4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_OSC
39 #include "../../include/linux/libcfs/libcfs.h"
41 #include "../include/lustre_dlm.h"
42 #include "../include/lustre_net.h"
43 #include "../include/lustre/lustre_user.h"
44 #include "../include/obd_cksum.h"
46 #include "../include/lustre_ha.h"
47 #include "../include/lprocfs_status.h"
48 #include "../include/lustre_debug.h"
49 #include "../include/lustre_param.h"
50 #include "../include/lustre_fid.h"
51 #include "../include/obd_class.h"
52 #include "../include/obd.h"
53 #include "osc_internal.h"
54 #include "osc_cl_internal.h"
56 atomic_t osc_pool_req_count
;
57 unsigned int osc_reqpool_maxreqcount
;
58 struct ptlrpc_request_pool
*osc_rq_pool
;
60 /* max memory used for request pool, unit is MB */
61 static unsigned int osc_reqpool_mem_max
= 5;
62 module_param(osc_reqpool_mem_max
, uint
, 0444);
64 struct osc_brw_async_args
{
70 struct brw_page
**aa_ppga
;
71 struct client_obd
*aa_cli
;
72 struct list_head aa_oaps
;
73 struct list_head aa_exts
;
74 struct cl_req
*aa_clerq
;
77 struct osc_async_args
{
78 struct obd_info
*aa_oi
;
81 struct osc_setattr_args
{
83 obd_enqueue_update_f sa_upcall
;
87 struct osc_fsync_args
{
88 struct obd_info
*fa_oi
;
89 obd_enqueue_update_f fa_upcall
;
93 struct osc_enqueue_args
{
94 struct obd_export
*oa_exp
;
95 enum ldlm_type oa_type
;
96 enum ldlm_mode oa_mode
;
98 osc_enqueue_upcall_f oa_upcall
;
100 struct ost_lvb
*oa_lvb
;
101 struct lustre_handle oa_lockh
;
102 unsigned int oa_agl
:1;
105 static void osc_release_ppga(struct brw_page
**ppga
, u32 count
);
106 static int brw_interpret(const struct lu_env
*env
,
107 struct ptlrpc_request
*req
, void *data
, int rc
);
109 /* Pack OSC object metadata for disk storage (LE byte order). */
110 static int osc_packmd(struct obd_export
*exp
, struct lov_mds_md
**lmmp
,
111 struct lov_stripe_md
*lsm
)
115 lmm_size
= sizeof(**lmmp
);
123 } else if (unlikely(lsm
&& ostid_id(&lsm
->lsm_oi
) == 0)) {
128 *lmmp
= kzalloc(lmm_size
, GFP_NOFS
);
134 ostid_cpu_to_le(&lsm
->lsm_oi
, &(*lmmp
)->lmm_oi
);
139 /* Unpack OSC object metadata from disk storage (LE byte order). */
140 static int osc_unpackmd(struct obd_export
*exp
, struct lov_stripe_md
**lsmp
,
141 struct lov_mds_md
*lmm
, int lmm_bytes
)
144 struct obd_import
*imp
= class_exp2cliimp(exp
);
147 if (lmm_bytes
< sizeof(*lmm
)) {
148 CERROR("%s: lov_mds_md too small: %d, need %d\n",
149 exp
->exp_obd
->obd_name
, lmm_bytes
,
153 /* XXX LOV_MAGIC etc check? */
155 if (unlikely(ostid_id(&lmm
->lmm_oi
) == 0)) {
156 CERROR("%s: zero lmm_object_id: rc = %d\n",
157 exp
->exp_obd
->obd_name
, -EINVAL
);
162 lsm_size
= lov_stripe_md_size(1);
167 kfree((*lsmp
)->lsm_oinfo
[0]);
174 *lsmp
= kzalloc(lsm_size
, GFP_NOFS
);
175 if (unlikely(!*lsmp
))
177 (*lsmp
)->lsm_oinfo
[0] = kzalloc(sizeof(struct lov_oinfo
),
179 if (unlikely(!(*lsmp
)->lsm_oinfo
[0])) {
183 loi_init((*lsmp
)->lsm_oinfo
[0]);
184 } else if (unlikely(ostid_id(&(*lsmp
)->lsm_oi
) == 0)) {
189 /* XXX zero *lsmp? */
190 ostid_le_to_cpu(&lmm
->lmm_oi
, &(*lsmp
)->lsm_oi
);
193 (imp
->imp_connect_data
.ocd_connect_flags
& OBD_CONNECT_MAXBYTES
))
194 (*lsmp
)->lsm_maxbytes
= imp
->imp_connect_data
.ocd_maxbytes
;
196 (*lsmp
)->lsm_maxbytes
= LUSTRE_STRIPE_MAXBYTES
;
201 static inline void osc_pack_req_body(struct ptlrpc_request
*req
,
202 struct obd_info
*oinfo
)
204 struct ost_body
*body
;
206 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
209 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
,
213 static int osc_getattr_interpret(const struct lu_env
*env
,
214 struct ptlrpc_request
*req
,
215 struct osc_async_args
*aa
, int rc
)
217 struct ost_body
*body
;
222 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
224 CDEBUG(D_INODE
, "mode: %o\n", body
->oa
.o_mode
);
225 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
,
226 aa
->aa_oi
->oi_oa
, &body
->oa
);
228 /* This should really be sent by the OST */
229 aa
->aa_oi
->oi_oa
->o_blksize
= DT_MAX_BRW_SIZE
;
230 aa
->aa_oi
->oi_oa
->o_valid
|= OBD_MD_FLBLKSZ
;
232 CDEBUG(D_INFO
, "can't unpack ost_body\n");
234 aa
->aa_oi
->oi_oa
->o_valid
= 0;
237 rc
= aa
->aa_oi
->oi_cb_up(aa
->aa_oi
, rc
);
241 static int osc_getattr_async(struct obd_export
*exp
, struct obd_info
*oinfo
,
242 struct ptlrpc_request_set
*set
)
244 struct ptlrpc_request
*req
;
245 struct osc_async_args
*aa
;
248 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_GETATTR
);
252 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_GETATTR
);
254 ptlrpc_request_free(req
);
258 osc_pack_req_body(req
, oinfo
);
260 ptlrpc_request_set_replen(req
);
261 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)osc_getattr_interpret
;
263 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
264 aa
= ptlrpc_req_async_args(req
);
267 ptlrpc_set_add_req(set
, req
);
271 static int osc_getattr(const struct lu_env
*env
, struct obd_export
*exp
,
272 struct obd_info
*oinfo
)
274 struct ptlrpc_request
*req
;
275 struct ost_body
*body
;
278 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_GETATTR
);
282 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_GETATTR
);
284 ptlrpc_request_free(req
);
288 osc_pack_req_body(req
, oinfo
);
290 ptlrpc_request_set_replen(req
);
292 rc
= ptlrpc_queue_wait(req
);
296 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
302 CDEBUG(D_INODE
, "mode: %o\n", body
->oa
.o_mode
);
303 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
, oinfo
->oi_oa
,
306 oinfo
->oi_oa
->o_blksize
= cli_brw_size(exp
->exp_obd
);
307 oinfo
->oi_oa
->o_valid
|= OBD_MD_FLBLKSZ
;
310 ptlrpc_req_finished(req
);
314 static int osc_setattr(const struct lu_env
*env
, struct obd_export
*exp
,
315 struct obd_info
*oinfo
, struct obd_trans_info
*oti
)
317 struct ptlrpc_request
*req
;
318 struct ost_body
*body
;
321 LASSERT(oinfo
->oi_oa
->o_valid
& OBD_MD_FLGROUP
);
323 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_SETATTR
);
327 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_SETATTR
);
329 ptlrpc_request_free(req
);
333 osc_pack_req_body(req
, oinfo
);
335 ptlrpc_request_set_replen(req
);
337 rc
= ptlrpc_queue_wait(req
);
341 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
347 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
, oinfo
->oi_oa
,
351 ptlrpc_req_finished(req
);
355 static int osc_setattr_interpret(const struct lu_env
*env
,
356 struct ptlrpc_request
*req
,
357 struct osc_setattr_args
*sa
, int rc
)
359 struct ost_body
*body
;
364 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
370 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
, sa
->sa_oa
,
373 rc
= sa
->sa_upcall(sa
->sa_cookie
, rc
);
377 int osc_setattr_async_base(struct obd_export
*exp
, struct obd_info
*oinfo
,
378 struct obd_trans_info
*oti
,
379 obd_enqueue_update_f upcall
, void *cookie
,
380 struct ptlrpc_request_set
*rqset
)
382 struct ptlrpc_request
*req
;
383 struct osc_setattr_args
*sa
;
386 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_SETATTR
);
390 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_SETATTR
);
392 ptlrpc_request_free(req
);
396 if (oti
&& oinfo
->oi_oa
->o_valid
& OBD_MD_FLCOOKIE
)
397 oinfo
->oi_oa
->o_lcookie
= *oti
->oti_logcookies
;
399 osc_pack_req_body(req
, oinfo
);
401 ptlrpc_request_set_replen(req
);
403 /* do mds to ost setattr asynchronously */
405 /* Do not wait for response. */
406 ptlrpcd_add_req(req
);
408 req
->rq_interpret_reply
=
409 (ptlrpc_interpterer_t
)osc_setattr_interpret
;
411 CLASSERT(sizeof(*sa
) <= sizeof(req
->rq_async_args
));
412 sa
= ptlrpc_req_async_args(req
);
413 sa
->sa_oa
= oinfo
->oi_oa
;
414 sa
->sa_upcall
= upcall
;
415 sa
->sa_cookie
= cookie
;
417 if (rqset
== PTLRPCD_SET
)
418 ptlrpcd_add_req(req
);
420 ptlrpc_set_add_req(rqset
, req
);
426 static int osc_setattr_async(struct obd_export
*exp
, struct obd_info
*oinfo
,
427 struct obd_trans_info
*oti
,
428 struct ptlrpc_request_set
*rqset
)
430 return osc_setattr_async_base(exp
, oinfo
, oti
,
431 oinfo
->oi_cb_up
, oinfo
, rqset
);
434 static int osc_real_create(struct obd_export
*exp
, struct obdo
*oa
,
435 struct lov_stripe_md
**ea
,
436 struct obd_trans_info
*oti
)
438 struct ptlrpc_request
*req
;
439 struct ost_body
*body
;
440 struct lov_stripe_md
*lsm
;
448 rc
= obd_alloc_memmd(exp
, &lsm
);
453 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_CREATE
);
459 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_CREATE
);
461 ptlrpc_request_free(req
);
465 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
468 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
, oa
);
470 ptlrpc_request_set_replen(req
);
472 if ((oa
->o_valid
& OBD_MD_FLFLAGS
) &&
473 oa
->o_flags
== OBD_FL_DELORPHAN
) {
475 "delorphan from OST integration");
476 /* Don't resend the delorphan req */
477 req
->rq_no_resend
= req
->rq_no_delay
= 1;
480 rc
= ptlrpc_queue_wait(req
);
484 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
490 CDEBUG(D_INFO
, "oa flags %x\n", oa
->o_flags
);
491 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
, oa
, &body
->oa
);
493 oa
->o_blksize
= cli_brw_size(exp
->exp_obd
);
494 oa
->o_valid
|= OBD_MD_FLBLKSZ
;
496 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
497 * have valid lsm_oinfo data structs, so don't go touching that.
498 * This needs to be fixed in a big way.
500 lsm
->lsm_oi
= oa
->o_oi
;
504 oti
->oti_transno
= lustre_msg_get_transno(req
->rq_repmsg
);
506 if (oa
->o_valid
& OBD_MD_FLCOOKIE
) {
507 if (!oti
->oti_logcookies
)
508 oti_alloc_cookies(oti
, 1);
509 *oti
->oti_logcookies
= oa
->o_lcookie
;
513 CDEBUG(D_HA
, "transno: %lld\n",
514 lustre_msg_get_transno(req
->rq_repmsg
));
516 ptlrpc_req_finished(req
);
519 obd_free_memmd(exp
, &lsm
);
523 int osc_punch_base(struct obd_export
*exp
, struct obd_info
*oinfo
,
524 obd_enqueue_update_f upcall
, void *cookie
,
525 struct ptlrpc_request_set
*rqset
)
527 struct ptlrpc_request
*req
;
528 struct osc_setattr_args
*sa
;
529 struct ost_body
*body
;
532 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_PUNCH
);
536 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_PUNCH
);
538 ptlrpc_request_free(req
);
541 req
->rq_request_portal
= OST_IO_PORTAL
; /* bug 7198 */
542 ptlrpc_at_set_req_timeout(req
);
544 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
546 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
,
549 ptlrpc_request_set_replen(req
);
551 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)osc_setattr_interpret
;
552 CLASSERT(sizeof(*sa
) <= sizeof(req
->rq_async_args
));
553 sa
= ptlrpc_req_async_args(req
);
554 sa
->sa_oa
= oinfo
->oi_oa
;
555 sa
->sa_upcall
= upcall
;
556 sa
->sa_cookie
= cookie
;
557 if (rqset
== PTLRPCD_SET
)
558 ptlrpcd_add_req(req
);
560 ptlrpc_set_add_req(rqset
, req
);
565 static int osc_sync_interpret(const struct lu_env
*env
,
566 struct ptlrpc_request
*req
,
569 struct osc_fsync_args
*fa
= arg
;
570 struct ost_body
*body
;
575 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
577 CERROR("can't unpack ost_body\n");
582 *fa
->fa_oi
->oi_oa
= body
->oa
;
584 rc
= fa
->fa_upcall(fa
->fa_cookie
, rc
);
588 int osc_sync_base(struct obd_export
*exp
, struct obd_info
*oinfo
,
589 obd_enqueue_update_f upcall
, void *cookie
,
590 struct ptlrpc_request_set
*rqset
)
592 struct ptlrpc_request
*req
;
593 struct ost_body
*body
;
594 struct osc_fsync_args
*fa
;
597 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_SYNC
);
601 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_SYNC
);
603 ptlrpc_request_free(req
);
607 /* overload the size and blocks fields in the oa with start/end */
608 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
610 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
,
613 ptlrpc_request_set_replen(req
);
614 req
->rq_interpret_reply
= osc_sync_interpret
;
616 CLASSERT(sizeof(*fa
) <= sizeof(req
->rq_async_args
));
617 fa
= ptlrpc_req_async_args(req
);
619 fa
->fa_upcall
= upcall
;
620 fa
->fa_cookie
= cookie
;
622 if (rqset
== PTLRPCD_SET
)
623 ptlrpcd_add_req(req
);
625 ptlrpc_set_add_req(rqset
, req
);
630 /* Find and cancel locally locks matched by @mode in the resource found by
631 * @objid. Found locks are added into @cancel list. Returns the amount of
632 * locks added to @cancels list.
634 static int osc_resource_get_unused(struct obd_export
*exp
, struct obdo
*oa
,
635 struct list_head
*cancels
,
636 enum ldlm_mode mode
, __u64 lock_flags
)
638 struct ldlm_namespace
*ns
= exp
->exp_obd
->obd_namespace
;
639 struct ldlm_res_id res_id
;
640 struct ldlm_resource
*res
;
643 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
644 * export) but disabled through procfs (flag in NS).
646 * This distinguishes from a case when ELC is not supported originally,
647 * when we still want to cancel locks in advance and just cancel them
648 * locally, without sending any RPC.
650 if (exp_connect_cancelset(exp
) && !ns_connect_cancelset(ns
))
653 ostid_build_res_name(&oa
->o_oi
, &res_id
);
654 res
= ldlm_resource_get(ns
, NULL
, &res_id
, 0, 0);
658 LDLM_RESOURCE_ADDREF(res
);
659 count
= ldlm_cancel_resource_local(res
, cancels
, NULL
, mode
,
660 lock_flags
, 0, NULL
);
661 LDLM_RESOURCE_DELREF(res
);
662 ldlm_resource_putref(res
);
666 static int osc_destroy_interpret(const struct lu_env
*env
,
667 struct ptlrpc_request
*req
, void *data
,
670 struct client_obd
*cli
= &req
->rq_import
->imp_obd
->u
.cli
;
672 atomic_dec(&cli
->cl_destroy_in_flight
);
673 wake_up(&cli
->cl_destroy_waitq
);
677 static int osc_can_send_destroy(struct client_obd
*cli
)
679 if (atomic_inc_return(&cli
->cl_destroy_in_flight
) <=
680 cli
->cl_max_rpcs_in_flight
) {
681 /* The destroy request can be sent */
684 if (atomic_dec_return(&cli
->cl_destroy_in_flight
) <
685 cli
->cl_max_rpcs_in_flight
) {
687 * The counter has been modified between the two atomic
690 wake_up(&cli
->cl_destroy_waitq
);
695 static int osc_create(const struct lu_env
*env
, struct obd_export
*exp
,
696 struct obdo
*oa
, struct lov_stripe_md
**ea
,
697 struct obd_trans_info
*oti
)
703 LASSERT(oa
->o_valid
& OBD_MD_FLGROUP
);
705 if ((oa
->o_valid
& OBD_MD_FLFLAGS
) &&
706 oa
->o_flags
== OBD_FL_RECREATE_OBJS
) {
707 return osc_real_create(exp
, oa
, ea
, oti
);
710 if (!fid_seq_is_mdt(ostid_seq(&oa
->o_oi
)))
711 return osc_real_create(exp
, oa
, ea
, oti
);
713 /* we should not get here anymore */
719 /* Destroy requests can be async always on the client, and we don't even really
720 * care about the return code since the client cannot do anything at all about
722 * When the MDS is unlinking a filename, it saves the file objects into a
723 * recovery llog, and these object records are cancelled when the OST reports
724 * they were destroyed and sync'd to disk (i.e. transaction committed).
725 * If the client dies, or the OST is down when the object should be destroyed,
726 * the records are not cancelled, and when the OST reconnects to the MDS next,
727 * it will retrieve the llog unlink logs and then sends the log cancellation
728 * cookies to the MDS after committing destroy transactions.
730 static int osc_destroy(const struct lu_env
*env
, struct obd_export
*exp
,
731 struct obdo
*oa
, struct lov_stripe_md
*ea
,
732 struct obd_trans_info
*oti
, struct obd_export
*md_export
)
734 struct client_obd
*cli
= &exp
->exp_obd
->u
.cli
;
735 struct ptlrpc_request
*req
;
736 struct ost_body
*body
;
741 CDEBUG(D_INFO
, "oa NULL\n");
745 count
= osc_resource_get_unused(exp
, oa
, &cancels
, LCK_PW
,
746 LDLM_FL_DISCARD_DATA
);
748 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
), &RQF_OST_DESTROY
);
750 ldlm_lock_list_put(&cancels
, l_bl_ast
, count
);
754 rc
= ldlm_prep_elc_req(exp
, req
, LUSTRE_OST_VERSION
, OST_DESTROY
,
757 ptlrpc_request_free(req
);
761 req
->rq_request_portal
= OST_IO_PORTAL
; /* bug 7198 */
762 ptlrpc_at_set_req_timeout(req
);
764 if (oti
&& oa
->o_valid
& OBD_MD_FLCOOKIE
)
765 oa
->o_lcookie
= *oti
->oti_logcookies
;
766 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
768 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
, oa
);
770 ptlrpc_request_set_replen(req
);
772 /* If osc_destroy is for destroying the unlink orphan,
773 * sent from MDT to OST, which should not be blocked here,
774 * because the process might be triggered by ptlrpcd, and
775 * it is not good to block ptlrpcd thread (b=16006
777 if (!(oa
->o_flags
& OBD_FL_DELORPHAN
)) {
778 req
->rq_interpret_reply
= osc_destroy_interpret
;
779 if (!osc_can_send_destroy(cli
)) {
780 struct l_wait_info lwi
= LWI_INTR(LWI_ON_SIGNAL_NOOP
,
784 * Wait until the number of on-going destroy RPCs drops
785 * under max_rpc_in_flight
787 l_wait_event_exclusive(cli
->cl_destroy_waitq
,
788 osc_can_send_destroy(cli
), &lwi
);
792 /* Do not wait for response */
793 ptlrpcd_add_req(req
);
797 static void osc_announce_cached(struct client_obd
*cli
, struct obdo
*oa
,
800 u32 bits
= OBD_MD_FLBLOCKS
|OBD_MD_FLGRANT
;
802 LASSERT(!(oa
->o_valid
& bits
));
805 spin_lock(&cli
->cl_loi_list_lock
);
806 oa
->o_dirty
= cli
->cl_dirty
;
807 if (unlikely(cli
->cl_dirty
- cli
->cl_dirty_transit
>
808 cli
->cl_dirty_max
)) {
809 CERROR("dirty %lu - %lu > dirty_max %lu\n",
810 cli
->cl_dirty
, cli
->cl_dirty_transit
, cli
->cl_dirty_max
);
812 } else if (unlikely(atomic_read(&obd_unstable_pages
) +
813 atomic_read(&obd_dirty_pages
) -
814 atomic_read(&obd_dirty_transit_pages
) >
815 (long)(obd_max_dirty_pages
+ 1))) {
816 /* The atomic_read() allowing the atomic_inc() are
817 * not covered by a lock thus they may safely race and trip
818 * this CERROR() unless we add in a small fudge factor (+1).
820 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
821 cli
->cl_import
->imp_obd
->obd_name
,
822 atomic_read(&obd_unstable_pages
),
823 atomic_read(&obd_dirty_pages
),
824 atomic_read(&obd_dirty_transit_pages
),
825 obd_max_dirty_pages
);
827 } else if (unlikely(cli
->cl_dirty_max
- cli
->cl_dirty
> 0x7fffffff)) {
828 CERROR("dirty %lu - dirty_max %lu too big???\n",
829 cli
->cl_dirty
, cli
->cl_dirty_max
);
832 long max_in_flight
= (cli
->cl_max_pages_per_rpc
<<
834 (cli
->cl_max_rpcs_in_flight
+ 1);
835 oa
->o_undirty
= max(cli
->cl_dirty_max
, max_in_flight
);
837 oa
->o_grant
= cli
->cl_avail_grant
+ cli
->cl_reserved_grant
;
838 oa
->o_dropped
= cli
->cl_lost_grant
;
839 cli
->cl_lost_grant
= 0;
840 spin_unlock(&cli
->cl_loi_list_lock
);
841 CDEBUG(D_CACHE
, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
842 oa
->o_dirty
, oa
->o_undirty
, oa
->o_dropped
, oa
->o_grant
);
845 void osc_update_next_shrink(struct client_obd
*cli
)
847 cli
->cl_next_shrink_grant
=
848 cfs_time_shift(cli
->cl_grant_shrink_interval
);
849 CDEBUG(D_CACHE
, "next time %ld to shrink grant\n",
850 cli
->cl_next_shrink_grant
);
853 static void __osc_update_grant(struct client_obd
*cli
, u64 grant
)
855 spin_lock(&cli
->cl_loi_list_lock
);
856 cli
->cl_avail_grant
+= grant
;
857 spin_unlock(&cli
->cl_loi_list_lock
);
860 static void osc_update_grant(struct client_obd
*cli
, struct ost_body
*body
)
862 if (body
->oa
.o_valid
& OBD_MD_FLGRANT
) {
863 CDEBUG(D_CACHE
, "got %llu extra grant\n", body
->oa
.o_grant
);
864 __osc_update_grant(cli
, body
->oa
.o_grant
);
868 static int osc_set_info_async(const struct lu_env
*env
, struct obd_export
*exp
,
869 u32 keylen
, void *key
, u32 vallen
,
870 void *val
, struct ptlrpc_request_set
*set
);
872 static int osc_shrink_grant_interpret(const struct lu_env
*env
,
873 struct ptlrpc_request
*req
,
876 struct client_obd
*cli
= &req
->rq_import
->imp_obd
->u
.cli
;
877 struct obdo
*oa
= ((struct osc_brw_async_args
*)aa
)->aa_oa
;
878 struct ost_body
*body
;
881 __osc_update_grant(cli
, oa
->o_grant
);
885 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
887 osc_update_grant(cli
, body
);
889 kmem_cache_free(obdo_cachep
, oa
);
893 static void osc_shrink_grant_local(struct client_obd
*cli
, struct obdo
*oa
)
895 spin_lock(&cli
->cl_loi_list_lock
);
896 oa
->o_grant
= cli
->cl_avail_grant
/ 4;
897 cli
->cl_avail_grant
-= oa
->o_grant
;
898 spin_unlock(&cli
->cl_loi_list_lock
);
899 if (!(oa
->o_valid
& OBD_MD_FLFLAGS
)) {
900 oa
->o_valid
|= OBD_MD_FLFLAGS
;
903 oa
->o_flags
|= OBD_FL_SHRINK_GRANT
;
904 osc_update_next_shrink(cli
);
907 /* Shrink the current grant, either from some large amount to enough for a
908 * full set of in-flight RPCs, or if we have already shrunk to that limit
909 * then to enough for a single RPC. This avoids keeping more grant than
910 * needed, and avoids shrinking the grant piecemeal.
912 static int osc_shrink_grant(struct client_obd
*cli
)
914 __u64 target_bytes
= (cli
->cl_max_rpcs_in_flight
+ 1) *
915 (cli
->cl_max_pages_per_rpc
<< PAGE_SHIFT
);
917 spin_lock(&cli
->cl_loi_list_lock
);
918 if (cli
->cl_avail_grant
<= target_bytes
)
919 target_bytes
= cli
->cl_max_pages_per_rpc
<< PAGE_SHIFT
;
920 spin_unlock(&cli
->cl_loi_list_lock
);
922 return osc_shrink_grant_to_target(cli
, target_bytes
);
925 int osc_shrink_grant_to_target(struct client_obd
*cli
, __u64 target_bytes
)
928 struct ost_body
*body
;
930 spin_lock(&cli
->cl_loi_list_lock
);
931 /* Don't shrink if we are already above or below the desired limit
932 * We don't want to shrink below a single RPC, as that will negatively
933 * impact block allocation and long-term performance.
935 if (target_bytes
< cli
->cl_max_pages_per_rpc
<< PAGE_SHIFT
)
936 target_bytes
= cli
->cl_max_pages_per_rpc
<< PAGE_SHIFT
;
938 if (target_bytes
>= cli
->cl_avail_grant
) {
939 spin_unlock(&cli
->cl_loi_list_lock
);
942 spin_unlock(&cli
->cl_loi_list_lock
);
944 body
= kzalloc(sizeof(*body
), GFP_NOFS
);
948 osc_announce_cached(cli
, &body
->oa
, 0);
950 spin_lock(&cli
->cl_loi_list_lock
);
951 body
->oa
.o_grant
= cli
->cl_avail_grant
- target_bytes
;
952 cli
->cl_avail_grant
= target_bytes
;
953 spin_unlock(&cli
->cl_loi_list_lock
);
954 if (!(body
->oa
.o_valid
& OBD_MD_FLFLAGS
)) {
955 body
->oa
.o_valid
|= OBD_MD_FLFLAGS
;
956 body
->oa
.o_flags
= 0;
958 body
->oa
.o_flags
|= OBD_FL_SHRINK_GRANT
;
959 osc_update_next_shrink(cli
);
961 rc
= osc_set_info_async(NULL
, cli
->cl_import
->imp_obd
->obd_self_export
,
962 sizeof(KEY_GRANT_SHRINK
), KEY_GRANT_SHRINK
,
963 sizeof(*body
), body
, NULL
);
965 __osc_update_grant(cli
, body
->oa
.o_grant
);
970 static int osc_should_shrink_grant(struct client_obd
*client
)
972 unsigned long time
= cfs_time_current();
973 unsigned long next_shrink
= client
->cl_next_shrink_grant
;
975 if ((client
->cl_import
->imp_connect_data
.ocd_connect_flags
&
976 OBD_CONNECT_GRANT_SHRINK
) == 0)
979 if (cfs_time_aftereq(time
, next_shrink
- 5 * CFS_TICK
)) {
980 /* Get the current RPC size directly, instead of going via:
981 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
982 * Keep comment here so that it can be found by searching.
984 int brw_size
= client
->cl_max_pages_per_rpc
<< PAGE_SHIFT
;
986 if (client
->cl_import
->imp_state
== LUSTRE_IMP_FULL
&&
987 client
->cl_avail_grant
> brw_size
)
990 osc_update_next_shrink(client
);
995 static int osc_grant_shrink_grant_cb(struct timeout_item
*item
, void *data
)
997 struct client_obd
*client
;
999 list_for_each_entry(client
, &item
->ti_obd_list
, cl_grant_shrink_list
) {
1000 if (osc_should_shrink_grant(client
))
1001 osc_shrink_grant(client
);
1006 static int osc_add_shrink_grant(struct client_obd
*client
)
1010 rc
= ptlrpc_add_timeout_client(client
->cl_grant_shrink_interval
,
1012 osc_grant_shrink_grant_cb
, NULL
,
1013 &client
->cl_grant_shrink_list
);
1015 CERROR("add grant client %s error %d\n",
1016 client
->cl_import
->imp_obd
->obd_name
, rc
);
1019 CDEBUG(D_CACHE
, "add grant client %s\n",
1020 client
->cl_import
->imp_obd
->obd_name
);
1021 osc_update_next_shrink(client
);
1025 static int osc_del_shrink_grant(struct client_obd
*client
)
1027 return ptlrpc_del_timeout_client(&client
->cl_grant_shrink_list
,
1031 static void osc_init_grant(struct client_obd
*cli
, struct obd_connect_data
*ocd
)
1034 * ocd_grant is the total grant amount we're expect to hold: if we've
1035 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1036 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1038 * race is tolerable here: if we're evicted, but imp_state already
1039 * left EVICTED state, then cl_dirty must be 0 already.
1041 spin_lock(&cli
->cl_loi_list_lock
);
1042 if (cli
->cl_import
->imp_state
== LUSTRE_IMP_EVICTED
)
1043 cli
->cl_avail_grant
= ocd
->ocd_grant
;
1045 cli
->cl_avail_grant
= ocd
->ocd_grant
- cli
->cl_dirty
;
1047 if (cli
->cl_avail_grant
< 0) {
1048 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1049 cli
->cl_import
->imp_obd
->obd_name
, cli
->cl_avail_grant
,
1050 ocd
->ocd_grant
, cli
->cl_dirty
);
1051 /* workaround for servers which do not have the patch from
1054 cli
->cl_avail_grant
= ocd
->ocd_grant
;
1057 /* determine the appropriate chunk size used by osc_extent. */
1058 cli
->cl_chunkbits
= max_t(int, PAGE_SHIFT
, ocd
->ocd_blocksize
);
1059 spin_unlock(&cli
->cl_loi_list_lock
);
1061 CDEBUG(D_CACHE
, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld chunk bits: %d\n",
1062 cli
->cl_import
->imp_obd
->obd_name
,
1063 cli
->cl_avail_grant
, cli
->cl_lost_grant
, cli
->cl_chunkbits
);
1065 if (ocd
->ocd_connect_flags
& OBD_CONNECT_GRANT_SHRINK
&&
1066 list_empty(&cli
->cl_grant_shrink_list
))
1067 osc_add_shrink_grant(cli
);
1070 /* We assume that the reason this OSC got a short read is because it read
1071 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1072 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1073 * this stripe never got written at or beyond this stripe offset yet.
1075 static void handle_short_read(int nob_read
, u32 page_count
,
1076 struct brw_page
**pga
)
1081 /* skip bytes read OK */
1082 while (nob_read
> 0) {
1083 LASSERT(page_count
> 0);
1085 if (pga
[i
]->count
> nob_read
) {
1086 /* EOF inside this page */
1087 ptr
= kmap(pga
[i
]->pg
) +
1088 (pga
[i
]->off
& ~PAGE_MASK
);
1089 memset(ptr
+ nob_read
, 0, pga
[i
]->count
- nob_read
);
1096 nob_read
-= pga
[i
]->count
;
1101 /* zero remaining pages */
1102 while (page_count
-- > 0) {
1103 ptr
= kmap(pga
[i
]->pg
) + (pga
[i
]->off
& ~PAGE_MASK
);
1104 memset(ptr
, 0, pga
[i
]->count
);
1110 static int check_write_rcs(struct ptlrpc_request
*req
,
1111 int requested_nob
, int niocount
,
1112 u32 page_count
, struct brw_page
**pga
)
1117 remote_rcs
= req_capsule_server_sized_get(&req
->rq_pill
, &RMF_RCS
,
1118 sizeof(*remote_rcs
) *
1121 CDEBUG(D_INFO
, "Missing/short RC vector on BRW_WRITE reply\n");
1125 /* return error if any niobuf was in error */
1126 for (i
= 0; i
< niocount
; i
++) {
1127 if ((int)remote_rcs
[i
] < 0)
1128 return remote_rcs
[i
];
1130 if (remote_rcs
[i
] != 0) {
1131 CDEBUG(D_INFO
, "rc[%d] invalid (%d) req %p\n",
1132 i
, remote_rcs
[i
], req
);
1137 if (req
->rq_bulk
->bd_nob_transferred
!= requested_nob
) {
1138 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1139 req
->rq_bulk
->bd_nob_transferred
, requested_nob
);
1146 static inline int can_merge_pages(struct brw_page
*p1
, struct brw_page
*p2
)
1148 if (p1
->flag
!= p2
->flag
) {
1149 unsigned mask
= ~(OBD_BRW_FROM_GRANT
| OBD_BRW_NOCACHE
|
1150 OBD_BRW_SYNC
| OBD_BRW_ASYNC
|
1151 OBD_BRW_NOQUOTA
| OBD_BRW_SOFT_SYNC
);
1153 /* warn if we try to combine flags that we don't know to be
1156 if (unlikely((p1
->flag
& mask
) != (p2
->flag
& mask
))) {
1157 CWARN("Saw flags 0x%x and 0x%x in the same brw, please report this at http://bugs.whamcloud.com/\n",
1158 p1
->flag
, p2
->flag
);
1163 return (p1
->off
+ p1
->count
== p2
->off
);
1166 static u32
osc_checksum_bulk(int nob
, u32 pg_count
,
1167 struct brw_page
**pga
, int opc
,
1168 enum cksum_type cksum_type
)
1172 struct cfs_crypto_hash_desc
*hdesc
;
1173 unsigned int bufsize
;
1175 unsigned char cfs_alg
= cksum_obd2cfs(cksum_type
);
1177 LASSERT(pg_count
> 0);
1179 hdesc
= cfs_crypto_hash_init(cfs_alg
, NULL
, 0);
1180 if (IS_ERR(hdesc
)) {
1181 CERROR("Unable to initialize checksum hash %s\n",
1182 cfs_crypto_hash_name(cfs_alg
));
1183 return PTR_ERR(hdesc
);
1186 while (nob
> 0 && pg_count
> 0) {
1187 int count
= pga
[i
]->count
> nob
? nob
: pga
[i
]->count
;
1189 /* corrupt the data before we compute the checksum, to
1190 * simulate an OST->client data error
1192 if (i
== 0 && opc
== OST_READ
&&
1193 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE
)) {
1194 unsigned char *ptr
= kmap(pga
[i
]->pg
);
1195 int off
= pga
[i
]->off
& ~PAGE_MASK
;
1197 memcpy(ptr
+ off
, "bad1", min(4, nob
));
1200 cfs_crypto_hash_update_page(hdesc
, pga
[i
]->pg
,
1201 pga
[i
]->off
& ~PAGE_MASK
,
1204 "page %p map %p index %lu flags %lx count %u priv %0lx: off %d\n",
1205 pga
[i
]->pg
, pga
[i
]->pg
->mapping
, pga
[i
]->pg
->index
,
1206 (long)pga
[i
]->pg
->flags
, page_count(pga
[i
]->pg
),
1207 page_private(pga
[i
]->pg
),
1208 (int)(pga
[i
]->off
& ~PAGE_MASK
));
1210 nob
-= pga
[i
]->count
;
1215 bufsize
= sizeof(cksum
);
1216 err
= cfs_crypto_hash_final(hdesc
, (unsigned char *)&cksum
, &bufsize
);
1218 /* For sending we only compute the wrong checksum instead
1219 * of corrupting the data so it is still correct on a redo
1221 if (opc
== OST_WRITE
&& OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND
))
1227 static int osc_brw_prep_request(int cmd
, struct client_obd
*cli
,
1229 struct lov_stripe_md
*lsm
, u32 page_count
,
1230 struct brw_page
**pga
,
1231 struct ptlrpc_request
**reqp
,
1235 struct ptlrpc_request
*req
;
1236 struct ptlrpc_bulk_desc
*desc
;
1237 struct ost_body
*body
;
1238 struct obd_ioobj
*ioobj
;
1239 struct niobuf_remote
*niobuf
;
1240 int niocount
, i
, requested_nob
, opc
, rc
;
1241 struct osc_brw_async_args
*aa
;
1242 struct req_capsule
*pill
;
1243 struct brw_page
*pg_prev
;
1245 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ
))
1246 return -ENOMEM
; /* Recoverable */
1247 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2
))
1248 return -EINVAL
; /* Fatal */
1250 if ((cmd
& OBD_BRW_WRITE
) != 0) {
1252 req
= ptlrpc_request_alloc_pool(cli
->cl_import
,
1254 &RQF_OST_BRW_WRITE
);
1257 req
= ptlrpc_request_alloc(cli
->cl_import
, &RQF_OST_BRW_READ
);
1262 for (niocount
= i
= 1; i
< page_count
; i
++) {
1263 if (!can_merge_pages(pga
[i
- 1], pga
[i
]))
1267 pill
= &req
->rq_pill
;
1268 req_capsule_set_size(pill
, &RMF_OBD_IOOBJ
, RCL_CLIENT
,
1270 req_capsule_set_size(pill
, &RMF_NIOBUF_REMOTE
, RCL_CLIENT
,
1271 niocount
* sizeof(*niobuf
));
1273 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, opc
);
1275 ptlrpc_request_free(req
);
1278 req
->rq_request_portal
= OST_IO_PORTAL
; /* bug 7198 */
1279 ptlrpc_at_set_req_timeout(req
);
1280 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1283 req
->rq_no_retry_einprogress
= 1;
1285 desc
= ptlrpc_prep_bulk_imp(req
, page_count
,
1286 cli
->cl_import
->imp_connect_data
.ocd_brw_size
>> LNET_MTU_BITS
,
1287 opc
== OST_WRITE
? BULK_GET_SOURCE
: BULK_PUT_SINK
,
1294 /* NB request now owns desc and will free it when it gets freed */
1296 body
= req_capsule_client_get(pill
, &RMF_OST_BODY
);
1297 ioobj
= req_capsule_client_get(pill
, &RMF_OBD_IOOBJ
);
1298 niobuf
= req_capsule_client_get(pill
, &RMF_NIOBUF_REMOTE
);
1299 LASSERT(body
&& ioobj
&& niobuf
);
1301 lustre_set_wire_obdo(&req
->rq_import
->imp_connect_data
, &body
->oa
, oa
);
1303 obdo_to_ioobj(oa
, ioobj
);
1304 ioobj
->ioo_bufcnt
= niocount
;
1305 /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1306 * that might be send for this request. The actual number is decided
1307 * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1308 * "max - 1" for old client compatibility sending "0", and also so the
1309 * the actual maximum is a power-of-two number, not one less. LU-1431
1311 ioobj_max_brw_set(ioobj
, desc
->bd_md_max_brw
);
1312 LASSERT(page_count
> 0);
1314 for (requested_nob
= i
= 0; i
< page_count
; i
++, niobuf
++) {
1315 struct brw_page
*pg
= pga
[i
];
1316 int poff
= pg
->off
& ~PAGE_MASK
;
1318 LASSERT(pg
->count
> 0);
1319 /* make sure there is no gap in the middle of page array */
1320 LASSERTF(page_count
== 1 ||
1321 (ergo(i
== 0, poff
+ pg
->count
== PAGE_SIZE
) &&
1322 ergo(i
> 0 && i
< page_count
- 1,
1323 poff
== 0 && pg
->count
== PAGE_SIZE
) &&
1324 ergo(i
== page_count
- 1, poff
== 0)),
1325 "i: %d/%d pg: %p off: %llu, count: %u\n",
1326 i
, page_count
, pg
, pg
->off
, pg
->count
);
1327 LASSERTF(i
== 0 || pg
->off
> pg_prev
->off
,
1328 "i %d p_c %u pg %p [pri %lu ind %lu] off %llu prev_pg %p [pri %lu ind %lu] off %llu\n",
1330 pg
->pg
, page_private(pg
->pg
), pg
->pg
->index
, pg
->off
,
1331 pg_prev
->pg
, page_private(pg_prev
->pg
),
1332 pg_prev
->pg
->index
, pg_prev
->off
);
1333 LASSERT((pga
[0]->flag
& OBD_BRW_SRVLOCK
) ==
1334 (pg
->flag
& OBD_BRW_SRVLOCK
));
1336 ptlrpc_prep_bulk_page_pin(desc
, pg
->pg
, poff
, pg
->count
);
1337 requested_nob
+= pg
->count
;
1339 if (i
> 0 && can_merge_pages(pg_prev
, pg
)) {
1341 niobuf
->len
+= pg
->count
;
1343 niobuf
->offset
= pg
->off
;
1344 niobuf
->len
= pg
->count
;
1345 niobuf
->flags
= pg
->flag
;
1350 LASSERTF((void *)(niobuf
- niocount
) ==
1351 req_capsule_client_get(&req
->rq_pill
, &RMF_NIOBUF_REMOTE
),
1352 "want %p - real %p\n", req_capsule_client_get(&req
->rq_pill
,
1353 &RMF_NIOBUF_REMOTE
), (void *)(niobuf
- niocount
));
1355 osc_announce_cached(cli
, &body
->oa
, opc
== OST_WRITE
? requested_nob
:0);
1357 if ((body
->oa
.o_valid
& OBD_MD_FLFLAGS
) == 0) {
1358 body
->oa
.o_valid
|= OBD_MD_FLFLAGS
;
1359 body
->oa
.o_flags
= 0;
1361 body
->oa
.o_flags
|= OBD_FL_RECOV_RESEND
;
1364 if (osc_should_shrink_grant(cli
))
1365 osc_shrink_grant_local(cli
, &body
->oa
);
1367 /* size[REQ_REC_OFF] still sizeof (*body) */
1368 if (opc
== OST_WRITE
) {
1369 if (cli
->cl_checksum
&&
1370 !sptlrpc_flavor_has_bulk(&req
->rq_flvr
)) {
1371 /* store cl_cksum_type in a local variable since
1372 * it can be changed via lprocfs
1374 enum cksum_type cksum_type
= cli
->cl_cksum_type
;
1376 if ((body
->oa
.o_valid
& OBD_MD_FLFLAGS
) == 0) {
1377 oa
->o_flags
&= OBD_FL_LOCAL_MASK
;
1378 body
->oa
.o_flags
= 0;
1380 body
->oa
.o_flags
|= cksum_type_pack(cksum_type
);
1381 body
->oa
.o_valid
|= OBD_MD_FLCKSUM
| OBD_MD_FLFLAGS
;
1382 body
->oa
.o_cksum
= osc_checksum_bulk(requested_nob
,
1386 CDEBUG(D_PAGE
, "checksum at write origin: %x\n",
1388 /* save this in 'oa', too, for later checking */
1389 oa
->o_valid
|= OBD_MD_FLCKSUM
| OBD_MD_FLFLAGS
;
1390 oa
->o_flags
|= cksum_type_pack(cksum_type
);
1392 /* clear out the checksum flag, in case this is a
1393 * resend but cl_checksum is no longer set. b=11238
1395 oa
->o_valid
&= ~OBD_MD_FLCKSUM
;
1397 oa
->o_cksum
= body
->oa
.o_cksum
;
1398 /* 1 RC per niobuf */
1399 req_capsule_set_size(pill
, &RMF_RCS
, RCL_SERVER
,
1400 sizeof(__u32
) * niocount
);
1402 if (cli
->cl_checksum
&&
1403 !sptlrpc_flavor_has_bulk(&req
->rq_flvr
)) {
1404 if ((body
->oa
.o_valid
& OBD_MD_FLFLAGS
) == 0)
1405 body
->oa
.o_flags
= 0;
1406 body
->oa
.o_flags
|= cksum_type_pack(cli
->cl_cksum_type
);
1407 body
->oa
.o_valid
|= OBD_MD_FLCKSUM
| OBD_MD_FLFLAGS
;
1410 ptlrpc_request_set_replen(req
);
1412 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
1413 aa
= ptlrpc_req_async_args(req
);
1415 aa
->aa_requested_nob
= requested_nob
;
1416 aa
->aa_nio_count
= niocount
;
1417 aa
->aa_page_count
= page_count
;
1421 INIT_LIST_HEAD(&aa
->aa_oaps
);
1427 ptlrpc_req_finished(req
);
1431 static int check_write_checksum(struct obdo
*oa
, const lnet_process_id_t
*peer
,
1432 __u32 client_cksum
, __u32 server_cksum
, int nob
,
1433 u32 page_count
, struct brw_page
**pga
,
1434 enum cksum_type client_cksum_type
)
1438 enum cksum_type cksum_type
;
1440 if (server_cksum
== client_cksum
) {
1441 CDEBUG(D_PAGE
, "checksum %x confirmed\n", client_cksum
);
1445 cksum_type
= cksum_type_unpack(oa
->o_valid
& OBD_MD_FLFLAGS
?
1447 new_cksum
= osc_checksum_bulk(nob
, page_count
, pga
, OST_WRITE
,
1450 if (cksum_type
!= client_cksum_type
)
1451 msg
= "the server did not use the checksum type specified in the original request - likely a protocol problem"
1453 else if (new_cksum
== server_cksum
)
1454 msg
= "changed on the client after we checksummed it - likely false positive due to mmap IO (bug 11742)"
1456 else if (new_cksum
== client_cksum
)
1457 msg
= "changed in transit before arrival at OST";
1459 msg
= "changed in transit AND doesn't match the original - likely false positive due to mmap IO (bug 11742)"
1462 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1463 " object "DOSTID
" extent [%llu-%llu]\n",
1464 msg
, libcfs_nid2str(peer
->nid
),
1465 oa
->o_valid
& OBD_MD_FLFID
? oa
->o_parent_seq
: (__u64
)0,
1466 oa
->o_valid
& OBD_MD_FLFID
? oa
->o_parent_oid
: 0,
1467 oa
->o_valid
& OBD_MD_FLFID
? oa
->o_parent_ver
: 0,
1468 POSTID(&oa
->o_oi
), pga
[0]->off
,
1469 pga
[page_count
-1]->off
+ pga
[page_count
-1]->count
- 1);
1470 CERROR("original client csum %x (type %x), server csum %x (type %x), client csum now %x\n",
1471 client_cksum
, client_cksum_type
,
1472 server_cksum
, cksum_type
, new_cksum
);
1476 /* Note rc enters this function as number of bytes transferred */
1477 static int osc_brw_fini_request(struct ptlrpc_request
*req
, int rc
)
1479 struct osc_brw_async_args
*aa
= (void *)&req
->rq_async_args
;
1480 const lnet_process_id_t
*peer
=
1481 &req
->rq_import
->imp_connection
->c_peer
;
1482 struct client_obd
*cli
= aa
->aa_cli
;
1483 struct ost_body
*body
;
1484 __u32 client_cksum
= 0;
1486 if (rc
< 0 && rc
!= -EDQUOT
) {
1487 DEBUG_REQ(D_INFO
, req
, "Failed request with rc = %d\n", rc
);
1491 LASSERTF(req
->rq_repmsg
, "rc = %d\n", rc
);
1492 body
= req_capsule_server_get(&req
->rq_pill
, &RMF_OST_BODY
);
1494 DEBUG_REQ(D_INFO
, req
, "Can't unpack body\n");
1498 /* set/clear over quota flag for a uid/gid */
1499 if (lustre_msg_get_opc(req
->rq_reqmsg
) == OST_WRITE
&&
1500 body
->oa
.o_valid
& (OBD_MD_FLUSRQUOTA
| OBD_MD_FLGRPQUOTA
)) {
1501 unsigned int qid
[MAXQUOTAS
] = { body
->oa
.o_uid
, body
->oa
.o_gid
};
1503 CDEBUG(D_QUOTA
, "setdq for [%u %u] with valid %#llx, flags %x\n",
1504 body
->oa
.o_uid
, body
->oa
.o_gid
, body
->oa
.o_valid
,
1506 osc_quota_setdq(cli
, qid
, body
->oa
.o_valid
, body
->oa
.o_flags
);
1509 osc_update_grant(cli
, body
);
1514 if (aa
->aa_oa
->o_valid
& OBD_MD_FLCKSUM
)
1515 client_cksum
= aa
->aa_oa
->o_cksum
; /* save for later */
1517 if (lustre_msg_get_opc(req
->rq_reqmsg
) == OST_WRITE
) {
1519 CERROR("Unexpected +ve rc %d\n", rc
);
1522 LASSERT(req
->rq_bulk
->bd_nob
== aa
->aa_requested_nob
);
1524 if (sptlrpc_cli_unwrap_bulk_write(req
, req
->rq_bulk
))
1527 if ((aa
->aa_oa
->o_valid
& OBD_MD_FLCKSUM
) && client_cksum
&&
1528 check_write_checksum(&body
->oa
, peer
, client_cksum
,
1529 body
->oa
.o_cksum
, aa
->aa_requested_nob
,
1530 aa
->aa_page_count
, aa
->aa_ppga
,
1531 cksum_type_unpack(aa
->aa_oa
->o_flags
)))
1534 rc
= check_write_rcs(req
, aa
->aa_requested_nob
,
1536 aa
->aa_page_count
, aa
->aa_ppga
);
1540 /* The rest of this function executes only for OST_READs */
1542 /* if unwrap_bulk failed, return -EAGAIN to retry */
1543 rc
= sptlrpc_cli_unwrap_bulk_read(req
, req
->rq_bulk
, rc
);
1549 if (rc
> aa
->aa_requested_nob
) {
1550 CERROR("Unexpected rc %d (%d requested)\n", rc
,
1551 aa
->aa_requested_nob
);
1555 if (rc
!= req
->rq_bulk
->bd_nob_transferred
) {
1556 CERROR("Unexpected rc %d (%d transferred)\n",
1557 rc
, req
->rq_bulk
->bd_nob_transferred
);
1561 if (rc
< aa
->aa_requested_nob
)
1562 handle_short_read(rc
, aa
->aa_page_count
, aa
->aa_ppga
);
1564 if (body
->oa
.o_valid
& OBD_MD_FLCKSUM
) {
1565 static int cksum_counter
;
1566 __u32 server_cksum
= body
->oa
.o_cksum
;
1569 enum cksum_type cksum_type
;
1571 cksum_type
= cksum_type_unpack(body
->oa
.o_valid
&OBD_MD_FLFLAGS
?
1572 body
->oa
.o_flags
: 0);
1573 client_cksum
= osc_checksum_bulk(rc
, aa
->aa_page_count
,
1574 aa
->aa_ppga
, OST_READ
,
1577 if (peer
->nid
!= req
->rq_bulk
->bd_sender
) {
1579 router
= libcfs_nid2str(req
->rq_bulk
->bd_sender
);
1582 if (server_cksum
!= client_cksum
) {
1583 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from %s%s%s inode " DFID
" object " DOSTID
" extent [%llu-%llu]\n",
1584 req
->rq_import
->imp_obd
->obd_name
,
1585 libcfs_nid2str(peer
->nid
),
1587 body
->oa
.o_valid
& OBD_MD_FLFID
?
1588 body
->oa
.o_parent_seq
: (__u64
)0,
1589 body
->oa
.o_valid
& OBD_MD_FLFID
?
1590 body
->oa
.o_parent_oid
: 0,
1591 body
->oa
.o_valid
& OBD_MD_FLFID
?
1592 body
->oa
.o_parent_ver
: 0,
1593 POSTID(&body
->oa
.o_oi
),
1594 aa
->aa_ppga
[0]->off
,
1595 aa
->aa_ppga
[aa
->aa_page_count
-1]->off
+
1596 aa
->aa_ppga
[aa
->aa_page_count
-1]->count
-
1598 CERROR("client %x, server %x, cksum_type %x\n",
1599 client_cksum
, server_cksum
, cksum_type
);
1601 aa
->aa_oa
->o_cksum
= client_cksum
;
1605 CDEBUG(D_PAGE
, "checksum %x confirmed\n", client_cksum
);
1608 } else if (unlikely(client_cksum
)) {
1609 static int cksum_missed
;
1612 if ((cksum_missed
& (-cksum_missed
)) == cksum_missed
)
1613 CERROR("Checksum %u requested from %s but not sent\n",
1614 cksum_missed
, libcfs_nid2str(peer
->nid
));
1620 lustre_get_wire_obdo(&req
->rq_import
->imp_connect_data
,
1621 aa
->aa_oa
, &body
->oa
);
1626 static int osc_brw_redo_request(struct ptlrpc_request
*request
,
1627 struct osc_brw_async_args
*aa
, int rc
)
1629 struct ptlrpc_request
*new_req
;
1630 struct osc_brw_async_args
*new_aa
;
1631 struct osc_async_page
*oap
;
1633 DEBUG_REQ(rc
== -EINPROGRESS
? D_RPCTRACE
: D_ERROR
, request
,
1634 "redo for recoverable error %d", rc
);
1636 rc
= osc_brw_prep_request(lustre_msg_get_opc(request
->rq_reqmsg
) ==
1637 OST_WRITE
? OBD_BRW_WRITE
: OBD_BRW_READ
,
1638 aa
->aa_cli
, aa
->aa_oa
,
1639 NULL
/* lsm unused by osc currently */,
1640 aa
->aa_page_count
, aa
->aa_ppga
,
1645 list_for_each_entry(oap
, &aa
->aa_oaps
, oap_rpc_item
) {
1646 if (oap
->oap_request
) {
1647 LASSERTF(request
== oap
->oap_request
,
1648 "request %p != oap_request %p\n",
1649 request
, oap
->oap_request
);
1650 if (oap
->oap_interrupted
) {
1651 ptlrpc_req_finished(new_req
);
1656 /* New request takes over pga and oaps from old request.
1657 * Note that copying a list_head doesn't work, need to move it...
1660 new_req
->rq_interpret_reply
= request
->rq_interpret_reply
;
1661 new_req
->rq_async_args
= request
->rq_async_args
;
1662 new_req
->rq_commit_cb
= request
->rq_commit_cb
;
1663 /* cap resend delay to the current request timeout, this is similar to
1664 * what ptlrpc does (see after_reply())
1666 if (aa
->aa_resends
> new_req
->rq_timeout
)
1667 new_req
->rq_sent
= ktime_get_real_seconds() + new_req
->rq_timeout
;
1669 new_req
->rq_sent
= ktime_get_real_seconds() + aa
->aa_resends
;
1670 new_req
->rq_generation_set
= 1;
1671 new_req
->rq_import_generation
= request
->rq_import_generation
;
1673 new_aa
= ptlrpc_req_async_args(new_req
);
1675 INIT_LIST_HEAD(&new_aa
->aa_oaps
);
1676 list_splice_init(&aa
->aa_oaps
, &new_aa
->aa_oaps
);
1677 INIT_LIST_HEAD(&new_aa
->aa_exts
);
1678 list_splice_init(&aa
->aa_exts
, &new_aa
->aa_exts
);
1679 new_aa
->aa_resends
= aa
->aa_resends
;
1681 list_for_each_entry(oap
, &new_aa
->aa_oaps
, oap_rpc_item
) {
1682 if (oap
->oap_request
) {
1683 ptlrpc_req_finished(oap
->oap_request
);
1684 oap
->oap_request
= ptlrpc_request_addref(new_req
);
1688 /* XXX: This code will run into problem if we're going to support
1689 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1690 * and wait for all of them to be finished. We should inherit request
1691 * set from old request.
1693 ptlrpcd_add_req(new_req
);
1695 DEBUG_REQ(D_INFO
, new_req
, "new request");
1700 * ugh, we want disk allocation on the target to happen in offset order. we'll
1701 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1702 * fine for our small page arrays and doesn't require allocation. its an
1703 * insertion sort that swaps elements that are strides apart, shrinking the
1704 * stride down until its '1' and the array is sorted.
1706 static void sort_brw_pages(struct brw_page
**array
, int num
)
1709 struct brw_page
*tmp
;
1713 for (stride
= 1; stride
< num
; stride
= (stride
* 3) + 1)
1718 for (i
= stride
; i
< num
; i
++) {
1721 while (j
>= stride
&& array
[j
- stride
]->off
> tmp
->off
) {
1722 array
[j
] = array
[j
- stride
];
1727 } while (stride
> 1);
1730 static void osc_release_ppga(struct brw_page
**ppga
, u32 count
)
1736 static int brw_interpret(const struct lu_env
*env
,
1737 struct ptlrpc_request
*req
, void *data
, int rc
)
1739 struct osc_brw_async_args
*aa
= data
;
1740 struct osc_extent
*ext
;
1741 struct osc_extent
*tmp
;
1742 struct client_obd
*cli
= aa
->aa_cli
;
1744 rc
= osc_brw_fini_request(req
, rc
);
1745 CDEBUG(D_INODE
, "request %p aa %p rc %d\n", req
, aa
, rc
);
1746 /* When server return -EINPROGRESS, client should always retry
1747 * regardless of the number of times the bulk was resent already.
1749 if (osc_recoverable_error(rc
)) {
1750 if (req
->rq_import_generation
!=
1751 req
->rq_import
->imp_generation
) {
1752 CDEBUG(D_HA
, "%s: resend cross eviction for object: " DOSTID
", rc = %d.\n",
1753 req
->rq_import
->imp_obd
->obd_name
,
1754 POSTID(&aa
->aa_oa
->o_oi
), rc
);
1755 } else if (rc
== -EINPROGRESS
||
1756 client_should_resend(aa
->aa_resends
, aa
->aa_cli
)) {
1757 rc
= osc_brw_redo_request(req
, aa
, rc
);
1759 CERROR("%s: too many resent retries for object: %llu:%llu, rc = %d.\n",
1760 req
->rq_import
->imp_obd
->obd_name
,
1761 POSTID(&aa
->aa_oa
->o_oi
), rc
);
1766 else if (rc
== -EAGAIN
|| rc
== -EINPROGRESS
)
1771 struct obdo
*oa
= aa
->aa_oa
;
1772 struct cl_attr
*attr
= &osc_env_info(env
)->oti_attr
;
1773 unsigned long valid
= 0;
1774 struct cl_object
*obj
;
1775 struct osc_async_page
*last
;
1777 last
= brw_page2oap(aa
->aa_ppga
[aa
->aa_page_count
- 1]);
1778 obj
= osc2cl(last
->oap_obj
);
1780 cl_object_attr_lock(obj
);
1781 if (oa
->o_valid
& OBD_MD_FLBLOCKS
) {
1782 attr
->cat_blocks
= oa
->o_blocks
;
1783 valid
|= CAT_BLOCKS
;
1785 if (oa
->o_valid
& OBD_MD_FLMTIME
) {
1786 attr
->cat_mtime
= oa
->o_mtime
;
1789 if (oa
->o_valid
& OBD_MD_FLATIME
) {
1790 attr
->cat_atime
= oa
->o_atime
;
1793 if (oa
->o_valid
& OBD_MD_FLCTIME
) {
1794 attr
->cat_ctime
= oa
->o_ctime
;
1798 if (lustre_msg_get_opc(req
->rq_reqmsg
) == OST_WRITE
) {
1799 struct lov_oinfo
*loi
= cl2osc(obj
)->oo_oinfo
;
1800 loff_t last_off
= last
->oap_count
+ last
->oap_obj_off
;
1802 /* Change file size if this is an out of quota or
1803 * direct IO write and it extends the file size
1805 if (loi
->loi_lvb
.lvb_size
< last_off
) {
1806 attr
->cat_size
= last_off
;
1809 /* Extend KMS if it's not a lockless write */
1810 if (loi
->loi_kms
< last_off
&&
1811 oap2osc_page(last
)->ops_srvlock
== 0) {
1812 attr
->cat_kms
= last_off
;
1818 cl_object_attr_set(env
, obj
, attr
, valid
);
1819 cl_object_attr_unlock(obj
);
1821 kmem_cache_free(obdo_cachep
, aa
->aa_oa
);
1823 list_for_each_entry_safe(ext
, tmp
, &aa
->aa_exts
, oe_link
) {
1824 list_del_init(&ext
->oe_link
);
1825 osc_extent_finish(env
, ext
, 1, rc
);
1827 LASSERT(list_empty(&aa
->aa_exts
));
1828 LASSERT(list_empty(&aa
->aa_oaps
));
1830 cl_req_completion(env
, aa
->aa_clerq
, rc
< 0 ? rc
:
1831 req
->rq_bulk
->bd_nob_transferred
);
1832 osc_release_ppga(aa
->aa_ppga
, aa
->aa_page_count
);
1833 ptlrpc_lprocfs_brw(req
, req
->rq_bulk
->bd_nob_transferred
);
1835 spin_lock(&cli
->cl_loi_list_lock
);
1836 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1837 * is called so we know whether to go to sync BRWs or wait for more
1840 if (lustre_msg_get_opc(req
->rq_reqmsg
) == OST_WRITE
)
1841 cli
->cl_w_in_flight
--;
1843 cli
->cl_r_in_flight
--;
1844 osc_wake_cache_waiters(cli
);
1845 spin_unlock(&cli
->cl_loi_list_lock
);
1847 osc_io_unplug(env
, cli
, NULL
);
1851 static void brw_commit(struct ptlrpc_request
*req
)
1853 spin_lock(&req
->rq_lock
);
1855 * If osc_inc_unstable_pages (via osc_extent_finish) races with
1856 * this called via the rq_commit_cb, I need to ensure
1857 * osc_dec_unstable_pages is still called. Otherwise unstable
1858 * pages may be leaked.
1860 if (req
->rq_unstable
) {
1861 spin_unlock(&req
->rq_lock
);
1862 osc_dec_unstable_pages(req
);
1863 spin_lock(&req
->rq_lock
);
1865 req
->rq_committed
= 1;
1867 spin_unlock(&req
->rq_lock
);
1871 * Build an RPC by the list of extent @ext_list. The caller must ensure
1872 * that the total pages in this list are NOT over max pages per RPC.
1873 * Extents in the list must be in OES_RPC state.
1875 int osc_build_rpc(const struct lu_env
*env
, struct client_obd
*cli
,
1876 struct list_head
*ext_list
, int cmd
)
1878 struct ptlrpc_request
*req
= NULL
;
1879 struct osc_extent
*ext
;
1880 struct brw_page
**pga
= NULL
;
1881 struct osc_brw_async_args
*aa
= NULL
;
1882 struct obdo
*oa
= NULL
;
1883 struct osc_async_page
*oap
;
1884 struct osc_async_page
*tmp
;
1885 struct cl_req
*clerq
= NULL
;
1886 enum cl_req_type crt
= (cmd
& OBD_BRW_WRITE
) ? CRT_WRITE
: CRT_READ
;
1887 struct ldlm_lock
*lock
= NULL
;
1888 struct cl_req_attr
*crattr
= NULL
;
1889 u64 starting_offset
= OBD_OBJECT_EOF
;
1890 u64 ending_offset
= 0;
1896 struct ost_body
*body
;
1897 LIST_HEAD(rpc_list
);
1899 LASSERT(!list_empty(ext_list
));
1901 /* add pages into rpc_list to build BRW rpc */
1902 list_for_each_entry(ext
, ext_list
, oe_link
) {
1903 LASSERT(ext
->oe_state
== OES_RPC
);
1904 mem_tight
|= ext
->oe_memalloc
;
1905 list_for_each_entry(oap
, &ext
->oe_pages
, oap_pending_item
) {
1907 list_add_tail(&oap
->oap_rpc_item
, &rpc_list
);
1908 if (starting_offset
> oap
->oap_obj_off
)
1909 starting_offset
= oap
->oap_obj_off
;
1911 LASSERT(oap
->oap_page_off
== 0);
1912 if (ending_offset
< oap
->oap_obj_off
+ oap
->oap_count
)
1913 ending_offset
= oap
->oap_obj_off
+
1916 LASSERT(oap
->oap_page_off
+ oap
->oap_count
==
1922 mpflag
= cfs_memory_pressure_get_and_set();
1924 crattr
= kzalloc(sizeof(*crattr
), GFP_NOFS
);
1930 pga
= kcalloc(page_count
, sizeof(*pga
), GFP_NOFS
);
1936 oa
= kmem_cache_zalloc(obdo_cachep
, GFP_NOFS
);
1943 list_for_each_entry(oap
, &rpc_list
, oap_rpc_item
) {
1944 struct cl_page
*page
= oap2cl_page(oap
);
1947 clerq
= cl_req_alloc(env
, page
, crt
,
1948 1 /* only 1-object rpcs for now */);
1949 if (IS_ERR(clerq
)) {
1950 rc
= PTR_ERR(clerq
);
1953 lock
= oap
->oap_ldlm_lock
;
1956 oap
->oap_brw_flags
|= OBD_BRW_MEMALLOC
;
1957 pga
[i
] = &oap
->oap_brw_page
;
1958 pga
[i
]->off
= oap
->oap_obj_off
+ oap
->oap_page_off
;
1959 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1960 pga
[i
]->pg
, oap
->oap_page
->index
, oap
,
1963 cl_req_page_add(env
, clerq
, page
);
1966 /* always get the data for the obdo for the rpc */
1968 crattr
->cra_oa
= oa
;
1969 cl_req_attr_set(env
, clerq
, crattr
, ~0ULL);
1971 oa
->o_handle
= lock
->l_remote_handle
;
1972 oa
->o_valid
|= OBD_MD_FLHANDLE
;
1975 rc
= cl_req_prep(env
, clerq
);
1977 CERROR("cl_req_prep failed: %d\n", rc
);
1981 sort_brw_pages(pga
, page_count
);
1982 rc
= osc_brw_prep_request(cmd
, cli
, oa
, NULL
, page_count
,
1985 CERROR("prep_req failed: %d\n", rc
);
1989 req
->rq_commit_cb
= brw_commit
;
1990 req
->rq_interpret_reply
= brw_interpret
;
1993 req
->rq_memalloc
= 1;
1995 /* Need to update the timestamps after the request is built in case
1996 * we race with setattr (locally or in queue at OST). If OST gets
1997 * later setattr before earlier BRW (as determined by the request xid),
1998 * the OST will not use BRW timestamps. Sadly, there is no obvious
1999 * way to do this in a single call. bug 10150
2001 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_OST_BODY
);
2002 crattr
->cra_oa
= &body
->oa
;
2003 cl_req_attr_set(env
, clerq
, crattr
,
2004 OBD_MD_FLMTIME
|OBD_MD_FLCTIME
|OBD_MD_FLATIME
);
2006 lustre_msg_set_jobid(req
->rq_reqmsg
, crattr
->cra_jobid
);
2008 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2009 aa
= ptlrpc_req_async_args(req
);
2010 INIT_LIST_HEAD(&aa
->aa_oaps
);
2011 list_splice_init(&rpc_list
, &aa
->aa_oaps
);
2012 INIT_LIST_HEAD(&aa
->aa_exts
);
2013 list_splice_init(ext_list
, &aa
->aa_exts
);
2014 aa
->aa_clerq
= clerq
;
2016 /* queued sync pages can be torn down while the pages
2017 * were between the pending list and the rpc
2020 list_for_each_entry(oap
, &aa
->aa_oaps
, oap_rpc_item
) {
2021 /* only one oap gets a request reference */
2024 if (oap
->oap_interrupted
&& !req
->rq_intr
) {
2025 CDEBUG(D_INODE
, "oap %p in req %p interrupted\n",
2027 ptlrpc_mark_interrupted(req
);
2031 tmp
->oap_request
= ptlrpc_request_addref(req
);
2033 spin_lock(&cli
->cl_loi_list_lock
);
2034 starting_offset
>>= PAGE_SHIFT
;
2035 if (cmd
== OBD_BRW_READ
) {
2036 cli
->cl_r_in_flight
++;
2037 lprocfs_oh_tally_log2(&cli
->cl_read_page_hist
, page_count
);
2038 lprocfs_oh_tally(&cli
->cl_read_rpc_hist
, cli
->cl_r_in_flight
);
2039 lprocfs_oh_tally_log2(&cli
->cl_read_offset_hist
,
2040 starting_offset
+ 1);
2042 cli
->cl_w_in_flight
++;
2043 lprocfs_oh_tally_log2(&cli
->cl_write_page_hist
, page_count
);
2044 lprocfs_oh_tally(&cli
->cl_write_rpc_hist
, cli
->cl_w_in_flight
);
2045 lprocfs_oh_tally_log2(&cli
->cl_write_offset_hist
,
2046 starting_offset
+ 1);
2048 spin_unlock(&cli
->cl_loi_list_lock
);
2050 DEBUG_REQ(D_INODE
, req
, "%d pages, aa %p. now %dr/%dw in flight",
2051 page_count
, aa
, cli
->cl_r_in_flight
,
2052 cli
->cl_w_in_flight
);
2054 ptlrpcd_add_req(req
);
2059 cfs_memory_pressure_restore(mpflag
);
2067 kmem_cache_free(obdo_cachep
, oa
);
2069 /* this should happen rarely and is pretty bad, it makes the
2070 * pending list not follow the dirty order
2072 while (!list_empty(ext_list
)) {
2073 ext
= list_entry(ext_list
->next
, struct osc_extent
,
2075 list_del_init(&ext
->oe_link
);
2076 osc_extent_finish(env
, ext
, 0, rc
);
2078 if (clerq
&& !IS_ERR(clerq
))
2079 cl_req_completion(env
, clerq
, rc
);
2084 static int osc_set_lock_data_with_check(struct ldlm_lock
*lock
,
2085 struct ldlm_enqueue_info
*einfo
)
2087 void *data
= einfo
->ei_cbdata
;
2090 LASSERT(lock
->l_blocking_ast
== einfo
->ei_cb_bl
);
2091 LASSERT(lock
->l_resource
->lr_type
== einfo
->ei_type
);
2092 LASSERT(lock
->l_completion_ast
== einfo
->ei_cb_cp
);
2093 LASSERT(lock
->l_glimpse_ast
== einfo
->ei_cb_gl
);
2095 lock_res_and_lock(lock
);
2097 if (!lock
->l_ast_data
)
2098 lock
->l_ast_data
= data
;
2099 if (lock
->l_ast_data
== data
)
2102 unlock_res_and_lock(lock
);
2107 static int osc_set_data_with_check(struct lustre_handle
*lockh
,
2108 struct ldlm_enqueue_info
*einfo
)
2110 struct ldlm_lock
*lock
= ldlm_handle2lock(lockh
);
2114 set
= osc_set_lock_data_with_check(lock
, einfo
);
2115 LDLM_LOCK_PUT(lock
);
2117 CERROR("lockh %p, data %p - client evicted?\n",
2118 lockh
, einfo
->ei_cbdata
);
2122 /* find any ldlm lock of the inode in osc
2127 static int osc_find_cbdata(struct obd_export
*exp
, struct lov_stripe_md
*lsm
,
2128 ldlm_iterator_t replace
, void *data
)
2130 struct ldlm_res_id res_id
;
2131 struct obd_device
*obd
= class_exp2obd(exp
);
2134 ostid_build_res_name(&lsm
->lsm_oi
, &res_id
);
2135 rc
= ldlm_resource_iterate(obd
->obd_namespace
, &res_id
, replace
, data
);
2136 if (rc
== LDLM_ITER_STOP
)
2138 if (rc
== LDLM_ITER_CONTINUE
)
2143 static int osc_enqueue_fini(struct ptlrpc_request
*req
,
2144 osc_enqueue_upcall_f upcall
, void *cookie
,
2145 struct lustre_handle
*lockh
, enum ldlm_mode mode
,
2146 __u64
*flags
, int agl
, int errcode
)
2148 bool intent
= *flags
& LDLM_FL_HAS_INTENT
;
2151 /* The request was created before ldlm_cli_enqueue call. */
2152 if (intent
&& errcode
== ELDLM_LOCK_ABORTED
) {
2153 struct ldlm_reply
*rep
;
2155 rep
= req_capsule_server_get(&req
->rq_pill
, &RMF_DLM_REP
);
2157 rep
->lock_policy_res1
=
2158 ptlrpc_status_ntoh(rep
->lock_policy_res1
);
2159 if (rep
->lock_policy_res1
)
2160 errcode
= rep
->lock_policy_res1
;
2162 *flags
|= LDLM_FL_LVB_READY
;
2163 } else if (errcode
== ELDLM_OK
) {
2164 *flags
|= LDLM_FL_LVB_READY
;
2167 /* Call the update callback. */
2168 rc
= (*upcall
)(cookie
, lockh
, errcode
);
2169 /* release the reference taken in ldlm_cli_enqueue() */
2170 if (errcode
== ELDLM_LOCK_MATCHED
)
2172 if (errcode
== ELDLM_OK
&& lustre_handle_is_used(lockh
))
2173 ldlm_lock_decref(lockh
, mode
);
2178 static int osc_enqueue_interpret(const struct lu_env
*env
,
2179 struct ptlrpc_request
*req
,
2180 struct osc_enqueue_args
*aa
, int rc
)
2182 struct ldlm_lock
*lock
;
2183 struct lustre_handle
*lockh
= &aa
->oa_lockh
;
2184 enum ldlm_mode mode
= aa
->oa_mode
;
2185 struct ost_lvb
*lvb
= aa
->oa_lvb
;
2186 __u32 lvb_len
= sizeof(*lvb
);
2190 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2193 lock
= ldlm_handle2lock(lockh
);
2194 LASSERTF(lock
, "lockh %llx, req %p, aa %p - client evicted?\n",
2195 lockh
->cookie
, req
, aa
);
2197 /* Take an additional reference so that a blocking AST that
2198 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2199 * to arrive after an upcall has been executed by
2200 * osc_enqueue_fini().
2202 ldlm_lock_addref(lockh
, mode
);
2204 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2205 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG
, 2);
2207 /* Let CP AST to grant the lock first. */
2208 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE
, 1);
2211 LASSERT(!aa
->oa_lvb
);
2212 LASSERT(!aa
->oa_flags
);
2213 aa
->oa_flags
= &flags
;
2216 /* Complete obtaining the lock procedure. */
2217 rc
= ldlm_cli_enqueue_fini(aa
->oa_exp
, req
, aa
->oa_type
, 1,
2218 aa
->oa_mode
, aa
->oa_flags
, lvb
, lvb_len
,
2220 /* Complete osc stuff. */
2221 rc
= osc_enqueue_fini(req
, aa
->oa_upcall
, aa
->oa_cookie
, lockh
, mode
,
2222 aa
->oa_flags
, aa
->oa_agl
, rc
);
2224 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE
, 10);
2226 ldlm_lock_decref(lockh
, mode
);
2227 LDLM_LOCK_PUT(lock
);
2231 struct ptlrpc_request_set
*PTLRPCD_SET
= (void *)1;
2233 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2234 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2235 * other synchronous requests, however keeping some locks and trying to obtain
2236 * others may take a considerable amount of time in a case of ost failure; and
2237 * when other sync requests do not get released lock from a client, the client
2238 * is evicted from the cluster -- such scenaries make the life difficult, so
2239 * release locks just after they are obtained.
2241 int osc_enqueue_base(struct obd_export
*exp
, struct ldlm_res_id
*res_id
,
2242 __u64
*flags
, ldlm_policy_data_t
*policy
,
2243 struct ost_lvb
*lvb
, int kms_valid
,
2244 osc_enqueue_upcall_f upcall
, void *cookie
,
2245 struct ldlm_enqueue_info
*einfo
,
2246 struct ptlrpc_request_set
*rqset
, int async
, int agl
)
2248 struct obd_device
*obd
= exp
->exp_obd
;
2249 struct lustre_handle lockh
= { 0 };
2250 struct ptlrpc_request
*req
= NULL
;
2251 int intent
= *flags
& LDLM_FL_HAS_INTENT
;
2252 __u64 match_lvb
= agl
? 0 : LDLM_FL_LVB_READY
;
2253 enum ldlm_mode mode
;
2256 /* Filesystem lock extents are extended to page boundaries so that
2257 * dealing with the page cache is a little smoother.
2259 policy
->l_extent
.start
-= policy
->l_extent
.start
& ~PAGE_MASK
;
2260 policy
->l_extent
.end
|= ~PAGE_MASK
;
2263 * kms is not valid when either object is completely fresh (so that no
2264 * locks are cached), or object was evicted. In the latter case cached
2265 * lock cannot be used, because it would prime inode state with
2266 * potentially stale LVB.
2271 /* Next, search for already existing extent locks that will cover us */
2272 /* If we're trying to read, we also search for an existing PW lock. The
2273 * VFS and page cache already protect us locally, so lots of readers/
2274 * writers can share a single PW lock.
2276 * There are problems with conversion deadlocks, so instead of
2277 * converting a read lock to a write lock, we'll just enqueue a new
2280 * At some point we should cancel the read lock instead of making them
2281 * send us a blocking callback, but there are problems with canceling
2282 * locks out from other users right now, too.
2284 mode
= einfo
->ei_mode
;
2285 if (einfo
->ei_mode
== LCK_PR
)
2287 mode
= ldlm_lock_match(obd
->obd_namespace
, *flags
| match_lvb
, res_id
,
2288 einfo
->ei_type
, policy
, mode
, &lockh
, 0);
2290 struct ldlm_lock
*matched
;
2292 if (*flags
& LDLM_FL_TEST_LOCK
)
2295 matched
= ldlm_handle2lock(&lockh
);
2297 /* AGL enqueues DLM locks speculatively. Therefore if
2298 * it already exists a DLM lock, it wll just inform the
2299 * caller to cancel the AGL process for this stripe.
2301 ldlm_lock_decref(&lockh
, mode
);
2302 LDLM_LOCK_PUT(matched
);
2304 } else if (osc_set_lock_data_with_check(matched
, einfo
)) {
2305 *flags
|= LDLM_FL_LVB_READY
;
2306 /* We already have a lock, and it's referenced. */
2307 (*upcall
)(cookie
, &lockh
, ELDLM_LOCK_MATCHED
);
2309 ldlm_lock_decref(&lockh
, mode
);
2310 LDLM_LOCK_PUT(matched
);
2313 ldlm_lock_decref(&lockh
, mode
);
2314 LDLM_LOCK_PUT(matched
);
2319 if (*flags
& LDLM_FL_TEST_LOCK
)
2322 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
),
2323 &RQF_LDLM_ENQUEUE_LVB
);
2327 rc
= ldlm_prep_enqueue_req(exp
, req
, NULL
, 0);
2329 ptlrpc_request_free(req
);
2333 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
,
2335 ptlrpc_request_set_replen(req
);
2338 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2339 *flags
&= ~LDLM_FL_BLOCK_GRANTED
;
2341 rc
= ldlm_cli_enqueue(exp
, &req
, einfo
, res_id
, policy
, flags
, lvb
,
2342 sizeof(*lvb
), LVB_T_OST
, &lockh
, async
);
2345 struct osc_enqueue_args
*aa
;
2347 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2348 aa
= ptlrpc_req_async_args(req
);
2350 aa
->oa_mode
= einfo
->ei_mode
;
2351 aa
->oa_type
= einfo
->ei_type
;
2352 lustre_handle_copy(&aa
->oa_lockh
, &lockh
);
2353 aa
->oa_upcall
= upcall
;
2354 aa
->oa_cookie
= cookie
;
2357 aa
->oa_flags
= flags
;
2360 /* AGL is essentially to enqueue an DLM lock
2361 * in advance, so we don't care about the
2362 * result of AGL enqueue.
2365 aa
->oa_flags
= NULL
;
2368 req
->rq_interpret_reply
=
2369 (ptlrpc_interpterer_t
)osc_enqueue_interpret
;
2370 if (rqset
== PTLRPCD_SET
)
2371 ptlrpcd_add_req(req
);
2373 ptlrpc_set_add_req(rqset
, req
);
2374 } else if (intent
) {
2375 ptlrpc_req_finished(req
);
2380 rc
= osc_enqueue_fini(req
, upcall
, cookie
, &lockh
, einfo
->ei_mode
,
2383 ptlrpc_req_finished(req
);
2388 int osc_match_base(struct obd_export
*exp
, struct ldlm_res_id
*res_id
,
2389 __u32 type
, ldlm_policy_data_t
*policy
, __u32 mode
,
2390 __u64
*flags
, void *data
, struct lustre_handle
*lockh
,
2393 struct obd_device
*obd
= exp
->exp_obd
;
2394 __u64 lflags
= *flags
;
2397 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH
))
2400 /* Filesystem lock extents are extended to page boundaries so that
2401 * dealing with the page cache is a little smoother
2403 policy
->l_extent
.start
-= policy
->l_extent
.start
& ~PAGE_MASK
;
2404 policy
->l_extent
.end
|= ~PAGE_MASK
;
2406 /* Next, search for already existing extent locks that will cover us */
2407 /* If we're trying to read, we also search for an existing PW lock. The
2408 * VFS and page cache already protect us locally, so lots of readers/
2409 * writers can share a single PW lock.
2414 rc
= ldlm_lock_match(obd
->obd_namespace
, lflags
,
2415 res_id
, type
, policy
, rc
, lockh
, unref
);
2418 if (!osc_set_data_with_check(lockh
, data
)) {
2419 if (!(lflags
& LDLM_FL_TEST_LOCK
))
2420 ldlm_lock_decref(lockh
, rc
);
2424 if (!(lflags
& LDLM_FL_TEST_LOCK
) && mode
!= rc
) {
2425 ldlm_lock_addref(lockh
, LCK_PR
);
2426 ldlm_lock_decref(lockh
, LCK_PW
);
2433 int osc_cancel_base(struct lustre_handle
*lockh
, __u32 mode
)
2435 if (unlikely(mode
== LCK_GROUP
))
2436 ldlm_lock_decref_and_cancel(lockh
, mode
);
2438 ldlm_lock_decref(lockh
, mode
);
2443 static int osc_statfs_interpret(const struct lu_env
*env
,
2444 struct ptlrpc_request
*req
,
2445 struct osc_async_args
*aa
, int rc
)
2447 struct obd_statfs
*msfs
;
2450 /* The request has in fact never been sent
2451 * due to issues at a higher level (LOV).
2452 * Exit immediately since the caller is
2453 * aware of the problem and takes care
2458 if ((rc
== -ENOTCONN
|| rc
== -EAGAIN
) &&
2459 (aa
->aa_oi
->oi_flags
& OBD_STATFS_NODELAY
)) {
2467 msfs
= req_capsule_server_get(&req
->rq_pill
, &RMF_OBD_STATFS
);
2473 *aa
->aa_oi
->oi_osfs
= *msfs
;
2475 rc
= aa
->aa_oi
->oi_cb_up(aa
->aa_oi
, rc
);
2479 static int osc_statfs_async(struct obd_export
*exp
,
2480 struct obd_info
*oinfo
, __u64 max_age
,
2481 struct ptlrpc_request_set
*rqset
)
2483 struct obd_device
*obd
= class_exp2obd(exp
);
2484 struct ptlrpc_request
*req
;
2485 struct osc_async_args
*aa
;
2488 /* We could possibly pass max_age in the request (as an absolute
2489 * timestamp or a "seconds.usec ago") so the target can avoid doing
2490 * extra calls into the filesystem if that isn't necessary (e.g.
2491 * during mount that would help a bit). Having relative timestamps
2492 * is not so great if request processing is slow, while absolute
2493 * timestamps are not ideal because they need time synchronization.
2495 req
= ptlrpc_request_alloc(obd
->u
.cli
.cl_import
, &RQF_OST_STATFS
);
2499 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_STATFS
);
2501 ptlrpc_request_free(req
);
2504 ptlrpc_request_set_replen(req
);
2505 req
->rq_request_portal
= OST_CREATE_PORTAL
;
2506 ptlrpc_at_set_req_timeout(req
);
2508 if (oinfo
->oi_flags
& OBD_STATFS_NODELAY
) {
2509 /* procfs requests not want stat in wait for avoid deadlock */
2510 req
->rq_no_resend
= 1;
2511 req
->rq_no_delay
= 1;
2514 req
->rq_interpret_reply
= (ptlrpc_interpterer_t
)osc_statfs_interpret
;
2515 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2516 aa
= ptlrpc_req_async_args(req
);
2519 ptlrpc_set_add_req(rqset
, req
);
2523 static int osc_statfs(const struct lu_env
*env
, struct obd_export
*exp
,
2524 struct obd_statfs
*osfs
, __u64 max_age
, __u32 flags
)
2526 struct obd_device
*obd
= class_exp2obd(exp
);
2527 struct obd_statfs
*msfs
;
2528 struct ptlrpc_request
*req
;
2529 struct obd_import
*imp
= NULL
;
2532 /* Since the request might also come from lprocfs, so we need
2533 * sync this with client_disconnect_export Bug15684
2535 down_read(&obd
->u
.cli
.cl_sem
);
2536 if (obd
->u
.cli
.cl_import
)
2537 imp
= class_import_get(obd
->u
.cli
.cl_import
);
2538 up_read(&obd
->u
.cli
.cl_sem
);
2542 /* We could possibly pass max_age in the request (as an absolute
2543 * timestamp or a "seconds.usec ago") so the target can avoid doing
2544 * extra calls into the filesystem if that isn't necessary (e.g.
2545 * during mount that would help a bit). Having relative timestamps
2546 * is not so great if request processing is slow, while absolute
2547 * timestamps are not ideal because they need time synchronization.
2549 req
= ptlrpc_request_alloc(imp
, &RQF_OST_STATFS
);
2551 class_import_put(imp
);
2556 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_STATFS
);
2558 ptlrpc_request_free(req
);
2561 ptlrpc_request_set_replen(req
);
2562 req
->rq_request_portal
= OST_CREATE_PORTAL
;
2563 ptlrpc_at_set_req_timeout(req
);
2565 if (flags
& OBD_STATFS_NODELAY
) {
2566 /* procfs requests not want stat in wait for avoid deadlock */
2567 req
->rq_no_resend
= 1;
2568 req
->rq_no_delay
= 1;
2571 rc
= ptlrpc_queue_wait(req
);
2575 msfs
= req_capsule_server_get(&req
->rq_pill
, &RMF_OBD_STATFS
);
2584 ptlrpc_req_finished(req
);
2588 /* Retrieve object striping information.
2590 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2591 * the maximum number of OST indices which will fit in the user buffer.
2592 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2594 static int osc_getstripe(struct lov_stripe_md
*lsm
,
2595 struct lov_user_md __user
*lump
)
2597 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2598 struct lov_user_md_v3 lum
, *lumk
;
2599 struct lov_user_ost_data_v1
*lmm_objects
;
2600 int rc
= 0, lum_size
;
2605 /* we only need the header part from user space to get lmm_magic and
2606 * lmm_stripe_count, (the header part is common to v1 and v3)
2608 lum_size
= sizeof(struct lov_user_md_v1
);
2609 if (copy_from_user(&lum
, lump
, lum_size
))
2612 if ((lum
.lmm_magic
!= LOV_USER_MAGIC_V1
) &&
2613 (lum
.lmm_magic
!= LOV_USER_MAGIC_V3
))
2616 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2617 LASSERT(sizeof(struct lov_user_md_v1
) == sizeof(struct lov_mds_md_v1
));
2618 LASSERT(sizeof(struct lov_user_md_v3
) == sizeof(struct lov_mds_md_v3
));
2619 LASSERT(sizeof(lum
.lmm_objects
[0]) == sizeof(lumk
->lmm_objects
[0]));
2621 /* we can use lov_mds_md_size() to compute lum_size
2622 * because lov_user_md_vX and lov_mds_md_vX have the same size
2624 if (lum
.lmm_stripe_count
> 0) {
2625 lum_size
= lov_mds_md_size(lum
.lmm_stripe_count
, lum
.lmm_magic
);
2626 lumk
= kzalloc(lum_size
, GFP_NOFS
);
2630 if (lum
.lmm_magic
== LOV_USER_MAGIC_V1
)
2632 &(((struct lov_user_md_v1
*)lumk
)->lmm_objects
[0]);
2634 lmm_objects
= &(lumk
->lmm_objects
[0]);
2635 lmm_objects
->l_ost_oi
= lsm
->lsm_oi
;
2637 lum_size
= lov_mds_md_size(0, lum
.lmm_magic
);
2641 lumk
->lmm_oi
= lsm
->lsm_oi
;
2642 lumk
->lmm_stripe_count
= 1;
2644 if (copy_to_user(lump
, lumk
, lum_size
))
2653 static int osc_iocontrol(unsigned int cmd
, struct obd_export
*exp
, int len
,
2654 void *karg
, void __user
*uarg
)
2656 struct obd_device
*obd
= exp
->exp_obd
;
2657 struct obd_ioctl_data
*data
= karg
;
2660 if (!try_module_get(THIS_MODULE
)) {
2661 CERROR("%s: cannot get module '%s'\n", obd
->obd_name
,
2662 module_name(THIS_MODULE
));
2666 case OBD_IOC_LOV_GET_CONFIG
: {
2668 struct lov_desc
*desc
;
2669 struct obd_uuid uuid
;
2673 if (obd_ioctl_getdata(&buf
, &len
, uarg
)) {
2678 data
= (struct obd_ioctl_data
*)buf
;
2680 if (sizeof(*desc
) > data
->ioc_inllen1
) {
2681 obd_ioctl_freedata(buf
, len
);
2686 if (data
->ioc_inllen2
< sizeof(uuid
)) {
2687 obd_ioctl_freedata(buf
, len
);
2692 desc
= (struct lov_desc
*)data
->ioc_inlbuf1
;
2693 desc
->ld_tgt_count
= 1;
2694 desc
->ld_active_tgt_count
= 1;
2695 desc
->ld_default_stripe_count
= 1;
2696 desc
->ld_default_stripe_size
= 0;
2697 desc
->ld_default_stripe_offset
= 0;
2698 desc
->ld_pattern
= 0;
2699 memcpy(&desc
->ld_uuid
, &obd
->obd_uuid
, sizeof(uuid
));
2701 memcpy(data
->ioc_inlbuf2
, &obd
->obd_uuid
, sizeof(uuid
));
2703 err
= copy_to_user(uarg
, buf
, len
);
2706 obd_ioctl_freedata(buf
, len
);
2709 case LL_IOC_LOV_SETSTRIPE
:
2710 err
= obd_alloc_memmd(exp
, karg
);
2714 case LL_IOC_LOV_GETSTRIPE
:
2715 err
= osc_getstripe(karg
, uarg
);
2717 case OBD_IOC_CLIENT_RECOVER
:
2718 err
= ptlrpc_recover_import(obd
->u
.cli
.cl_import
,
2719 data
->ioc_inlbuf1
, 0);
2723 case IOC_OSC_SET_ACTIVE
:
2724 err
= ptlrpc_set_import_active(obd
->u
.cli
.cl_import
,
2727 case OBD_IOC_POLL_QUOTACHECK
:
2728 err
= osc_quota_poll_check(exp
, karg
);
2730 case OBD_IOC_PING_TARGET
:
2731 err
= ptlrpc_obd_ping(obd
);
2734 CDEBUG(D_INODE
, "unrecognised ioctl %#x by %s\n",
2735 cmd
, current_comm());
2740 module_put(THIS_MODULE
);
2744 static int osc_get_info(const struct lu_env
*env
, struct obd_export
*exp
,
2745 u32 keylen
, void *key
, __u32
*vallen
, void *val
,
2746 struct lov_stripe_md
*lsm
)
2748 if (!vallen
|| !val
)
2751 if (KEY_IS(KEY_LOCK_TO_STRIPE
)) {
2752 __u32
*stripe
= val
;
2753 *vallen
= sizeof(*stripe
);
2756 } else if (KEY_IS(KEY_LAST_ID
)) {
2757 struct ptlrpc_request
*req
;
2762 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
),
2763 &RQF_OST_GET_INFO_LAST_ID
);
2767 req_capsule_set_size(&req
->rq_pill
, &RMF_SETINFO_KEY
,
2768 RCL_CLIENT
, keylen
);
2769 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_GET_INFO
);
2771 ptlrpc_request_free(req
);
2775 tmp
= req_capsule_client_get(&req
->rq_pill
, &RMF_SETINFO_KEY
);
2776 memcpy(tmp
, key
, keylen
);
2778 req
->rq_no_delay
= req
->rq_no_resend
= 1;
2779 ptlrpc_request_set_replen(req
);
2780 rc
= ptlrpc_queue_wait(req
);
2784 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_OBD_ID
);
2790 *((u64
*)val
) = *reply
;
2792 ptlrpc_req_finished(req
);
2794 } else if (KEY_IS(KEY_FIEMAP
)) {
2795 struct ll_fiemap_info_key
*fm_key
= key
;
2796 struct ldlm_res_id res_id
;
2797 ldlm_policy_data_t policy
;
2798 struct lustre_handle lockh
;
2799 enum ldlm_mode mode
= 0;
2800 struct ptlrpc_request
*req
;
2801 struct ll_user_fiemap
*reply
;
2805 if (!(fm_key
->fiemap
.fm_flags
& FIEMAP_FLAG_SYNC
))
2808 policy
.l_extent
.start
= fm_key
->fiemap
.fm_start
&
2811 if (OBD_OBJECT_EOF
- fm_key
->fiemap
.fm_length
<=
2812 fm_key
->fiemap
.fm_start
+ PAGE_SIZE
- 1)
2813 policy
.l_extent
.end
= OBD_OBJECT_EOF
;
2815 policy
.l_extent
.end
= (fm_key
->fiemap
.fm_start
+
2816 fm_key
->fiemap
.fm_length
+
2817 PAGE_SIZE
- 1) & PAGE_MASK
;
2819 ostid_build_res_name(&fm_key
->oa
.o_oi
, &res_id
);
2820 mode
= ldlm_lock_match(exp
->exp_obd
->obd_namespace
,
2821 LDLM_FL_BLOCK_GRANTED
|
2823 &res_id
, LDLM_EXTENT
, &policy
,
2824 LCK_PR
| LCK_PW
, &lockh
, 0);
2825 if (mode
) { /* lock is cached on client */
2826 if (mode
!= LCK_PR
) {
2827 ldlm_lock_addref(&lockh
, LCK_PR
);
2828 ldlm_lock_decref(&lockh
, LCK_PW
);
2830 } else { /* no cached lock, needs acquire lock on server side */
2831 fm_key
->oa
.o_valid
|= OBD_MD_FLFLAGS
;
2832 fm_key
->oa
.o_flags
|= OBD_FL_SRVLOCK
;
2836 req
= ptlrpc_request_alloc(class_exp2cliimp(exp
),
2837 &RQF_OST_GET_INFO_FIEMAP
);
2843 req_capsule_set_size(&req
->rq_pill
, &RMF_FIEMAP_KEY
,
2844 RCL_CLIENT
, keylen
);
2845 req_capsule_set_size(&req
->rq_pill
, &RMF_FIEMAP_VAL
,
2846 RCL_CLIENT
, *vallen
);
2847 req_capsule_set_size(&req
->rq_pill
, &RMF_FIEMAP_VAL
,
2848 RCL_SERVER
, *vallen
);
2850 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_GET_INFO
);
2852 ptlrpc_request_free(req
);
2856 tmp
= req_capsule_client_get(&req
->rq_pill
, &RMF_FIEMAP_KEY
);
2857 memcpy(tmp
, key
, keylen
);
2858 tmp
= req_capsule_client_get(&req
->rq_pill
, &RMF_FIEMAP_VAL
);
2859 memcpy(tmp
, val
, *vallen
);
2861 ptlrpc_request_set_replen(req
);
2862 rc
= ptlrpc_queue_wait(req
);
2866 reply
= req_capsule_server_get(&req
->rq_pill
, &RMF_FIEMAP_VAL
);
2872 memcpy(val
, reply
, *vallen
);
2874 ptlrpc_req_finished(req
);
2877 ldlm_lock_decref(&lockh
, LCK_PR
);
2884 static int osc_set_info_async(const struct lu_env
*env
, struct obd_export
*exp
,
2885 u32 keylen
, void *key
, u32 vallen
,
2886 void *val
, struct ptlrpc_request_set
*set
)
2888 struct ptlrpc_request
*req
;
2889 struct obd_device
*obd
= exp
->exp_obd
;
2890 struct obd_import
*imp
= class_exp2cliimp(exp
);
2894 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN
, 10);
2896 if (KEY_IS(KEY_CHECKSUM
)) {
2897 if (vallen
!= sizeof(int))
2899 exp
->exp_obd
->u
.cli
.cl_checksum
= (*(int *)val
) ? 1 : 0;
2903 if (KEY_IS(KEY_SPTLRPC_CONF
)) {
2904 sptlrpc_conf_client_adapt(obd
);
2908 if (KEY_IS(KEY_FLUSH_CTX
)) {
2909 sptlrpc_import_flush_my_ctx(imp
);
2913 if (KEY_IS(KEY_CACHE_SET
)) {
2914 struct client_obd
*cli
= &obd
->u
.cli
;
2916 LASSERT(!cli
->cl_cache
); /* only once */
2917 cli
->cl_cache
= val
;
2918 atomic_inc(&cli
->cl_cache
->ccc_users
);
2919 cli
->cl_lru_left
= &cli
->cl_cache
->ccc_lru_left
;
2921 /* add this osc into entity list */
2922 LASSERT(list_empty(&cli
->cl_lru_osc
));
2923 spin_lock(&cli
->cl_cache
->ccc_lru_lock
);
2924 list_add(&cli
->cl_lru_osc
, &cli
->cl_cache
->ccc_lru
);
2925 spin_unlock(&cli
->cl_cache
->ccc_lru_lock
);
2930 if (KEY_IS(KEY_CACHE_LRU_SHRINK
)) {
2931 struct client_obd
*cli
= &obd
->u
.cli
;
2932 int nr
= atomic_read(&cli
->cl_lru_in_list
) >> 1;
2933 int target
= *(int *)val
;
2935 nr
= osc_lru_shrink(env
, cli
, min(nr
, target
), true);
2940 if (!set
&& !KEY_IS(KEY_GRANT_SHRINK
))
2943 /* We pass all other commands directly to OST. Since nobody calls osc
2944 * methods directly and everybody is supposed to go through LOV, we
2945 * assume lov checked invalid values for us.
2946 * The only recognised values so far are evict_by_nid and mds_conn.
2947 * Even if something bad goes through, we'd get a -EINVAL from OST
2951 req
= ptlrpc_request_alloc(imp
, KEY_IS(KEY_GRANT_SHRINK
) ?
2952 &RQF_OST_SET_GRANT_INFO
:
2957 req_capsule_set_size(&req
->rq_pill
, &RMF_SETINFO_KEY
,
2958 RCL_CLIENT
, keylen
);
2959 if (!KEY_IS(KEY_GRANT_SHRINK
))
2960 req_capsule_set_size(&req
->rq_pill
, &RMF_SETINFO_VAL
,
2961 RCL_CLIENT
, vallen
);
2962 rc
= ptlrpc_request_pack(req
, LUSTRE_OST_VERSION
, OST_SET_INFO
);
2964 ptlrpc_request_free(req
);
2968 tmp
= req_capsule_client_get(&req
->rq_pill
, &RMF_SETINFO_KEY
);
2969 memcpy(tmp
, key
, keylen
);
2970 tmp
= req_capsule_client_get(&req
->rq_pill
, KEY_IS(KEY_GRANT_SHRINK
) ?
2973 memcpy(tmp
, val
, vallen
);
2975 if (KEY_IS(KEY_GRANT_SHRINK
)) {
2976 struct osc_brw_async_args
*aa
;
2979 CLASSERT(sizeof(*aa
) <= sizeof(req
->rq_async_args
));
2980 aa
= ptlrpc_req_async_args(req
);
2981 oa
= kmem_cache_zalloc(obdo_cachep
, GFP_NOFS
);
2983 ptlrpc_req_finished(req
);
2986 *oa
= ((struct ost_body
*)val
)->oa
;
2988 req
->rq_interpret_reply
= osc_shrink_grant_interpret
;
2991 ptlrpc_request_set_replen(req
);
2992 if (!KEY_IS(KEY_GRANT_SHRINK
)) {
2994 ptlrpc_set_add_req(set
, req
);
2995 ptlrpc_check_set(NULL
, set
);
2997 ptlrpcd_add_req(req
);
3003 static int osc_reconnect(const struct lu_env
*env
,
3004 struct obd_export
*exp
, struct obd_device
*obd
,
3005 struct obd_uuid
*cluuid
,
3006 struct obd_connect_data
*data
,
3009 struct client_obd
*cli
= &obd
->u
.cli
;
3011 if (data
&& (data
->ocd_connect_flags
& OBD_CONNECT_GRANT
)) {
3014 spin_lock(&cli
->cl_loi_list_lock
);
3015 data
->ocd_grant
= (cli
->cl_avail_grant
+ cli
->cl_dirty
) ?:
3016 2 * cli_brw_size(obd
);
3017 lost_grant
= cli
->cl_lost_grant
;
3018 cli
->cl_lost_grant
= 0;
3019 spin_unlock(&cli
->cl_loi_list_lock
);
3021 CDEBUG(D_RPCTRACE
, "ocd_connect_flags: %#llx ocd_version: %d ocd_grant: %d, lost: %ld.\n",
3022 data
->ocd_connect_flags
,
3023 data
->ocd_version
, data
->ocd_grant
, lost_grant
);
3029 static int osc_disconnect(struct obd_export
*exp
)
3031 struct obd_device
*obd
= class_exp2obd(exp
);
3034 rc
= client_disconnect_export(exp
);
3036 * Initially we put del_shrink_grant before disconnect_export, but it
3037 * causes the following problem if setup (connect) and cleanup
3038 * (disconnect) are tangled together.
3039 * connect p1 disconnect p2
3040 * ptlrpc_connect_import
3041 * ............... class_manual_cleanup
3044 * ptlrpc_connect_interrupt
3046 * add this client to shrink list
3048 * Bang! pinger trigger the shrink.
3049 * So the osc should be disconnected from the shrink list, after we
3050 * are sure the import has been destroyed. BUG18662
3052 if (!obd
->u
.cli
.cl_import
)
3053 osc_del_shrink_grant(&obd
->u
.cli
);
3057 static int osc_import_event(struct obd_device
*obd
,
3058 struct obd_import
*imp
,
3059 enum obd_import_event event
)
3061 struct client_obd
*cli
;
3064 LASSERT(imp
->imp_obd
== obd
);
3067 case IMP_EVENT_DISCON
: {
3069 spin_lock(&cli
->cl_loi_list_lock
);
3070 cli
->cl_avail_grant
= 0;
3071 cli
->cl_lost_grant
= 0;
3072 spin_unlock(&cli
->cl_loi_list_lock
);
3075 case IMP_EVENT_INACTIVE
: {
3076 rc
= obd_notify_observer(obd
, obd
, OBD_NOTIFY_INACTIVE
, NULL
);
3079 case IMP_EVENT_INVALIDATE
: {
3080 struct ldlm_namespace
*ns
= obd
->obd_namespace
;
3084 env
= cl_env_get(&refcheck
);
3088 /* all pages go to failing rpcs due to the invalid
3091 osc_io_unplug(env
, cli
, NULL
);
3093 ldlm_namespace_cleanup(ns
, LDLM_FL_LOCAL_ONLY
);
3094 cl_env_put(env
, &refcheck
);
3100 case IMP_EVENT_ACTIVE
: {
3101 rc
= obd_notify_observer(obd
, obd
, OBD_NOTIFY_ACTIVE
, NULL
);
3104 case IMP_EVENT_OCD
: {
3105 struct obd_connect_data
*ocd
= &imp
->imp_connect_data
;
3107 if (ocd
->ocd_connect_flags
& OBD_CONNECT_GRANT
)
3108 osc_init_grant(&obd
->u
.cli
, ocd
);
3111 if (ocd
->ocd_connect_flags
& OBD_CONNECT_REQPORTAL
)
3112 imp
->imp_client
->cli_request_portal
= OST_REQUEST_PORTAL
;
3114 rc
= obd_notify_observer(obd
, obd
, OBD_NOTIFY_OCD
, NULL
);
3117 case IMP_EVENT_DEACTIVATE
: {
3118 rc
= obd_notify_observer(obd
, obd
, OBD_NOTIFY_DEACTIVATE
, NULL
);
3121 case IMP_EVENT_ACTIVATE
: {
3122 rc
= obd_notify_observer(obd
, obd
, OBD_NOTIFY_ACTIVATE
, NULL
);
3126 CERROR("Unknown import event %d\n", event
);
3133 * Determine whether the lock can be canceled before replaying the lock
3134 * during recovery, see bug16774 for detailed information.
3136 * \retval zero the lock can't be canceled
3137 * \retval other ok to cancel
3139 static int osc_cancel_weight(struct ldlm_lock
*lock
)
3142 * Cancel all unused and granted extent lock.
3144 if (lock
->l_resource
->lr_type
== LDLM_EXTENT
&&
3145 lock
->l_granted_mode
== lock
->l_req_mode
&&
3146 osc_ldlm_weigh_ast(lock
) == 0)
3152 static int brw_queue_work(const struct lu_env
*env
, void *data
)
3154 struct client_obd
*cli
= data
;
3156 CDEBUG(D_CACHE
, "Run writeback work for client obd %p.\n", cli
);
3158 osc_io_unplug(env
, cli
, NULL
);
3162 int osc_setup(struct obd_device
*obd
, struct lustre_cfg
*lcfg
)
3164 struct lprocfs_static_vars lvars
= { NULL
};
3165 struct client_obd
*cli
= &obd
->u
.cli
;
3172 rc
= ptlrpcd_addref();
3176 rc
= client_obd_setup(obd
, lcfg
);
3180 handler
= ptlrpcd_alloc_work(cli
->cl_import
, brw_queue_work
, cli
);
3181 if (IS_ERR(handler
)) {
3182 rc
= PTR_ERR(handler
);
3183 goto out_client_setup
;
3185 cli
->cl_writeback_work
= handler
;
3187 handler
= ptlrpcd_alloc_work(cli
->cl_import
, lru_queue_work
, cli
);
3188 if (IS_ERR(handler
)) {
3189 rc
= PTR_ERR(handler
);
3190 goto out_ptlrpcd_work
;
3193 cli
->cl_lru_work
= handler
;
3195 rc
= osc_quota_setup(obd
);
3197 goto out_ptlrpcd_work
;
3199 cli
->cl_grant_shrink_interval
= GRANT_SHRINK_INTERVAL
;
3200 lprocfs_osc_init_vars(&lvars
);
3201 if (lprocfs_obd_setup(obd
, lvars
.obd_vars
, lvars
.sysfs_vars
) == 0) {
3202 lproc_osc_attach_seqstat(obd
);
3203 sptlrpc_lprocfs_cliobd_attach(obd
);
3204 ptlrpc_lprocfs_register_obd(obd
);
3208 * We try to control the total number of requests with a upper limit
3209 * osc_reqpool_maxreqcount. There might be some race which will cause
3210 * over-limit allocation, but it is fine.
3212 req_count
= atomic_read(&osc_pool_req_count
);
3213 if (req_count
< osc_reqpool_maxreqcount
) {
3214 adding
= cli
->cl_max_rpcs_in_flight
+ 2;
3215 if (req_count
+ adding
> osc_reqpool_maxreqcount
)
3216 adding
= osc_reqpool_maxreqcount
- req_count
;
3218 added
= ptlrpc_add_rqs_to_pool(osc_rq_pool
, adding
);
3219 atomic_add(added
, &osc_pool_req_count
);
3222 INIT_LIST_HEAD(&cli
->cl_grant_shrink_list
);
3223 ns_register_cancel(obd
->obd_namespace
, osc_cancel_weight
);
3227 if (cli
->cl_writeback_work
) {
3228 ptlrpcd_destroy_work(cli
->cl_writeback_work
);
3229 cli
->cl_writeback_work
= NULL
;
3231 if (cli
->cl_lru_work
) {
3232 ptlrpcd_destroy_work(cli
->cl_lru_work
);
3233 cli
->cl_lru_work
= NULL
;
3236 client_obd_cleanup(obd
);
3242 static int osc_precleanup(struct obd_device
*obd
, enum obd_cleanup_stage stage
)
3245 case OBD_CLEANUP_EARLY
: {
3246 struct obd_import
*imp
;
3248 imp
= obd
->u
.cli
.cl_import
;
3249 CDEBUG(D_HA
, "Deactivating import %s\n", obd
->obd_name
);
3250 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3251 ptlrpc_deactivate_import(imp
);
3252 spin_lock(&imp
->imp_lock
);
3253 imp
->imp_pingable
= 0;
3254 spin_unlock(&imp
->imp_lock
);
3257 case OBD_CLEANUP_EXPORTS
: {
3258 struct client_obd
*cli
= &obd
->u
.cli
;
3260 * for echo client, export may be on zombie list, wait for
3261 * zombie thread to cull it, because cli.cl_import will be
3262 * cleared in client_disconnect_export():
3263 * class_export_destroy() -> obd_cleanup() ->
3264 * echo_device_free() -> echo_client_cleanup() ->
3265 * obd_disconnect() -> osc_disconnect() ->
3266 * client_disconnect_export()
3268 obd_zombie_barrier();
3269 if (cli
->cl_writeback_work
) {
3270 ptlrpcd_destroy_work(cli
->cl_writeback_work
);
3271 cli
->cl_writeback_work
= NULL
;
3273 if (cli
->cl_lru_work
) {
3274 ptlrpcd_destroy_work(cli
->cl_lru_work
);
3275 cli
->cl_lru_work
= NULL
;
3277 obd_cleanup_client_import(obd
);
3278 ptlrpc_lprocfs_unregister_obd(obd
);
3279 lprocfs_obd_cleanup(obd
);
3286 static int osc_cleanup(struct obd_device
*obd
)
3288 struct client_obd
*cli
= &obd
->u
.cli
;
3292 if (cli
->cl_cache
) {
3293 LASSERT(atomic_read(&cli
->cl_cache
->ccc_users
) > 0);
3294 spin_lock(&cli
->cl_cache
->ccc_lru_lock
);
3295 list_del_init(&cli
->cl_lru_osc
);
3296 spin_unlock(&cli
->cl_cache
->ccc_lru_lock
);
3297 cli
->cl_lru_left
= NULL
;
3298 atomic_dec(&cli
->cl_cache
->ccc_users
);
3299 cli
->cl_cache
= NULL
;
3302 /* free memory of osc quota cache */
3303 osc_quota_cleanup(obd
);
3305 rc
= client_obd_cleanup(obd
);
3311 int osc_process_config_base(struct obd_device
*obd
, struct lustre_cfg
*lcfg
)
3313 struct lprocfs_static_vars lvars
= { NULL
};
3316 lprocfs_osc_init_vars(&lvars
);
3318 switch (lcfg
->lcfg_command
) {
3320 rc
= class_process_proc_param(PARAM_OSC
, lvars
.obd_vars
,
3330 static int osc_process_config(struct obd_device
*obd
, u32 len
, void *buf
)
3332 return osc_process_config_base(obd
, buf
);
3335 static struct obd_ops osc_obd_ops
= {
3336 .owner
= THIS_MODULE
,
3338 .precleanup
= osc_precleanup
,
3339 .cleanup
= osc_cleanup
,
3340 .add_conn
= client_import_add_conn
,
3341 .del_conn
= client_import_del_conn
,
3342 .connect
= client_connect_import
,
3343 .reconnect
= osc_reconnect
,
3344 .disconnect
= osc_disconnect
,
3345 .statfs
= osc_statfs
,
3346 .statfs_async
= osc_statfs_async
,
3347 .packmd
= osc_packmd
,
3348 .unpackmd
= osc_unpackmd
,
3349 .create
= osc_create
,
3350 .destroy
= osc_destroy
,
3351 .getattr
= osc_getattr
,
3352 .getattr_async
= osc_getattr_async
,
3353 .setattr
= osc_setattr
,
3354 .setattr_async
= osc_setattr_async
,
3355 .find_cbdata
= osc_find_cbdata
,
3356 .iocontrol
= osc_iocontrol
,
3357 .get_info
= osc_get_info
,
3358 .set_info_async
= osc_set_info_async
,
3359 .import_event
= osc_import_event
,
3360 .process_config
= osc_process_config
,
3361 .quotactl
= osc_quotactl
,
3362 .quotacheck
= osc_quotacheck
,
3365 extern struct lu_kmem_descr osc_caches
[];
3366 extern struct lock_class_key osc_ast_guard_class
;
3368 static int __init
osc_init(void)
3370 struct lprocfs_static_vars lvars
= { NULL
};
3371 unsigned int reqpool_size
;
3372 unsigned int reqsize
;
3375 /* print an address of _any_ initialized kernel symbol from this
3376 * module, to allow debugging with gdb that doesn't support data
3377 * symbols from modules.
3379 CDEBUG(D_INFO
, "Lustre OSC module (%p).\n", &osc_caches
);
3381 rc
= lu_kmem_init(osc_caches
);
3385 lprocfs_osc_init_vars(&lvars
);
3387 rc
= class_register_type(&osc_obd_ops
, NULL
,
3388 LUSTRE_OSC_NAME
, &osc_device_type
);
3392 /* This is obviously too much memory, only prevent overflow here */
3393 if (osc_reqpool_mem_max
>= 1 << 12 || osc_reqpool_mem_max
== 0) {
3398 reqpool_size
= osc_reqpool_mem_max
<< 20;
3401 while (reqsize
< OST_MAXREQSIZE
)
3402 reqsize
= reqsize
<< 1;
3405 * We don't enlarge the request count in OSC pool according to
3406 * cl_max_rpcs_in_flight. The allocation from the pool will only be
3407 * tried after normal allocation failed. So a small OSC pool won't
3408 * cause much performance degression in most of cases.
3410 osc_reqpool_maxreqcount
= reqpool_size
/ reqsize
;
3412 atomic_set(&osc_pool_req_count
, 0);
3413 osc_rq_pool
= ptlrpc_init_rq_pool(0, OST_MAXREQSIZE
,
3414 ptlrpc_add_rqs_to_pool
);
3422 class_unregister_type(LUSTRE_OSC_NAME
);
3424 lu_kmem_fini(osc_caches
);
3428 static void /*__exit*/ osc_exit(void)
3430 class_unregister_type(LUSTRE_OSC_NAME
);
3431 lu_kmem_fini(osc_caches
);
3432 ptlrpc_free_rq_pool(osc_rq_pool
);
3435 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3436 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3437 MODULE_LICENSE("GPL");
3438 MODULE_VERSION(LUSTRE_VERSION_STRING
);
3440 module_init(osc_init
);
3441 module_exit(osc_exit
);