920 spin_unlock(&imp->imp_lock);
3923 case OBD_CLEANUP_EXPORTS: {
3924 /* If we set up but never connected, the
3925 client import will not have been cleaned. */
3926 if (obd->u.cli.cl_import) {
3927 struct obd_import *imp;
3928 down_write(&obd->u.cli.cl_sem);
3929 imp = obd->u.cli.cl_import;
3930 CDEBUG(D_CONFIG, "%s: client import never connected\n",
3932 ptlrpc_invalidate_import(imp);
3933 if (imp->imp_rq_pool) {
3934 ptlrpc_free_rq_pool(imp->imp_rq_pool);
3935 imp->imp_rq_pool = NULL;
3937 class_destroy_import(imp);
3938 up_write(&obd->u.cli.cl_sem);
3939 obd->u.cli.cl_import = NULL;
3941 rc = obd_llog_finish(obd, 0);
3943 CERROR("failed to cleanup llogging subsystems\n");
3950 int osc_cleanup(struct obd_device *obd)
3952 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3956 ptlrpc_lprocfs_unregister_obd(obd);
3957 lprocfs_obd_cleanup(obd);
3959 spin_lock(&oscc->oscc_lock);
3960 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3961 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3962 spin_unlock(&oscc->oscc_lock);
3964 /* free memory of osc quota cache */
3965 lquota_cleanup(quota_interface, obd);
3967 rc = client_obd_cleanup(obd);
3973 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3975 struct lprocfs_static_vars lvars = { 0 };
3978 lprocfs_osc_init_vars(&lvars);
3980 switch (lcfg->lcfg_command) {
3982 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3992 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3994 return osc_process_config_base(obd, buf);
3997 struct obd_ops osc_obd_ops = {
3998 .o_owner = THIS_MODULE,
3999 .o_setup = osc_setup,
4000 .o_precleanup = osc_precleanup,
4001 .o_cleanup = osc_cleanup,
4002 .o_add_conn = client_import_add_conn,
4003 .o_del_conn = client_import_del_conn,
4004 .o_connect = client_connect_import,
4005 .o_reconnect = osc_reconnect,
4006 .o_disconnect = osc_disconnect,
4007 .o_statfs = osc_statfs,
4008 .o_statfs_async = osc_statfs_async,
4009 .o_packmd = osc_packmd,
4010 .o_unpackmd = osc_unpackmd,
4011 .o_precreate = osc_precreate,
4012 .o_create = osc_create,
4013 .o_destroy = osc_destroy,
4014 .o_getattr = osc_getattr,
4015 .o_getattr_async = osc_getattr_async,
4016 .o_setattr = osc_setattr,
4017 .o_setattr_async = osc_setattr_async,
4019 .o_punch = osc_punch,
4021 .o_enqueue = osc_enqueue,
4022 .o_change_cbdata = osc_change_cbdata,
4023 .o_cancel = osc_cancel,
4024 .o_cancel_unused = osc_cancel_unused,
4025 .o_iocontrol = osc_iocontrol,
4026 .o_get_info = osc_get_info,
4027 .o_set_info_async = osc_set_info_async,
4028 .o_import_event = osc_import_event,
4029 .o_llog_init = osc_llog_init,
4030 .o_llog_finish = osc_llog_finish,
4031 .o_process_config = osc_process_config,
4034 extern struct lu_kmem_descr osc_caches[];
4035 extern spinlock_t osc_ast_guard;
4036 extern struct lock_class_key osc_ast_guard_class;
4038 int __init osc_init(void)
4040 struct lprocfs_static_vars lvars = { 0 };
4044 /* print an address of _any_ initialized kernel symbol from this
4045 * module, to allow debugging with gdb that doesn't support data
4046 * symbols from modules.*/
4047 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4049 rc = lu_kmem_init(osc_caches);
4051 lprocfs_osc_init_vars(&lvars);
4053 request_module("lquota");
4054 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4055 lquota_init(quota_interface);
4056 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4058 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4059 LUSTRE_OSC_NAME, &osc_device_type);
4061 if (quota_interface)
4062 PORTAL_SYMBOL_PUT(osc_quota_interface);
4063 lu_kmem_fini(osc_caches);
4067 spin_lock_init(&osc_ast_guard);
4068 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4074 static void /*__exit*/ osc_exit(void)
4076 lu_device_type_fini(&osc_device_type);
4078 lquota_exit(quota_interface);
4079 if (quota_interface)
4080 PORTAL_SYMBOL_PUT(osc_quota_interface);
4082 class_unregister_type(LUSTRE_OSC_NAME);
4083 lu_kmem_fini(osc_caches);
4086 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4087 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4088 MODULE_LICENSE("GPL");
4090 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);