1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_OSC
42 #include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include <lustre_cache.h>
65 #include "osc_internal.h"
67 static quota_interface_t *quota_interface = NULL;
68 extern quota_interface_t osc_quota_interface;
70 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
71 static int brw_interpret(struct ptlrpc_request *request, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76 struct lov_stripe_md *lsm)
81 lmm_size = sizeof(**lmmp);
86 OBD_FREE(*lmmp, lmm_size);
92 OBD_ALLOC(*lmmp, lmm_size);
98 LASSERT(lsm->lsm_object_id);
99 LASSERT(lsm->lsm_object_gr);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109 struct lov_mds_md *lmm, int lmm_bytes)
115 if (lmm_bytes < sizeof (*lmm)) {
116 CERROR("lov_mds_md too small: %d, need %d\n",
117 lmm_bytes, (int)sizeof(*lmm));
120 /* XXX LOV_MAGIC etc check? */
122 if (lmm->lmm_object_id == 0) {
123 CERROR("lov_mds_md: zero lmm_object_id\n");
128 lsm_size = lov_stripe_md_size(1);
132 if (*lsmp != NULL && lmm == NULL) {
133 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134 OBD_FREE(*lsmp, lsm_size);
140 OBD_ALLOC(*lsmp, lsm_size);
143 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145 OBD_FREE(*lsmp, lsm_size);
148 loi_init((*lsmp)->lsm_oinfo[0]);
152 /* XXX zero *lsmp? */
153 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT((*lsmp)->lsm_object_gr);
159 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165 struct ost_body *body, void *capa)
167 struct obd_capa *oc = (struct obd_capa *)capa;
168 struct lustre_capa *c;
173 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
176 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177 DEBUG_CAPA(D_SEC, c, "pack");
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181 struct obd_info *oinfo)
183 struct ost_body *body;
185 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
188 body->oa = *oinfo->oi_oa;
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(struct ptlrpc_request *req,
204 struct osc_async_args *aa, int rc)
206 struct ost_body *body;
212 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
213 lustre_swab_ost_body);
215 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
218 /* This should really be sent by the OST */
219 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222 CDEBUG(D_INFO, "can't unpack ost_body\n");
224 aa->aa_oi->oi_oa->o_valid = 0;
227 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232 struct ptlrpc_request_set *set)
234 struct ptlrpc_request *req;
235 struct osc_async_args *aa;
239 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
243 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246 ptlrpc_request_free(req);
250 osc_pack_req_body(req, oinfo);
252 ptlrpc_request_set_replen(req);
253 req->rq_interpret_reply = osc_getattr_interpret;
255 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256 aa = ptlrpc_req_async_args(req);
259 ptlrpc_set_add_req(set, req);
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
265 struct ptlrpc_request *req;
266 struct ost_body *body;
270 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
274 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
277 ptlrpc_request_free(req);
281 osc_pack_req_body(req, oinfo);
283 ptlrpc_request_set_replen(req);
285 rc = ptlrpc_queue_wait(req);
289 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
291 GOTO(out, rc = -EPROTO);
293 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294 *oinfo->oi_oa = body->oa;
296 /* This should really be sent by the OST */
297 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
302 ptlrpc_req_finished(req);
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307 struct obd_trans_info *oti)
309 struct ptlrpc_request *req;
310 struct ost_body *body;
314 LASSERT(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
315 oinfo->oi_oa->o_gr > 0);
317 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
321 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
322 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
324 ptlrpc_request_free(req);
328 osc_pack_req_body(req, oinfo);
330 ptlrpc_request_set_replen(req);
332 rc = ptlrpc_queue_wait(req);
336 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
338 GOTO(out, rc = -EPROTO);
340 *oinfo->oi_oa = body->oa;
344 ptlrpc_req_finished(req);
348 static int osc_setattr_interpret(struct ptlrpc_request *req,
349 struct osc_async_args *aa, int rc)
351 struct ost_body *body;
357 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
359 GOTO(out, rc = -EPROTO);
361 *aa->aa_oi->oi_oa = body->oa;
363 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
367 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
368 struct obd_trans_info *oti,
369 struct ptlrpc_request_set *rqset)
371 struct ptlrpc_request *req;
372 struct osc_async_args *aa;
376 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
380 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
381 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
383 ptlrpc_request_free(req);
387 osc_pack_req_body(req, oinfo);
389 ptlrpc_request_set_replen(req);
391 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
393 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
396 /* do mds to ost setattr asynchronouly */
398 /* Do not wait for response. */
399 ptlrpcd_add_req(req);
401 req->rq_interpret_reply = osc_setattr_interpret;
403 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
404 aa = ptlrpc_req_async_args(req);
407 ptlrpc_set_add_req(rqset, req);
413 int osc_real_create(struct obd_export *exp, struct obdo *oa,
414 struct lov_stripe_md **ea, struct obd_trans_info *oti)
416 struct ptlrpc_request *req;
417 struct ost_body *body;
418 struct lov_stripe_md *lsm;
427 rc = obd_alloc_memmd(exp, &lsm);
432 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
434 GOTO(out, rc = -ENOMEM);
436 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
438 ptlrpc_request_free(req);
442 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
446 ptlrpc_request_set_replen(req);
448 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
449 oa->o_flags == OBD_FL_DELORPHAN) {
451 "delorphan from OST integration");
452 /* Don't resend the delorphan req */
453 req->rq_no_resend = req->rq_no_delay = 1;
456 rc = ptlrpc_queue_wait(req);
460 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
462 GOTO(out_req, rc = -EPROTO);
466 /* This should really be sent by the OST */
467 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
468 oa->o_valid |= OBD_MD_FLBLKSZ;
470 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
471 * have valid lsm_oinfo data structs, so don't go touching that.
472 * This needs to be fixed in a big way.
474 lsm->lsm_object_id = oa->o_id;
475 lsm->lsm_object_gr = oa->o_gr;
479 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
481 if (oa->o_valid & OBD_MD_FLCOOKIE) {
482 if (!oti->oti_logcookies)
483 oti_alloc_cookies(oti, 1);
484 *oti->oti_logcookies = oa->o_lcookie;
488 CDEBUG(D_HA, "transno: "LPD64"\n",
489 lustre_msg_get_transno(req->rq_repmsg));
491 ptlrpc_req_finished(req);
494 obd_free_memmd(exp, &lsm);
498 static int osc_punch_interpret(struct ptlrpc_request *req,
499 struct osc_async_args *aa, int rc)
501 struct ost_body *body;
507 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
509 GOTO(out, rc = -EPROTO);
511 *aa->aa_oi->oi_oa = body->oa;
513 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
517 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
518 struct obd_trans_info *oti,
519 struct ptlrpc_request_set *rqset)
521 struct ptlrpc_request *req;
522 struct osc_async_args *aa;
523 struct ost_body *body;
528 CDEBUG(D_INFO, "oa NULL\n");
532 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
536 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
537 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
539 ptlrpc_request_free(req);
542 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
543 ptlrpc_at_set_req_timeout(req);
544 osc_pack_req_body(req, oinfo);
546 /* overload the size and blocks fields in the oa with start/end */
547 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
549 body->oa.o_size = oinfo->oi_policy.l_extent.start;
550 body->oa.o_blocks = oinfo->oi_policy.l_extent.end;
551 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
552 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = osc_punch_interpret;
556 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
557 aa = ptlrpc_req_async_args(req);
559 ptlrpc_set_add_req(rqset, req);
564 static int osc_sync(struct obd_export *exp, struct obdo *oa,
565 struct lov_stripe_md *md, obd_size start, obd_size end,
568 struct ptlrpc_request *req;
569 struct ost_body *body;
574 CDEBUG(D_INFO, "oa NULL\n");
578 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
582 osc_set_capa_size(req, &RMF_CAPA1, capa);
583 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
585 ptlrpc_request_free(req);
589 /* overload the size and blocks fields in the oa with start/end */
590 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
593 body->oa.o_size = start;
594 body->oa.o_blocks = end;
595 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
596 osc_pack_capa(req, body, capa);
598 ptlrpc_request_set_replen(req);
600 rc = ptlrpc_queue_wait(req);
604 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
606 GOTO(out, rc = -EPROTO);
612 ptlrpc_req_finished(req);
616 /* Find and cancel locally locks matched by @mode in the resource found by
617 * @objid. Found locks are added into @cancel list. Returns the amount of
618 * locks added to @cancels list. */
619 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
620 struct list_head *cancels, ldlm_mode_t mode,
623 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
624 struct ldlm_res_id res_id;
625 struct ldlm_resource *res;
629 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
630 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
634 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
635 lock_flags, 0, NULL);
636 ldlm_resource_putref(res);
640 static int osc_destroy_interpret(struct ptlrpc_request *req, void *data,
643 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
645 atomic_dec(&cli->cl_destroy_in_flight);
646 cfs_waitq_signal(&cli->cl_destroy_waitq);
650 static int osc_can_send_destroy(struct client_obd *cli)
652 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
653 cli->cl_max_rpcs_in_flight) {
654 /* The destroy request can be sent */
657 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
658 cli->cl_max_rpcs_in_flight) {
660 * The counter has been modified between the two atomic
663 cfs_waitq_signal(&cli->cl_destroy_waitq);
668 /* Destroy requests can be async always on the client, and we don't even really
669 * care about the return code since the client cannot do anything at all about
671 * When the MDS is unlinking a filename, it saves the file objects into a
672 * recovery llog, and these object records are cancelled when the OST reports
673 * they were destroyed and sync'd to disk (i.e. transaction committed).
674 * If the client dies, or the OST is down when the object should be destroyed,
675 * the records are not cancelled, and when the OST reconnects to the MDS next,
676 * it will retrieve the llog unlink logs and then sends the log cancellation
677 * cookies to the MDS after committing destroy transactions. */
678 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
679 struct lov_stripe_md *ea, struct obd_trans_info *oti,
680 struct obd_export *md_export)
682 struct client_obd *cli = &exp->exp_obd->u.cli;
683 struct ptlrpc_request *req;
684 struct ost_body *body;
685 CFS_LIST_HEAD(cancels);
690 CDEBUG(D_INFO, "oa NULL\n");
694 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
695 LDLM_FL_DISCARD_DATA);
697 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
699 ldlm_lock_list_put(&cancels, l_bl_ast, count);
703 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
706 ptlrpc_request_free(req);
710 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
711 req->rq_interpret_reply = osc_destroy_interpret;
712 ptlrpc_at_set_req_timeout(req);
714 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
715 oa->o_lcookie = *oti->oti_logcookies;
716 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
720 ptlrpc_request_set_replen(req);
722 if (!osc_can_send_destroy(cli)) {
723 struct l_wait_info lwi = { 0 };
726 * Wait until the number of on-going destroy RPCs drops
727 * under max_rpc_in_flight
729 l_wait_event_exclusive(cli->cl_destroy_waitq,
730 osc_can_send_destroy(cli), &lwi);
733 /* Do not wait for response */
734 ptlrpcd_add_req(req);
738 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
741 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
743 LASSERT(!(oa->o_valid & bits));
746 client_obd_list_lock(&cli->cl_loi_list_lock);
747 oa->o_dirty = cli->cl_dirty;
748 if (cli->cl_dirty > cli->cl_dirty_max) {
749 CERROR("dirty %lu > dirty_max %lu\n",
750 cli->cl_dirty, cli->cl_dirty_max);
752 } else if (atomic_read(&obd_dirty_pages) > obd_max_dirty_pages) {
753 CERROR("dirty %d > system dirty_max %d\n",
754 atomic_read(&obd_dirty_pages), obd_max_dirty_pages);
756 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
757 CERROR("dirty %lu - dirty_max %lu too big???\n",
758 cli->cl_dirty, cli->cl_dirty_max);
761 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
762 (cli->cl_max_rpcs_in_flight + 1);
763 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
765 oa->o_grant = cli->cl_avail_grant;
766 oa->o_dropped = cli->cl_lost_grant;
767 cli->cl_lost_grant = 0;
768 client_obd_list_unlock(&cli->cl_loi_list_lock);
769 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
770 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
773 /* caller must hold loi_list_lock */
774 static void osc_consume_write_grant(struct client_obd *cli,
775 struct brw_page *pga)
777 atomic_inc(&obd_dirty_pages);
778 cli->cl_dirty += CFS_PAGE_SIZE;
779 cli->cl_avail_grant -= CFS_PAGE_SIZE;
780 pga->flag |= OBD_BRW_FROM_GRANT;
781 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
782 CFS_PAGE_SIZE, pga, pga->pg);
783 LASSERT(cli->cl_avail_grant >= 0);
786 /* the companion to osc_consume_write_grant, called when a brw has completed.
787 * must be called with the loi lock held. */
788 static void osc_release_write_grant(struct client_obd *cli,
789 struct brw_page *pga, int sent)
791 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
794 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
799 pga->flag &= ~OBD_BRW_FROM_GRANT;
800 atomic_dec(&obd_dirty_pages);
801 cli->cl_dirty -= CFS_PAGE_SIZE;
803 cli->cl_lost_grant += CFS_PAGE_SIZE;
804 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
805 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
806 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
807 /* For short writes we shouldn't count parts of pages that
808 * span a whole block on the OST side, or our accounting goes
809 * wrong. Should match the code in filter_grant_check. */
810 int offset = pga->off & ~CFS_PAGE_MASK;
811 int count = pga->count + (offset & (blocksize - 1));
812 int end = (offset + pga->count) & (blocksize - 1);
814 count += blocksize - end;
816 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
817 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
818 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
819 cli->cl_avail_grant, cli->cl_dirty);
825 static unsigned long rpcs_in_flight(struct client_obd *cli)
827 return cli->cl_r_in_flight + cli->cl_w_in_flight;
830 /* caller must hold loi_list_lock */
831 void osc_wake_cache_waiters(struct client_obd *cli)
833 struct list_head *l, *tmp;
834 struct osc_cache_waiter *ocw;
837 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
838 /* if we can't dirty more, we must wait until some is written */
839 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
840 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
841 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
842 "osc max %ld, sys max %d\n", cli->cl_dirty,
843 cli->cl_dirty_max, obd_max_dirty_pages);
847 /* if still dirty cache but no grant wait for pending RPCs that
848 * may yet return us some grant before doing sync writes */
849 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
850 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
851 cli->cl_w_in_flight);
855 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
856 list_del_init(&ocw->ocw_entry);
857 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
858 /* no more RPCs in flight to return grant, do sync IO */
859 ocw->ocw_rc = -EDQUOT;
860 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
862 osc_consume_write_grant(cli,
863 &ocw->ocw_oap->oap_brw_page);
866 cfs_waitq_signal(&ocw->ocw_waitq);
872 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
874 client_obd_list_lock(&cli->cl_loi_list_lock);
875 cli->cl_avail_grant = ocd->ocd_grant;
876 client_obd_list_unlock(&cli->cl_loi_list_lock);
878 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
879 cli->cl_avail_grant, cli->cl_lost_grant);
880 LASSERT(cli->cl_avail_grant >= 0);
883 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
885 client_obd_list_lock(&cli->cl_loi_list_lock);
886 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
887 if (body->oa.o_valid & OBD_MD_FLGRANT)
888 cli->cl_avail_grant += body->oa.o_grant;
889 /* waiters are woken in brw_interpret */
890 client_obd_list_unlock(&cli->cl_loi_list_lock);
893 /* We assume that the reason this OSC got a short read is because it read
894 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
895 * via the LOV, and it _knows_ it's reading inside the file, it's just that
896 * this stripe never got written at or beyond this stripe offset yet. */
897 static void handle_short_read(int nob_read, obd_count page_count,
898 struct brw_page **pga)
903 /* skip bytes read OK */
904 while (nob_read > 0) {
905 LASSERT (page_count > 0);
907 if (pga[i]->count > nob_read) {
908 /* EOF inside this page */
909 ptr = cfs_kmap(pga[i]->pg) +
910 (pga[i]->off & ~CFS_PAGE_MASK);
911 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
912 cfs_kunmap(pga[i]->pg);
918 nob_read -= pga[i]->count;
923 /* zero remaining pages */
924 while (page_count-- > 0) {
925 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
926 memset(ptr, 0, pga[i]->count);
927 cfs_kunmap(pga[i]->pg);
932 static int check_write_rcs(struct ptlrpc_request *req,
933 int requested_nob, int niocount,
934 obd_count page_count, struct brw_page **pga)
938 /* return error if any niobuf was in error */
939 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
940 sizeof(*remote_rcs) * niocount, NULL);
941 if (remote_rcs == NULL) {
942 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
945 if (lustre_msg_swabbed(req->rq_repmsg))
946 for (i = 0; i < niocount; i++)
947 __swab32s(&remote_rcs[i]);
949 for (i = 0; i < niocount; i++) {
950 if (remote_rcs[i] < 0)
951 return(remote_rcs[i]);
953 if (remote_rcs[i] != 0) {
954 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
955 i, remote_rcs[i], req);
960 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
961 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
962 requested_nob, req->rq_bulk->bd_nob_transferred);
969 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
971 if (p1->flag != p2->flag) {
972 unsigned mask = ~OBD_BRW_FROM_GRANT;
974 /* warn if we try to combine flags that we don't know to be
976 if ((p1->flag & mask) != (p2->flag & mask))
977 CERROR("is it ok to have flags 0x%x and 0x%x in the "
978 "same brw?\n", p1->flag, p2->flag);
982 return (p1->off + p1->count == p2->off);
985 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
986 struct brw_page **pga, int opc,
987 cksum_type_t cksum_type)
992 LASSERT (pg_count > 0);
993 cksum = init_checksum(cksum_type);
994 while (nob > 0 && pg_count > 0) {
995 unsigned char *ptr = cfs_kmap(pga[i]->pg);
996 int off = pga[i]->off & ~CFS_PAGE_MASK;
997 int count = pga[i]->count > nob ? nob : pga[i]->count;
999 /* corrupt the data before we compute the checksum, to
1000 * simulate an OST->client data error */
1001 if (i == 0 && opc == OST_READ &&
1002 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1003 memcpy(ptr + off, "bad1", min(4, nob));
1004 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1005 cfs_kunmap(pga[i]->pg);
1006 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1009 nob -= pga[i]->count;
1013 /* For sending we only compute the wrong checksum instead
1014 * of corrupting the data so it is still correct on a redo */
1015 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1021 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1022 struct lov_stripe_md *lsm, obd_count page_count,
1023 struct brw_page **pga,
1024 struct ptlrpc_request **reqp,
1025 struct obd_capa *ocapa)
1027 struct ptlrpc_request *req;
1028 struct ptlrpc_bulk_desc *desc;
1029 struct ost_body *body;
1030 struct obd_ioobj *ioobj;
1031 struct niobuf_remote *niobuf;
1032 int niocount, i, requested_nob, opc, rc;
1033 struct osc_brw_async_args *aa;
1034 struct req_capsule *pill;
1035 struct brw_page *pg_prev;
1038 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1039 RETURN(-ENOMEM); /* Recoverable */
1040 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1041 RETURN(-EINVAL); /* Fatal */
1043 if ((cmd & OBD_BRW_WRITE) != 0) {
1045 req = ptlrpc_request_alloc_pool(cli->cl_import,
1046 cli->cl_import->imp_rq_pool,
1050 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1056 for (niocount = i = 1; i < page_count; i++) {
1057 if (!can_merge_pages(pga[i - 1], pga[i]))
1061 pill = &req->rq_pill;
1062 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1063 niocount * sizeof(*niobuf));
1064 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1066 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1068 ptlrpc_request_free(req);
1071 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1072 ptlrpc_at_set_req_timeout(req);
1074 if (opc == OST_WRITE)
1075 desc = ptlrpc_prep_bulk_imp(req, page_count,
1076 BULK_GET_SOURCE, OST_BULK_PORTAL);
1078 desc = ptlrpc_prep_bulk_imp(req, page_count,
1079 BULK_PUT_SINK, OST_BULK_PORTAL);
1082 GOTO(out, rc = -ENOMEM);
1083 /* NB request now owns desc and will free it when it gets freed */
1085 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1086 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1087 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1088 LASSERT(body && ioobj && niobuf);
1092 obdo_to_ioobj(oa, ioobj);
1093 ioobj->ioo_bufcnt = niocount;
1094 osc_pack_capa(req, body, ocapa);
1095 LASSERT (page_count > 0);
1097 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1098 struct brw_page *pg = pga[i];
1100 LASSERT(pg->count > 0);
1101 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1102 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1103 pg->off, pg->count);
1105 LASSERTF(i == 0 || pg->off > pg_prev->off,
1106 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1107 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1109 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1110 pg_prev->pg, page_private(pg_prev->pg),
1111 pg_prev->pg->index, pg_prev->off);
1113 LASSERTF(i == 0 || pg->off > pg_prev->off,
1114 "i %d p_c %u\n", i, page_count);
1116 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1117 (pg->flag & OBD_BRW_SRVLOCK));
1119 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1121 requested_nob += pg->count;
1123 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1125 niobuf->len += pg->count;
1127 niobuf->offset = pg->off;
1128 niobuf->len = pg->count;
1129 niobuf->flags = pg->flag;
1134 LASSERTF((void *)(niobuf - niocount) ==
1135 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1136 niocount * sizeof(*niobuf)),
1137 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1138 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1139 (void *)(niobuf - niocount));
1141 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1143 /* size[REQ_REC_OFF] still sizeof (*body) */
1144 if (opc == OST_WRITE) {
1145 if (unlikely(cli->cl_checksum) &&
1146 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1147 /* store cl_cksum_type in a local variable since
1148 * it can be changed via lprocfs */
1149 cksum_type_t cksum_type = cli->cl_cksum_type;
1151 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1152 oa->o_flags = body->oa.o_flags = 0;
1153 body->oa.o_flags |= cksum_type_pack(cksum_type);
1154 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1155 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1159 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1161 /* save this in 'oa', too, for later checking */
1162 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1163 oa->o_flags |= cksum_type_pack(cksum_type);
1165 /* clear out the checksum flag, in case this is a
1166 * resend but cl_checksum is no longer set. b=11238 */
1167 oa->o_valid &= ~OBD_MD_FLCKSUM;
1169 oa->o_cksum = body->oa.o_cksum;
1170 /* 1 RC per niobuf */
1171 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1172 sizeof(__u32) * niocount);
1174 if (unlikely(cli->cl_checksum) &&
1175 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1176 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1177 body->oa.o_flags = 0;
1178 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1179 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1181 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1182 /* 1 RC for the whole I/O */
1184 ptlrpc_request_set_replen(req);
1186 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1187 aa = ptlrpc_req_async_args(req);
1189 aa->aa_requested_nob = requested_nob;
1190 aa->aa_nio_count = niocount;
1191 aa->aa_page_count = page_count;
1195 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1201 ptlrpc_req_finished(req);
1205 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1206 __u32 client_cksum, __u32 server_cksum, int nob,
1207 obd_count page_count, struct brw_page **pga,
1208 cksum_type_t client_cksum_type)
1212 cksum_type_t cksum_type;
1214 if (server_cksum == client_cksum) {
1215 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1219 if (oa->o_valid & OBD_MD_FLFLAGS)
1220 cksum_type = cksum_type_unpack(oa->o_flags);
1222 cksum_type = OBD_CKSUM_CRC32;
1224 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1227 if (cksum_type != client_cksum_type)
1228 msg = "the server did not use the checksum type specified in "
1229 "the original request - likely a protocol problem";
1230 else if (new_cksum == server_cksum)
1231 msg = "changed on the client after we checksummed it - "
1232 "likely false positive due to mmap IO (bug 11742)";
1233 else if (new_cksum == client_cksum)
1234 msg = "changed in transit before arrival at OST";
1236 msg = "changed in transit AND doesn't match the original - "
1237 "likely false positive due to mmap IO (bug 11742)";
1239 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1240 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1241 "["LPU64"-"LPU64"]\n",
1242 msg, libcfs_nid2str(peer->nid),
1243 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1244 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1247 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1249 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1250 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1251 "client csum now %x\n", client_cksum, client_cksum_type,
1252 server_cksum, cksum_type, new_cksum);
1256 /* Note rc enters this function as number of bytes transferred */
1257 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1259 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1260 const lnet_process_id_t *peer =
1261 &req->rq_import->imp_connection->c_peer;
1262 struct client_obd *cli = aa->aa_cli;
1263 struct ost_body *body;
1264 __u32 client_cksum = 0;
1267 if (rc < 0 && rc != -EDQUOT)
1270 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1271 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1272 lustre_swab_ost_body);
1274 CDEBUG(D_INFO, "Can't unpack body\n");
1278 /* set/clear over quota flag for a uid/gid */
1279 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1280 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1281 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1282 body->oa.o_gid, body->oa.o_valid,
1288 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1289 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1291 osc_update_grant(cli, body);
1293 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1295 CERROR("Unexpected +ve rc %d\n", rc);
1298 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1300 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1301 check_write_checksum(&body->oa, peer, client_cksum,
1302 body->oa.o_cksum, aa->aa_requested_nob,
1303 aa->aa_page_count, aa->aa_ppga,
1304 cksum_type_unpack(aa->aa_oa->o_flags)))
1307 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1310 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1311 aa->aa_page_count, aa->aa_ppga);
1315 /* The rest of this function executes only for OST_READs */
1316 if (rc > aa->aa_requested_nob) {
1317 CERROR("Unexpected rc %d (%d requested)\n", rc,
1318 aa->aa_requested_nob);
1322 if (rc != req->rq_bulk->bd_nob_transferred) {
1323 CERROR ("Unexpected rc %d (%d transferred)\n",
1324 rc, req->rq_bulk->bd_nob_transferred);
1328 if (rc < aa->aa_requested_nob)
1329 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1331 if (sptlrpc_cli_unwrap_bulk_read(req, rc, aa->aa_page_count,
1333 GOTO(out, rc = -EAGAIN);
1335 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1336 static int cksum_counter;
1337 __u32 server_cksum = body->oa.o_cksum;
1340 cksum_type_t cksum_type;
1342 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1343 cksum_type = cksum_type_unpack(body->oa.o_flags);
1345 cksum_type = OBD_CKSUM_CRC32;
1346 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1347 aa->aa_ppga, OST_READ,
1350 if (peer->nid == req->rq_bulk->bd_sender) {
1354 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1357 if (server_cksum == ~0 && rc > 0) {
1358 CERROR("Protocol error: server %s set the 'checksum' "
1359 "bit, but didn't send a checksum. Not fatal, "
1360 "but please notify on http://bugzilla.lustre.org/\n",
1361 libcfs_nid2str(peer->nid));
1362 } else if (server_cksum != client_cksum) {
1363 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1364 "%s%s%s inum "LPU64"/"LPU64" object "
1365 LPU64"/"LPU64" extent "
1366 "["LPU64"-"LPU64"]\n",
1367 req->rq_import->imp_obd->obd_name,
1368 libcfs_nid2str(peer->nid),
1370 body->oa.o_valid & OBD_MD_FLFID ?
1371 body->oa.o_fid : (__u64)0,
1372 body->oa.o_valid & OBD_MD_FLFID ?
1373 body->oa.o_generation :(__u64)0,
1375 body->oa.o_valid & OBD_MD_FLGROUP ?
1376 body->oa.o_gr : (__u64)0,
1377 aa->aa_ppga[0]->off,
1378 aa->aa_ppga[aa->aa_page_count-1]->off +
1379 aa->aa_ppga[aa->aa_page_count-1]->count -
1381 CERROR("client %x, server %x, cksum_type %x\n",
1382 client_cksum, server_cksum, cksum_type);
1384 aa->aa_oa->o_cksum = client_cksum;
1388 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1391 } else if (unlikely(client_cksum)) {
1392 static int cksum_missed;
1395 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1396 CERROR("Checksum %u requested from %s but not sent\n",
1397 cksum_missed, libcfs_nid2str(peer->nid));
1403 *aa->aa_oa = body->oa;
1408 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1409 struct lov_stripe_md *lsm,
1410 obd_count page_count, struct brw_page **pga,
1411 struct obd_capa *ocapa)
1413 struct ptlrpc_request *req;
1417 struct l_wait_info lwi;
1421 cfs_waitq_init(&waitq);
1424 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1425 page_count, pga, &req, ocapa);
1429 rc = ptlrpc_queue_wait(req);
1431 if (rc == -ETIMEDOUT && req->rq_resend) {
1432 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1433 ptlrpc_req_finished(req);
1437 rc = osc_brw_fini_request(req, rc);
1439 ptlrpc_req_finished(req);
1440 if (osc_recoverable_error(rc)) {
1442 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1443 CERROR("too many resend retries, returning error\n");
1447 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1448 l_wait_event(waitq, 0, &lwi);
1456 int osc_brw_redo_request(struct ptlrpc_request *request,
1457 struct osc_brw_async_args *aa)
1459 struct ptlrpc_request *new_req;
1460 struct ptlrpc_request_set *set = request->rq_set;
1461 struct osc_brw_async_args *new_aa;
1462 struct osc_async_page *oap;
1466 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1467 CERROR("too many resend retries, returning error\n");
1471 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1473 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
1474 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1475 ocapa = lustre_unpack_capa(request->rq_reqmsg,
1478 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1479 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1480 aa->aa_cli, aa->aa_oa,
1481 NULL /* lsm unused by osc currently */,
1482 aa->aa_page_count, aa->aa_ppga,
1483 &new_req, NULL /* ocapa */);
1487 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1489 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1490 if (oap->oap_request != NULL) {
1491 LASSERTF(request == oap->oap_request,
1492 "request %p != oap_request %p\n",
1493 request, oap->oap_request);
1494 if (oap->oap_interrupted) {
1495 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1496 ptlrpc_req_finished(new_req);
1501 /* New request takes over pga and oaps from old request.
1502 * Note that copying a list_head doesn't work, need to move it... */
1504 new_req->rq_interpret_reply = request->rq_interpret_reply;
1505 new_req->rq_async_args = request->rq_async_args;
1506 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1508 new_aa = ptlrpc_req_async_args(new_req);
1510 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1511 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1512 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1514 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1515 if (oap->oap_request) {
1516 ptlrpc_req_finished(oap->oap_request);
1517 oap->oap_request = ptlrpc_request_addref(new_req);
1521 /* use ptlrpc_set_add_req is safe because interpret functions work
1522 * in check_set context. only one way exist with access to request
1523 * from different thread got -EINTR - this way protected with
1524 * cl_loi_list_lock */
1525 ptlrpc_set_add_req(set, new_req);
1527 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1529 DEBUG_REQ(D_INFO, new_req, "new request");
1533 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1534 struct lov_stripe_md *lsm, obd_count page_count,
1535 struct brw_page **pga, struct ptlrpc_request_set *set,
1536 struct obd_capa *ocapa)
1538 struct ptlrpc_request *req;
1539 struct client_obd *cli = &exp->exp_obd->u.cli;
1541 struct osc_brw_async_args *aa;
1544 /* Consume write credits even if doing a sync write -
1545 * otherwise we may run out of space on OST due to grant. */
1546 if (cmd == OBD_BRW_WRITE) {
1547 spin_lock(&cli->cl_loi_list_lock);
1548 for (i = 0; i < page_count; i++) {
1549 if (cli->cl_avail_grant >= CFS_PAGE_SIZE)
1550 osc_consume_write_grant(cli, pga[i]);
1552 spin_unlock(&cli->cl_loi_list_lock);
1555 rc = osc_brw_prep_request(cmd, cli, oa, lsm, page_count, pga,
1558 aa = ptlrpc_req_async_args(req);
1559 if (cmd == OBD_BRW_READ) {
1560 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1561 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1563 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1564 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1565 cli->cl_w_in_flight);
1567 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
1569 LASSERT(list_empty(&aa->aa_oaps));
1571 req->rq_interpret_reply = brw_interpret;
1572 ptlrpc_set_add_req(set, req);
1573 client_obd_list_lock(&cli->cl_loi_list_lock);
1574 if (cmd == OBD_BRW_READ)
1575 cli->cl_r_in_flight++;
1577 cli->cl_w_in_flight++;
1578 client_obd_list_unlock(&cli->cl_loi_list_lock);
1579 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DIO_PAUSE, 3);
1580 } else if (cmd == OBD_BRW_WRITE) {
1581 client_obd_list_lock(&cli->cl_loi_list_lock);
1582 for (i = 0; i < page_count; i++)
1583 osc_release_write_grant(cli, pga[i], 0);
1584 osc_wake_cache_waiters(cli);
1585 client_obd_list_unlock(&cli->cl_loi_list_lock);
1591 * ugh, we want disk allocation on the target to happen in offset order. we'll
1592 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1593 * fine for our small page arrays and doesn't require allocation. its an
1594 * insertion sort that swaps elements that are strides apart, shrinking the
1595 * stride down until its '1' and the array is sorted.
1597 static void sort_brw_pages(struct brw_page **array, int num)
1600 struct brw_page *tmp;
1604 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1609 for (i = stride ; i < num ; i++) {
1612 while (j >= stride && array[j - stride]->off > tmp->off) {
1613 array[j] = array[j - stride];
1618 } while (stride > 1);
1621 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1627 LASSERT (pages > 0);
1628 offset = pg[i]->off & ~CFS_PAGE_MASK;
1632 if (pages == 0) /* that's all */
1635 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1636 return count; /* doesn't end on page boundary */
1639 offset = pg[i]->off & ~CFS_PAGE_MASK;
1640 if (offset != 0) /* doesn't start on page boundary */
1647 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1649 struct brw_page **ppga;
1652 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1656 for (i = 0; i < count; i++)
1661 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1663 LASSERT(ppga != NULL);
1664 OBD_FREE(ppga, sizeof(*ppga) * count);
1667 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1668 obd_count page_count, struct brw_page *pga,
1669 struct obd_trans_info *oti)
1671 struct obdo *saved_oa = NULL;
1672 struct brw_page **ppga, **orig;
1673 struct obd_import *imp = class_exp2cliimp(exp);
1674 struct client_obd *cli = &imp->imp_obd->u.cli;
1675 int rc, page_count_orig;
1678 if (cmd & OBD_BRW_CHECK) {
1679 /* The caller just wants to know if there's a chance that this
1680 * I/O can succeed */
1682 if (imp == NULL || imp->imp_invalid)
1687 /* test_brw with a failed create can trip this, maybe others. */
1688 LASSERT(cli->cl_max_pages_per_rpc);
1692 orig = ppga = osc_build_ppga(pga, page_count);
1695 page_count_orig = page_count;
1697 sort_brw_pages(ppga, page_count);
1698 while (page_count) {
1699 obd_count pages_per_brw;
1701 if (page_count > cli->cl_max_pages_per_rpc)
1702 pages_per_brw = cli->cl_max_pages_per_rpc;
1704 pages_per_brw = page_count;
1706 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1708 if (saved_oa != NULL) {
1709 /* restore previously saved oa */
1710 *oinfo->oi_oa = *saved_oa;
1711 } else if (page_count > pages_per_brw) {
1712 /* save a copy of oa (brw will clobber it) */
1713 OBDO_ALLOC(saved_oa);
1714 if (saved_oa == NULL)
1715 GOTO(out, rc = -ENOMEM);
1716 *saved_oa = *oinfo->oi_oa;
1719 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1720 pages_per_brw, ppga, oinfo->oi_capa);
1725 page_count -= pages_per_brw;
1726 ppga += pages_per_brw;
1730 osc_release_ppga(orig, page_count_orig);
1732 if (saved_oa != NULL)
1733 OBDO_FREE(saved_oa);
1738 static int osc_brw_async(int cmd, struct obd_export *exp,
1739 struct obd_info *oinfo, obd_count page_count,
1740 struct brw_page *pga, struct obd_trans_info *oti,
1741 struct ptlrpc_request_set *set)
1743 struct brw_page **ppga, **orig;
1744 struct client_obd *cli = &exp->exp_obd->u.cli;
1745 int page_count_orig;
1749 if (cmd & OBD_BRW_CHECK) {
1750 struct obd_import *imp = class_exp2cliimp(exp);
1751 /* The caller just wants to know if there's a chance that this
1752 * I/O can succeed */
1754 if (imp == NULL || imp->imp_invalid)
1759 orig = ppga = osc_build_ppga(pga, page_count);
1762 page_count_orig = page_count;
1764 sort_brw_pages(ppga, page_count);
1765 while (page_count) {
1766 struct brw_page **copy;
1767 obd_count pages_per_brw;
1769 pages_per_brw = min_t(obd_count, page_count,
1770 cli->cl_max_pages_per_rpc);
1772 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1774 /* use ppga only if single RPC is going to fly */
1775 if (pages_per_brw != page_count_orig || ppga != orig) {
1776 OBD_ALLOC(copy, sizeof(*copy) * pages_per_brw);
1778 GOTO(out, rc = -ENOMEM);
1779 memcpy(copy, ppga, sizeof(*copy) * pages_per_brw);
1783 rc = async_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1784 pages_per_brw, copy, set, oinfo->oi_capa);
1788 OBD_FREE(copy, sizeof(*copy) * pages_per_brw);
1792 /* we passed it to async_internal() which is
1793 * now responsible for releasing memory */
1797 page_count -= pages_per_brw;
1798 ppga += pages_per_brw;
1802 osc_release_ppga(orig, page_count_orig);
1806 static void osc_check_rpcs(struct client_obd *cli);
1808 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1809 * the dirty accounting. Writeback completes or truncate happens before
1810 * writing starts. Must be called with the loi lock held. */
1811 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1814 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1818 /* This maintains the lists of pending pages to read/write for a given object
1819 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1820 * to quickly find objects that are ready to send an RPC. */
1821 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1827 if (lop->lop_num_pending == 0)
1830 /* if we have an invalid import we want to drain the queued pages
1831 * by forcing them through rpcs that immediately fail and complete
1832 * the pages. recovery relies on this to empty the queued pages
1833 * before canceling the locks and evicting down the llite pages */
1834 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1837 /* stream rpcs in queue order as long as as there is an urgent page
1838 * queued. this is our cheap solution for good batching in the case
1839 * where writepage marks some random page in the middle of the file
1840 * as urgent because of, say, memory pressure */
1841 if (!list_empty(&lop->lop_urgent)) {
1842 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1845 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1846 optimal = cli->cl_max_pages_per_rpc;
1847 if (cmd & OBD_BRW_WRITE) {
1848 /* trigger a write rpc stream as long as there are dirtiers
1849 * waiting for space. as they're waiting, they're not going to
1850 * create more pages to coallesce with what's waiting.. */
1851 if (!list_empty(&cli->cl_cache_waiters)) {
1852 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1855 /* +16 to avoid triggering rpcs that would want to include pages
1856 * that are being queued but which can't be made ready until
1857 * the queuer finishes with the page. this is a wart for
1858 * llite::commit_write() */
1861 if (lop->lop_num_pending >= optimal)
1867 static void on_list(struct list_head *item, struct list_head *list,
1870 if (list_empty(item) && should_be_on)
1871 list_add_tail(item, list);
1872 else if (!list_empty(item) && !should_be_on)
1873 list_del_init(item);
1876 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1877 * can find pages to build into rpcs quickly */
1878 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1880 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1881 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1882 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1884 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1885 loi->loi_write_lop.lop_num_pending);
1887 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1888 loi->loi_read_lop.lop_num_pending);
1891 static void lop_update_pending(struct client_obd *cli,
1892 struct loi_oap_pages *lop, int cmd, int delta)
1894 lop->lop_num_pending += delta;
1895 if (cmd & OBD_BRW_WRITE)
1896 cli->cl_pending_w_pages += delta;
1898 cli->cl_pending_r_pages += delta;
1901 /* this is called when a sync waiter receives an interruption. Its job is to
1902 * get the caller woken as soon as possible. If its page hasn't been put in an
1903 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1904 * desiring interruption which will forcefully complete the rpc once the rpc
1906 static void osc_occ_interrupted(struct oig_callback_context *occ)
1908 struct osc_async_page *oap;
1909 struct loi_oap_pages *lop;
1910 struct lov_oinfo *loi;
1913 /* XXX member_of() */
1914 oap = list_entry(occ, struct osc_async_page, oap_occ);
1916 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
1918 oap->oap_interrupted = 1;
1920 /* ok, it's been put in an rpc. only one oap gets a request reference */
1921 if (oap->oap_request != NULL) {
1922 ptlrpc_mark_interrupted(oap->oap_request);
1923 ptlrpcd_wake(oap->oap_request);
1927 /* we don't get interruption callbacks until osc_trigger_group_io()
1928 * has been called and put the sync oaps in the pending/urgent lists.*/
1929 if (!list_empty(&oap->oap_pending_item)) {
1930 list_del_init(&oap->oap_pending_item);
1931 list_del_init(&oap->oap_urgent_item);
1934 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1935 &loi->loi_write_lop : &loi->loi_read_lop;
1936 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1937 loi_list_maint(oap->oap_cli, oap->oap_loi);
1939 oig_complete_one(oap->oap_oig, &oap->oap_occ, -EINTR);
1940 oap->oap_oig = NULL;
1944 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
1947 /* this is trying to propogate async writeback errors back up to the
1948 * application. As an async write fails we record the error code for later if
1949 * the app does an fsync. As long as errors persist we force future rpcs to be
1950 * sync so that the app can get a sync error and break the cycle of queueing
1951 * pages for which writeback will fail. */
1952 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1959 ar->ar_force_sync = 1;
1960 ar->ar_min_xid = ptlrpc_sample_next_xid();
1965 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1966 ar->ar_force_sync = 0;
1969 static void osc_oap_to_pending(struct osc_async_page *oap)
1971 struct loi_oap_pages *lop;
1973 if (oap->oap_cmd & OBD_BRW_WRITE)
1974 lop = &oap->oap_loi->loi_write_lop;
1976 lop = &oap->oap_loi->loi_read_lop;
1978 if (oap->oap_async_flags & ASYNC_URGENT)
1979 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1980 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1981 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1984 /* this must be called holding the loi list lock to give coverage to exit_cache,
1985 * async_flag maintenance, and oap_request */
1986 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1987 struct osc_async_page *oap, int sent, int rc)
1992 if (oap->oap_request != NULL) {
1993 xid = ptlrpc_req_xid(oap->oap_request);
1994 ptlrpc_req_finished(oap->oap_request);
1995 oap->oap_request = NULL;
1998 oap->oap_async_flags = 0;
1999 oap->oap_interrupted = 0;
2001 if (oap->oap_cmd & OBD_BRW_WRITE) {
2002 osc_process_ar(&cli->cl_ar, xid, rc);
2003 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2006 if (rc == 0 && oa != NULL) {
2007 if (oa->o_valid & OBD_MD_FLBLOCKS)
2008 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2009 if (oa->o_valid & OBD_MD_FLMTIME)
2010 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2011 if (oa->o_valid & OBD_MD_FLATIME)
2012 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2013 if (oa->o_valid & OBD_MD_FLCTIME)
2014 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2018 osc_exit_cache(cli, oap, sent);
2019 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
2020 oap->oap_oig = NULL;
2025 rc = oap->oap_caller_ops->ap_completion(oap->oap_caller_data,
2026 oap->oap_cmd, oa, rc);
2028 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2029 * I/O on the page could start, but OSC calls it under lock
2030 * and thus we can add oap back to pending safely */
2032 /* upper layer wants to leave the page on pending queue */
2033 osc_oap_to_pending(oap);
2035 osc_exit_cache(cli, oap, sent);
2039 static int brw_interpret(struct ptlrpc_request *req, void *data, int rc)
2041 struct osc_brw_async_args *aa = data;
2042 struct client_obd *cli;
2045 rc = osc_brw_fini_request(req, rc);
2046 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2047 if (osc_recoverable_error(rc)) {
2048 rc = osc_brw_redo_request(req, aa);
2055 client_obd_list_lock(&cli->cl_loi_list_lock);
2057 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2058 * is called so we know whether to go to sync BRWs or wait for more
2059 * RPCs to complete */
2060 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2061 cli->cl_w_in_flight--;
2063 cli->cl_r_in_flight--;
2065 if (!list_empty(&aa->aa_oaps)) { /* from osc_send_oap_rpc() */
2066 struct osc_async_page *oap, *tmp;
2067 /* the caller may re-use the oap after the completion call so
2068 * we need to clean it up a little */
2069 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2070 list_del_init(&oap->oap_rpc_item);
2071 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
2073 OBDO_FREE(aa->aa_oa);
2074 } else { /* from async_internal() */
2076 for (i = 0; i < aa->aa_page_count; i++)
2077 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2079 osc_wake_cache_waiters(cli);
2080 osc_check_rpcs(cli);
2081 client_obd_list_unlock(&cli->cl_loi_list_lock);
2083 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2087 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
2088 struct list_head *rpc_list,
2089 int page_count, int cmd)
2091 struct ptlrpc_request *req;
2092 struct brw_page **pga = NULL;
2093 struct osc_brw_async_args *aa;
2094 struct obdo *oa = NULL;
2095 struct obd_async_page_ops *ops = NULL;
2096 void *caller_data = NULL;
2097 struct obd_capa *ocapa;
2098 struct osc_async_page *oap;
2099 struct ldlm_lock *lock = NULL;
2103 LASSERT(!list_empty(rpc_list));
2105 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2107 RETURN(ERR_PTR(-ENOMEM));
2111 GOTO(out, req = ERR_PTR(-ENOMEM));
2114 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2116 ops = oap->oap_caller_ops;
2117 caller_data = oap->oap_caller_data;
2118 lock = oap->oap_ldlm_lock;
2120 pga[i] = &oap->oap_brw_page;
2121 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2122 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2123 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2127 /* always get the data for the obdo for the rpc */
2128 LASSERT(ops != NULL);
2129 ops->ap_fill_obdo(caller_data, cmd, oa);
2130 ocapa = ops->ap_lookup_capa(caller_data, cmd);
2132 oa->o_handle = lock->l_remote_handle;
2133 oa->o_valid |= OBD_MD_FLHANDLE;
2136 sort_brw_pages(pga, page_count);
2137 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2141 CERROR("prep_req failed: %d\n", rc);
2142 GOTO(out, req = ERR_PTR(rc));
2145 /* Need to update the timestamps after the request is built in case
2146 * we race with setattr (locally or in queue at OST). If OST gets
2147 * later setattr before earlier BRW (as determined by the request xid),
2148 * the OST will not use BRW timestamps. Sadly, there is no obvious
2149 * way to do this in a single call. bug 10150 */
2150 ops->ap_update_obdo(caller_data, cmd, oa,
2151 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME);
2153 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2154 aa = ptlrpc_req_async_args(req);
2155 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2156 list_splice(rpc_list, &aa->aa_oaps);
2157 CFS_INIT_LIST_HEAD(rpc_list);
2164 OBD_FREE(pga, sizeof(*pga) * page_count);
2169 /* the loi lock is held across this function but it's allowed to release
2170 * and reacquire it during its work */
2172 * prepare pages for ASYNC io and put pages in send queue.
2176 * \param cmd - OBD_BRW_* macroses
2177 * \param lop - pending pages
2179 * \return zero if pages successfully add to send queue.
2180 * \return not zere if error occurring.
2182 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
2183 int cmd, struct loi_oap_pages *lop)
2185 struct ptlrpc_request *req;
2186 obd_count page_count = 0;
2187 struct osc_async_page *oap = NULL, *tmp;
2188 struct osc_brw_async_args *aa;
2189 struct obd_async_page_ops *ops;
2190 CFS_LIST_HEAD(rpc_list);
2191 unsigned int ending_offset;
2192 unsigned starting_offset = 0;
2196 /* first we find the pages we're allowed to work with */
2197 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2199 ops = oap->oap_caller_ops;
2201 LASSERT(oap->oap_magic == OAP_MAGIC);
2203 if (page_count != 0 &&
2204 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2205 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2206 " oap %p, page %p, srvlock %u\n",
2207 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2210 /* in llite being 'ready' equates to the page being locked
2211 * until completion unlocks it. commit_write submits a page
2212 * as not ready because its unlock will happen unconditionally
2213 * as the call returns. if we race with commit_write giving
2214 * us that page we dont' want to create a hole in the page
2215 * stream, so we stop and leave the rpc to be fired by
2216 * another dirtier or kupdated interval (the not ready page
2217 * will still be on the dirty list). we could call in
2218 * at the end of ll_file_write to process the queue again. */
2219 if (!(oap->oap_async_flags & ASYNC_READY)) {
2220 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
2222 CDEBUG(D_INODE, "oap %p page %p returned %d "
2223 "instead of ready\n", oap,
2227 /* llite is telling us that the page is still
2228 * in commit_write and that we should try
2229 * and put it in an rpc again later. we
2230 * break out of the loop so we don't create
2231 * a hole in the sequence of pages in the rpc
2236 /* the io isn't needed.. tell the checks
2237 * below to complete the rpc with EINTR */
2238 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2239 oap->oap_count = -EINTR;
2242 oap->oap_async_flags |= ASYNC_READY;
2245 LASSERTF(0, "oap %p page %p returned %d "
2246 "from make_ready\n", oap,
2254 * Page submitted for IO has to be locked. Either by
2255 * ->ap_make_ready() or by higher layers.
2257 #if defined(__KERNEL__) && defined(__linux__)
2258 if(!(PageLocked(oap->oap_page) &&
2259 (CheckWriteback(oap->oap_page, cmd) || oap->oap_oig !=NULL))) {
2260 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2261 oap->oap_page, (long)oap->oap_page->flags, oap->oap_async_flags);
2265 /* If there is a gap at the start of this page, it can't merge
2266 * with any previous page, so we'll hand the network a
2267 * "fragmented" page array that it can't transfer in 1 RDMA */
2268 if (page_count != 0 && oap->oap_page_off != 0)
2271 /* take the page out of our book-keeping */
2272 list_del_init(&oap->oap_pending_item);
2273 lop_update_pending(cli, lop, cmd, -1);
2274 list_del_init(&oap->oap_urgent_item);
2276 if (page_count == 0)
2277 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2278 (PTLRPC_MAX_BRW_SIZE - 1);
2280 /* ask the caller for the size of the io as the rpc leaves. */
2281 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
2283 ops->ap_refresh_count(oap->oap_caller_data,cmd);
2284 if (oap->oap_count <= 0) {
2285 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2287 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
2291 /* now put the page back in our accounting */
2292 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2293 if (page_count == 0)
2294 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2295 if (++page_count >= cli->cl_max_pages_per_rpc)
2298 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2299 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2300 * have the same alignment as the initial writes that allocated
2301 * extents on the server. */
2302 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2303 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2304 if (ending_offset == 0)
2307 /* If there is a gap at the end of this page, it can't merge
2308 * with any subsequent pages, so we'll hand the network a
2309 * "fragmented" page array that it can't transfer in 1 RDMA */
2310 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2314 osc_wake_cache_waiters(cli);
2316 if (page_count == 0)
2319 loi_list_maint(cli, loi);
2321 client_obd_list_unlock(&cli->cl_loi_list_lock);
2323 req = osc_build_req(cli, &rpc_list, page_count, cmd);
2325 /* this should happen rarely and is pretty bad, it makes the
2326 * pending list not follow the dirty order */
2327 client_obd_list_lock(&cli->cl_loi_list_lock);
2328 list_for_each_entry_safe(oap, tmp, &rpc_list, oap_rpc_item) {
2329 list_del_init(&oap->oap_rpc_item);
2331 /* queued sync pages can be torn down while the pages
2332 * were between the pending list and the rpc */
2333 if (oap->oap_interrupted) {
2334 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2335 osc_ap_completion(cli, NULL, oap, 0,
2339 osc_ap_completion(cli, NULL, oap, 0, PTR_ERR(req));
2341 loi_list_maint(cli, loi);
2342 RETURN(PTR_ERR(req));
2345 aa = ptlrpc_req_async_args(req);
2347 if (cmd == OBD_BRW_READ) {
2348 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2349 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2350 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2351 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2353 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2354 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2355 cli->cl_w_in_flight);
2356 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2357 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2359 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2361 client_obd_list_lock(&cli->cl_loi_list_lock);
2363 if (cmd == OBD_BRW_READ)
2364 cli->cl_r_in_flight++;
2366 cli->cl_w_in_flight++;
2368 /* queued sync pages can be torn down while the pages
2369 * were between the pending list and the rpc */
2371 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2372 /* only one oap gets a request reference */
2375 if (oap->oap_interrupted && !req->rq_intr) {
2376 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2378 ptlrpc_mark_interrupted(req);
2382 tmp->oap_request = ptlrpc_request_addref(req);
2384 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2385 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2387 req->rq_interpret_reply = brw_interpret;
2388 ptlrpcd_add_req(req);
2392 #define LOI_DEBUG(LOI, STR, args...) \
2393 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2394 !list_empty(&(LOI)->loi_cli_item), \
2395 (LOI)->loi_write_lop.lop_num_pending, \
2396 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2397 (LOI)->loi_read_lop.lop_num_pending, \
2398 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2401 /* This is called by osc_check_rpcs() to find which objects have pages that
2402 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2403 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2406 /* first return all objects which we already know to have
2407 * pages ready to be stuffed into rpcs */
2408 if (!list_empty(&cli->cl_loi_ready_list))
2409 RETURN(list_entry(cli->cl_loi_ready_list.next,
2410 struct lov_oinfo, loi_cli_item));
2412 /* then if we have cache waiters, return all objects with queued
2413 * writes. This is especially important when many small files
2414 * have filled up the cache and not been fired into rpcs because
2415 * they don't pass the nr_pending/object threshhold */
2416 if (!list_empty(&cli->cl_cache_waiters) &&
2417 !list_empty(&cli->cl_loi_write_list))
2418 RETURN(list_entry(cli->cl_loi_write_list.next,
2419 struct lov_oinfo, loi_write_item));
2421 /* then return all queued objects when we have an invalid import
2422 * so that they get flushed */
2423 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2424 if (!list_empty(&cli->cl_loi_write_list))
2425 RETURN(list_entry(cli->cl_loi_write_list.next,
2426 struct lov_oinfo, loi_write_item));
2427 if (!list_empty(&cli->cl_loi_read_list))
2428 RETURN(list_entry(cli->cl_loi_read_list.next,
2429 struct lov_oinfo, loi_read_item));
2434 /* called with the loi list lock held */
2435 static void osc_check_rpcs(struct client_obd *cli)
2437 struct lov_oinfo *loi;
2438 int rc = 0, race_counter = 0;
2441 while ((loi = osc_next_loi(cli)) != NULL) {
2442 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2444 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2447 /* attempt some read/write balancing by alternating between
2448 * reads and writes in an object. The makes_rpc checks here
2449 * would be redundant if we were getting read/write work items
2450 * instead of objects. we don't want send_oap_rpc to drain a
2451 * partial read pending queue when we're given this object to
2452 * do io on writes while there are cache waiters */
2453 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2454 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
2455 &loi->loi_write_lop);
2463 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2464 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
2465 &loi->loi_read_lop);
2474 /* attempt some inter-object balancing by issueing rpcs
2475 * for each object in turn */
2476 if (!list_empty(&loi->loi_cli_item))
2477 list_del_init(&loi->loi_cli_item);
2478 if (!list_empty(&loi->loi_write_item))
2479 list_del_init(&loi->loi_write_item);
2480 if (!list_empty(&loi->loi_read_item))
2481 list_del_init(&loi->loi_read_item);
2483 loi_list_maint(cli, loi);
2485 /* send_oap_rpc fails with 0 when make_ready tells it to
2486 * back off. llite's make_ready does this when it tries
2487 * to lock a page queued for write that is already locked.
2488 * we want to try sending rpcs from many objects, but we
2489 * don't want to spin failing with 0. */
2490 if (race_counter == 10)
2496 /* we're trying to queue a page in the osc so we're subject to the
2497 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2498 * If the osc's queued pages are already at that limit, then we want to sleep
2499 * until there is space in the osc's queue for us. We also may be waiting for
2500 * write credits from the OST if there are RPCs in flight that may return some
2501 * before we fall back to sync writes.
2503 * We need this know our allocation was granted in the presence of signals */
2504 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2508 client_obd_list_lock(&cli->cl_loi_list_lock);
2509 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2510 client_obd_list_unlock(&cli->cl_loi_list_lock);
2514 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2515 * grant or cache space. */
2516 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
2517 struct osc_async_page *oap)
2519 struct osc_cache_waiter ocw;
2520 struct l_wait_info lwi = { 0 };
2524 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2525 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2526 cli->cl_dirty_max, obd_max_dirty_pages,
2527 cli->cl_lost_grant, cli->cl_avail_grant);
2529 /* force the caller to try sync io. this can jump the list
2530 * of queued writes and create a discontiguous rpc stream */
2531 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2532 loi->loi_ar.ar_force_sync)
2535 /* Hopefully normal case - cache space and write credits available */
2536 if ((cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max) &&
2537 (atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) &&
2538 (cli->cl_avail_grant >= CFS_PAGE_SIZE)) {
2539 /* account for ourselves */
2540 osc_consume_write_grant(cli, &oap->oap_brw_page);
2544 /* Make sure that there are write rpcs in flight to wait for. This
2545 * is a little silly as this object may not have any pending but
2546 * other objects sure might. */
2547 if (cli->cl_w_in_flight) {
2548 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2549 cfs_waitq_init(&ocw.ocw_waitq);
2553 loi_list_maint(cli, loi);
2554 osc_check_rpcs(cli);
2555 client_obd_list_unlock(&cli->cl_loi_list_lock);
2557 CDEBUG(D_CACHE, "sleeping for cache space\n");
2558 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2560 client_obd_list_lock(&cli->cl_loi_list_lock);
2561 if (!list_empty(&ocw.ocw_entry)) {
2562 list_del(&ocw.ocw_entry);
2572 * Checks if requested extent lock is compatible with a lock under the page.
2574 * Checks if the lock under \a page is compatible with a read or write lock
2575 * (specified by \a rw) for an extent [\a start , \a end].
2577 * \param exp osc export
2578 * \param lsm striping information for the file
2579 * \param res osc_async_page placeholder
2580 * \param rw OBD_BRW_READ if requested for reading,
2581 * OBD_BRW_WRITE if requested for writing
2582 * \param start start of the requested extent
2583 * \param end end of the requested extent
2584 * \param cookie transparent parameter for passing locking context
2586 * \post result == 1, *cookie == context, appropriate lock is referenced or
2589 * \retval 1 owned lock is reused for the request
2590 * \retval 0 no lock reused for the request
2592 * \see osc_release_short_lock
2594 static int osc_reget_short_lock(struct obd_export *exp,
2595 struct lov_stripe_md *lsm,
2597 obd_off start, obd_off end,
2600 struct osc_async_page *oap = *res;
2605 spin_lock(&oap->oap_lock);
2606 rc = ldlm_lock_fast_match(oap->oap_ldlm_lock, rw,
2607 start, end, cookie);
2608 spin_unlock(&oap->oap_lock);
2614 * Releases a reference to a lock taken in a "fast" way.
2616 * Releases a read or a write (specified by \a rw) lock
2617 * referenced by \a cookie.
2619 * \param exp osc export
2620 * \param lsm striping information for the file
2621 * \param end end of the locked extent
2622 * \param rw OBD_BRW_READ if requested for reading,
2623 * OBD_BRW_WRITE if requested for writing
2624 * \param cookie transparent parameter for passing locking context
2626 * \post appropriate lock is dereferenced
2628 * \see osc_reget_short_lock
2630 static int osc_release_short_lock(struct obd_export *exp,
2631 struct lov_stripe_md *lsm, obd_off end,
2632 void *cookie, int rw)
2635 ldlm_lock_fast_release(cookie, rw);
2636 /* no error could have happened at this layer */
2640 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2641 struct lov_oinfo *loi, cfs_page_t *page,
2642 obd_off offset, struct obd_async_page_ops *ops,
2643 void *data, void **res, int nocache,
2644 struct lustre_handle *lockh)
2646 struct osc_async_page *oap;
2647 struct ldlm_res_id oid;
2652 return size_round(sizeof(*oap));
2655 oap->oap_magic = OAP_MAGIC;
2656 oap->oap_cli = &exp->exp_obd->u.cli;
2659 oap->oap_caller_ops = ops;
2660 oap->oap_caller_data = data;
2662 oap->oap_page = page;
2663 oap->oap_obj_off = offset;
2665 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2666 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2667 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2668 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2670 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2672 spin_lock_init(&oap->oap_lock);
2674 /* If the page was marked as notcacheable - don't add to any locks */
2676 osc_build_res_name(loi->loi_id, loi->loi_gr, &oid);
2677 /* This is the only place where we can call cache_add_extent
2678 without oap_lock, because this page is locked now, and
2679 the lock we are adding it to is referenced, so cannot lose
2680 any pages either. */
2681 rc = cache_add_extent(oap->oap_cli->cl_cache, &oid, oap, lockh);
2686 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2690 struct osc_async_page *oap_from_cookie(void *cookie)
2692 struct osc_async_page *oap = cookie;
2693 if (oap->oap_magic != OAP_MAGIC)
2694 return ERR_PTR(-EINVAL);
2698 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2699 struct lov_oinfo *loi, void *cookie,
2700 int cmd, obd_off off, int count,
2701 obd_flag brw_flags, enum async_flags async_flags)
2703 struct client_obd *cli = &exp->exp_obd->u.cli;
2704 struct osc_async_page *oap;
2708 oap = oap_from_cookie(cookie);
2710 RETURN(PTR_ERR(oap));
2712 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2715 if (!list_empty(&oap->oap_pending_item) ||
2716 !list_empty(&oap->oap_urgent_item) ||
2717 !list_empty(&oap->oap_rpc_item))
2720 /* check if the file's owner/group is over quota */
2721 #ifdef HAVE_QUOTA_SUPPORT
2722 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)){
2723 struct obd_async_page_ops *ops;
2730 ops = oap->oap_caller_ops;
2731 ops->ap_fill_obdo(oap->oap_caller_data, cmd, oa);
2732 if (lquota_chkdq(quota_interface, cli, oa->o_uid, oa->o_gid) ==
2743 loi = lsm->lsm_oinfo[0];
2745 client_obd_list_lock(&cli->cl_loi_list_lock);
2748 oap->oap_page_off = off;
2749 oap->oap_count = count;
2750 oap->oap_brw_flags = brw_flags;
2751 oap->oap_async_flags = async_flags;
2753 if (cmd & OBD_BRW_WRITE) {
2754 rc = osc_enter_cache(cli, loi, oap);
2756 client_obd_list_unlock(&cli->cl_loi_list_lock);
2761 osc_oap_to_pending(oap);
2762 loi_list_maint(cli, loi);
2764 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2767 osc_check_rpcs(cli);
2768 client_obd_list_unlock(&cli->cl_loi_list_lock);
2773 /* aka (~was & now & flag), but this is more clear :) */
2774 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2776 static int osc_set_async_flags(struct obd_export *exp,
2777 struct lov_stripe_md *lsm,
2778 struct lov_oinfo *loi, void *cookie,
2779 obd_flag async_flags)
2781 struct client_obd *cli = &exp->exp_obd->u.cli;
2782 struct loi_oap_pages *lop;
2783 struct osc_async_page *oap;
2787 oap = oap_from_cookie(cookie);
2789 RETURN(PTR_ERR(oap));
2792 * bug 7311: OST-side locking is only supported for liblustre for now
2793 * (and liblustre never calls obd_set_async_flags(). I hope.), generic
2794 * implementation has to handle case where OST-locked page was picked
2795 * up by, e.g., ->writepage().
2797 LASSERT(!(oap->oap_brw_flags & OBD_BRW_SRVLOCK));
2798 LASSERT(!LIBLUSTRE_CLIENT); /* check that liblustre angels do fear to
2801 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2805 loi = lsm->lsm_oinfo[0];
2807 if (oap->oap_cmd & OBD_BRW_WRITE) {
2808 lop = &loi->loi_write_lop;
2810 lop = &loi->loi_read_lop;
2813 client_obd_list_lock(&cli->cl_loi_list_lock);
2815 if (list_empty(&oap->oap_pending_item))
2816 GOTO(out, rc = -EINVAL);
2818 if ((oap->oap_async_flags & async_flags) == async_flags)
2821 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2822 oap->oap_async_flags |= ASYNC_READY;
2824 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2825 if (list_empty(&oap->oap_rpc_item)) {
2826 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2827 loi_list_maint(cli, loi);
2831 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2832 oap->oap_async_flags);
2834 osc_check_rpcs(cli);
2835 client_obd_list_unlock(&cli->cl_loi_list_lock);
2839 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2840 struct lov_oinfo *loi,
2841 struct obd_io_group *oig, void *cookie,
2842 int cmd, obd_off off, int count,
2844 obd_flag async_flags)
2846 struct client_obd *cli = &exp->exp_obd->u.cli;
2847 struct osc_async_page *oap;
2848 struct loi_oap_pages *lop;
2852 oap = oap_from_cookie(cookie);
2854 RETURN(PTR_ERR(oap));
2856 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2859 if (!list_empty(&oap->oap_pending_item) ||
2860 !list_empty(&oap->oap_urgent_item) ||
2861 !list_empty(&oap->oap_rpc_item))
2865 loi = lsm->lsm_oinfo[0];
2867 client_obd_list_lock(&cli->cl_loi_list_lock);
2870 oap->oap_page_off = off;
2871 oap->oap_count = count;
2872 oap->oap_brw_flags = brw_flags;
2873 oap->oap_async_flags = async_flags;
2875 if (cmd & OBD_BRW_WRITE)
2876 lop = &loi->loi_write_lop;
2878 lop = &loi->loi_read_lop;
2880 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2881 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2883 rc = oig_add_one(oig, &oap->oap_occ);
2886 LOI_DEBUG(loi, "oap %p page %p on group pending: rc %d\n",
2887 oap, oap->oap_page, rc);
2889 client_obd_list_unlock(&cli->cl_loi_list_lock);
2894 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2895 struct loi_oap_pages *lop, int cmd)
2897 struct list_head *pos, *tmp;
2898 struct osc_async_page *oap;
2900 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2901 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2902 list_del(&oap->oap_pending_item);
2903 osc_oap_to_pending(oap);
2905 loi_list_maint(cli, loi);
2908 static int osc_trigger_group_io(struct obd_export *exp,
2909 struct lov_stripe_md *lsm,
2910 struct lov_oinfo *loi,
2911 struct obd_io_group *oig)
2913 struct client_obd *cli = &exp->exp_obd->u.cli;
2917 loi = lsm->lsm_oinfo[0];
2919 client_obd_list_lock(&cli->cl_loi_list_lock);
2921 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2922 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2924 osc_check_rpcs(cli);
2925 client_obd_list_unlock(&cli->cl_loi_list_lock);
2930 static int osc_teardown_async_page(struct obd_export *exp,
2931 struct lov_stripe_md *lsm,
2932 struct lov_oinfo *loi, void *cookie)
2934 struct client_obd *cli = &exp->exp_obd->u.cli;
2935 struct loi_oap_pages *lop;
2936 struct osc_async_page *oap;
2940 oap = oap_from_cookie(cookie);
2942 RETURN(PTR_ERR(oap));
2945 loi = lsm->lsm_oinfo[0];
2947 if (oap->oap_cmd & OBD_BRW_WRITE) {
2948 lop = &loi->loi_write_lop;
2950 lop = &loi->loi_read_lop;
2953 client_obd_list_lock(&cli->cl_loi_list_lock);
2955 if (!list_empty(&oap->oap_rpc_item))
2956 GOTO(out, rc = -EBUSY);
2958 osc_exit_cache(cli, oap, 0);
2959 osc_wake_cache_waiters(cli);
2961 if (!list_empty(&oap->oap_urgent_item)) {
2962 list_del_init(&oap->oap_urgent_item);
2963 oap->oap_async_flags &= ~ASYNC_URGENT;
2965 if (!list_empty(&oap->oap_pending_item)) {
2966 list_del_init(&oap->oap_pending_item);
2967 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2969 loi_list_maint(cli, loi);
2970 cache_remove_extent(cli->cl_cache, oap);
2972 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2974 client_obd_list_unlock(&cli->cl_loi_list_lock);
2978 int osc_extent_blocking_cb(struct ldlm_lock *lock,
2979 struct ldlm_lock_desc *new, void *data,
2982 struct lustre_handle lockh = { 0 };
2986 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
2987 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
2992 case LDLM_CB_BLOCKING:
2993 ldlm_lock2handle(lock, &lockh);
2994 rc = ldlm_cli_cancel(&lockh);
2996 CERROR("ldlm_cli_cancel failed: %d\n", rc);
2998 case LDLM_CB_CANCELING: {
3000 ldlm_lock2handle(lock, &lockh);
3001 /* This lock wasn't granted, don't try to do anything */
3002 if (lock->l_req_mode != lock->l_granted_mode)
3005 cache_remove_lock(lock->l_conn_export->exp_obd->u.cli.cl_cache,
3008 if (lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb)
3009 lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb(
3010 lock, new, data,flag);
3019 EXPORT_SYMBOL(osc_extent_blocking_cb);
3021 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data,
3024 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3027 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
3030 lock_res_and_lock(lock);
3031 #if defined (__KERNEL__) && defined (__linux__)
3032 /* Liang XXX: Darwin and Winnt checking should be added */
3033 if (lock->l_ast_data && lock->l_ast_data != data) {
3034 struct inode *new_inode = data;
3035 struct inode *old_inode = lock->l_ast_data;
3036 if (!(old_inode->i_state & I_FREEING))
3037 LDLM_ERROR(lock, "inconsistent l_ast_data found");
3038 LASSERTF(old_inode->i_state & I_FREEING,
3039 "Found existing inode %p/%lu/%u state %lu in lock: "
3040 "setting data to %p/%lu/%u\n", old_inode,
3041 old_inode->i_ino, old_inode->i_generation,
3043 new_inode, new_inode->i_ino, new_inode->i_generation);
3046 lock->l_ast_data = data;
3047 lock->l_flags |= (flags & LDLM_FL_NO_LRU);
3048 unlock_res_and_lock(lock);
3049 LDLM_LOCK_PUT(lock);
3052 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3053 ldlm_iterator_t replace, void *data)
3055 struct ldlm_res_id res_id;
3056 struct obd_device *obd = class_exp2obd(exp);
3058 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3059 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3063 static int osc_enqueue_fini(struct obd_device *obd, struct ptlrpc_request *req,
3064 struct obd_info *oinfo, int intent, int rc)
3069 /* The request was created before ldlm_cli_enqueue call. */
3070 if (rc == ELDLM_LOCK_ABORTED) {
3071 struct ldlm_reply *rep;
3072 rep = req_capsule_server_get(&req->rq_pill,
3075 LASSERT(rep != NULL);
3076 if (rep->lock_policy_res1)
3077 rc = rep->lock_policy_res1;
3081 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3082 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3083 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_size,
3084 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_blocks,
3085 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_mtime);
3089 cache_add_lock(obd->u.cli.cl_cache, oinfo->oi_lockh);
3091 /* Call the update callback. */
3092 rc = oinfo->oi_cb_up(oinfo, rc);
3096 static int osc_enqueue_interpret(struct ptlrpc_request *req,
3097 struct osc_enqueue_args *aa, int rc)
3099 int intent = aa->oa_oi->oi_flags & LDLM_FL_HAS_INTENT;
3100 struct lov_stripe_md *lsm = aa->oa_oi->oi_md;
3101 struct ldlm_lock *lock;
3103 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3105 lock = ldlm_handle2lock(aa->oa_oi->oi_lockh);
3107 /* Complete obtaining the lock procedure. */
3108 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3110 &aa->oa_oi->oi_flags,
3111 &lsm->lsm_oinfo[0]->loi_lvb,
3112 sizeof(lsm->lsm_oinfo[0]->loi_lvb),
3113 lustre_swab_ost_lvb,
3114 aa->oa_oi->oi_lockh, rc);
3116 /* Complete osc stuff. */
3117 rc = osc_enqueue_fini(aa->oa_exp->exp_obd, req, aa->oa_oi, intent, rc);
3119 /* Release the lock for async request. */
3120 if (lustre_handle_is_used(aa->oa_oi->oi_lockh) && rc == ELDLM_OK)
3121 ldlm_lock_decref(aa->oa_oi->oi_lockh, aa->oa_ei->ei_mode);
3123 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3124 aa->oa_oi->oi_lockh, req, aa);
3125 LDLM_LOCK_PUT(lock);
3129 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3130 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3131 * other synchronous requests, however keeping some locks and trying to obtain
3132 * others may take a considerable amount of time in a case of ost failure; and
3133 * when other sync requests do not get released lock from a client, the client
3134 * is excluded from the cluster -- such scenarious make the life difficult, so
3135 * release locks just after they are obtained. */
3136 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3137 struct ldlm_enqueue_info *einfo,
3138 struct ptlrpc_request_set *rqset)
3140 struct ldlm_res_id res_id;
3141 struct obd_device *obd = exp->exp_obd;
3142 struct ptlrpc_request *req = NULL;
3143 int intent = oinfo->oi_flags & LDLM_FL_HAS_INTENT;
3149 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3150 oinfo->oi_md->lsm_object_gr, &res_id);
3151 /* Filesystem lock extents are extended to page boundaries so that
3152 * dealing with the page cache is a little smoother. */
3153 oinfo->oi_policy.l_extent.start -=
3154 oinfo->oi_policy.l_extent.start & ~CFS_PAGE_MASK;
3155 oinfo->oi_policy.l_extent.end |= ~CFS_PAGE_MASK;
3157 if (oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid == 0)
3160 /* Next, search for already existing extent locks that will cover us */
3161 /* If we're trying to read, we also search for an existing PW lock. The
3162 * VFS and page cache already protect us locally, so lots of readers/
3163 * writers can share a single PW lock.
3165 * There are problems with conversion deadlocks, so instead of
3166 * converting a read lock to a write lock, we'll just enqueue a new
3169 * At some point we should cancel the read lock instead of making them
3170 * send us a blocking callback, but there are problems with canceling
3171 * locks out from other users right now, too. */
3172 mode = einfo->ei_mode;
3173 if (einfo->ei_mode == LCK_PR)
3175 mode = ldlm_lock_match(obd->obd_namespace,
3176 oinfo->oi_flags | LDLM_FL_LVB_READY, &res_id,
3177 einfo->ei_type, &oinfo->oi_policy, mode,
3180 /* addref the lock only if not async requests and PW lock is
3181 * matched whereas we asked for PR. */
3182 if (!rqset && einfo->ei_mode != mode)
3183 ldlm_lock_addref(oinfo->oi_lockh, LCK_PR);
3184 osc_set_data_with_check(oinfo->oi_lockh, einfo->ei_cbdata,
3187 /* I would like to be able to ASSERT here that rss <=
3188 * kms, but I can't, for reasons which are explained in
3192 /* We already have a lock, and it's referenced */
3193 oinfo->oi_cb_up(oinfo, ELDLM_OK);
3195 /* For async requests, decref the lock. */
3196 if (einfo->ei_mode != mode)
3197 ldlm_lock_decref(oinfo->oi_lockh, LCK_PW);
3199 ldlm_lock_decref(oinfo->oi_lockh, einfo->ei_mode);
3206 CFS_LIST_HEAD(cancels);
3207 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3208 &RQF_LDLM_ENQUEUE_LVB);
3212 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3216 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3217 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb));
3218 ptlrpc_request_set_replen(req);
3221 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3222 oinfo->oi_flags &= ~LDLM_FL_BLOCK_GRANTED;
3224 rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id,
3225 &oinfo->oi_policy, &oinfo->oi_flags,
3226 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3227 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb),
3228 lustre_swab_ost_lvb, oinfo->oi_lockh,
3232 struct osc_enqueue_args *aa;
3233 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3234 aa = ptlrpc_req_async_args(req);
3239 req->rq_interpret_reply = osc_enqueue_interpret;
3240 ptlrpc_set_add_req(rqset, req);
3241 } else if (intent) {
3242 ptlrpc_req_finished(req);
3247 rc = osc_enqueue_fini(obd, req, oinfo, intent, rc);
3249 ptlrpc_req_finished(req);
3254 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
3255 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3256 int *flags, void *data, struct lustre_handle *lockh)
3258 struct ldlm_res_id res_id;
3259 struct obd_device *obd = exp->exp_obd;
3260 int lflags = *flags;
3264 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3266 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3269 /* Filesystem lock extents are extended to page boundaries so that
3270 * dealing with the page cache is a little smoother */
3271 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3272 policy->l_extent.end |= ~CFS_PAGE_MASK;
3274 /* Next, search for already existing extent locks that will cover us */
3275 /* If we're trying to read, we also search for an existing PW lock. The
3276 * VFS and page cache already protect us locally, so lots of readers/
3277 * writers can share a single PW lock. */
3281 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3282 &res_id, type, policy, rc, lockh);
3284 osc_set_data_with_check(lockh, data, lflags);
3285 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3286 ldlm_lock_addref(lockh, LCK_PR);
3287 ldlm_lock_decref(lockh, LCK_PW);
3294 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3295 __u32 mode, struct lustre_handle *lockh)
3299 if (unlikely(mode == LCK_GROUP))
3300 ldlm_lock_decref_and_cancel(lockh, mode);
3302 ldlm_lock_decref(lockh, mode);
3307 static int osc_cancel_unused(struct obd_export *exp,
3308 struct lov_stripe_md *lsm, int flags,
3311 struct obd_device *obd = class_exp2obd(exp);
3312 struct ldlm_res_id res_id, *resp = NULL;
3315 resp = osc_build_res_name(lsm->lsm_object_id,
3316 lsm->lsm_object_gr, &res_id);
3319 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3322 static int osc_join_lru(struct obd_export *exp,
3323 struct lov_stripe_md *lsm, int join)
3325 struct obd_device *obd = class_exp2obd(exp);
3326 struct ldlm_res_id res_id, *resp = NULL;
3329 resp = osc_build_res_name(lsm->lsm_object_id,
3330 lsm->lsm_object_gr, &res_id);
3333 return ldlm_cli_join_lru(obd->obd_namespace, resp, join);
3336 static int osc_statfs_interpret(struct ptlrpc_request *req,
3337 struct osc_async_args *aa, int rc)
3339 struct obd_statfs *msfs;
3345 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3347 GOTO(out, rc = -EPROTO);
3350 *aa->aa_oi->oi_osfs = *msfs;
3352 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3356 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3357 __u64 max_age, struct ptlrpc_request_set *rqset)
3359 struct ptlrpc_request *req;
3360 struct osc_async_args *aa;
3364 /* We could possibly pass max_age in the request (as an absolute
3365 * timestamp or a "seconds.usec ago") so the target can avoid doing
3366 * extra calls into the filesystem if that isn't necessary (e.g.
3367 * during mount that would help a bit). Having relative timestamps
3368 * is not so great if request processing is slow, while absolute
3369 * timestamps are not ideal because they need time synchronization. */
3370 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3374 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3376 ptlrpc_request_free(req);
3379 ptlrpc_request_set_replen(req);
3380 req->rq_request_portal = OST_CREATE_PORTAL;
3381 ptlrpc_at_set_req_timeout(req);
3383 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3384 /* procfs requests not want stat in wait for avoid deadlock */
3385 req->rq_no_resend = 1;
3386 req->rq_no_delay = 1;
3389 req->rq_interpret_reply = osc_statfs_interpret;
3390 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3391 aa = ptlrpc_req_async_args(req);
3394 ptlrpc_set_add_req(rqset, req);
3398 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3399 __u64 max_age, __u32 flags)
3401 struct obd_statfs *msfs;
3402 struct ptlrpc_request *req;
3403 struct obd_import *imp = NULL;
3407 /*Since the request might also come from lprocfs, so we need
3408 *sync this with client_disconnect_export Bug15684*/
3409 down_read(&obd->u.cli.cl_sem);
3410 if (obd->u.cli.cl_import)
3411 imp = class_import_get(obd->u.cli.cl_import);
3412 up_read(&obd->u.cli.cl_sem);
3416 /* We could possibly pass max_age in the request (as an absolute
3417 * timestamp or a "seconds.usec ago") so the target can avoid doing
3418 * extra calls into the filesystem if that isn't necessary (e.g.
3419 * during mount that would help a bit). Having relative timestamps
3420 * is not so great if request processing is slow, while absolute
3421 * timestamps are not ideal because they need time synchronization. */
3422 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3424 class_import_put(imp);
3429 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3431 ptlrpc_request_free(req);
3434 ptlrpc_request_set_replen(req);
3435 req->rq_request_portal = OST_CREATE_PORTAL;
3436 ptlrpc_at_set_req_timeout(req);
3438 if (flags & OBD_STATFS_NODELAY) {
3439 /* procfs requests not want stat in wait for avoid deadlock */
3440 req->rq_no_resend = 1;
3441 req->rq_no_delay = 1;
3444 rc = ptlrpc_queue_wait(req);
3448 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3450 GOTO(out, rc = -EPROTO);
3457 ptlrpc_req_finished(req);
3461 /* Retrieve object striping information.
3463 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3464 * the maximum number of OST indices which will fit in the user buffer.
3465 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3467 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3469 struct lov_user_md lum, *lumk;
3470 int rc = 0, lum_size;
3476 if (copy_from_user(&lum, lump, sizeof(lum)))
3479 if (lum.lmm_magic != LOV_USER_MAGIC)
3482 if (lum.lmm_stripe_count > 0) {
3483 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
3484 OBD_ALLOC(lumk, lum_size);
3488 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
3489 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
3491 lum_size = sizeof(lum);
3495 lumk->lmm_object_id = lsm->lsm_object_id;
3496 lumk->lmm_object_gr = lsm->lsm_object_gr;
3497 lumk->lmm_stripe_count = 1;
3499 if (copy_to_user(lump, lumk, lum_size))
3503 OBD_FREE(lumk, lum_size);
3509 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3510 void *karg, void *uarg)
3512 struct obd_device *obd = exp->exp_obd;
3513 struct obd_ioctl_data *data = karg;
3517 if (!try_module_get(THIS_MODULE)) {
3518 CERROR("Can't get module. Is it alive?");
3522 case OBD_IOC_LOV_GET_CONFIG: {
3524 struct lov_desc *desc;
3525 struct obd_uuid uuid;
3529 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3530 GOTO(out, err = -EINVAL);
3532 data = (struct obd_ioctl_data *)buf;
3534 if (sizeof(*desc) > data->ioc_inllen1) {
3535 obd_ioctl_freedata(buf, len);
3536 GOTO(out, err = -EINVAL);
3539 if (data->ioc_inllen2 < sizeof(uuid)) {
3540 obd_ioctl_freedata(buf, len);
3541 GOTO(out, err = -EINVAL);
3544 desc = (struct lov_desc *)data->ioc_inlbuf1;
3545 desc->ld_tgt_count = 1;
3546 desc->ld_active_tgt_count = 1;
3547 desc->ld_default_stripe_count = 1;
3548 desc->ld_default_stripe_size = 0;
3549 desc->ld_default_stripe_offset = 0;
3550 desc->ld_pattern = 0;
3551 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3553 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3555 err = copy_to_user((void *)uarg, buf, len);
3558 obd_ioctl_freedata(buf, len);
3561 case LL_IOC_LOV_SETSTRIPE:
3562 err = obd_alloc_memmd(exp, karg);
3566 case LL_IOC_LOV_GETSTRIPE:
3567 err = osc_getstripe(karg, uarg);
3569 case OBD_IOC_CLIENT_RECOVER:
3570 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3575 case IOC_OSC_SET_ACTIVE:
3576 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3579 case OBD_IOC_POLL_QUOTACHECK:
3580 err = lquota_poll_check(quota_interface, exp,
3581 (struct if_quotacheck *)karg);
3584 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3585 cmd, cfs_curproc_comm());
3586 GOTO(out, err = -ENOTTY);
3589 module_put(THIS_MODULE);
3593 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3594 void *key, __u32 *vallen, void *val,
3595 struct lov_stripe_md *lsm)
3598 if (!vallen || !val)
3601 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3602 __u32 *stripe = val;
3603 *vallen = sizeof(*stripe);
3606 } else if (KEY_IS(KEY_LAST_ID)) {
3607 struct ptlrpc_request *req;
3612 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3613 &RQF_OST_GET_INFO_LAST_ID);
3617 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3618 RCL_CLIENT, keylen);
3619 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3621 ptlrpc_request_free(req);
3625 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3626 memcpy(tmp, key, keylen);
3628 ptlrpc_request_set_replen(req);
3629 rc = ptlrpc_queue_wait(req);
3633 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3635 GOTO(out, rc = -EPROTO);
3637 *((obd_id *)val) = *reply;
3639 ptlrpc_req_finished(req);
3641 } else if (KEY_IS(KEY_FIEMAP)) {
3642 struct ptlrpc_request *req;
3643 struct ll_user_fiemap *reply;
3647 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3648 &RQF_OST_GET_INFO_FIEMAP);
3652 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3653 RCL_CLIENT, keylen);
3654 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3655 RCL_CLIENT, *vallen);
3656 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3657 RCL_SERVER, *vallen);
3659 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3661 ptlrpc_request_free(req);
3665 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3666 memcpy(tmp, key, keylen);
3667 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3668 memcpy(tmp, val, *vallen);
3670 ptlrpc_request_set_replen(req);
3671 rc = ptlrpc_queue_wait(req);
3675 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3677 GOTO(out1, rc = -EPROTO);
3679 memcpy(val, reply, *vallen);
3681 ptlrpc_req_finished(req);
3689 static int osc_setinfo_mds_conn_interpret(struct ptlrpc_request *req,
3692 struct llog_ctxt *ctxt;
3693 struct obd_import *imp = req->rq_import;
3699 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3702 rc = llog_initiator_connect(ctxt);
3704 CERROR("cannot establish connection for "
3705 "ctxt %p: %d\n", ctxt, rc);
3708 llog_ctxt_put(ctxt);
3709 spin_lock(&imp->imp_lock);
3710 imp->imp_server_timeout = 1;
3711 imp->imp_pingable = 1;
3712 spin_unlock(&imp->imp_lock);
3713 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3718 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3719 void *key, obd_count vallen, void *val,
3720 struct ptlrpc_request_set *set)
3722 struct ptlrpc_request *req;
3723 struct obd_device *obd = exp->exp_obd;
3724 struct obd_import *imp = class_exp2cliimp(exp);
3729 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3731 if (KEY_IS(KEY_NEXT_ID)) {
3732 if (vallen != sizeof(obd_id))
3736 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3737 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3738 exp->exp_obd->obd_name,
3739 obd->u.cli.cl_oscc.oscc_next_id);
3744 if (KEY_IS(KEY_UNLINKED)) {
3745 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3746 spin_lock(&oscc->oscc_lock);
3747 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3748 spin_unlock(&oscc->oscc_lock);
3752 if (KEY_IS(KEY_INIT_RECOV)) {
3753 if (vallen != sizeof(int))
3755 spin_lock(&imp->imp_lock);
3756 imp->imp_initial_recov = *(int *)val;
3757 spin_unlock(&imp->imp_lock);
3758 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3759 exp->exp_obd->obd_name,
3760 imp->imp_initial_recov);
3764 if (KEY_IS(KEY_CHECKSUM)) {
3765 if (vallen != sizeof(int))
3767 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3771 if (KEY_IS(KEY_FLUSH_CTX)) {
3772 sptlrpc_import_flush_my_ctx(imp);
3779 /* We pass all other commands directly to OST. Since nobody calls osc
3780 methods directly and everybody is supposed to go through LOV, we
3781 assume lov checked invalid values for us.
3782 The only recognised values so far are evict_by_nid and mds_conn.
3783 Even if something bad goes through, we'd get a -EINVAL from OST
3787 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3791 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3792 RCL_CLIENT, keylen);
3793 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3794 RCL_CLIENT, vallen);
3795 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3797 ptlrpc_request_free(req);
3801 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3802 memcpy(tmp, key, keylen);
3803 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3804 memcpy(tmp, val, vallen);
3806 if (KEY_IS(KEY_MDS_CONN)) {
3807 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3809 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3810 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3811 LASSERT(oscc->oscc_oa.o_gr > 0);
3812 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3815 ptlrpc_request_set_replen(req);
3816 ptlrpc_set_add_req(set, req);
3817 ptlrpc_check_set(set);
3823 static struct llog_operations osc_size_repl_logops = {
3824 lop_cancel: llog_obd_repl_cancel
3827 static struct llog_operations osc_mds_ost_orig_logops;
3828 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3829 struct obd_device *tgt, int count,
3830 struct llog_catid *catid, struct obd_uuid *uuid)
3835 LASSERT(olg == &obd->obd_olg);
3836 spin_lock(&obd->obd_dev_lock);
3837 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3838 osc_mds_ost_orig_logops = llog_lvfs_ops;
3839 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3840 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3841 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3842 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3844 spin_unlock(&obd->obd_dev_lock);
3846 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3847 &catid->lci_logid, &osc_mds_ost_orig_logops);
3849 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3853 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3854 NULL, &osc_size_repl_logops);
3856 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3859 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3860 obd->obd_name, tgt->obd_name, count, catid, rc);
3861 CERROR("logid "LPX64":0x%x\n",
3862 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3867 static int osc_llog_finish(struct obd_device *obd, int count)
3869 struct llog_ctxt *ctxt;
3870 int rc = 0, rc2 = 0;
3873 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3875 rc = llog_cleanup(ctxt);
3877 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3879 rc2 = llog_cleanup(ctxt);
3886 static int osc_reconnect(const struct lu_env *env,
3887 struct obd_export *exp, struct obd_device *obd,
3888 struct obd_uuid *cluuid,
3889 struct obd_connect_data *data)
3891 struct client_obd *cli = &obd->u.cli;
3893 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3896 client_obd_list_lock(&cli->cl_loi_list_lock);
3897 data->ocd_grant = cli->cl_avail_grant ?:
3898 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3899 lost_grant = cli->cl_lost_grant;
3900 cli->cl_lost_grant = 0;
3901 client_obd_list_unlock(&cli->cl_loi_list_lock);
3903 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3904 "cl_lost_grant: %ld\n", data->ocd_grant,
3905 cli->cl_avail_grant, lost_grant);
3906 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3907 " ocd_grant: %d\n", data->ocd_connect_flags,
3908 data->ocd_version, data->ocd_grant);
3914 static int osc_disconnect(struct obd_export *exp)
3916 struct obd_device *obd = class_exp2obd(exp);
3917 struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3920 if (obd->u.cli.cl_conn_count == 1)
3921 /* flush any remaining cancel messages out to the target */
3922 llog_sync(ctxt, exp);
3924 llog_ctxt_put(ctxt);
3926 rc = client_disconnect_export(exp);
3930 static int osc_import_event(struct obd_device *obd,
3931 struct obd_import *imp,
3932 enum obd_import_event event)
3934 struct client_obd *cli;
3938 LASSERT(imp->imp_obd == obd);
3941 case IMP_EVENT_DISCON: {
3942 /* Only do this on the MDS OSC's */
3943 if (imp->imp_server_timeout) {
3944 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3946 spin_lock(&oscc->oscc_lock);
3947 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3948 spin_unlock(&oscc->oscc_lock);
3951 client_obd_list_lock(&cli->cl_loi_list_lock);
3952 cli->cl_avail_grant = 0;
3953 cli->cl_lost_grant = 0;
3954 client_obd_list_unlock(&cli->cl_loi_list_lock);
3957 case IMP_EVENT_INACTIVE: {
3958 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3961 case IMP_EVENT_INVALIDATE: {
3962 struct ldlm_namespace *ns = obd->obd_namespace;
3966 client_obd_list_lock(&cli->cl_loi_list_lock);
3967 /* all pages go to failing rpcs due to the invalid import */
3968 osc_check_rpcs(cli);
3969 client_obd_list_unlock(&cli->cl_loi_list_lock);
3971 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3975 case IMP_EVENT_ACTIVE: {
3976 /* Only do this on the MDS OSC's */
3977 if (imp->imp_server_timeout) {
3978 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3980 spin_lock(&oscc->oscc_lock);
3981 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3982 spin_unlock(&oscc->oscc_lock);
3984 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3987 case IMP_EVENT_OCD: {
3988 struct obd_connect_data *ocd = &imp->imp_connect_data;
3990 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3991 osc_init_grant(&obd->u.cli, ocd);
3994 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3995 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3997 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4001 CERROR("Unknown import event %d\n", event);
4007 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4013 rc = ptlrpcd_addref();
4017 rc = client_obd_setup(obd, lcfg);
4021 struct lprocfs_static_vars lvars = { 0 };
4022 struct client_obd *cli = &obd->u.cli;
4024 lprocfs_osc_init_vars(&lvars);
4025 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4026 lproc_osc_attach_seqstat(obd);
4027 sptlrpc_lprocfs_cliobd_attach(obd);
4028 ptlrpc_lprocfs_register_obd(obd);
4032 /* We need to allocate a few requests more, because
4033 brw_interpret tries to create new requests before freeing
4034 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4035 reserved, but I afraid that might be too much wasted RAM
4036 in fact, so 2 is just my guess and still should work. */
4037 cli->cl_import->imp_rq_pool =
4038 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4040 ptlrpc_add_rqs_to_pool);
4041 cli->cl_cache = cache_create(obd);
4042 if (!cli->cl_cache) {
4051 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4057 case OBD_CLEANUP_EARLY: {
4058 struct obd_import *imp;
4059 imp = obd->u.cli.cl_import;
4060 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4061 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4062 ptlrpc_deactivate_import(imp);
4063 spin_lock(&imp->imp_lock);
4064 imp->imp_pingable = 0;
4065 spin_unlock(&imp->imp_lock);
4068 case OBD_CLEANUP_EXPORTS: {
4069 /* If we set up but never connected, the
4070 client import will not have been cleaned. */
4071 if (obd->u.cli.cl_import) {
4072 struct obd_import *imp;
4073 imp = obd->u.cli.cl_import;
4074 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4076 ptlrpc_invalidate_import(imp);
4077 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4078 class_destroy_import(imp);
4079 obd->u.cli.cl_import = NULL;
4081 rc = obd_llog_finish(obd, 0);
4083 CERROR("failed to cleanup llogging subsystems\n");
4090 int osc_cleanup(struct obd_device *obd)
4092 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4096 ptlrpc_lprocfs_unregister_obd(obd);
4097 lprocfs_obd_cleanup(obd);
4099 spin_lock(&oscc->oscc_lock);
4100 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
4101 oscc->oscc_flags |= OSCC_FLAG_EXITING;
4102 spin_unlock(&oscc->oscc_lock);
4104 /* free memory of osc quota cache */
4105 lquota_cleanup(quota_interface, obd);
4107 cache_destroy(obd->u.cli.cl_cache);
4108 rc = client_obd_cleanup(obd);
4114 static int osc_register_page_removal_cb(struct obd_export *exp,
4115 obd_page_removal_cb_t func,
4116 obd_pin_extent_cb pin_cb)
4118 return cache_add_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func,
4122 static int osc_unregister_page_removal_cb(struct obd_export *exp,
4123 obd_page_removal_cb_t func)
4125 return cache_del_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func);
4128 static int osc_register_lock_cancel_cb(struct obd_export *exp,
4129 obd_lock_cancel_cb cb)
4131 LASSERT(exp->exp_obd->u.cli.cl_ext_lock_cancel_cb == NULL);
4133 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = cb;
4137 static int osc_unregister_lock_cancel_cb(struct obd_export *exp,
4138 obd_lock_cancel_cb cb)
4140 if (exp->exp_obd->u.cli.cl_ext_lock_cancel_cb != cb) {
4141 CERROR("Unregistering cancel cb %p, while only %p was "
4143 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb);
4147 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = NULL;
4151 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4153 struct lustre_cfg *lcfg = buf;
4154 struct lprocfs_static_vars lvars = { 0 };
4157 lprocfs_osc_init_vars(&lvars);
4159 switch (lcfg->lcfg_command) {
4160 case LCFG_SPTLRPC_CONF:
4161 rc = sptlrpc_cliobd_process_config(obd, lcfg);
4164 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4172 struct obd_ops osc_obd_ops = {
4173 .o_owner = THIS_MODULE,
4174 .o_setup = osc_setup,
4175 .o_precleanup = osc_precleanup,
4176 .o_cleanup = osc_cleanup,
4177 .o_add_conn = client_import_add_conn,
4178 .o_del_conn = client_import_del_conn,
4179 .o_connect = client_connect_import,
4180 .o_reconnect = osc_reconnect,
4181 .o_disconnect = osc_disconnect,
4182 .o_statfs = osc_statfs,
4183 .o_statfs_async = osc_statfs_async,
4184 .o_packmd = osc_packmd,
4185 .o_unpackmd = osc_unpackmd,
4186 .o_precreate = osc_precreate,
4187 .o_create = osc_create,
4188 .o_destroy = osc_destroy,
4189 .o_getattr = osc_getattr,
4190 .o_getattr_async = osc_getattr_async,
4191 .o_setattr = osc_setattr,
4192 .o_setattr_async = osc_setattr_async,
4194 .o_brw_async = osc_brw_async,
4195 .o_prep_async_page = osc_prep_async_page,
4196 .o_reget_short_lock = osc_reget_short_lock,
4197 .o_release_short_lock = osc_release_short_lock,
4198 .o_queue_async_io = osc_queue_async_io,
4199 .o_set_async_flags = osc_set_async_flags,
4200 .o_queue_group_io = osc_queue_group_io,
4201 .o_trigger_group_io = osc_trigger_group_io,
4202 .o_teardown_async_page = osc_teardown_async_page,
4203 .o_punch = osc_punch,
4205 .o_enqueue = osc_enqueue,
4206 .o_match = osc_match,
4207 .o_change_cbdata = osc_change_cbdata,
4208 .o_cancel = osc_cancel,
4209 .o_cancel_unused = osc_cancel_unused,
4210 .o_join_lru = osc_join_lru,
4211 .o_iocontrol = osc_iocontrol,
4212 .o_get_info = osc_get_info,
4213 .o_set_info_async = osc_set_info_async,
4214 .o_import_event = osc_import_event,
4215 .o_llog_init = osc_llog_init,
4216 .o_llog_finish = osc_llog_finish,
4217 .o_process_config = osc_process_config,
4218 .o_register_page_removal_cb = osc_register_page_removal_cb,
4219 .o_unregister_page_removal_cb = osc_unregister_page_removal_cb,
4220 .o_register_lock_cancel_cb = osc_register_lock_cancel_cb,
4221 .o_unregister_lock_cancel_cb = osc_unregister_lock_cancel_cb,
4223 int __init osc_init(void)
4225 struct lprocfs_static_vars lvars = { 0 };
4229 lprocfs_osc_init_vars(&lvars);
4231 request_module("lquota");
4232 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4233 lquota_init(quota_interface);
4234 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4236 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4237 LUSTRE_OSC_NAME, NULL);
4239 if (quota_interface)
4240 PORTAL_SYMBOL_PUT(osc_quota_interface);
4248 static void /*__exit*/ osc_exit(void)
4250 lquota_exit(quota_interface);
4251 if (quota_interface)
4252 PORTAL_SYMBOL_PUT(osc_quota_interface);
4254 class_unregister_type(LUSTRE_OSC_NAME);
4257 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4258 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4259 MODULE_LICENSE("GPL");
4261 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);