1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
5 * Author Peter Braam <braam@clusterfs.com>
7 * This file is part of the Lustre file system, http://www.lustre.org
8 * Lustre is a trademark of Cluster File Systems, Inc.
10 * You may have signed or agreed to another license before downloading
11 * this software. If so, you are bound by the terms and conditions
12 * of that agreement, and the following does not apply to you. See the
13 * LICENSE file included with this distribution for more information.
15 * If you did not agree to a different license, then this copy of Lustre
16 * is open source software; you can redistribute it and/or modify it
17 * under the terms of version 2 of the GNU General Public License as
18 * published by the Free Software Foundation.
20 * In either case, Lustre is distributed in the hope that it will be
21 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
22 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 * license text for more details.
25 * For testing and management it is treated as an obd_device,
26 * although * it does not export a full OBD method table (the
27 * requests are coming * in over the wire, so object target modules
28 * do not have a full * method table.)
33 # define EXPORT_SYMTAB
35 #define DEBUG_SUBSYSTEM S_OSC
38 # include <libcfs/libcfs.h>
39 #else /* __KERNEL__ */
40 # include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <libcfs/kp30.h>
45 #include <lustre_net.h>
46 #include <lustre/lustre_user.h>
47 #include <obd_cksum.h>
55 #include <lustre_ha.h>
56 #include <lprocfs_status.h>
57 #include <lustre_log.h>
58 #include <lustre_debug.h>
59 #include <lustre_param.h>
60 #include <lustre_cache.h>
61 #include "osc_internal.h"
63 static quota_interface_t *quota_interface = NULL;
64 extern quota_interface_t osc_quota_interface;
66 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
67 int osc_cleanup(struct obd_device *obd);
69 /* Pack OSC object metadata for disk storage (LE byte order). */
70 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
71 struct lov_stripe_md *lsm)
76 lmm_size = sizeof(**lmmp);
81 OBD_FREE(*lmmp, lmm_size);
87 OBD_ALLOC(*lmmp, lmm_size);
93 LASSERT(lsm->lsm_object_id);
94 LASSERT(lsm->lsm_object_gr);
95 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
96 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
102 /* Unpack OSC object metadata from disk storage (LE byte order). */
103 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
104 struct lov_mds_md *lmm, int lmm_bytes)
110 if (lmm_bytes < sizeof (*lmm)) {
111 CERROR("lov_mds_md too small: %d, need %d\n",
112 lmm_bytes, (int)sizeof(*lmm));
115 /* XXX LOV_MAGIC etc check? */
117 if (lmm->lmm_object_id == 0) {
118 CERROR("lov_mds_md: zero lmm_object_id\n");
123 lsm_size = lov_stripe_md_size(1);
127 if (*lsmp != NULL && lmm == NULL) {
128 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
129 OBD_FREE(*lsmp, lsm_size);
135 OBD_ALLOC(*lsmp, lsm_size);
138 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
139 if ((*lsmp)->lsm_oinfo[0] == NULL) {
140 OBD_FREE(*lsmp, lsm_size);
143 loi_init((*lsmp)->lsm_oinfo[0]);
147 /* XXX zero *lsmp? */
148 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
149 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
150 LASSERT((*lsmp)->lsm_object_id);
151 LASSERT((*lsmp)->lsm_object_gr);
154 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
159 static inline void osc_pack_capa(struct ptlrpc_request *req,
160 struct ost_body *body, void *capa)
162 struct obd_capa *oc = (struct obd_capa *)capa;
163 struct lustre_capa *c;
168 c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
171 body->oa.o_valid |= OBD_MD_FLOSSCAPA;
172 DEBUG_CAPA(D_SEC, c, "pack");
175 static inline void osc_pack_req_body(struct ptlrpc_request *req,
176 struct obd_info *oinfo)
178 struct ost_body *body;
180 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
183 body->oa = *oinfo->oi_oa;
184 osc_pack_capa(req, body, oinfo->oi_capa);
187 static inline void osc_set_capa_size(struct ptlrpc_request *req,
188 const struct req_msg_field *field,
192 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
194 /* it is already calculated as sizeof struct obd_capa */
198 static int osc_getattr_interpret(struct ptlrpc_request *req,
199 struct osc_async_args *aa, int rc)
201 struct ost_body *body;
207 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
208 lustre_swab_ost_body);
210 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
211 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
213 /* This should really be sent by the OST */
214 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
215 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
217 CDEBUG(D_INFO, "can't unpack ost_body\n");
219 aa->aa_oi->oi_oa->o_valid = 0;
222 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
226 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
227 struct ptlrpc_request_set *set)
229 struct ptlrpc_request *req;
230 struct osc_async_args *aa;
234 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
238 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
239 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
241 ptlrpc_request_free(req);
245 osc_pack_req_body(req, oinfo);
247 ptlrpc_request_set_replen(req);
248 req->rq_interpret_reply = osc_getattr_interpret;
250 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
251 aa = (struct osc_async_args *)&req->rq_async_args;
254 ptlrpc_set_add_req(set, req);
258 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
260 struct ptlrpc_request *req;
261 struct ost_body *body;
265 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
269 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
270 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
272 ptlrpc_request_free(req);
276 osc_pack_req_body(req, oinfo);
278 ptlrpc_request_set_replen(req);
280 rc = ptlrpc_queue_wait(req);
284 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
286 GOTO(out, rc = -EPROTO);
288 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
289 *oinfo->oi_oa = body->oa;
291 /* This should really be sent by the OST */
292 oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
293 oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
297 ptlrpc_req_finished(req);
301 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
302 struct obd_trans_info *oti)
304 struct ptlrpc_request *req;
305 struct ost_body *body;
309 LASSERT(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
310 oinfo->oi_oa->o_gr > 0);
312 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
316 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
317 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
319 ptlrpc_request_free(req);
323 osc_pack_req_body(req, oinfo);
325 ptlrpc_request_set_replen(req);
328 rc = ptlrpc_queue_wait(req);
332 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
334 GOTO(out, rc = -EPROTO);
336 *oinfo->oi_oa = body->oa;
340 ptlrpc_req_finished(req);
344 static int osc_setattr_interpret(struct ptlrpc_request *req,
345 struct osc_async_args *aa, int rc)
347 struct ost_body *body;
353 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
355 GOTO(out, rc = -EPROTO);
357 *aa->aa_oi->oi_oa = body->oa;
359 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
363 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
364 struct obd_trans_info *oti,
365 struct ptlrpc_request_set *rqset)
367 struct ptlrpc_request *req;
368 struct osc_async_args *aa;
372 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
376 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
377 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
379 ptlrpc_request_free(req);
383 osc_pack_req_body(req, oinfo);
385 ptlrpc_request_set_replen(req);
387 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
389 *obdo_logcookie(oinfo->oi_oa) = *oti->oti_logcookies;
392 /* do mds to ost setattr asynchronouly */
394 /* Do not wait for response. */
395 ptlrpcd_add_req(req);
397 req->rq_interpret_reply = osc_setattr_interpret;
399 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
400 aa = (struct osc_async_args *)&req->rq_async_args;
403 ptlrpc_set_add_req(rqset, req);
409 int osc_real_create(struct obd_export *exp, struct obdo *oa,
410 struct lov_stripe_md **ea, struct obd_trans_info *oti)
412 struct ptlrpc_request *req;
413 struct ost_body *body;
414 struct lov_stripe_md *lsm;
423 rc = obd_alloc_memmd(exp, &lsm);
428 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
430 GOTO(out, rc = -ENOMEM);
432 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
434 ptlrpc_request_free(req);
438 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
442 ptlrpc_request_set_replen(req);
444 if (oa->o_valid & OBD_MD_FLINLINE) {
445 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
446 oa->o_flags == OBD_FL_DELORPHAN);
448 "delorphan from OST integration");
449 /* Don't resend the delorphan req */
450 req->rq_no_resend = req->rq_no_delay = 1;
453 rc = ptlrpc_queue_wait(req);
457 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
459 GOTO(out_req, rc = -EPROTO);
463 /* This should really be sent by the OST */
464 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
465 oa->o_valid |= OBD_MD_FLBLKSZ;
467 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
468 * have valid lsm_oinfo data structs, so don't go touching that.
469 * This needs to be fixed in a big way.
471 lsm->lsm_object_id = oa->o_id;
472 lsm->lsm_object_gr = oa->o_gr;
476 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
478 if (oa->o_valid & OBD_MD_FLCOOKIE) {
479 if (!oti->oti_logcookies)
480 oti_alloc_cookies(oti, 1);
481 *oti->oti_logcookies = *obdo_logcookie(oa);
485 CDEBUG(D_HA, "transno: "LPD64"\n",
486 lustre_msg_get_transno(req->rq_repmsg));
488 ptlrpc_req_finished(req);
491 obd_free_memmd(exp, &lsm);
495 static int osc_punch_interpret(struct ptlrpc_request *req,
496 struct osc_async_args *aa, int rc)
498 struct ost_body *body;
504 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
506 GOTO(out, rc = -EPROTO);
508 *aa->aa_oi->oi_oa = body->oa;
510 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
514 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
515 struct obd_trans_info *oti,
516 struct ptlrpc_request_set *rqset)
518 struct ptlrpc_request *req;
519 struct osc_async_args *aa;
520 struct ost_body *body;
525 CDEBUG(D_INFO, "oa NULL\n");
529 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
533 osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
534 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
536 ptlrpc_request_free(req);
539 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
540 osc_pack_req_body(req, oinfo);
542 /* overload the size and blocks fields in the oa with start/end */
543 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
545 body->oa.o_size = oinfo->oi_policy.l_extent.start;
546 body->oa.o_blocks = oinfo->oi_policy.l_extent.end;
547 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
548 ptlrpc_request_set_replen(req);
551 req->rq_interpret_reply = osc_punch_interpret;
552 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
553 aa = (struct osc_async_args *)&req->rq_async_args;
555 ptlrpc_set_add_req(rqset, req);
560 static int osc_sync(struct obd_export *exp, struct obdo *oa,
561 struct lov_stripe_md *md, obd_size start, obd_size end,
564 struct ptlrpc_request *req;
565 struct ost_body *body;
570 CDEBUG(D_INFO, "oa NULL\n");
574 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
578 osc_set_capa_size(req, &RMF_CAPA1, capa);
579 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
581 ptlrpc_request_free(req);
585 /* overload the size and blocks fields in the oa with start/end */
586 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
589 body->oa.o_size = start;
590 body->oa.o_blocks = end;
591 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
592 osc_pack_capa(req, body, capa);
594 ptlrpc_request_set_replen(req);
596 rc = ptlrpc_queue_wait(req);
600 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
602 GOTO(out, rc = -EPROTO);
608 ptlrpc_req_finished(req);
612 /* Find and cancel locally locks matched by @mode in the resource found by
613 * @objid. Found locks are added into @cancel list. Returns the amount of
614 * locks added to @cancels list. */
615 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
616 struct list_head *cancels, ldlm_mode_t mode,
619 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
620 struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0 } };
621 struct ldlm_resource *res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
628 count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
629 lock_flags, 0, NULL);
630 ldlm_resource_putref(res);
634 static int osc_destroy_interpret(struct ptlrpc_request *req, void *data,
637 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
639 atomic_dec(&cli->cl_destroy_in_flight);
640 cfs_waitq_signal(&cli->cl_destroy_waitq);
644 static int osc_can_send_destroy(struct client_obd *cli)
646 if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
647 cli->cl_max_rpcs_in_flight) {
648 /* The destroy request can be sent */
651 if (atomic_dec_return(&cli->cl_destroy_in_flight) <
652 cli->cl_max_rpcs_in_flight) {
654 * The counter has been modified between the two atomic
657 cfs_waitq_signal(&cli->cl_destroy_waitq);
662 /* Destroy requests can be async always on the client, and we don't even really
663 * care about the return code since the client cannot do anything at all about
665 * When the MDS is unlinking a filename, it saves the file objects into a
666 * recovery llog, and these object records are cancelled when the OST reports
667 * they were destroyed and sync'd to disk (i.e. transaction committed).
668 * If the client dies, or the OST is down when the object should be destroyed,
669 * the records are not cancelled, and when the OST reconnects to the MDS next,
670 * it will retrieve the llog unlink logs and then sends the log cancellation
671 * cookies to the MDS after committing destroy transactions. */
672 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
673 struct lov_stripe_md *ea, struct obd_trans_info *oti,
674 struct obd_export *md_export)
676 struct client_obd *cli = &exp->exp_obd->u.cli;
677 struct ptlrpc_request *req;
678 struct ost_body *body;
679 CFS_LIST_HEAD(cancels);
684 CDEBUG(D_INFO, "oa NULL\n");
688 count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
689 LDLM_FL_DISCARD_DATA);
691 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
693 ldlm_lock_list_put(&cancels, l_bl_ast, count);
697 rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
700 ptlrpc_request_free(req);
704 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
705 req->rq_interpret_reply = osc_destroy_interpret;
707 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
708 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
709 sizeof(*oti->oti_logcookies));
710 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
714 ptlrpc_request_set_replen(req);
716 if (!osc_can_send_destroy(cli)) {
717 struct l_wait_info lwi = { 0 };
720 * Wait until the number of on-going destroy RPCs drops
721 * under max_rpc_in_flight
723 l_wait_event_exclusive(cli->cl_destroy_waitq,
724 osc_can_send_destroy(cli), &lwi);
727 /* Do not wait for response */
728 ptlrpcd_add_req(req);
732 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
735 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
737 LASSERT(!(oa->o_valid & bits));
740 client_obd_list_lock(&cli->cl_loi_list_lock);
741 oa->o_dirty = cli->cl_dirty;
742 if (cli->cl_dirty > cli->cl_dirty_max) {
743 CERROR("dirty %lu > dirty_max %lu\n",
744 cli->cl_dirty, cli->cl_dirty_max);
746 } else if (atomic_read(&obd_dirty_pages) > obd_max_dirty_pages) {
747 CERROR("dirty %d > system dirty_max %d\n",
748 atomic_read(&obd_dirty_pages), obd_max_dirty_pages);
750 } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
751 CERROR("dirty %lu - dirty_max %lu too big???\n",
752 cli->cl_dirty, cli->cl_dirty_max);
755 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
756 (cli->cl_max_rpcs_in_flight + 1);
757 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
759 oa->o_grant = cli->cl_avail_grant;
760 oa->o_dropped = cli->cl_lost_grant;
761 cli->cl_lost_grant = 0;
762 client_obd_list_unlock(&cli->cl_loi_list_lock);
763 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
764 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
767 /* caller must hold loi_list_lock */
768 static void osc_consume_write_grant(struct client_obd *cli,
769 struct brw_page *pga)
771 atomic_inc(&obd_dirty_pages);
772 cli->cl_dirty += CFS_PAGE_SIZE;
773 cli->cl_avail_grant -= CFS_PAGE_SIZE;
774 pga->flag |= OBD_BRW_FROM_GRANT;
775 CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
776 CFS_PAGE_SIZE, pga, pga->pg);
777 LASSERT(cli->cl_avail_grant >= 0);
780 /* the companion to osc_consume_write_grant, called when a brw has completed.
781 * must be called with the loi lock held. */
782 static void osc_release_write_grant(struct client_obd *cli,
783 struct brw_page *pga, int sent)
785 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
788 if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
793 pga->flag &= ~OBD_BRW_FROM_GRANT;
794 atomic_dec(&obd_dirty_pages);
795 cli->cl_dirty -= CFS_PAGE_SIZE;
797 cli->cl_lost_grant += CFS_PAGE_SIZE;
798 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
799 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
800 } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
801 /* For short writes we shouldn't count parts of pages that
802 * span a whole block on the OST side, or our accounting goes
803 * wrong. Should match the code in filter_grant_check. */
804 int offset = pga->off & ~CFS_PAGE_MASK;
805 int count = pga->count + (offset & (blocksize - 1));
806 int end = (offset + pga->count) & (blocksize - 1);
808 count += blocksize - end;
810 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
811 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
812 CFS_PAGE_SIZE - count, cli->cl_lost_grant,
813 cli->cl_avail_grant, cli->cl_dirty);
819 static unsigned long rpcs_in_flight(struct client_obd *cli)
821 return cli->cl_r_in_flight + cli->cl_w_in_flight;
824 /* caller must hold loi_list_lock */
825 void osc_wake_cache_waiters(struct client_obd *cli)
827 struct list_head *l, *tmp;
828 struct osc_cache_waiter *ocw;
831 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
832 /* if we can't dirty more, we must wait until some is written */
833 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
834 (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
835 CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
836 "osc max %ld, sys max %d\n", cli->cl_dirty,
837 cli->cl_dirty_max, obd_max_dirty_pages);
841 /* if still dirty cache but no grant wait for pending RPCs that
842 * may yet return us some grant before doing sync writes */
843 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
844 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
845 cli->cl_w_in_flight);
849 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
850 list_del_init(&ocw->ocw_entry);
851 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
852 /* no more RPCs in flight to return grant, do sync IO */
853 ocw->ocw_rc = -EDQUOT;
854 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
856 osc_consume_write_grant(cli,
857 &ocw->ocw_oap->oap_brw_page);
860 cfs_waitq_signal(&ocw->ocw_waitq);
866 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
868 client_obd_list_lock(&cli->cl_loi_list_lock);
869 cli->cl_avail_grant = ocd->ocd_grant;
870 client_obd_list_unlock(&cli->cl_loi_list_lock);
872 CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
873 cli->cl_avail_grant, cli->cl_lost_grant);
874 LASSERT(cli->cl_avail_grant >= 0);
877 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
879 client_obd_list_lock(&cli->cl_loi_list_lock);
880 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
881 if (body->oa.o_valid & OBD_MD_FLGRANT)
882 cli->cl_avail_grant += body->oa.o_grant;
883 /* waiters are woken in brw_interpret_oap */
884 client_obd_list_unlock(&cli->cl_loi_list_lock);
887 /* We assume that the reason this OSC got a short read is because it read
888 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
889 * via the LOV, and it _knows_ it's reading inside the file, it's just that
890 * this stripe never got written at or beyond this stripe offset yet. */
891 static void handle_short_read(int nob_read, obd_count page_count,
892 struct brw_page **pga)
897 /* skip bytes read OK */
898 while (nob_read > 0) {
899 LASSERT (page_count > 0);
901 if (pga[i]->count > nob_read) {
902 /* EOF inside this page */
903 ptr = cfs_kmap(pga[i]->pg) +
904 (pga[i]->off & ~CFS_PAGE_MASK);
905 memset(ptr + nob_read, 0, pga[i]->count - nob_read);
906 cfs_kunmap(pga[i]->pg);
912 nob_read -= pga[i]->count;
917 /* zero remaining pages */
918 while (page_count-- > 0) {
919 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
920 memset(ptr, 0, pga[i]->count);
921 cfs_kunmap(pga[i]->pg);
926 static int check_write_rcs(struct ptlrpc_request *req,
927 int requested_nob, int niocount,
928 obd_count page_count, struct brw_page **pga)
932 /* return error if any niobuf was in error */
933 remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
934 sizeof(*remote_rcs) * niocount, NULL);
935 if (remote_rcs == NULL) {
936 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
939 if (lustre_msg_swabbed(req->rq_repmsg))
940 for (i = 0; i < niocount; i++)
941 __swab32s(&remote_rcs[i]);
943 for (i = 0; i < niocount; i++) {
944 if (remote_rcs[i] < 0)
945 return(remote_rcs[i]);
947 if (remote_rcs[i] != 0) {
948 CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
949 i, remote_rcs[i], req);
954 if (req->rq_bulk->bd_nob_transferred != requested_nob) {
955 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
956 requested_nob, req->rq_bulk->bd_nob_transferred);
963 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
965 if (p1->flag != p2->flag) {
966 unsigned mask = ~OBD_BRW_FROM_GRANT;
968 /* warn if we try to combine flags that we don't know to be
970 if ((p1->flag & mask) != (p2->flag & mask))
971 CERROR("is it ok to have flags 0x%x and 0x%x in the "
972 "same brw?\n", p1->flag, p2->flag);
976 return (p1->off + p1->count == p2->off);
979 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
980 struct brw_page **pga, int opc,
981 cksum_type_t cksum_type)
986 LASSERT (pg_count > 0);
987 cksum = init_checksum(cksum_type);
988 while (nob > 0 && pg_count > 0) {
989 unsigned char *ptr = cfs_kmap(pga[i]->pg);
990 int off = pga[i]->off & ~CFS_PAGE_MASK;
991 int count = pga[i]->count > nob ? nob : pga[i]->count;
993 /* corrupt the data before we compute the checksum, to
994 * simulate an OST->client data error */
995 if (i == 0 && opc == OST_READ &&
996 OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
997 memcpy(ptr + off, "bad1", min(4, nob));
998 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
999 cfs_kunmap(pga[i]->pg);
1000 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1003 nob -= pga[i]->count;
1007 /* For sending we only compute the wrong checksum instead
1008 * of corrupting the data so it is still correct on a redo */
1009 if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1015 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1016 struct lov_stripe_md *lsm, obd_count page_count,
1017 struct brw_page **pga,
1018 struct ptlrpc_request **reqp,
1019 struct obd_capa *ocapa)
1021 struct ptlrpc_request *req;
1022 struct ptlrpc_bulk_desc *desc;
1023 struct ost_body *body;
1024 struct obd_ioobj *ioobj;
1025 struct niobuf_remote *niobuf;
1026 int niocount, i, requested_nob, opc, rc;
1027 struct osc_brw_async_args *aa;
1028 struct req_capsule *pill;
1031 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1032 RETURN(-ENOMEM); /* Recoverable */
1033 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1034 RETURN(-EINVAL); /* Fatal */
1036 if ((cmd & OBD_BRW_WRITE) != 0) {
1038 req = ptlrpc_request_alloc_pool(cli->cl_import,
1039 cli->cl_import->imp_rq_pool,
1043 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1049 for (niocount = i = 1; i < page_count; i++) {
1050 if (!can_merge_pages(pga[i - 1], pga[i]))
1054 pill = &req->rq_pill;
1055 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1056 niocount * sizeof(*niobuf));
1057 osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1059 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1061 ptlrpc_request_free(req);
1064 req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1066 if (opc == OST_WRITE)
1067 desc = ptlrpc_prep_bulk_imp(req, page_count,
1068 BULK_GET_SOURCE, OST_BULK_PORTAL);
1070 desc = ptlrpc_prep_bulk_imp(req, page_count,
1071 BULK_PUT_SINK, OST_BULK_PORTAL);
1074 GOTO(out, rc = -ENOMEM);
1075 /* NB request now owns desc and will free it when it gets freed */
1077 body = req_capsule_client_get(pill, &RMF_OST_BODY);
1078 ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1079 niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1080 LASSERT(body && ioobj && niobuf);
1084 obdo_to_ioobj(oa, ioobj);
1085 ioobj->ioo_bufcnt = niocount;
1086 osc_pack_capa(req, body, ocapa);
1087 LASSERT (page_count > 0);
1088 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1089 struct brw_page *pg = pga[i];
1090 struct brw_page *pg_prev = pga[i - 1];
1092 LASSERT(pg->count > 0);
1093 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1094 "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1095 pg->off, pg->count);
1097 LASSERTF(i == 0 || pg->off > pg_prev->off,
1098 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1099 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1101 pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1102 pg_prev->pg, page_private(pg_prev->pg),
1103 pg_prev->pg->index, pg_prev->off);
1105 LASSERTF(i == 0 || pg->off > pg_prev->off,
1106 "i %d p_c %u\n", i, page_count);
1108 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1109 (pg->flag & OBD_BRW_SRVLOCK));
1111 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1113 requested_nob += pg->count;
1115 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1117 niobuf->len += pg->count;
1119 niobuf->offset = pg->off;
1120 niobuf->len = pg->count;
1121 niobuf->flags = pg->flag;
1125 LASSERT((void *)(niobuf - niocount) ==
1126 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1127 niocount * sizeof(*niobuf)));
1128 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1130 /* size[REQ_REC_OFF] still sizeof (*body) */
1131 if (opc == OST_WRITE) {
1132 if (unlikely(cli->cl_checksum) &&
1133 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1134 /* store cl_cksum_type in a local variable since
1135 * it can be changed via lprocfs */
1136 cksum_type_t cksum_type = cli->cl_cksum_type;
1138 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1139 oa->o_flags = body->oa.o_flags = 0;
1140 body->oa.o_flags |= cksum_type_pack(cksum_type);
1141 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1142 body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1146 CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1148 /* save this in 'oa', too, for later checking */
1149 oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1150 oa->o_flags |= cksum_type_pack(cksum_type);
1152 /* clear out the checksum flag, in case this is a
1153 * resend but cl_checksum is no longer set. b=11238 */
1154 oa->o_valid &= ~OBD_MD_FLCKSUM;
1156 oa->o_cksum = body->oa.o_cksum;
1157 /* 1 RC per niobuf */
1158 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1159 sizeof(__u32) * niocount);
1161 if (unlikely(cli->cl_checksum) &&
1162 req->rq_flvr.sf_bulk_hash == BULK_HASH_ALG_NULL) {
1163 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1164 body->oa.o_flags = 0;
1165 body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1166 body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1168 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1169 /* 1 RC for the whole I/O */
1171 ptlrpc_request_set_replen(req);
1173 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1174 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1176 aa->aa_requested_nob = requested_nob;
1177 aa->aa_nio_count = niocount;
1178 aa->aa_page_count = page_count;
1182 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1188 ptlrpc_req_finished(req);
1192 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1193 __u32 client_cksum, __u32 server_cksum, int nob,
1194 obd_count page_count, struct brw_page **pga,
1195 cksum_type_t client_cksum_type)
1199 cksum_type_t cksum_type;
1201 if (server_cksum == client_cksum) {
1202 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1206 if (oa->o_valid & OBD_MD_FLFLAGS)
1207 cksum_type = cksum_type_unpack(oa->o_flags);
1209 cksum_type = OBD_CKSUM_CRC32;
1211 new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1214 if (cksum_type != client_cksum_type)
1215 msg = "the server did not use the checksum type specified in "
1216 "the original request - likely a protocol problem";
1217 else if (new_cksum == server_cksum)
1218 msg = "changed on the client after we checksummed it - "
1219 "likely false positive due to mmap IO (bug 11742)";
1220 else if (new_cksum == client_cksum)
1221 msg = "changed in transit before arrival at OST";
1223 msg = "changed in transit AND doesn't match the original - "
1224 "likely false positive due to mmap IO (bug 11742)";
1226 LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1227 LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1228 "["LPU64"-"LPU64"]\n",
1229 msg, libcfs_nid2str(peer->nid),
1230 oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1231 oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1234 oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1236 pga[page_count-1]->off + pga[page_count-1]->count - 1);
1237 CERROR("original client csum %x (type %x), server csum %x (type %x), "
1238 "client csum now %x\n", client_cksum, client_cksum_type,
1239 server_cksum, cksum_type, new_cksum);
1243 /* Note rc enters this function as number of bytes transferred */
1244 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1246 struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1247 const lnet_process_id_t *peer =
1248 &req->rq_import->imp_connection->c_peer;
1249 struct client_obd *cli = aa->aa_cli;
1250 struct ost_body *body;
1251 __u32 client_cksum = 0;
1254 if (rc < 0 && rc != -EDQUOT)
1257 LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1258 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1259 lustre_swab_ost_body);
1261 CDEBUG(D_INFO, "Can't unpack body\n");
1265 /* set/clear over quota flag for a uid/gid */
1266 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1267 body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1268 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1269 body->oa.o_gid, body->oa.o_valid,
1275 if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1276 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1278 osc_update_grant(cli, body);
1280 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1282 CERROR("Unexpected +ve rc %d\n", rc);
1285 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1287 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1288 check_write_checksum(&body->oa, peer, client_cksum,
1289 body->oa.o_cksum, aa->aa_requested_nob,
1290 aa->aa_page_count, aa->aa_ppga,
1291 cksum_type_unpack(aa->aa_oa->o_flags)))
1294 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1297 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1298 aa->aa_page_count, aa->aa_ppga);
1302 /* The rest of this function executes only for OST_READs */
1303 if (rc > aa->aa_requested_nob) {
1304 CERROR("Unexpected rc %d (%d requested)\n", rc,
1305 aa->aa_requested_nob);
1309 if (rc != req->rq_bulk->bd_nob_transferred) {
1310 CERROR ("Unexpected rc %d (%d transferred)\n",
1311 rc, req->rq_bulk->bd_nob_transferred);
1315 if (rc < aa->aa_requested_nob)
1316 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1318 if (sptlrpc_cli_unwrap_bulk_read(req, rc, aa->aa_page_count,
1320 GOTO(out, rc = -EAGAIN);
1322 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1323 static int cksum_counter;
1324 __u32 server_cksum = body->oa.o_cksum;
1327 cksum_type_t cksum_type;
1329 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1330 cksum_type = cksum_type_unpack(body->oa.o_flags);
1332 cksum_type = OBD_CKSUM_CRC32;
1333 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1334 aa->aa_ppga, OST_READ,
1337 if (peer->nid == req->rq_bulk->bd_sender) {
1341 router = libcfs_nid2str(req->rq_bulk->bd_sender);
1344 if (server_cksum == ~0 && rc > 0) {
1345 CERROR("Protocol error: server %s set the 'checksum' "
1346 "bit, but didn't send a checksum. Not fatal, "
1347 "but please tell CFS.\n",
1348 libcfs_nid2str(peer->nid));
1349 } else if (server_cksum != client_cksum) {
1350 LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1351 "%s%s%s inum "LPU64"/"LPU64" object "
1352 LPU64"/"LPU64" extent "
1353 "["LPU64"-"LPU64"]\n",
1354 req->rq_import->imp_obd->obd_name,
1355 libcfs_nid2str(peer->nid),
1357 body->oa.o_valid & OBD_MD_FLFID ?
1358 body->oa.o_fid : (__u64)0,
1359 body->oa.o_valid & OBD_MD_FLFID ?
1360 body->oa.o_generation :(__u64)0,
1362 body->oa.o_valid & OBD_MD_FLGROUP ?
1363 body->oa.o_gr : (__u64)0,
1364 aa->aa_ppga[0]->off,
1365 aa->aa_ppga[aa->aa_page_count-1]->off +
1366 aa->aa_ppga[aa->aa_page_count-1]->count -
1368 CERROR("client %x, server %x, cksum_type %x\n",
1369 client_cksum, server_cksum, cksum_type);
1371 aa->aa_oa->o_cksum = client_cksum;
1375 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1378 } else if (unlikely(client_cksum)) {
1379 static int cksum_missed;
1382 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1383 CERROR("Checksum %u requested from %s but not sent\n",
1384 cksum_missed, libcfs_nid2str(peer->nid));
1390 *aa->aa_oa = body->oa;
1395 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1396 struct lov_stripe_md *lsm,
1397 obd_count page_count, struct brw_page **pga,
1398 struct obd_capa *ocapa)
1400 struct ptlrpc_request *req;
1404 struct l_wait_info lwi;
1408 cfs_waitq_init(&waitq);
1411 rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1412 page_count, pga, &req, ocapa);
1416 rc = ptlrpc_queue_wait(req);
1418 if (rc == -ETIMEDOUT && req->rq_resend) {
1419 DEBUG_REQ(D_HA, req, "BULK TIMEOUT");
1420 ptlrpc_req_finished(req);
1424 rc = osc_brw_fini_request(req, rc);
1426 ptlrpc_req_finished(req);
1427 if (osc_recoverable_error(rc)) {
1429 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1430 CERROR("too many resend retries, returning error\n");
1434 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1435 l_wait_event(waitq, 0, &lwi);
1443 int osc_brw_redo_request(struct ptlrpc_request *request,
1444 struct osc_brw_async_args *aa)
1446 struct ptlrpc_request *new_req;
1447 struct ptlrpc_request_set *set = request->rq_set;
1448 struct osc_brw_async_args *new_aa;
1449 struct osc_async_page *oap;
1453 if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1454 CERROR("too many resend retries, returning error\n");
1458 DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1460 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
1461 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1462 ocapa = lustre_unpack_capa(request->rq_reqmsg,
1465 rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1466 OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1467 aa->aa_cli, aa->aa_oa,
1468 NULL /* lsm unused by osc currently */,
1469 aa->aa_page_count, aa->aa_ppga,
1470 &new_req, NULL /* ocapa */);
1474 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1476 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1477 if (oap->oap_request != NULL) {
1478 LASSERTF(request == oap->oap_request,
1479 "request %p != oap_request %p\n",
1480 request, oap->oap_request);
1481 if (oap->oap_interrupted) {
1482 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1483 ptlrpc_req_finished(new_req);
1488 /* New request takes over pga and oaps from old request.
1489 * Note that copying a list_head doesn't work, need to move it... */
1491 new_req->rq_interpret_reply = request->rq_interpret_reply;
1492 new_req->rq_async_args = request->rq_async_args;
1493 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1495 new_aa = (struct osc_brw_async_args *)&new_req->rq_async_args;
1497 CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1498 list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1499 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1501 list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1502 if (oap->oap_request) {
1503 ptlrpc_req_finished(oap->oap_request);
1504 oap->oap_request = ptlrpc_request_addref(new_req);
1508 /* use ptlrpc_set_add_req is safe because interpret functions work
1509 * in check_set context. only one way exist with access to request
1510 * from different thread got -EINTR - this way protected with
1511 * cl_loi_list_lock */
1512 ptlrpc_set_add_req(set, new_req);
1514 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1516 DEBUG_REQ(D_INFO, new_req, "new request");
1520 static int brw_interpret(struct ptlrpc_request *req, void *data, int rc)
1522 struct osc_brw_async_args *aa = data;
1526 rc = osc_brw_fini_request(req, rc);
1527 if (osc_recoverable_error(rc)) {
1528 rc = osc_brw_redo_request(req, aa);
1533 client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1534 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1535 aa->aa_cli->cl_w_in_flight--;
1537 aa->aa_cli->cl_r_in_flight--;
1538 for (i = 0; i < aa->aa_page_count; i++)
1539 osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
1540 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1542 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1547 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1548 struct lov_stripe_md *lsm, obd_count page_count,
1549 struct brw_page **pga, struct ptlrpc_request_set *set,
1550 struct obd_capa *ocapa)
1552 struct ptlrpc_request *req;
1553 struct client_obd *cli = &exp->exp_obd->u.cli;
1555 struct osc_brw_async_args *aa;
1558 /* Consume write credits even if doing a sync write -
1559 * otherwise we may run out of space on OST due to grant. */
1560 if (cmd == OBD_BRW_WRITE) {
1561 spin_lock(&cli->cl_loi_list_lock);
1562 for (i = 0; i < page_count; i++) {
1563 if (cli->cl_avail_grant >= CFS_PAGE_SIZE)
1564 osc_consume_write_grant(cli, pga[i]);
1566 spin_unlock(&cli->cl_loi_list_lock);
1569 rc = osc_brw_prep_request(cmd, cli, oa, lsm, page_count, pga,
1572 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1573 if (cmd == OBD_BRW_READ) {
1574 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1575 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1576 ptlrpc_lprocfs_brw(req, OST_READ, aa->aa_requested_nob);
1578 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1579 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1580 cli->cl_w_in_flight);
1581 ptlrpc_lprocfs_brw(req, OST_WRITE, aa->aa_requested_nob);
1585 req->rq_interpret_reply = brw_interpret;
1586 ptlrpc_set_add_req(set, req);
1587 client_obd_list_lock(&cli->cl_loi_list_lock);
1588 if (cmd == OBD_BRW_READ)
1589 cli->cl_r_in_flight++;
1591 cli->cl_w_in_flight++;
1592 client_obd_list_unlock(&cli->cl_loi_list_lock);
1593 } else if (cmd == OBD_BRW_WRITE) {
1594 client_obd_list_lock(&cli->cl_loi_list_lock);
1595 for (i = 0; i < page_count; i++)
1596 osc_release_write_grant(cli, pga[i], 0);
1597 client_obd_list_unlock(&cli->cl_loi_list_lock);
1603 * ugh, we want disk allocation on the target to happen in offset order. we'll
1604 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1605 * fine for our small page arrays and doesn't require allocation. its an
1606 * insertion sort that swaps elements that are strides apart, shrinking the
1607 * stride down until its '1' and the array is sorted.
1609 static void sort_brw_pages(struct brw_page **array, int num)
1612 struct brw_page *tmp;
1616 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1621 for (i = stride ; i < num ; i++) {
1624 while (j >= stride && array[j - stride]->off > tmp->off) {
1625 array[j] = array[j - stride];
1630 } while (stride > 1);
1633 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1639 LASSERT (pages > 0);
1640 offset = pg[i]->off & ~CFS_PAGE_MASK;
1644 if (pages == 0) /* that's all */
1647 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1648 return count; /* doesn't end on page boundary */
1651 offset = pg[i]->off & ~CFS_PAGE_MASK;
1652 if (offset != 0) /* doesn't start on page boundary */
1659 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1661 struct brw_page **ppga;
1664 OBD_ALLOC(ppga, sizeof(*ppga) * count);
1668 for (i = 0; i < count; i++)
1673 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1675 LASSERT(ppga != NULL);
1676 OBD_FREE(ppga, sizeof(*ppga) * count);
1679 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1680 obd_count page_count, struct brw_page *pga,
1681 struct obd_trans_info *oti)
1683 struct obdo *saved_oa = NULL;
1684 struct brw_page **ppga, **orig;
1685 struct obd_import *imp = class_exp2cliimp(exp);
1686 struct client_obd *cli = &imp->imp_obd->u.cli;
1687 int rc, page_count_orig;
1690 if (cmd & OBD_BRW_CHECK) {
1691 /* The caller just wants to know if there's a chance that this
1692 * I/O can succeed */
1694 if (imp == NULL || imp->imp_invalid)
1699 /* test_brw with a failed create can trip this, maybe others. */
1700 LASSERT(cli->cl_max_pages_per_rpc);
1704 orig = ppga = osc_build_ppga(pga, page_count);
1707 page_count_orig = page_count;
1709 sort_brw_pages(ppga, page_count);
1710 while (page_count) {
1711 obd_count pages_per_brw;
1713 if (page_count > cli->cl_max_pages_per_rpc)
1714 pages_per_brw = cli->cl_max_pages_per_rpc;
1716 pages_per_brw = page_count;
1718 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1720 if (saved_oa != NULL) {
1721 /* restore previously saved oa */
1722 *oinfo->oi_oa = *saved_oa;
1723 } else if (page_count > pages_per_brw) {
1724 /* save a copy of oa (brw will clobber it) */
1725 OBDO_ALLOC(saved_oa);
1726 if (saved_oa == NULL)
1727 GOTO(out, rc = -ENOMEM);
1728 *saved_oa = *oinfo->oi_oa;
1731 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1732 pages_per_brw, ppga, oinfo->oi_capa);
1737 page_count -= pages_per_brw;
1738 ppga += pages_per_brw;
1742 osc_release_ppga(orig, page_count_orig);
1744 if (saved_oa != NULL)
1745 OBDO_FREE(saved_oa);
1750 static int osc_brw_async(int cmd, struct obd_export *exp,
1751 struct obd_info *oinfo, obd_count page_count,
1752 struct brw_page *pga, struct obd_trans_info *oti,
1753 struct ptlrpc_request_set *set)
1755 struct brw_page **ppga, **orig;
1756 struct client_obd *cli = &exp->exp_obd->u.cli;
1757 int page_count_orig;
1761 if (cmd & OBD_BRW_CHECK) {
1762 struct obd_import *imp = class_exp2cliimp(exp);
1763 /* The caller just wants to know if there's a chance that this
1764 * I/O can succeed */
1766 if (imp == NULL || imp->imp_invalid)
1771 orig = ppga = osc_build_ppga(pga, page_count);
1774 page_count_orig = page_count;
1776 sort_brw_pages(ppga, page_count);
1777 while (page_count) {
1778 struct brw_page **copy;
1779 obd_count pages_per_brw;
1781 pages_per_brw = min_t(obd_count, page_count,
1782 cli->cl_max_pages_per_rpc);
1784 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1786 /* use ppga only if single RPC is going to fly */
1787 if (pages_per_brw != page_count_orig || ppga != orig) {
1788 OBD_ALLOC(copy, sizeof(*copy) * pages_per_brw);
1790 GOTO(out, rc = -ENOMEM);
1791 memcpy(copy, ppga, sizeof(*copy) * pages_per_brw);
1795 rc = async_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1796 pages_per_brw, copy, set, oinfo->oi_capa);
1800 OBD_FREE(copy, sizeof(*copy) * pages_per_brw);
1804 /* we passed it to async_internal() which is
1805 * now responsible for releasing memory */
1809 page_count -= pages_per_brw;
1810 ppga += pages_per_brw;
1814 osc_release_ppga(orig, page_count_orig);
1818 static void osc_check_rpcs(struct client_obd *cli);
1820 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1821 * the dirty accounting. Writeback completes or truncate happens before
1822 * writing starts. Must be called with the loi lock held. */
1823 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1826 osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1830 /* This maintains the lists of pending pages to read/write for a given object
1831 * (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1832 * to quickly find objects that are ready to send an RPC. */
1833 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1839 if (lop->lop_num_pending == 0)
1842 /* if we have an invalid import we want to drain the queued pages
1843 * by forcing them through rpcs that immediately fail and complete
1844 * the pages. recovery relies on this to empty the queued pages
1845 * before canceling the locks and evicting down the llite pages */
1846 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1849 /* stream rpcs in queue order as long as as there is an urgent page
1850 * queued. this is our cheap solution for good batching in the case
1851 * where writepage marks some random page in the middle of the file
1852 * as urgent because of, say, memory pressure */
1853 if (!list_empty(&lop->lop_urgent)) {
1854 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1857 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1858 optimal = cli->cl_max_pages_per_rpc;
1859 if (cmd & OBD_BRW_WRITE) {
1860 /* trigger a write rpc stream as long as there are dirtiers
1861 * waiting for space. as they're waiting, they're not going to
1862 * create more pages to coallesce with what's waiting.. */
1863 if (!list_empty(&cli->cl_cache_waiters)) {
1864 CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1867 /* +16 to avoid triggering rpcs that would want to include pages
1868 * that are being queued but which can't be made ready until
1869 * the queuer finishes with the page. this is a wart for
1870 * llite::commit_write() */
1873 if (lop->lop_num_pending >= optimal)
1879 static void on_list(struct list_head *item, struct list_head *list,
1882 if (list_empty(item) && should_be_on)
1883 list_add_tail(item, list);
1884 else if (!list_empty(item) && !should_be_on)
1885 list_del_init(item);
1888 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1889 * can find pages to build into rpcs quickly */
1890 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1892 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1893 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1894 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1896 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1897 loi->loi_write_lop.lop_num_pending);
1899 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1900 loi->loi_read_lop.lop_num_pending);
1903 static void lop_update_pending(struct client_obd *cli,
1904 struct loi_oap_pages *lop, int cmd, int delta)
1906 lop->lop_num_pending += delta;
1907 if (cmd & OBD_BRW_WRITE)
1908 cli->cl_pending_w_pages += delta;
1910 cli->cl_pending_r_pages += delta;
1913 /* this is called when a sync waiter receives an interruption. Its job is to
1914 * get the caller woken as soon as possible. If its page hasn't been put in an
1915 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1916 * desiring interruption which will forcefully complete the rpc once the rpc
1918 static void osc_occ_interrupted(struct oig_callback_context *occ)
1920 struct osc_async_page *oap;
1921 struct loi_oap_pages *lop;
1922 struct lov_oinfo *loi;
1925 /* XXX member_of() */
1926 oap = list_entry(occ, struct osc_async_page, oap_occ);
1928 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
1930 oap->oap_interrupted = 1;
1932 /* ok, it's been put in an rpc. only one oap gets a request reference */
1933 if (oap->oap_request != NULL) {
1934 ptlrpc_mark_interrupted(oap->oap_request);
1935 ptlrpcd_wake(oap->oap_request);
1939 /* we don't get interruption callbacks until osc_trigger_group_io()
1940 * has been called and put the sync oaps in the pending/urgent lists.*/
1941 if (!list_empty(&oap->oap_pending_item)) {
1942 list_del_init(&oap->oap_pending_item);
1943 list_del_init(&oap->oap_urgent_item);
1946 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
1947 &loi->loi_write_lop : &loi->loi_read_lop;
1948 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1949 loi_list_maint(oap->oap_cli, oap->oap_loi);
1951 oig_complete_one(oap->oap_oig, &oap->oap_occ, -EINTR);
1952 oap->oap_oig = NULL;
1956 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
1959 /* this is trying to propogate async writeback errors back up to the
1960 * application. As an async write fails we record the error code for later if
1961 * the app does an fsync. As long as errors persist we force future rpcs to be
1962 * sync so that the app can get a sync error and break the cycle of queueing
1963 * pages for which writeback will fail. */
1964 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1971 ar->ar_force_sync = 1;
1972 ar->ar_min_xid = ptlrpc_sample_next_xid();
1977 if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1978 ar->ar_force_sync = 0;
1981 static void osc_oap_to_pending(struct osc_async_page *oap)
1983 struct loi_oap_pages *lop;
1985 if (oap->oap_cmd & OBD_BRW_WRITE)
1986 lop = &oap->oap_loi->loi_write_lop;
1988 lop = &oap->oap_loi->loi_read_lop;
1990 if (oap->oap_async_flags & ASYNC_URGENT)
1991 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1992 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1993 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
1996 /* this must be called holding the loi list lock to give coverage to exit_cache,
1997 * async_flag maintenance, and oap_request */
1998 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1999 struct osc_async_page *oap, int sent, int rc)
2004 if (oap->oap_request != NULL) {
2005 xid = ptlrpc_req_xid(oap->oap_request);
2006 ptlrpc_req_finished(oap->oap_request);
2007 oap->oap_request = NULL;
2010 oap->oap_async_flags = 0;
2011 oap->oap_interrupted = 0;
2013 if (oap->oap_cmd & OBD_BRW_WRITE) {
2014 osc_process_ar(&cli->cl_ar, xid, rc);
2015 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2018 if (rc == 0 && oa != NULL) {
2019 if (oa->o_valid & OBD_MD_FLBLOCKS)
2020 oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2021 if (oa->o_valid & OBD_MD_FLMTIME)
2022 oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2023 if (oa->o_valid & OBD_MD_FLATIME)
2024 oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2025 if (oa->o_valid & OBD_MD_FLCTIME)
2026 oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2030 osc_exit_cache(cli, oap, sent);
2031 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
2032 oap->oap_oig = NULL;
2037 rc = oap->oap_caller_ops->ap_completion(oap->oap_caller_data,
2038 oap->oap_cmd, oa, rc);
2040 /* ll_ap_completion (from llite) drops PG_locked. so, a new
2041 * I/O on the page could start, but OSC calls it under lock
2042 * and thus we can add oap back to pending safely */
2044 /* upper layer wants to leave the page on pending queue */
2045 osc_oap_to_pending(oap);
2047 osc_exit_cache(cli, oap, sent);
2051 static int brw_interpret_oap(struct ptlrpc_request *req, void *data, int rc)
2053 struct osc_async_page *oap, *tmp;
2054 struct osc_brw_async_args *aa = data;
2055 struct client_obd *cli;
2058 rc = osc_brw_fini_request(req, rc);
2059 CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2060 if (osc_recoverable_error(rc)) {
2061 rc = osc_brw_redo_request(req, aa);
2068 client_obd_list_lock(&cli->cl_loi_list_lock);
2070 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2071 * is called so we know whether to go to sync BRWs or wait for more
2072 * RPCs to complete */
2073 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2074 cli->cl_w_in_flight--;
2076 cli->cl_r_in_flight--;
2078 /* the caller may re-use the oap after the completion call so
2079 * we need to clean it up a little */
2080 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2081 list_del_init(&oap->oap_rpc_item);
2082 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
2085 osc_wake_cache_waiters(cli);
2086 osc_check_rpcs(cli);
2088 client_obd_list_unlock(&cli->cl_loi_list_lock);
2090 OBDO_FREE(aa->aa_oa);
2092 osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2096 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
2097 struct list_head *rpc_list,
2098 int page_count, int cmd)
2100 struct ptlrpc_request *req;
2101 struct brw_page **pga = NULL;
2102 struct osc_brw_async_args *aa;
2103 struct obdo *oa = NULL;
2104 struct obd_async_page_ops *ops = NULL;
2105 void *caller_data = NULL;
2106 struct obd_capa *ocapa;
2107 struct osc_async_page *oap;
2111 LASSERT(!list_empty(rpc_list));
2113 OBD_ALLOC(pga, sizeof(*pga) * page_count);
2115 RETURN(ERR_PTR(-ENOMEM));
2119 GOTO(out, req = ERR_PTR(-ENOMEM));
2122 list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2124 ops = oap->oap_caller_ops;
2125 caller_data = oap->oap_caller_data;
2127 pga[i] = &oap->oap_brw_page;
2128 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2129 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2130 pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2134 /* always get the data for the obdo for the rpc */
2135 LASSERT(ops != NULL);
2136 ops->ap_fill_obdo(caller_data, cmd, oa);
2137 ocapa = ops->ap_lookup_capa(caller_data, cmd);
2139 sort_brw_pages(pga, page_count);
2140 rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2144 CERROR("prep_req failed: %d\n", rc);
2145 GOTO(out, req = ERR_PTR(rc));
2148 /* Need to update the timestamps after the request is built in case
2149 * we race with setattr (locally or in queue at OST). If OST gets
2150 * later setattr before earlier BRW (as determined by the request xid),
2151 * the OST will not use BRW timestamps. Sadly, there is no obvious
2152 * way to do this in a single call. bug 10150 */
2153 ops->ap_update_obdo(caller_data, cmd, oa,
2154 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME);
2156 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2157 aa = (struct osc_brw_async_args *)&req->rq_async_args;
2158 CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2159 list_splice(rpc_list, &aa->aa_oaps);
2160 CFS_INIT_LIST_HEAD(rpc_list);
2167 OBD_FREE(pga, sizeof(*pga) * page_count);
2172 /* the loi lock is held across this function but it's allowed to release
2173 * and reacquire it during its work */
2175 * prepare pages for ASYNC io and put pages in send queue.
2179 * \param cmd - OBD_BRW_* macroses
2180 * \param lop - pending pages
2182 * \return zero if pages successfully add to send queue.
2183 * \return not zere if error occurring.
2185 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
2186 int cmd, struct loi_oap_pages *lop)
2188 struct ptlrpc_request *req;
2189 obd_count page_count = 0;
2190 struct osc_async_page *oap = NULL, *tmp;
2191 struct osc_brw_async_args *aa;
2192 struct obd_async_page_ops *ops;
2193 CFS_LIST_HEAD(rpc_list);
2194 unsigned int ending_offset;
2195 unsigned starting_offset = 0;
2199 /* first we find the pages we're allowed to work with */
2200 list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2202 ops = oap->oap_caller_ops;
2204 LASSERT(oap->oap_magic == OAP_MAGIC);
2206 if (page_count != 0 &&
2207 srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2208 CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2209 " oap %p, page %p, srvlock %u\n",
2210 oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2213 /* in llite being 'ready' equates to the page being locked
2214 * until completion unlocks it. commit_write submits a page
2215 * as not ready because its unlock will happen unconditionally
2216 * as the call returns. if we race with commit_write giving
2217 * us that page we dont' want to create a hole in the page
2218 * stream, so we stop and leave the rpc to be fired by
2219 * another dirtier or kupdated interval (the not ready page
2220 * will still be on the dirty list). we could call in
2221 * at the end of ll_file_write to process the queue again. */
2222 if (!(oap->oap_async_flags & ASYNC_READY)) {
2223 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
2225 CDEBUG(D_INODE, "oap %p page %p returned %d "
2226 "instead of ready\n", oap,
2230 /* llite is telling us that the page is still
2231 * in commit_write and that we should try
2232 * and put it in an rpc again later. we
2233 * break out of the loop so we don't create
2234 * a hole in the sequence of pages in the rpc
2239 /* the io isn't needed.. tell the checks
2240 * below to complete the rpc with EINTR */
2241 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2242 oap->oap_count = -EINTR;
2245 oap->oap_async_flags |= ASYNC_READY;
2248 LASSERTF(0, "oap %p page %p returned %d "
2249 "from make_ready\n", oap,
2257 * Page submitted for IO has to be locked. Either by
2258 * ->ap_make_ready() or by higher layers.
2260 #if defined(__KERNEL__) && defined(__linux__)
2261 if(!(PageLocked(oap->oap_page) &&
2262 (CheckWriteback(oap->oap_page, cmd) || oap->oap_oig !=NULL))) {
2263 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2264 oap->oap_page, (long)oap->oap_page->flags, oap->oap_async_flags);
2268 /* If there is a gap at the start of this page, it can't merge
2269 * with any previous page, so we'll hand the network a
2270 * "fragmented" page array that it can't transfer in 1 RDMA */
2271 if (page_count != 0 && oap->oap_page_off != 0)
2274 /* take the page out of our book-keeping */
2275 list_del_init(&oap->oap_pending_item);
2276 lop_update_pending(cli, lop, cmd, -1);
2277 list_del_init(&oap->oap_urgent_item);
2279 if (page_count == 0)
2280 starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2281 (PTLRPC_MAX_BRW_SIZE - 1);
2283 /* ask the caller for the size of the io as the rpc leaves. */
2284 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
2286 ops->ap_refresh_count(oap->oap_caller_data,cmd);
2287 if (oap->oap_count <= 0) {
2288 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2290 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
2294 /* now put the page back in our accounting */
2295 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2296 if (page_count == 0)
2297 srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2298 if (++page_count >= cli->cl_max_pages_per_rpc)
2301 /* End on a PTLRPC_MAX_BRW_SIZE boundary. We want full-sized
2302 * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2303 * have the same alignment as the initial writes that allocated
2304 * extents on the server. */
2305 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2306 oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2307 if (ending_offset == 0)
2310 /* If there is a gap at the end of this page, it can't merge
2311 * with any subsequent pages, so we'll hand the network a
2312 * "fragmented" page array that it can't transfer in 1 RDMA */
2313 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2317 osc_wake_cache_waiters(cli);
2319 if (page_count == 0)
2322 loi_list_maint(cli, loi);
2324 client_obd_list_unlock(&cli->cl_loi_list_lock);
2326 req = osc_build_req(cli, &rpc_list, page_count, cmd);
2328 /* this should happen rarely and is pretty bad, it makes the
2329 * pending list not follow the dirty order */
2330 client_obd_list_lock(&cli->cl_loi_list_lock);
2331 list_for_each_entry_safe(oap, tmp, &rpc_list, oap_rpc_item) {
2332 list_del_init(&oap->oap_rpc_item);
2334 /* queued sync pages can be torn down while the pages
2335 * were between the pending list and the rpc */
2336 if (oap->oap_interrupted) {
2337 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2338 osc_ap_completion(cli, NULL, oap, 0,
2342 osc_ap_completion(cli, NULL, oap, 0, PTR_ERR(req));
2344 loi_list_maint(cli, loi);
2345 RETURN(PTR_ERR(req));
2348 aa = (struct osc_brw_async_args *)&req->rq_async_args;
2350 if (cmd == OBD_BRW_READ) {
2351 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2352 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2353 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2354 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2355 ptlrpc_lprocfs_brw(req, OST_READ, aa->aa_requested_nob);
2357 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2358 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2359 cli->cl_w_in_flight);
2360 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2361 (starting_offset >> CFS_PAGE_SHIFT) + 1);
2362 ptlrpc_lprocfs_brw(req, OST_WRITE, aa->aa_requested_nob);
2365 client_obd_list_lock(&cli->cl_loi_list_lock);
2367 if (cmd == OBD_BRW_READ)
2368 cli->cl_r_in_flight++;
2370 cli->cl_w_in_flight++;
2372 /* queued sync pages can be torn down while the pages
2373 * were between the pending list and the rpc */
2375 list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2376 /* only one oap gets a request reference */
2379 if (oap->oap_interrupted && !req->rq_intr) {
2380 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2382 ptlrpc_mark_interrupted(req);
2386 tmp->oap_request = ptlrpc_request_addref(req);
2388 DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2389 page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2391 req->rq_interpret_reply = brw_interpret_oap;
2392 ptlrpcd_add_req(req);
2396 #define LOI_DEBUG(LOI, STR, args...) \
2397 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
2398 !list_empty(&(LOI)->loi_cli_item), \
2399 (LOI)->loi_write_lop.lop_num_pending, \
2400 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
2401 (LOI)->loi_read_lop.lop_num_pending, \
2402 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
2405 /* This is called by osc_check_rpcs() to find which objects have pages that
2406 * we could be sending. These lists are maintained by lop_makes_rpc(). */
2407 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2410 /* first return all objects which we already know to have
2411 * pages ready to be stuffed into rpcs */
2412 if (!list_empty(&cli->cl_loi_ready_list))
2413 RETURN(list_entry(cli->cl_loi_ready_list.next,
2414 struct lov_oinfo, loi_cli_item));
2416 /* then if we have cache waiters, return all objects with queued
2417 * writes. This is especially important when many small files
2418 * have filled up the cache and not been fired into rpcs because
2419 * they don't pass the nr_pending/object threshhold */
2420 if (!list_empty(&cli->cl_cache_waiters) &&
2421 !list_empty(&cli->cl_loi_write_list))
2422 RETURN(list_entry(cli->cl_loi_write_list.next,
2423 struct lov_oinfo, loi_write_item));
2425 /* then return all queued objects when we have an invalid import
2426 * so that they get flushed */
2427 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2428 if (!list_empty(&cli->cl_loi_write_list))
2429 RETURN(list_entry(cli->cl_loi_write_list.next,
2430 struct lov_oinfo, loi_write_item));
2431 if (!list_empty(&cli->cl_loi_read_list))
2432 RETURN(list_entry(cli->cl_loi_read_list.next,
2433 struct lov_oinfo, loi_read_item));
2438 /* called with the loi list lock held */
2439 static void osc_check_rpcs(struct client_obd *cli)
2441 struct lov_oinfo *loi;
2442 int rc = 0, race_counter = 0;
2445 while ((loi = osc_next_loi(cli)) != NULL) {
2446 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2448 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
2451 /* attempt some read/write balancing by alternating between
2452 * reads and writes in an object. The makes_rpc checks here
2453 * would be redundant if we were getting read/write work items
2454 * instead of objects. we don't want send_oap_rpc to drain a
2455 * partial read pending queue when we're given this object to
2456 * do io on writes while there are cache waiters */
2457 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2458 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
2459 &loi->loi_write_lop);
2467 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2468 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
2469 &loi->loi_read_lop);
2478 /* attempt some inter-object balancing by issueing rpcs
2479 * for each object in turn */
2480 if (!list_empty(&loi->loi_cli_item))
2481 list_del_init(&loi->loi_cli_item);
2482 if (!list_empty(&loi->loi_write_item))
2483 list_del_init(&loi->loi_write_item);
2484 if (!list_empty(&loi->loi_read_item))
2485 list_del_init(&loi->loi_read_item);
2487 loi_list_maint(cli, loi);
2489 /* send_oap_rpc fails with 0 when make_ready tells it to
2490 * back off. llite's make_ready does this when it tries
2491 * to lock a page queued for write that is already locked.
2492 * we want to try sending rpcs from many objects, but we
2493 * don't want to spin failing with 0. */
2494 if (race_counter == 10)
2500 /* we're trying to queue a page in the osc so we're subject to the
2501 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2502 * If the osc's queued pages are already at that limit, then we want to sleep
2503 * until there is space in the osc's queue for us. We also may be waiting for
2504 * write credits from the OST if there are RPCs in flight that may return some
2505 * before we fall back to sync writes.
2507 * We need this know our allocation was granted in the presence of signals */
2508 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2512 client_obd_list_lock(&cli->cl_loi_list_lock);
2513 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2514 client_obd_list_unlock(&cli->cl_loi_list_lock);
2518 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2519 * grant or cache space. */
2520 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
2521 struct osc_async_page *oap)
2523 struct osc_cache_waiter ocw;
2524 struct l_wait_info lwi = { 0 };
2528 CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2529 "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2530 cli->cl_dirty_max, obd_max_dirty_pages,
2531 cli->cl_lost_grant, cli->cl_avail_grant);
2533 /* force the caller to try sync io. this can jump the list
2534 * of queued writes and create a discontiguous rpc stream */
2535 if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2536 loi->loi_ar.ar_force_sync)
2539 /* Hopefully normal case - cache space and write credits available */
2540 if ((cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max) &&
2541 (atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) &&
2542 (cli->cl_avail_grant >= CFS_PAGE_SIZE)) {
2543 /* account for ourselves */
2544 osc_consume_write_grant(cli, &oap->oap_brw_page);
2548 /* Make sure that there are write rpcs in flight to wait for. This
2549 * is a little silly as this object may not have any pending but
2550 * other objects sure might. */
2551 if (cli->cl_w_in_flight) {
2552 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2553 cfs_waitq_init(&ocw.ocw_waitq);
2557 loi_list_maint(cli, loi);
2558 osc_check_rpcs(cli);
2559 client_obd_list_unlock(&cli->cl_loi_list_lock);
2561 CDEBUG(D_CACHE, "sleeping for cache space\n");
2562 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2564 client_obd_list_lock(&cli->cl_loi_list_lock);
2565 if (!list_empty(&ocw.ocw_entry)) {
2566 list_del(&ocw.ocw_entry);
2575 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2576 struct lov_oinfo *loi, cfs_page_t *page,
2577 obd_off offset, struct obd_async_page_ops *ops,
2578 void *data, void **res, int nocache,
2579 struct lustre_handle *lockh)
2581 struct osc_async_page *oap;
2582 struct ldlm_res_id oid = {{0}};
2587 return size_round(sizeof(*oap));
2590 oap->oap_magic = OAP_MAGIC;
2591 oap->oap_cli = &exp->exp_obd->u.cli;
2594 oap->oap_caller_ops = ops;
2595 oap->oap_caller_data = data;
2597 oap->oap_page = page;
2598 oap->oap_obj_off = offset;
2600 CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2601 CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2602 CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2603 CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2605 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2607 spin_lock_init(&oap->oap_lock);
2609 /* If the page was marked as notcacheable - don't add to any locks */
2611 oid.name[0] = loi->loi_id;
2612 oid.name[2] = loi->loi_gr;
2613 /* This is the only place where we can call cache_add_extent
2614 without oap_lock, because this page is locked now, and
2615 the lock we are adding it to is referenced, so cannot lose
2616 any pages either. */
2617 rc = cache_add_extent(oap->oap_cli->cl_cache, &oid, oap, lockh);
2622 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2626 struct osc_async_page *oap_from_cookie(void *cookie)
2628 struct osc_async_page *oap = cookie;
2629 if (oap->oap_magic != OAP_MAGIC)
2630 return ERR_PTR(-EINVAL);
2634 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2635 struct lov_oinfo *loi, void *cookie,
2636 int cmd, obd_off off, int count,
2637 obd_flag brw_flags, enum async_flags async_flags)
2639 struct client_obd *cli = &exp->exp_obd->u.cli;
2640 struct osc_async_page *oap;
2644 oap = oap_from_cookie(cookie);
2646 RETURN(PTR_ERR(oap));
2648 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2651 if (!list_empty(&oap->oap_pending_item) ||
2652 !list_empty(&oap->oap_urgent_item) ||
2653 !list_empty(&oap->oap_rpc_item))
2656 /* check if the file's owner/group is over quota */
2657 #ifdef HAVE_QUOTA_SUPPORT
2658 if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)){
2659 struct obd_async_page_ops *ops;
2666 ops = oap->oap_caller_ops;
2667 ops->ap_fill_obdo(oap->oap_caller_data, cmd, oa);
2668 if (lquota_chkdq(quota_interface, cli, oa->o_uid, oa->o_gid) ==
2679 loi = lsm->lsm_oinfo[0];
2681 client_obd_list_lock(&cli->cl_loi_list_lock);
2684 oap->oap_page_off = off;
2685 oap->oap_count = count;
2686 oap->oap_brw_flags = brw_flags;
2687 oap->oap_async_flags = async_flags;
2689 if (cmd & OBD_BRW_WRITE) {
2690 rc = osc_enter_cache(cli, loi, oap);
2692 client_obd_list_unlock(&cli->cl_loi_list_lock);
2697 osc_oap_to_pending(oap);
2698 loi_list_maint(cli, loi);
2700 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2703 osc_check_rpcs(cli);
2704 client_obd_list_unlock(&cli->cl_loi_list_lock);
2709 /* aka (~was & now & flag), but this is more clear :) */
2710 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2712 static int osc_set_async_flags(struct obd_export *exp,
2713 struct lov_stripe_md *lsm,
2714 struct lov_oinfo *loi, void *cookie,
2715 obd_flag async_flags)
2717 struct client_obd *cli = &exp->exp_obd->u.cli;
2718 struct loi_oap_pages *lop;
2719 struct osc_async_page *oap;
2723 oap = oap_from_cookie(cookie);
2725 RETURN(PTR_ERR(oap));
2728 * bug 7311: OST-side locking is only supported for liblustre for now
2729 * (and liblustre never calls obd_set_async_flags(). I hope.), generic
2730 * implementation has to handle case where OST-locked page was picked
2731 * up by, e.g., ->writepage().
2733 LASSERT(!(oap->oap_brw_flags & OBD_BRW_SRVLOCK));
2734 LASSERT(!LIBLUSTRE_CLIENT); /* check that liblustre angels do fear to
2737 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2741 loi = lsm->lsm_oinfo[0];
2743 if (oap->oap_cmd & OBD_BRW_WRITE) {
2744 lop = &loi->loi_write_lop;
2746 lop = &loi->loi_read_lop;
2749 client_obd_list_lock(&cli->cl_loi_list_lock);
2751 if (list_empty(&oap->oap_pending_item))
2752 GOTO(out, rc = -EINVAL);
2754 if ((oap->oap_async_flags & async_flags) == async_flags)
2757 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2758 oap->oap_async_flags |= ASYNC_READY;
2760 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2761 if (list_empty(&oap->oap_rpc_item)) {
2762 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2763 loi_list_maint(cli, loi);
2767 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2768 oap->oap_async_flags);
2770 osc_check_rpcs(cli);
2771 client_obd_list_unlock(&cli->cl_loi_list_lock);
2775 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2776 struct lov_oinfo *loi,
2777 struct obd_io_group *oig, void *cookie,
2778 int cmd, obd_off off, int count,
2780 obd_flag async_flags)
2782 struct client_obd *cli = &exp->exp_obd->u.cli;
2783 struct osc_async_page *oap;
2784 struct loi_oap_pages *lop;
2788 oap = oap_from_cookie(cookie);
2790 RETURN(PTR_ERR(oap));
2792 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2795 if (!list_empty(&oap->oap_pending_item) ||
2796 !list_empty(&oap->oap_urgent_item) ||
2797 !list_empty(&oap->oap_rpc_item))
2801 loi = lsm->lsm_oinfo[0];
2803 client_obd_list_lock(&cli->cl_loi_list_lock);
2806 oap->oap_page_off = off;
2807 oap->oap_count = count;
2808 oap->oap_brw_flags = brw_flags;
2809 oap->oap_async_flags = async_flags;
2811 if (cmd & OBD_BRW_WRITE)
2812 lop = &loi->loi_write_lop;
2814 lop = &loi->loi_read_lop;
2816 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2817 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2819 rc = oig_add_one(oig, &oap->oap_occ);
2822 LOI_DEBUG(loi, "oap %p page %p on group pending: rc %d\n",
2823 oap, oap->oap_page, rc);
2825 client_obd_list_unlock(&cli->cl_loi_list_lock);
2830 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2831 struct loi_oap_pages *lop, int cmd)
2833 struct list_head *pos, *tmp;
2834 struct osc_async_page *oap;
2836 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2837 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2838 list_del(&oap->oap_pending_item);
2839 osc_oap_to_pending(oap);
2841 loi_list_maint(cli, loi);
2844 static int osc_trigger_group_io(struct obd_export *exp,
2845 struct lov_stripe_md *lsm,
2846 struct lov_oinfo *loi,
2847 struct obd_io_group *oig)
2849 struct client_obd *cli = &exp->exp_obd->u.cli;
2853 loi = lsm->lsm_oinfo[0];
2855 client_obd_list_lock(&cli->cl_loi_list_lock);
2857 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2858 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2860 osc_check_rpcs(cli);
2861 client_obd_list_unlock(&cli->cl_loi_list_lock);
2866 static int osc_teardown_async_page(struct obd_export *exp,
2867 struct lov_stripe_md *lsm,
2868 struct lov_oinfo *loi, void *cookie)
2870 struct client_obd *cli = &exp->exp_obd->u.cli;
2871 struct loi_oap_pages *lop;
2872 struct osc_async_page *oap;
2876 oap = oap_from_cookie(cookie);
2878 RETURN(PTR_ERR(oap));
2881 loi = lsm->lsm_oinfo[0];
2883 if (oap->oap_cmd & OBD_BRW_WRITE) {
2884 lop = &loi->loi_write_lop;
2886 lop = &loi->loi_read_lop;
2889 client_obd_list_lock(&cli->cl_loi_list_lock);
2891 if (!list_empty(&oap->oap_rpc_item))
2892 GOTO(out, rc = -EBUSY);
2894 osc_exit_cache(cli, oap, 0);
2895 osc_wake_cache_waiters(cli);
2897 if (!list_empty(&oap->oap_urgent_item)) {
2898 list_del_init(&oap->oap_urgent_item);
2899 oap->oap_async_flags &= ~ASYNC_URGENT;
2901 if (!list_empty(&oap->oap_pending_item)) {
2902 list_del_init(&oap->oap_pending_item);
2903 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2905 loi_list_maint(cli, loi);
2906 cache_remove_extent(cli->cl_cache, oap);
2908 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2910 client_obd_list_unlock(&cli->cl_loi_list_lock);
2914 int osc_extent_blocking_cb(struct ldlm_lock *lock,
2915 struct ldlm_lock_desc *new, void *data,
2918 struct lustre_handle lockh = { 0 };
2922 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
2923 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
2928 case LDLM_CB_BLOCKING:
2929 ldlm_lock2handle(lock, &lockh);
2930 rc = ldlm_cli_cancel(&lockh);
2932 CERROR("ldlm_cli_cancel failed: %d\n", rc);
2934 case LDLM_CB_CANCELING: {
2936 ldlm_lock2handle(lock, &lockh);
2937 /* This lock wasn't granted, don't try to do anything */
2938 if (lock->l_req_mode != lock->l_granted_mode)
2941 cache_remove_lock(lock->l_conn_export->exp_obd->u.cli.cl_cache,
2944 if (lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb)
2945 lock->l_conn_export->exp_obd->u.cli.cl_ext_lock_cancel_cb(
2946 lock, new, data,flag);
2955 EXPORT_SYMBOL(osc_extent_blocking_cb);
2957 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data,
2960 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2963 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2966 lock_res_and_lock(lock);
2967 #if defined (__KERNEL__) && defined (__linux__)
2968 /* Liang XXX: Darwin and Winnt checking should be added */
2969 if (lock->l_ast_data && lock->l_ast_data != data) {
2970 struct inode *new_inode = data;
2971 struct inode *old_inode = lock->l_ast_data;
2972 if (!(old_inode->i_state & I_FREEING))
2973 LDLM_ERROR(lock, "inconsistent l_ast_data found");
2974 LASSERTF(old_inode->i_state & I_FREEING,
2975 "Found existing inode %p/%lu/%u state %lu in lock: "
2976 "setting data to %p/%lu/%u\n", old_inode,
2977 old_inode->i_ino, old_inode->i_generation,
2979 new_inode, new_inode->i_ino, new_inode->i_generation);
2982 lock->l_ast_data = data;
2983 lock->l_flags |= (flags & LDLM_FL_NO_LRU);
2984 unlock_res_and_lock(lock);
2985 LDLM_LOCK_PUT(lock);
2988 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2989 ldlm_iterator_t replace, void *data)
2991 struct ldlm_res_id res_id = { .name = {0} };
2992 struct obd_device *obd = class_exp2obd(exp);
2994 res_id.name[0] = lsm->lsm_object_id;
2995 res_id.name[2] = lsm->lsm_object_gr;
2997 ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3001 static int osc_enqueue_fini(struct obd_device *obd, struct ptlrpc_request *req,
3002 struct obd_info *oinfo, int intent, int rc)
3007 /* The request was created before ldlm_cli_enqueue call. */
3008 if (rc == ELDLM_LOCK_ABORTED) {
3009 struct ldlm_reply *rep;
3010 rep = req_capsule_server_get(&req->rq_pill,
3013 LASSERT(rep != NULL);
3014 if (rep->lock_policy_res1)
3015 rc = rep->lock_policy_res1;
3019 if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3020 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3021 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_size,
3022 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_blocks,
3023 oinfo->oi_md->lsm_oinfo[0]->loi_lvb.lvb_mtime);
3027 cache_add_lock(obd->u.cli.cl_cache, oinfo->oi_lockh);
3029 /* Call the update callback. */
3030 rc = oinfo->oi_cb_up(oinfo, rc);
3034 static int osc_enqueue_interpret(struct ptlrpc_request *req,
3035 struct osc_enqueue_args *aa, int rc)
3037 int intent = aa->oa_oi->oi_flags & LDLM_FL_HAS_INTENT;
3038 struct lov_stripe_md *lsm = aa->oa_oi->oi_md;
3039 struct ldlm_lock *lock;
3041 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3043 lock = ldlm_handle2lock(aa->oa_oi->oi_lockh);
3045 /* Complete obtaining the lock procedure. */
3046 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3048 &aa->oa_oi->oi_flags,
3049 &lsm->lsm_oinfo[0]->loi_lvb,
3050 sizeof(lsm->lsm_oinfo[0]->loi_lvb),
3051 lustre_swab_ost_lvb,
3052 aa->oa_oi->oi_lockh, rc);
3054 /* Complete osc stuff. */
3055 rc = osc_enqueue_fini(aa->oa_exp->exp_obd, req, aa->oa_oi, intent, rc);
3057 /* Release the lock for async request. */
3058 if (lustre_handle_is_used(aa->oa_oi->oi_lockh) && rc == ELDLM_OK)
3059 ldlm_lock_decref(aa->oa_oi->oi_lockh, aa->oa_ei->ei_mode);
3061 LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3062 aa->oa_oi->oi_lockh, req, aa);
3063 LDLM_LOCK_PUT(lock);
3067 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3068 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3069 * other synchronous requests, however keeping some locks and trying to obtain
3070 * others may take a considerable amount of time in a case of ost failure; and
3071 * when other sync requests do not get released lock from a client, the client
3072 * is excluded from the cluster -- such scenarious make the life difficult, so
3073 * release locks just after they are obtained. */
3074 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3075 struct ldlm_enqueue_info *einfo,
3076 struct ptlrpc_request_set *rqset)
3078 struct ldlm_res_id res_id = { .name = {0} };
3079 struct obd_device *obd = exp->exp_obd;
3080 struct ptlrpc_request *req = NULL;
3081 int intent = oinfo->oi_flags & LDLM_FL_HAS_INTENT;
3086 res_id.name[0] = oinfo->oi_md->lsm_object_id;
3087 res_id.name[2] = oinfo->oi_md->lsm_object_gr;
3089 /* Filesystem lock extents are extended to page boundaries so that
3090 * dealing with the page cache is a little smoother. */
3091 oinfo->oi_policy.l_extent.start -=
3092 oinfo->oi_policy.l_extent.start & ~CFS_PAGE_MASK;
3093 oinfo->oi_policy.l_extent.end |= ~CFS_PAGE_MASK;
3095 if (oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid == 0)
3098 /* Next, search for already existing extent locks that will cover us */
3099 /* If we're trying to read, we also search for an existing PW lock. The
3100 * VFS and page cache already protect us locally, so lots of readers/
3101 * writers can share a single PW lock.
3103 * There are problems with conversion deadlocks, so instead of
3104 * converting a read lock to a write lock, we'll just enqueue a new
3107 * At some point we should cancel the read lock instead of making them
3108 * send us a blocking callback, but there are problems with canceling
3109 * locks out from other users right now, too. */
3110 mode = einfo->ei_mode;
3111 if (einfo->ei_mode == LCK_PR)
3113 mode = ldlm_lock_match(obd->obd_namespace,
3114 oinfo->oi_flags | LDLM_FL_LVB_READY, &res_id,
3115 einfo->ei_type, &oinfo->oi_policy, mode,
3118 /* addref the lock only if not async requests and PW lock is
3119 * matched whereas we asked for PR. */
3120 if (!rqset && einfo->ei_mode != mode)
3121 ldlm_lock_addref(oinfo->oi_lockh, LCK_PR);
3122 osc_set_data_with_check(oinfo->oi_lockh, einfo->ei_cbdata,
3125 /* I would like to be able to ASSERT here that rss <=
3126 * kms, but I can't, for reasons which are explained in
3130 /* We already have a lock, and it's referenced */
3131 oinfo->oi_cb_up(oinfo, ELDLM_OK);
3133 /* For async requests, decref the lock. */
3134 if (einfo->ei_mode != mode)
3135 ldlm_lock_decref(oinfo->oi_lockh, LCK_PW);
3137 ldlm_lock_decref(oinfo->oi_lockh, einfo->ei_mode);
3144 CFS_LIST_HEAD(cancels);
3145 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3146 &RQF_LDLM_ENQUEUE_LVB);
3150 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3154 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3155 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb));
3156 ptlrpc_request_set_replen(req);
3159 /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3160 oinfo->oi_flags &= ~LDLM_FL_BLOCK_GRANTED;
3162 rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id,
3163 &oinfo->oi_policy, &oinfo->oi_flags,
3164 &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3165 sizeof(oinfo->oi_md->lsm_oinfo[0]->loi_lvb),
3166 lustre_swab_ost_lvb, oinfo->oi_lockh,
3170 struct osc_enqueue_args *aa;
3171 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3172 aa = (struct osc_enqueue_args *)&req->rq_async_args;
3177 req->rq_interpret_reply = osc_enqueue_interpret;
3178 ptlrpc_set_add_req(rqset, req);
3179 } else if (intent) {
3180 ptlrpc_req_finished(req);
3185 rc = osc_enqueue_fini(obd, req, oinfo, intent, rc);
3187 ptlrpc_req_finished(req);
3192 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
3193 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3194 int *flags, void *data, struct lustre_handle *lockh)
3196 struct ldlm_res_id res_id = { .name = {0} };
3197 struct obd_device *obd = exp->exp_obd;
3198 int lflags = *flags;
3202 res_id.name[0] = lsm->lsm_object_id;
3203 res_id.name[2] = lsm->lsm_object_gr;
3205 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3208 /* Filesystem lock extents are extended to page boundaries so that
3209 * dealing with the page cache is a little smoother */
3210 policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3211 policy->l_extent.end |= ~CFS_PAGE_MASK;
3213 /* Next, search for already existing extent locks that will cover us */
3214 /* If we're trying to read, we also search for an existing PW lock. The
3215 * VFS and page cache already protect us locally, so lots of readers/
3216 * writers can share a single PW lock. */
3220 rc = ldlm_lock_match(obd->obd_namespace, lflags | LDLM_FL_LVB_READY,
3221 &res_id, type, policy, rc, lockh);
3223 osc_set_data_with_check(lockh, data, lflags);
3224 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3225 ldlm_lock_addref(lockh, LCK_PR);
3226 ldlm_lock_decref(lockh, LCK_PW);
3233 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3234 __u32 mode, struct lustre_handle *lockh)
3238 if (unlikely(mode == LCK_GROUP))
3239 ldlm_lock_decref_and_cancel(lockh, mode);
3241 ldlm_lock_decref(lockh, mode);
3246 static int osc_cancel_unused(struct obd_export *exp,
3247 struct lov_stripe_md *lsm, int flags,
3250 struct obd_device *obd = class_exp2obd(exp);
3251 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
3254 res_id.name[0] = lsm->lsm_object_id;
3255 res_id.name[2] = lsm->lsm_object_gr;
3259 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3262 static int osc_join_lru(struct obd_export *exp,
3263 struct lov_stripe_md *lsm, int join)
3265 struct obd_device *obd = class_exp2obd(exp);
3266 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
3269 res_id.name[0] = lsm->lsm_object_id;
3270 res_id.name[2] = lsm->lsm_object_gr;
3274 return ldlm_cli_join_lru(obd->obd_namespace, resp, join);
3277 static int osc_statfs_interpret(struct ptlrpc_request *req,
3278 struct osc_async_args *aa, int rc)
3280 struct obd_statfs *msfs;
3286 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3288 GOTO(out, rc = -EPROTO);
3291 *aa->aa_oi->oi_osfs = *msfs;
3293 rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3297 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3298 __u64 max_age, struct ptlrpc_request_set *rqset)
3300 struct ptlrpc_request *req;
3301 struct osc_async_args *aa;
3305 /* We could possibly pass max_age in the request (as an absolute
3306 * timestamp or a "seconds.usec ago") so the target can avoid doing
3307 * extra calls into the filesystem if that isn't necessary (e.g.
3308 * during mount that would help a bit). Having relative timestamps
3309 * is not so great if request processing is slow, while absolute
3310 * timestamps are not ideal because they need time synchronization. */
3311 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3315 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3317 ptlrpc_request_free(req);
3320 ptlrpc_request_set_replen(req);
3321 req->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
3322 if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3323 /* procfs requests not want stat in wait for avoid deadlock */
3324 req->rq_no_resend = 1;
3325 req->rq_no_delay = 1;
3328 req->rq_interpret_reply = osc_statfs_interpret;
3329 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3330 aa = (struct osc_async_args *)&req->rq_async_args;
3333 ptlrpc_set_add_req(rqset, req);
3337 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3338 __u64 max_age, __u32 flags)
3340 struct obd_statfs *msfs;
3341 struct ptlrpc_request *req;
3345 /* We could possibly pass max_age in the request (as an absolute
3346 * timestamp or a "seconds.usec ago") so the target can avoid doing
3347 * extra calls into the filesystem if that isn't necessary (e.g.
3348 * during mount that would help a bit). Having relative timestamps
3349 * is not so great if request processing is slow, while absolute
3350 * timestamps are not ideal because they need time synchronization. */
3351 req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3355 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3357 ptlrpc_request_free(req);
3360 ptlrpc_request_set_replen(req);
3361 req->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
3363 if (flags & OBD_STATFS_NODELAY) {
3364 /* procfs requests not want stat in wait for avoid deadlock */
3365 req->rq_no_resend = 1;
3366 req->rq_no_delay = 1;
3369 rc = ptlrpc_queue_wait(req);
3373 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3375 GOTO(out, rc = -EPROTO);
3382 ptlrpc_req_finished(req);
3386 /* Retrieve object striping information.
3388 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3389 * the maximum number of OST indices which will fit in the user buffer.
3390 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3392 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3394 struct lov_user_md lum, *lumk;
3395 int rc = 0, lum_size;
3401 if (copy_from_user(&lum, lump, sizeof(lum)))
3404 if (lum.lmm_magic != LOV_USER_MAGIC)
3407 if (lum.lmm_stripe_count > 0) {
3408 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
3409 OBD_ALLOC(lumk, lum_size);
3413 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
3414 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
3416 lum_size = sizeof(lum);
3420 lumk->lmm_object_id = lsm->lsm_object_id;
3421 lumk->lmm_object_gr = lsm->lsm_object_gr;
3422 lumk->lmm_stripe_count = 1;
3424 if (copy_to_user(lump, lumk, lum_size))
3428 OBD_FREE(lumk, lum_size);
3434 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3435 void *karg, void *uarg)
3437 struct obd_device *obd = exp->exp_obd;
3438 struct obd_ioctl_data *data = karg;
3442 if (!try_module_get(THIS_MODULE)) {
3443 CERROR("Can't get module. Is it alive?");
3447 case OBD_IOC_LOV_GET_CONFIG: {
3449 struct lov_desc *desc;
3450 struct obd_uuid uuid;
3454 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3455 GOTO(out, err = -EINVAL);
3457 data = (struct obd_ioctl_data *)buf;
3459 if (sizeof(*desc) > data->ioc_inllen1) {
3460 obd_ioctl_freedata(buf, len);
3461 GOTO(out, err = -EINVAL);
3464 if (data->ioc_inllen2 < sizeof(uuid)) {
3465 obd_ioctl_freedata(buf, len);
3466 GOTO(out, err = -EINVAL);
3469 desc = (struct lov_desc *)data->ioc_inlbuf1;
3470 desc->ld_tgt_count = 1;
3471 desc->ld_active_tgt_count = 1;
3472 desc->ld_default_stripe_count = 1;
3473 desc->ld_default_stripe_size = 0;
3474 desc->ld_default_stripe_offset = 0;
3475 desc->ld_pattern = 0;
3476 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3478 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3480 err = copy_to_user((void *)uarg, buf, len);
3483 obd_ioctl_freedata(buf, len);
3486 case LL_IOC_LOV_SETSTRIPE:
3487 err = obd_alloc_memmd(exp, karg);
3491 case LL_IOC_LOV_GETSTRIPE:
3492 err = osc_getstripe(karg, uarg);
3494 case OBD_IOC_CLIENT_RECOVER:
3495 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3500 case IOC_OSC_SET_ACTIVE:
3501 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3504 case OBD_IOC_POLL_QUOTACHECK:
3505 err = lquota_poll_check(quota_interface, exp,
3506 (struct if_quotacheck *)karg);
3509 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3510 cmd, cfs_curproc_comm());
3511 GOTO(out, err = -ENOTTY);
3514 module_put(THIS_MODULE);
3518 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3519 void *key, __u32 *vallen, void *val)
3522 if (!vallen || !val)
3525 if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3526 __u32 *stripe = val;
3527 *vallen = sizeof(*stripe);
3530 } else if (KEY_IS(KEY_LAST_ID)) {
3531 struct ptlrpc_request *req;
3536 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3537 &RQF_OST_GET_INFO_LAST_ID);
3541 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3542 RCL_CLIENT, keylen);
3543 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3545 ptlrpc_request_free(req);
3549 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3550 memcpy(tmp, key, keylen);
3552 ptlrpc_request_set_replen(req);
3553 rc = ptlrpc_queue_wait(req);
3557 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3559 GOTO(out, rc = -EPROTO);
3561 *((obd_id *)val) = *reply;
3563 ptlrpc_req_finished(req);
3569 static int osc_setinfo_mds_conn_interpret(struct ptlrpc_request *req,
3572 struct llog_ctxt *ctxt;
3573 struct obd_import *imp = req->rq_import;
3579 ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
3582 rc = llog_initiator_connect(ctxt);
3584 CERROR("cannot establish connection for "
3585 "ctxt %p: %d\n", ctxt, rc);
3588 llog_ctxt_put(ctxt);
3589 spin_lock(&imp->imp_lock);
3590 imp->imp_server_timeout = 1;
3591 imp->imp_pingable = 1;
3592 spin_unlock(&imp->imp_lock);
3593 CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
3598 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
3599 void *key, obd_count vallen, void *val,
3600 struct ptlrpc_request_set *set)
3602 struct ptlrpc_request *req;
3603 struct obd_device *obd = exp->exp_obd;
3604 struct obd_import *imp = class_exp2cliimp(exp);
3609 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3611 if (KEY_IS(KEY_NEXT_ID)) {
3612 if (vallen != sizeof(obd_id))
3616 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
3617 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
3618 exp->exp_obd->obd_name,
3619 obd->u.cli.cl_oscc.oscc_next_id);
3624 if (KEY_IS(KEY_UNLINKED)) {
3625 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3626 spin_lock(&oscc->oscc_lock);
3627 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3628 spin_unlock(&oscc->oscc_lock);
3632 if (KEY_IS(KEY_INIT_RECOV)) {
3633 if (vallen != sizeof(int))
3635 spin_lock(&imp->imp_lock);
3636 imp->imp_initial_recov = *(int *)val;
3637 spin_unlock(&imp->imp_lock);
3638 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
3639 exp->exp_obd->obd_name,
3640 imp->imp_initial_recov);
3644 if (KEY_IS(KEY_CHECKSUM)) {
3645 if (vallen != sizeof(int))
3647 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3651 if (KEY_IS(KEY_FLUSH_CTX)) {
3652 sptlrpc_import_flush_my_ctx(imp);
3659 /* We pass all other commands directly to OST. Since nobody calls osc
3660 methods directly and everybody is supposed to go through LOV, we
3661 assume lov checked invalid values for us.
3662 The only recognised values so far are evict_by_nid and mds_conn.
3663 Even if something bad goes through, we'd get a -EINVAL from OST
3667 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
3671 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3672 RCL_CLIENT, keylen);
3673 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3674 RCL_CLIENT, vallen);
3675 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3677 ptlrpc_request_free(req);
3681 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3682 memcpy(tmp, key, keylen);
3683 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
3684 memcpy(tmp, val, vallen);
3686 if (KEY_IS(KEY_MDS_CONN)) {
3687 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3689 oscc->oscc_oa.o_gr = (*(__u32 *)val);
3690 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
3691 LASSERT(oscc->oscc_oa.o_gr > 0);
3692 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
3695 ptlrpc_request_set_replen(req);
3696 ptlrpc_set_add_req(set, req);
3697 ptlrpc_check_set(set);
3703 static struct llog_operations osc_size_repl_logops = {
3704 lop_cancel: llog_obd_repl_cancel
3707 static struct llog_operations osc_mds_ost_orig_logops;
3708 static int osc_llog_init(struct obd_device *obd, int group,
3709 struct obd_device *tgt, int count,
3710 struct llog_catid *catid, struct obd_uuid *uuid)
3714 LASSERT(group == OBD_LLOG_GROUP);
3715 spin_lock(&obd->obd_dev_lock);
3716 if (osc_mds_ost_orig_logops.lop_setup != llog_obd_origin_setup) {
3717 osc_mds_ost_orig_logops = llog_lvfs_ops;
3718 osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
3719 osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
3720 osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
3721 osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
3723 spin_unlock(&obd->obd_dev_lock);
3725 rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, count,
3726 &catid->lci_logid, &osc_mds_ost_orig_logops);
3728 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
3732 rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, count,
3733 NULL, &osc_size_repl_logops);
3735 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
3738 CERROR("osc '%s' tgt '%s' cnt %d catid %p rc=%d\n",
3739 obd->obd_name, tgt->obd_name, count, catid, rc);
3740 CERROR("logid "LPX64":0x%x\n",
3741 catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
3746 static int osc_llog_finish(struct obd_device *obd, int count)
3748 struct llog_ctxt *ctxt;
3749 int rc = 0, rc2 = 0;
3752 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3754 rc = llog_cleanup(ctxt);
3756 ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3758 rc2 = llog_cleanup(ctxt);
3765 static int osc_reconnect(const struct lu_env *env,
3766 struct obd_export *exp, struct obd_device *obd,
3767 struct obd_uuid *cluuid,
3768 struct obd_connect_data *data)
3770 struct client_obd *cli = &obd->u.cli;
3772 if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3775 client_obd_list_lock(&cli->cl_loi_list_lock);
3776 data->ocd_grant = cli->cl_avail_grant ?:
3777 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
3778 lost_grant = cli->cl_lost_grant;
3779 cli->cl_lost_grant = 0;
3780 client_obd_list_unlock(&cli->cl_loi_list_lock);
3782 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
3783 "cl_lost_grant: %ld\n", data->ocd_grant,
3784 cli->cl_avail_grant, lost_grant);
3785 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3786 " ocd_grant: %d\n", data->ocd_connect_flags,
3787 data->ocd_version, data->ocd_grant);
3793 static int osc_disconnect(struct obd_export *exp)
3795 struct obd_device *obd = class_exp2obd(exp);
3796 struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3799 if (obd->u.cli.cl_conn_count == 1)
3800 /* flush any remaining cancel messages out to the target */
3801 llog_sync(ctxt, exp);
3803 llog_ctxt_put(ctxt);
3805 rc = client_disconnect_export(exp);
3809 static int osc_import_event(struct obd_device *obd,
3810 struct obd_import *imp,
3811 enum obd_import_event event)
3813 struct client_obd *cli;
3817 LASSERT(imp->imp_obd == obd);
3820 case IMP_EVENT_DISCON: {
3821 /* Only do this on the MDS OSC's */
3822 if (imp->imp_server_timeout) {
3823 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3825 spin_lock(&oscc->oscc_lock);
3826 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3827 spin_unlock(&oscc->oscc_lock);
3830 client_obd_list_lock(&cli->cl_loi_list_lock);
3831 cli->cl_avail_grant = 0;
3832 cli->cl_lost_grant = 0;
3833 client_obd_list_unlock(&cli->cl_loi_list_lock);
3836 case IMP_EVENT_INACTIVE: {
3837 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3840 case IMP_EVENT_INVALIDATE: {
3841 struct ldlm_namespace *ns = obd->obd_namespace;
3845 client_obd_list_lock(&cli->cl_loi_list_lock);
3846 /* all pages go to failing rpcs due to the invalid import */
3847 osc_check_rpcs(cli);
3848 client_obd_list_unlock(&cli->cl_loi_list_lock);
3850 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3854 case IMP_EVENT_ACTIVE: {
3855 /* Only do this on the MDS OSC's */
3856 if (imp->imp_server_timeout) {
3857 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3859 spin_lock(&oscc->oscc_lock);
3860 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3861 spin_unlock(&oscc->oscc_lock);
3863 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3866 case IMP_EVENT_OCD: {
3867 struct obd_connect_data *ocd = &imp->imp_connect_data;
3869 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3870 osc_init_grant(&obd->u.cli, ocd);
3873 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3874 imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3876 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3880 CERROR("Unknown import event %d\n", event);
3886 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3892 rc = ptlrpcd_addref();
3896 rc = client_obd_setup(obd, lcfg);
3900 struct lprocfs_static_vars lvars = { 0 };
3901 struct client_obd *cli = &obd->u.cli;
3903 lprocfs_osc_init_vars(&lvars);
3904 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3905 lproc_osc_attach_seqstat(obd);
3906 sptlrpc_lprocfs_cliobd_attach(obd);
3907 ptlrpc_lprocfs_register_obd(obd);
3911 /* We need to allocate a few requests more, because
3912 brw_interpret_oap tries to create new requests before freeing
3913 previous ones. Ideally we want to have 2x max_rpcs_in_flight
3914 reserved, but I afraid that might be too much wasted RAM
3915 in fact, so 2 is just my guess and still should work. */
3916 cli->cl_import->imp_rq_pool =
3917 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3919 ptlrpc_add_rqs_to_pool);
3920 cli->cl_cache = cache_create(obd);
3921 if (!cli->cl_cache) {
3930 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3936 case OBD_CLEANUP_EARLY: {
3937 struct obd_import *imp;
3938 imp = obd->u.cli.cl_import;
3939 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3940 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3941 ptlrpc_deactivate_import(imp);
3942 spin_lock(&imp->imp_lock);
3943 imp->imp_pingable = 0;
3944 spin_unlock(&imp->imp_lock);
3947 case OBD_CLEANUP_EXPORTS: {
3948 /* If we set up but never connected, the
3949 client import will not have been cleaned. */
3950 if (obd->u.cli.cl_import) {
3951 struct obd_import *imp;
3952 imp = obd->u.cli.cl_import;
3953 CDEBUG(D_CONFIG, "%s: client import never connected\n",
3955 ptlrpc_invalidate_import(imp);
3956 ptlrpc_free_rq_pool(imp->imp_rq_pool);
3957 class_destroy_import(imp);
3958 obd->u.cli.cl_import = NULL;
3962 case OBD_CLEANUP_SELF_EXP:
3963 rc = obd_llog_finish(obd, 0);
3965 CERROR("failed to cleanup llogging subsystems\n");
3967 case OBD_CLEANUP_OBD:
3973 int osc_cleanup(struct obd_device *obd)
3975 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3979 ptlrpc_lprocfs_unregister_obd(obd);
3980 lprocfs_obd_cleanup(obd);
3982 spin_lock(&oscc->oscc_lock);
3983 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3984 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3985 spin_unlock(&oscc->oscc_lock);
3987 /* free memory of osc quota cache */
3988 lquota_cleanup(quota_interface, obd);
3990 cache_destroy(obd->u.cli.cl_cache);
3991 rc = client_obd_cleanup(obd);
3997 static int osc_register_page_removal_cb(struct obd_export *exp,
3998 obd_page_removal_cb_t func,
3999 obd_pin_extent_cb pin_cb)
4001 return cache_add_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func,
4005 static int osc_unregister_page_removal_cb(struct obd_export *exp,
4006 obd_page_removal_cb_t func)
4008 return cache_del_extent_removal_cb(exp->exp_obd->u.cli.cl_cache, func);
4011 static int osc_register_lock_cancel_cb(struct obd_export *exp,
4012 obd_lock_cancel_cb cb)
4014 LASSERT(exp->exp_obd->u.cli.cl_ext_lock_cancel_cb == NULL);
4016 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = cb;
4020 static int osc_unregister_lock_cancel_cb(struct obd_export *exp,
4021 obd_lock_cancel_cb cb)
4023 if (exp->exp_obd->u.cli.cl_ext_lock_cancel_cb != cb) {
4024 CERROR("Unregistering cancel cb %p, while only %p was "
4026 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb);
4030 exp->exp_obd->u.cli.cl_ext_lock_cancel_cb = NULL;
4034 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4036 struct lustre_cfg *lcfg = buf;
4037 struct lprocfs_static_vars lvars = { 0 };
4040 lprocfs_osc_init_vars(&lvars);
4042 switch (lcfg->lcfg_command) {
4043 case LCFG_SPTLRPC_CONF:
4044 rc = sptlrpc_cliobd_process_config(obd, lcfg);
4047 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4055 struct obd_ops osc_obd_ops = {
4056 .o_owner = THIS_MODULE,
4057 .o_setup = osc_setup,
4058 .o_precleanup = osc_precleanup,
4059 .o_cleanup = osc_cleanup,
4060 .o_add_conn = client_import_add_conn,
4061 .o_del_conn = client_import_del_conn,
4062 .o_connect = client_connect_import,
4063 .o_reconnect = osc_reconnect,
4064 .o_disconnect = osc_disconnect,
4065 .o_statfs = osc_statfs,
4066 .o_statfs_async = osc_statfs_async,
4067 .o_packmd = osc_packmd,
4068 .o_unpackmd = osc_unpackmd,
4069 .o_precreate = osc_precreate,
4070 .o_create = osc_create,
4071 .o_destroy = osc_destroy,
4072 .o_getattr = osc_getattr,
4073 .o_getattr_async = osc_getattr_async,
4074 .o_setattr = osc_setattr,
4075 .o_setattr_async = osc_setattr_async,
4077 .o_brw_async = osc_brw_async,
4078 .o_prep_async_page = osc_prep_async_page,
4079 .o_queue_async_io = osc_queue_async_io,
4080 .o_set_async_flags = osc_set_async_flags,
4081 .o_queue_group_io = osc_queue_group_io,
4082 .o_trigger_group_io = osc_trigger_group_io,
4083 .o_teardown_async_page = osc_teardown_async_page,
4084 .o_punch = osc_punch,
4086 .o_enqueue = osc_enqueue,
4087 .o_match = osc_match,
4088 .o_change_cbdata = osc_change_cbdata,
4089 .o_cancel = osc_cancel,
4090 .o_cancel_unused = osc_cancel_unused,
4091 .o_join_lru = osc_join_lru,
4092 .o_iocontrol = osc_iocontrol,
4093 .o_get_info = osc_get_info,
4094 .o_set_info_async = osc_set_info_async,
4095 .o_import_event = osc_import_event,
4096 .o_llog_init = osc_llog_init,
4097 .o_llog_finish = osc_llog_finish,
4098 .o_process_config = osc_process_config,
4099 .o_register_page_removal_cb = osc_register_page_removal_cb,
4100 .o_unregister_page_removal_cb = osc_unregister_page_removal_cb,
4101 .o_register_lock_cancel_cb = osc_register_lock_cancel_cb,
4102 .o_unregister_lock_cancel_cb = osc_unregister_lock_cancel_cb,
4104 int __init osc_init(void)
4106 struct lprocfs_static_vars lvars = { 0 };
4110 lprocfs_osc_init_vars(&lvars);
4112 request_module("lquota");
4113 quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4114 lquota_init(quota_interface);
4115 init_obd_quota_ops(quota_interface, &osc_obd_ops);
4117 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4118 LUSTRE_OSC_NAME, NULL);
4120 if (quota_interface)
4121 PORTAL_SYMBOL_PUT(osc_quota_interface);
4129 static void /*__exit*/ osc_exit(void)
4131 lquota_exit(quota_interface);
4132 if (quota_interface)
4133 PORTAL_SYMBOL_PUT(osc_quota_interface);
4135 class_unregister_type(LUSTRE_OSC_NAME);
4138 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4139 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4140 MODULE_LICENSE("GPL");
4142 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);