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_async_args *aa, 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(aa->aa_oi->oi_oa, &body->oa);
363 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
367 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
368 struct obd_trans_info *oti,
369 struct ptlrpc_request_set *rqset)
371 struct ptlrpc_request *req;
372 struct osc_async_args *aa;
376 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
380 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
381 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
383 ptlrpc_request_free(req);
387 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
388 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
390 osc_pack_req_body(req, oinfo);
392 ptlrpc_request_set_replen(req);
394 /* do mds to ost setattr asynchronously */
396 /* Do not wait for response. */
397 ptlrpcd_add_req(req, PSCOPE_OTHER);
399 req->rq_interpret_reply =
400 (ptlrpc_interpterer_t)osc_setattr_interpret;
402 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
403 aa = ptlrpc_req_async_args(req);
406 ptlrpc_set_add_req(rqset, req);
412 int osc_real_create(struct obd_export *exp, struct obdo *oa,
413 struct lov_stripe_md **ea, struct obd_trans_info *oti)
415 struct ptlrpc_request *req;
416 struct ost_body *body;
417 struct lov_stripe_md *lsm;
426 rc = obd_alloc_memmd(exp, &lsm);
431 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
433 GOTO(out, rc = -ENOMEM);
435 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
437 ptlrpc_request_free(req);
441 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
443 lustre_set_wire_obdo(&body->oa, oa);
445 ptlrpc_request_set_replen(req);
447 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
448 oa->o_flags == OBD_FL_DELORPHAN) {
450 "delorphan from OST integration");
451 /* Don't resend the delorphan req */
452 req->rq_no_resend = req->rq_no_delay = 1;
455 rc = ptlrpc_queue_wait(req);
459 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
461 GOTO(out_req, rc = -EPROTO);
463 lustre_get_wire_obdo(oa, &body->oa);
465 /* This should really be sent by the OST */
466 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
467 oa->o_valid |= OBD_MD_FLBLKSZ;
469 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
470 * have valid lsm_oinfo data structs, so don't go touching that.
471 * This needs to be fixed in a big way.
473 lsm->lsm_object_id = oa->o_id;
474 lsm->lsm_object_gr = oa->o_gr;
478 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
480 if (oa->o_valid & OBD_MD_FLCOOKIE) {
481 if (!oti->oti_logcookies)
482 oti_alloc_cookies(oti, 1);
483 *oti->oti_logcookies = oa->o_lcookie;
487 CDEBUG(D_HA, "transno: "LPD64"\n",
488 lustre_msg_get_transno(req->rq_repmsg));
490 ptlrpc_req_finished(req);
493 obd_free_memmd(exp, &lsm);
497 static int osc_punch_interpret(const struct lu_env *env,
498 struct ptlrpc_request *req,
499 struct osc_punch_args *aa, int rc)
501 struct ost_body *body;
507 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
509 GOTO(out, rc = -EPROTO);
511 lustre_get_wire_obdo(aa->pa_oa, &body->oa);
513 rc = aa->pa_upcall(aa->pa_cookie, rc);
517 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
518 struct obd_capa *capa,
519 obd_enqueue_update_f upcall, void *cookie,
520 struct ptlrpc_request_set *rqset)
522 struct ptlrpc_request *req;
523 struct osc_punch_args *aa;
524 struct ost_body *body;
528 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
532 osc_set_capa_size(req, &RMF_CAPA1, capa);
533 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
535 ptlrpc_request_free(req);
538 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
539 ptlrpc_at_set_req_timeout(req);
541 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
543 lustre_set_wire_obdo(&body->oa, oa);
544 osc_pack_capa(req, body, capa);
546 ptlrpc_request_set_replen(req);
549 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_punch_interpret;
550 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
551 aa = ptlrpc_req_async_args(req);
553 aa->pa_upcall = upcall;
554 aa->pa_cookie = cookie;
555 if (rqset == PTLRPCD_SET)
556 ptlrpcd_add_req(req, PSCOPE_OTHER);
558 ptlrpc_set_add_req(rqset, req);
563 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
564 struct obd_trans_info *oti,
565 struct ptlrpc_request_set *rqset)
567 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
568 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
569 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
570 return osc_punch_base(exp, oinfo->oi_oa, oinfo->oi_capa,
571 oinfo->oi_cb_up, oinfo, rqset);
574 static int osc_sync(struct obd_export *exp, struct obdo *oa,
575 struct lov_stripe_md *md, obd_size start, obd_size end,
578 struct ptlrpc_request *req;
579 struct ost_body *body;
584 CDEBUG(D_INFO, "oa NULL\n");
588 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
592 osc_set_capa_size(req, &RMF_CAPA1, capa);
593 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
595 ptlrpc_request_free(req);
599 /* overload the size and blocks fields in the oa with start/end */
600 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
602 lustre_set_wire_obdo(&body->oa, oa);
603 body->oa.o_size = start;
604 body->oa.o_blocks = end;
605 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
606 osc_pack_capa(req, body, capa);
608 ptlrpc_request_set_replen(req);
610 rc = ptlrpc_queue_wait(req);
614 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
616 GOTO(out, rc = -EPROTO);
618 lustre_get_wire_obdo(oa, &body->oa);
622 ptlrpc_req_finished(req);
626 /* Find and cancel locally locks matched by @mode in the resource found by
627 * @objid. Found locks are added into @cancel list. Returns the amount of
628 * locks added to @cancels list. */
629 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
630 struct list_head *cancels, ldlm_mode_t mode,
633 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
634 struct ldlm_res_id res_id;
635 struct ldlm_resource *res;
639 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
640 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
644 LDLM_RESOURCE_ADDREF(res);
645 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
646 lock_flags, 0, NULL);
647 LDLM_RESOURCE_DELREF(res);
648 ldlm_resource_putref(res);
652 static int osc_destroy_interpret(const struct lu_env *env,
653 struct ptlrpc_request *req, void *data,
656 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
658 atomic_dec(&cli->cl_destroy_in_flight);
659 cfs_waitq_signal(&cli->cl_destroy_waitq);
663 static int osc_can_send_destroy(struct client_obd *cli)
665 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
666 cli->cl_max_rpcs_in_flight) {
667 /* The destroy request can be sent */
670 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
671 cli->cl_max_rpcs_in_flight) {
673 * The counter has been modified between the two atomic
676 cfs_waitq_signal(&cli->cl_destroy_waitq);
681 /* Destroy requests can be async always on the client, and we don't even really
682 * care about the return code since the client cannot do anything at all about
684 * When the MDS is unlinking a filename, it saves the file objects into a
685 * recovery llog, and these object records are cancelled when the OST reports
686 * they were destroyed and sync'd to disk (i.e. transaction committed).
687 * If the client dies, or the OST is down when the object should be destroyed,
688 * the records are not cancelled, and when the OST reconnects to the MDS next,
689 * it will retrieve the llog unlink logs and then sends the log cancellation
690 * cookies to the MDS after committing destroy transactions. */
691 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
692 struct lov_stripe_md *ea, struct obd_trans_info *oti,
693 struct obd_export *md_export, void *capa)
695 struct client_obd *cli = &exp->exp_obd->u.cli;
696 struct ptlrpc_request *req;
697 struct ost_body *body;
698 CFS_LIST_HEAD(cancels);
703 CDEBUG(D_INFO, "oa NULL\n");
707 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
708 LDLM_FL_DISCARD_DATA);
710 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
712 ldlm_lock_list_put(&cancels, l_bl_ast, count);
716 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
717 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
720 ptlrpc_request_free(req);
724 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
725 ptlrpc_at_set_req_timeout(req);
727 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
728 oa->o_lcookie = *oti->oti_logcookies;
729 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
731 lustre_set_wire_obdo(&body->oa, oa);
733 osc_pack_capa(req, body, (struct obd_capa *)capa);
734 ptlrpc_request_set_replen(req);
736 /* don't throttle destroy RPCs for the MDT */
737 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
738 req->rq_interpret_reply = osc_destroy_interpret;
739 if (!osc_can_send_destroy(cli)) {
740 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
744 * Wait until the number of on-going destroy RPCs drops
745 * under max_rpc_in_flight
747 l_wait_event_exclusive(cli->cl_destroy_waitq,
748 osc_can_send_destroy(cli), &lwi);
752 /* Do not wait for response */
753 ptlrpcd_add_req(req, PSCOPE_OTHER);
757 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
760 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
762 LASSERT(!(oa->o_valid & bits));
765 client_obd_list_lock(&cli->cl_loi_list_lock);
766 oa->o_dirty = cli->cl_dirty;
767 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
768 CERROR("dirty %lu - %lu > dirty_max %lu\n",
769 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
771 } else if (atomic_read(&obd_dirty_pages) -
772 atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages + 1){
773 /* The atomic_read() allowing the atomic_inc() are not covered
774 * by a lock thus they may safely race and trip this CERROR()
775 * unless we add in a small fudge factor (+1). */
776 CERROR("dirty %d - %d > system dirty_max %d\n",
777 atomic_read(&obd_dirty_pages),
778 atomic_read(&obd_dirty_transit_pages),
779 obd_max_dirty_pages);
781 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
782 CERROR("dirty %lu - dirty_max %lu too big???\n",
783 cli->cl_dirty, cli->cl_dirty_max);
786 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
787 (cli->cl_max_rpcs_in_flight + 1);
788 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
790 oa->o_grant = cli->cl_avail_grant;
791 oa->o_dropped = cli->cl_lost_grant;
792 cli->cl_lost_grant = 0;
793 client_obd_list_unlock(&cli->cl_loi_list_lock);
794 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
795 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
799 static void osc_update_next_shrink(struct client_obd *cli)
801 cli->cl_next_shrink_grant =
802 cfs_time_shift(cli->cl_grant_shrink_interval);
803 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
804 cli->cl_next_shrink_grant);
807 /* caller must hold loi_list_lock */
808 static void osc_consume_write_grant(struct client_obd *cli,
809 struct brw_page *pga)
811 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
812 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
813 atomic_inc(&obd_dirty_pages);
814 cli->cl_dirty += CFS_PAGE_SIZE;
815 cli->cl_avail_grant -= CFS_PAGE_SIZE;
816 pga->flag |= OBD_BRW_FROM_GRANT;
817 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
818 CFS_PAGE_SIZE, pga, pga->pg);
819 LASSERT(cli->cl_avail_grant >= 0);
820 osc_update_next_shrink(cli);
823 /* the companion to osc_consume_write_grant, called when a brw has completed.
824 * must be called with the loi lock held. */
825 static void osc_release_write_grant(struct client_obd *cli,
826 struct brw_page *pga, int sent)
828 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
831 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
832 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
837 pga->flag &= ~OBD_BRW_FROM_GRANT;
838 atomic_dec(&obd_dirty_pages);
839 cli->cl_dirty -= CFS_PAGE_SIZE;
840 if (pga->flag & OBD_BRW_NOCACHE) {
841 pga->flag &= ~OBD_BRW_NOCACHE;
842 atomic_dec(&obd_dirty_transit_pages);
843 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
846 cli->cl_lost_grant += CFS_PAGE_SIZE;
847 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
848 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
849 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
850 /* For short writes we shouldn't count parts of pages that
851 * span a whole block on the OST side, or our accounting goes
852 * wrong. Should match the code in filter_grant_check. */
853 int offset = pga->off & ~CFS_PAGE_MASK;
854 int count = pga->count + (offset & (blocksize - 1));
855 int end = (offset + pga->count) & (blocksize - 1);
857 count += blocksize - end;
859 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
860 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
861 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
862 cli->cl_avail_grant, cli->cl_dirty);
868 static unsigned long rpcs_in_flight(struct client_obd *cli)
870 return cli->cl_r_in_flight + cli->cl_w_in_flight;
873 /* caller must hold loi_list_lock */
874 void osc_wake_cache_waiters(struct client_obd *cli)
876 struct list_head *l, *tmp;
877 struct osc_cache_waiter *ocw;
880 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
881 /* if we can't dirty more, we must wait until some is written */
882 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
883 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
884 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
885 "osc max %ld, sys max %d\n", cli->cl_dirty,
886 cli->cl_dirty_max, obd_max_dirty_pages);
890 /* if still dirty cache but no grant wait for pending RPCs that
891 * may yet return us some grant before doing sync writes */
892 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
893 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
894 cli->cl_w_in_flight);
898 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
899 list_del_init(&ocw->ocw_entry);
900 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
901 /* no more RPCs in flight to return grant, do sync IO */
902 ocw->ocw_rc = -EDQUOT;
903 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
905 osc_consume_write_grant(cli,
906 &ocw->ocw_oap->oap_brw_page);
909 cfs_waitq_signal(&ocw->ocw_waitq);
915 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
917 client_obd_list_lock(&cli->cl_loi_list_lock);
918 cli->cl_avail_grant += grant;
919 client_obd_list_unlock(&cli->cl_loi_list_lock);
922 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
924 if (body->oa.o_valid & OBD_MD_FLGRANT) {
925 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
926 __osc_update_grant(cli, body->oa.o_grant);
930 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
931 void *key, obd_count vallen, void *val,
932 struct ptlrpc_request_set *set);
934 static int osc_shrink_grant_interpret(const struct lu_env *env,
935 struct ptlrpc_request *req,
938 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
939 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
940 struct ost_body *body;
943 __osc_update_grant(cli, oa->o_grant);
947 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
949 osc_update_grant(cli, body);
955 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
957 client_obd_list_lock(&cli->cl_loi_list_lock);
958 oa->o_grant = cli->cl_avail_grant / 4;
959 cli->cl_avail_grant -= oa->o_grant;
960 client_obd_list_unlock(&cli->cl_loi_list_lock);
961 oa->o_flags |= OBD_FL_SHRINK_GRANT;
962 osc_update_next_shrink(cli);
965 /* Shrink the current grant, either from some large amount to enough for a
966 * full set of in-flight RPCs, or if we have already shrunk to that limit
967 * then to enough for a single RPC. This avoids keeping more grant than
968 * needed, and avoids shrinking the grant piecemeal. */
969 static int osc_shrink_grant(struct client_obd *cli)
971 long target = (cli->cl_max_rpcs_in_flight + 1) *
972 cli->cl_max_pages_per_rpc;
974 client_obd_list_lock(&cli->cl_loi_list_lock);
975 if (cli->cl_avail_grant <= target)
976 target = cli->cl_max_pages_per_rpc;
977 client_obd_list_unlock(&cli->cl_loi_list_lock);
979 return osc_shrink_grant_to_target(cli, target);
982 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
985 struct ost_body *body;
988 client_obd_list_lock(&cli->cl_loi_list_lock);
989 /* Don't shrink if we are already above or below the desired limit
990 * We don't want to shrink below a single RPC, as that will negatively
991 * impact block allocation and long-term performance. */
992 if (target < cli->cl_max_pages_per_rpc)
993 target = cli->cl_max_pages_per_rpc;
995 if (target >= cli->cl_avail_grant) {
996 client_obd_list_unlock(&cli->cl_loi_list_lock);
999 client_obd_list_unlock(&cli->cl_loi_list_lock);
1001 OBD_ALLOC_PTR(body);
1005 osc_announce_cached(cli, &body->oa, 0);
1007 client_obd_list_lock(&cli->cl_loi_list_lock);
1008 body->oa.o_grant = cli->cl_avail_grant - target;
1009 cli->cl_avail_grant = target;
1010 client_obd_list_unlock(&cli->cl_loi_list_lock);
1011 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1012 osc_update_next_shrink(cli);
1014 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1015 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1016 sizeof(*body), body, NULL);
1018 __osc_update_grant(cli, body->oa.o_grant);
1023 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1024 static int osc_should_shrink_grant(struct client_obd *client)
1026 cfs_time_t time = cfs_time_current();
1027 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1028 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1029 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1030 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1033 osc_update_next_shrink(client);
1038 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1040 struct client_obd *client;
1042 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1043 if (osc_should_shrink_grant(client))
1044 osc_shrink_grant(client);
1049 static int osc_add_shrink_grant(struct client_obd *client)
1053 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1055 osc_grant_shrink_grant_cb, NULL,
1056 &client->cl_grant_shrink_list);
1058 CERROR("add grant client %s error %d\n",
1059 client->cl_import->imp_obd->obd_name, rc);
1062 CDEBUG(D_CACHE, "add grant client %s \n",
1063 client->cl_import->imp_obd->obd_name);
1064 osc_update_next_shrink(client);
1068 static int osc_del_shrink_grant(struct client_obd *client)
1070 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1074 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1077 * ocd_grant is the total grant amount we're expect to hold: if we've
1078 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1079 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1081 * race is tolerable here: if we're evicted, but imp_state already
1082 * left EVICTED state, then cl_dirty must be 0 already.
1084 client_obd_list_lock(&cli->cl_loi_list_lock);
1085 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1086 cli->cl_avail_grant = ocd->ocd_grant;
1088 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1089 client_obd_list_unlock(&cli->cl_loi_list_lock);
1091 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1092 cli->cl_avail_grant, cli->cl_lost_grant);
1093 LASSERT(cli->cl_avail_grant >= 0);
1095 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1096 list_empty(&cli->cl_grant_shrink_list))
1097 osc_add_shrink_grant(cli);
1100 /* We assume that the reason this OSC got a short read is because it read
1101 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1102 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1103 * this stripe never got written at or beyond this stripe offset yet. */
1104 static void handle_short_read(int nob_read, obd_count page_count,
1105 struct brw_page **pga)
1110 /* skip bytes read OK */
1111 while (nob_read > 0) {
1112 LASSERT (page_count > 0);
1114 if (pga[i]->count > nob_read) {
1115 /* EOF inside this page */
1116 ptr = cfs_kmap(pga[i]->pg) +
1117 (pga[i]->off & ~CFS_PAGE_MASK);
1118 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1119 cfs_kunmap(pga[i]->pg);
1125 nob_read -= pga[i]->count;
1130 /* zero remaining pages */
1131 while (page_count-- > 0) {
1132 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1133 memset(ptr, 0, pga[i]->count);
1134 cfs_kunmap(pga[i]->pg);
1139 static int check_write_rcs(struct ptlrpc_request *req,
1140 int requested_nob, int niocount,
1141 obd_count page_count, struct brw_page **pga)
1146 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1147 sizeof(*remote_rcs) *
1149 if (remote_rcs == NULL) {
1150 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1154 /* return error if any niobuf was in error */
1155 for (i = 0; i < niocount; i++) {
1156 if (remote_rcs[i] < 0)
1157 return(remote_rcs[i]);
1159 if (remote_rcs[i] != 0) {
1160 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1161 i, remote_rcs[i], req);
1166 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1167 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1168 req->rq_bulk->bd_nob_transferred, requested_nob);
1175 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1177 if (p1->flag != p2->flag) {
1178 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1179 OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1181 /* warn if we try to combine flags that we don't know to be
1182 * safe to combine */
1183 if ((p1->flag & mask) != (p2->flag & mask))
1184 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1185 "same brw?\n", p1->flag, p2->flag);
1189 return (p1->off + p1->count == p2->off);
1192 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1193 struct brw_page **pga, int opc,
1194 cksum_type_t cksum_type)
1199 LASSERT (pg_count > 0);
1200 cksum = init_checksum(cksum_type);
1201 while (nob > 0 && pg_count > 0) {
1202 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1203 int off = pga[i]->off & ~CFS_PAGE_MASK;
1204 int count = pga[i]->count > nob ? nob : pga[i]->count;
1206 /* corrupt the data before we compute the checksum, to
1207 * simulate an OST->client data error */
1208 if (i == 0 && opc == OST_READ &&
1209 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1210 memcpy(ptr + off, "bad1", min(4, nob));
1211 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1212 cfs_kunmap(pga[i]->pg);
1213 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1216 nob -= pga[i]->count;
1220 /* For sending we only compute the wrong checksum instead
1221 * of corrupting the data so it is still correct on a redo */
1222 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1228 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1229 struct lov_stripe_md *lsm, obd_count page_count,
1230 struct brw_page **pga,
1231 struct ptlrpc_request **reqp,
1232 struct obd_capa *ocapa, int reserve)
1234 struct ptlrpc_request *req;
1235 struct ptlrpc_bulk_desc *desc;
1236 struct ost_body *body;
1237 struct obd_ioobj *ioobj;
1238 struct niobuf_remote *niobuf;
1239 int niocount, i, requested_nob, opc, rc;
1240 struct osc_brw_async_args *aa;
1241 struct req_capsule *pill;
1242 struct brw_page *pg_prev;
1245 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1246 RETURN(-ENOMEM); /* Recoverable */
1247 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1248 RETURN(-EINVAL); /* Fatal */
1250 if ((cmd & OBD_BRW_WRITE) != 0) {
1252 req = ptlrpc_request_alloc_pool(cli->cl_import,
1253 cli->cl_import->imp_rq_pool,
1257 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1262 for (niocount = i = 1; i < page_count; i++) {
1263 if (!can_merge_pages(pga[i - 1], pga[i]))
1267 pill = &req->rq_pill;
1268 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1270 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1271 niocount * sizeof(*niobuf));
1272 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1274 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1276 ptlrpc_request_free(req);
1279 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1280 ptlrpc_at_set_req_timeout(req);
1282 if (opc == OST_WRITE)
1283 desc = ptlrpc_prep_bulk_imp(req, page_count,
1284 BULK_GET_SOURCE, OST_BULK_PORTAL);
1286 desc = ptlrpc_prep_bulk_imp(req, page_count,
1287 BULK_PUT_SINK, OST_BULK_PORTAL);
1290 GOTO(out, rc = -ENOMEM);
1291 /* NB request now owns desc and will free it when it gets freed */
1293 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1294 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1295 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1296 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1298 lustre_set_wire_obdo(&body->oa, oa);
1300 obdo_to_ioobj(oa, ioobj);
1301 ioobj->ioo_bufcnt = niocount;
1302 osc_pack_capa(req, body, ocapa);
1303 LASSERT (page_count > 0);
1305 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1306 struct brw_page *pg = pga[i];
1308 LASSERT(pg->count > 0);
1309 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1310 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1311 pg->off, pg->count);
1313 LASSERTF(i == 0 || pg->off > pg_prev->off,
1314 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1315 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1317 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1318 pg_prev->pg, page_private(pg_prev->pg),
1319 pg_prev->pg->index, pg_prev->off);
1321 LASSERTF(i == 0 || pg->off > pg_prev->off,
1322 "i %d p_c %u\n", i, page_count);
1324 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1325 (pg->flag & OBD_BRW_SRVLOCK));
1327 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1329 requested_nob += pg->count;
1331 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1333 niobuf->len += pg->count;
1335 niobuf->offset = pg->off;
1336 niobuf->len = pg->count;
1337 niobuf->flags = pg->flag;
1342 LASSERTF((void *)(niobuf - niocount) ==
1343 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1344 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1345 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1347 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1348 if (osc_should_shrink_grant(cli))
1349 osc_shrink_grant_local(cli, &body->oa);
1351 /* size[REQ_REC_OFF] still sizeof (*body) */
1352 if (opc == OST_WRITE) {
1353 if (unlikely(cli->cl_checksum) &&
1354 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1355 /* store cl_cksum_type in a local variable since
1356 * it can be changed via lprocfs */
1357 cksum_type_t cksum_type = cli->cl_cksum_type;
1359 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1360 oa->o_flags &= OBD_FL_LOCAL_MASK;
1361 body->oa.o_flags = 0;
1363 body->oa.o_flags |= cksum_type_pack(cksum_type);
1364 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1365 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1369 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1371 /* save this in 'oa', too, for later checking */
1372 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1373 oa->o_flags |= cksum_type_pack(cksum_type);
1375 /* clear out the checksum flag, in case this is a
1376 * resend but cl_checksum is no longer set. b=11238 */
1377 oa->o_valid &= ~OBD_MD_FLCKSUM;
1379 oa->o_cksum = body->oa.o_cksum;
1380 /* 1 RC per niobuf */
1381 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1382 sizeof(__u32) * niocount);
1384 if (unlikely(cli->cl_checksum) &&
1385 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1386 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1387 body->oa.o_flags = 0;
1388 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1389 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1391 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER, 0);
1392 /* 1 RC for the whole I/O */
1394 ptlrpc_request_set_replen(req);
1396 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1397 aa = ptlrpc_req_async_args(req);
1399 aa->aa_requested_nob = requested_nob;
1400 aa->aa_nio_count = niocount;
1401 aa->aa_page_count = page_count;
1405 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1406 if (ocapa && reserve)
1407 aa->aa_ocapa = capa_get(ocapa);
1413 ptlrpc_req_finished(req);
1417 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1418 __u32 client_cksum, __u32 server_cksum, int nob,
1419 obd_count page_count, struct brw_page **pga,
1420 cksum_type_t client_cksum_type)
1424 cksum_type_t cksum_type;
1426 if (server_cksum == client_cksum) {
1427 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1431 if (oa->o_valid & OBD_MD_FLFLAGS)
1432 cksum_type = cksum_type_unpack(oa->o_flags);
1434 cksum_type = OBD_CKSUM_CRC32;
1436 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1439 if (cksum_type != client_cksum_type)
1440 msg = "the server did not use the checksum type specified in "
1441 "the original request - likely a protocol problem";
1442 else if (new_cksum == server_cksum)
1443 msg = "changed on the client after we checksummed it - "
1444 "likely false positive due to mmap IO (bug 11742)";
1445 else if (new_cksum == client_cksum)
1446 msg = "changed in transit before arrival at OST";
1448 msg = "changed in transit AND doesn't match the original - "
1449 "likely false positive due to mmap IO (bug 11742)";
1451 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1452 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1453 "["LPU64"-"LPU64"]\n",
1454 msg, libcfs_nid2str(peer->nid),
1455 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1456 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1459 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1461 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1462 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1463 "client csum now %x\n", client_cksum, client_cksum_type,
1464 server_cksum, cksum_type, new_cksum);
1468 /* Note rc enters this function as number of bytes transferred */
1469 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1471 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1472 const lnet_process_id_t *peer =
1473 &req->rq_import->imp_connection->c_peer;
1474 struct client_obd *cli = aa->aa_cli;
1475 struct ost_body *body;
1476 __u32 client_cksum = 0;
1479 if (rc < 0 && rc != -EDQUOT)
1482 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1483 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1485 CDEBUG(D_INFO, "Can't unpack body\n");
1489 /* set/clear over quota flag for a uid/gid */
1490 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1491 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1492 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1494 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1501 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1502 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1504 osc_update_grant(cli, body);
1506 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1508 CERROR("Unexpected +ve rc %d\n", rc);
1511 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1513 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1516 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1517 check_write_checksum(&body->oa, peer, client_cksum,
1518 body->oa.o_cksum, aa->aa_requested_nob,
1519 aa->aa_page_count, aa->aa_ppga,
1520 cksum_type_unpack(aa->aa_oa->o_flags)))
1523 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1524 aa->aa_page_count, aa->aa_ppga);
1528 /* The rest of this function executes only for OST_READs */
1530 /* if unwrap_bulk failed, return -EAGAIN to retry */
1531 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1533 GOTO(out, rc = -EAGAIN);
1535 if (rc > aa->aa_requested_nob) {
1536 CERROR("Unexpected rc %d (%d requested)\n", rc,
1537 aa->aa_requested_nob);
1541 if (rc != req->rq_bulk->bd_nob_transferred) {
1542 CERROR ("Unexpected rc %d (%d transferred)\n",
1543 rc, req->rq_bulk->bd_nob_transferred);
1547 if (rc < aa->aa_requested_nob)
1548 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1550 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1551 static int cksum_counter;
1552 __u32 server_cksum = body->oa.o_cksum;
1555 cksum_type_t cksum_type;
1557 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1558 cksum_type = cksum_type_unpack(body->oa.o_flags);
1560 cksum_type = OBD_CKSUM_CRC32;
1561 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1562 aa->aa_ppga, OST_READ,
1565 if (peer->nid == req->rq_bulk->bd_sender) {
1569 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1572 if (server_cksum == ~0 && rc > 0) {
1573 CERROR("Protocol error: server %s set the 'checksum' "
1574 "bit, but didn't send a checksum. Not fatal, "
1575 "but please notify on http://bugzilla.lustre.org/\n",
1576 libcfs_nid2str(peer->nid));
1577 } else if (server_cksum != client_cksum) {
1578 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1579 "%s%s%s inum "LPU64"/"LPU64" object "
1580 LPU64"/"LPU64" extent "
1581 "["LPU64"-"LPU64"]\n",
1582 req->rq_import->imp_obd->obd_name,
1583 libcfs_nid2str(peer->nid),
1585 body->oa.o_valid & OBD_MD_FLFID ?
1586 body->oa.o_fid : (__u64)0,
1587 body->oa.o_valid & OBD_MD_FLFID ?
1588 body->oa.o_generation :(__u64)0,
1590 body->oa.o_valid & OBD_MD_FLGROUP ?
1591 body->oa.o_gr : (__u64)0,
1592 aa->aa_ppga[0]->off,
1593 aa->aa_ppga[aa->aa_page_count-1]->off +
1594 aa->aa_ppga[aa->aa_page_count-1]->count -
1596 CERROR("client %x, server %x, cksum_type %x\n",
1597 client_cksum, server_cksum, cksum_type);
1599 aa->aa_oa->o_cksum = client_cksum;
1603 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1606 } else if (unlikely(client_cksum)) {
1607 static int cksum_missed;
1610 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1611 CERROR("Checksum %u requested from %s but not sent\n",
1612 cksum_missed, libcfs_nid2str(peer->nid));
1618 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1623 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1624 struct lov_stripe_md *lsm,
1625 obd_count page_count, struct brw_page **pga,
1626 struct obd_capa *ocapa)
1628 struct ptlrpc_request *req;
1632 struct l_wait_info lwi;
1636 cfs_waitq_init(&waitq);
1639 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1640 page_count, pga, &req, ocapa, 0);
1644 rc = ptlrpc_queue_wait(req);
1646 if (rc == -ETIMEDOUT && req->rq_resend) {
1647 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1648 ptlrpc_req_finished(req);
1652 rc = osc_brw_fini_request(req, rc);
1654 ptlrpc_req_finished(req);
1655 if (osc_recoverable_error(rc)) {
1657 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1658 CERROR("too many resend retries, returning error\n");
1662 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1663 l_wait_event(waitq, 0, &lwi);
1671 int osc_brw_redo_request(struct ptlrpc_request *request,
1672 struct osc_brw_async_args *aa)
1674 struct ptlrpc_request *new_req;
1675 struct ptlrpc_request_set *set = request->rq_set;
1676 struct osc_brw_async_args *new_aa;
1677 struct osc_async_page *oap;
1681 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1682 CERROR("too many resend retries, returning error\n");
1686 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1688 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1689 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1690 aa->aa_cli, aa->aa_oa,
1691 NULL /* lsm unused by osc currently */,
1692 aa->aa_page_count, aa->aa_ppga,
1693 &new_req, aa->aa_ocapa, 0);
1697 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1699 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1700 if (oap->oap_request != NULL) {
1701 LASSERTF(request == oap->oap_request,
1702 "request %p != oap_request %p\n",
1703 request, oap->oap_request);
1704 if (oap->oap_interrupted) {
1705 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1706 ptlrpc_req_finished(new_req);
1711 /* New request takes over pga and oaps from old request.
1712 * Note that copying a list_head doesn't work, need to move it... */
1714 new_req->rq_interpret_reply = request->rq_interpret_reply;
1715 new_req->rq_async_args = request->rq_async_args;
1716 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1718 new_aa = ptlrpc_req_async_args(new_req);
1720 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1721 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1722 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1724 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1725 if (oap->oap_request) {
1726 ptlrpc_req_finished(oap->oap_request);
1727 oap->oap_request = ptlrpc_request_addref(new_req);
1731 new_aa->aa_ocapa = aa->aa_ocapa;
1732 aa->aa_ocapa = NULL;
1734 /* use ptlrpc_set_add_req is safe because interpret functions work
1735 * in check_set context. only one way exist with access to request
1736 * from different thread got -EINTR - this way protected with
1737 * cl_loi_list_lock */
1738 ptlrpc_set_add_req(set, new_req);
1740 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1742 DEBUG_REQ(D_INFO, new_req, "new request");
1747 * ugh, we want disk allocation on the target to happen in offset order. we'll
1748 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1749 * fine for our small page arrays and doesn't require allocation. its an
1750 * insertion sort that swaps elements that are strides apart, shrinking the
1751 * stride down until its '1' and the array is sorted.
1753 static void sort_brw_pages(struct brw_page **array, int num)
1756 struct brw_page *tmp;
1760 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1765 for (i = stride ; i < num ; i++) {
1768 while (j >= stride && array[j - stride]->off > tmp->off) {
1769 array[j] = array[j - stride];
1774 } while (stride > 1);
1777 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1783 LASSERT (pages > 0);
1784 offset = pg[i]->off & ~CFS_PAGE_MASK;
1788 if (pages == 0) /* that's all */
1791 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1792 return count; /* doesn't end on page boundary */
1795 offset = pg[i]->off & ~CFS_PAGE_MASK;
1796 if (offset != 0) /* doesn't start on page boundary */
1803 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1805 struct brw_page **ppga;
1808 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1812 for (i = 0; i < count; i++)
1817 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1819 LASSERT(ppga != NULL);
1820 OBD_FREE(ppga, sizeof(*ppga) * count);
1823 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1824 obd_count page_count, struct brw_page *pga,
1825 struct obd_trans_info *oti)
1827 struct obdo *saved_oa = NULL;
1828 struct brw_page **ppga, **orig;
1829 struct obd_import *imp = class_exp2cliimp(exp);
1830 struct client_obd *cli;
1831 int rc, page_count_orig;
1834 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1835 cli = &imp->imp_obd->u.cli;
1837 if (cmd & OBD_BRW_CHECK) {
1838 /* The caller just wants to know if there's a chance that this
1839 * I/O can succeed */
1841 if (imp->imp_invalid)
1846 /* test_brw with a failed create can trip this, maybe others. */
1847 LASSERT(cli->cl_max_pages_per_rpc);
1851 orig = ppga = osc_build_ppga(pga, page_count);
1854 page_count_orig = page_count;
1856 sort_brw_pages(ppga, page_count);
1857 while (page_count) {
1858 obd_count pages_per_brw;
1860 if (page_count > cli->cl_max_pages_per_rpc)
1861 pages_per_brw = cli->cl_max_pages_per_rpc;
1863 pages_per_brw = page_count;
1865 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1867 if (saved_oa != NULL) {
1868 /* restore previously saved oa */
1869 *oinfo->oi_oa = *saved_oa;
1870 } else if (page_count > pages_per_brw) {
1871 /* save a copy of oa (brw will clobber it) */
1872 OBDO_ALLOC(saved_oa);
1873 if (saved_oa == NULL)
1874 GOTO(out, rc = -ENOMEM);
1875 *saved_oa = *oinfo->oi_oa;
1878 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1879 pages_per_brw, ppga, oinfo->oi_capa);
1884 page_count -= pages_per_brw;
1885 ppga += pages_per_brw;
1889 osc_release_ppga(orig, page_count_orig);
1891 if (saved_oa != NULL)
1892 OBDO_FREE(saved_oa);
1897 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1898 * the dirty accounting. Writeback completes or truncate happens before
1899 * writing starts. Must be called with the loi lock held. */
1900 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1903 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1907 /* This maintains the lists of pending pages to read/write for a given object
1908 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1909 * to quickly find objects that are ready to send an RPC. */
1910 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1916 if (lop->lop_num_pending == 0)
1919 /* if we have an invalid import we want to drain the queued pages
1920 * by forcing them through rpcs that immediately fail and complete
1921 * the pages. recovery relies on this to empty the queued pages
1922 * before canceling the locks and evicting down the llite pages */
1923 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1926 /* stream rpcs in queue order as long as as there is an urgent page
1927 * queued. this is our cheap solution for good batching in the case
1928 * where writepage marks some random page in the middle of the file
1929 * as urgent because of, say, memory pressure */
1930 if (!list_empty(&lop->lop_urgent)) {
1931 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1934 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1935 optimal = cli->cl_max_pages_per_rpc;
1936 if (cmd & OBD_BRW_WRITE) {
1937 /* trigger a write rpc stream as long as there are dirtiers
1938 * waiting for space. as they're waiting, they're not going to
1939 * create more pages to coallesce with what's waiting.. */
1940 if (!list_empty(&cli->cl_cache_waiters)) {
1941 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1944 /* +16 to avoid triggering rpcs that would want to include pages
1945 * that are being queued but which can't be made ready until
1946 * the queuer finishes with the page. this is a wart for
1947 * llite::commit_write() */
1950 if (lop->lop_num_pending >= optimal)
1956 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1958 struct osc_async_page *oap;
1961 if (list_empty(&lop->lop_urgent))
1964 oap = list_entry(lop->lop_urgent.next,
1965 struct osc_async_page, oap_urgent_item);
1967 if (oap->oap_async_flags & ASYNC_HP) {
1968 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1975 static void on_list(struct list_head *item, struct list_head *list,
1978 if (list_empty(item) && should_be_on)
1979 list_add_tail(item, list);
1980 else if (!list_empty(item) && !should_be_on)
1981 list_del_init(item);
1984 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1985 * can find pages to build into rpcs quickly */
1986 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1988 if (lop_makes_hprpc(&loi->loi_write_lop) ||
1989 lop_makes_hprpc(&loi->loi_read_lop)) {
1991 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1992 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1994 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1995 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1996 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1997 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2000 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2001 loi->loi_write_lop.lop_num_pending);
2003 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2004 loi->loi_read_lop.lop_num_pending);
2007 static void lop_update_pending(struct client_obd *cli,
2008 struct loi_oap_pages *lop, int cmd, int delta)
2010 lop->lop_num_pending += delta;
2011 if (cmd & OBD_BRW_WRITE)
2012 cli->cl_pending_w_pages += delta;
2014 cli->cl_pending_r_pages += delta;
2018 * this is called when a sync waiter receives an interruption. Its job is to
2019 * get the caller woken as soon as possible. If its page hasn't been put in an
2020 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2021 * desiring interruption which will forcefully complete the rpc once the rpc
2024 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2026 struct loi_oap_pages *lop;
2027 struct lov_oinfo *loi;
2031 LASSERT(!oap->oap_interrupted);
2032 oap->oap_interrupted = 1;
2034 /* ok, it's been put in an rpc. only one oap gets a request reference */
2035 if (oap->oap_request != NULL) {
2036 ptlrpc_mark_interrupted(oap->oap_request);
2037 ptlrpcd_wake(oap->oap_request);
2038 ptlrpc_req_finished(oap->oap_request);
2039 oap->oap_request = NULL;
2043 * page completion may be called only if ->cpo_prep() method was
2044 * executed by osc_io_submit(), that also adds page the to pending list
2046 if (!list_empty(&oap->oap_pending_item)) {
2047 list_del_init(&oap->oap_pending_item);
2048 list_del_init(&oap->oap_urgent_item);
2051 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2052 &loi->loi_write_lop : &loi->loi_read_lop;
2053 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2054 loi_list_maint(oap->oap_cli, oap->oap_loi);
2055 rc = oap->oap_caller_ops->ap_completion(env,
2056 oap->oap_caller_data,
2057 oap->oap_cmd, NULL, -EINTR);
2063 /* this is trying to propogate async writeback errors back up to the
2064 * application. As an async write fails we record the error code for later if
2065 * the app does an fsync. As long as errors persist we force future rpcs to be
2066 * sync so that the app can get a sync error and break the cycle of queueing
2067 * pages for which writeback will fail. */
2068 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2075 ar->ar_force_sync = 1;
2076 ar->ar_min_xid = ptlrpc_sample_next_xid();
2081 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2082 ar->ar_force_sync = 0;
2085 void osc_oap_to_pending(struct osc_async_page *oap)
2087 struct loi_oap_pages *lop;
2089 if (oap->oap_cmd & OBD_BRW_WRITE)
2090 lop = &oap->oap_loi->loi_write_lop;
2092 lop = &oap->oap_loi->loi_read_lop;
2094 if (oap->oap_async_flags & ASYNC_HP)
2095 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2096 else if (oap->oap_async_flags & ASYNC_URGENT)
2097 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2098 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2099 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2102 /* this must be called holding the loi list lock to give coverage to exit_cache,
2103 * async_flag maintenance, and oap_request */
2104 static void osc_ap_completion(const struct lu_env *env,
2105 struct client_obd *cli, struct obdo *oa,
2106 struct osc_async_page *oap, int sent, int rc)
2111 if (oap->oap_request != NULL) {
2112 xid = ptlrpc_req_xid(oap->oap_request);
2113 ptlrpc_req_finished(oap->oap_request);
2114 oap->oap_request = NULL;
2117 spin_lock(&oap->oap_lock);
2118 oap->oap_async_flags = 0;
2119 spin_unlock(&oap->oap_lock);
2120 oap->oap_interrupted = 0;
2122 if (oap->oap_cmd & OBD_BRW_WRITE) {
2123 osc_process_ar(&cli->cl_ar, xid, rc);
2124 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2127 if (rc == 0 && oa != NULL) {
2128 if (oa->o_valid & OBD_MD_FLBLOCKS)
2129 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2130 if (oa->o_valid & OBD_MD_FLMTIME)
2131 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2132 if (oa->o_valid & OBD_MD_FLATIME)
2133 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2134 if (oa->o_valid & OBD_MD_FLCTIME)
2135 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2138 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2139 oap->oap_cmd, oa, rc);
2141 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2142 * I/O on the page could start, but OSC calls it under lock
2143 * and thus we can add oap back to pending safely */
2145 /* upper layer wants to leave the page on pending queue */
2146 osc_oap_to_pending(oap);
2148 osc_exit_cache(cli, oap, sent);
2152 static int brw_interpret(const struct lu_env *env,
2153 struct ptlrpc_request *req, void *data, int rc)
2155 struct osc_brw_async_args *aa = data;
2156 struct client_obd *cli;
2160 rc = osc_brw_fini_request(req, rc);
2161 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2162 if (osc_recoverable_error(rc)) {
2163 rc = osc_brw_redo_request(req, aa);
2169 capa_put(aa->aa_ocapa);
2170 aa->aa_ocapa = NULL;
2175 client_obd_list_lock(&cli->cl_loi_list_lock);
2177 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2178 * is called so we know whether to go to sync BRWs or wait for more
2179 * RPCs to complete */
2180 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2181 cli->cl_w_in_flight--;
2183 cli->cl_r_in_flight--;
2185 async = list_empty(&aa->aa_oaps);
2186 if (!async) { /* from osc_send_oap_rpc() */
2187 struct osc_async_page *oap, *tmp;
2188 /* the caller may re-use the oap after the completion call so
2189 * we need to clean it up a little */
2190 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2191 list_del_init(&oap->oap_rpc_item);
2192 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2194 OBDO_FREE(aa->aa_oa);
2195 } else { /* from async_internal() */
2197 for (i = 0; i < aa->aa_page_count; i++)
2198 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2200 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2201 OBDO_FREE(aa->aa_oa);
2203 osc_wake_cache_waiters(cli);
2204 osc_check_rpcs(env, cli);
2205 client_obd_list_unlock(&cli->cl_loi_list_lock);
2207 cl_req_completion(env, aa->aa_clerq, rc);
2208 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2212 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2213 struct client_obd *cli,
2214 struct list_head *rpc_list,
2215 int page_count, int cmd)
2217 struct ptlrpc_request *req;
2218 struct brw_page **pga = NULL;
2219 struct osc_brw_async_args *aa;
2220 struct obdo *oa = NULL;
2221 const struct obd_async_page_ops *ops = NULL;
2222 void *caller_data = NULL;
2223 struct osc_async_page *oap;
2224 struct osc_async_page *tmp;
2225 struct ost_body *body;
2226 struct cl_req *clerq = NULL;
2227 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2228 struct ldlm_lock *lock = NULL;
2229 struct cl_req_attr crattr;
2233 LASSERT(!list_empty(rpc_list));
2235 memset(&crattr, 0, sizeof crattr);
2236 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2238 GOTO(out, req = ERR_PTR(-ENOMEM));
2242 GOTO(out, req = ERR_PTR(-ENOMEM));
2245 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2246 struct cl_page *page = osc_oap2cl_page(oap);
2248 ops = oap->oap_caller_ops;
2249 caller_data = oap->oap_caller_data;
2251 clerq = cl_req_alloc(env, page, crt,
2252 1 /* only 1-object rpcs for
2255 GOTO(out, req = (void *)clerq);
2256 lock = oap->oap_ldlm_lock;
2258 pga[i] = &oap->oap_brw_page;
2259 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2260 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2261 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2263 cl_req_page_add(env, clerq, page);
2266 /* always get the data for the obdo for the rpc */
2267 LASSERT(ops != NULL);
2269 crattr.cra_capa = NULL;
2270 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2272 oa->o_handle = lock->l_remote_handle;
2273 oa->o_valid |= OBD_MD_FLHANDLE;
2276 rc = cl_req_prep(env, clerq);
2278 CERROR("cl_req_prep failed: %d\n", rc);
2279 GOTO(out, req = ERR_PTR(rc));
2282 sort_brw_pages(pga, page_count);
2283 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2284 pga, &req, crattr.cra_capa, 1);
2286 CERROR("prep_req failed: %d\n", rc);
2287 GOTO(out, req = ERR_PTR(rc));
2290 /* Need to update the timestamps after the request is built in case
2291 * we race with setattr (locally or in queue at OST). If OST gets
2292 * later setattr before earlier BRW (as determined by the request xid),
2293 * the OST will not use BRW timestamps. Sadly, there is no obvious
2294 * way to do this in a single call. bug 10150 */
2295 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2296 cl_req_attr_set(env, clerq, &crattr,
2297 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2299 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2300 aa = ptlrpc_req_async_args(req);
2301 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2302 list_splice(rpc_list, &aa->aa_oaps);
2303 CFS_INIT_LIST_HEAD(rpc_list);
2304 aa->aa_clerq = clerq;
2306 capa_put(crattr.cra_capa);
2311 OBD_FREE(pga, sizeof(*pga) * page_count);
2312 /* this should happen rarely and is pretty bad, it makes the
2313 * pending list not follow the dirty order */
2314 client_obd_list_lock(&cli->cl_loi_list_lock);
2315 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2316 list_del_init(&oap->oap_rpc_item);
2318 /* queued sync pages can be torn down while the pages
2319 * were between the pending list and the rpc */
2320 if (oap->oap_interrupted) {
2321 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2322 osc_ap_completion(env, cli, NULL, oap, 0,
2326 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2328 if (clerq && !IS_ERR(clerq))
2329 cl_req_completion(env, clerq, PTR_ERR(req));
2335 * prepare pages for ASYNC io and put pages in send queue.
2337 * \param cmd OBD_BRW_* macroses
2338 * \param lop pending pages
2340 * \return zero if pages successfully add to send queue.
2341 * \return not zere if error occurring.
2344 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2345 struct lov_oinfo *loi,
2346 int cmd, struct loi_oap_pages *lop)
2348 struct ptlrpc_request *req;
2349 obd_count page_count = 0;
2350 struct osc_async_page *oap = NULL, *tmp;
2351 struct osc_brw_async_args *aa;
2352 const struct obd_async_page_ops *ops;
2353 CFS_LIST_HEAD(rpc_list);
2354 CFS_LIST_HEAD(tmp_list);
2355 unsigned int ending_offset;
2356 unsigned starting_offset = 0;
2358 struct cl_object *clob = NULL;
2361 /* ASYNC_HP pages first. At present, when the lock the pages is
2362 * to be canceled, the pages covered by the lock will be sent out
2363 * with ASYNC_HP. We have to send out them as soon as possible. */
2364 list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2365 if (oap->oap_async_flags & ASYNC_HP)
2366 list_move(&oap->oap_pending_item, &tmp_list);
2368 list_move_tail(&oap->oap_pending_item, &tmp_list);
2369 if (++page_count >= cli->cl_max_pages_per_rpc)
2373 list_splice(&tmp_list, &lop->lop_pending);
2376 /* first we find the pages we're allowed to work with */
2377 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2379 ops = oap->oap_caller_ops;
2381 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2382 "magic 0x%x\n", oap, oap->oap_magic);
2385 /* pin object in memory, so that completion call-backs
2386 * can be safely called under client_obd_list lock. */
2387 clob = osc_oap2cl_page(oap)->cp_obj;
2388 cl_object_get(clob);
2391 if (page_count != 0 &&
2392 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2393 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2394 " oap %p, page %p, srvlock %u\n",
2395 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2399 /* If there is a gap at the start of this page, it can't merge
2400 * with any previous page, so we'll hand the network a
2401 * "fragmented" page array that it can't transfer in 1 RDMA */
2402 if (page_count != 0 && oap->oap_page_off != 0)
2405 /* in llite being 'ready' equates to the page being locked
2406 * until completion unlocks it. commit_write submits a page
2407 * as not ready because its unlock will happen unconditionally
2408 * as the call returns. if we race with commit_write giving
2409 * us that page we dont' want to create a hole in the page
2410 * stream, so we stop and leave the rpc to be fired by
2411 * another dirtier or kupdated interval (the not ready page
2412 * will still be on the dirty list). we could call in
2413 * at the end of ll_file_write to process the queue again. */
2414 if (!(oap->oap_async_flags & ASYNC_READY)) {
2415 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2418 CDEBUG(D_INODE, "oap %p page %p returned %d "
2419 "instead of ready\n", oap,
2423 /* llite is telling us that the page is still
2424 * in commit_write and that we should try
2425 * and put it in an rpc again later. we
2426 * break out of the loop so we don't create
2427 * a hole in the sequence of pages in the rpc
2432 /* the io isn't needed.. tell the checks
2433 * below to complete the rpc with EINTR */
2434 spin_lock(&oap->oap_lock);
2435 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2436 spin_unlock(&oap->oap_lock);
2437 oap->oap_count = -EINTR;
2440 spin_lock(&oap->oap_lock);
2441 oap->oap_async_flags |= ASYNC_READY;
2442 spin_unlock(&oap->oap_lock);
2445 LASSERTF(0, "oap %p page %p returned %d "
2446 "from make_ready\n", oap,
2454 * Page submitted for IO has to be locked. Either by
2455 * ->ap_make_ready() or by higher layers.
2457 #if defined(__KERNEL__) && defined(__linux__)
2459 struct cl_page *page;
2461 page = osc_oap2cl_page(oap);
2463 if (page->cp_type == CPT_CACHEABLE &&
2464 !(PageLocked(oap->oap_page) &&
2465 (CheckWriteback(oap->oap_page, cmd)))) {
2466 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2468 (long)oap->oap_page->flags,
2469 oap->oap_async_flags);
2475 /* take the page out of our book-keeping */
2476 list_del_init(&oap->oap_pending_item);
2477 lop_update_pending(cli, lop, cmd, -1);
2478 list_del_init(&oap->oap_urgent_item);
2480 if (page_count == 0)
2481 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2482 (PTLRPC_MAX_BRW_SIZE - 1);
2484 /* ask the caller for the size of the io as the rpc leaves. */
2485 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2487 ops->ap_refresh_count(env, oap->oap_caller_data,
2489 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2491 if (oap->oap_count <= 0) {
2492 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2494 osc_ap_completion(env, cli, NULL,
2495 oap, 0, oap->oap_count);
2499 /* now put the page back in our accounting */
2500 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2501 if (page_count == 0)
2502 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2503 if (++page_count >= cli->cl_max_pages_per_rpc)
2506 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2507 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2508 * have the same alignment as the initial writes that allocated
2509 * extents on the server. */
2510 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2511 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2512 if (ending_offset == 0)
2515 /* If there is a gap at the end of this page, it can't merge
2516 * with any subsequent pages, so we'll hand the network a
2517 * "fragmented" page array that it can't transfer in 1 RDMA */
2518 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2522 osc_wake_cache_waiters(cli);
2524 loi_list_maint(cli, loi);
2526 client_obd_list_unlock(&cli->cl_loi_list_lock);
2529 cl_object_put(env, clob);
2531 if (page_count == 0) {
2532 client_obd_list_lock(&cli->cl_loi_list_lock);
2536 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2538 LASSERT(list_empty(&rpc_list));
2539 loi_list_maint(cli, loi);
2540 RETURN(PTR_ERR(req));
2543 aa = ptlrpc_req_async_args(req);
2545 if (cmd == OBD_BRW_READ) {
2546 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2547 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2548 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2549 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2551 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2552 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2553 cli->cl_w_in_flight);
2554 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2555 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2557 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2559 client_obd_list_lock(&cli->cl_loi_list_lock);
2561 if (cmd == OBD_BRW_READ)
2562 cli->cl_r_in_flight++;
2564 cli->cl_w_in_flight++;
2566 /* queued sync pages can be torn down while the pages
2567 * were between the pending list and the rpc */
2569 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2570 /* only one oap gets a request reference */
2573 if (oap->oap_interrupted && !req->rq_intr) {
2574 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2576 ptlrpc_mark_interrupted(req);
2580 tmp->oap_request = ptlrpc_request_addref(req);
2582 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2583 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2585 req->rq_interpret_reply = brw_interpret;
2586 ptlrpcd_add_req(req, PSCOPE_BRW);
2590 #define LOI_DEBUG(LOI, STR, args...) \
2591 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2592 !list_empty(&(LOI)->loi_ready_item) || \
2593 !list_empty(&(LOI)->loi_hp_ready_item), \
2594 (LOI)->loi_write_lop.lop_num_pending, \
2595 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2596 (LOI)->loi_read_lop.lop_num_pending, \
2597 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2600 /* This is called by osc_check_rpcs() to find which objects have pages that
2601 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2602 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2606 /* First return objects that have blocked locks so that they
2607 * will be flushed quickly and other clients can get the lock,
2608 * then objects which have pages ready to be stuffed into RPCs */
2609 if (!list_empty(&cli->cl_loi_hp_ready_list))
2610 RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
2611 struct lov_oinfo, loi_hp_ready_item));
2612 if (!list_empty(&cli->cl_loi_ready_list))
2613 RETURN(list_entry(cli->cl_loi_ready_list.next,
2614 struct lov_oinfo, loi_ready_item));
2616 /* then if we have cache waiters, return all objects with queued
2617 * writes. This is especially important when many small files
2618 * have filled up the cache and not been fired into rpcs because
2619 * they don't pass the nr_pending/object threshhold */
2620 if (!list_empty(&cli->cl_cache_waiters) &&
2621 !list_empty(&cli->cl_loi_write_list))
2622 RETURN(list_entry(cli->cl_loi_write_list.next,
2623 struct lov_oinfo, loi_write_item));
2625 /* then return all queued objects when we have an invalid import
2626 * so that they get flushed */
2627 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2628 if (!list_empty(&cli->cl_loi_write_list))
2629 RETURN(list_entry(cli->cl_loi_write_list.next,
2630 struct lov_oinfo, loi_write_item));
2631 if (!list_empty(&cli->cl_loi_read_list))
2632 RETURN(list_entry(cli->cl_loi_read_list.next,
2633 struct lov_oinfo, loi_read_item));
2638 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2640 struct osc_async_page *oap;
2643 if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
2644 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2645 struct osc_async_page, oap_urgent_item);
2646 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2649 if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
2650 oap = list_entry(loi->loi_read_lop.lop_urgent.next,
2651 struct osc_async_page, oap_urgent_item);
2652 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2655 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2658 /* called with the loi list lock held */
2659 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2661 struct lov_oinfo *loi;
2662 int rc = 0, race_counter = 0;
2665 while ((loi = osc_next_loi(cli)) != NULL) {
2666 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2668 if (osc_max_rpc_in_flight(cli, loi))
2671 /* attempt some read/write balancing by alternating between
2672 * reads and writes in an object. The makes_rpc checks here
2673 * would be redundant if we were getting read/write work items
2674 * instead of objects. we don't want send_oap_rpc to drain a
2675 * partial read pending queue when we're given this object to
2676 * do io on writes while there are cache waiters */
2677 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2678 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2679 &loi->loi_write_lop);
2681 CERROR("Write request failed with %d\n", rc);
2683 /* osc_send_oap_rpc failed, mostly because of
2686 * It can't break here, because if:
2687 * - a page was submitted by osc_io_submit, so
2689 * - no request in flight
2690 * - no subsequent request
2691 * The system will be in live-lock state,
2692 * because there is no chance to call
2693 * osc_io_unplug() and osc_check_rpcs() any
2694 * more. pdflush can't help in this case,
2695 * because it might be blocked at grabbing
2696 * the page lock as we mentioned.
2698 * Anyway, continue to drain pages. */
2707 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2708 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2709 &loi->loi_read_lop);
2711 CERROR("Read request failed with %d\n", rc);
2719 /* attempt some inter-object balancing by issueing rpcs
2720 * for each object in turn */
2721 if (!list_empty(&loi->loi_hp_ready_item))
2722 list_del_init(&loi->loi_hp_ready_item);
2723 if (!list_empty(&loi->loi_ready_item))
2724 list_del_init(&loi->loi_ready_item);
2725 if (!list_empty(&loi->loi_write_item))
2726 list_del_init(&loi->loi_write_item);
2727 if (!list_empty(&loi->loi_read_item))
2728 list_del_init(&loi->loi_read_item);
2730 loi_list_maint(cli, loi);
2732 /* send_oap_rpc fails with 0 when make_ready tells it to
2733 * back off. llite's make_ready does this when it tries
2734 * to lock a page queued for write that is already locked.
2735 * we want to try sending rpcs from many objects, but we
2736 * don't want to spin failing with 0. */
2737 if (race_counter == 10)
2743 /* we're trying to queue a page in the osc so we're subject to the
2744 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2745 * If the osc's queued pages are already at that limit, then we want to sleep
2746 * until there is space in the osc's queue for us. We also may be waiting for
2747 * write credits from the OST if there are RPCs in flight that may return some
2748 * before we fall back to sync writes.
2750 * We need this know our allocation was granted in the presence of signals */
2751 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2755 client_obd_list_lock(&cli->cl_loi_list_lock);
2756 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2757 client_obd_list_unlock(&cli->cl_loi_list_lock);
2762 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2765 int osc_enter_cache_try(const struct lu_env *env,
2766 struct client_obd *cli, struct lov_oinfo *loi,
2767 struct osc_async_page *oap, int transient)
2771 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2773 osc_consume_write_grant(cli, &oap->oap_brw_page);
2775 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2776 atomic_inc(&obd_dirty_transit_pages);
2777 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2783 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2784 * grant or cache space. */
2785 static int osc_enter_cache(const struct lu_env *env,
2786 struct client_obd *cli, struct lov_oinfo *loi,
2787 struct osc_async_page *oap)
2789 struct osc_cache_waiter ocw;
2790 struct l_wait_info lwi = { 0 };
2794 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2795 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2796 cli->cl_dirty_max, obd_max_dirty_pages,
2797 cli->cl_lost_grant, cli->cl_avail_grant);
2799 /* force the caller to try sync io. this can jump the list
2800 * of queued writes and create a discontiguous rpc stream */
2801 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2802 loi->loi_ar.ar_force_sync)
2805 /* Hopefully normal case - cache space and write credits available */
2806 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2807 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2808 osc_enter_cache_try(env, cli, loi, oap, 0))
2811 /* Make sure that there are write rpcs in flight to wait for. This
2812 * is a little silly as this object may not have any pending but
2813 * other objects sure might. */
2814 if (cli->cl_w_in_flight) {
2815 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2816 cfs_waitq_init(&ocw.ocw_waitq);
2820 loi_list_maint(cli, loi);
2821 osc_check_rpcs(env, cli);
2822 client_obd_list_unlock(&cli->cl_loi_list_lock);
2824 CDEBUG(D_CACHE, "sleeping for cache space\n");
2825 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2827 client_obd_list_lock(&cli->cl_loi_list_lock);
2828 if (!list_empty(&ocw.ocw_entry)) {
2829 list_del(&ocw.ocw_entry);
2839 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2840 struct lov_oinfo *loi, cfs_page_t *page,
2841 obd_off offset, const struct obd_async_page_ops *ops,
2842 void *data, void **res, int nocache,
2843 struct lustre_handle *lockh)
2845 struct osc_async_page *oap;
2850 return size_round(sizeof(*oap));
2853 oap->oap_magic = OAP_MAGIC;
2854 oap->oap_cli = &exp->exp_obd->u.cli;
2857 oap->oap_caller_ops = ops;
2858 oap->oap_caller_data = data;
2860 oap->oap_page = page;
2861 oap->oap_obj_off = offset;
2862 if (!client_is_remote(exp) &&
2863 cfs_capable(CFS_CAP_SYS_RESOURCE))
2864 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2866 LASSERT(!(offset & ~CFS_PAGE_MASK));
2868 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2869 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2870 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2871 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2873 spin_lock_init(&oap->oap_lock);
2874 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2878 struct osc_async_page *oap_from_cookie(void *cookie)
2880 struct osc_async_page *oap = cookie;
2881 if (oap->oap_magic != OAP_MAGIC)
2882 return ERR_PTR(-EINVAL);
2886 int osc_queue_async_io(const struct lu_env *env,
2887 struct obd_export *exp, struct lov_stripe_md *lsm,
2888 struct lov_oinfo *loi, void *cookie,
2889 int cmd, obd_off off, int count,
2890 obd_flag brw_flags, enum async_flags async_flags)
2892 struct client_obd *cli = &exp->exp_obd->u.cli;
2893 struct osc_async_page *oap;
2897 oap = oap_from_cookie(cookie);
2899 RETURN(PTR_ERR(oap));
2901 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2904 if (!list_empty(&oap->oap_pending_item) ||
2905 !list_empty(&oap->oap_urgent_item) ||
2906 !list_empty(&oap->oap_rpc_item))
2909 /* check if the file's owner/group is over quota */
2910 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2911 struct cl_object *obj;
2912 struct cl_attr attr; /* XXX put attr into thread info */
2913 unsigned int qid[MAXQUOTAS];
2915 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2917 cl_object_attr_lock(obj);
2918 rc = cl_object_attr_get(env, obj, &attr);
2919 cl_object_attr_unlock(obj);
2921 qid[USRQUOTA] = attr.cat_uid;
2922 qid[GRPQUOTA] = attr.cat_gid;
2924 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2931 loi = lsm->lsm_oinfo[0];
2933 client_obd_list_lock(&cli->cl_loi_list_lock);
2935 LASSERT(off + count <= CFS_PAGE_SIZE);
2937 oap->oap_page_off = off;
2938 oap->oap_count = count;
2939 oap->oap_brw_flags = brw_flags;
2940 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2941 if (libcfs_memory_pressure_get())
2942 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2943 spin_lock(&oap->oap_lock);
2944 oap->oap_async_flags = async_flags;
2945 spin_unlock(&oap->oap_lock);
2947 if (cmd & OBD_BRW_WRITE) {
2948 rc = osc_enter_cache(env, cli, loi, oap);
2950 client_obd_list_unlock(&cli->cl_loi_list_lock);
2955 osc_oap_to_pending(oap);
2956 loi_list_maint(cli, loi);
2958 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2961 osc_check_rpcs(env, cli);
2962 client_obd_list_unlock(&cli->cl_loi_list_lock);
2967 /* aka (~was & now & flag), but this is more clear :) */
2968 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2970 int osc_set_async_flags_base(struct client_obd *cli,
2971 struct lov_oinfo *loi, struct osc_async_page *oap,
2972 obd_flag async_flags)
2974 struct loi_oap_pages *lop;
2978 LASSERT(!list_empty(&oap->oap_pending_item));
2980 if (oap->oap_cmd & OBD_BRW_WRITE) {
2981 lop = &loi->loi_write_lop;
2983 lop = &loi->loi_read_lop;
2986 if ((oap->oap_async_flags & async_flags) == async_flags)
2989 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2990 flags |= ASYNC_READY;
2992 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
2993 list_empty(&oap->oap_rpc_item)) {
2994 if (oap->oap_async_flags & ASYNC_HP)
2995 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2997 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2998 flags |= ASYNC_URGENT;
2999 loi_list_maint(cli, loi);
3001 spin_lock(&oap->oap_lock);
3002 oap->oap_async_flags |= flags;
3003 spin_unlock(&oap->oap_lock);
3005 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3006 oap->oap_async_flags);
3010 int osc_teardown_async_page(struct obd_export *exp,
3011 struct lov_stripe_md *lsm,
3012 struct lov_oinfo *loi, void *cookie)
3014 struct client_obd *cli = &exp->exp_obd->u.cli;
3015 struct loi_oap_pages *lop;
3016 struct osc_async_page *oap;
3020 oap = oap_from_cookie(cookie);
3022 RETURN(PTR_ERR(oap));
3025 loi = lsm->lsm_oinfo[0];
3027 if (oap->oap_cmd & OBD_BRW_WRITE) {
3028 lop = &loi->loi_write_lop;
3030 lop = &loi->loi_read_lop;
3033 client_obd_list_lock(&cli->cl_loi_list_lock);
3035 if (!list_empty(&oap->oap_rpc_item))
3036 GOTO(out, rc = -EBUSY);
3038 osc_exit_cache(cli, oap, 0);
3039 osc_wake_cache_waiters(cli);
3041 if (!list_empty(&oap->oap_urgent_item)) {
3042 list_del_init(&oap->oap_urgent_item);
3043 spin_lock(&oap->oap_lock);
3044 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3045 spin_unlock(&oap->oap_lock);
3047 if (!list_empty(&oap->oap_pending_item)) {
3048 list_del_init(&oap->oap_pending_item);
3049 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3051 loi_list_maint(cli, loi);
3052 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3054 client_obd_list_unlock(&cli->cl_loi_list_lock);
3058 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3059 struct ldlm_enqueue_info *einfo,
3062 void *data = einfo->ei_cbdata;
3064 LASSERT(lock != NULL);
3065 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3066 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3067 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3068 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3070 lock_res_and_lock(lock);
3071 spin_lock(&osc_ast_guard);
3072 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3073 lock->l_ast_data = data;
3074 spin_unlock(&osc_ast_guard);
3075 unlock_res_and_lock(lock);
3078 static void osc_set_data_with_check(struct lustre_handle *lockh,
3079 struct ldlm_enqueue_info *einfo,
3082 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3085 osc_set_lock_data_with_check(lock, einfo, flags);
3086 LDLM_LOCK_PUT(lock);
3088 CERROR("lockh %p, data %p - client evicted?\n",
3089 lockh, einfo->ei_cbdata);
3092 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3093 ldlm_iterator_t replace, void *data)
3095 struct ldlm_res_id res_id;
3096 struct obd_device *obd = class_exp2obd(exp);
3098 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3099 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3103 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3104 obd_enqueue_update_f upcall, void *cookie,
3107 int intent = *flags & LDLM_FL_HAS_INTENT;
3111 /* The request was created before ldlm_cli_enqueue call. */
3112 if (rc == ELDLM_LOCK_ABORTED) {
3113 struct ldlm_reply *rep;
3114 rep = req_capsule_server_get(&req->rq_pill,
3117 LASSERT(rep != NULL);
3118 if (rep->lock_policy_res1)
3119 rc = rep->lock_policy_res1;
3123 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3124 *flags |= LDLM_FL_LVB_READY;
3125 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3126 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3129 /* Call the update callback. */
3130 rc = (*upcall)(cookie, rc);
3134 static int osc_enqueue_interpret(const struct lu_env *env,
3135 struct ptlrpc_request *req,
3136 struct osc_enqueue_args *aa, int rc)
3138 struct ldlm_lock *lock;
3139 struct lustre_handle handle;
3142 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3143 * might be freed anytime after lock upcall has been called. */
3144 lustre_handle_copy(&handle, aa->oa_lockh);
3145 mode = aa->oa_ei->ei_mode;
3147 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3149 lock = ldlm_handle2lock(&handle);
3151 /* Take an additional reference so that a blocking AST that
3152 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3153 * to arrive after an upcall has been executed by
3154 * osc_enqueue_fini(). */
3155 ldlm_lock_addref(&handle, mode);
3157 /* Complete obtaining the lock procedure. */
3158 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3159 mode, aa->oa_flags, aa->oa_lvb,
3160 sizeof(*aa->oa_lvb), &handle, rc);
3161 /* Complete osc stuff. */
3162 rc = osc_enqueue_fini(req, aa->oa_lvb,
3163 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3165 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3167 /* Release the lock for async request. */
3168 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3170 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3171 * not already released by
3172 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3174 ldlm_lock_decref(&handle, mode);
3176 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3177 aa->oa_lockh, req, aa);
3178 ldlm_lock_decref(&handle, mode);
3179 LDLM_LOCK_PUT(lock);
3183 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3184 struct lov_oinfo *loi, int flags,
3185 struct ost_lvb *lvb, __u32 mode, int rc)
3187 if (rc == ELDLM_OK) {
3188 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3191 LASSERT(lock != NULL);
3192 loi->loi_lvb = *lvb;
3193 tmp = loi->loi_lvb.lvb_size;
3194 /* Extend KMS up to the end of this lock and no further
3195 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3196 if (tmp > lock->l_policy_data.l_extent.end)
3197 tmp = lock->l_policy_data.l_extent.end + 1;
3198 if (tmp >= loi->loi_kms) {
3199 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3200 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3201 loi_kms_set(loi, tmp);
3203 LDLM_DEBUG(lock, "lock acquired, setting rss="
3204 LPU64"; leaving kms="LPU64", end="LPU64,
3205 loi->loi_lvb.lvb_size, loi->loi_kms,
3206 lock->l_policy_data.l_extent.end);
3208 ldlm_lock_allow_match(lock);
3209 LDLM_LOCK_PUT(lock);
3210 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3211 loi->loi_lvb = *lvb;
3212 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3213 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3217 EXPORT_SYMBOL(osc_update_enqueue);
3219 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3221 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3222 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3223 * other synchronous requests, however keeping some locks and trying to obtain
3224 * others may take a considerable amount of time in a case of ost failure; and
3225 * when other sync requests do not get released lock from a client, the client
3226 * is excluded from the cluster -- such scenarious make the life difficult, so
3227 * release locks just after they are obtained. */
3228 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3229 int *flags, ldlm_policy_data_t *policy,
3230 struct ost_lvb *lvb, int kms_valid,
3231 obd_enqueue_update_f upcall, void *cookie,
3232 struct ldlm_enqueue_info *einfo,
3233 struct lustre_handle *lockh,
3234 struct ptlrpc_request_set *rqset, int async)
3236 struct obd_device *obd = exp->exp_obd;
3237 struct ptlrpc_request *req = NULL;
3238 int intent = *flags & LDLM_FL_HAS_INTENT;
3243 /* Filesystem lock extents are extended to page boundaries so that
3244 * dealing with the page cache is a little smoother. */
3245 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3246 policy->l_extent.end |= ~CFS_PAGE_MASK;
3249 * kms is not valid when either object is completely fresh (so that no
3250 * locks are cached), or object was evicted. In the latter case cached
3251 * lock cannot be used, because it would prime inode state with
3252 * potentially stale LVB.
3257 /* Next, search for already existing extent locks that will cover us */
3258 /* If we're trying to read, we also search for an existing PW lock. The
3259 * VFS and page cache already protect us locally, so lots of readers/
3260 * writers can share a single PW lock.
3262 * There are problems with conversion deadlocks, so instead of
3263 * converting a read lock to a write lock, we'll just enqueue a new
3266 * At some point we should cancel the read lock instead of making them
3267 * send us a blocking callback, but there are problems with canceling
3268 * locks out from other users right now, too. */
3269 mode = einfo->ei_mode;
3270 if (einfo->ei_mode == LCK_PR)
3272 mode = ldlm_lock_match(obd->obd_namespace,
3273 *flags | LDLM_FL_LVB_READY, res_id,
3274 einfo->ei_type, policy, mode, lockh, 0);
3276 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3278 if (matched->l_ast_data == NULL ||
3279 matched->l_ast_data == einfo->ei_cbdata) {
3280 /* addref the lock only if not async requests and PW
3281 * lock is matched whereas we asked for PR. */
3282 if (!rqset && einfo->ei_mode != mode)
3283 ldlm_lock_addref(lockh, LCK_PR);
3284 osc_set_lock_data_with_check(matched, einfo, *flags);
3286 /* I would like to be able to ASSERT here that
3287 * rss <= kms, but I can't, for reasons which
3288 * are explained in lov_enqueue() */
3291 /* We already have a lock, and it's referenced */
3292 (*upcall)(cookie, ELDLM_OK);
3294 /* For async requests, decref the lock. */
3295 if (einfo->ei_mode != mode)
3296 ldlm_lock_decref(lockh, LCK_PW);
3298 ldlm_lock_decref(lockh, einfo->ei_mode);
3299 LDLM_LOCK_PUT(matched);
3302 ldlm_lock_decref(lockh, mode);
3303 LDLM_LOCK_PUT(matched);
3308 CFS_LIST_HEAD(cancels);
3309 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3310 &RQF_LDLM_ENQUEUE_LVB);
3314 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3318 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3320 ptlrpc_request_set_replen(req);
3323 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3324 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3326 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3327 sizeof(*lvb), lockh, async);
3330 struct osc_enqueue_args *aa;
3331 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3332 aa = ptlrpc_req_async_args(req);
3335 aa->oa_flags = flags;
3336 aa->oa_upcall = upcall;
3337 aa->oa_cookie = cookie;
3339 aa->oa_lockh = lockh;
3341 req->rq_interpret_reply =
3342 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3343 if (rqset == PTLRPCD_SET)
3344 ptlrpcd_add_req(req, PSCOPE_OTHER);
3346 ptlrpc_set_add_req(rqset, req);
3347 } else if (intent) {
3348 ptlrpc_req_finished(req);
3353 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3355 ptlrpc_req_finished(req);
3360 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3361 struct ldlm_enqueue_info *einfo,
3362 struct ptlrpc_request_set *rqset)
3364 struct ldlm_res_id res_id;
3368 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3369 oinfo->oi_md->lsm_object_gr, &res_id);
3371 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3372 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3373 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3374 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3375 rqset, rqset != NULL);
3379 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3380 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3381 int *flags, void *data, struct lustre_handle *lockh,
3384 struct obd_device *obd = exp->exp_obd;
3385 int lflags = *flags;
3389 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3392 /* Filesystem lock extents are extended to page boundaries so that
3393 * dealing with the page cache is a little smoother */
3394 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3395 policy->l_extent.end |= ~CFS_PAGE_MASK;
3397 /* Next, search for already existing extent locks that will cover us */
3398 /* If we're trying to read, we also search for an existing PW lock. The
3399 * VFS and page cache already protect us locally, so lots of readers/
3400 * writers can share a single PW lock. */
3404 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3405 res_id, type, policy, rc, lockh, unref);
3408 osc_set_data_with_check(lockh, data, lflags);
3409 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3410 ldlm_lock_addref(lockh, LCK_PR);
3411 ldlm_lock_decref(lockh, LCK_PW);
3418 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3422 if (unlikely(mode == LCK_GROUP))
3423 ldlm_lock_decref_and_cancel(lockh, mode);
3425 ldlm_lock_decref(lockh, mode);
3430 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3431 __u32 mode, struct lustre_handle *lockh)
3434 RETURN(osc_cancel_base(lockh, mode));
3437 static int osc_cancel_unused(struct obd_export *exp,
3438 struct lov_stripe_md *lsm, int flags,
3441 struct obd_device *obd = class_exp2obd(exp);
3442 struct ldlm_res_id res_id, *resp = NULL;
3445 resp = osc_build_res_name(lsm->lsm_object_id,
3446 lsm->lsm_object_gr, &res_id);
3449 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3452 static int osc_statfs_interpret(const struct lu_env *env,
3453 struct ptlrpc_request *req,
3454 struct osc_async_args *aa, int rc)
3456 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3457 struct obd_statfs *msfs;
3462 /* The request has in fact never been sent
3463 * due to issues at a higher level (LOV).
3464 * Exit immediately since the caller is
3465 * aware of the problem and takes care
3466 * of the clean up */
3469 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3470 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3476 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3478 GOTO(out, rc = -EPROTO);
3481 /* Reinitialize the RDONLY and DEGRADED flags at the client
3482 * on each statfs, so they don't stay set permanently. */
3483 spin_lock(&cli->cl_oscc.oscc_lock);
3485 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3486 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3487 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3488 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3490 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3491 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3492 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3493 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3495 /* Add a bit of hysteresis so this flag isn't continually flapping,
3496 * and ensure that new files don't get extremely fragmented due to
3497 * only a small amount of available space in the filesystem.
3498 * We want to set the NOSPC flag when there is less than ~0.1% free
3499 * and clear it when there is at least ~0.2% free space, so:
3500 * avail < ~0.1% max max = avail + used
3501 * 1025 * avail < avail + used used = blocks - free
3502 * 1024 * avail < used
3503 * 1024 * avail < blocks - free
3504 * avail < ((blocks - free) >> 10)
3506 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3507 * lose that amount of space so in those cases we report no space left
3508 * if their is less than 1 GB left. */
3509 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3510 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3511 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3512 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3513 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3514 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3515 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3517 spin_unlock(&cli->cl_oscc.oscc_lock);
3519 *aa->aa_oi->oi_osfs = *msfs;
3521 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3525 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3526 __u64 max_age, struct ptlrpc_request_set *rqset)
3528 struct ptlrpc_request *req;
3529 struct osc_async_args *aa;
3533 /* We could possibly pass max_age in the request (as an absolute
3534 * timestamp or a "seconds.usec ago") so the target can avoid doing
3535 * extra calls into the filesystem if that isn't necessary (e.g.
3536 * during mount that would help a bit). Having relative timestamps
3537 * is not so great if request processing is slow, while absolute
3538 * timestamps are not ideal because they need time synchronization. */
3539 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3543 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3545 ptlrpc_request_free(req);
3548 ptlrpc_request_set_replen(req);
3549 req->rq_request_portal = OST_CREATE_PORTAL;
3550 ptlrpc_at_set_req_timeout(req);
3552 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3553 /* procfs requests not want stat in wait for avoid deadlock */
3554 req->rq_no_resend = 1;
3555 req->rq_no_delay = 1;
3558 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3559 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3560 aa = ptlrpc_req_async_args(req);
3563 ptlrpc_set_add_req(rqset, req);
3567 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3568 __u64 max_age, __u32 flags)
3570 struct obd_statfs *msfs;
3571 struct ptlrpc_request *req;
3572 struct obd_import *imp = NULL;
3576 /*Since the request might also come from lprocfs, so we need
3577 *sync this with client_disconnect_export Bug15684*/
3578 down_read(&obd->u.cli.cl_sem);
3579 if (obd->u.cli.cl_import)
3580 imp = class_import_get(obd->u.cli.cl_import);
3581 up_read(&obd->u.cli.cl_sem);
3585 /* We could possibly pass max_age in the request (as an absolute
3586 * timestamp or a "seconds.usec ago") so the target can avoid doing
3587 * extra calls into the filesystem if that isn't necessary (e.g.
3588 * during mount that would help a bit). Having relative timestamps
3589 * is not so great if request processing is slow, while absolute
3590 * timestamps are not ideal because they need time synchronization. */
3591 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3593 class_import_put(imp);
3598 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3600 ptlrpc_request_free(req);
3603 ptlrpc_request_set_replen(req);
3604 req->rq_request_portal = OST_CREATE_PORTAL;
3605 ptlrpc_at_set_req_timeout(req);
3607 if (flags & OBD_STATFS_NODELAY) {
3608 /* procfs requests not want stat in wait for avoid deadlock */
3609 req->rq_no_resend = 1;
3610 req->rq_no_delay = 1;
3613 rc = ptlrpc_queue_wait(req);
3617 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3619 GOTO(out, rc = -EPROTO);
3626 ptlrpc_req_finished(req);
3630 /* Retrieve object striping information.
3632 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3633 * the maximum number of OST indices which will fit in the user buffer.
3634 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3636 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3638 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3639 struct lov_user_md_v3 lum, *lumk;
3640 struct lov_user_ost_data_v1 *lmm_objects;
3641 int rc = 0, lum_size;
3647 /* we only need the header part from user space to get lmm_magic and
3648 * lmm_stripe_count, (the header part is common to v1 and v3) */
3649 lum_size = sizeof(struct lov_user_md_v1);
3650 if (copy_from_user(&lum, lump, lum_size))
3653 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3654 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3657 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3658 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3659 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3660 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3662 /* we can use lov_mds_md_size() to compute lum_size
3663 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3664 if (lum.lmm_stripe_count > 0) {
3665 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3666 OBD_ALLOC(lumk, lum_size);
3670 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3671 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3673 lmm_objects = &(lumk->lmm_objects[0]);
3674 lmm_objects->l_object_id = lsm->lsm_object_id;
3676 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3680 lumk->lmm_object_id = lsm->lsm_object_id;
3681 lumk->lmm_object_gr = lsm->lsm_object_gr;
3682 lumk->lmm_stripe_count = 1;
3684 if (copy_to_user(lump, lumk, lum_size))
3688 OBD_FREE(lumk, lum_size);
3694 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3695 void *karg, void *uarg)
3697 struct obd_device *obd = exp->exp_obd;
3698 struct obd_ioctl_data *data = karg;
3702 if (!try_module_get(THIS_MODULE)) {
3703 CERROR("Can't get module. Is it alive?");
3707 case OBD_IOC_LOV_GET_CONFIG: {
3709 struct lov_desc *desc;
3710 struct obd_uuid uuid;
3714 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3715 GOTO(out, err = -EINVAL);
3717 data = (struct obd_ioctl_data *)buf;
3719 if (sizeof(*desc) > data->ioc_inllen1) {
3720 obd_ioctl_freedata(buf, len);
3721 GOTO(out, err = -EINVAL);
3724 if (data->ioc_inllen2 < sizeof(uuid)) {
3725 obd_ioctl_freedata(buf, len);
3726 GOTO(out, err = -EINVAL);
3729 desc = (struct lov_desc *)data->ioc_inlbuf1;
3730 desc->ld_tgt_count = 1;
3731 desc->ld_active_tgt_count = 1;
3732 desc->ld_default_stripe_count = 1;
3733 desc->ld_default_stripe_size = 0;
3734 desc->ld_default_stripe_offset = 0;
3735 desc->ld_pattern = 0;
3736 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3738 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3740 err = copy_to_user((void *)uarg, buf, len);
3743 obd_ioctl_freedata(buf, len);
3746 case LL_IOC_LOV_SETSTRIPE:
3747 err = obd_alloc_memmd(exp, karg);
3751 case LL_IOC_LOV_GETSTRIPE:
3752 err = osc_getstripe(karg, uarg);
3754 case OBD_IOC_CLIENT_RECOVER:
3755 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3760 case IOC_OSC_SET_ACTIVE:
3761 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3764 case OBD_IOC_POLL_QUOTACHECK:
3765 err = lquota_poll_check(quota_interface, exp,
3766 (struct if_quotacheck *)karg);
3768 case OBD_IOC_PING_TARGET:
3769 err = ptlrpc_obd_ping(obd);
3772 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3773 cmd, cfs_curproc_comm());
3774 GOTO(out, err = -ENOTTY);
3777 module_put(THIS_MODULE);
3781 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3782 void *key, __u32 *vallen, void *val,
3783 struct lov_stripe_md *lsm)
3786 if (!vallen || !val)
3789 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3790 __u32 *stripe = val;
3791 *vallen = sizeof(*stripe);
3794 } else if (KEY_IS(KEY_LAST_ID)) {
3795 struct ptlrpc_request *req;
3800 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3801 &RQF_OST_GET_INFO_LAST_ID);
3805 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3806 RCL_CLIENT, keylen);
3807 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3809 ptlrpc_request_free(req);
3813 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3814 memcpy(tmp, key, keylen);
3816 req->rq_no_delay = req->rq_no_resend = 1;
3817 ptlrpc_request_set_replen(req);
3818 rc = ptlrpc_queue_wait(req);
3822 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3824 GOTO(out, rc = -EPROTO);
3826 *((obd_id *)val) = *reply;
3828 ptlrpc_req_finished(req);
3830 } else if (KEY_IS(KEY_FIEMAP)) {
3831 struct ptlrpc_request *req;
3832 struct ll_user_fiemap *reply;
3836 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3837 &RQF_OST_GET_INFO_FIEMAP);
3841 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3842 RCL_CLIENT, keylen);
3843 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3844 RCL_CLIENT, *vallen);
3845 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3846 RCL_SERVER, *vallen);
3848 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3850 ptlrpc_request_free(req);
3854 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3855 memcpy(tmp, key, keylen);
3856 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3857 memcpy(tmp, val, *vallen);
3859 ptlrpc_request_set_replen(req);
3860 rc = ptlrpc_queue_wait(req);
3864 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3866 GOTO(out1, rc = -EPROTO);
3868 memcpy(val, reply, *vallen);
3870 ptlrpc_req_finished(req);
3878 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3880 struct llog_ctxt *ctxt;
3884 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3886 rc = llog_initiator_connect(ctxt);
3887 llog_ctxt_put(ctxt);
3889 /* XXX return an error? skip setting below flags? */
3892 spin_lock(&imp->imp_lock);
3893 imp->imp_server_timeout = 1;
3894 imp->imp_pingable = 1;
3895 spin_unlock(&imp->imp_lock);
3896 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3901 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3902 struct ptlrpc_request *req,
3909 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3912 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3913 void *key, obd_count vallen, void *val,
3914 struct ptlrpc_request_set *set)
3916 struct ptlrpc_request *req;
3917 struct obd_device *obd = exp->exp_obd;
3918 struct obd_import *imp = class_exp2cliimp(exp);
3923 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3925 if (KEY_IS(KEY_NEXT_ID)) {
3927 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3929 if (vallen != sizeof(obd_id))
3934 if (vallen != sizeof(obd_id))
3937 /* avoid race between allocate new object and set next id
3938 * from ll_sync thread */
3939 spin_lock(&oscc->oscc_lock);
3940 new_val = *((obd_id*)val) + 1;
3941 if (new_val > oscc->oscc_next_id)
3942 oscc->oscc_next_id = new_val;
3943 spin_unlock(&oscc->oscc_lock);
3944 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3945 exp->exp_obd->obd_name,
3946 obd->u.cli.cl_oscc.oscc_next_id);
3951 if (KEY_IS(KEY_INIT_RECOV)) {
3952 if (vallen != sizeof(int))
3954 spin_lock(&imp->imp_lock);
3955 imp->imp_initial_recov = *(int *)val;
3956 spin_unlock(&imp->imp_lock);
3957 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3958 exp->exp_obd->obd_name,
3959 imp->imp_initial_recov);
3963 if (KEY_IS(KEY_CHECKSUM)) {
3964 if (vallen != sizeof(int))
3966 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3970 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3971 sptlrpc_conf_client_adapt(obd);
3975 if (KEY_IS(KEY_FLUSH_CTX)) {
3976 sptlrpc_import_flush_my_ctx(imp);
3980 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3983 /* We pass all other commands directly to OST. Since nobody calls osc
3984 methods directly and everybody is supposed to go through LOV, we
3985 assume lov checked invalid values for us.
3986 The only recognised values so far are evict_by_nid and mds_conn.
3987 Even if something bad goes through, we'd get a -EINVAL from OST
3990 if (KEY_IS(KEY_GRANT_SHRINK))
3991 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
3993 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
3998 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3999 RCL_CLIENT, keylen);
4000 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4001 RCL_CLIENT, vallen);
4002 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4004 ptlrpc_request_free(req);
4008 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4009 memcpy(tmp, key, keylen);
4010 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4011 memcpy(tmp, val, vallen);
4013 if (KEY_IS(KEY_MDS_CONN)) {
4014 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4016 oscc->oscc_oa.o_gr = (*(__u32 *)val);
4017 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4018 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
4019 req->rq_no_delay = req->rq_no_resend = 1;
4020 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4021 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4022 struct osc_grant_args *aa;
4025 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4026 aa = ptlrpc_req_async_args(req);
4029 ptlrpc_req_finished(req);
4032 *oa = ((struct ost_body *)val)->oa;
4034 req->rq_interpret_reply = osc_shrink_grant_interpret;
4037 ptlrpc_request_set_replen(req);
4038 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4039 LASSERT(set != NULL);
4040 ptlrpc_set_add_req(set, req);
4041 ptlrpc_check_set(NULL, set);
4043 ptlrpcd_add_req(req, PSCOPE_OTHER);
4049 static struct llog_operations osc_size_repl_logops = {
4050 lop_cancel: llog_obd_repl_cancel
4053 static struct llog_operations osc_mds_ost_orig_logops;
4055 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4056 struct obd_device *tgt, struct llog_catid *catid)
4061 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4062 &catid->lci_logid, &osc_mds_ost_orig_logops);
4064 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4068 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4069 NULL, &osc_size_repl_logops);
4071 struct llog_ctxt *ctxt =
4072 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4075 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4080 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4081 obd->obd_name, tgt->obd_name, catid, rc);
4082 CERROR("logid "LPX64":0x%x\n",
4083 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4088 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4089 struct obd_device *disk_obd, int *index)
4091 struct llog_catid catid;
4092 static char name[32] = CATLIST;
4096 LASSERT(olg == &obd->obd_olg);
4098 mutex_down(&olg->olg_cat_processing);
4099 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4101 CERROR("rc: %d\n", rc);
4105 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4106 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4107 catid.lci_logid.lgl_ogr, catid.lci_logid.lgl_ogen);
4109 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4111 CERROR("rc: %d\n", rc);
4115 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4117 CERROR("rc: %d\n", rc);
4122 mutex_up(&olg->olg_cat_processing);
4127 static int osc_llog_finish(struct obd_device *obd, int count)
4129 struct llog_ctxt *ctxt;
4130 int rc = 0, rc2 = 0;
4133 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4135 rc = llog_cleanup(ctxt);
4137 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4139 rc2 = llog_cleanup(ctxt);
4146 static int osc_reconnect(const struct lu_env *env,
4147 struct obd_export *exp, struct obd_device *obd,
4148 struct obd_uuid *cluuid,
4149 struct obd_connect_data *data,
4152 struct client_obd *cli = &obd->u.cli;
4154 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4157 client_obd_list_lock(&cli->cl_loi_list_lock);
4158 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4159 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4160 lost_grant = cli->cl_lost_grant;
4161 cli->cl_lost_grant = 0;
4162 client_obd_list_unlock(&cli->cl_loi_list_lock);
4164 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4165 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4166 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4167 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4168 " ocd_grant: %d\n", data->ocd_connect_flags,
4169 data->ocd_version, data->ocd_grant);
4175 static int osc_disconnect(struct obd_export *exp)
4177 struct obd_device *obd = class_exp2obd(exp);
4178 struct llog_ctxt *ctxt;
4181 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4183 if (obd->u.cli.cl_conn_count == 1) {
4184 /* Flush any remaining cancel messages out to the
4186 llog_sync(ctxt, exp);
4188 llog_ctxt_put(ctxt);
4190 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4194 rc = client_disconnect_export(exp);
4196 * Initially we put del_shrink_grant before disconnect_export, but it
4197 * causes the following problem if setup (connect) and cleanup
4198 * (disconnect) are tangled together.
4199 * connect p1 disconnect p2
4200 * ptlrpc_connect_import
4201 * ............... class_manual_cleanup
4204 * ptlrpc_connect_interrupt
4206 * add this client to shrink list
4208 * Bang! pinger trigger the shrink.
4209 * So the osc should be disconnected from the shrink list, after we
4210 * are sure the import has been destroyed. BUG18662
4212 if (obd->u.cli.cl_import == NULL)
4213 osc_del_shrink_grant(&obd->u.cli);
4217 static int osc_import_event(struct obd_device *obd,
4218 struct obd_import *imp,
4219 enum obd_import_event event)
4221 struct client_obd *cli;
4225 LASSERT(imp->imp_obd == obd);
4228 case IMP_EVENT_DISCON: {
4229 /* Only do this on the MDS OSC's */
4230 if (imp->imp_server_timeout) {
4231 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4233 spin_lock(&oscc->oscc_lock);
4234 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4235 spin_unlock(&oscc->oscc_lock);
4238 client_obd_list_lock(&cli->cl_loi_list_lock);
4239 cli->cl_avail_grant = 0;
4240 cli->cl_lost_grant = 0;
4241 client_obd_list_unlock(&cli->cl_loi_list_lock);
4244 case IMP_EVENT_INACTIVE: {
4245 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4248 case IMP_EVENT_INVALIDATE: {
4249 struct ldlm_namespace *ns = obd->obd_namespace;
4253 env = cl_env_get(&refcheck);
4257 client_obd_list_lock(&cli->cl_loi_list_lock);
4258 /* all pages go to failing rpcs due to the invalid
4260 osc_check_rpcs(env, cli);
4261 client_obd_list_unlock(&cli->cl_loi_list_lock);
4263 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4264 cl_env_put(env, &refcheck);
4269 case IMP_EVENT_ACTIVE: {
4270 /* Only do this on the MDS OSC's */
4271 if (imp->imp_server_timeout) {
4272 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4274 spin_lock(&oscc->oscc_lock);
4275 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4276 spin_unlock(&oscc->oscc_lock);
4278 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4281 case IMP_EVENT_OCD: {
4282 struct obd_connect_data *ocd = &imp->imp_connect_data;
4284 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4285 osc_init_grant(&obd->u.cli, ocd);
4288 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4289 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4291 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4295 CERROR("Unknown import event %d\n", event);
4301 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4307 rc = ptlrpcd_addref();
4311 rc = client_obd_setup(obd, lcfg);
4315 struct lprocfs_static_vars lvars = { 0 };
4316 struct client_obd *cli = &obd->u.cli;
4318 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4319 lprocfs_osc_init_vars(&lvars);
4320 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4321 lproc_osc_attach_seqstat(obd);
4322 sptlrpc_lprocfs_cliobd_attach(obd);
4323 ptlrpc_lprocfs_register_obd(obd);
4327 /* We need to allocate a few requests more, because
4328 brw_interpret tries to create new requests before freeing
4329 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4330 reserved, but I afraid that might be too much wasted RAM
4331 in fact, so 2 is just my guess and still should work. */
4332 cli->cl_import->imp_rq_pool =
4333 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4335 ptlrpc_add_rqs_to_pool);
4337 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4338 sema_init(&cli->cl_grant_sem, 1);
4344 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4350 case OBD_CLEANUP_EARLY: {
4351 struct obd_import *imp;
4352 imp = obd->u.cli.cl_import;
4353 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4354 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4355 ptlrpc_deactivate_import(imp);
4356 spin_lock(&imp->imp_lock);
4357 imp->imp_pingable = 0;
4358 spin_unlock(&imp->imp_lock);
4361 case OBD_CLEANUP_EXPORTS: {
4362 /* If we set up but never connected, the
4363 client import will not have been cleaned. */
4364 if (obd->u.cli.cl_import) {
4365 struct obd_import *imp;
4366 down_write(&obd->u.cli.cl_sem);
4367 imp = obd->u.cli.cl_import;
4368 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4370 ptlrpc_invalidate_import(imp);
4371 if (imp->imp_rq_pool) {
4372 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4373 imp->imp_rq_pool = NULL;
4375 class_destroy_import(imp);
4376 up_write(&obd->u.cli.cl_sem);
4377 obd->u.cli.cl_import = NULL;
4379 rc = obd_llog_finish(obd, 0);
4381 CERROR("failed to cleanup llogging subsystems\n");
4388 int osc_cleanup(struct obd_device *obd)
4393 ptlrpc_lprocfs_unregister_obd(obd);
4394 lprocfs_obd_cleanup(obd);
4396 /* free memory of osc quota cache */
4397 lquota_cleanup(quota_interface, obd);
4399 rc = client_obd_cleanup(obd);
4405 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4407 struct lprocfs_static_vars lvars = { 0 };
4410 lprocfs_osc_init_vars(&lvars);
4412 switch (lcfg->lcfg_command) {
4414 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4424 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4426 return osc_process_config_base(obd, buf);
4429 struct obd_ops osc_obd_ops = {
4430 .o_owner = THIS_MODULE,
4431 .o_setup = osc_setup,
4432 .o_precleanup = osc_precleanup,
4433 .o_cleanup = osc_cleanup,
4434 .o_add_conn = client_import_add_conn,
4435 .o_del_conn = client_import_del_conn,
4436 .o_connect = client_connect_import,
4437 .o_reconnect = osc_reconnect,
4438 .o_disconnect = osc_disconnect,
4439 .o_statfs = osc_statfs,
4440 .o_statfs_async = osc_statfs_async,
4441 .o_packmd = osc_packmd,
4442 .o_unpackmd = osc_unpackmd,
4443 .o_precreate = osc_precreate,
4444 .o_create = osc_create,
4445 .o_create_async = osc_create_async,
4446 .o_destroy = osc_destroy,
4447 .o_getattr = osc_getattr,
4448 .o_getattr_async = osc_getattr_async,
4449 .o_setattr = osc_setattr,
4450 .o_setattr_async = osc_setattr_async,
4452 .o_punch = osc_punch,
4454 .o_enqueue = osc_enqueue,
4455 .o_change_cbdata = osc_change_cbdata,
4456 .o_cancel = osc_cancel,
4457 .o_cancel_unused = osc_cancel_unused,
4458 .o_iocontrol = osc_iocontrol,
4459 .o_get_info = osc_get_info,
4460 .o_set_info_async = osc_set_info_async,
4461 .o_import_event = osc_import_event,
4462 .o_llog_init = osc_llog_init,
4463 .o_llog_finish = osc_llog_finish,
4464 .o_process_config = osc_process_config,
4467 extern struct lu_kmem_descr osc_caches[];
4468 extern spinlock_t osc_ast_guard;
4469 extern struct lock_class_key osc_ast_guard_class;
4471 int __init osc_init(void)
4473 struct lprocfs_static_vars lvars = { 0 };
4477 /* print an address of _any_ initialized kernel symbol from this
4478 * module, to allow debugging with gdb that doesn't support data
4479 * symbols from modules.*/
4480 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4482 rc = lu_kmem_init(osc_caches);
4484 lprocfs_osc_init_vars(&lvars);
4486 request_module("lquota");
4487 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4488 lquota_init(quota_interface);
4489 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4491 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4492 LUSTRE_OSC_NAME, &osc_device_type);
4494 if (quota_interface)
4495 PORTAL_SYMBOL_PUT(osc_quota_interface);
4496 lu_kmem_fini(osc_caches);
4500 spin_lock_init(&osc_ast_guard);
4501 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4503 osc_mds_ost_orig_logops = llog_lvfs_ops;
4504 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4505 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4506 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4507 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4513 static void /*__exit*/ osc_exit(void)
4515 lu_device_type_fini(&osc_device_type);
4517 lquota_exit(quota_interface);
4518 if (quota_interface)
4519 PORTAL_SYMBOL_PUT(osc_quota_interface);
4521 class_unregister_type(LUSTRE_OSC_NAME);
4522 lu_kmem_fini(osc_caches);
4525 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4526 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4527 MODULE_LICENSE("GPL");
4529 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);