1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_OSC
42 #include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71 struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76 struct lov_stripe_md *lsm)
81 lmm_size = sizeof(**lmmp);
86 OBD_FREE(*lmmp, lmm_size);
92 OBD_ALLOC(*lmmp, lmm_size);
98 LASSERT(lsm->lsm_object_id);
99 LASSERT_MDS_GROUP(lsm->lsm_object_gr);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109 struct lov_mds_md *lmm, int lmm_bytes)
115 if (lmm_bytes < sizeof (*lmm)) {
116 CERROR("lov_mds_md too small: %d, need %d\n",
117 lmm_bytes, (int)sizeof(*lmm));
120 /* XXX LOV_MAGIC etc check? */
122 if (lmm->lmm_object_id == 0) {
123 CERROR("lov_mds_md: zero lmm_object_id\n");
128 lsm_size = lov_stripe_md_size(1);
132 if (*lsmp != NULL && lmm == NULL) {
133 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134 OBD_FREE(*lsmp, lsm_size);
140 OBD_ALLOC(*lsmp, lsm_size);
143 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145 OBD_FREE(*lsmp, lsm_size);
148 loi_init((*lsmp)->lsm_oinfo[0]);
152 /* XXX zero *lsmp? */
153 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT_MDS_GROUP((*lsmp)->lsm_object_gr);
159 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165 struct ost_body *body, void *capa)
167 struct obd_capa *oc = (struct obd_capa *)capa;
168 struct lustre_capa *c;
173 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
176 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177 DEBUG_CAPA(D_SEC, c, "pack");
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181 struct obd_info *oinfo)
183 struct ost_body *body;
185 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
188 body->oa = *oinfo->oi_oa;
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(const struct lu_env *env,
204 struct ptlrpc_request *req,
205 struct osc_async_args *aa, int rc)
207 struct ost_body *body;
213 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
214 lustre_swab_ost_body);
216 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
217 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
219 /* This should really be sent by the OST */
220 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
221 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
223 CDEBUG(D_INFO, "can't unpack ost_body\n");
225 aa->aa_oi->oi_oa->o_valid = 0;
228 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
232 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
233 struct ptlrpc_request_set *set)
235 struct ptlrpc_request *req;
236 struct osc_async_args *aa;
240 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
244 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
245 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
247 ptlrpc_request_free(req);
251 osc_pack_req_body(req, oinfo);
253 ptlrpc_request_set_replen(req);
254 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
256 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
257 aa = ptlrpc_req_async_args(req);
260 ptlrpc_set_add_req(set, req);
264 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
266 struct ptlrpc_request *req;
267 struct ost_body *body;
271 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
275 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
278 ptlrpc_request_free(req);
282 osc_pack_req_body(req, oinfo);
284 ptlrpc_request_set_replen(req);
286 rc = ptlrpc_queue_wait(req);
290 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
292 GOTO(out, rc = -EPROTO);
294 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
295 *oinfo->oi_oa = body->oa;
297 /* This should really be sent by the OST */
298 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
299 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
303 ptlrpc_req_finished(req);
307 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
308 struct obd_trans_info *oti)
310 struct ptlrpc_request *req;
311 struct ost_body *body;
315 LASSERTF(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
316 CHECK_MDS_GROUP(oinfo->oi_oa->o_gr),
317 "oinfo->oi_oa->o_valid="LPU64" oinfo->oi_oa->o_gr="LPU64"\n",
318 oinfo->oi_oa->o_valid, oinfo->oi_oa->o_gr);
320 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
324 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
325 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
327 ptlrpc_request_free(req);
331 osc_pack_req_body(req, oinfo);
333 ptlrpc_request_set_replen(req);
335 rc = ptlrpc_queue_wait(req);
339 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
341 GOTO(out, rc = -EPROTO);
343 *oinfo->oi_oa = body->oa;
347 ptlrpc_req_finished(req);
351 static int osc_setattr_interpret(const struct lu_env *env,
352 struct ptlrpc_request *req,
353 struct osc_async_args *aa, int rc)
355 struct ost_body *body;
361 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
363 GOTO(out, rc = -EPROTO);
365 *aa->aa_oi->oi_oa = body->oa;
367 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
371 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
372 struct obd_trans_info *oti,
373 struct ptlrpc_request_set *rqset)
375 struct ptlrpc_request *req;
376 struct osc_async_args *aa;
380 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
384 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
385 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
387 ptlrpc_request_free(req);
391 osc_pack_req_body(req, oinfo);
393 ptlrpc_request_set_replen(req);
395 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
397 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
400 /* do mds to ost setattr asynchronously */
402 /* Do not wait for response. */
403 ptlrpcd_add_req(req, PSCOPE_OTHER);
405 req->rq_interpret_reply =
406 (ptlrpc_interpterer_t)osc_setattr_interpret;
408 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
409 aa = ptlrpc_req_async_args(req);
412 ptlrpc_set_add_req(rqset, req);
418 int osc_real_create(struct obd_export *exp, struct obdo *oa,
419 struct lov_stripe_md **ea, struct obd_trans_info *oti)
421 struct ptlrpc_request *req;
422 struct ost_body *body;
423 struct lov_stripe_md *lsm;
432 rc = obd_alloc_memmd(exp, &lsm);
437 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
439 GOTO(out, rc = -ENOMEM);
441 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
443 ptlrpc_request_free(req);
447 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
451 ptlrpc_request_set_replen(req);
453 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
454 oa->o_flags == OBD_FL_DELORPHAN) {
456 "delorphan from OST integration");
457 /* Don't resend the delorphan req */
458 req->rq_no_resend = req->rq_no_delay = 1;
461 rc = ptlrpc_queue_wait(req);
465 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
467 GOTO(out_req, rc = -EPROTO);
471 /* This should really be sent by the OST */
472 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
473 oa->o_valid |= OBD_MD_FLBLKSZ;
475 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
476 * have valid lsm_oinfo data structs, so don't go touching that.
477 * This needs to be fixed in a big way.
479 lsm->lsm_object_id = oa->o_id;
480 lsm->lsm_object_gr = oa->o_gr;
484 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
486 if (oa->o_valid & OBD_MD_FLCOOKIE) {
487 if (!oti->oti_logcookies)
488 oti_alloc_cookies(oti, 1);
489 *oti->oti_logcookies = oa->o_lcookie;
493 CDEBUG(D_HA, "transno: "LPD64"\n",
494 lustre_msg_get_transno(req->rq_repmsg));
496 ptlrpc_req_finished(req);
499 obd_free_memmd(exp, &lsm);
503 static int osc_punch_interpret(const struct lu_env *env,
504 struct ptlrpc_request *req,
505 struct osc_punch_args *aa, int rc)
507 struct ost_body *body;
513 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
515 GOTO(out, rc = -EPROTO);
517 *aa->pa_oa = body->oa;
519 rc = aa->pa_upcall(aa->pa_cookie, rc);
523 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
524 struct obd_capa *capa,
525 obd_enqueue_update_f upcall, void *cookie,
526 struct ptlrpc_request_set *rqset)
528 struct ptlrpc_request *req;
529 struct osc_punch_args *aa;
530 struct ost_body *body;
534 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
538 osc_set_capa_size(req, &RMF_CAPA1, capa);
539 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
541 ptlrpc_request_free(req);
544 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
545 ptlrpc_at_set_req_timeout(req);
547 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
550 osc_pack_capa(req, body, capa);
552 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_punch_interpret;
556 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
557 aa = ptlrpc_req_async_args(req);
559 aa->pa_upcall = upcall;
560 aa->pa_cookie = cookie;
561 if (rqset == PTLRPCD_SET)
562 ptlrpcd_add_req(req, PSCOPE_OTHER);
564 ptlrpc_set_add_req(rqset, req);
569 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
570 struct obd_trans_info *oti,
571 struct ptlrpc_request_set *rqset)
573 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
574 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
575 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
576 return osc_punch_base(exp, oinfo->oi_oa, oinfo->oi_capa,
577 oinfo->oi_cb_up, oinfo, rqset);
580 static int osc_sync(struct obd_export *exp, struct obdo *oa,
581 struct lov_stripe_md *md, obd_size start, obd_size end,
584 struct ptlrpc_request *req;
585 struct ost_body *body;
590 CDEBUG(D_INFO, "oa NULL\n");
594 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
598 osc_set_capa_size(req, &RMF_CAPA1, capa);
599 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
601 ptlrpc_request_free(req);
605 /* overload the size and blocks fields in the oa with start/end */
606 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
609 body->oa.o_size = start;
610 body->oa.o_blocks = end;
611 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
612 osc_pack_capa(req, body, capa);
614 ptlrpc_request_set_replen(req);
616 rc = ptlrpc_queue_wait(req);
620 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
622 GOTO(out, rc = -EPROTO);
628 ptlrpc_req_finished(req);
632 /* Find and cancel locally locks matched by @mode in the resource found by
633 * @objid. Found locks are added into @cancel list. Returns the amount of
634 * locks added to @cancels list. */
635 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
636 struct list_head *cancels, ldlm_mode_t mode,
639 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
640 struct ldlm_res_id res_id;
641 struct ldlm_resource *res;
645 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
646 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
650 LDLM_RESOURCE_ADDREF(res);
651 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
652 lock_flags, 0, NULL);
653 LDLM_RESOURCE_DELREF(res);
654 ldlm_resource_putref(res);
658 static int osc_destroy_interpret(const struct lu_env *env,
659 struct ptlrpc_request *req, void *data,
662 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
664 atomic_dec(&cli->cl_destroy_in_flight);
665 cfs_waitq_signal(&cli->cl_destroy_waitq);
669 static int osc_can_send_destroy(struct client_obd *cli)
671 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
672 cli->cl_max_rpcs_in_flight) {
673 /* The destroy request can be sent */
676 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
677 cli->cl_max_rpcs_in_flight) {
679 * The counter has been modified between the two atomic
682 cfs_waitq_signal(&cli->cl_destroy_waitq);
687 /* Destroy requests can be async always on the client, and we don't even really
688 * care about the return code since the client cannot do anything at all about
690 * When the MDS is unlinking a filename, it saves the file objects into a
691 * recovery llog, and these object records are cancelled when the OST reports
692 * they were destroyed and sync'd to disk (i.e. transaction committed).
693 * If the client dies, or the OST is down when the object should be destroyed,
694 * the records are not cancelled, and when the OST reconnects to the MDS next,
695 * it will retrieve the llog unlink logs and then sends the log cancellation
696 * cookies to the MDS after committing destroy transactions. */
697 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
698 struct lov_stripe_md *ea, struct obd_trans_info *oti,
699 struct obd_export *md_export, void *capa)
701 struct client_obd *cli = &exp->exp_obd->u.cli;
702 struct ptlrpc_request *req;
703 struct ost_body *body;
704 CFS_LIST_HEAD(cancels);
709 CDEBUG(D_INFO, "oa NULL\n");
713 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
714 LDLM_FL_DISCARD_DATA);
716 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
718 ldlm_lock_list_put(&cancels, l_bl_ast, count);
722 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
723 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
726 ptlrpc_request_free(req);
730 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
731 req->rq_interpret_reply = osc_destroy_interpret;
732 ptlrpc_at_set_req_timeout(req);
734 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
735 oa->o_lcookie = *oti->oti_logcookies;
736 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
740 osc_pack_capa(req, body, (struct obd_capa *)capa);
741 ptlrpc_request_set_replen(req);
743 if (!osc_can_send_destroy(cli)) {
744 struct l_wait_info lwi = { 0 };
747 * Wait until the number of on-going destroy RPCs drops
748 * under max_rpc_in_flight
750 l_wait_event_exclusive(cli->cl_destroy_waitq,
751 osc_can_send_destroy(cli), &lwi);
754 /* Do not wait for response */
755 ptlrpcd_add_req(req, PSCOPE_OTHER);
759 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
762 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
764 LASSERT(!(oa->o_valid & bits));
767 client_obd_list_lock(&cli->cl_loi_list_lock);
768 oa->o_dirty = cli->cl_dirty;
769 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
770 CERROR("dirty %lu - %lu > dirty_max %lu\n",
771 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
773 } else if (atomic_read(&obd_dirty_pages) -
774 atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages){
775 CERROR("dirty %d - %d > system dirty_max %d\n",
776 atomic_read(&obd_dirty_pages),
777 atomic_read(&obd_dirty_transit_pages),
778 obd_max_dirty_pages);
780 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
781 CERROR("dirty %lu - dirty_max %lu too big???\n",
782 cli->cl_dirty, cli->cl_dirty_max);
785 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
786 (cli->cl_max_rpcs_in_flight + 1);
787 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
789 oa->o_grant = cli->cl_avail_grant;
790 oa->o_dropped = cli->cl_lost_grant;
791 cli->cl_lost_grant = 0;
792 client_obd_list_unlock(&cli->cl_loi_list_lock);
793 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
794 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
797 /* caller must hold loi_list_lock */
798 static void osc_consume_write_grant(struct client_obd *cli,
799 struct brw_page *pga)
801 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
802 atomic_inc(&obd_dirty_pages);
803 cli->cl_dirty += CFS_PAGE_SIZE;
804 cli->cl_avail_grant -= CFS_PAGE_SIZE;
805 pga->flag |= OBD_BRW_FROM_GRANT;
806 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
807 CFS_PAGE_SIZE, pga, pga->pg);
808 LASSERT(cli->cl_avail_grant >= 0);
811 /* the companion to osc_consume_write_grant, called when a brw has completed.
812 * must be called with the loi lock held. */
813 static void osc_release_write_grant(struct client_obd *cli,
814 struct brw_page *pga, int sent)
816 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
819 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
824 pga->flag &= ~OBD_BRW_FROM_GRANT;
825 atomic_dec(&obd_dirty_pages);
826 cli->cl_dirty -= CFS_PAGE_SIZE;
827 if (pga->flag & OBD_BRW_NOCACHE) {
828 pga->flag &= ~OBD_BRW_NOCACHE;
829 atomic_dec(&obd_dirty_transit_pages);
830 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
833 cli->cl_lost_grant += CFS_PAGE_SIZE;
834 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
835 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
836 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
837 /* For short writes we shouldn't count parts of pages that
838 * span a whole block on the OST side, or our accounting goes
839 * wrong. Should match the code in filter_grant_check. */
840 int offset = pga->off & ~CFS_PAGE_MASK;
841 int count = pga->count + (offset & (blocksize - 1));
842 int end = (offset + pga->count) & (blocksize - 1);
844 count += blocksize - end;
846 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
847 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
848 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
849 cli->cl_avail_grant, cli->cl_dirty);
855 static unsigned long rpcs_in_flight(struct client_obd *cli)
857 return cli->cl_r_in_flight + cli->cl_w_in_flight;
860 /* caller must hold loi_list_lock */
861 void osc_wake_cache_waiters(struct client_obd *cli)
863 struct list_head *l, *tmp;
864 struct osc_cache_waiter *ocw;
867 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
868 /* if we can't dirty more, we must wait until some is written */
869 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
870 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
871 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
872 "osc max %ld, sys max %d\n", cli->cl_dirty,
873 cli->cl_dirty_max, obd_max_dirty_pages);
877 /* if still dirty cache but no grant wait for pending RPCs that
878 * may yet return us some grant before doing sync writes */
879 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
880 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
881 cli->cl_w_in_flight);
885 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
886 list_del_init(&ocw->ocw_entry);
887 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
888 /* no more RPCs in flight to return grant, do sync IO */
889 ocw->ocw_rc = -EDQUOT;
890 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
892 osc_consume_write_grant(cli,
893 &ocw->ocw_oap->oap_brw_page);
896 cfs_waitq_signal(&ocw->ocw_waitq);
902 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
904 client_obd_list_lock(&cli->cl_loi_list_lock);
905 cli->cl_avail_grant = ocd->ocd_grant;
906 client_obd_list_unlock(&cli->cl_loi_list_lock);
908 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
909 cli->cl_avail_grant, cli->cl_lost_grant);
910 LASSERT(cli->cl_avail_grant >= 0);
913 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
915 client_obd_list_lock(&cli->cl_loi_list_lock);
916 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
917 if (body->oa.o_valid & OBD_MD_FLGRANT)
918 cli->cl_avail_grant += body->oa.o_grant;
919 /* waiters are woken in brw_interpret */
920 client_obd_list_unlock(&cli->cl_loi_list_lock);
923 /* We assume that the reason this OSC got a short read is because it read
924 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
925 * via the LOV, and it _knows_ it's reading inside the file, it's just that
926 * this stripe never got written at or beyond this stripe offset yet. */
927 static void handle_short_read(int nob_read, obd_count page_count,
928 struct brw_page **pga)
933 /* skip bytes read OK */
934 while (nob_read > 0) {
935 LASSERT (page_count > 0);
937 if (pga[i]->count > nob_read) {
938 /* EOF inside this page */
939 ptr = cfs_kmap(pga[i]->pg) +
940 (pga[i]->off & ~CFS_PAGE_MASK);
941 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
942 cfs_kunmap(pga[i]->pg);
948 nob_read -= pga[i]->count;
953 /* zero remaining pages */
954 while (page_count-- > 0) {
955 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
956 memset(ptr, 0, pga[i]->count);
957 cfs_kunmap(pga[i]->pg);
962 static int check_write_rcs(struct ptlrpc_request *req,
963 int requested_nob, int niocount,
964 obd_count page_count, struct brw_page **pga)
968 /* return error if any niobuf was in error */
969 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
970 sizeof(*remote_rcs) * niocount, NULL);
971 if (remote_rcs == NULL) {
972 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
975 if (lustre_msg_swabbed(req->rq_repmsg))
976 for (i = 0; i < niocount; i++)
977 __swab32s(&remote_rcs[i]);
979 for (i = 0; i < niocount; i++) {
980 if (remote_rcs[i] < 0)
981 return(remote_rcs[i]);
983 if (remote_rcs[i] != 0) {
984 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
985 i, remote_rcs[i], req);
990 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
991 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
992 req->rq_bulk->bd_nob_transferred, requested_nob);
999 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1001 if (p1->flag != p2->flag) {
1002 unsigned mask = ~(OBD_BRW_FROM_GRANT|OBD_BRW_NOCACHE);
1004 /* warn if we try to combine flags that we don't know to be
1005 * safe to combine */
1006 if ((p1->flag & mask) != (p2->flag & mask))
1007 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1008 "same brw?\n", p1->flag, p2->flag);
1012 return (p1->off + p1->count == p2->off);
1015 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1016 struct brw_page **pga, int opc,
1017 cksum_type_t cksum_type)
1022 LASSERT (pg_count > 0);
1023 cksum = init_checksum(cksum_type);
1024 while (nob > 0 && pg_count > 0) {
1025 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1026 int off = pga[i]->off & ~CFS_PAGE_MASK;
1027 int count = pga[i]->count > nob ? nob : pga[i]->count;
1029 /* corrupt the data before we compute the checksum, to
1030 * simulate an OST->client data error */
1031 if (i == 0 && opc == OST_READ &&
1032 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1033 memcpy(ptr + off, "bad1", min(4, nob));
1034 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1035 cfs_kunmap(pga[i]->pg);
1036 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1039 nob -= pga[i]->count;
1043 /* For sending we only compute the wrong checksum instead
1044 * of corrupting the data so it is still correct on a redo */
1045 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1051 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1052 struct lov_stripe_md *lsm, obd_count page_count,
1053 struct brw_page **pga,
1054 struct ptlrpc_request **reqp,
1055 struct obd_capa *ocapa, int reserve)
1057 struct ptlrpc_request *req;
1058 struct ptlrpc_bulk_desc *desc;
1059 struct ost_body *body;
1060 struct obd_ioobj *ioobj;
1061 struct niobuf_remote *niobuf;
1062 int niocount, i, requested_nob, opc, rc;
1063 struct osc_brw_async_args *aa;
1064 struct req_capsule *pill;
1065 struct brw_page *pg_prev;
1068 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1069 RETURN(-ENOMEM); /* Recoverable */
1070 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1071 RETURN(-EINVAL); /* Fatal */
1073 if ((cmd & OBD_BRW_WRITE) != 0) {
1075 req = ptlrpc_request_alloc_pool(cli->cl_import,
1076 cli->cl_import->imp_rq_pool,
1080 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1085 for (niocount = i = 1; i < page_count; i++) {
1086 if (!can_merge_pages(pga[i - 1], pga[i]))
1090 pill = &req->rq_pill;
1091 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1092 niocount * sizeof(*niobuf));
1093 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1095 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1097 ptlrpc_request_free(req);
1100 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1101 ptlrpc_at_set_req_timeout(req);
1103 if (opc == OST_WRITE)
1104 desc = ptlrpc_prep_bulk_imp(req, page_count,
1105 BULK_GET_SOURCE, OST_BULK_PORTAL);
1107 desc = ptlrpc_prep_bulk_imp(req, page_count,
1108 BULK_PUT_SINK, OST_BULK_PORTAL);
1111 GOTO(out, rc = -ENOMEM);
1112 /* NB request now owns desc and will free it when it gets freed */
1114 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1115 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1116 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1117 LASSERT(body && ioobj && niobuf);
1121 obdo_to_ioobj(oa, ioobj);
1122 ioobj->ioo_bufcnt = niocount;
1123 osc_pack_capa(req, body, ocapa);
1124 LASSERT (page_count > 0);
1126 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1127 struct brw_page *pg = pga[i];
1129 LASSERT(pg->count > 0);
1130 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1131 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1132 pg->off, pg->count);
1134 LASSERTF(i == 0 || pg->off > pg_prev->off,
1135 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1136 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1138 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1139 pg_prev->pg, page_private(pg_prev->pg),
1140 pg_prev->pg->index, pg_prev->off);
1142 LASSERTF(i == 0 || pg->off > pg_prev->off,
1143 "i %d p_c %u\n", i, page_count);
1145 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1146 (pg->flag & OBD_BRW_SRVLOCK));
1148 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1150 requested_nob += pg->count;
1152 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1154 niobuf->len += pg->count;
1156 niobuf->offset = pg->off;
1157 niobuf->len = pg->count;
1158 niobuf->flags = pg->flag;
1163 LASSERTF((void *)(niobuf - niocount) ==
1164 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1165 niocount * sizeof(*niobuf)),
1166 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1167 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1168 (void *)(niobuf - niocount));
1170 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1172 /* size[REQ_REC_OFF] still sizeof (*body) */
1173 if (opc == OST_WRITE) {
1174 if (unlikely(cli->cl_checksum) &&
1175 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1176 /* store cl_cksum_type in a local variable since
1177 * it can be changed via lprocfs */
1178 cksum_type_t cksum_type = cli->cl_cksum_type;
1180 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1181 oa->o_flags = body->oa.o_flags = 0;
1182 body->oa.o_flags |= cksum_type_pack(cksum_type);
1183 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1184 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1188 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1190 /* save this in 'oa', too, for later checking */
1191 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1192 oa->o_flags |= cksum_type_pack(cksum_type);
1194 /* clear out the checksum flag, in case this is a
1195 * resend but cl_checksum is no longer set. b=11238 */
1196 oa->o_valid &= ~OBD_MD_FLCKSUM;
1198 oa->o_cksum = body->oa.o_cksum;
1199 /* 1 RC per niobuf */
1200 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1201 sizeof(__u32) * niocount);
1203 if (unlikely(cli->cl_checksum) &&
1204 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1205 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1206 body->oa.o_flags = 0;
1207 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1208 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1210 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1211 /* 1 RC for the whole I/O */
1213 ptlrpc_request_set_replen(req);
1215 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1216 aa = ptlrpc_req_async_args(req);
1218 aa->aa_requested_nob = requested_nob;
1219 aa->aa_nio_count = niocount;
1220 aa->aa_page_count = page_count;
1224 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1225 if (ocapa && reserve)
1226 aa->aa_ocapa = capa_get(ocapa);
1232 ptlrpc_req_finished(req);
1236 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1237 __u32 client_cksum, __u32 server_cksum, int nob,
1238 obd_count page_count, struct brw_page **pga,
1239 cksum_type_t client_cksum_type)
1243 cksum_type_t cksum_type;
1245 if (server_cksum == client_cksum) {
1246 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1250 if (oa->o_valid & OBD_MD_FLFLAGS)
1251 cksum_type = cksum_type_unpack(oa->o_flags);
1253 cksum_type = OBD_CKSUM_CRC32;
1255 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1258 if (cksum_type != client_cksum_type)
1259 msg = "the server did not use the checksum type specified in "
1260 "the original request - likely a protocol problem";
1261 else if (new_cksum == server_cksum)
1262 msg = "changed on the client after we checksummed it - "
1263 "likely false positive due to mmap IO (bug 11742)";
1264 else if (new_cksum == client_cksum)
1265 msg = "changed in transit before arrival at OST";
1267 msg = "changed in transit AND doesn't match the original - "
1268 "likely false positive due to mmap IO (bug 11742)";
1270 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1271 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1272 "["LPU64"-"LPU64"]\n",
1273 msg, libcfs_nid2str(peer->nid),
1274 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1275 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1278 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1280 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1281 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1282 "client csum now %x\n", client_cksum, client_cksum_type,
1283 server_cksum, cksum_type, new_cksum);
1287 /* Note rc enters this function as number of bytes transferred */
1288 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1290 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1291 const lnet_process_id_t *peer =
1292 &req->rq_import->imp_connection->c_peer;
1293 struct client_obd *cli = aa->aa_cli;
1294 struct ost_body *body;
1295 __u32 client_cksum = 0;
1298 if (rc < 0 && rc != -EDQUOT)
1301 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1302 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1303 lustre_swab_ost_body);
1305 CDEBUG(D_INFO, "Can't unpack body\n");
1309 /* set/clear over quota flag for a uid/gid */
1310 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1311 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1312 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1313 body->oa.o_gid, body->oa.o_valid,
1319 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1320 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1322 osc_update_grant(cli, body);
1324 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1326 CERROR("Unexpected +ve rc %d\n", rc);
1329 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1331 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1332 check_write_checksum(&body->oa, peer, client_cksum,
1333 body->oa.o_cksum, aa->aa_requested_nob,
1334 aa->aa_page_count, aa->aa_ppga,
1335 cksum_type_unpack(aa->aa_oa->o_flags)))
1338 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1341 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1342 aa->aa_page_count, aa->aa_ppga);
1346 /* The rest of this function executes only for OST_READs */
1347 if (rc > aa->aa_requested_nob) {
1348 CERROR("Unexpected rc %d (%d requested)\n", rc,
1349 aa->aa_requested_nob);
1353 if (rc != req->rq_bulk->bd_nob_transferred) {
1354 CERROR ("Unexpected rc %d (%d transferred)\n",
1355 rc, req->rq_bulk->bd_nob_transferred);
1359 if (rc < aa->aa_requested_nob)
1360 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1362 if (sptlrpc_cli_unwrap_bulk_read(req, rc, aa->aa_page_count,
1364 GOTO(out, rc = -EAGAIN);
1366 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1367 static int cksum_counter;
1368 __u32 server_cksum = body->oa.o_cksum;
1371 cksum_type_t cksum_type;
1373 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1374 cksum_type = cksum_type_unpack(body->oa.o_flags);
1376 cksum_type = OBD_CKSUM_CRC32;
1377 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1378 aa->aa_ppga, OST_READ,
1381 if (peer->nid == req->rq_bulk->bd_sender) {
1385 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1388 if (server_cksum == ~0 && rc > 0) {
1389 CERROR("Protocol error: server %s set the 'checksum' "
1390 "bit, but didn't send a checksum. Not fatal, "
1391 "but please notify on http://bugzilla.lustre.org/\n",
1392 libcfs_nid2str(peer->nid));
1393 } else if (server_cksum != client_cksum) {
1394 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1395 "%s%s%s inum "LPU64"/"LPU64" object "
1396 LPU64"/"LPU64" extent "
1397 "["LPU64"-"LPU64"]\n",
1398 req->rq_import->imp_obd->obd_name,
1399 libcfs_nid2str(peer->nid),
1401 body->oa.o_valid & OBD_MD_FLFID ?
1402 body->oa.o_fid : (__u64)0,
1403 body->oa.o_valid & OBD_MD_FLFID ?
1404 body->oa.o_generation :(__u64)0,
1406 body->oa.o_valid & OBD_MD_FLGROUP ?
1407 body->oa.o_gr : (__u64)0,
1408 aa->aa_ppga[0]->off,
1409 aa->aa_ppga[aa->aa_page_count-1]->off +
1410 aa->aa_ppga[aa->aa_page_count-1]->count -
1412 CERROR("client %x, server %x, cksum_type %x\n",
1413 client_cksum, server_cksum, cksum_type);
1415 aa->aa_oa->o_cksum = client_cksum;
1419 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1422 } else if (unlikely(client_cksum)) {
1423 static int cksum_missed;
1426 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1427 CERROR("Checksum %u requested from %s but not sent\n",
1428 cksum_missed, libcfs_nid2str(peer->nid));
1434 *aa->aa_oa = body->oa;
1439 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1440 struct lov_stripe_md *lsm,
1441 obd_count page_count, struct brw_page **pga,
1442 struct obd_capa *ocapa)
1444 struct ptlrpc_request *req;
1448 struct l_wait_info lwi;
1452 cfs_waitq_init(&waitq);
1455 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1456 page_count, pga, &req, ocapa, 0);
1460 rc = ptlrpc_queue_wait(req);
1462 if (rc == -ETIMEDOUT && req->rq_resend) {
1463 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1464 ptlrpc_req_finished(req);
1468 rc = osc_brw_fini_request(req, rc);
1470 ptlrpc_req_finished(req);
1471 if (osc_recoverable_error(rc)) {
1473 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1474 CERROR("too many resend retries, returning error\n");
1478 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1479 l_wait_event(waitq, 0, &lwi);
1487 int osc_brw_redo_request(struct ptlrpc_request *request,
1488 struct osc_brw_async_args *aa)
1490 struct ptlrpc_request *new_req;
1491 struct ptlrpc_request_set *set = request->rq_set;
1492 struct osc_brw_async_args *new_aa;
1493 struct osc_async_page *oap;
1497 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1498 CERROR("too many resend retries, returning error\n");
1502 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1504 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1505 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1506 aa->aa_cli, aa->aa_oa,
1507 NULL /* lsm unused by osc currently */,
1508 aa->aa_page_count, aa->aa_ppga,
1509 &new_req, aa->aa_ocapa, 0);
1513 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1515 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1516 if (oap->oap_request != NULL) {
1517 LASSERTF(request == oap->oap_request,
1518 "request %p != oap_request %p\n",
1519 request, oap->oap_request);
1520 if (oap->oap_interrupted) {
1521 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1522 ptlrpc_req_finished(new_req);
1527 /* New request takes over pga and oaps from old request.
1528 * Note that copying a list_head doesn't work, need to move it... */
1530 new_req->rq_interpret_reply = request->rq_interpret_reply;
1531 new_req->rq_async_args = request->rq_async_args;
1532 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1534 new_aa = ptlrpc_req_async_args(new_req);
1536 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1537 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1538 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1540 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1541 if (oap->oap_request) {
1542 ptlrpc_req_finished(oap->oap_request);
1543 oap->oap_request = ptlrpc_request_addref(new_req);
1547 new_aa->aa_ocapa = aa->aa_ocapa;
1548 aa->aa_ocapa = NULL;
1550 /* use ptlrpc_set_add_req is safe because interpret functions work
1551 * in check_set context. only one way exist with access to request
1552 * from different thread got -EINTR - this way protected with
1553 * cl_loi_list_lock */
1554 ptlrpc_set_add_req(set, new_req);
1556 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1558 DEBUG_REQ(D_INFO, new_req, "new request");
1563 * ugh, we want disk allocation on the target to happen in offset order. we'll
1564 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1565 * fine for our small page arrays and doesn't require allocation. its an
1566 * insertion sort that swaps elements that are strides apart, shrinking the
1567 * stride down until its '1' and the array is sorted.
1569 static void sort_brw_pages(struct brw_page **array, int num)
1572 struct brw_page *tmp;
1576 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1581 for (i = stride ; i < num ; i++) {
1584 while (j >= stride && array[j - stride]->off > tmp->off) {
1585 array[j] = array[j - stride];
1590 } while (stride > 1);
1593 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1599 LASSERT (pages > 0);
1600 offset = pg[i]->off & ~CFS_PAGE_MASK;
1604 if (pages == 0) /* that's all */
1607 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1608 return count; /* doesn't end on page boundary */
1611 offset = pg[i]->off & ~CFS_PAGE_MASK;
1612 if (offset != 0) /* doesn't start on page boundary */
1619 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1621 struct brw_page **ppga;
1624 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1628 for (i = 0; i < count; i++)
1633 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1635 LASSERT(ppga != NULL);
1636 OBD_FREE(ppga, sizeof(*ppga) * count);
1639 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1640 obd_count page_count, struct brw_page *pga,
1641 struct obd_trans_info *oti)
1643 struct obdo *saved_oa = NULL;
1644 struct brw_page **ppga, **orig;
1645 struct obd_import *imp = class_exp2cliimp(exp);
1646 struct client_obd *cli = &imp->imp_obd->u.cli;
1647 int rc, page_count_orig;
1650 if (cmd & OBD_BRW_CHECK) {
1651 /* The caller just wants to know if there's a chance that this
1652 * I/O can succeed */
1654 if (imp == NULL || imp->imp_invalid)
1659 /* test_brw with a failed create can trip this, maybe others. */
1660 LASSERT(cli->cl_max_pages_per_rpc);
1664 orig = ppga = osc_build_ppga(pga, page_count);
1667 page_count_orig = page_count;
1669 sort_brw_pages(ppga, page_count);
1670 while (page_count) {
1671 obd_count pages_per_brw;
1673 if (page_count > cli->cl_max_pages_per_rpc)
1674 pages_per_brw = cli->cl_max_pages_per_rpc;
1676 pages_per_brw = page_count;
1678 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1680 if (saved_oa != NULL) {
1681 /* restore previously saved oa */
1682 *oinfo->oi_oa = *saved_oa;
1683 } else if (page_count > pages_per_brw) {
1684 /* save a copy of oa (brw will clobber it) */
1685 OBDO_ALLOC(saved_oa);
1686 if (saved_oa == NULL)
1687 GOTO(out, rc = -ENOMEM);
1688 *saved_oa = *oinfo->oi_oa;
1691 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1692 pages_per_brw, ppga, oinfo->oi_capa);
1697 page_count -= pages_per_brw;
1698 ppga += pages_per_brw;
1702 osc_release_ppga(orig, page_count_orig);
1704 if (saved_oa != NULL)
1705 OBDO_FREE(saved_oa);
1710 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1711 * the dirty accounting. Writeback completes or truncate happens before
1712 * writing starts. Must be called with the loi lock held. */
1713 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1716 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1720 /* This maintains the lists of pending pages to read/write for a given object
1721 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1722 * to quickly find objects that are ready to send an RPC. */
1723 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1729 if (lop->lop_num_pending == 0)
1732 /* if we have an invalid import we want to drain the queued pages
1733 * by forcing them through rpcs that immediately fail and complete
1734 * the pages. recovery relies on this to empty the queued pages
1735 * before canceling the locks and evicting down the llite pages */
1736 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1739 /* stream rpcs in queue order as long as as there is an urgent page
1740 * queued. this is our cheap solution for good batching in the case
1741 * where writepage marks some random page in the middle of the file
1742 * as urgent because of, say, memory pressure */
1743 if (!list_empty(&lop->lop_urgent)) {
1744 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1747 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1748 optimal = cli->cl_max_pages_per_rpc;
1749 if (cmd & OBD_BRW_WRITE) {
1750 /* trigger a write rpc stream as long as there are dirtiers
1751 * waiting for space. as they're waiting, they're not going to
1752 * create more pages to coallesce with what's waiting.. */
1753 if (!list_empty(&cli->cl_cache_waiters)) {
1754 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1757 /* +16 to avoid triggering rpcs that would want to include pages
1758 * that are being queued but which can't be made ready until
1759 * the queuer finishes with the page. this is a wart for
1760 * llite::commit_write() */
1763 if (lop->lop_num_pending >= optimal)
1769 static void on_list(struct list_head *item, struct list_head *list,
1772 if (list_empty(item) && should_be_on)
1773 list_add_tail(item, list);
1774 else if (!list_empty(item) && !should_be_on)
1775 list_del_init(item);
1778 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1779 * can find pages to build into rpcs quickly */
1780 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1782 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1783 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1784 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1786 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1787 loi->loi_write_lop.lop_num_pending);
1789 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1790 loi->loi_read_lop.lop_num_pending);
1793 static void lop_update_pending(struct client_obd *cli,
1794 struct loi_oap_pages *lop, int cmd, int delta)
1796 lop->lop_num_pending += delta;
1797 if (cmd & OBD_BRW_WRITE)
1798 cli->cl_pending_w_pages += delta;
1800 cli->cl_pending_r_pages += delta;
1804 * this is called when a sync waiter receives an interruption. Its job is to
1805 * get the caller woken as soon as possible. If its page hasn't been put in an
1806 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1807 * desiring interruption which will forcefully complete the rpc once the rpc
1810 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
1812 struct loi_oap_pages *lop;
1813 struct lov_oinfo *loi;
1817 LASSERT(!oap->oap_interrupted);
1818 oap->oap_interrupted = 1;
1820 /* ok, it's been put in an rpc. only one oap gets a request reference */
1821 if (oap->oap_request != NULL) {
1822 ptlrpc_mark_interrupted(oap->oap_request);
1823 ptlrpcd_wake(oap->oap_request);
1824 ptlrpc_req_finished(oap->oap_request);
1825 oap->oap_request = NULL;
1829 * page completion may be called only if ->cpo_prep() method was
1830 * executed by osc_io_submit(), that also adds page the to pending list
1832 if (!list_empty(&oap->oap_pending_item)) {
1833 list_del_init(&oap->oap_pending_item);
1834 list_del_init(&oap->oap_urgent_item);
1837 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1838 &loi->loi_write_lop : &loi->loi_read_lop;
1839 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1840 loi_list_maint(oap->oap_cli, oap->oap_loi);
1841 rc = oap->oap_caller_ops->ap_completion(env,
1842 oap->oap_caller_data,
1843 oap->oap_cmd, NULL, -EINTR);
1849 /* this is trying to propogate async writeback errors back up to the
1850 * application. As an async write fails we record the error code for later if
1851 * the app does an fsync. As long as errors persist we force future rpcs to be
1852 * sync so that the app can get a sync error and break the cycle of queueing
1853 * pages for which writeback will fail. */
1854 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1861 ar->ar_force_sync = 1;
1862 ar->ar_min_xid = ptlrpc_sample_next_xid();
1867 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1868 ar->ar_force_sync = 0;
1871 void osc_oap_to_pending(struct osc_async_page *oap)
1873 struct loi_oap_pages *lop;
1875 if (oap->oap_cmd & OBD_BRW_WRITE)
1876 lop = &oap->oap_loi->loi_write_lop;
1878 lop = &oap->oap_loi->loi_read_lop;
1880 if (oap->oap_async_flags & ASYNC_URGENT)
1881 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1882 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1883 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1886 /* this must be called holding the loi list lock to give coverage to exit_cache,
1887 * async_flag maintenance, and oap_request */
1888 static void osc_ap_completion(const struct lu_env *env,
1889 struct client_obd *cli, struct obdo *oa,
1890 struct osc_async_page *oap, int sent, int rc)
1895 if (oap->oap_request != NULL) {
1896 xid = ptlrpc_req_xid(oap->oap_request);
1897 ptlrpc_req_finished(oap->oap_request);
1898 oap->oap_request = NULL;
1901 oap->oap_async_flags = 0;
1902 oap->oap_interrupted = 0;
1904 if (oap->oap_cmd & OBD_BRW_WRITE) {
1905 osc_process_ar(&cli->cl_ar, xid, rc);
1906 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
1909 if (rc == 0 && oa != NULL) {
1910 if (oa->o_valid & OBD_MD_FLBLOCKS)
1911 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
1912 if (oa->o_valid & OBD_MD_FLMTIME)
1913 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
1914 if (oa->o_valid & OBD_MD_FLATIME)
1915 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
1916 if (oa->o_valid & OBD_MD_FLCTIME)
1917 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
1920 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
1921 oap->oap_cmd, oa, rc);
1923 /* ll_ap_completion (from llite) drops PG_locked. so, a new
1924 * I/O on the page could start, but OSC calls it under lock
1925 * and thus we can add oap back to pending safely */
1927 /* upper layer wants to leave the page on pending queue */
1928 osc_oap_to_pending(oap);
1930 osc_exit_cache(cli, oap, sent);
1934 static int brw_interpret(const struct lu_env *env,
1935 struct ptlrpc_request *req, void *data, int rc)
1937 struct osc_brw_async_args *aa = data;
1938 struct client_obd *cli;
1942 rc = osc_brw_fini_request(req, rc);
1943 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1944 if (osc_recoverable_error(rc)) {
1945 rc = osc_brw_redo_request(req, aa);
1951 capa_put(aa->aa_ocapa);
1952 aa->aa_ocapa = NULL;
1957 client_obd_list_lock(&cli->cl_loi_list_lock);
1959 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1960 * is called so we know whether to go to sync BRWs or wait for more
1961 * RPCs to complete */
1962 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1963 cli->cl_w_in_flight--;
1965 cli->cl_r_in_flight--;
1967 async = list_empty(&aa->aa_oaps);
1968 if (!async) { /* from osc_send_oap_rpc() */
1969 struct osc_async_page *oap, *tmp;
1970 /* the caller may re-use the oap after the completion call so
1971 * we need to clean it up a little */
1972 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
1973 list_del_init(&oap->oap_rpc_item);
1974 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
1976 OBDO_FREE(aa->aa_oa);
1977 } else { /* from async_internal() */
1979 for (i = 0; i < aa->aa_page_count; i++)
1980 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
1982 osc_wake_cache_waiters(cli);
1983 osc_check_rpcs(env, cli);
1984 client_obd_list_unlock(&cli->cl_loi_list_lock);
1986 cl_req_completion(env, aa->aa_clerq, rc);
1987 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1991 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
1992 struct client_obd *cli,
1993 struct list_head *rpc_list,
1994 int page_count, int cmd)
1996 struct ptlrpc_request *req;
1997 struct brw_page **pga = NULL;
1998 struct osc_brw_async_args *aa;
1999 struct obdo *oa = NULL;
2000 const struct obd_async_page_ops *ops = NULL;
2001 void *caller_data = NULL;
2002 struct osc_async_page *oap;
2003 struct osc_async_page *tmp;
2004 struct ost_body *body;
2005 struct cl_req *clerq = NULL;
2006 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2007 struct ldlm_lock *lock = NULL;
2008 struct cl_req_attr crattr;
2012 LASSERT(!list_empty(rpc_list));
2014 memset(&crattr, 0, sizeof crattr);
2015 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2017 GOTO(out, req = ERR_PTR(-ENOMEM));
2021 GOTO(out, req = ERR_PTR(-ENOMEM));
2024 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2025 struct cl_page *page = osc_oap2cl_page(oap);
2027 ops = oap->oap_caller_ops;
2028 caller_data = oap->oap_caller_data;
2030 clerq = cl_req_alloc(env, page, crt,
2031 1 /* only 1-object rpcs for
2034 GOTO(out, req = (void *)clerq);
2035 lock = oap->oap_ldlm_lock;
2037 pga[i] = &oap->oap_brw_page;
2038 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2039 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2040 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2042 cl_req_page_add(env, clerq, page);
2045 /* always get the data for the obdo for the rpc */
2046 LASSERT(ops != NULL);
2048 crattr.cra_capa = NULL;
2049 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2051 oa->o_handle = lock->l_remote_handle;
2052 oa->o_valid |= OBD_MD_FLHANDLE;
2055 rc = cl_req_prep(env, clerq);
2057 CERROR("cl_req_prep failed: %d\n", rc);
2058 GOTO(out, req = ERR_PTR(rc));
2061 sort_brw_pages(pga, page_count);
2062 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2063 pga, &req, crattr.cra_capa, 1);
2065 CERROR("prep_req failed: %d\n", rc);
2066 GOTO(out, req = ERR_PTR(rc));
2069 /* Need to update the timestamps after the request is built in case
2070 * we race with setattr (locally or in queue at OST). If OST gets
2071 * later setattr before earlier BRW (as determined by the request xid),
2072 * the OST will not use BRW timestamps. Sadly, there is no obvious
2073 * way to do this in a single call. bug 10150 */
2074 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2075 cl_req_attr_set(env, clerq, &crattr,
2076 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2078 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2079 aa = ptlrpc_req_async_args(req);
2080 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2081 list_splice(rpc_list, &aa->aa_oaps);
2082 CFS_INIT_LIST_HEAD(rpc_list);
2083 aa->aa_clerq = clerq;
2085 capa_put(crattr.cra_capa);
2090 OBD_FREE(pga, sizeof(*pga) * page_count);
2091 /* this should happen rarely and is pretty bad, it makes the
2092 * pending list not follow the dirty order */
2093 client_obd_list_lock(&cli->cl_loi_list_lock);
2094 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2095 list_del_init(&oap->oap_rpc_item);
2097 /* queued sync pages can be torn down while the pages
2098 * were between the pending list and the rpc */
2099 if (oap->oap_interrupted) {
2100 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2101 osc_ap_completion(env, cli, NULL, oap, 0,
2105 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2107 if (clerq && !IS_ERR(clerq))
2108 cl_req_completion(env, clerq, PTR_ERR(req));
2114 * prepare pages for ASYNC io and put pages in send queue.
2118 * \param cmd - OBD_BRW_* macroses
2119 * \param lop - pending pages
2121 * \return zero if pages successfully add to send queue.
2122 * \return not zere if error occurring.
2125 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2126 struct lov_oinfo *loi,
2127 int cmd, struct loi_oap_pages *lop)
2129 struct ptlrpc_request *req;
2130 obd_count page_count = 0;
2131 struct osc_async_page *oap = NULL, *tmp;
2132 struct osc_brw_async_args *aa;
2133 const struct obd_async_page_ops *ops;
2134 CFS_LIST_HEAD(rpc_list);
2135 unsigned int ending_offset;
2136 unsigned starting_offset = 0;
2138 struct cl_object *clob = NULL;
2141 /* first we find the pages we're allowed to work with */
2142 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2144 ops = oap->oap_caller_ops;
2146 LASSERT(oap->oap_magic == OAP_MAGIC);
2149 /* pin object in memory, so that completion call-backs
2150 * can be safely called under client_obd_list lock. */
2151 clob = osc_oap2cl_page(oap)->cp_obj;
2152 cl_object_get(clob);
2155 if (page_count != 0 &&
2156 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2157 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2158 " oap %p, page %p, srvlock %u\n",
2159 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2162 /* in llite being 'ready' equates to the page being locked
2163 * until completion unlocks it. commit_write submits a page
2164 * as not ready because its unlock will happen unconditionally
2165 * as the call returns. if we race with commit_write giving
2166 * us that page we dont' want to create a hole in the page
2167 * stream, so we stop and leave the rpc to be fired by
2168 * another dirtier or kupdated interval (the not ready page
2169 * will still be on the dirty list). we could call in
2170 * at the end of ll_file_write to process the queue again. */
2171 if (!(oap->oap_async_flags & ASYNC_READY)) {
2172 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2175 CDEBUG(D_INODE, "oap %p page %p returned %d "
2176 "instead of ready\n", oap,
2180 /* llite is telling us that the page is still
2181 * in commit_write and that we should try
2182 * and put it in an rpc again later. we
2183 * break out of the loop so we don't create
2184 * a hole in the sequence of pages in the rpc
2189 /* the io isn't needed.. tell the checks
2190 * below to complete the rpc with EINTR */
2191 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2192 oap->oap_count = -EINTR;
2195 oap->oap_async_flags |= ASYNC_READY;
2198 LASSERTF(0, "oap %p page %p returned %d "
2199 "from make_ready\n", oap,
2207 * Page submitted for IO has to be locked. Either by
2208 * ->ap_make_ready() or by higher layers.
2210 #if defined(__KERNEL__) && defined(__linux__)
2212 struct cl_page *page;
2214 page = osc_oap2cl_page(oap);
2216 if (page->cp_type == CPT_CACHEABLE &&
2217 !(PageLocked(oap->oap_page) &&
2218 (CheckWriteback(oap->oap_page, cmd)))) {
2219 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2221 (long)oap->oap_page->flags,
2222 oap->oap_async_flags);
2227 /* If there is a gap at the start of this page, it can't merge
2228 * with any previous page, so we'll hand the network a
2229 * "fragmented" page array that it can't transfer in 1 RDMA */
2230 if (page_count != 0 && oap->oap_page_off != 0)
2233 /* take the page out of our book-keeping */
2234 list_del_init(&oap->oap_pending_item);
2235 lop_update_pending(cli, lop, cmd, -1);
2236 list_del_init(&oap->oap_urgent_item);
2238 if (page_count == 0)
2239 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2240 (PTLRPC_MAX_BRW_SIZE - 1);
2242 /* ask the caller for the size of the io as the rpc leaves. */
2243 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2245 ops->ap_refresh_count(env, oap->oap_caller_data,
2247 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2249 if (oap->oap_count <= 0) {
2250 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2252 osc_ap_completion(env, cli, NULL,
2253 oap, 0, oap->oap_count);
2257 /* now put the page back in our accounting */
2258 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2259 if (page_count == 0)
2260 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2261 if (++page_count >= cli->cl_max_pages_per_rpc)
2264 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2265 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2266 * have the same alignment as the initial writes that allocated
2267 * extents on the server. */
2268 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2269 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2270 if (ending_offset == 0)
2273 /* If there is a gap at the end of this page, it can't merge
2274 * with any subsequent pages, so we'll hand the network a
2275 * "fragmented" page array that it can't transfer in 1 RDMA */
2276 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2280 osc_wake_cache_waiters(cli);
2282 loi_list_maint(cli, loi);
2284 client_obd_list_unlock(&cli->cl_loi_list_lock);
2287 cl_object_put(env, clob);
2289 if (page_count == 0) {
2290 client_obd_list_lock(&cli->cl_loi_list_lock);
2294 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2296 LASSERT(list_empty(&rpc_list));
2297 loi_list_maint(cli, loi);
2298 RETURN(PTR_ERR(req));
2301 aa = ptlrpc_req_async_args(req);
2303 if (cmd == OBD_BRW_READ) {
2304 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2305 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2306 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2307 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2309 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2310 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2311 cli->cl_w_in_flight);
2312 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2313 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2315 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2317 client_obd_list_lock(&cli->cl_loi_list_lock);
2319 if (cmd == OBD_BRW_READ)
2320 cli->cl_r_in_flight++;
2322 cli->cl_w_in_flight++;
2324 /* queued sync pages can be torn down while the pages
2325 * were between the pending list and the rpc */
2327 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2328 /* only one oap gets a request reference */
2331 if (oap->oap_interrupted && !req->rq_intr) {
2332 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2334 ptlrpc_mark_interrupted(req);
2338 tmp->oap_request = ptlrpc_request_addref(req);
2340 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2341 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2343 req->rq_interpret_reply = brw_interpret;
2344 ptlrpcd_add_req(req, PSCOPE_BRW);
2348 #define LOI_DEBUG(LOI, STR, args...) \
2349 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2350 !list_empty(&(LOI)->loi_cli_item), \
2351 (LOI)->loi_write_lop.lop_num_pending, \
2352 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2353 (LOI)->loi_read_lop.lop_num_pending, \
2354 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2357 /* This is called by osc_check_rpcs() to find which objects have pages that
2358 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2359 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2362 /* first return all objects which we already know to have
2363 * pages ready to be stuffed into rpcs */
2364 if (!list_empty(&cli->cl_loi_ready_list))
2365 RETURN(list_entry(cli->cl_loi_ready_list.next,
2366 struct lov_oinfo, loi_cli_item));
2368 /* then if we have cache waiters, return all objects with queued
2369 * writes. This is especially important when many small files
2370 * have filled up the cache and not been fired into rpcs because
2371 * they don't pass the nr_pending/object threshhold */
2372 if (!list_empty(&cli->cl_cache_waiters) &&
2373 !list_empty(&cli->cl_loi_write_list))
2374 RETURN(list_entry(cli->cl_loi_write_list.next,
2375 struct lov_oinfo, loi_write_item));
2377 /* then return all queued objects when we have an invalid import
2378 * so that they get flushed */
2379 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2380 if (!list_empty(&cli->cl_loi_write_list))
2381 RETURN(list_entry(cli->cl_loi_write_list.next,
2382 struct lov_oinfo, loi_write_item));
2383 if (!list_empty(&cli->cl_loi_read_list))
2384 RETURN(list_entry(cli->cl_loi_read_list.next,
2385 struct lov_oinfo, loi_read_item));
2390 /* called with the loi list lock held */
2391 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2393 struct lov_oinfo *loi;
2394 int rc = 0, race_counter = 0;
2397 while ((loi = osc_next_loi(cli)) != NULL) {
2398 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2400 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2403 /* attempt some read/write balancing by alternating between
2404 * reads and writes in an object. The makes_rpc checks here
2405 * would be redundant if we were getting read/write work items
2406 * instead of objects. we don't want send_oap_rpc to drain a
2407 * partial read pending queue when we're given this object to
2408 * do io on writes while there are cache waiters */
2409 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2410 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2411 &loi->loi_write_lop);
2419 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2420 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2421 &loi->loi_read_lop);
2430 /* attempt some inter-object balancing by issueing rpcs
2431 * for each object in turn */
2432 if (!list_empty(&loi->loi_cli_item))
2433 list_del_init(&loi->loi_cli_item);
2434 if (!list_empty(&loi->loi_write_item))
2435 list_del_init(&loi->loi_write_item);
2436 if (!list_empty(&loi->loi_read_item))
2437 list_del_init(&loi->loi_read_item);
2439 loi_list_maint(cli, loi);
2441 /* send_oap_rpc fails with 0 when make_ready tells it to
2442 * back off. llite's make_ready does this when it tries
2443 * to lock a page queued for write that is already locked.
2444 * we want to try sending rpcs from many objects, but we
2445 * don't want to spin failing with 0. */
2446 if (race_counter == 10)
2452 /* we're trying to queue a page in the osc so we're subject to the
2453 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2454 * If the osc's queued pages are already at that limit, then we want to sleep
2455 * until there is space in the osc's queue for us. We also may be waiting for
2456 * write credits from the OST if there are RPCs in flight that may return some
2457 * before we fall back to sync writes.
2459 * We need this know our allocation was granted in the presence of signals */
2460 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2464 client_obd_list_lock(&cli->cl_loi_list_lock);
2465 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2466 client_obd_list_unlock(&cli->cl_loi_list_lock);
2471 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2474 int osc_enter_cache_try(const struct lu_env *env,
2475 struct client_obd *cli, struct lov_oinfo *loi,
2476 struct osc_async_page *oap, int transient)
2480 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2482 osc_consume_write_grant(cli, &oap->oap_brw_page);
2484 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2485 atomic_inc(&obd_dirty_transit_pages);
2486 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2492 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2493 * grant or cache space. */
2494 static int osc_enter_cache(const struct lu_env *env,
2495 struct client_obd *cli, struct lov_oinfo *loi,
2496 struct osc_async_page *oap)
2498 struct osc_cache_waiter ocw;
2499 struct l_wait_info lwi = { 0 };
2503 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2504 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2505 cli->cl_dirty_max, obd_max_dirty_pages,
2506 cli->cl_lost_grant, cli->cl_avail_grant);
2508 /* force the caller to try sync io. this can jump the list
2509 * of queued writes and create a discontiguous rpc stream */
2510 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2511 loi->loi_ar.ar_force_sync)
2514 /* Hopefully normal case - cache space and write credits available */
2515 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2516 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2517 osc_enter_cache_try(env, cli, loi, oap, 0))
2520 /* Make sure that there are write rpcs in flight to wait for. This
2521 * is a little silly as this object may not have any pending but
2522 * other objects sure might. */
2523 if (cli->cl_w_in_flight) {
2524 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2525 cfs_waitq_init(&ocw.ocw_waitq);
2529 loi_list_maint(cli, loi);
2530 osc_check_rpcs(env, cli);
2531 client_obd_list_unlock(&cli->cl_loi_list_lock);
2533 CDEBUG(D_CACHE, "sleeping for cache space\n");
2534 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2536 client_obd_list_lock(&cli->cl_loi_list_lock);
2537 if (!list_empty(&ocw.ocw_entry)) {
2538 list_del(&ocw.ocw_entry);
2548 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2549 struct lov_oinfo *loi, cfs_page_t *page,
2550 obd_off offset, const struct obd_async_page_ops *ops,
2551 void *data, void **res, int nocache,
2552 struct lustre_handle *lockh)
2554 struct osc_async_page *oap;
2559 return size_round(sizeof(*oap));
2562 oap->oap_magic = OAP_MAGIC;
2563 oap->oap_cli = &exp->exp_obd->u.cli;
2566 oap->oap_caller_ops = ops;
2567 oap->oap_caller_data = data;
2569 oap->oap_page = page;
2570 oap->oap_obj_off = offset;
2571 if (!client_is_remote(exp) &&
2572 cfs_capable(CFS_CAP_SYS_RESOURCE))
2573 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2575 LASSERT(!(offset & ~CFS_PAGE_MASK));
2577 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2578 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2579 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2580 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2582 spin_lock_init(&oap->oap_lock);
2583 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2587 struct osc_async_page *oap_from_cookie(void *cookie)
2589 struct osc_async_page *oap = cookie;
2590 if (oap->oap_magic != OAP_MAGIC)
2591 return ERR_PTR(-EINVAL);
2595 int osc_queue_async_io(const struct lu_env *env,
2596 struct obd_export *exp, struct lov_stripe_md *lsm,
2597 struct lov_oinfo *loi, void *cookie,
2598 int cmd, obd_off off, int count,
2599 obd_flag brw_flags, enum async_flags async_flags)
2601 struct client_obd *cli = &exp->exp_obd->u.cli;
2602 struct osc_async_page *oap;
2606 oap = oap_from_cookie(cookie);
2608 RETURN(PTR_ERR(oap));
2610 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2613 if (!list_empty(&oap->oap_pending_item) ||
2614 !list_empty(&oap->oap_urgent_item) ||
2615 !list_empty(&oap->oap_rpc_item))
2618 /* check if the file's owner/group is over quota */
2619 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2620 struct cl_object *obj;
2621 struct cl_attr attr; /* XXX put attr into thread info */
2623 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2625 cl_object_attr_lock(obj);
2626 rc = cl_object_attr_get(env, obj, &attr);
2627 cl_object_attr_unlock(obj);
2629 if (rc == 0 && lquota_chkdq(quota_interface, cli, attr.cat_uid,
2630 attr.cat_gid) == NO_QUOTA)
2637 loi = lsm->lsm_oinfo[0];
2639 client_obd_list_lock(&cli->cl_loi_list_lock);
2641 LASSERT(off + count <= CFS_PAGE_SIZE);
2643 oap->oap_page_off = off;
2644 oap->oap_count = count;
2645 oap->oap_brw_flags = brw_flags;
2646 oap->oap_async_flags = async_flags;
2648 if (cmd & OBD_BRW_WRITE) {
2649 rc = osc_enter_cache(env, cli, loi, oap);
2651 client_obd_list_unlock(&cli->cl_loi_list_lock);
2656 osc_oap_to_pending(oap);
2657 loi_list_maint(cli, loi);
2659 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2662 osc_check_rpcs(env, cli);
2663 client_obd_list_unlock(&cli->cl_loi_list_lock);
2668 /* aka (~was & now & flag), but this is more clear :) */
2669 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2671 int osc_set_async_flags_base(struct client_obd *cli,
2672 struct lov_oinfo *loi, struct osc_async_page *oap,
2673 obd_flag async_flags)
2675 struct loi_oap_pages *lop;
2678 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2681 if (oap->oap_cmd & OBD_BRW_WRITE) {
2682 lop = &loi->loi_write_lop;
2684 lop = &loi->loi_read_lop;
2687 if (list_empty(&oap->oap_pending_item))
2690 if ((oap->oap_async_flags & async_flags) == async_flags)
2693 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2694 oap->oap_async_flags |= ASYNC_READY;
2696 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2697 if (list_empty(&oap->oap_rpc_item)) {
2698 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2699 loi_list_maint(cli, loi);
2703 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2704 oap->oap_async_flags);
2708 int osc_teardown_async_page(struct obd_export *exp,
2709 struct lov_stripe_md *lsm,
2710 struct lov_oinfo *loi, void *cookie)
2712 struct client_obd *cli = &exp->exp_obd->u.cli;
2713 struct loi_oap_pages *lop;
2714 struct osc_async_page *oap;
2718 oap = oap_from_cookie(cookie);
2720 RETURN(PTR_ERR(oap));
2723 loi = lsm->lsm_oinfo[0];
2725 if (oap->oap_cmd & OBD_BRW_WRITE) {
2726 lop = &loi->loi_write_lop;
2728 lop = &loi->loi_read_lop;
2731 client_obd_list_lock(&cli->cl_loi_list_lock);
2733 if (!list_empty(&oap->oap_rpc_item))
2734 GOTO(out, rc = -EBUSY);
2736 osc_exit_cache(cli, oap, 0);
2737 osc_wake_cache_waiters(cli);
2739 if (!list_empty(&oap->oap_urgent_item)) {
2740 list_del_init(&oap->oap_urgent_item);
2741 oap->oap_async_flags &= ~ASYNC_URGENT;
2743 if (!list_empty(&oap->oap_pending_item)) {
2744 list_del_init(&oap->oap_pending_item);
2745 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2747 loi_list_maint(cli, loi);
2748 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2750 client_obd_list_unlock(&cli->cl_loi_list_lock);
2754 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
2755 struct ldlm_enqueue_info *einfo,
2758 void *data = einfo->ei_cbdata;
2760 LASSERT(lock != NULL);
2761 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2762 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2763 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2764 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2766 lock_res_and_lock(lock);
2767 spin_lock(&osc_ast_guard);
2768 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
2769 lock->l_ast_data = data;
2770 spin_unlock(&osc_ast_guard);
2771 unlock_res_and_lock(lock);
2774 static void osc_set_data_with_check(struct lustre_handle *lockh,
2775 struct ldlm_enqueue_info *einfo,
2778 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2781 osc_set_lock_data_with_check(lock, einfo, flags);
2782 LDLM_LOCK_PUT(lock);
2784 CERROR("lockh %p, data %p - client evicted?\n",
2785 lockh, einfo->ei_cbdata);
2788 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2789 ldlm_iterator_t replace, void *data)
2791 struct ldlm_res_id res_id;
2792 struct obd_device *obd = class_exp2obd(exp);
2794 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
2795 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2799 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2800 obd_enqueue_update_f upcall, void *cookie,
2803 int intent = *flags & LDLM_FL_HAS_INTENT;
2807 /* The request was created before ldlm_cli_enqueue call. */
2808 if (rc == ELDLM_LOCK_ABORTED) {
2809 struct ldlm_reply *rep;
2810 rep = req_capsule_server_get(&req->rq_pill,
2813 LASSERT(rep != NULL);
2814 if (rep->lock_policy_res1)
2815 rc = rep->lock_policy_res1;
2819 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
2820 *flags |= LDLM_FL_LVB_READY;
2821 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2822 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2825 /* Call the update callback. */
2826 rc = (*upcall)(cookie, rc);
2830 static int osc_enqueue_interpret(const struct lu_env *env,
2831 struct ptlrpc_request *req,
2832 struct osc_enqueue_args *aa, int rc)
2834 struct ldlm_lock *lock;
2835 struct lustre_handle handle;
2838 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2839 * might be freed anytime after lock upcall has been called. */
2840 lustre_handle_copy(&handle, aa->oa_lockh);
2841 mode = aa->oa_ei->ei_mode;
2843 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2845 lock = ldlm_handle2lock(&handle);
2847 /* Take an additional reference so that a blocking AST that
2848 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2849 * to arrive after an upcall has been executed by
2850 * osc_enqueue_fini(). */
2851 ldlm_lock_addref(&handle, mode);
2853 /* Complete obtaining the lock procedure. */
2854 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2855 mode, aa->oa_flags, aa->oa_lvb,
2856 sizeof(*aa->oa_lvb), lustre_swab_ost_lvb,
2858 /* Complete osc stuff. */
2859 rc = osc_enqueue_fini(req, aa->oa_lvb,
2860 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
2861 /* Release the lock for async request. */
2862 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2864 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2865 * not already released by
2866 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2868 ldlm_lock_decref(&handle, mode);
2870 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2871 aa->oa_lockh, req, aa);
2872 ldlm_lock_decref(&handle, mode);
2873 LDLM_LOCK_PUT(lock);
2877 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2878 struct lov_oinfo *loi, int flags,
2879 struct ost_lvb *lvb, __u32 mode, int rc)
2881 if (rc == ELDLM_OK) {
2882 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2885 LASSERT(lock != NULL);
2886 loi->loi_lvb = *lvb;
2887 tmp = loi->loi_lvb.lvb_size;
2888 /* Extend KMS up to the end of this lock and no further
2889 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2890 if (tmp > lock->l_policy_data.l_extent.end)
2891 tmp = lock->l_policy_data.l_extent.end + 1;
2892 if (tmp >= loi->loi_kms) {
2893 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2894 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2895 loi_kms_set(loi, tmp);
2897 LDLM_DEBUG(lock, "lock acquired, setting rss="
2898 LPU64"; leaving kms="LPU64", end="LPU64,
2899 loi->loi_lvb.lvb_size, loi->loi_kms,
2900 lock->l_policy_data.l_extent.end);
2902 ldlm_lock_allow_match(lock);
2903 LDLM_LOCK_PUT(lock);
2904 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2905 loi->loi_lvb = *lvb;
2906 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2907 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2911 EXPORT_SYMBOL(osc_update_enqueue);
2913 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2915 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2916 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2917 * other synchronous requests, however keeping some locks and trying to obtain
2918 * others may take a considerable amount of time in a case of ost failure; and
2919 * when other sync requests do not get released lock from a client, the client
2920 * is excluded from the cluster -- such scenarious make the life difficult, so
2921 * release locks just after they are obtained. */
2922 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2923 int *flags, ldlm_policy_data_t *policy,
2924 struct ost_lvb *lvb, int kms_valid,
2925 obd_enqueue_update_f upcall, void *cookie,
2926 struct ldlm_enqueue_info *einfo,
2927 struct lustre_handle *lockh,
2928 struct ptlrpc_request_set *rqset, int async)
2930 struct obd_device *obd = exp->exp_obd;
2931 struct ptlrpc_request *req = NULL;
2932 int intent = *flags & LDLM_FL_HAS_INTENT;
2937 /* Filesystem lock extents are extended to page boundaries so that
2938 * dealing with the page cache is a little smoother. */
2939 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2940 policy->l_extent.end |= ~CFS_PAGE_MASK;
2943 * kms is not valid when either object is completely fresh (so that no
2944 * locks are cached), or object was evicted. In the latter case cached
2945 * lock cannot be used, because it would prime inode state with
2946 * potentially stale LVB.
2951 /* Next, search for already existing extent locks that will cover us */
2952 /* If we're trying to read, we also search for an existing PW lock. The
2953 * VFS and page cache already protect us locally, so lots of readers/
2954 * writers can share a single PW lock.
2956 * There are problems with conversion deadlocks, so instead of
2957 * converting a read lock to a write lock, we'll just enqueue a new
2960 * At some point we should cancel the read lock instead of making them
2961 * send us a blocking callback, but there are problems with canceling
2962 * locks out from other users right now, too. */
2963 mode = einfo->ei_mode;
2964 if (einfo->ei_mode == LCK_PR)
2966 mode = ldlm_lock_match(obd->obd_namespace,
2967 *flags | LDLM_FL_LVB_READY, res_id,
2968 einfo->ei_type, policy, mode, lockh, 0);
2970 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2972 if (matched->l_ast_data == NULL ||
2973 matched->l_ast_data == einfo->ei_cbdata) {
2974 /* addref the lock only if not async requests and PW
2975 * lock is matched whereas we asked for PR. */
2976 if (!rqset && einfo->ei_mode != mode)
2977 ldlm_lock_addref(lockh, LCK_PR);
2978 osc_set_lock_data_with_check(matched, einfo, *flags);
2980 /* I would like to be able to ASSERT here that
2981 * rss <= kms, but I can't, for reasons which
2982 * are explained in lov_enqueue() */
2985 /* We already have a lock, and it's referenced */
2986 (*upcall)(cookie, ELDLM_OK);
2988 /* For async requests, decref the lock. */
2989 if (einfo->ei_mode != mode)
2990 ldlm_lock_decref(lockh, LCK_PW);
2992 ldlm_lock_decref(lockh, einfo->ei_mode);
2993 LDLM_LOCK_PUT(matched);
2996 ldlm_lock_decref(lockh, mode);
2997 LDLM_LOCK_PUT(matched);
3002 CFS_LIST_HEAD(cancels);
3003 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3004 &RQF_LDLM_ENQUEUE_LVB);
3008 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3012 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3014 ptlrpc_request_set_replen(req);
3017 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3018 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3020 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3021 sizeof(*lvb), lustre_swab_ost_lvb, lockh, async);
3024 struct osc_enqueue_args *aa;
3025 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3026 aa = ptlrpc_req_async_args(req);
3029 aa->oa_flags = flags;
3030 aa->oa_upcall = upcall;
3031 aa->oa_cookie = cookie;
3033 aa->oa_lockh = lockh;
3035 req->rq_interpret_reply =
3036 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3037 if (rqset == PTLRPCD_SET)
3038 ptlrpcd_add_req(req, PSCOPE_OTHER);
3040 ptlrpc_set_add_req(rqset, req);
3041 } else if (intent) {
3042 ptlrpc_req_finished(req);
3047 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3049 ptlrpc_req_finished(req);
3054 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3055 struct ldlm_enqueue_info *einfo,
3056 struct ptlrpc_request_set *rqset)
3058 struct ldlm_res_id res_id;
3062 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3063 oinfo->oi_md->lsm_object_gr, &res_id);
3065 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3066 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3067 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3068 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3069 rqset, rqset != NULL);
3073 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3074 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3075 int *flags, void *data, struct lustre_handle *lockh,
3078 struct obd_device *obd = exp->exp_obd;
3079 int lflags = *flags;
3083 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3086 /* Filesystem lock extents are extended to page boundaries so that
3087 * dealing with the page cache is a little smoother */
3088 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3089 policy->l_extent.end |= ~CFS_PAGE_MASK;
3091 /* Next, search for already existing extent locks that will cover us */
3092 /* If we're trying to read, we also search for an existing PW lock. The
3093 * VFS and page cache already protect us locally, so lots of readers/
3094 * writers can share a single PW lock. */
3098 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3099 res_id, type, policy, rc, lockh, unref);
3102 osc_set_data_with_check(lockh, data, lflags);
3103 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3104 ldlm_lock_addref(lockh, LCK_PR);
3105 ldlm_lock_decref(lockh, LCK_PW);
3112 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3116 if (unlikely(mode == LCK_GROUP))
3117 ldlm_lock_decref_and_cancel(lockh, mode);
3119 ldlm_lock_decref(lockh, mode);
3124 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3125 __u32 mode, struct lustre_handle *lockh)
3128 RETURN(osc_cancel_base(lockh, mode));
3131 static int osc_cancel_unused(struct obd_export *exp,
3132 struct lov_stripe_md *lsm, int flags,
3135 struct obd_device *obd = class_exp2obd(exp);
3136 struct ldlm_res_id res_id, *resp = NULL;
3139 resp = osc_build_res_name(lsm->lsm_object_id,
3140 lsm->lsm_object_gr, &res_id);
3143 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3146 static int osc_statfs_interpret(const struct lu_env *env,
3147 struct ptlrpc_request *req,
3148 struct osc_async_args *aa, int rc)
3150 struct obd_statfs *msfs;
3156 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3158 GOTO(out, rc = -EPROTO);
3161 *aa->aa_oi->oi_osfs = *msfs;
3163 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3167 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3168 __u64 max_age, struct ptlrpc_request_set *rqset)
3170 struct ptlrpc_request *req;
3171 struct osc_async_args *aa;
3175 /* We could possibly pass max_age in the request (as an absolute
3176 * timestamp or a "seconds.usec ago") so the target can avoid doing
3177 * extra calls into the filesystem if that isn't necessary (e.g.
3178 * during mount that would help a bit). Having relative timestamps
3179 * is not so great if request processing is slow, while absolute
3180 * timestamps are not ideal because they need time synchronization. */
3181 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3185 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3187 ptlrpc_request_free(req);
3190 ptlrpc_request_set_replen(req);
3191 req->rq_request_portal = OST_CREATE_PORTAL;
3192 ptlrpc_at_set_req_timeout(req);
3194 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3195 /* procfs requests not want stat in wait for avoid deadlock */
3196 req->rq_no_resend = 1;
3197 req->rq_no_delay = 1;
3200 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3201 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3202 aa = ptlrpc_req_async_args(req);
3205 ptlrpc_set_add_req(rqset, req);
3209 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3210 __u64 max_age, __u32 flags)
3212 struct obd_statfs *msfs;
3213 struct ptlrpc_request *req;
3214 struct obd_import *imp = NULL;
3218 /*Since the request might also come from lprocfs, so we need
3219 *sync this with client_disconnect_export Bug15684*/
3220 down_read(&obd->u.cli.cl_sem);
3221 if (obd->u.cli.cl_import)
3222 imp = class_import_get(obd->u.cli.cl_import);
3223 up_read(&obd->u.cli.cl_sem);
3227 /* We could possibly pass max_age in the request (as an absolute
3228 * timestamp or a "seconds.usec ago") so the target can avoid doing
3229 * extra calls into the filesystem if that isn't necessary (e.g.
3230 * during mount that would help a bit). Having relative timestamps
3231 * is not so great if request processing is slow, while absolute
3232 * timestamps are not ideal because they need time synchronization. */
3233 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3235 class_import_put(imp);
3240 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3242 ptlrpc_request_free(req);
3245 ptlrpc_request_set_replen(req);
3246 req->rq_request_portal = OST_CREATE_PORTAL;
3247 ptlrpc_at_set_req_timeout(req);
3249 if (flags & OBD_STATFS_NODELAY) {
3250 /* procfs requests not want stat in wait for avoid deadlock */
3251 req->rq_no_resend = 1;
3252 req->rq_no_delay = 1;
3255 rc = ptlrpc_queue_wait(req);
3259 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3261 GOTO(out, rc = -EPROTO);
3268 ptlrpc_req_finished(req);
3272 /* Retrieve object striping information.
3274 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3275 * the maximum number of OST indices which will fit in the user buffer.
3276 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3278 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3280 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3281 struct lov_user_md_v3 lum, *lumk;
3282 struct lov_user_ost_data_v1 *lmm_objects;
3283 int rc = 0, lum_size;
3289 /* we only need the header part from user space to get lmm_magic and
3290 * lmm_stripe_count, (the header part is common to v1 and v3) */
3291 lum_size = sizeof(struct lov_user_md_v1);
3292 if (copy_from_user(&lum, lump, lum_size))
3295 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3296 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3299 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3300 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3301 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3302 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3304 /* we can use lov_mds_md_size() to compute lum_size
3305 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3306 if (lum.lmm_stripe_count > 0) {
3307 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3308 OBD_ALLOC(lumk, lum_size);
3312 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3313 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3315 lmm_objects = &(lumk->lmm_objects[0]);
3316 lmm_objects->l_object_id = lsm->lsm_object_id;
3318 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3322 lumk->lmm_object_id = lsm->lsm_object_id;
3323 lumk->lmm_object_gr = lsm->lsm_object_gr;
3324 lumk->lmm_stripe_count = 1;
3326 if (copy_to_user(lump, lumk, lum_size))
3330 OBD_FREE(lumk, lum_size);
3336 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3337 void *karg, void *uarg)
3339 struct obd_device *obd = exp->exp_obd;
3340 struct obd_ioctl_data *data = karg;
3344 if (!try_module_get(THIS_MODULE)) {
3345 CERROR("Can't get module. Is it alive?");
3349 case OBD_IOC_LOV_GET_CONFIG: {
3351 struct lov_desc *desc;
3352 struct obd_uuid uuid;
3356 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3357 GOTO(out, err = -EINVAL);
3359 data = (struct obd_ioctl_data *)buf;
3361 if (sizeof(*desc) > data->ioc_inllen1) {
3362 obd_ioctl_freedata(buf, len);
3363 GOTO(out, err = -EINVAL);
3366 if (data->ioc_inllen2 < sizeof(uuid)) {
3367 obd_ioctl_freedata(buf, len);
3368 GOTO(out, err = -EINVAL);
3371 desc = (struct lov_desc *)data->ioc_inlbuf1;
3372 desc->ld_tgt_count = 1;
3373 desc->ld_active_tgt_count = 1;
3374 desc->ld_default_stripe_count = 1;
3375 desc->ld_default_stripe_size = 0;
3376 desc->ld_default_stripe_offset = 0;
3377 desc->ld_pattern = 0;
3378 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3380 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3382 err = copy_to_user((void *)uarg, buf, len);
3385 obd_ioctl_freedata(buf, len);
3388 case LL_IOC_LOV_SETSTRIPE:
3389 err = obd_alloc_memmd(exp, karg);
3393 case LL_IOC_LOV_GETSTRIPE:
3394 err = osc_getstripe(karg, uarg);
3396 case OBD_IOC_CLIENT_RECOVER:
3397 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3402 case IOC_OSC_SET_ACTIVE:
3403 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3406 case OBD_IOC_POLL_QUOTACHECK:
3407 err = lquota_poll_check(quota_interface, exp,
3408 (struct if_quotacheck *)karg);
3411 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3412 cmd, cfs_curproc_comm());
3413 GOTO(out, err = -ENOTTY);
3416 module_put(THIS_MODULE);
3420 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3421 void *key, __u32 *vallen, void *val,
3422 struct lov_stripe_md *lsm)
3425 if (!vallen || !val)
3428 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3429 __u32 *stripe = val;
3430 *vallen = sizeof(*stripe);
3433 } else if (KEY_IS(KEY_LAST_ID)) {
3434 struct ptlrpc_request *req;
3439 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3440 &RQF_OST_GET_INFO_LAST_ID);
3444 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3445 RCL_CLIENT, keylen);
3446 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3448 ptlrpc_request_free(req);
3452 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3453 memcpy(tmp, key, keylen);
3455 ptlrpc_request_set_replen(req);
3456 rc = ptlrpc_queue_wait(req);
3460 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3462 GOTO(out, rc = -EPROTO);
3464 *((obd_id *)val) = *reply;
3466 ptlrpc_req_finished(req);
3468 } else if (KEY_IS(KEY_FIEMAP)) {
3469 struct ptlrpc_request *req;
3470 struct ll_user_fiemap *reply;
3474 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3475 &RQF_OST_GET_INFO_FIEMAP);
3479 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3480 RCL_CLIENT, keylen);
3481 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3482 RCL_CLIENT, *vallen);
3483 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3484 RCL_SERVER, *vallen);
3486 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3488 ptlrpc_request_free(req);
3492 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3493 memcpy(tmp, key, keylen);
3494 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3495 memcpy(tmp, val, *vallen);
3497 ptlrpc_request_set_replen(req);
3498 rc = ptlrpc_queue_wait(req);
3502 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3504 GOTO(out1, rc = -EPROTO);
3506 memcpy(val, reply, *vallen);
3508 ptlrpc_req_finished(req);
3516 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3517 struct ptlrpc_request *req,
3520 struct llog_ctxt *ctxt;
3521 struct obd_import *imp = req->rq_import;
3527 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3530 rc = llog_initiator_connect(ctxt);
3532 CERROR("cannot establish connection for "
3533 "ctxt %p: %d\n", ctxt, rc);
3536 llog_ctxt_put(ctxt);
3537 spin_lock(&imp->imp_lock);
3538 imp->imp_server_timeout = 1;
3539 imp->imp_pingable = 1;
3540 spin_unlock(&imp->imp_lock);
3541 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3546 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3547 void *key, obd_count vallen, void *val,
3548 struct ptlrpc_request_set *set)
3550 struct ptlrpc_request *req;
3551 struct obd_device *obd = exp->exp_obd;
3552 struct obd_import *imp = class_exp2cliimp(exp);
3557 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3559 if (KEY_IS(KEY_NEXT_ID)) {
3560 if (vallen != sizeof(obd_id))
3564 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3565 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3566 exp->exp_obd->obd_name,
3567 obd->u.cli.cl_oscc.oscc_next_id);
3572 if (KEY_IS(KEY_UNLINKED)) {
3573 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3574 spin_lock(&oscc->oscc_lock);
3575 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3576 spin_unlock(&oscc->oscc_lock);
3580 if (KEY_IS(KEY_INIT_RECOV)) {
3581 if (vallen != sizeof(int))
3583 spin_lock(&imp->imp_lock);
3584 imp->imp_initial_recov = *(int *)val;
3585 spin_unlock(&imp->imp_lock);
3586 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3587 exp->exp_obd->obd_name,
3588 imp->imp_initial_recov);
3592 if (KEY_IS(KEY_CHECKSUM)) {
3593 if (vallen != sizeof(int))
3595 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3599 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3600 sptlrpc_conf_client_adapt(obd);
3604 if (KEY_IS(KEY_FLUSH_CTX)) {
3605 sptlrpc_import_flush_my_ctx(imp);
3612 /* We pass all other commands directly to OST. Since nobody calls osc
3613 methods directly and everybody is supposed to go through LOV, we
3614 assume lov checked invalid values for us.
3615 The only recognised values so far are evict_by_nid and mds_conn.
3616 Even if something bad goes through, we'd get a -EINVAL from OST
3620 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3624 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3625 RCL_CLIENT, keylen);
3626 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3627 RCL_CLIENT, vallen);
3628 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3630 ptlrpc_request_free(req);
3634 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3635 memcpy(tmp, key, keylen);
3636 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3637 memcpy(tmp, val, vallen);
3639 if (KEY_IS(KEY_MDS_CONN)) {
3640 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3642 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3643 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3644 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
3645 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3648 ptlrpc_request_set_replen(req);
3649 ptlrpc_set_add_req(set, req);
3650 ptlrpc_check_set(NULL, set);
3656 static struct llog_operations osc_size_repl_logops = {
3657 lop_cancel: llog_obd_repl_cancel
3660 static struct llog_operations osc_mds_ost_orig_logops;
3661 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3662 struct obd_device *tgt, int count,
3663 struct llog_catid *catid, struct obd_uuid *uuid)
3668 LASSERT(olg == &obd->obd_olg);
3669 spin_lock(&obd->obd_dev_lock);
3670 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3671 osc_mds_ost_orig_logops = llog_lvfs_ops;
3672 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3673 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3674 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3675 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3677 spin_unlock(&obd->obd_dev_lock);
3679 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3680 &catid->lci_logid, &osc_mds_ost_orig_logops);
3682 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3686 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3687 NULL, &osc_size_repl_logops);
3689 struct llog_ctxt *ctxt =
3690 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3693 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3698 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3699 obd->obd_name, tgt->obd_name, count, catid, rc);
3700 CERROR("logid "LPX64":0x%x\n",
3701 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3706 static int osc_llog_finish(struct obd_device *obd, int count)
3708 struct llog_ctxt *ctxt;
3709 int rc = 0, rc2 = 0;
3712 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3714 rc = llog_cleanup(ctxt);
3716 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3718 rc2 = llog_cleanup(ctxt);
3725 static int osc_reconnect(const struct lu_env *env,
3726 struct obd_export *exp, struct obd_device *obd,
3727 struct obd_uuid *cluuid,
3728 struct obd_connect_data *data,
3731 struct client_obd *cli = &obd->u.cli;
3733 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3736 client_obd_list_lock(&cli->cl_loi_list_lock);
3737 data->ocd_grant = cli->cl_avail_grant ?:
3738 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3739 lost_grant = cli->cl_lost_grant;
3740 cli->cl_lost_grant = 0;
3741 client_obd_list_unlock(&cli->cl_loi_list_lock);
3743 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3744 "cl_lost_grant: %ld\n", data->ocd_grant,
3745 cli->cl_avail_grant, lost_grant);
3746 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3747 " ocd_grant: %d\n", data->ocd_connect_flags,
3748 data->ocd_version, data->ocd_grant);
3754 static int osc_disconnect(struct obd_export *exp)
3756 struct obd_device *obd = class_exp2obd(exp);
3757 struct llog_ctxt *ctxt;
3760 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3762 if (obd->u.cli.cl_conn_count == 1) {
3763 /* Flush any remaining cancel messages out to the
3765 llog_sync(ctxt, exp);
3767 llog_ctxt_put(ctxt);
3769 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3773 rc = client_disconnect_export(exp);
3777 static int osc_import_event(struct obd_device *obd,
3778 struct obd_import *imp,
3779 enum obd_import_event event)
3781 struct client_obd *cli;
3785 LASSERT(imp->imp_obd == obd);
3788 case IMP_EVENT_DISCON: {
3789 /* Only do this on the MDS OSC's */
3790 if (imp->imp_server_timeout) {
3791 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3793 spin_lock(&oscc->oscc_lock);
3794 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3795 spin_unlock(&oscc->oscc_lock);
3798 client_obd_list_lock(&cli->cl_loi_list_lock);
3799 cli->cl_avail_grant = 0;
3800 cli->cl_lost_grant = 0;
3801 client_obd_list_unlock(&cli->cl_loi_list_lock);
3804 case IMP_EVENT_INACTIVE: {
3805 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3808 case IMP_EVENT_INVALIDATE: {
3809 struct ldlm_namespace *ns = obd->obd_namespace;
3813 env = cl_env_get(&refcheck);
3817 client_obd_list_lock(&cli->cl_loi_list_lock);
3818 /* all pages go to failing rpcs due to the invalid
3820 osc_check_rpcs(env, cli);
3821 client_obd_list_unlock(&cli->cl_loi_list_lock);
3823 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3824 cl_env_put(env, &refcheck);
3829 case IMP_EVENT_ACTIVE: {
3830 /* Only do this on the MDS OSC's */
3831 if (imp->imp_server_timeout) {
3832 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3834 spin_lock(&oscc->oscc_lock);
3835 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3836 spin_unlock(&oscc->oscc_lock);
3838 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3841 case IMP_EVENT_OCD: {
3842 struct obd_connect_data *ocd = &imp->imp_connect_data;
3844 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3845 osc_init_grant(&obd->u.cli, ocd);
3848 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3849 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3851 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3855 CERROR("Unknown import event %d\n", event);
3861 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3867 rc = ptlrpcd_addref();
3871 rc = client_obd_setup(obd, lcfg);
3875 struct lprocfs_static_vars lvars = { 0 };
3876 struct client_obd *cli = &obd->u.cli;
3878 lprocfs_osc_init_vars(&lvars);
3879 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3880 lproc_osc_attach_seqstat(obd);
3881 sptlrpc_lprocfs_cliobd_attach(obd);
3882 ptlrpc_lprocfs_register_obd(obd);
3886 /* We need to allocate a few requests more, because
3887 brw_interpret tries to create new requests before freeing
3888 previous ones. Ideally we want to have 2x max_rpcs_in_flight
3889 reserved, but I afraid that might be too much wasted RAM
3890 in fact, so 2 is just my guess and still should work. */
3891 cli->cl_import->imp_rq_pool =
3892 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3894 ptlrpc_add_rqs_to_pool);
3900 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3906 case OBD_CLEANUP_EARLY: {
3907 struct obd_import *imp;
3908 imp = obd->u.cli.cl_import;
3909 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3910 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3911 ptlrpc_deactivate_import(imp);
3912 spin_lock(&imp->imp_lock);
3913 imp->imp_pingable = 0;
3914 spin_unlock(&imp->imp_lock);
3917 case OBD_CLEANUP_EXPORTS: {
3918 /* If we set up but never connected, the
3919 client import will not have been cleaned. */
3920 if (obd->u.cli.cl_import) {
3921 struct obd_import *imp;
3922 down_write(&obd->u.cli.cl_sem);
3923 imp = obd->u.cli.cl_import;
3924 CDEBUG(D_CONFIG, "%s: client import never connected\n",
3926 ptlrpc_invalidate_import(imp);
3927 if (imp->imp_rq_pool) {
3928 ptlrpc_free_rq_pool(imp->imp_rq_pool);
3929 imp->imp_rq_pool = NULL;
3931 class_destroy_import(imp);
3932 up_write(&obd->u.cli.cl_sem);
3933 obd->u.cli.cl_import = NULL;
3935 rc = obd_llog_finish(obd, 0);
3937 CERROR("failed to cleanup llogging subsystems\n");
3944 int osc_cleanup(struct obd_device *obd)
3946 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3950 ptlrpc_lprocfs_unregister_obd(obd);
3951 lprocfs_obd_cleanup(obd);
3953 spin_lock(&oscc->oscc_lock);
3954 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3955 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3956 spin_unlock(&oscc->oscc_lock);
3958 /* free memory of osc quota cache */
3959 lquota_cleanup(quota_interface, obd);
3961 rc = client_obd_cleanup(obd);
3967 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3969 struct lprocfs_static_vars lvars = { 0 };
3972 lprocfs_osc_init_vars(&lvars);
3974 switch (lcfg->lcfg_command) {
3976 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3986 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3988 return osc_process_config_base(obd, buf);
3991 struct obd_ops osc_obd_ops = {
3992 .o_owner = THIS_MODULE,
3993 .o_setup = osc_setup,
3994 .o_precleanup = osc_precleanup,
3995 .o_cleanup = osc_cleanup,
3996 .o_add_conn = client_import_add_conn,
3997 .o_del_conn = client_import_del_conn,
3998 .o_connect = client_connect_import,
3999 .o_reconnect = osc_reconnect,
4000 .o_disconnect = osc_disconnect,
4001 .o_statfs = osc_statfs,
4002 .o_statfs_async = osc_statfs_async,
4003 .o_packmd = osc_packmd,
4004 .o_unpackmd = osc_unpackmd,
4005 .o_precreate = osc_precreate,
4006 .o_create = osc_create,
4007 .o_destroy = osc_destroy,
4008 .o_getattr = osc_getattr,
4009 .o_getattr_async = osc_getattr_async,
4010 .o_setattr = osc_setattr,
4011 .o_setattr_async = osc_setattr_async,
4013 .o_punch = osc_punch,
4015 .o_enqueue = osc_enqueue,
4016 .o_change_cbdata = osc_change_cbdata,
4017 .o_cancel = osc_cancel,
4018 .o_cancel_unused = osc_cancel_unused,
4019 .o_iocontrol = osc_iocontrol,
4020 .o_get_info = osc_get_info,
4021 .o_set_info_async = osc_set_info_async,
4022 .o_import_event = osc_import_event,
4023 .o_llog_init = osc_llog_init,
4024 .o_llog_finish = osc_llog_finish,
4025 .o_process_config = osc_process_config,
4028 extern struct lu_kmem_descr osc_caches[];
4029 extern spinlock_t osc_ast_guard;
4030 extern struct lock_class_key osc_ast_guard_class;
4032 int __init osc_init(void)
4034 struct lprocfs_static_vars lvars = { 0 };
4038 /* print an address of _any_ initialized kernel symbol from this
4039 * module, to allow debugging with gdb that doesn't support data
4040 * symbols from modules.*/
4041 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4043 rc = lu_kmem_init(osc_caches);
4045 lprocfs_osc_init_vars(&lvars);
4047 request_module("lquota");
4048 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4049 lquota_init(quota_interface);
4050 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4052 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4053 LUSTRE_OSC_NAME, &osc_device_type);
4055 if (quota_interface)
4056 PORTAL_SYMBOL_PUT(osc_quota_interface);
4057 lu_kmem_fini(osc_caches);
4061 spin_lock_init(&osc_ast_guard);
4062 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4068 static void /*__exit*/ osc_exit(void)
4070 lu_device_type_fini(&osc_device_type);
4072 lquota_exit(quota_interface);
4073 if (quota_interface)
4074 PORTAL_SYMBOL_PUT(osc_quota_interface);
4076 class_unregister_type(LUSTRE_OSC_NAME);
4077 lu_kmem_fini(osc_caches);
4080 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4081 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4082 MODULE_LICENSE("GPL");
4084 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);