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 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(const struct lu_env *env,
204 struct ptlrpc_request *req,
205 struct osc_async_args *aa, int rc)
207 struct ost_body *body;
213 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
215 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
218 /* This should really be sent by the OST */
219 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222 CDEBUG(D_INFO, "can't unpack ost_body\n");
224 aa->aa_oi->oi_oa->o_valid = 0;
227 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232 struct ptlrpc_request_set *set)
234 struct ptlrpc_request *req;
235 struct osc_async_args *aa;
239 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
243 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246 ptlrpc_request_free(req);
250 osc_pack_req_body(req, oinfo);
252 ptlrpc_request_set_replen(req);
253 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
255 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256 aa = ptlrpc_req_async_args(req);
259 ptlrpc_set_add_req(set, req);
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
265 struct ptlrpc_request *req;
266 struct ost_body *body;
270 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
274 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
277 ptlrpc_request_free(req);
281 osc_pack_req_body(req, oinfo);
283 ptlrpc_request_set_replen(req);
285 rc = ptlrpc_queue_wait(req);
289 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
291 GOTO(out, rc = -EPROTO);
293 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
296 /* This should really be sent by the OST */
297 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
302 ptlrpc_req_finished(req);
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307 struct obd_trans_info *oti)
309 struct ptlrpc_request *req;
310 struct ost_body *body;
314 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
316 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
320 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
321 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
323 ptlrpc_request_free(req);
327 osc_pack_req_body(req, oinfo);
329 ptlrpc_request_set_replen(req);
331 rc = ptlrpc_queue_wait(req);
335 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
337 GOTO(out, rc = -EPROTO);
339 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
343 ptlrpc_req_finished(req);
347 static int osc_setattr_interpret(const struct lu_env *env,
348 struct ptlrpc_request *req,
349 struct osc_setattr_args *sa, int rc)
351 struct ost_body *body;
357 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
359 GOTO(out, rc = -EPROTO);
361 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
363 rc = sa->sa_upcall(sa->sa_cookie, rc);
367 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
368 struct obd_trans_info *oti,
369 obd_enqueue_update_f upcall, void *cookie,
370 struct ptlrpc_request_set *rqset)
372 struct ptlrpc_request *req;
373 struct osc_setattr_args *sa;
377 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
381 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
382 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
384 ptlrpc_request_free(req);
388 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
389 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
391 osc_pack_req_body(req, oinfo);
393 ptlrpc_request_set_replen(req);
395 /* do mds to ost setattr asynchronously */
397 /* Do not wait for response. */
398 ptlrpcd_add_req(req, PSCOPE_OTHER);
400 req->rq_interpret_reply =
401 (ptlrpc_interpterer_t)osc_setattr_interpret;
403 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
404 sa = ptlrpc_req_async_args(req);
405 sa->sa_oa = oinfo->oi_oa;
406 sa->sa_upcall = upcall;
407 sa->sa_cookie = cookie;
409 if (rqset == PTLRPCD_SET)
410 ptlrpcd_add_req(req, PSCOPE_OTHER);
412 ptlrpc_set_add_req(rqset, req);
418 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
419 struct obd_trans_info *oti,
420 struct ptlrpc_request_set *rqset)
422 return osc_setattr_async_base(exp, oinfo, oti,
423 oinfo->oi_cb_up, oinfo, rqset);
426 int osc_real_create(struct obd_export *exp, struct obdo *oa,
427 struct lov_stripe_md **ea, struct obd_trans_info *oti)
429 struct ptlrpc_request *req;
430 struct ost_body *body;
431 struct lov_stripe_md *lsm;
440 rc = obd_alloc_memmd(exp, &lsm);
445 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
447 GOTO(out, rc = -ENOMEM);
449 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
451 ptlrpc_request_free(req);
455 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
457 lustre_set_wire_obdo(&body->oa, oa);
459 ptlrpc_request_set_replen(req);
461 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462 oa->o_flags == OBD_FL_DELORPHAN) {
464 "delorphan from OST integration");
465 /* Don't resend the delorphan req */
466 req->rq_no_resend = req->rq_no_delay = 1;
469 rc = ptlrpc_queue_wait(req);
473 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
475 GOTO(out_req, rc = -EPROTO);
477 lustre_get_wire_obdo(oa, &body->oa);
479 /* This should really be sent by the OST */
480 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
481 oa->o_valid |= OBD_MD_FLBLKSZ;
483 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
484 * have valid lsm_oinfo data structs, so don't go touching that.
485 * This needs to be fixed in a big way.
487 lsm->lsm_object_id = oa->o_id;
488 lsm->lsm_object_gr = oa->o_gr;
492 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
494 if (oa->o_valid & OBD_MD_FLCOOKIE) {
495 if (!oti->oti_logcookies)
496 oti_alloc_cookies(oti, 1);
497 *oti->oti_logcookies = oa->o_lcookie;
501 CDEBUG(D_HA, "transno: "LPD64"\n",
502 lustre_msg_get_transno(req->rq_repmsg));
504 ptlrpc_req_finished(req);
507 obd_free_memmd(exp, &lsm);
511 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
512 obd_enqueue_update_f upcall, void *cookie,
513 struct ptlrpc_request_set *rqset)
515 struct ptlrpc_request *req;
516 struct osc_setattr_args *sa;
517 struct ost_body *body;
521 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
525 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
526 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
528 ptlrpc_request_free(req);
531 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
532 ptlrpc_at_set_req_timeout(req);
534 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
536 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
537 osc_pack_capa(req, body, oinfo->oi_capa);
539 ptlrpc_request_set_replen(req);
542 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
543 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
544 sa = ptlrpc_req_async_args(req);
545 sa->sa_oa = oinfo->oi_oa;
546 sa->sa_upcall = upcall;
547 sa->sa_cookie = cookie;
548 if (rqset == PTLRPCD_SET)
549 ptlrpcd_add_req(req, PSCOPE_OTHER);
551 ptlrpc_set_add_req(rqset, req);
556 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
557 struct obd_trans_info *oti,
558 struct ptlrpc_request_set *rqset)
560 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
561 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
562 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
563 return osc_punch_base(exp, oinfo,
564 oinfo->oi_cb_up, oinfo, rqset);
567 static int osc_sync(struct obd_export *exp, struct obdo *oa,
568 struct lov_stripe_md *md, obd_size start, obd_size end,
571 struct ptlrpc_request *req;
572 struct ost_body *body;
577 CDEBUG(D_INFO, "oa NULL\n");
581 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
585 osc_set_capa_size(req, &RMF_CAPA1, capa);
586 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
588 ptlrpc_request_free(req);
592 /* overload the size and blocks fields in the oa with start/end */
593 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
595 lustre_set_wire_obdo(&body->oa, oa);
596 body->oa.o_size = start;
597 body->oa.o_blocks = end;
598 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
599 osc_pack_capa(req, body, capa);
601 ptlrpc_request_set_replen(req);
603 rc = ptlrpc_queue_wait(req);
607 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
609 GOTO(out, rc = -EPROTO);
611 lustre_get_wire_obdo(oa, &body->oa);
615 ptlrpc_req_finished(req);
619 /* Find and cancel locally locks matched by @mode in the resource found by
620 * @objid. Found locks are added into @cancel list. Returns the amount of
621 * locks added to @cancels list. */
622 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
624 ldlm_mode_t mode, int lock_flags)
626 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
627 struct ldlm_res_id res_id;
628 struct ldlm_resource *res;
632 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
633 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
637 LDLM_RESOURCE_ADDREF(res);
638 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
639 lock_flags, 0, NULL);
640 LDLM_RESOURCE_DELREF(res);
641 ldlm_resource_putref(res);
645 static int osc_destroy_interpret(const struct lu_env *env,
646 struct ptlrpc_request *req, void *data,
649 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
651 cfs_atomic_dec(&cli->cl_destroy_in_flight);
652 cfs_waitq_signal(&cli->cl_destroy_waitq);
656 static int osc_can_send_destroy(struct client_obd *cli)
658 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
659 cli->cl_max_rpcs_in_flight) {
660 /* The destroy request can be sent */
663 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
664 cli->cl_max_rpcs_in_flight) {
666 * The counter has been modified between the two atomic
669 cfs_waitq_signal(&cli->cl_destroy_waitq);
674 /* Destroy requests can be async always on the client, and we don't even really
675 * care about the return code since the client cannot do anything at all about
677 * When the MDS is unlinking a filename, it saves the file objects into a
678 * recovery llog, and these object records are cancelled when the OST reports
679 * they were destroyed and sync'd to disk (i.e. transaction committed).
680 * If the client dies, or the OST is down when the object should be destroyed,
681 * the records are not cancelled, and when the OST reconnects to the MDS next,
682 * it will retrieve the llog unlink logs and then sends the log cancellation
683 * cookies to the MDS after committing destroy transactions. */
684 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
685 struct lov_stripe_md *ea, struct obd_trans_info *oti,
686 struct obd_export *md_export, void *capa)
688 struct client_obd *cli = &exp->exp_obd->u.cli;
689 struct ptlrpc_request *req;
690 struct ost_body *body;
691 CFS_LIST_HEAD(cancels);
696 CDEBUG(D_INFO, "oa NULL\n");
700 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
701 LDLM_FL_DISCARD_DATA);
703 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
705 ldlm_lock_list_put(&cancels, l_bl_ast, count);
709 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
710 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
713 ptlrpc_request_free(req);
717 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
718 ptlrpc_at_set_req_timeout(req);
720 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
721 oa->o_lcookie = *oti->oti_logcookies;
722 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
724 lustre_set_wire_obdo(&body->oa, oa);
726 osc_pack_capa(req, body, (struct obd_capa *)capa);
727 ptlrpc_request_set_replen(req);
729 /* don't throttle destroy RPCs for the MDT */
730 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
731 req->rq_interpret_reply = osc_destroy_interpret;
732 if (!osc_can_send_destroy(cli)) {
733 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
737 * Wait until the number of on-going destroy RPCs drops
738 * under max_rpc_in_flight
740 l_wait_event_exclusive(cli->cl_destroy_waitq,
741 osc_can_send_destroy(cli), &lwi);
745 /* Do not wait for response */
746 ptlrpcd_add_req(req, PSCOPE_OTHER);
750 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
753 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
755 LASSERT(!(oa->o_valid & bits));
758 client_obd_list_lock(&cli->cl_loi_list_lock);
759 oa->o_dirty = cli->cl_dirty;
760 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
761 CERROR("dirty %lu - %lu > dirty_max %lu\n",
762 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
764 } else if (cfs_atomic_read(&obd_dirty_pages) -
765 cfs_atomic_read(&obd_dirty_transit_pages) >
766 obd_max_dirty_pages + 1){
767 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
768 * not covered by a lock thus they may safely race and trip
769 * this CERROR() unless we add in a small fudge factor (+1). */
770 CERROR("dirty %d - %d > system dirty_max %d\n",
771 cfs_atomic_read(&obd_dirty_pages),
772 cfs_atomic_read(&obd_dirty_transit_pages),
773 obd_max_dirty_pages);
775 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
776 CERROR("dirty %lu - dirty_max %lu too big???\n",
777 cli->cl_dirty, cli->cl_dirty_max);
780 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
781 (cli->cl_max_rpcs_in_flight + 1);
782 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
784 oa->o_grant = cli->cl_avail_grant;
785 oa->o_dropped = cli->cl_lost_grant;
786 cli->cl_lost_grant = 0;
787 client_obd_list_unlock(&cli->cl_loi_list_lock);
788 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
789 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
793 static void osc_update_next_shrink(struct client_obd *cli)
795 cli->cl_next_shrink_grant =
796 cfs_time_shift(cli->cl_grant_shrink_interval);
797 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
798 cli->cl_next_shrink_grant);
801 /* caller must hold loi_list_lock */
802 static void osc_consume_write_grant(struct client_obd *cli,
803 struct brw_page *pga)
805 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
806 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
807 cfs_atomic_inc(&obd_dirty_pages);
808 cli->cl_dirty += CFS_PAGE_SIZE;
809 cli->cl_avail_grant -= CFS_PAGE_SIZE;
810 pga->flag |= OBD_BRW_FROM_GRANT;
811 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
812 CFS_PAGE_SIZE, pga, pga->pg);
813 LASSERT(cli->cl_avail_grant >= 0);
814 osc_update_next_shrink(cli);
817 /* the companion to osc_consume_write_grant, called when a brw has completed.
818 * must be called with the loi lock held. */
819 static void osc_release_write_grant(struct client_obd *cli,
820 struct brw_page *pga, int sent)
822 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
825 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
826 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
831 pga->flag &= ~OBD_BRW_FROM_GRANT;
832 cfs_atomic_dec(&obd_dirty_pages);
833 cli->cl_dirty -= CFS_PAGE_SIZE;
834 if (pga->flag & OBD_BRW_NOCACHE) {
835 pga->flag &= ~OBD_BRW_NOCACHE;
836 cfs_atomic_dec(&obd_dirty_transit_pages);
837 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
840 cli->cl_lost_grant += CFS_PAGE_SIZE;
841 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
842 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
843 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
844 /* For short writes we shouldn't count parts of pages that
845 * span a whole block on the OST side, or our accounting goes
846 * wrong. Should match the code in filter_grant_check. */
847 int offset = pga->off & ~CFS_PAGE_MASK;
848 int count = pga->count + (offset & (blocksize - 1));
849 int end = (offset + pga->count) & (blocksize - 1);
851 count += blocksize - end;
853 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
854 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
855 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
856 cli->cl_avail_grant, cli->cl_dirty);
862 static unsigned long rpcs_in_flight(struct client_obd *cli)
864 return cli->cl_r_in_flight + cli->cl_w_in_flight;
867 /* caller must hold loi_list_lock */
868 void osc_wake_cache_waiters(struct client_obd *cli)
871 struct osc_cache_waiter *ocw;
874 cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
875 /* if we can't dirty more, we must wait until some is written */
876 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
877 (cfs_atomic_read(&obd_dirty_pages) + 1 >
878 obd_max_dirty_pages)) {
879 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
880 "osc max %ld, sys max %d\n", cli->cl_dirty,
881 cli->cl_dirty_max, obd_max_dirty_pages);
885 /* if still dirty cache but no grant wait for pending RPCs that
886 * may yet return us some grant before doing sync writes */
887 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
888 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
889 cli->cl_w_in_flight);
893 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
894 cfs_list_del_init(&ocw->ocw_entry);
895 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
896 /* no more RPCs in flight to return grant, do sync IO */
897 ocw->ocw_rc = -EDQUOT;
898 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
900 osc_consume_write_grant(cli,
901 &ocw->ocw_oap->oap_brw_page);
904 cfs_waitq_signal(&ocw->ocw_waitq);
910 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
912 client_obd_list_lock(&cli->cl_loi_list_lock);
913 cli->cl_avail_grant += grant;
914 client_obd_list_unlock(&cli->cl_loi_list_lock);
917 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
919 if (body->oa.o_valid & OBD_MD_FLGRANT) {
920 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
921 __osc_update_grant(cli, body->oa.o_grant);
925 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
926 void *key, obd_count vallen, void *val,
927 struct ptlrpc_request_set *set);
929 static int osc_shrink_grant_interpret(const struct lu_env *env,
930 struct ptlrpc_request *req,
933 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
934 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
935 struct ost_body *body;
938 __osc_update_grant(cli, oa->o_grant);
942 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
944 osc_update_grant(cli, body);
950 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
952 client_obd_list_lock(&cli->cl_loi_list_lock);
953 oa->o_grant = cli->cl_avail_grant / 4;
954 cli->cl_avail_grant -= oa->o_grant;
955 client_obd_list_unlock(&cli->cl_loi_list_lock);
956 oa->o_flags |= OBD_FL_SHRINK_GRANT;
957 osc_update_next_shrink(cli);
960 /* Shrink the current grant, either from some large amount to enough for a
961 * full set of in-flight RPCs, or if we have already shrunk to that limit
962 * then to enough for a single RPC. This avoids keeping more grant than
963 * needed, and avoids shrinking the grant piecemeal. */
964 static int osc_shrink_grant(struct client_obd *cli)
966 long target = (cli->cl_max_rpcs_in_flight + 1) *
967 cli->cl_max_pages_per_rpc;
969 client_obd_list_lock(&cli->cl_loi_list_lock);
970 if (cli->cl_avail_grant <= target)
971 target = cli->cl_max_pages_per_rpc;
972 client_obd_list_unlock(&cli->cl_loi_list_lock);
974 return osc_shrink_grant_to_target(cli, target);
977 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
980 struct ost_body *body;
983 client_obd_list_lock(&cli->cl_loi_list_lock);
984 /* Don't shrink if we are already above or below the desired limit
985 * We don't want to shrink below a single RPC, as that will negatively
986 * impact block allocation and long-term performance. */
987 if (target < cli->cl_max_pages_per_rpc)
988 target = cli->cl_max_pages_per_rpc;
990 if (target >= cli->cl_avail_grant) {
991 client_obd_list_unlock(&cli->cl_loi_list_lock);
994 client_obd_list_unlock(&cli->cl_loi_list_lock);
1000 osc_announce_cached(cli, &body->oa, 0);
1002 client_obd_list_lock(&cli->cl_loi_list_lock);
1003 body->oa.o_grant = cli->cl_avail_grant - target;
1004 cli->cl_avail_grant = target;
1005 client_obd_list_unlock(&cli->cl_loi_list_lock);
1006 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1007 osc_update_next_shrink(cli);
1009 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1010 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1011 sizeof(*body), body, NULL);
1013 __osc_update_grant(cli, body->oa.o_grant);
1018 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1019 static int osc_should_shrink_grant(struct client_obd *client)
1021 cfs_time_t time = cfs_time_current();
1022 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1023 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1024 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1025 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1028 osc_update_next_shrink(client);
1033 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1035 struct client_obd *client;
1037 cfs_list_for_each_entry(client, &item->ti_obd_list,
1038 cl_grant_shrink_list) {
1039 if (osc_should_shrink_grant(client))
1040 osc_shrink_grant(client);
1045 static int osc_add_shrink_grant(struct client_obd *client)
1049 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1051 osc_grant_shrink_grant_cb, NULL,
1052 &client->cl_grant_shrink_list);
1054 CERROR("add grant client %s error %d\n",
1055 client->cl_import->imp_obd->obd_name, rc);
1058 CDEBUG(D_CACHE, "add grant client %s \n",
1059 client->cl_import->imp_obd->obd_name);
1060 osc_update_next_shrink(client);
1064 static int osc_del_shrink_grant(struct client_obd *client)
1066 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1070 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1073 * ocd_grant is the total grant amount we're expect to hold: if we've
1074 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1075 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1077 * race is tolerable here: if we're evicted, but imp_state already
1078 * left EVICTED state, then cl_dirty must be 0 already.
1080 client_obd_list_lock(&cli->cl_loi_list_lock);
1081 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1082 cli->cl_avail_grant = ocd->ocd_grant;
1084 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1085 client_obd_list_unlock(&cli->cl_loi_list_lock);
1087 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1088 cli->cl_avail_grant, cli->cl_lost_grant);
1089 LASSERT(cli->cl_avail_grant >= 0);
1091 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1092 cfs_list_empty(&cli->cl_grant_shrink_list))
1093 osc_add_shrink_grant(cli);
1096 /* We assume that the reason this OSC got a short read is because it read
1097 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1098 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1099 * this stripe never got written at or beyond this stripe offset yet. */
1100 static void handle_short_read(int nob_read, obd_count page_count,
1101 struct brw_page **pga)
1106 /* skip bytes read OK */
1107 while (nob_read > 0) {
1108 LASSERT (page_count > 0);
1110 if (pga[i]->count > nob_read) {
1111 /* EOF inside this page */
1112 ptr = cfs_kmap(pga[i]->pg) +
1113 (pga[i]->off & ~CFS_PAGE_MASK);
1114 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1115 cfs_kunmap(pga[i]->pg);
1121 nob_read -= pga[i]->count;
1126 /* zero remaining pages */
1127 while (page_count-- > 0) {
1128 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1129 memset(ptr, 0, pga[i]->count);
1130 cfs_kunmap(pga[i]->pg);
1135 static int check_write_rcs(struct ptlrpc_request *req,
1136 int requested_nob, int niocount,
1137 obd_count page_count, struct brw_page **pga)
1142 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1143 sizeof(*remote_rcs) *
1145 if (remote_rcs == NULL) {
1146 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1150 /* return error if any niobuf was in error */
1151 for (i = 0; i < niocount; i++) {
1152 if (remote_rcs[i] < 0)
1153 return(remote_rcs[i]);
1155 if (remote_rcs[i] != 0) {
1156 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1157 i, remote_rcs[i], req);
1162 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1163 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1164 req->rq_bulk->bd_nob_transferred, requested_nob);
1171 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1173 if (p1->flag != p2->flag) {
1174 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1175 OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1177 /* warn if we try to combine flags that we don't know to be
1178 * safe to combine */
1179 if ((p1->flag & mask) != (p2->flag & mask))
1180 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1181 "same brw?\n", p1->flag, p2->flag);
1185 return (p1->off + p1->count == p2->off);
1188 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1189 struct brw_page **pga, int opc,
1190 cksum_type_t cksum_type)
1195 LASSERT (pg_count > 0);
1196 cksum = init_checksum(cksum_type);
1197 while (nob > 0 && pg_count > 0) {
1198 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1199 int off = pga[i]->off & ~CFS_PAGE_MASK;
1200 int count = pga[i]->count > nob ? nob : pga[i]->count;
1202 /* corrupt the data before we compute the checksum, to
1203 * simulate an OST->client data error */
1204 if (i == 0 && opc == OST_READ &&
1205 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1206 memcpy(ptr + off, "bad1", min(4, nob));
1207 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1208 cfs_kunmap(pga[i]->pg);
1209 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1212 nob -= pga[i]->count;
1216 /* For sending we only compute the wrong checksum instead
1217 * of corrupting the data so it is still correct on a redo */
1218 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1224 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1225 struct lov_stripe_md *lsm, obd_count page_count,
1226 struct brw_page **pga,
1227 struct ptlrpc_request **reqp,
1228 struct obd_capa *ocapa, int reserve)
1230 struct ptlrpc_request *req;
1231 struct ptlrpc_bulk_desc *desc;
1232 struct ost_body *body;
1233 struct obd_ioobj *ioobj;
1234 struct niobuf_remote *niobuf;
1235 int niocount, i, requested_nob, opc, rc;
1236 struct osc_brw_async_args *aa;
1237 struct req_capsule *pill;
1238 struct brw_page *pg_prev;
1241 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1242 RETURN(-ENOMEM); /* Recoverable */
1243 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1244 RETURN(-EINVAL); /* Fatal */
1246 if ((cmd & OBD_BRW_WRITE) != 0) {
1248 req = ptlrpc_request_alloc_pool(cli->cl_import,
1249 cli->cl_import->imp_rq_pool,
1250 &RQF_OST_BRW_WRITE);
1253 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1258 for (niocount = i = 1; i < page_count; i++) {
1259 if (!can_merge_pages(pga[i - 1], pga[i]))
1263 pill = &req->rq_pill;
1264 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1266 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1267 niocount * sizeof(*niobuf));
1268 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1270 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1272 ptlrpc_request_free(req);
1275 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1276 ptlrpc_at_set_req_timeout(req);
1278 if (opc == OST_WRITE)
1279 desc = ptlrpc_prep_bulk_imp(req, page_count,
1280 BULK_GET_SOURCE, OST_BULK_PORTAL);
1282 desc = ptlrpc_prep_bulk_imp(req, page_count,
1283 BULK_PUT_SINK, OST_BULK_PORTAL);
1286 GOTO(out, rc = -ENOMEM);
1287 /* NB request now owns desc and will free it when it gets freed */
1289 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1290 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1291 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1292 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1294 lustre_set_wire_obdo(&body->oa, oa);
1296 obdo_to_ioobj(oa, ioobj);
1297 ioobj->ioo_bufcnt = niocount;
1298 osc_pack_capa(req, body, ocapa);
1299 LASSERT (page_count > 0);
1301 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1302 struct brw_page *pg = pga[i];
1304 LASSERT(pg->count > 0);
1305 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1306 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1307 pg->off, pg->count);
1309 LASSERTF(i == 0 || pg->off > pg_prev->off,
1310 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1311 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1313 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1314 pg_prev->pg, page_private(pg_prev->pg),
1315 pg_prev->pg->index, pg_prev->off);
1317 LASSERTF(i == 0 || pg->off > pg_prev->off,
1318 "i %d p_c %u\n", i, page_count);
1320 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1321 (pg->flag & OBD_BRW_SRVLOCK));
1323 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1325 requested_nob += pg->count;
1327 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1329 niobuf->len += pg->count;
1331 niobuf->offset = pg->off;
1332 niobuf->len = pg->count;
1333 niobuf->flags = pg->flag;
1338 LASSERTF((void *)(niobuf - niocount) ==
1339 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1340 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1341 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1343 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1344 if (osc_should_shrink_grant(cli))
1345 osc_shrink_grant_local(cli, &body->oa);
1347 /* size[REQ_REC_OFF] still sizeof (*body) */
1348 if (opc == OST_WRITE) {
1349 if (unlikely(cli->cl_checksum) &&
1350 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1351 /* store cl_cksum_type in a local variable since
1352 * it can be changed via lprocfs */
1353 cksum_type_t cksum_type = cli->cl_cksum_type;
1355 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1356 oa->o_flags &= OBD_FL_LOCAL_MASK;
1357 body->oa.o_flags = 0;
1359 body->oa.o_flags |= cksum_type_pack(cksum_type);
1360 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1361 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1365 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1367 /* save this in 'oa', too, for later checking */
1368 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1369 oa->o_flags |= cksum_type_pack(cksum_type);
1371 /* clear out the checksum flag, in case this is a
1372 * resend but cl_checksum is no longer set. b=11238 */
1373 oa->o_valid &= ~OBD_MD_FLCKSUM;
1375 oa->o_cksum = body->oa.o_cksum;
1376 /* 1 RC per niobuf */
1377 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1378 sizeof(__u32) * niocount);
1380 if (unlikely(cli->cl_checksum) &&
1381 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1382 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1383 body->oa.o_flags = 0;
1384 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1385 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1388 ptlrpc_request_set_replen(req);
1390 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1391 aa = ptlrpc_req_async_args(req);
1393 aa->aa_requested_nob = requested_nob;
1394 aa->aa_nio_count = niocount;
1395 aa->aa_page_count = page_count;
1399 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1400 if (ocapa && reserve)
1401 aa->aa_ocapa = capa_get(ocapa);
1407 ptlrpc_req_finished(req);
1411 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1412 __u32 client_cksum, __u32 server_cksum, int nob,
1413 obd_count page_count, struct brw_page **pga,
1414 cksum_type_t client_cksum_type)
1418 cksum_type_t cksum_type;
1420 if (server_cksum == client_cksum) {
1421 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1425 if (oa->o_valid & OBD_MD_FLFLAGS)
1426 cksum_type = cksum_type_unpack(oa->o_flags);
1428 cksum_type = OBD_CKSUM_CRC32;
1430 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1433 if (cksum_type != client_cksum_type)
1434 msg = "the server did not use the checksum type specified in "
1435 "the original request - likely a protocol problem";
1436 else if (new_cksum == server_cksum)
1437 msg = "changed on the client after we checksummed it - "
1438 "likely false positive due to mmap IO (bug 11742)";
1439 else if (new_cksum == client_cksum)
1440 msg = "changed in transit before arrival at OST";
1442 msg = "changed in transit AND doesn't match the original - "
1443 "likely false positive due to mmap IO (bug 11742)";
1445 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1446 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1447 "["LPU64"-"LPU64"]\n",
1448 msg, libcfs_nid2str(peer->nid),
1449 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1450 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1453 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1455 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1456 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1457 "client csum now %x\n", client_cksum, client_cksum_type,
1458 server_cksum, cksum_type, new_cksum);
1462 /* Note rc enters this function as number of bytes transferred */
1463 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1465 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1466 const lnet_process_id_t *peer =
1467 &req->rq_import->imp_connection->c_peer;
1468 struct client_obd *cli = aa->aa_cli;
1469 struct ost_body *body;
1470 __u32 client_cksum = 0;
1473 if (rc < 0 && rc != -EDQUOT) {
1474 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1478 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1479 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1481 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1485 #ifdef HAVE_QUOTA_SUPPORT
1486 /* set/clear over quota flag for a uid/gid */
1487 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1488 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1489 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1491 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1492 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1494 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1502 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1503 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1505 osc_update_grant(cli, body);
1507 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1509 CERROR("Unexpected +ve rc %d\n", rc);
1512 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1514 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1517 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1518 check_write_checksum(&body->oa, peer, client_cksum,
1519 body->oa.o_cksum, aa->aa_requested_nob,
1520 aa->aa_page_count, aa->aa_ppga,
1521 cksum_type_unpack(aa->aa_oa->o_flags)))
1524 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1525 aa->aa_page_count, aa->aa_ppga);
1529 /* The rest of this function executes only for OST_READs */
1531 /* if unwrap_bulk failed, return -EAGAIN to retry */
1532 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1534 GOTO(out, rc = -EAGAIN);
1536 if (rc > aa->aa_requested_nob) {
1537 CERROR("Unexpected rc %d (%d requested)\n", rc,
1538 aa->aa_requested_nob);
1542 if (rc != req->rq_bulk->bd_nob_transferred) {
1543 CERROR ("Unexpected rc %d (%d transferred)\n",
1544 rc, req->rq_bulk->bd_nob_transferred);
1548 if (rc < aa->aa_requested_nob)
1549 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1551 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1552 static int cksum_counter;
1553 __u32 server_cksum = body->oa.o_cksum;
1556 cksum_type_t cksum_type;
1558 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1559 cksum_type = cksum_type_unpack(body->oa.o_flags);
1561 cksum_type = OBD_CKSUM_CRC32;
1562 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1563 aa->aa_ppga, OST_READ,
1566 if (peer->nid == req->rq_bulk->bd_sender) {
1570 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1573 if (server_cksum == ~0 && rc > 0) {
1574 CERROR("Protocol error: server %s set the 'checksum' "
1575 "bit, but didn't send a checksum. Not fatal, "
1576 "but please notify on http://bugzilla.lustre.org/\n",
1577 libcfs_nid2str(peer->nid));
1578 } else if (server_cksum != client_cksum) {
1579 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1580 "%s%s%s inum "LPU64"/"LPU64" object "
1581 LPU64"/"LPU64" extent "
1582 "["LPU64"-"LPU64"]\n",
1583 req->rq_import->imp_obd->obd_name,
1584 libcfs_nid2str(peer->nid),
1586 body->oa.o_valid & OBD_MD_FLFID ?
1587 body->oa.o_fid : (__u64)0,
1588 body->oa.o_valid & OBD_MD_FLFID ?
1589 body->oa.o_generation :(__u64)0,
1591 body->oa.o_valid & OBD_MD_FLGROUP ?
1592 body->oa.o_gr : (__u64)0,
1593 aa->aa_ppga[0]->off,
1594 aa->aa_ppga[aa->aa_page_count-1]->off +
1595 aa->aa_ppga[aa->aa_page_count-1]->count -
1597 CERROR("client %x, server %x, cksum_type %x\n",
1598 client_cksum, server_cksum, cksum_type);
1600 aa->aa_oa->o_cksum = client_cksum;
1604 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1607 } else if (unlikely(client_cksum)) {
1608 static int cksum_missed;
1611 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1612 CERROR("Checksum %u requested from %s but not sent\n",
1613 cksum_missed, libcfs_nid2str(peer->nid));
1619 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1624 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1625 struct lov_stripe_md *lsm,
1626 obd_count page_count, struct brw_page **pga,
1627 struct obd_capa *ocapa)
1629 struct ptlrpc_request *req;
1633 struct l_wait_info lwi;
1637 cfs_waitq_init(&waitq);
1640 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1641 page_count, pga, &req, ocapa, 0);
1645 rc = ptlrpc_queue_wait(req);
1647 if (rc == -ETIMEDOUT && req->rq_resend) {
1648 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1649 ptlrpc_req_finished(req);
1653 rc = osc_brw_fini_request(req, rc);
1655 ptlrpc_req_finished(req);
1656 if (osc_recoverable_error(rc)) {
1658 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1659 CERROR("too many resend retries, returning error\n");
1663 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1664 l_wait_event(waitq, 0, &lwi);
1672 int osc_brw_redo_request(struct ptlrpc_request *request,
1673 struct osc_brw_async_args *aa)
1675 struct ptlrpc_request *new_req;
1676 struct ptlrpc_request_set *set = request->rq_set;
1677 struct osc_brw_async_args *new_aa;
1678 struct osc_async_page *oap;
1682 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1683 CERROR("too many resend retries, returning error\n");
1687 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1689 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1690 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1691 aa->aa_cli, aa->aa_oa,
1692 NULL /* lsm unused by osc currently */,
1693 aa->aa_page_count, aa->aa_ppga,
1694 &new_req, aa->aa_ocapa, 0);
1698 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1700 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1701 if (oap->oap_request != NULL) {
1702 LASSERTF(request == oap->oap_request,
1703 "request %p != oap_request %p\n",
1704 request, oap->oap_request);
1705 if (oap->oap_interrupted) {
1706 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1707 ptlrpc_req_finished(new_req);
1712 /* New request takes over pga and oaps from old request.
1713 * Note that copying a list_head doesn't work, need to move it... */
1715 new_req->rq_interpret_reply = request->rq_interpret_reply;
1716 new_req->rq_async_args = request->rq_async_args;
1717 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1719 new_aa = ptlrpc_req_async_args(new_req);
1721 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1722 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1723 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1725 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1726 if (oap->oap_request) {
1727 ptlrpc_req_finished(oap->oap_request);
1728 oap->oap_request = ptlrpc_request_addref(new_req);
1732 new_aa->aa_ocapa = aa->aa_ocapa;
1733 aa->aa_ocapa = NULL;
1735 /* use ptlrpc_set_add_req is safe because interpret functions work
1736 * in check_set context. only one way exist with access to request
1737 * from different thread got -EINTR - this way protected with
1738 * cl_loi_list_lock */
1739 ptlrpc_set_add_req(set, new_req);
1741 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1743 DEBUG_REQ(D_INFO, new_req, "new request");
1748 * ugh, we want disk allocation on the target to happen in offset order. we'll
1749 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1750 * fine for our small page arrays and doesn't require allocation. its an
1751 * insertion sort that swaps elements that are strides apart, shrinking the
1752 * stride down until its '1' and the array is sorted.
1754 static void sort_brw_pages(struct brw_page **array, int num)
1757 struct brw_page *tmp;
1761 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1766 for (i = stride ; i < num ; i++) {
1769 while (j >= stride && array[j - stride]->off > tmp->off) {
1770 array[j] = array[j - stride];
1775 } while (stride > 1);
1778 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1784 LASSERT (pages > 0);
1785 offset = pg[i]->off & ~CFS_PAGE_MASK;
1789 if (pages == 0) /* that's all */
1792 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1793 return count; /* doesn't end on page boundary */
1796 offset = pg[i]->off & ~CFS_PAGE_MASK;
1797 if (offset != 0) /* doesn't start on page boundary */
1804 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1806 struct brw_page **ppga;
1809 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1813 for (i = 0; i < count; i++)
1818 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1820 LASSERT(ppga != NULL);
1821 OBD_FREE(ppga, sizeof(*ppga) * count);
1824 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1825 obd_count page_count, struct brw_page *pga,
1826 struct obd_trans_info *oti)
1828 struct obdo *saved_oa = NULL;
1829 struct brw_page **ppga, **orig;
1830 struct obd_import *imp = class_exp2cliimp(exp);
1831 struct client_obd *cli;
1832 int rc, page_count_orig;
1835 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1836 cli = &imp->imp_obd->u.cli;
1838 if (cmd & OBD_BRW_CHECK) {
1839 /* The caller just wants to know if there's a chance that this
1840 * I/O can succeed */
1842 if (imp->imp_invalid)
1847 /* test_brw with a failed create can trip this, maybe others. */
1848 LASSERT(cli->cl_max_pages_per_rpc);
1852 orig = ppga = osc_build_ppga(pga, page_count);
1855 page_count_orig = page_count;
1857 sort_brw_pages(ppga, page_count);
1858 while (page_count) {
1859 obd_count pages_per_brw;
1861 if (page_count > cli->cl_max_pages_per_rpc)
1862 pages_per_brw = cli->cl_max_pages_per_rpc;
1864 pages_per_brw = page_count;
1866 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1868 if (saved_oa != NULL) {
1869 /* restore previously saved oa */
1870 *oinfo->oi_oa = *saved_oa;
1871 } else if (page_count > pages_per_brw) {
1872 /* save a copy of oa (brw will clobber it) */
1873 OBDO_ALLOC(saved_oa);
1874 if (saved_oa == NULL)
1875 GOTO(out, rc = -ENOMEM);
1876 *saved_oa = *oinfo->oi_oa;
1879 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1880 pages_per_brw, ppga, oinfo->oi_capa);
1885 page_count -= pages_per_brw;
1886 ppga += pages_per_brw;
1890 osc_release_ppga(orig, page_count_orig);
1892 if (saved_oa != NULL)
1893 OBDO_FREE(saved_oa);
1898 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1899 * the dirty accounting. Writeback completes or truncate happens before
1900 * writing starts. Must be called with the loi lock held. */
1901 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1904 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1908 /* This maintains the lists of pending pages to read/write for a given object
1909 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1910 * to quickly find objects that are ready to send an RPC. */
1911 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1917 if (lop->lop_num_pending == 0)
1920 /* if we have an invalid import we want to drain the queued pages
1921 * by forcing them through rpcs that immediately fail and complete
1922 * the pages. recovery relies on this to empty the queued pages
1923 * before canceling the locks and evicting down the llite pages */
1924 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1927 /* stream rpcs in queue order as long as as there is an urgent page
1928 * queued. this is our cheap solution for good batching in the case
1929 * where writepage marks some random page in the middle of the file
1930 * as urgent because of, say, memory pressure */
1931 if (!cfs_list_empty(&lop->lop_urgent)) {
1932 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1935 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1936 optimal = cli->cl_max_pages_per_rpc;
1937 if (cmd & OBD_BRW_WRITE) {
1938 /* trigger a write rpc stream as long as there are dirtiers
1939 * waiting for space. as they're waiting, they're not going to
1940 * create more pages to coallesce with what's waiting.. */
1941 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1942 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1945 /* +16 to avoid triggering rpcs that would want to include pages
1946 * that are being queued but which can't be made ready until
1947 * the queuer finishes with the page. this is a wart for
1948 * llite::commit_write() */
1951 if (lop->lop_num_pending >= optimal)
1957 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1959 struct osc_async_page *oap;
1962 if (cfs_list_empty(&lop->lop_urgent))
1965 oap = cfs_list_entry(lop->lop_urgent.next,
1966 struct osc_async_page, oap_urgent_item);
1968 if (oap->oap_async_flags & ASYNC_HP) {
1969 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1976 static void on_list(cfs_list_t *item, cfs_list_t *list,
1979 if (cfs_list_empty(item) && should_be_on)
1980 cfs_list_add_tail(item, list);
1981 else if (!cfs_list_empty(item) && !should_be_on)
1982 cfs_list_del_init(item);
1985 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1986 * can find pages to build into rpcs quickly */
1987 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1989 if (lop_makes_hprpc(&loi->loi_write_lop) ||
1990 lop_makes_hprpc(&loi->loi_read_lop)) {
1992 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1993 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1995 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1996 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1997 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1998 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2001 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2002 loi->loi_write_lop.lop_num_pending);
2004 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2005 loi->loi_read_lop.lop_num_pending);
2008 static void lop_update_pending(struct client_obd *cli,
2009 struct loi_oap_pages *lop, int cmd, int delta)
2011 lop->lop_num_pending += delta;
2012 if (cmd & OBD_BRW_WRITE)
2013 cli->cl_pending_w_pages += delta;
2015 cli->cl_pending_r_pages += delta;
2019 * this is called when a sync waiter receives an interruption. Its job is to
2020 * get the caller woken as soon as possible. If its page hasn't been put in an
2021 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2022 * desiring interruption which will forcefully complete the rpc once the rpc
2025 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2027 struct loi_oap_pages *lop;
2028 struct lov_oinfo *loi;
2032 LASSERT(!oap->oap_interrupted);
2033 oap->oap_interrupted = 1;
2035 /* ok, it's been put in an rpc. only one oap gets a request reference */
2036 if (oap->oap_request != NULL) {
2037 ptlrpc_mark_interrupted(oap->oap_request);
2038 ptlrpcd_wake(oap->oap_request);
2039 ptlrpc_req_finished(oap->oap_request);
2040 oap->oap_request = NULL;
2044 * page completion may be called only if ->cpo_prep() method was
2045 * executed by osc_io_submit(), that also adds page the to pending list
2047 if (!cfs_list_empty(&oap->oap_pending_item)) {
2048 cfs_list_del_init(&oap->oap_pending_item);
2049 cfs_list_del_init(&oap->oap_urgent_item);
2052 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2053 &loi->loi_write_lop : &loi->loi_read_lop;
2054 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2055 loi_list_maint(oap->oap_cli, oap->oap_loi);
2056 rc = oap->oap_caller_ops->ap_completion(env,
2057 oap->oap_caller_data,
2058 oap->oap_cmd, NULL, -EINTR);
2064 /* this is trying to propogate async writeback errors back up to the
2065 * application. As an async write fails we record the error code for later if
2066 * the app does an fsync. As long as errors persist we force future rpcs to be
2067 * sync so that the app can get a sync error and break the cycle of queueing
2068 * pages for which writeback will fail. */
2069 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2076 ar->ar_force_sync = 1;
2077 ar->ar_min_xid = ptlrpc_sample_next_xid();
2082 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2083 ar->ar_force_sync = 0;
2086 void osc_oap_to_pending(struct osc_async_page *oap)
2088 struct loi_oap_pages *lop;
2090 if (oap->oap_cmd & OBD_BRW_WRITE)
2091 lop = &oap->oap_loi->loi_write_lop;
2093 lop = &oap->oap_loi->loi_read_lop;
2095 if (oap->oap_async_flags & ASYNC_HP)
2096 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2097 else if (oap->oap_async_flags & ASYNC_URGENT)
2098 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2099 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2100 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2103 /* this must be called holding the loi list lock to give coverage to exit_cache,
2104 * async_flag maintenance, and oap_request */
2105 static void osc_ap_completion(const struct lu_env *env,
2106 struct client_obd *cli, struct obdo *oa,
2107 struct osc_async_page *oap, int sent, int rc)
2112 if (oap->oap_request != NULL) {
2113 xid = ptlrpc_req_xid(oap->oap_request);
2114 ptlrpc_req_finished(oap->oap_request);
2115 oap->oap_request = NULL;
2118 cfs_spin_lock(&oap->oap_lock);
2119 oap->oap_async_flags = 0;
2120 cfs_spin_unlock(&oap->oap_lock);
2121 oap->oap_interrupted = 0;
2123 if (oap->oap_cmd & OBD_BRW_WRITE) {
2124 osc_process_ar(&cli->cl_ar, xid, rc);
2125 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2128 if (rc == 0 && oa != NULL) {
2129 if (oa->o_valid & OBD_MD_FLBLOCKS)
2130 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2131 if (oa->o_valid & OBD_MD_FLMTIME)
2132 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2133 if (oa->o_valid & OBD_MD_FLATIME)
2134 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2135 if (oa->o_valid & OBD_MD_FLCTIME)
2136 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2139 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2140 oap->oap_cmd, oa, rc);
2142 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2143 * I/O on the page could start, but OSC calls it under lock
2144 * and thus we can add oap back to pending safely */
2146 /* upper layer wants to leave the page on pending queue */
2147 osc_oap_to_pending(oap);
2149 osc_exit_cache(cli, oap, sent);
2153 static int brw_interpret(const struct lu_env *env,
2154 struct ptlrpc_request *req, void *data, int rc)
2156 struct osc_brw_async_args *aa = data;
2157 struct client_obd *cli;
2161 rc = osc_brw_fini_request(req, rc);
2162 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2163 if (osc_recoverable_error(rc)) {
2164 rc = osc_brw_redo_request(req, aa);
2170 capa_put(aa->aa_ocapa);
2171 aa->aa_ocapa = NULL;
2176 client_obd_list_lock(&cli->cl_loi_list_lock);
2178 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2179 * is called so we know whether to go to sync BRWs or wait for more
2180 * RPCs to complete */
2181 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2182 cli->cl_w_in_flight--;
2184 cli->cl_r_in_flight--;
2186 async = cfs_list_empty(&aa->aa_oaps);
2187 if (!async) { /* from osc_send_oap_rpc() */
2188 struct osc_async_page *oap, *tmp;
2189 /* the caller may re-use the oap after the completion call so
2190 * we need to clean it up a little */
2191 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2193 cfs_list_del_init(&oap->oap_rpc_item);
2194 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2196 OBDO_FREE(aa->aa_oa);
2197 } else { /* from async_internal() */
2199 for (i = 0; i < aa->aa_page_count; i++)
2200 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2202 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2203 OBDO_FREE(aa->aa_oa);
2205 osc_wake_cache_waiters(cli);
2206 osc_check_rpcs(env, cli);
2207 client_obd_list_unlock(&cli->cl_loi_list_lock);
2209 cl_req_completion(env, aa->aa_clerq, rc);
2210 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2214 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2215 struct client_obd *cli,
2216 cfs_list_t *rpc_list,
2217 int page_count, int cmd)
2219 struct ptlrpc_request *req;
2220 struct brw_page **pga = NULL;
2221 struct osc_brw_async_args *aa;
2222 struct obdo *oa = NULL;
2223 const struct obd_async_page_ops *ops = NULL;
2224 void *caller_data = NULL;
2225 struct osc_async_page *oap;
2226 struct osc_async_page *tmp;
2227 struct ost_body *body;
2228 struct cl_req *clerq = NULL;
2229 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2230 struct ldlm_lock *lock = NULL;
2231 struct cl_req_attr crattr;
2235 LASSERT(!cfs_list_empty(rpc_list));
2237 memset(&crattr, 0, sizeof crattr);
2238 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2240 GOTO(out, req = ERR_PTR(-ENOMEM));
2244 GOTO(out, req = ERR_PTR(-ENOMEM));
2247 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2248 struct cl_page *page = osc_oap2cl_page(oap);
2250 ops = oap->oap_caller_ops;
2251 caller_data = oap->oap_caller_data;
2253 clerq = cl_req_alloc(env, page, crt,
2254 1 /* only 1-object rpcs for
2257 GOTO(out, req = (void *)clerq);
2258 lock = oap->oap_ldlm_lock;
2260 pga[i] = &oap->oap_brw_page;
2261 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2262 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2263 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2265 cl_req_page_add(env, clerq, page);
2268 /* always get the data for the obdo for the rpc */
2269 LASSERT(ops != NULL);
2271 crattr.cra_capa = NULL;
2272 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2274 oa->o_handle = lock->l_remote_handle;
2275 oa->o_valid |= OBD_MD_FLHANDLE;
2278 rc = cl_req_prep(env, clerq);
2280 CERROR("cl_req_prep failed: %d\n", rc);
2281 GOTO(out, req = ERR_PTR(rc));
2284 sort_brw_pages(pga, page_count);
2285 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2286 pga, &req, crattr.cra_capa, 1);
2288 CERROR("prep_req failed: %d\n", rc);
2289 GOTO(out, req = ERR_PTR(rc));
2292 /* Need to update the timestamps after the request is built in case
2293 * we race with setattr (locally or in queue at OST). If OST gets
2294 * later setattr before earlier BRW (as determined by the request xid),
2295 * the OST will not use BRW timestamps. Sadly, there is no obvious
2296 * way to do this in a single call. bug 10150 */
2297 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2298 cl_req_attr_set(env, clerq, &crattr,
2299 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2301 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2302 aa = ptlrpc_req_async_args(req);
2303 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2304 cfs_list_splice(rpc_list, &aa->aa_oaps);
2305 CFS_INIT_LIST_HEAD(rpc_list);
2306 aa->aa_clerq = clerq;
2308 capa_put(crattr.cra_capa);
2313 OBD_FREE(pga, sizeof(*pga) * page_count);
2314 /* this should happen rarely and is pretty bad, it makes the
2315 * pending list not follow the dirty order */
2316 client_obd_list_lock(&cli->cl_loi_list_lock);
2317 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2318 cfs_list_del_init(&oap->oap_rpc_item);
2320 /* queued sync pages can be torn down while the pages
2321 * were between the pending list and the rpc */
2322 if (oap->oap_interrupted) {
2323 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2324 osc_ap_completion(env, cli, NULL, oap, 0,
2328 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2330 if (clerq && !IS_ERR(clerq))
2331 cl_req_completion(env, clerq, PTR_ERR(req));
2337 * prepare pages for ASYNC io and put pages in send queue.
2339 * \param cmd OBD_BRW_* macroses
2340 * \param lop pending pages
2342 * \return zero if pages successfully add to send queue.
2343 * \return not zere if error occurring.
2346 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2347 struct lov_oinfo *loi,
2348 int cmd, struct loi_oap_pages *lop)
2350 struct ptlrpc_request *req;
2351 obd_count page_count = 0;
2352 struct osc_async_page *oap = NULL, *tmp;
2353 struct osc_brw_async_args *aa;
2354 const struct obd_async_page_ops *ops;
2355 CFS_LIST_HEAD(rpc_list);
2356 CFS_LIST_HEAD(tmp_list);
2357 unsigned int ending_offset;
2358 unsigned starting_offset = 0;
2360 struct cl_object *clob = NULL;
2363 /* ASYNC_HP pages first. At present, when the lock the pages is
2364 * to be canceled, the pages covered by the lock will be sent out
2365 * with ASYNC_HP. We have to send out them as soon as possible. */
2366 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2367 if (oap->oap_async_flags & ASYNC_HP)
2368 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2370 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2371 if (++page_count >= cli->cl_max_pages_per_rpc)
2375 cfs_list_splice(&tmp_list, &lop->lop_pending);
2378 /* first we find the pages we're allowed to work with */
2379 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2381 ops = oap->oap_caller_ops;
2383 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2384 "magic 0x%x\n", oap, oap->oap_magic);
2387 /* pin object in memory, so that completion call-backs
2388 * can be safely called under client_obd_list lock. */
2389 clob = osc_oap2cl_page(oap)->cp_obj;
2390 cl_object_get(clob);
2393 if (page_count != 0 &&
2394 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2395 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2396 " oap %p, page %p, srvlock %u\n",
2397 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2401 /* If there is a gap at the start of this page, it can't merge
2402 * with any previous page, so we'll hand the network a
2403 * "fragmented" page array that it can't transfer in 1 RDMA */
2404 if (page_count != 0 && oap->oap_page_off != 0)
2407 /* in llite being 'ready' equates to the page being locked
2408 * until completion unlocks it. commit_write submits a page
2409 * as not ready because its unlock will happen unconditionally
2410 * as the call returns. if we race with commit_write giving
2411 * us that page we dont' want to create a hole in the page
2412 * stream, so we stop and leave the rpc to be fired by
2413 * another dirtier or kupdated interval (the not ready page
2414 * will still be on the dirty list). we could call in
2415 * at the end of ll_file_write to process the queue again. */
2416 if (!(oap->oap_async_flags & ASYNC_READY)) {
2417 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2420 CDEBUG(D_INODE, "oap %p page %p returned %d "
2421 "instead of ready\n", oap,
2425 /* llite is telling us that the page is still
2426 * in commit_write and that we should try
2427 * and put it in an rpc again later. we
2428 * break out of the loop so we don't create
2429 * a hole in the sequence of pages in the rpc
2434 /* the io isn't needed.. tell the checks
2435 * below to complete the rpc with EINTR */
2436 cfs_spin_lock(&oap->oap_lock);
2437 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2438 cfs_spin_unlock(&oap->oap_lock);
2439 oap->oap_count = -EINTR;
2442 cfs_spin_lock(&oap->oap_lock);
2443 oap->oap_async_flags |= ASYNC_READY;
2444 cfs_spin_unlock(&oap->oap_lock);
2447 LASSERTF(0, "oap %p page %p returned %d "
2448 "from make_ready\n", oap,
2456 * Page submitted for IO has to be locked. Either by
2457 * ->ap_make_ready() or by higher layers.
2459 #if defined(__KERNEL__) && defined(__linux__)
2461 struct cl_page *page;
2463 page = osc_oap2cl_page(oap);
2465 if (page->cp_type == CPT_CACHEABLE &&
2466 !(PageLocked(oap->oap_page) &&
2467 (CheckWriteback(oap->oap_page, cmd)))) {
2468 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2470 (long)oap->oap_page->flags,
2471 oap->oap_async_flags);
2477 /* take the page out of our book-keeping */
2478 cfs_list_del_init(&oap->oap_pending_item);
2479 lop_update_pending(cli, lop, cmd, -1);
2480 cfs_list_del_init(&oap->oap_urgent_item);
2482 if (page_count == 0)
2483 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2484 (PTLRPC_MAX_BRW_SIZE - 1);
2486 /* ask the caller for the size of the io as the rpc leaves. */
2487 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2489 ops->ap_refresh_count(env, oap->oap_caller_data,
2491 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2493 if (oap->oap_count <= 0) {
2494 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2496 osc_ap_completion(env, cli, NULL,
2497 oap, 0, oap->oap_count);
2501 /* now put the page back in our accounting */
2502 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2503 if (page_count == 0)
2504 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2505 if (++page_count >= cli->cl_max_pages_per_rpc)
2508 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2509 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2510 * have the same alignment as the initial writes that allocated
2511 * extents on the server. */
2512 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2513 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2514 if (ending_offset == 0)
2517 /* If there is a gap at the end of this page, it can't merge
2518 * with any subsequent pages, so we'll hand the network a
2519 * "fragmented" page array that it can't transfer in 1 RDMA */
2520 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2524 osc_wake_cache_waiters(cli);
2526 loi_list_maint(cli, loi);
2528 client_obd_list_unlock(&cli->cl_loi_list_lock);
2531 cl_object_put(env, clob);
2533 if (page_count == 0) {
2534 client_obd_list_lock(&cli->cl_loi_list_lock);
2538 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2540 LASSERT(cfs_list_empty(&rpc_list));
2541 loi_list_maint(cli, loi);
2542 RETURN(PTR_ERR(req));
2545 aa = ptlrpc_req_async_args(req);
2547 if (cmd == OBD_BRW_READ) {
2548 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2549 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2550 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2551 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2553 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2554 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2555 cli->cl_w_in_flight);
2556 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2557 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2559 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2561 client_obd_list_lock(&cli->cl_loi_list_lock);
2563 if (cmd == OBD_BRW_READ)
2564 cli->cl_r_in_flight++;
2566 cli->cl_w_in_flight++;
2568 /* queued sync pages can be torn down while the pages
2569 * were between the pending list and the rpc */
2571 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2572 /* only one oap gets a request reference */
2575 if (oap->oap_interrupted && !req->rq_intr) {
2576 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2578 ptlrpc_mark_interrupted(req);
2582 tmp->oap_request = ptlrpc_request_addref(req);
2584 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2585 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2587 req->rq_interpret_reply = brw_interpret;
2588 ptlrpcd_add_req(req, PSCOPE_BRW);
2592 #define LOI_DEBUG(LOI, STR, args...) \
2593 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2594 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2595 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2596 (LOI)->loi_write_lop.lop_num_pending, \
2597 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2598 (LOI)->loi_read_lop.lop_num_pending, \
2599 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2602 /* This is called by osc_check_rpcs() to find which objects have pages that
2603 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2604 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2608 /* First return objects that have blocked locks so that they
2609 * will be flushed quickly and other clients can get the lock,
2610 * then objects which have pages ready to be stuffed into RPCs */
2611 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2612 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2613 struct lov_oinfo, loi_hp_ready_item));
2614 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2615 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2616 struct lov_oinfo, loi_ready_item));
2618 /* then if we have cache waiters, return all objects with queued
2619 * writes. This is especially important when many small files
2620 * have filled up the cache and not been fired into rpcs because
2621 * they don't pass the nr_pending/object threshhold */
2622 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2623 !cfs_list_empty(&cli->cl_loi_write_list))
2624 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2625 struct lov_oinfo, loi_write_item));
2627 /* then return all queued objects when we have an invalid import
2628 * so that they get flushed */
2629 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2630 if (!cfs_list_empty(&cli->cl_loi_write_list))
2631 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2634 if (!cfs_list_empty(&cli->cl_loi_read_list))
2635 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2636 struct lov_oinfo, loi_read_item));
2641 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2643 struct osc_async_page *oap;
2646 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2647 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2648 struct osc_async_page, oap_urgent_item);
2649 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2652 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2653 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2654 struct osc_async_page, oap_urgent_item);
2655 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2658 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2661 /* called with the loi list lock held */
2662 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2664 struct lov_oinfo *loi;
2665 int rc = 0, race_counter = 0;
2668 while ((loi = osc_next_loi(cli)) != NULL) {
2669 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2671 if (osc_max_rpc_in_flight(cli, loi))
2674 /* attempt some read/write balancing by alternating between
2675 * reads and writes in an object. The makes_rpc checks here
2676 * would be redundant if we were getting read/write work items
2677 * instead of objects. we don't want send_oap_rpc to drain a
2678 * partial read pending queue when we're given this object to
2679 * do io on writes while there are cache waiters */
2680 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2681 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2682 &loi->loi_write_lop);
2684 CERROR("Write request failed with %d\n", rc);
2686 /* osc_send_oap_rpc failed, mostly because of
2689 * It can't break here, because if:
2690 * - a page was submitted by osc_io_submit, so
2692 * - no request in flight
2693 * - no subsequent request
2694 * The system will be in live-lock state,
2695 * because there is no chance to call
2696 * osc_io_unplug() and osc_check_rpcs() any
2697 * more. pdflush can't help in this case,
2698 * because it might be blocked at grabbing
2699 * the page lock as we mentioned.
2701 * Anyway, continue to drain pages. */
2710 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2711 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2712 &loi->loi_read_lop);
2714 CERROR("Read request failed with %d\n", rc);
2722 /* attempt some inter-object balancing by issueing rpcs
2723 * for each object in turn */
2724 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2725 cfs_list_del_init(&loi->loi_hp_ready_item);
2726 if (!cfs_list_empty(&loi->loi_ready_item))
2727 cfs_list_del_init(&loi->loi_ready_item);
2728 if (!cfs_list_empty(&loi->loi_write_item))
2729 cfs_list_del_init(&loi->loi_write_item);
2730 if (!cfs_list_empty(&loi->loi_read_item))
2731 cfs_list_del_init(&loi->loi_read_item);
2733 loi_list_maint(cli, loi);
2735 /* send_oap_rpc fails with 0 when make_ready tells it to
2736 * back off. llite's make_ready does this when it tries
2737 * to lock a page queued for write that is already locked.
2738 * we want to try sending rpcs from many objects, but we
2739 * don't want to spin failing with 0. */
2740 if (race_counter == 10)
2746 /* we're trying to queue a page in the osc so we're subject to the
2747 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2748 * If the osc's queued pages are already at that limit, then we want to sleep
2749 * until there is space in the osc's queue for us. We also may be waiting for
2750 * write credits from the OST if there are RPCs in flight that may return some
2751 * before we fall back to sync writes.
2753 * We need this know our allocation was granted in the presence of signals */
2754 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2758 client_obd_list_lock(&cli->cl_loi_list_lock);
2759 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2760 client_obd_list_unlock(&cli->cl_loi_list_lock);
2765 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2768 int osc_enter_cache_try(const struct lu_env *env,
2769 struct client_obd *cli, struct lov_oinfo *loi,
2770 struct osc_async_page *oap, int transient)
2774 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2776 osc_consume_write_grant(cli, &oap->oap_brw_page);
2778 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2779 cfs_atomic_inc(&obd_dirty_transit_pages);
2780 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2786 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2787 * grant or cache space. */
2788 static int osc_enter_cache(const struct lu_env *env,
2789 struct client_obd *cli, struct lov_oinfo *loi,
2790 struct osc_async_page *oap)
2792 struct osc_cache_waiter ocw;
2793 struct l_wait_info lwi = { 0 };
2797 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2798 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2799 cli->cl_dirty_max, obd_max_dirty_pages,
2800 cli->cl_lost_grant, cli->cl_avail_grant);
2802 /* force the caller to try sync io. this can jump the list
2803 * of queued writes and create a discontiguous rpc stream */
2804 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2805 loi->loi_ar.ar_force_sync)
2808 /* Hopefully normal case - cache space and write credits available */
2809 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2810 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2811 osc_enter_cache_try(env, cli, loi, oap, 0))
2814 /* It is safe to block as a cache waiter as long as there is grant
2815 * space available or the hope of additional grant being returned
2816 * when an in flight write completes. Using the write back cache
2817 * if possible is preferable to sending the data synchronously
2818 * because write pages can then be merged in to large requests.
2819 * The addition of this cache waiter will causing pending write
2820 * pages to be sent immediately. */
2821 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2822 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2823 cfs_waitq_init(&ocw.ocw_waitq);
2827 loi_list_maint(cli, loi);
2828 osc_check_rpcs(env, cli);
2829 client_obd_list_unlock(&cli->cl_loi_list_lock);
2831 CDEBUG(D_CACHE, "sleeping for cache space\n");
2832 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2834 client_obd_list_lock(&cli->cl_loi_list_lock);
2835 if (!cfs_list_empty(&ocw.ocw_entry)) {
2836 cfs_list_del(&ocw.ocw_entry);
2846 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2847 struct lov_oinfo *loi, cfs_page_t *page,
2848 obd_off offset, const struct obd_async_page_ops *ops,
2849 void *data, void **res, int nocache,
2850 struct lustre_handle *lockh)
2852 struct osc_async_page *oap;
2857 return cfs_size_round(sizeof(*oap));
2860 oap->oap_magic = OAP_MAGIC;
2861 oap->oap_cli = &exp->exp_obd->u.cli;
2864 oap->oap_caller_ops = ops;
2865 oap->oap_caller_data = data;
2867 oap->oap_page = page;
2868 oap->oap_obj_off = offset;
2869 if (!client_is_remote(exp) &&
2870 cfs_capable(CFS_CAP_SYS_RESOURCE))
2871 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2873 LASSERT(!(offset & ~CFS_PAGE_MASK));
2875 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2876 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2877 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2878 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2880 cfs_spin_lock_init(&oap->oap_lock);
2881 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2885 struct osc_async_page *oap_from_cookie(void *cookie)
2887 struct osc_async_page *oap = cookie;
2888 if (oap->oap_magic != OAP_MAGIC)
2889 return ERR_PTR(-EINVAL);
2893 int osc_queue_async_io(const struct lu_env *env,
2894 struct obd_export *exp, struct lov_stripe_md *lsm,
2895 struct lov_oinfo *loi, void *cookie,
2896 int cmd, obd_off off, int count,
2897 obd_flag brw_flags, enum async_flags async_flags)
2899 struct client_obd *cli = &exp->exp_obd->u.cli;
2900 struct osc_async_page *oap;
2904 oap = oap_from_cookie(cookie);
2906 RETURN(PTR_ERR(oap));
2908 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2911 if (!cfs_list_empty(&oap->oap_pending_item) ||
2912 !cfs_list_empty(&oap->oap_urgent_item) ||
2913 !cfs_list_empty(&oap->oap_rpc_item))
2916 /* check if the file's owner/group is over quota */
2917 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2918 struct cl_object *obj;
2919 struct cl_attr attr; /* XXX put attr into thread info */
2920 unsigned int qid[MAXQUOTAS];
2922 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2924 cl_object_attr_lock(obj);
2925 rc = cl_object_attr_get(env, obj, &attr);
2926 cl_object_attr_unlock(obj);
2928 qid[USRQUOTA] = attr.cat_uid;
2929 qid[GRPQUOTA] = attr.cat_gid;
2931 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2938 loi = lsm->lsm_oinfo[0];
2940 client_obd_list_lock(&cli->cl_loi_list_lock);
2942 LASSERT(off + count <= CFS_PAGE_SIZE);
2944 oap->oap_page_off = off;
2945 oap->oap_count = count;
2946 oap->oap_brw_flags = brw_flags;
2947 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2948 if (libcfs_memory_pressure_get())
2949 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2950 cfs_spin_lock(&oap->oap_lock);
2951 oap->oap_async_flags = async_flags;
2952 cfs_spin_unlock(&oap->oap_lock);
2954 if (cmd & OBD_BRW_WRITE) {
2955 rc = osc_enter_cache(env, cli, loi, oap);
2957 client_obd_list_unlock(&cli->cl_loi_list_lock);
2962 osc_oap_to_pending(oap);
2963 loi_list_maint(cli, loi);
2965 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2968 osc_check_rpcs(env, cli);
2969 client_obd_list_unlock(&cli->cl_loi_list_lock);
2974 /* aka (~was & now & flag), but this is more clear :) */
2975 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2977 int osc_set_async_flags_base(struct client_obd *cli,
2978 struct lov_oinfo *loi, struct osc_async_page *oap,
2979 obd_flag async_flags)
2981 struct loi_oap_pages *lop;
2985 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
2987 if (oap->oap_cmd & OBD_BRW_WRITE) {
2988 lop = &loi->loi_write_lop;
2990 lop = &loi->loi_read_lop;
2993 if ((oap->oap_async_flags & async_flags) == async_flags)
2996 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2997 flags |= ASYNC_READY;
2999 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3000 cfs_list_empty(&oap->oap_rpc_item)) {
3001 if (oap->oap_async_flags & ASYNC_HP)
3002 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3004 cfs_list_add_tail(&oap->oap_urgent_item,
3006 flags |= ASYNC_URGENT;
3007 loi_list_maint(cli, loi);
3009 cfs_spin_lock(&oap->oap_lock);
3010 oap->oap_async_flags |= flags;
3011 cfs_spin_unlock(&oap->oap_lock);
3013 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3014 oap->oap_async_flags);
3018 int osc_teardown_async_page(struct obd_export *exp,
3019 struct lov_stripe_md *lsm,
3020 struct lov_oinfo *loi, void *cookie)
3022 struct client_obd *cli = &exp->exp_obd->u.cli;
3023 struct loi_oap_pages *lop;
3024 struct osc_async_page *oap;
3028 oap = oap_from_cookie(cookie);
3030 RETURN(PTR_ERR(oap));
3033 loi = lsm->lsm_oinfo[0];
3035 if (oap->oap_cmd & OBD_BRW_WRITE) {
3036 lop = &loi->loi_write_lop;
3038 lop = &loi->loi_read_lop;
3041 client_obd_list_lock(&cli->cl_loi_list_lock);
3043 if (!cfs_list_empty(&oap->oap_rpc_item))
3044 GOTO(out, rc = -EBUSY);
3046 osc_exit_cache(cli, oap, 0);
3047 osc_wake_cache_waiters(cli);
3049 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3050 cfs_list_del_init(&oap->oap_urgent_item);
3051 cfs_spin_lock(&oap->oap_lock);
3052 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3053 cfs_spin_unlock(&oap->oap_lock);
3055 if (!cfs_list_empty(&oap->oap_pending_item)) {
3056 cfs_list_del_init(&oap->oap_pending_item);
3057 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3059 loi_list_maint(cli, loi);
3060 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3062 client_obd_list_unlock(&cli->cl_loi_list_lock);
3066 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3067 struct ldlm_enqueue_info *einfo,
3070 void *data = einfo->ei_cbdata;
3072 LASSERT(lock != NULL);
3073 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3074 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3075 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3076 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3078 lock_res_and_lock(lock);
3079 cfs_spin_lock(&osc_ast_guard);
3080 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3081 lock->l_ast_data = data;
3082 cfs_spin_unlock(&osc_ast_guard);
3083 unlock_res_and_lock(lock);
3086 static void osc_set_data_with_check(struct lustre_handle *lockh,
3087 struct ldlm_enqueue_info *einfo,
3090 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3093 osc_set_lock_data_with_check(lock, einfo, flags);
3094 LDLM_LOCK_PUT(lock);
3096 CERROR("lockh %p, data %p - client evicted?\n",
3097 lockh, einfo->ei_cbdata);
3100 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3101 ldlm_iterator_t replace, void *data)
3103 struct ldlm_res_id res_id;
3104 struct obd_device *obd = class_exp2obd(exp);
3106 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3107 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3111 /* find any ldlm lock of the inode in osc
3115 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3116 ldlm_iterator_t replace, void *data)
3118 struct ldlm_res_id res_id;
3119 struct obd_device *obd = class_exp2obd(exp);
3122 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3123 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3124 if (rc == LDLM_ITER_STOP)
3126 if (rc == LDLM_ITER_CONTINUE)
3131 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3132 obd_enqueue_update_f upcall, void *cookie,
3135 int intent = *flags & LDLM_FL_HAS_INTENT;
3139 /* The request was created before ldlm_cli_enqueue call. */
3140 if (rc == ELDLM_LOCK_ABORTED) {
3141 struct ldlm_reply *rep;
3142 rep = req_capsule_server_get(&req->rq_pill,
3145 LASSERT(rep != NULL);
3146 if (rep->lock_policy_res1)
3147 rc = rep->lock_policy_res1;
3151 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3152 *flags |= LDLM_FL_LVB_READY;
3153 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3154 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3157 /* Call the update callback. */
3158 rc = (*upcall)(cookie, rc);
3162 static int osc_enqueue_interpret(const struct lu_env *env,
3163 struct ptlrpc_request *req,
3164 struct osc_enqueue_args *aa, int rc)
3166 struct ldlm_lock *lock;
3167 struct lustre_handle handle;
3170 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3171 * might be freed anytime after lock upcall has been called. */
3172 lustre_handle_copy(&handle, aa->oa_lockh);
3173 mode = aa->oa_ei->ei_mode;
3175 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3177 lock = ldlm_handle2lock(&handle);
3179 /* Take an additional reference so that a blocking AST that
3180 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3181 * to arrive after an upcall has been executed by
3182 * osc_enqueue_fini(). */
3183 ldlm_lock_addref(&handle, mode);
3185 /* Complete obtaining the lock procedure. */
3186 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3187 mode, aa->oa_flags, aa->oa_lvb,
3188 sizeof(*aa->oa_lvb), &handle, rc);
3189 /* Complete osc stuff. */
3190 rc = osc_enqueue_fini(req, aa->oa_lvb,
3191 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3193 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3195 /* Release the lock for async request. */
3196 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3198 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3199 * not already released by
3200 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3202 ldlm_lock_decref(&handle, mode);
3204 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3205 aa->oa_lockh, req, aa);
3206 ldlm_lock_decref(&handle, mode);
3207 LDLM_LOCK_PUT(lock);
3211 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3212 struct lov_oinfo *loi, int flags,
3213 struct ost_lvb *lvb, __u32 mode, int rc)
3215 if (rc == ELDLM_OK) {
3216 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3219 LASSERT(lock != NULL);
3220 loi->loi_lvb = *lvb;
3221 tmp = loi->loi_lvb.lvb_size;
3222 /* Extend KMS up to the end of this lock and no further
3223 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3224 if (tmp > lock->l_policy_data.l_extent.end)
3225 tmp = lock->l_policy_data.l_extent.end + 1;
3226 if (tmp >= loi->loi_kms) {
3227 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3228 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3229 loi_kms_set(loi, tmp);
3231 LDLM_DEBUG(lock, "lock acquired, setting rss="
3232 LPU64"; leaving kms="LPU64", end="LPU64,
3233 loi->loi_lvb.lvb_size, loi->loi_kms,
3234 lock->l_policy_data.l_extent.end);
3236 ldlm_lock_allow_match(lock);
3237 LDLM_LOCK_PUT(lock);
3238 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3239 loi->loi_lvb = *lvb;
3240 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3241 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3245 EXPORT_SYMBOL(osc_update_enqueue);
3247 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3249 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3250 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3251 * other synchronous requests, however keeping some locks and trying to obtain
3252 * others may take a considerable amount of time in a case of ost failure; and
3253 * when other sync requests do not get released lock from a client, the client
3254 * is excluded from the cluster -- such scenarious make the life difficult, so
3255 * release locks just after they are obtained. */
3256 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3257 int *flags, ldlm_policy_data_t *policy,
3258 struct ost_lvb *lvb, int kms_valid,
3259 obd_enqueue_update_f upcall, void *cookie,
3260 struct ldlm_enqueue_info *einfo,
3261 struct lustre_handle *lockh,
3262 struct ptlrpc_request_set *rqset, int async)
3264 struct obd_device *obd = exp->exp_obd;
3265 struct ptlrpc_request *req = NULL;
3266 int intent = *flags & LDLM_FL_HAS_INTENT;
3271 /* Filesystem lock extents are extended to page boundaries so that
3272 * dealing with the page cache is a little smoother. */
3273 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3274 policy->l_extent.end |= ~CFS_PAGE_MASK;
3277 * kms is not valid when either object is completely fresh (so that no
3278 * locks are cached), or object was evicted. In the latter case cached
3279 * lock cannot be used, because it would prime inode state with
3280 * potentially stale LVB.
3285 /* Next, search for already existing extent locks that will cover us */
3286 /* If we're trying to read, we also search for an existing PW lock. The
3287 * VFS and page cache already protect us locally, so lots of readers/
3288 * writers can share a single PW lock.
3290 * There are problems with conversion deadlocks, so instead of
3291 * converting a read lock to a write lock, we'll just enqueue a new
3294 * At some point we should cancel the read lock instead of making them
3295 * send us a blocking callback, but there are problems with canceling
3296 * locks out from other users right now, too. */
3297 mode = einfo->ei_mode;
3298 if (einfo->ei_mode == LCK_PR)
3300 mode = ldlm_lock_match(obd->obd_namespace,
3301 *flags | LDLM_FL_LVB_READY, res_id,
3302 einfo->ei_type, policy, mode, lockh, 0);
3304 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3306 if (matched->l_ast_data == NULL ||
3307 matched->l_ast_data == einfo->ei_cbdata) {
3308 /* addref the lock only if not async requests and PW
3309 * lock is matched whereas we asked for PR. */
3310 if (!rqset && einfo->ei_mode != mode)
3311 ldlm_lock_addref(lockh, LCK_PR);
3312 osc_set_lock_data_with_check(matched, einfo, *flags);
3314 /* I would like to be able to ASSERT here that
3315 * rss <= kms, but I can't, for reasons which
3316 * are explained in lov_enqueue() */
3319 /* We already have a lock, and it's referenced */
3320 (*upcall)(cookie, ELDLM_OK);
3322 /* For async requests, decref the lock. */
3323 if (einfo->ei_mode != mode)
3324 ldlm_lock_decref(lockh, LCK_PW);
3326 ldlm_lock_decref(lockh, einfo->ei_mode);
3327 LDLM_LOCK_PUT(matched);
3330 ldlm_lock_decref(lockh, mode);
3331 LDLM_LOCK_PUT(matched);
3336 CFS_LIST_HEAD(cancels);
3337 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3338 &RQF_LDLM_ENQUEUE_LVB);
3342 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3346 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3348 ptlrpc_request_set_replen(req);
3351 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3352 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3354 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3355 sizeof(*lvb), lockh, async);
3358 struct osc_enqueue_args *aa;
3359 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3360 aa = ptlrpc_req_async_args(req);
3363 aa->oa_flags = flags;
3364 aa->oa_upcall = upcall;
3365 aa->oa_cookie = cookie;
3367 aa->oa_lockh = lockh;
3369 req->rq_interpret_reply =
3370 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3371 if (rqset == PTLRPCD_SET)
3372 ptlrpcd_add_req(req, PSCOPE_OTHER);
3374 ptlrpc_set_add_req(rqset, req);
3375 } else if (intent) {
3376 ptlrpc_req_finished(req);
3381 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3383 ptlrpc_req_finished(req);
3388 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3389 struct ldlm_enqueue_info *einfo,
3390 struct ptlrpc_request_set *rqset)
3392 struct ldlm_res_id res_id;
3396 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3397 oinfo->oi_md->lsm_object_gr, &res_id);
3399 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3400 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3401 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3402 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3403 rqset, rqset != NULL);
3407 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3408 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3409 int *flags, void *data, struct lustre_handle *lockh,
3412 struct obd_device *obd = exp->exp_obd;
3413 int lflags = *flags;
3417 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3420 /* Filesystem lock extents are extended to page boundaries so that
3421 * dealing with the page cache is a little smoother */
3422 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3423 policy->l_extent.end |= ~CFS_PAGE_MASK;
3425 /* Next, search for already existing extent locks that will cover us */
3426 /* If we're trying to read, we also search for an existing PW lock. The
3427 * VFS and page cache already protect us locally, so lots of readers/
3428 * writers can share a single PW lock. */
3432 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3433 res_id, type, policy, rc, lockh, unref);
3436 osc_set_data_with_check(lockh, data, lflags);
3437 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3438 ldlm_lock_addref(lockh, LCK_PR);
3439 ldlm_lock_decref(lockh, LCK_PW);
3446 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3450 if (unlikely(mode == LCK_GROUP))
3451 ldlm_lock_decref_and_cancel(lockh, mode);
3453 ldlm_lock_decref(lockh, mode);
3458 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3459 __u32 mode, struct lustre_handle *lockh)
3462 RETURN(osc_cancel_base(lockh, mode));
3465 static int osc_cancel_unused(struct obd_export *exp,
3466 struct lov_stripe_md *lsm,
3467 ldlm_cancel_flags_t flags,
3470 struct obd_device *obd = class_exp2obd(exp);
3471 struct ldlm_res_id res_id, *resp = NULL;
3474 resp = osc_build_res_name(lsm->lsm_object_id,
3475 lsm->lsm_object_gr, &res_id);
3478 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3481 static int osc_statfs_interpret(const struct lu_env *env,
3482 struct ptlrpc_request *req,
3483 struct osc_async_args *aa, int rc)
3485 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3486 struct obd_statfs *msfs;
3491 /* The request has in fact never been sent
3492 * due to issues at a higher level (LOV).
3493 * Exit immediately since the caller is
3494 * aware of the problem and takes care
3495 * of the clean up */
3498 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3499 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3505 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3507 GOTO(out, rc = -EPROTO);
3510 /* Reinitialize the RDONLY and DEGRADED flags at the client
3511 * on each statfs, so they don't stay set permanently. */
3512 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3514 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3515 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3516 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3517 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3519 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3520 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3521 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3522 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3524 /* Add a bit of hysteresis so this flag isn't continually flapping,
3525 * and ensure that new files don't get extremely fragmented due to
3526 * only a small amount of available space in the filesystem.
3527 * We want to set the NOSPC flag when there is less than ~0.1% free
3528 * and clear it when there is at least ~0.2% free space, so:
3529 * avail < ~0.1% max max = avail + used
3530 * 1025 * avail < avail + used used = blocks - free
3531 * 1024 * avail < used
3532 * 1024 * avail < blocks - free
3533 * avail < ((blocks - free) >> 10)
3535 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3536 * lose that amount of space so in those cases we report no space left
3537 * if their is less than 1 GB left. */
3538 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3539 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3540 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3541 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3542 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3543 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3544 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3546 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3548 *aa->aa_oi->oi_osfs = *msfs;
3550 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3554 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3555 __u64 max_age, struct ptlrpc_request_set *rqset)
3557 struct ptlrpc_request *req;
3558 struct osc_async_args *aa;
3562 /* We could possibly pass max_age in the request (as an absolute
3563 * timestamp or a "seconds.usec ago") so the target can avoid doing
3564 * extra calls into the filesystem if that isn't necessary (e.g.
3565 * during mount that would help a bit). Having relative timestamps
3566 * is not so great if request processing is slow, while absolute
3567 * timestamps are not ideal because they need time synchronization. */
3568 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3572 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3574 ptlrpc_request_free(req);
3577 ptlrpc_request_set_replen(req);
3578 req->rq_request_portal = OST_CREATE_PORTAL;
3579 ptlrpc_at_set_req_timeout(req);
3581 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3582 /* procfs requests not want stat in wait for avoid deadlock */
3583 req->rq_no_resend = 1;
3584 req->rq_no_delay = 1;
3587 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3588 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3589 aa = ptlrpc_req_async_args(req);
3592 ptlrpc_set_add_req(rqset, req);
3596 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3597 __u64 max_age, __u32 flags)
3599 struct obd_statfs *msfs;
3600 struct ptlrpc_request *req;
3601 struct obd_import *imp = NULL;
3605 /*Since the request might also come from lprocfs, so we need
3606 *sync this with client_disconnect_export Bug15684*/
3607 cfs_down_read(&obd->u.cli.cl_sem);
3608 if (obd->u.cli.cl_import)
3609 imp = class_import_get(obd->u.cli.cl_import);
3610 cfs_up_read(&obd->u.cli.cl_sem);
3614 /* We could possibly pass max_age in the request (as an absolute
3615 * timestamp or a "seconds.usec ago") so the target can avoid doing
3616 * extra calls into the filesystem if that isn't necessary (e.g.
3617 * during mount that would help a bit). Having relative timestamps
3618 * is not so great if request processing is slow, while absolute
3619 * timestamps are not ideal because they need time synchronization. */
3620 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3622 class_import_put(imp);
3627 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3629 ptlrpc_request_free(req);
3632 ptlrpc_request_set_replen(req);
3633 req->rq_request_portal = OST_CREATE_PORTAL;
3634 ptlrpc_at_set_req_timeout(req);
3636 if (flags & OBD_STATFS_NODELAY) {
3637 /* procfs requests not want stat in wait for avoid deadlock */
3638 req->rq_no_resend = 1;
3639 req->rq_no_delay = 1;
3642 rc = ptlrpc_queue_wait(req);
3646 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3648 GOTO(out, rc = -EPROTO);
3655 ptlrpc_req_finished(req);
3659 /* Retrieve object striping information.
3661 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3662 * the maximum number of OST indices which will fit in the user buffer.
3663 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3665 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3667 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3668 struct lov_user_md_v3 lum, *lumk;
3669 struct lov_user_ost_data_v1 *lmm_objects;
3670 int rc = 0, lum_size;
3676 /* we only need the header part from user space to get lmm_magic and
3677 * lmm_stripe_count, (the header part is common to v1 and v3) */
3678 lum_size = sizeof(struct lov_user_md_v1);
3679 if (cfs_copy_from_user(&lum, lump, lum_size))
3682 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3683 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3686 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3687 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3688 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3689 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3691 /* we can use lov_mds_md_size() to compute lum_size
3692 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3693 if (lum.lmm_stripe_count > 0) {
3694 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3695 OBD_ALLOC(lumk, lum_size);
3699 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3700 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3702 lmm_objects = &(lumk->lmm_objects[0]);
3703 lmm_objects->l_object_id = lsm->lsm_object_id;
3705 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3709 lumk->lmm_object_id = lsm->lsm_object_id;
3710 lumk->lmm_object_gr = lsm->lsm_object_gr;
3711 lumk->lmm_stripe_count = 1;
3713 if (cfs_copy_to_user(lump, lumk, lum_size))
3717 OBD_FREE(lumk, lum_size);
3723 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3724 void *karg, void *uarg)
3726 struct obd_device *obd = exp->exp_obd;
3727 struct obd_ioctl_data *data = karg;
3731 if (!cfs_try_module_get(THIS_MODULE)) {
3732 CERROR("Can't get module. Is it alive?");
3736 case OBD_IOC_LOV_GET_CONFIG: {
3738 struct lov_desc *desc;
3739 struct obd_uuid uuid;
3743 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3744 GOTO(out, err = -EINVAL);
3746 data = (struct obd_ioctl_data *)buf;
3748 if (sizeof(*desc) > data->ioc_inllen1) {
3749 obd_ioctl_freedata(buf, len);
3750 GOTO(out, err = -EINVAL);
3753 if (data->ioc_inllen2 < sizeof(uuid)) {
3754 obd_ioctl_freedata(buf, len);
3755 GOTO(out, err = -EINVAL);
3758 desc = (struct lov_desc *)data->ioc_inlbuf1;
3759 desc->ld_tgt_count = 1;
3760 desc->ld_active_tgt_count = 1;
3761 desc->ld_default_stripe_count = 1;
3762 desc->ld_default_stripe_size = 0;
3763 desc->ld_default_stripe_offset = 0;
3764 desc->ld_pattern = 0;
3765 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3767 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3769 err = cfs_copy_to_user((void *)uarg, buf, len);
3772 obd_ioctl_freedata(buf, len);
3775 case LL_IOC_LOV_SETSTRIPE:
3776 err = obd_alloc_memmd(exp, karg);
3780 case LL_IOC_LOV_GETSTRIPE:
3781 err = osc_getstripe(karg, uarg);
3783 case OBD_IOC_CLIENT_RECOVER:
3784 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3789 case IOC_OSC_SET_ACTIVE:
3790 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3793 case OBD_IOC_POLL_QUOTACHECK:
3794 err = lquota_poll_check(quota_interface, exp,
3795 (struct if_quotacheck *)karg);
3797 case OBD_IOC_PING_TARGET:
3798 err = ptlrpc_obd_ping(obd);
3801 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3802 cmd, cfs_curproc_comm());
3803 GOTO(out, err = -ENOTTY);
3806 cfs_module_put(THIS_MODULE);
3810 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3811 void *key, __u32 *vallen, void *val,
3812 struct lov_stripe_md *lsm)
3815 if (!vallen || !val)
3818 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3819 __u32 *stripe = val;
3820 *vallen = sizeof(*stripe);
3823 } else if (KEY_IS(KEY_LAST_ID)) {
3824 struct ptlrpc_request *req;
3829 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3830 &RQF_OST_GET_INFO_LAST_ID);
3834 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3835 RCL_CLIENT, keylen);
3836 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3838 ptlrpc_request_free(req);
3842 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3843 memcpy(tmp, key, keylen);
3845 req->rq_no_delay = req->rq_no_resend = 1;
3846 ptlrpc_request_set_replen(req);
3847 rc = ptlrpc_queue_wait(req);
3851 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3853 GOTO(out, rc = -EPROTO);
3855 *((obd_id *)val) = *reply;
3857 ptlrpc_req_finished(req);
3859 } else if (KEY_IS(KEY_FIEMAP)) {
3860 struct ptlrpc_request *req;
3861 struct ll_user_fiemap *reply;
3865 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3866 &RQF_OST_GET_INFO_FIEMAP);
3870 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3871 RCL_CLIENT, keylen);
3872 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3873 RCL_CLIENT, *vallen);
3874 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3875 RCL_SERVER, *vallen);
3877 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3879 ptlrpc_request_free(req);
3883 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3884 memcpy(tmp, key, keylen);
3885 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3886 memcpy(tmp, val, *vallen);
3888 ptlrpc_request_set_replen(req);
3889 rc = ptlrpc_queue_wait(req);
3893 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3895 GOTO(out1, rc = -EPROTO);
3897 memcpy(val, reply, *vallen);
3899 ptlrpc_req_finished(req);
3907 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3909 struct llog_ctxt *ctxt;
3913 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3915 rc = llog_initiator_connect(ctxt);
3916 llog_ctxt_put(ctxt);
3918 /* XXX return an error? skip setting below flags? */
3921 cfs_spin_lock(&imp->imp_lock);
3922 imp->imp_server_timeout = 1;
3923 imp->imp_pingable = 1;
3924 cfs_spin_unlock(&imp->imp_lock);
3925 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3930 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3931 struct ptlrpc_request *req,
3938 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3941 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3942 void *key, obd_count vallen, void *val,
3943 struct ptlrpc_request_set *set)
3945 struct ptlrpc_request *req;
3946 struct obd_device *obd = exp->exp_obd;
3947 struct obd_import *imp = class_exp2cliimp(exp);
3952 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3954 if (KEY_IS(KEY_NEXT_ID)) {
3956 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3958 if (vallen != sizeof(obd_id))
3963 if (vallen != sizeof(obd_id))
3966 /* avoid race between allocate new object and set next id
3967 * from ll_sync thread */
3968 cfs_spin_lock(&oscc->oscc_lock);
3969 new_val = *((obd_id*)val) + 1;
3970 if (new_val > oscc->oscc_next_id)
3971 oscc->oscc_next_id = new_val;
3972 cfs_spin_unlock(&oscc->oscc_lock);
3973 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3974 exp->exp_obd->obd_name,
3975 obd->u.cli.cl_oscc.oscc_next_id);
3980 if (KEY_IS(KEY_INIT_RECOV)) {
3981 if (vallen != sizeof(int))
3983 cfs_spin_lock(&imp->imp_lock);
3984 imp->imp_initial_recov = *(int *)val;
3985 cfs_spin_unlock(&imp->imp_lock);
3986 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3987 exp->exp_obd->obd_name,
3988 imp->imp_initial_recov);
3992 if (KEY_IS(KEY_CHECKSUM)) {
3993 if (vallen != sizeof(int))
3995 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3999 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4000 sptlrpc_conf_client_adapt(obd);
4004 if (KEY_IS(KEY_FLUSH_CTX)) {
4005 sptlrpc_import_flush_my_ctx(imp);
4009 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4012 /* We pass all other commands directly to OST. Since nobody calls osc
4013 methods directly and everybody is supposed to go through LOV, we
4014 assume lov checked invalid values for us.
4015 The only recognised values so far are evict_by_nid and mds_conn.
4016 Even if something bad goes through, we'd get a -EINVAL from OST
4019 if (KEY_IS(KEY_GRANT_SHRINK))
4020 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4022 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4027 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4028 RCL_CLIENT, keylen);
4029 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4030 RCL_CLIENT, vallen);
4031 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4033 ptlrpc_request_free(req);
4037 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4038 memcpy(tmp, key, keylen);
4039 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4040 memcpy(tmp, val, vallen);
4042 if (KEY_IS(KEY_MDS_CONN)) {
4043 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4045 oscc->oscc_oa.o_gr = (*(__u32 *)val);
4046 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4047 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
4048 req->rq_no_delay = req->rq_no_resend = 1;
4049 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4050 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4051 struct osc_grant_args *aa;
4054 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4055 aa = ptlrpc_req_async_args(req);
4058 ptlrpc_req_finished(req);
4061 *oa = ((struct ost_body *)val)->oa;
4063 req->rq_interpret_reply = osc_shrink_grant_interpret;
4066 ptlrpc_request_set_replen(req);
4067 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4068 LASSERT(set != NULL);
4069 ptlrpc_set_add_req(set, req);
4070 ptlrpc_check_set(NULL, set);
4072 ptlrpcd_add_req(req, PSCOPE_OTHER);
4078 static struct llog_operations osc_size_repl_logops = {
4079 lop_cancel: llog_obd_repl_cancel
4082 static struct llog_operations osc_mds_ost_orig_logops;
4084 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4085 struct obd_device *tgt, struct llog_catid *catid)
4090 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4091 &catid->lci_logid, &osc_mds_ost_orig_logops);
4093 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4097 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4098 NULL, &osc_size_repl_logops);
4100 struct llog_ctxt *ctxt =
4101 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4104 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4109 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4110 obd->obd_name, tgt->obd_name, catid, rc);
4111 CERROR("logid "LPX64":0x%x\n",
4112 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4117 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4118 struct obd_device *disk_obd, int *index)
4120 struct llog_catid catid;
4121 static char name[32] = CATLIST;
4125 LASSERT(olg == &obd->obd_olg);
4127 cfs_mutex_down(&olg->olg_cat_processing);
4128 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4130 CERROR("rc: %d\n", rc);
4134 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4135 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4136 catid.lci_logid.lgl_ogr, catid.lci_logid.lgl_ogen);
4138 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4140 CERROR("rc: %d\n", rc);
4144 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4146 CERROR("rc: %d\n", rc);
4151 cfs_mutex_up(&olg->olg_cat_processing);
4156 static int osc_llog_finish(struct obd_device *obd, int count)
4158 struct llog_ctxt *ctxt;
4159 int rc = 0, rc2 = 0;
4162 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4164 rc = llog_cleanup(ctxt);
4166 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4168 rc2 = llog_cleanup(ctxt);
4175 static int osc_reconnect(const struct lu_env *env,
4176 struct obd_export *exp, struct obd_device *obd,
4177 struct obd_uuid *cluuid,
4178 struct obd_connect_data *data,
4181 struct client_obd *cli = &obd->u.cli;
4183 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4186 client_obd_list_lock(&cli->cl_loi_list_lock);
4187 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4188 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4189 lost_grant = cli->cl_lost_grant;
4190 cli->cl_lost_grant = 0;
4191 client_obd_list_unlock(&cli->cl_loi_list_lock);
4193 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4194 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4195 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4196 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4197 " ocd_grant: %d\n", data->ocd_connect_flags,
4198 data->ocd_version, data->ocd_grant);
4204 static int osc_disconnect(struct obd_export *exp)
4206 struct obd_device *obd = class_exp2obd(exp);
4207 struct llog_ctxt *ctxt;
4210 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4212 if (obd->u.cli.cl_conn_count == 1) {
4213 /* Flush any remaining cancel messages out to the
4215 llog_sync(ctxt, exp);
4217 llog_ctxt_put(ctxt);
4219 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4223 rc = client_disconnect_export(exp);
4225 * Initially we put del_shrink_grant before disconnect_export, but it
4226 * causes the following problem if setup (connect) and cleanup
4227 * (disconnect) are tangled together.
4228 * connect p1 disconnect p2
4229 * ptlrpc_connect_import
4230 * ............... class_manual_cleanup
4233 * ptlrpc_connect_interrupt
4235 * add this client to shrink list
4237 * Bang! pinger trigger the shrink.
4238 * So the osc should be disconnected from the shrink list, after we
4239 * are sure the import has been destroyed. BUG18662
4241 if (obd->u.cli.cl_import == NULL)
4242 osc_del_shrink_grant(&obd->u.cli);
4246 static int osc_import_event(struct obd_device *obd,
4247 struct obd_import *imp,
4248 enum obd_import_event event)
4250 struct client_obd *cli;
4254 LASSERT(imp->imp_obd == obd);
4257 case IMP_EVENT_DISCON: {
4258 /* Only do this on the MDS OSC's */
4259 if (imp->imp_server_timeout) {
4260 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4262 cfs_spin_lock(&oscc->oscc_lock);
4263 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4264 cfs_spin_unlock(&oscc->oscc_lock);
4267 client_obd_list_lock(&cli->cl_loi_list_lock);
4268 cli->cl_avail_grant = 0;
4269 cli->cl_lost_grant = 0;
4270 client_obd_list_unlock(&cli->cl_loi_list_lock);
4273 case IMP_EVENT_INACTIVE: {
4274 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4277 case IMP_EVENT_INVALIDATE: {
4278 struct ldlm_namespace *ns = obd->obd_namespace;
4282 env = cl_env_get(&refcheck);
4286 client_obd_list_lock(&cli->cl_loi_list_lock);
4287 /* all pages go to failing rpcs due to the invalid
4289 osc_check_rpcs(env, cli);
4290 client_obd_list_unlock(&cli->cl_loi_list_lock);
4292 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4293 cl_env_put(env, &refcheck);
4298 case IMP_EVENT_ACTIVE: {
4299 /* Only do this on the MDS OSC's */
4300 if (imp->imp_server_timeout) {
4301 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4303 cfs_spin_lock(&oscc->oscc_lock);
4304 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4305 cfs_spin_unlock(&oscc->oscc_lock);
4307 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4310 case IMP_EVENT_OCD: {
4311 struct obd_connect_data *ocd = &imp->imp_connect_data;
4313 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4314 osc_init_grant(&obd->u.cli, ocd);
4317 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4318 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4320 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4324 CERROR("Unknown import event %d\n", event);
4330 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4336 rc = ptlrpcd_addref();
4340 rc = client_obd_setup(obd, lcfg);
4344 struct lprocfs_static_vars lvars = { 0 };
4345 struct client_obd *cli = &obd->u.cli;
4347 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4348 lprocfs_osc_init_vars(&lvars);
4349 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4350 lproc_osc_attach_seqstat(obd);
4351 sptlrpc_lprocfs_cliobd_attach(obd);
4352 ptlrpc_lprocfs_register_obd(obd);
4356 /* We need to allocate a few requests more, because
4357 brw_interpret tries to create new requests before freeing
4358 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4359 reserved, but I afraid that might be too much wasted RAM
4360 in fact, so 2 is just my guess and still should work. */
4361 cli->cl_import->imp_rq_pool =
4362 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4364 ptlrpc_add_rqs_to_pool);
4366 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4367 cfs_sema_init(&cli->cl_grant_sem, 1);
4373 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4379 case OBD_CLEANUP_EARLY: {
4380 struct obd_import *imp;
4381 imp = obd->u.cli.cl_import;
4382 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4383 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4384 ptlrpc_deactivate_import(imp);
4385 cfs_spin_lock(&imp->imp_lock);
4386 imp->imp_pingable = 0;
4387 cfs_spin_unlock(&imp->imp_lock);
4390 case OBD_CLEANUP_EXPORTS: {
4391 /* If we set up but never connected, the
4392 client import will not have been cleaned. */
4393 if (obd->u.cli.cl_import) {
4394 struct obd_import *imp;
4395 cfs_down_write(&obd->u.cli.cl_sem);
4396 imp = obd->u.cli.cl_import;
4397 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4399 ptlrpc_invalidate_import(imp);
4400 if (imp->imp_rq_pool) {
4401 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4402 imp->imp_rq_pool = NULL;
4404 class_destroy_import(imp);
4405 cfs_up_write(&obd->u.cli.cl_sem);
4406 obd->u.cli.cl_import = NULL;
4408 rc = obd_llog_finish(obd, 0);
4410 CERROR("failed to cleanup llogging subsystems\n");
4417 int osc_cleanup(struct obd_device *obd)
4422 ptlrpc_lprocfs_unregister_obd(obd);
4423 lprocfs_obd_cleanup(obd);
4425 /* free memory of osc quota cache */
4426 lquota_cleanup(quota_interface, obd);
4428 rc = client_obd_cleanup(obd);
4434 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4436 struct lprocfs_static_vars lvars = { 0 };
4439 lprocfs_osc_init_vars(&lvars);
4441 switch (lcfg->lcfg_command) {
4443 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4453 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4455 return osc_process_config_base(obd, buf);
4458 struct obd_ops osc_obd_ops = {
4459 .o_owner = THIS_MODULE,
4460 .o_setup = osc_setup,
4461 .o_precleanup = osc_precleanup,
4462 .o_cleanup = osc_cleanup,
4463 .o_add_conn = client_import_add_conn,
4464 .o_del_conn = client_import_del_conn,
4465 .o_connect = client_connect_import,
4466 .o_reconnect = osc_reconnect,
4467 .o_disconnect = osc_disconnect,
4468 .o_statfs = osc_statfs,
4469 .o_statfs_async = osc_statfs_async,
4470 .o_packmd = osc_packmd,
4471 .o_unpackmd = osc_unpackmd,
4472 .o_precreate = osc_precreate,
4473 .o_create = osc_create,
4474 .o_create_async = osc_create_async,
4475 .o_destroy = osc_destroy,
4476 .o_getattr = osc_getattr,
4477 .o_getattr_async = osc_getattr_async,
4478 .o_setattr = osc_setattr,
4479 .o_setattr_async = osc_setattr_async,
4481 .o_punch = osc_punch,
4483 .o_enqueue = osc_enqueue,
4484 .o_change_cbdata = osc_change_cbdata,
4485 .o_find_cbdata = osc_find_cbdata,
4486 .o_cancel = osc_cancel,
4487 .o_cancel_unused = osc_cancel_unused,
4488 .o_iocontrol = osc_iocontrol,
4489 .o_get_info = osc_get_info,
4490 .o_set_info_async = osc_set_info_async,
4491 .o_import_event = osc_import_event,
4492 .o_llog_init = osc_llog_init,
4493 .o_llog_finish = osc_llog_finish,
4494 .o_process_config = osc_process_config,
4497 extern struct lu_kmem_descr osc_caches[];
4498 extern cfs_spinlock_t osc_ast_guard;
4499 extern cfs_lock_class_key_t osc_ast_guard_class;
4501 int __init osc_init(void)
4503 struct lprocfs_static_vars lvars = { 0 };
4507 /* print an address of _any_ initialized kernel symbol from this
4508 * module, to allow debugging with gdb that doesn't support data
4509 * symbols from modules.*/
4510 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4512 rc = lu_kmem_init(osc_caches);
4514 lprocfs_osc_init_vars(&lvars);
4516 cfs_request_module("lquota");
4517 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4518 lquota_init(quota_interface);
4519 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4521 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4522 LUSTRE_OSC_NAME, &osc_device_type);
4524 if (quota_interface)
4525 PORTAL_SYMBOL_PUT(osc_quota_interface);
4526 lu_kmem_fini(osc_caches);
4530 cfs_spin_lock_init(&osc_ast_guard);
4531 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4533 osc_mds_ost_orig_logops = llog_lvfs_ops;
4534 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4535 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4536 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4537 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4543 static void /*__exit*/ osc_exit(void)
4545 lu_device_type_fini(&osc_device_type);
4547 lquota_exit(quota_interface);
4548 if (quota_interface)
4549 PORTAL_SYMBOL_PUT(osc_quota_interface);
4551 class_unregister_type(LUSTRE_OSC_NAME);
4552 lu_kmem_fini(osc_caches);
4555 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4556 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4557 MODULE_LICENSE("GPL");
4559 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);