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
43 # include <libcfs/libcfs.h>
44 #else /* __KERNEL__ */
45 # include <liblustre.h>
48 # include <lustre_dlm.h>
49 #include <libcfs/kp30.h>
50 #include <lustre_net.h>
51 #include <lustre/lustre_user.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 static quota_interface_t *quota_interface;
75 extern quota_interface_t osc_quota_interface;
78 atomic_t osc_resend_time;
80 /* Pack OSC object metadata for disk storage (LE byte order). */
81 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
82 struct lov_stripe_md *lsm)
87 lmm_size = sizeof(**lmmp);
92 OBD_FREE(*lmmp, lmm_size);
98 OBD_ALLOC(*lmmp, lmm_size);
104 LASSERT(lsm->lsm_object_id);
105 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
111 /* Unpack OSC object metadata from disk storage (LE byte order). */
112 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
113 struct lov_mds_md *lmm, int lmm_bytes)
119 if (lmm_bytes < sizeof (*lmm)) {
120 CERROR("lov_mds_md too small: %d, need %d\n",
121 lmm_bytes, (int)sizeof(*lmm));
124 /* XXX LOV_MAGIC etc check? */
126 if (lmm->lmm_object_id == 0) {
127 CERROR("lov_mds_md: zero lmm_object_id\n");
132 lsm_size = lov_stripe_md_size(1);
136 if (*lsmp != NULL && lmm == NULL) {
137 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
138 OBD_FREE(*lsmp, lsm_size);
144 OBD_ALLOC(*lsmp, lsm_size);
147 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
148 if ((*lsmp)->lsm_oinfo[0] == NULL) {
149 OBD_FREE(*lsmp, lsm_size);
152 loi_init((*lsmp)->lsm_oinfo[0]);
156 /* XXX zero *lsmp? */
157 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
158 LASSERT((*lsmp)->lsm_object_id);
161 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
166 static int osc_getattr_interpret(struct ptlrpc_request *req,
167 struct osc_async_args *aa, int rc)
169 struct ost_body *body;
175 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
176 lustre_swab_ost_body);
178 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
179 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
181 /* This should really be sent by the OST */
182 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
183 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
185 CERROR("can't unpack ost_body\n");
187 aa->aa_oi->oi_oa->o_valid = 0;
190 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
194 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
195 struct ptlrpc_request_set *set)
197 struct ptlrpc_request *req;
198 struct ost_body *body;
199 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
200 struct osc_async_args *aa;
203 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
204 OST_GETATTR, 2, size,NULL);
208 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
209 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
211 ptlrpc_req_set_repsize(req, 2, size);
212 req->rq_interpret_reply = osc_getattr_interpret;
214 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
215 aa = ptlrpc_req_async_args(req);
218 ptlrpc_set_add_req(set, req);
222 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
224 struct ptlrpc_request *req;
225 struct ost_body *body;
226 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
230 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
231 OST_GETATTR, 2, size, NULL);
235 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
236 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
238 ptlrpc_req_set_repsize(req, 2, size);
240 rc = ptlrpc_queue_wait(req);
242 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
246 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
247 lustre_swab_ost_body);
249 CERROR ("can't unpack ost_body\n");
250 GOTO (out, rc = -EPROTO);
253 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
254 memcpy(oinfo->oi_oa, &body->oa, sizeof(*oinfo->oi_oa));
256 /* This should really be sent by the OST */
257 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
258 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
262 ptlrpc_req_finished(req);
266 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
267 struct obd_trans_info *oti)
269 struct ptlrpc_request *req;
270 struct ost_body *body;
271 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
275 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
276 OST_SETATTR, 2, size, NULL);
280 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
281 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
283 ptlrpc_req_set_repsize(req, 2, size);
285 rc = ptlrpc_queue_wait(req);
289 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
290 lustre_swab_ost_body);
292 GOTO(out, rc = -EPROTO);
294 memcpy(oinfo->oi_oa, &body->oa, sizeof(*oinfo->oi_oa));
298 ptlrpc_req_finished(req);
302 static int osc_setattr_interpret(struct ptlrpc_request *req,
303 struct osc_async_args *aa, int rc)
305 struct ost_body *body;
311 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
312 lustre_swab_ost_body);
314 CERROR("can't unpack ost_body\n");
315 GOTO(out, rc = -EPROTO);
318 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
320 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
324 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
325 struct obd_trans_info *oti,
326 struct ptlrpc_request_set *rqset)
328 struct ptlrpc_request *req;
329 struct ost_body *body;
330 __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body), 0 };
332 struct osc_async_args *aa;
335 if (osc_exp_is_2_0_server(exp)) {
339 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
340 OST_SETATTR, bufcount, size, NULL);
344 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
346 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
348 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
351 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
352 ptlrpc_req_set_repsize(req, 2, size);
353 /* do mds to ost setattr asynchronouly */
355 /* Do not wait for response. */
356 ptlrpcd_add_req(req);
358 req->rq_interpret_reply = osc_setattr_interpret;
360 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
361 aa = ptlrpc_req_async_args(req);
364 ptlrpc_set_add_req(rqset, req);
370 int osc_real_create(struct obd_export *exp, struct obdo *oa,
371 struct lov_stripe_md **ea, struct obd_trans_info *oti)
373 struct ptlrpc_request *req;
374 struct ost_body *body;
375 struct lov_stripe_md *lsm;
376 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
385 rc = obd_alloc_memmd(exp, &lsm);
390 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
391 OST_CREATE, 2, size, NULL);
393 GOTO(out, rc = -ENOMEM);
395 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
396 memcpy(&body->oa, oa, sizeof(body->oa));
398 ptlrpc_req_set_repsize(req, 2, size);
399 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
400 oa->o_flags == OBD_FL_DELORPHAN) {
402 "delorphan from OST integration");
403 /* Don't resend the delorphan req */
404 req->rq_no_resend = req->rq_no_delay = 1;
407 rc = ptlrpc_queue_wait(req);
411 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
412 lustre_swab_ost_body);
414 CERROR ("can't unpack ost_body\n");
415 GOTO (out_req, rc = -EPROTO);
418 memcpy(oa, &body->oa, sizeof(*oa));
420 /* This should really be sent by the OST */
421 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
422 oa->o_valid |= OBD_MD_FLBLKSZ;
424 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
425 * have valid lsm_oinfo data structs, so don't go touching that.
426 * This needs to be fixed in a big way.
428 lsm->lsm_object_id = oa->o_id;
432 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
434 if (oa->o_valid & OBD_MD_FLCOOKIE) {
435 if (!oti->oti_logcookies)
436 oti_alloc_cookies(oti, 1);
437 *oti->oti_logcookies = oa->o_lcookie;
441 CDEBUG(D_HA, "transno: "LPD64"\n",
442 lustre_msg_get_transno(req->rq_repmsg));
444 ptlrpc_req_finished(req);
447 obd_free_memmd(exp, &lsm);
451 static int osc_punch_interpret(struct ptlrpc_request *req,
452 struct osc_async_args *aa, int rc)
454 struct ost_body *body;
460 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof (*body),
461 lustre_swab_ost_body);
463 CERROR ("can't unpack ost_body\n");
464 GOTO(out, rc = -EPROTO);
467 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
469 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
473 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
474 struct obd_trans_info *oti,
475 struct ptlrpc_request_set *rqset)
477 struct ptlrpc_request *req;
478 struct osc_async_args *aa;
479 struct ost_body *body;
480 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
488 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
489 OST_PUNCH, 2, size, NULL);
493 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
494 ptlrpc_at_set_req_timeout(req);
496 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
497 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
499 /* overload the size and blocks fields in the oa with start/end */
500 body->oa.o_size = oinfo->oi_policy.l_extent.start;
501 body->oa.o_blocks = oinfo->oi_policy.l_extent.end;
502 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
504 ptlrpc_req_set_repsize(req, 2, size);
506 req->rq_interpret_reply = osc_punch_interpret;
507 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
508 aa = ptlrpc_req_async_args(req);
510 ptlrpc_set_add_req(rqset, req);
515 static int osc_sync_interpret(struct ptlrpc_request *req,
516 struct osc_async_args *aa, int rc)
518 struct ost_body *body;
524 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
525 lustre_swab_ost_body);
527 CERROR ("can't unpack ost_body\n");
528 GOTO(out, rc = -EPROTO);
531 *aa->aa_oi->oi_oa = body->oa;
533 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
537 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
538 obd_size start, obd_size end,
539 struct ptlrpc_request_set *set)
541 struct ptlrpc_request *req;
542 struct ost_body *body;
543 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
544 struct osc_async_args *aa;
552 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
553 OST_SYNC, 2, size, NULL);
557 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
558 memcpy(&body->oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
560 /* overload the size and blocks fields in the oa with start/end */
561 body->oa.o_size = start;
562 body->oa.o_blocks = end;
563 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
565 ptlrpc_req_set_repsize(req, 2, size);
566 req->rq_interpret_reply = osc_sync_interpret;
568 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
569 aa = ptlrpc_req_async_args(req);
572 ptlrpc_set_add_req(set, req);
576 /* Find and cancel locally locks matched by @mode in the resource found by
577 * @objid. Found locks are added into @cancel list. Returns the amount of
578 * locks added to @cancels list. */
579 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
580 struct list_head *cancels, ldlm_mode_t mode,
583 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
584 struct ldlm_res_id res_id;
585 struct ldlm_resource *res;
589 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
590 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
594 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
595 lock_flags, 0, NULL);
596 ldlm_resource_putref(res);
600 static int osc_destroy_interpret(struct ptlrpc_request *req, void *data,
603 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
605 atomic_dec(&cli->cl_destroy_in_flight);
606 cfs_waitq_signal(&cli->cl_destroy_waitq);
610 static int osc_can_send_destroy(struct client_obd *cli)
612 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
613 cli->cl_max_rpcs_in_flight) {
614 /* The destroy request can be sent */
617 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
618 cli->cl_max_rpcs_in_flight) {
620 * The counter has been modified between the two atomic
623 cfs_waitq_signal(&cli->cl_destroy_waitq);
628 /* Destroy requests can be async always on the client, and we don't even really
629 * care about the return code since the client cannot do anything at all about
631 * When the MDS is unlinking a filename, it saves the file objects into a
632 * recovery llog, and these object records are cancelled when the OST reports
633 * they were destroyed and sync'd to disk (i.e. transaction committed).
634 * If the client dies, or the OST is down when the object should be destroyed,
635 * the records are not cancelled, and when the OST reconnects to the MDS next,
636 * it will retrieve the llog unlink logs and then sends the log cancellation
637 * cookies to the MDS after committing destroy transactions. */
638 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
639 struct lov_stripe_md *ea, struct obd_trans_info *oti,
640 struct obd_export *md_export)
642 CFS_LIST_HEAD(cancels);
643 struct ptlrpc_request *req;
644 struct ost_body *body;
645 __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body),
646 sizeof(struct ldlm_request) };
647 int count, bufcount = 2;
648 struct client_obd *cli = &exp->exp_obd->u.cli;
656 LASSERT(oa->o_id != 0);
658 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
659 LDLM_FL_DISCARD_DATA);
660 if (exp_connect_cancelset(exp))
662 req = ldlm_prep_elc_req(exp, LUSTRE_OST_VERSION, OST_DESTROY, bufcount,
663 size, REQ_REC_OFF + 1, 0, &cancels, count);
667 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
668 ptlrpc_at_set_req_timeout(req);
670 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
672 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
673 oa->o_lcookie = *oti->oti_logcookies;
676 memcpy(&body->oa, oa, sizeof(*oa));
677 ptlrpc_req_set_repsize(req, 2, size);
679 /* don't throttle destroy RPCs for the MDT */
680 if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
681 req->rq_interpret_reply = osc_destroy_interpret;
682 if (!osc_can_send_destroy(cli)) {
683 struct l_wait_info lwi = { 0 };
686 * Wait until the number of on-going destroy RPCs drops
687 * under max_rpc_in_flight
689 l_wait_event_exclusive(cli->cl_destroy_waitq,
690 osc_can_send_destroy(cli), &lwi);
694 /* Do not wait for response */
695 ptlrpcd_add_req(req);
699 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
702 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
704 LASSERT(!(oa->o_valid & bits));
707 client_obd_list_lock(&cli->cl_loi_list_lock);
708 oa->o_dirty = cli->cl_dirty;
709 if (cli->cl_dirty > cli->cl_dirty_max) {
710 CERROR("dirty %lu > dirty_max %lu\n",
711 cli->cl_dirty, cli->cl_dirty_max);
713 } else if (atomic_read(&obd_dirty_pages) > obd_max_dirty_pages) {
714 CERROR("dirty %d > system dirty_max %d\n",
715 atomic_read(&obd_dirty_pages), obd_max_dirty_pages);
717 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
718 CERROR("dirty %lu - dirty_max %lu too big???\n",
719 cli->cl_dirty, cli->cl_dirty_max);
722 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
723 (cli->cl_max_rpcs_in_flight + 1);
724 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
726 oa->o_grant = cli->cl_avail_grant;
727 oa->o_dropped = cli->cl_lost_grant;
728 cli->cl_lost_grant = 0;
729 client_obd_list_unlock(&cli->cl_loi_list_lock);
730 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
731 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
734 /* caller must hold loi_list_lock */
735 static void osc_consume_write_grant(struct client_obd *cli,struct brw_page *pga)
737 atomic_inc(&obd_dirty_pages);
738 cli->cl_dirty += CFS_PAGE_SIZE;
739 cli->cl_avail_grant -= CFS_PAGE_SIZE;
740 pga->flag |= OBD_BRW_FROM_GRANT;
741 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
742 CFS_PAGE_SIZE, pga, pga->pg);
743 LASSERTF(cli->cl_avail_grant >= 0, "invalid avail grant is %ld \n",
744 cli->cl_avail_grant);
747 /* the companion to osc_consume_write_grant, called when a brw has completed.
748 * must be called with the loi lock held. */
749 static void osc_release_write_grant(struct client_obd *cli,
750 struct brw_page *pga, int sent)
752 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
755 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
760 pga->flag &= ~OBD_BRW_FROM_GRANT;
761 atomic_dec(&obd_dirty_pages);
762 cli->cl_dirty -= CFS_PAGE_SIZE;
764 cli->cl_lost_grant += CFS_PAGE_SIZE;
765 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
766 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
767 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
768 /* For short writes we shouldn't count parts of pages that
769 * span a whole block on the OST side, or our accounting goes
770 * wrong. Should match the code in filter_grant_check. */
771 int offset = pga->off & ~CFS_PAGE_MASK;
772 int count = pga->count + (offset & (blocksize - 1));
773 int end = (offset + pga->count) & (blocksize - 1);
775 count += blocksize - end;
777 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
778 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
779 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
780 cli->cl_avail_grant, cli->cl_dirty);
786 static unsigned long rpcs_in_flight(struct client_obd *cli)
788 return cli->cl_r_in_flight + cli->cl_w_in_flight;
791 /* caller must hold loi_list_lock */
792 void osc_wake_cache_waiters(struct client_obd *cli)
794 struct list_head *l, *tmp;
795 struct osc_cache_waiter *ocw;
798 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
799 /* if we can't dirty more, we must wait until some is written */
800 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
801 ((atomic_read(&obd_dirty_pages)+1)>(obd_max_dirty_pages))) {
802 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
803 "osc max %ld, sys max %d\n", cli->cl_dirty,
804 cli->cl_dirty_max, obd_max_dirty_pages);
808 /* if still dirty cache but no grant wait for pending RPCs that
809 * may yet return us some grant before doing sync writes */
810 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
811 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
812 cli->cl_w_in_flight);
816 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
817 list_del_init(&ocw->ocw_entry);
818 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
819 /* no more RPCs in flight to return grant, do sync IO */
820 ocw->ocw_rc = -EDQUOT;
821 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
823 osc_consume_write_grant(cli,
824 &ocw->ocw_oap->oap_brw_page);
827 cfs_waitq_signal(&ocw->ocw_waitq);
833 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
835 client_obd_list_lock(&cli->cl_loi_list_lock);
836 cli->cl_avail_grant = ocd->ocd_grant;
837 client_obd_list_unlock(&cli->cl_loi_list_lock);
839 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
840 cli->cl_avail_grant, cli->cl_lost_grant);
841 LASSERT(cli->cl_avail_grant >= 0);
844 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
846 client_obd_list_lock(&cli->cl_loi_list_lock);
847 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
848 if (body->oa.o_valid & OBD_MD_FLGRANT)
849 cli->cl_avail_grant += body->oa.o_grant;
850 /* waiters are woken in brw_interpret */
851 client_obd_list_unlock(&cli->cl_loi_list_lock);
854 /* We assume that the reason this OSC got a short read is because it read
855 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
856 * via the LOV, and it _knows_ it's reading inside the file, it's just that
857 * this stripe never got written at or beyond this stripe offset yet. */
858 static void handle_short_read(int nob_read, obd_count page_count,
859 struct brw_page **pga)
864 /* skip bytes read OK */
865 while (nob_read > 0) {
866 LASSERT (page_count > 0);
868 if (pga[i]->count > nob_read) {
869 /* EOF inside this page */
870 ptr = cfs_kmap(pga[i]->pg) +
871 (pga[i]->off & ~CFS_PAGE_MASK);
872 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
873 cfs_kunmap(pga[i]->pg);
879 nob_read -= pga[i]->count;
884 /* zero remaining pages */
885 while (page_count-- > 0) {
886 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
887 memset(ptr, 0, pga[i]->count);
888 cfs_kunmap(pga[i]->pg);
893 static int check_write_rcs(struct ptlrpc_request *req,
894 int requested_nob, int niocount,
895 obd_count page_count, struct brw_page **pga)
899 /* return error if any niobuf was in error */
900 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
901 sizeof(*remote_rcs) * niocount, NULL);
902 if (remote_rcs == NULL) {
903 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
906 if (lustre_rep_need_swab(req))
907 for (i = 0; i < niocount; i++)
908 __swab32s(&remote_rcs[i]);
910 for (i = 0; i < niocount; i++) {
911 if (remote_rcs[i] < 0)
912 return(remote_rcs[i]);
914 if (remote_rcs[i] != 0) {
915 CERROR("rc[%d] invalid (%d) req %p\n",
916 i, remote_rcs[i], req);
921 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
922 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
923 req->rq_bulk->bd_nob_transferred, requested_nob);
930 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
932 if (p1->flag != p2->flag) {
933 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_ASYNC);
935 /* warn if we try to combine flags that we don't know to be
937 if ((p1->flag & mask) != (p2->flag & mask))
938 CERROR("is it ok to have flags 0x%x and 0x%x in the "
939 "same brw?\n", p1->flag, p2->flag);
943 return (p1->off + p1->count == p2->off);
946 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
947 struct brw_page **pga, int opc,
948 cksum_type_t cksum_type)
953 LASSERT (pg_count > 0);
954 cksum = init_checksum(cksum_type);
955 while (nob > 0 && pg_count > 0) {
956 unsigned char *ptr = cfs_kmap(pga[i]->pg);
957 int off = pga[i]->off & ~CFS_PAGE_MASK;
958 int count = pga[i]->count > nob ? nob : pga[i]->count;
960 /* corrupt the data before we compute the checksum, to
961 * simulate an OST->client data error */
962 if (i == 0 && opc == OST_READ &&
963 OBD_FAIL_CHECK_ONCE(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
964 memcpy(ptr + off, "bad1", min(4, nob));
965 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
966 cfs_kunmap(pga[i]->pg);
967 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
970 nob -= pga[i]->count;
974 /* For sending we only compute the wrong checksum instead
975 * of corrupting the data so it is still correct on a redo */
976 if (opc == OST_WRITE && OBD_FAIL_CHECK_ONCE(OBD_FAIL_OSC_CHECKSUM_SEND))
982 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
983 struct lov_stripe_md *lsm, obd_count page_count,
984 struct brw_page **pga,
985 struct ptlrpc_request **reqp)
987 struct ptlrpc_request *req;
988 struct ptlrpc_bulk_desc *desc;
989 struct ost_body *body;
990 struct obd_ioobj *ioobj;
991 struct niobuf_remote *niobuf;
992 __u32 size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
993 int niocount, i, requested_nob, opc, rc;
994 struct ptlrpc_request_pool *pool;
995 struct osc_brw_async_args *aa;
996 struct brw_page *pg_prev;
999 OBD_FAIL_RETURN(OBD_FAIL_OSC_BRW_PREP_REQ, -ENOMEM); /* Recoverable */
1000 OBD_FAIL_RETURN(OBD_FAIL_OSC_BRW_PREP_REQ2, -EINVAL); /* Fatal */
1002 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
1003 pool = ((cmd & OBD_BRW_WRITE) != 0) ? cli->cl_import->imp_rq_pool :NULL;
1005 for (niocount = i = 1; i < page_count; i++) {
1006 if (!can_merge_pages(pga[i - 1], pga[i]))
1010 size[REQ_REC_OFF + 1] = sizeof(*ioobj);
1011 size[REQ_REC_OFF + 2] = niocount * sizeof(*niobuf);
1013 req = ptlrpc_prep_req_pool(cli->cl_import, LUSTRE_OST_VERSION, opc, 4, size,
1018 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1019 ptlrpc_at_set_req_timeout(req);
1021 if (opc == OST_WRITE)
1022 desc = ptlrpc_prep_bulk_imp (req, page_count,
1023 BULK_GET_SOURCE, OST_BULK_PORTAL);
1025 desc = ptlrpc_prep_bulk_imp (req, page_count,
1026 BULK_PUT_SINK, OST_BULK_PORTAL);
1028 GOTO(out, rc = -ENOMEM);
1029 /* NB request now owns desc and will free it when it gets freed */
1031 body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
1032 ioobj = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, sizeof(*ioobj));
1033 niobuf = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1034 niocount * sizeof(*niobuf));
1036 memcpy(&body->oa, oa, sizeof(*oa));
1038 obdo_to_ioobj(oa, ioobj);
1039 ioobj->ioo_bufcnt = niocount;
1041 LASSERT (page_count > 0);
1043 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1044 struct brw_page *pg = pga[i];
1046 LASSERT(pg->count > 0);
1047 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1048 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1049 pg->off, pg->count);
1051 LASSERTF(i == 0 || pg->off > pg_prev->off,
1052 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1053 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1055 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1056 pg_prev->pg, page_private(pg_prev->pg),
1057 pg_prev->pg->index, pg_prev->off);
1059 LASSERTF(i == 0 || pg->off > pg_prev->off,
1060 "i %d p_c %u\n", i, page_count);
1062 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1063 (pg->flag & OBD_BRW_SRVLOCK));
1065 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1067 requested_nob += pg->count;
1069 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1071 niobuf->len += pg->count;
1073 niobuf->offset = pg->off;
1074 niobuf->len = pg->count;
1075 niobuf->flags = pg->flag;
1080 LASSERTF((void *)(niobuf - niocount) ==
1081 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1082 niocount * sizeof(*niobuf)),
1083 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1084 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1085 (void *)(niobuf - niocount));
1087 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1089 /* size[REQ_REC_OFF] still sizeof (*body) */
1090 if (opc == OST_WRITE) {
1091 if (cli->cl_checksum) {
1092 /* store cl_cksum_type in a local variable since
1093 * it can be changed via lprocfs */
1094 cksum_type_t cksum_type = cli->cl_cksum_type;
1096 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1097 oa->o_flags = body->oa.o_flags = 0;
1098 body->oa.o_flags |= cksum_type_pack(cksum_type);
1099 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1100 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1104 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1106 /* save this in 'oa', too, for later checking */
1107 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1108 oa->o_flags |= cksum_type_pack(cksum_type);
1110 /* clear out the checksum flag, in case this is a
1111 * resend but cl_checksum is no longer set. b=11238 */
1112 oa->o_valid &= ~OBD_MD_FLCKSUM;
1114 oa->o_cksum = body->oa.o_cksum;
1115 /* 1 RC per niobuf */
1116 size[REPLY_REC_OFF + 1] = sizeof(__u32) * niocount;
1117 ptlrpc_req_set_repsize(req, 3, size);
1119 if (cli->cl_checksum) {
1120 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1121 body->oa.o_flags = 0;
1122 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1123 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1125 /* 1 RC for the whole I/O */
1126 ptlrpc_req_set_repsize(req, 2, size);
1129 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1130 aa = ptlrpc_req_async_args(req);
1132 aa->aa_requested_nob = requested_nob;
1133 aa->aa_nio_count = niocount;
1134 aa->aa_page_count = page_count;
1138 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1144 ptlrpc_req_finished (req);
1148 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1149 __u32 client_cksum, __u32 server_cksum, int nob,
1150 obd_count page_count, struct brw_page **pga,
1151 cksum_type_t client_cksum_type)
1155 cksum_type_t cksum_type;
1157 if (server_cksum == client_cksum) {
1158 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1162 if (oa->o_valid & OBD_MD_FLFLAGS)
1163 cksum_type = cksum_type_unpack(oa->o_flags);
1165 cksum_type = OBD_CKSUM_CRC32;
1167 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1170 if (cksum_type != client_cksum_type)
1171 msg = "the server did not use the checksum type specified in "
1172 "the original request - likely a protocol problem";
1173 else if (new_cksum == server_cksum)
1174 msg = "changed on the client after we checksummed it - "
1175 "likely false positive due to mmap IO (bug 11742)";
1176 else if (new_cksum == client_cksum)
1177 msg = "changed in transit before arrival at OST";
1179 msg = "changed in transit AND doesn't match the original - "
1180 "likely false positive due to mmap IO (bug 11742)";
1182 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1183 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1184 "["LPU64"-"LPU64"]\n",
1185 msg, libcfs_nid2str(peer->nid),
1186 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1187 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1190 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1192 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1193 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1194 "client csum now %x\n", client_cksum, client_cksum_type,
1195 server_cksum, cksum_type, new_cksum);
1200 /* Note rc enters this function as number of bytes transferred */
1201 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1203 struct osc_brw_async_args *aa = ptlrpc_req_async_args(req);
1204 const lnet_process_id_t *peer =
1205 &req->rq_import->imp_connection->c_peer;
1206 struct client_obd *cli = aa->aa_cli;
1207 struct ost_body *body;
1208 __u32 client_cksum = 0;
1211 if (rc < 0 && rc != -EDQUOT)
1214 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1215 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1216 lustre_swab_ost_body);
1218 CERROR ("Can't unpack body\n");
1222 /* set/clear over quota flag for a uid/gid */
1223 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1224 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1225 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1226 body->oa.o_gid, body->oa.o_valid,
1232 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1233 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1235 osc_update_grant(cli, body);
1237 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1239 CERROR ("Unexpected +ve rc %d\n", rc);
1242 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1244 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1245 check_write_checksum(&body->oa, peer, client_cksum,
1246 body->oa.o_cksum, aa->aa_requested_nob,
1247 aa->aa_page_count, aa->aa_ppga,
1248 cksum_type_unpack(aa->aa_oa->o_flags)))
1251 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1252 aa->aa_page_count, aa->aa_ppga);
1256 /* The rest of this function executes only for OST_READs */
1257 if (rc > aa->aa_requested_nob) {
1258 CERROR("Unexpected rc %d (%d requested)\n", rc,
1259 aa->aa_requested_nob);
1263 if (rc != req->rq_bulk->bd_nob_transferred) {
1264 CERROR ("Unexpected rc %d (%d transferred)\n",
1265 rc, req->rq_bulk->bd_nob_transferred);
1269 if (rc < aa->aa_requested_nob)
1270 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1272 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1273 static int cksum_counter;
1274 __u32 server_cksum = body->oa.o_cksum;
1277 cksum_type_t cksum_type;
1279 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1280 cksum_type = cksum_type_unpack(body->oa.o_flags);
1282 cksum_type = OBD_CKSUM_CRC32;
1283 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1284 aa->aa_ppga, OST_READ,
1287 if (peer->nid == req->rq_bulk->bd_sender) {
1291 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1294 if (server_cksum == ~0 && rc > 0) {
1295 CERROR("Protocol error: server %s set the 'checksum' "
1296 "bit, but didn't send a checksum. Not fatal, "
1297 "but please notify on http://bugzilla.lustre.org/\n",
1298 libcfs_nid2str(peer->nid));
1299 } else if (server_cksum != client_cksum) {
1300 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1301 "%s%s%s inum "LPU64"/"LPU64" object "
1302 LPU64"/"LPU64" extent "
1303 "["LPU64"-"LPU64"]\n",
1304 req->rq_import->imp_obd->obd_name,
1305 libcfs_nid2str(peer->nid),
1307 body->oa.o_valid & OBD_MD_FLFID ?
1308 body->oa.o_fid : (__u64)0,
1309 body->oa.o_valid & OBD_MD_FLFID ?
1310 body->oa.o_generation :(__u64)0,
1312 body->oa.o_valid & OBD_MD_FLGROUP ?
1313 body->oa.o_gr : (__u64)0,
1314 aa->aa_ppga[0]->off,
1315 aa->aa_ppga[aa->aa_page_count-1]->off +
1316 aa->aa_ppga[aa->aa_page_count-1]->count -
1318 CERROR("client %x, server %x, cksum_type %x\n",
1319 client_cksum, server_cksum, cksum_type);
1321 aa->aa_oa->o_cksum = client_cksum;
1325 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1328 } else if (unlikely(client_cksum)) {
1329 static int cksum_missed;
1332 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1333 CERROR("Checksum %u requested from %s but not sent\n",
1334 cksum_missed, libcfs_nid2str(peer->nid));
1340 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
1345 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
1346 struct lov_stripe_md *lsm,
1347 obd_count page_count, struct brw_page **pga)
1349 struct ptlrpc_request *request;
1353 struct l_wait_info lwi;
1356 init_waitqueue_head(&waitq);
1359 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1360 page_count, pga, &request);
1364 rc = ptlrpc_queue_wait(request);
1366 if (rc == -ETIMEDOUT && request->rq_resend) {
1367 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
1368 ptlrpc_req_finished(request);
1372 rc = osc_brw_fini_request(request, rc);
1374 ptlrpc_req_finished(request);
1375 if (osc_recoverable_error(rc)) {
1377 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1378 CERROR("too many resend retries, returning error\n");
1382 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1383 l_wait_event(waitq, 0, &lwi);
1390 int osc_brw_redo_request(struct ptlrpc_request *request,
1391 struct osc_brw_async_args *aa)
1393 struct ptlrpc_request *new_req;
1394 struct ptlrpc_request_set *set = request->rq_set;
1395 struct osc_brw_async_args *new_aa;
1396 struct osc_async_page *oap;
1400 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1401 CERROR("too many resend retries, returning error\n");
1405 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1407 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1408 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1409 aa->aa_cli, aa->aa_oa,
1410 NULL /* lsm unused by osc currently */,
1411 aa->aa_page_count, aa->aa_ppga, &new_req);
1415 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1417 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1418 if (oap->oap_request != NULL) {
1419 LASSERTF(request == oap->oap_request,
1420 "request %p != oap_request %p\n",
1421 request, oap->oap_request);
1422 if (oap->oap_interrupted) {
1423 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1424 ptlrpc_req_finished(new_req);
1429 /* New request takes over pga and oaps from old request.
1430 * Note that copying a list_head doesn't work, need to move it... */
1432 new_req->rq_interpret_reply = request->rq_interpret_reply;
1433 new_req->rq_async_args = request->rq_async_args;
1434 new_req->rq_sent = CURRENT_SECONDS + aa->aa_resends;
1436 new_aa = ptlrpc_req_async_args(new_req);
1438 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1439 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1440 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1442 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1443 if (oap->oap_request) {
1444 ptlrpc_req_finished(oap->oap_request);
1445 oap->oap_request = ptlrpc_request_addref(new_req);
1449 /* use ptlrpc_set_add_req is safe because interpret functions work
1450 * in check_set context. only one way exist with access to request
1451 * from different thread got -EINTR - this way protected with
1452 * cl_loi_list_lock */
1453 ptlrpc_set_add_req(set, new_req);
1455 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1457 DEBUG_REQ(D_INFO, new_req, "new request");
1461 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1462 struct lov_stripe_md *lsm, obd_count page_count,
1463 struct brw_page **pga, struct ptlrpc_request_set *set)
1465 struct ptlrpc_request *request;
1466 struct client_obd *cli = &exp->exp_obd->u.cli;
1468 struct osc_brw_async_args *aa;
1471 /* Consume write credits even if doing a sync write -
1472 * otherwise we may run out of space on OST due to grant. */
1473 if (cmd == OBD_BRW_WRITE) {
1474 client_obd_list_lock(&cli->cl_loi_list_lock);
1475 for (i = 0; i < page_count; i++) {
1476 if (cli->cl_avail_grant >= CFS_PAGE_SIZE)
1477 osc_consume_write_grant(cli, pga[i]);
1479 client_obd_list_unlock(&cli->cl_loi_list_lock);
1482 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1483 page_count, pga, &request);
1485 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1486 aa = ptlrpc_req_async_args(request);
1487 if (cmd == OBD_BRW_READ) {
1488 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1489 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1491 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1492 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1493 cli->cl_w_in_flight);
1495 ptlrpc_lprocfs_brw(request, aa->aa_requested_nob);
1497 LASSERT(list_empty(&aa->aa_oaps));
1500 request->rq_interpret_reply = brw_interpret;
1501 ptlrpc_set_add_req(set, request);
1502 client_obd_list_lock(&cli->cl_loi_list_lock);
1503 if (cmd == OBD_BRW_READ)
1504 cli->cl_r_in_flight++;
1506 cli->cl_w_in_flight++;
1507 client_obd_list_unlock(&cli->cl_loi_list_lock);
1508 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DIO_PAUSE, 3);
1509 } else if (cmd == OBD_BRW_WRITE) {
1510 client_obd_list_lock(&cli->cl_loi_list_lock);
1511 for (i = 0; i < page_count; i++)
1512 osc_release_write_grant(cli, pga[i], 0);
1513 osc_wake_cache_waiters(cli);
1514 client_obd_list_unlock(&cli->cl_loi_list_lock);
1521 * ugh, we want disk allocation on the target to happen in offset order. we'll
1522 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1523 * fine for our small page arrays and doesn't require allocation. its an
1524 * insertion sort that swaps elements that are strides apart, shrinking the
1525 * stride down until its '1' and the array is sorted.
1527 static void sort_brw_pages(struct brw_page **array, int num)
1530 struct brw_page *tmp;
1534 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1539 for (i = stride ; i < num ; i++) {
1542 while (j >= stride && array[j-stride]->off > tmp->off) {
1543 array[j] = array[j - stride];
1548 } while (stride > 1);
1551 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1557 LASSERT (pages > 0);
1558 offset = pg[i]->off & (~CFS_PAGE_MASK);
1562 if (pages == 0) /* that's all */
1565 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1566 return count; /* doesn't end on page boundary */
1569 offset = pg[i]->off & (~CFS_PAGE_MASK);
1570 if (offset != 0) /* doesn't start on page boundary */
1577 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1579 struct brw_page **ppga;
1582 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1586 for (i = 0; i < count; i++)
1591 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1593 LASSERT(ppga != NULL);
1594 OBD_FREE(ppga, sizeof(*ppga) * count);
1597 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1598 obd_count page_count, struct brw_page *pga,
1599 struct obd_trans_info *oti)
1601 struct obdo *saved_oa = NULL;
1602 struct brw_page **ppga, **orig;
1603 struct obd_import *imp = class_exp2cliimp(exp);
1604 struct client_obd *cli = &imp->imp_obd->u.cli;
1605 int rc, page_count_orig;
1608 if (cmd & OBD_BRW_CHECK) {
1609 /* The caller just wants to know if there's a chance that this
1610 * I/O can succeed */
1612 if (imp == NULL || imp->imp_invalid)
1617 /* test_brw with a failed create can trip this, maybe others. */
1618 LASSERT(cli->cl_max_pages_per_rpc);
1622 orig = ppga = osc_build_ppga(pga, page_count);
1625 page_count_orig = page_count;
1627 sort_brw_pages(ppga, page_count);
1628 while (page_count) {
1629 obd_count pages_per_brw;
1631 if (page_count > cli->cl_max_pages_per_rpc)
1632 pages_per_brw = cli->cl_max_pages_per_rpc;
1634 pages_per_brw = page_count;
1636 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1638 if (saved_oa != NULL) {
1639 /* restore previously saved oa */
1640 *oinfo->oi_oa = *saved_oa;
1641 } else if (page_count > pages_per_brw) {
1642 /* save a copy of oa (brw will clobber it) */
1643 OBDO_ALLOC(saved_oa);
1644 if (saved_oa == NULL)
1645 GOTO(out, rc = -ENOMEM);
1646 *saved_oa = *oinfo->oi_oa;
1649 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1650 pages_per_brw, ppga);
1655 page_count -= pages_per_brw;
1656 ppga += pages_per_brw;
1660 osc_release_ppga(orig, page_count_orig);
1662 if (saved_oa != NULL)
1663 OBDO_FREE(saved_oa);
1668 static int osc_brw_async(int cmd, struct obd_export *exp,
1669 struct obd_info *oinfo, obd_count page_count,
1670 struct brw_page *pga, struct obd_trans_info *oti,
1671 struct ptlrpc_request_set *set)
1673 struct brw_page **ppga, **orig;
1674 int page_count_orig;
1678 if (cmd & OBD_BRW_CHECK) {
1679 /* The caller just wants to know if there's a chance that this
1680 * I/O can succeed */
1681 struct obd_import *imp = class_exp2cliimp(exp);
1683 if (imp == NULL || imp->imp_invalid)
1688 orig = ppga = osc_build_ppga(pga, page_count);
1691 page_count_orig = page_count;
1693 sort_brw_pages(ppga, page_count);
1694 while (page_count) {
1695 struct brw_page **copy;
1696 obd_count pages_per_brw;
1698 pages_per_brw = min_t(obd_count, page_count,
1699 class_exp2cliimp(exp)->imp_obd->u.cli.cl_max_pages_per_rpc);
1701 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1703 /* use ppga only if single RPC is going to fly */
1704 if (pages_per_brw != page_count_orig || ppga != orig) {
1705 OBD_ALLOC(copy, pages_per_brw * sizeof(*copy));
1707 GOTO(out, rc = -ENOMEM);
1708 memcpy(copy, ppga, pages_per_brw * sizeof(*copy));
1712 rc = async_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1713 pages_per_brw, copy, set);
1717 OBD_FREE(copy, pages_per_brw * sizeof(*copy));
1722 /* we passed it to async_internal() which is
1723 * now responsible for releasing memory */
1727 page_count -= pages_per_brw;
1728 ppga += pages_per_brw;
1732 osc_release_ppga(orig, page_count_orig);
1736 static void osc_check_rpcs(struct client_obd *cli);
1738 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1739 * the dirty accounting. Writeback completes or truncate happens before
1740 * writing starts. Must be called with the loi lock held. */
1741 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1744 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1747 /* This maintains the lists of pending pages to read/write for a given object
1748 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1749 * to quickly find objects that are ready to send an RPC. */
1750 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1756 if (lop->lop_num_pending == 0)
1759 /* if we have an invalid import we want to drain the queued pages
1760 * by forcing them through rpcs that immediately fail and complete
1761 * the pages. recovery relies on this to empty the queued pages
1762 * before canceling the locks and evicting down the llite pages */
1763 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1766 /* stream rpcs in queue order as long as as there is an urgent page
1767 * queued. this is our cheap solution for good batching in the case
1768 * where writepage marks some random page in the middle of the file
1769 * as urgent because of, say, memory pressure */
1770 if (!list_empty(&lop->lop_urgent)) {
1771 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1775 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1776 optimal = cli->cl_max_pages_per_rpc;
1777 if (cmd & OBD_BRW_WRITE) {
1778 /* trigger a write rpc stream as long as there are dirtiers
1779 * waiting for space. as they're waiting, they're not going to
1780 * create more pages to coallesce with what's waiting.. */
1781 if (!list_empty(&cli->cl_cache_waiters)) {
1782 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1786 /* +16 to avoid triggering rpcs that would want to include pages
1787 * that are being queued but which can't be made ready until
1788 * the queuer finishes with the page. this is a wart for
1789 * llite::commit_write() */
1792 if (lop->lop_num_pending >= optimal)
1798 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1800 struct osc_async_page *oap;
1803 if (list_empty(&lop->lop_urgent))
1806 oap = list_entry(lop->lop_urgent.next,
1807 struct osc_async_page, oap_urgent_item);
1809 if (oap->oap_async_flags & ASYNC_HP) {
1810 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1817 static void on_list(struct list_head *item, struct list_head *list,
1820 if (list_empty(item) && should_be_on)
1821 list_add_tail(item, list);
1822 else if (!list_empty(item) && !should_be_on)
1823 list_del_init(item);
1826 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1827 * can find pages to build into rpcs quickly */
1828 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1830 if (lop_makes_hprpc(&loi->loi_write_lop) ||
1831 lop_makes_hprpc(&loi->loi_read_lop)) {
1833 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1834 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1836 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1837 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1838 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1839 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1842 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1843 loi->loi_write_lop.lop_num_pending);
1845 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1846 loi->loi_read_lop.lop_num_pending);
1849 static void lop_update_pending(struct client_obd *cli,
1850 struct loi_oap_pages *lop, int cmd, int delta)
1852 lop->lop_num_pending += delta;
1853 if (cmd & OBD_BRW_WRITE)
1854 cli->cl_pending_w_pages += delta;
1856 cli->cl_pending_r_pages += delta;
1859 /* this is called when a sync waiter receives an interruption. Its job is to
1860 * get the caller woken as soon as possible. If its page hasn't been put in an
1861 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1862 * desiring interruption which will forcefully complete the rpc once the rpc
1864 static void osc_occ_interrupted(struct oig_callback_context *occ)
1866 struct osc_async_page *oap;
1867 struct loi_oap_pages *lop;
1868 struct lov_oinfo *loi;
1871 /* XXX member_of() */
1872 oap = list_entry(occ, struct osc_async_page, oap_occ);
1874 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
1876 oap->oap_interrupted = 1;
1878 /* ok, it's been put in an rpc. only one oap gets a request reference */
1879 if (oap->oap_request != NULL) {
1880 ptlrpc_mark_interrupted(oap->oap_request);
1881 ptlrpcd_wake(oap->oap_request);
1885 /* we don't get interruption callbacks until osc_trigger_group_io()
1886 * has been called and put the sync oaps in the pending/urgent lists.*/
1887 if (!list_empty(&oap->oap_pending_item)) {
1888 list_del_init(&oap->oap_pending_item);
1889 list_del_init(&oap->oap_urgent_item);
1892 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1893 &loi->loi_write_lop : &loi->loi_read_lop;
1894 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1895 loi_list_maint(oap->oap_cli, oap->oap_loi);
1897 oig_complete_one(oap->oap_oig, &oap->oap_occ, -EINTR);
1898 oap->oap_oig = NULL;
1902 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
1905 /* this is trying to propogate async writeback errors back up to the
1906 * application. As an async write fails we record the error code for later if
1907 * the app does an fsync. As long as errors persist we force future rpcs to be
1908 * sync so that the app can get a sync error and break the cycle of queueing
1909 * pages for which writeback will fail. */
1910 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1917 ar->ar_force_sync = 1;
1918 ar->ar_min_xid = ptlrpc_sample_next_xid();
1923 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1924 ar->ar_force_sync = 0;
1927 static void osc_oap_to_pending(struct osc_async_page *oap)
1929 struct loi_oap_pages *lop;
1931 if (oap->oap_cmd & OBD_BRW_WRITE)
1932 lop = &oap->oap_loi->loi_write_lop;
1934 lop = &oap->oap_loi->loi_read_lop;
1936 if (oap->oap_async_flags & ASYNC_HP)
1937 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1938 else if (oap->oap_async_flags & ASYNC_URGENT)
1939 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
1940 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1941 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1944 /* this must be called holding the loi list lock to give coverage to exit_cache,
1945 * async_flag maintenance, and oap_request */
1946 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1947 struct osc_async_page *oap, int sent, int rc)
1952 if (oap->oap_request != NULL) {
1953 xid = ptlrpc_req_xid(oap->oap_request);
1954 ptlrpc_req_finished(oap->oap_request);
1955 oap->oap_request = NULL;
1958 oap->oap_async_flags = 0;
1959 oap->oap_interrupted = 0;
1961 if (oap->oap_cmd & OBD_BRW_WRITE) {
1962 osc_process_ar(&cli->cl_ar, xid, rc);
1963 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
1966 if (rc == 0 && oa != NULL) {
1967 if (oa->o_valid & OBD_MD_FLBLOCKS)
1968 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
1969 if (oa->o_valid & OBD_MD_FLMTIME)
1970 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
1971 if (oa->o_valid & OBD_MD_FLATIME)
1972 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
1973 if (oa->o_valid & OBD_MD_FLCTIME)
1974 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
1978 osc_exit_cache(cli, oap, sent);
1979 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1980 oap->oap_oig = NULL;
1985 rc = oap->oap_caller_ops->ap_completion(oap->oap_caller_data,
1986 oap->oap_cmd, oa, rc);
1988 /* ll_ap_completion (from llite) drops PG_locked. so, a new
1989 * I/O on the page could start, but OSC calls it under lock
1990 * and thus we can add oap back to pending safely */
1992 /* upper layer wants to leave the page on pending queue */
1993 osc_oap_to_pending(oap);
1995 osc_exit_cache(cli, oap, sent);
1999 static int brw_interpret(struct ptlrpc_request *request, void *data, int rc)
2001 struct osc_brw_async_args *aa = data;
2002 struct client_obd *cli;
2005 rc = osc_brw_fini_request(request, rc);
2006 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
2008 if (osc_recoverable_error(rc)) {
2009 rc = osc_brw_redo_request(request, aa);
2015 client_obd_list_lock(&cli->cl_loi_list_lock);
2016 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2017 * is called so we know whether to go to sync BRWs or wait for more
2018 * RPCs to complete */
2019 if (lustre_msg_get_opc(request->rq_reqmsg) == OST_WRITE)
2020 cli->cl_w_in_flight--;
2022 cli->cl_r_in_flight--;
2024 if (!list_empty(&aa->aa_oaps)) { /* from osc_send_oap_rpc() */
2025 struct osc_async_page *oap, *tmp;
2026 /* the caller may re-use the oap after the completion call so
2027 * we need to clean it up a little */
2028 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2029 list_del_init(&oap->oap_rpc_item);
2030 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
2032 OBDO_FREE(aa->aa_oa);
2033 } else { /* from async_internal() */
2035 for (i = 0; i < aa->aa_page_count; i++)
2036 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2038 osc_wake_cache_waiters(cli);
2039 osc_check_rpcs(cli);
2040 client_obd_list_unlock(&cli->cl_loi_list_lock);
2042 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2047 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
2048 struct list_head *rpc_list,
2049 int page_count, int cmd)
2051 struct ptlrpc_request *req;
2052 struct brw_page **pga = NULL;
2053 struct osc_brw_async_args *aa;
2054 struct obdo *oa = NULL;
2055 struct obd_async_page_ops *ops = NULL;
2056 void *caller_data = NULL;
2057 struct osc_async_page *oap;
2058 struct ldlm_lock *lock = NULL;
2063 LASSERT(!list_empty(rpc_list));
2065 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2067 RETURN(ERR_PTR(-ENOMEM));
2071 GOTO(out, req = ERR_PTR(-ENOMEM));
2074 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2076 ops = oap->oap_caller_ops;
2077 caller_data = oap->oap_caller_data;
2078 lock = oap->oap_ldlm_lock;
2080 pga[i] = &oap->oap_brw_page;
2081 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2082 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2083 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2087 /* always get the data for the obdo for the rpc */
2088 LASSERT(ops != NULL);
2089 ops->ap_fill_obdo(caller_data, cmd, oa);
2091 oa->o_handle = lock->l_remote_handle;
2092 oa->o_valid |= OBD_MD_FLHANDLE;
2095 sort_brw_pages(pga, page_count);
2096 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count, pga, &req);
2098 CERROR("prep_req failed: %d\n", rc);
2099 GOTO(out, req = ERR_PTR(rc));
2101 oa = &((struct ost_body *)lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
2102 sizeof(struct ost_body)))->oa;
2104 /* Need to update the timestamps after the request is built in case
2105 * we race with setattr (locally or in queue at OST). If OST gets
2106 * later setattr before earlier BRW (as determined by the request xid),
2107 * the OST will not use BRW timestamps. Sadly, there is no obvious
2108 * way to do this in a single call. bug 10150 */
2109 if (pga[0]->flag & OBD_BRW_SRVLOCK) {
2110 /* in case of lockless read/write do not use inode's
2111 * timestamps because concurrent stat might fill the
2112 * inode with out-of-date times, send current
2114 if (cmd & OBD_BRW_WRITE) {
2115 oa->o_mtime = oa->o_ctime = LTIME_S(CURRENT_TIME);
2116 oa->o_valid |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
2117 valid = OBD_MD_FLATIME;
2119 oa->o_atime = LTIME_S(CURRENT_TIME);
2120 oa->o_valid |= OBD_MD_FLATIME;
2121 valid = OBD_MD_FLMTIME | OBD_MD_FLCTIME;
2124 valid = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2126 ops->ap_update_obdo(caller_data, cmd, oa, valid);
2128 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2129 aa = ptlrpc_req_async_args(req);
2130 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2131 list_splice(rpc_list, &aa->aa_oaps);
2132 CFS_INIT_LIST_HEAD(rpc_list);
2139 OBD_FREE(pga, sizeof(*pga) * page_count);
2144 /* the loi lock is held across this function but it's allowed to release
2145 * and reacquire it during its work */
2147 * prepare pages for ASYNC io and put pages in send queue.
2151 * \param cmd - OBD_BRW_* macroses
2152 * \param lop - pending pages
2154 * \return zero if pages successfully add to send queue.
2155 * \return not zere if error occurring.
2157 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
2158 int cmd, struct loi_oap_pages *lop)
2160 struct ptlrpc_request *req;
2161 obd_count page_count = 0;
2162 struct osc_async_page *oap = NULL, *tmp;
2163 struct osc_brw_async_args *aa;
2164 struct obd_async_page_ops *ops;
2165 CFS_LIST_HEAD(rpc_list);
2166 unsigned int ending_offset;
2167 unsigned starting_offset = 0;
2171 /* If there are HP OAPs we need to handle at least 1 of them,
2172 * move it the beginning of the pending list for that. */
2173 if (!list_empty(&lop->lop_urgent)) {
2174 oap = list_entry(lop->lop_urgent.next,
2175 struct osc_async_page, oap_urgent_item);
2176 if (oap->oap_async_flags & ASYNC_HP)
2177 list_move(&oap->oap_pending_item, &lop->lop_pending);
2180 /* first we find the pages we're allowed to work with */
2181 list_for_each_entry_safe(oap, tmp, &lop->lop_pending, oap_pending_item){
2182 ops = oap->oap_caller_ops;
2184 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2185 "magic 0x%x\n", oap, oap->oap_magic);
2187 if (page_count != 0 &&
2188 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2189 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2190 " oap %p, page %p, srvlock %u\n",
2191 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2194 /* in llite being 'ready' equates to the page being locked
2195 * until completion unlocks it. commit_write submits a page
2196 * as not ready because its unlock will happen unconditionally
2197 * as the call returns. if we race with commit_write giving
2198 * us that page we dont' want to create a hole in the page
2199 * stream, so we stop and leave the rpc to be fired by
2200 * another dirtier or kupdated interval (the not ready page
2201 * will still be on the dirty list). we could call in
2202 * at the end of ll_file_write to process the queue again. */
2203 if (!(oap->oap_async_flags & ASYNC_READY)) {
2204 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
2206 CDEBUG(D_INODE, "oap %p page %p returned %d "
2207 "instead of ready\n", oap,
2211 /* llite is telling us that the page is still
2212 * in commit_write and that we should try
2213 * and put it in an rpc again later. we
2214 * break out of the loop so we don't create
2215 * a hole in the sequence of pages in the rpc
2220 /* the io isn't needed.. tell the checks
2221 * below to complete the rpc with EINTR */
2222 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2223 oap->oap_count = -EINTR;
2226 oap->oap_async_flags |= ASYNC_READY;
2229 LASSERTF(0, "oap %p page %p returned %d "
2230 "from make_ready\n", oap,
2238 * Page submitted for IO has to be locked. Either by
2239 * ->ap_make_ready() or by higher layers.
2241 #if defined(__KERNEL__) && defined(__linux__)
2242 if(!(PageLocked(oap->oap_page) &&
2243 (CheckWriteback(oap->oap_page, cmd) || oap->oap_oig !=NULL))) {
2244 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2245 oap->oap_page, (long)oap->oap_page->flags, oap->oap_async_flags);
2249 /* If there is a gap at the start of this page, it can't merge
2250 * with any previous page, so we'll hand the network a
2251 * "fragmented" page array that it can't transfer in 1 RDMA */
2252 if (page_count != 0 && oap->oap_page_off != 0)
2255 /* take the page out of our book-keeping */
2256 list_del_init(&oap->oap_pending_item);
2257 lop_update_pending(cli, lop, cmd, -1);
2258 list_del_init(&oap->oap_urgent_item);
2260 if (page_count == 0)
2261 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2262 (PTLRPC_MAX_BRW_SIZE - 1);
2264 /* ask the caller for the size of the io as the rpc leaves. */
2265 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
2267 ops->ap_refresh_count(oap->oap_caller_data,cmd);
2268 if (oap->oap_count <= 0) {
2269 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2271 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
2275 /* now put the page back in our accounting */
2276 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2277 if (page_count == 0)
2278 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2279 if (++page_count >= cli->cl_max_pages_per_rpc)
2282 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2283 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2284 * have the same alignment as the initial writes that allocated
2285 * extents on the server. */
2286 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2287 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2288 if (ending_offset == 0)
2291 /* If there is a gap at the end of this page, it can't merge
2292 * with any subsequent pages, so we'll hand the network a
2293 * "fragmented" page array that it can't transfer in 1 RDMA */
2294 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2298 osc_wake_cache_waiters(cli);
2300 if (page_count == 0)
2303 loi_list_maint(cli, loi);
2305 client_obd_list_unlock(&cli->cl_loi_list_lock);
2307 req = osc_build_req(cli, &rpc_list, page_count, cmd);
2309 /* this should happen rarely and is pretty bad, it makes the
2310 * pending list not follow the dirty order */
2311 client_obd_list_lock(&cli->cl_loi_list_lock);
2312 list_for_each_entry_safe(oap, tmp, &rpc_list, oap_rpc_item) {
2313 list_del_init(&oap->oap_rpc_item);
2315 /* queued sync pages can be torn down while the pages
2316 * were between the pending list and the rpc */
2317 if (oap->oap_interrupted) {
2318 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2319 osc_ap_completion(cli, NULL, oap, 0,
2323 osc_ap_completion(cli, NULL, oap, 0, PTR_ERR(req));
2325 loi_list_maint(cli, loi);
2326 RETURN(PTR_ERR(req));
2329 aa = ptlrpc_req_async_args(req);
2330 if (cmd == OBD_BRW_READ) {
2331 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2332 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2333 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2334 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2336 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2337 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2338 cli->cl_w_in_flight);
2339 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2340 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2342 ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2344 client_obd_list_lock(&cli->cl_loi_list_lock);
2346 if (cmd == OBD_BRW_READ)
2347 cli->cl_r_in_flight++;
2349 cli->cl_w_in_flight++;
2351 /* queued sync pages can be torn down while the pages
2352 * were between the pending list and the rpc */
2354 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2355 /* only one oap gets a request reference */
2358 if (oap->oap_interrupted && !req->rq_intr) {
2359 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2361 ptlrpc_mark_interrupted(req);
2365 tmp->oap_request = ptlrpc_request_addref(req);
2367 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2368 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2370 req->rq_interpret_reply = brw_interpret;
2371 ptlrpcd_add_req(req);
2375 #define LOI_DEBUG(LOI, STR, args...) \
2376 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2377 !list_empty(&(LOI)->loi_ready_item) || \
2378 !list_empty(&(LOI)->loi_hp_ready_item), \
2379 (LOI)->loi_write_lop.lop_num_pending, \
2380 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2381 (LOI)->loi_read_lop.lop_num_pending, \
2382 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2385 /* This is called by osc_check_rpcs() to find which objects have pages that
2386 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2387 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2390 /* First return objects that have blocked locks so that they
2391 * will be flushed quickly and other clients can get the lock,
2392 * then objects which have pages ready to be stuffed into RPCs */
2393 if (!list_empty(&cli->cl_loi_hp_ready_list))
2394 RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
2395 struct lov_oinfo, loi_hp_ready_item));
2396 if (!list_empty(&cli->cl_loi_ready_list))
2397 RETURN(list_entry(cli->cl_loi_ready_list.next,
2398 struct lov_oinfo, loi_ready_item));
2400 /* then if we have cache waiters, return all objects with queued
2401 * writes. This is especially important when many small files
2402 * have filled up the cache and not been fired into rpcs because
2403 * they don't pass the nr_pending/object threshhold */
2404 if (!list_empty(&cli->cl_cache_waiters) &&
2405 !list_empty(&cli->cl_loi_write_list))
2406 RETURN(list_entry(cli->cl_loi_write_list.next,
2407 struct lov_oinfo, loi_write_item));
2409 /* then return all queued objects when we have an invalid import
2410 * so that they get flushed */
2411 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2412 if (!list_empty(&cli->cl_loi_write_list))
2413 RETURN(list_entry(cli->cl_loi_write_list.next,
2414 struct lov_oinfo, loi_write_item));
2415 if (!list_empty(&cli->cl_loi_read_list))
2416 RETURN(list_entry(cli->cl_loi_read_list.next,
2417 struct lov_oinfo, loi_read_item));
2422 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2424 struct osc_async_page *oap;
2427 if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
2428 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2429 struct osc_async_page, oap_urgent_item);
2430 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2433 if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
2434 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2435 struct osc_async_page, oap_urgent_item);
2436 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2439 return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2442 /* called with the loi list lock held */
2443 static void osc_check_rpcs(struct client_obd *cli)
2445 struct lov_oinfo *loi;
2446 int rc = 0, race_counter = 0;
2449 while ((loi = osc_next_loi(cli)) != NULL) {
2450 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2452 if (osc_max_rpc_in_flight(cli, loi))
2455 /* attempt some read/write balancing by alternating between
2456 * reads and writes in an object. The makes_rpc checks here
2457 * would be redundant if we were getting read/write work items
2458 * instead of objects. we don't want send_oap_rpc to drain a
2459 * partial read pending queue when we're given this object to
2460 * do io on writes while there are cache waiters */
2461 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2462 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
2463 &loi->loi_write_lop);
2471 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2472 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
2473 &loi->loi_read_lop);
2482 /* attempt some inter-object balancing by issueing rpcs
2483 * for each object in turn */
2484 if (!list_empty(&loi->loi_hp_ready_item))
2485 list_del_init(&loi->loi_hp_ready_item);
2486 if (!list_empty(&loi->loi_ready_item))
2487 list_del_init(&loi->loi_ready_item);
2488 if (!list_empty(&loi->loi_write_item))
2489 list_del_init(&loi->loi_write_item);
2490 if (!list_empty(&loi->loi_read_item))
2491 list_del_init(&loi->loi_read_item);
2493 loi_list_maint(cli, loi);
2495 /* send_oap_rpc fails with 0 when make_ready tells it to
2496 * back off. llite's make_ready does this when it tries
2497 * to lock a page queued for write that is already locked.
2498 * we want to try sending rpcs from many objects, but we
2499 * don't want to spin failing with 0. */
2500 if (race_counter == 10)
2506 /* we're trying to queue a page in the osc so we're subject to the
2507 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2508 * If the osc's queued pages are already at that limit, then we want to sleep
2509 * until there is space in the osc's queue for us. We also may be waiting for
2510 * write credits from the OST if there are RPCs in flight that may return some
2511 * before we fall back to sync writes.
2513 * We need this know our allocation was granted in the presence of signals */
2514 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2518 client_obd_list_lock(&cli->cl_loi_list_lock);
2519 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2520 client_obd_list_unlock(&cli->cl_loi_list_lock);
2524 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2525 * grant or cache space. */
2526 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
2527 struct osc_async_page *oap)
2529 struct osc_cache_waiter ocw;
2530 struct l_wait_info lwi = { 0 };
2533 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2534 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2535 cli->cl_dirty_max, obd_max_dirty_pages,
2536 cli->cl_lost_grant, cli->cl_avail_grant);
2538 /* force the caller to try sync io. this can jump the list
2539 * of queued writes and create a discontiguous rpc stream */
2540 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2541 loi->loi_ar.ar_force_sync)
2544 /* Hopefully normal case - cache space and write credits available */
2545 if ((cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max) &&
2546 (atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) &&
2547 (cli->cl_avail_grant >= CFS_PAGE_SIZE)) {
2548 /* account for ourselves */
2549 osc_consume_write_grant(cli, &oap->oap_brw_page);
2553 /* Make sure that there are write rpcs in flight to wait for. This
2554 * is a little silly as this object may not have any pending but
2555 * other objects sure might. */
2556 if (cli->cl_w_in_flight) {
2557 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2558 cfs_waitq_init(&ocw.ocw_waitq);
2562 loi_list_maint(cli, loi);
2563 osc_check_rpcs(cli);
2564 client_obd_list_unlock(&cli->cl_loi_list_lock);
2566 CDEBUG(D_CACHE, "sleeping for cache space\n");
2567 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2569 client_obd_list_lock(&cli->cl_loi_list_lock);
2570 if (!list_empty(&ocw.ocw_entry)) {
2571 list_del(&ocw.ocw_entry);
2580 static int osc_reget_short_lock(struct obd_export *exp,
2581 struct lov_stripe_md *lsm,
2583 obd_off start, obd_off end,
2586 struct osc_async_page *oap = *res;
2591 spin_lock(&oap->oap_lock);
2592 rc = ldlm_lock_fast_match(oap->oap_ldlm_lock, rw,
2593 start, end, cookie);
2594 spin_unlock(&oap->oap_lock);
2599 static int osc_release_short_lock(struct obd_export *exp,
2600 struct lov_stripe_md *lsm, obd_off end,
2601 void *cookie, int rw)
2604 ldlm_lock_fast_release(cookie, rw);
2605 /* no error could have happened at this layer */
2609 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2610 struct lov_oinfo *loi, cfs_page_t *page,
2611 obd_off offset, struct obd_async_page_ops *ops,
2612 void *data, void **res, int nocache,
2613 struct lustre_handle *lockh)
2615 struct osc_async_page *oap;
2616 struct ldlm_res_id oid = {{0}};
2622 return size_round(sizeof(*oap));
2625 oap->oap_magic = OAP_MAGIC;
2626 oap->oap_cli = &exp->exp_obd->u.cli;
2629 oap->oap_caller_ops = ops;
2630 oap->oap_caller_data = data;
2632 oap->oap_page = page;
2633 oap->oap_obj_off = offset;
2635 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2636 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2637 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2638 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2640 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2642 spin_lock_init(&oap->oap_lock);
2644 /* If the page was marked as notcacheable - don't add to any locks */
2646 osc_build_res_name(loi->loi_id, loi->loi_gr, &oid);
2647 /* This is the only place where we can call cache_add_extent
2648 without oap_lock, because this page is locked now, and
2649 the lock we are adding it to is referenced, so cannot lose
2650 any pages either. */
2651 rc = cache_add_extent(oap->oap_cli->cl_cache, &oid, oap, lockh);
2656 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2660 struct osc_async_page *oap_from_cookie(void *cookie)
2662 struct osc_async_page *oap = cookie;
2663 if (oap->oap_magic != OAP_MAGIC)
2664 return ERR_PTR(-EINVAL);
2668 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2669 struct lov_oinfo *loi, void *cookie,
2670 int cmd, obd_off off, int count,
2671 obd_flag brw_flags, enum async_flags async_flags)
2673 struct client_obd *cli = &exp->exp_obd->u.cli;
2674 struct osc_async_page *oap;
2678 oap = oap_from_cookie(cookie);
2680 RETURN(PTR_ERR(oap));
2682 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2685 if (!list_empty(&oap->oap_pending_item) ||
2686 !list_empty(&oap->oap_urgent_item) ||
2687 !list_empty(&oap->oap_rpc_item))
2690 /* check if the file's owner/group is over quota */
2691 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)){
2692 struct obd_async_page_ops *ops;
2699 ops = oap->oap_caller_ops;
2700 ops->ap_fill_obdo(oap->oap_caller_data, cmd, oa);
2701 if (lquota_chkdq(quota_interface, cli, oa->o_uid, oa->o_gid) ==
2711 loi = lsm->lsm_oinfo[0];
2713 client_obd_list_lock(&cli->cl_loi_list_lock);
2716 oap->oap_page_off = off;
2717 oap->oap_count = count;
2718 oap->oap_brw_flags = brw_flags;
2719 oap->oap_async_flags = async_flags;
2721 if (cmd & OBD_BRW_WRITE) {
2722 rc = osc_enter_cache(cli, loi, oap);
2724 client_obd_list_unlock(&cli->cl_loi_list_lock);
2729 osc_oap_to_pending(oap);
2730 loi_list_maint(cli, loi);
2732 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2735 osc_check_rpcs(cli);
2736 client_obd_list_unlock(&cli->cl_loi_list_lock);
2741 /* aka (~was & now & flag), but this is more clear :) */
2742 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2744 static int osc_set_async_flags(struct obd_export *exp,
2745 struct lov_stripe_md *lsm,
2746 struct lov_oinfo *loi, void *cookie,
2747 obd_flag async_flags)
2749 struct client_obd *cli = &exp->exp_obd->u.cli;
2750 struct loi_oap_pages *lop;
2751 struct osc_async_page *oap;
2755 oap = oap_from_cookie(cookie);
2757 RETURN(PTR_ERR(oap));
2760 * bug 7311: OST-side locking is only supported for liblustre for now
2761 * (and liblustre never calls obd_set_async_flags(). I hope.), generic
2762 * implementation has to handle case where OST-locked page was picked
2763 * up by, e.g., ->writepage().
2765 LASSERT(!(oap->oap_brw_flags & OBD_BRW_SRVLOCK));
2766 LASSERT(!LIBLUSTRE_CLIENT); /* check that liblustre angels do fear to
2769 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2773 loi = lsm->lsm_oinfo[0];
2775 if (oap->oap_cmd & OBD_BRW_WRITE) {
2776 lop = &loi->loi_write_lop;
2778 lop = &loi->loi_read_lop;
2781 client_obd_list_lock(&cli->cl_loi_list_lock);
2783 if (list_empty(&oap->oap_pending_item))
2784 GOTO(out, rc = -EINVAL);
2786 if ((oap->oap_async_flags & async_flags) == async_flags)
2789 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2790 oap->oap_async_flags |= ASYNC_READY;
2792 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
2793 list_empty(&oap->oap_rpc_item)) {
2794 if (oap->oap_async_flags & ASYNC_HP)
2795 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2797 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2798 oap->oap_async_flags |= ASYNC_URGENT;
2799 loi_list_maint(cli, loi);
2802 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2803 oap->oap_async_flags);
2805 osc_check_rpcs(cli);
2806 client_obd_list_unlock(&cli->cl_loi_list_lock);
2810 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2811 struct lov_oinfo *loi,
2812 struct obd_io_group *oig, void *cookie,
2813 int cmd, obd_off off, int count,
2815 obd_flag async_flags)
2817 struct client_obd *cli = &exp->exp_obd->u.cli;
2818 struct osc_async_page *oap;
2819 struct loi_oap_pages *lop;
2823 oap = oap_from_cookie(cookie);
2825 RETURN(PTR_ERR(oap));
2827 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2830 if (!list_empty(&oap->oap_pending_item) ||
2831 !list_empty(&oap->oap_urgent_item) ||
2832 !list_empty(&oap->oap_rpc_item))
2836 loi = lsm->lsm_oinfo[0];
2838 client_obd_list_lock(&cli->cl_loi_list_lock);
2841 oap->oap_page_off = off;
2842 oap->oap_count = count;
2843 oap->oap_brw_flags = brw_flags;
2844 oap->oap_async_flags = async_flags;
2846 if (cmd & OBD_BRW_WRITE)
2847 lop = &loi->loi_write_lop;
2849 lop = &loi->loi_read_lop;
2851 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2852 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2854 rc = oig_add_one(oig, &oap->oap_occ);
2857 LOI_DEBUG(loi, "oap %p page %p on group pending: rc %d\n",
2858 oap, oap->oap_page, rc);
2860 client_obd_list_unlock(&cli->cl_loi_list_lock);
2865 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2866 struct loi_oap_pages *lop, int cmd)
2868 struct list_head *pos, *tmp;
2869 struct osc_async_page *oap;
2871 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2872 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2873 list_del(&oap->oap_pending_item);
2874 osc_oap_to_pending(oap);
2876 loi_list_maint(cli, loi);
2879 static int osc_trigger_group_io(struct obd_export *exp,
2880 struct lov_stripe_md *lsm,
2881 struct lov_oinfo *loi,
2882 struct obd_io_group *oig)
2884 struct client_obd *cli = &exp->exp_obd->u.cli;
2888 loi = lsm->lsm_oinfo[0];
2890 client_obd_list_lock(&cli->cl_loi_list_lock);
2892 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2893 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2895 osc_check_rpcs(cli);
2896 client_obd_list_unlock(&cli->cl_loi_list_lock);
2901 static int osc_teardown_async_page(struct obd_export *exp,
2902 struct lov_stripe_md *lsm,
2903 struct lov_oinfo *loi, void *cookie)
2905 struct client_obd *cli = &exp->exp_obd->u.cli;
2906 struct loi_oap_pages *lop;
2907 struct osc_async_page *oap;
2911 oap = oap_from_cookie(cookie);
2913 RETURN(PTR_ERR(oap));
2916 loi = lsm->lsm_oinfo[0];
2918 if (oap->oap_cmd & OBD_BRW_WRITE) {
2919 lop = &loi->loi_write_lop;
2921 lop = &loi->loi_read_lop;
2924 client_obd_list_lock(&cli->cl_loi_list_lock);
2926 if (!list_empty(&oap->oap_rpc_item))
2927 GOTO(out, rc = -EBUSY);
2929 osc_exit_cache(cli, oap, 0);
2930 osc_wake_cache_waiters(cli);
2932 if (!list_empty(&oap->oap_urgent_item)) {
2933 list_del_init(&oap->oap_urgent_item);
2934 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
2937 if (!list_empty(&oap->oap_pending_item)) {
2938 list_del_init(&oap->oap_pending_item);
2939 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2941 loi_list_maint(cli, loi);
2942 cache_remove_extent(cli->cl_cache, oap);
2944 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2946 client_obd_list_unlock(&cli->cl_loi_list_lock);
2950 int osc_extent_blocking_cb(struct ldlm_lock *lock,
2951 struct ldlm_lock_desc *new, void *data,
2954 struct lustre_handle lockh = { 0 };
2958 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
2959 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
2964 case LDLM_CB_BLOCKING:
2965 ldlm_lock2handle(lock, &lockh);
2966 rc = ldlm_cli_cancel(&lockh);
2968 CERROR("ldlm_cli_cancel failed: %d\n", rc);
2970 case LDLM_CB_CANCELING: {
2972 ldlm_lock2handle(lock, &lockh);
2973 /* This lock wasn't granted, don't try to do anything */
2974 if (lock->l_req_mode != lock->l_granted_mode)
2977 cache_remove_lock(lock->l_conn_export->exp_obd->u.cli.cl_cache,
2980 if (lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb)
2981 lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb(
2982 lock, new, data,flag);
2991 EXPORT_SYMBOL(osc_extent_blocking_cb);
2993 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data,
2996 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2999 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
3002 lock_res_and_lock(lock);
3003 #if defined (__KERNEL__) && defined (__linux__)
3004 /* Liang XXX: Darwin and Winnt checking should be added */
3005 if (lock->l_ast_data && lock->l_ast_data != data) {
3006 struct inode *new_inode = data;
3007 struct inode *old_inode = lock->l_ast_data;
3008 if (!(old_inode->i_state & I_FREEING))
3009 LDLM_ERROR(lock, "inconsistent l_ast_data found");
3010 LASSERTF(old_inode->i_state & I_FREEING,
3011 "Found existing inode %p/%lu/%u state %lu in lock: "
3012 "setting data to %p/%lu/%u\n", old_inode,
3013 old_inode->i_ino, old_inode->i_generation,
3015 new_inode, new_inode->i_ino, new_inode->i_generation);
3018 lock->l_ast_data = data;
3019 lock->l_flags |= (flags & LDLM_FL_NO_LRU);
3020 unlock_res_and_lock(lock);
3021 LDLM_LOCK_PUT(lock);
3024 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3025 ldlm_iterator_t replace, void *data)
3027 struct ldlm_res_id res_id;
3028 struct obd_device *obd = class_exp2obd(exp);
3030 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3031 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3035 static int osc_enqueue_fini(struct obd_device *obd, struct ptlrpc_request *req,
3036 struct obd_info *oinfo, int intent, int rc)
3041 /* The request was created before ldlm_cli_enqueue call. */
3042 if (rc == ELDLM_LOCK_ABORTED) {
3043 struct ldlm_reply *rep;
3045 /* swabbed by ldlm_cli_enqueue() */
3046 LASSERT(lustre_rep_swabbed(req, DLM_LOCKREPLY_OFF));
3047 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
3049 LASSERT(rep != NULL);
3050 if (rep->lock_policy_res1)
3051 rc = rep->lock_policy_res1;
3055 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3056 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3057 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_size,
3058 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_blocks,
3059 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_mtime);
3063 cache_add_lock(obd->u.cli.cl_cache, oinfo->oi_lockh);
3065 /* Call the update callback. */
3066 rc = oinfo->oi_cb_up(oinfo, rc);
3070 static int osc_enqueue_interpret(struct ptlrpc_request *req,
3071 struct osc_enqueue_args *aa, int rc)
3073 int intent = aa->oa_oi->oi_flags & LDLM_FL_HAS_INTENT;
3074 struct lov_stripe_md *lsm = aa->oa_oi->oi_md;
3075 struct ldlm_lock *lock;
3077 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3079 lock = ldlm_handle2lock(aa->oa_oi->oi_lockh);
3081 /* Complete obtaining the lock procedure. */
3082 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3084 &aa->oa_oi->oi_flags,
3085 &lsm->lsm_oinfo[0]->loi_lvb,
3086 sizeof(lsm->lsm_oinfo[0]->loi_lvb),
3087 lustre_swab_ost_lvb,
3088 aa->oa_oi->oi_lockh, rc);
3090 /* Complete osc stuff. */
3091 rc = osc_enqueue_fini(aa->oa_exp->exp_obd, req, aa->oa_oi, intent, rc);
3093 /* Release the lock for async request. */
3094 if (lustre_handle_is_used(aa->oa_oi->oi_lockh) && rc == ELDLM_OK)
3095 ldlm_lock_decref(aa->oa_oi->oi_lockh, aa->oa_ei->ei_mode);
3097 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3098 aa->oa_oi->oi_lockh, req, aa);
3099 LDLM_LOCK_PUT(lock);
3103 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3104 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3105 * other synchronous requests, however keeping some locks and trying to obtain
3106 * others may take a considerable amount of time in a case of ost failure; and
3107 * when other sync requests do not get released lock from a client, the client
3108 * is excluded from the cluster -- such scenarious make the life difficult, so
3109 * release locks just after they are obtained. */
3110 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3111 struct ldlm_enqueue_info *einfo,
3112 struct ptlrpc_request_set *rqset)
3114 struct ldlm_res_id res_id;
3115 struct obd_device *obd = exp->exp_obd;
3116 struct ldlm_reply *rep;
3117 struct ptlrpc_request *req = NULL;
3118 int intent = oinfo->oi_flags & LDLM_FL_HAS_INTENT;
3123 osc_build_res_name(oinfo->oi_md->lsm_object_id,
3124 oinfo->oi_md->lsm_object_gr, &res_id);
3125 /* Filesystem lock extents are extended to page boundaries so that
3126 * dealing with the page cache is a little smoother. */
3127 oinfo->oi_policy.l_extent.start -=
3128 oinfo->oi_policy.l_extent.start & ~CFS_PAGE_MASK;
3129 oinfo->oi_policy.l_extent.end |= ~CFS_PAGE_MASK;
3131 if (oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid == 0)
3134 /* Next, search for already existing extent locks that will cover us */
3135 /* If we're trying to read, we also search for an existing PW lock. The
3136 * VFS and page cache already protect us locally, so lots of readers/
3137 * writers can share a single PW lock.
3139 * There are problems with conversion deadlocks, so instead of
3140 * converting a read lock to a write lock, we'll just enqueue a new
3143 * At some point we should cancel the read lock instead of making them
3144 * send us a blocking callback, but there are problems with canceling
3145 * locks out from other users right now, too. */
3146 mode = einfo->ei_mode;
3147 if (einfo->ei_mode == LCK_PR)
3149 mode = ldlm_lock_match(obd->obd_namespace,
3150 oinfo->oi_flags | LDLM_FL_LVB_READY, &res_id,
3151 einfo->ei_type, &oinfo->oi_policy, mode,
3154 /* addref the lock only if not async requests and PW lock is
3155 * matched whereas we asked for PR. */
3156 if (!rqset && einfo->ei_mode != mode)
3157 ldlm_lock_addref(oinfo->oi_lockh, LCK_PR);
3158 osc_set_data_with_check(oinfo->oi_lockh, einfo->ei_cbdata,
3161 /* I would like to be able to ASSERT here that rss <=
3162 * kms, but I can't, for reasons which are explained in
3166 /* We already have a lock, and it's referenced */
3167 oinfo->oi_cb_up(oinfo, ELDLM_OK);
3169 /* For async requests, decref the lock. */
3170 if (einfo->ei_mode != mode)
3171 ldlm_lock_decref(oinfo->oi_lockh, LCK_PW);
3173 ldlm_lock_decref(oinfo->oi_lockh, einfo->ei_mode);
3181 [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
3182 [DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request),
3183 [DLM_LOCKREQ_OFF + 1] = 0 };
3185 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
3189 size[DLM_LOCKREPLY_OFF] = sizeof(*rep);
3190 size[DLM_REPLY_REC_OFF] =
3191 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb);
3192 ptlrpc_req_set_repsize(req, 3, size);
3195 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3196 oinfo->oi_flags &= ~LDLM_FL_BLOCK_GRANTED;
3198 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id,
3199 &oinfo->oi_policy, &oinfo->oi_flags,
3200 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3201 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb),
3202 lustre_swab_ost_lvb, oinfo->oi_lockh,
3206 struct osc_enqueue_args *aa;
3207 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3208 aa = ptlrpc_req_async_args(req);
3213 req->rq_interpret_reply = osc_enqueue_interpret;
3214 ptlrpc_set_add_req(rqset, req);
3215 } else if (intent) {
3216 ptlrpc_req_finished(req);
3221 rc = osc_enqueue_fini(obd, req, oinfo, intent, rc);
3223 ptlrpc_req_finished(req);
3228 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
3229 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3230 int *flags, void *data, struct lustre_handle *lockh,
3233 struct ldlm_res_id res_id;
3234 struct obd_device *obd = exp->exp_obd;
3235 int lflags = *flags;
3239 osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3241 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
3243 /* Filesystem lock extents are extended to page boundaries so that
3244 * dealing with the page cache is a little smoother */
3245 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3246 policy->l_extent.end |= ~CFS_PAGE_MASK;
3248 /* Next, search for already existing extent locks that will cover us */
3249 /* If we're trying to read, we also search for an existing PW lock. The
3250 * VFS and page cache already protect us locally, so lots of readers/
3251 * writers can share a single PW lock. */
3255 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3256 &res_id, type, policy, rc, lockh);
3258 osc_set_data_with_check(lockh, data, lflags);
3259 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3260 ldlm_lock_addref(lockh, LCK_PR);
3261 ldlm_lock_decref(lockh, LCK_PW);
3263 if (n_matches != NULL)
3270 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3271 __u32 mode, struct lustre_handle *lockh)
3275 if (unlikely(mode == LCK_GROUP))
3276 ldlm_lock_decref_and_cancel(lockh, mode);
3278 ldlm_lock_decref(lockh, mode);
3283 static int osc_cancel_unused(struct obd_export *exp,
3284 struct lov_stripe_md *lsm, int flags, void *opaque)
3286 struct obd_device *obd = class_exp2obd(exp);
3287 struct ldlm_res_id res_id, *resp = NULL;
3290 resp = osc_build_res_name(lsm->lsm_object_id,
3291 lsm->lsm_object_gr, &res_id);
3294 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3298 static int osc_join_lru(struct obd_export *exp,
3299 struct lov_stripe_md *lsm, int join)
3301 struct obd_device *obd = class_exp2obd(exp);
3302 struct ldlm_res_id res_id, *resp = NULL;
3305 resp = osc_build_res_name(lsm->lsm_object_id,
3306 lsm->lsm_object_gr, &res_id);
3309 return ldlm_cli_join_lru(obd->obd_namespace, resp, join);
3313 static int osc_statfs_interpret(struct ptlrpc_request *req,
3314 struct osc_async_args *aa, int rc)
3316 struct obd_statfs *msfs;
3322 msfs = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*msfs),
3323 lustre_swab_obd_statfs);
3325 CERROR("Can't unpack obd_statfs\n");
3326 GOTO(out, rc = -EPROTO);
3329 memcpy(aa->aa_oi->oi_osfs, msfs, sizeof(*msfs));
3331 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3335 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3336 __u64 max_age, struct ptlrpc_request_set *rqset)
3338 struct ptlrpc_request *req;
3339 struct osc_async_args *aa;
3340 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*oinfo->oi_osfs) };
3343 /* We could possibly pass max_age in the request (as an absolute
3344 * timestamp or a "seconds.usec ago") so the target can avoid doing
3345 * extra calls into the filesystem if that isn't necessary (e.g.
3346 * during mount that would help a bit). Having relative timestamps
3347 * is not so great if request processing is slow, while absolute
3348 * timestamps are not ideal because they need time synchronization. */
3349 req = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OST_VERSION,
3350 OST_STATFS, 1, NULL, NULL);
3354 ptlrpc_req_set_repsize(req, 2, size);
3355 req->rq_request_portal = OST_CREATE_PORTAL;
3356 ptlrpc_at_set_req_timeout(req);
3357 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3358 /* procfs requests not want stat in wait for avoid deadlock */
3359 req->rq_no_resend = 1;
3360 req->rq_no_delay = 1;
3363 req->rq_interpret_reply = osc_statfs_interpret;
3364 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3365 aa = ptlrpc_req_async_args(req);
3368 ptlrpc_set_add_req(rqset, req);
3372 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3373 __u64 max_age, __u32 flags)
3375 struct obd_statfs *msfs;
3376 struct ptlrpc_request *req;
3377 struct obd_import *imp = NULL;
3378 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) };
3382 /*Since the request might also come from lprocfs, so we need
3383 *sync this with client_disconnect_export Bug15684*/
3384 down_read(&obd->u.cli.cl_sem);
3385 if (obd->u.cli.cl_import)
3386 imp = class_import_get(obd->u.cli.cl_import);
3387 up_read(&obd->u.cli.cl_sem);
3391 /* We could possibly pass max_age in the request (as an absolute
3392 * timestamp or a "seconds.usec ago") so the target can avoid doing
3393 * extra calls into the filesystem if that isn't necessary (e.g.
3394 * during mount that would help a bit). Having relative timestamps
3395 * is not so great if request processing is slow, while absolute
3396 * timestamps are not ideal because they need time synchronization. */
3397 req = ptlrpc_prep_req(imp, LUSTRE_OST_VERSION,
3398 OST_STATFS, 1, NULL, NULL);
3400 class_import_put(imp);
3404 ptlrpc_req_set_repsize(req, 2, size);
3405 req->rq_request_portal = OST_CREATE_PORTAL;
3406 ptlrpc_at_set_req_timeout(req);
3408 if (flags & OBD_STATFS_NODELAY) {
3409 /* procfs requests not want stat in wait for avoid deadlock */
3410 req->rq_no_resend = 1;
3411 req->rq_no_delay = 1;
3414 rc = ptlrpc_queue_wait(req);
3418 msfs = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*msfs),
3419 lustre_swab_obd_statfs);
3421 CERROR("Can't unpack obd_statfs\n");
3422 GOTO(out, rc = -EPROTO);
3425 memcpy(osfs, msfs, sizeof(*osfs));
3429 ptlrpc_req_finished(req);
3433 /* Retrieve object striping information.
3435 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3436 * the maximum number of OST indices which will fit in the user buffer.
3437 * lmm_magic must be LOV_MAGIC_V1 or LOV_MAGIC_V3 (we only use 1 slot here).
3439 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3441 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3442 struct lov_user_md_v3 lum, *lumk;
3443 int rc = 0, lum_size;
3444 struct lov_user_ost_data_v1 *lmm_objects;
3450 /* we only need the header part from user space to get lmm_magic and
3451 * lmm_stripe_count, (the header part is common to v1 and v3) */
3452 lum_size = sizeof(struct lov_user_md_v1);
3453 if (copy_from_user(&lum, lump, lum_size))
3456 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3457 (lum.lmm_magic != LOV_USER_MAGIC_V3))
3460 /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3461 LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3462 LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3463 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3465 /* we can use lov_mds_md_size() to compute lum_size
3466 * because lov_user_md_vX and lov_mds_md_vX have the same size */
3467 if (lum.lmm_stripe_count > 0) {
3468 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3469 OBD_ALLOC(lumk, lum_size);
3472 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3473 lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3475 lmm_objects = &(lumk->lmm_objects[0]);
3476 lmm_objects->l_object_id = lsm->lsm_object_id;
3478 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3482 lumk->lmm_object_id = lsm->lsm_object_id;
3483 lumk->lmm_stripe_count = 1;
3485 if (copy_to_user(lump, lumk, lum_size))
3489 OBD_FREE(lumk, lum_size);
3495 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3496 void *karg, void *uarg)
3498 struct obd_device *obd = exp->exp_obd;
3499 struct obd_ioctl_data *data = karg;
3503 if (!try_module_get(THIS_MODULE)) {
3504 CERROR("Can't get module. Is it alive?");
3508 case OBD_IOC_LOV_GET_CONFIG: {
3510 struct lov_desc *desc;
3511 struct obd_uuid uuid;
3515 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3516 GOTO(out, err = -EINVAL);
3518 data = (struct obd_ioctl_data *)buf;
3520 if (sizeof(*desc) > data->ioc_inllen1) {
3521 obd_ioctl_freedata(buf, len);
3522 GOTO(out, err = -EINVAL);
3525 if (data->ioc_inllen2 < sizeof(uuid)) {
3526 obd_ioctl_freedata(buf, len);
3527 GOTO(out, err = -EINVAL);
3530 desc = (struct lov_desc *)data->ioc_inlbuf1;
3531 desc->ld_tgt_count = 1;
3532 desc->ld_active_tgt_count = 1;
3533 desc->ld_default_stripe_count = 1;
3534 desc->ld_default_stripe_size = 0;
3535 desc->ld_default_stripe_offset = 0;
3536 desc->ld_pattern = 0;
3537 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3539 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3541 err = copy_to_user((void *)uarg, buf, len);
3544 obd_ioctl_freedata(buf, len);
3547 case LL_IOC_LOV_SETSTRIPE:
3548 err = obd_alloc_memmd(exp, karg);
3552 case LL_IOC_LOV_GETSTRIPE:
3553 err = osc_getstripe(karg, uarg);
3555 case OBD_IOC_CLIENT_RECOVER:
3556 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3561 case IOC_OSC_SET_ACTIVE:
3562 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3565 case OBD_IOC_POLL_QUOTACHECK:
3566 err = lquota_poll_check(quota_interface, exp,
3567 (struct if_quotacheck *)karg);
3569 case OBD_IOC_DESTROY: {
3572 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
3573 GOTO (out, err = -EPERM);
3574 oa = &data->ioc_obdo1;
3577 GOTO(out, err = -EINVAL);
3579 oa->o_valid |= OBD_MD_FLGROUP;
3581 err = osc_destroy(exp, oa, NULL, NULL, NULL);
3584 case OBD_IOC_PING_TARGET:
3585 err = ptlrpc_obd_ping(obd);
3588 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3589 cmd, cfs_curproc_comm());
3590 GOTO(out, err = -ENOTTY);
3593 module_put(THIS_MODULE);
3597 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3598 void *key, __u32 *vallen, void *val, struct lov_stripe_md *lsm)
3601 if (!vallen || !val)
3604 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3605 __u32 *stripe = val;
3606 *vallen = sizeof(*stripe);
3609 } else if (KEY_IS(KEY_LAST_ID)) {
3610 struct ptlrpc_request *req;
3612 char *bufs[2] = { NULL, key };
3613 __u32 size[2] = { sizeof(struct ptlrpc_body), keylen };
3616 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
3617 OST_GET_INFO, 2, size, bufs);
3621 size[REPLY_REC_OFF] = *vallen;
3622 ptlrpc_req_set_repsize(req, 2, size);
3623 rc = ptlrpc_queue_wait(req);
3627 reply = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*reply),
3628 lustre_swab_ost_last_id);
3629 if (reply == NULL) {
3630 CERROR("Can't unpack OST last ID\n");
3631 GOTO(out, rc = -EPROTO);
3633 *((obd_id *)val) = *reply;
3635 ptlrpc_req_finished(req);
3637 } else if (KEY_IS(KEY_FIEMAP)) {
3638 struct ptlrpc_request *req;
3639 struct ll_user_fiemap *reply;
3640 char *bufs[2] = { NULL, key };
3641 __u32 size[2] = { sizeof(struct ptlrpc_body), keylen };
3644 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
3645 OST_GET_INFO, 2, size, bufs);
3649 size[REPLY_REC_OFF] = *vallen;
3650 ptlrpc_req_set_repsize(req, 2, size);
3652 rc = ptlrpc_queue_wait(req);
3655 reply = lustre_swab_repbuf(req, REPLY_REC_OFF, *vallen,
3656 lustre_swab_fiemap);
3657 if (reply == NULL) {
3658 CERROR("Can't unpack FIEMAP reply.\n");
3659 GOTO(out1, rc = -EPROTO);
3662 memcpy(val, reply, *vallen);
3665 ptlrpc_req_finished(req);
3673 static int osc_setinfo_mds_conn_interpret(struct ptlrpc_request *req,
3676 struct llog_ctxt *ctxt;
3677 struct obd_import *imp = req->rq_import;
3683 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3686 rc = llog_initiator_connect(ctxt);
3688 CERROR("cannot establish connection for "
3689 "ctxt %p: %d\n", ctxt, rc);
3692 llog_ctxt_put(ctxt);
3693 spin_lock(&imp->imp_lock);
3694 imp->imp_server_timeout = 1;
3695 imp->imp_pingable = 1;
3696 spin_unlock(&imp->imp_lock);
3697 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3702 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3703 void *key, obd_count vallen, void *val,
3704 struct ptlrpc_request_set *set)
3706 struct ptlrpc_request *req;
3707 struct obd_device *obd = exp->exp_obd;
3708 struct obd_import *imp = class_exp2cliimp(exp);
3709 __u32 size[3] = { sizeof(struct ptlrpc_body), keylen, vallen };
3710 char *bufs[3] = { NULL, key, val };
3713 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3715 if (KEY_IS(KEY_NEXT_ID)) {
3716 if (vallen != sizeof(obd_id))
3718 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3719 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3720 exp->exp_obd->obd_name,
3721 obd->u.cli.cl_oscc.oscc_next_id);
3726 if (KEY_IS(KEY_UNLINKED)) {
3727 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3728 spin_lock(&oscc->oscc_lock);
3729 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3730 spin_unlock(&oscc->oscc_lock);
3734 if (KEY_IS(KEY_INIT_RECOV)) {
3735 if (vallen != sizeof(int))
3737 spin_lock(&imp->imp_lock);
3738 imp->imp_initial_recov = *(int *)val;
3739 spin_unlock(&imp->imp_lock);
3740 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3741 exp->exp_obd->obd_name,
3742 imp->imp_initial_recov);
3746 if (KEY_IS(KEY_CHECKSUM)) {
3747 if (vallen != sizeof(int))
3749 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3756 /* We pass all other commands directly to OST. Since nobody calls osc
3757 methods directly and everybody is supposed to go through LOV, we
3758 assume lov checked invalid values for us.
3759 The only recognised values so far are evict_by_nid and mds_conn.
3760 Even if something bad goes through, we'd get a -EINVAL from OST
3763 req = ptlrpc_prep_req(imp, LUSTRE_OST_VERSION, OST_SET_INFO, 3, size,
3768 if (KEY_IS(KEY_MDS_CONN))
3769 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3771 ptlrpc_req_set_repsize(req, 1, NULL);
3772 ptlrpc_set_add_req(set, req);
3773 ptlrpc_check_set(set);
3779 static struct llog_operations osc_size_repl_logops = {
3780 lop_cancel: llog_obd_repl_cancel
3783 static struct llog_operations osc_mds_ost_orig_logops;
3784 static int osc_llog_init(struct obd_device *obd, struct obd_device *tgt,
3785 int count, struct llog_catid *catid,
3786 struct obd_uuid *uuid)
3791 spin_lock(&obd->obd_dev_lock);
3792 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3793 osc_mds_ost_orig_logops = llog_lvfs_ops;
3794 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3795 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3796 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3797 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3799 spin_unlock(&obd->obd_dev_lock);
3801 rc = llog_setup(obd, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3802 &catid->lci_logid, &osc_mds_ost_orig_logops);
3804 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3808 rc = llog_setup(obd, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3809 &osc_size_repl_logops);
3811 struct llog_ctxt *ctxt =
3812 llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3815 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3819 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3820 obd->obd_name, tgt->obd_name, count, catid, rc);
3821 CERROR("logid "LPX64":0x%x\n",
3822 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3827 static int osc_llog_finish(struct obd_device *obd, int count)
3829 struct llog_ctxt *ctxt;
3830 int rc = 0, rc2 = 0;
3833 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3835 rc = llog_cleanup(ctxt);
3837 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3839 rc2 = llog_cleanup(ctxt);
3846 static int osc_reconnect(struct obd_export *exp, struct obd_device *obd,
3847 struct obd_uuid *cluuid,
3848 struct obd_connect_data *data,
3851 struct client_obd *cli = &obd->u.cli;
3853 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3856 client_obd_list_lock(&cli->cl_loi_list_lock);
3857 data->ocd_grant = cli->cl_avail_grant ?:
3858 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3859 lost_grant = cli->cl_lost_grant;
3860 cli->cl_lost_grant = 0;
3861 client_obd_list_unlock(&cli->cl_loi_list_lock);
3863 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3864 "cl_lost_grant: %ld\n", data->ocd_grant,
3865 cli->cl_avail_grant, lost_grant);
3866 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3867 " ocd_grant: %d\n", data->ocd_connect_flags,
3868 data->ocd_version, data->ocd_grant);
3874 static int osc_disconnect(struct obd_export *exp)
3876 struct obd_device *obd = class_exp2obd(exp);
3877 struct llog_ctxt *ctxt;
3880 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3882 if (obd->u.cli.cl_conn_count == 1) {
3883 /* Flush any remaining cancel messages out to the
3885 llog_sync(ctxt, exp);
3887 llog_ctxt_put(ctxt);
3889 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3893 rc = client_disconnect_export(exp);
3897 static int osc_import_event(struct obd_device *obd,
3898 struct obd_import *imp,
3899 enum obd_import_event event)
3901 struct client_obd *cli;
3905 LASSERT(imp->imp_obd == obd);
3908 case IMP_EVENT_DISCON: {
3909 /* Only do this on the MDS OSC's */
3910 if (imp->imp_server_timeout) {
3911 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3913 spin_lock(&oscc->oscc_lock);
3914 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3915 spin_unlock(&oscc->oscc_lock);
3918 client_obd_list_lock(&cli->cl_loi_list_lock);
3919 cli->cl_avail_grant = 0;
3920 cli->cl_lost_grant = 0;
3921 client_obd_list_unlock(&cli->cl_loi_list_lock);
3922 ptlrpc_import_setasync(imp, -1);
3926 case IMP_EVENT_INACTIVE: {
3927 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3930 case IMP_EVENT_INVALIDATE: {
3931 struct ldlm_namespace *ns = obd->obd_namespace;
3935 client_obd_list_lock(&cli->cl_loi_list_lock);
3936 /* all pages go to failing rpcs due to the invalid import */
3937 osc_check_rpcs(cli);
3938 client_obd_list_unlock(&cli->cl_loi_list_lock);
3940 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3944 case IMP_EVENT_ACTIVE: {
3945 /* Only do this on the MDS OSC's */
3946 if (imp->imp_server_timeout) {
3947 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3949 spin_lock(&oscc->oscc_lock);
3950 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3951 spin_unlock(&oscc->oscc_lock);
3953 CDEBUG(D_INFO, "notify server \n");
3954 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3957 case IMP_EVENT_OCD: {
3958 struct obd_connect_data *ocd = &imp->imp_connect_data;
3960 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3961 osc_init_grant(&obd->u.cli, ocd);
3964 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3965 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3967 ptlrpc_import_setasync(imp, 1);
3968 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3972 CERROR("Unknown import event %d\n", event);
3978 int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3984 rc = ptlrpcd_addref();
3988 rc = client_obd_setup(obd, len, buf);
3992 struct lprocfs_static_vars lvars = { 0 };
3993 struct client_obd *cli = &obd->u.cli;
3995 lprocfs_osc_init_vars(&lvars);
3996 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3997 lproc_osc_attach_seqstat(obd);
3998 ptlrpc_lprocfs_register_obd(obd);
4002 /* We need to allocate a few requests more, because
4003 brw_interpret tries to create new requests before freeing
4004 previous ones. Ideally we want to have 2x max_rpcs_in_flight
4005 reserved, but I afraid that might be too much wasted RAM
4006 in fact, so 2 is just my guess and still should work. */
4007 cli->cl_import->imp_rq_pool =
4008 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4010 ptlrpc_add_rqs_to_pool);
4011 cli->cl_cache = cache_create(obd);
4012 if (!cli->cl_cache) {
4021 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4027 case OBD_CLEANUP_EARLY: {
4028 struct obd_import *imp;
4029 imp = obd->u.cli.cl_import;
4030 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4031 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4032 ptlrpc_deactivate_import(imp);
4035 case OBD_CLEANUP_EXPORTS: {
4036 /* If we set up but never connected, the
4037 client import will not have been cleaned. */
4038 if (obd->u.cli.cl_import) {
4039 struct obd_import *imp;
4040 down_write(&obd->u.cli.cl_sem);
4041 imp = obd->u.cli.cl_import;
4042 CDEBUG(D_CONFIG, "%s: client import never connected\n",
4044 ptlrpc_invalidate_import(imp);
4045 if (imp->imp_rq_pool) {
4046 ptlrpc_free_rq_pool(imp->imp_rq_pool);
4047 imp->imp_rq_pool = NULL;
4049 class_destroy_import(imp);
4050 up_write(&obd->u.cli.cl_sem);
4051 obd->u.cli.cl_import = NULL;
4053 rc = obd_llog_finish(obd, 0);
4055 CERROR("failed to cleanup llogging subsystems\n");
4058 case OBD_CLEANUP_SELF_EXP:
4060 case OBD_CLEANUP_OBD:
4066 int osc_cleanup(struct obd_device *obd)
4068 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4072 ptlrpc_lprocfs_unregister_obd(obd);
4073 lprocfs_obd_cleanup(obd);
4075 spin_lock(&oscc->oscc_lock);
4076 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
4077 oscc->oscc_flags |= OSCC_FLAG_EXITING;
4078 spin_unlock(&oscc->oscc_lock);
4080 /* free memory of osc quota cache */
4081 lquota_cleanup(quota_interface, obd);
4083 cache_destroy(obd->u.cli.cl_cache);
4084 rc = client_obd_cleanup(obd);
4090 static int osc_register_page_removal_cb(struct obd_device *obd,
4091 obd_page_removal_cb_t func,
4092 obd_pin_extent_cb pin_cb)
4096 /* this server - not need init */
4100 return cache_add_extent_removal_cb(obd->u.cli.cl_cache, func,
4104 static int osc_unregister_page_removal_cb(struct obd_device *obd,
4105 obd_page_removal_cb_t func)
4108 return cache_del_extent_removal_cb(obd->u.cli.cl_cache, func);
4111 static int osc_register_lock_cancel_cb(struct obd_device *obd,
4112 obd_lock_cancel_cb cb)
4115 LASSERT(obd->u.cli.cl_ext_lock_cancel_cb == NULL);
4117 /* this server - not need init */
4121 obd->u.cli.cl_ext_lock_cancel_cb = cb;
4125 static int osc_unregister_lock_cancel_cb(struct obd_device *obd,
4126 obd_lock_cancel_cb cb)
4130 if (obd->u.cli.cl_ext_lock_cancel_cb != cb) {
4131 CERROR("Unregistering cancel cb %p, while only %p was "
4133 obd->u.cli.cl_ext_lock_cancel_cb);
4137 obd->u.cli.cl_ext_lock_cancel_cb = NULL;
4141 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4143 struct lustre_cfg *lcfg = buf;
4144 struct lprocfs_static_vars lvars = { 0 };
4147 lprocfs_osc_init_vars(&lvars);
4149 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars, lcfg, obd);
4153 struct obd_ops osc_obd_ops = {
4154 .o_owner = THIS_MODULE,
4155 .o_setup = osc_setup,
4156 .o_precleanup = osc_precleanup,
4157 .o_cleanup = osc_cleanup,
4158 .o_add_conn = client_import_add_conn,
4159 .o_del_conn = client_import_del_conn,
4160 .o_connect = client_connect_import,
4161 .o_reconnect = osc_reconnect,
4162 .o_disconnect = osc_disconnect,
4163 .o_statfs = osc_statfs,
4164 .o_statfs_async = osc_statfs_async,
4165 .o_packmd = osc_packmd,
4166 .o_unpackmd = osc_unpackmd,
4167 .o_precreate = osc_precreate,
4168 .o_create = osc_create,
4169 .o_destroy = osc_destroy,
4170 .o_getattr = osc_getattr,
4171 .o_getattr_async = osc_getattr_async,
4172 .o_setattr = osc_setattr,
4173 .o_setattr_async = osc_setattr_async,
4175 .o_brw_async = osc_brw_async,
4176 .o_prep_async_page = osc_prep_async_page,
4177 .o_reget_short_lock = osc_reget_short_lock,
4178 .o_release_short_lock = osc_release_short_lock,
4179 .o_queue_async_io = osc_queue_async_io,
4180 .o_set_async_flags = osc_set_async_flags,
4181 .o_queue_group_io = osc_queue_group_io,
4182 .o_trigger_group_io = osc_trigger_group_io,
4183 .o_teardown_async_page = osc_teardown_async_page,
4184 .o_punch = osc_punch,
4186 .o_enqueue = osc_enqueue,
4187 .o_match = osc_match,
4188 .o_change_cbdata = osc_change_cbdata,
4189 .o_cancel = osc_cancel,
4190 .o_cancel_unused = osc_cancel_unused,
4191 .o_join_lru = osc_join_lru,
4192 .o_iocontrol = osc_iocontrol,
4193 .o_get_info = osc_get_info,
4194 .o_set_info_async = osc_set_info_async,
4195 .o_import_event = osc_import_event,
4196 .o_llog_init = osc_llog_init,
4197 .o_llog_finish = osc_llog_finish,
4198 .o_process_config = osc_process_config,
4199 .o_register_page_removal_cb = osc_register_page_removal_cb,
4200 .o_unregister_page_removal_cb = osc_unregister_page_removal_cb,
4201 .o_register_lock_cancel_cb = osc_register_lock_cancel_cb,
4202 .o_unregister_lock_cancel_cb = osc_unregister_lock_cancel_cb,
4204 int __init osc_init(void)
4206 struct lprocfs_static_vars lvars = { 0 };
4210 lprocfs_osc_init_vars(&lvars);
4212 request_module("lquota");
4213 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4214 lquota_init(quota_interface);
4215 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4217 rc = class_register_type(&osc_obd_ops, lvars.module_vars,
4220 if (quota_interface)
4221 PORTAL_SYMBOL_PUT(osc_quota_interface);
4229 static void /*__exit*/ osc_exit(void)
4231 lquota_exit(quota_interface);
4232 if (quota_interface)
4233 PORTAL_SYMBOL_PUT(osc_quota_interface);
4235 class_unregister_type(LUSTRE_OSC_NAME);
4238 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4239 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4240 MODULE_LICENSE("GPL");
4242 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);