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, 2013, 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 <libcfs/libcfs.h>
42 # include <liblustre.h>
45 #include <lustre_dlm.h>
46 #include <lustre_net.h>
47 #include <lustre/lustre_user.h>
48 #include <obd_cksum.h>
54 #include <lustre_ha.h>
55 #include <lprocfs_status.h>
56 #include <lustre_log.h>
57 #include <lustre_debug.h>
58 #include <lustre_param.h>
59 #include <lustre_fid.h>
60 #include "osc_internal.h"
61 #include "osc_cl_internal.h"
63 struct osc_brw_async_args {
67 obd_count aa_page_count;
69 struct brw_page **aa_ppga;
70 struct client_obd *aa_cli;
71 struct list_head aa_oaps;
72 struct list_head aa_exts;
73 struct obd_capa *aa_ocapa;
74 struct cl_req *aa_clerq;
77 #define osc_grant_args osc_brw_async_args
79 struct osc_async_args {
80 struct obd_info *aa_oi;
83 struct osc_setattr_args {
85 obd_enqueue_update_f sa_upcall;
89 struct osc_fsync_args {
90 struct obd_info *fa_oi;
91 obd_enqueue_update_f fa_upcall;
95 struct osc_enqueue_args {
96 struct obd_export *oa_exp;
98 obd_enqueue_update_f oa_upcall;
100 struct ost_lvb *oa_lvb;
101 struct lustre_handle *oa_lockh;
102 struct ldlm_enqueue_info *oa_ei;
103 unsigned int oa_agl:1;
106 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
107 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
110 /* Unpack OSC object metadata from disk storage (LE byte order). */
111 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
112 struct lov_mds_md *lmm, int lmm_bytes)
115 struct obd_import *imp = class_exp2cliimp(exp);
119 if (lmm_bytes < sizeof(*lmm)) {
120 CERROR("%s: lov_mds_md too small: %d, need %d\n",
121 exp->exp_obd->obd_name, lmm_bytes,
125 /* XXX LOV_MAGIC etc check? */
127 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
128 CERROR("%s: zero lmm_object_id: rc = %d\n",
129 exp->exp_obd->obd_name, -EINVAL);
134 lsm_size = lov_stripe_md_size(1);
138 if (*lsmp != NULL && lmm == NULL) {
139 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
140 OBD_FREE(*lsmp, lsm_size);
146 OBD_ALLOC(*lsmp, lsm_size);
147 if (unlikely(*lsmp == NULL))
149 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
150 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
151 OBD_FREE(*lsmp, lsm_size);
154 loi_init((*lsmp)->lsm_oinfo[0]);
155 } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
160 /* XXX zero *lsmp? */
161 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
164 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
165 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
167 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
172 static inline void osc_pack_capa(struct ptlrpc_request *req,
173 struct ost_body *body, void *capa)
175 struct obd_capa *oc = (struct obd_capa *)capa;
176 struct lustre_capa *c;
181 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
184 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
185 DEBUG_CAPA(D_SEC, c, "pack");
188 static inline void osc_pack_req_body(struct ptlrpc_request *req,
189 struct obd_info *oinfo)
191 struct ost_body *body;
193 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
196 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
198 osc_pack_capa(req, body, oinfo->oi_capa);
201 static inline void osc_set_capa_size(struct ptlrpc_request *req,
202 const struct req_msg_field *field,
206 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
208 /* it is already calculated as sizeof struct obd_capa */
212 static int osc_getattr_interpret(const struct lu_env *env,
213 struct ptlrpc_request *req,
214 struct osc_async_args *aa, int rc)
216 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;
249 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
253 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
254 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
256 ptlrpc_request_free(req);
260 osc_pack_req_body(req, oinfo);
262 ptlrpc_request_set_replen(req);
263 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
265 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
266 aa = ptlrpc_req_async_args(req);
269 ptlrpc_set_add_req(set, req);
273 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
274 struct obd_info *oinfo)
276 struct ptlrpc_request *req;
277 struct ost_body *body;
281 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
285 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
286 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
288 ptlrpc_request_free(req);
292 osc_pack_req_body(req, oinfo);
294 ptlrpc_request_set_replen(req);
296 rc = ptlrpc_queue_wait(req);
300 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
302 GOTO(out, rc = -EPROTO);
304 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
305 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
308 oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
309 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
313 ptlrpc_req_finished(req);
317 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
318 struct obd_info *oinfo, struct obd_trans_info *oti)
320 struct ptlrpc_request *req;
321 struct ost_body *body;
325 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
327 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
331 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
332 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
334 ptlrpc_request_free(req);
338 osc_pack_req_body(req, oinfo);
340 ptlrpc_request_set_replen(req);
342 rc = ptlrpc_queue_wait(req);
346 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
348 GOTO(out, rc = -EPROTO);
350 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
355 ptlrpc_req_finished(req);
359 static int osc_setattr_interpret(const struct lu_env *env,
360 struct ptlrpc_request *req,
361 struct osc_setattr_args *sa, int rc)
363 struct ost_body *body;
369 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
371 GOTO(out, rc = -EPROTO);
373 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
376 rc = sa->sa_upcall(sa->sa_cookie, rc);
380 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
381 struct obd_trans_info *oti,
382 obd_enqueue_update_f upcall, void *cookie,
383 struct ptlrpc_request_set *rqset)
385 struct ptlrpc_request *req;
386 struct osc_setattr_args *sa;
390 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
394 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
395 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
397 ptlrpc_request_free(req);
401 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
402 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
404 osc_pack_req_body(req, oinfo);
406 ptlrpc_request_set_replen(req);
408 /* do mds to ost setattr asynchronously */
410 /* Do not wait for response. */
411 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
413 req->rq_interpret_reply =
414 (ptlrpc_interpterer_t)osc_setattr_interpret;
416 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
417 sa = ptlrpc_req_async_args(req);
418 sa->sa_oa = oinfo->oi_oa;
419 sa->sa_upcall = upcall;
420 sa->sa_cookie = cookie;
422 if (rqset == PTLRPCD_SET)
423 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
425 ptlrpc_set_add_req(rqset, req);
431 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
432 struct obd_trans_info *oti,
433 struct ptlrpc_request_set *rqset)
435 return osc_setattr_async_base(exp, oinfo, oti,
436 oinfo->oi_cb_up, oinfo, rqset);
439 int osc_real_create(struct obd_export *exp, struct obdo *oa,
440 struct lov_stripe_md **ea, struct obd_trans_info *oti)
442 struct ptlrpc_request *req;
443 struct ost_body *body;
444 struct lov_stripe_md *lsm;
453 rc = obd_alloc_memmd(exp, &lsm);
458 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
460 GOTO(out, rc = -ENOMEM);
462 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
464 ptlrpc_request_free(req);
468 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
471 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
473 ptlrpc_request_set_replen(req);
475 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
476 oa->o_flags == OBD_FL_DELORPHAN) {
478 "delorphan from OST integration");
479 /* Don't resend the delorphan req */
480 req->rq_no_resend = req->rq_no_delay = 1;
483 rc = ptlrpc_queue_wait(req);
487 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
489 GOTO(out_req, rc = -EPROTO);
491 CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
492 lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
494 oa->o_blksize = cli_brw_size(exp->exp_obd);
495 oa->o_valid |= OBD_MD_FLBLKSZ;
497 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
498 * have valid lsm_oinfo data structs, so don't go touching that.
499 * This needs to be fixed in a big way.
501 lsm->lsm_oi = oa->o_oi;
505 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
507 if (oa->o_valid & OBD_MD_FLCOOKIE) {
508 if (!oti->oti_logcookies)
509 oti_alloc_cookies(oti, 1);
510 *oti->oti_logcookies = oa->o_lcookie;
514 CDEBUG(D_HA, "transno: "LPD64"\n",
515 lustre_msg_get_transno(req->rq_repmsg));
517 ptlrpc_req_finished(req);
520 obd_free_memmd(exp, &lsm);
524 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
525 obd_enqueue_update_f upcall, void *cookie,
526 struct ptlrpc_request_set *rqset)
528 struct ptlrpc_request *req;
529 struct osc_setattr_args *sa;
530 struct ost_body *body;
534 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
538 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_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(&req->rq_import->imp_connect_data, &body->oa,
551 osc_pack_capa(req, body, oinfo->oi_capa);
553 ptlrpc_request_set_replen(req);
555 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
556 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
557 sa = ptlrpc_req_async_args(req);
558 sa->sa_oa = oinfo->oi_oa;
559 sa->sa_upcall = upcall;
560 sa->sa_cookie = cookie;
561 if (rqset == PTLRPCD_SET)
562 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
564 ptlrpc_set_add_req(rqset, req);
569 static int osc_sync_interpret(const struct lu_env *env,
570 struct ptlrpc_request *req,
573 struct osc_fsync_args *fa = arg;
574 struct ost_body *body;
580 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
582 CERROR ("can't unpack ost_body\n");
583 GOTO(out, rc = -EPROTO);
586 *fa->fa_oi->oi_oa = body->oa;
588 rc = fa->fa_upcall(fa->fa_cookie, rc);
592 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
593 obd_enqueue_update_f upcall, void *cookie,
594 struct ptlrpc_request_set *rqset)
596 struct ptlrpc_request *req;
597 struct ost_body *body;
598 struct osc_fsync_args *fa;
602 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
606 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
607 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
609 ptlrpc_request_free(req);
613 /* overload the size and blocks fields in the oa with start/end */
614 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
616 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
618 osc_pack_capa(req, body, oinfo->oi_capa);
620 ptlrpc_request_set_replen(req);
621 req->rq_interpret_reply = osc_sync_interpret;
623 CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
624 fa = ptlrpc_req_async_args(req);
626 fa->fa_upcall = upcall;
627 fa->fa_cookie = cookie;
629 if (rqset == PTLRPCD_SET)
630 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
632 ptlrpc_set_add_req(rqset, req);
637 /* Find and cancel locally locks matched by @mode in the resource found by
638 * @objid. Found locks are added into @cancel list. Returns the amount of
639 * locks added to @cancels list. */
640 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
642 ldlm_mode_t mode, __u64 lock_flags)
644 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
645 struct ldlm_res_id res_id;
646 struct ldlm_resource *res;
650 /* Return, i.e. cancel nothing, only if ELC is supported (flag in
651 * export) but disabled through procfs (flag in NS).
653 * This distinguishes from a case when ELC is not supported originally,
654 * when we still want to cancel locks in advance and just cancel them
655 * locally, without sending any RPC. */
656 if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
659 ostid_build_res_name(&oa->o_oi, &res_id);
660 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
664 LDLM_RESOURCE_ADDREF(res);
665 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
666 lock_flags, 0, NULL);
667 LDLM_RESOURCE_DELREF(res);
668 ldlm_resource_putref(res);
672 static int osc_destroy_interpret(const struct lu_env *env,
673 struct ptlrpc_request *req, void *data,
676 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
678 atomic_dec(&cli->cl_destroy_in_flight);
679 wake_up(&cli->cl_destroy_waitq);
683 static int osc_can_send_destroy(struct client_obd *cli)
685 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
686 cli->cl_max_rpcs_in_flight) {
687 /* The destroy request can be sent */
690 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
691 cli->cl_max_rpcs_in_flight) {
693 * The counter has been modified between the two atomic
696 wake_up(&cli->cl_destroy_waitq);
701 int osc_create(const struct lu_env *env, struct obd_export *exp,
702 struct obdo *oa, struct lov_stripe_md **ea,
703 struct obd_trans_info *oti)
710 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
712 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
713 oa->o_flags == OBD_FL_RECREATE_OBJS) {
714 RETURN(osc_real_create(exp, oa, ea, oti));
717 if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
718 RETURN(osc_real_create(exp, oa, ea, oti));
720 /* we should not get here anymore */
726 /* Destroy requests can be async always on the client, and we don't even really
727 * care about the return code since the client cannot do anything at all about
729 * When the MDS is unlinking a filename, it saves the file objects into a
730 * recovery llog, and these object records are cancelled when the OST reports
731 * they were destroyed and sync'd to disk (i.e. transaction committed).
732 * If the client dies, or the OST is down when the object should be destroyed,
733 * the records are not cancelled, and when the OST reconnects to the MDS next,
734 * it will retrieve the llog unlink logs and then sends the log cancellation
735 * cookies to the MDS after committing destroy transactions. */
736 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
737 struct obdo *oa, struct lov_stripe_md *ea,
738 struct obd_trans_info *oti, struct obd_export *md_export,
741 struct client_obd *cli = &exp->exp_obd->u.cli;
742 struct ptlrpc_request *req;
743 struct ost_body *body;
744 CFS_LIST_HEAD(cancels);
749 CDEBUG(D_INFO, "oa NULL\n");
753 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
754 LDLM_FL_DISCARD_DATA);
756 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
758 ldlm_lock_list_put(&cancels, l_bl_ast, count);
762 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
763 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
766 ptlrpc_request_free(req);
770 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
771 ptlrpc_at_set_req_timeout(req);
773 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
774 oa->o_lcookie = *oti->oti_logcookies;
775 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
777 lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
779 osc_pack_capa(req, body, (struct obd_capa *)capa);
780 ptlrpc_request_set_replen(req);
782 /* If osc_destory is for destroying the unlink orphan,
783 * sent from MDT to OST, which should not be blocked here,
784 * because the process might be triggered by ptlrpcd, and
785 * it is not good to block ptlrpcd thread (b=16006)*/
786 if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
787 req->rq_interpret_reply = osc_destroy_interpret;
788 if (!osc_can_send_destroy(cli)) {
789 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
793 * Wait until the number of on-going destroy RPCs drops
794 * under max_rpc_in_flight
796 l_wait_event_exclusive(cli->cl_destroy_waitq,
797 osc_can_send_destroy(cli), &lwi);
801 /* Do not wait for response */
802 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
806 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
809 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
811 LASSERT(!(oa->o_valid & bits));
814 client_obd_list_lock(&cli->cl_loi_list_lock);
815 oa->o_dirty = cli->cl_dirty;
816 if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
817 cli->cl_dirty_max)) {
818 CERROR("dirty %lu - %lu > dirty_max %lu\n",
819 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
821 } else if (unlikely(atomic_read(&obd_unstable_pages) +
822 atomic_read(&obd_dirty_pages) -
823 atomic_read(&obd_dirty_transit_pages) >
824 (long)(obd_max_dirty_pages + 1))) {
825 /* The atomic_read() allowing the atomic_inc() are
826 * not covered by a lock thus they may safely race and trip
827 * this CERROR() unless we add in a small fudge factor (+1). */
828 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
829 cli->cl_import->imp_obd->obd_name,
830 atomic_read(&obd_unstable_pages),
831 atomic_read(&obd_dirty_pages),
832 atomic_read(&obd_dirty_transit_pages),
833 obd_max_dirty_pages);
835 } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
836 CERROR("dirty %lu - dirty_max %lu too big???\n",
837 cli->cl_dirty, cli->cl_dirty_max);
840 long max_in_flight = (cli->cl_max_pages_per_rpc <<
842 (cli->cl_max_rpcs_in_flight + 1);
843 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
845 oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
846 oa->o_dropped = cli->cl_lost_grant;
847 cli->cl_lost_grant = 0;
848 client_obd_list_unlock(&cli->cl_loi_list_lock);
849 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
850 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
854 void osc_update_next_shrink(struct client_obd *cli)
856 cli->cl_next_shrink_grant =
857 cfs_time_shift(cli->cl_grant_shrink_interval);
858 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
859 cli->cl_next_shrink_grant);
862 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
864 client_obd_list_lock(&cli->cl_loi_list_lock);
865 cli->cl_avail_grant += grant;
866 client_obd_list_unlock(&cli->cl_loi_list_lock);
869 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
871 if (body->oa.o_valid & OBD_MD_FLGRANT) {
872 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
873 __osc_update_grant(cli, body->oa.o_grant);
877 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
878 obd_count keylen, void *key, obd_count vallen,
879 void *val, struct ptlrpc_request_set *set);
881 static int osc_shrink_grant_interpret(const struct lu_env *env,
882 struct ptlrpc_request *req,
885 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
886 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
887 struct ost_body *body;
890 __osc_update_grant(cli, oa->o_grant);
894 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
896 osc_update_grant(cli, body);
902 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
904 client_obd_list_lock(&cli->cl_loi_list_lock);
905 oa->o_grant = cli->cl_avail_grant / 4;
906 cli->cl_avail_grant -= oa->o_grant;
907 client_obd_list_unlock(&cli->cl_loi_list_lock);
908 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
909 oa->o_valid |= OBD_MD_FLFLAGS;
912 oa->o_flags |= OBD_FL_SHRINK_GRANT;
913 osc_update_next_shrink(cli);
916 /* Shrink the current grant, either from some large amount to enough for a
917 * full set of in-flight RPCs, or if we have already shrunk to that limit
918 * then to enough for a single RPC. This avoids keeping more grant than
919 * needed, and avoids shrinking the grant piecemeal. */
920 static int osc_shrink_grant(struct client_obd *cli)
922 __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
923 (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
925 client_obd_list_lock(&cli->cl_loi_list_lock);
926 if (cli->cl_avail_grant <= target_bytes)
927 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
928 client_obd_list_unlock(&cli->cl_loi_list_lock);
930 return osc_shrink_grant_to_target(cli, target_bytes);
933 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
936 struct ost_body *body;
939 client_obd_list_lock(&cli->cl_loi_list_lock);
940 /* Don't shrink if we are already above or below the desired limit
941 * We don't want to shrink below a single RPC, as that will negatively
942 * impact block allocation and long-term performance. */
943 if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
944 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
946 if (target_bytes >= cli->cl_avail_grant) {
947 client_obd_list_unlock(&cli->cl_loi_list_lock);
950 client_obd_list_unlock(&cli->cl_loi_list_lock);
956 osc_announce_cached(cli, &body->oa, 0);
958 client_obd_list_lock(&cli->cl_loi_list_lock);
959 body->oa.o_grant = cli->cl_avail_grant - target_bytes;
960 cli->cl_avail_grant = target_bytes;
961 client_obd_list_unlock(&cli->cl_loi_list_lock);
962 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
963 body->oa.o_valid |= OBD_MD_FLFLAGS;
964 body->oa.o_flags = 0;
966 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
967 osc_update_next_shrink(cli);
969 rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
970 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
971 sizeof(*body), body, NULL);
973 __osc_update_grant(cli, body->oa.o_grant);
978 static int osc_should_shrink_grant(struct client_obd *client)
980 cfs_time_t time = cfs_time_current();
981 cfs_time_t next_shrink = client->cl_next_shrink_grant;
983 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
984 OBD_CONNECT_GRANT_SHRINK) == 0)
987 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
988 /* Get the current RPC size directly, instead of going via:
989 * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
990 * Keep comment here so that it can be found by searching. */
991 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
993 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
994 client->cl_avail_grant > brw_size)
997 osc_update_next_shrink(client);
1002 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1004 struct client_obd *client;
1006 cfs_list_for_each_entry(client, &item->ti_obd_list,
1007 cl_grant_shrink_list) {
1008 if (osc_should_shrink_grant(client))
1009 osc_shrink_grant(client);
1014 static int osc_add_shrink_grant(struct client_obd *client)
1018 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1020 osc_grant_shrink_grant_cb, NULL,
1021 &client->cl_grant_shrink_list);
1023 CERROR("add grant client %s error %d\n",
1024 client->cl_import->imp_obd->obd_name, rc);
1027 CDEBUG(D_CACHE, "add grant client %s \n",
1028 client->cl_import->imp_obd->obd_name);
1029 osc_update_next_shrink(client);
1033 static int osc_del_shrink_grant(struct client_obd *client)
1035 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1039 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1042 * ocd_grant is the total grant amount we're expect to hold: if we've
1043 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1044 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1046 * race is tolerable here: if we're evicted, but imp_state already
1047 * left EVICTED state, then cl_dirty must be 0 already.
1049 client_obd_list_lock(&cli->cl_loi_list_lock);
1050 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1051 cli->cl_avail_grant = ocd->ocd_grant;
1053 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1055 if (cli->cl_avail_grant < 0) {
1056 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1057 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1058 ocd->ocd_grant, cli->cl_dirty);
1059 /* workaround for servers which do not have the patch from
1061 cli->cl_avail_grant = ocd->ocd_grant;
1064 /* determine the appropriate chunk size used by osc_extent. */
1065 cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1066 client_obd_list_unlock(&cli->cl_loi_list_lock);
1068 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1069 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1070 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1072 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1073 cfs_list_empty(&cli->cl_grant_shrink_list))
1074 osc_add_shrink_grant(cli);
1077 /* We assume that the reason this OSC got a short read is because it read
1078 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1079 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1080 * this stripe never got written at or beyond this stripe offset yet. */
1081 static void handle_short_read(int nob_read, obd_count page_count,
1082 struct brw_page **pga)
1087 /* skip bytes read OK */
1088 while (nob_read > 0) {
1089 LASSERT (page_count > 0);
1091 if (pga[i]->count > nob_read) {
1092 /* EOF inside this page */
1093 ptr = kmap(pga[i]->pg) +
1094 (pga[i]->off & ~CFS_PAGE_MASK);
1095 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1102 nob_read -= pga[i]->count;
1107 /* zero remaining pages */
1108 while (page_count-- > 0) {
1109 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1110 memset(ptr, 0, pga[i]->count);
1116 static int check_write_rcs(struct ptlrpc_request *req,
1117 int requested_nob, int niocount,
1118 obd_count page_count, struct brw_page **pga)
1123 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1124 sizeof(*remote_rcs) *
1126 if (remote_rcs == NULL) {
1127 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1131 /* return error if any niobuf was in error */
1132 for (i = 0; i < niocount; i++) {
1133 if ((int)remote_rcs[i] < 0)
1134 return(remote_rcs[i]);
1136 if (remote_rcs[i] != 0) {
1137 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1138 i, remote_rcs[i], req);
1143 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1144 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1145 req->rq_bulk->bd_nob_transferred, requested_nob);
1152 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1154 if (p1->flag != p2->flag) {
1155 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1156 OBD_BRW_SYNC | OBD_BRW_ASYNC |
1157 OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
1159 /* warn if we try to combine flags that we don't know to be
1160 * safe to combine */
1161 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1162 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1163 "report this at http://bugs.whamcloud.com/\n",
1164 p1->flag, p2->flag);
1169 return (p1->off + p1->count == p2->off);
1172 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1173 struct brw_page **pga, int opc,
1174 cksum_type_t cksum_type)
1178 struct cfs_crypto_hash_desc *hdesc;
1179 unsigned int bufsize;
1181 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1183 LASSERT(pg_count > 0);
1185 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1186 if (IS_ERR(hdesc)) {
1187 CERROR("Unable to initialize checksum hash %s\n",
1188 cfs_crypto_hash_name(cfs_alg));
1189 return PTR_ERR(hdesc);
1192 while (nob > 0 && pg_count > 0) {
1193 int count = pga[i]->count > nob ? nob : pga[i]->count;
1195 /* corrupt the data before we compute the checksum, to
1196 * simulate an OST->client data error */
1197 if (i == 0 && opc == OST_READ &&
1198 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1199 unsigned char *ptr = kmap(pga[i]->pg);
1200 int off = pga[i]->off & ~CFS_PAGE_MASK;
1201 memcpy(ptr + off, "bad1", min(4, nob));
1204 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1205 pga[i]->off & ~CFS_PAGE_MASK,
1207 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1208 (int)(pga[i]->off & ~CFS_PAGE_MASK));
1210 nob -= pga[i]->count;
1216 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1219 cfs_crypto_hash_final(hdesc, NULL, NULL);
1221 /* For sending we only compute the wrong checksum instead
1222 * of corrupting the data so it is still correct on a redo */
1223 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1229 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1230 struct lov_stripe_md *lsm, obd_count page_count,
1231 struct brw_page **pga,
1232 struct ptlrpc_request **reqp,
1233 struct obd_capa *ocapa, int reserve,
1236 struct ptlrpc_request *req;
1237 struct ptlrpc_bulk_desc *desc;
1238 struct ost_body *body;
1239 struct obd_ioobj *ioobj;
1240 struct niobuf_remote *niobuf;
1241 int niocount, i, requested_nob, opc, rc;
1242 struct osc_brw_async_args *aa;
1243 struct req_capsule *pill;
1244 struct brw_page *pg_prev;
1247 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1248 RETURN(-ENOMEM); /* Recoverable */
1249 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1250 RETURN(-EINVAL); /* Fatal */
1252 if ((cmd & OBD_BRW_WRITE) != 0) {
1254 req = ptlrpc_request_alloc_pool(cli->cl_import,
1255 cli->cl_import->imp_rq_pool,
1256 &RQF_OST_BRW_WRITE);
1259 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1264 for (niocount = i = 1; i < page_count; i++) {
1265 if (!can_merge_pages(pga[i - 1], pga[i]))
1269 pill = &req->rq_pill;
1270 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1272 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1273 niocount * sizeof(*niobuf));
1274 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1276 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1278 ptlrpc_request_free(req);
1281 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1282 ptlrpc_at_set_req_timeout(req);
1283 /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1285 req->rq_no_retry_einprogress = 1;
1287 desc = ptlrpc_prep_bulk_imp(req, page_count,
1288 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1289 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1293 GOTO(out, rc = -ENOMEM);
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 != NULL && ioobj != NULL && niobuf != NULL);
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 */
1310 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1311 osc_pack_capa(req, body, ocapa);
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 & ~CFS_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_CACHE_SIZE) &&
1322 ergo(i > 0 && i < page_count - 1,
1323 poff == 0 && pg->count == PAGE_CACHE_SIZE) &&
1324 ergo(i == page_count - 1, poff == 0)),
1325 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1326 i, page_count, pg, pg->off, pg->count);
1328 LASSERTF(i == 0 || pg->off > pg_prev->off,
1329 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1330 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1332 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1333 pg_prev->pg, page_private(pg_prev->pg),
1334 pg_prev->pg->index, pg_prev->off);
1336 LASSERTF(i == 0 || pg->off > pg_prev->off,
1337 "i %d p_c %u\n", i, page_count);
1339 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1340 (pg->flag & OBD_BRW_SRVLOCK));
1342 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1343 requested_nob += pg->count;
1345 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1347 niobuf->len += pg->count;
1349 niobuf->offset = pg->off;
1350 niobuf->len = pg->count;
1351 niobuf->flags = pg->flag;
1356 LASSERTF((void *)(niobuf - niocount) ==
1357 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1358 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1359 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1361 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1363 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1364 body->oa.o_valid |= OBD_MD_FLFLAGS;
1365 body->oa.o_flags = 0;
1367 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1370 if (osc_should_shrink_grant(cli))
1371 osc_shrink_grant_local(cli, &body->oa);
1373 /* size[REQ_REC_OFF] still sizeof (*body) */
1374 if (opc == OST_WRITE) {
1375 if (cli->cl_checksum &&
1376 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1377 /* store cl_cksum_type in a local variable since
1378 * it can be changed via lprocfs */
1379 cksum_type_t cksum_type = cli->cl_cksum_type;
1381 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1382 oa->o_flags &= OBD_FL_LOCAL_MASK;
1383 body->oa.o_flags = 0;
1385 body->oa.o_flags |= cksum_type_pack(cksum_type);
1386 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1387 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1391 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1393 /* save this in 'oa', too, for later checking */
1394 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1395 oa->o_flags |= cksum_type_pack(cksum_type);
1397 /* clear out the checksum flag, in case this is a
1398 * resend but cl_checksum is no longer set. b=11238 */
1399 oa->o_valid &= ~OBD_MD_FLCKSUM;
1401 oa->o_cksum = body->oa.o_cksum;
1402 /* 1 RC per niobuf */
1403 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1404 sizeof(__u32) * niocount);
1406 if (cli->cl_checksum &&
1407 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1408 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1409 body->oa.o_flags = 0;
1410 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1411 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1414 ptlrpc_request_set_replen(req);
1416 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1417 aa = ptlrpc_req_async_args(req);
1419 aa->aa_requested_nob = requested_nob;
1420 aa->aa_nio_count = niocount;
1421 aa->aa_page_count = page_count;
1425 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1426 if (ocapa && reserve)
1427 aa->aa_ocapa = capa_get(ocapa);
1433 ptlrpc_req_finished(req);
1437 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1438 __u32 client_cksum, __u32 server_cksum, int nob,
1439 obd_count page_count, struct brw_page **pga,
1440 cksum_type_t client_cksum_type)
1444 cksum_type_t cksum_type;
1446 if (server_cksum == client_cksum) {
1447 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1451 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1453 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1456 if (cksum_type != client_cksum_type)
1457 msg = "the server did not use the checksum type specified in "
1458 "the original request - likely a protocol problem";
1459 else if (new_cksum == server_cksum)
1460 msg = "changed on the client after we checksummed it - "
1461 "likely false positive due to mmap IO (bug 11742)";
1462 else if (new_cksum == client_cksum)
1463 msg = "changed in transit before arrival at OST";
1465 msg = "changed in transit AND doesn't match the original - "
1466 "likely false positive due to mmap IO (bug 11742)";
1468 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1469 " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1470 msg, libcfs_nid2str(peer->nid),
1471 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1472 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1473 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1474 POSTID(&oa->o_oi), pga[0]->off,
1475 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1476 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1477 "client csum now %x\n", client_cksum, client_cksum_type,
1478 server_cksum, cksum_type, new_cksum);
1482 /* Note rc enters this function as number of bytes transferred */
1483 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1485 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1486 const lnet_process_id_t *peer =
1487 &req->rq_import->imp_connection->c_peer;
1488 struct client_obd *cli = aa->aa_cli;
1489 struct ost_body *body;
1490 __u32 client_cksum = 0;
1493 if (rc < 0 && rc != -EDQUOT) {
1494 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1498 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1499 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1501 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1505 /* set/clear over quota flag for a uid/gid */
1506 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1507 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1508 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1510 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1511 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1513 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1516 osc_update_grant(cli, body);
1521 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1522 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1524 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1526 CERROR("Unexpected +ve rc %d\n", rc);
1529 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1531 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1534 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1535 check_write_checksum(&body->oa, peer, client_cksum,
1536 body->oa.o_cksum, aa->aa_requested_nob,
1537 aa->aa_page_count, aa->aa_ppga,
1538 cksum_type_unpack(aa->aa_oa->o_flags)))
1541 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1542 aa->aa_page_count, aa->aa_ppga);
1546 /* The rest of this function executes only for OST_READs */
1548 /* if unwrap_bulk failed, return -EAGAIN to retry */
1549 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1551 GOTO(out, rc = -EAGAIN);
1553 if (rc > aa->aa_requested_nob) {
1554 CERROR("Unexpected rc %d (%d requested)\n", rc,
1555 aa->aa_requested_nob);
1559 if (rc != req->rq_bulk->bd_nob_transferred) {
1560 CERROR ("Unexpected rc %d (%d transferred)\n",
1561 rc, req->rq_bulk->bd_nob_transferred);
1565 if (rc < aa->aa_requested_nob)
1566 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1568 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1569 static int cksum_counter;
1570 __u32 server_cksum = body->oa.o_cksum;
1573 cksum_type_t cksum_type;
1575 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1576 body->oa.o_flags : 0);
1577 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1578 aa->aa_ppga, OST_READ,
1581 if (peer->nid == req->rq_bulk->bd_sender) {
1585 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1588 if (server_cksum == ~0 && rc > 0) {
1589 CERROR("Protocol error: server %s set the 'checksum' "
1590 "bit, but didn't send a checksum. Not fatal, "
1591 "but please notify on http://bugs.whamcloud.com/\n",
1592 libcfs_nid2str(peer->nid));
1593 } else if (server_cksum != client_cksum) {
1594 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1595 "%s%s%s inode "DFID" object "DOSTID
1596 " extent ["LPU64"-"LPU64"]\n",
1597 req->rq_import->imp_obd->obd_name,
1598 libcfs_nid2str(peer->nid),
1600 body->oa.o_valid & OBD_MD_FLFID ?
1601 body->oa.o_parent_seq : (__u64)0,
1602 body->oa.o_valid & OBD_MD_FLFID ?
1603 body->oa.o_parent_oid : 0,
1604 body->oa.o_valid & OBD_MD_FLFID ?
1605 body->oa.o_parent_ver : 0,
1606 POSTID(&body->oa.o_oi),
1607 aa->aa_ppga[0]->off,
1608 aa->aa_ppga[aa->aa_page_count-1]->off +
1609 aa->aa_ppga[aa->aa_page_count-1]->count -
1611 CERROR("client %x, server %x, cksum_type %x\n",
1612 client_cksum, server_cksum, cksum_type);
1614 aa->aa_oa->o_cksum = client_cksum;
1618 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1621 } else if (unlikely(client_cksum)) {
1622 static int cksum_missed;
1625 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1626 CERROR("Checksum %u requested from %s but not sent\n",
1627 cksum_missed, libcfs_nid2str(peer->nid));
1633 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1634 aa->aa_oa, &body->oa);
1639 static int osc_brw_redo_request(struct ptlrpc_request *request,
1640 struct osc_brw_async_args *aa, int rc)
1642 struct ptlrpc_request *new_req;
1643 struct osc_brw_async_args *new_aa;
1644 struct osc_async_page *oap;
1647 DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1648 "redo for recoverable error %d", rc);
1650 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1651 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1652 aa->aa_cli, aa->aa_oa,
1653 NULL /* lsm unused by osc currently */,
1654 aa->aa_page_count, aa->aa_ppga,
1655 &new_req, aa->aa_ocapa, 0, 1);
1659 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1660 if (oap->oap_request != NULL) {
1661 LASSERTF(request == oap->oap_request,
1662 "request %p != oap_request %p\n",
1663 request, oap->oap_request);
1664 if (oap->oap_interrupted) {
1665 ptlrpc_req_finished(new_req);
1670 /* New request takes over pga and oaps from old request.
1671 * Note that copying a list_head doesn't work, need to move it... */
1673 new_req->rq_interpret_reply = request->rq_interpret_reply;
1674 new_req->rq_async_args = request->rq_async_args;
1675 new_req->rq_commit_cb = request->rq_commit_cb;
1676 /* cap resend delay to the current request timeout, this is similar to
1677 * what ptlrpc does (see after_reply()) */
1678 if (aa->aa_resends > new_req->rq_timeout)
1679 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1681 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1682 new_req->rq_generation_set = 1;
1683 new_req->rq_import_generation = request->rq_import_generation;
1685 new_aa = ptlrpc_req_async_args(new_req);
1687 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1688 cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1689 CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1690 cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1691 new_aa->aa_resends = aa->aa_resends;
1693 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1694 if (oap->oap_request) {
1695 ptlrpc_req_finished(oap->oap_request);
1696 oap->oap_request = ptlrpc_request_addref(new_req);
1700 new_aa->aa_ocapa = aa->aa_ocapa;
1701 aa->aa_ocapa = NULL;
1703 /* XXX: This code will run into problem if we're going to support
1704 * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1705 * and wait for all of them to be finished. We should inherit request
1706 * set from old request. */
1707 ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1709 DEBUG_REQ(D_INFO, new_req, "new request");
1714 * ugh, we want disk allocation on the target to happen in offset order. we'll
1715 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1716 * fine for our small page arrays and doesn't require allocation. its an
1717 * insertion sort that swaps elements that are strides apart, shrinking the
1718 * stride down until its '1' and the array is sorted.
1720 static void sort_brw_pages(struct brw_page **array, int num)
1723 struct brw_page *tmp;
1727 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1732 for (i = stride ; i < num ; i++) {
1735 while (j >= stride && array[j - stride]->off > tmp->off) {
1736 array[j] = array[j - stride];
1741 } while (stride > 1);
1744 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1746 LASSERT(ppga != NULL);
1747 OBD_FREE(ppga, sizeof(*ppga) * count);
1750 static int brw_interpret(const struct lu_env *env,
1751 struct ptlrpc_request *req, void *data, int rc)
1753 struct osc_brw_async_args *aa = data;
1754 struct osc_extent *ext;
1755 struct osc_extent *tmp;
1756 struct client_obd *cli = aa->aa_cli;
1759 rc = osc_brw_fini_request(req, rc);
1760 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1761 /* When server return -EINPROGRESS, client should always retry
1762 * regardless of the number of times the bulk was resent already. */
1763 if (osc_recoverable_error(rc)) {
1764 if (req->rq_import_generation !=
1765 req->rq_import->imp_generation) {
1766 CDEBUG(D_HA, "%s: resend cross eviction for object: "
1767 ""DOSTID", rc = %d.\n",
1768 req->rq_import->imp_obd->obd_name,
1769 POSTID(&aa->aa_oa->o_oi), rc);
1770 } else if (rc == -EINPROGRESS ||
1771 client_should_resend(aa->aa_resends, aa->aa_cli)) {
1772 rc = osc_brw_redo_request(req, aa, rc);
1774 CERROR("%s: too many resent retries for object: "
1775 ""LPU64":"LPU64", rc = %d.\n",
1776 req->rq_import->imp_obd->obd_name,
1777 POSTID(&aa->aa_oa->o_oi), rc);
1782 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1787 capa_put(aa->aa_ocapa);
1788 aa->aa_ocapa = NULL;
1792 struct obdo *oa = aa->aa_oa;
1793 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1794 unsigned long valid = 0;
1795 struct cl_object *obj;
1796 struct osc_async_page *last;
1798 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1799 obj = osc2cl(last->oap_obj);
1801 cl_object_attr_lock(obj);
1802 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1803 attr->cat_blocks = oa->o_blocks;
1804 valid |= CAT_BLOCKS;
1806 if (oa->o_valid & OBD_MD_FLMTIME) {
1807 attr->cat_mtime = oa->o_mtime;
1810 if (oa->o_valid & OBD_MD_FLATIME) {
1811 attr->cat_atime = oa->o_atime;
1814 if (oa->o_valid & OBD_MD_FLCTIME) {
1815 attr->cat_ctime = oa->o_ctime;
1819 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1820 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1821 loff_t last_off = last->oap_count + last->oap_obj_off;
1823 /* Change file size if this is an out of quota or
1824 * direct IO write and it extends the file size */
1825 if (loi->loi_lvb.lvb_size < last_off) {
1826 attr->cat_size = last_off;
1829 /* Extend KMS if it's not a lockless write */
1830 if (loi->loi_kms < last_off &&
1831 oap2osc_page(last)->ops_srvlock == 0) {
1832 attr->cat_kms = last_off;
1838 cl_object_attr_set(env, obj, attr, valid);
1839 cl_object_attr_unlock(obj);
1841 OBDO_FREE(aa->aa_oa);
1843 cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1844 cfs_list_del_init(&ext->oe_link);
1845 osc_extent_finish(env, ext, 1, rc);
1847 LASSERT(cfs_list_empty(&aa->aa_exts));
1848 LASSERT(cfs_list_empty(&aa->aa_oaps));
1850 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1851 req->rq_bulk->bd_nob_transferred);
1852 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1853 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1855 client_obd_list_lock(&cli->cl_loi_list_lock);
1856 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1857 * is called so we know whether to go to sync BRWs or wait for more
1858 * RPCs to complete */
1859 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1860 cli->cl_w_in_flight--;
1862 cli->cl_r_in_flight--;
1863 osc_wake_cache_waiters(cli);
1864 client_obd_list_unlock(&cli->cl_loi_list_lock);
1866 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
1870 static void brw_commit(struct ptlrpc_request *req)
1872 /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1873 * this called via the rq_commit_cb, I need to ensure
1874 * osc_dec_unstable_pages is still called. Otherwise unstable
1875 * pages may be leaked. */
1876 spin_lock(&req->rq_lock);
1877 if (likely(req->rq_unstable)) {
1878 req->rq_unstable = 0;
1879 spin_unlock(&req->rq_lock);
1881 osc_dec_unstable_pages(req);
1883 req->rq_committed = 1;
1884 spin_unlock(&req->rq_lock);
1889 * Build an RPC by the list of extent @ext_list. The caller must ensure
1890 * that the total pages in this list are NOT over max pages per RPC.
1891 * Extents in the list must be in OES_RPC state.
1893 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1894 cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
1896 struct ptlrpc_request *req = NULL;
1897 struct osc_extent *ext;
1898 struct brw_page **pga = NULL;
1899 struct osc_brw_async_args *aa = NULL;
1900 struct obdo *oa = NULL;
1901 struct osc_async_page *oap;
1902 struct osc_async_page *tmp;
1903 struct cl_req *clerq = NULL;
1904 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
1906 struct ldlm_lock *lock = NULL;
1907 struct cl_req_attr *crattr = NULL;
1908 obd_off starting_offset = OBD_OBJECT_EOF;
1909 obd_off ending_offset = 0;
1915 CFS_LIST_HEAD(rpc_list);
1918 LASSERT(!cfs_list_empty(ext_list));
1920 /* add pages into rpc_list to build BRW rpc */
1921 cfs_list_for_each_entry(ext, ext_list, oe_link) {
1922 LASSERT(ext->oe_state == OES_RPC);
1923 mem_tight |= ext->oe_memalloc;
1924 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1926 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
1927 if (starting_offset > oap->oap_obj_off)
1928 starting_offset = oap->oap_obj_off;
1930 LASSERT(oap->oap_page_off == 0);
1931 if (ending_offset < oap->oap_obj_off + oap->oap_count)
1932 ending_offset = oap->oap_obj_off +
1935 LASSERT(oap->oap_page_off + oap->oap_count ==
1941 mpflag = cfs_memory_pressure_get_and_set();
1943 OBD_ALLOC(crattr, sizeof(*crattr));
1945 GOTO(out, rc = -ENOMEM);
1947 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1949 GOTO(out, rc = -ENOMEM);
1953 GOTO(out, rc = -ENOMEM);
1956 cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1957 struct cl_page *page = oap2cl_page(oap);
1958 if (clerq == NULL) {
1959 clerq = cl_req_alloc(env, page, crt,
1960 1 /* only 1-object rpcs for now */);
1962 GOTO(out, rc = PTR_ERR(clerq));
1963 lock = oap->oap_ldlm_lock;
1966 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1967 pga[i] = &oap->oap_brw_page;
1968 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1969 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1970 pga[i]->pg, page_index(oap->oap_page), oap,
1973 cl_req_page_add(env, clerq, page);
1976 /* always get the data for the obdo for the rpc */
1977 LASSERT(clerq != NULL);
1978 crattr->cra_oa = oa;
1979 cl_req_attr_set(env, clerq, crattr, ~0ULL);
1981 oa->o_handle = lock->l_remote_handle;
1982 oa->o_valid |= OBD_MD_FLHANDLE;
1985 rc = cl_req_prep(env, clerq);
1987 CERROR("cl_req_prep failed: %d\n", rc);
1991 sort_brw_pages(pga, page_count);
1992 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1993 pga, &req, crattr->cra_capa, 1, 0);
1995 CERROR("prep_req failed: %d\n", rc);
1999 req->rq_commit_cb = brw_commit;
2000 req->rq_interpret_reply = brw_interpret;
2003 req->rq_memalloc = 1;
2005 /* Need to update the timestamps after the request is built in case
2006 * we race with setattr (locally or in queue at OST). If OST gets
2007 * later setattr before earlier BRW (as determined by the request xid),
2008 * the OST will not use BRW timestamps. Sadly, there is no obvious
2009 * way to do this in a single call. bug 10150 */
2010 cl_req_attr_set(env, clerq, crattr,
2011 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2013 lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2015 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2016 aa = ptlrpc_req_async_args(req);
2017 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2018 cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2019 CFS_INIT_LIST_HEAD(&aa->aa_exts);
2020 cfs_list_splice_init(ext_list, &aa->aa_exts);
2021 aa->aa_clerq = clerq;
2023 /* queued sync pages can be torn down while the pages
2024 * were between the pending list and the rpc */
2026 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2027 /* only one oap gets a request reference */
2030 if (oap->oap_interrupted && !req->rq_intr) {
2031 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2033 ptlrpc_mark_interrupted(req);
2037 tmp->oap_request = ptlrpc_request_addref(req);
2039 client_obd_list_lock(&cli->cl_loi_list_lock);
2040 starting_offset >>= PAGE_CACHE_SHIFT;
2041 if (cmd == OBD_BRW_READ) {
2042 cli->cl_r_in_flight++;
2043 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2044 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2045 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2046 starting_offset + 1);
2048 cli->cl_w_in_flight++;
2049 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2050 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2051 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2052 starting_offset + 1);
2054 client_obd_list_unlock(&cli->cl_loi_list_lock);
2056 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2057 page_count, aa, cli->cl_r_in_flight,
2058 cli->cl_w_in_flight);
2060 /* XXX: Maybe the caller can check the RPC bulk descriptor to
2061 * see which CPU/NUMA node the majority of pages were allocated
2062 * on, and try to assign the async RPC to the CPU core
2063 * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2065 * But on the other hand, we expect that multiple ptlrpcd
2066 * threads and the initial write sponsor can run in parallel,
2067 * especially when data checksum is enabled, which is CPU-bound
2068 * operation and single ptlrpcd thread cannot process in time.
2069 * So more ptlrpcd threads sharing BRW load
2070 * (with PDL_POLICY_ROUND) seems better.
2072 ptlrpcd_add_req(req, pol, -1);
2078 cfs_memory_pressure_restore(mpflag);
2080 if (crattr != NULL) {
2081 capa_put(crattr->cra_capa);
2082 OBD_FREE(crattr, sizeof(*crattr));
2086 LASSERT(req == NULL);
2091 OBD_FREE(pga, sizeof(*pga) * page_count);
2092 /* this should happen rarely and is pretty bad, it makes the
2093 * pending list not follow the dirty order */
2094 while (!cfs_list_empty(ext_list)) {
2095 ext = cfs_list_entry(ext_list->next, struct osc_extent,
2097 cfs_list_del_init(&ext->oe_link);
2098 osc_extent_finish(env, ext, 0, rc);
2100 if (clerq && !IS_ERR(clerq))
2101 cl_req_completion(env, clerq, rc);
2106 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2107 struct ldlm_enqueue_info *einfo)
2109 void *data = einfo->ei_cbdata;
2112 LASSERT(lock != NULL);
2113 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2114 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2115 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2116 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2118 lock_res_and_lock(lock);
2119 spin_lock(&osc_ast_guard);
2121 if (lock->l_ast_data == NULL)
2122 lock->l_ast_data = data;
2123 if (lock->l_ast_data == data)
2126 spin_unlock(&osc_ast_guard);
2127 unlock_res_and_lock(lock);
2132 static int osc_set_data_with_check(struct lustre_handle *lockh,
2133 struct ldlm_enqueue_info *einfo)
2135 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2139 set = osc_set_lock_data_with_check(lock, einfo);
2140 LDLM_LOCK_PUT(lock);
2142 CERROR("lockh %p, data %p - client evicted?\n",
2143 lockh, einfo->ei_cbdata);
2147 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2148 ldlm_iterator_t replace, void *data)
2150 struct ldlm_res_id res_id;
2151 struct obd_device *obd = class_exp2obd(exp);
2153 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2154 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2158 /* find any ldlm lock of the inode in osc
2162 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2163 ldlm_iterator_t replace, void *data)
2165 struct ldlm_res_id res_id;
2166 struct obd_device *obd = class_exp2obd(exp);
2169 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2170 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2171 if (rc == LDLM_ITER_STOP)
2173 if (rc == LDLM_ITER_CONTINUE)
2178 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2179 obd_enqueue_update_f upcall, void *cookie,
2180 __u64 *flags, int agl, int rc)
2182 int intent = *flags & LDLM_FL_HAS_INTENT;
2186 /* The request was created before ldlm_cli_enqueue call. */
2187 if (rc == ELDLM_LOCK_ABORTED) {
2188 struct ldlm_reply *rep;
2189 rep = req_capsule_server_get(&req->rq_pill,
2192 LASSERT(rep != NULL);
2193 rep->lock_policy_res1 =
2194 ptlrpc_status_ntoh(rep->lock_policy_res1);
2195 if (rep->lock_policy_res1)
2196 rc = rep->lock_policy_res1;
2200 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2202 *flags |= LDLM_FL_LVB_READY;
2203 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2204 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2207 /* Call the update callback. */
2208 rc = (*upcall)(cookie, rc);
2212 static int osc_enqueue_interpret(const struct lu_env *env,
2213 struct ptlrpc_request *req,
2214 struct osc_enqueue_args *aa, int rc)
2216 struct ldlm_lock *lock;
2217 struct lustre_handle handle;
2219 struct ost_lvb *lvb;
2221 __u64 *flags = aa->oa_flags;
2223 /* Make a local copy of a lock handle and a mode, because aa->oa_*
2224 * might be freed anytime after lock upcall has been called. */
2225 lustre_handle_copy(&handle, aa->oa_lockh);
2226 mode = aa->oa_ei->ei_mode;
2228 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2230 lock = ldlm_handle2lock(&handle);
2232 /* Take an additional reference so that a blocking AST that
2233 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2234 * to arrive after an upcall has been executed by
2235 * osc_enqueue_fini(). */
2236 ldlm_lock_addref(&handle, mode);
2238 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2239 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2241 /* Let CP AST to grant the lock first. */
2242 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2244 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2249 lvb_len = sizeof(*aa->oa_lvb);
2252 /* Complete obtaining the lock procedure. */
2253 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2254 mode, flags, lvb, lvb_len, &handle, rc);
2255 /* Complete osc stuff. */
2256 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2257 flags, aa->oa_agl, rc);
2259 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2261 /* Release the lock for async request. */
2262 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2264 * Releases a reference taken by ldlm_cli_enqueue(), if it is
2265 * not already released by
2266 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2268 ldlm_lock_decref(&handle, mode);
2270 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2271 aa->oa_lockh, req, aa);
2272 ldlm_lock_decref(&handle, mode);
2273 LDLM_LOCK_PUT(lock);
2277 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2279 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2280 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2281 * other synchronous requests, however keeping some locks and trying to obtain
2282 * others may take a considerable amount of time in a case of ost failure; and
2283 * when other sync requests do not get released lock from a client, the client
2284 * is excluded from the cluster -- such scenarious make the life difficult, so
2285 * release locks just after they are obtained. */
2286 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2287 __u64 *flags, ldlm_policy_data_t *policy,
2288 struct ost_lvb *lvb, int kms_valid,
2289 obd_enqueue_update_f upcall, void *cookie,
2290 struct ldlm_enqueue_info *einfo,
2291 struct lustre_handle *lockh,
2292 struct ptlrpc_request_set *rqset, int async, int agl)
2294 struct obd_device *obd = exp->exp_obd;
2295 struct ptlrpc_request *req = NULL;
2296 int intent = *flags & LDLM_FL_HAS_INTENT;
2297 __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2302 /* Filesystem lock extents are extended to page boundaries so that
2303 * dealing with the page cache is a little smoother. */
2304 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2305 policy->l_extent.end |= ~CFS_PAGE_MASK;
2308 * kms is not valid when either object is completely fresh (so that no
2309 * locks are cached), or object was evicted. In the latter case cached
2310 * lock cannot be used, because it would prime inode state with
2311 * potentially stale LVB.
2316 /* Next, search for already existing extent locks that will cover us */
2317 /* If we're trying to read, we also search for an existing PW lock. The
2318 * VFS and page cache already protect us locally, so lots of readers/
2319 * writers can share a single PW lock.
2321 * There are problems with conversion deadlocks, so instead of
2322 * converting a read lock to a write lock, we'll just enqueue a new
2325 * At some point we should cancel the read lock instead of making them
2326 * send us a blocking callback, but there are problems with canceling
2327 * locks out from other users right now, too. */
2328 mode = einfo->ei_mode;
2329 if (einfo->ei_mode == LCK_PR)
2331 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2332 einfo->ei_type, policy, mode, lockh, 0);
2334 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2336 if ((agl != 0) && !ldlm_is_lvb_ready(matched)) {
2337 /* For AGL, if enqueue RPC is sent but the lock is not
2338 * granted, then skip to process this strpe.
2339 * Return -ECANCELED to tell the caller. */
2340 ldlm_lock_decref(lockh, mode);
2341 LDLM_LOCK_PUT(matched);
2343 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2344 *flags |= LDLM_FL_LVB_READY;
2345 /* addref the lock only if not async requests and PW
2346 * lock is matched whereas we asked for PR. */
2347 if (!rqset && einfo->ei_mode != mode)
2348 ldlm_lock_addref(lockh, LCK_PR);
2350 /* I would like to be able to ASSERT here that
2351 * rss <= kms, but I can't, for reasons which
2352 * are explained in lov_enqueue() */
2355 /* We already have a lock, and it's referenced.
2357 * At this point, the cl_lock::cll_state is CLS_QUEUING,
2358 * AGL upcall may change it to CLS_HELD directly. */
2359 (*upcall)(cookie, ELDLM_OK);
2361 if (einfo->ei_mode != mode)
2362 ldlm_lock_decref(lockh, LCK_PW);
2364 /* For async requests, decref the lock. */
2365 ldlm_lock_decref(lockh, einfo->ei_mode);
2366 LDLM_LOCK_PUT(matched);
2369 ldlm_lock_decref(lockh, mode);
2370 LDLM_LOCK_PUT(matched);
2376 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2377 &RQF_LDLM_ENQUEUE_LVB);
2381 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2383 ptlrpc_request_free(req);
2387 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2389 ptlrpc_request_set_replen(req);
2392 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2393 *flags &= ~LDLM_FL_BLOCK_GRANTED;
2395 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2396 sizeof(*lvb), LVB_T_OST, lockh, async);
2399 struct osc_enqueue_args *aa;
2400 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2401 aa = ptlrpc_req_async_args(req);
2404 aa->oa_flags = flags;
2405 aa->oa_upcall = upcall;
2406 aa->oa_cookie = cookie;
2408 aa->oa_lockh = lockh;
2411 req->rq_interpret_reply =
2412 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2413 if (rqset == PTLRPCD_SET)
2414 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2416 ptlrpc_set_add_req(rqset, req);
2417 } else if (intent) {
2418 ptlrpc_req_finished(req);
2423 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2425 ptlrpc_req_finished(req);
2430 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2431 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2432 __u64 *flags, void *data, struct lustre_handle *lockh,
2435 struct obd_device *obd = exp->exp_obd;
2436 __u64 lflags = *flags;
2440 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2443 /* Filesystem lock extents are extended to page boundaries so that
2444 * dealing with the page cache is a little smoother */
2445 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2446 policy->l_extent.end |= ~CFS_PAGE_MASK;
2448 /* Next, search for already existing extent locks that will cover us */
2449 /* If we're trying to read, we also search for an existing PW lock. The
2450 * VFS and page cache already protect us locally, so lots of readers/
2451 * writers can share a single PW lock. */
2455 rc = ldlm_lock_match(obd->obd_namespace, lflags,
2456 res_id, type, policy, rc, lockh, unref);
2459 if (!osc_set_data_with_check(lockh, data)) {
2460 if (!(lflags & LDLM_FL_TEST_LOCK))
2461 ldlm_lock_decref(lockh, rc);
2465 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2466 ldlm_lock_addref(lockh, LCK_PR);
2467 ldlm_lock_decref(lockh, LCK_PW);
2474 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2478 if (unlikely(mode == LCK_GROUP))
2479 ldlm_lock_decref_and_cancel(lockh, mode);
2481 ldlm_lock_decref(lockh, mode);
2486 static int osc_statfs_interpret(const struct lu_env *env,
2487 struct ptlrpc_request *req,
2488 struct osc_async_args *aa, int rc)
2490 struct obd_statfs *msfs;
2494 /* The request has in fact never been sent
2495 * due to issues at a higher level (LOV).
2496 * Exit immediately since the caller is
2497 * aware of the problem and takes care
2498 * of the clean up */
2501 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2502 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2508 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2510 GOTO(out, rc = -EPROTO);
2513 *aa->aa_oi->oi_osfs = *msfs;
2515 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2519 static int osc_statfs_async(struct obd_export *exp,
2520 struct obd_info *oinfo, __u64 max_age,
2521 struct ptlrpc_request_set *rqset)
2523 struct obd_device *obd = class_exp2obd(exp);
2524 struct ptlrpc_request *req;
2525 struct osc_async_args *aa;
2529 /* We could possibly pass max_age in the request (as an absolute
2530 * timestamp or a "seconds.usec ago") so the target can avoid doing
2531 * extra calls into the filesystem if that isn't necessary (e.g.
2532 * during mount that would help a bit). Having relative timestamps
2533 * is not so great if request processing is slow, while absolute
2534 * timestamps are not ideal because they need time synchronization. */
2535 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2539 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2541 ptlrpc_request_free(req);
2544 ptlrpc_request_set_replen(req);
2545 req->rq_request_portal = OST_CREATE_PORTAL;
2546 ptlrpc_at_set_req_timeout(req);
2548 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2549 /* procfs requests not want stat in wait for avoid deadlock */
2550 req->rq_no_resend = 1;
2551 req->rq_no_delay = 1;
2554 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2555 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2556 aa = ptlrpc_req_async_args(req);
2559 ptlrpc_set_add_req(rqset, req);
2563 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2564 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2566 struct obd_device *obd = class_exp2obd(exp);
2567 struct obd_statfs *msfs;
2568 struct ptlrpc_request *req;
2569 struct obd_import *imp = NULL;
2573 /*Since the request might also come from lprocfs, so we need
2574 *sync this with client_disconnect_export Bug15684*/
2575 down_read(&obd->u.cli.cl_sem);
2576 if (obd->u.cli.cl_import)
2577 imp = class_import_get(obd->u.cli.cl_import);
2578 up_read(&obd->u.cli.cl_sem);
2582 /* We could possibly pass max_age in the request (as an absolute
2583 * timestamp or a "seconds.usec ago") so the target can avoid doing
2584 * extra calls into the filesystem if that isn't necessary (e.g.
2585 * during mount that would help a bit). Having relative timestamps
2586 * is not so great if request processing is slow, while absolute
2587 * timestamps are not ideal because they need time synchronization. */
2588 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2590 class_import_put(imp);
2595 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2597 ptlrpc_request_free(req);
2600 ptlrpc_request_set_replen(req);
2601 req->rq_request_portal = OST_CREATE_PORTAL;
2602 ptlrpc_at_set_req_timeout(req);
2604 if (flags & OBD_STATFS_NODELAY) {
2605 /* procfs requests not want stat in wait for avoid deadlock */
2606 req->rq_no_resend = 1;
2607 req->rq_no_delay = 1;
2610 rc = ptlrpc_queue_wait(req);
2614 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2616 GOTO(out, rc = -EPROTO);
2623 ptlrpc_req_finished(req);
2627 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2628 void *karg, void *uarg)
2630 struct obd_device *obd = exp->exp_obd;
2631 struct obd_ioctl_data *data = karg;
2635 if (!try_module_get(THIS_MODULE)) {
2636 CERROR("Can't get module. Is it alive?");
2640 case OBD_IOC_CLIENT_RECOVER:
2641 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2642 data->ioc_inlbuf1, 0);
2646 case IOC_OSC_SET_ACTIVE:
2647 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2650 case OBD_IOC_POLL_QUOTACHECK:
2651 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
2653 case OBD_IOC_PING_TARGET:
2654 err = ptlrpc_obd_ping(obd);
2657 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2658 cmd, current_comm());
2659 GOTO(out, err = -ENOTTY);
2662 module_put(THIS_MODULE);
2666 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2667 obd_count keylen, void *key, __u32 *vallen, void *val,
2668 struct lov_stripe_md *lsm)
2671 if (!vallen || !val)
2674 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2675 __u32 *stripe = val;
2676 *vallen = sizeof(*stripe);
2679 } else if (KEY_IS(KEY_LAST_ID)) {
2680 struct ptlrpc_request *req;
2685 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2686 &RQF_OST_GET_INFO_LAST_ID);
2690 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2691 RCL_CLIENT, keylen);
2692 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2694 ptlrpc_request_free(req);
2698 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2699 memcpy(tmp, key, keylen);
2701 req->rq_no_delay = req->rq_no_resend = 1;
2702 ptlrpc_request_set_replen(req);
2703 rc = ptlrpc_queue_wait(req);
2707 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2709 GOTO(out, rc = -EPROTO);
2711 *((obd_id *)val) = *reply;
2713 ptlrpc_req_finished(req);
2715 } else if (KEY_IS(KEY_FIEMAP)) {
2716 struct ll_fiemap_info_key *fm_key =
2717 (struct ll_fiemap_info_key *)key;
2718 struct ldlm_res_id res_id;
2719 ldlm_policy_data_t policy;
2720 struct lustre_handle lockh;
2721 ldlm_mode_t mode = 0;
2722 struct ptlrpc_request *req;
2723 struct ll_user_fiemap *reply;
2727 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2730 policy.l_extent.start = fm_key->fiemap.fm_start &
2733 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2734 fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
2735 policy.l_extent.end = OBD_OBJECT_EOF;
2737 policy.l_extent.end = (fm_key->fiemap.fm_start +
2738 fm_key->fiemap.fm_length +
2739 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
2741 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2742 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2743 LDLM_FL_BLOCK_GRANTED |
2745 &res_id, LDLM_EXTENT, &policy,
2746 LCK_PR | LCK_PW, &lockh, 0);
2747 if (mode) { /* lock is cached on client */
2748 if (mode != LCK_PR) {
2749 ldlm_lock_addref(&lockh, LCK_PR);
2750 ldlm_lock_decref(&lockh, LCK_PW);
2752 } else { /* no cached lock, needs acquire lock on server side */
2753 fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2754 fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2758 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2759 &RQF_OST_GET_INFO_FIEMAP);
2761 GOTO(drop_lock, rc = -ENOMEM);
2763 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2764 RCL_CLIENT, keylen);
2765 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2766 RCL_CLIENT, *vallen);
2767 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2768 RCL_SERVER, *vallen);
2770 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2772 ptlrpc_request_free(req);
2773 GOTO(drop_lock, rc);
2776 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2777 memcpy(tmp, key, keylen);
2778 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2779 memcpy(tmp, val, *vallen);
2781 ptlrpc_request_set_replen(req);
2782 rc = ptlrpc_queue_wait(req);
2786 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2788 GOTO(fini_req, rc = -EPROTO);
2790 memcpy(val, reply, *vallen);
2792 ptlrpc_req_finished(req);
2795 ldlm_lock_decref(&lockh, LCK_PR);
2802 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2803 obd_count keylen, void *key, obd_count vallen,
2804 void *val, struct ptlrpc_request_set *set)
2806 struct ptlrpc_request *req;
2807 struct obd_device *obd = exp->exp_obd;
2808 struct obd_import *imp = class_exp2cliimp(exp);
2813 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2815 if (KEY_IS(KEY_CHECKSUM)) {
2816 if (vallen != sizeof(int))
2818 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2822 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2823 sptlrpc_conf_client_adapt(obd);
2827 if (KEY_IS(KEY_FLUSH_CTX)) {
2828 sptlrpc_import_flush_my_ctx(imp);
2832 if (KEY_IS(KEY_CACHE_SET)) {
2833 struct client_obd *cli = &obd->u.cli;
2835 LASSERT(cli->cl_cache == NULL); /* only once */
2836 cli->cl_cache = (struct cl_client_cache *)val;
2837 atomic_inc(&cli->cl_cache->ccc_users);
2838 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2840 /* add this osc into entity list */
2841 LASSERT(cfs_list_empty(&cli->cl_lru_osc));
2842 spin_lock(&cli->cl_cache->ccc_lru_lock);
2843 cfs_list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2844 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2849 if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2850 struct client_obd *cli = &obd->u.cli;
2851 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2852 int target = *(int *)val;
2854 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2859 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2862 /* We pass all other commands directly to OST. Since nobody calls osc
2863 methods directly and everybody is supposed to go through LOV, we
2864 assume lov checked invalid values for us.
2865 The only recognised values so far are evict_by_nid and mds_conn.
2866 Even if something bad goes through, we'd get a -EINVAL from OST
2869 req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2870 &RQF_OST_SET_GRANT_INFO :
2875 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2876 RCL_CLIENT, keylen);
2877 if (!KEY_IS(KEY_GRANT_SHRINK))
2878 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2879 RCL_CLIENT, vallen);
2880 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2882 ptlrpc_request_free(req);
2886 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2887 memcpy(tmp, key, keylen);
2888 tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2891 memcpy(tmp, val, vallen);
2893 if (KEY_IS(KEY_GRANT_SHRINK)) {
2894 struct osc_grant_args *aa;
2897 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2898 aa = ptlrpc_req_async_args(req);
2901 ptlrpc_req_finished(req);
2904 *oa = ((struct ost_body *)val)->oa;
2906 req->rq_interpret_reply = osc_shrink_grant_interpret;
2909 ptlrpc_request_set_replen(req);
2910 if (!KEY_IS(KEY_GRANT_SHRINK)) {
2911 LASSERT(set != NULL);
2912 ptlrpc_set_add_req(set, req);
2913 ptlrpc_check_set(NULL, set);
2915 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2920 static int osc_reconnect(const struct lu_env *env,
2921 struct obd_export *exp, struct obd_device *obd,
2922 struct obd_uuid *cluuid,
2923 struct obd_connect_data *data,
2926 struct client_obd *cli = &obd->u.cli;
2928 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2931 client_obd_list_lock(&cli->cl_loi_list_lock);
2932 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2933 2 * cli_brw_size(obd);
2934 lost_grant = cli->cl_lost_grant;
2935 cli->cl_lost_grant = 0;
2936 client_obd_list_unlock(&cli->cl_loi_list_lock);
2938 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
2939 " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2940 data->ocd_version, data->ocd_grant, lost_grant);
2946 static int osc_disconnect(struct obd_export *exp)
2948 struct obd_device *obd = class_exp2obd(exp);
2949 struct llog_ctxt *ctxt;
2952 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2954 if (obd->u.cli.cl_conn_count == 1) {
2955 /* Flush any remaining cancel messages out to the
2957 llog_sync(ctxt, exp, 0);
2959 llog_ctxt_put(ctxt);
2961 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
2965 rc = client_disconnect_export(exp);
2967 * Initially we put del_shrink_grant before disconnect_export, but it
2968 * causes the following problem if setup (connect) and cleanup
2969 * (disconnect) are tangled together.
2970 * connect p1 disconnect p2
2971 * ptlrpc_connect_import
2972 * ............... class_manual_cleanup
2975 * ptlrpc_connect_interrupt
2977 * add this client to shrink list
2979 * Bang! pinger trigger the shrink.
2980 * So the osc should be disconnected from the shrink list, after we
2981 * are sure the import has been destroyed. BUG18662
2983 if (obd->u.cli.cl_import == NULL)
2984 osc_del_shrink_grant(&obd->u.cli);
2988 static int osc_import_event(struct obd_device *obd,
2989 struct obd_import *imp,
2990 enum obd_import_event event)
2992 struct client_obd *cli;
2996 LASSERT(imp->imp_obd == obd);
2999 case IMP_EVENT_DISCON: {
3001 client_obd_list_lock(&cli->cl_loi_list_lock);
3002 cli->cl_avail_grant = 0;
3003 cli->cl_lost_grant = 0;
3004 client_obd_list_unlock(&cli->cl_loi_list_lock);
3007 case IMP_EVENT_INACTIVE: {
3008 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3011 case IMP_EVENT_INVALIDATE: {
3012 struct ldlm_namespace *ns = obd->obd_namespace;
3016 env = cl_env_get(&refcheck);
3020 /* all pages go to failing rpcs due to the invalid
3022 osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3024 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3025 cl_env_put(env, &refcheck);
3030 case IMP_EVENT_ACTIVE: {
3031 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3034 case IMP_EVENT_OCD: {
3035 struct obd_connect_data *ocd = &imp->imp_connect_data;
3037 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3038 osc_init_grant(&obd->u.cli, ocd);
3041 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3042 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3044 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3047 case IMP_EVENT_DEACTIVATE: {
3048 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3051 case IMP_EVENT_ACTIVATE: {
3052 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3056 CERROR("Unknown import event %d\n", event);
3063 * Determine whether the lock can be canceled before replaying the lock
3064 * during recovery, see bug16774 for detailed information.
3066 * \retval zero the lock can't be canceled
3067 * \retval other ok to cancel
3069 static int osc_cancel_weight(struct ldlm_lock *lock)
3072 * Cancel all unused and granted extent lock.
3074 if (lock->l_resource->lr_type == LDLM_EXTENT &&
3075 lock->l_granted_mode == lock->l_req_mode &&
3076 osc_ldlm_weigh_ast(lock) == 0)
3082 static int brw_queue_work(const struct lu_env *env, void *data)
3084 struct client_obd *cli = data;
3086 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3088 osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3092 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3094 struct client_obd *cli = &obd->u.cli;
3099 rc = ptlrpcd_addref();
3103 rc = client_obd_setup(obd, lcfg);
3105 GOTO(out_ptlrpcd, rc);
3107 handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3108 if (IS_ERR(handler))
3109 GOTO(out_client_setup, rc = PTR_ERR(handler));
3110 cli->cl_writeback_work = handler;
3112 handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3113 if (IS_ERR(handler))
3114 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3115 cli->cl_lru_work = handler;
3117 rc = osc_quota_setup(obd);
3119 GOTO(out_ptlrpcd_work, rc);
3121 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3123 obd->obd_vars = lprocfs_osc_obd_vars;
3125 if (lprocfs_seq_obd_setup(obd) == 0) {
3126 lproc_osc_attach_seqstat(obd);
3127 sptlrpc_lprocfs_cliobd_attach(obd);
3128 ptlrpc_lprocfs_register_obd(obd);
3131 /* We need to allocate a few requests more, because
3132 * brw_interpret tries to create new requests before freeing
3133 * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3134 * reserved, but I'm afraid that might be too much wasted RAM
3135 * in fact, so 2 is just my guess and still should work. */
3136 cli->cl_import->imp_rq_pool =
3137 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3139 ptlrpc_add_rqs_to_pool);
3141 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3142 ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3146 if (cli->cl_writeback_work != NULL) {
3147 ptlrpcd_destroy_work(cli->cl_writeback_work);
3148 cli->cl_writeback_work = NULL;
3150 if (cli->cl_lru_work != NULL) {
3151 ptlrpcd_destroy_work(cli->cl_lru_work);
3152 cli->cl_lru_work = NULL;
3155 client_obd_cleanup(obd);
3161 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3167 case OBD_CLEANUP_EARLY: {
3168 struct obd_import *imp;
3169 imp = obd->u.cli.cl_import;
3170 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3171 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3172 ptlrpc_deactivate_import(imp);
3173 spin_lock(&imp->imp_lock);
3174 imp->imp_pingable = 0;
3175 spin_unlock(&imp->imp_lock);
3178 case OBD_CLEANUP_EXPORTS: {
3179 struct client_obd *cli = &obd->u.cli;
3181 * for echo client, export may be on zombie list, wait for
3182 * zombie thread to cull it, because cli.cl_import will be
3183 * cleared in client_disconnect_export():
3184 * class_export_destroy() -> obd_cleanup() ->
3185 * echo_device_free() -> echo_client_cleanup() ->
3186 * obd_disconnect() -> osc_disconnect() ->
3187 * client_disconnect_export()
3189 obd_zombie_barrier();
3190 if (cli->cl_writeback_work) {
3191 ptlrpcd_destroy_work(cli->cl_writeback_work);
3192 cli->cl_writeback_work = NULL;
3194 if (cli->cl_lru_work) {
3195 ptlrpcd_destroy_work(cli->cl_lru_work);
3196 cli->cl_lru_work = NULL;
3198 obd_cleanup_client_import(obd);
3199 ptlrpc_lprocfs_unregister_obd(obd);
3200 lprocfs_obd_cleanup(obd);
3201 rc = obd_llog_finish(obd, 0);
3203 CERROR("failed to cleanup llogging subsystems\n");
3210 int osc_cleanup(struct obd_device *obd)
3212 struct client_obd *cli = &obd->u.cli;
3218 if (cli->cl_cache != NULL) {
3219 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3220 spin_lock(&cli->cl_cache->ccc_lru_lock);
3221 cfs_list_del_init(&cli->cl_lru_osc);
3222 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3223 cli->cl_lru_left = NULL;
3224 atomic_dec(&cli->cl_cache->ccc_users);
3225 cli->cl_cache = NULL;
3228 /* free memory of osc quota cache */
3229 osc_quota_cleanup(obd);
3231 rc = client_obd_cleanup(obd);
3237 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3239 int rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
3241 return rc > 0 ? 0: rc;
3244 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
3246 return osc_process_config_base(obd, buf);
3249 struct obd_ops osc_obd_ops = {
3250 .o_owner = THIS_MODULE,
3251 .o_setup = osc_setup,
3252 .o_precleanup = osc_precleanup,
3253 .o_cleanup = osc_cleanup,
3254 .o_add_conn = client_import_add_conn,
3255 .o_del_conn = client_import_del_conn,
3256 .o_connect = client_connect_import,
3257 .o_reconnect = osc_reconnect,
3258 .o_disconnect = osc_disconnect,
3259 .o_statfs = osc_statfs,
3260 .o_statfs_async = osc_statfs_async,
3261 .o_unpackmd = osc_unpackmd,
3262 .o_create = osc_create,
3263 .o_destroy = osc_destroy,
3264 .o_getattr = osc_getattr,
3265 .o_getattr_async = osc_getattr_async,
3266 .o_setattr = osc_setattr,
3267 .o_setattr_async = osc_setattr_async,
3268 .o_change_cbdata = osc_change_cbdata,
3269 .o_find_cbdata = osc_find_cbdata,
3270 .o_iocontrol = osc_iocontrol,
3271 .o_get_info = osc_get_info,
3272 .o_set_info_async = osc_set_info_async,
3273 .o_import_event = osc_import_event,
3274 .o_process_config = osc_process_config,
3275 .o_quotactl = osc_quotactl,
3276 .o_quotacheck = osc_quotacheck,
3279 extern struct lu_kmem_descr osc_caches[];
3280 extern spinlock_t osc_ast_guard;
3281 extern struct lock_class_key osc_ast_guard_class;
3283 int __init osc_init(void)
3288 /* print an address of _any_ initialized kernel symbol from this
3289 * module, to allow debugging with gdb that doesn't support data
3290 * symbols from modules.*/
3291 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3293 rc = lu_kmem_init(osc_caches);
3297 rc = class_register_type(&osc_obd_ops, NULL, NULL,
3298 #ifndef HAVE_ONLY_PROCFS_SEQ
3301 LUSTRE_OSC_NAME, &osc_device_type);
3303 lu_kmem_fini(osc_caches);
3307 spin_lock_init(&osc_ast_guard);
3308 lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3314 static void /*__exit*/ osc_exit(void)
3316 class_unregister_type(LUSTRE_OSC_NAME);
3317 lu_kmem_fini(osc_caches);
3320 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3321 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3322 MODULE_LICENSE("GPL");
3324 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);