1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_OSC
42 #include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include <lustre_cache.h>
65 #include "osc_internal.h"
67 static quota_interface_t *quota_interface = NULL;
68 extern quota_interface_t osc_quota_interface;
70 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
71 static int brw_interpret(struct ptlrpc_request *request, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76 struct lov_stripe_md *lsm)
81 lmm_size = sizeof(**lmmp);
86 OBD_FREE(*lmmp, lmm_size);
92 OBD_ALLOC(*lmmp, lmm_size);
98 LASSERT(lsm->lsm_object_id);
99 LASSERT(lsm->lsm_object_gr);
100 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109 struct lov_mds_md *lmm, int lmm_bytes)
115 if (lmm_bytes < sizeof (*lmm)) {
116 CERROR("lov_mds_md too small: %d, need %d\n",
117 lmm_bytes, (int)sizeof(*lmm));
120 /* XXX LOV_MAGIC etc check? */
122 if (lmm->lmm_object_id == 0) {
123 CERROR("lov_mds_md: zero lmm_object_id\n");
128 lsm_size = lov_stripe_md_size(1);
132 if (*lsmp != NULL && lmm == NULL) {
133 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134 OBD_FREE(*lsmp, lsm_size);
140 OBD_ALLOC(*lsmp, lsm_size);
143 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145 OBD_FREE(*lsmp, lsm_size);
148 loi_init((*lsmp)->lsm_oinfo[0]);
152 /* XXX zero *lsmp? */
153 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155 LASSERT((*lsmp)->lsm_object_id);
156 LASSERT((*lsmp)->lsm_object_gr);
159 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165 struct ost_body *body, void *capa)
167 struct obd_capa *oc = (struct obd_capa *)capa;
168 struct lustre_capa *c;
173 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
176 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177 DEBUG_CAPA(D_SEC, c, "pack");
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181 struct obd_info *oinfo)
183 struct ost_body *body;
185 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
188 body->oa = *oinfo->oi_oa;
189 osc_pack_capa(req, body, oinfo->oi_capa);
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193 const struct req_msg_field *field,
197 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
199 /* it is already calculated as sizeof struct obd_capa */
203 static int osc_getattr_interpret(struct ptlrpc_request *req,
204 struct osc_async_args *aa, int rc)
206 struct ost_body *body;
212 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
213 lustre_swab_ost_body);
215 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
218 /* This should really be sent by the OST */
219 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222 CDEBUG(D_INFO, "can't unpack ost_body\n");
224 aa->aa_oi->oi_oa->o_valid = 0;
227 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232 struct ptlrpc_request_set *set)
234 struct ptlrpc_request *req;
235 struct osc_async_args *aa;
239 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
243 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246 ptlrpc_request_free(req);
250 osc_pack_req_body(req, oinfo);
252 ptlrpc_request_set_replen(req);
253 req->rq_interpret_reply = osc_getattr_interpret;
255 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256 aa = ptlrpc_req_async_args(req);
259 ptlrpc_set_add_req(set, req);
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
265 struct ptlrpc_request *req;
266 struct ost_body *body;
270 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
274 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
277 ptlrpc_request_free(req);
281 osc_pack_req_body(req, oinfo);
283 ptlrpc_request_set_replen(req);
285 rc = ptlrpc_queue_wait(req);
289 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
291 GOTO(out, rc = -EPROTO);
293 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294 *oinfo->oi_oa = body->oa;
296 /* This should really be sent by the OST */
297 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
302 ptlrpc_req_finished(req);
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307 struct obd_trans_info *oti)
309 struct ptlrpc_request *req;
310 struct ost_body *body;
314 LASSERT(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
315 oinfo->oi_oa->o_gr > 0);
317 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
321 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
322 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
324 ptlrpc_request_free(req);
328 osc_pack_req_body(req, oinfo);
330 ptlrpc_request_set_replen(req);
333 rc = ptlrpc_queue_wait(req);
337 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
339 GOTO(out, rc = -EPROTO);
341 *oinfo->oi_oa = body->oa;
345 ptlrpc_req_finished(req);
349 static int osc_setattr_interpret(struct ptlrpc_request *req,
350 struct osc_async_args *aa, int rc)
352 struct ost_body *body;
358 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
360 GOTO(out, rc = -EPROTO);
362 *aa->aa_oi->oi_oa = body->oa;
364 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
368 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
369 struct obd_trans_info *oti,
370 struct ptlrpc_request_set *rqset)
372 struct ptlrpc_request *req;
373 struct osc_async_args *aa;
377 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
381 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
382 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
384 ptlrpc_request_free(req);
388 osc_pack_req_body(req, oinfo);
390 ptlrpc_request_set_replen(req);
392 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
394 *obdo_logcookie(oinfo->oi_oa) = *oti->oti_logcookies;
397 /* do mds to ost setattr asynchronouly */
399 /* Do not wait for response. */
400 ptlrpcd_add_req(req);
402 req->rq_interpret_reply = osc_setattr_interpret;
404 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
405 aa = ptlrpc_req_async_args(req);
408 ptlrpc_set_add_req(rqset, req);
414 int osc_real_create(struct obd_export *exp, struct obdo *oa,
415 struct lov_stripe_md **ea, struct obd_trans_info *oti)
417 struct ptlrpc_request *req;
418 struct ost_body *body;
419 struct lov_stripe_md *lsm;
428 rc = obd_alloc_memmd(exp, &lsm);
433 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
435 GOTO(out, rc = -ENOMEM);
437 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
439 ptlrpc_request_free(req);
443 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
447 ptlrpc_request_set_replen(req);
449 if (oa->o_valid & OBD_MD_FLINLINE) {
450 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
451 oa->o_flags == OBD_FL_DELORPHAN);
453 "delorphan from OST integration");
454 /* Don't resend the delorphan req */
455 req->rq_no_resend = req->rq_no_delay = 1;
458 rc = ptlrpc_queue_wait(req);
462 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
464 GOTO(out_req, rc = -EPROTO);
468 /* This should really be sent by the OST */
469 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
470 oa->o_valid |= OBD_MD_FLBLKSZ;
472 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
473 * have valid lsm_oinfo data structs, so don't go touching that.
474 * This needs to be fixed in a big way.
476 lsm->lsm_object_id = oa->o_id;
477 lsm->lsm_object_gr = oa->o_gr;
481 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
483 if (oa->o_valid & OBD_MD_FLCOOKIE) {
484 if (!oti->oti_logcookies)
485 oti_alloc_cookies(oti, 1);
486 *oti->oti_logcookies = *obdo_logcookie(oa);
490 CDEBUG(D_HA, "transno: "LPD64"\n",
491 lustre_msg_get_transno(req->rq_repmsg));
493 ptlrpc_req_finished(req);
496 obd_free_memmd(exp, &lsm);
500 static int osc_punch_interpret(struct ptlrpc_request *req,
501 struct osc_async_args *aa, int rc)
503 struct ost_body *body;
509 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
511 GOTO(out, rc = -EPROTO);
513 *aa->aa_oi->oi_oa = body->oa;
515 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
519 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
520 struct obd_trans_info *oti,
521 struct ptlrpc_request_set *rqset)
523 struct ptlrpc_request *req;
524 struct osc_async_args *aa;
525 struct ost_body *body;
530 CDEBUG(D_INFO, "oa NULL\n");
534 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
538 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
539 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
541 ptlrpc_request_free(req);
544 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
545 ptlrpc_at_set_req_timeout(req);
546 osc_pack_req_body(req, oinfo);
548 /* overload the size and blocks fields in the oa with start/end */
549 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
551 body->oa.o_size = oinfo->oi_policy.l_extent.start;
552 body->oa.o_blocks = oinfo->oi_policy.l_extent.end;
553 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
554 ptlrpc_request_set_replen(req);
557 req->rq_interpret_reply = osc_punch_interpret;
558 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
559 aa = ptlrpc_req_async_args(req);
561 ptlrpc_set_add_req(rqset, req);
566 static int osc_sync(struct obd_export *exp, struct obdo *oa,
567 struct lov_stripe_md *md, obd_size start, obd_size end,
570 struct ptlrpc_request *req;
571 struct ost_body *body;
576 CDEBUG(D_INFO, "oa NULL\n");
580 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
584 osc_set_capa_size(req, &RMF_CAPA1, capa);
585 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
587 ptlrpc_request_free(req);
591 /* overload the size and blocks fields in the oa with start/end */
592 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
595 body->oa.o_size = start;
596 body->oa.o_blocks = end;
597 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
598 osc_pack_capa(req, body, capa);
600 ptlrpc_request_set_replen(req);
602 rc = ptlrpc_queue_wait(req);
606 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
608 GOTO(out, rc = -EPROTO);
614 ptlrpc_req_finished(req);
618 /* Find and cancel locally locks matched by @mode in the resource found by
619 * @objid. Found locks are added into @cancel list. Returns the amount of
620 * locks added to @cancels list. */
621 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
622 struct list_head *cancels, ldlm_mode_t mode,
625 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
626 struct ldlm_res_id res_id;
627 struct ldlm_resource *res;
631 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
632 res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
636 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
637 lock_flags, 0, NULL);
638 ldlm_resource_putref(res);
642 static int osc_destroy_interpret(struct ptlrpc_request *req, void *data,
645 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
647 atomic_dec(&cli->cl_destroy_in_flight);
648 cfs_waitq_signal(&cli->cl_destroy_waitq);
652 static int osc_can_send_destroy(struct client_obd *cli)
654 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
655 cli->cl_max_rpcs_in_flight) {
656 /* The destroy request can be sent */
659 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
660 cli->cl_max_rpcs_in_flight) {
662 * The counter has been modified between the two atomic
665 cfs_waitq_signal(&cli->cl_destroy_waitq);
670 /* Destroy requests can be async always on the client, and we don't even really
671 * care about the return code since the client cannot do anything at all about
673 * When the MDS is unlinking a filename, it saves the file objects into a
674 * recovery llog, and these object records are cancelled when the OST reports
675 * they were destroyed and sync'd to disk (i.e. transaction committed).
676 * If the client dies, or the OST is down when the object should be destroyed,
677 * the records are not cancelled, and when the OST reconnects to the MDS next,
678 * it will retrieve the llog unlink logs and then sends the log cancellation
679 * cookies to the MDS after committing destroy transactions. */
680 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
681 struct lov_stripe_md *ea, struct obd_trans_info *oti,
682 struct obd_export *md_export)
684 struct client_obd *cli = &exp->exp_obd->u.cli;
685 struct ptlrpc_request *req;
686 struct ost_body *body;
687 CFS_LIST_HEAD(cancels);
692 CDEBUG(D_INFO, "oa NULL\n");
696 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
697 LDLM_FL_DISCARD_DATA);
699 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
701 ldlm_lock_list_put(&cancels, l_bl_ast, count);
705 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
708 ptlrpc_request_free(req);
712 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
713 req->rq_interpret_reply = osc_destroy_interpret;
714 ptlrpc_at_set_req_timeout(req);
716 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
717 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
718 sizeof(*oti->oti_logcookies));
719 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
723 ptlrpc_request_set_replen(req);
725 if (!osc_can_send_destroy(cli)) {
726 struct l_wait_info lwi = { 0 };
729 * Wait until the number of on-going destroy RPCs drops
730 * under max_rpc_in_flight
732 l_wait_event_exclusive(cli->cl_destroy_waitq,
733 osc_can_send_destroy(cli), &lwi);
736 /* Do not wait for response */
737 ptlrpcd_add_req(req);
741 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
744 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
746 LASSERT(!(oa->o_valid & bits));
749 client_obd_list_lock(&cli->cl_loi_list_lock);
750 oa->o_dirty = cli->cl_dirty;
751 if (cli->cl_dirty > cli->cl_dirty_max) {
752 CERROR("dirty %lu > dirty_max %lu\n",
753 cli->cl_dirty, cli->cl_dirty_max);
755 } else if (atomic_read(&obd_dirty_pages) > obd_max_dirty_pages) {
756 CERROR("dirty %d > system dirty_max %d\n",
757 atomic_read(&obd_dirty_pages), obd_max_dirty_pages);
759 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
760 CERROR("dirty %lu - dirty_max %lu too big???\n",
761 cli->cl_dirty, cli->cl_dirty_max);
764 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
765 (cli->cl_max_rpcs_in_flight + 1);
766 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
768 oa->o_grant = cli->cl_avail_grant;
769 oa->o_dropped = cli->cl_lost_grant;
770 cli->cl_lost_grant = 0;
771 client_obd_list_unlock(&cli->cl_loi_list_lock);
772 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
773 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
776 /* caller must hold loi_list_lock */
777 static void osc_consume_write_grant(struct client_obd *cli,
778 struct brw_page *pga)
780 atomic_inc(&obd_dirty_pages);
781 cli->cl_dirty += CFS_PAGE_SIZE;
782 cli->cl_avail_grant -= CFS_PAGE_SIZE;
783 pga->flag |= OBD_BRW_FROM_GRANT;
784 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
785 CFS_PAGE_SIZE, pga, pga->pg);
786 LASSERT(cli->cl_avail_grant >= 0);
789 /* the companion to osc_consume_write_grant, called when a brw has completed.
790 * must be called with the loi lock held. */
791 static void osc_release_write_grant(struct client_obd *cli,
792 struct brw_page *pga, int sent)
794 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
797 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
802 pga->flag &= ~OBD_BRW_FROM_GRANT;
803 atomic_dec(&obd_dirty_pages);
804 cli->cl_dirty -= CFS_PAGE_SIZE;
806 cli->cl_lost_grant += CFS_PAGE_SIZE;
807 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
808 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
809 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
810 /* For short writes we shouldn't count parts of pages that
811 * span a whole block on the OST side, or our accounting goes
812 * wrong. Should match the code in filter_grant_check. */
813 int offset = pga->off & ~CFS_PAGE_MASK;
814 int count = pga->count + (offset & (blocksize - 1));
815 int end = (offset + pga->count) & (blocksize - 1);
817 count += blocksize - end;
819 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
820 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
821 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
822 cli->cl_avail_grant, cli->cl_dirty);
828 static unsigned long rpcs_in_flight(struct client_obd *cli)
830 return cli->cl_r_in_flight + cli->cl_w_in_flight;
833 /* caller must hold loi_list_lock */
834 void osc_wake_cache_waiters(struct client_obd *cli)
836 struct list_head *l, *tmp;
837 struct osc_cache_waiter *ocw;
840 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
841 /* if we can't dirty more, we must wait until some is written */
842 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
843 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
844 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
845 "osc max %ld, sys max %d\n", cli->cl_dirty,
846 cli->cl_dirty_max, obd_max_dirty_pages);
850 /* if still dirty cache but no grant wait for pending RPCs that
851 * may yet return us some grant before doing sync writes */
852 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
853 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
854 cli->cl_w_in_flight);
858 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
859 list_del_init(&ocw->ocw_entry);
860 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
861 /* no more RPCs in flight to return grant, do sync IO */
862 ocw->ocw_rc = -EDQUOT;
863 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
865 osc_consume_write_grant(cli,
866 &ocw->ocw_oap->oap_brw_page);
869 cfs_waitq_signal(&ocw->ocw_waitq);
875 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
877 client_obd_list_lock(&cli->cl_loi_list_lock);
878 cli->cl_avail_grant = ocd->ocd_grant;
879 client_obd_list_unlock(&cli->cl_loi_list_lock);
881 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
882 cli->cl_avail_grant, cli->cl_lost_grant);
883 LASSERT(cli->cl_avail_grant >= 0);
886 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
888 client_obd_list_lock(&cli->cl_loi_list_lock);
889 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
890 if (body->oa.o_valid & OBD_MD_FLGRANT)
891 cli->cl_avail_grant += body->oa.o_grant;
892 /* waiters are woken in brw_interpret */
893 client_obd_list_unlock(&cli->cl_loi_list_lock);
896 /* We assume that the reason this OSC got a short read is because it read
897 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
898 * via the LOV, and it _knows_ it's reading inside the file, it's just that
899 * this stripe never got written at or beyond this stripe offset yet. */
900 static void handle_short_read(int nob_read, obd_count page_count,
901 struct brw_page **pga)
906 /* skip bytes read OK */
907 while (nob_read > 0) {
908 LASSERT (page_count > 0);
910 if (pga[i]->count > nob_read) {
911 /* EOF inside this page */
912 ptr = cfs_kmap(pga[i]->pg) +
913 (pga[i]->off & ~CFS_PAGE_MASK);
914 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
915 cfs_kunmap(pga[i]->pg);
921 nob_read -= pga[i]->count;
926 /* zero remaining pages */
927 while (page_count-- > 0) {
928 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
929 memset(ptr, 0, pga[i]->count);
930 cfs_kunmap(pga[i]->pg);
935 static int check_write_rcs(struct ptlrpc_request *req,
936 int requested_nob, int niocount,
937 obd_count page_count, struct brw_page **pga)
941 /* return error if any niobuf was in error */
942 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
943 sizeof(*remote_rcs) * niocount, NULL);
944 if (remote_rcs == NULL) {
945 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
948 if (lustre_msg_swabbed(req->rq_repmsg))
949 for (i = 0; i < niocount; i++)
950 __swab32s(&remote_rcs[i]);
952 for (i = 0; i < niocount; i++) {
953 if (remote_rcs[i] < 0)
954 return(remote_rcs[i]);
956 if (remote_rcs[i] != 0) {
957 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
958 i, remote_rcs[i], req);
963 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
964 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
965 requested_nob, req->rq_bulk->bd_nob_transferred);
972 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
974 if (p1->flag != p2->flag) {
975 unsigned mask = ~OBD_BRW_FROM_GRANT;
977 /* warn if we try to combine flags that we don't know to be
979 if ((p1->flag & mask) != (p2->flag & mask))
980 CERROR("is it ok to have flags 0x%x and 0x%x in the "
981 "same brw?\n", p1->flag, p2->flag);
985 return (p1->off + p1->count == p2->off);
988 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
989 struct brw_page **pga, int opc,
990 cksum_type_t cksum_type)
995 LASSERT (pg_count > 0);
996 cksum = init_checksum(cksum_type);
997 while (nob > 0 && pg_count > 0) {
998 unsigned char *ptr = cfs_kmap(pga[i]->pg);
999 int off = pga[i]->off & ~CFS_PAGE_MASK;
1000 int count = pga[i]->count > nob ? nob : pga[i]->count;
1002 /* corrupt the data before we compute the checksum, to
1003 * simulate an OST->client data error */
1004 if (i == 0 && opc == OST_READ &&
1005 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1006 memcpy(ptr + off, "bad1", min(4, nob));
1007 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1008 cfs_kunmap(pga[i]->pg);
1009 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1012 nob -= pga[i]->count;
1016 /* For sending we only compute the wrong checksum instead
1017 * of corrupting the data so it is still correct on a redo */
1018 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1024 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1025 struct lov_stripe_md *lsm, obd_count page_count,
1026 struct brw_page **pga,
1027 struct ptlrpc_request **reqp,
1028 struct obd_capa *ocapa)
1030 struct ptlrpc_request *req;
1031 struct ptlrpc_bulk_desc *desc;
1032 struct ost_body *body;
1033 struct obd_ioobj *ioobj;
1034 struct niobuf_remote *niobuf;
1035 int niocount, i, requested_nob, opc, rc;
1036 struct osc_brw_async_args *aa;
1037 struct req_capsule *pill;
1038 struct brw_page *pg_prev;
1041 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1042 RETURN(-ENOMEM); /* Recoverable */
1043 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1044 RETURN(-EINVAL); /* Fatal */
1046 if ((cmd & OBD_BRW_WRITE) != 0) {
1048 req = ptlrpc_request_alloc_pool(cli->cl_import,
1049 cli->cl_import->imp_rq_pool,
1053 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1059 for (niocount = i = 1; i < page_count; i++) {
1060 if (!can_merge_pages(pga[i - 1], pga[i]))
1064 pill = &req->rq_pill;
1065 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1066 niocount * sizeof(*niobuf));
1067 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1069 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1071 ptlrpc_request_free(req);
1074 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1075 ptlrpc_at_set_req_timeout(req);
1077 if (opc == OST_WRITE)
1078 desc = ptlrpc_prep_bulk_imp(req, page_count,
1079 BULK_GET_SOURCE, OST_BULK_PORTAL);
1081 desc = ptlrpc_prep_bulk_imp(req, page_count,
1082 BULK_PUT_SINK, OST_BULK_PORTAL);
1085 GOTO(out, rc = -ENOMEM);
1086 /* NB request now owns desc and will free it when it gets freed */
1088 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1089 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1090 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1091 LASSERT(body && ioobj && niobuf);
1095 obdo_to_ioobj(oa, ioobj);
1096 ioobj->ioo_bufcnt = niocount;
1097 osc_pack_capa(req, body, ocapa);
1098 LASSERT (page_count > 0);
1100 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1101 struct brw_page *pg = pga[i];
1103 LASSERT(pg->count > 0);
1104 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1105 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1106 pg->off, pg->count);
1108 LASSERTF(i == 0 || pg->off > pg_prev->off,
1109 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1110 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1112 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1113 pg_prev->pg, page_private(pg_prev->pg),
1114 pg_prev->pg->index, pg_prev->off);
1116 LASSERTF(i == 0 || pg->off > pg_prev->off,
1117 "i %d p_c %u\n", i, page_count);
1119 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1120 (pg->flag & OBD_BRW_SRVLOCK));
1122 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1124 requested_nob += pg->count;
1126 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1128 niobuf->len += pg->count;
1130 niobuf->offset = pg->off;
1131 niobuf->len = pg->count;
1132 niobuf->flags = pg->flag;
1137 LASSERTF((void *)(niobuf - niocount) ==
1138 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1139 niocount * sizeof(*niobuf)),
1140 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1141 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1142 (void *)(niobuf - niocount));
1144 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1146 /* size[REQ_REC_OFF] still sizeof (*body) */
1147 if (opc == OST_WRITE) {
1148 if (unlikely(cli->cl_checksum) &&
1149 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1150 /* store cl_cksum_type in a local variable since
1151 * it can be changed via lprocfs */
1152 cksum_type_t cksum_type = cli->cl_cksum_type;
1154 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1155 oa->o_flags = body->oa.o_flags = 0;
1156 body->oa.o_flags |= cksum_type_pack(cksum_type);
1157 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1158 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1162 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1164 /* save this in 'oa', too, for later checking */
1165 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1166 oa->o_flags |= cksum_type_pack(cksum_type);
1168 /* clear out the checksum flag, in case this is a
1169 * resend but cl_checksum is no longer set. b=11238 */
1170 oa->o_valid &= ~OBD_MD_FLCKSUM;
1172 oa->o_cksum = body->oa.o_cksum;
1173 /* 1 RC per niobuf */
1174 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1175 sizeof(__u32) * niocount);
1177 if (unlikely(cli->cl_checksum) &&
1178 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1179 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1180 body->oa.o_flags = 0;
1181 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1182 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1184 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1185 /* 1 RC for the whole I/O */
1187 ptlrpc_request_set_replen(req);
1189 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1190 aa = ptlrpc_req_async_args(req);
1192 aa->aa_requested_nob = requested_nob;
1193 aa->aa_nio_count = niocount;
1194 aa->aa_page_count = page_count;
1198 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1204 ptlrpc_req_finished(req);
1208 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1209 __u32 client_cksum, __u32 server_cksum, int nob,
1210 obd_count page_count, struct brw_page **pga,
1211 cksum_type_t client_cksum_type)
1215 cksum_type_t cksum_type;
1217 if (server_cksum == client_cksum) {
1218 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1222 if (oa->o_valid & OBD_MD_FLFLAGS)
1223 cksum_type = cksum_type_unpack(oa->o_flags);
1225 cksum_type = OBD_CKSUM_CRC32;
1227 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1230 if (cksum_type != client_cksum_type)
1231 msg = "the server did not use the checksum type specified in "
1232 "the original request - likely a protocol problem";
1233 else if (new_cksum == server_cksum)
1234 msg = "changed on the client after we checksummed it - "
1235 "likely false positive due to mmap IO (bug 11742)";
1236 else if (new_cksum == client_cksum)
1237 msg = "changed in transit before arrival at OST";
1239 msg = "changed in transit AND doesn't match the original - "
1240 "likely false positive due to mmap IO (bug 11742)";
1242 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1243 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1244 "["LPU64"-"LPU64"]\n",
1245 msg, libcfs_nid2str(peer->nid),
1246 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1247 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1250 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1252 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1253 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1254 "client csum now %x\n", client_cksum, client_cksum_type,
1255 server_cksum, cksum_type, new_cksum);
1259 /* Note rc enters this function as number of bytes transferred */
1260 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1262 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1263 const lnet_process_id_t *peer =
1264 &req->rq_import->imp_connection->c_peer;
1265 struct client_obd *cli = aa->aa_cli;
1266 struct ost_body *body;
1267 __u32 client_cksum = 0;
1270 if (rc < 0 && rc != -EDQUOT)
1273 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1274 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1275 lustre_swab_ost_body);
1277 CDEBUG(D_INFO, "Can't unpack body\n");
1281 /* set/clear over quota flag for a uid/gid */
1282 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1283 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1284 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1285 body->oa.o_gid, body->oa.o_valid,
1291 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1292 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1294 osc_update_grant(cli, body);
1296 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1298 CERROR("Unexpected +ve rc %d\n", rc);
1301 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1303 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1304 check_write_checksum(&body->oa, peer, client_cksum,
1305 body->oa.o_cksum, aa->aa_requested_nob,
1306 aa->aa_page_count, aa->aa_ppga,
1307 cksum_type_unpack(aa->aa_oa->o_flags)))
1310 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1313 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1314 aa->aa_page_count, aa->aa_ppga);
1318 /* The rest of this function executes only for OST_READs */
1319 if (rc > aa->aa_requested_nob) {
1320 CERROR("Unexpected rc %d (%d requested)\n", rc,
1321 aa->aa_requested_nob);
1325 if (rc != req->rq_bulk->bd_nob_transferred) {
1326 CERROR ("Unexpected rc %d (%d transferred)\n",
1327 rc, req->rq_bulk->bd_nob_transferred);
1331 if (rc < aa->aa_requested_nob)
1332 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1334 if (sptlrpc_cli_unwrap_bulk_read(req, rc, aa->aa_page_count,
1336 GOTO(out, rc = -EAGAIN);
1338 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1339 static int cksum_counter;
1340 __u32 server_cksum = body->oa.o_cksum;
1343 cksum_type_t cksum_type;
1345 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1346 cksum_type = cksum_type_unpack(body->oa.o_flags);
1348 cksum_type = OBD_CKSUM_CRC32;
1349 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1350 aa->aa_ppga, OST_READ,
1353 if (peer->nid == req->rq_bulk->bd_sender) {
1357 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1360 if (server_cksum == ~0 && rc > 0) {
1361 CERROR("Protocol error: server %s set the 'checksum' "
1362 "bit, but didn't send a checksum. Not fatal, "
1363 "but please notify on http://bugzilla.lustre.org/\n",
1364 libcfs_nid2str(peer->nid));
1365 } else if (server_cksum != client_cksum) {
1366 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1367 "%s%s%s inum "LPU64"/"LPU64" object "
1368 LPU64"/"LPU64" extent "
1369 "["LPU64"-"LPU64"]\n",
1370 req->rq_import->imp_obd->obd_name,
1371 libcfs_nid2str(peer->nid),
1373 body->oa.o_valid & OBD_MD_FLFID ?
1374 body->oa.o_fid : (__u64)0,
1375 body->oa.o_valid & OBD_MD_FLFID ?
1376 body->oa.o_generation :(__u64)0,
1378 body->oa.o_valid & OBD_MD_FLGROUP ?
1379 body->oa.o_gr : (__u64)0,
1380 aa->aa_ppga[0]->off,
1381 aa->aa_ppga[aa->aa_page_count-1]->off +
1382 aa->aa_ppga[aa->aa_page_count-1]->count -
1384 CERROR("client %x, server %x, cksum_type %x\n",
1385 client_cksum, server_cksum, cksum_type);
1387 aa->aa_oa->o_cksum = client_cksum;
1391 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1394 } else if (unlikely(client_cksum)) {
1395 static int cksum_missed;
1398 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1399 CERROR("Checksum %u requested from %s but not sent\n",
1400 cksum_missed, libcfs_nid2str(peer->nid));
1406 *aa->aa_oa = body->oa;
1411 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1412 struct lov_stripe_md *lsm,
1413 obd_count page_count, struct brw_page **pga,
1414 struct obd_capa *ocapa)
1416 struct ptlrpc_request *req;
1420 struct l_wait_info lwi;
1424 cfs_waitq_init(&waitq);
1427 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1428 page_count, pga, &req, ocapa);
1432 rc = ptlrpc_queue_wait(req);
1434 if (rc == -ETIMEDOUT && req->rq_resend) {
1435 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1436 ptlrpc_req_finished(req);
1440 rc = osc_brw_fini_request(req, rc);
1442 ptlrpc_req_finished(req);
1443 if (osc_recoverable_error(rc)) {
1445 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1446 CERROR("too many resend retries, returning error\n");
1450 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1451 l_wait_event(waitq, 0, &lwi);
1459 int osc_brw_redo_request(struct ptlrpc_request *request,
1460 struct osc_brw_async_args *aa)
1462 struct ptlrpc_request *new_req;
1463 struct ptlrpc_request_set *set = request->rq_set;
1464 struct osc_brw_async_args *new_aa;
1465 struct osc_async_page *oap;
1469 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1470 CERROR("too many resend retries, returning error\n");
1474 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1476 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
1477 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1478 ocapa = lustre_unpack_capa(request->rq_reqmsg,
1481 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1482 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1483 aa->aa_cli, aa->aa_oa,
1484 NULL /* lsm unused by osc currently */,
1485 aa->aa_page_count, aa->aa_ppga,
1486 &new_req, NULL /* ocapa */);
1490 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1492 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1493 if (oap->oap_request != NULL) {
1494 LASSERTF(request == oap->oap_request,
1495 "request %p != oap_request %p\n",
1496 request, oap->oap_request);
1497 if (oap->oap_interrupted) {
1498 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1499 ptlrpc_req_finished(new_req);
1504 /* New request takes over pga and oaps from old request.
1505 * Note that copying a list_head doesn't work, need to move it... */
1507 new_req->rq_interpret_reply = request->rq_interpret_reply;
1508 new_req->rq_async_args = request->rq_async_args;
1509 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1511 new_aa = ptlrpc_req_async_args(new_req);
1513 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1514 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1515 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1517 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1518 if (oap->oap_request) {
1519 ptlrpc_req_finished(oap->oap_request);
1520 oap->oap_request = ptlrpc_request_addref(new_req);
1524 /* use ptlrpc_set_add_req is safe because interpret functions work
1525 * in check_set context. only one way exist with access to request
1526 * from different thread got -EINTR - this way protected with
1527 * cl_loi_list_lock */
1528 ptlrpc_set_add_req(set, new_req);
1530 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1532 DEBUG_REQ(D_INFO, new_req, "new request");
1536 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1537 struct lov_stripe_md *lsm, obd_count page_count,
1538 struct brw_page **pga, struct ptlrpc_request_set *set,
1539 struct obd_capa *ocapa)
1541 struct ptlrpc_request *req;
1542 struct client_obd *cli = &exp->exp_obd->u.cli;
1544 struct osc_brw_async_args *aa;
1547 /* Consume write credits even if doing a sync write -
1548 * otherwise we may run out of space on OST due to grant. */
1549 if (cmd == OBD_BRW_WRITE) {
1550 spin_lock(&cli->cl_loi_list_lock);
1551 for (i = 0; i < page_count; i++) {
1552 if (cli->cl_avail_grant >= CFS_PAGE_SIZE)
1553 osc_consume_write_grant(cli, pga[i]);
1555 spin_unlock(&cli->cl_loi_list_lock);
1558 rc = osc_brw_prep_request(cmd, cli, oa, lsm, page_count, pga,
1561 aa = ptlrpc_req_async_args(req);
1562 if (cmd == OBD_BRW_READ) {
1563 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1564 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1565 ptlrpc_lprocfs_brw(req, OST_READ, aa->aa_requested_nob);
1567 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1568 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1569 cli->cl_w_in_flight);
1570 ptlrpc_lprocfs_brw(req, OST_WRITE, aa->aa_requested_nob);
1573 LASSERT(list_empty(&aa->aa_oaps));
1575 req->rq_interpret_reply = brw_interpret;
1576 ptlrpc_set_add_req(set, req);
1577 client_obd_list_lock(&cli->cl_loi_list_lock);
1578 if (cmd == OBD_BRW_READ)
1579 cli->cl_r_in_flight++;
1581 cli->cl_w_in_flight++;
1582 client_obd_list_unlock(&cli->cl_loi_list_lock);
1583 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DIO_PAUSE, 3);
1584 } else if (cmd == OBD_BRW_WRITE) {
1585 client_obd_list_lock(&cli->cl_loi_list_lock);
1586 for (i = 0; i < page_count; i++)
1587 osc_release_write_grant(cli, pga[i], 0);
1588 osc_wake_cache_waiters(cli);
1589 client_obd_list_unlock(&cli->cl_loi_list_lock);
1595 * ugh, we want disk allocation on the target to happen in offset order. we'll
1596 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1597 * fine for our small page arrays and doesn't require allocation. its an
1598 * insertion sort that swaps elements that are strides apart, shrinking the
1599 * stride down until its '1' and the array is sorted.
1601 static void sort_brw_pages(struct brw_page **array, int num)
1604 struct brw_page *tmp;
1608 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1613 for (i = stride ; i < num ; i++) {
1616 while (j >= stride && array[j - stride]->off > tmp->off) {
1617 array[j] = array[j - stride];
1622 } while (stride > 1);
1625 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1631 LASSERT (pages > 0);
1632 offset = pg[i]->off & ~CFS_PAGE_MASK;
1636 if (pages == 0) /* that's all */
1639 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1640 return count; /* doesn't end on page boundary */
1643 offset = pg[i]->off & ~CFS_PAGE_MASK;
1644 if (offset != 0) /* doesn't start on page boundary */
1651 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1653 struct brw_page **ppga;
1656 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1660 for (i = 0; i < count; i++)
1665 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1667 LASSERT(ppga != NULL);
1668 OBD_FREE(ppga, sizeof(*ppga) * count);
1671 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1672 obd_count page_count, struct brw_page *pga,
1673 struct obd_trans_info *oti)
1675 struct obdo *saved_oa = NULL;
1676 struct brw_page **ppga, **orig;
1677 struct obd_import *imp = class_exp2cliimp(exp);
1678 struct client_obd *cli = &imp->imp_obd->u.cli;
1679 int rc, page_count_orig;
1682 if (cmd & OBD_BRW_CHECK) {
1683 /* The caller just wants to know if there's a chance that this
1684 * I/O can succeed */
1686 if (imp == NULL || imp->imp_invalid)
1691 /* test_brw with a failed create can trip this, maybe others. */
1692 LASSERT(cli->cl_max_pages_per_rpc);
1696 orig = ppga = osc_build_ppga(pga, page_count);
1699 page_count_orig = page_count;
1701 sort_brw_pages(ppga, page_count);
1702 while (page_count) {
1703 obd_count pages_per_brw;
1705 if (page_count > cli->cl_max_pages_per_rpc)
1706 pages_per_brw = cli->cl_max_pages_per_rpc;
1708 pages_per_brw = page_count;
1710 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1712 if (saved_oa != NULL) {
1713 /* restore previously saved oa */
1714 *oinfo->oi_oa = *saved_oa;
1715 } else if (page_count > pages_per_brw) {
1716 /* save a copy of oa (brw will clobber it) */
1717 OBDO_ALLOC(saved_oa);
1718 if (saved_oa == NULL)
1719 GOTO(out, rc = -ENOMEM);
1720 *saved_oa = *oinfo->oi_oa;
1723 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1724 pages_per_brw, ppga, oinfo->oi_capa);
1729 page_count -= pages_per_brw;
1730 ppga += pages_per_brw;
1734 osc_release_ppga(orig, page_count_orig);
1736 if (saved_oa != NULL)
1737 OBDO_FREE(saved_oa);
1742 static int osc_brw_async(int cmd, struct obd_export *exp,
1743 struct obd_info *oinfo, obd_count page_count,
1744 struct brw_page *pga, struct obd_trans_info *oti,
1745 struct ptlrpc_request_set *set)
1747 struct brw_page **ppga, **orig;
1748 struct client_obd *cli = &exp->exp_obd->u.cli;
1749 int page_count_orig;
1753 if (cmd & OBD_BRW_CHECK) {
1754 struct obd_import *imp = class_exp2cliimp(exp);
1755 /* The caller just wants to know if there's a chance that this
1756 * I/O can succeed */
1758 if (imp == NULL || imp->imp_invalid)
1763 orig = ppga = osc_build_ppga(pga, page_count);
1766 page_count_orig = page_count;
1768 sort_brw_pages(ppga, page_count);
1769 while (page_count) {
1770 struct brw_page **copy;
1771 obd_count pages_per_brw;
1773 pages_per_brw = min_t(obd_count, page_count,
1774 cli->cl_max_pages_per_rpc);
1776 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1778 /* use ppga only if single RPC is going to fly */
1779 if (pages_per_brw != page_count_orig || ppga != orig) {
1780 OBD_ALLOC(copy, sizeof(*copy) * pages_per_brw);
1782 GOTO(out, rc = -ENOMEM);
1783 memcpy(copy, ppga, sizeof(*copy) * pages_per_brw);
1787 rc = async_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1788 pages_per_brw, copy, set, oinfo->oi_capa);
1792 OBD_FREE(copy, sizeof(*copy) * pages_per_brw);
1796 /* we passed it to async_internal() which is
1797 * now responsible for releasing memory */
1801 page_count -= pages_per_brw;
1802 ppga += pages_per_brw;
1806 osc_release_ppga(orig, page_count_orig);
1810 static void osc_check_rpcs(struct client_obd *cli);
1812 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1813 * the dirty accounting. Writeback completes or truncate happens before
1814 * writing starts. Must be called with the loi lock held. */
1815 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1818 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1822 /* This maintains the lists of pending pages to read/write for a given object
1823 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1824 * to quickly find objects that are ready to send an RPC. */
1825 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1831 if (lop->lop_num_pending == 0)
1834 /* if we have an invalid import we want to drain the queued pages
1835 * by forcing them through rpcs that immediately fail and complete
1836 * the pages. recovery relies on this to empty the queued pages
1837 * before canceling the locks and evicting down the llite pages */
1838 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1841 /* stream rpcs in queue order as long as as there is an urgent page
1842 * queued. this is our cheap solution for good batching in the case
1843 * where writepage marks some random page in the middle of the file
1844 * as urgent because of, say, memory pressure */
1845 if (!list_empty(&lop->lop_urgent)) {
1846 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1849 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1850 optimal = cli->cl_max_pages_per_rpc;
1851 if (cmd & OBD_BRW_WRITE) {
1852 /* trigger a write rpc stream as long as there are dirtiers
1853 * waiting for space. as they're waiting, they're not going to
1854 * create more pages to coallesce with what's waiting.. */
1855 if (!list_empty(&cli->cl_cache_waiters)) {
1856 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1859 /* +16 to avoid triggering rpcs that would want to include pages
1860 * that are being queued but which can't be made ready until
1861 * the queuer finishes with the page. this is a wart for
1862 * llite::commit_write() */
1865 if (lop->lop_num_pending >= optimal)
1871 static void on_list(struct list_head *item, struct list_head *list,
1874 if (list_empty(item) && should_be_on)
1875 list_add_tail(item, list);
1876 else if (!list_empty(item) && !should_be_on)
1877 list_del_init(item);
1880 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1881 * can find pages to build into rpcs quickly */
1882 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1884 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1885 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1886 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1888 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1889 loi->loi_write_lop.lop_num_pending);
1891 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1892 loi->loi_read_lop.lop_num_pending);
1895 static void lop_update_pending(struct client_obd *cli,
1896 struct loi_oap_pages *lop, int cmd, int delta)
1898 lop->lop_num_pending += delta;
1899 if (cmd & OBD_BRW_WRITE)
1900 cli->cl_pending_w_pages += delta;
1902 cli->cl_pending_r_pages += delta;
1905 /* this is called when a sync waiter receives an interruption. Its job is to
1906 * get the caller woken as soon as possible. If its page hasn't been put in an
1907 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1908 * desiring interruption which will forcefully complete the rpc once the rpc
1910 static void osc_occ_interrupted(struct oig_callback_context *occ)
1912 struct osc_async_page *oap;
1913 struct loi_oap_pages *lop;
1914 struct lov_oinfo *loi;
1917 /* XXX member_of() */
1918 oap = list_entry(occ, struct osc_async_page, oap_occ);
1920 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
1922 oap->oap_interrupted = 1;
1924 /* ok, it's been put in an rpc. only one oap gets a request reference */
1925 if (oap->oap_request != NULL) {
1926 ptlrpc_mark_interrupted(oap->oap_request);
1927 ptlrpcd_wake(oap->oap_request);
1931 /* we don't get interruption callbacks until osc_trigger_group_io()
1932 * has been called and put the sync oaps in the pending/urgent lists.*/
1933 if (!list_empty(&oap->oap_pending_item)) {
1934 list_del_init(&oap->oap_pending_item);
1935 list_del_init(&oap->oap_urgent_item);
1938 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1939 &loi->loi_write_lop : &loi->loi_read_lop;
1940 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1941 loi_list_maint(oap->oap_cli, oap->oap_loi);
1943 oig_complete_one(oap->oap_oig, &oap->oap_occ, -EINTR);
1944 oap->oap_oig = NULL;
1948 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
1951 /* this is trying to propogate async writeback errors back up to the
1952 * application. As an async write fails we record the error code for later if
1953 * the app does an fsync. As long as errors persist we force future rpcs to be
1954 * sync so that the app can get a sync error and break the cycle of queueing
1955 * pages for which writeback will fail. */
1956 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1963 ar->ar_force_sync = 1;
1964 ar->ar_min_xid = ptlrpc_sample_next_xid();
1969 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1970 ar->ar_force_sync = 0;
1973 static void osc_oap_to_pending(struct osc_async_page *oap)
1975 struct loi_oap_pages *lop;
1977 if (oap->oap_cmd & OBD_BRW_WRITE)
1978 lop = &oap->oap_loi->loi_write_lop;
1980 lop = &oap->oap_loi->loi_read_lop;
1982 if (oap->oap_async_flags & ASYNC_URGENT)
1983 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1984 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1985 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1988 /* this must be called holding the loi list lock to give coverage to exit_cache,
1989 * async_flag maintenance, and oap_request */
1990 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1991 struct osc_async_page *oap, int sent, int rc)
1996 if (oap->oap_request != NULL) {
1997 xid = ptlrpc_req_xid(oap->oap_request);
1998 ptlrpc_req_finished(oap->oap_request);
1999 oap->oap_request = NULL;
2002 oap->oap_async_flags = 0;
2003 oap->oap_interrupted = 0;
2005 if (oap->oap_cmd & OBD_BRW_WRITE) {
2006 osc_process_ar(&cli->cl_ar, xid, rc);
2007 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2010 if (rc == 0 && oa != NULL) {
2011 if (oa->o_valid & OBD_MD_FLBLOCKS)
2012 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2013 if (oa->o_valid & OBD_MD_FLMTIME)
2014 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2015 if (oa->o_valid & OBD_MD_FLATIME)
2016 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2017 if (oa->o_valid & OBD_MD_FLCTIME)
2018 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2022 osc_exit_cache(cli, oap, sent);
2023 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
2024 oap->oap_oig = NULL;
2029 rc = oap->oap_caller_ops->ap_completion(oap->oap_caller_data,
2030 oap->oap_cmd, oa, rc);
2032 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2033 * I/O on the page could start, but OSC calls it under lock
2034 * and thus we can add oap back to pending safely */
2036 /* upper layer wants to leave the page on pending queue */
2037 osc_oap_to_pending(oap);
2039 osc_exit_cache(cli, oap, sent);
2043 static int brw_interpret(struct ptlrpc_request *req, void *data, int rc)
2045 struct osc_brw_async_args *aa = data;
2046 struct client_obd *cli;
2049 rc = osc_brw_fini_request(req, rc);
2050 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2051 if (osc_recoverable_error(rc)) {
2052 rc = osc_brw_redo_request(req, aa);
2059 client_obd_list_lock(&cli->cl_loi_list_lock);
2061 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2062 * is called so we know whether to go to sync BRWs or wait for more
2063 * RPCs to complete */
2064 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2065 cli->cl_w_in_flight--;
2067 cli->cl_r_in_flight--;
2069 if (!list_empty(&aa->aa_oaps)) { /* from osc_send_oap_rpc() */
2070 struct osc_async_page *oap, *tmp;
2071 /* the caller may re-use the oap after the completion call so
2072 * we need to clean it up a little */
2073 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2074 list_del_init(&oap->oap_rpc_item);
2075 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
2077 OBDO_FREE(aa->aa_oa);
2078 } else { /* from async_internal() */
2080 for (i = 0; i < aa->aa_page_count; i++)
2081 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2083 osc_wake_cache_waiters(cli);
2084 osc_check_rpcs(cli);
2085 client_obd_list_unlock(&cli->cl_loi_list_lock);
2087 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2091 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
2092 struct list_head *rpc_list,
2093 int page_count, int cmd)
2095 struct ptlrpc_request *req;
2096 struct brw_page **pga = NULL;
2097 struct osc_brw_async_args *aa;
2098 struct obdo *oa = NULL;
2099 struct obd_async_page_ops *ops = NULL;
2100 void *caller_data = NULL;
2101 struct obd_capa *ocapa;
2102 struct osc_async_page *oap;
2106 LASSERT(!list_empty(rpc_list));
2108 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2110 RETURN(ERR_PTR(-ENOMEM));
2114 GOTO(out, req = ERR_PTR(-ENOMEM));
2117 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2119 ops = oap->oap_caller_ops;
2120 caller_data = oap->oap_caller_data;
2122 pga[i] = &oap->oap_brw_page;
2123 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2124 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2125 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2129 /* always get the data for the obdo for the rpc */
2130 LASSERT(ops != NULL);
2131 ops->ap_fill_obdo(caller_data, cmd, oa);
2132 ocapa = ops->ap_lookup_capa(caller_data, cmd);
2134 sort_brw_pages(pga, page_count);
2135 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2139 CERROR("prep_req failed: %d\n", rc);
2140 GOTO(out, req = ERR_PTR(rc));
2143 /* Need to update the timestamps after the request is built in case
2144 * we race with setattr (locally or in queue at OST). If OST gets
2145 * later setattr before earlier BRW (as determined by the request xid),
2146 * the OST will not use BRW timestamps. Sadly, there is no obvious
2147 * way to do this in a single call. bug 10150 */
2148 ops->ap_update_obdo(caller_data, cmd, oa,
2149 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME);
2151 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2152 aa = ptlrpc_req_async_args(req);
2153 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2154 list_splice(rpc_list, &aa->aa_oaps);
2155 CFS_INIT_LIST_HEAD(rpc_list);
2162 OBD_FREE(pga, sizeof(*pga) * page_count);
2167 /* the loi lock is held across this function but it's allowed to release
2168 * and reacquire it during its work */
2170 * prepare pages for ASYNC io and put pages in send queue.
2174 * \param cmd - OBD_BRW_* macroses
2175 * \param lop - pending pages
2177 * \return zero if pages successfully add to send queue.
2178 * \return not zere if error occurring.
2180 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
2181 int cmd, struct loi_oap_pages *lop)
2183 struct ptlrpc_request *req;
2184 obd_count page_count = 0;
2185 struct osc_async_page *oap = NULL, *tmp;
2186 struct osc_brw_async_args *aa;
2187 struct obd_async_page_ops *ops;
2188 CFS_LIST_HEAD(rpc_list);
2189 unsigned int ending_offset;
2190 unsigned starting_offset = 0;
2194 /* first we find the pages we're allowed to work with */
2195 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2197 ops = oap->oap_caller_ops;
2199 LASSERT(oap->oap_magic == OAP_MAGIC);
2201 if (page_count != 0 &&
2202 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2203 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2204 " oap %p, page %p, srvlock %u\n",
2205 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2208 /* in llite being 'ready' equates to the page being locked
2209 * until completion unlocks it. commit_write submits a page
2210 * as not ready because its unlock will happen unconditionally
2211 * as the call returns. if we race with commit_write giving
2212 * us that page we dont' want to create a hole in the page
2213 * stream, so we stop and leave the rpc to be fired by
2214 * another dirtier or kupdated interval (the not ready page
2215 * will still be on the dirty list). we could call in
2216 * at the end of ll_file_write to process the queue again. */
2217 if (!(oap->oap_async_flags & ASYNC_READY)) {
2218 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
2220 CDEBUG(D_INODE, "oap %p page %p returned %d "
2221 "instead of ready\n", oap,
2225 /* llite is telling us that the page is still
2226 * in commit_write and that we should try
2227 * and put it in an rpc again later. we
2228 * break out of the loop so we don't create
2229 * a hole in the sequence of pages in the rpc
2234 /* the io isn't needed.. tell the checks
2235 * below to complete the rpc with EINTR */
2236 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2237 oap->oap_count = -EINTR;
2240 oap->oap_async_flags |= ASYNC_READY;
2243 LASSERTF(0, "oap %p page %p returned %d "
2244 "from make_ready\n", oap,
2252 * Page submitted for IO has to be locked. Either by
2253 * ->ap_make_ready() or by higher layers.
2255 #if defined(__KERNEL__) && defined(__linux__)
2256 if(!(PageLocked(oap->oap_page) &&
2257 (CheckWriteback(oap->oap_page, cmd) || oap->oap_oig !=NULL))) {
2258 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2259 oap->oap_page, (long)oap->oap_page->flags, oap->oap_async_flags);
2263 /* If there is a gap at the start of this page, it can't merge
2264 * with any previous page, so we'll hand the network a
2265 * "fragmented" page array that it can't transfer in 1 RDMA */
2266 if (page_count != 0 && oap->oap_page_off != 0)
2269 /* take the page out of our book-keeping */
2270 list_del_init(&oap->oap_pending_item);
2271 lop_update_pending(cli, lop, cmd, -1);
2272 list_del_init(&oap->oap_urgent_item);
2274 if (page_count == 0)
2275 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2276 (PTLRPC_MAX_BRW_SIZE - 1);
2278 /* ask the caller for the size of the io as the rpc leaves. */
2279 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
2281 ops->ap_refresh_count(oap->oap_caller_data,cmd);
2282 if (oap->oap_count <= 0) {
2283 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2285 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
2289 /* now put the page back in our accounting */
2290 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2291 if (page_count == 0)
2292 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2293 if (++page_count >= cli->cl_max_pages_per_rpc)
2296 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2297 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2298 * have the same alignment as the initial writes that allocated
2299 * extents on the server. */
2300 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2301 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2302 if (ending_offset == 0)
2305 /* If there is a gap at the end of this page, it can't merge
2306 * with any subsequent pages, so we'll hand the network a
2307 * "fragmented" page array that it can't transfer in 1 RDMA */
2308 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2312 osc_wake_cache_waiters(cli);
2314 if (page_count == 0)
2317 loi_list_maint(cli, loi);
2319 client_obd_list_unlock(&cli->cl_loi_list_lock);
2321 req = osc_build_req(cli, &rpc_list, page_count, cmd);
2323 /* this should happen rarely and is pretty bad, it makes the
2324 * pending list not follow the dirty order */
2325 client_obd_list_lock(&cli->cl_loi_list_lock);
2326 list_for_each_entry_safe(oap, tmp, &rpc_list, oap_rpc_item) {
2327 list_del_init(&oap->oap_rpc_item);
2329 /* queued sync pages can be torn down while the pages
2330 * were between the pending list and the rpc */
2331 if (oap->oap_interrupted) {
2332 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2333 osc_ap_completion(cli, NULL, oap, 0,
2337 osc_ap_completion(cli, NULL, oap, 0, PTR_ERR(req));
2339 loi_list_maint(cli, loi);
2340 RETURN(PTR_ERR(req));
2343 aa = ptlrpc_req_async_args(req);
2345 if (cmd == OBD_BRW_READ) {
2346 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2347 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2348 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2349 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2350 ptlrpc_lprocfs_brw(req, OST_READ, aa->aa_requested_nob);
2352 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2353 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2354 cli->cl_w_in_flight);
2355 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2356 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2357 ptlrpc_lprocfs_brw(req, OST_WRITE, aa->aa_requested_nob);
2360 client_obd_list_lock(&cli->cl_loi_list_lock);
2362 if (cmd == OBD_BRW_READ)
2363 cli->cl_r_in_flight++;
2365 cli->cl_w_in_flight++;
2367 /* queued sync pages can be torn down while the pages
2368 * were between the pending list and the rpc */
2370 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2371 /* only one oap gets a request reference */
2374 if (oap->oap_interrupted && !req->rq_intr) {
2375 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2377 ptlrpc_mark_interrupted(req);
2381 tmp->oap_request = ptlrpc_request_addref(req);
2383 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2384 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2386 req->rq_interpret_reply = brw_interpret;
2387 ptlrpcd_add_req(req);
2391 #define LOI_DEBUG(LOI, STR, args...) \
2392 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2393 !list_empty(&(LOI)->loi_cli_item), \
2394 (LOI)->loi_write_lop.lop_num_pending, \
2395 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2396 (LOI)->loi_read_lop.lop_num_pending, \
2397 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2400 /* This is called by osc_check_rpcs() to find which objects have pages that
2401 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2402 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2405 /* first return all objects which we already know to have
2406 * pages ready to be stuffed into rpcs */
2407 if (!list_empty(&cli->cl_loi_ready_list))
2408 RETURN(list_entry(cli->cl_loi_ready_list.next,
2409 struct lov_oinfo, loi_cli_item));
2411 /* then if we have cache waiters, return all objects with queued
2412 * writes. This is especially important when many small files
2413 * have filled up the cache and not been fired into rpcs because
2414 * they don't pass the nr_pending/object threshhold */
2415 if (!list_empty(&cli->cl_cache_waiters) &&
2416 !list_empty(&cli->cl_loi_write_list))
2417 RETURN(list_entry(cli->cl_loi_write_list.next,
2418 struct lov_oinfo, loi_write_item));
2420 /* then return all queued objects when we have an invalid import
2421 * so that they get flushed */
2422 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2423 if (!list_empty(&cli->cl_loi_write_list))
2424 RETURN(list_entry(cli->cl_loi_write_list.next,
2425 struct lov_oinfo, loi_write_item));
2426 if (!list_empty(&cli->cl_loi_read_list))
2427 RETURN(list_entry(cli->cl_loi_read_list.next,
2428 struct lov_oinfo, loi_read_item));
2433 /* called with the loi list lock held */
2434 static void osc_check_rpcs(struct client_obd *cli)
2436 struct lov_oinfo *loi;
2437 int rc = 0, race_counter = 0;
2440 while ((loi = osc_next_loi(cli)) != NULL) {
2441 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2443 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2446 /* attempt some read/write balancing by alternating between
2447 * reads and writes in an object. The makes_rpc checks here
2448 * would be redundant if we were getting read/write work items
2449 * instead of objects. we don't want send_oap_rpc to drain a
2450 * partial read pending queue when we're given this object to
2451 * do io on writes while there are cache waiters */
2452 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2453 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
2454 &loi->loi_write_lop);
2462 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2463 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
2464 &loi->loi_read_lop);
2473 /* attempt some inter-object balancing by issueing rpcs
2474 * for each object in turn */
2475 if (!list_empty(&loi->loi_cli_item))
2476 list_del_init(&loi->loi_cli_item);
2477 if (!list_empty(&loi->loi_write_item))
2478 list_del_init(&loi->loi_write_item);
2479 if (!list_empty(&loi->loi_read_item))
2480 list_del_init(&loi->loi_read_item);
2482 loi_list_maint(cli, loi);
2484 /* send_oap_rpc fails with 0 when make_ready tells it to
2485 * back off. llite's make_ready does this when it tries
2486 * to lock a page queued for write that is already locked.
2487 * we want to try sending rpcs from many objects, but we
2488 * don't want to spin failing with 0. */
2489 if (race_counter == 10)
2495 /* we're trying to queue a page in the osc so we're subject to the
2496 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2497 * If the osc's queued pages are already at that limit, then we want to sleep
2498 * until there is space in the osc's queue for us. We also may be waiting for
2499 * write credits from the OST if there are RPCs in flight that may return some
2500 * before we fall back to sync writes.
2502 * We need this know our allocation was granted in the presence of signals */
2503 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2507 client_obd_list_lock(&cli->cl_loi_list_lock);
2508 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2509 client_obd_list_unlock(&cli->cl_loi_list_lock);
2513 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2514 * grant or cache space. */
2515 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
2516 struct osc_async_page *oap)
2518 struct osc_cache_waiter ocw;
2519 struct l_wait_info lwi = { 0 };
2523 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2524 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2525 cli->cl_dirty_max, obd_max_dirty_pages,
2526 cli->cl_lost_grant, cli->cl_avail_grant);
2528 /* force the caller to try sync io. this can jump the list
2529 * of queued writes and create a discontiguous rpc stream */
2530 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2531 loi->loi_ar.ar_force_sync)
2534 /* Hopefully normal case - cache space and write credits available */
2535 if ((cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max) &&
2536 (atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) &&
2537 (cli->cl_avail_grant >= CFS_PAGE_SIZE)) {
2538 /* account for ourselves */
2539 osc_consume_write_grant(cli, &oap->oap_brw_page);
2543 /* Make sure that there are write rpcs in flight to wait for. This
2544 * is a little silly as this object may not have any pending but
2545 * other objects sure might. */
2546 if (cli->cl_w_in_flight) {
2547 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2548 cfs_waitq_init(&ocw.ocw_waitq);
2552 loi_list_maint(cli, loi);
2553 osc_check_rpcs(cli);
2554 client_obd_list_unlock(&cli->cl_loi_list_lock);
2556 CDEBUG(D_CACHE, "sleeping for cache space\n");
2557 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2559 client_obd_list_lock(&cli->cl_loi_list_lock);
2560 if (!list_empty(&ocw.ocw_entry)) {
2561 list_del(&ocw.ocw_entry);
2571 * Checks if requested extent lock is compatible with a lock under the page.
2573 * Checks if the lock under \a page is compatible with a read or write lock
2574 * (specified by \a rw) for an extent [\a start , \a end].
2576 * \param exp osc export
2577 * \param lsm striping information for the file
2578 * \param res osc_async_page placeholder
2579 * \param rw OBD_BRW_READ if requested for reading,
2580 * OBD_BRW_WRITE if requested for writing
2581 * \param start start of the requested extent
2582 * \param end end of the requested extent
2583 * \param cookie transparent parameter for passing locking context
2585 * \post result == 1, *cookie == context, appropriate lock is referenced or
2588 * \retval 1 owned lock is reused for the request
2589 * \retval 0 no lock reused for the request
2591 * \see osc_release_short_lock
2593 static int osc_reget_short_lock(struct obd_export *exp,
2594 struct lov_stripe_md *lsm,
2596 obd_off start, obd_off end,
2599 struct osc_async_page *oap = *res;
2604 spin_lock(&oap->oap_lock);
2605 rc = ldlm_lock_fast_match(oap->oap_ldlm_lock, rw,
2606 start, end, cookie);
2607 spin_unlock(&oap->oap_lock);
2613 * Releases a reference to a lock taken in a "fast" way.
2615 * Releases a read or a write (specified by \a rw) lock
2616 * referenced by \a cookie.
2618 * \param exp osc export
2619 * \param lsm striping information for the file
2620 * \param end end of the locked extent
2621 * \param rw OBD_BRW_READ if requested for reading,
2622 * OBD_BRW_WRITE if requested for writing
2623 * \param cookie transparent parameter for passing locking context
2625 * \post appropriate lock is dereferenced
2627 * \see osc_reget_short_lock
2629 static int osc_release_short_lock(struct obd_export *exp,
2630 struct lov_stripe_md *lsm, obd_off end,
2631 void *cookie, int rw)
2634 ldlm_lock_fast_release(cookie, rw);
2635 /* no error could have happened at this layer */
2639 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2640 struct lov_oinfo *loi, cfs_page_t *page,
2641 obd_off offset, struct obd_async_page_ops *ops,
2642 void *data, void **res, int nocache,
2643 struct lustre_handle *lockh)
2645 struct osc_async_page *oap;
2646 struct ldlm_res_id oid;
2651 return size_round(sizeof(*oap));
2654 oap->oap_magic = OAP_MAGIC;
2655 oap->oap_cli = &exp->exp_obd->u.cli;
2658 oap->oap_caller_ops = ops;
2659 oap->oap_caller_data = data;
2661 oap->oap_page = page;
2662 oap->oap_obj_off = offset;
2664 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2665 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2666 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2667 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2669 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2671 spin_lock_init(&oap->oap_lock);
2673 /* If the page was marked as notcacheable - don't add to any locks */
2675 osc_build_res_name(loi->loi_id, loi->loi_gr, &oid);
2676 /* This is the only place where we can call cache_add_extent
2677 without oap_lock, because this page is locked now, and
2678 the lock we are adding it to is referenced, so cannot lose
2679 any pages either. */
2680 rc = cache_add_extent(oap->oap_cli->cl_cache, &oid, oap, lockh);
2685 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2689 struct osc_async_page *oap_from_cookie(void *cookie)
2691 struct osc_async_page *oap = cookie;
2692 if (oap->oap_magic != OAP_MAGIC)
2693 return ERR_PTR(-EINVAL);
2697 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2698 struct lov_oinfo *loi, void *cookie,
2699 int cmd, obd_off off, int count,
2700 obd_flag brw_flags, enum async_flags async_flags)
2702 struct client_obd *cli = &exp->exp_obd->u.cli;
2703 struct osc_async_page *oap;
2707 oap = oap_from_cookie(cookie);
2709 RETURN(PTR_ERR(oap));
2711 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2714 if (!list_empty(&oap->oap_pending_item) ||
2715 !list_empty(&oap->oap_urgent_item) ||
2716 !list_empty(&oap->oap_rpc_item))
2719 /* check if the file's owner/group is over quota */
2720 #ifdef HAVE_QUOTA_SUPPORT
2721 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)){
2722 struct obd_async_page_ops *ops;
2729 ops = oap->oap_caller_ops;
2730 ops->ap_fill_obdo(oap->oap_caller_data, cmd, oa);
2731 if (lquota_chkdq(quota_interface, cli, oa->o_uid, oa->o_gid) ==
2742 loi = lsm->lsm_oinfo[0];
2744 client_obd_list_lock(&cli->cl_loi_list_lock);
2747 oap->oap_page_off = off;
2748 oap->oap_count = count;
2749 oap->oap_brw_flags = brw_flags;
2750 oap->oap_async_flags = async_flags;
2752 if (cmd & OBD_BRW_WRITE) {
2753 rc = osc_enter_cache(cli, loi, oap);
2755 client_obd_list_unlock(&cli->cl_loi_list_lock);
2760 osc_oap_to_pending(oap);
2761 loi_list_maint(cli, loi);
2763 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2766 osc_check_rpcs(cli);
2767 client_obd_list_unlock(&cli->cl_loi_list_lock);
2772 /* aka (~was & now & flag), but this is more clear :) */
2773 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2775 static int osc_set_async_flags(struct obd_export *exp,
2776 struct lov_stripe_md *lsm,
2777 struct lov_oinfo *loi, void *cookie,
2778 obd_flag async_flags)
2780 struct client_obd *cli = &exp->exp_obd->u.cli;
2781 struct loi_oap_pages *lop;
2782 struct osc_async_page *oap;
2786 oap = oap_from_cookie(cookie);
2788 RETURN(PTR_ERR(oap));
2791 * bug 7311: OST-side locking is only supported for liblustre for now
2792 * (and liblustre never calls obd_set_async_flags(). I hope.), generic
2793 * implementation has to handle case where OST-locked page was picked
2794 * up by, e.g., ->writepage().
2796 LASSERT(!(oap->oap_brw_flags & OBD_BRW_SRVLOCK));
2797 LASSERT(!LIBLUSTRE_CLIENT); /* check that liblustre angels do fear to
2800 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2804 loi = lsm->lsm_oinfo[0];
2806 if (oap->oap_cmd & OBD_BRW_WRITE) {
2807 lop = &loi->loi_write_lop;
2809 lop = &loi->loi_read_lop;
2812 client_obd_list_lock(&cli->cl_loi_list_lock);
2814 if (list_empty(&oap->oap_pending_item))
2815 GOTO(out, rc = -EINVAL);
2817 if ((oap->oap_async_flags & async_flags) == async_flags)
2820 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2821 oap->oap_async_flags |= ASYNC_READY;
2823 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2824 if (list_empty(&oap->oap_rpc_item)) {
2825 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2826 loi_list_maint(cli, loi);
2830 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2831 oap->oap_async_flags);
2833 osc_check_rpcs(cli);
2834 client_obd_list_unlock(&cli->cl_loi_list_lock);
2838 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2839 struct lov_oinfo *loi,
2840 struct obd_io_group *oig, void *cookie,
2841 int cmd, obd_off off, int count,
2843 obd_flag async_flags)
2845 struct client_obd *cli = &exp->exp_obd->u.cli;
2846 struct osc_async_page *oap;
2847 struct loi_oap_pages *lop;
2851 oap = oap_from_cookie(cookie);
2853 RETURN(PTR_ERR(oap));
2855 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2858 if (!list_empty(&oap->oap_pending_item) ||
2859 !list_empty(&oap->oap_urgent_item) ||
2860 !list_empty(&oap->oap_rpc_item))
2864 loi = lsm->lsm_oinfo[0];
2866 client_obd_list_lock(&cli->cl_loi_list_lock);
2869 oap->oap_page_off = off;
2870 oap->oap_count = count;
2871 oap->oap_brw_flags = brw_flags;
2872 oap->oap_async_flags = async_flags;
2874 if (cmd & OBD_BRW_WRITE)
2875 lop = &loi->loi_write_lop;
2877 lop = &loi->loi_read_lop;
2879 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2880 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2882 rc = oig_add_one(oig, &oap->oap_occ);
2885 LOI_DEBUG(loi, "oap %p page %p on group pending: rc %d\n",
2886 oap, oap->oap_page, rc);
2888 client_obd_list_unlock(&cli->cl_loi_list_lock);
2893 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2894 struct loi_oap_pages *lop, int cmd)
2896 struct list_head *pos, *tmp;
2897 struct osc_async_page *oap;
2899 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2900 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2901 list_del(&oap->oap_pending_item);
2902 osc_oap_to_pending(oap);
2904 loi_list_maint(cli, loi);
2907 static int osc_trigger_group_io(struct obd_export *exp,
2908 struct lov_stripe_md *lsm,
2909 struct lov_oinfo *loi,
2910 struct obd_io_group *oig)
2912 struct client_obd *cli = &exp->exp_obd->u.cli;
2916 loi = lsm->lsm_oinfo[0];
2918 client_obd_list_lock(&cli->cl_loi_list_lock);
2920 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2921 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2923 osc_check_rpcs(cli);
2924 client_obd_list_unlock(&cli->cl_loi_list_lock);
2929 static int osc_teardown_async_page(struct obd_export *exp,
2930 struct lov_stripe_md *lsm,
2931 struct lov_oinfo *loi, void *cookie)
2933 struct client_obd *cli = &exp->exp_obd->u.cli;
2934 struct loi_oap_pages *lop;
2935 struct osc_async_page *oap;
2939 oap = oap_from_cookie(cookie);
2941 RETURN(PTR_ERR(oap));
2944 loi = lsm->lsm_oinfo[0];
2946 if (oap->oap_cmd & OBD_BRW_WRITE) {
2947 lop = &loi->loi_write_lop;
2949 lop = &loi->loi_read_lop;
2952 client_obd_list_lock(&cli->cl_loi_list_lock);
2954 if (!list_empty(&oap->oap_rpc_item))
2955 GOTO(out, rc = -EBUSY);
2957 osc_exit_cache(cli, oap, 0);
2958 osc_wake_cache_waiters(cli);
2960 if (!list_empty(&oap->oap_urgent_item)) {
2961 list_del_init(&oap->oap_urgent_item);
2962 oap->oap_async_flags &= ~ASYNC_URGENT;
2964 if (!list_empty(&oap->oap_pending_item)) {
2965 list_del_init(&oap->oap_pending_item);
2966 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2968 loi_list_maint(cli, loi);
2969 cache_remove_extent(cli->cl_cache, oap);
2971 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2973 client_obd_list_unlock(&cli->cl_loi_list_lock);
2977 int osc_extent_blocking_cb(struct ldlm_lock *lock,
2978 struct ldlm_lock_desc *new, void *data,
2981 struct lustre_handle lockh = { 0 };
2985 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
2986 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
2991 case LDLM_CB_BLOCKING:
2992 ldlm_lock2handle(lock, &lockh);
2993 rc = ldlm_cli_cancel(&lockh);
2995 CERROR("ldlm_cli_cancel failed: %d\n", rc);
2997 case LDLM_CB_CANCELING: {
2999 ldlm_lock2handle(lock, &lockh);
3000 /* This lock wasn't granted, don't try to do anything */
3001 if (lock->l_req_mode != lock->l_granted_mode)
3004 cache_remove_lock(lock->l_conn_export->exp_obd->u.cli.cl_cache,
3007 if (lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb)
3008 lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb(
3009 lock, new, data,flag);
3018 EXPORT_SYMBOL(osc_extent_blocking_cb);
3020 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data,
3023 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3026 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
3029 lock_res_and_lock(lock);
3030 #if defined (__KERNEL__) && defined (__linux__)
3031 /* Liang XXX: Darwin and Winnt checking should be added */
3032 if (lock->l_ast_data && lock->l_ast_data != data) {
3033 struct inode *new_inode = data;
3034 struct inode *old_inode = lock->l_ast_data;
3035 if (!(old_inode->i_state & I_FREEING))
3036 LDLM_ERROR(lock, "inconsistent l_ast_data found");
3037 LASSERTF(old_inode->i_state & I_FREEING,
3038 "Found existing inode %p/%lu/%u state %lu in lock: "
3039 "setting data to %p/%lu/%u\n", old_inode,
3040 old_inode->i_ino, old_inode->i_generation,
3042 new_inode, new_inode->i_ino, new_inode->i_generation);
3045 lock->l_ast_data = data;
3046 lock->l_flags |= (flags & LDLM_FL_NO_LRU);
3047 unlock_res_and_lock(lock);
3048 LDLM_LOCK_PUT(lock);
3051 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3052 ldlm_iterator_t replace, void *data)
3054 struct ldlm_res_id res_id;
3055 struct obd_device *obd = class_exp2obd(exp);
3057 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3058 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3062 static int osc_enqueue_fini(struct obd_device *obd, struct ptlrpc_request *req,
3063 struct obd_info *oinfo, int intent, int rc)
3068 /* The request was created before ldlm_cli_enqueue call. */
3069 if (rc == ELDLM_LOCK_ABORTED) {
3070 struct ldlm_reply *rep;
3071 rep = req_capsule_server_get(&req->rq_pill,
3074 LASSERT(rep != NULL);
3075 if (rep->lock_policy_res1)
3076 rc = rep->lock_policy_res1;
3080 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3081 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3082 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_size,
3083 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_blocks,
3084 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_mtime);
3088 cache_add_lock(obd->u.cli.cl_cache, oinfo->oi_lockh);
3090 /* Call the update callback. */
3091 rc = oinfo->oi_cb_up(oinfo, rc);
3095 static int osc_enqueue_interpret(struct ptlrpc_request *req,
3096 struct osc_enqueue_args *aa, int rc)
3098 int intent = aa->oa_oi->oi_flags & LDLM_FL_HAS_INTENT;
3099 struct lov_stripe_md *lsm = aa->oa_oi->oi_md;
3100 struct ldlm_lock *lock;
3102 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3104 lock = ldlm_handle2lock(aa->oa_oi->oi_lockh);
3106 /* Complete obtaining the lock procedure. */
3107 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3109 &aa->oa_oi->oi_flags,
3110 &lsm->lsm_oinfo[0]->loi_lvb,
3111 sizeof(lsm->lsm_oinfo[0]->loi_lvb),
3112 lustre_swab_ost_lvb,
3113 aa->oa_oi->oi_lockh, rc);
3115 /* Complete osc stuff. */
3116 rc = osc_enqueue_fini(aa->oa_exp->exp_obd, req, aa->oa_oi, intent, rc);
3118 /* Release the lock for async request. */
3119 if (lustre_handle_is_used(aa->oa_oi->oi_lockh) && rc == ELDLM_OK)
3120 ldlm_lock_decref(aa->oa_oi->oi_lockh, aa->oa_ei->ei_mode);
3122 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3123 aa->oa_oi->oi_lockh, req, aa);
3124 LDLM_LOCK_PUT(lock);
3128 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3129 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3130 * other synchronous requests, however keeping some locks and trying to obtain
3131 * others may take a considerable amount of time in a case of ost failure; and
3132 * when other sync requests do not get released lock from a client, the client
3133 * is excluded from the cluster -- such scenarious make the life difficult, so
3134 * release locks just after they are obtained. */
3135 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3136 struct ldlm_enqueue_info *einfo,
3137 struct ptlrpc_request_set *rqset)
3139 struct ldlm_res_id res_id;
3140 struct obd_device *obd = exp->exp_obd;
3141 struct ptlrpc_request *req = NULL;
3142 int intent = oinfo->oi_flags & LDLM_FL_HAS_INTENT;
3148 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3149 oinfo->oi_md->lsm_object_gr, &res_id);
3150 /* Filesystem lock extents are extended to page boundaries so that
3151 * dealing with the page cache is a little smoother. */
3152 oinfo->oi_policy.l_extent.start -=
3153 oinfo->oi_policy.l_extent.start & ~CFS_PAGE_MASK;
3154 oinfo->oi_policy.l_extent.end |= ~CFS_PAGE_MASK;
3156 if (oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid == 0)
3159 /* Next, search for already existing extent locks that will cover us */
3160 /* If we're trying to read, we also search for an existing PW lock. The
3161 * VFS and page cache already protect us locally, so lots of readers/
3162 * writers can share a single PW lock.
3164 * There are problems with conversion deadlocks, so instead of
3165 * converting a read lock to a write lock, we'll just enqueue a new
3168 * At some point we should cancel the read lock instead of making them
3169 * send us a blocking callback, but there are problems with canceling
3170 * locks out from other users right now, too. */
3171 mode = einfo->ei_mode;
3172 if (einfo->ei_mode == LCK_PR)
3174 mode = ldlm_lock_match(obd->obd_namespace,
3175 oinfo->oi_flags | LDLM_FL_LVB_READY, &res_id,
3176 einfo->ei_type, &oinfo->oi_policy, mode,
3179 /* addref the lock only if not async requests and PW lock is
3180 * matched whereas we asked for PR. */
3181 if (!rqset && einfo->ei_mode != mode)
3182 ldlm_lock_addref(oinfo->oi_lockh, LCK_PR);
3183 osc_set_data_with_check(oinfo->oi_lockh, einfo->ei_cbdata,
3186 /* I would like to be able to ASSERT here that rss <=
3187 * kms, but I can't, for reasons which are explained in
3191 /* We already have a lock, and it's referenced */
3192 oinfo->oi_cb_up(oinfo, ELDLM_OK);
3194 /* For async requests, decref the lock. */
3195 if (einfo->ei_mode != mode)
3196 ldlm_lock_decref(oinfo->oi_lockh, LCK_PW);
3198 ldlm_lock_decref(oinfo->oi_lockh, einfo->ei_mode);
3205 CFS_LIST_HEAD(cancels);
3206 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3207 &RQF_LDLM_ENQUEUE_LVB);
3211 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3215 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3216 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb));
3217 ptlrpc_request_set_replen(req);
3220 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3221 oinfo->oi_flags &= ~LDLM_FL_BLOCK_GRANTED;
3223 rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id,
3224 &oinfo->oi_policy, &oinfo->oi_flags,
3225 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3226 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb),
3227 lustre_swab_ost_lvb, oinfo->oi_lockh,
3231 struct osc_enqueue_args *aa;
3232 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3233 aa = ptlrpc_req_async_args(req);
3238 req->rq_interpret_reply = osc_enqueue_interpret;
3239 ptlrpc_set_add_req(rqset, req);
3240 } else if (intent) {
3241 ptlrpc_req_finished(req);
3246 rc = osc_enqueue_fini(obd, req, oinfo, intent, rc);
3248 ptlrpc_req_finished(req);
3253 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
3254 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3255 int *flags, void *data, struct lustre_handle *lockh)
3257 struct ldlm_res_id res_id;
3258 struct obd_device *obd = exp->exp_obd;
3259 int lflags = *flags;
3263 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3265 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3268 /* Filesystem lock extents are extended to page boundaries so that
3269 * dealing with the page cache is a little smoother */
3270 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3271 policy->l_extent.end |= ~CFS_PAGE_MASK;
3273 /* Next, search for already existing extent locks that will cover us */
3274 /* If we're trying to read, we also search for an existing PW lock. The
3275 * VFS and page cache already protect us locally, so lots of readers/
3276 * writers can share a single PW lock. */
3280 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3281 &res_id, type, policy, rc, lockh);
3283 osc_set_data_with_check(lockh, data, lflags);
3284 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3285 ldlm_lock_addref(lockh, LCK_PR);
3286 ldlm_lock_decref(lockh, LCK_PW);
3293 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3294 __u32 mode, struct lustre_handle *lockh)
3298 if (unlikely(mode == LCK_GROUP))
3299 ldlm_lock_decref_and_cancel(lockh, mode);
3301 ldlm_lock_decref(lockh, mode);
3306 static int osc_cancel_unused(struct obd_export *exp,
3307 struct lov_stripe_md *lsm, int flags,
3310 struct obd_device *obd = class_exp2obd(exp);
3311 struct ldlm_res_id res_id, *resp = NULL;
3314 resp = osc_build_res_name(lsm->lsm_object_id,
3315 lsm->lsm_object_gr, &res_id);
3318 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3321 static int osc_join_lru(struct obd_export *exp,
3322 struct lov_stripe_md *lsm, int join)
3324 struct obd_device *obd = class_exp2obd(exp);
3325 struct ldlm_res_id res_id, *resp = NULL;
3328 resp = osc_build_res_name(lsm->lsm_object_id,
3329 lsm->lsm_object_gr, &res_id);
3332 return ldlm_cli_join_lru(obd->obd_namespace, resp, join);
3335 static int osc_statfs_interpret(struct ptlrpc_request *req,
3336 struct osc_async_args *aa, int rc)
3338 struct obd_statfs *msfs;
3344 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3346 GOTO(out, rc = -EPROTO);
3349 *aa->aa_oi->oi_osfs = *msfs;
3351 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3355 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3356 __u64 max_age, struct ptlrpc_request_set *rqset)
3358 struct ptlrpc_request *req;
3359 struct osc_async_args *aa;
3363 /* We could possibly pass max_age in the request (as an absolute
3364 * timestamp or a "seconds.usec ago") so the target can avoid doing
3365 * extra calls into the filesystem if that isn't necessary (e.g.
3366 * during mount that would help a bit). Having relative timestamps
3367 * is not so great if request processing is slow, while absolute
3368 * timestamps are not ideal because they need time synchronization. */
3369 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3373 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3375 ptlrpc_request_free(req);
3378 ptlrpc_request_set_replen(req);
3379 req->rq_request_portal = OST_CREATE_PORTAL;
3380 ptlrpc_at_set_req_timeout(req);
3382 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3383 /* procfs requests not want stat in wait for avoid deadlock */
3384 req->rq_no_resend = 1;
3385 req->rq_no_delay = 1;
3388 req->rq_interpret_reply = osc_statfs_interpret;
3389 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3390 aa = ptlrpc_req_async_args(req);
3393 ptlrpc_set_add_req(rqset, req);
3397 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3398 __u64 max_age, __u32 flags)
3400 struct obd_statfs *msfs;
3401 struct ptlrpc_request *req;
3402 struct obd_import *imp = NULL;
3406 /*Since the request might also come from lprocfs, so we need
3407 *sync this with client_disconnect_export Bug15684*/
3408 down_read(&obd->u.cli.cl_sem);
3409 if (obd->u.cli.cl_import)
3410 imp = class_import_get(obd->u.cli.cl_import);
3411 up_read(&obd->u.cli.cl_sem);
3415 /* We could possibly pass max_age in the request (as an absolute
3416 * timestamp or a "seconds.usec ago") so the target can avoid doing
3417 * extra calls into the filesystem if that isn't necessary (e.g.
3418 * during mount that would help a bit). Having relative timestamps
3419 * is not so great if request processing is slow, while absolute
3420 * timestamps are not ideal because they need time synchronization. */
3421 req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3423 class_import_put(imp);
3428 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3430 ptlrpc_request_free(req);
3433 ptlrpc_request_set_replen(req);
3434 req->rq_request_portal = OST_CREATE_PORTAL;
3435 ptlrpc_at_set_req_timeout(req);
3437 if (flags & OBD_STATFS_NODELAY) {
3438 /* procfs requests not want stat in wait for avoid deadlock */
3439 req->rq_no_resend = 1;
3440 req->rq_no_delay = 1;
3443 rc = ptlrpc_queue_wait(req);
3447 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3449 GOTO(out, rc = -EPROTO);
3456 ptlrpc_req_finished(req);
3460 /* Retrieve object striping information.
3462 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3463 * the maximum number of OST indices which will fit in the user buffer.
3464 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3466 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3468 struct lov_user_md lum, *lumk;
3469 int rc = 0, lum_size;
3475 if (copy_from_user(&lum, lump, sizeof(lum)))
3478 if (lum.lmm_magic != LOV_USER_MAGIC)
3481 if (lum.lmm_stripe_count > 0) {
3482 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
3483 OBD_ALLOC(lumk, lum_size);
3487 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
3488 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
3490 lum_size = sizeof(lum);
3494 lumk->lmm_object_id = lsm->lsm_object_id;
3495 lumk->lmm_object_gr = lsm->lsm_object_gr;
3496 lumk->lmm_stripe_count = 1;
3498 if (copy_to_user(lump, lumk, lum_size))
3502 OBD_FREE(lumk, lum_size);
3508 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3509 void *karg, void *uarg)
3511 struct obd_device *obd = exp->exp_obd;
3512 struct obd_ioctl_data *data = karg;
3516 if (!try_module_get(THIS_MODULE)) {
3517 CERROR("Can't get module. Is it alive?");
3521 case OBD_IOC_LOV_GET_CONFIG: {
3523 struct lov_desc *desc;
3524 struct obd_uuid uuid;
3528 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3529 GOTO(out, err = -EINVAL);
3531 data = (struct obd_ioctl_data *)buf;
3533 if (sizeof(*desc) > data->ioc_inllen1) {
3534 obd_ioctl_freedata(buf, len);
3535 GOTO(out, err = -EINVAL);
3538 if (data->ioc_inllen2 < sizeof(uuid)) {
3539 obd_ioctl_freedata(buf, len);
3540 GOTO(out, err = -EINVAL);
3543 desc = (struct lov_desc *)data->ioc_inlbuf1;
3544 desc->ld_tgt_count = 1;
3545 desc->ld_active_tgt_count = 1;
3546 desc->ld_default_stripe_count = 1;
3547 desc->ld_default_stripe_size = 0;
3548 desc->ld_default_stripe_offset = 0;
3549 desc->ld_pattern = 0;
3550 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3552 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3554 err = copy_to_user((void *)uarg, buf, len);
3557 obd_ioctl_freedata(buf, len);
3560 case LL_IOC_LOV_SETSTRIPE:
3561 err = obd_alloc_memmd(exp, karg);
3565 case LL_IOC_LOV_GETSTRIPE:
3566 err = osc_getstripe(karg, uarg);
3568 case OBD_IOC_CLIENT_RECOVER:
3569 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3574 case IOC_OSC_SET_ACTIVE:
3575 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3578 case OBD_IOC_POLL_QUOTACHECK:
3579 err = lquota_poll_check(quota_interface, exp,
3580 (struct if_quotacheck *)karg);
3583 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3584 cmd, cfs_curproc_comm());
3585 GOTO(out, err = -ENOTTY);
3588 module_put(THIS_MODULE);
3592 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3593 void *key, __u32 *vallen, void *val,
3594 struct lov_stripe_md *lsm)
3597 if (!vallen || !val)
3600 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3601 __u32 *stripe = val;
3602 *vallen = sizeof(*stripe);
3605 } else if (KEY_IS(KEY_LAST_ID)) {
3606 struct ptlrpc_request *req;
3611 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3612 &RQF_OST_GET_INFO_LAST_ID);
3616 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3617 RCL_CLIENT, keylen);
3618 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3620 ptlrpc_request_free(req);
3624 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3625 memcpy(tmp, key, keylen);
3627 ptlrpc_request_set_replen(req);
3628 rc = ptlrpc_queue_wait(req);
3632 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3634 GOTO(out, rc = -EPROTO);
3636 *((obd_id *)val) = *reply;
3638 ptlrpc_req_finished(req);
3640 } else if (KEY_IS(KEY_FIEMAP)) {
3641 struct ptlrpc_request *req;
3642 struct ll_user_fiemap *reply;
3646 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3647 &RQF_OST_GET_INFO_FIEMAP);
3651 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3652 RCL_CLIENT, keylen);
3653 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3654 RCL_CLIENT, *vallen);
3655 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3656 RCL_SERVER, *vallen);
3658 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3660 ptlrpc_request_free(req);
3664 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3665 memcpy(tmp, key, keylen);
3666 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3667 memcpy(tmp, val, *vallen);
3669 ptlrpc_request_set_replen(req);
3670 rc = ptlrpc_queue_wait(req);
3674 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3676 GOTO(out1, rc = -EPROTO);
3678 memcpy(val, reply, *vallen);
3680 ptlrpc_req_finished(req);
3688 static int osc_setinfo_mds_conn_interpret(struct ptlrpc_request *req,
3691 struct llog_ctxt *ctxt;
3692 struct obd_import *imp = req->rq_import;
3698 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3701 rc = llog_initiator_connect(ctxt);
3703 CERROR("cannot establish connection for "
3704 "ctxt %p: %d\n", ctxt, rc);
3707 llog_ctxt_put(ctxt);
3708 spin_lock(&imp->imp_lock);
3709 imp->imp_server_timeout = 1;
3710 imp->imp_pingable = 1;
3711 spin_unlock(&imp->imp_lock);
3712 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3717 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3718 void *key, obd_count vallen, void *val,
3719 struct ptlrpc_request_set *set)
3721 struct ptlrpc_request *req;
3722 struct obd_device *obd = exp->exp_obd;
3723 struct obd_import *imp = class_exp2cliimp(exp);
3728 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3730 if (KEY_IS(KEY_NEXT_ID)) {
3731 if (vallen != sizeof(obd_id))
3735 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3736 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3737 exp->exp_obd->obd_name,
3738 obd->u.cli.cl_oscc.oscc_next_id);
3743 if (KEY_IS(KEY_UNLINKED)) {
3744 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3745 spin_lock(&oscc->oscc_lock);
3746 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3747 spin_unlock(&oscc->oscc_lock);
3751 if (KEY_IS(KEY_INIT_RECOV)) {
3752 if (vallen != sizeof(int))
3754 spin_lock(&imp->imp_lock);
3755 imp->imp_initial_recov = *(int *)val;
3756 spin_unlock(&imp->imp_lock);
3757 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3758 exp->exp_obd->obd_name,
3759 imp->imp_initial_recov);
3763 if (KEY_IS(KEY_CHECKSUM)) {
3764 if (vallen != sizeof(int))
3766 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3770 if (KEY_IS(KEY_FLUSH_CTX)) {
3771 sptlrpc_import_flush_my_ctx(imp);
3778 /* We pass all other commands directly to OST. Since nobody calls osc
3779 methods directly and everybody is supposed to go through LOV, we
3780 assume lov checked invalid values for us.
3781 The only recognised values so far are evict_by_nid and mds_conn.
3782 Even if something bad goes through, we'd get a -EINVAL from OST
3786 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3790 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3791 RCL_CLIENT, keylen);
3792 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3793 RCL_CLIENT, vallen);
3794 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3796 ptlrpc_request_free(req);
3800 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3801 memcpy(tmp, key, keylen);
3802 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3803 memcpy(tmp, val, vallen);
3805 if (KEY_IS(KEY_MDS_CONN)) {
3806 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3808 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3809 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3810 LASSERT(oscc->oscc_oa.o_gr > 0);
3811 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3814 ptlrpc_request_set_replen(req);
3815 ptlrpc_set_add_req(set, req);
3816 ptlrpc_check_set(set);
3822 static struct llog_operations osc_size_repl_logops = {
3823 lop_cancel: llog_obd_repl_cancel
3826 static struct llog_operations osc_mds_ost_orig_logops;
3827 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3828 struct obd_device *tgt, int count,
3829 struct llog_catid *catid, struct obd_uuid *uuid)
3834 LASSERT(olg == &obd->obd_olg);
3835 spin_lock(&obd->obd_dev_lock);
3836 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3837 osc_mds_ost_orig_logops = llog_lvfs_ops;
3838 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3839 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3840 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3841 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3843 spin_unlock(&obd->obd_dev_lock);
3845 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3846 &catid->lci_logid, &osc_mds_ost_orig_logops);
3848 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3852 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3853 NULL, &osc_size_repl_logops);
3855 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3858 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3859 obd->obd_name, tgt->obd_name, count, catid, rc);
3860 CERROR("logid "LPX64":0x%x\n",
3861 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3866 static int osc_llog_finish(struct obd_device *obd, int count)
3868 struct llog_ctxt *ctxt;
3869 int rc = 0, rc2 = 0;
3872 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3874 rc = llog_cleanup(ctxt);
3876 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3878 rc2 = llog_cleanup(ctxt);
3885 static int osc_reconnect(const struct lu_env *env,
3886 struct obd_export *exp, struct obd_device *obd,
3887 struct obd_uuid *cluuid,
3888 struct obd_connect_data *data)
3890 struct client_obd *cli = &obd->u.cli;
3892 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3895 client_obd_list_lock(&cli->cl_loi_list_lock);
3896 data->ocd_grant = cli->cl_avail_grant ?:
3897 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3898 lost_grant = cli->cl_lost_grant;
3899 cli->cl_lost_grant = 0;
3900 client_obd_list_unlock(&cli->cl_loi_list_lock);
3902 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3903 "cl_lost_grant: %ld\n", data->ocd_grant,
3904 cli->cl_avail_grant, lost_grant);
3905 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3906 " ocd_grant: %d\n", data->ocd_connect_flags,
3907 data->ocd_version, data->ocd_grant);
3913 static int osc_disconnect(struct obd_export *exp)
3915 struct obd_device *obd = class_exp2obd(exp);
3916 struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3919 if (obd->u.cli.cl_conn_count == 1)
3920 /* flush any remaining cancel messages out to the target */
3921 llog_sync(ctxt, exp);
3923 llog_ctxt_put(ctxt);
3925 rc = client_disconnect_export(exp);
3929 static int osc_import_event(struct obd_device *obd,
3930 struct obd_import *imp,
3931 enum obd_import_event event)
3933 struct client_obd *cli;
3937 LASSERT(imp->imp_obd == obd);
3940 case IMP_EVENT_DISCON: {
3941 /* Only do this on the MDS OSC's */
3942 if (imp->imp_server_timeout) {
3943 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3945 spin_lock(&oscc->oscc_lock);
3946 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3947 spin_unlock(&oscc->oscc_lock);
3950 client_obd_list_lock(&cli->cl_loi_list_lock);
3951 cli->cl_avail_grant = 0;
3952 cli->cl_lost_grant = 0;
3953 client_obd_list_unlock(&cli->cl_loi_list_lock);
3956 case IMP_EVENT_INACTIVE: {
3957 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3960 case IMP_EVENT_INVALIDATE: {
3961 struct ldlm_namespace *ns = obd->obd_namespace;
3965 client_obd_list_lock(&cli->cl_loi_list_lock);
3966 /* all pages go to failing rpcs due to the invalid import */
3967 osc_check_rpcs(cli);
3968 client_obd_list_unlock(&cli->cl_loi_list_lock);
3970 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3974 case IMP_EVENT_ACTIVE: {
3975 /* Only do this on the MDS OSC's */
3976 if (imp->imp_server_timeout) {
3977 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3979 spin_lock(&oscc->oscc_lock);
3980 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3981 spin_unlock(&oscc->oscc_lock);
3983 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3986 case IMP_EVENT_OCD: {
3987 struct obd_connect_data *ocd = &imp->imp_connect_data;
3989 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3990 osc_init_grant(&obd->u.cli, ocd);
3993 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3994 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3996 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4000 CERROR("Unknown import event %d\n", event);
4006 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4012 rc = ptlrpcd_addref();
4016 rc = client_obd_setup(obd, lcfg);
4020 struct lprocfs_static_vars lvars = { 0 };
4021 struct client_obd *cli = &obd->u.cli;
4023 lprocfs_osc_init_vars(&lvars);
4024 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4025 lproc_osc_attach_seqstat(obd);
4026 sptlrpc_lprocfs_cliobd_attach(obd);
4027 ptlrpc_lprocfs_register_obd(obd);
4031 /* We need to allocate a few requests more, because
4032 brw_interpret tries to create new requests before freeing
4033 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4034 reserved, but I afraid that might be too much wasted RAM
4035 in fact, so 2 is just my guess and still should work. */
4036 cli->cl_import->imp_rq_pool =
4037 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4039 ptlrpc_add_rqs_to_pool);
4040 cli->cl_cache = cache_create(obd);
4041 if (!cli->cl_cache) {
4050 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4056 case OBD_CLEANUP_EARLY: {
4057 struct obd_import *imp;
4058 imp = obd->u.cli.cl_import;
4059 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4060 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4061 ptlrpc_deactivate_import(imp);
4062 spin_lock(&imp->imp_lock);
4063 imp->imp_pingable = 0;
4064 spin_unlock(&imp->imp_lock);
4067 case OBD_CLEANUP_EXPORTS: {
4068 /* If we set up but never connected, the
4069 client import will not have been cleaned. */
4070 if (obd->u.cli.cl_import) {
4071 struct obd_import *imp;
4072 imp = obd->u.cli.cl_import;
4073 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4075 ptlrpc_invalidate_import(imp);
4076 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4077 class_destroy_import(imp);
4078 obd->u.cli.cl_import = NULL;
4080 rc = obd_llog_finish(obd, 0);
4082 CERROR("failed to cleanup llogging subsystems\n");
4089 int osc_cleanup(struct obd_device *obd)
4091 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4095 ptlrpc_lprocfs_unregister_obd(obd);
4096 lprocfs_obd_cleanup(obd);
4098 spin_lock(&oscc->oscc_lock);
4099 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
4100 oscc->oscc_flags |= OSCC_FLAG_EXITING;
4101 spin_unlock(&oscc->oscc_lock);
4103 /* free memory of osc quota cache */
4104 lquota_cleanup(quota_interface, obd);
4106 cache_destroy(obd->u.cli.cl_cache);
4107 rc = client_obd_cleanup(obd);
4113 static int osc_register_page_removal_cb(struct obd_export *exp,
4114 obd_page_removal_cb_t func,
4115 obd_pin_extent_cb pin_cb)
4117 return cache_add_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func,
4121 static int osc_unregister_page_removal_cb(struct obd_export *exp,
4122 obd_page_removal_cb_t func)
4124 return cache_del_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func);
4127 static int osc_register_lock_cancel_cb(struct obd_export *exp,
4128 obd_lock_cancel_cb cb)
4130 LASSERT(exp->exp_obd->u.cli.cl_ext_lock_cancel_cb == NULL);
4132 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = cb;
4136 static int osc_unregister_lock_cancel_cb(struct obd_export *exp,
4137 obd_lock_cancel_cb cb)
4139 if (exp->exp_obd->u.cli.cl_ext_lock_cancel_cb != cb) {
4140 CERROR("Unregistering cancel cb %p, while only %p was "
4142 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb);
4146 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = NULL;
4150 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4152 struct lustre_cfg *lcfg = buf;
4153 struct lprocfs_static_vars lvars = { 0 };
4156 lprocfs_osc_init_vars(&lvars);
4158 switch (lcfg->lcfg_command) {
4159 case LCFG_SPTLRPC_CONF:
4160 rc = sptlrpc_cliobd_process_config(obd, lcfg);
4163 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4171 struct obd_ops osc_obd_ops = {
4172 .o_owner = THIS_MODULE,
4173 .o_setup = osc_setup,
4174 .o_precleanup = osc_precleanup,
4175 .o_cleanup = osc_cleanup,
4176 .o_add_conn = client_import_add_conn,
4177 .o_del_conn = client_import_del_conn,
4178 .o_connect = client_connect_import,
4179 .o_reconnect = osc_reconnect,
4180 .o_disconnect = osc_disconnect,
4181 .o_statfs = osc_statfs,
4182 .o_statfs_async = osc_statfs_async,
4183 .o_packmd = osc_packmd,
4184 .o_unpackmd = osc_unpackmd,
4185 .o_precreate = osc_precreate,
4186 .o_create = osc_create,
4187 .o_destroy = osc_destroy,
4188 .o_getattr = osc_getattr,
4189 .o_getattr_async = osc_getattr_async,
4190 .o_setattr = osc_setattr,
4191 .o_setattr_async = osc_setattr_async,
4193 .o_brw_async = osc_brw_async,
4194 .o_prep_async_page = osc_prep_async_page,
4195 .o_reget_short_lock = osc_reget_short_lock,
4196 .o_release_short_lock = osc_release_short_lock,
4197 .o_queue_async_io = osc_queue_async_io,
4198 .o_set_async_flags = osc_set_async_flags,
4199 .o_queue_group_io = osc_queue_group_io,
4200 .o_trigger_group_io = osc_trigger_group_io,
4201 .o_teardown_async_page = osc_teardown_async_page,
4202 .o_punch = osc_punch,
4204 .o_enqueue = osc_enqueue,
4205 .o_match = osc_match,
4206 .o_change_cbdata = osc_change_cbdata,
4207 .o_cancel = osc_cancel,
4208 .o_cancel_unused = osc_cancel_unused,
4209 .o_join_lru = osc_join_lru,
4210 .o_iocontrol = osc_iocontrol,
4211 .o_get_info = osc_get_info,
4212 .o_set_info_async = osc_set_info_async,
4213 .o_import_event = osc_import_event,
4214 .o_llog_init = osc_llog_init,
4215 .o_llog_finish = osc_llog_finish,
4216 .o_process_config = osc_process_config,
4217 .o_register_page_removal_cb = osc_register_page_removal_cb,
4218 .o_unregister_page_removal_cb = osc_unregister_page_removal_cb,
4219 .o_register_lock_cancel_cb = osc_register_lock_cancel_cb,
4220 .o_unregister_lock_cancel_cb = osc_unregister_lock_cancel_cb,
4222 int __init osc_init(void)
4224 struct lprocfs_static_vars lvars = { 0 };
4228 lprocfs_osc_init_vars(&lvars);
4230 request_module("lquota");
4231 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4232 lquota_init(quota_interface);
4233 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4235 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4236 LUSTRE_OSC_NAME, NULL);
4238 if (quota_interface)
4239 PORTAL_SYMBOL_PUT(osc_quota_interface);
4247 static void /*__exit*/ osc_exit(void)
4249 lquota_exit(quota_interface);
4250 if (quota_interface)
4251 PORTAL_SYMBOL_PUT(osc_quota_interface);
4253 class_unregister_type(LUSTRE_OSC_NAME);
4256 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4257 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4258 MODULE_LICENSE("GPL");
4260 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);