1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_OSC
42 #include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71 struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76 struct lov_stripe_md *lsm)
81 lmm_size = sizeof(**lmmp);
86 OBD_FREE(*lmmp, lmm_size);
92 OBD_ALLOC(*lmmp, lmm_size);
98 LASSERT(lsm->lsm_object_id);
99 LASSERT_MDS_GROUP(lsm->lsm_object_gr);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109 struct lov_mds_md *lmm, int lmm_bytes)
115 if (lmm_bytes < sizeof (*lmm)) {
116 CERROR("lov_mds_md too small: %d, need %d\n",
117 lmm_bytes, (int)sizeof(*lmm));
120 /* XXX LOV_MAGIC etc check? */
122 if (lmm->lmm_object_id == 0) {
123 CERROR("lov_mds_md: zero lmm_object_id\n");
128 lsm_size = lov_stripe_md_size(1);
132 if (*lsmp != NULL && lmm == NULL) {
133 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134 OBD_FREE(*lsmp, lsm_size);
140 OBD_ALLOC(*lsmp, lsm_size);
143 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145 OBD_FREE(*lsmp, lsm_size);
148 loi_init((*lsmp)->lsm_oinfo[0]);
152 /* XXX zero *lsmp? */
153 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT_MDS_GROUP((*lsmp)->lsm_object_gr);
159 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165 struct ost_body *body, void *capa)
167 struct obd_capa *oc = (struct obd_capa *)capa;
168 struct lustre_capa *c;
173 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
176 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177 DEBUG_CAPA(D_SEC, c, "pack");
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181 struct obd_info *oinfo)
183 struct ost_body *body;
185 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
188 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(const struct lu_env *env,
204 struct ptlrpc_request *req,
205 struct osc_async_args *aa, int rc)
207 struct ost_body *body;
213 body = 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 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->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 lustre_get_wire_obdo(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 lustre_get_wire_obdo(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 lustre_get_wire_obdo(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);
449 lustre_set_wire_obdo(&body->oa, oa);
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);
469 lustre_get_wire_obdo(oa, &body->oa);
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 lustre_get_wire_obdo(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);
549 lustre_set_wire_obdo(&body->oa, oa);
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);
608 lustre_set_wire_obdo(&body->oa, oa);
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);
624 lustre_get_wire_obdo(oa, &body->oa);
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 ptlrpc_at_set_req_timeout(req);
733 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
734 oa->o_lcookie = *oti->oti_logcookies;
735 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
737 lustre_set_wire_obdo(&body->oa, oa);
739 osc_pack_capa(req, body, (struct obd_capa *)capa);
740 ptlrpc_request_set_replen(req);
742 /* don't throttle destroy RPCs for the MDT */
743 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
744 req->rq_interpret_reply = osc_destroy_interpret;
745 if (!osc_can_send_destroy(cli)) {
746 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
750 * Wait until the number of on-going destroy RPCs drops
751 * under max_rpc_in_flight
753 l_wait_event_exclusive(cli->cl_destroy_waitq,
754 osc_can_send_destroy(cli), &lwi);
758 /* Do not wait for response */
759 ptlrpcd_add_req(req, PSCOPE_OTHER);
763 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
766 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
768 LASSERT(!(oa->o_valid & bits));
771 client_obd_list_lock(&cli->cl_loi_list_lock);
772 oa->o_dirty = cli->cl_dirty;
773 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
774 CERROR("dirty %lu - %lu > dirty_max %lu\n",
775 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
777 } else if (atomic_read(&obd_dirty_pages) -
778 atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages){
779 CERROR("dirty %d - %d > system dirty_max %d\n",
780 atomic_read(&obd_dirty_pages),
781 atomic_read(&obd_dirty_transit_pages),
782 obd_max_dirty_pages);
784 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
785 CERROR("dirty %lu - dirty_max %lu too big???\n",
786 cli->cl_dirty, cli->cl_dirty_max);
789 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
790 (cli->cl_max_rpcs_in_flight + 1);
791 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
793 oa->o_grant = cli->cl_avail_grant;
794 oa->o_dropped = cli->cl_lost_grant;
795 cli->cl_lost_grant = 0;
796 client_obd_list_unlock(&cli->cl_loi_list_lock);
797 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
798 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
802 static void osc_update_next_shrink(struct client_obd *cli)
804 cli->cl_next_shrink_grant =
805 cfs_time_shift(cli->cl_grant_shrink_interval);
806 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
807 cli->cl_next_shrink_grant);
810 /* caller must hold loi_list_lock */
811 static void osc_consume_write_grant(struct client_obd *cli,
812 struct brw_page *pga)
814 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock);
815 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
816 atomic_inc(&obd_dirty_pages);
817 cli->cl_dirty += CFS_PAGE_SIZE;
818 cli->cl_avail_grant -= CFS_PAGE_SIZE;
819 pga->flag |= OBD_BRW_FROM_GRANT;
820 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
821 CFS_PAGE_SIZE, pga, pga->pg);
822 LASSERT(cli->cl_avail_grant >= 0);
823 osc_update_next_shrink(cli);
826 /* the companion to osc_consume_write_grant, called when a brw has completed.
827 * must be called with the loi lock held. */
828 static void osc_release_write_grant(struct client_obd *cli,
829 struct brw_page *pga, int sent)
831 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
834 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock);
835 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
840 pga->flag &= ~OBD_BRW_FROM_GRANT;
841 atomic_dec(&obd_dirty_pages);
842 cli->cl_dirty -= CFS_PAGE_SIZE;
843 if (pga->flag & OBD_BRW_NOCACHE) {
844 pga->flag &= ~OBD_BRW_NOCACHE;
845 atomic_dec(&obd_dirty_transit_pages);
846 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
849 cli->cl_lost_grant += CFS_PAGE_SIZE;
850 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
851 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
852 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
853 /* For short writes we shouldn't count parts of pages that
854 * span a whole block on the OST side, or our accounting goes
855 * wrong. Should match the code in filter_grant_check. */
856 int offset = pga->off & ~CFS_PAGE_MASK;
857 int count = pga->count + (offset & (blocksize - 1));
858 int end = (offset + pga->count) & (blocksize - 1);
860 count += blocksize - end;
862 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
863 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
864 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
865 cli->cl_avail_grant, cli->cl_dirty);
871 static unsigned long rpcs_in_flight(struct client_obd *cli)
873 return cli->cl_r_in_flight + cli->cl_w_in_flight;
876 /* caller must hold loi_list_lock */
877 void osc_wake_cache_waiters(struct client_obd *cli)
879 struct list_head *l, *tmp;
880 struct osc_cache_waiter *ocw;
883 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
884 /* if we can't dirty more, we must wait until some is written */
885 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
886 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
887 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
888 "osc max %ld, sys max %d\n", cli->cl_dirty,
889 cli->cl_dirty_max, obd_max_dirty_pages);
893 /* if still dirty cache but no grant wait for pending RPCs that
894 * may yet return us some grant before doing sync writes */
895 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
896 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
897 cli->cl_w_in_flight);
901 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
902 list_del_init(&ocw->ocw_entry);
903 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
904 /* no more RPCs in flight to return grant, do sync IO */
905 ocw->ocw_rc = -EDQUOT;
906 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
908 osc_consume_write_grant(cli,
909 &ocw->ocw_oap->oap_brw_page);
912 cfs_waitq_signal(&ocw->ocw_waitq);
918 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
920 client_obd_list_lock(&cli->cl_loi_list_lock);
921 cli->cl_avail_grant += grant;
922 client_obd_list_unlock(&cli->cl_loi_list_lock);
925 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
927 if (body->oa.o_valid & OBD_MD_FLGRANT) {
928 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
929 __osc_update_grant(cli, body->oa.o_grant);
933 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
934 void *key, obd_count vallen, void *val,
935 struct ptlrpc_request_set *set);
937 static int osc_shrink_grant_interpret(const struct lu_env *env,
938 struct ptlrpc_request *req,
941 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
942 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
943 struct ost_body *body;
946 __osc_update_grant(cli, oa->o_grant);
950 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
952 osc_update_grant(cli, body);
958 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
960 client_obd_list_lock(&cli->cl_loi_list_lock);
961 oa->o_grant = cli->cl_avail_grant / 4;
962 cli->cl_avail_grant -= oa->o_grant;
963 client_obd_list_unlock(&cli->cl_loi_list_lock);
964 oa->o_flags |= OBD_FL_SHRINK_GRANT;
965 osc_update_next_shrink(cli);
968 /* Shrink the current grant, either from some large amount to enough for a
969 * full set of in-flight RPCs, or if we have already shrunk to that limit
970 * then to enough for a single RPC. This avoids keeping more grant than
971 * needed, and avoids shrinking the grant piecemeal. */
972 static int osc_shrink_grant(struct client_obd *cli)
974 long target = (cli->cl_max_rpcs_in_flight + 1) *
975 cli->cl_max_pages_per_rpc;
977 client_obd_list_lock(&cli->cl_loi_list_lock);
978 if (cli->cl_avail_grant <= target)
979 target = cli->cl_max_pages_per_rpc;
980 client_obd_list_unlock(&cli->cl_loi_list_lock);
982 return osc_shrink_grant_to_target(cli, target);
985 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
988 struct ost_body *body;
991 client_obd_list_lock(&cli->cl_loi_list_lock);
992 /* Don't shrink if we are already above or below the desired limit
993 * We don't want to shrink below a single RPC, as that will negatively
994 * impact block allocation and long-term performance. */
995 if (target < cli->cl_max_pages_per_rpc)
996 target = cli->cl_max_pages_per_rpc;
998 if (target >= cli->cl_avail_grant) {
999 client_obd_list_unlock(&cli->cl_loi_list_lock);
1002 client_obd_list_unlock(&cli->cl_loi_list_lock);
1004 OBD_ALLOC_PTR(body);
1008 osc_announce_cached(cli, &body->oa, 0);
1010 client_obd_list_lock(&cli->cl_loi_list_lock);
1011 body->oa.o_grant = cli->cl_avail_grant - target;
1012 cli->cl_avail_grant = target;
1013 client_obd_list_unlock(&cli->cl_loi_list_lock);
1014 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1015 osc_update_next_shrink(cli);
1017 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1018 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1019 sizeof(*body), body, NULL);
1021 __osc_update_grant(cli, body->oa.o_grant);
1026 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1027 static int osc_should_shrink_grant(struct client_obd *client)
1029 cfs_time_t time = cfs_time_current();
1030 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1031 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1032 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1033 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1036 osc_update_next_shrink(client);
1041 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1043 struct client_obd *client;
1045 list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1046 if (osc_should_shrink_grant(client))
1047 osc_shrink_grant(client);
1052 static int osc_add_shrink_grant(struct client_obd *client)
1056 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1058 osc_grant_shrink_grant_cb, NULL,
1059 &client->cl_grant_shrink_list);
1061 CERROR("add grant client %s error %d\n",
1062 client->cl_import->imp_obd->obd_name, rc);
1065 CDEBUG(D_CACHE, "add grant client %s \n",
1066 client->cl_import->imp_obd->obd_name);
1067 osc_update_next_shrink(client);
1071 static int osc_del_shrink_grant(struct client_obd *client)
1073 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1077 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1079 client_obd_list_lock(&cli->cl_loi_list_lock);
1080 cli->cl_avail_grant = ocd->ocd_grant;
1081 client_obd_list_unlock(&cli->cl_loi_list_lock);
1083 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1084 list_empty(&cli->cl_grant_shrink_list))
1085 osc_add_shrink_grant(cli);
1087 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1088 cli->cl_avail_grant, cli->cl_lost_grant);
1089 LASSERT(cli->cl_avail_grant >= 0);
1092 /* We assume that the reason this OSC got a short read is because it read
1093 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1094 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1095 * this stripe never got written at or beyond this stripe offset yet. */
1096 static void handle_short_read(int nob_read, obd_count page_count,
1097 struct brw_page **pga)
1102 /* skip bytes read OK */
1103 while (nob_read > 0) {
1104 LASSERT (page_count > 0);
1106 if (pga[i]->count > nob_read) {
1107 /* EOF inside this page */
1108 ptr = cfs_kmap(pga[i]->pg) +
1109 (pga[i]->off & ~CFS_PAGE_MASK);
1110 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1111 cfs_kunmap(pga[i]->pg);
1117 nob_read -= pga[i]->count;
1122 /* zero remaining pages */
1123 while (page_count-- > 0) {
1124 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1125 memset(ptr, 0, pga[i]->count);
1126 cfs_kunmap(pga[i]->pg);
1131 static int check_write_rcs(struct ptlrpc_request *req,
1132 int requested_nob, int niocount,
1133 obd_count page_count, struct brw_page **pga)
1137 /* return error if any niobuf was in error */
1138 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
1139 sizeof(*remote_rcs) * niocount, NULL);
1140 if (remote_rcs == NULL) {
1141 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1144 if (lustre_msg_swabbed(req->rq_repmsg))
1145 for (i = 0; i < niocount; i++)
1146 __swab32s(&remote_rcs[i]);
1148 for (i = 0; i < niocount; i++) {
1149 if (remote_rcs[i] < 0)
1150 return(remote_rcs[i]);
1152 if (remote_rcs[i] != 0) {
1153 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1154 i, remote_rcs[i], req);
1159 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1160 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1161 req->rq_bulk->bd_nob_transferred, requested_nob);
1168 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1170 if (p1->flag != p2->flag) {
1171 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1172 OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1174 /* warn if we try to combine flags that we don't know to be
1175 * safe to combine */
1176 if ((p1->flag & mask) != (p2->flag & mask))
1177 CERROR("is it ok to have flags 0x%x and 0x%x in the "
1178 "same brw?\n", p1->flag, p2->flag);
1182 return (p1->off + p1->count == p2->off);
1185 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1186 struct brw_page **pga, int opc,
1187 cksum_type_t cksum_type)
1192 LASSERT (pg_count > 0);
1193 cksum = init_checksum(cksum_type);
1194 while (nob > 0 && pg_count > 0) {
1195 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1196 int off = pga[i]->off & ~CFS_PAGE_MASK;
1197 int count = pga[i]->count > nob ? nob : pga[i]->count;
1199 /* corrupt the data before we compute the checksum, to
1200 * simulate an OST->client data error */
1201 if (i == 0 && opc == OST_READ &&
1202 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1203 memcpy(ptr + off, "bad1", min(4, nob));
1204 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1205 cfs_kunmap(pga[i]->pg);
1206 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1209 nob -= pga[i]->count;
1213 /* For sending we only compute the wrong checksum instead
1214 * of corrupting the data so it is still correct on a redo */
1215 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1221 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1222 struct lov_stripe_md *lsm, obd_count page_count,
1223 struct brw_page **pga,
1224 struct ptlrpc_request **reqp,
1225 struct obd_capa *ocapa, int reserve)
1227 struct ptlrpc_request *req;
1228 struct ptlrpc_bulk_desc *desc;
1229 struct ost_body *body;
1230 struct obd_ioobj *ioobj;
1231 struct niobuf_remote *niobuf;
1232 int niocount, i, requested_nob, opc, rc;
1233 struct osc_brw_async_args *aa;
1234 struct req_capsule *pill;
1235 struct brw_page *pg_prev;
1238 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1239 RETURN(-ENOMEM); /* Recoverable */
1240 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1241 RETURN(-EINVAL); /* Fatal */
1243 if ((cmd & OBD_BRW_WRITE) != 0) {
1245 req = ptlrpc_request_alloc_pool(cli->cl_import,
1246 cli->cl_import->imp_rq_pool,
1250 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1255 for (niocount = i = 1; i < page_count; i++) {
1256 if (!can_merge_pages(pga[i - 1], pga[i]))
1260 pill = &req->rq_pill;
1261 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1262 niocount * sizeof(*niobuf));
1263 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1265 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1267 ptlrpc_request_free(req);
1270 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1271 ptlrpc_at_set_req_timeout(req);
1273 if (opc == OST_WRITE)
1274 desc = ptlrpc_prep_bulk_imp(req, page_count,
1275 BULK_GET_SOURCE, OST_BULK_PORTAL);
1277 desc = ptlrpc_prep_bulk_imp(req, page_count,
1278 BULK_PUT_SINK, OST_BULK_PORTAL);
1281 GOTO(out, rc = -ENOMEM);
1282 /* NB request now owns desc and will free it when it gets freed */
1284 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1285 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1286 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1287 LASSERT(body && ioobj && niobuf);
1289 lustre_set_wire_obdo(&body->oa, oa);
1291 obdo_to_ioobj(oa, ioobj);
1292 ioobj->ioo_bufcnt = niocount;
1293 osc_pack_capa(req, body, ocapa);
1294 LASSERT (page_count > 0);
1296 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1297 struct brw_page *pg = pga[i];
1299 LASSERT(pg->count > 0);
1300 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1301 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1302 pg->off, pg->count);
1304 LASSERTF(i == 0 || pg->off > pg_prev->off,
1305 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1306 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1308 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1309 pg_prev->pg, page_private(pg_prev->pg),
1310 pg_prev->pg->index, pg_prev->off);
1312 LASSERTF(i == 0 || pg->off > pg_prev->off,
1313 "i %d p_c %u\n", i, page_count);
1315 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1316 (pg->flag & OBD_BRW_SRVLOCK));
1318 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1320 requested_nob += pg->count;
1322 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1324 niobuf->len += pg->count;
1326 niobuf->offset = pg->off;
1327 niobuf->len = pg->count;
1328 niobuf->flags = pg->flag;
1333 LASSERTF((void *)(niobuf - niocount) ==
1334 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1335 niocount * sizeof(*niobuf)),
1336 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1337 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1338 (void *)(niobuf - niocount));
1340 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1341 if (osc_should_shrink_grant(cli))
1342 osc_shrink_grant_local(cli, &body->oa);
1344 /* size[REQ_REC_OFF] still sizeof (*body) */
1345 if (opc == OST_WRITE) {
1346 if (unlikely(cli->cl_checksum) &&
1347 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1348 /* store cl_cksum_type in a local variable since
1349 * it can be changed via lprocfs */
1350 cksum_type_t cksum_type = cli->cl_cksum_type;
1352 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1353 oa->o_flags &= OBD_FL_LOCAL_MASK;
1354 body->oa.o_flags = 0;
1356 body->oa.o_flags |= cksum_type_pack(cksum_type);
1357 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1358 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1362 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1364 /* save this in 'oa', too, for later checking */
1365 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1366 oa->o_flags |= cksum_type_pack(cksum_type);
1368 /* clear out the checksum flag, in case this is a
1369 * resend but cl_checksum is no longer set. b=11238 */
1370 oa->o_valid &= ~OBD_MD_FLCKSUM;
1372 oa->o_cksum = body->oa.o_cksum;
1373 /* 1 RC per niobuf */
1374 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1375 sizeof(__u32) * niocount);
1377 if (unlikely(cli->cl_checksum) &&
1378 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1379 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1380 body->oa.o_flags = 0;
1381 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1382 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1384 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1385 /* 1 RC for the whole I/O */
1387 ptlrpc_request_set_replen(req);
1389 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1390 aa = ptlrpc_req_async_args(req);
1392 aa->aa_requested_nob = requested_nob;
1393 aa->aa_nio_count = niocount;
1394 aa->aa_page_count = page_count;
1398 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1399 if (ocapa && reserve)
1400 aa->aa_ocapa = capa_get(ocapa);
1406 ptlrpc_req_finished(req);
1410 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1411 __u32 client_cksum, __u32 server_cksum, int nob,
1412 obd_count page_count, struct brw_page **pga,
1413 cksum_type_t client_cksum_type)
1417 cksum_type_t cksum_type;
1419 if (server_cksum == client_cksum) {
1420 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1424 if (oa->o_valid & OBD_MD_FLFLAGS)
1425 cksum_type = cksum_type_unpack(oa->o_flags);
1427 cksum_type = OBD_CKSUM_CRC32;
1429 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1432 if (cksum_type != client_cksum_type)
1433 msg = "the server did not use the checksum type specified in "
1434 "the original request - likely a protocol problem";
1435 else if (new_cksum == server_cksum)
1436 msg = "changed on the client after we checksummed it - "
1437 "likely false positive due to mmap IO (bug 11742)";
1438 else if (new_cksum == client_cksum)
1439 msg = "changed in transit before arrival at OST";
1441 msg = "changed in transit AND doesn't match the original - "
1442 "likely false positive due to mmap IO (bug 11742)";
1444 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1445 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1446 "["LPU64"-"LPU64"]\n",
1447 msg, libcfs_nid2str(peer->nid),
1448 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1449 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1452 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1454 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1455 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1456 "client csum now %x\n", client_cksum, client_cksum_type,
1457 server_cksum, cksum_type, new_cksum);
1461 /* Note rc enters this function as number of bytes transferred */
1462 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1464 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1465 const lnet_process_id_t *peer =
1466 &req->rq_import->imp_connection->c_peer;
1467 struct client_obd *cli = aa->aa_cli;
1468 struct ost_body *body;
1469 __u32 client_cksum = 0;
1472 if (rc < 0 && rc != -EDQUOT)
1475 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1476 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1477 lustre_swab_ost_body);
1479 CDEBUG(D_INFO, "Can't unpack body\n");
1483 /* set/clear over quota flag for a uid/gid */
1484 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1485 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1486 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1488 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1495 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1496 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1498 osc_update_grant(cli, body);
1500 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1502 CERROR("Unexpected +ve rc %d\n", rc);
1505 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1507 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1510 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1511 check_write_checksum(&body->oa, peer, client_cksum,
1512 body->oa.o_cksum, aa->aa_requested_nob,
1513 aa->aa_page_count, aa->aa_ppga,
1514 cksum_type_unpack(aa->aa_oa->o_flags)))
1517 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1518 aa->aa_page_count, aa->aa_ppga);
1522 /* The rest of this function executes only for OST_READs */
1524 /* if unwrap_bulk failed, return -EAGAIN to retry */
1525 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1527 GOTO(out, rc = -EAGAIN);
1529 if (rc > aa->aa_requested_nob) {
1530 CERROR("Unexpected rc %d (%d requested)\n", rc,
1531 aa->aa_requested_nob);
1535 if (rc != req->rq_bulk->bd_nob_transferred) {
1536 CERROR ("Unexpected rc %d (%d transferred)\n",
1537 rc, req->rq_bulk->bd_nob_transferred);
1541 if (rc < aa->aa_requested_nob)
1542 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1544 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1545 static int cksum_counter;
1546 __u32 server_cksum = body->oa.o_cksum;
1549 cksum_type_t cksum_type;
1551 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1552 cksum_type = cksum_type_unpack(body->oa.o_flags);
1554 cksum_type = OBD_CKSUM_CRC32;
1555 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1556 aa->aa_ppga, OST_READ,
1559 if (peer->nid == req->rq_bulk->bd_sender) {
1563 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1566 if (server_cksum == ~0 && rc > 0) {
1567 CERROR("Protocol error: server %s set the 'checksum' "
1568 "bit, but didn't send a checksum. Not fatal, "
1569 "but please notify on http://bugzilla.lustre.org/\n",
1570 libcfs_nid2str(peer->nid));
1571 } else if (server_cksum != client_cksum) {
1572 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1573 "%s%s%s inum "LPU64"/"LPU64" object "
1574 LPU64"/"LPU64" extent "
1575 "["LPU64"-"LPU64"]\n",
1576 req->rq_import->imp_obd->obd_name,
1577 libcfs_nid2str(peer->nid),
1579 body->oa.o_valid & OBD_MD_FLFID ?
1580 body->oa.o_fid : (__u64)0,
1581 body->oa.o_valid & OBD_MD_FLFID ?
1582 body->oa.o_generation :(__u64)0,
1584 body->oa.o_valid & OBD_MD_FLGROUP ?
1585 body->oa.o_gr : (__u64)0,
1586 aa->aa_ppga[0]->off,
1587 aa->aa_ppga[aa->aa_page_count-1]->off +
1588 aa->aa_ppga[aa->aa_page_count-1]->count -
1590 CERROR("client %x, server %x, cksum_type %x\n",
1591 client_cksum, server_cksum, cksum_type);
1593 aa->aa_oa->o_cksum = client_cksum;
1597 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1600 } else if (unlikely(client_cksum)) {
1601 static int cksum_missed;
1604 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1605 CERROR("Checksum %u requested from %s but not sent\n",
1606 cksum_missed, libcfs_nid2str(peer->nid));
1612 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1617 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1618 struct lov_stripe_md *lsm,
1619 obd_count page_count, struct brw_page **pga,
1620 struct obd_capa *ocapa)
1622 struct ptlrpc_request *req;
1626 struct l_wait_info lwi;
1630 cfs_waitq_init(&waitq);
1633 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1634 page_count, pga, &req, ocapa, 0);
1638 rc = ptlrpc_queue_wait(req);
1640 if (rc == -ETIMEDOUT && req->rq_resend) {
1641 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1642 ptlrpc_req_finished(req);
1646 rc = osc_brw_fini_request(req, rc);
1648 ptlrpc_req_finished(req);
1649 if (osc_recoverable_error(rc)) {
1651 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1652 CERROR("too many resend retries, returning error\n");
1656 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1657 l_wait_event(waitq, 0, &lwi);
1665 int osc_brw_redo_request(struct ptlrpc_request *request,
1666 struct osc_brw_async_args *aa)
1668 struct ptlrpc_request *new_req;
1669 struct ptlrpc_request_set *set = request->rq_set;
1670 struct osc_brw_async_args *new_aa;
1671 struct osc_async_page *oap;
1675 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1676 CERROR("too many resend retries, returning error\n");
1680 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1682 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1683 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1684 aa->aa_cli, aa->aa_oa,
1685 NULL /* lsm unused by osc currently */,
1686 aa->aa_page_count, aa->aa_ppga,
1687 &new_req, aa->aa_ocapa, 0);
1691 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1693 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1694 if (oap->oap_request != NULL) {
1695 LASSERTF(request == oap->oap_request,
1696 "request %p != oap_request %p\n",
1697 request, oap->oap_request);
1698 if (oap->oap_interrupted) {
1699 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1700 ptlrpc_req_finished(new_req);
1705 /* New request takes over pga and oaps from old request.
1706 * Note that copying a list_head doesn't work, need to move it... */
1708 new_req->rq_interpret_reply = request->rq_interpret_reply;
1709 new_req->rq_async_args = request->rq_async_args;
1710 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1712 new_aa = ptlrpc_req_async_args(new_req);
1714 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1715 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1716 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1718 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1719 if (oap->oap_request) {
1720 ptlrpc_req_finished(oap->oap_request);
1721 oap->oap_request = ptlrpc_request_addref(new_req);
1725 new_aa->aa_ocapa = aa->aa_ocapa;
1726 aa->aa_ocapa = NULL;
1728 /* use ptlrpc_set_add_req is safe because interpret functions work
1729 * in check_set context. only one way exist with access to request
1730 * from different thread got -EINTR - this way protected with
1731 * cl_loi_list_lock */
1732 ptlrpc_set_add_req(set, new_req);
1734 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1736 DEBUG_REQ(D_INFO, new_req, "new request");
1741 * ugh, we want disk allocation on the target to happen in offset order. we'll
1742 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1743 * fine for our small page arrays and doesn't require allocation. its an
1744 * insertion sort that swaps elements that are strides apart, shrinking the
1745 * stride down until its '1' and the array is sorted.
1747 static void sort_brw_pages(struct brw_page **array, int num)
1750 struct brw_page *tmp;
1754 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1759 for (i = stride ; i < num ; i++) {
1762 while (j >= stride && array[j - stride]->off > tmp->off) {
1763 array[j] = array[j - stride];
1768 } while (stride > 1);
1771 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1777 LASSERT (pages > 0);
1778 offset = pg[i]->off & ~CFS_PAGE_MASK;
1782 if (pages == 0) /* that's all */
1785 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1786 return count; /* doesn't end on page boundary */
1789 offset = pg[i]->off & ~CFS_PAGE_MASK;
1790 if (offset != 0) /* doesn't start on page boundary */
1797 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1799 struct brw_page **ppga;
1802 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1806 for (i = 0; i < count; i++)
1811 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1813 LASSERT(ppga != NULL);
1814 OBD_FREE(ppga, sizeof(*ppga) * count);
1817 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1818 obd_count page_count, struct brw_page *pga,
1819 struct obd_trans_info *oti)
1821 struct obdo *saved_oa = NULL;
1822 struct brw_page **ppga, **orig;
1823 struct obd_import *imp = class_exp2cliimp(exp);
1824 struct client_obd *cli;
1825 int rc, page_count_orig;
1828 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1829 cli = &imp->imp_obd->u.cli;
1831 if (cmd & OBD_BRW_CHECK) {
1832 /* The caller just wants to know if there's a chance that this
1833 * I/O can succeed */
1835 if (imp->imp_invalid)
1840 /* test_brw with a failed create can trip this, maybe others. */
1841 LASSERT(cli->cl_max_pages_per_rpc);
1845 orig = ppga = osc_build_ppga(pga, page_count);
1848 page_count_orig = page_count;
1850 sort_brw_pages(ppga, page_count);
1851 while (page_count) {
1852 obd_count pages_per_brw;
1854 if (page_count > cli->cl_max_pages_per_rpc)
1855 pages_per_brw = cli->cl_max_pages_per_rpc;
1857 pages_per_brw = page_count;
1859 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1861 if (saved_oa != NULL) {
1862 /* restore previously saved oa */
1863 *oinfo->oi_oa = *saved_oa;
1864 } else if (page_count > pages_per_brw) {
1865 /* save a copy of oa (brw will clobber it) */
1866 OBDO_ALLOC(saved_oa);
1867 if (saved_oa == NULL)
1868 GOTO(out, rc = -ENOMEM);
1869 *saved_oa = *oinfo->oi_oa;
1872 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1873 pages_per_brw, ppga, oinfo->oi_capa);
1878 page_count -= pages_per_brw;
1879 ppga += pages_per_brw;
1883 osc_release_ppga(orig, page_count_orig);
1885 if (saved_oa != NULL)
1886 OBDO_FREE(saved_oa);
1891 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1892 * the dirty accounting. Writeback completes or truncate happens before
1893 * writing starts. Must be called with the loi lock held. */
1894 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1897 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1901 /* This maintains the lists of pending pages to read/write for a given object
1902 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1903 * to quickly find objects that are ready to send an RPC. */
1904 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1910 if (lop->lop_num_pending == 0)
1913 /* if we have an invalid import we want to drain the queued pages
1914 * by forcing them through rpcs that immediately fail and complete
1915 * the pages. recovery relies on this to empty the queued pages
1916 * before canceling the locks and evicting down the llite pages */
1917 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1920 /* stream rpcs in queue order as long as as there is an urgent page
1921 * queued. this is our cheap solution for good batching in the case
1922 * where writepage marks some random page in the middle of the file
1923 * as urgent because of, say, memory pressure */
1924 if (!list_empty(&lop->lop_urgent)) {
1925 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1928 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1929 optimal = cli->cl_max_pages_per_rpc;
1930 if (cmd & OBD_BRW_WRITE) {
1931 /* trigger a write rpc stream as long as there are dirtiers
1932 * waiting for space. as they're waiting, they're not going to
1933 * create more pages to coallesce with what's waiting.. */
1934 if (!list_empty(&cli->cl_cache_waiters)) {
1935 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1938 /* +16 to avoid triggering rpcs that would want to include pages
1939 * that are being queued but which can't be made ready until
1940 * the queuer finishes with the page. this is a wart for
1941 * llite::commit_write() */
1944 if (lop->lop_num_pending >= optimal)
1950 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1952 struct osc_async_page *oap;
1955 if (list_empty(&lop->lop_urgent))
1958 oap = list_entry(lop->lop_urgent.next,
1959 struct osc_async_page, oap_urgent_item);
1961 if (oap->oap_async_flags & ASYNC_HP) {
1962 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1969 static void on_list(struct list_head *item, struct list_head *list,
1972 if (list_empty(item) && should_be_on)
1973 list_add_tail(item, list);
1974 else if (!list_empty(item) && !should_be_on)
1975 list_del_init(item);
1978 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1979 * can find pages to build into rpcs quickly */
1980 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1982 if (lop_makes_hprpc(&loi->loi_write_lop) ||
1983 lop_makes_hprpc(&loi->loi_read_lop)) {
1985 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1986 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1988 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1989 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1990 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1991 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1994 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1995 loi->loi_write_lop.lop_num_pending);
1997 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1998 loi->loi_read_lop.lop_num_pending);
2001 static void lop_update_pending(struct client_obd *cli,
2002 struct loi_oap_pages *lop, int cmd, int delta)
2004 lop->lop_num_pending += delta;
2005 if (cmd & OBD_BRW_WRITE)
2006 cli->cl_pending_w_pages += delta;
2008 cli->cl_pending_r_pages += delta;
2012 * this is called when a sync waiter receives an interruption. Its job is to
2013 * get the caller woken as soon as possible. If its page hasn't been put in an
2014 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2015 * desiring interruption which will forcefully complete the rpc once the rpc
2018 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2020 struct loi_oap_pages *lop;
2021 struct lov_oinfo *loi;
2025 LASSERT(!oap->oap_interrupted);
2026 oap->oap_interrupted = 1;
2028 /* ok, it's been put in an rpc. only one oap gets a request reference */
2029 if (oap->oap_request != NULL) {
2030 ptlrpc_mark_interrupted(oap->oap_request);
2031 ptlrpcd_wake(oap->oap_request);
2032 ptlrpc_req_finished(oap->oap_request);
2033 oap->oap_request = NULL;
2037 * page completion may be called only if ->cpo_prep() method was
2038 * executed by osc_io_submit(), that also adds page the to pending list
2040 if (!list_empty(&oap->oap_pending_item)) {
2041 list_del_init(&oap->oap_pending_item);
2042 list_del_init(&oap->oap_urgent_item);
2045 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2046 &loi->loi_write_lop : &loi->loi_read_lop;
2047 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2048 loi_list_maint(oap->oap_cli, oap->oap_loi);
2049 rc = oap->oap_caller_ops->ap_completion(env,
2050 oap->oap_caller_data,
2051 oap->oap_cmd, NULL, -EINTR);
2057 /* this is trying to propogate async writeback errors back up to the
2058 * application. As an async write fails we record the error code for later if
2059 * the app does an fsync. As long as errors persist we force future rpcs to be
2060 * sync so that the app can get a sync error and break the cycle of queueing
2061 * pages for which writeback will fail. */
2062 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2069 ar->ar_force_sync = 1;
2070 ar->ar_min_xid = ptlrpc_sample_next_xid();
2075 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2076 ar->ar_force_sync = 0;
2079 void osc_oap_to_pending(struct osc_async_page *oap)
2081 struct loi_oap_pages *lop;
2083 if (oap->oap_cmd & OBD_BRW_WRITE)
2084 lop = &oap->oap_loi->loi_write_lop;
2086 lop = &oap->oap_loi->loi_read_lop;
2088 if (oap->oap_async_flags & ASYNC_HP)
2089 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2090 else if (oap->oap_async_flags & ASYNC_URGENT)
2091 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2092 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2093 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2096 /* this must be called holding the loi list lock to give coverage to exit_cache,
2097 * async_flag maintenance, and oap_request */
2098 static void osc_ap_completion(const struct lu_env *env,
2099 struct client_obd *cli, struct obdo *oa,
2100 struct osc_async_page *oap, int sent, int rc)
2105 if (oap->oap_request != NULL) {
2106 xid = ptlrpc_req_xid(oap->oap_request);
2107 ptlrpc_req_finished(oap->oap_request);
2108 oap->oap_request = NULL;
2111 oap->oap_async_flags = 0;
2112 oap->oap_interrupted = 0;
2114 if (oap->oap_cmd & OBD_BRW_WRITE) {
2115 osc_process_ar(&cli->cl_ar, xid, rc);
2116 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2119 if (rc == 0 && oa != NULL) {
2120 if (oa->o_valid & OBD_MD_FLBLOCKS)
2121 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2122 if (oa->o_valid & OBD_MD_FLMTIME)
2123 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2124 if (oa->o_valid & OBD_MD_FLATIME)
2125 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2126 if (oa->o_valid & OBD_MD_FLCTIME)
2127 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2130 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2131 oap->oap_cmd, oa, rc);
2133 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2134 * I/O on the page could start, but OSC calls it under lock
2135 * and thus we can add oap back to pending safely */
2137 /* upper layer wants to leave the page on pending queue */
2138 osc_oap_to_pending(oap);
2140 osc_exit_cache(cli, oap, sent);
2144 static int brw_interpret(const struct lu_env *env,
2145 struct ptlrpc_request *req, void *data, int rc)
2147 struct osc_brw_async_args *aa = data;
2148 struct client_obd *cli;
2152 rc = osc_brw_fini_request(req, rc);
2153 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2154 if (osc_recoverable_error(rc)) {
2155 rc = osc_brw_redo_request(req, aa);
2161 capa_put(aa->aa_ocapa);
2162 aa->aa_ocapa = NULL;
2167 client_obd_list_lock(&cli->cl_loi_list_lock);
2169 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2170 * is called so we know whether to go to sync BRWs or wait for more
2171 * RPCs to complete */
2172 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2173 cli->cl_w_in_flight--;
2175 cli->cl_r_in_flight--;
2177 async = list_empty(&aa->aa_oaps);
2178 if (!async) { /* from osc_send_oap_rpc() */
2179 struct osc_async_page *oap, *tmp;
2180 /* the caller may re-use the oap after the completion call so
2181 * we need to clean it up a little */
2182 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2183 list_del_init(&oap->oap_rpc_item);
2184 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2186 OBDO_FREE(aa->aa_oa);
2187 } else { /* from async_internal() */
2189 for (i = 0; i < aa->aa_page_count; i++)
2190 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2192 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2193 OBDO_FREE(aa->aa_oa);
2195 osc_wake_cache_waiters(cli);
2196 osc_check_rpcs(env, cli);
2197 client_obd_list_unlock(&cli->cl_loi_list_lock);
2199 cl_req_completion(env, aa->aa_clerq, rc);
2200 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2204 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2205 struct client_obd *cli,
2206 struct list_head *rpc_list,
2207 int page_count, int cmd)
2209 struct ptlrpc_request *req;
2210 struct brw_page **pga = NULL;
2211 struct osc_brw_async_args *aa;
2212 struct obdo *oa = NULL;
2213 const struct obd_async_page_ops *ops = NULL;
2214 void *caller_data = NULL;
2215 struct osc_async_page *oap;
2216 struct osc_async_page *tmp;
2217 struct ost_body *body;
2218 struct cl_req *clerq = NULL;
2219 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2220 struct ldlm_lock *lock = NULL;
2221 struct cl_req_attr crattr;
2225 LASSERT(!list_empty(rpc_list));
2227 memset(&crattr, 0, sizeof crattr);
2228 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2230 GOTO(out, req = ERR_PTR(-ENOMEM));
2234 GOTO(out, req = ERR_PTR(-ENOMEM));
2237 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2238 struct cl_page *page = osc_oap2cl_page(oap);
2240 ops = oap->oap_caller_ops;
2241 caller_data = oap->oap_caller_data;
2243 clerq = cl_req_alloc(env, page, crt,
2244 1 /* only 1-object rpcs for
2247 GOTO(out, req = (void *)clerq);
2248 lock = oap->oap_ldlm_lock;
2250 pga[i] = &oap->oap_brw_page;
2251 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2252 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2253 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2255 cl_req_page_add(env, clerq, page);
2258 /* always get the data for the obdo for the rpc */
2259 LASSERT(ops != NULL);
2261 crattr.cra_capa = NULL;
2262 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2264 oa->o_handle = lock->l_remote_handle;
2265 oa->o_valid |= OBD_MD_FLHANDLE;
2268 rc = cl_req_prep(env, clerq);
2270 CERROR("cl_req_prep failed: %d\n", rc);
2271 GOTO(out, req = ERR_PTR(rc));
2274 sort_brw_pages(pga, page_count);
2275 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2276 pga, &req, crattr.cra_capa, 1);
2278 CERROR("prep_req failed: %d\n", rc);
2279 GOTO(out, req = ERR_PTR(rc));
2282 /* Need to update the timestamps after the request is built in case
2283 * we race with setattr (locally or in queue at OST). If OST gets
2284 * later setattr before earlier BRW (as determined by the request xid),
2285 * the OST will not use BRW timestamps. Sadly, there is no obvious
2286 * way to do this in a single call. bug 10150 */
2287 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2288 cl_req_attr_set(env, clerq, &crattr,
2289 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2291 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2292 aa = ptlrpc_req_async_args(req);
2293 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2294 list_splice(rpc_list, &aa->aa_oaps);
2295 CFS_INIT_LIST_HEAD(rpc_list);
2296 aa->aa_clerq = clerq;
2298 capa_put(crattr.cra_capa);
2303 OBD_FREE(pga, sizeof(*pga) * page_count);
2304 /* this should happen rarely and is pretty bad, it makes the
2305 * pending list not follow the dirty order */
2306 client_obd_list_lock(&cli->cl_loi_list_lock);
2307 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2308 list_del_init(&oap->oap_rpc_item);
2310 /* queued sync pages can be torn down while the pages
2311 * were between the pending list and the rpc */
2312 if (oap->oap_interrupted) {
2313 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2314 osc_ap_completion(env, cli, NULL, oap, 0,
2318 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2320 if (clerq && !IS_ERR(clerq))
2321 cl_req_completion(env, clerq, PTR_ERR(req));
2327 * prepare pages for ASYNC io and put pages in send queue.
2331 * \param cmd - OBD_BRW_* macroses
2332 * \param lop - pending pages
2334 * \return zero if pages successfully add to send queue.
2335 * \return not zere if error occurring.
2338 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2339 struct lov_oinfo *loi,
2340 int cmd, struct loi_oap_pages *lop)
2342 struct ptlrpc_request *req;
2343 obd_count page_count = 0;
2344 struct osc_async_page *oap = NULL, *tmp;
2345 struct osc_brw_async_args *aa;
2346 const struct obd_async_page_ops *ops;
2347 CFS_LIST_HEAD(rpc_list);
2348 unsigned int ending_offset;
2349 unsigned starting_offset = 0;
2351 struct cl_object *clob = NULL;
2354 /* If there are HP OAPs we need to handle at least 1 of them,
2355 * move it the beginning of the pending list for that. */
2356 if (!list_empty(&lop->lop_urgent)) {
2357 oap = list_entry(lop->lop_urgent.next,
2358 struct osc_async_page, oap_urgent_item);
2359 if (oap->oap_async_flags & ASYNC_HP)
2360 list_move(&oap->oap_pending_item, &lop->lop_pending);
2363 /* first we find the pages we're allowed to work with */
2364 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2366 ops = oap->oap_caller_ops;
2368 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2369 "magic 0x%x\n", oap, oap->oap_magic);
2372 /* pin object in memory, so that completion call-backs
2373 * can be safely called under client_obd_list lock. */
2374 clob = osc_oap2cl_page(oap)->cp_obj;
2375 cl_object_get(clob);
2378 if (page_count != 0 &&
2379 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2380 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2381 " oap %p, page %p, srvlock %u\n",
2382 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2385 /* in llite being 'ready' equates to the page being locked
2386 * until completion unlocks it. commit_write submits a page
2387 * as not ready because its unlock will happen unconditionally
2388 * as the call returns. if we race with commit_write giving
2389 * us that page we dont' want to create a hole in the page
2390 * stream, so we stop and leave the rpc to be fired by
2391 * another dirtier or kupdated interval (the not ready page
2392 * will still be on the dirty list). we could call in
2393 * at the end of ll_file_write to process the queue again. */
2394 if (!(oap->oap_async_flags & ASYNC_READY)) {
2395 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2398 CDEBUG(D_INODE, "oap %p page %p returned %d "
2399 "instead of ready\n", oap,
2403 /* llite is telling us that the page is still
2404 * in commit_write and that we should try
2405 * and put it in an rpc again later. we
2406 * break out of the loop so we don't create
2407 * a hole in the sequence of pages in the rpc
2412 /* the io isn't needed.. tell the checks
2413 * below to complete the rpc with EINTR */
2414 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2415 oap->oap_count = -EINTR;
2418 oap->oap_async_flags |= ASYNC_READY;
2421 LASSERTF(0, "oap %p page %p returned %d "
2422 "from make_ready\n", oap,
2430 * Page submitted for IO has to be locked. Either by
2431 * ->ap_make_ready() or by higher layers.
2433 #if defined(__KERNEL__) && defined(__linux__)
2435 struct cl_page *page;
2437 page = osc_oap2cl_page(oap);
2439 if (page->cp_type == CPT_CACHEABLE &&
2440 !(PageLocked(oap->oap_page) &&
2441 (CheckWriteback(oap->oap_page, cmd)))) {
2442 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2444 (long)oap->oap_page->flags,
2445 oap->oap_async_flags);
2450 /* If there is a gap at the start of this page, it can't merge
2451 * with any previous page, so we'll hand the network a
2452 * "fragmented" page array that it can't transfer in 1 RDMA */
2453 if (page_count != 0 && oap->oap_page_off != 0)
2456 /* take the page out of our book-keeping */
2457 list_del_init(&oap->oap_pending_item);
2458 lop_update_pending(cli, lop, cmd, -1);
2459 list_del_init(&oap->oap_urgent_item);
2461 if (page_count == 0)
2462 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2463 (PTLRPC_MAX_BRW_SIZE - 1);
2465 /* ask the caller for the size of the io as the rpc leaves. */
2466 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2468 ops->ap_refresh_count(env, oap->oap_caller_data,
2470 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2472 if (oap->oap_count <= 0) {
2473 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2475 osc_ap_completion(env, cli, NULL,
2476 oap, 0, oap->oap_count);
2480 /* now put the page back in our accounting */
2481 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2482 if (page_count == 0)
2483 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2484 if (++page_count >= cli->cl_max_pages_per_rpc)
2487 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2488 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2489 * have the same alignment as the initial writes that allocated
2490 * extents on the server. */
2491 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2492 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2493 if (ending_offset == 0)
2496 /* If there is a gap at the end of this page, it can't merge
2497 * with any subsequent pages, so we'll hand the network a
2498 * "fragmented" page array that it can't transfer in 1 RDMA */
2499 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2503 osc_wake_cache_waiters(cli);
2505 loi_list_maint(cli, loi);
2507 client_obd_list_unlock(&cli->cl_loi_list_lock);
2510 cl_object_put(env, clob);
2512 if (page_count == 0) {
2513 client_obd_list_lock(&cli->cl_loi_list_lock);
2517 req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2519 LASSERT(list_empty(&rpc_list));
2520 loi_list_maint(cli, loi);
2521 RETURN(PTR_ERR(req));
2524 aa = ptlrpc_req_async_args(req);
2526 if (cmd == OBD_BRW_READ) {
2527 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2528 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2529 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2530 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2532 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2533 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2534 cli->cl_w_in_flight);
2535 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2536 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2538 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2540 client_obd_list_lock(&cli->cl_loi_list_lock);
2542 if (cmd == OBD_BRW_READ)
2543 cli->cl_r_in_flight++;
2545 cli->cl_w_in_flight++;
2547 /* queued sync pages can be torn down while the pages
2548 * were between the pending list and the rpc */
2550 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2551 /* only one oap gets a request reference */
2554 if (oap->oap_interrupted && !req->rq_intr) {
2555 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2557 ptlrpc_mark_interrupted(req);
2561 tmp->oap_request = ptlrpc_request_addref(req);
2563 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2564 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2566 req->rq_interpret_reply = brw_interpret;
2567 ptlrpcd_add_req(req, PSCOPE_BRW);
2571 #define LOI_DEBUG(LOI, STR, args...) \
2572 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2573 !list_empty(&(LOI)->loi_ready_item) || \
2574 !list_empty(&(LOI)->loi_hp_ready_item), \
2575 (LOI)->loi_write_lop.lop_num_pending, \
2576 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2577 (LOI)->loi_read_lop.lop_num_pending, \
2578 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2581 /* This is called by osc_check_rpcs() to find which objects have pages that
2582 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2583 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2587 /* First return objects that have blocked locks so that they
2588 * will be flushed quickly and other clients can get the lock,
2589 * then objects which have pages ready to be stuffed into RPCs */
2590 if (!list_empty(&cli->cl_loi_hp_ready_list))
2591 RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
2592 struct lov_oinfo, loi_hp_ready_item));
2593 if (!list_empty(&cli->cl_loi_ready_list))
2594 RETURN(list_entry(cli->cl_loi_ready_list.next,
2595 struct lov_oinfo, loi_ready_item));
2597 /* then if we have cache waiters, return all objects with queued
2598 * writes. This is especially important when many small files
2599 * have filled up the cache and not been fired into rpcs because
2600 * they don't pass the nr_pending/object threshhold */
2601 if (!list_empty(&cli->cl_cache_waiters) &&
2602 !list_empty(&cli->cl_loi_write_list))
2603 RETURN(list_entry(cli->cl_loi_write_list.next,
2604 struct lov_oinfo, loi_write_item));
2606 /* then return all queued objects when we have an invalid import
2607 * so that they get flushed */
2608 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2609 if (!list_empty(&cli->cl_loi_write_list))
2610 RETURN(list_entry(cli->cl_loi_write_list.next,
2611 struct lov_oinfo, loi_write_item));
2612 if (!list_empty(&cli->cl_loi_read_list))
2613 RETURN(list_entry(cli->cl_loi_read_list.next,
2614 struct lov_oinfo, loi_read_item));
2619 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2621 struct osc_async_page *oap;
2624 if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
2625 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2626 struct osc_async_page, oap_urgent_item);
2627 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2630 if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
2631 oap = list_entry(loi->loi_read_lop.lop_urgent.next,
2632 struct osc_async_page, oap_urgent_item);
2633 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2636 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2639 /* called with the loi list lock held */
2640 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2642 struct lov_oinfo *loi;
2643 int rc = 0, race_counter = 0;
2646 while ((loi = osc_next_loi(cli)) != NULL) {
2647 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2649 if (osc_max_rpc_in_flight(cli, loi))
2652 /* attempt some read/write balancing by alternating between
2653 * reads and writes in an object. The makes_rpc checks here
2654 * would be redundant if we were getting read/write work items
2655 * instead of objects. we don't want send_oap_rpc to drain a
2656 * partial read pending queue when we're given this object to
2657 * do io on writes while there are cache waiters */
2658 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2659 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2660 &loi->loi_write_lop);
2668 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2669 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2670 &loi->loi_read_lop);
2679 /* attempt some inter-object balancing by issueing rpcs
2680 * for each object in turn */
2681 if (!list_empty(&loi->loi_hp_ready_item))
2682 list_del_init(&loi->loi_hp_ready_item);
2683 if (!list_empty(&loi->loi_ready_item))
2684 list_del_init(&loi->loi_ready_item);
2685 if (!list_empty(&loi->loi_write_item))
2686 list_del_init(&loi->loi_write_item);
2687 if (!list_empty(&loi->loi_read_item))
2688 list_del_init(&loi->loi_read_item);
2690 loi_list_maint(cli, loi);
2692 /* send_oap_rpc fails with 0 when make_ready tells it to
2693 * back off. llite's make_ready does this when it tries
2694 * to lock a page queued for write that is already locked.
2695 * we want to try sending rpcs from many objects, but we
2696 * don't want to spin failing with 0. */
2697 if (race_counter == 10)
2703 /* we're trying to queue a page in the osc so we're subject to the
2704 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2705 * If the osc's queued pages are already at that limit, then we want to sleep
2706 * until there is space in the osc's queue for us. We also may be waiting for
2707 * write credits from the OST if there are RPCs in flight that may return some
2708 * before we fall back to sync writes.
2710 * We need this know our allocation was granted in the presence of signals */
2711 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2715 client_obd_list_lock(&cli->cl_loi_list_lock);
2716 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2717 client_obd_list_unlock(&cli->cl_loi_list_lock);
2722 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2725 int osc_enter_cache_try(const struct lu_env *env,
2726 struct client_obd *cli, struct lov_oinfo *loi,
2727 struct osc_async_page *oap, int transient)
2731 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2733 osc_consume_write_grant(cli, &oap->oap_brw_page);
2735 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2736 atomic_inc(&obd_dirty_transit_pages);
2737 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2743 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2744 * grant or cache space. */
2745 static int osc_enter_cache(const struct lu_env *env,
2746 struct client_obd *cli, struct lov_oinfo *loi,
2747 struct osc_async_page *oap)
2749 struct osc_cache_waiter ocw;
2750 struct l_wait_info lwi = { 0 };
2754 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2755 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2756 cli->cl_dirty_max, obd_max_dirty_pages,
2757 cli->cl_lost_grant, cli->cl_avail_grant);
2759 /* force the caller to try sync io. this can jump the list
2760 * of queued writes and create a discontiguous rpc stream */
2761 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2762 loi->loi_ar.ar_force_sync)
2765 /* Hopefully normal case - cache space and write credits available */
2766 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2767 atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2768 osc_enter_cache_try(env, cli, loi, oap, 0))
2771 /* Make sure that there are write rpcs in flight to wait for. This
2772 * is a little silly as this object may not have any pending but
2773 * other objects sure might. */
2774 if (cli->cl_w_in_flight) {
2775 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2776 cfs_waitq_init(&ocw.ocw_waitq);
2780 loi_list_maint(cli, loi);
2781 osc_check_rpcs(env, cli);
2782 client_obd_list_unlock(&cli->cl_loi_list_lock);
2784 CDEBUG(D_CACHE, "sleeping for cache space\n");
2785 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2787 client_obd_list_lock(&cli->cl_loi_list_lock);
2788 if (!list_empty(&ocw.ocw_entry)) {
2789 list_del(&ocw.ocw_entry);
2799 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2800 struct lov_oinfo *loi, cfs_page_t *page,
2801 obd_off offset, const struct obd_async_page_ops *ops,
2802 void *data, void **res, int nocache,
2803 struct lustre_handle *lockh)
2805 struct osc_async_page *oap;
2810 return size_round(sizeof(*oap));
2813 oap->oap_magic = OAP_MAGIC;
2814 oap->oap_cli = &exp->exp_obd->u.cli;
2817 oap->oap_caller_ops = ops;
2818 oap->oap_caller_data = data;
2820 oap->oap_page = page;
2821 oap->oap_obj_off = offset;
2822 if (!client_is_remote(exp) &&
2823 cfs_capable(CFS_CAP_SYS_RESOURCE))
2824 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2826 LASSERT(!(offset & ~CFS_PAGE_MASK));
2828 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2829 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2830 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2831 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2833 spin_lock_init(&oap->oap_lock);
2834 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2838 struct osc_async_page *oap_from_cookie(void *cookie)
2840 struct osc_async_page *oap = cookie;
2841 if (oap->oap_magic != OAP_MAGIC)
2842 return ERR_PTR(-EINVAL);
2846 int osc_queue_async_io(const struct lu_env *env,
2847 struct obd_export *exp, struct lov_stripe_md *lsm,
2848 struct lov_oinfo *loi, void *cookie,
2849 int cmd, obd_off off, int count,
2850 obd_flag brw_flags, enum async_flags async_flags)
2852 struct client_obd *cli = &exp->exp_obd->u.cli;
2853 struct osc_async_page *oap;
2857 oap = oap_from_cookie(cookie);
2859 RETURN(PTR_ERR(oap));
2861 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2864 if (!list_empty(&oap->oap_pending_item) ||
2865 !list_empty(&oap->oap_urgent_item) ||
2866 !list_empty(&oap->oap_rpc_item))
2869 /* check if the file's owner/group is over quota */
2870 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2871 struct cl_object *obj;
2872 struct cl_attr attr; /* XXX put attr into thread info */
2873 unsigned int qid[MAXQUOTAS];
2875 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2877 cl_object_attr_lock(obj);
2878 rc = cl_object_attr_get(env, obj, &attr);
2879 cl_object_attr_unlock(obj);
2881 qid[USRQUOTA] = attr.cat_uid;
2882 qid[GRPQUOTA] = attr.cat_gid;
2884 lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2891 loi = lsm->lsm_oinfo[0];
2893 client_obd_list_lock(&cli->cl_loi_list_lock);
2895 LASSERT(off + count <= CFS_PAGE_SIZE);
2897 oap->oap_page_off = off;
2898 oap->oap_count = count;
2899 oap->oap_brw_flags = brw_flags;
2900 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2901 if (libcfs_memory_pressure_get())
2902 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2903 oap->oap_async_flags = async_flags;
2905 if (cmd & OBD_BRW_WRITE) {
2906 rc = osc_enter_cache(env, cli, loi, oap);
2908 client_obd_list_unlock(&cli->cl_loi_list_lock);
2913 osc_oap_to_pending(oap);
2914 loi_list_maint(cli, loi);
2916 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2919 osc_check_rpcs(env, cli);
2920 client_obd_list_unlock(&cli->cl_loi_list_lock);
2925 /* aka (~was & now & flag), but this is more clear :) */
2926 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2928 int osc_set_async_flags_base(struct client_obd *cli,
2929 struct lov_oinfo *loi, struct osc_async_page *oap,
2930 obd_flag async_flags)
2932 struct loi_oap_pages *lop;
2935 LASSERT(!list_empty(&oap->oap_pending_item));
2937 if (oap->oap_cmd & OBD_BRW_WRITE) {
2938 lop = &loi->loi_write_lop;
2940 lop = &loi->loi_read_lop;
2943 if ((oap->oap_async_flags & async_flags) == async_flags)
2946 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2947 oap->oap_async_flags |= ASYNC_READY;
2949 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
2950 list_empty(&oap->oap_rpc_item)) {
2951 if (oap->oap_async_flags & ASYNC_HP)
2952 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2954 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2955 oap->oap_async_flags |= ASYNC_URGENT;
2956 loi_list_maint(cli, loi);
2959 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2960 oap->oap_async_flags);
2964 int osc_teardown_async_page(struct obd_export *exp,
2965 struct lov_stripe_md *lsm,
2966 struct lov_oinfo *loi, void *cookie)
2968 struct client_obd *cli = &exp->exp_obd->u.cli;
2969 struct loi_oap_pages *lop;
2970 struct osc_async_page *oap;
2974 oap = oap_from_cookie(cookie);
2976 RETURN(PTR_ERR(oap));
2979 loi = lsm->lsm_oinfo[0];
2981 if (oap->oap_cmd & OBD_BRW_WRITE) {
2982 lop = &loi->loi_write_lop;
2984 lop = &loi->loi_read_lop;
2987 client_obd_list_lock(&cli->cl_loi_list_lock);
2989 if (!list_empty(&oap->oap_rpc_item))
2990 GOTO(out, rc = -EBUSY);
2992 osc_exit_cache(cli, oap, 0);
2993 osc_wake_cache_waiters(cli);
2995 if (!list_empty(&oap->oap_urgent_item)) {
2996 list_del_init(&oap->oap_urgent_item);
2997 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
2999 if (!list_empty(&oap->oap_pending_item)) {
3000 list_del_init(&oap->oap_pending_item);
3001 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3003 loi_list_maint(cli, loi);
3004 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3006 client_obd_list_unlock(&cli->cl_loi_list_lock);
3010 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3011 struct ldlm_enqueue_info *einfo,
3014 void *data = einfo->ei_cbdata;
3016 LASSERT(lock != NULL);
3017 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3018 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3019 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3020 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3022 lock_res_and_lock(lock);
3023 spin_lock(&osc_ast_guard);
3024 LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3025 lock->l_ast_data = data;
3026 spin_unlock(&osc_ast_guard);
3027 unlock_res_and_lock(lock);
3030 static void osc_set_data_with_check(struct lustre_handle *lockh,
3031 struct ldlm_enqueue_info *einfo,
3034 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3037 osc_set_lock_data_with_check(lock, einfo, flags);
3038 LDLM_LOCK_PUT(lock);
3040 CERROR("lockh %p, data %p - client evicted?\n",
3041 lockh, einfo->ei_cbdata);
3044 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3045 ldlm_iterator_t replace, void *data)
3047 struct ldlm_res_id res_id;
3048 struct obd_device *obd = class_exp2obd(exp);
3050 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3051 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3055 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3056 obd_enqueue_update_f upcall, void *cookie,
3059 int intent = *flags & LDLM_FL_HAS_INTENT;
3063 /* The request was created before ldlm_cli_enqueue call. */
3064 if (rc == ELDLM_LOCK_ABORTED) {
3065 struct ldlm_reply *rep;
3066 rep = req_capsule_server_get(&req->rq_pill,
3069 LASSERT(rep != NULL);
3070 if (rep->lock_policy_res1)
3071 rc = rep->lock_policy_res1;
3075 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3076 *flags |= LDLM_FL_LVB_READY;
3077 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3078 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3081 /* Call the update callback. */
3082 rc = (*upcall)(cookie, rc);
3086 static int osc_enqueue_interpret(const struct lu_env *env,
3087 struct ptlrpc_request *req,
3088 struct osc_enqueue_args *aa, int rc)
3090 struct ldlm_lock *lock;
3091 struct lustre_handle handle;
3094 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3095 * might be freed anytime after lock upcall has been called. */
3096 lustre_handle_copy(&handle, aa->oa_lockh);
3097 mode = aa->oa_ei->ei_mode;
3099 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3101 lock = ldlm_handle2lock(&handle);
3103 /* Take an additional reference so that a blocking AST that
3104 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3105 * to arrive after an upcall has been executed by
3106 * osc_enqueue_fini(). */
3107 ldlm_lock_addref(&handle, mode);
3109 /* Complete obtaining the lock procedure. */
3110 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3111 mode, aa->oa_flags, aa->oa_lvb,
3112 sizeof(*aa->oa_lvb), lustre_swab_ost_lvb,
3114 /* Complete osc stuff. */
3115 rc = osc_enqueue_fini(req, aa->oa_lvb,
3116 aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3118 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3120 /* Release the lock for async request. */
3121 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3123 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3124 * not already released by
3125 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3127 ldlm_lock_decref(&handle, mode);
3129 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3130 aa->oa_lockh, req, aa);
3131 ldlm_lock_decref(&handle, mode);
3132 LDLM_LOCK_PUT(lock);
3136 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3137 struct lov_oinfo *loi, int flags,
3138 struct ost_lvb *lvb, __u32 mode, int rc)
3140 if (rc == ELDLM_OK) {
3141 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3144 LASSERT(lock != NULL);
3145 loi->loi_lvb = *lvb;
3146 tmp = loi->loi_lvb.lvb_size;
3147 /* Extend KMS up to the end of this lock and no further
3148 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3149 if (tmp > lock->l_policy_data.l_extent.end)
3150 tmp = lock->l_policy_data.l_extent.end + 1;
3151 if (tmp >= loi->loi_kms) {
3152 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3153 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3154 loi_kms_set(loi, tmp);
3156 LDLM_DEBUG(lock, "lock acquired, setting rss="
3157 LPU64"; leaving kms="LPU64", end="LPU64,
3158 loi->loi_lvb.lvb_size, loi->loi_kms,
3159 lock->l_policy_data.l_extent.end);
3161 ldlm_lock_allow_match(lock);
3162 LDLM_LOCK_PUT(lock);
3163 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3164 loi->loi_lvb = *lvb;
3165 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3166 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3170 EXPORT_SYMBOL(osc_update_enqueue);
3172 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3174 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3175 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3176 * other synchronous requests, however keeping some locks and trying to obtain
3177 * others may take a considerable amount of time in a case of ost failure; and
3178 * when other sync requests do not get released lock from a client, the client
3179 * is excluded from the cluster -- such scenarious make the life difficult, so
3180 * release locks just after they are obtained. */
3181 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3182 int *flags, ldlm_policy_data_t *policy,
3183 struct ost_lvb *lvb, int kms_valid,
3184 obd_enqueue_update_f upcall, void *cookie,
3185 struct ldlm_enqueue_info *einfo,
3186 struct lustre_handle *lockh,
3187 struct ptlrpc_request_set *rqset, int async)
3189 struct obd_device *obd = exp->exp_obd;
3190 struct ptlrpc_request *req = NULL;
3191 int intent = *flags & LDLM_FL_HAS_INTENT;
3196 /* Filesystem lock extents are extended to page boundaries so that
3197 * dealing with the page cache is a little smoother. */
3198 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3199 policy->l_extent.end |= ~CFS_PAGE_MASK;
3202 * kms is not valid when either object is completely fresh (so that no
3203 * locks are cached), or object was evicted. In the latter case cached
3204 * lock cannot be used, because it would prime inode state with
3205 * potentially stale LVB.
3210 /* Next, search for already existing extent locks that will cover us */
3211 /* If we're trying to read, we also search for an existing PW lock. The
3212 * VFS and page cache already protect us locally, so lots of readers/
3213 * writers can share a single PW lock.
3215 * There are problems with conversion deadlocks, so instead of
3216 * converting a read lock to a write lock, we'll just enqueue a new
3219 * At some point we should cancel the read lock instead of making them
3220 * send us a blocking callback, but there are problems with canceling
3221 * locks out from other users right now, too. */
3222 mode = einfo->ei_mode;
3223 if (einfo->ei_mode == LCK_PR)
3225 mode = ldlm_lock_match(obd->obd_namespace,
3226 *flags | LDLM_FL_LVB_READY, res_id,
3227 einfo->ei_type, policy, mode, lockh, 0);
3229 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3231 if (matched->l_ast_data == NULL ||
3232 matched->l_ast_data == einfo->ei_cbdata) {
3233 /* addref the lock only if not async requests and PW
3234 * lock is matched whereas we asked for PR. */
3235 if (!rqset && einfo->ei_mode != mode)
3236 ldlm_lock_addref(lockh, LCK_PR);
3237 osc_set_lock_data_with_check(matched, einfo, *flags);
3239 /* I would like to be able to ASSERT here that
3240 * rss <= kms, but I can't, for reasons which
3241 * are explained in lov_enqueue() */
3244 /* We already have a lock, and it's referenced */
3245 (*upcall)(cookie, ELDLM_OK);
3247 /* For async requests, decref the lock. */
3248 if (einfo->ei_mode != mode)
3249 ldlm_lock_decref(lockh, LCK_PW);
3251 ldlm_lock_decref(lockh, einfo->ei_mode);
3252 LDLM_LOCK_PUT(matched);
3255 ldlm_lock_decref(lockh, mode);
3256 LDLM_LOCK_PUT(matched);
3261 CFS_LIST_HEAD(cancels);
3262 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3263 &RQF_LDLM_ENQUEUE_LVB);
3267 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3271 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3273 ptlrpc_request_set_replen(req);
3276 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3277 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3279 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3280 sizeof(*lvb), lustre_swab_ost_lvb, lockh, async);
3283 struct osc_enqueue_args *aa;
3284 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3285 aa = ptlrpc_req_async_args(req);
3288 aa->oa_flags = flags;
3289 aa->oa_upcall = upcall;
3290 aa->oa_cookie = cookie;
3292 aa->oa_lockh = lockh;
3294 req->rq_interpret_reply =
3295 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3296 if (rqset == PTLRPCD_SET)
3297 ptlrpcd_add_req(req, PSCOPE_OTHER);
3299 ptlrpc_set_add_req(rqset, req);
3300 } else if (intent) {
3301 ptlrpc_req_finished(req);
3306 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3308 ptlrpc_req_finished(req);
3313 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3314 struct ldlm_enqueue_info *einfo,
3315 struct ptlrpc_request_set *rqset)
3317 struct ldlm_res_id res_id;
3321 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3322 oinfo->oi_md->lsm_object_gr, &res_id);
3324 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3325 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3326 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3327 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3328 rqset, rqset != NULL);
3332 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3333 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3334 int *flags, void *data, struct lustre_handle *lockh,
3337 struct obd_device *obd = exp->exp_obd;
3338 int lflags = *flags;
3342 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3345 /* Filesystem lock extents are extended to page boundaries so that
3346 * dealing with the page cache is a little smoother */
3347 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3348 policy->l_extent.end |= ~CFS_PAGE_MASK;
3350 /* Next, search for already existing extent locks that will cover us */
3351 /* If we're trying to read, we also search for an existing PW lock. The
3352 * VFS and page cache already protect us locally, so lots of readers/
3353 * writers can share a single PW lock. */
3357 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3358 res_id, type, policy, rc, lockh, unref);
3361 osc_set_data_with_check(lockh, data, lflags);
3362 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3363 ldlm_lock_addref(lockh, LCK_PR);
3364 ldlm_lock_decref(lockh, LCK_PW);
3371 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3375 if (unlikely(mode == LCK_GROUP))
3376 ldlm_lock_decref_and_cancel(lockh, mode);
3378 ldlm_lock_decref(lockh, mode);
3383 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3384 __u32 mode, struct lustre_handle *lockh)
3387 RETURN(osc_cancel_base(lockh, mode));
3390 static int osc_cancel_unused(struct obd_export *exp,
3391 struct lov_stripe_md *lsm, int flags,
3394 struct obd_device *obd = class_exp2obd(exp);
3395 struct ldlm_res_id res_id, *resp = NULL;
3398 resp = osc_build_res_name(lsm->lsm_object_id,
3399 lsm->lsm_object_gr, &res_id);
3402 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3405 static int osc_statfs_interpret(const struct lu_env *env,
3406 struct ptlrpc_request *req,
3407 struct osc_async_args *aa, int rc)
3409 struct obd_statfs *msfs;
3412 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3413 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3419 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3421 GOTO(out, rc = -EPROTO);
3424 *aa->aa_oi->oi_osfs = *msfs;
3426 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3430 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3431 __u64 max_age, struct ptlrpc_request_set *rqset)
3433 struct ptlrpc_request *req;
3434 struct osc_async_args *aa;
3438 /* We could possibly pass max_age in the request (as an absolute
3439 * timestamp or a "seconds.usec ago") so the target can avoid doing
3440 * extra calls into the filesystem if that isn't necessary (e.g.
3441 * during mount that would help a bit). Having relative timestamps
3442 * is not so great if request processing is slow, while absolute
3443 * timestamps are not ideal because they need time synchronization. */
3444 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3448 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3450 ptlrpc_request_free(req);
3453 ptlrpc_request_set_replen(req);
3454 req->rq_request_portal = OST_CREATE_PORTAL;
3455 ptlrpc_at_set_req_timeout(req);
3457 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3458 /* procfs requests not want stat in wait for avoid deadlock */
3459 req->rq_no_resend = 1;
3460 req->rq_no_delay = 1;
3463 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3464 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3465 aa = ptlrpc_req_async_args(req);
3468 ptlrpc_set_add_req(rqset, req);
3472 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3473 __u64 max_age, __u32 flags)
3475 struct obd_statfs *msfs;
3476 struct ptlrpc_request *req;
3477 struct obd_import *imp = NULL;
3481 /*Since the request might also come from lprocfs, so we need
3482 *sync this with client_disconnect_export Bug15684*/
3483 down_read(&obd->u.cli.cl_sem);
3484 if (obd->u.cli.cl_import)
3485 imp = class_import_get(obd->u.cli.cl_import);
3486 up_read(&obd->u.cli.cl_sem);
3490 /* We could possibly pass max_age in the request (as an absolute
3491 * timestamp or a "seconds.usec ago") so the target can avoid doing
3492 * extra calls into the filesystem if that isn't necessary (e.g.
3493 * during mount that would help a bit). Having relative timestamps
3494 * is not so great if request processing is slow, while absolute
3495 * timestamps are not ideal because they need time synchronization. */
3496 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3498 class_import_put(imp);
3503 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3505 ptlrpc_request_free(req);
3508 ptlrpc_request_set_replen(req);
3509 req->rq_request_portal = OST_CREATE_PORTAL;
3510 ptlrpc_at_set_req_timeout(req);
3512 if (flags & OBD_STATFS_NODELAY) {
3513 /* procfs requests not want stat in wait for avoid deadlock */
3514 req->rq_no_resend = 1;
3515 req->rq_no_delay = 1;
3518 rc = ptlrpc_queue_wait(req);
3522 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3524 GOTO(out, rc = -EPROTO);
3531 ptlrpc_req_finished(req);
3535 /* Retrieve object striping information.
3537 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3538 * the maximum number of OST indices which will fit in the user buffer.
3539 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3541 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3543 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3544 struct lov_user_md_v3 lum, *lumk;
3545 struct lov_user_ost_data_v1 *lmm_objects;
3546 int rc = 0, lum_size;
3552 /* we only need the header part from user space to get lmm_magic and
3553 * lmm_stripe_count, (the header part is common to v1 and v3) */
3554 lum_size = sizeof(struct lov_user_md_v1);
3555 if (copy_from_user(&lum, lump, lum_size))
3558 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3559 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3562 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3563 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3564 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3565 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3567 /* we can use lov_mds_md_size() to compute lum_size
3568 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3569 if (lum.lmm_stripe_count > 0) {
3570 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3571 OBD_ALLOC(lumk, lum_size);
3575 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3576 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3578 lmm_objects = &(lumk->lmm_objects[0]);
3579 lmm_objects->l_object_id = lsm->lsm_object_id;
3581 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3585 lumk->lmm_object_id = lsm->lsm_object_id;
3586 lumk->lmm_object_gr = lsm->lsm_object_gr;
3587 lumk->lmm_stripe_count = 1;
3589 if (copy_to_user(lump, lumk, lum_size))
3593 OBD_FREE(lumk, lum_size);
3599 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3600 void *karg, void *uarg)
3602 struct obd_device *obd = exp->exp_obd;
3603 struct obd_ioctl_data *data = karg;
3607 if (!try_module_get(THIS_MODULE)) {
3608 CERROR("Can't get module. Is it alive?");
3612 case OBD_IOC_LOV_GET_CONFIG: {
3614 struct lov_desc *desc;
3615 struct obd_uuid uuid;
3619 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3620 GOTO(out, err = -EINVAL);
3622 data = (struct obd_ioctl_data *)buf;
3624 if (sizeof(*desc) > data->ioc_inllen1) {
3625 obd_ioctl_freedata(buf, len);
3626 GOTO(out, err = -EINVAL);
3629 if (data->ioc_inllen2 < sizeof(uuid)) {
3630 obd_ioctl_freedata(buf, len);
3631 GOTO(out, err = -EINVAL);
3634 desc = (struct lov_desc *)data->ioc_inlbuf1;
3635 desc->ld_tgt_count = 1;
3636 desc->ld_active_tgt_count = 1;
3637 desc->ld_default_stripe_count = 1;
3638 desc->ld_default_stripe_size = 0;
3639 desc->ld_default_stripe_offset = 0;
3640 desc->ld_pattern = 0;
3641 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3643 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3645 err = copy_to_user((void *)uarg, buf, len);
3648 obd_ioctl_freedata(buf, len);
3651 case LL_IOC_LOV_SETSTRIPE:
3652 err = obd_alloc_memmd(exp, karg);
3656 case LL_IOC_LOV_GETSTRIPE:
3657 err = osc_getstripe(karg, uarg);
3659 case OBD_IOC_CLIENT_RECOVER:
3660 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3665 case IOC_OSC_SET_ACTIVE:
3666 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3669 case OBD_IOC_POLL_QUOTACHECK:
3670 err = lquota_poll_check(quota_interface, exp,
3671 (struct if_quotacheck *)karg);
3673 case OBD_IOC_PING_TARGET:
3674 err = ptlrpc_obd_ping(obd);
3677 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3678 cmd, cfs_curproc_comm());
3679 GOTO(out, err = -ENOTTY);
3682 module_put(THIS_MODULE);
3686 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3687 void *key, __u32 *vallen, void *val,
3688 struct lov_stripe_md *lsm)
3691 if (!vallen || !val)
3694 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3695 __u32 *stripe = val;
3696 *vallen = sizeof(*stripe);
3699 } else if (KEY_IS(KEY_LAST_ID)) {
3700 struct ptlrpc_request *req;
3705 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3706 &RQF_OST_GET_INFO_LAST_ID);
3710 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3711 RCL_CLIENT, keylen);
3712 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3714 ptlrpc_request_free(req);
3718 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3719 memcpy(tmp, key, keylen);
3721 req->rq_no_delay = req->rq_no_resend = 1;
3722 ptlrpc_request_set_replen(req);
3723 rc = ptlrpc_queue_wait(req);
3727 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3729 GOTO(out, rc = -EPROTO);
3731 *((obd_id *)val) = *reply;
3733 ptlrpc_req_finished(req);
3735 } else if (KEY_IS(KEY_FIEMAP)) {
3736 struct ptlrpc_request *req;
3737 struct ll_user_fiemap *reply;
3741 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3742 &RQF_OST_GET_INFO_FIEMAP);
3746 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3747 RCL_CLIENT, keylen);
3748 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3749 RCL_CLIENT, *vallen);
3750 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3751 RCL_SERVER, *vallen);
3753 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3755 ptlrpc_request_free(req);
3759 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3760 memcpy(tmp, key, keylen);
3761 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3762 memcpy(tmp, val, *vallen);
3764 ptlrpc_request_set_replen(req);
3765 rc = ptlrpc_queue_wait(req);
3769 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3771 GOTO(out1, rc = -EPROTO);
3773 memcpy(val, reply, *vallen);
3775 ptlrpc_req_finished(req);
3783 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
3785 struct llog_ctxt *ctxt;
3789 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3791 rc = llog_initiator_connect(ctxt);
3792 llog_ctxt_put(ctxt);
3794 /* XXX return an error? skip setting below flags? */
3797 spin_lock(&imp->imp_lock);
3798 imp->imp_server_timeout = 1;
3799 imp->imp_pingable = 1;
3800 spin_unlock(&imp->imp_lock);
3801 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3806 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
3807 struct ptlrpc_request *req,
3814 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
3817 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3818 void *key, obd_count vallen, void *val,
3819 struct ptlrpc_request_set *set)
3821 struct ptlrpc_request *req;
3822 struct obd_device *obd = exp->exp_obd;
3823 struct obd_import *imp = class_exp2cliimp(exp);
3828 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3830 if (KEY_IS(KEY_NEXT_ID)) {
3831 if (vallen != sizeof(obd_id))
3835 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3836 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3837 exp->exp_obd->obd_name,
3838 obd->u.cli.cl_oscc.oscc_next_id);
3843 if (KEY_IS(KEY_UNLINKED)) {
3844 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3845 spin_lock(&oscc->oscc_lock);
3846 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3847 spin_unlock(&oscc->oscc_lock);
3851 if (KEY_IS(KEY_INIT_RECOV)) {
3852 if (vallen != sizeof(int))
3854 spin_lock(&imp->imp_lock);
3855 imp->imp_initial_recov = *(int *)val;
3856 spin_unlock(&imp->imp_lock);
3857 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3858 exp->exp_obd->obd_name,
3859 imp->imp_initial_recov);
3863 if (KEY_IS(KEY_CHECKSUM)) {
3864 if (vallen != sizeof(int))
3866 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3870 if (KEY_IS(KEY_SPTLRPC_CONF)) {
3871 sptlrpc_conf_client_adapt(obd);
3875 if (KEY_IS(KEY_FLUSH_CTX)) {
3876 sptlrpc_import_flush_my_ctx(imp);
3880 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3883 /* We pass all other commands directly to OST. Since nobody calls osc
3884 methods directly and everybody is supposed to go through LOV, we
3885 assume lov checked invalid values for us.
3886 The only recognised values so far are evict_by_nid and mds_conn.
3887 Even if something bad goes through, we'd get a -EINVAL from OST
3890 if (KEY_IS(KEY_GRANT_SHRINK))
3891 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
3893 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3898 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3899 RCL_CLIENT, keylen);
3900 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3901 RCL_CLIENT, vallen);
3902 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3904 ptlrpc_request_free(req);
3908 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3909 memcpy(tmp, key, keylen);
3910 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3911 memcpy(tmp, val, vallen);
3913 if (KEY_IS(KEY_MDS_CONN)) {
3914 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3916 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3917 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3918 LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
3919 req->rq_no_delay = req->rq_no_resend = 1;
3920 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3921 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
3922 struct osc_grant_args *aa;
3925 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3926 aa = ptlrpc_req_async_args(req);
3929 ptlrpc_req_finished(req);
3932 *oa = ((struct ost_body *)val)->oa;
3934 req->rq_interpret_reply = osc_shrink_grant_interpret;
3937 ptlrpc_request_set_replen(req);
3938 if (!KEY_IS(KEY_GRANT_SHRINK)) {
3939 LASSERT(set != NULL);
3940 ptlrpc_set_add_req(set, req);
3941 ptlrpc_check_set(NULL, set);
3943 ptlrpcd_add_req(req, PSCOPE_OTHER);
3949 static struct llog_operations osc_size_repl_logops = {
3950 lop_cancel: llog_obd_repl_cancel
3953 static struct llog_operations osc_mds_ost_orig_logops;
3954 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3955 struct obd_device *tgt, int count,
3956 struct llog_catid *catid, struct obd_uuid *uuid)
3961 LASSERT(olg == &obd->obd_olg);
3962 spin_lock(&obd->obd_dev_lock);
3963 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3964 osc_mds_ost_orig_logops = llog_lvfs_ops;
3965 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3966 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3967 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3968 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3970 spin_unlock(&obd->obd_dev_lock);
3972 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3973 &catid->lci_logid, &osc_mds_ost_orig_logops);
3975 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3979 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3980 NULL, &osc_size_repl_logops);
3982 struct llog_ctxt *ctxt =
3983 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3986 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3991 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3992 obd->obd_name, tgt->obd_name, count, catid, rc);
3993 CERROR("logid "LPX64":0x%x\n",
3994 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3999 static int osc_llog_finish(struct obd_device *obd, int count)
4001 struct llog_ctxt *ctxt;
4002 int rc = 0, rc2 = 0;
4005 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4007 rc = llog_cleanup(ctxt);
4009 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4011 rc2 = llog_cleanup(ctxt);
4018 static int osc_reconnect(const struct lu_env *env,
4019 struct obd_export *exp, struct obd_device *obd,
4020 struct obd_uuid *cluuid,
4021 struct obd_connect_data *data,
4024 struct client_obd *cli = &obd->u.cli;
4026 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4029 client_obd_list_lock(&cli->cl_loi_list_lock);
4030 data->ocd_grant = cli->cl_avail_grant ?:
4031 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4032 lost_grant = cli->cl_lost_grant;
4033 cli->cl_lost_grant = 0;
4034 client_obd_list_unlock(&cli->cl_loi_list_lock);
4036 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4037 "cl_lost_grant: %ld\n", data->ocd_grant,
4038 cli->cl_avail_grant, lost_grant);
4039 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4040 " ocd_grant: %d\n", data->ocd_connect_flags,
4041 data->ocd_version, data->ocd_grant);
4047 static int osc_disconnect(struct obd_export *exp)
4049 struct obd_device *obd = class_exp2obd(exp);
4050 struct llog_ctxt *ctxt;
4053 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4055 if (obd->u.cli.cl_conn_count == 1) {
4056 /* Flush any remaining cancel messages out to the
4058 llog_sync(ctxt, exp);
4060 llog_ctxt_put(ctxt);
4062 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4066 rc = client_disconnect_export(exp);
4068 * Initially we put del_shrink_grant before disconnect_export, but it
4069 * causes the following problem if setup (connect) and cleanup
4070 * (disconnect) are tangled together.
4071 * connect p1 disconnect p2
4072 * ptlrpc_connect_import
4073 * ............... class_manual_cleanup
4076 * ptlrpc_connect_interrupt
4078 * add this client to shrink list
4080 * Bang! pinger trigger the shrink.
4081 * So the osc should be disconnected from the shrink list, after we
4082 * are sure the import has been destroyed. BUG18662
4084 if (obd->u.cli.cl_import == NULL)
4085 osc_del_shrink_grant(&obd->u.cli);
4089 static int osc_import_event(struct obd_device *obd,
4090 struct obd_import *imp,
4091 enum obd_import_event event)
4093 struct client_obd *cli;
4097 LASSERT(imp->imp_obd == obd);
4100 case IMP_EVENT_DISCON: {
4101 /* Only do this on the MDS OSC's */
4102 if (imp->imp_server_timeout) {
4103 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4105 spin_lock(&oscc->oscc_lock);
4106 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4107 spin_unlock(&oscc->oscc_lock);
4110 client_obd_list_lock(&cli->cl_loi_list_lock);
4111 cli->cl_avail_grant = 0;
4112 cli->cl_lost_grant = 0;
4113 client_obd_list_unlock(&cli->cl_loi_list_lock);
4116 case IMP_EVENT_INACTIVE: {
4117 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4120 case IMP_EVENT_INVALIDATE: {
4121 struct ldlm_namespace *ns = obd->obd_namespace;
4125 env = cl_env_get(&refcheck);
4129 client_obd_list_lock(&cli->cl_loi_list_lock);
4130 /* all pages go to failing rpcs due to the invalid
4132 osc_check_rpcs(env, cli);
4133 client_obd_list_unlock(&cli->cl_loi_list_lock);
4135 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4136 cl_env_put(env, &refcheck);
4141 case IMP_EVENT_ACTIVE: {
4142 /* Only do this on the MDS OSC's */
4143 if (imp->imp_server_timeout) {
4144 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4146 spin_lock(&oscc->oscc_lock);
4147 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
4148 spin_unlock(&oscc->oscc_lock);
4150 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4153 case IMP_EVENT_OCD: {
4154 struct obd_connect_data *ocd = &imp->imp_connect_data;
4156 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4157 osc_init_grant(&obd->u.cli, ocd);
4160 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4161 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4163 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4167 CERROR("Unknown import event %d\n", event);
4173 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4179 rc = ptlrpcd_addref();
4183 rc = client_obd_setup(obd, lcfg);
4187 struct lprocfs_static_vars lvars = { 0 };
4188 struct client_obd *cli = &obd->u.cli;
4190 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4191 lprocfs_osc_init_vars(&lvars);
4192 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4193 lproc_osc_attach_seqstat(obd);
4194 sptlrpc_lprocfs_cliobd_attach(obd);
4195 ptlrpc_lprocfs_register_obd(obd);
4199 /* We need to allocate a few requests more, because
4200 brw_interpret tries to create new requests before freeing
4201 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4202 reserved, but I afraid that might be too much wasted RAM
4203 in fact, so 2 is just my guess and still should work. */
4204 cli->cl_import->imp_rq_pool =
4205 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4207 ptlrpc_add_rqs_to_pool);
4209 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4210 sema_init(&cli->cl_grant_sem, 1);
4216 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4222 case OBD_CLEANUP_EARLY: {
4223 struct obd_import *imp;
4224 imp = obd->u.cli.cl_import;
4225 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4226 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4227 ptlrpc_deactivate_import(imp);
4228 spin_lock(&imp->imp_lock);
4229 imp->imp_pingable = 0;
4230 spin_unlock(&imp->imp_lock);
4233 case OBD_CLEANUP_EXPORTS: {
4234 /* If we set up but never connected, the
4235 client import will not have been cleaned. */
4236 if (obd->u.cli.cl_import) {
4237 struct obd_import *imp;
4238 down_write(&obd->u.cli.cl_sem);
4239 imp = obd->u.cli.cl_import;
4240 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4242 ptlrpc_invalidate_import(imp);
4243 if (imp->imp_rq_pool) {
4244 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4245 imp->imp_rq_pool = NULL;
4247 class_destroy_import(imp);
4248 up_write(&obd->u.cli.cl_sem);
4249 obd->u.cli.cl_import = NULL;
4251 rc = obd_llog_finish(obd, 0);
4253 CERROR("failed to cleanup llogging subsystems\n");
4260 int osc_cleanup(struct obd_device *obd)
4265 ptlrpc_lprocfs_unregister_obd(obd);
4266 lprocfs_obd_cleanup(obd);
4268 /* free memory of osc quota cache */
4269 lquota_cleanup(quota_interface, obd);
4271 rc = client_obd_cleanup(obd);
4277 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4279 struct lprocfs_static_vars lvars = { 0 };
4282 lprocfs_osc_init_vars(&lvars);
4284 switch (lcfg->lcfg_command) {
4286 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4296 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4298 return osc_process_config_base(obd, buf);
4301 struct obd_ops osc_obd_ops = {
4302 .o_owner = THIS_MODULE,
4303 .o_setup = osc_setup,
4304 .o_precleanup = osc_precleanup,
4305 .o_cleanup = osc_cleanup,
4306 .o_add_conn = client_import_add_conn,
4307 .o_del_conn = client_import_del_conn,
4308 .o_connect = client_connect_import,
4309 .o_reconnect = osc_reconnect,
4310 .o_disconnect = osc_disconnect,
4311 .o_statfs = osc_statfs,
4312 .o_statfs_async = osc_statfs_async,
4313 .o_packmd = osc_packmd,
4314 .o_unpackmd = osc_unpackmd,
4315 .o_precreate = osc_precreate,
4316 .o_create = osc_create,
4317 .o_create_async = osc_create_async,
4318 .o_destroy = osc_destroy,
4319 .o_getattr = osc_getattr,
4320 .o_getattr_async = osc_getattr_async,
4321 .o_setattr = osc_setattr,
4322 .o_setattr_async = osc_setattr_async,
4324 .o_punch = osc_punch,
4326 .o_enqueue = osc_enqueue,
4327 .o_change_cbdata = osc_change_cbdata,
4328 .o_cancel = osc_cancel,
4329 .o_cancel_unused = osc_cancel_unused,
4330 .o_iocontrol = osc_iocontrol,
4331 .o_get_info = osc_get_info,
4332 .o_set_info_async = osc_set_info_async,
4333 .o_import_event = osc_import_event,
4334 .o_llog_init = osc_llog_init,
4335 .o_llog_finish = osc_llog_finish,
4336 .o_process_config = osc_process_config,
4339 extern struct lu_kmem_descr osc_caches[];
4340 extern spinlock_t osc_ast_guard;
4341 extern struct lock_class_key osc_ast_guard_class;
4343 int __init osc_init(void)
4345 struct lprocfs_static_vars lvars = { 0 };
4349 /* print an address of _any_ initialized kernel symbol from this
4350 * module, to allow debugging with gdb that doesn't support data
4351 * symbols from modules.*/
4352 CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4354 rc = lu_kmem_init(osc_caches);
4356 lprocfs_osc_init_vars(&lvars);
4358 request_module("lquota");
4359 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4360 lquota_init(quota_interface);
4361 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4363 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4364 LUSTRE_OSC_NAME, &osc_device_type);
4366 if (quota_interface)
4367 PORTAL_SYMBOL_PUT(osc_quota_interface);
4368 lu_kmem_fini(osc_caches);
4372 spin_lock_init(&osc_ast_guard);
4373 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4379 static void /*__exit*/ osc_exit(void)
4381 lu_device_type_fini(&osc_device_type);
4383 lquota_exit(quota_interface);
4384 if (quota_interface)
4385 PORTAL_SYMBOL_PUT(osc_quota_interface);
4387 class_unregister_type(LUSTRE_OSC_NAME);
4388 lu_kmem_fini(osc_caches);
4391 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4392 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4393 MODULE_LICENSE("GPL");
4395 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);