1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 # define EXPORT_SYMTAB
42 #define DEBUG_SUBSYSTEM S_OSC
44 #include <libcfs/libcfs.h>
47 # include <liblustre.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre/lustre_user.h>
53 #include <obd_cksum.h>
61 #include <lustre_ha.h>
62 #include <lprocfs_status.h>
63 #include <lustre_log.h>
64 #include <lustre_debug.h>
65 #include <lustre_param.h>
66 #include "osc_internal.h"
68 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
69 static int brw_interpret(const struct lu_env *env,
70 struct ptlrpc_request *req, void *data, int rc);
71 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli,
73 int osc_cleanup(struct obd_device *obd);
75 /* Pack OSC object metadata for disk storage (LE byte order). */
76 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
77 struct lov_stripe_md *lsm)
82 lmm_size = sizeof(**lmmp);
87 OBD_FREE(*lmmp, lmm_size);
93 OBD_ALLOC(*lmmp, lmm_size);
99 LASSERT(lsm->lsm_object_id);
100 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
101 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
102 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
108 /* Unpack OSC object metadata from disk storage (LE byte order). */
109 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
110 struct lov_mds_md *lmm, int lmm_bytes)
113 struct obd_import *imp = class_exp2cliimp(exp);
117 if (lmm_bytes < sizeof (*lmm)) {
118 CERROR("lov_mds_md too small: %d, need %d\n",
119 lmm_bytes, (int)sizeof(*lmm));
122 /* XXX LOV_MAGIC etc check? */
124 if (lmm->lmm_object_id == 0) {
125 CERROR("lov_mds_md: zero lmm_object_id\n");
130 lsm_size = lov_stripe_md_size(1);
134 if (*lsmp != NULL && lmm == NULL) {
135 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
136 OBD_FREE(*lsmp, lsm_size);
142 OBD_ALLOC(*lsmp, lsm_size);
145 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
146 if ((*lsmp)->lsm_oinfo[0] == NULL) {
147 OBD_FREE(*lsmp, lsm_size);
150 loi_init((*lsmp)->lsm_oinfo[0]);
154 /* XXX zero *lsmp? */
155 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
156 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
157 LASSERT((*lsmp)->lsm_object_id);
158 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
162 (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
163 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
165 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
170 static inline void osc_pack_capa(struct ptlrpc_request *req,
171 struct ost_body *body, void *capa)
173 struct obd_capa *oc = (struct obd_capa *)capa;
174 struct lustre_capa *c;
179 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
182 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
183 DEBUG_CAPA(D_SEC, c, "pack");
186 static inline void osc_pack_req_body(struct ptlrpc_request *req,
187 struct obd_info *oinfo)
189 struct ost_body *body;
191 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
194 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
195 osc_pack_capa(req, body, oinfo->oi_capa);
198 static inline void osc_set_capa_size(struct ptlrpc_request *req,
199 const struct req_msg_field *field,
203 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
205 /* it is already calculated as sizeof struct obd_capa */
209 static int osc_getattr_interpret(const struct lu_env *env,
210 struct ptlrpc_request *req,
211 struct osc_async_args *aa, int rc)
213 struct ost_body *body;
219 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
221 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
222 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
224 /* This should really be sent by the OST */
225 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
226 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
228 CDEBUG(D_INFO, "can't unpack ost_body\n");
230 aa->aa_oi->oi_oa->o_valid = 0;
233 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
237 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
238 struct ptlrpc_request_set *set)
240 struct ptlrpc_request *req;
241 struct osc_async_args *aa;
245 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
249 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
250 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
252 ptlrpc_request_free(req);
256 osc_pack_req_body(req, oinfo);
258 ptlrpc_request_set_replen(req);
259 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
261 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
262 aa = ptlrpc_req_async_args(req);
265 ptlrpc_set_add_req(set, req);
269 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
271 struct ptlrpc_request *req;
272 struct ost_body *body;
276 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
280 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
281 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
283 ptlrpc_request_free(req);
287 osc_pack_req_body(req, oinfo);
289 ptlrpc_request_set_replen(req);
291 rc = ptlrpc_queue_wait(req);
295 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
297 GOTO(out, rc = -EPROTO);
299 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
300 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
302 /* This should really be sent by the OST */
303 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
304 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
308 ptlrpc_req_finished(req);
312 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
313 struct obd_trans_info *oti)
315 struct ptlrpc_request *req;
316 struct ost_body *body;
320 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
322 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
326 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
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);
343 GOTO(out, rc = -EPROTO);
345 lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
349 ptlrpc_req_finished(req);
353 static int osc_setattr_interpret(const struct lu_env *env,
354 struct ptlrpc_request *req,
355 struct osc_setattr_args *sa, int rc)
357 struct ost_body *body;
363 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
365 GOTO(out, rc = -EPROTO);
367 lustre_get_wire_obdo(sa->sa_oa, &body->oa);
369 rc = sa->sa_upcall(sa->sa_cookie, rc);
373 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
374 struct obd_trans_info *oti,
375 obd_enqueue_update_f upcall, void *cookie,
376 struct ptlrpc_request_set *rqset)
378 struct ptlrpc_request *req;
379 struct osc_setattr_args *sa;
383 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
387 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
388 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
390 ptlrpc_request_free(req);
394 if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
395 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
397 osc_pack_req_body(req, oinfo);
399 ptlrpc_request_set_replen(req);
401 /* do mds to ost setattr asynchronously */
403 /* Do not wait for response. */
404 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
406 req->rq_interpret_reply =
407 (ptlrpc_interpterer_t)osc_setattr_interpret;
409 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
410 sa = ptlrpc_req_async_args(req);
411 sa->sa_oa = oinfo->oi_oa;
412 sa->sa_upcall = upcall;
413 sa->sa_cookie = cookie;
415 if (rqset == PTLRPCD_SET)
416 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
418 ptlrpc_set_add_req(rqset, req);
424 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
425 struct obd_trans_info *oti,
426 struct ptlrpc_request_set *rqset)
428 return osc_setattr_async_base(exp, oinfo, oti,
429 oinfo->oi_cb_up, oinfo, rqset);
432 int osc_real_create(struct obd_export *exp, struct obdo *oa,
433 struct lov_stripe_md **ea, struct obd_trans_info *oti)
435 struct ptlrpc_request *req;
436 struct ost_body *body;
437 struct lov_stripe_md *lsm;
446 rc = obd_alloc_memmd(exp, &lsm);
451 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
453 GOTO(out, rc = -ENOMEM);
455 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
457 ptlrpc_request_free(req);
461 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
463 lustre_set_wire_obdo(&body->oa, oa);
465 ptlrpc_request_set_replen(req);
467 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
468 oa->o_flags == OBD_FL_DELORPHAN) {
470 "delorphan from OST integration");
471 /* Don't resend the delorphan req */
472 req->rq_no_resend = req->rq_no_delay = 1;
475 rc = ptlrpc_queue_wait(req);
479 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
481 GOTO(out_req, rc = -EPROTO);
483 lustre_get_wire_obdo(oa, &body->oa);
485 /* This should really be sent by the OST */
486 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
487 oa->o_valid |= OBD_MD_FLBLKSZ;
489 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
490 * have valid lsm_oinfo data structs, so don't go touching that.
491 * This needs to be fixed in a big way.
493 lsm->lsm_object_id = oa->o_id;
494 lsm->lsm_object_seq = oa->o_seq;
498 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
500 if (oa->o_valid & OBD_MD_FLCOOKIE) {
501 if (!oti->oti_logcookies)
502 oti_alloc_cookies(oti, 1);
503 *oti->oti_logcookies = oa->o_lcookie;
507 CDEBUG(D_HA, "transno: "LPD64"\n",
508 lustre_msg_get_transno(req->rq_repmsg));
510 ptlrpc_req_finished(req);
513 obd_free_memmd(exp, &lsm);
517 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
518 obd_enqueue_update_f upcall, void *cookie,
519 struct ptlrpc_request_set *rqset)
521 struct ptlrpc_request *req;
522 struct osc_setattr_args *sa;
523 struct ost_body *body;
527 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
531 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
532 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
534 ptlrpc_request_free(req);
537 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
538 ptlrpc_at_set_req_timeout(req);
540 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
542 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
543 osc_pack_capa(req, body, oinfo->oi_capa);
545 ptlrpc_request_set_replen(req);
548 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
549 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
550 sa = ptlrpc_req_async_args(req);
551 sa->sa_oa = oinfo->oi_oa;
552 sa->sa_upcall = upcall;
553 sa->sa_cookie = cookie;
554 if (rqset == PTLRPCD_SET)
555 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
557 ptlrpc_set_add_req(rqset, req);
562 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
563 struct obd_trans_info *oti,
564 struct ptlrpc_request_set *rqset)
566 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
567 oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
568 oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
569 return osc_punch_base(exp, oinfo,
570 oinfo->oi_cb_up, oinfo, rqset);
573 static int osc_sync_interpret(const struct lu_env *env,
574 struct ptlrpc_request *req,
577 struct osc_async_args *aa = arg;
578 struct ost_body *body;
584 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
586 CERROR ("can't unpack ost_body\n");
587 GOTO(out, rc = -EPROTO);
590 *aa->aa_oi->oi_oa = body->oa;
592 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
596 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
597 obd_size start, obd_size end,
598 struct ptlrpc_request_set *set)
600 struct ptlrpc_request *req;
601 struct ost_body *body;
602 struct osc_async_args *aa;
607 CDEBUG(D_INFO, "oa NULL\n");
611 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
615 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
616 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
618 ptlrpc_request_free(req);
622 /* overload the size and blocks fields in the oa with start/end */
623 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
625 lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
626 body->oa.o_size = start;
627 body->oa.o_blocks = end;
628 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
629 osc_pack_capa(req, body, oinfo->oi_capa);
631 ptlrpc_request_set_replen(req);
632 req->rq_interpret_reply = osc_sync_interpret;
634 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
635 aa = ptlrpc_req_async_args(req);
638 ptlrpc_set_add_req(set, req);
642 /* Find and cancel locally locks matched by @mode in the resource found by
643 * @objid. Found locks are added into @cancel list. Returns the amount of
644 * locks added to @cancels list. */
645 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
647 ldlm_mode_t mode, int lock_flags)
649 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
650 struct ldlm_res_id res_id;
651 struct ldlm_resource *res;
655 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
656 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
660 LDLM_RESOURCE_ADDREF(res);
661 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
662 lock_flags, 0, NULL);
663 LDLM_RESOURCE_DELREF(res);
664 ldlm_resource_putref(res);
668 static int osc_destroy_interpret(const struct lu_env *env,
669 struct ptlrpc_request *req, void *data,
672 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
674 cfs_atomic_dec(&cli->cl_destroy_in_flight);
675 cfs_waitq_signal(&cli->cl_destroy_waitq);
679 static int osc_can_send_destroy(struct client_obd *cli)
681 if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
682 cli->cl_max_rpcs_in_flight) {
683 /* The destroy request can be sent */
686 if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
687 cli->cl_max_rpcs_in_flight) {
689 * The counter has been modified between the two atomic
692 cfs_waitq_signal(&cli->cl_destroy_waitq);
697 /* Destroy requests can be async always on the client, and we don't even really
698 * care about the return code since the client cannot do anything at all about
700 * When the MDS is unlinking a filename, it saves the file objects into a
701 * recovery llog, and these object records are cancelled when the OST reports
702 * they were destroyed and sync'd to disk (i.e. transaction committed).
703 * If the client dies, or the OST is down when the object should be destroyed,
704 * the records are not cancelled, and when the OST reconnects to the MDS next,
705 * it will retrieve the llog unlink logs and then sends the log cancellation
706 * cookies to the MDS after committing destroy transactions. */
707 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
708 struct lov_stripe_md *ea, struct obd_trans_info *oti,
709 struct obd_export *md_export, void *capa)
711 struct client_obd *cli = &exp->exp_obd->u.cli;
712 struct ptlrpc_request *req;
713 struct ost_body *body;
714 CFS_LIST_HEAD(cancels);
719 CDEBUG(D_INFO, "oa NULL\n");
723 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
724 LDLM_FL_DISCARD_DATA);
726 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
728 ldlm_lock_list_put(&cancels, l_bl_ast, count);
732 osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
733 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
736 ptlrpc_request_free(req);
740 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
741 ptlrpc_at_set_req_timeout(req);
743 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
744 oa->o_lcookie = *oti->oti_logcookies;
745 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
747 lustre_set_wire_obdo(&body->oa, oa);
749 osc_pack_capa(req, body, (struct obd_capa *)capa);
750 ptlrpc_request_set_replen(req);
752 /* don't throttle destroy RPCs for the MDT */
753 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
754 req->rq_interpret_reply = osc_destroy_interpret;
755 if (!osc_can_send_destroy(cli)) {
756 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
760 * Wait until the number of on-going destroy RPCs drops
761 * under max_rpc_in_flight
763 l_wait_event_exclusive(cli->cl_destroy_waitq,
764 osc_can_send_destroy(cli), &lwi);
768 /* Do not wait for response */
769 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
773 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
776 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
778 LASSERT(!(oa->o_valid & bits));
781 client_obd_list_lock(&cli->cl_loi_list_lock);
782 oa->o_dirty = cli->cl_dirty;
783 if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
784 CERROR("dirty %lu - %lu > dirty_max %lu\n",
785 cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
787 } else if (cfs_atomic_read(&obd_dirty_pages) -
788 cfs_atomic_read(&obd_dirty_transit_pages) >
789 obd_max_dirty_pages + 1){
790 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
791 * not covered by a lock thus they may safely race and trip
792 * this CERROR() unless we add in a small fudge factor (+1). */
793 CERROR("dirty %d - %d > system dirty_max %d\n",
794 cfs_atomic_read(&obd_dirty_pages),
795 cfs_atomic_read(&obd_dirty_transit_pages),
796 obd_max_dirty_pages);
798 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
799 CERROR("dirty %lu - dirty_max %lu too big???\n",
800 cli->cl_dirty, cli->cl_dirty_max);
803 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
804 (cli->cl_max_rpcs_in_flight + 1);
805 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
807 oa->o_grant = cli->cl_avail_grant;
808 oa->o_dropped = cli->cl_lost_grant;
809 cli->cl_lost_grant = 0;
810 client_obd_list_unlock(&cli->cl_loi_list_lock);
811 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
812 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
816 static void osc_update_next_shrink(struct client_obd *cli)
818 cli->cl_next_shrink_grant =
819 cfs_time_shift(cli->cl_grant_shrink_interval);
820 CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
821 cli->cl_next_shrink_grant);
824 /* caller must hold loi_list_lock */
825 static void osc_consume_write_grant(struct client_obd *cli,
826 struct brw_page *pga)
828 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
829 LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
830 cfs_atomic_inc(&obd_dirty_pages);
831 cli->cl_dirty += CFS_PAGE_SIZE;
832 cli->cl_avail_grant -= CFS_PAGE_SIZE;
833 pga->flag |= OBD_BRW_FROM_GRANT;
834 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
835 CFS_PAGE_SIZE, pga, pga->pg);
836 LASSERT(cli->cl_avail_grant >= 0);
837 osc_update_next_shrink(cli);
840 /* the companion to osc_consume_write_grant, called when a brw has completed.
841 * must be called with the loi lock held. */
842 static void osc_release_write_grant(struct client_obd *cli,
843 struct brw_page *pga, int sent)
845 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
848 LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
849 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
854 pga->flag &= ~OBD_BRW_FROM_GRANT;
855 cfs_atomic_dec(&obd_dirty_pages);
856 cli->cl_dirty -= CFS_PAGE_SIZE;
857 if (pga->flag & OBD_BRW_NOCACHE) {
858 pga->flag &= ~OBD_BRW_NOCACHE;
859 cfs_atomic_dec(&obd_dirty_transit_pages);
860 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
863 /* Reclaim grant from truncated pages. This is used to solve
864 * write-truncate and grant all gone(to lost_grant) problem.
865 * For a vfs write this problem can be easily solved by a sync
866 * write, however, this is not an option for page_mkwrite()
867 * because grant has to be allocated before a page becomes
869 if (cli->cl_avail_grant < PTLRPC_MAX_BRW_SIZE)
870 cli->cl_avail_grant += CFS_PAGE_SIZE;
872 cli->cl_lost_grant += CFS_PAGE_SIZE;
873 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
874 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
875 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
876 /* For short writes we shouldn't count parts of pages that
877 * span a whole block on the OST side, or our accounting goes
878 * wrong. Should match the code in filter_grant_check. */
879 int offset = pga->off & ~CFS_PAGE_MASK;
880 int count = pga->count + (offset & (blocksize - 1));
881 int end = (offset + pga->count) & (blocksize - 1);
883 count += blocksize - end;
885 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
886 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
887 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
888 cli->cl_avail_grant, cli->cl_dirty);
894 static unsigned long rpcs_in_flight(struct client_obd *cli)
896 return cli->cl_r_in_flight + cli->cl_w_in_flight;
899 /* caller must hold loi_list_lock */
900 void osc_wake_cache_waiters(struct client_obd *cli)
903 struct osc_cache_waiter *ocw;
906 cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
907 /* if we can't dirty more, we must wait until some is written */
908 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
909 (cfs_atomic_read(&obd_dirty_pages) + 1 >
910 obd_max_dirty_pages)) {
911 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
912 "osc max %ld, sys max %d\n", cli->cl_dirty,
913 cli->cl_dirty_max, obd_max_dirty_pages);
917 /* if still dirty cache but no grant wait for pending RPCs that
918 * may yet return us some grant before doing sync writes */
919 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
920 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
921 cli->cl_w_in_flight);
925 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
926 cfs_list_del_init(&ocw->ocw_entry);
927 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
928 /* no more RPCs in flight to return grant, do sync IO */
929 ocw->ocw_rc = -EDQUOT;
930 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
932 osc_consume_write_grant(cli,
933 &ocw->ocw_oap->oap_brw_page);
936 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld\n",
937 ocw, ocw->ocw_oap, cli->cl_avail_grant);
939 cfs_waitq_signal(&ocw->ocw_waitq);
945 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
947 client_obd_list_lock(&cli->cl_loi_list_lock);
948 cli->cl_avail_grant += grant;
949 client_obd_list_unlock(&cli->cl_loi_list_lock);
952 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
954 if (body->oa.o_valid & OBD_MD_FLGRANT) {
955 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
956 __osc_update_grant(cli, body->oa.o_grant);
960 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
961 void *key, obd_count vallen, void *val,
962 struct ptlrpc_request_set *set);
964 static int osc_shrink_grant_interpret(const struct lu_env *env,
965 struct ptlrpc_request *req,
968 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
969 struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
970 struct ost_body *body;
973 __osc_update_grant(cli, oa->o_grant);
977 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
979 osc_update_grant(cli, body);
985 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
987 client_obd_list_lock(&cli->cl_loi_list_lock);
988 oa->o_grant = cli->cl_avail_grant / 4;
989 cli->cl_avail_grant -= oa->o_grant;
990 client_obd_list_unlock(&cli->cl_loi_list_lock);
991 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
992 oa->o_valid |= OBD_MD_FLFLAGS;
995 oa->o_flags |= OBD_FL_SHRINK_GRANT;
996 osc_update_next_shrink(cli);
999 /* Shrink the current grant, either from some large amount to enough for a
1000 * full set of in-flight RPCs, or if we have already shrunk to that limit
1001 * then to enough for a single RPC. This avoids keeping more grant than
1002 * needed, and avoids shrinking the grant piecemeal. */
1003 static int osc_shrink_grant(struct client_obd *cli)
1005 long target = (cli->cl_max_rpcs_in_flight + 1) *
1006 cli->cl_max_pages_per_rpc;
1008 client_obd_list_lock(&cli->cl_loi_list_lock);
1009 if (cli->cl_avail_grant <= target)
1010 target = cli->cl_max_pages_per_rpc;
1011 client_obd_list_unlock(&cli->cl_loi_list_lock);
1013 return osc_shrink_grant_to_target(cli, target);
1016 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1019 struct ost_body *body;
1022 client_obd_list_lock(&cli->cl_loi_list_lock);
1023 /* Don't shrink if we are already above or below the desired limit
1024 * We don't want to shrink below a single RPC, as that will negatively
1025 * impact block allocation and long-term performance. */
1026 if (target < cli->cl_max_pages_per_rpc)
1027 target = cli->cl_max_pages_per_rpc;
1029 if (target >= cli->cl_avail_grant) {
1030 client_obd_list_unlock(&cli->cl_loi_list_lock);
1033 client_obd_list_unlock(&cli->cl_loi_list_lock);
1035 OBD_ALLOC_PTR(body);
1039 osc_announce_cached(cli, &body->oa, 0);
1041 client_obd_list_lock(&cli->cl_loi_list_lock);
1042 body->oa.o_grant = cli->cl_avail_grant - target;
1043 cli->cl_avail_grant = target;
1044 client_obd_list_unlock(&cli->cl_loi_list_lock);
1045 if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1046 body->oa.o_valid |= OBD_MD_FLFLAGS;
1047 body->oa.o_flags = 0;
1049 body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1050 osc_update_next_shrink(cli);
1052 rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1053 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1054 sizeof(*body), body, NULL);
1056 __osc_update_grant(cli, body->oa.o_grant);
1061 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1062 static int osc_should_shrink_grant(struct client_obd *client)
1064 cfs_time_t time = cfs_time_current();
1065 cfs_time_t next_shrink = client->cl_next_shrink_grant;
1067 if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1068 OBD_CONNECT_GRANT_SHRINK) == 0)
1071 if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1072 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1073 client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1076 osc_update_next_shrink(client);
1081 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1083 struct client_obd *client;
1085 cfs_list_for_each_entry(client, &item->ti_obd_list,
1086 cl_grant_shrink_list) {
1087 if (osc_should_shrink_grant(client))
1088 osc_shrink_grant(client);
1093 static int osc_add_shrink_grant(struct client_obd *client)
1097 rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1099 osc_grant_shrink_grant_cb, NULL,
1100 &client->cl_grant_shrink_list);
1102 CERROR("add grant client %s error %d\n",
1103 client->cl_import->imp_obd->obd_name, rc);
1106 CDEBUG(D_CACHE, "add grant client %s \n",
1107 client->cl_import->imp_obd->obd_name);
1108 osc_update_next_shrink(client);
1112 static int osc_del_shrink_grant(struct client_obd *client)
1114 return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1118 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1121 * ocd_grant is the total grant amount we're expect to hold: if we've
1122 * been evicted, it's the new avail_grant amount, cl_dirty will drop
1123 * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1125 * race is tolerable here: if we're evicted, but imp_state already
1126 * left EVICTED state, then cl_dirty must be 0 already.
1128 client_obd_list_lock(&cli->cl_loi_list_lock);
1129 if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1130 cli->cl_avail_grant = ocd->ocd_grant;
1132 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1134 if (cli->cl_avail_grant < 0) {
1135 CWARN("%s: available grant < 0, the OSS is probably not running"
1136 " with patch from bug20278 (%ld) \n",
1137 cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1138 /* workaround for 1.6 servers which do not have
1139 * the patch from bug20278 */
1140 cli->cl_avail_grant = ocd->ocd_grant;
1143 client_obd_list_unlock(&cli->cl_loi_list_lock);
1145 CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1146 cli->cl_import->imp_obd->obd_name,
1147 cli->cl_avail_grant, cli->cl_lost_grant);
1149 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1150 cfs_list_empty(&cli->cl_grant_shrink_list))
1151 osc_add_shrink_grant(cli);
1154 /* We assume that the reason this OSC got a short read is because it read
1155 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1156 * via the LOV, and it _knows_ it's reading inside the file, it's just that
1157 * this stripe never got written at or beyond this stripe offset yet. */
1158 static void handle_short_read(int nob_read, obd_count page_count,
1159 struct brw_page **pga)
1164 /* skip bytes read OK */
1165 while (nob_read > 0) {
1166 LASSERT (page_count > 0);
1168 if (pga[i]->count > nob_read) {
1169 /* EOF inside this page */
1170 ptr = cfs_kmap(pga[i]->pg) +
1171 (pga[i]->off & ~CFS_PAGE_MASK);
1172 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1173 cfs_kunmap(pga[i]->pg);
1179 nob_read -= pga[i]->count;
1184 /* zero remaining pages */
1185 while (page_count-- > 0) {
1186 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1187 memset(ptr, 0, pga[i]->count);
1188 cfs_kunmap(pga[i]->pg);
1193 static int check_write_rcs(struct ptlrpc_request *req,
1194 int requested_nob, int niocount,
1195 obd_count page_count, struct brw_page **pga)
1200 remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1201 sizeof(*remote_rcs) *
1203 if (remote_rcs == NULL) {
1204 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1208 /* return error if any niobuf was in error */
1209 for (i = 0; i < niocount; i++) {
1210 if ((int)remote_rcs[i] < 0)
1211 return(remote_rcs[i]);
1213 if (remote_rcs[i] != 0) {
1214 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1215 i, remote_rcs[i], req);
1220 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1221 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1222 req->rq_bulk->bd_nob_transferred, requested_nob);
1229 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1231 if (p1->flag != p2->flag) {
1232 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1233 OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1235 /* warn if we try to combine flags that we don't know to be
1236 * safe to combine */
1237 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1238 CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1239 "report this at http://bugs.whamcloud.com/\n",
1240 p1->flag, p2->flag);
1245 return (p1->off + p1->count == p2->off);
1248 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1249 struct brw_page **pga, int opc,
1250 cksum_type_t cksum_type)
1255 LASSERT (pg_count > 0);
1256 cksum = init_checksum(cksum_type);
1257 while (nob > 0 && pg_count > 0) {
1258 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1259 int off = pga[i]->off & ~CFS_PAGE_MASK;
1260 int count = pga[i]->count > nob ? nob : pga[i]->count;
1262 /* corrupt the data before we compute the checksum, to
1263 * simulate an OST->client data error */
1264 if (i == 0 && opc == OST_READ &&
1265 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1266 memcpy(ptr + off, "bad1", min(4, nob));
1267 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1268 cfs_kunmap(pga[i]->pg);
1269 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1272 nob -= pga[i]->count;
1276 /* For sending we only compute the wrong checksum instead
1277 * of corrupting the data so it is still correct on a redo */
1278 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1281 return fini_checksum(cksum, cksum_type);
1284 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1285 struct lov_stripe_md *lsm, obd_count page_count,
1286 struct brw_page **pga,
1287 struct ptlrpc_request **reqp,
1288 struct obd_capa *ocapa, int reserve,
1291 struct ptlrpc_request *req;
1292 struct ptlrpc_bulk_desc *desc;
1293 struct ost_body *body;
1294 struct obd_ioobj *ioobj;
1295 struct niobuf_remote *niobuf;
1296 int niocount, i, requested_nob, opc, rc;
1297 struct osc_brw_async_args *aa;
1298 struct req_capsule *pill;
1299 struct brw_page *pg_prev;
1302 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1303 RETURN(-ENOMEM); /* Recoverable */
1304 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1305 RETURN(-EINVAL); /* Fatal */
1307 if ((cmd & OBD_BRW_WRITE) != 0) {
1309 req = ptlrpc_request_alloc_pool(cli->cl_import,
1310 cli->cl_import->imp_rq_pool,
1311 &RQF_OST_BRW_WRITE);
1314 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1319 for (niocount = i = 1; i < page_count; i++) {
1320 if (!can_merge_pages(pga[i - 1], pga[i]))
1324 pill = &req->rq_pill;
1325 req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1327 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1328 niocount * sizeof(*niobuf));
1329 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1331 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1333 ptlrpc_request_free(req);
1336 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1337 ptlrpc_at_set_req_timeout(req);
1339 if (opc == OST_WRITE)
1340 desc = ptlrpc_prep_bulk_imp(req, page_count,
1341 BULK_GET_SOURCE, OST_BULK_PORTAL);
1343 desc = ptlrpc_prep_bulk_imp(req, page_count,
1344 BULK_PUT_SINK, OST_BULK_PORTAL);
1347 GOTO(out, rc = -ENOMEM);
1348 /* NB request now owns desc and will free it when it gets freed */
1350 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1351 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1352 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1353 LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1355 lustre_set_wire_obdo(&body->oa, oa);
1357 obdo_to_ioobj(oa, ioobj);
1358 ioobj->ioo_bufcnt = niocount;
1359 osc_pack_capa(req, body, ocapa);
1360 LASSERT (page_count > 0);
1362 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1363 struct brw_page *pg = pga[i];
1364 int poff = pg->off & ~CFS_PAGE_MASK;
1366 LASSERT(pg->count > 0);
1367 /* make sure there is no gap in the middle of page array */
1368 LASSERTF(page_count == 1 ||
1369 (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1370 ergo(i > 0 && i < page_count - 1,
1371 poff == 0 && pg->count == CFS_PAGE_SIZE) &&
1372 ergo(i == page_count - 1, poff == 0)),
1373 "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1374 i, page_count, pg, pg->off, pg->count);
1376 LASSERTF(i == 0 || pg->off > pg_prev->off,
1377 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1378 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1380 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1381 pg_prev->pg, page_private(pg_prev->pg),
1382 pg_prev->pg->index, pg_prev->off);
1384 LASSERTF(i == 0 || pg->off > pg_prev->off,
1385 "i %d p_c %u\n", i, page_count);
1387 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1388 (pg->flag & OBD_BRW_SRVLOCK));
1390 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1391 requested_nob += pg->count;
1393 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1395 niobuf->len += pg->count;
1397 niobuf->offset = pg->off;
1398 niobuf->len = pg->count;
1399 niobuf->flags = pg->flag;
1404 LASSERTF((void *)(niobuf - niocount) ==
1405 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1406 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1407 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1409 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1411 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1412 body->oa.o_valid |= OBD_MD_FLFLAGS;
1413 body->oa.o_flags = 0;
1415 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1418 if (osc_should_shrink_grant(cli))
1419 osc_shrink_grant_local(cli, &body->oa);
1421 /* size[REQ_REC_OFF] still sizeof (*body) */
1422 if (opc == OST_WRITE) {
1423 if (cli->cl_checksum &&
1424 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1425 /* store cl_cksum_type in a local variable since
1426 * it can be changed via lprocfs */
1427 cksum_type_t cksum_type = cli->cl_cksum_type;
1429 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1430 oa->o_flags &= OBD_FL_LOCAL_MASK;
1431 body->oa.o_flags = 0;
1433 body->oa.o_flags |= cksum_type_pack(cksum_type);
1434 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1435 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1439 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1441 /* save this in 'oa', too, for later checking */
1442 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1443 oa->o_flags |= cksum_type_pack(cksum_type);
1445 /* clear out the checksum flag, in case this is a
1446 * resend but cl_checksum is no longer set. b=11238 */
1447 oa->o_valid &= ~OBD_MD_FLCKSUM;
1449 oa->o_cksum = body->oa.o_cksum;
1450 /* 1 RC per niobuf */
1451 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1452 sizeof(__u32) * niocount);
1454 if (cli->cl_checksum &&
1455 !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1456 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1457 body->oa.o_flags = 0;
1458 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1459 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1462 ptlrpc_request_set_replen(req);
1464 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1465 aa = ptlrpc_req_async_args(req);
1467 aa->aa_requested_nob = requested_nob;
1468 aa->aa_nio_count = niocount;
1469 aa->aa_page_count = page_count;
1473 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1474 if (ocapa && reserve)
1475 aa->aa_ocapa = capa_get(ocapa);
1481 ptlrpc_req_finished(req);
1485 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1486 __u32 client_cksum, __u32 server_cksum, int nob,
1487 obd_count page_count, struct brw_page **pga,
1488 cksum_type_t client_cksum_type)
1492 cksum_type_t cksum_type;
1494 if (server_cksum == client_cksum) {
1495 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1499 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1501 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1504 if (cksum_type != client_cksum_type)
1505 msg = "the server did not use the checksum type specified in "
1506 "the original request - likely a protocol problem";
1507 else if (new_cksum == server_cksum)
1508 msg = "changed on the client after we checksummed it - "
1509 "likely false positive due to mmap IO (bug 11742)";
1510 else if (new_cksum == client_cksum)
1511 msg = "changed in transit before arrival at OST";
1513 msg = "changed in transit AND doesn't match the original - "
1514 "likely false positive due to mmap IO (bug 11742)";
1516 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1517 " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1518 msg, libcfs_nid2str(peer->nid),
1519 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1520 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1521 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1523 oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1525 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1526 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1527 "client csum now %x\n", client_cksum, client_cksum_type,
1528 server_cksum, cksum_type, new_cksum);
1532 /* Note rc enters this function as number of bytes transferred */
1533 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1535 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1536 const lnet_process_id_t *peer =
1537 &req->rq_import->imp_connection->c_peer;
1538 struct client_obd *cli = aa->aa_cli;
1539 struct ost_body *body;
1540 __u32 client_cksum = 0;
1543 if (rc < 0 && rc != -EDQUOT) {
1544 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1548 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1549 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1551 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1555 /* set/clear over quota flag for a uid/gid */
1556 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1557 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1558 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1560 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1561 body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1563 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1566 osc_update_grant(cli, body);
1571 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1572 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1574 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1576 CERROR("Unexpected +ve rc %d\n", rc);
1579 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1581 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1584 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1585 check_write_checksum(&body->oa, peer, client_cksum,
1586 body->oa.o_cksum, aa->aa_requested_nob,
1587 aa->aa_page_count, aa->aa_ppga,
1588 cksum_type_unpack(aa->aa_oa->o_flags)))
1591 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1592 aa->aa_page_count, aa->aa_ppga);
1596 /* The rest of this function executes only for OST_READs */
1598 /* if unwrap_bulk failed, return -EAGAIN to retry */
1599 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1601 GOTO(out, rc = -EAGAIN);
1603 if (rc > aa->aa_requested_nob) {
1604 CERROR("Unexpected rc %d (%d requested)\n", rc,
1605 aa->aa_requested_nob);
1609 if (rc != req->rq_bulk->bd_nob_transferred) {
1610 CERROR ("Unexpected rc %d (%d transferred)\n",
1611 rc, req->rq_bulk->bd_nob_transferred);
1615 if (rc < aa->aa_requested_nob)
1616 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1618 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1619 static int cksum_counter;
1620 __u32 server_cksum = body->oa.o_cksum;
1623 cksum_type_t cksum_type;
1625 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1626 body->oa.o_flags : 0);
1627 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1628 aa->aa_ppga, OST_READ,
1631 if (peer->nid == req->rq_bulk->bd_sender) {
1635 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1638 if (server_cksum == ~0 && rc > 0) {
1639 CERROR("Protocol error: server %s set the 'checksum' "
1640 "bit, but didn't send a checksum. Not fatal, "
1641 "but please notify on http://bugs.whamcloud.com/\n",
1642 libcfs_nid2str(peer->nid));
1643 } else if (server_cksum != client_cksum) {
1644 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1645 "%s%s%s inode "DFID" object "
1646 LPU64"/"LPU64" extent "
1647 "["LPU64"-"LPU64"]\n",
1648 req->rq_import->imp_obd->obd_name,
1649 libcfs_nid2str(peer->nid),
1651 body->oa.o_valid & OBD_MD_FLFID ?
1652 body->oa.o_parent_seq : (__u64)0,
1653 body->oa.o_valid & OBD_MD_FLFID ?
1654 body->oa.o_parent_oid : 0,
1655 body->oa.o_valid & OBD_MD_FLFID ?
1656 body->oa.o_parent_ver : 0,
1658 body->oa.o_valid & OBD_MD_FLGROUP ?
1659 body->oa.o_seq : (__u64)0,
1660 aa->aa_ppga[0]->off,
1661 aa->aa_ppga[aa->aa_page_count-1]->off +
1662 aa->aa_ppga[aa->aa_page_count-1]->count -
1664 CERROR("client %x, server %x, cksum_type %x\n",
1665 client_cksum, server_cksum, cksum_type);
1667 aa->aa_oa->o_cksum = client_cksum;
1671 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1674 } else if (unlikely(client_cksum)) {
1675 static int cksum_missed;
1678 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1679 CERROR("Checksum %u requested from %s but not sent\n",
1680 cksum_missed, libcfs_nid2str(peer->nid));
1686 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1691 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1692 struct lov_stripe_md *lsm,
1693 obd_count page_count, struct brw_page **pga,
1694 struct obd_capa *ocapa)
1696 struct ptlrpc_request *req;
1700 struct l_wait_info lwi;
1704 cfs_waitq_init(&waitq);
1707 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1708 page_count, pga, &req, ocapa, 0, resends);
1712 rc = ptlrpc_queue_wait(req);
1714 if (rc == -ETIMEDOUT && req->rq_resend) {
1715 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1716 ptlrpc_req_finished(req);
1720 rc = osc_brw_fini_request(req, rc);
1722 ptlrpc_req_finished(req);
1723 if (osc_recoverable_error(rc)) {
1725 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1726 CERROR("too many resend retries, returning error\n");
1730 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1731 l_wait_event(waitq, 0, &lwi);
1739 int osc_brw_redo_request(struct ptlrpc_request *request,
1740 struct osc_brw_async_args *aa)
1742 struct ptlrpc_request *new_req;
1743 struct ptlrpc_request_set *set = request->rq_set;
1744 struct osc_brw_async_args *new_aa;
1745 struct osc_async_page *oap;
1749 if (!client_should_resend(aa->aa_resends, aa->aa_cli)) {
1750 CERROR("too many resent retries, returning error\n");
1754 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1756 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1757 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1758 aa->aa_cli, aa->aa_oa,
1759 NULL /* lsm unused by osc currently */,
1760 aa->aa_page_count, aa->aa_ppga,
1761 &new_req, aa->aa_ocapa, 0, 1);
1765 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1767 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1768 if (oap->oap_request != NULL) {
1769 LASSERTF(request == oap->oap_request,
1770 "request %p != oap_request %p\n",
1771 request, oap->oap_request);
1772 if (oap->oap_interrupted) {
1773 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1774 ptlrpc_req_finished(new_req);
1779 /* New request takes over pga and oaps from old request.
1780 * Note that copying a list_head doesn't work, need to move it... */
1782 new_req->rq_interpret_reply = request->rq_interpret_reply;
1783 new_req->rq_async_args = request->rq_async_args;
1784 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1786 new_aa = ptlrpc_req_async_args(new_req);
1788 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1789 cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1790 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1792 cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1793 if (oap->oap_request) {
1794 ptlrpc_req_finished(oap->oap_request);
1795 oap->oap_request = ptlrpc_request_addref(new_req);
1799 new_aa->aa_ocapa = aa->aa_ocapa;
1800 aa->aa_ocapa = NULL;
1802 /* use ptlrpc_set_add_req is safe because interpret functions work
1803 * in check_set context. only one way exist with access to request
1804 * from different thread got -EINTR - this way protected with
1805 * cl_loi_list_lock */
1806 ptlrpc_set_add_req(set, new_req);
1808 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1810 DEBUG_REQ(D_INFO, new_req, "new request");
1815 * ugh, we want disk allocation on the target to happen in offset order. we'll
1816 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1817 * fine for our small page arrays and doesn't require allocation. its an
1818 * insertion sort that swaps elements that are strides apart, shrinking the
1819 * stride down until its '1' and the array is sorted.
1821 static void sort_brw_pages(struct brw_page **array, int num)
1824 struct brw_page *tmp;
1828 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1833 for (i = stride ; i < num ; i++) {
1836 while (j >= stride && array[j - stride]->off > tmp->off) {
1837 array[j] = array[j - stride];
1842 } while (stride > 1);
1845 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1851 LASSERT (pages > 0);
1852 offset = pg[i]->off & ~CFS_PAGE_MASK;
1856 if (pages == 0) /* that's all */
1859 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1860 return count; /* doesn't end on page boundary */
1863 offset = pg[i]->off & ~CFS_PAGE_MASK;
1864 if (offset != 0) /* doesn't start on page boundary */
1871 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1873 struct brw_page **ppga;
1876 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1880 for (i = 0; i < count; i++)
1885 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1887 LASSERT(ppga != NULL);
1888 OBD_FREE(ppga, sizeof(*ppga) * count);
1891 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1892 obd_count page_count, struct brw_page *pga,
1893 struct obd_trans_info *oti)
1895 struct obdo *saved_oa = NULL;
1896 struct brw_page **ppga, **orig;
1897 struct obd_import *imp = class_exp2cliimp(exp);
1898 struct client_obd *cli;
1899 int rc, page_count_orig;
1902 LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1903 cli = &imp->imp_obd->u.cli;
1905 if (cmd & OBD_BRW_CHECK) {
1906 /* The caller just wants to know if there's a chance that this
1907 * I/O can succeed */
1909 if (imp->imp_invalid)
1914 /* test_brw with a failed create can trip this, maybe others. */
1915 LASSERT(cli->cl_max_pages_per_rpc);
1919 orig = ppga = osc_build_ppga(pga, page_count);
1922 page_count_orig = page_count;
1924 sort_brw_pages(ppga, page_count);
1925 while (page_count) {
1926 obd_count pages_per_brw;
1928 if (page_count > cli->cl_max_pages_per_rpc)
1929 pages_per_brw = cli->cl_max_pages_per_rpc;
1931 pages_per_brw = page_count;
1933 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1935 if (saved_oa != NULL) {
1936 /* restore previously saved oa */
1937 *oinfo->oi_oa = *saved_oa;
1938 } else if (page_count > pages_per_brw) {
1939 /* save a copy of oa (brw will clobber it) */
1940 OBDO_ALLOC(saved_oa);
1941 if (saved_oa == NULL)
1942 GOTO(out, rc = -ENOMEM);
1943 *saved_oa = *oinfo->oi_oa;
1946 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1947 pages_per_brw, ppga, oinfo->oi_capa);
1952 page_count -= pages_per_brw;
1953 ppga += pages_per_brw;
1957 osc_release_ppga(orig, page_count_orig);
1959 if (saved_oa != NULL)
1960 OBDO_FREE(saved_oa);
1965 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1966 * the dirty accounting. Writeback completes or truncate happens before
1967 * writing starts. Must be called with the loi lock held. */
1968 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1971 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1975 /* This maintains the lists of pending pages to read/write for a given object
1976 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1977 * to quickly find objects that are ready to send an RPC. */
1978 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1983 if (lop->lop_num_pending == 0)
1986 /* if we have an invalid import we want to drain the queued pages
1987 * by forcing them through rpcs that immediately fail and complete
1988 * the pages. recovery relies on this to empty the queued pages
1989 * before canceling the locks and evicting down the llite pages */
1990 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1993 /* stream rpcs in queue order as long as as there is an urgent page
1994 * queued. this is our cheap solution for good batching in the case
1995 * where writepage marks some random page in the middle of the file
1996 * as urgent because of, say, memory pressure */
1997 if (!cfs_list_empty(&lop->lop_urgent)) {
1998 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
2002 if (cmd & OBD_BRW_WRITE) {
2003 /* trigger a write rpc stream as long as there are dirtiers
2004 * waiting for space. as they're waiting, they're not going to
2005 * create more pages to coalesce with what's waiting.. */
2006 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2007 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2011 if (lop->lop_num_pending >= cli->cl_max_pages_per_rpc)
2017 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2019 struct osc_async_page *oap;
2022 if (cfs_list_empty(&lop->lop_urgent))
2025 oap = cfs_list_entry(lop->lop_urgent.next,
2026 struct osc_async_page, oap_urgent_item);
2028 if (oap->oap_async_flags & ASYNC_HP) {
2029 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2036 static void on_list(cfs_list_t *item, cfs_list_t *list,
2039 if (cfs_list_empty(item) && should_be_on)
2040 cfs_list_add_tail(item, list);
2041 else if (!cfs_list_empty(item) && !should_be_on)
2042 cfs_list_del_init(item);
2045 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2046 * can find pages to build into rpcs quickly */
2047 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2049 if (lop_makes_hprpc(&loi->loi_write_lop) ||
2050 lop_makes_hprpc(&loi->loi_read_lop)) {
2052 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2053 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2055 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2056 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2057 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2058 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2061 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2062 loi->loi_write_lop.lop_num_pending);
2064 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2065 loi->loi_read_lop.lop_num_pending);
2068 static void lop_update_pending(struct client_obd *cli,
2069 struct loi_oap_pages *lop, int cmd, int delta)
2071 lop->lop_num_pending += delta;
2072 if (cmd & OBD_BRW_WRITE)
2073 cli->cl_pending_w_pages += delta;
2075 cli->cl_pending_r_pages += delta;
2079 * this is called when a sync waiter receives an interruption. Its job is to
2080 * get the caller woken as soon as possible. If its page hasn't been put in an
2081 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
2082 * desiring interruption which will forcefully complete the rpc once the rpc
2085 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2087 struct loi_oap_pages *lop;
2088 struct lov_oinfo *loi;
2092 LASSERT(!oap->oap_interrupted);
2093 oap->oap_interrupted = 1;
2095 /* ok, it's been put in an rpc. only one oap gets a request reference */
2096 if (oap->oap_request != NULL) {
2097 ptlrpc_mark_interrupted(oap->oap_request);
2098 ptlrpcd_wake(oap->oap_request);
2099 ptlrpc_req_finished(oap->oap_request);
2100 oap->oap_request = NULL;
2104 * page completion may be called only if ->cpo_prep() method was
2105 * executed by osc_io_submit(), that also adds page the to pending list
2107 if (!cfs_list_empty(&oap->oap_pending_item)) {
2108 cfs_list_del_init(&oap->oap_pending_item);
2109 cfs_list_del_init(&oap->oap_urgent_item);
2112 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2113 &loi->loi_write_lop : &loi->loi_read_lop;
2114 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2115 loi_list_maint(oap->oap_cli, oap->oap_loi);
2116 rc = oap->oap_caller_ops->ap_completion(env,
2117 oap->oap_caller_data,
2118 oap->oap_cmd, NULL, -EINTR);
2124 /* this is trying to propogate async writeback errors back up to the
2125 * application. As an async write fails we record the error code for later if
2126 * the app does an fsync. As long as errors persist we force future rpcs to be
2127 * sync so that the app can get a sync error and break the cycle of queueing
2128 * pages for which writeback will fail. */
2129 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2136 ar->ar_force_sync = 1;
2137 ar->ar_min_xid = ptlrpc_sample_next_xid();
2142 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2143 ar->ar_force_sync = 0;
2146 void osc_oap_to_pending(struct osc_async_page *oap)
2148 struct loi_oap_pages *lop;
2150 if (oap->oap_cmd & OBD_BRW_WRITE)
2151 lop = &oap->oap_loi->loi_write_lop;
2153 lop = &oap->oap_loi->loi_read_lop;
2155 if (oap->oap_async_flags & ASYNC_HP)
2156 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2157 else if (oap->oap_async_flags & ASYNC_URGENT)
2158 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2159 cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2160 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2163 /* this must be called holding the loi list lock to give coverage to exit_cache,
2164 * async_flag maintenance, and oap_request */
2165 static void osc_ap_completion(const struct lu_env *env,
2166 struct client_obd *cli, struct obdo *oa,
2167 struct osc_async_page *oap, int sent, int rc)
2172 if (oap->oap_request != NULL) {
2173 xid = ptlrpc_req_xid(oap->oap_request);
2174 ptlrpc_req_finished(oap->oap_request);
2175 oap->oap_request = NULL;
2178 cfs_spin_lock(&oap->oap_lock);
2179 oap->oap_async_flags = 0;
2180 cfs_spin_unlock(&oap->oap_lock);
2181 oap->oap_interrupted = 0;
2183 if (oap->oap_cmd & OBD_BRW_WRITE) {
2184 osc_process_ar(&cli->cl_ar, xid, rc);
2185 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2188 if (rc == 0 && oa != NULL) {
2189 if (oa->o_valid & OBD_MD_FLBLOCKS)
2190 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2191 if (oa->o_valid & OBD_MD_FLMTIME)
2192 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2193 if (oa->o_valid & OBD_MD_FLATIME)
2194 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2195 if (oa->o_valid & OBD_MD_FLCTIME)
2196 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2199 rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2200 oap->oap_cmd, oa, rc);
2202 /* cl_page_completion() drops PG_locked. so, a new I/O on the page could
2203 * start, but OSC calls it under lock and thus we can add oap back to
2206 /* upper layer wants to leave the page on pending queue */
2207 osc_oap_to_pending(oap);
2209 osc_exit_cache(cli, oap, sent);
2213 static int brw_queue_work(const struct lu_env *env, void *data)
2215 struct client_obd *cli = data;
2217 CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2219 client_obd_list_lock(&cli->cl_loi_list_lock);
2220 osc_check_rpcs0(env, cli, 1);
2221 client_obd_list_unlock(&cli->cl_loi_list_lock);
2225 static int brw_interpret(const struct lu_env *env,
2226 struct ptlrpc_request *req, void *data, int rc)
2228 struct osc_brw_async_args *aa = data;
2229 struct client_obd *cli;
2233 rc = osc_brw_fini_request(req, rc);
2234 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2235 if (osc_recoverable_error(rc)) {
2236 rc = osc_brw_redo_request(req, aa);
2242 capa_put(aa->aa_ocapa);
2243 aa->aa_ocapa = NULL;
2247 client_obd_list_lock(&cli->cl_loi_list_lock);
2249 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2250 * is called so we know whether to go to sync BRWs or wait for more
2251 * RPCs to complete */
2252 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2253 cli->cl_w_in_flight--;
2255 cli->cl_r_in_flight--;
2257 async = cfs_list_empty(&aa->aa_oaps);
2258 if (!async) { /* from osc_send_oap_rpc() */
2259 struct osc_async_page *oap, *tmp;
2260 /* the caller may re-use the oap after the completion call so
2261 * we need to clean it up a little */
2262 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2264 cfs_list_del_init(&oap->oap_rpc_item);
2265 osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2267 OBDO_FREE(aa->aa_oa);
2268 } else { /* from async_internal() */
2270 for (i = 0; i < aa->aa_page_count; i++)
2271 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2273 osc_wake_cache_waiters(cli);
2274 osc_check_rpcs0(env, cli, 1);
2275 client_obd_list_unlock(&cli->cl_loi_list_lock);
2278 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2279 req->rq_bulk->bd_nob_transferred);
2280 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2281 ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2286 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2287 struct client_obd *cli,
2288 cfs_list_t *rpc_list,
2289 int page_count, int cmd)
2291 struct ptlrpc_request *req;
2292 struct brw_page **pga = NULL;
2293 struct osc_brw_async_args *aa;
2294 struct obdo *oa = NULL;
2295 const struct obd_async_page_ops *ops = NULL;
2296 struct osc_async_page *oap;
2297 struct osc_async_page *tmp;
2298 struct cl_req *clerq = NULL;
2299 enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2300 struct ldlm_lock *lock = NULL;
2301 struct cl_req_attr crattr;
2302 int i, rc, mpflag = 0;
2305 LASSERT(!cfs_list_empty(rpc_list));
2307 if (cmd & OBD_BRW_MEMALLOC)
2308 mpflag = cfs_memory_pressure_get_and_set();
2310 memset(&crattr, 0, sizeof crattr);
2311 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2313 GOTO(out, req = ERR_PTR(-ENOMEM));
2317 GOTO(out, req = ERR_PTR(-ENOMEM));
2320 cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2321 struct cl_page *page = osc_oap2cl_page(oap);
2323 ops = oap->oap_caller_ops;
2325 clerq = cl_req_alloc(env, page, crt,
2326 1 /* only 1-object rpcs for
2329 GOTO(out, req = (void *)clerq);
2330 lock = oap->oap_ldlm_lock;
2332 pga[i] = &oap->oap_brw_page;
2333 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2334 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2335 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2337 cl_req_page_add(env, clerq, page);
2340 /* always get the data for the obdo for the rpc */
2341 LASSERT(ops != NULL);
2343 crattr.cra_capa = NULL;
2344 cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2346 oa->o_handle = lock->l_remote_handle;
2347 oa->o_valid |= OBD_MD_FLHANDLE;
2350 rc = cl_req_prep(env, clerq);
2352 CERROR("cl_req_prep failed: %d\n", rc);
2353 GOTO(out, req = ERR_PTR(rc));
2356 sort_brw_pages(pga, page_count);
2357 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2358 pga, &req, crattr.cra_capa, 1, 0);
2360 CERROR("prep_req failed: %d\n", rc);
2361 GOTO(out, req = ERR_PTR(rc));
2364 if (cmd & OBD_BRW_MEMALLOC)
2365 req->rq_memalloc = 1;
2367 /* Need to update the timestamps after the request is built in case
2368 * we race with setattr (locally or in queue at OST). If OST gets
2369 * later setattr before earlier BRW (as determined by the request xid),
2370 * the OST will not use BRW timestamps. Sadly, there is no obvious
2371 * way to do this in a single call. bug 10150 */
2372 cl_req_attr_set(env, clerq, &crattr,
2373 OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2375 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2376 aa = ptlrpc_req_async_args(req);
2377 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2378 cfs_list_splice(rpc_list, &aa->aa_oaps);
2379 CFS_INIT_LIST_HEAD(rpc_list);
2380 aa->aa_clerq = clerq;
2382 if (cmd & OBD_BRW_MEMALLOC)
2383 cfs_memory_pressure_restore(mpflag);
2385 capa_put(crattr.cra_capa);
2390 OBD_FREE(pga, sizeof(*pga) * page_count);
2391 /* this should happen rarely and is pretty bad, it makes the
2392 * pending list not follow the dirty order */
2393 client_obd_list_lock(&cli->cl_loi_list_lock);
2394 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2395 cfs_list_del_init(&oap->oap_rpc_item);
2397 /* queued sync pages can be torn down while the pages
2398 * were between the pending list and the rpc */
2399 if (oap->oap_interrupted) {
2400 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2401 osc_ap_completion(env, cli, NULL, oap, 0,
2405 osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2407 if (clerq && !IS_ERR(clerq))
2408 cl_req_completion(env, clerq, PTR_ERR(req));
2414 * prepare pages for ASYNC io and put pages in send queue.
2416 * \param cmd OBD_BRW_* macroses
2417 * \param lop pending pages
2419 * \return zero if no page added to send queue.
2420 * \return 1 if pages successfully added to send queue.
2421 * \return negative on errors.
2424 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2425 struct lov_oinfo *loi, int cmd,
2426 struct loi_oap_pages *lop, pdl_policy_t pol)
2428 struct ptlrpc_request *req;
2429 obd_count page_count = 0;
2430 struct osc_async_page *oap = NULL, *tmp;
2431 struct osc_brw_async_args *aa;
2432 const struct obd_async_page_ops *ops;
2433 CFS_LIST_HEAD(rpc_list);
2434 int srvlock = 0, mem_tight = 0;
2435 struct cl_object *clob = NULL;
2436 obd_off starting_offset = OBD_OBJECT_EOF;
2437 unsigned int ending_offset;
2438 int starting_page_off = 0;
2441 /* ASYNC_HP pages first. At present, when the lock the pages is
2442 * to be canceled, the pages covered by the lock will be sent out
2443 * with ASYNC_HP. We have to send out them as soon as possible. */
2444 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2445 if (oap->oap_async_flags & ASYNC_HP)
2446 cfs_list_move(&oap->oap_pending_item, &rpc_list);
2447 else if (!(oap->oap_brw_flags & OBD_BRW_SYNC))
2448 /* only do this for writeback pages. */
2449 cfs_list_move_tail(&oap->oap_pending_item, &rpc_list);
2450 if (++page_count >= cli->cl_max_pages_per_rpc)
2453 cfs_list_splice_init(&rpc_list, &lop->lop_pending);
2456 /* first we find the pages we're allowed to work with */
2457 cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2459 ops = oap->oap_caller_ops;
2461 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2462 "magic 0x%x\n", oap, oap->oap_magic);
2465 /* pin object in memory, so that completion call-backs
2466 * can be safely called under client_obd_list lock. */
2467 clob = osc_oap2cl_page(oap)->cp_obj;
2468 cl_object_get(clob);
2471 if (page_count != 0 &&
2472 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2473 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2474 " oap %p, page %p, srvlock %u\n",
2475 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2479 /* If there is a gap at the start of this page, it can't merge
2480 * with any previous page, so we'll hand the network a
2481 * "fragmented" page array that it can't transfer in 1 RDMA */
2482 if (oap->oap_obj_off < starting_offset) {
2483 if (starting_page_off != 0)
2486 starting_page_off = oap->oap_page_off;
2487 starting_offset = oap->oap_obj_off + starting_page_off;
2488 } else if (oap->oap_page_off != 0)
2491 /* in llite being 'ready' equates to the page being locked
2492 * until completion unlocks it. commit_write submits a page
2493 * as not ready because its unlock will happen unconditionally
2494 * as the call returns. if we race with commit_write giving
2495 * us that page we don't want to create a hole in the page
2496 * stream, so we stop and leave the rpc to be fired by
2497 * another dirtier or kupdated interval (the not ready page
2498 * will still be on the dirty list). we could call in
2499 * at the end of ll_file_write to process the queue again. */
2500 if (!(oap->oap_async_flags & ASYNC_READY)) {
2501 int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2504 CDEBUG(D_INODE, "oap %p page %p returned %d "
2505 "instead of ready\n", oap,
2509 /* llite is telling us that the page is still
2510 * in commit_write and that we should try
2511 * and put it in an rpc again later. we
2512 * break out of the loop so we don't create
2513 * a hole in the sequence of pages in the rpc
2518 /* the io isn't needed.. tell the checks
2519 * below to complete the rpc with EINTR */
2520 cfs_spin_lock(&oap->oap_lock);
2521 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2522 cfs_spin_unlock(&oap->oap_lock);
2523 oap->oap_count = -EINTR;
2526 cfs_spin_lock(&oap->oap_lock);
2527 oap->oap_async_flags |= ASYNC_READY;
2528 cfs_spin_unlock(&oap->oap_lock);
2531 LASSERTF(0, "oap %p page %p returned %d "
2532 "from make_ready\n", oap,
2540 /* take the page out of our book-keeping */
2541 cfs_list_del_init(&oap->oap_pending_item);
2542 lop_update_pending(cli, lop, cmd, -1);
2543 cfs_list_del_init(&oap->oap_urgent_item);
2545 /* ask the caller for the size of the io as the rpc leaves. */
2546 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2548 ops->ap_refresh_count(env, oap->oap_caller_data,
2550 LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2552 if (oap->oap_count <= 0) {
2553 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2555 osc_ap_completion(env, cli, NULL,
2556 oap, 0, oap->oap_count);
2560 /* now put the page back in our accounting */
2561 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2562 if (page_count++ == 0)
2563 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2565 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2568 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2569 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2570 * have the same alignment as the initial writes that allocated
2571 * extents on the server. */
2572 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2574 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2577 if (page_count >= cli->cl_max_pages_per_rpc)
2580 /* If there is a gap at the end of this page, it can't merge
2581 * with any subsequent pages, so we'll hand the network a
2582 * "fragmented" page array that it can't transfer in 1 RDMA */
2583 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2587 loi_list_maint(cli, loi);
2589 client_obd_list_unlock(&cli->cl_loi_list_lock);
2592 cl_object_put(env, clob);
2594 if (page_count == 0) {
2595 client_obd_list_lock(&cli->cl_loi_list_lock);
2599 req = osc_build_req(env, cli, &rpc_list, page_count,
2600 mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2602 LASSERT(cfs_list_empty(&rpc_list));
2603 loi_list_maint(cli, loi);
2604 RETURN(PTR_ERR(req));
2607 aa = ptlrpc_req_async_args(req);
2609 starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2610 if (cmd == OBD_BRW_READ) {
2611 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2612 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2613 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2614 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2616 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2617 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2618 cli->cl_w_in_flight);
2619 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2620 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2623 client_obd_list_lock(&cli->cl_loi_list_lock);
2625 if (cmd == OBD_BRW_READ)
2626 cli->cl_r_in_flight++;
2628 cli->cl_w_in_flight++;
2630 /* queued sync pages can be torn down while the pages
2631 * were between the pending list and the rpc */
2633 cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2634 /* only one oap gets a request reference */
2637 if (oap->oap_interrupted && !req->rq_intr) {
2638 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2640 ptlrpc_mark_interrupted(req);
2644 tmp->oap_request = ptlrpc_request_addref(req);
2646 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2647 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2649 req->rq_interpret_reply = brw_interpret;
2651 /* XXX: Maybe the caller can check the RPC bulk descriptor to see which
2652 * CPU/NUMA node the majority of pages were allocated on, and try
2653 * to assign the async RPC to the CPU core (PDL_POLICY_PREFERRED)
2654 * to reduce cross-CPU memory traffic.
2656 * But on the other hand, we expect that multiple ptlrpcd threads
2657 * and the initial write sponsor can run in parallel, especially
2658 * when data checksum is enabled, which is CPU-bound operation and
2659 * single ptlrpcd thread cannot process in time. So more ptlrpcd
2660 * threads sharing BRW load (with PDL_POLICY_ROUND) seems better.
2662 ptlrpcd_add_req(req, pol, -1);
2666 #define LOI_DEBUG(LOI, STR, args...) \
2667 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2668 !cfs_list_empty(&(LOI)->loi_ready_item) || \
2669 !cfs_list_empty(&(LOI)->loi_hp_ready_item), \
2670 (LOI)->loi_write_lop.lop_num_pending, \
2671 !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2672 (LOI)->loi_read_lop.lop_num_pending, \
2673 !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2676 /* This is called by osc_check_rpcs() to find which objects have pages that
2677 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2678 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2682 /* First return objects that have blocked locks so that they
2683 * will be flushed quickly and other clients can get the lock,
2684 * then objects which have pages ready to be stuffed into RPCs */
2685 if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2686 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2687 struct lov_oinfo, loi_hp_ready_item));
2688 if (!cfs_list_empty(&cli->cl_loi_ready_list))
2689 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2690 struct lov_oinfo, loi_ready_item));
2692 /* then if we have cache waiters, return all objects with queued
2693 * writes. This is especially important when many small files
2694 * have filled up the cache and not been fired into rpcs because
2695 * they don't pass the nr_pending/object threshhold */
2696 if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2697 !cfs_list_empty(&cli->cl_loi_write_list))
2698 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2699 struct lov_oinfo, loi_write_item));
2701 /* then return all queued objects when we have an invalid import
2702 * so that they get flushed */
2703 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2704 if (!cfs_list_empty(&cli->cl_loi_write_list))
2705 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2708 if (!cfs_list_empty(&cli->cl_loi_read_list))
2709 RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2710 struct lov_oinfo, loi_read_item));
2715 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2717 struct osc_async_page *oap;
2720 if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2721 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2722 struct osc_async_page, oap_urgent_item);
2723 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2726 if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2727 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2728 struct osc_async_page, oap_urgent_item);
2729 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2732 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2735 /* called with the loi list lock held */
2736 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli, int ptlrpc)
2738 struct lov_oinfo *loi;
2739 int rc = 0, race_counter = 0;
2743 pol = ptlrpc ? PDL_POLICY_SAME : PDL_POLICY_ROUND;
2745 while ((loi = osc_next_loi(cli)) != NULL) {
2746 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2748 if (osc_max_rpc_in_flight(cli, loi))
2751 /* attempt some read/write balancing by alternating between
2752 * reads and writes in an object. The makes_rpc checks here
2753 * would be redundant if we were getting read/write work items
2754 * instead of objects. we don't want send_oap_rpc to drain a
2755 * partial read pending queue when we're given this object to
2756 * do io on writes while there are cache waiters */
2757 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2758 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2759 &loi->loi_write_lop, pol);
2761 CERROR("Write request failed with %d\n", rc);
2763 /* osc_send_oap_rpc failed, mostly because of
2766 * It can't break here, because if:
2767 * - a page was submitted by osc_io_submit, so
2769 * - no request in flight
2770 * - no subsequent request
2771 * The system will be in live-lock state,
2772 * because there is no chance to call
2773 * osc_io_unplug() and osc_check_rpcs() any
2774 * more. pdflush can't help in this case,
2775 * because it might be blocked at grabbing
2776 * the page lock as we mentioned.
2778 * Anyway, continue to drain pages. */
2787 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2788 rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2789 &loi->loi_read_lop, pol);
2791 CERROR("Read request failed with %d\n", rc);
2799 /* attempt some inter-object balancing by issuing rpcs
2800 * for each object in turn */
2801 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2802 cfs_list_del_init(&loi->loi_hp_ready_item);
2803 if (!cfs_list_empty(&loi->loi_ready_item))
2804 cfs_list_del_init(&loi->loi_ready_item);
2805 if (!cfs_list_empty(&loi->loi_write_item))
2806 cfs_list_del_init(&loi->loi_write_item);
2807 if (!cfs_list_empty(&loi->loi_read_item))
2808 cfs_list_del_init(&loi->loi_read_item);
2810 loi_list_maint(cli, loi);
2812 /* send_oap_rpc fails with 0 when make_ready tells it to
2813 * back off. llite's make_ready does this when it tries
2814 * to lock a page queued for write that is already locked.
2815 * we want to try sending rpcs from many objects, but we
2816 * don't want to spin failing with 0. */
2817 if (race_counter == 10)
2822 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2824 osc_check_rpcs0(env, cli, 0);
2828 * Non-blocking version of osc_enter_cache() that consumes grant only when it
2831 int osc_enter_cache_try(const struct lu_env *env,
2832 struct client_obd *cli, struct lov_oinfo *loi,
2833 struct osc_async_page *oap, int transient)
2837 has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2839 osc_consume_write_grant(cli, &oap->oap_brw_page);
2841 cli->cl_dirty_transit += CFS_PAGE_SIZE;
2842 cfs_atomic_inc(&obd_dirty_transit_pages);
2843 oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2849 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2850 * grant or cache space. */
2851 static int osc_enter_cache(const struct lu_env *env,
2852 struct client_obd *cli, struct lov_oinfo *loi,
2853 struct osc_async_page *oap)
2855 struct osc_cache_waiter ocw;
2856 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2860 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2861 "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2862 cli->cl_dirty_max, obd_max_dirty_pages,
2863 cli->cl_lost_grant, cli->cl_avail_grant);
2865 /* force the caller to try sync io. this can jump the list
2866 * of queued writes and create a discontiguous rpc stream */
2867 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2868 cli->cl_dirty_max < CFS_PAGE_SIZE ||
2869 cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2872 /* Hopefully normal case - cache space and write credits available */
2873 if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2874 cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2875 osc_enter_cache_try(env, cli, loi, oap, 0))
2878 /* We can get here for two reasons: too many dirty pages in cache, or
2879 * run out of grants. In both cases we should write dirty pages out.
2880 * Adding a cache waiter will trigger urgent write-out no matter what
2882 * The exiting condition is no avail grants and no dirty pages caching,
2883 * that really means there is no space on the OST. */
2884 cfs_waitq_init(&ocw.ocw_waitq);
2886 while (cli->cl_dirty > 0) {
2887 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2890 loi_list_maint(cli, loi);
2891 osc_check_rpcs(env, cli);
2892 client_obd_list_unlock(&cli->cl_loi_list_lock);
2894 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
2895 cli->cl_import->imp_obd->obd_name, &ocw, oap);
2897 rc = l_wait_event(ocw.ocw_waitq, cfs_list_empty(&ocw.ocw_entry), &lwi);
2899 client_obd_list_lock(&cli->cl_loi_list_lock);
2900 cfs_list_del_init(&ocw.ocw_entry);
2913 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2914 struct lov_oinfo *loi, cfs_page_t *page,
2915 obd_off offset, const struct obd_async_page_ops *ops,
2916 void *data, void **res, int nocache,
2917 struct lustre_handle *lockh)
2919 struct osc_async_page *oap;
2924 return cfs_size_round(sizeof(*oap));
2927 oap->oap_magic = OAP_MAGIC;
2928 oap->oap_cli = &exp->exp_obd->u.cli;
2931 oap->oap_caller_ops = ops;
2932 oap->oap_caller_data = data;
2934 oap->oap_page = page;
2935 oap->oap_obj_off = offset;
2936 if (!client_is_remote(exp) &&
2937 cfs_capable(CFS_CAP_SYS_RESOURCE))
2938 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2940 LASSERT(!(offset & ~CFS_PAGE_MASK));
2942 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2943 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2944 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2945 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2947 cfs_spin_lock_init(&oap->oap_lock);
2948 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2952 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2953 struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2954 struct osc_async_page *oap, int cmd, int off,
2955 int count, obd_flag brw_flags, enum async_flags async_flags)
2957 struct client_obd *cli = &exp->exp_obd->u.cli;
2961 if (oap->oap_magic != OAP_MAGIC)
2964 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2967 if (!cfs_list_empty(&oap->oap_pending_item) ||
2968 !cfs_list_empty(&oap->oap_urgent_item) ||
2969 !cfs_list_empty(&oap->oap_rpc_item))
2972 /* check if the file's owner/group is over quota */
2973 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2974 struct cl_object *obj;
2975 struct cl_attr attr; /* XXX put attr into thread info */
2976 unsigned int qid[MAXQUOTAS];
2978 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2980 cl_object_attr_lock(obj);
2981 rc = cl_object_attr_get(env, obj, &attr);
2982 cl_object_attr_unlock(obj);
2984 qid[USRQUOTA] = attr.cat_uid;
2985 qid[GRPQUOTA] = attr.cat_gid;
2987 osc_quota_chkdq(cli, qid) == NO_QUOTA)
2994 loi = lsm->lsm_oinfo[0];
2996 client_obd_list_lock(&cli->cl_loi_list_lock);
2998 LASSERT(off + count <= CFS_PAGE_SIZE);
3000 oap->oap_page_off = off;
3001 oap->oap_count = count;
3002 oap->oap_brw_flags = brw_flags;
3003 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3004 if (cfs_memory_pressure_get())
3005 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3006 cfs_spin_lock(&oap->oap_lock);
3007 oap->oap_async_flags = async_flags;
3008 cfs_spin_unlock(&oap->oap_lock);
3010 if (cmd & OBD_BRW_WRITE) {
3011 rc = osc_enter_cache(env, cli, loi, oap);
3013 client_obd_list_unlock(&cli->cl_loi_list_lock);
3018 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3021 osc_oap_to_pending(oap);
3022 loi_list_maint(cli, loi);
3023 if (!osc_max_rpc_in_flight(cli, loi) &&
3024 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
3025 LASSERT(cli->cl_writeback_work != NULL);
3026 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
3028 CDEBUG(D_CACHE, "Queued writeback work for client obd %p/%d.\n",
3031 client_obd_list_unlock(&cli->cl_loi_list_lock);
3036 /* aka (~was & now & flag), but this is more clear :) */
3037 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3039 int osc_set_async_flags_base(struct client_obd *cli,
3040 struct lov_oinfo *loi, struct osc_async_page *oap,
3041 obd_flag async_flags)
3043 struct loi_oap_pages *lop;
3047 LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3049 if (oap->oap_cmd & OBD_BRW_WRITE) {
3050 lop = &loi->loi_write_lop;
3052 lop = &loi->loi_read_lop;
3055 if ((oap->oap_async_flags & async_flags) == async_flags)
3058 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3059 flags |= ASYNC_READY;
3061 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3062 cfs_list_empty(&oap->oap_rpc_item)) {
3063 if (oap->oap_async_flags & ASYNC_HP)
3064 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3066 cfs_list_add_tail(&oap->oap_urgent_item,
3068 flags |= ASYNC_URGENT;
3069 loi_list_maint(cli, loi);
3071 cfs_spin_lock(&oap->oap_lock);
3072 oap->oap_async_flags |= flags;
3073 cfs_spin_unlock(&oap->oap_lock);
3075 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3076 oap->oap_async_flags);
3080 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3081 struct lov_oinfo *loi, struct osc_async_page *oap)
3083 struct client_obd *cli = &exp->exp_obd->u.cli;
3084 struct loi_oap_pages *lop;
3088 if (oap->oap_magic != OAP_MAGIC)
3092 loi = lsm->lsm_oinfo[0];
3094 if (oap->oap_cmd & OBD_BRW_WRITE) {
3095 lop = &loi->loi_write_lop;
3097 lop = &loi->loi_read_lop;
3100 client_obd_list_lock(&cli->cl_loi_list_lock);
3102 if (!cfs_list_empty(&oap->oap_rpc_item))
3103 GOTO(out, rc = -EBUSY);
3105 osc_exit_cache(cli, oap, 0);
3106 osc_wake_cache_waiters(cli);
3108 if (!cfs_list_empty(&oap->oap_urgent_item)) {
3109 cfs_list_del_init(&oap->oap_urgent_item);
3110 cfs_spin_lock(&oap->oap_lock);
3111 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3112 cfs_spin_unlock(&oap->oap_lock);
3114 if (!cfs_list_empty(&oap->oap_pending_item)) {
3115 cfs_list_del_init(&oap->oap_pending_item);
3116 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3118 loi_list_maint(cli, loi);
3119 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3121 client_obd_list_unlock(&cli->cl_loi_list_lock);
3125 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3126 struct ldlm_enqueue_info *einfo)
3128 void *data = einfo->ei_cbdata;
3131 LASSERT(lock != NULL);
3132 LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3133 LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3134 LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3135 LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3137 lock_res_and_lock(lock);
3138 cfs_spin_lock(&osc_ast_guard);
3140 if (lock->l_ast_data == NULL)
3141 lock->l_ast_data = data;
3142 if (lock->l_ast_data == data)
3145 cfs_spin_unlock(&osc_ast_guard);
3146 unlock_res_and_lock(lock);
3151 static int osc_set_data_with_check(struct lustre_handle *lockh,
3152 struct ldlm_enqueue_info *einfo)
3154 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3158 set = osc_set_lock_data_with_check(lock, einfo);
3159 LDLM_LOCK_PUT(lock);
3161 CERROR("lockh %p, data %p - client evicted?\n",
3162 lockh, einfo->ei_cbdata);
3166 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3167 ldlm_iterator_t replace, void *data)
3169 struct ldlm_res_id res_id;
3170 struct obd_device *obd = class_exp2obd(exp);
3172 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3173 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3177 /* find any ldlm lock of the inode in osc
3181 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3182 ldlm_iterator_t replace, void *data)
3184 struct ldlm_res_id res_id;
3185 struct obd_device *obd = class_exp2obd(exp);
3188 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3189 rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3190 if (rc == LDLM_ITER_STOP)
3192 if (rc == LDLM_ITER_CONTINUE)
3197 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3198 obd_enqueue_update_f upcall, void *cookie,
3199 int *flags, int agl, int rc)
3201 int intent = *flags & LDLM_FL_HAS_INTENT;
3205 /* The request was created before ldlm_cli_enqueue call. */
3206 if (rc == ELDLM_LOCK_ABORTED) {
3207 struct ldlm_reply *rep;
3208 rep = req_capsule_server_get(&req->rq_pill,
3211 LASSERT(rep != NULL);
3212 if (rep->lock_policy_res1)
3213 rc = rep->lock_policy_res1;
3217 if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
3219 *flags |= LDLM_FL_LVB_READY;
3220 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3221 lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3224 /* Call the update callback. */
3225 rc = (*upcall)(cookie, rc);
3229 static int osc_enqueue_interpret(const struct lu_env *env,
3230 struct ptlrpc_request *req,
3231 struct osc_enqueue_args *aa, int rc)
3233 struct ldlm_lock *lock;
3234 struct lustre_handle handle;
3236 struct ost_lvb *lvb;
3238 int *flags = aa->oa_flags;
3240 /* Make a local copy of a lock handle and a mode, because aa->oa_*
3241 * might be freed anytime after lock upcall has been called. */
3242 lustre_handle_copy(&handle, aa->oa_lockh);
3243 mode = aa->oa_ei->ei_mode;
3245 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3247 lock = ldlm_handle2lock(&handle);
3249 /* Take an additional reference so that a blocking AST that
3250 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3251 * to arrive after an upcall has been executed by
3252 * osc_enqueue_fini(). */
3253 ldlm_lock_addref(&handle, mode);
3255 /* Let CP AST to grant the lock first. */
3256 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3258 if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
3263 lvb_len = sizeof(*aa->oa_lvb);
3266 /* Complete obtaining the lock procedure. */
3267 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3268 mode, flags, lvb, lvb_len, &handle, rc);
3269 /* Complete osc stuff. */
3270 rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
3271 flags, aa->oa_agl, rc);
3273 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3275 /* Release the lock for async request. */
3276 if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3278 * Releases a reference taken by ldlm_cli_enqueue(), if it is
3279 * not already released by
3280 * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3282 ldlm_lock_decref(&handle, mode);
3284 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3285 aa->oa_lockh, req, aa);
3286 ldlm_lock_decref(&handle, mode);
3287 LDLM_LOCK_PUT(lock);
3291 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3292 struct lov_oinfo *loi, int flags,
3293 struct ost_lvb *lvb, __u32 mode, int rc)
3295 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3297 if (rc == ELDLM_OK) {
3300 LASSERT(lock != NULL);
3301 loi->loi_lvb = *lvb;
3302 tmp = loi->loi_lvb.lvb_size;
3303 /* Extend KMS up to the end of this lock and no further
3304 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3305 if (tmp > lock->l_policy_data.l_extent.end)
3306 tmp = lock->l_policy_data.l_extent.end + 1;
3307 if (tmp >= loi->loi_kms) {
3308 LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3309 ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3310 loi_kms_set(loi, tmp);
3312 LDLM_DEBUG(lock, "lock acquired, setting rss="
3313 LPU64"; leaving kms="LPU64", end="LPU64,
3314 loi->loi_lvb.lvb_size, loi->loi_kms,
3315 lock->l_policy_data.l_extent.end);
3317 ldlm_lock_allow_match(lock);
3318 } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3319 LASSERT(lock != NULL);
3320 loi->loi_lvb = *lvb;
3321 ldlm_lock_allow_match(lock);
3322 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3323 " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3329 ldlm_lock_fail_match(lock);
3331 LDLM_LOCK_PUT(lock);
3334 EXPORT_SYMBOL(osc_update_enqueue);
3336 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3338 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3339 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3340 * other synchronous requests, however keeping some locks and trying to obtain
3341 * others may take a considerable amount of time in a case of ost failure; and
3342 * when other sync requests do not get released lock from a client, the client
3343 * is excluded from the cluster -- such scenarious make the life difficult, so
3344 * release locks just after they are obtained. */
3345 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3346 int *flags, ldlm_policy_data_t *policy,
3347 struct ost_lvb *lvb, int kms_valid,
3348 obd_enqueue_update_f upcall, void *cookie,
3349 struct ldlm_enqueue_info *einfo,
3350 struct lustre_handle *lockh,
3351 struct ptlrpc_request_set *rqset, int async, int agl)
3353 struct obd_device *obd = exp->exp_obd;
3354 struct ptlrpc_request *req = NULL;
3355 int intent = *flags & LDLM_FL_HAS_INTENT;
3356 int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
3361 /* Filesystem lock extents are extended to page boundaries so that
3362 * dealing with the page cache is a little smoother. */
3363 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3364 policy->l_extent.end |= ~CFS_PAGE_MASK;
3367 * kms is not valid when either object is completely fresh (so that no
3368 * locks are cached), or object was evicted. In the latter case cached
3369 * lock cannot be used, because it would prime inode state with
3370 * potentially stale LVB.
3375 /* Next, search for already existing extent locks that will cover us */
3376 /* If we're trying to read, we also search for an existing PW lock. The
3377 * VFS and page cache already protect us locally, so lots of readers/
3378 * writers can share a single PW lock.
3380 * There are problems with conversion deadlocks, so instead of
3381 * converting a read lock to a write lock, we'll just enqueue a new
3384 * At some point we should cancel the read lock instead of making them
3385 * send us a blocking callback, but there are problems with canceling
3386 * locks out from other users right now, too. */
3387 mode = einfo->ei_mode;
3388 if (einfo->ei_mode == LCK_PR)
3390 mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
3391 einfo->ei_type, policy, mode, lockh, 0);
3393 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3395 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
3396 /* For AGL, if enqueue RPC is sent but the lock is not
3397 * granted, then skip to process this strpe.
3398 * Return -ECANCELED to tell the caller. */
3399 ldlm_lock_decref(lockh, mode);
3400 LDLM_LOCK_PUT(matched);
3402 } else if (osc_set_lock_data_with_check(matched, einfo)) {
3403 *flags |= LDLM_FL_LVB_READY;
3404 /* addref the lock only if not async requests and PW
3405 * lock is matched whereas we asked for PR. */
3406 if (!rqset && einfo->ei_mode != mode)
3407 ldlm_lock_addref(lockh, LCK_PR);
3409 /* I would like to be able to ASSERT here that
3410 * rss <= kms, but I can't, for reasons which
3411 * are explained in lov_enqueue() */
3414 /* We already have a lock, and it's referenced */
3415 (*upcall)(cookie, ELDLM_OK);
3417 if (einfo->ei_mode != mode)
3418 ldlm_lock_decref(lockh, LCK_PW);
3420 /* For async requests, decref the lock. */
3421 ldlm_lock_decref(lockh, einfo->ei_mode);
3422 LDLM_LOCK_PUT(matched);
3425 ldlm_lock_decref(lockh, mode);
3426 LDLM_LOCK_PUT(matched);
3432 CFS_LIST_HEAD(cancels);
3433 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3434 &RQF_LDLM_ENQUEUE_LVB);
3438 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3440 ptlrpc_request_free(req);
3444 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3446 ptlrpc_request_set_replen(req);
3449 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3450 *flags &= ~LDLM_FL_BLOCK_GRANTED;
3452 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3453 sizeof(*lvb), lockh, async);
3456 struct osc_enqueue_args *aa;
3457 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3458 aa = ptlrpc_req_async_args(req);
3461 aa->oa_flags = flags;
3462 aa->oa_upcall = upcall;
3463 aa->oa_cookie = cookie;
3465 aa->oa_lockh = lockh;
3468 req->rq_interpret_reply =
3469 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3470 if (rqset == PTLRPCD_SET)
3471 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3473 ptlrpc_set_add_req(rqset, req);
3474 } else if (intent) {
3475 ptlrpc_req_finished(req);
3480 rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
3482 ptlrpc_req_finished(req);
3487 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3488 struct ldlm_enqueue_info *einfo,
3489 struct ptlrpc_request_set *rqset)
3491 struct ldlm_res_id res_id;
3495 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3496 oinfo->oi_md->lsm_object_seq, &res_id);
3498 rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3499 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3500 oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3501 oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3502 rqset, rqset != NULL, 0);
3506 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3507 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3508 int *flags, void *data, struct lustre_handle *lockh,
3511 struct obd_device *obd = exp->exp_obd;
3512 int lflags = *flags;
3516 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3519 /* Filesystem lock extents are extended to page boundaries so that
3520 * dealing with the page cache is a little smoother */
3521 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3522 policy->l_extent.end |= ~CFS_PAGE_MASK;
3524 /* Next, search for already existing extent locks that will cover us */
3525 /* If we're trying to read, we also search for an existing PW lock. The
3526 * VFS and page cache already protect us locally, so lots of readers/
3527 * writers can share a single PW lock. */
3531 rc = ldlm_lock_match(obd->obd_namespace, lflags,
3532 res_id, type, policy, rc, lockh, unref);
3535 if (!osc_set_data_with_check(lockh, data)) {
3536 if (!(lflags & LDLM_FL_TEST_LOCK))
3537 ldlm_lock_decref(lockh, rc);
3541 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3542 ldlm_lock_addref(lockh, LCK_PR);
3543 ldlm_lock_decref(lockh, LCK_PW);
3550 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3554 if (unlikely(mode == LCK_GROUP))
3555 ldlm_lock_decref_and_cancel(lockh, mode);
3557 ldlm_lock_decref(lockh, mode);
3562 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3563 __u32 mode, struct lustre_handle *lockh)
3566 RETURN(osc_cancel_base(lockh, mode));
3569 static int osc_cancel_unused(struct obd_export *exp,
3570 struct lov_stripe_md *lsm,
3571 ldlm_cancel_flags_t flags,
3574 struct obd_device *obd = class_exp2obd(exp);
3575 struct ldlm_res_id res_id, *resp = NULL;
3578 resp = osc_build_res_name(lsm->lsm_object_id,
3579 lsm->lsm_object_seq, &res_id);
3582 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3585 static int osc_statfs_interpret(const struct lu_env *env,
3586 struct ptlrpc_request *req,
3587 struct osc_async_args *aa, int rc)
3589 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3590 struct obd_statfs *msfs;
3595 /* The request has in fact never been sent
3596 * due to issues at a higher level (LOV).
3597 * Exit immediately since the caller is
3598 * aware of the problem and takes care
3599 * of the clean up */
3602 if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3603 (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3609 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3611 GOTO(out, rc = -EPROTO);
3614 /* Reinitialize the RDONLY and DEGRADED flags at the client
3615 * on each statfs, so they don't stay set permanently. */
3616 cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3618 if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3619 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3620 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3621 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3623 if (unlikely(msfs->os_state & OS_STATE_READONLY))
3624 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3625 else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3626 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3628 /* Add a bit of hysteresis so this flag isn't continually flapping,
3629 * and ensure that new files don't get extremely fragmented due to
3630 * only a small amount of available space in the filesystem.
3631 * We want to set the NOSPC flag when there is less than ~0.1% free
3632 * and clear it when there is at least ~0.2% free space, so:
3633 * avail < ~0.1% max max = avail + used
3634 * 1025 * avail < avail + used used = blocks - free
3635 * 1024 * avail < used
3636 * 1024 * avail < blocks - free
3637 * avail < ((blocks - free) >> 10)
3639 * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3640 * lose that amount of space so in those cases we report no space left
3641 * if their is less than 1 GB left. */
3642 used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3643 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3644 ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3645 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3646 else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3647 (msfs->os_ffree > 64) &&
3648 (msfs->os_bavail > (used << 1)))) {
3649 cli->cl_oscc.oscc_flags &= ~(OSCC_FLAG_NOSPC |
3650 OSCC_FLAG_NOSPC_BLK);
3653 if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3654 (msfs->os_bavail < used)))
3655 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC_BLK;
3657 cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3659 *aa->aa_oi->oi_osfs = *msfs;
3661 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3665 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3666 __u64 max_age, struct ptlrpc_request_set *rqset)
3668 struct ptlrpc_request *req;
3669 struct osc_async_args *aa;
3673 /* We could possibly pass max_age in the request (as an absolute
3674 * timestamp or a "seconds.usec ago") so the target can avoid doing
3675 * extra calls into the filesystem if that isn't necessary (e.g.
3676 * during mount that would help a bit). Having relative timestamps
3677 * is not so great if request processing is slow, while absolute
3678 * timestamps are not ideal because they need time synchronization. */
3679 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3683 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3685 ptlrpc_request_free(req);
3688 ptlrpc_request_set_replen(req);
3689 req->rq_request_portal = OST_CREATE_PORTAL;
3690 ptlrpc_at_set_req_timeout(req);
3692 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3693 /* procfs requests not want stat in wait for avoid deadlock */
3694 req->rq_no_resend = 1;
3695 req->rq_no_delay = 1;
3698 req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3699 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3700 aa = ptlrpc_req_async_args(req);
3703 ptlrpc_set_add_req(rqset, req);
3707 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3708 __u64 max_age, __u32 flags)
3710 struct obd_statfs *msfs;
3711 struct ptlrpc_request *req;
3712 struct obd_import *imp = NULL;
3716 /*Since the request might also come from lprocfs, so we need
3717 *sync this with client_disconnect_export Bug15684*/
3718 cfs_down_read(&obd->u.cli.cl_sem);
3719 if (obd->u.cli.cl_import)
3720 imp = class_import_get(obd->u.cli.cl_import);
3721 cfs_up_read(&obd->u.cli.cl_sem);
3725 /* We could possibly pass max_age in the request (as an absolute
3726 * timestamp or a "seconds.usec ago") so the target can avoid doing
3727 * extra calls into the filesystem if that isn't necessary (e.g.
3728 * during mount that would help a bit). Having relative timestamps
3729 * is not so great if request processing is slow, while absolute
3730 * timestamps are not ideal because they need time synchronization. */
3731 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3733 class_import_put(imp);
3738 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3740 ptlrpc_request_free(req);
3743 ptlrpc_request_set_replen(req);
3744 req->rq_request_portal = OST_CREATE_PORTAL;
3745 ptlrpc_at_set_req_timeout(req);
3747 if (flags & OBD_STATFS_NODELAY) {
3748 /* procfs requests not want stat in wait for avoid deadlock */
3749 req->rq_no_resend = 1;
3750 req->rq_no_delay = 1;
3753 rc = ptlrpc_queue_wait(req);
3757 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3759 GOTO(out, rc = -EPROTO);
3766 ptlrpc_req_finished(req);
3770 /* Retrieve object striping information.
3772 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3773 * the maximum number of OST indices which will fit in the user buffer.
3774 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3776 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3778 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3779 struct lov_user_md_v3 lum, *lumk;
3780 struct lov_user_ost_data_v1 *lmm_objects;
3781 int rc = 0, lum_size;
3787 /* we only need the header part from user space to get lmm_magic and
3788 * lmm_stripe_count, (the header part is common to v1 and v3) */
3789 lum_size = sizeof(struct lov_user_md_v1);
3790 if (cfs_copy_from_user(&lum, lump, lum_size))
3793 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3794 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3797 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3798 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3799 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3800 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3802 /* we can use lov_mds_md_size() to compute lum_size
3803 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3804 if (lum.lmm_stripe_count > 0) {
3805 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3806 OBD_ALLOC(lumk, lum_size);
3810 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3811 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3813 lmm_objects = &(lumk->lmm_objects[0]);
3814 lmm_objects->l_object_id = lsm->lsm_object_id;
3816 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3820 lumk->lmm_object_id = lsm->lsm_object_id;
3821 lumk->lmm_object_seq = lsm->lsm_object_seq;
3822 lumk->lmm_stripe_count = 1;
3824 if (cfs_copy_to_user(lump, lumk, lum_size))
3828 OBD_FREE(lumk, lum_size);
3834 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3835 void *karg, void *uarg)
3837 struct obd_device *obd = exp->exp_obd;
3838 struct obd_ioctl_data *data = karg;
3842 if (!cfs_try_module_get(THIS_MODULE)) {
3843 CERROR("Can't get module. Is it alive?");
3847 case OBD_IOC_LOV_GET_CONFIG: {
3849 struct lov_desc *desc;
3850 struct obd_uuid uuid;
3854 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3855 GOTO(out, err = -EINVAL);
3857 data = (struct obd_ioctl_data *)buf;
3859 if (sizeof(*desc) > data->ioc_inllen1) {
3860 obd_ioctl_freedata(buf, len);
3861 GOTO(out, err = -EINVAL);
3864 if (data->ioc_inllen2 < sizeof(uuid)) {
3865 obd_ioctl_freedata(buf, len);
3866 GOTO(out, err = -EINVAL);
3869 desc = (struct lov_desc *)data->ioc_inlbuf1;
3870 desc->ld_tgt_count = 1;
3871 desc->ld_active_tgt_count = 1;
3872 desc->ld_default_stripe_count = 1;
3873 desc->ld_default_stripe_size = 0;
3874 desc->ld_default_stripe_offset = 0;
3875 desc->ld_pattern = 0;
3876 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3878 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3880 err = cfs_copy_to_user((void *)uarg, buf, len);
3883 obd_ioctl_freedata(buf, len);
3886 case LL_IOC_LOV_SETSTRIPE:
3887 err = obd_alloc_memmd(exp, karg);
3891 case LL_IOC_LOV_GETSTRIPE:
3892 err = osc_getstripe(karg, uarg);
3894 case OBD_IOC_CLIENT_RECOVER:
3895 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3896 data->ioc_inlbuf1, 0);
3900 case IOC_OSC_SET_ACTIVE:
3901 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3904 case OBD_IOC_POLL_QUOTACHECK:
3905 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
3907 case OBD_IOC_PING_TARGET:
3908 err = ptlrpc_obd_ping(obd);
3911 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3912 cmd, cfs_curproc_comm());
3913 GOTO(out, err = -ENOTTY);
3916 cfs_module_put(THIS_MODULE);
3920 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3921 void *key, __u32 *vallen, void *val,
3922 struct lov_stripe_md *lsm)
3925 if (!vallen || !val)
3928 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3929 __u32 *stripe = val;
3930 *vallen = sizeof(*stripe);
3933 } else if (KEY_IS(KEY_LAST_ID)) {
3934 struct ptlrpc_request *req;
3939 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3940 &RQF_OST_GET_INFO_LAST_ID);
3944 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3945 RCL_CLIENT, keylen);
3946 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3948 ptlrpc_request_free(req);
3952 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3953 memcpy(tmp, key, keylen);
3955 req->rq_no_delay = req->rq_no_resend = 1;
3956 ptlrpc_request_set_replen(req);
3957 rc = ptlrpc_queue_wait(req);
3961 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3963 GOTO(out, rc = -EPROTO);
3965 *((obd_id *)val) = *reply;
3967 ptlrpc_req_finished(req);
3969 } else if (KEY_IS(KEY_FIEMAP)) {
3970 struct ptlrpc_request *req;
3971 struct ll_user_fiemap *reply;
3975 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3976 &RQF_OST_GET_INFO_FIEMAP);
3980 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3981 RCL_CLIENT, keylen);
3982 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3983 RCL_CLIENT, *vallen);
3984 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3985 RCL_SERVER, *vallen);
3987 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3989 ptlrpc_request_free(req);
3993 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3994 memcpy(tmp, key, keylen);
3995 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3996 memcpy(tmp, val, *vallen);
3998 ptlrpc_request_set_replen(req);
3999 rc = ptlrpc_queue_wait(req);
4003 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
4005 GOTO(out1, rc = -EPROTO);
4007 memcpy(val, reply, *vallen);
4009 ptlrpc_req_finished(req);
4017 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
4019 struct llog_ctxt *ctxt;
4023 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
4025 rc = llog_initiator_connect(ctxt);
4026 llog_ctxt_put(ctxt);
4028 /* XXX return an error? skip setting below flags? */
4031 cfs_spin_lock(&imp->imp_lock);
4032 imp->imp_server_timeout = 1;
4033 imp->imp_pingable = 1;
4034 cfs_spin_unlock(&imp->imp_lock);
4035 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
4040 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
4041 struct ptlrpc_request *req,
4048 RETURN(osc_setinfo_mds_connect_import(req->rq_import));
4051 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
4052 void *key, obd_count vallen, void *val,
4053 struct ptlrpc_request_set *set)
4055 struct ptlrpc_request *req;
4056 struct obd_device *obd = exp->exp_obd;
4057 struct obd_import *imp = class_exp2cliimp(exp);
4062 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
4064 if (KEY_IS(KEY_NEXT_ID)) {
4066 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4068 if (vallen != sizeof(obd_id))
4073 if (vallen != sizeof(obd_id))
4076 /* avoid race between allocate new object and set next id
4077 * from ll_sync thread */
4078 cfs_spin_lock(&oscc->oscc_lock);
4079 new_val = *((obd_id*)val) + 1;
4080 if (new_val > oscc->oscc_next_id)
4081 oscc->oscc_next_id = new_val;
4082 cfs_spin_unlock(&oscc->oscc_lock);
4083 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
4084 exp->exp_obd->obd_name,
4085 obd->u.cli.cl_oscc.oscc_next_id);
4090 if (KEY_IS(KEY_CHECKSUM)) {
4091 if (vallen != sizeof(int))
4093 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
4097 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4098 sptlrpc_conf_client_adapt(obd);
4102 if (KEY_IS(KEY_FLUSH_CTX)) {
4103 sptlrpc_import_flush_my_ctx(imp);
4107 if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4110 /* We pass all other commands directly to OST. Since nobody calls osc
4111 methods directly and everybody is supposed to go through LOV, we
4112 assume lov checked invalid values for us.
4113 The only recognised values so far are evict_by_nid and mds_conn.
4114 Even if something bad goes through, we'd get a -EINVAL from OST
4117 if (KEY_IS(KEY_GRANT_SHRINK))
4118 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4120 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4125 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4126 RCL_CLIENT, keylen);
4127 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4128 RCL_CLIENT, vallen);
4129 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4131 ptlrpc_request_free(req);
4135 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4136 memcpy(tmp, key, keylen);
4137 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4138 memcpy(tmp, val, vallen);
4140 if (KEY_IS(KEY_MDS_CONN)) {
4141 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4143 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4144 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4145 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4146 req->rq_no_delay = req->rq_no_resend = 1;
4147 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4148 } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4149 struct osc_grant_args *aa;
4152 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4153 aa = ptlrpc_req_async_args(req);
4156 ptlrpc_req_finished(req);
4159 *oa = ((struct ost_body *)val)->oa;
4161 req->rq_interpret_reply = osc_shrink_grant_interpret;
4164 ptlrpc_request_set_replen(req);
4165 if (!KEY_IS(KEY_GRANT_SHRINK)) {
4166 LASSERT(set != NULL);
4167 ptlrpc_set_add_req(set, req);
4168 ptlrpc_check_set(NULL, set);
4170 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
4176 static struct llog_operations osc_size_repl_logops = {
4177 lop_cancel: llog_obd_repl_cancel
4180 static struct llog_operations osc_mds_ost_orig_logops;
4182 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4183 struct obd_device *tgt, struct llog_catid *catid)
4188 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4189 &catid->lci_logid, &osc_mds_ost_orig_logops);
4191 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4195 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4196 NULL, &osc_size_repl_logops);
4198 struct llog_ctxt *ctxt =
4199 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4202 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4207 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4208 obd->obd_name, tgt->obd_name, catid, rc);
4209 CERROR("logid "LPX64":0x%x\n",
4210 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4215 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4216 struct obd_device *disk_obd, int *index)
4218 struct llog_catid catid;
4219 static char name[32] = CATLIST;
4223 LASSERT(olg == &obd->obd_olg);
4225 cfs_mutex_lock(&olg->olg_cat_processing);
4226 rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4228 CERROR("rc: %d\n", rc);
4232 CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4233 obd->obd_name, *index, catid.lci_logid.lgl_oid,
4234 catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4236 rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4238 CERROR("rc: %d\n", rc);
4242 rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4244 CERROR("rc: %d\n", rc);
4249 cfs_mutex_unlock(&olg->olg_cat_processing);
4254 static int osc_llog_finish(struct obd_device *obd, int count)
4256 struct llog_ctxt *ctxt;
4257 int rc = 0, rc2 = 0;
4260 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4262 rc = llog_cleanup(ctxt);
4264 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4266 rc2 = llog_cleanup(ctxt);
4273 static int osc_reconnect(const struct lu_env *env,
4274 struct obd_export *exp, struct obd_device *obd,
4275 struct obd_uuid *cluuid,
4276 struct obd_connect_data *data,
4279 struct client_obd *cli = &obd->u.cli;
4281 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4284 client_obd_list_lock(&cli->cl_loi_list_lock);
4285 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4286 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4287 lost_grant = cli->cl_lost_grant;
4288 cli->cl_lost_grant = 0;
4289 client_obd_list_unlock(&cli->cl_loi_list_lock);
4291 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4292 "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4293 cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4294 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4295 " ocd_grant: %d\n", data->ocd_connect_flags,
4296 data->ocd_version, data->ocd_grant);
4302 static int osc_disconnect(struct obd_export *exp)
4304 struct obd_device *obd = class_exp2obd(exp);
4305 struct llog_ctxt *ctxt;
4308 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4310 if (obd->u.cli.cl_conn_count == 1) {
4311 /* Flush any remaining cancel messages out to the
4313 llog_sync(ctxt, exp);
4315 llog_ctxt_put(ctxt);
4317 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4321 rc = client_disconnect_export(exp);
4323 * Initially we put del_shrink_grant before disconnect_export, but it
4324 * causes the following problem if setup (connect) and cleanup
4325 * (disconnect) are tangled together.
4326 * connect p1 disconnect p2
4327 * ptlrpc_connect_import
4328 * ............... class_manual_cleanup
4331 * ptlrpc_connect_interrupt
4333 * add this client to shrink list
4335 * Bang! pinger trigger the shrink.
4336 * So the osc should be disconnected from the shrink list, after we
4337 * are sure the import has been destroyed. BUG18662
4339 if (obd->u.cli.cl_import == NULL)
4340 osc_del_shrink_grant(&obd->u.cli);
4344 static int osc_import_event(struct obd_device *obd,
4345 struct obd_import *imp,
4346 enum obd_import_event event)
4348 struct client_obd *cli;
4352 LASSERT(imp->imp_obd == obd);
4355 case IMP_EVENT_DISCON: {
4356 /* Only do this on the MDS OSC's */
4357 if (imp->imp_server_timeout) {
4358 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4360 cfs_spin_lock(&oscc->oscc_lock);
4361 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4362 cfs_spin_unlock(&oscc->oscc_lock);
4365 client_obd_list_lock(&cli->cl_loi_list_lock);
4366 cli->cl_avail_grant = 0;
4367 cli->cl_lost_grant = 0;
4368 client_obd_list_unlock(&cli->cl_loi_list_lock);
4371 case IMP_EVENT_INACTIVE: {
4372 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4375 case IMP_EVENT_INVALIDATE: {
4376 struct ldlm_namespace *ns = obd->obd_namespace;
4380 env = cl_env_get(&refcheck);
4384 client_obd_list_lock(&cli->cl_loi_list_lock);
4385 /* all pages go to failing rpcs due to the invalid
4387 osc_check_rpcs(env, cli);
4388 client_obd_list_unlock(&cli->cl_loi_list_lock);
4390 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4391 cl_env_put(env, &refcheck);
4396 case IMP_EVENT_ACTIVE: {
4397 /* Only do this on the MDS OSC's */
4398 if (imp->imp_server_timeout) {
4399 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4401 cfs_spin_lock(&oscc->oscc_lock);
4402 oscc->oscc_flags &= ~(OSCC_FLAG_NOSPC |
4403 OSCC_FLAG_NOSPC_BLK);
4404 cfs_spin_unlock(&oscc->oscc_lock);
4406 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4409 case IMP_EVENT_OCD: {
4410 struct obd_connect_data *ocd = &imp->imp_connect_data;
4412 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4413 osc_init_grant(&obd->u.cli, ocd);
4416 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4417 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4419 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4422 case IMP_EVENT_DEACTIVATE: {
4423 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
4426 case IMP_EVENT_ACTIVATE: {
4427 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
4431 CERROR("Unknown import event %d\n", event);
4438 * Determine whether the lock can be canceled before replaying the lock
4439 * during recovery, see bug16774 for detailed information.
4441 * \retval zero the lock can't be canceled
4442 * \retval other ok to cancel
4444 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
4446 check_res_locked(lock->l_resource);
4449 * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
4451 * XXX as a future improvement, we can also cancel unused write lock
4452 * if it doesn't have dirty data and active mmaps.
4454 if (lock->l_resource->lr_type == LDLM_EXTENT &&
4455 (lock->l_granted_mode == LCK_PR ||
4456 lock->l_granted_mode == LCK_CR) &&
4457 (osc_dlm_lock_pageref(lock) == 0))
4463 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4465 struct client_obd *cli = &obd->u.cli;
4470 rc = ptlrpcd_addref();
4474 rc = client_obd_setup(obd, lcfg);
4477 handler = ptlrpcd_alloc_work(cli->cl_import,
4478 brw_queue_work, cli);
4479 if (!IS_ERR(handler))
4480 cli->cl_writeback_work = handler;
4482 rc = PTR_ERR(handler);
4486 struct lprocfs_static_vars lvars = { 0 };
4488 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4489 lprocfs_osc_init_vars(&lvars);
4490 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4491 lproc_osc_attach_seqstat(obd);
4492 sptlrpc_lprocfs_cliobd_attach(obd);
4493 ptlrpc_lprocfs_register_obd(obd);
4497 /* We need to allocate a few requests more, because
4498 brw_interpret tries to create new requests before freeing
4499 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4500 reserved, but I afraid that might be too much wasted RAM
4501 in fact, so 2 is just my guess and still should work. */
4502 cli->cl_import->imp_rq_pool =
4503 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4505 ptlrpc_add_rqs_to_pool);
4507 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4509 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
4517 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4523 case OBD_CLEANUP_EARLY: {
4524 struct obd_import *imp;
4525 imp = obd->u.cli.cl_import;
4526 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4527 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4528 ptlrpc_deactivate_import(imp);
4529 cfs_spin_lock(&imp->imp_lock);
4530 imp->imp_pingable = 0;
4531 cfs_spin_unlock(&imp->imp_lock);
4534 case OBD_CLEANUP_EXPORTS: {
4535 struct client_obd *cli = &obd->u.cli;
4537 * for echo client, export may be on zombie list, wait for
4538 * zombie thread to cull it, because cli.cl_import will be
4539 * cleared in client_disconnect_export():
4540 * class_export_destroy() -> obd_cleanup() ->
4541 * echo_device_free() -> echo_client_cleanup() ->
4542 * obd_disconnect() -> osc_disconnect() ->
4543 * client_disconnect_export()
4545 obd_zombie_barrier();
4546 if (cli->cl_writeback_work) {
4547 ptlrpcd_destroy_work(cli->cl_writeback_work);
4548 cli->cl_writeback_work = NULL;
4550 obd_cleanup_client_import(obd);
4551 ptlrpc_lprocfs_unregister_obd(obd);
4552 lprocfs_obd_cleanup(obd);
4553 rc = obd_llog_finish(obd, 0);
4555 CERROR("failed to cleanup llogging subsystems\n");
4562 int osc_cleanup(struct obd_device *obd)
4568 /* free memory of osc quota cache */
4569 osc_quota_cleanup(obd);
4571 rc = client_obd_cleanup(obd);
4577 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4579 struct lprocfs_static_vars lvars = { 0 };
4582 lprocfs_osc_init_vars(&lvars);
4584 switch (lcfg->lcfg_command) {
4586 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4596 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4598 return osc_process_config_base(obd, buf);
4601 struct obd_ops osc_obd_ops = {
4602 .o_owner = THIS_MODULE,
4603 .o_setup = osc_setup,
4604 .o_precleanup = osc_precleanup,
4605 .o_cleanup = osc_cleanup,
4606 .o_add_conn = client_import_add_conn,
4607 .o_del_conn = client_import_del_conn,
4608 .o_connect = client_connect_import,
4609 .o_reconnect = osc_reconnect,
4610 .o_disconnect = osc_disconnect,
4611 .o_statfs = osc_statfs,
4612 .o_statfs_async = osc_statfs_async,
4613 .o_packmd = osc_packmd,
4614 .o_unpackmd = osc_unpackmd,
4615 .o_precreate = osc_precreate,
4616 .o_create = osc_create,
4617 .o_create_async = osc_create_async,
4618 .o_destroy = osc_destroy,
4619 .o_getattr = osc_getattr,
4620 .o_getattr_async = osc_getattr_async,
4621 .o_setattr = osc_setattr,
4622 .o_setattr_async = osc_setattr_async,
4624 .o_punch = osc_punch,
4626 .o_enqueue = osc_enqueue,
4627 .o_change_cbdata = osc_change_cbdata,
4628 .o_find_cbdata = osc_find_cbdata,
4629 .o_cancel = osc_cancel,
4630 .o_cancel_unused = osc_cancel_unused,
4631 .o_iocontrol = osc_iocontrol,
4632 .o_get_info = osc_get_info,
4633 .o_set_info_async = osc_set_info_async,
4634 .o_import_event = osc_import_event,
4635 .o_llog_init = osc_llog_init,
4636 .o_llog_finish = osc_llog_finish,
4637 .o_process_config = osc_process_config,
4638 .o_quotactl = osc_quotactl,
4639 .o_quotacheck = osc_quotacheck,
4640 .o_quota_adjust_qunit = osc_quota_adjust_qunit,
4643 extern struct lu_kmem_descr osc_caches[];
4644 extern cfs_spinlock_t osc_ast_guard;
4645 extern cfs_lock_class_key_t osc_ast_guard_class;
4647 int __init osc_init(void)
4649 struct lprocfs_static_vars lvars = { 0 };
4653 /* print an address of _any_ initialized kernel symbol from this
4654 * module, to allow debugging with gdb that doesn't support data
4655 * symbols from modules.*/
4656 CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
4658 rc = lu_kmem_init(osc_caches);
4660 lprocfs_osc_init_vars(&lvars);
4663 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4664 LUSTRE_OSC_NAME, &osc_device_type);
4666 lu_kmem_fini(osc_caches);
4670 cfs_spin_lock_init(&osc_ast_guard);
4671 cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4673 osc_mds_ost_orig_logops = llog_lvfs_ops;
4674 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4675 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4676 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4677 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4683 static void /*__exit*/ osc_exit(void)
4685 lu_device_type_fini(&osc_device_type);
4688 class_unregister_type(LUSTRE_OSC_NAME);
4689 lu_kmem_fini(osc_caches);
4692 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4693 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4694 MODULE_LICENSE("GPL");
4696 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);