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 Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 * For testing and management it is treated as an obd_device,
23 * although * it does not export a full OBD method table (the
24 * requests are coming * in over the wire, so object target modules
25 * do not have a full * method table.)
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_OSC
35 # include <linux/version.h>
36 # include <linux/module.h>
37 # include <linux/mm.h>
38 # include <linux/highmem.h>
39 # include <linux/ctype.h>
40 # include <linux/init.h>
41 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
42 # include <linux/workqueue.h>
43 # include <linux/smp_lock.h>
45 # include <linux/locks.h>
47 #else /* __KERNEL__ */
48 # include <liblustre.h>
51 # include <linux/lustre_dlm.h>
52 #include <linux/kp30.h>
53 #include <linux/lustre_net.h>
54 #include <lustre/lustre_user.h>
55 #include <linux/obd_ost.h>
56 #include <linux/obd_lov.h>
62 #include <linux/lustre_ha.h>
63 #include <linux/lprocfs_status.h>
64 #include <linux/lustre_log.h>
65 #include "osc_internal.h"
67 /* Pack OSC object metadata for disk storage (LE byte order). */
68 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
69 struct lov_stripe_md *lsm)
74 lmm_size = sizeof(**lmmp);
79 OBD_FREE(*lmmp, lmm_size);
85 OBD_ALLOC(*lmmp, lmm_size);
91 LASSERT(lsm->lsm_object_id);
92 LASSERT(lsm->lsm_object_gr);
93 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
94 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
100 /* Unpack OSC object metadata from disk storage (LE byte order). */
101 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
102 struct lov_mds_md *lmm, int lmm_bytes)
108 if (lmm_bytes < sizeof (*lmm)) {
109 CERROR("lov_mds_md too small: %d, need %d\n",
110 lmm_bytes, (int)sizeof(*lmm));
113 /* XXX LOV_MAGIC etc check? */
115 if (lmm->lmm_object_id == 0) {
116 CERROR("lov_mds_md: zero lmm_object_id\n");
121 lsm_size = lov_stripe_md_size(1);
125 if (*lsmp != NULL && lmm == NULL) {
126 OBD_FREE(*lsmp, lsm_size);
132 OBD_ALLOC(*lsmp, lsm_size);
135 loi_init((*lsmp)->lsm_oinfo);
139 /* XXX zero *lsmp? */
140 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
141 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
142 LASSERT((*lsmp)->lsm_object_id);
143 LASSERT((*lsmp)->lsm_object_gr);
146 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
151 static int osc_getattr_interpret(struct ptlrpc_request *req,
152 struct osc_getattr_async_args *aa, int rc)
154 struct ost_body *body;
160 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
162 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
163 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
165 /* This should really be sent by the OST */
166 aa->aa_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
167 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
169 CERROR("can't unpack ost_body\n");
171 aa->aa_oa->o_valid = 0;
177 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
178 struct lov_stripe_md *md,
179 struct ptlrpc_request_set *set)
181 struct ptlrpc_request *request;
182 struct ost_body *body;
183 int size = sizeof(*body);
184 struct osc_getattr_async_args *aa;
187 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
192 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
193 memcpy(&body->oa, oa, sizeof(*oa));
195 request->rq_replen = lustre_msg_size(1, &size);
196 request->rq_interpret_reply = osc_getattr_interpret;
198 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
199 aa = (struct osc_getattr_async_args *)&request->rq_async_args;
202 ptlrpc_set_add_req (set, request);
206 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
207 struct lov_stripe_md *md)
209 struct ptlrpc_request *request;
210 struct ost_body *body;
211 int rc, size = sizeof(*body);
214 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
219 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
220 memcpy(&body->oa, oa, sizeof(*oa));
222 request->rq_replen = lustre_msg_size(1, &size);
224 rc = ptlrpc_queue_wait(request);
226 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
230 body = lustre_swab_repbuf(request, 0, sizeof (*body),
231 lustre_swab_ost_body);
233 CERROR ("can't unpack ost_body\n");
234 GOTO (out, rc = -EPROTO);
237 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
238 memcpy(oa, &body->oa, sizeof(*oa));
240 /* This should really be sent by the OST */
241 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
242 oa->o_valid |= OBD_MD_FLBLKSZ;
246 ptlrpc_req_finished(request);
250 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
251 struct lov_stripe_md *md, struct obd_trans_info *oti)
253 struct ptlrpc_request *request;
254 struct ost_body *body;
255 int rc, size = sizeof(*body);
258 LASSERT(!(oa->o_valid & OBD_MD_FLGROUP) || oa->o_gr > 0);
260 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SETATTR, 1, &size,
265 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
266 memcpy(&body->oa, oa, sizeof(*oa));
268 request->rq_replen = lustre_msg_size(1, &size);
270 rc = ptlrpc_queue_wait(request);
274 body = lustre_swab_repbuf(request, 0, sizeof(*body),
275 lustre_swab_ost_body);
277 GOTO(out, rc = -EPROTO);
279 memcpy(oa, &body->oa, sizeof(*oa));
283 ptlrpc_req_finished(request);
287 int osc_real_create(struct obd_export *exp, struct obdo *oa,
288 struct lov_stripe_md **ea, struct obd_trans_info *oti)
290 struct ptlrpc_request *request;
291 struct ost_body *body;
292 struct lov_stripe_md *lsm;
293 int rc, size = sizeof(*body);
301 rc = obd_alloc_memmd(exp, &lsm);
306 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_CREATE, 1, &size,
309 GOTO(out, rc = -ENOMEM);
311 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
312 memcpy(&body->oa, oa, sizeof(body->oa));
314 request->rq_replen = lustre_msg_size(1, &size);
315 if (oa->o_valid & OBD_MD_FLINLINE) {
316 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
317 oa->o_flags == OBD_FL_DELORPHAN);
318 DEBUG_REQ(D_HA, request,
319 "delorphan from OST integration");
320 /* Don't resend the delorphan request */
321 request->rq_no_resend = request->rq_no_delay = 1;
324 rc = ptlrpc_queue_wait(request);
328 body = lustre_swab_repbuf(request, 0, sizeof(*body),
329 lustre_swab_ost_body);
331 CERROR ("can't unpack ost_body\n");
332 GOTO (out_req, rc = -EPROTO);
335 memcpy(oa, &body->oa, sizeof(*oa));
337 /* This should really be sent by the OST */
338 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
339 oa->o_valid |= OBD_MD_FLBLKSZ;
341 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
342 * have valid lsm_oinfo data structs, so don't go touching that.
343 * This needs to be fixed in a big way.
345 lsm->lsm_object_id = oa->o_id;
346 lsm->lsm_object_gr = oa->o_gr;
350 oti->oti_transno = request->rq_repmsg->transno;
352 if (oa->o_valid & OBD_MD_FLCOOKIE) {
353 if (!oti->oti_logcookies)
354 oti_alloc_cookies(oti, 1);
355 memcpy(oti->oti_logcookies, obdo_logcookie(oa),
356 sizeof(oti->oti_onecookie));
360 CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
363 ptlrpc_req_finished(request);
366 obd_free_memmd(exp, &lsm);
370 static int osc_punch(struct obd_export *exp, struct obdo *oa,
371 struct lov_stripe_md *md, obd_size start,
372 obd_size end, struct obd_trans_info *oti)
374 struct ptlrpc_request *request;
375 struct ost_body *body;
376 int rc, size = sizeof(*body);
384 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_PUNCH, 1, &size,
389 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
390 memcpy(&body->oa, oa, sizeof(*oa));
392 /* overload the size and blocks fields in the oa with start/end */
393 body->oa.o_size = start;
394 body->oa.o_blocks = end;
395 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
397 request->rq_replen = lustre_msg_size(1, &size);
399 rc = ptlrpc_queue_wait(request);
403 body = lustre_swab_repbuf (request, 0, sizeof (*body),
404 lustre_swab_ost_body);
406 CERROR ("can't unpack ost_body\n");
407 GOTO (out, rc = -EPROTO);
410 memcpy(oa, &body->oa, sizeof(*oa));
414 ptlrpc_req_finished(request);
418 static int osc_sync(struct obd_export *exp, struct obdo *oa,
419 struct lov_stripe_md *md, obd_size start, obd_size end)
421 struct ptlrpc_request *request;
422 struct ost_body *body;
423 int rc, size = sizeof(*body);
431 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SYNC, 1, &size,
436 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
437 memcpy(&body->oa, oa, sizeof(*oa));
439 /* overload the size and blocks fields in the oa with start/end */
440 body->oa.o_size = start;
441 body->oa.o_blocks = end;
442 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
444 request->rq_replen = lustre_msg_size(1, &size);
446 rc = ptlrpc_queue_wait(request);
450 body = lustre_swab_repbuf(request, 0, sizeof(*body),
451 lustre_swab_ost_body);
453 CERROR ("can't unpack ost_body\n");
454 GOTO (out, rc = -EPROTO);
457 memcpy(oa, &body->oa, sizeof(*oa));
461 ptlrpc_req_finished(request);
465 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
466 struct lov_stripe_md *ea, struct obd_trans_info *oti)
468 struct ptlrpc_request *request;
469 struct ost_body *body;
470 int rc, size = sizeof(*body);
478 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_DESTROY, 1,
483 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
485 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
486 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
487 sizeof(*oti->oti_logcookies));
488 oti->oti_logcookies++;
491 memcpy(&body->oa, oa, sizeof(*oa));
492 request->rq_replen = lustre_msg_size(1, &size);
494 rc = ptlrpc_queue_wait(request);
498 body = lustre_swab_repbuf(request, 0, sizeof(*body),
499 lustre_swab_ost_body);
501 CERROR ("Can't unpack body\n");
502 GOTO (out, rc = -EPROTO);
505 memcpy(oa, &body->oa, sizeof(*oa));
509 ptlrpc_req_finished(request);
513 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
516 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
518 LASSERT(!(oa->o_valid & bits));
521 spin_lock(&cli->cl_loi_list_lock);
522 oa->o_dirty = cli->cl_dirty;
523 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
524 oa->o_grant = cli->cl_avail_grant;
525 oa->o_dropped = cli->cl_lost_grant;
526 cli->cl_lost_grant = 0;
527 spin_unlock(&cli->cl_loi_list_lock);
528 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
529 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
532 /* caller must hold loi_list_lock */
533 static void osc_consume_write_grant(struct client_obd *cli,
534 struct osc_async_page *oap)
536 cli->cl_dirty += PAGE_SIZE;
537 cli->cl_avail_grant -= PAGE_SIZE;
538 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
539 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
540 LASSERT(cli->cl_avail_grant >= 0);
543 /* caller must hold loi_list_lock */
544 void osc_wake_cache_waiters(struct client_obd *cli)
546 struct list_head *l, *tmp;
547 struct osc_cache_waiter *ocw;
549 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
550 /* if we can't dirty more, we must wait until some is written */
551 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
552 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
553 cli->cl_dirty, cli->cl_dirty_max);
557 /* if still dirty cache but no grant wait for pending RPCs that
558 * may yet return us some grant before doing sync writes */
559 if (cli->cl_brw_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
560 CDEBUG(D_CACHE, "%d BRWs in flight, no grant\n",
561 cli->cl_brw_in_flight);
565 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
566 list_del_init(&ocw->ocw_entry);
567 if (cli->cl_avail_grant < PAGE_SIZE) {
568 /* no more RPCs in flight to return grant, do sync IO */
569 ocw->ocw_rc = -EDQUOT;
570 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
572 osc_consume_write_grant(cli, ocw->ocw_oap);
575 wake_up(&ocw->ocw_waitq);
581 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
583 spin_lock(&cli->cl_loi_list_lock);
584 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
585 cli->cl_avail_grant += body->oa.o_grant;
586 /* waiters are woken in brw_interpret_oap */
587 spin_unlock(&cli->cl_loi_list_lock);
590 /* We assume that the reason this OSC got a short read is because it read
591 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
592 * via the LOV, and it _knows_ it's reading inside the file, it's just that
593 * this stripe never got written at or beyond this stripe offset yet. */
594 static void handle_short_read(int nob_read, obd_count page_count,
595 struct brw_page *pga)
599 /* skip bytes read OK */
600 while (nob_read > 0) {
601 LASSERT (page_count > 0);
603 if (pga->count > nob_read) {
604 /* EOF inside this page */
605 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
606 memset(ptr + nob_read, 0, pga->count - nob_read);
613 nob_read -= pga->count;
618 /* zero remaining pages */
619 while (page_count-- > 0) {
620 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
621 memset(ptr, 0, pga->count);
627 static int check_write_rcs(struct ptlrpc_request *request,
628 int requested_nob, int niocount,
629 obd_count page_count, struct brw_page *pga)
633 /* return error if any niobuf was in error */
634 remote_rcs = lustre_swab_repbuf(request, 1,
635 sizeof(*remote_rcs) * niocount, NULL);
636 if (remote_rcs == NULL) {
637 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
640 if (lustre_msg_swabbed(request->rq_repmsg))
641 for (i = 0; i < niocount; i++)
642 __swab32s(&remote_rcs[i]);
644 for (i = 0; i < niocount; i++) {
645 if (remote_rcs[i] < 0)
646 return(remote_rcs[i]);
648 if (remote_rcs[i] != 0) {
649 CERROR("rc[%d] invalid (%d) req %p\n",
650 i, remote_rcs[i], request);
655 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
656 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
657 requested_nob, request->rq_bulk->bd_nob_transferred);
664 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
666 if (p1->flag != p2->flag) {
667 unsigned mask = ~OBD_BRW_FROM_GRANT;
669 /* warn if we try to combine flags that we don't know to be
671 if ((p1->flag & mask) != (p2->flag & mask))
672 CERROR("is it ok to have flags 0x%x and 0x%x in the "
673 "same brw?\n", p1->flag, p2->flag);
677 return (p1->off + p1->count == p2->off);
681 static obd_count cksum_pages(int nob, obd_count page_count,
682 struct brw_page *pga)
688 LASSERT (page_count > 0);
691 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
692 pga->count > nob ? nob : pga->count);
704 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
705 struct lov_stripe_md *lsm, obd_count page_count,
706 struct brw_page *pga, int *requested_nobp,
707 int *niocountp, struct ptlrpc_request **reqp)
709 struct ptlrpc_request *req;
710 struct ptlrpc_bulk_desc *desc;
711 struct client_obd *cli = &imp->imp_obd->u.cli;
712 struct ost_body *body;
713 struct obd_ioobj *ioobj;
714 struct niobuf_remote *niobuf;
723 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
725 for (niocount = i = 1; i < page_count; i++)
726 if (!can_merge_pages(&pga[i - 1], &pga[i]))
729 size[0] = sizeof(*body);
730 size[1] = sizeof(*ioobj);
731 size[2] = niocount * sizeof(*niobuf);
733 req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
737 if (opc == OST_WRITE)
738 desc = ptlrpc_prep_bulk_imp (req, page_count,
739 BULK_GET_SOURCE, OST_BULK_PORTAL);
741 desc = ptlrpc_prep_bulk_imp (req, page_count,
742 BULK_PUT_SINK, OST_BULK_PORTAL);
744 GOTO(out, rc = -ENOMEM);
745 /* NB request now owns desc and will free it when it gets freed */
747 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
748 ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
749 niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
751 memcpy(&body->oa, oa, sizeof(*oa));
753 obdo_to_ioobj(oa, ioobj);
754 ioobj->ioo_bufcnt = niocount;
756 LASSERT (page_count > 0);
757 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
758 struct brw_page *pg = &pga[i];
759 struct brw_page *pg_prev = pg - 1;
761 LASSERT(pg->count > 0);
762 LASSERT((pg->off & ~PAGE_MASK)+ pg->count <= PAGE_SIZE);
763 LASSERTF(i == 0 || pg->off > pg_prev->off,
764 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
765 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
767 pg->pg, pg->pg->private, pg->pg->index, pg->off,
768 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
771 ptlrpc_prep_bulk_page(desc, pg->pg,
772 pg->off & ~PAGE_MASK, pg->count);
773 requested_nob += pg->count;
775 if (i > 0 && can_merge_pages(pg_prev, pg)) {
777 niobuf->len += pg->count;
779 niobuf->offset = pg->off;
780 niobuf->len = pg->count;
781 niobuf->flags = pg->flag;
785 LASSERT((void *)(niobuf - niocount) ==
786 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
787 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
789 /* size[0] still sizeof (*body) */
790 if (opc == OST_WRITE) {
792 body->oa.o_valid |= OBD_MD_FLCKSUM;
793 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
795 /* 1 RC per niobuf */
796 size[1] = sizeof(__u32) * niocount;
797 req->rq_replen = lustre_msg_size(2, size);
799 /* 1 RC for the whole I/O */
800 req->rq_replen = lustre_msg_size(1, size);
803 *niocountp = niocount;
804 *requested_nobp = requested_nob;
809 ptlrpc_req_finished (req);
813 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
814 int requested_nob, int niocount,
815 obd_count page_count, struct brw_page *pga,
818 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
819 struct ost_body *body;
825 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
827 CERROR ("Can't unpack body\n");
831 osc_update_grant(cli, body);
832 memcpy(oa, &body->oa, sizeof(*oa));
834 if (req->rq_reqmsg->opc == OST_WRITE) {
836 CERROR ("Unexpected +ve rc %d\n", rc);
839 LASSERT (req->rq_bulk->bd_nob == requested_nob);
841 RETURN(check_write_rcs(req, requested_nob, niocount,
845 if (rc > requested_nob) {
846 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
850 if (rc != req->rq_bulk->bd_nob_transferred) {
851 CERROR ("Unexpected rc %d (%d transferred)\n",
852 rc, req->rq_bulk->bd_nob_transferred);
856 if (rc < requested_nob)
857 handle_short_read(rc, page_count, pga);
860 if (oa->o_valid & OBD_MD_FLCKSUM) {
861 const struct ptlrpc_peer *peer =
862 &req->rq_import->imp_connection->c_peer;
863 static int cksum_counter;
864 obd_count server_cksum = oa->o_cksum;
865 obd_count cksum = cksum_pages(rc, page_count, pga);
866 char str[PTL_NALFMT_SIZE];
868 ptlrpc_peernid2str(peer, str);
871 if (server_cksum != cksum) {
872 CERROR("Bad checksum: server %x, client %x, server NID "
873 LPX64" (%s)\n", server_cksum, cksum,
874 peer->peer_nid, str);
877 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
878 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
879 cksum_counter, peer->peer_nid, str, cksum);
882 static int cksum_missed;
885 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
886 CERROR("Request checksum %u from "LPX64", no reply\n",
888 req->rq_import->imp_connection->c_peer.peer_nid);
894 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
895 struct lov_stripe_md *lsm,
896 obd_count page_count, struct brw_page *pga)
900 struct ptlrpc_request *request;
905 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
906 page_count, pga, &requested_nob, &niocount,
911 rc = ptlrpc_queue_wait(request);
913 if (rc == -ETIMEDOUT && request->rq_resend) {
914 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
915 ptlrpc_req_finished(request);
919 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
920 page_count, pga, rc);
922 ptlrpc_req_finished(request);
926 static int brw_interpret(struct ptlrpc_request *request,
927 struct osc_brw_async_args *aa, int rc)
929 struct obdo *oa = aa->aa_oa;
930 int requested_nob = aa->aa_requested_nob;
931 int niocount = aa->aa_nio_count;
932 obd_count page_count = aa->aa_page_count;
933 struct brw_page *pga = aa->aa_pga;
936 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
937 page_count, pga, rc);
941 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
942 struct lov_stripe_md *lsm, obd_count page_count,
943 struct brw_page *pga, struct ptlrpc_request_set *set)
945 struct ptlrpc_request *request;
948 struct osc_brw_async_args *aa;
952 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
953 page_count, pga, &requested_nob, &nio_count,
956 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
957 aa = (struct osc_brw_async_args *)&request->rq_async_args;
959 aa->aa_requested_nob = requested_nob;
960 aa->aa_nio_count = nio_count;
961 aa->aa_page_count = page_count;
964 request->rq_interpret_reply = brw_interpret;
965 ptlrpc_set_add_req(set, request);
971 #define min_t(type,x,y) \
972 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
976 * ugh, we want disk allocation on the target to happen in offset order. we'll
977 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
978 * fine for our small page arrays and doesn't require allocation. its an
979 * insertion sort that swaps elements that are strides apart, shrinking the
980 * stride down until its '1' and the array is sorted.
982 static void sort_brw_pages(struct brw_page *array, int num)
989 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
994 for (i = stride ; i < num ; i++) {
997 while (j >= stride && array[j - stride].off >
999 array[j] = array[j - stride];
1004 } while (stride > 1);
1007 /* make sure we the regions we're passing to elan don't violate its '4
1008 * fragments' constraint. portal headers are a fragment, all full
1009 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1010 * counts as a fragment. I think. see bug 934. */
1011 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1014 int saw_whole_frag = 0;
1017 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1018 if (pg->count == PAGE_SIZE) {
1019 if (!saw_whole_frag) {
1030 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1031 struct lov_stripe_md *md, obd_count page_count,
1032 struct brw_page *pga, struct obd_trans_info *oti)
1036 if (cmd == OBD_BRW_CHECK) {
1037 /* The caller just wants to know if there's a chance that this
1038 * I/O can succeed */
1039 struct obd_import *imp = class_exp2cliimp(exp);
1041 if (imp == NULL || imp->imp_invalid)
1046 while (page_count) {
1047 obd_count pages_per_brw;
1050 if (page_count > PTLRPC_MAX_BRW_PAGES)
1051 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1053 pages_per_brw = page_count;
1055 sort_brw_pages(pga, pages_per_brw);
1056 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1058 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1063 page_count -= pages_per_brw;
1064 pga += pages_per_brw;
1069 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1070 struct lov_stripe_md *md, obd_count page_count,
1071 struct brw_page *pga, struct ptlrpc_request_set *set,
1072 struct obd_trans_info *oti)
1076 if (cmd == OBD_BRW_CHECK) {
1077 /* The caller just wants to know if there's a chance that this
1078 * I/O can succeed */
1079 struct obd_import *imp = class_exp2cliimp(exp);
1081 if (imp == NULL || imp->imp_invalid)
1086 while (page_count) {
1087 obd_count pages_per_brw;
1090 if (page_count > PTLRPC_MAX_BRW_PAGES)
1091 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1093 pages_per_brw = page_count;
1095 sort_brw_pages(pga, pages_per_brw);
1096 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1098 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1103 page_count -= pages_per_brw;
1104 pga += pages_per_brw;
1109 static void osc_check_rpcs(struct client_obd *cli);
1110 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1112 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1113 static void lop_update_pending(struct client_obd *cli,
1114 struct loi_oap_pages *lop, int cmd, int delta);
1116 /* this is called when a sync waiter receives an interruption. Its job is to
1117 * get the caller woken as soon as possible. If its page hasn't been put in an
1118 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1119 * desiring interruption which will forcefully complete the rpc once the rpc
1121 static void osc_occ_interrupted(struct oig_callback_context *occ)
1123 struct osc_async_page *oap;
1124 struct loi_oap_pages *lop;
1125 struct lov_oinfo *loi;
1128 /* XXX member_of() */
1129 oap = list_entry(occ, struct osc_async_page, oap_occ);
1131 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1133 oap->oap_interrupted = 1;
1135 /* ok, it's been put in an rpc. */
1136 if (oap->oap_request != NULL) {
1137 ptlrpc_mark_interrupted(oap->oap_request);
1138 ptlrpcd_wake(oap->oap_request);
1142 /* we don't get interruption callbacks until osc_trigger_sync_io()
1143 * has been called and put the sync oaps in the pending/urgent lists.*/
1144 if (!list_empty(&oap->oap_pending_item)) {
1145 list_del_init(&oap->oap_pending_item);
1146 if (oap->oap_async_flags & ASYNC_URGENT)
1147 list_del_init(&oap->oap_urgent_item);
1150 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1151 &loi->loi_write_lop : &loi->loi_read_lop;
1152 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1153 loi_list_maint(oap->oap_cli, oap->oap_loi);
1155 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1156 oap->oap_oig = NULL;
1160 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1163 /* this must be called holding the loi list lock to give coverage to exit_cache,
1164 * async_flag maintenance, and oap_request */
1165 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1166 struct osc_async_page *oap, int sent, int rc)
1168 osc_exit_cache(cli, oap, sent);
1169 oap->oap_async_flags = 0;
1170 oap->oap_interrupted = 0;
1172 if (oap->oap_request != NULL) {
1173 ptlrpc_req_finished(oap->oap_request);
1174 oap->oap_request = NULL;
1177 if (rc == 0 && oa != NULL)
1178 oap->oap_loi->loi_blocks = oa->o_blocks;
1181 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1182 oap->oap_oig = NULL;
1187 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1191 static int brw_interpret_oap(struct ptlrpc_request *request,
1192 struct osc_brw_async_args *aa, int rc)
1194 struct osc_async_page *oap;
1195 struct client_obd *cli;
1196 struct list_head *pos, *n;
1200 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1201 aa->aa_nio_count, aa->aa_page_count,
1204 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1207 /* in failout recovery we ignore writeback failure and want
1208 * to just tell llite to unlock the page and continue */
1209 if (request->rq_reqmsg->opc == OST_WRITE &&
1210 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1211 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1213 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1217 spin_lock(&cli->cl_loi_list_lock);
1219 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1220 * is called so we know whether to go to sync BRWs or wait for more
1221 * RPCs to complete */
1222 cli->cl_brw_in_flight--;
1224 /* the caller may re-use the oap after the completion call so
1225 * we need to clean it up a little */
1226 list_for_each_safe(pos, n, &aa->aa_oaps) {
1227 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1229 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1230 //oap->oap_page, oap->oap_page->index, oap);
1232 list_del_init(&oap->oap_rpc_item);
1233 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1236 osc_wake_cache_waiters(cli);
1237 osc_check_rpcs(cli);
1239 spin_unlock(&cli->cl_loi_list_lock);
1241 obdo_free(aa->aa_oa);
1242 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1247 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1248 struct list_head *rpc_list,
1249 int page_count, int cmd)
1251 struct ptlrpc_request *req;
1252 struct brw_page *pga = NULL;
1253 int requested_nob, nio_count;
1254 struct osc_brw_async_args *aa;
1255 struct obdo *oa = NULL;
1256 struct obd_async_page_ops *ops = NULL;
1257 void *caller_data = NULL;
1258 struct list_head *pos;
1261 LASSERT(!list_empty(rpc_list));
1263 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1265 RETURN(ERR_PTR(-ENOMEM));
1269 GOTO(out, req = ERR_PTR(-ENOMEM));
1272 list_for_each(pos, rpc_list) {
1273 struct osc_async_page *oap;
1275 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1277 ops = oap->oap_caller_ops;
1278 caller_data = oap->oap_caller_data;
1280 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1281 pga[i].pg = oap->oap_page;
1282 pga[i].count = oap->oap_count;
1283 pga[i].flag = oap->oap_brw_flags;
1284 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1285 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1289 /* always get the data for the obdo for the rpc */
1290 LASSERT(ops != NULL);
1291 ops->ap_fill_obdo(caller_data, cmd, oa);
1293 sort_brw_pages(pga, page_count);
1294 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1295 pga, &requested_nob, &nio_count, &req);
1297 CERROR("prep_req failed: %d\n", rc);
1298 GOTO(out, req = ERR_PTR(rc));
1301 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1302 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1304 aa->aa_requested_nob = requested_nob;
1305 aa->aa_nio_count = nio_count;
1306 aa->aa_page_count = page_count;
1315 OBD_FREE(pga, sizeof(*pga) * page_count);
1320 static void lop_update_pending(struct client_obd *cli,
1321 struct loi_oap_pages *lop, int cmd, int delta)
1323 lop->lop_num_pending += delta;
1324 if (cmd == OBD_BRW_WRITE)
1325 cli->cl_pending_w_pages += delta;
1327 cli->cl_pending_r_pages += delta;
1330 /* the loi lock is held across this function but it's allowed to release
1331 * and reacquire it during its work */
1332 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1333 int cmd, struct loi_oap_pages *lop)
1335 struct ptlrpc_request *request;
1336 obd_count page_count = 0;
1337 struct list_head *tmp, *pos;
1338 struct osc_async_page *oap = NULL;
1339 struct osc_brw_async_args *aa;
1340 struct obd_async_page_ops *ops;
1341 LIST_HEAD(rpc_list);
1344 /* first we find the pages we're allowed to work with */
1345 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1346 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1347 ops = oap->oap_caller_ops;
1349 LASSERT(oap->oap_magic == OAP_MAGIC);
1351 /* in llite being 'ready' equates to the page being locked
1352 * until completion unlocks it. commit_write submits a page
1353 * as not ready because its unlock will happen unconditionally
1354 * as the call returns. if we race with commit_write giving
1355 * us that page we dont' want to create a hole in the page
1356 * stream, so we stop and leave the rpc to be fired by
1357 * another dirtier or kupdated interval (the not ready page
1358 * will still be on the dirty list). we could call in
1359 * at the end of ll_file_write to process the queue again. */
1360 if (!(oap->oap_async_flags & ASYNC_READY)) {
1361 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1363 CDEBUG(D_INODE, "oap %p page %p returned %d "
1364 "instead of ready\n", oap,
1368 /* llite is telling us that the page is still
1369 * in commit_write and that we should try
1370 * and put it in an rpc again later. we
1371 * break out of the loop so we don't create
1372 * a hole in the sequence of pages in the rpc
1377 /* the io isn't needed.. tell the checks
1378 * below to complete the rpc with EINTR */
1379 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1380 oap->oap_count = -EINTR;
1383 oap->oap_async_flags |= ASYNC_READY;
1386 LASSERTF(0, "oap %p page %p returned %d "
1387 "from make_ready\n", oap,
1395 /* take the page out of our book-keeping */
1396 list_del_init(&oap->oap_pending_item);
1397 lop_update_pending(cli, lop, cmd, -1);
1398 list_del_init(&oap->oap_urgent_item);
1400 /* ask the caller for the size of the io as the rpc leaves. */
1401 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1403 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1404 if (oap->oap_count <= 0) {
1405 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1407 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1411 /* now put the page back in our accounting */
1412 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1413 if (++page_count >= cli->cl_max_pages_per_rpc)
1417 osc_wake_cache_waiters(cli);
1419 if (page_count == 0)
1422 loi_list_maint(cli, loi);
1423 spin_unlock(&cli->cl_loi_list_lock);
1425 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1426 if (IS_ERR(request)) {
1427 /* this should happen rarely and is pretty bad, it makes the
1428 * pending list not follow the dirty order */
1429 spin_lock(&cli->cl_loi_list_lock);
1430 list_for_each_safe(pos, tmp, &rpc_list) {
1431 oap = list_entry(pos, struct osc_async_page,
1433 list_del_init(&oap->oap_rpc_item);
1435 /* queued sync pages can be torn down while the pages
1436 * were between the pending list and the rpc */
1437 if (oap->oap_interrupted) {
1438 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1439 osc_ap_completion(cli, NULL, oap, 0,
1444 /* put the page back in the loi/lop lists */
1445 list_add_tail(&oap->oap_pending_item,
1447 lop_update_pending(cli, lop, cmd, 1);
1448 if (oap->oap_async_flags & ASYNC_URGENT)
1449 list_add(&oap->oap_urgent_item,
1452 loi_list_maint(cli, loi);
1453 RETURN(PTR_ERR(request));
1456 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1457 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1458 INIT_LIST_HEAD(&aa->aa_oaps);
1459 list_splice(&rpc_list, &aa->aa_oaps);
1460 INIT_LIST_HEAD(&rpc_list);
1463 if (cmd == OBD_BRW_READ) {
1464 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1465 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1467 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1468 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1469 cli->cl_brw_in_flight);
1473 spin_lock(&cli->cl_loi_list_lock);
1475 cli->cl_brw_in_flight++;
1476 /* queued sync pages can be torn down while the pages
1477 * were between the pending list and the rpc */
1478 list_for_each(pos, &aa->aa_oaps) {
1479 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1480 if (oap->oap_interrupted) {
1481 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1483 ptlrpc_mark_interrupted(request);
1488 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %d in flight\n", request,
1489 page_count, aa, cli->cl_brw_in_flight);
1491 oap->oap_request = ptlrpc_request_addref(request);
1492 request->rq_interpret_reply = brw_interpret_oap;
1493 ptlrpcd_add_req(request);
1497 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1503 if (lop->lop_num_pending == 0)
1506 /* if we have an invalid import we want to drain the queued pages
1507 * by forcing them through rpcs that immediately fail and complete
1508 * the pages. recovery relies on this to empty the queued pages
1509 * before canceling the locks and evicting down the llite pages */
1510 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1513 /* stream rpcs in queue order as long as as there is an urgent page
1514 * queued. this is our cheap solution for good batching in the case
1515 * where writepage marks some random page in the middle of the file as
1516 * urgent because of, say, memory pressure */
1517 if (!list_empty(&lop->lop_urgent))
1520 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1521 optimal = cli->cl_max_pages_per_rpc;
1522 if (cmd == OBD_BRW_WRITE) {
1523 /* trigger a write rpc stream as long as there are dirtiers
1524 * waiting for space. as they're waiting, they're not going to
1525 * create more pages to coallesce with what's waiting.. */
1526 if (!list_empty(&cli->cl_cache_waiters))
1529 /* *2 to avoid triggering rpcs that would want to include pages
1530 * that are being queued but which can't be made ready until
1531 * the queuer finishes with the page. this is a wart for
1532 * llite::commit_write() */
1535 if (lop->lop_num_pending >= optimal)
1541 static void on_list(struct list_head *item, struct list_head *list,
1544 if (list_empty(item) && should_be_on)
1545 list_add_tail(item, list);
1546 else if (!list_empty(item) && !should_be_on)
1547 list_del_init(item);
1550 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1551 * can find pages to build into rpcs quickly */
1552 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1554 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1555 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1556 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1558 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1559 loi->loi_write_lop.lop_num_pending);
1561 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1562 loi->loi_read_lop.lop_num_pending);
1565 #define LOI_DEBUG(LOI, STR, args...) \
1566 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1567 !list_empty(&(LOI)->loi_cli_item), \
1568 (LOI)->loi_write_lop.lop_num_pending, \
1569 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1570 (LOI)->loi_read_lop.lop_num_pending, \
1571 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1574 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1577 /* first return all objects which we already know to have
1578 * pages ready to be stuffed into rpcs */
1579 if (!list_empty(&cli->cl_loi_ready_list))
1580 RETURN(list_entry(cli->cl_loi_ready_list.next,
1581 struct lov_oinfo, loi_cli_item));
1583 /* then if we have cache waiters, return all objects with queued
1584 * writes. This is especially important when many small files
1585 * have filled up the cache and not been fired into rpcs because
1586 * they don't pass the nr_pending/object threshhold */
1587 if (!list_empty(&cli->cl_cache_waiters) &&
1588 !list_empty(&cli->cl_loi_write_list))
1589 RETURN(list_entry(cli->cl_loi_write_list.next,
1590 struct lov_oinfo, loi_write_item));
1592 /* then return all queued objects when we have an invalid import
1593 * so that they get flushed */
1594 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1595 if (!list_empty(&cli->cl_loi_write_list))
1596 RETURN(list_entry(cli->cl_loi_write_list.next,
1597 struct lov_oinfo, loi_write_item));
1598 if (!list_empty(&cli->cl_loi_read_list))
1599 RETURN(list_entry(cli->cl_loi_read_list.next,
1600 struct lov_oinfo, loi_read_item));
1605 /* called with the loi list lock held */
1606 static void osc_check_rpcs(struct client_obd *cli)
1608 struct lov_oinfo *loi;
1609 int rc = 0, race_counter = 0;
1612 while ((loi = osc_next_loi(cli)) != NULL) {
1613 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1615 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1618 /* attempt some read/write balancing by alternating between
1619 * reads and writes in an object. The makes_rpc checks here
1620 * would be redundant if we were getting read/write work items
1621 * instead of objects. we don't want send_oap_rpc to drain a
1622 * partial read pending queue when we're given this object to
1623 * do io on writes while there are cache waiters */
1624 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1625 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1626 &loi->loi_write_lop);
1634 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1635 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1636 &loi->loi_read_lop);
1645 /* attempt some inter-object balancing by issueing rpcs
1646 * for each object in turn */
1647 if (!list_empty(&loi->loi_cli_item))
1648 list_del_init(&loi->loi_cli_item);
1649 if (!list_empty(&loi->loi_write_item))
1650 list_del_init(&loi->loi_write_item);
1651 if (!list_empty(&loi->loi_read_item))
1652 list_del_init(&loi->loi_read_item);
1654 loi_list_maint(cli, loi);
1656 /* send_oap_rpc fails with 0 when make_ready tells it to
1657 * back off. llite's make_ready does this when it tries
1658 * to lock a page queued for write that is already locked.
1659 * we want to try sending rpcs from many objects, but we
1660 * don't want to spin failing with 0. */
1661 if (race_counter == 10)
1667 /* we're trying to queue a page in the osc so we're subject to the
1668 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1669 * If the osc's queued pages are already at that limit, then we want to sleep
1670 * until there is space in the osc's queue for us. We also may be waiting for
1671 * write credits from the OST if there are RPCs in flight that may return some
1672 * before we fall back to sync writes.
1674 * We need this know our allocation was granted in the presence of signals */
1675 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1679 spin_lock(&cli->cl_loi_list_lock);
1680 rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1681 spin_unlock(&cli->cl_loi_list_lock);
1685 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1686 * grant or cache space. */
1687 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1688 struct osc_async_page *oap)
1690 struct osc_cache_waiter ocw;
1691 struct l_wait_info lwi = { 0 };
1693 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1694 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1695 cli->cl_avail_grant);
1697 if (cli->cl_dirty_max < PAGE_SIZE)
1700 /* Hopefully normal case - cache space and write credits available */
1701 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1702 cli->cl_avail_grant >= PAGE_SIZE) {
1703 /* account for ourselves */
1704 osc_consume_write_grant(cli, oap);
1708 /* Make sure that there are write rpcs in flight to wait for. This
1709 * is a little silly as this object may not have any pending but
1710 * other objects sure might. */
1711 if (cli->cl_brw_in_flight) {
1712 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1713 init_waitqueue_head(&ocw.ocw_waitq);
1717 loi_list_maint(cli, loi);
1718 osc_check_rpcs(cli);
1719 spin_unlock(&cli->cl_loi_list_lock);
1721 CDEBUG(0, "sleeping for cache space\n");
1722 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1724 spin_lock(&cli->cl_loi_list_lock);
1725 if (!list_empty(&ocw.ocw_entry)) {
1726 list_del(&ocw.ocw_entry);
1735 /* the companion to enter_cache, called when an oap is no longer part of the
1736 * dirty accounting.. so writeback completes or truncate happens before writing
1737 * starts. must be called with the loi lock held. */
1738 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1743 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1748 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1749 cli->cl_dirty -= PAGE_SIZE;
1751 cli->cl_lost_grant += PAGE_SIZE;
1752 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1753 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1759 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1760 struct lov_oinfo *loi, struct page *page,
1761 obd_off offset, struct obd_async_page_ops *ops,
1762 void *data, void **res)
1764 struct osc_async_page *oap;
1767 OBD_ALLOC(oap, sizeof(*oap));
1771 oap->oap_magic = OAP_MAGIC;
1772 oap->oap_cli = &exp->exp_obd->u.cli;
1775 oap->oap_caller_ops = ops;
1776 oap->oap_caller_data = data;
1778 oap->oap_page = page;
1779 oap->oap_obj_off = offset;
1781 INIT_LIST_HEAD(&oap->oap_pending_item);
1782 INIT_LIST_HEAD(&oap->oap_urgent_item);
1783 INIT_LIST_HEAD(&oap->oap_rpc_item);
1785 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1787 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1792 struct osc_async_page *oap_from_cookie(void *cookie)
1794 struct osc_async_page *oap = cookie;
1795 if (oap->oap_magic != OAP_MAGIC)
1796 return ERR_PTR(-EINVAL);
1800 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1801 struct lov_oinfo *loi, void *cookie,
1802 int cmd, obd_off off, int count,
1803 obd_flag brw_flags, enum async_flags async_flags)
1805 struct client_obd *cli = &exp->exp_obd->u.cli;
1806 struct osc_async_page *oap;
1807 struct loi_oap_pages *lop;
1811 oap = oap_from_cookie(cookie);
1813 RETURN(PTR_ERR(oap));
1815 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1818 if (!list_empty(&oap->oap_pending_item) ||
1819 !list_empty(&oap->oap_urgent_item) ||
1820 !list_empty(&oap->oap_rpc_item))
1824 loi = &lsm->lsm_oinfo[0];
1826 spin_lock(&cli->cl_loi_list_lock);
1829 oap->oap_async_flags = async_flags;
1830 oap->oap_page_off = off;
1831 oap->oap_count = count;
1832 oap->oap_brw_flags = brw_flags;
1834 if (cmd == OBD_BRW_WRITE) {
1835 rc = osc_enter_cache(cli, loi, oap);
1837 spin_unlock(&cli->cl_loi_list_lock);
1840 lop = &loi->loi_write_lop;
1842 lop = &loi->loi_read_lop;
1845 if (oap->oap_async_flags & ASYNC_URGENT)
1846 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1847 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1848 lop_update_pending(cli, lop, cmd, 1);
1850 loi_list_maint(cli, loi);
1852 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1855 osc_check_rpcs(cli);
1856 spin_unlock(&cli->cl_loi_list_lock);
1861 /* aka (~was & now & flag), but this is more clear :) */
1862 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1864 static int osc_set_async_flags(struct obd_export *exp,
1865 struct lov_stripe_md *lsm,
1866 struct lov_oinfo *loi, void *cookie,
1867 obd_flag async_flags)
1869 struct client_obd *cli = &exp->exp_obd->u.cli;
1870 struct loi_oap_pages *lop;
1871 struct osc_async_page *oap;
1875 oap = oap_from_cookie(cookie);
1877 RETURN(PTR_ERR(oap));
1879 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1883 loi = &lsm->lsm_oinfo[0];
1885 if (oap->oap_cmd == OBD_BRW_WRITE) {
1886 lop = &loi->loi_write_lop;
1888 lop = &loi->loi_read_lop;
1891 spin_lock(&cli->cl_loi_list_lock);
1893 if (list_empty(&oap->oap_pending_item))
1894 GOTO(out, rc = -EINVAL);
1896 if ((oap->oap_async_flags & async_flags) == async_flags)
1899 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1900 oap->oap_async_flags |= ASYNC_READY;
1902 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1903 if (list_empty(&oap->oap_rpc_item)) {
1904 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1905 loi_list_maint(cli, loi);
1909 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1910 oap->oap_async_flags);
1912 osc_check_rpcs(cli);
1913 spin_unlock(&cli->cl_loi_list_lock);
1917 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1918 struct lov_oinfo *loi,
1919 struct obd_io_group *oig, void *cookie,
1920 int cmd, obd_off off, int count,
1922 obd_flag async_flags)
1924 struct client_obd *cli = &exp->exp_obd->u.cli;
1925 struct osc_async_page *oap;
1926 struct loi_oap_pages *lop;
1929 oap = oap_from_cookie(cookie);
1931 RETURN(PTR_ERR(oap));
1933 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1936 if (!list_empty(&oap->oap_pending_item) ||
1937 !list_empty(&oap->oap_urgent_item) ||
1938 !list_empty(&oap->oap_rpc_item))
1942 loi = &lsm->lsm_oinfo[0];
1944 spin_lock(&cli->cl_loi_list_lock);
1947 oap->oap_page_off = off;
1948 oap->oap_count = count;
1949 oap->oap_brw_flags = brw_flags;
1950 oap->oap_async_flags = async_flags;
1952 if (cmd == OBD_BRW_WRITE)
1953 lop = &loi->loi_write_lop;
1955 lop = &loi->loi_read_lop;
1957 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
1958 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
1960 oig_add_one(oig, &oap->oap_occ);
1963 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
1965 spin_unlock(&cli->cl_loi_list_lock);
1970 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1971 struct loi_oap_pages *lop, int cmd)
1973 struct list_head *pos, *tmp;
1974 struct osc_async_page *oap;
1976 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
1977 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1978 list_del(&oap->oap_pending_item);
1979 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1980 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1981 lop_update_pending(cli, lop, cmd, 1);
1983 loi_list_maint(cli, loi);
1986 static int osc_trigger_group_io(struct obd_export *exp,
1987 struct lov_stripe_md *lsm,
1988 struct lov_oinfo *loi,
1989 struct obd_io_group *oig)
1991 struct client_obd *cli = &exp->exp_obd->u.cli;
1995 loi = &lsm->lsm_oinfo[0];
1997 spin_lock(&cli->cl_loi_list_lock);
1999 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2000 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2002 osc_check_rpcs(cli);
2003 spin_unlock(&cli->cl_loi_list_lock);
2008 static int osc_teardown_async_page(struct obd_export *exp,
2009 struct lov_stripe_md *lsm,
2010 struct lov_oinfo *loi, void *cookie)
2012 struct client_obd *cli = &exp->exp_obd->u.cli;
2013 struct loi_oap_pages *lop;
2014 struct osc_async_page *oap;
2018 oap = oap_from_cookie(cookie);
2020 RETURN(PTR_ERR(oap));
2023 loi = &lsm->lsm_oinfo[0];
2025 if (oap->oap_cmd == OBD_BRW_WRITE) {
2026 lop = &loi->loi_write_lop;
2028 lop = &loi->loi_read_lop;
2031 spin_lock(&cli->cl_loi_list_lock);
2033 if (!list_empty(&oap->oap_rpc_item))
2034 GOTO(out, rc = -EBUSY);
2036 osc_exit_cache(cli, oap, 0);
2037 osc_wake_cache_waiters(cli);
2039 if (!list_empty(&oap->oap_urgent_item)) {
2040 list_del_init(&oap->oap_urgent_item);
2041 oap->oap_async_flags &= ~ASYNC_URGENT;
2043 if (!list_empty(&oap->oap_pending_item)) {
2044 list_del_init(&oap->oap_pending_item);
2045 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2047 loi_list_maint(cli, loi);
2049 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2051 spin_unlock(&cli->cl_loi_list_lock);
2053 OBD_FREE(oap, sizeof(*oap));
2058 /* Note: caller will lock/unlock, and set uptodate on the pages */
2059 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2060 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2061 struct lov_stripe_md *lsm, obd_count page_count,
2062 struct brw_page *pga)
2064 struct ptlrpc_request *request = NULL;
2065 struct ost_body *body;
2066 struct niobuf_remote *nioptr;
2067 struct obd_ioobj *iooptr;
2068 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2072 /* XXX does not handle 'new' brw protocol */
2074 size[1] = sizeof(struct obd_ioobj);
2075 size[2] = page_count * sizeof(*nioptr);
2077 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2082 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2083 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2084 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2085 sizeof(*nioptr) * page_count);
2087 memcpy(&body->oa, oa, sizeof(body->oa));
2089 obdo_to_ioobj(oa, iooptr);
2090 iooptr->ioo_bufcnt = page_count;
2092 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2093 LASSERT(PageLocked(pga[mapped].pg));
2094 LASSERT(mapped == 0 ||
2095 pga[mapped].off > pga[mapped - 1].off);
2097 nioptr->offset = pga[mapped].off;
2098 nioptr->len = pga[mapped].count;
2099 nioptr->flags = pga[mapped].flag;
2102 size[1] = page_count * sizeof(*nioptr);
2103 request->rq_replen = lustre_msg_size(2, size);
2105 rc = ptlrpc_queue_wait(request);
2109 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2110 lustre_swab_ost_body);
2112 CERROR("Can't unpack body\n");
2113 GOTO(out_req, rc = -EPROTO);
2116 memcpy(oa, &body->oa, sizeof(*oa));
2118 swab = lustre_msg_swabbed(request->rq_repmsg);
2119 LASSERT_REPSWAB(request, 1);
2120 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2122 /* nioptr missing or short */
2123 GOTO(out_req, rc = -EPROTO);
2127 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2128 struct page *page = pga[mapped].pg;
2129 struct buffer_head *bh;
2133 lustre_swab_niobuf_remote (nioptr);
2135 /* got san device associated */
2136 LASSERT(exp->exp_obd != NULL);
2137 dev = exp->exp_obd->u.cli.cl_sandev;
2140 if (!nioptr->offset) {
2141 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2142 page->mapping->host->i_ino,
2144 memset(page_address(page), 0, PAGE_SIZE);
2148 if (!page->buffers) {
2149 create_empty_buffers(page, dev, PAGE_SIZE);
2152 clear_bit(BH_New, &bh->b_state);
2153 set_bit(BH_Mapped, &bh->b_state);
2154 bh->b_blocknr = (unsigned long)nioptr->offset;
2156 clear_bit(BH_Uptodate, &bh->b_state);
2158 ll_rw_block(READ, 1, &bh);
2162 /* if buffer already existed, it must be the
2163 * one we mapped before, check it */
2164 LASSERT(!test_bit(BH_New, &bh->b_state));
2165 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2166 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2168 /* wait it's io completion */
2169 if (test_bit(BH_Lock, &bh->b_state))
2172 if (!test_bit(BH_Uptodate, &bh->b_state))
2173 ll_rw_block(READ, 1, &bh);
2177 /* must do syncronous write here */
2179 if (!buffer_uptodate(bh)) {
2187 ptlrpc_req_finished(request);
2191 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2192 struct lov_stripe_md *lsm, obd_count page_count,
2193 struct brw_page *pga)
2195 struct ptlrpc_request *request = NULL;
2196 struct ost_body *body;
2197 struct niobuf_remote *nioptr;
2198 struct obd_ioobj *iooptr;
2199 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2203 size[1] = sizeof(struct obd_ioobj);
2204 size[2] = page_count * sizeof(*nioptr);
2206 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2211 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2212 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2213 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2214 sizeof (*nioptr) * page_count);
2216 memcpy(&body->oa, oa, sizeof(body->oa));
2218 obdo_to_ioobj(oa, iooptr);
2219 iooptr->ioo_bufcnt = page_count;
2222 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2223 LASSERT(PageLocked(pga[mapped].pg));
2224 LASSERT(mapped == 0 ||
2225 pga[mapped].off > pga[mapped - 1].off);
2227 nioptr->offset = pga[mapped].off;
2228 nioptr->len = pga[mapped].count;
2229 nioptr->flags = pga[mapped].flag;
2232 size[1] = page_count * sizeof(*nioptr);
2233 request->rq_replen = lustre_msg_size(2, size);
2235 rc = ptlrpc_queue_wait(request);
2239 swab = lustre_msg_swabbed (request->rq_repmsg);
2240 LASSERT_REPSWAB (request, 1);
2241 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2243 CERROR("absent/short niobuf array\n");
2244 GOTO(out_req, rc = -EPROTO);
2248 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2249 struct page *page = pga[mapped].pg;
2250 struct buffer_head *bh;
2254 lustre_swab_niobuf_remote (nioptr);
2256 /* got san device associated */
2257 LASSERT(exp->exp_obd != NULL);
2258 dev = exp->exp_obd->u.cli.cl_sandev;
2260 if (!page->buffers) {
2261 create_empty_buffers(page, dev, PAGE_SIZE);
2264 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2265 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2266 LASSERT(page->buffers->b_blocknr ==
2267 (unsigned long)nioptr->offset);
2273 /* if buffer locked, wait it's io completion */
2274 if (test_bit(BH_Lock, &bh->b_state))
2277 clear_bit(BH_New, &bh->b_state);
2278 set_bit(BH_Mapped, &bh->b_state);
2280 /* override the block nr */
2281 bh->b_blocknr = (unsigned long)nioptr->offset;
2283 /* we are about to write it, so set it
2285 * page lock should garentee no race condition here */
2286 set_bit(BH_Uptodate, &bh->b_state);
2287 set_bit(BH_Dirty, &bh->b_state);
2289 ll_rw_block(WRITE, 1, &bh);
2291 /* must do syncronous write here */
2293 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2301 ptlrpc_req_finished(request);
2305 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2306 struct lov_stripe_md *lsm, obd_count page_count,
2307 struct brw_page *pga, struct obd_trans_info *oti)
2311 while (page_count) {
2312 obd_count pages_per_brw;
2315 if (page_count > PTLRPC_MAX_BRW_PAGES)
2316 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2318 pages_per_brw = page_count;
2320 if (cmd & OBD_BRW_WRITE)
2321 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2323 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2328 page_count -= pages_per_brw;
2329 pga += pages_per_brw;
2336 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2338 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2340 LASSERT(lock != NULL);
2341 l_lock(&lock->l_resource->lr_namespace->ns_lock);
2343 if (lock->l_ast_data && lock->l_ast_data != data) {
2344 struct inode *new_inode = data;
2345 struct inode *old_inode = lock->l_ast_data;
2346 LASSERTF(old_inode->i_state & I_FREEING,
2347 "Found existing inode %p/%lu/%u state %lu in lock: "
2348 "setting data to %p/%lu/%u\n", old_inode,
2349 old_inode->i_ino, old_inode->i_generation,
2351 new_inode, new_inode->i_ino, new_inode->i_generation);
2354 lock->l_ast_data = data;
2355 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2356 LDLM_LOCK_PUT(lock);
2359 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2360 ldlm_iterator_t replace, void *data)
2362 struct ldlm_res_id res_id = { .name = {0} };
2363 struct obd_device *obd = class_exp2obd(exp);
2365 res_id.name[0] = lsm->lsm_object_id;
2366 res_id.name[2] = lsm->lsm_object_gr;
2367 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2371 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2372 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2373 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2374 void *data, __u32 lvb_len, void *lvb_swabber,
2375 struct lustre_handle *lockh)
2377 struct obd_device *obd = exp->exp_obd;
2378 struct ldlm_res_id res_id = { .name = {0} };
2383 res_id.name[0] = lsm->lsm_object_id;
2384 res_id.name[2] = lsm->lsm_object_gr;
2386 /* Filesystem lock extents are extended to page boundaries so that
2387 * dealing with the page cache is a little smoother. */
2388 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2389 policy->l_extent.end |= ~PAGE_MASK;
2391 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2394 /* Next, search for already existing extent locks that will cover us */
2395 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2398 osc_set_data_with_check(lockh, data);
2399 if (*flags & LDLM_FL_HAS_INTENT) {
2400 /* I would like to be able to ASSERT here that rss <=
2401 * kms, but I can't, for reasons which are explained in
2404 /* We already have a lock, and it's referenced */
2408 /* If we're trying to read, we also search for an existing PW lock. The
2409 * VFS and page cache already protect us locally, so lots of readers/
2410 * writers can share a single PW lock.
2412 * There are problems with conversion deadlocks, so instead of
2413 * converting a read lock to a write lock, we'll just enqueue a new
2416 * At some point we should cancel the read lock instead of making them
2417 * send us a blocking callback, but there are problems with canceling
2418 * locks out from other users right now, too. */
2420 if (mode == LCK_PR) {
2421 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2422 policy, LCK_PW, lockh);
2424 /* FIXME: This is not incredibly elegant, but it might
2425 * be more elegant than adding another parameter to
2426 * lock_match. I want a second opinion. */
2427 ldlm_lock_addref(lockh, LCK_PR);
2428 ldlm_lock_decref(lockh, LCK_PW);
2429 osc_set_data_with_check(lockh, data);
2435 rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, res_id, type,
2436 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2437 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2439 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2440 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2441 lvb.lvb_size, lvb.lvb_blocks);
2442 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2443 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2449 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2450 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2451 int *flags, void *data, struct lustre_handle *lockh)
2453 struct ldlm_res_id res_id = { .name = {0} };
2454 struct obd_device *obd = exp->exp_obd;
2458 res_id.name[0] = lsm->lsm_object_id;
2459 res_id.name[2] = lsm->lsm_object_gr;
2461 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2463 /* Filesystem lock extents are extended to page boundaries so that
2464 * dealing with the page cache is a little smoother */
2465 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2466 policy->l_extent.end |= ~PAGE_MASK;
2468 /* Next, search for already existing extent locks that will cover us */
2469 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2470 policy, mode, lockh);
2472 if (!(*flags & LDLM_FL_TEST_LOCK))
2473 osc_set_data_with_check(lockh, data);
2476 /* If we're trying to read, we also search for an existing PW lock. The
2477 * VFS and page cache already protect us locally, so lots of readers/
2478 * writers can share a single PW lock. */
2479 if (mode == LCK_PR) {
2480 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2481 policy, LCK_PW, lockh);
2482 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2483 /* FIXME: This is not incredibly elegant, but it might
2484 * be more elegant than adding another parameter to
2485 * lock_match. I want a second opinion. */
2486 osc_set_data_with_check(lockh, data);
2487 ldlm_lock_addref(lockh, LCK_PR);
2488 ldlm_lock_decref(lockh, LCK_PW);
2494 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2495 __u32 mode, struct lustre_handle *lockh)
2499 if (mode == LCK_GROUP)
2500 ldlm_lock_decref_and_cancel(lockh, mode);
2502 ldlm_lock_decref(lockh, mode);
2507 static int osc_cancel_unused(struct obd_export *exp,
2508 struct lov_stripe_md *lsm, int flags, void *opaque)
2510 struct obd_device *obd = class_exp2obd(exp);
2511 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2514 res_id.name[0] = lsm->lsm_object_id;
2515 res_id.name[2] = lsm->lsm_object_gr;
2519 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2522 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2523 unsigned long max_age)
2525 struct obd_statfs *msfs;
2526 struct ptlrpc_request *request;
2527 int rc, size = sizeof(*osfs);
2530 /* We could possibly pass max_age in the request (as an absolute
2531 * timestamp or a "seconds.usec ago") so the target can avoid doing
2532 * extra calls into the filesystem if that isn't necessary (e.g.
2533 * during mount that would help a bit). Having relative timestamps
2534 * is not so great if request processing is slow, while absolute
2535 * timestamps are not ideal because they need time synchronization. */
2536 request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2540 request->rq_replen = lustre_msg_size(1, &size);
2541 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2543 rc = ptlrpc_queue_wait(request);
2547 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2548 lustre_swab_obd_statfs);
2550 CERROR("Can't unpack obd_statfs\n");
2551 GOTO(out, rc = -EPROTO);
2554 memcpy(osfs, msfs, sizeof(*osfs));
2558 ptlrpc_req_finished(request);
2562 /* Retrieve object striping information.
2564 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2565 * the maximum number of OST indices which will fit in the user buffer.
2566 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2568 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2570 struct lov_user_md lum, *lumk;
2577 rc = copy_from_user(&lum, lump, sizeof(lum));
2581 if (lum.lmm_magic != LOV_USER_MAGIC)
2584 if (lum.lmm_stripe_count > 0) {
2585 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2586 OBD_ALLOC(lumk, lum_size);
2590 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2591 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2593 lum_size = sizeof(lum);
2597 lumk->lmm_object_id = lsm->lsm_object_id;
2598 lumk->lmm_object_gr = lsm->lsm_object_gr;
2599 lumk->lmm_stripe_count = 1;
2601 if (copy_to_user(lump, lumk, lum_size))
2605 OBD_FREE(lumk, lum_size);
2610 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2611 void *karg, void *uarg)
2613 struct obd_device *obd = exp->exp_obd;
2614 struct obd_ioctl_data *data = karg;
2618 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2621 if (!try_module_get(THIS_MODULE)) {
2622 CERROR("Can't get module. Is it alive?");
2627 case OBD_IOC_LOV_GET_CONFIG: {
2629 struct lov_desc *desc;
2630 struct obd_uuid uuid;
2634 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2635 GOTO(out, err = -EINVAL);
2637 data = (struct obd_ioctl_data *)buf;
2639 if (sizeof(*desc) > data->ioc_inllen1) {
2641 GOTO(out, err = -EINVAL);
2644 if (data->ioc_inllen2 < sizeof(uuid)) {
2646 GOTO(out, err = -EINVAL);
2649 if (data->ioc_inllen3 < sizeof(__u32)) {
2651 GOTO(out, err = -EINVAL);
2654 desc = (struct lov_desc *)data->ioc_inlbuf1;
2655 desc->ld_tgt_count = 1;
2656 desc->ld_active_tgt_count = 1;
2657 desc->ld_default_stripe_count = 1;
2658 desc->ld_default_stripe_size = 0;
2659 desc->ld_default_stripe_offset = 0;
2660 desc->ld_pattern = 0;
2661 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2662 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2663 *((__u32 *)data->ioc_inlbuf3) = 1;
2665 err = copy_to_user((void *)uarg, buf, len);
2668 obd_ioctl_freedata(buf, len);
2671 case LL_IOC_LOV_SETSTRIPE:
2672 err = obd_alloc_memmd(exp, karg);
2676 case LL_IOC_LOV_GETSTRIPE:
2677 err = osc_getstripe(karg, uarg);
2679 case OBD_IOC_CLIENT_RECOVER:
2680 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2685 case IOC_OSC_SET_ACTIVE:
2686 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2690 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2691 GOTO(out, err = -ENOTTY);
2694 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2697 module_put(THIS_MODULE);
2702 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2703 void *key, __u32 *vallen, void *val)
2706 if (!vallen || !val)
2709 if (keylen > strlen("lock_to_stripe") &&
2710 strcmp(key, "lock_to_stripe") == 0) {
2711 __u32 *stripe = val;
2712 *vallen = sizeof(*stripe);
2715 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2716 struct ptlrpc_request *req;
2718 char *bufs[1] = {key};
2720 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2725 req->rq_replen = lustre_msg_size(1, vallen);
2726 rc = ptlrpc_queue_wait(req);
2730 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2731 lustre_swab_ost_last_id);
2732 if (reply == NULL) {
2733 CERROR("Can't unpack OST last ID\n");
2734 GOTO(out, rc = -EPROTO);
2736 *((obd_id *)val) = *reply;
2738 ptlrpc_req_finished(req);
2744 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2745 void *key, obd_count vallen, void *val)
2747 struct ptlrpc_request *req;
2748 struct obd_device *obd = exp->exp_obd;
2749 struct obd_import *imp = class_exp2cliimp(exp);
2750 struct llog_ctxt *ctxt;
2751 int rc, size[2] = {keylen, vallen};
2752 char *bufs[2] = {key, val};
2755 if (keylen == strlen("next_id") &&
2756 memcmp(key, "next_id", strlen("next_id")) == 0) {
2757 if (vallen != sizeof(obd_id))
2759 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
2760 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
2761 exp->exp_obd->obd_name,
2762 obd->u.cli.cl_oscc.oscc_next_id);
2767 if (keylen == strlen("growth_count") &&
2768 memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2769 if (vallen != sizeof(int))
2771 obd->u.cli.cl_oscc.oscc_grow_count = *((int*)val);
2775 if (keylen == strlen("unlinked") &&
2776 memcmp(key, "unlinked", keylen) == 0) {
2777 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2778 spin_lock(&oscc->oscc_lock);
2779 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2780 spin_unlock(&oscc->oscc_lock);
2783 if (keylen == strlen("unrecovery") &&
2784 memcmp(key, "unrecovery", keylen) == 0) {
2785 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2786 spin_lock(&oscc->oscc_lock);
2787 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
2788 spin_unlock(&oscc->oscc_lock);
2791 if (keylen == strlen("initial_recov") &&
2792 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2793 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2794 if (vallen != sizeof(int))
2796 imp->imp_initial_recov = *(int *)val;
2797 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2798 exp->exp_obd->obd_name,
2799 imp->imp_initial_recov);
2803 if (keylen < strlen("mds_conn") ||
2804 memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2807 req = ptlrpc_prep_req(imp, OST_SET_INFO, 2, size, bufs);
2811 req->rq_replen = lustre_msg_size(0, NULL);
2812 rc = ptlrpc_queue_wait(req);
2813 ptlrpc_req_finished(req);
2815 ctxt = llog_get_context(&exp->exp_obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2817 rc = llog_initiator_connect(ctxt);
2822 imp->imp_server_timeout = 1;
2823 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2824 imp->imp_pingable = 1;
2830 static struct llog_operations osc_size_repl_logops = {
2831 lop_cancel: llog_obd_repl_cancel
2834 static struct llog_operations osc_unlink_orig_logops;
2835 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2836 struct obd_device *tgt, int count,
2837 struct llog_catid *catid)
2842 osc_unlink_orig_logops = llog_lvfs_ops;
2843 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2844 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
2845 osc_unlink_orig_logops.lop_add = llog_catalog_add;
2846 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2848 rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
2849 &catid->lci_logid, &osc_unlink_orig_logops);
2853 rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2854 &osc_size_repl_logops);
2858 static int osc_llog_finish(struct obd_device *obd,
2859 struct obd_llogs *llogs, int count)
2864 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
2868 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
2873 static int osc_connect(struct lustre_handle *exph,
2874 struct obd_device *obd, struct obd_uuid *cluuid)
2878 rc = client_connect_import(exph, obd, cluuid);
2883 static int osc_disconnect(struct obd_export *exp, int flags)
2885 struct obd_device *obd = class_exp2obd(exp);
2886 struct llog_ctxt *ctxt;
2889 ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
2890 if (obd->u.cli.cl_conn_count == 1)
2891 /* flush any remaining cancel messages out to the target */
2892 llog_sync(ctxt, exp);
2894 rc = client_disconnect_export(exp, flags);
2898 static int osc_import_event(struct obd_device *obd,
2899 struct obd_import *imp,
2900 enum obd_import_event event)
2902 struct client_obd *cli;
2905 LASSERT(imp->imp_obd == obd);
2908 case IMP_EVENT_DISCON: {
2909 /* Only do this on the MDS OSC's */
2910 if (imp->imp_server_timeout) {
2911 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2913 spin_lock(&oscc->oscc_lock);
2914 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
2915 spin_unlock(&oscc->oscc_lock);
2919 case IMP_EVENT_INACTIVE: {
2920 if (obd->obd_observer)
2921 rc = obd_notify(obd->obd_observer, obd, 0);
2924 case IMP_EVENT_INVALIDATE: {
2925 struct ldlm_namespace *ns = obd->obd_namespace;
2929 spin_lock(&cli->cl_loi_list_lock);
2930 cli->cl_avail_grant = 0;
2931 cli->cl_lost_grant = 0;
2932 /* all pages go to failing rpcs due to the invalid import */
2933 osc_check_rpcs(cli);
2934 spin_unlock(&cli->cl_loi_list_lock);
2936 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2940 case IMP_EVENT_ACTIVE: {
2941 if (obd->obd_observer)
2942 rc = obd_notify(obd->obd_observer, obd, 1);
2946 CERROR("Unknown import event %d\n", event);
2952 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
2954 struct lprocfs_static_vars lvars;
2958 lprocfs_init_vars(osc,&lvars);
2959 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
2963 rc = lproc_osc_attach_seqstat(dev);
2965 lprocfs_obd_detach(dev);
2969 ptlrpc_lprocfs_register_obd(dev);
2973 static int osc_detach(struct obd_device *dev)
2975 ptlrpc_lprocfs_unregister_obd(dev);
2976 return lprocfs_obd_detach(dev);
2979 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2983 rc = ptlrpcd_addref();
2987 rc = client_obd_setup(obd, len, buf);
2996 static int osc_cleanup(struct obd_device *obd, int flags)
3000 rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3001 LDLM_FL_CONFIG_CHANGE, NULL);
3005 rc = client_obd_cleanup(obd, flags);
3010 struct obd_ops osc_obd_ops = {
3011 .o_owner = THIS_MODULE,
3012 .o_attach = osc_attach,
3013 .o_detach = osc_detach,
3014 .o_setup = osc_setup,
3015 .o_cleanup = osc_cleanup,
3016 .o_connect = osc_connect,
3017 .o_disconnect = osc_disconnect,
3018 .o_statfs = osc_statfs,
3019 .o_packmd = osc_packmd,
3020 .o_unpackmd = osc_unpackmd,
3021 .o_create = osc_create,
3022 .o_destroy = osc_destroy,
3023 .o_getattr = osc_getattr,
3024 .o_getattr_async = osc_getattr_async,
3025 .o_setattr = osc_setattr,
3027 .o_brw_async = osc_brw_async,
3028 .o_prep_async_page = osc_prep_async_page,
3029 .o_queue_async_io = osc_queue_async_io,
3030 .o_set_async_flags = osc_set_async_flags,
3031 .o_queue_group_io = osc_queue_group_io,
3032 .o_trigger_group_io = osc_trigger_group_io,
3033 .o_teardown_async_page = osc_teardown_async_page,
3034 .o_punch = osc_punch,
3036 .o_enqueue = osc_enqueue,
3037 .o_match = osc_match,
3038 .o_change_cbdata = osc_change_cbdata,
3039 .o_cancel = osc_cancel,
3040 .o_cancel_unused = osc_cancel_unused,
3041 .o_iocontrol = osc_iocontrol,
3042 .o_get_info = osc_get_info,
3043 .o_set_info = osc_set_info,
3044 .o_import_event = osc_import_event,
3045 .o_llog_init = osc_llog_init,
3046 .o_llog_finish = osc_llog_finish,
3049 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3050 struct obd_ops sanosc_obd_ops = {
3051 .o_owner = THIS_MODULE,
3052 .o_attach = osc_attach,
3053 .o_detach = osc_detach,
3054 .o_cleanup = client_obd_cleanup,
3055 .o_connect = osc_connect,
3056 .o_disconnect = client_disconnect_export,
3057 .o_statfs = osc_statfs,
3058 .o_packmd = osc_packmd,
3059 .o_unpackmd = osc_unpackmd,
3060 .o_create = osc_real_create,
3061 .o_destroy = osc_destroy,
3062 .o_getattr = osc_getattr,
3063 .o_getattr_async = osc_getattr_async,
3064 .o_setattr = osc_setattr,
3065 .o_setup = client_sanobd_setup,
3066 .o_brw = sanosc_brw,
3067 .o_punch = osc_punch,
3069 .o_enqueue = osc_enqueue,
3070 .o_match = osc_match,
3071 .o_change_cbdata = osc_change_cbdata,
3072 .o_cancel = osc_cancel,
3073 .o_cancel_unused = osc_cancel_unused,
3074 .o_iocontrol = osc_iocontrol,
3075 .o_import_event = osc_import_event,
3076 .o_llog_init = osc_llog_init,
3077 .o_llog_finish = osc_llog_finish,
3081 int __init osc_init(void)
3083 struct lprocfs_static_vars lvars;
3084 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3085 struct lprocfs_static_vars sanlvars;
3090 lprocfs_init_vars(osc, &lvars);
3091 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3092 lprocfs_init_vars(osc, &sanlvars);
3095 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3100 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3101 rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3102 LUSTRE_SANOSC_NAME);
3104 class_unregister_type(LUSTRE_OSC_NAME);
3111 static void /*__exit*/ osc_exit(void)
3113 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3114 class_unregister_type(LUSTRE_SANOSC_NAME);
3116 class_unregister_type(LUSTRE_OSC_NAME);
3119 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3120 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3121 MODULE_LICENSE("GPL");
3123 module_init(osc_init);
3124 module_exit(osc_exit);