1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 # define EXPORT_SYMTAB
43 #define DEBUG_SUBSYSTEM S_OSC
45 #include <libcfs/libcfs.h>
48 # include <liblustre.h>
51 #include <lustre_dlm.h>
52 #include <lustre_net.h>
53 #include <lustre/lustre_user.h>
54 #include <obd_cksum.h>
62 #include <lustre_ha.h>
63 #include <lprocfs_status.h>
64 #include <lustre_log.h>
65 #include <lustre_debug.h>
66 #include <lustre_param.h>
67 #include "osc_internal.h"
69 static quota_interface_t *quota_interface = NULL;
70 extern quota_interface_t osc_quota_interface;
72 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
73 static int brw_interpret(const struct lu_env *env,
74 struct ptlrpc_request *req, void *data, int rc);
75 int osc_cleanup(struct obd_device *obd);
77 /* Pack OSC object metadata for disk storage (LE byte order). */
78 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
79 struct lov_stripe_md *lsm)
84 lmm_size = sizeof(**lmmp);
89 OBD_FREE(*lmmp, lmm_size);
95 OBD_ALLOC(*lmmp, lmm_size);
101 LASSERT(lsm->lsm_object_id);
102 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
103 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
104 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
110 /* Unpack OSC object metadata from disk storage (LE byte order). */
111 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
112 struct lov_mds_md *lmm, int lmm_bytes)
115 struct obd_import *imp = class_exp2cliimp(exp);
119 if (lmm_bytes < sizeof (*lmm)) {
120 CERROR("lov_mds_md too small: %d, need %d\n",
121 lmm_bytes, (int)sizeof(*lmm));
124 /* XXX LOV_MAGIC etc check? */
126 if (lmm->lmm_object_id == 0) {
127 CERROR("lov_mds_md: zero lmm_object_id\n");
132 lsm_size = lov_stripe_md_size(1);
136 if (*lsmp != NULL && lmm == NULL) {
137 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
138 OBD_FREE(*lsmp, lsm_size);
144 OBD_ALLOC(*lsmp, lsm_size);
147 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
148 if ((*lsmp)->lsm_oinfo[0] == NULL) {
149 OBD_FREE(*lsmp, lsm_size);
152 loi_init((*lsmp)->lsm_oinfo[0]);
156 /* XXX zero *lsmp? */
157 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
158 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
159 LASSERT((*lsmp)->lsm_object_id);
160 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
164 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
165 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
167 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
172 static inline void osc_pack_capa(struct ptlrpc_request *req,
173 struct ost_body *body, void *capa)
175 struct obd_capa *oc = (struct obd_capa *)capa;
176 struct lustre_capa *c;
181 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
184 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
185 DEBUG_CAPA(D_SEC, c, "pack");
188 static inline void osc_pack_req_body(struct ptlrpc_request *req,
189 struct obd_info *oinfo)
191 struct ost_body *body;
193 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
196 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
197 osc_pack_capa(req, body, oinfo->oi_capa);
200 static inline void osc_set_capa_size(struct ptlrpc_request *req,
201 const struct req_msg_field *field,
205 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
207 /* it is already calculated as sizeof struct obd_capa */
211 static int osc_getattr_interpret(const struct lu_env *env,
212 struct ptlrpc_request *req,
213 struct osc_async_args *aa, int rc)
215 struct ost_body *body;
221 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
223 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
224 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
226 /* This should really be sent by the OST */
227 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
228 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
230 CDEBUG(D_INFO, "can't unpack ost_body\n");
232 aa->aa_oi->oi_oa->o_valid = 0;
235 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
239 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
240 struct ptlrpc_request_set *set)
242 struct ptlrpc_request *req;
243 struct osc_async_args *aa;
247 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
251 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
252 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
254 ptlrpc_request_free(req);
258 osc_pack_req_body(req, oinfo);
260 ptlrpc_request_set_replen(req);
261 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
263 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
264 aa = ptlrpc_req_async_args(req);
267 ptlrpc_set_add_req(set, req);
271 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
273 struct ptlrpc_request *req;
274 struct ost_body *body;
278 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
282 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
283 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
285 ptlrpc_request_free(req);
289 osc_pack_req_body(req, oinfo);
291 ptlrpc_request_set_replen(req);
293 rc = ptlrpc_queue_wait(req);
297 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
299 GOTO(out, rc = -EPROTO);
301 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
302 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
304 /* This should really be sent by the OST */
305 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
306 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
310 ptlrpc_req_finished(req);
314 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
315 struct obd_trans_info *oti)
317 struct ptlrpc_request *req;
318 struct ost_body *body;
322 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
324 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
328 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
329 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
331 ptlrpc_request_free(req);
335 osc_pack_req_body(req, oinfo);
337 ptlrpc_request_set_replen(req);
339 rc = ptlrpc_queue_wait(req);
343 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
345 GOTO(out, rc = -EPROTO);
347 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
351 ptlrpc_req_finished(req);
355 static int osc_setattr_interpret(const struct lu_env *env,
356 struct ptlrpc_request *req,
357 struct osc_setattr_args *sa, int rc)
359 struct ost_body *body;
365 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
367 GOTO(out, rc = -EPROTO);
369 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
371 rc = sa->sa_upcall(sa->sa_cookie, rc);
375 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
376 struct obd_trans_info *oti,
377 obd_enqueue_update_f upcall, void *cookie,
378 struct ptlrpc_request_set *rqset)
380 struct ptlrpc_request *req;
381 struct osc_setattr_args *sa;
385 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
389 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
390 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
392 ptlrpc_request_free(req);
396 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
397 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
399 osc_pack_req_body(req, oinfo);
401 ptlrpc_request_set_replen(req);
403 /* do mds to ost setattr asynchronously */
405 /* Do not wait for response. */
406 ptlrpcd_add_req(req, PSCOPE_OTHER);
408 req->rq_interpret_reply =
409 (ptlrpc_interpterer_t)osc_setattr_interpret;
411 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
412 sa = ptlrpc_req_async_args(req);
413 sa->sa_oa = oinfo->oi_oa;
414 sa->sa_upcall = upcall;
415 sa->sa_cookie = cookie;
417 if (rqset == PTLRPCD_SET)
418 ptlrpcd_add_req(req, PSCOPE_OTHER);
420 ptlrpc_set_add_req(rqset, req);
426 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
427 struct obd_trans_info *oti,
428 struct ptlrpc_request_set *rqset)
430 return osc_setattr_async_base(exp, oinfo, oti,
431 oinfo->oi_cb_up, oinfo, rqset);
434 int osc_real_create(struct obd_export *exp, struct obdo *oa,
435 struct lov_stripe_md **ea, struct obd_trans_info *oti)
437 struct ptlrpc_request *req;
438 struct ost_body *body;
439 struct lov_stripe_md *lsm;
448 rc = obd_alloc_memmd(exp, &lsm);
453 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
455 GOTO(out, rc = -ENOMEM);
457 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
459 ptlrpc_request_free(req);
463 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
465 lustre_set_wire_obdo(&body->oa, oa);
467 ptlrpc_request_set_replen(req);
469 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
470 oa->o_flags == OBD_FL_DELORPHAN) {
472 "delorphan from OST integration");
473 /* Don't resend the delorphan req */
474 req->rq_no_resend = req->rq_no_delay = 1;
477 rc = ptlrpc_queue_wait(req);
481 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
483 GOTO(out_req, rc = -EPROTO);
485 lustre_get_wire_obdo(oa, &body->oa);
487 /* This should really be sent by the OST */
488 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
489 oa->o_valid |= OBD_MD_FLBLKSZ;
491 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
492 * have valid lsm_oinfo data structs, so don't go touching that.
493 * This needs to be fixed in a big way.
495 lsm->lsm_object_id = oa->o_id;
496 lsm->lsm_object_seq = oa->o_seq;
500 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
502 if (oa->o_valid & OBD_MD_FLCOOKIE) {
503 if (!oti->oti_logcookies)
504 oti_alloc_cookies(oti, 1);
505 *oti->oti_logcookies = oa->o_lcookie;
509 CDEBUG(D_HA, "transno: "LPD64"\n",
510 lustre_msg_get_transno(req->rq_repmsg));
512 ptlrpc_req_finished(req);
515 obd_free_memmd(exp, &lsm);
519 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
520 obd_enqueue_update_f upcall, void *cookie,
521 struct ptlrpc_request_set *rqset)
523 struct ptlrpc_request *req;
524 struct osc_setattr_args *sa;
525 struct ost_body *body;
529 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
533 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
534 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
536 ptlrpc_request_free(req);
539 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
540 ptlrpc_at_set_req_timeout(req);
542 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
544 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
545 osc_pack_capa(req, body, oinfo->oi_capa);
547 ptlrpc_request_set_replen(req);
550 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
551 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
552 sa = ptlrpc_req_async_args(req);
553 sa->sa_oa = oinfo->oi_oa;
554 sa->sa_upcall = upcall;
555 sa->sa_cookie = cookie;
556 if (rqset == PTLRPCD_SET)
557 ptlrpcd_add_req(req, PSCOPE_OTHER);
559 ptlrpc_set_add_req(rqset, req);
564 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
565 struct obd_trans_info *oti,
566 struct ptlrpc_request_set *rqset)
568 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
569 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
570 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
571 return osc_punch_base(exp, oinfo,
572 oinfo->oi_cb_up, oinfo, rqset);
575 static int osc_sync_interpret(const struct lu_env *env,
576 struct ptlrpc_request *req,
579 struct osc_async_args *aa = arg;
580 struct ost_body *body;
586 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
588 CERROR ("can't unpack ost_body\n");
589 GOTO(out, rc = -EPROTO);
592 *aa->aa_oi->oi_oa = body->oa;
594 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
598 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
599 obd_size start, obd_size end,
600 struct ptlrpc_request_set *set)
602 struct ptlrpc_request *req;
603 struct ost_body *body;
604 struct osc_async_args *aa;
609 CDEBUG(D_INFO, "oa NULL\n");
613 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
617 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
618 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
620 ptlrpc_request_free(req);
624 /* overload the size and blocks fields in the oa with start/end */
625 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
627 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
628 body->oa.o_size = start;
629 body->oa.o_blocks = end;
630 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
631 osc_pack_capa(req, body, oinfo->oi_capa);
633 ptlrpc_request_set_replen(req);
634 req->rq_interpret_reply = osc_sync_interpret;
636 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
637 aa = ptlrpc_req_async_args(req);
640 ptlrpc_set_add_req(set, req);
644 /* Find and cancel locally locks matched by @mode in the resource found by
645 * @objid. Found locks are added into @cancel list. Returns the amount of
646 * locks added to @cancels list. */
647 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
649 ldlm_mode_t mode, int lock_flags)
651 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
652 struct ldlm_res_id res_id;
653 struct ldlm_resource *res;
657 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
658 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
662 LDLM_RESOURCE_ADDREF(res);
663 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
664 lock_flags, 0, NULL);
665 LDLM_RESOURCE_DELREF(res);
666 ldlm_resource_putref(res);
670 static int osc_destroy_interpret(const struct lu_env *env,
671 struct ptlrpc_request *req, void *data,
674 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
676 cfs_atomic_dec(&cli->cl_destroy_in_flight);
677 cfs_waitq_signal(&cli->cl_destroy_waitq);
681 static int osc_can_send_destroy(struct client_obd *cli)
683 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
684 cli->cl_max_rpcs_in_flight) {
685 /* The destroy request can be sent */
688 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
689 cli->cl_max_rpcs_in_flight) {
691 * The counter has been modified between the two atomic
694 cfs_waitq_signal(&cli->cl_destroy_waitq);
699 /* Destroy requests can be async always on the client, and we don't even really
700 * care about the return code since the client cannot do anything at all about
702 * When the MDS is unlinking a filename, it saves the file objects into a
703 * recovery llog, and these object records are cancelled when the OST reports
704 * they were destroyed and sync'd to disk (i.e. transaction committed).
705 * If the client dies, or the OST is down when the object should be destroyed,
706 * the records are not cancelled, and when the OST reconnects to the MDS next,
707 * it will retrieve the llog unlink logs and then sends the log cancellation
708 * cookies to the MDS after committing destroy transactions. */
709 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
710 struct lov_stripe_md *ea, struct obd_trans_info *oti,
711 struct obd_export *md_export, void *capa)
713 struct client_obd *cli = &exp->exp_obd->u.cli;
714 struct ptlrpc_request *req;
715 struct ost_body *body;
716 CFS_LIST_HEAD(cancels);
721 CDEBUG(D_INFO, "oa NULL\n");
725 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
726 LDLM_FL_DISCARD_DATA);
728 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
730 ldlm_lock_list_put(&cancels, l_bl_ast, count);
734 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
735 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
738 ptlrpc_request_free(req);
742 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
743 ptlrpc_at_set_req_timeout(req);
745 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
746 oa->o_lcookie = *oti->oti_logcookies;
747 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
749 lustre_set_wire_obdo(&body->oa, oa);
751 osc_pack_capa(req, body, (struct obd_capa *)capa);
752 ptlrpc_request_set_replen(req);
754 /* don't throttle destroy RPCs for the MDT */
755 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
756 req->rq_interpret_reply = osc_destroy_interpret;
757 if (!osc_can_send_destroy(cli)) {
758 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
762 * Wait until the number of on-going destroy RPCs drops
763 * under max_rpc_in_flight
765 l_wait_event_exclusive(cli->cl_destroy_waitq,
766 osc_can_send_destroy(cli), &lwi);
770 /* Do not wait for response */
771 ptlrpcd_add_req(req, PSCOPE_OTHER);
775 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
778 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
780 LASSERT(!(oa->o_valid & bits));
783 client_obd_list_lock(&cli->cl_loi_list_lock);
784 oa->o_dirty = cli->cl_dirty;
785 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
786 CERROR("dirty %lu - %lu > dirty_max %lu\n",
787 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
789 } else if (cfs_atomic_read(&obd_dirty_pages) -
790 cfs_atomic_read(&obd_dirty_transit_pages) >
791 obd_max_dirty_pages + 1){
792 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
793 * not covered by a lock thus they may safely race and trip
794 * this CERROR() unless we add in a small fudge factor (+1). */
795 CERROR("dirty %d - %d > system dirty_max %d\n",
796 cfs_atomic_read(&obd_dirty_pages),
797 cfs_atomic_read(&obd_dirty_transit_pages),
798 obd_max_dirty_pages);
800 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
801 CERROR("dirty %lu - dirty_max %lu too big???\n",
802 cli->cl_dirty, cli->cl_dirty_max);
805 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
806 (cli->cl_max_rpcs_in_flight + 1);
807 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
809 oa->o_grant = cli->cl_avail_grant;
810 oa->o_dropped = cli->cl_lost_grant;
811 cli->cl_lost_grant = 0;
812 client_obd_list_unlock(&cli->cl_loi_list_lock);
813 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
814 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
818 static void osc_update_next_shrink(struct client_obd *cli)
820 cli->cl_next_shrink_grant =
821 cfs_time_shift(cli->cl_grant_shrink_interval);
822 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
823 cli->cl_next_shrink_grant);
826 /* caller must hold loi_list_lock */
827 static void osc_consume_write_grant(struct client_obd *cli,
828 struct brw_page *pga)
830 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
831 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
832 cfs_atomic_inc(&obd_dirty_pages);
833 cli->cl_dirty += CFS_PAGE_SIZE;
834 cli->cl_avail_grant -= CFS_PAGE_SIZE;
835 pga->flag |= OBD_BRW_FROM_GRANT;
836 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
837 CFS_PAGE_SIZE, pga, pga->pg);
838 LASSERT(cli->cl_avail_grant >= 0);
839 osc_update_next_shrink(cli);
842 /* the companion to osc_consume_write_grant, called when a brw has completed.
843 * must be called with the loi lock held. */
844 static void osc_release_write_grant(struct client_obd *cli,
845 struct brw_page *pga, int sent)
847 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
850 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
851 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
856 pga->flag &= ~OBD_BRW_FROM_GRANT;
857 cfs_atomic_dec(&obd_dirty_pages);
858 cli->cl_dirty -= CFS_PAGE_SIZE;
859 if (pga->flag & OBD_BRW_NOCACHE) {
860 pga->flag &= ~OBD_BRW_NOCACHE;
861 cfs_atomic_dec(&obd_dirty_transit_pages);
862 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
865 cli->cl_lost_grant += CFS_PAGE_SIZE;
866 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
867 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
868 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
869 /* For short writes we shouldn't count parts of pages that
870 * span a whole block on the OST side, or our accounting goes
871 * wrong. Should match the code in filter_grant_check. */
872 int offset = pga->off & ~CFS_PAGE_MASK;
873 int count = pga->count + (offset & (blocksize - 1));
874 int end = (offset + pga->count) & (blocksize - 1);
876 count += blocksize - end;
878 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
879 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
880 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
881 cli->cl_avail_grant, cli->cl_dirty);
887 static unsigned long rpcs_in_flight(struct client_obd *cli)
889 return cli->cl_r_in_flight + cli->cl_w_in_flight;
892 /* caller must hold loi_list_lock */
893 void osc_wake_cache_waiters(struct client_obd *cli)
896 struct osc_cache_waiter *ocw;
899 cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
900 /* if we can't dirty more, we must wait until some is written */
901 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
902 (cfs_atomic_read(&obd_dirty_pages) + 1 >
903 obd_max_dirty_pages)) {
904 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
905 "osc max %ld, sys max %d\n", cli->cl_dirty,
906 cli->cl_dirty_max, obd_max_dirty_pages);
910 /* if still dirty cache but no grant wait for pending RPCs that
911 * may yet return us some grant before doing sync writes */
912 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
913 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
914 cli->cl_w_in_flight);
918 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
919 cfs_list_del_init(&ocw->ocw_entry);
920 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
921 /* no more RPCs in flight to return grant, do sync IO */
922 ocw->ocw_rc = -EDQUOT;
923 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
925 osc_consume_write_grant(cli,
926 &ocw->ocw_oap->oap_brw_page);
929 cfs_waitq_signal(&ocw->ocw_waitq);
935 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
937 client_obd_list_lock(&cli->cl_loi_list_lock);
938 cli->cl_avail_grant += grant;
939 client_obd_list_unlock(&cli->cl_loi_list_lock);
942 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
944 if (body->oa.o_valid & OBD_MD_FLGRANT) {
945 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
946 __osc_update_grant(cli, body->oa.o_grant);
950 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
951 void *key, obd_count vallen, void *val,
952 struct ptlrpc_request_set *set);
954 static int osc_shrink_grant_interpret(const struct lu_env *env,
955 struct ptlrpc_request *req,
958 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
959 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
960 struct ost_body *body;
963 __osc_update_grant(cli, oa->o_grant);
967 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
969 osc_update_grant(cli, body);
975 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
977 client_obd_list_lock(&cli->cl_loi_list_lock);
978 oa->o_grant = cli->cl_avail_grant / 4;
979 cli->cl_avail_grant -= oa->o_grant;
980 client_obd_list_unlock(&cli->cl_loi_list_lock);
981 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
982 oa->o_valid |= OBD_MD_FLFLAGS;
985 oa->o_flags |= OBD_FL_SHRINK_GRANT;
986 osc_update_next_shrink(cli);
989 /* Shrink the current grant, either from some large amount to enough for a
990 * full set of in-flight RPCs, or if we have already shrunk to that limit
991 * then to enough for a single RPC. This avoids keeping more grant than
992 * needed, and avoids shrinking the grant piecemeal. */
993 static int osc_shrink_grant(struct client_obd *cli)
995 long target = (cli->cl_max_rpcs_in_flight + 1) *
996 cli->cl_max_pages_per_rpc;
998 client_obd_list_lock(&cli->cl_loi_list_lock);
999 if (cli->cl_avail_grant <= target)
1000 target = cli->cl_max_pages_per_rpc;
1001 client_obd_list_unlock(&cli->cl_loi_list_lock);
1003 return osc_shrink_grant_to_target(cli, target);
1006 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1009 struct ost_body *body;
1012 client_obd_list_lock(&cli->cl_loi_list_lock);
1013 /* Don't shrink if we are already above or below the desired limit
1014 * We don't want to shrink below a single RPC, as that will negatively
1015 * impact block allocation and long-term performance. */
1016 if (target < cli->cl_max_pages_per_rpc)
1017 target = cli->cl_max_pages_per_rpc;
1019 if (target >= cli->cl_avail_grant) {
1020 client_obd_list_unlock(&cli->cl_loi_list_lock);
1023 client_obd_list_unlock(&cli->cl_loi_list_lock);
1025 OBD_ALLOC_PTR(body);
1029 osc_announce_cached(cli, &body->oa, 0);
1031 client_obd_list_lock(&cli->cl_loi_list_lock);
1032 body->oa.o_grant = cli->cl_avail_grant - target;
1033 cli->cl_avail_grant = target;
1034 client_obd_list_unlock(&cli->cl_loi_list_lock);
1035 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1036 body->oa.o_valid |= OBD_MD_FLFLAGS;
1037 body->oa.o_flags = 0;
1039 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1040 osc_update_next_shrink(cli);
1042 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1043 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1044 sizeof(*body), body, NULL);
1046 __osc_update_grant(cli, body->oa.o_grant);
1051 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1052 static int osc_should_shrink_grant(struct client_obd *client)
1054 cfs_time_t time = cfs_time_current();
1055 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1057 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1058 OBD_CONNECT_GRANT_SHRINK) == 0)
1061 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1062 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1063 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1066 osc_update_next_shrink(client);
1071 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1073 struct client_obd *client;
1075 cfs_list_for_each_entry(client, &item->ti_obd_list,
1076 cl_grant_shrink_list) {
1077 if (osc_should_shrink_grant(client))
1078 osc_shrink_grant(client);
1083 static int osc_add_shrink_grant(struct client_obd *client)
1087 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1089 osc_grant_shrink_grant_cb, NULL,
1090 &client->cl_grant_shrink_list);
1092 CERROR("add grant client %s error %d\n",
1093 client->cl_import->imp_obd->obd_name, rc);
1096 CDEBUG(D_CACHE, "add grant client %s \n",
1097 client->cl_import->imp_obd->obd_name);
1098 osc_update_next_shrink(client);
1102 static int osc_del_shrink_grant(struct client_obd *client)
1104 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1108 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1111 * ocd_grant is the total grant amount we're expect to hold: if we've
1112 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1113 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1115 * race is tolerable here: if we're evicted, but imp_state already
1116 * left EVICTED state, then cl_dirty must be 0 already.
1118 client_obd_list_lock(&cli->cl_loi_list_lock);
1119 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1120 cli->cl_avail_grant = ocd->ocd_grant;
1122 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1124 if (cli->cl_avail_grant < 0) {
1125 CWARN("%s: available grant < 0, the OSS is probably not running"
1126 " with patch from bug20278 (%ld) \n",
1127 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1128 /* workaround for 1.6 servers which do not have
1129 * the patch from bug20278 */
1130 cli->cl_avail_grant = ocd->ocd_grant;
1133 client_obd_list_unlock(&cli->cl_loi_list_lock);
1135 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1136 cli->cl_import->imp_obd->obd_name,
1137 cli->cl_avail_grant, cli->cl_lost_grant);
1139 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1140 cfs_list_empty(&cli->cl_grant_shrink_list))
1141 osc_add_shrink_grant(cli);
1144 /* We assume that the reason this OSC got a short read is because it read
1145 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1146 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1147 * this stripe never got written at or beyond this stripe offset yet. */
1148 static void handle_short_read(int nob_read, obd_count page_count,
1149 struct brw_page **pga)
1154 /* skip bytes read OK */
1155 while (nob_read > 0) {
1156 LASSERT (page_count > 0);
1158 if (pga[i]->count > nob_read) {
1159 /* EOF inside this page */
1160 ptr = cfs_kmap(pga[i]->pg) +
1161 (pga[i]->off & ~CFS_PAGE_MASK);
1162 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1163 cfs_kunmap(pga[i]->pg);
1169 nob_read -= pga[i]->count;
1174 /* zero remaining pages */
1175 while (page_count-- > 0) {
1176 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1177 memset(ptr, 0, pga[i]->count);
1178 cfs_kunmap(pga[i]->pg);
1183 static int check_write_rcs(struct ptlrpc_request *req,
1184 int requested_nob, int niocount,
1185 obd_count page_count, struct brw_page **pga)
1190 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1191 sizeof(*remote_rcs) *
1193 if (remote_rcs == NULL) {
1194 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1198 /* return error if any niobuf was in error */
1199 for (i = 0; i < niocount; i++) {
1200 if ((int)remote_rcs[i] < 0)
1201 return(remote_rcs[i]);
1203 if (remote_rcs[i] != 0) {
1204 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1205 i, remote_rcs[i], req);
1210 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1211 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1212 req->rq_bulk->bd_nob_transferred, requested_nob);
1219 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1221 if (p1->flag != p2->flag) {
1222 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1223 OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
1225 /* warn if we try to combine flags that we don't know to be
1226 * safe to combine */
1227 if ((p1->flag & mask) != (p2->flag & mask))
1228 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1229 "same brw?\n", p1->flag, p2->flag);
1233 return (p1->off + p1->count == p2->off);
1236 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1237 struct brw_page **pga, int opc,
1238 cksum_type_t cksum_type)
1243 LASSERT (pg_count > 0);
1244 cksum = init_checksum(cksum_type);
1245 while (nob > 0 && pg_count > 0) {
1246 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1247 int off = pga[i]->off & ~CFS_PAGE_MASK;
1248 int count = pga[i]->count > nob ? nob : pga[i]->count;
1250 /* corrupt the data before we compute the checksum, to
1251 * simulate an OST->client data error */
1252 if (i == 0 && opc == OST_READ &&
1253 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1254 memcpy(ptr + off, "bad1", min(4, nob));
1255 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1256 cfs_kunmap(pga[i]->pg);
1257 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1260 nob -= pga[i]->count;
1264 /* For sending we only compute the wrong checksum instead
1265 * of corrupting the data so it is still correct on a redo */
1266 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1272 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1273 struct lov_stripe_md *lsm, obd_count page_count,
1274 struct brw_page **pga,
1275 struct ptlrpc_request **reqp,
1276 struct obd_capa *ocapa, int reserve,
1279 struct ptlrpc_request *req;
1280 struct ptlrpc_bulk_desc *desc;
1281 struct ost_body *body;
1282 struct obd_ioobj *ioobj;
1283 struct niobuf_remote *niobuf;
1284 int niocount, i, requested_nob, opc, rc;
1285 struct osc_brw_async_args *aa;
1286 struct req_capsule *pill;
1287 struct brw_page *pg_prev;
1290 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1291 RETURN(-ENOMEM); /* Recoverable */
1292 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1293 RETURN(-EINVAL); /* Fatal */
1295 if ((cmd & OBD_BRW_WRITE) != 0) {
1297 req = ptlrpc_request_alloc_pool(cli->cl_import,
1298 cli->cl_import->imp_rq_pool,
1299 &RQF_OST_BRW_WRITE);
1302 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1307 for (niocount = i = 1; i < page_count; i++) {
1308 if (!can_merge_pages(pga[i - 1], pga[i]))
1312 pill = &req->rq_pill;
1313 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1315 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1316 niocount * sizeof(*niobuf));
1317 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1319 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1321 ptlrpc_request_free(req);
1324 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1325 ptlrpc_at_set_req_timeout(req);
1327 if (opc == OST_WRITE)
1328 desc = ptlrpc_prep_bulk_imp(req, page_count,
1329 BULK_GET_SOURCE, OST_BULK_PORTAL);
1331 desc = ptlrpc_prep_bulk_imp(req, page_count,
1332 BULK_PUT_SINK, OST_BULK_PORTAL);
1335 GOTO(out, rc = -ENOMEM);
1336 /* NB request now owns desc and will free it when it gets freed */
1338 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1339 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1340 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1341 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1343 lustre_set_wire_obdo(&body->oa, oa);
1345 obdo_to_ioobj(oa, ioobj);
1346 ioobj->ioo_bufcnt = niocount;
1347 osc_pack_capa(req, body, ocapa);
1348 LASSERT (page_count > 0);
1350 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1351 struct brw_page *pg = pga[i];
1353 LASSERT(pg->count > 0);
1354 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1355 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1356 pg->off, pg->count);
1358 LASSERTF(i == 0 || pg->off > pg_prev->off,
1359 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1360 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1362 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1363 pg_prev->pg, page_private(pg_prev->pg),
1364 pg_prev->pg->index, pg_prev->off);
1366 LASSERTF(i == 0 || pg->off > pg_prev->off,
1367 "i %d p_c %u\n", i, page_count);
1369 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1370 (pg->flag & OBD_BRW_SRVLOCK));
1372 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1374 requested_nob += pg->count;
1376 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1378 niobuf->len += pg->count;
1380 niobuf->offset = pg->off;
1381 niobuf->len = pg->count;
1382 niobuf->flags = pg->flag;
1387 LASSERTF((void *)(niobuf - niocount) ==
1388 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1389 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1390 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1392 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1394 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1395 body->oa.o_valid |= OBD_MD_FLFLAGS;
1396 body->oa.o_flags = 0;
1398 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1401 if (osc_should_shrink_grant(cli))
1402 osc_shrink_grant_local(cli, &body->oa);
1404 /* size[REQ_REC_OFF] still sizeof (*body) */
1405 if (opc == OST_WRITE) {
1406 if (unlikely(cli->cl_checksum) &&
1407 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1408 /* store cl_cksum_type in a local variable since
1409 * it can be changed via lprocfs */
1410 cksum_type_t cksum_type = cli->cl_cksum_type;
1412 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1413 oa->o_flags &= OBD_FL_LOCAL_MASK;
1414 body->oa.o_flags = 0;
1416 body->oa.o_flags |= cksum_type_pack(cksum_type);
1417 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1418 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1422 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1424 /* save this in 'oa', too, for later checking */
1425 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1426 oa->o_flags |= cksum_type_pack(cksum_type);
1428 /* clear out the checksum flag, in case this is a
1429 * resend but cl_checksum is no longer set. b=11238 */
1430 oa->o_valid &= ~OBD_MD_FLCKSUM;
1432 oa->o_cksum = body->oa.o_cksum;
1433 /* 1 RC per niobuf */
1434 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1435 sizeof(__u32) * niocount);
1437 if (unlikely(cli->cl_checksum) &&
1438 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1439 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1440 body->oa.o_flags = 0;
1441 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1442 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1445 ptlrpc_request_set_replen(req);
1447 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1448 aa = ptlrpc_req_async_args(req);
1450 aa->aa_requested_nob = requested_nob;
1451 aa->aa_nio_count = niocount;
1452 aa->aa_page_count = page_count;
1456 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1457 if (ocapa && reserve)
1458 aa->aa_ocapa = capa_get(ocapa);
1464 ptlrpc_req_finished(req);
1468 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1469 __u32 client_cksum, __u32 server_cksum, int nob,
1470 obd_count page_count, struct brw_page **pga,
1471 cksum_type_t client_cksum_type)
1475 cksum_type_t cksum_type;
1477 if (server_cksum == client_cksum) {
1478 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1482 /* If this is mmaped file - it can be changed at any time */
1483 if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1486 if (oa->o_valid & OBD_MD_FLFLAGS)
1487 cksum_type = cksum_type_unpack(oa->o_flags);
1489 cksum_type = OBD_CKSUM_CRC32;
1491 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1494 if (cksum_type != client_cksum_type)
1495 msg = "the server did not use the checksum type specified in "
1496 "the original request - likely a protocol problem";
1497 else if (new_cksum == server_cksum)
1498 msg = "changed on the client after we checksummed it - "
1499 "likely false positive due to mmap IO (bug 11742)";
1500 else if (new_cksum == client_cksum)
1501 msg = "changed in transit before arrival at OST";
1503 msg = "changed in transit AND doesn't match the original - "
1504 "likely false positive due to mmap IO (bug 11742)";
1506 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1507 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1508 msg, libcfs_nid2str(peer->nid),
1509 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1510 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1511 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1513 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1515 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1516 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1517 "client csum now %x\n", client_cksum, client_cksum_type,
1518 server_cksum, cksum_type, new_cksum);
1522 /* Note rc enters this function as number of bytes transferred */
1523 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1525 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1526 const lnet_process_id_t *peer =
1527 &req->rq_import->imp_connection->c_peer;
1528 struct client_obd *cli = aa->aa_cli;
1529 struct ost_body *body;
1530 __u32 client_cksum = 0;
1533 if (rc < 0 && rc != -EDQUOT) {
1534 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1538 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1539 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1541 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1545 #ifdef HAVE_QUOTA_SUPPORT
1546 /* set/clear over quota flag for a uid/gid */
1547 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1548 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1549 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1551 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1552 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1554 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1559 osc_update_grant(cli, body);
1564 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1565 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1567 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1569 CERROR("Unexpected +ve rc %d\n", rc);
1572 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1574 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1577 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1578 check_write_checksum(&body->oa, peer, client_cksum,
1579 body->oa.o_cksum, aa->aa_requested_nob,
1580 aa->aa_page_count, aa->aa_ppga,
1581 cksum_type_unpack(aa->aa_oa->o_flags)))
1584 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1585 aa->aa_page_count, aa->aa_ppga);
1589 /* The rest of this function executes only for OST_READs */
1591 /* if unwrap_bulk failed, return -EAGAIN to retry */
1592 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1594 GOTO(out, rc = -EAGAIN);
1596 if (rc > aa->aa_requested_nob) {
1597 CERROR("Unexpected rc %d (%d requested)\n", rc,
1598 aa->aa_requested_nob);
1602 if (rc != req->rq_bulk->bd_nob_transferred) {
1603 CERROR ("Unexpected rc %d (%d transferred)\n",
1604 rc, req->rq_bulk->bd_nob_transferred);
1608 if (rc < aa->aa_requested_nob)
1609 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1611 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1612 static int cksum_counter;
1613 __u32 server_cksum = body->oa.o_cksum;
1616 cksum_type_t cksum_type;
1618 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1619 cksum_type = cksum_type_unpack(body->oa.o_flags);
1621 cksum_type = OBD_CKSUM_CRC32;
1622 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1623 aa->aa_ppga, OST_READ,
1626 if (peer->nid == req->rq_bulk->bd_sender) {
1630 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1633 if (server_cksum == ~0 && rc > 0) {
1634 CERROR("Protocol error: server %s set the 'checksum' "
1635 "bit, but didn't send a checksum. Not fatal, "
1636 "but please notify on http://bugs.whamcloud.com/\n",
1637 libcfs_nid2str(peer->nid));
1638 } else if (server_cksum != client_cksum) {
1639 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1640 "%s%s%s inode "DFID" object "
1641 LPU64"/"LPU64" extent "
1642 "["LPU64"-"LPU64"]\n",
1643 req->rq_import->imp_obd->obd_name,
1644 libcfs_nid2str(peer->nid),
1646 body->oa.o_valid & OBD_MD_FLFID ?
1647 body->oa.o_parent_seq : (__u64)0,
1648 body->oa.o_valid & OBD_MD_FLFID ?
1649 body->oa.o_parent_oid : 0,
1650 body->oa.o_valid & OBD_MD_FLFID ?
1651 body->oa.o_parent_ver : 0,
1653 body->oa.o_valid & OBD_MD_FLGROUP ?
1654 body->oa.o_seq : (__u64)0,
1655 aa->aa_ppga[0]->off,
1656 aa->aa_ppga[aa->aa_page_count-1]->off +
1657 aa->aa_ppga[aa->aa_page_count-1]->count -
1659 CERROR("client %x, server %x, cksum_type %x\n",
1660 client_cksum, server_cksum, cksum_type);
1662 aa->aa_oa->o_cksum = client_cksum;
1666 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1669 } else if (unlikely(client_cksum)) {
1670 static int cksum_missed;
1673 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1674 CERROR("Checksum %u requested from %s but not sent\n",
1675 cksum_missed, libcfs_nid2str(peer->nid));
1681 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1686 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1687 struct lov_stripe_md *lsm,
1688 obd_count page_count, struct brw_page **pga,
1689 struct obd_capa *ocapa)
1691 struct ptlrpc_request *req;
1695 struct l_wait_info lwi;
1699 cfs_waitq_init(&waitq);
1702 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1703 page_count, pga, &req, ocapa, 0, resends);
1707 rc = ptlrpc_queue_wait(req);
1709 if (rc == -ETIMEDOUT && req->rq_resend) {
1710 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1711 ptlrpc_req_finished(req);
1715 rc = osc_brw_fini_request(req, rc);
1717 ptlrpc_req_finished(req);
1718 if (osc_recoverable_error(rc)) {
1720 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1721 CERROR("too many resend retries, returning error\n");
1725 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1726 l_wait_event(waitq, 0, &lwi);
1734 int osc_brw_redo_request(struct ptlrpc_request *request,
1735 struct osc_brw_async_args *aa)
1737 struct ptlrpc_request *new_req;
1738 struct ptlrpc_request_set *set = request->rq_set;
1739 struct osc_brw_async_args *new_aa;
1740 struct osc_async_page *oap;
1744 if (!client_should_resend(aa->aa_resends, aa->aa_cli)) {
1745 CERROR("too many resent retries, returning error\n");
1749 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1751 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1752 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1753 aa->aa_cli, aa->aa_oa,
1754 NULL /* lsm unused by osc currently */,
1755 aa->aa_page_count, aa->aa_ppga,
1756 &new_req, aa->aa_ocapa, 0, 1);
1760 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1762 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1763 if (oap->oap_request != NULL) {
1764 LASSERTF(request == oap->oap_request,
1765 "request %p != oap_request %p\n",
1766 request, oap->oap_request);
1767 if (oap->oap_interrupted) {
1768 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1769 ptlrpc_req_finished(new_req);
1774 /* New request takes over pga and oaps from old request.
1775 * Note that copying a list_head doesn't work, need to move it... */
1777 new_req->rq_interpret_reply = request->rq_interpret_reply;
1778 new_req->rq_async_args = request->rq_async_args;
1779 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1781 new_aa = ptlrpc_req_async_args(new_req);
1783 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1784 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1785 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1787 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1788 if (oap->oap_request) {
1789 ptlrpc_req_finished(oap->oap_request);
1790 oap->oap_request = ptlrpc_request_addref(new_req);
1794 new_aa->aa_ocapa = aa->aa_ocapa;
1795 aa->aa_ocapa = NULL;
1797 /* use ptlrpc_set_add_req is safe because interpret functions work
1798 * in check_set context. only one way exist with access to request
1799 * from different thread got -EINTR - this way protected with
1800 * cl_loi_list_lock */
1801 ptlrpc_set_add_req(set, new_req);
1803 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1805 DEBUG_REQ(D_INFO, new_req, "new request");
1810 * ugh, we want disk allocation on the target to happen in offset order. we'll
1811 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1812 * fine for our small page arrays and doesn't require allocation. its an
1813 * insertion sort that swaps elements that are strides apart, shrinking the
1814 * stride down until its '1' and the array is sorted.
1816 static void sort_brw_pages(struct brw_page **array, int num)
1819 struct brw_page *tmp;
1823 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1828 for (i = stride ; i < num ; i++) {
1831 while (j >= stride && array[j - stride]->off > tmp->off) {
1832 array[j] = array[j - stride];
1837 } while (stride > 1);
1840 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1846 LASSERT (pages > 0);
1847 offset = pg[i]->off & ~CFS_PAGE_MASK;
1851 if (pages == 0) /* that's all */
1854 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1855 return count; /* doesn't end on page boundary */
1858 offset = pg[i]->off & ~CFS_PAGE_MASK;
1859 if (offset != 0) /* doesn't start on page boundary */
1866 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1868 struct brw_page **ppga;
1871 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1875 for (i = 0; i < count; i++)
1880 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1882 LASSERT(ppga != NULL);
1883 OBD_FREE(ppga, sizeof(*ppga) * count);
1886 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1887 obd_count page_count, struct brw_page *pga,
1888 struct obd_trans_info *oti)
1890 struct obdo *saved_oa = NULL;
1891 struct brw_page **ppga, **orig;
1892 struct obd_import *imp = class_exp2cliimp(exp);
1893 struct client_obd *cli;
1894 int rc, page_count_orig;
1897 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1898 cli = &imp->imp_obd->u.cli;
1900 if (cmd & OBD_BRW_CHECK) {
1901 /* The caller just wants to know if there's a chance that this
1902 * I/O can succeed */
1904 if (imp->imp_invalid)
1909 /* test_brw with a failed create can trip this, maybe others. */
1910 LASSERT(cli->cl_max_pages_per_rpc);
1914 orig = ppga = osc_build_ppga(pga, page_count);
1917 page_count_orig = page_count;
1919 sort_brw_pages(ppga, page_count);
1920 while (page_count) {
1921 obd_count pages_per_brw;
1923 if (page_count > cli->cl_max_pages_per_rpc)
1924 pages_per_brw = cli->cl_max_pages_per_rpc;
1926 pages_per_brw = page_count;
1928 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1930 if (saved_oa != NULL) {
1931 /* restore previously saved oa */
1932 *oinfo->oi_oa = *saved_oa;
1933 } else if (page_count > pages_per_brw) {
1934 /* save a copy of oa (brw will clobber it) */
1935 OBDO_ALLOC(saved_oa);
1936 if (saved_oa == NULL)
1937 GOTO(out, rc = -ENOMEM);
1938 *saved_oa = *oinfo->oi_oa;
1941 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1942 pages_per_brw, ppga, oinfo->oi_capa);
1947 page_count -= pages_per_brw;
1948 ppga += pages_per_brw;
1952 osc_release_ppga(orig, page_count_orig);
1954 if (saved_oa != NULL)
1955 OBDO_FREE(saved_oa);
1960 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1961 * the dirty accounting. Writeback completes or truncate happens before
1962 * writing starts. Must be called with the loi lock held. */
1963 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1966 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1970 /* This maintains the lists of pending pages to read/write for a given object
1971 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1972 * to quickly find objects that are ready to send an RPC. */
1973 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1979 if (lop->lop_num_pending == 0)
1982 /* if we have an invalid import we want to drain the queued pages
1983 * by forcing them through rpcs that immediately fail and complete
1984 * the pages. recovery relies on this to empty the queued pages
1985 * before canceling the locks and evicting down the llite pages */
1986 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1989 /* stream rpcs in queue order as long as as there is an urgent page
1990 * queued. this is our cheap solution for good batching in the case
1991 * where writepage marks some random page in the middle of the file
1992 * as urgent because of, say, memory pressure */
1993 if (!cfs_list_empty(&lop->lop_urgent)) {
1994 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1997 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1998 optimal = cli->cl_max_pages_per_rpc;
1999 if (cmd & OBD_BRW_WRITE) {
2000 /* trigger a write rpc stream as long as there are dirtiers
2001 * waiting for space. as they're waiting, they're not going to
2002 * create more pages to coalesce with what's waiting.. */
2003 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2004 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2007 /* +16 to avoid triggering rpcs that would want to include pages
2008 * that are being queued but which can't be made ready until
2009 * the queuer finishes with the page. this is a wart for
2010 * llite::commit_write() */
2013 if (lop->lop_num_pending >= optimal)
2019 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2021 struct osc_async_page *oap;
2024 if (cfs_list_empty(&lop->lop_urgent))
2027 oap = cfs_list_entry(lop->lop_urgent.next,
2028 struct osc_async_page, oap_urgent_item);
2030 if (oap->oap_async_flags & ASYNC_HP) {
2031 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2038 static void on_list(cfs_list_t *item, cfs_list_t *list,
2041 if (cfs_list_empty(item) && should_be_on)
2042 cfs_list_add_tail(item, list);
2043 else if (!cfs_list_empty(item) && !should_be_on)
2044 cfs_list_del_init(item);
2047 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2048 * can find pages to build into rpcs quickly */
2049 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2051 if (lop_makes_hprpc(&loi->loi_write_lop) ||
2052 lop_makes_hprpc(&loi->loi_read_lop)) {
2054 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2055 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2057 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2058 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2059 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2060 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2063 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2064 loi->loi_write_lop.lop_num_pending);
2066 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2067 loi->loi_read_lop.lop_num_pending);
2070 static void lop_update_pending(struct client_obd *cli,
2071 struct loi_oap_pages *lop, int cmd, int delta)
2073 lop->lop_num_pending += delta;
2074 if (cmd & OBD_BRW_WRITE)
2075 cli->cl_pending_w_pages += delta;
2077 cli->cl_pending_r_pages += delta;
2081 * this is called when a sync waiter receives an interruption. Its job is to
2082 * get the caller woken as soon as possible. If its page hasn't been put in an
2083 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2084 * desiring interruption which will forcefully complete the rpc once the rpc
2087 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2089 struct loi_oap_pages *lop;
2090 struct lov_oinfo *loi;
2094 LASSERT(!oap->oap_interrupted);
2095 oap->oap_interrupted = 1;
2097 /* ok, it's been put in an rpc. only one oap gets a request reference */
2098 if (oap->oap_request != NULL) {
2099 ptlrpc_mark_interrupted(oap->oap_request);
2100 ptlrpcd_wake(oap->oap_request);
2101 ptlrpc_req_finished(oap->oap_request);
2102 oap->oap_request = NULL;
2106 * page completion may be called only if ->cpo_prep() method was
2107 * executed by osc_io_submit(), that also adds page the to pending list
2109 if (!cfs_list_empty(&oap->oap_pending_item)) {
2110 cfs_list_del_init(&oap->oap_pending_item);
2111 cfs_list_del_init(&oap->oap_urgent_item);
2114 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2115 &loi->loi_write_lop : &loi->loi_read_lop;
2116 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2117 loi_list_maint(oap->oap_cli, oap->oap_loi);
2118 rc = oap->oap_caller_ops->ap_completion(env,
2119 oap->oap_caller_data,
2120 oap->oap_cmd, NULL, -EINTR);
2126 /* this is trying to propogate async writeback errors back up to the
2127 * application. As an async write fails we record the error code for later if
2128 * the app does an fsync. As long as errors persist we force future rpcs to be
2129 * sync so that the app can get a sync error and break the cycle of queueing
2130 * pages for which writeback will fail. */
2131 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2138 ar->ar_force_sync = 1;
2139 ar->ar_min_xid = ptlrpc_sample_next_xid();
2144 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2145 ar->ar_force_sync = 0;
2148 void osc_oap_to_pending(struct osc_async_page *oap)
2150 struct loi_oap_pages *lop;
2152 if (oap->oap_cmd & OBD_BRW_WRITE)
2153 lop = &oap->oap_loi->loi_write_lop;
2155 lop = &oap->oap_loi->loi_read_lop;
2157 if (oap->oap_async_flags & ASYNC_HP)
2158 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2159 else if (oap->oap_async_flags & ASYNC_URGENT)
2160 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2161 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2162 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2165 /* this must be called holding the loi list lock to give coverage to exit_cache,
2166 * async_flag maintenance, and oap_request */
2167 static void osc_ap_completion(const struct lu_env *env,
2168 struct client_obd *cli, struct obdo *oa,
2169 struct osc_async_page *oap, int sent, int rc)
2174 if (oap->oap_request != NULL) {
2175 xid = ptlrpc_req_xid(oap->oap_request);
2176 ptlrpc_req_finished(oap->oap_request);
2177 oap->oap_request = NULL;
2180 cfs_spin_lock(&oap->oap_lock);
2181 oap->oap_async_flags = 0;
2182 cfs_spin_unlock(&oap->oap_lock);
2183 oap->oap_interrupted = 0;
2185 if (oap->oap_cmd & OBD_BRW_WRITE) {
2186 osc_process_ar(&cli->cl_ar, xid, rc);
2187 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2190 if (rc == 0 && oa != NULL) {
2191 if (oa->o_valid & OBD_MD_FLBLOCKS)
2192 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2193 if (oa->o_valid & OBD_MD_FLMTIME)
2194 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2195 if (oa->o_valid & OBD_MD_FLATIME)
2196 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2197 if (oa->o_valid & OBD_MD_FLCTIME)
2198 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2201 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2202 oap->oap_cmd, oa, rc);
2204 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2205 * I/O on the page could start, but OSC calls it under lock
2206 * and thus we can add oap back to pending safely */
2208 /* upper layer wants to leave the page on pending queue */
2209 osc_oap_to_pending(oap);
2211 osc_exit_cache(cli, oap, sent);
2215 static int brw_interpret(const struct lu_env *env,
2216 struct ptlrpc_request *req, void *data, int rc)
2218 struct osc_brw_async_args *aa = data;
2219 struct client_obd *cli;
2223 rc = osc_brw_fini_request(req, rc);
2224 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2225 if (osc_recoverable_error(rc)) {
2226 /* Only retry once for mmaped files since the mmaped page
2227 * might be modified at anytime. We have to retry at least
2228 * once in case there WAS really a corruption of the page
2229 * on the network, that was not caused by mmap() modifying
2230 * the page. Bug11742 */
2231 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2232 aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2233 aa->aa_oa->o_flags & OBD_FL_MMAP) {
2236 rc = osc_brw_redo_request(req, aa);
2243 capa_put(aa->aa_ocapa);
2244 aa->aa_ocapa = NULL;
2249 client_obd_list_lock(&cli->cl_loi_list_lock);
2251 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2252 * is called so we know whether to go to sync BRWs or wait for more
2253 * RPCs to complete */
2254 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2255 cli->cl_w_in_flight--;
2257 cli->cl_r_in_flight--;
2259 async = cfs_list_empty(&aa->aa_oaps);
2260 if (!async) { /* from osc_send_oap_rpc() */
2261 struct osc_async_page *oap, *tmp;
2262 /* the caller may re-use the oap after the completion call so
2263 * we need to clean it up a little */
2264 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2266 cfs_list_del_init(&oap->oap_rpc_item);
2267 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2269 OBDO_FREE(aa->aa_oa);
2270 } else { /* from async_internal() */
2272 for (i = 0; i < aa->aa_page_count; i++)
2273 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2275 osc_wake_cache_waiters(cli);
2276 osc_check_rpcs(env, cli);
2277 client_obd_list_unlock(&cli->cl_loi_list_lock);
2279 cl_req_completion(env, aa->aa_clerq, rc);
2280 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2285 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2286 struct client_obd *cli,
2287 cfs_list_t *rpc_list,
2288 int page_count, int cmd)
2290 struct ptlrpc_request *req;
2291 struct brw_page **pga = NULL;
2292 struct osc_brw_async_args *aa;
2293 struct obdo *oa = NULL;
2294 const struct obd_async_page_ops *ops = NULL;
2295 void *caller_data = NULL;
2296 struct osc_async_page *oap;
2297 struct osc_async_page *tmp;
2298 struct ost_body *body;
2299 struct cl_req *clerq = NULL;
2300 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2301 struct ldlm_lock *lock = NULL;
2302 struct cl_req_attr crattr;
2303 int i, rc, mpflag = 0;
2306 LASSERT(!cfs_list_empty(rpc_list));
2308 if (cmd & OBD_BRW_MEMALLOC)
2309 mpflag = cfs_memory_pressure_get_and_set();
2311 memset(&crattr, 0, sizeof crattr);
2312 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2314 GOTO(out, req = ERR_PTR(-ENOMEM));
2318 GOTO(out, req = ERR_PTR(-ENOMEM));
2321 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2322 struct cl_page *page = osc_oap2cl_page(oap);
2324 ops = oap->oap_caller_ops;
2325 caller_data = oap->oap_caller_data;
2327 clerq = cl_req_alloc(env, page, crt,
2328 1 /* only 1-object rpcs for
2331 GOTO(out, req = (void *)clerq);
2332 lock = oap->oap_ldlm_lock;
2334 pga[i] = &oap->oap_brw_page;
2335 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2336 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2337 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2339 cl_req_page_add(env, clerq, page);
2342 /* always get the data for the obdo for the rpc */
2343 LASSERT(ops != NULL);
2345 crattr.cra_capa = NULL;
2346 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2348 oa->o_handle = lock->l_remote_handle;
2349 oa->o_valid |= OBD_MD_FLHANDLE;
2352 rc = cl_req_prep(env, clerq);
2354 CERROR("cl_req_prep failed: %d\n", rc);
2355 GOTO(out, req = ERR_PTR(rc));
2358 sort_brw_pages(pga, page_count);
2359 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2360 pga, &req, crattr.cra_capa, 1, 0);
2362 CERROR("prep_req failed: %d\n", rc);
2363 GOTO(out, req = ERR_PTR(rc));
2366 if (cmd & OBD_BRW_MEMALLOC)
2367 req->rq_memalloc = 1;
2369 /* Need to update the timestamps after the request is built in case
2370 * we race with setattr (locally or in queue at OST). If OST gets
2371 * later setattr before earlier BRW (as determined by the request xid),
2372 * the OST will not use BRW timestamps. Sadly, there is no obvious
2373 * way to do this in a single call. bug 10150 */
2374 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2375 cl_req_attr_set(env, clerq, &crattr,
2376 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2378 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2379 aa = ptlrpc_req_async_args(req);
2380 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2381 cfs_list_splice(rpc_list, &aa->aa_oaps);
2382 CFS_INIT_LIST_HEAD(rpc_list);
2383 aa->aa_clerq = clerq;
2385 if (cmd & OBD_BRW_MEMALLOC)
2386 cfs_memory_pressure_restore(mpflag);
2388 capa_put(crattr.cra_capa);
2393 OBD_FREE(pga, sizeof(*pga) * page_count);
2394 /* this should happen rarely and is pretty bad, it makes the
2395 * pending list not follow the dirty order */
2396 client_obd_list_lock(&cli->cl_loi_list_lock);
2397 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2398 cfs_list_del_init(&oap->oap_rpc_item);
2400 /* queued sync pages can be torn down while the pages
2401 * were between the pending list and the rpc */
2402 if (oap->oap_interrupted) {
2403 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2404 osc_ap_completion(env, cli, NULL, oap, 0,
2408 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2410 if (clerq && !IS_ERR(clerq))
2411 cl_req_completion(env, clerq, PTR_ERR(req));
2417 * prepare pages for ASYNC io and put pages in send queue.
2419 * \param cmd OBD_BRW_* macroses
2420 * \param lop pending pages
2422 * \return zero if no page added to send queue.
2423 * \return 1 if pages successfully added to send queue.
2424 * \return negative on errors.
2427 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2428 struct lov_oinfo *loi,
2429 int cmd, struct loi_oap_pages *lop)
2431 struct ptlrpc_request *req;
2432 obd_count page_count = 0;
2433 struct osc_async_page *oap = NULL, *tmp;
2434 struct osc_brw_async_args *aa;
2435 const struct obd_async_page_ops *ops;
2436 CFS_LIST_HEAD(rpc_list);
2437 CFS_LIST_HEAD(tmp_list);
2438 unsigned int ending_offset;
2439 obd_off starting_offset = OBD_OBJECT_EOF;
2440 int starting_page_off = 0;
2441 int srvlock = 0, mem_tight = 0;
2442 struct cl_object *clob = NULL;
2445 /* ASYNC_HP pages first. At present, when the lock the pages is
2446 * to be canceled, the pages covered by the lock will be sent out
2447 * with ASYNC_HP. We have to send out them as soon as possible. */
2448 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2449 if (oap->oap_async_flags & ASYNC_HP)
2450 cfs_list_move(&oap->oap_pending_item, &tmp_list);
2452 cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2453 if (++page_count >= cli->cl_max_pages_per_rpc)
2457 cfs_list_splice(&tmp_list, &lop->lop_pending);
2460 /* first we find the pages we're allowed to work with */
2461 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2463 ops = oap->oap_caller_ops;
2465 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2466 "magic 0x%x\n", oap, oap->oap_magic);
2469 /* pin object in memory, so that completion call-backs
2470 * can be safely called under client_obd_list lock. */
2471 clob = osc_oap2cl_page(oap)->cp_obj;
2472 cl_object_get(clob);
2475 if (page_count != 0 &&
2476 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2477 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2478 " oap %p, page %p, srvlock %u\n",
2479 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2483 /* If there is a gap at the start of this page, it can't merge
2484 * with any previous page, so we'll hand the network a
2485 * "fragmented" page array that it can't transfer in 1 RDMA */
2486 if (oap->oap_obj_off < starting_offset) {
2487 if (starting_page_off != 0)
2490 starting_page_off = oap->oap_page_off;
2491 starting_offset = oap->oap_obj_off + starting_page_off;
2492 } else if (oap->oap_page_off != 0)
2495 /* in llite being 'ready' equates to the page being locked
2496 * until completion unlocks it. commit_write submits a page
2497 * as not ready because its unlock will happen unconditionally
2498 * as the call returns. if we race with commit_write giving
2499 * us that page we don't want to create a hole in the page
2500 * stream, so we stop and leave the rpc to be fired by
2501 * another dirtier or kupdated interval (the not ready page
2502 * will still be on the dirty list). we could call in
2503 * at the end of ll_file_write to process the queue again. */
2504 if (!(oap->oap_async_flags & ASYNC_READY)) {
2505 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2508 CDEBUG(D_INODE, "oap %p page %p returned %d "
2509 "instead of ready\n", oap,
2513 /* llite is telling us that the page is still
2514 * in commit_write and that we should try
2515 * and put it in an rpc again later. we
2516 * break out of the loop so we don't create
2517 * a hole in the sequence of pages in the rpc
2522 /* the io isn't needed.. tell the checks
2523 * below to complete the rpc with EINTR */
2524 cfs_spin_lock(&oap->oap_lock);
2525 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2526 cfs_spin_unlock(&oap->oap_lock);
2527 oap->oap_count = -EINTR;
2530 cfs_spin_lock(&oap->oap_lock);
2531 oap->oap_async_flags |= ASYNC_READY;
2532 cfs_spin_unlock(&oap->oap_lock);
2535 LASSERTF(0, "oap %p page %p returned %d "
2536 "from make_ready\n", oap,
2544 * Page submitted for IO has to be locked. Either by
2545 * ->ap_make_ready() or by higher layers.
2547 #if defined(__KERNEL__) && defined(__linux__)
2549 struct cl_page *page;
2551 page = osc_oap2cl_page(oap);
2553 if (page->cp_type == CPT_CACHEABLE &&
2554 !(PageLocked(oap->oap_page) &&
2555 (CheckWriteback(oap->oap_page, cmd)))) {
2556 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2558 (long)oap->oap_page->flags,
2559 oap->oap_async_flags);
2565 /* take the page out of our book-keeping */
2566 cfs_list_del_init(&oap->oap_pending_item);
2567 lop_update_pending(cli, lop, cmd, -1);
2568 cfs_list_del_init(&oap->oap_urgent_item);
2570 /* ask the caller for the size of the io as the rpc leaves. */
2571 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2573 ops->ap_refresh_count(env, oap->oap_caller_data,
2575 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2577 if (oap->oap_count <= 0) {
2578 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2580 osc_ap_completion(env, cli, NULL,
2581 oap, 0, oap->oap_count);
2585 /* now put the page back in our accounting */
2586 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2587 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2589 if (page_count == 0)
2590 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2591 if (++page_count >= cli->cl_max_pages_per_rpc)
2594 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2595 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2596 * have the same alignment as the initial writes that allocated
2597 * extents on the server. */
2598 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2599 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2600 if (ending_offset == 0)
2603 /* If there is a gap at the end of this page, it can't merge
2604 * with any subsequent pages, so we'll hand the network a
2605 * "fragmented" page array that it can't transfer in 1 RDMA */
2606 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2610 osc_wake_cache_waiters(cli);
2612 loi_list_maint(cli, loi);
2614 client_obd_list_unlock(&cli->cl_loi_list_lock);
2617 cl_object_put(env, clob);
2619 if (page_count == 0) {
2620 client_obd_list_lock(&cli->cl_loi_list_lock);
2624 req = osc_build_req(env, cli, &rpc_list, page_count,
2625 mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2627 LASSERT(cfs_list_empty(&rpc_list));
2628 loi_list_maint(cli, loi);
2629 RETURN(PTR_ERR(req));
2632 aa = ptlrpc_req_async_args(req);
2634 starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2635 if (cmd == OBD_BRW_READ) {
2636 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2637 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2638 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2639 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2641 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2642 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2643 cli->cl_w_in_flight);
2644 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2645 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2647 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2649 client_obd_list_lock(&cli->cl_loi_list_lock);
2651 if (cmd == OBD_BRW_READ)
2652 cli->cl_r_in_flight++;
2654 cli->cl_w_in_flight++;
2656 /* queued sync pages can be torn down while the pages
2657 * were between the pending list and the rpc */
2659 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2660 /* only one oap gets a request reference */
2663 if (oap->oap_interrupted && !req->rq_intr) {
2664 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2666 ptlrpc_mark_interrupted(req);
2670 tmp->oap_request = ptlrpc_request_addref(req);
2672 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2673 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2675 req->rq_interpret_reply = brw_interpret;
2676 ptlrpcd_add_req(req, PSCOPE_BRW);
2680 #define LOI_DEBUG(LOI, STR, args...) \
2681 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2682 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2683 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2684 (LOI)->loi_write_lop.lop_num_pending, \
2685 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2686 (LOI)->loi_read_lop.lop_num_pending, \
2687 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2690 /* This is called by osc_check_rpcs() to find which objects have pages that
2691 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2692 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2696 /* First return objects that have blocked locks so that they
2697 * will be flushed quickly and other clients can get the lock,
2698 * then objects which have pages ready to be stuffed into RPCs */
2699 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2700 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2701 struct lov_oinfo, loi_hp_ready_item));
2702 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2703 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2704 struct lov_oinfo, loi_ready_item));
2706 /* then if we have cache waiters, return all objects with queued
2707 * writes. This is especially important when many small files
2708 * have filled up the cache and not been fired into rpcs because
2709 * they don't pass the nr_pending/object threshhold */
2710 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2711 !cfs_list_empty(&cli->cl_loi_write_list))
2712 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2713 struct lov_oinfo, loi_write_item));
2715 /* then return all queued objects when we have an invalid import
2716 * so that they get flushed */
2717 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2718 if (!cfs_list_empty(&cli->cl_loi_write_list))
2719 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2722 if (!cfs_list_empty(&cli->cl_loi_read_list))
2723 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2724 struct lov_oinfo, loi_read_item));
2729 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2731 struct osc_async_page *oap;
2734 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2735 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2736 struct osc_async_page, oap_urgent_item);
2737 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2740 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2741 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2742 struct osc_async_page, oap_urgent_item);
2743 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2746 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2749 /* called with the loi list lock held */
2750 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2752 struct lov_oinfo *loi;
2753 int rc = 0, race_counter = 0;
2756 while ((loi = osc_next_loi(cli)) != NULL) {
2757 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2759 if (osc_max_rpc_in_flight(cli, loi))
2762 /* attempt some read/write balancing by alternating between
2763 * reads and writes in an object. The makes_rpc checks here
2764 * would be redundant if we were getting read/write work items
2765 * instead of objects. we don't want send_oap_rpc to drain a
2766 * partial read pending queue when we're given this object to
2767 * do io on writes while there are cache waiters */
2768 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2769 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2770 &loi->loi_write_lop);
2772 CERROR("Write request failed with %d\n", rc);
2774 /* osc_send_oap_rpc failed, mostly because of
2777 * It can't break here, because if:
2778 * - a page was submitted by osc_io_submit, so
2780 * - no request in flight
2781 * - no subsequent request
2782 * The system will be in live-lock state,
2783 * because there is no chance to call
2784 * osc_io_unplug() and osc_check_rpcs() any
2785 * more. pdflush can't help in this case,
2786 * because it might be blocked at grabbing
2787 * the page lock as we mentioned.
2789 * Anyway, continue to drain pages. */
2798 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2799 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2800 &loi->loi_read_lop);
2802 CERROR("Read request failed with %d\n", rc);
2810 /* attempt some inter-object balancing by issuing rpcs
2811 * for each object in turn */
2812 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2813 cfs_list_del_init(&loi->loi_hp_ready_item);
2814 if (!cfs_list_empty(&loi->loi_ready_item))
2815 cfs_list_del_init(&loi->loi_ready_item);
2816 if (!cfs_list_empty(&loi->loi_write_item))
2817 cfs_list_del_init(&loi->loi_write_item);
2818 if (!cfs_list_empty(&loi->loi_read_item))
2819 cfs_list_del_init(&loi->loi_read_item);
2821 loi_list_maint(cli, loi);
2823 /* send_oap_rpc fails with 0 when make_ready tells it to
2824 * back off. llite's make_ready does this when it tries
2825 * to lock a page queued for write that is already locked.
2826 * we want to try sending rpcs from many objects, but we
2827 * don't want to spin failing with 0. */
2828 if (race_counter == 10)
2834 /* we're trying to queue a page in the osc so we're subject to the
2835 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2836 * If the osc's queued pages are already at that limit, then we want to sleep
2837 * until there is space in the osc's queue for us. We also may be waiting for
2838 * write credits from the OST if there are RPCs in flight that may return some
2839 * before we fall back to sync writes.
2841 * We need this know our allocation was granted in the presence of signals */
2842 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2846 client_obd_list_lock(&cli->cl_loi_list_lock);
2847 rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2848 client_obd_list_unlock(&cli->cl_loi_list_lock);
2853 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2856 int osc_enter_cache_try(const struct lu_env *env,
2857 struct client_obd *cli, struct lov_oinfo *loi,
2858 struct osc_async_page *oap, int transient)
2862 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2864 osc_consume_write_grant(cli, &oap->oap_brw_page);
2866 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2867 cfs_atomic_inc(&obd_dirty_transit_pages);
2868 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2874 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2875 * grant or cache space. */
2876 static int osc_enter_cache(const struct lu_env *env,
2877 struct client_obd *cli, struct lov_oinfo *loi,
2878 struct osc_async_page *oap)
2880 struct osc_cache_waiter ocw;
2881 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2885 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2886 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2887 cli->cl_dirty_max, obd_max_dirty_pages,
2888 cli->cl_lost_grant, cli->cl_avail_grant);
2890 /* force the caller to try sync io. this can jump the list
2891 * of queued writes and create a discontiguous rpc stream */
2892 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2893 loi->loi_ar.ar_force_sync)
2896 /* Hopefully normal case - cache space and write credits available */
2897 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2898 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2899 osc_enter_cache_try(env, cli, loi, oap, 0))
2902 /* It is safe to block as a cache waiter as long as there is grant
2903 * space available or the hope of additional grant being returned
2904 * when an in flight write completes. Using the write back cache
2905 * if possible is preferable to sending the data synchronously
2906 * because write pages can then be merged in to large requests.
2907 * The addition of this cache waiter will causing pending write
2908 * pages to be sent immediately. */
2909 if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2910 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2911 cfs_waitq_init(&ocw.ocw_waitq);
2915 loi_list_maint(cli, loi);
2916 osc_check_rpcs(env, cli);
2917 client_obd_list_unlock(&cli->cl_loi_list_lock);
2919 CDEBUG(D_CACHE, "sleeping for cache space\n");
2920 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2922 client_obd_list_lock(&cli->cl_loi_list_lock);
2923 if (!cfs_list_empty(&ocw.ocw_entry)) {
2924 cfs_list_del(&ocw.ocw_entry);
2934 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2935 struct lov_oinfo *loi, cfs_page_t *page,
2936 obd_off offset, const struct obd_async_page_ops *ops,
2937 void *data, void **res, int nocache,
2938 struct lustre_handle *lockh)
2940 struct osc_async_page *oap;
2945 return cfs_size_round(sizeof(*oap));
2948 oap->oap_magic = OAP_MAGIC;
2949 oap->oap_cli = &exp->exp_obd->u.cli;
2952 oap->oap_caller_ops = ops;
2953 oap->oap_caller_data = data;
2955 oap->oap_page = page;
2956 oap->oap_obj_off = offset;
2957 if (!client_is_remote(exp) &&
2958 cfs_capable(CFS_CAP_SYS_RESOURCE))
2959 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2961 LASSERT(!(offset & ~CFS_PAGE_MASK));
2963 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2964 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2965 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2966 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2968 cfs_spin_lock_init(&oap->oap_lock);
2969 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2973 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2974 struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2975 struct osc_async_page *oap, int cmd, int off,
2976 int count, obd_flag brw_flags, enum async_flags async_flags)
2978 struct client_obd *cli = &exp->exp_obd->u.cli;
2982 if (oap->oap_magic != OAP_MAGIC)
2985 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2988 if (!cfs_list_empty(&oap->oap_pending_item) ||
2989 !cfs_list_empty(&oap->oap_urgent_item) ||
2990 !cfs_list_empty(&oap->oap_rpc_item))
2993 /* check if the file's owner/group is over quota */
2994 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2995 struct cl_object *obj;
2996 struct cl_attr attr; /* XXX put attr into thread info */
2997 unsigned int qid[MAXQUOTAS];
2999 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
3001 cl_object_attr_lock(obj);
3002 rc = cl_object_attr_get(env, obj, &attr);
3003 cl_object_attr_unlock(obj);
3005 qid[USRQUOTA] = attr.cat_uid;
3006 qid[GRPQUOTA] = attr.cat_gid;
3008 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
3015 loi = lsm->lsm_oinfo[0];
3017 client_obd_list_lock(&cli->cl_loi_list_lock);
3019 LASSERT(off + count <= CFS_PAGE_SIZE);
3021 oap->oap_page_off = off;
3022 oap->oap_count = count;
3023 oap->oap_brw_flags = brw_flags;
3024 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3025 if (cfs_memory_pressure_get())
3026 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3027 cfs_spin_lock(&oap->oap_lock);
3028 oap->oap_async_flags = async_flags;
3029 cfs_spin_unlock(&oap->oap_lock);
3031 if (cmd & OBD_BRW_WRITE) {
3032 rc = osc_enter_cache(env, cli, loi, oap);
3034 client_obd_list_unlock(&cli->cl_loi_list_lock);
3039 osc_oap_to_pending(oap);
3040 loi_list_maint(cli, loi);
3042 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3045 osc_check_rpcs(env, cli);
3046 client_obd_list_unlock(&cli->cl_loi_list_lock);
3051 /* aka (~was & now & flag), but this is more clear :) */
3052 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3054 int osc_set_async_flags_base(struct client_obd *cli,
3055 struct lov_oinfo *loi, struct osc_async_page *oap,
3056 obd_flag async_flags)
3058 struct loi_oap_pages *lop;
3062 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3064 if (oap->oap_cmd & OBD_BRW_WRITE) {
3065 lop = &loi->loi_write_lop;
3067 lop = &loi->loi_read_lop;
3070 if ((oap->oap_async_flags & async_flags) == async_flags)
3073 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3074 flags |= ASYNC_READY;
3076 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3077 cfs_list_empty(&oap->oap_rpc_item)) {
3078 if (oap->oap_async_flags & ASYNC_HP)
3079 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3081 cfs_list_add_tail(&oap->oap_urgent_item,
3083 flags |= ASYNC_URGENT;
3084 loi_list_maint(cli, loi);
3086 cfs_spin_lock(&oap->oap_lock);
3087 oap->oap_async_flags |= flags;
3088 cfs_spin_unlock(&oap->oap_lock);
3090 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3091 oap->oap_async_flags);
3095 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3096 struct lov_oinfo *loi, struct osc_async_page *oap)
3098 struct client_obd *cli = &exp->exp_obd->u.cli;
3099 struct loi_oap_pages *lop;
3103 if (oap->oap_magic != OAP_MAGIC)
3107 loi = lsm->lsm_oinfo[0];
3109 if (oap->oap_cmd & OBD_BRW_WRITE) {
3110 lop = &loi->loi_write_lop;
3112 lop = &loi->loi_read_lop;
3115 client_obd_list_lock(&cli->cl_loi_list_lock);
3117 if (!cfs_list_empty(&oap->oap_rpc_item))
3118 GOTO(out, rc = -EBUSY);
3120 osc_exit_cache(cli, oap, 0);
3121 osc_wake_cache_waiters(cli);
3123 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3124 cfs_list_del_init(&oap->oap_urgent_item);
3125 cfs_spin_lock(&oap->oap_lock);
3126 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3127 cfs_spin_unlock(&oap->oap_lock);
3129 if (!cfs_list_empty(&oap->oap_pending_item)) {
3130 cfs_list_del_init(&oap->oap_pending_item);
3131 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3133 loi_list_maint(cli, loi);
3134 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3136 client_obd_list_unlock(&cli->cl_loi_list_lock);
3140 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3141 struct ldlm_enqueue_info *einfo)
3143 void *data = einfo->ei_cbdata;
3146 LASSERT(lock != NULL);
3147 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3148 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3149 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3150 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3152 lock_res_and_lock(lock);
3153 cfs_spin_lock(&osc_ast_guard);
3155 if (lock->l_ast_data == NULL)
3156 lock->l_ast_data = data;
3157 if (lock->l_ast_data == data)
3160 cfs_spin_unlock(&osc_ast_guard);
3161 unlock_res_and_lock(lock);
3166 static int osc_set_data_with_check(struct lustre_handle *lockh,
3167 struct ldlm_enqueue_info *einfo)
3169 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3173 set = osc_set_lock_data_with_check(lock, einfo);
3174 LDLM_LOCK_PUT(lock);
3176 CERROR("lockh %p, data %p - client evicted?\n",
3177 lockh, einfo->ei_cbdata);
3181 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3182 ldlm_iterator_t replace, void *data)
3184 struct ldlm_res_id res_id;
3185 struct obd_device *obd = class_exp2obd(exp);
3187 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3188 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3192 /* find any ldlm lock of the inode in osc
3196 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3197 ldlm_iterator_t replace, void *data)
3199 struct ldlm_res_id res_id;
3200 struct obd_device *obd = class_exp2obd(exp);
3203 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3204 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3205 if (rc == LDLM_ITER_STOP)
3207 if (rc == LDLM_ITER_CONTINUE)
3212 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3213 obd_enqueue_update_f upcall, void *cookie,
3216 int intent = *flags & LDLM_FL_HAS_INTENT;
3220 /* The request was created before ldlm_cli_enqueue call. */
3221 if (rc == ELDLM_LOCK_ABORTED) {
3222 struct ldlm_reply *rep;
3223 rep = req_capsule_server_get(&req->rq_pill,
3226 LASSERT(rep != NULL);
3227 if (rep->lock_policy_res1)
3228 rc = rep->lock_policy_res1;
3232 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3233 *flags |= LDLM_FL_LVB_READY;
3234 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3235 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3238 /* Call the update callback. */
3239 rc = (*upcall)(cookie, rc);
3243 static int osc_enqueue_interpret(const struct lu_env *env,
3244 struct ptlrpc_request *req,
3245 struct osc_enqueue_args *aa, int rc)
3247 struct ldlm_lock *lock;
3248 struct lustre_handle handle;
3251 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3252 * might be freed anytime after lock upcall has been called. */
3253 lustre_handle_copy(&handle, aa->oa_lockh);
3254 mode = aa->oa_ei->ei_mode;
3256 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3258 lock = ldlm_handle2lock(&handle);
3260 /* Take an additional reference so that a blocking AST that
3261 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3262 * to arrive after an upcall has been executed by
3263 * osc_enqueue_fini(). */
3264 ldlm_lock_addref(&handle, mode);
3266 /* Let CP AST to grant the lock first. */
3267 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3269 /* Complete obtaining the lock procedure. */
3270 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3271 mode, aa->oa_flags, aa->oa_lvb,
3272 sizeof(*aa->oa_lvb), &handle, rc);
3273 /* Complete osc stuff. */
3274 rc = osc_enqueue_fini(req, aa->oa_lvb,
3275 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3277 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3279 /* Release the lock for async request. */
3280 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3282 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3283 * not already released by
3284 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3286 ldlm_lock_decref(&handle, mode);
3288 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3289 aa->oa_lockh, req, aa);
3290 ldlm_lock_decref(&handle, mode);
3291 LDLM_LOCK_PUT(lock);
3295 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3296 struct lov_oinfo *loi, int flags,
3297 struct ost_lvb *lvb, __u32 mode, int rc)
3299 if (rc == ELDLM_OK) {
3300 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3303 LASSERT(lock != NULL);
3304 loi->loi_lvb = *lvb;
3305 tmp = loi->loi_lvb.lvb_size;
3306 /* Extend KMS up to the end of this lock and no further
3307 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3308 if (tmp > lock->l_policy_data.l_extent.end)
3309 tmp = lock->l_policy_data.l_extent.end + 1;
3310 if (tmp >= loi->loi_kms) {
3311 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3312 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3313 loi_kms_set(loi, tmp);
3315 LDLM_DEBUG(lock, "lock acquired, setting rss="
3316 LPU64"; leaving kms="LPU64", end="LPU64,
3317 loi->loi_lvb.lvb_size, loi->loi_kms,
3318 lock->l_policy_data.l_extent.end);
3320 ldlm_lock_allow_match(lock);
3321 LDLM_LOCK_PUT(lock);
3322 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3323 loi->loi_lvb = *lvb;
3324 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3325 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3329 EXPORT_SYMBOL(osc_update_enqueue);
3331 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3333 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3334 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3335 * other synchronous requests, however keeping some locks and trying to obtain
3336 * others may take a considerable amount of time in a case of ost failure; and
3337 * when other sync requests do not get released lock from a client, the client
3338 * is excluded from the cluster -- such scenarious make the life difficult, so
3339 * release locks just after they are obtained. */
3340 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3341 int *flags, ldlm_policy_data_t *policy,
3342 struct ost_lvb *lvb, int kms_valid,
3343 obd_enqueue_update_f upcall, void *cookie,
3344 struct ldlm_enqueue_info *einfo,
3345 struct lustre_handle *lockh,
3346 struct ptlrpc_request_set *rqset, int async)
3348 struct obd_device *obd = exp->exp_obd;
3349 struct ptlrpc_request *req = NULL;
3350 int intent = *flags & LDLM_FL_HAS_INTENT;
3355 /* Filesystem lock extents are extended to page boundaries so that
3356 * dealing with the page cache is a little smoother. */
3357 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3358 policy->l_extent.end |= ~CFS_PAGE_MASK;
3361 * kms is not valid when either object is completely fresh (so that no
3362 * locks are cached), or object was evicted. In the latter case cached
3363 * lock cannot be used, because it would prime inode state with
3364 * potentially stale LVB.
3369 /* Next, search for already existing extent locks that will cover us */
3370 /* If we're trying to read, we also search for an existing PW lock. The
3371 * VFS and page cache already protect us locally, so lots of readers/
3372 * writers can share a single PW lock.
3374 * There are problems with conversion deadlocks, so instead of
3375 * converting a read lock to a write lock, we'll just enqueue a new
3378 * At some point we should cancel the read lock instead of making them
3379 * send us a blocking callback, but there are problems with canceling
3380 * locks out from other users right now, too. */
3381 mode = einfo->ei_mode;
3382 if (einfo->ei_mode == LCK_PR)
3384 mode = ldlm_lock_match(obd->obd_namespace,
3385 *flags | LDLM_FL_LVB_READY, res_id,
3386 einfo->ei_type, policy, mode, lockh, 0);
3388 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3390 if (osc_set_lock_data_with_check(matched, einfo)) {
3391 /* addref the lock only if not async requests and PW
3392 * lock is matched whereas we asked for PR. */
3393 if (!rqset && einfo->ei_mode != mode)
3394 ldlm_lock_addref(lockh, LCK_PR);
3396 /* I would like to be able to ASSERT here that
3397 * rss <= kms, but I can't, for reasons which
3398 * are explained in lov_enqueue() */
3401 /* We already have a lock, and it's referenced */
3402 (*upcall)(cookie, ELDLM_OK);
3404 /* For async requests, decref the lock. */
3405 if (einfo->ei_mode != mode)
3406 ldlm_lock_decref(lockh, LCK_PW);
3408 ldlm_lock_decref(lockh, einfo->ei_mode);
3409 LDLM_LOCK_PUT(matched);
3412 ldlm_lock_decref(lockh, mode);
3413 LDLM_LOCK_PUT(matched);
3418 CFS_LIST_HEAD(cancels);
3419 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3420 &RQF_LDLM_ENQUEUE_LVB);
3424 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3426 ptlrpc_request_free(req);
3430 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3432 ptlrpc_request_set_replen(req);
3435 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3436 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3438 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3439 sizeof(*lvb), lockh, async);
3442 struct osc_enqueue_args *aa;
3443 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3444 aa = ptlrpc_req_async_args(req);
3447 aa->oa_flags = flags;
3448 aa->oa_upcall = upcall;
3449 aa->oa_cookie = cookie;
3451 aa->oa_lockh = lockh;
3453 req->rq_interpret_reply =
3454 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3455 if (rqset == PTLRPCD_SET)
3456 ptlrpcd_add_req(req, PSCOPE_OTHER);
3458 ptlrpc_set_add_req(rqset, req);
3459 } else if (intent) {
3460 ptlrpc_req_finished(req);
3465 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3467 ptlrpc_req_finished(req);
3472 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3473 struct ldlm_enqueue_info *einfo,
3474 struct ptlrpc_request_set *rqset)
3476 struct ldlm_res_id res_id;
3480 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3481 oinfo->oi_md->lsm_object_seq, &res_id);
3483 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3484 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3485 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3486 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3487 rqset, rqset != NULL);
3491 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3492 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3493 int *flags, void *data, struct lustre_handle *lockh,
3496 struct obd_device *obd = exp->exp_obd;
3497 int lflags = *flags;
3501 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3504 /* Filesystem lock extents are extended to page boundaries so that
3505 * dealing with the page cache is a little smoother */
3506 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3507 policy->l_extent.end |= ~CFS_PAGE_MASK;
3509 /* Next, search for already existing extent locks that will cover us */
3510 /* If we're trying to read, we also search for an existing PW lock. The
3511 * VFS and page cache already protect us locally, so lots of readers/
3512 * writers can share a single PW lock. */
3516 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3517 res_id, type, policy, rc, lockh, unref);
3520 if (!osc_set_data_with_check(lockh, data)) {
3521 if (!(lflags & LDLM_FL_TEST_LOCK))
3522 ldlm_lock_decref(lockh, rc);
3526 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3527 ldlm_lock_addref(lockh, LCK_PR);
3528 ldlm_lock_decref(lockh, LCK_PW);
3535 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3539 if (unlikely(mode == LCK_GROUP))
3540 ldlm_lock_decref_and_cancel(lockh, mode);
3542 ldlm_lock_decref(lockh, mode);
3547 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3548 __u32 mode, struct lustre_handle *lockh)
3551 RETURN(osc_cancel_base(lockh, mode));
3554 static int osc_cancel_unused(struct obd_export *exp,
3555 struct lov_stripe_md *lsm,
3556 ldlm_cancel_flags_t flags,
3559 struct obd_device *obd = class_exp2obd(exp);
3560 struct ldlm_res_id res_id, *resp = NULL;
3563 resp = osc_build_res_name(lsm->lsm_object_id,
3564 lsm->lsm_object_seq, &res_id);
3567 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3570 static int osc_statfs_interpret(const struct lu_env *env,
3571 struct ptlrpc_request *req,
3572 struct osc_async_args *aa, int rc)
3574 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3575 struct obd_statfs *msfs;
3580 /* The request has in fact never been sent
3581 * due to issues at a higher level (LOV).
3582 * Exit immediately since the caller is
3583 * aware of the problem and takes care
3584 * of the clean up */
3587 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3588 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3594 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3596 GOTO(out, rc = -EPROTO);
3599 /* Reinitialize the RDONLY and DEGRADED flags at the client
3600 * on each statfs, so they don't stay set permanently. */
3601 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3603 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3604 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3605 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3606 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3608 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3609 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3610 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3611 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3613 /* Add a bit of hysteresis so this flag isn't continually flapping,
3614 * and ensure that new files don't get extremely fragmented due to
3615 * only a small amount of available space in the filesystem.
3616 * We want to set the NOSPC flag when there is less than ~0.1% free
3617 * and clear it when there is at least ~0.2% free space, so:
3618 * avail < ~0.1% max max = avail + used
3619 * 1025 * avail < avail + used used = blocks - free
3620 * 1024 * avail < used
3621 * 1024 * avail < blocks - free
3622 * avail < ((blocks - free) >> 10)
3624 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3625 * lose that amount of space so in those cases we report no space left
3626 * if their is less than 1 GB left. */
3627 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3628 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3629 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3630 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3631 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3632 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3633 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3635 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3637 *aa->aa_oi->oi_osfs = *msfs;
3639 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3643 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3644 __u64 max_age, struct ptlrpc_request_set *rqset)
3646 struct ptlrpc_request *req;
3647 struct osc_async_args *aa;
3651 /* We could possibly pass max_age in the request (as an absolute
3652 * timestamp or a "seconds.usec ago") so the target can avoid doing
3653 * extra calls into the filesystem if that isn't necessary (e.g.
3654 * during mount that would help a bit). Having relative timestamps
3655 * is not so great if request processing is slow, while absolute
3656 * timestamps are not ideal because they need time synchronization. */
3657 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3661 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3663 ptlrpc_request_free(req);
3666 ptlrpc_request_set_replen(req);
3667 req->rq_request_portal = OST_CREATE_PORTAL;
3668 ptlrpc_at_set_req_timeout(req);
3670 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3671 /* procfs requests not want stat in wait for avoid deadlock */
3672 req->rq_no_resend = 1;
3673 req->rq_no_delay = 1;
3676 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3677 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3678 aa = ptlrpc_req_async_args(req);
3681 ptlrpc_set_add_req(rqset, req);
3685 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3686 __u64 max_age, __u32 flags)
3688 struct obd_statfs *msfs;
3689 struct ptlrpc_request *req;
3690 struct obd_import *imp = NULL;
3694 /*Since the request might also come from lprocfs, so we need
3695 *sync this with client_disconnect_export Bug15684*/
3696 cfs_down_read(&obd->u.cli.cl_sem);
3697 if (obd->u.cli.cl_import)
3698 imp = class_import_get(obd->u.cli.cl_import);
3699 cfs_up_read(&obd->u.cli.cl_sem);
3703 /* We could possibly pass max_age in the request (as an absolute
3704 * timestamp or a "seconds.usec ago") so the target can avoid doing
3705 * extra calls into the filesystem if that isn't necessary (e.g.
3706 * during mount that would help a bit). Having relative timestamps
3707 * is not so great if request processing is slow, while absolute
3708 * timestamps are not ideal because they need time synchronization. */
3709 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3711 class_import_put(imp);
3716 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3718 ptlrpc_request_free(req);
3721 ptlrpc_request_set_replen(req);
3722 req->rq_request_portal = OST_CREATE_PORTAL;
3723 ptlrpc_at_set_req_timeout(req);
3725 if (flags & OBD_STATFS_NODELAY) {
3726 /* procfs requests not want stat in wait for avoid deadlock */
3727 req->rq_no_resend = 1;
3728 req->rq_no_delay = 1;
3731 rc = ptlrpc_queue_wait(req);
3735 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3737 GOTO(out, rc = -EPROTO);
3744 ptlrpc_req_finished(req);
3748 /* Retrieve object striping information.
3750 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3751 * the maximum number of OST indices which will fit in the user buffer.
3752 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3754 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3756 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3757 struct lov_user_md_v3 lum, *lumk;
3758 struct lov_user_ost_data_v1 *lmm_objects;
3759 int rc = 0, lum_size;
3765 /* we only need the header part from user space to get lmm_magic and
3766 * lmm_stripe_count, (the header part is common to v1 and v3) */
3767 lum_size = sizeof(struct lov_user_md_v1);
3768 if (cfs_copy_from_user(&lum, lump, lum_size))
3771 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3772 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3775 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3776 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3777 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3778 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3780 /* we can use lov_mds_md_size() to compute lum_size
3781 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3782 if (lum.lmm_stripe_count > 0) {
3783 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3784 OBD_ALLOC(lumk, lum_size);
3788 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3789 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3791 lmm_objects = &(lumk->lmm_objects[0]);
3792 lmm_objects->l_object_id = lsm->lsm_object_id;
3794 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3798 lumk->lmm_object_id = lsm->lsm_object_id;
3799 lumk->lmm_object_seq = lsm->lsm_object_seq;
3800 lumk->lmm_stripe_count = 1;
3802 if (cfs_copy_to_user(lump, lumk, lum_size))
3806 OBD_FREE(lumk, lum_size);
3812 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3813 void *karg, void *uarg)
3815 struct obd_device *obd = exp->exp_obd;
3816 struct obd_ioctl_data *data = karg;
3820 if (!cfs_try_module_get(THIS_MODULE)) {
3821 CERROR("Can't get module. Is it alive?");
3825 case OBD_IOC_LOV_GET_CONFIG: {
3827 struct lov_desc *desc;
3828 struct obd_uuid uuid;
3832 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3833 GOTO(out, err = -EINVAL);
3835 data = (struct obd_ioctl_data *)buf;
3837 if (sizeof(*desc) > data->ioc_inllen1) {
3838 obd_ioctl_freedata(buf, len);
3839 GOTO(out, err = -EINVAL);
3842 if (data->ioc_inllen2 < sizeof(uuid)) {
3843 obd_ioctl_freedata(buf, len);
3844 GOTO(out, err = -EINVAL);
3847 desc = (struct lov_desc *)data->ioc_inlbuf1;
3848 desc->ld_tgt_count = 1;
3849 desc->ld_active_tgt_count = 1;
3850 desc->ld_default_stripe_count = 1;
3851 desc->ld_default_stripe_size = 0;
3852 desc->ld_default_stripe_offset = 0;
3853 desc->ld_pattern = 0;
3854 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3856 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3858 err = cfs_copy_to_user((void *)uarg, buf, len);
3861 obd_ioctl_freedata(buf, len);
3864 case LL_IOC_LOV_SETSTRIPE:
3865 err = obd_alloc_memmd(exp, karg);
3869 case LL_IOC_LOV_GETSTRIPE:
3870 err = osc_getstripe(karg, uarg);
3872 case OBD_IOC_CLIENT_RECOVER:
3873 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3878 case IOC_OSC_SET_ACTIVE:
3879 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3882 case OBD_IOC_POLL_QUOTACHECK:
3883 err = lquota_poll_check(quota_interface, exp,
3884 (struct if_quotacheck *)karg);
3886 case OBD_IOC_PING_TARGET:
3887 err = ptlrpc_obd_ping(obd);
3890 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3891 cmd, cfs_curproc_comm());
3892 GOTO(out, err = -ENOTTY);
3895 cfs_module_put(THIS_MODULE);
3899 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3900 void *key, __u32 *vallen, void *val,
3901 struct lov_stripe_md *lsm)
3904 if (!vallen || !val)
3907 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3908 __u32 *stripe = val;
3909 *vallen = sizeof(*stripe);
3912 } else if (KEY_IS(KEY_LAST_ID)) {
3913 struct ptlrpc_request *req;
3918 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3919 &RQF_OST_GET_INFO_LAST_ID);
3923 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3924 RCL_CLIENT, keylen);
3925 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3927 ptlrpc_request_free(req);
3931 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3932 memcpy(tmp, key, keylen);
3934 req->rq_no_delay = req->rq_no_resend = 1;
3935 ptlrpc_request_set_replen(req);
3936 rc = ptlrpc_queue_wait(req);
3940 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3942 GOTO(out, rc = -EPROTO);
3944 *((obd_id *)val) = *reply;
3946 ptlrpc_req_finished(req);
3948 } else if (KEY_IS(KEY_FIEMAP)) {
3949 struct ptlrpc_request *req;
3950 struct ll_user_fiemap *reply;
3954 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3955 &RQF_OST_GET_INFO_FIEMAP);
3959 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3960 RCL_CLIENT, keylen);
3961 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3962 RCL_CLIENT, *vallen);
3963 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3964 RCL_SERVER, *vallen);
3966 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3968 ptlrpc_request_free(req);
3972 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3973 memcpy(tmp, key, keylen);
3974 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3975 memcpy(tmp, val, *vallen);
3977 ptlrpc_request_set_replen(req);
3978 rc = ptlrpc_queue_wait(req);
3982 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3984 GOTO(out1, rc = -EPROTO);
3986 memcpy(val, reply, *vallen);
3988 ptlrpc_req_finished(req);
3996 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3998 struct llog_ctxt *ctxt;
4002 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
4004 rc = llog_initiator_connect(ctxt);
4005 llog_ctxt_put(ctxt);
4007 /* XXX return an error? skip setting below flags? */
4010 cfs_spin_lock(&imp->imp_lock);
4011 imp->imp_server_timeout = 1;
4012 imp->imp_pingable = 1;
4013 cfs_spin_unlock(&imp->imp_lock);
4014 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
4019 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
4020 struct ptlrpc_request *req,
4027 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
4030 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
4031 void *key, obd_count vallen, void *val,
4032 struct ptlrpc_request_set *set)
4034 struct ptlrpc_request *req;
4035 struct obd_device *obd = exp->exp_obd;
4036 struct obd_import *imp = class_exp2cliimp(exp);
4041 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
4043 if (KEY_IS(KEY_NEXT_ID)) {
4045 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4047 if (vallen != sizeof(obd_id))
4052 if (vallen != sizeof(obd_id))
4055 /* avoid race between allocate new object and set next id
4056 * from ll_sync thread */
4057 cfs_spin_lock(&oscc->oscc_lock);
4058 new_val = *((obd_id*)val) + 1;
4059 if (new_val > oscc->oscc_next_id)
4060 oscc->oscc_next_id = new_val;
4061 cfs_spin_unlock(&oscc->oscc_lock);
4062 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
4063 exp->exp_obd->obd_name,
4064 obd->u.cli.cl_oscc.oscc_next_id);
4069 if (KEY_IS(KEY_CHECKSUM)) {
4070 if (vallen != sizeof(int))
4072 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
4076 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4077 sptlrpc_conf_client_adapt(obd);
4081 if (KEY_IS(KEY_FLUSH_CTX)) {
4082 sptlrpc_import_flush_my_ctx(imp);
4086 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4089 /* We pass all other commands directly to OST. Since nobody calls osc
4090 methods directly and everybody is supposed to go through LOV, we
4091 assume lov checked invalid values for us.
4092 The only recognised values so far are evict_by_nid and mds_conn.
4093 Even if something bad goes through, we'd get a -EINVAL from OST
4096 if (KEY_IS(KEY_GRANT_SHRINK))
4097 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4099 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4104 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4105 RCL_CLIENT, keylen);
4106 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4107 RCL_CLIENT, vallen);
4108 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4110 ptlrpc_request_free(req);
4114 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4115 memcpy(tmp, key, keylen);
4116 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4117 memcpy(tmp, val, vallen);
4119 if (KEY_IS(KEY_MDS_CONN)) {
4120 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4122 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4123 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4124 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4125 req->rq_no_delay = req->rq_no_resend = 1;
4126 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4127 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4128 struct osc_grant_args *aa;
4131 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4132 aa = ptlrpc_req_async_args(req);
4135 ptlrpc_req_finished(req);
4138 *oa = ((struct ost_body *)val)->oa;
4140 req->rq_interpret_reply = osc_shrink_grant_interpret;
4143 ptlrpc_request_set_replen(req);
4144 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4145 LASSERT(set != NULL);
4146 ptlrpc_set_add_req(set, req);
4147 ptlrpc_check_set(NULL, set);
4149 ptlrpcd_add_req(req, PSCOPE_OTHER);
4155 static struct llog_operations osc_size_repl_logops = {
4156 lop_cancel: llog_obd_repl_cancel
4159 static struct llog_operations osc_mds_ost_orig_logops;
4161 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4162 struct obd_device *tgt, struct llog_catid *catid)
4167 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4168 &catid->lci_logid, &osc_mds_ost_orig_logops);
4170 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4174 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4175 NULL, &osc_size_repl_logops);
4177 struct llog_ctxt *ctxt =
4178 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4181 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4186 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4187 obd->obd_name, tgt->obd_name, catid, rc);
4188 CERROR("logid "LPX64":0x%x\n",
4189 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4194 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4195 struct obd_device *disk_obd, int *index)
4197 struct llog_catid catid;
4198 static char name[32] = CATLIST;
4202 LASSERT(olg == &obd->obd_olg);
4204 cfs_mutex_down(&olg->olg_cat_processing);
4205 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4207 CERROR("rc: %d\n", rc);
4211 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4212 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4213 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4215 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4217 CERROR("rc: %d\n", rc);
4221 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4223 CERROR("rc: %d\n", rc);
4228 cfs_mutex_up(&olg->olg_cat_processing);
4233 static int osc_llog_finish(struct obd_device *obd, int count)
4235 struct llog_ctxt *ctxt;
4236 int rc = 0, rc2 = 0;
4239 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4241 rc = llog_cleanup(ctxt);
4243 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4245 rc2 = llog_cleanup(ctxt);
4252 static int osc_reconnect(const struct lu_env *env,
4253 struct obd_export *exp, struct obd_device *obd,
4254 struct obd_uuid *cluuid,
4255 struct obd_connect_data *data,
4258 struct client_obd *cli = &obd->u.cli;
4260 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4263 client_obd_list_lock(&cli->cl_loi_list_lock);
4264 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4265 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4266 lost_grant = cli->cl_lost_grant;
4267 cli->cl_lost_grant = 0;
4268 client_obd_list_unlock(&cli->cl_loi_list_lock);
4270 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4271 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4272 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4273 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4274 " ocd_grant: %d\n", data->ocd_connect_flags,
4275 data->ocd_version, data->ocd_grant);
4281 static int osc_disconnect(struct obd_export *exp)
4283 struct obd_device *obd = class_exp2obd(exp);
4284 struct llog_ctxt *ctxt;
4287 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4289 if (obd->u.cli.cl_conn_count == 1) {
4290 /* Flush any remaining cancel messages out to the
4292 llog_sync(ctxt, exp);
4294 llog_ctxt_put(ctxt);
4296 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4300 rc = client_disconnect_export(exp);
4302 * Initially we put del_shrink_grant before disconnect_export, but it
4303 * causes the following problem if setup (connect) and cleanup
4304 * (disconnect) are tangled together.
4305 * connect p1 disconnect p2
4306 * ptlrpc_connect_import
4307 * ............... class_manual_cleanup
4310 * ptlrpc_connect_interrupt
4312 * add this client to shrink list
4314 * Bang! pinger trigger the shrink.
4315 * So the osc should be disconnected from the shrink list, after we
4316 * are sure the import has been destroyed. BUG18662
4318 if (obd->u.cli.cl_import == NULL)
4319 osc_del_shrink_grant(&obd->u.cli);
4323 static int osc_import_event(struct obd_device *obd,
4324 struct obd_import *imp,
4325 enum obd_import_event event)
4327 struct client_obd *cli;
4331 LASSERT(imp->imp_obd == obd);
4334 case IMP_EVENT_DISCON: {
4335 /* Only do this on the MDS OSC's */
4336 if (imp->imp_server_timeout) {
4337 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4339 cfs_spin_lock(&oscc->oscc_lock);
4340 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4341 cfs_spin_unlock(&oscc->oscc_lock);
4344 client_obd_list_lock(&cli->cl_loi_list_lock);
4345 cli->cl_avail_grant = 0;
4346 cli->cl_lost_grant = 0;
4347 client_obd_list_unlock(&cli->cl_loi_list_lock);
4350 case IMP_EVENT_INACTIVE: {
4351 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4354 case IMP_EVENT_INVALIDATE: {
4355 struct ldlm_namespace *ns = obd->obd_namespace;
4359 env = cl_env_get(&refcheck);
4363 client_obd_list_lock(&cli->cl_loi_list_lock);
4364 /* all pages go to failing rpcs due to the invalid
4366 osc_check_rpcs(env, cli);
4367 client_obd_list_unlock(&cli->cl_loi_list_lock);
4369 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4370 cl_env_put(env, &refcheck);
4375 case IMP_EVENT_ACTIVE: {
4376 /* Only do this on the MDS OSC's */
4377 if (imp->imp_server_timeout) {
4378 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4380 cfs_spin_lock(&oscc->oscc_lock);
4381 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4382 cfs_spin_unlock(&oscc->oscc_lock);
4384 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4387 case IMP_EVENT_OCD: {
4388 struct obd_connect_data *ocd = &imp->imp_connect_data;
4390 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4391 osc_init_grant(&obd->u.cli, ocd);
4394 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4395 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4397 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4400 case IMP_EVENT_DEACTIVATE: {
4401 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
4404 case IMP_EVENT_ACTIVATE: {
4405 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
4409 CERROR("Unknown import event %d\n", event);
4416 * Determine whether the lock can be canceled before replaying the lock
4417 * during recovery, see bug16774 for detailed information.
4419 * \retval zero the lock can't be canceled
4420 * \retval other ok to cancel
4422 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
4424 check_res_locked(lock->l_resource);
4427 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
4429 * XXX as a future improvement, we can also cancel unused write lock
4430 * if it doesn't have dirty data and active mmaps.
4432 if (lock->l_resource->lr_type == LDLM_EXTENT &&
4433 (lock->l_granted_mode == LCK_PR ||
4434 lock->l_granted_mode == LCK_CR) &&
4435 (osc_dlm_lock_pageref(lock) == 0))
4441 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4447 rc = ptlrpcd_addref();
4451 rc = client_obd_setup(obd, lcfg);
4455 struct lprocfs_static_vars lvars = { 0 };
4456 struct client_obd *cli = &obd->u.cli;
4458 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4459 lprocfs_osc_init_vars(&lvars);
4460 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4461 lproc_osc_attach_seqstat(obd);
4462 sptlrpc_lprocfs_cliobd_attach(obd);
4463 ptlrpc_lprocfs_register_obd(obd);
4467 /* We need to allocate a few requests more, because
4468 brw_interpret tries to create new requests before freeing
4469 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4470 reserved, but I afraid that might be too much wasted RAM
4471 in fact, so 2 is just my guess and still should work. */
4472 cli->cl_import->imp_rq_pool =
4473 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4475 ptlrpc_add_rqs_to_pool);
4477 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4478 cfs_sema_init(&cli->cl_grant_sem, 1);
4480 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
4486 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4492 case OBD_CLEANUP_EARLY: {
4493 struct obd_import *imp;
4494 imp = obd->u.cli.cl_import;
4495 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4496 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4497 ptlrpc_deactivate_import(imp);
4498 cfs_spin_lock(&imp->imp_lock);
4499 imp->imp_pingable = 0;
4500 cfs_spin_unlock(&imp->imp_lock);
4503 case OBD_CLEANUP_EXPORTS: {
4505 * for echo client, export may be on zombie list, wait for
4506 * zombie thread to cull it, because cli.cl_import will be
4507 * cleared in client_disconnect_export():
4508 * class_export_destroy() -> obd_cleanup() ->
4509 * echo_device_free() -> echo_client_cleanup() ->
4510 * obd_disconnect() -> osc_disconnect() ->
4511 * client_disconnect_export()
4513 obd_zombie_barrier();
4514 /* If we set up but never connected, the
4515 client import will not have been cleaned. */
4516 if (obd->u.cli.cl_import) {
4517 struct obd_import *imp;
4518 cfs_down_write(&obd->u.cli.cl_sem);
4519 imp = obd->u.cli.cl_import;
4520 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4522 ptlrpc_invalidate_import(imp);
4523 if (imp->imp_rq_pool) {
4524 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4525 imp->imp_rq_pool = NULL;
4527 class_destroy_import(imp);
4528 cfs_up_write(&obd->u.cli.cl_sem);
4529 obd->u.cli.cl_import = NULL;
4531 rc = obd_llog_finish(obd, 0);
4533 CERROR("failed to cleanup llogging subsystems\n");
4540 int osc_cleanup(struct obd_device *obd)
4545 ptlrpc_lprocfs_unregister_obd(obd);
4546 lprocfs_obd_cleanup(obd);
4548 /* free memory of osc quota cache */
4549 lquota_cleanup(quota_interface, obd);
4551 rc = client_obd_cleanup(obd);
4557 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4559 struct lprocfs_static_vars lvars = { 0 };
4562 lprocfs_osc_init_vars(&lvars);
4564 switch (lcfg->lcfg_command) {
4566 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4576 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4578 return osc_process_config_base(obd, buf);
4581 struct obd_ops osc_obd_ops = {
4582 .o_owner = THIS_MODULE,
4583 .o_setup = osc_setup,
4584 .o_precleanup = osc_precleanup,
4585 .o_cleanup = osc_cleanup,
4586 .o_add_conn = client_import_add_conn,
4587 .o_del_conn = client_import_del_conn,
4588 .o_connect = client_connect_import,
4589 .o_reconnect = osc_reconnect,
4590 .o_disconnect = osc_disconnect,
4591 .o_statfs = osc_statfs,
4592 .o_statfs_async = osc_statfs_async,
4593 .o_packmd = osc_packmd,
4594 .o_unpackmd = osc_unpackmd,
4595 .o_precreate = osc_precreate,
4596 .o_create = osc_create,
4597 .o_create_async = osc_create_async,
4598 .o_destroy = osc_destroy,
4599 .o_getattr = osc_getattr,
4600 .o_getattr_async = osc_getattr_async,
4601 .o_setattr = osc_setattr,
4602 .o_setattr_async = osc_setattr_async,
4604 .o_punch = osc_punch,
4606 .o_enqueue = osc_enqueue,
4607 .o_change_cbdata = osc_change_cbdata,
4608 .o_find_cbdata = osc_find_cbdata,
4609 .o_cancel = osc_cancel,
4610 .o_cancel_unused = osc_cancel_unused,
4611 .o_iocontrol = osc_iocontrol,
4612 .o_get_info = osc_get_info,
4613 .o_set_info_async = osc_set_info_async,
4614 .o_import_event = osc_import_event,
4615 .o_llog_init = osc_llog_init,
4616 .o_llog_finish = osc_llog_finish,
4617 .o_process_config = osc_process_config,
4620 extern struct lu_kmem_descr osc_caches[];
4621 extern cfs_spinlock_t osc_ast_guard;
4622 extern cfs_lock_class_key_t osc_ast_guard_class;
4624 int __init osc_init(void)
4626 struct lprocfs_static_vars lvars = { 0 };
4630 /* print an address of _any_ initialized kernel symbol from this
4631 * module, to allow debugging with gdb that doesn't support data
4632 * symbols from modules.*/
4633 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4635 rc = lu_kmem_init(osc_caches);
4637 lprocfs_osc_init_vars(&lvars);
4639 cfs_request_module("lquota");
4640 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4641 lquota_init(quota_interface);
4642 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4644 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4645 LUSTRE_OSC_NAME, &osc_device_type);
4647 if (quota_interface)
4648 PORTAL_SYMBOL_PUT(osc_quota_interface);
4649 lu_kmem_fini(osc_caches);
4653 cfs_spin_lock_init(&osc_ast_guard);
4654 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4656 osc_mds_ost_orig_logops = llog_lvfs_ops;
4657 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4658 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4659 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4660 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4666 static void /*__exit*/ osc_exit(void)
4668 lu_device_type_fini(&osc_device_type);
4670 lquota_exit(quota_interface);
4671 if (quota_interface)
4672 PORTAL_SYMBOL_PUT(osc_quota_interface);
4674 class_unregister_type(LUSTRE_OSC_NAME);
4675 lu_kmem_fini(osc_caches);
4678 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4679 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4680 MODULE_LICENSE("GPL");
4682 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);