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;
722 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
724 for (niocount = i = 1; i < page_count; i++)
725 if (!can_merge_pages(&pga[i - 1], &pga[i]))
728 size[0] = sizeof(*body);
729 size[1] = sizeof(*ioobj);
730 size[2] = niocount * sizeof(*niobuf);
732 req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
736 if (opc == OST_WRITE)
737 desc = ptlrpc_prep_bulk_imp (req, page_count,
738 BULK_GET_SOURCE, OST_BULK_PORTAL);
740 desc = ptlrpc_prep_bulk_imp (req, page_count,
741 BULK_PUT_SINK, OST_BULK_PORTAL);
743 GOTO(out, rc = -ENOMEM);
744 /* NB request now owns desc and will free it when it gets freed */
746 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
747 ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
748 niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
750 memcpy(&body->oa, oa, sizeof(*oa));
752 obdo_to_ioobj(oa, ioobj);
753 ioobj->ioo_bufcnt = niocount;
755 LASSERT (page_count > 0);
756 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
757 struct brw_page *pg = &pga[i];
758 struct brw_page *pg_prev = pg - 1;
760 LASSERT(pg->count > 0);
761 LASSERT((pg->off & ~PAGE_MASK)+ pg->count <= PAGE_SIZE);
762 LASSERTF(i == 0 || pg->off > pg_prev->off,
763 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
764 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
766 pg->pg, pg->pg->private, pg->pg->index, pg->off,
767 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
770 ptlrpc_prep_bulk_page(desc, pg->pg,
771 pg->off & ~PAGE_MASK, pg->count);
772 requested_nob += pg->count;
774 if (i > 0 && can_merge_pages(pg_prev, pg)) {
776 niobuf->len += pg->count;
778 niobuf->offset = pg->off;
779 niobuf->len = pg->count;
780 niobuf->flags = pg->flag;
784 LASSERT((void *)(niobuf - niocount) ==
785 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
786 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
788 /* size[0] still sizeof (*body) */
789 if (opc == OST_WRITE) {
791 body->oa.o_valid |= OBD_MD_FLCKSUM;
792 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
794 /* 1 RC per niobuf */
795 size[1] = sizeof(__u32) * niocount;
796 req->rq_replen = lustre_msg_size(2, size);
798 /* 1 RC for the whole I/O */
799 req->rq_replen = lustre_msg_size(1, size);
802 *niocountp = niocount;
803 *requested_nobp = requested_nob;
808 ptlrpc_req_finished (req);
812 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
813 int requested_nob, int niocount,
814 obd_count page_count, struct brw_page *pga,
817 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
818 struct ost_body *body;
824 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
826 CERROR ("Can't unpack body\n");
830 osc_update_grant(cli, body);
831 memcpy(oa, &body->oa, sizeof(*oa));
833 if (req->rq_reqmsg->opc == OST_WRITE) {
835 CERROR ("Unexpected +ve rc %d\n", rc);
838 LASSERT (req->rq_bulk->bd_nob == requested_nob);
840 RETURN(check_write_rcs(req, requested_nob, niocount,
844 if (rc > requested_nob) {
845 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
849 if (rc != req->rq_bulk->bd_nob_transferred) {
850 CERROR ("Unexpected rc %d (%d transferred)\n",
851 rc, req->rq_bulk->bd_nob_transferred);
855 if (rc < requested_nob)
856 handle_short_read(rc, page_count, pga);
859 if (oa->o_valid & OBD_MD_FLCKSUM) {
860 const struct ptlrpc_peer *peer =
861 &req->rq_import->imp_connection->c_peer;
862 static int cksum_counter;
863 obd_count server_cksum = oa->o_cksum;
864 obd_count cksum = cksum_pages(rc, page_count, pga);
865 char str[PTL_NALFMT_SIZE];
867 ptlrpc_peernid2str(peer, str);
870 if (server_cksum != cksum) {
871 CERROR("Bad checksum: server %x, client %x, server NID "
872 LPX64" (%s)\n", server_cksum, cksum,
873 peer->peer_nid, str);
876 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
877 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
878 cksum_counter, peer->peer_nid, str, cksum);
881 static int cksum_missed;
884 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
885 CERROR("Request checksum %u from "LPX64", no reply\n",
887 req->rq_import->imp_connection->c_peer.peer_nid);
893 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
894 struct lov_stripe_md *lsm,
895 obd_count page_count, struct brw_page *pga)
899 struct ptlrpc_request *request;
904 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
905 page_count, pga, &requested_nob, &niocount,
910 rc = ptlrpc_queue_wait(request);
912 if (rc == -ETIMEDOUT && request->rq_resend) {
913 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
914 ptlrpc_req_finished(request);
918 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
919 page_count, pga, rc);
921 ptlrpc_req_finished(request);
925 static int brw_interpret(struct ptlrpc_request *request,
926 struct osc_brw_async_args *aa, int rc)
928 struct obdo *oa = aa->aa_oa;
929 int requested_nob = aa->aa_requested_nob;
930 int niocount = aa->aa_nio_count;
931 obd_count page_count = aa->aa_page_count;
932 struct brw_page *pga = aa->aa_pga;
935 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
936 page_count, pga, rc);
940 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
941 struct lov_stripe_md *lsm, obd_count page_count,
942 struct brw_page *pga, struct ptlrpc_request_set *set)
944 struct ptlrpc_request *request;
947 struct osc_brw_async_args *aa;
951 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
952 page_count, pga, &requested_nob, &nio_count,
955 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
956 aa = (struct osc_brw_async_args *)&request->rq_async_args;
958 aa->aa_requested_nob = requested_nob;
959 aa->aa_nio_count = nio_count;
960 aa->aa_page_count = page_count;
963 request->rq_interpret_reply = brw_interpret;
964 ptlrpc_set_add_req(set, request);
970 #define min_t(type,x,y) \
971 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
975 * ugh, we want disk allocation on the target to happen in offset order. we'll
976 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
977 * fine for our small page arrays and doesn't require allocation. its an
978 * insertion sort that swaps elements that are strides apart, shrinking the
979 * stride down until its '1' and the array is sorted.
981 static void sort_brw_pages(struct brw_page *array, int num)
988 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
993 for (i = stride ; i < num ; i++) {
996 while (j >= stride && array[j - stride].off >
998 array[j] = array[j - stride];
1003 } while (stride > 1);
1006 /* make sure we the regions we're passing to elan don't violate its '4
1007 * fragments' constraint. portal headers are a fragment, all full
1008 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1009 * counts as a fragment. I think. see bug 934. */
1010 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1013 int saw_whole_frag = 0;
1016 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1017 if (pg->count == PAGE_SIZE) {
1018 if (!saw_whole_frag) {
1029 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1030 struct lov_stripe_md *md, obd_count page_count,
1031 struct brw_page *pga, struct obd_trans_info *oti)
1035 if (cmd == OBD_BRW_CHECK) {
1036 /* The caller just wants to know if there's a chance that this
1037 * I/O can succeed */
1038 struct obd_import *imp = class_exp2cliimp(exp);
1040 if (imp == NULL || imp->imp_invalid)
1045 while (page_count) {
1046 obd_count pages_per_brw;
1049 if (page_count > PTLRPC_MAX_BRW_PAGES)
1050 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1052 pages_per_brw = page_count;
1054 sort_brw_pages(pga, pages_per_brw);
1055 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1057 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1062 page_count -= pages_per_brw;
1063 pga += pages_per_brw;
1068 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1069 struct lov_stripe_md *md, obd_count page_count,
1070 struct brw_page *pga, struct ptlrpc_request_set *set,
1071 struct obd_trans_info *oti)
1075 if (cmd == OBD_BRW_CHECK) {
1076 /* The caller just wants to know if there's a chance that this
1077 * I/O can succeed */
1078 struct obd_import *imp = class_exp2cliimp(exp);
1080 if (imp == NULL || imp->imp_invalid)
1085 while (page_count) {
1086 obd_count pages_per_brw;
1089 if (page_count > PTLRPC_MAX_BRW_PAGES)
1090 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1092 pages_per_brw = page_count;
1094 sort_brw_pages(pga, pages_per_brw);
1095 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1097 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1102 page_count -= pages_per_brw;
1103 pga += pages_per_brw;
1108 static void osc_check_rpcs(struct client_obd *cli);
1109 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1111 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1112 static void lop_update_pending(struct client_obd *cli,
1113 struct loi_oap_pages *lop, int cmd, int delta);
1115 /* this is called when a sync waiter receives an interruption. Its job is to
1116 * get the caller woken as soon as possible. If its page hasn't been put in an
1117 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1118 * desiring interruption which will forcefully complete the rpc once the rpc
1120 static void osc_occ_interrupted(struct oig_callback_context *occ)
1122 struct osc_async_page *oap;
1123 struct loi_oap_pages *lop;
1124 struct lov_oinfo *loi;
1127 /* XXX member_of() */
1128 oap = list_entry(occ, struct osc_async_page, oap_occ);
1130 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1132 oap->oap_interrupted = 1;
1134 /* ok, it's been put in an rpc. */
1135 if (oap->oap_request != NULL) {
1136 ptlrpc_mark_interrupted(oap->oap_request);
1137 ptlrpcd_wake(oap->oap_request);
1141 /* we don't get interruption callbacks until osc_trigger_sync_io()
1142 * has been called and put the sync oaps in the pending/urgent lists.*/
1143 if (!list_empty(&oap->oap_pending_item)) {
1144 list_del_init(&oap->oap_pending_item);
1145 if (oap->oap_async_flags & ASYNC_URGENT)
1146 list_del_init(&oap->oap_urgent_item);
1149 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1150 &loi->loi_write_lop : &loi->loi_read_lop;
1151 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1152 loi_list_maint(oap->oap_cli, oap->oap_loi);
1154 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1155 oap->oap_oig = NULL;
1159 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1162 /* this must be called holding the loi list lock to give coverage to exit_cache,
1163 * async_flag maintenance, and oap_request */
1164 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1165 struct osc_async_page *oap, int sent, int rc)
1167 osc_exit_cache(cli, oap, sent);
1168 oap->oap_async_flags = 0;
1169 oap->oap_interrupted = 0;
1171 if (oap->oap_request != NULL) {
1172 ptlrpc_req_finished(oap->oap_request);
1173 oap->oap_request = NULL;
1176 if (rc == 0 && oa != NULL)
1177 oap->oap_loi->loi_blocks = oa->o_blocks;
1180 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1181 oap->oap_oig = NULL;
1186 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1190 static int brw_interpret_oap(struct ptlrpc_request *request,
1191 struct osc_brw_async_args *aa, int rc)
1193 struct osc_async_page *oap;
1194 struct client_obd *cli;
1195 struct list_head *pos, *n;
1199 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1200 aa->aa_nio_count, aa->aa_page_count,
1203 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1206 /* in failout recovery we ignore writeback failure and want
1207 * to just tell llite to unlock the page and continue */
1208 if (request->rq_reqmsg->opc == OST_WRITE &&
1209 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1210 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1212 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1216 spin_lock(&cli->cl_loi_list_lock);
1218 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1219 * is called so we know whether to go to sync BRWs or wait for more
1220 * RPCs to complete */
1221 cli->cl_brw_in_flight--;
1223 /* the caller may re-use the oap after the completion call so
1224 * we need to clean it up a little */
1225 list_for_each_safe(pos, n, &aa->aa_oaps) {
1226 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1228 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1229 //oap->oap_page, oap->oap_page->index, oap);
1231 list_del_init(&oap->oap_rpc_item);
1232 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1235 osc_wake_cache_waiters(cli);
1236 osc_check_rpcs(cli);
1238 spin_unlock(&cli->cl_loi_list_lock);
1240 obdo_free(aa->aa_oa);
1241 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1246 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1247 struct list_head *rpc_list,
1248 int page_count, int cmd)
1250 struct ptlrpc_request *req;
1251 struct brw_page *pga = NULL;
1252 int requested_nob, nio_count;
1253 struct osc_brw_async_args *aa;
1254 struct obdo *oa = NULL;
1255 struct obd_async_page_ops *ops = NULL;
1256 void *caller_data = NULL;
1257 struct list_head *pos;
1260 LASSERT(!list_empty(rpc_list));
1262 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1264 RETURN(ERR_PTR(-ENOMEM));
1268 GOTO(out, req = ERR_PTR(-ENOMEM));
1271 list_for_each(pos, rpc_list) {
1272 struct osc_async_page *oap;
1274 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1276 ops = oap->oap_caller_ops;
1277 caller_data = oap->oap_caller_data;
1279 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1280 pga[i].pg = oap->oap_page;
1281 pga[i].count = oap->oap_count;
1282 pga[i].flag = oap->oap_brw_flags;
1283 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1284 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1288 /* always get the data for the obdo for the rpc */
1289 LASSERT(ops != NULL);
1290 ops->ap_fill_obdo(caller_data, cmd, oa);
1292 sort_brw_pages(pga, page_count);
1293 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1294 pga, &requested_nob, &nio_count, &req);
1296 CERROR("prep_req failed: %d\n", rc);
1297 GOTO(out, req = ERR_PTR(rc));
1300 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1301 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1303 aa->aa_requested_nob = requested_nob;
1304 aa->aa_nio_count = nio_count;
1305 aa->aa_page_count = page_count;
1314 OBD_FREE(pga, sizeof(*pga) * page_count);
1319 static void lop_update_pending(struct client_obd *cli,
1320 struct loi_oap_pages *lop, int cmd, int delta)
1322 lop->lop_num_pending += delta;
1323 if (cmd == OBD_BRW_WRITE)
1324 cli->cl_pending_w_pages += delta;
1326 cli->cl_pending_r_pages += delta;
1329 /* the loi lock is held across this function but it's allowed to release
1330 * and reacquire it during its work */
1331 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1332 int cmd, struct loi_oap_pages *lop)
1334 struct ptlrpc_request *request;
1335 obd_count page_count = 0;
1336 struct list_head *tmp, *pos;
1337 struct osc_async_page *oap = NULL;
1338 struct osc_brw_async_args *aa;
1339 struct obd_async_page_ops *ops;
1340 LIST_HEAD(rpc_list);
1343 /* first we find the pages we're allowed to work with */
1344 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1345 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1346 ops = oap->oap_caller_ops;
1348 LASSERT(oap->oap_magic == OAP_MAGIC);
1350 /* in llite being 'ready' equates to the page being locked
1351 * until completion unlocks it. commit_write submits a page
1352 * as not ready because its unlock will happen unconditionally
1353 * as the call returns. if we race with commit_write giving
1354 * us that page we dont' want to create a hole in the page
1355 * stream, so we stop and leave the rpc to be fired by
1356 * another dirtier or kupdated interval (the not ready page
1357 * will still be on the dirty list). we could call in
1358 * at the end of ll_file_write to process the queue again. */
1359 if (!(oap->oap_async_flags & ASYNC_READY)) {
1360 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1362 CDEBUG(D_INODE, "oap %p page %p returned %d "
1363 "instead of ready\n", oap,
1367 /* llite is telling us that the page is still
1368 * in commit_write and that we should try
1369 * and put it in an rpc again later. we
1370 * break out of the loop so we don't create
1371 * a hole in the sequence of pages in the rpc
1376 /* the io isn't needed.. tell the checks
1377 * below to complete the rpc with EINTR */
1378 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1379 oap->oap_count = -EINTR;
1382 oap->oap_async_flags |= ASYNC_READY;
1385 LASSERTF(0, "oap %p page %p returned %d "
1386 "from make_ready\n", oap,
1394 /* take the page out of our book-keeping */
1395 list_del_init(&oap->oap_pending_item);
1396 lop_update_pending(cli, lop, cmd, -1);
1397 list_del_init(&oap->oap_urgent_item);
1399 /* ask the caller for the size of the io as the rpc leaves. */
1400 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1402 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1403 if (oap->oap_count <= 0) {
1404 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1406 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1410 /* now put the page back in our accounting */
1411 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1412 if (++page_count >= cli->cl_max_pages_per_rpc)
1416 osc_wake_cache_waiters(cli);
1418 if (page_count == 0)
1421 loi_list_maint(cli, loi);
1422 spin_unlock(&cli->cl_loi_list_lock);
1424 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1425 if (IS_ERR(request)) {
1426 /* this should happen rarely and is pretty bad, it makes the
1427 * pending list not follow the dirty order */
1428 spin_lock(&cli->cl_loi_list_lock);
1429 list_for_each_safe(pos, tmp, &rpc_list) {
1430 oap = list_entry(pos, struct osc_async_page,
1432 list_del_init(&oap->oap_rpc_item);
1434 /* queued sync pages can be torn down while the pages
1435 * were between the pending list and the rpc */
1436 if (oap->oap_interrupted) {
1437 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1438 osc_ap_completion(cli, NULL, oap, 0,
1443 /* put the page back in the loi/lop lists */
1444 list_add_tail(&oap->oap_pending_item,
1446 lop_update_pending(cli, lop, cmd, 1);
1447 if (oap->oap_async_flags & ASYNC_URGENT)
1448 list_add(&oap->oap_urgent_item,
1451 loi_list_maint(cli, loi);
1452 RETURN(PTR_ERR(request));
1455 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1456 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1457 INIT_LIST_HEAD(&aa->aa_oaps);
1458 list_splice(&rpc_list, &aa->aa_oaps);
1459 INIT_LIST_HEAD(&rpc_list);
1462 if (cmd == OBD_BRW_READ) {
1463 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1464 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1466 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1467 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1468 cli->cl_brw_in_flight);
1472 spin_lock(&cli->cl_loi_list_lock);
1474 cli->cl_brw_in_flight++;
1475 /* queued sync pages can be torn down while the pages
1476 * were between the pending list and the rpc */
1477 list_for_each(pos, &aa->aa_oaps) {
1478 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1479 if (oap->oap_interrupted) {
1480 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1482 ptlrpc_mark_interrupted(request);
1487 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %d in flight\n", request,
1488 page_count, aa, cli->cl_brw_in_flight);
1490 oap->oap_request = ptlrpc_request_addref(request);
1491 request->rq_interpret_reply = brw_interpret_oap;
1492 ptlrpcd_add_req(request);
1496 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1502 if (lop->lop_num_pending == 0)
1505 /* if we have an invalid import we want to drain the queued pages
1506 * by forcing them through rpcs that immediately fail and complete
1507 * the pages. recovery relies on this to empty the queued pages
1508 * before canceling the locks and evicting down the llite pages */
1509 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1512 /* stream rpcs in queue order as long as as there is an urgent page
1513 * queued. this is our cheap solution for good batching in the case
1514 * where writepage marks some random page in the middle of the file as
1515 * urgent because of, say, memory pressure */
1516 if (!list_empty(&lop->lop_urgent))
1519 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1520 optimal = cli->cl_max_pages_per_rpc;
1521 if (cmd == OBD_BRW_WRITE) {
1522 /* trigger a write rpc stream as long as there are dirtiers
1523 * waiting for space. as they're waiting, they're not going to
1524 * create more pages to coallesce with what's waiting.. */
1525 if (!list_empty(&cli->cl_cache_waiters))
1528 /* *2 to avoid triggering rpcs that would want to include pages
1529 * that are being queued but which can't be made ready until
1530 * the queuer finishes with the page. this is a wart for
1531 * llite::commit_write() */
1534 if (lop->lop_num_pending >= optimal)
1540 static void on_list(struct list_head *item, struct list_head *list,
1543 if (list_empty(item) && should_be_on)
1544 list_add_tail(item, list);
1545 else if (!list_empty(item) && !should_be_on)
1546 list_del_init(item);
1549 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1550 * can find pages to build into rpcs quickly */
1551 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1553 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1554 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1555 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1557 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1558 loi->loi_write_lop.lop_num_pending);
1560 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1561 loi->loi_read_lop.lop_num_pending);
1564 #define LOI_DEBUG(LOI, STR, args...) \
1565 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1566 !list_empty(&(LOI)->loi_cli_item), \
1567 (LOI)->loi_write_lop.lop_num_pending, \
1568 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1569 (LOI)->loi_read_lop.lop_num_pending, \
1570 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1573 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1576 /* first return all objects which we already know to have
1577 * pages ready to be stuffed into rpcs */
1578 if (!list_empty(&cli->cl_loi_ready_list))
1579 RETURN(list_entry(cli->cl_loi_ready_list.next,
1580 struct lov_oinfo, loi_cli_item));
1582 /* then if we have cache waiters, return all objects with queued
1583 * writes. This is especially important when many small files
1584 * have filled up the cache and not been fired into rpcs because
1585 * they don't pass the nr_pending/object threshhold */
1586 if (!list_empty(&cli->cl_cache_waiters) &&
1587 !list_empty(&cli->cl_loi_write_list))
1588 RETURN(list_entry(cli->cl_loi_write_list.next,
1589 struct lov_oinfo, loi_write_item));
1591 /* then return all queued objects when we have an invalid import
1592 * so that they get flushed */
1593 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1594 if (!list_empty(&cli->cl_loi_write_list))
1595 RETURN(list_entry(cli->cl_loi_write_list.next,
1596 struct lov_oinfo, loi_write_item));
1597 if (!list_empty(&cli->cl_loi_read_list))
1598 RETURN(list_entry(cli->cl_loi_read_list.next,
1599 struct lov_oinfo, loi_read_item));
1604 /* called with the loi list lock held */
1605 static void osc_check_rpcs(struct client_obd *cli)
1607 struct lov_oinfo *loi;
1608 int rc = 0, race_counter = 0;
1611 while ((loi = osc_next_loi(cli)) != NULL) {
1612 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1614 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1617 /* attempt some read/write balancing by alternating between
1618 * reads and writes in an object. The makes_rpc checks here
1619 * would be redundant if we were getting read/write work items
1620 * instead of objects. we don't want send_oap_rpc to drain a
1621 * partial read pending queue when we're given this object to
1622 * do io on writes while there are cache waiters */
1623 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1624 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1625 &loi->loi_write_lop);
1633 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1634 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1635 &loi->loi_read_lop);
1644 /* attempt some inter-object balancing by issueing rpcs
1645 * for each object in turn */
1646 if (!list_empty(&loi->loi_cli_item))
1647 list_del_init(&loi->loi_cli_item);
1648 if (!list_empty(&loi->loi_write_item))
1649 list_del_init(&loi->loi_write_item);
1650 if (!list_empty(&loi->loi_read_item))
1651 list_del_init(&loi->loi_read_item);
1653 loi_list_maint(cli, loi);
1655 /* send_oap_rpc fails with 0 when make_ready tells it to
1656 * back off. llite's make_ready does this when it tries
1657 * to lock a page queued for write that is already locked.
1658 * we want to try sending rpcs from many objects, but we
1659 * don't want to spin failing with 0. */
1660 if (race_counter == 10)
1666 /* we're trying to queue a page in the osc so we're subject to the
1667 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1668 * If the osc's queued pages are already at that limit, then we want to sleep
1669 * until there is space in the osc's queue for us. We also may be waiting for
1670 * write credits from the OST if there are RPCs in flight that may return some
1671 * before we fall back to sync writes.
1673 * We need this know our allocation was granted in the presence of signals */
1674 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1678 spin_lock(&cli->cl_loi_list_lock);
1679 rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1680 spin_unlock(&cli->cl_loi_list_lock);
1684 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1685 * grant or cache space. */
1686 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1687 struct osc_async_page *oap)
1689 struct osc_cache_waiter ocw;
1690 struct l_wait_info lwi = { 0 };
1692 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1693 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1694 cli->cl_avail_grant);
1696 if (cli->cl_dirty_max < PAGE_SIZE)
1699 /* Hopefully normal case - cache space and write credits available */
1700 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1701 cli->cl_avail_grant >= PAGE_SIZE) {
1702 /* account for ourselves */
1703 osc_consume_write_grant(cli, oap);
1707 /* Make sure that there are write rpcs in flight to wait for. This
1708 * is a little silly as this object may not have any pending but
1709 * other objects sure might. */
1710 if (cli->cl_brw_in_flight) {
1711 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1712 init_waitqueue_head(&ocw.ocw_waitq);
1716 loi_list_maint(cli, loi);
1717 osc_check_rpcs(cli);
1718 spin_unlock(&cli->cl_loi_list_lock);
1720 CDEBUG(0, "sleeping for cache space\n");
1721 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1723 spin_lock(&cli->cl_loi_list_lock);
1724 if (!list_empty(&ocw.ocw_entry)) {
1725 list_del(&ocw.ocw_entry);
1734 /* the companion to enter_cache, called when an oap is no longer part of the
1735 * dirty accounting.. so writeback completes or truncate happens before writing
1736 * starts. must be called with the loi lock held. */
1737 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1742 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1747 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1748 cli->cl_dirty -= PAGE_SIZE;
1750 cli->cl_lost_grant += PAGE_SIZE;
1751 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1752 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1758 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1759 struct lov_oinfo *loi, struct page *page,
1760 obd_off offset, struct obd_async_page_ops *ops,
1761 void *data, void **res)
1763 struct osc_async_page *oap;
1766 OBD_ALLOC(oap, sizeof(*oap));
1770 oap->oap_magic = OAP_MAGIC;
1771 oap->oap_cli = &exp->exp_obd->u.cli;
1774 oap->oap_caller_ops = ops;
1775 oap->oap_caller_data = data;
1777 oap->oap_page = page;
1778 oap->oap_obj_off = offset;
1780 INIT_LIST_HEAD(&oap->oap_pending_item);
1781 INIT_LIST_HEAD(&oap->oap_urgent_item);
1782 INIT_LIST_HEAD(&oap->oap_rpc_item);
1784 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1786 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1791 struct osc_async_page *oap_from_cookie(void *cookie)
1793 struct osc_async_page *oap = cookie;
1794 if (oap->oap_magic != OAP_MAGIC)
1795 return ERR_PTR(-EINVAL);
1799 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1800 struct lov_oinfo *loi, void *cookie,
1801 int cmd, obd_off off, int count,
1802 obd_flag brw_flags, enum async_flags async_flags)
1804 struct client_obd *cli = &exp->exp_obd->u.cli;
1805 struct osc_async_page *oap;
1806 struct loi_oap_pages *lop;
1810 oap = oap_from_cookie(cookie);
1812 RETURN(PTR_ERR(oap));
1814 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1817 if (!list_empty(&oap->oap_pending_item) ||
1818 !list_empty(&oap->oap_urgent_item) ||
1819 !list_empty(&oap->oap_rpc_item))
1823 loi = &lsm->lsm_oinfo[0];
1825 spin_lock(&cli->cl_loi_list_lock);
1828 oap->oap_async_flags = async_flags;
1829 oap->oap_page_off = off;
1830 oap->oap_count = count;
1831 oap->oap_brw_flags = brw_flags;
1833 if (cmd == OBD_BRW_WRITE) {
1834 rc = osc_enter_cache(cli, loi, oap);
1836 spin_unlock(&cli->cl_loi_list_lock);
1839 lop = &loi->loi_write_lop;
1841 lop = &loi->loi_read_lop;
1844 if (oap->oap_async_flags & ASYNC_URGENT)
1845 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1846 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1847 lop_update_pending(cli, lop, cmd, 1);
1849 loi_list_maint(cli, loi);
1851 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1854 osc_check_rpcs(cli);
1855 spin_unlock(&cli->cl_loi_list_lock);
1860 /* aka (~was & now & flag), but this is more clear :) */
1861 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1863 static int osc_set_async_flags(struct obd_export *exp,
1864 struct lov_stripe_md *lsm,
1865 struct lov_oinfo *loi, void *cookie,
1866 obd_flag async_flags)
1868 struct client_obd *cli = &exp->exp_obd->u.cli;
1869 struct loi_oap_pages *lop;
1870 struct osc_async_page *oap;
1874 oap = oap_from_cookie(cookie);
1876 RETURN(PTR_ERR(oap));
1878 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1882 loi = &lsm->lsm_oinfo[0];
1884 if (oap->oap_cmd == OBD_BRW_WRITE) {
1885 lop = &loi->loi_write_lop;
1887 lop = &loi->loi_read_lop;
1890 spin_lock(&cli->cl_loi_list_lock);
1892 if (list_empty(&oap->oap_pending_item))
1893 GOTO(out, rc = -EINVAL);
1895 if ((oap->oap_async_flags & async_flags) == async_flags)
1898 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1899 oap->oap_async_flags |= ASYNC_READY;
1901 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1902 if (list_empty(&oap->oap_rpc_item)) {
1903 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1904 loi_list_maint(cli, loi);
1908 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1909 oap->oap_async_flags);
1911 osc_check_rpcs(cli);
1912 spin_unlock(&cli->cl_loi_list_lock);
1916 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1917 struct lov_oinfo *loi,
1918 struct obd_io_group *oig, void *cookie,
1919 int cmd, obd_off off, int count,
1921 obd_flag async_flags)
1923 struct client_obd *cli = &exp->exp_obd->u.cli;
1924 struct osc_async_page *oap;
1925 struct loi_oap_pages *lop;
1928 oap = oap_from_cookie(cookie);
1930 RETURN(PTR_ERR(oap));
1932 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1935 if (!list_empty(&oap->oap_pending_item) ||
1936 !list_empty(&oap->oap_urgent_item) ||
1937 !list_empty(&oap->oap_rpc_item))
1941 loi = &lsm->lsm_oinfo[0];
1943 spin_lock(&cli->cl_loi_list_lock);
1946 oap->oap_page_off = off;
1947 oap->oap_count = count;
1948 oap->oap_brw_flags = brw_flags;
1949 oap->oap_async_flags = async_flags;
1951 if (cmd == OBD_BRW_WRITE)
1952 lop = &loi->loi_write_lop;
1954 lop = &loi->loi_read_lop;
1956 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
1957 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
1959 oig_add_one(oig, &oap->oap_occ);
1962 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
1964 spin_unlock(&cli->cl_loi_list_lock);
1969 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1970 struct loi_oap_pages *lop, int cmd)
1972 struct list_head *pos, *tmp;
1973 struct osc_async_page *oap;
1975 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
1976 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1977 list_del(&oap->oap_pending_item);
1978 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1979 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1980 lop_update_pending(cli, lop, cmd, 1);
1982 loi_list_maint(cli, loi);
1985 static int osc_trigger_group_io(struct obd_export *exp,
1986 struct lov_stripe_md *lsm,
1987 struct lov_oinfo *loi,
1988 struct obd_io_group *oig)
1990 struct client_obd *cli = &exp->exp_obd->u.cli;
1994 loi = &lsm->lsm_oinfo[0];
1996 spin_lock(&cli->cl_loi_list_lock);
1998 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
1999 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2001 osc_check_rpcs(cli);
2002 spin_unlock(&cli->cl_loi_list_lock);
2007 static int osc_teardown_async_page(struct obd_export *exp,
2008 struct lov_stripe_md *lsm,
2009 struct lov_oinfo *loi, void *cookie)
2011 struct client_obd *cli = &exp->exp_obd->u.cli;
2012 struct loi_oap_pages *lop;
2013 struct osc_async_page *oap;
2017 oap = oap_from_cookie(cookie);
2019 RETURN(PTR_ERR(oap));
2022 loi = &lsm->lsm_oinfo[0];
2024 if (oap->oap_cmd == OBD_BRW_WRITE) {
2025 lop = &loi->loi_write_lop;
2027 lop = &loi->loi_read_lop;
2030 spin_lock(&cli->cl_loi_list_lock);
2032 if (!list_empty(&oap->oap_rpc_item))
2033 GOTO(out, rc = -EBUSY);
2035 osc_exit_cache(cli, oap, 0);
2036 osc_wake_cache_waiters(cli);
2038 if (!list_empty(&oap->oap_urgent_item)) {
2039 list_del_init(&oap->oap_urgent_item);
2040 oap->oap_async_flags &= ~ASYNC_URGENT;
2042 if (!list_empty(&oap->oap_pending_item)) {
2043 list_del_init(&oap->oap_pending_item);
2044 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2046 loi_list_maint(cli, loi);
2048 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2050 spin_unlock(&cli->cl_loi_list_lock);
2052 OBD_FREE(oap, sizeof(*oap));
2057 /* Note: caller will lock/unlock, and set uptodate on the pages */
2058 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2059 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2060 struct lov_stripe_md *lsm, obd_count page_count,
2061 struct brw_page *pga)
2063 struct ptlrpc_request *request = NULL;
2064 struct ost_body *body;
2065 struct niobuf_remote *nioptr;
2066 struct obd_ioobj *iooptr;
2067 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2071 /* XXX does not handle 'new' brw protocol */
2073 size[1] = sizeof(struct obd_ioobj);
2074 size[2] = page_count * sizeof(*nioptr);
2076 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2081 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2082 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2083 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2084 sizeof(*nioptr) * page_count);
2086 memcpy(&body->oa, oa, sizeof(body->oa));
2088 obdo_to_ioobj(oa, iooptr);
2089 iooptr->ioo_bufcnt = page_count;
2091 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2092 LASSERT(PageLocked(pga[mapped].pg));
2093 LASSERT(mapped == 0 ||
2094 pga[mapped].off > pga[mapped - 1].off);
2096 nioptr->offset = pga[mapped].off;
2097 nioptr->len = pga[mapped].count;
2098 nioptr->flags = pga[mapped].flag;
2101 size[1] = page_count * sizeof(*nioptr);
2102 request->rq_replen = lustre_msg_size(2, size);
2104 rc = ptlrpc_queue_wait(request);
2108 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2109 lustre_swab_ost_body);
2111 CERROR("Can't unpack body\n");
2112 GOTO(out_req, rc = -EPROTO);
2115 memcpy(oa, &body->oa, sizeof(*oa));
2117 swab = lustre_msg_swabbed(request->rq_repmsg);
2118 LASSERT_REPSWAB(request, 1);
2119 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2121 /* nioptr missing or short */
2122 GOTO(out_req, rc = -EPROTO);
2126 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2127 struct page *page = pga[mapped].pg;
2128 struct buffer_head *bh;
2132 lustre_swab_niobuf_remote (nioptr);
2134 /* got san device associated */
2135 LASSERT(exp->exp_obd != NULL);
2136 dev = exp->exp_obd->u.cli.cl_sandev;
2139 if (!nioptr->offset) {
2140 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2141 page->mapping->host->i_ino,
2143 memset(page_address(page), 0, PAGE_SIZE);
2147 if (!page->buffers) {
2148 create_empty_buffers(page, dev, PAGE_SIZE);
2151 clear_bit(BH_New, &bh->b_state);
2152 set_bit(BH_Mapped, &bh->b_state);
2153 bh->b_blocknr = (unsigned long)nioptr->offset;
2155 clear_bit(BH_Uptodate, &bh->b_state);
2157 ll_rw_block(READ, 1, &bh);
2161 /* if buffer already existed, it must be the
2162 * one we mapped before, check it */
2163 LASSERT(!test_bit(BH_New, &bh->b_state));
2164 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2165 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2167 /* wait it's io completion */
2168 if (test_bit(BH_Lock, &bh->b_state))
2171 if (!test_bit(BH_Uptodate, &bh->b_state))
2172 ll_rw_block(READ, 1, &bh);
2176 /* must do syncronous write here */
2178 if (!buffer_uptodate(bh)) {
2186 ptlrpc_req_finished(request);
2190 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2191 struct lov_stripe_md *lsm, obd_count page_count,
2192 struct brw_page *pga)
2194 struct ptlrpc_request *request = NULL;
2195 struct ost_body *body;
2196 struct niobuf_remote *nioptr;
2197 struct obd_ioobj *iooptr;
2198 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2202 size[1] = sizeof(struct obd_ioobj);
2203 size[2] = page_count * sizeof(*nioptr);
2205 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2210 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2211 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2212 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2213 sizeof (*nioptr) * page_count);
2215 memcpy(&body->oa, oa, sizeof(body->oa));
2217 obdo_to_ioobj(oa, iooptr);
2218 iooptr->ioo_bufcnt = page_count;
2221 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2222 LASSERT(PageLocked(pga[mapped].pg));
2223 LASSERT(mapped == 0 ||
2224 pga[mapped].off > pga[mapped - 1].off);
2226 nioptr->offset = pga[mapped].off;
2227 nioptr->len = pga[mapped].count;
2228 nioptr->flags = pga[mapped].flag;
2231 size[1] = page_count * sizeof(*nioptr);
2232 request->rq_replen = lustre_msg_size(2, size);
2234 rc = ptlrpc_queue_wait(request);
2238 swab = lustre_msg_swabbed (request->rq_repmsg);
2239 LASSERT_REPSWAB (request, 1);
2240 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2242 CERROR("absent/short niobuf array\n");
2243 GOTO(out_req, rc = -EPROTO);
2247 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2248 struct page *page = pga[mapped].pg;
2249 struct buffer_head *bh;
2253 lustre_swab_niobuf_remote (nioptr);
2255 /* got san device associated */
2256 LASSERT(exp->exp_obd != NULL);
2257 dev = exp->exp_obd->u.cli.cl_sandev;
2259 if (!page->buffers) {
2260 create_empty_buffers(page, dev, PAGE_SIZE);
2263 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2264 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2265 LASSERT(page->buffers->b_blocknr ==
2266 (unsigned long)nioptr->offset);
2272 /* if buffer locked, wait it's io completion */
2273 if (test_bit(BH_Lock, &bh->b_state))
2276 clear_bit(BH_New, &bh->b_state);
2277 set_bit(BH_Mapped, &bh->b_state);
2279 /* override the block nr */
2280 bh->b_blocknr = (unsigned long)nioptr->offset;
2282 /* we are about to write it, so set it
2284 * page lock should garentee no race condition here */
2285 set_bit(BH_Uptodate, &bh->b_state);
2286 set_bit(BH_Dirty, &bh->b_state);
2288 ll_rw_block(WRITE, 1, &bh);
2290 /* must do syncronous write here */
2292 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2300 ptlrpc_req_finished(request);
2304 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2305 struct lov_stripe_md *lsm, obd_count page_count,
2306 struct brw_page *pga, struct obd_trans_info *oti)
2310 while (page_count) {
2311 obd_count pages_per_brw;
2314 if (page_count > PTLRPC_MAX_BRW_PAGES)
2315 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2317 pages_per_brw = page_count;
2319 if (cmd & OBD_BRW_WRITE)
2320 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2322 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2327 page_count -= pages_per_brw;
2328 pga += pages_per_brw;
2335 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2337 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2339 LASSERT(lock != NULL);
2340 l_lock(&lock->l_resource->lr_namespace->ns_lock);
2342 if (lock->l_ast_data && lock->l_ast_data != data) {
2343 struct inode *new_inode = data;
2344 struct inode *old_inode = lock->l_ast_data;
2345 LASSERTF(old_inode->i_state & I_FREEING,
2346 "Found existing inode %p/%lu/%u state %lu in lock: "
2347 "setting data to %p/%lu/%u\n", old_inode,
2348 old_inode->i_ino, old_inode->i_generation,
2350 new_inode, new_inode->i_ino, new_inode->i_generation);
2353 lock->l_ast_data = data;
2354 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2355 LDLM_LOCK_PUT(lock);
2358 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2359 ldlm_iterator_t replace, void *data)
2361 struct ldlm_res_id res_id = { .name = {0} };
2362 struct obd_device *obd = class_exp2obd(exp);
2364 res_id.name[0] = lsm->lsm_object_id;
2365 res_id.name[2] = lsm->lsm_object_gr;
2366 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2370 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2371 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2372 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2373 void *data, __u32 lvb_len, void *lvb_swabber,
2374 struct lustre_handle *lockh)
2376 struct obd_device *obd = exp->exp_obd;
2377 struct ldlm_res_id res_id = { .name = {0} };
2382 res_id.name[0] = lsm->lsm_object_id;
2383 res_id.name[2] = lsm->lsm_object_gr;
2385 /* Filesystem lock extents are extended to page boundaries so that
2386 * dealing with the page cache is a little smoother. */
2387 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2388 policy->l_extent.end |= ~PAGE_MASK;
2390 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2393 /* Next, search for already existing extent locks that will cover us */
2394 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2397 osc_set_data_with_check(lockh, data);
2398 if (*flags & LDLM_FL_HAS_INTENT) {
2399 /* I would like to be able to ASSERT here that rss <=
2400 * kms, but I can't, for reasons which are explained in
2403 /* We already have a lock, and it's referenced */
2407 /* If we're trying to read, we also search for an existing PW lock. The
2408 * VFS and page cache already protect us locally, so lots of readers/
2409 * writers can share a single PW lock.
2411 * There are problems with conversion deadlocks, so instead of
2412 * converting a read lock to a write lock, we'll just enqueue a new
2415 * At some point we should cancel the read lock instead of making them
2416 * send us a blocking callback, but there are problems with canceling
2417 * locks out from other users right now, too. */
2419 if (mode == LCK_PR) {
2420 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2421 policy, LCK_PW, lockh);
2423 /* FIXME: This is not incredibly elegant, but it might
2424 * be more elegant than adding another parameter to
2425 * lock_match. I want a second opinion. */
2426 ldlm_lock_addref(lockh, LCK_PR);
2427 ldlm_lock_decref(lockh, LCK_PW);
2428 osc_set_data_with_check(lockh, data);
2434 rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, res_id, type,
2435 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2436 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2438 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2439 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2440 lvb.lvb_size, lvb.lvb_blocks);
2441 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2442 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2448 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2449 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2450 int *flags, void *data, struct lustre_handle *lockh)
2452 struct ldlm_res_id res_id = { .name = {0} };
2453 struct obd_device *obd = exp->exp_obd;
2457 res_id.name[0] = lsm->lsm_object_id;
2458 res_id.name[2] = lsm->lsm_object_gr;
2460 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2462 /* Filesystem lock extents are extended to page boundaries so that
2463 * dealing with the page cache is a little smoother */
2464 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2465 policy->l_extent.end |= ~PAGE_MASK;
2467 /* Next, search for already existing extent locks that will cover us */
2468 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2469 policy, mode, lockh);
2471 if (!(*flags & LDLM_FL_TEST_LOCK))
2472 osc_set_data_with_check(lockh, data);
2475 /* If we're trying to read, we also search for an existing PW lock. The
2476 * VFS and page cache already protect us locally, so lots of readers/
2477 * writers can share a single PW lock. */
2478 if (mode == LCK_PR) {
2479 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2480 policy, LCK_PW, lockh);
2481 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2482 /* FIXME: This is not incredibly elegant, but it might
2483 * be more elegant than adding another parameter to
2484 * lock_match. I want a second opinion. */
2485 osc_set_data_with_check(lockh, data);
2486 ldlm_lock_addref(lockh, LCK_PR);
2487 ldlm_lock_decref(lockh, LCK_PW);
2493 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2494 __u32 mode, struct lustre_handle *lockh)
2498 if (mode == LCK_GROUP)
2499 ldlm_lock_decref_and_cancel(lockh, mode);
2501 ldlm_lock_decref(lockh, mode);
2506 static int osc_cancel_unused(struct obd_export *exp,
2507 struct lov_stripe_md *lsm, int flags, void *opaque)
2509 struct obd_device *obd = class_exp2obd(exp);
2510 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2513 res_id.name[0] = lsm->lsm_object_id;
2514 res_id.name[2] = lsm->lsm_object_gr;
2518 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2521 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2522 unsigned long max_age)
2524 struct obd_statfs *msfs;
2525 struct ptlrpc_request *request;
2526 int rc, size = sizeof(*osfs);
2529 /* We could possibly pass max_age in the request (as an absolute
2530 * timestamp or a "seconds.usec ago") so the target can avoid doing
2531 * extra calls into the filesystem if that isn't necessary (e.g.
2532 * during mount that would help a bit). Having relative timestamps
2533 * is not so great if request processing is slow, while absolute
2534 * timestamps are not ideal because they need time synchronization. */
2535 request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2539 request->rq_replen = lustre_msg_size(1, &size);
2540 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2542 rc = ptlrpc_queue_wait(request);
2546 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2547 lustre_swab_obd_statfs);
2549 CERROR("Can't unpack obd_statfs\n");
2550 GOTO(out, rc = -EPROTO);
2553 memcpy(osfs, msfs, sizeof(*osfs));
2557 ptlrpc_req_finished(request);
2561 /* Retrieve object striping information.
2563 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2564 * the maximum number of OST indices which will fit in the user buffer.
2565 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2567 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2569 struct lov_user_md lum, *lumk;
2576 rc = copy_from_user(&lum, lump, sizeof(lum));
2580 if (lum.lmm_magic != LOV_USER_MAGIC)
2583 if (lum.lmm_stripe_count > 0) {
2584 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2585 OBD_ALLOC(lumk, lum_size);
2589 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2590 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2592 lum_size = sizeof(lum);
2596 lumk->lmm_object_id = lsm->lsm_object_id;
2597 lumk->lmm_object_gr = lsm->lsm_object_gr;
2598 lumk->lmm_stripe_count = 1;
2600 if (copy_to_user(lump, lumk, lum_size))
2604 OBD_FREE(lumk, lum_size);
2609 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2610 void *karg, void *uarg)
2612 struct obd_device *obd = exp->exp_obd;
2613 struct obd_ioctl_data *data = karg;
2617 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2620 if (!try_module_get(THIS_MODULE)) {
2621 CERROR("Can't get module. Is it alive?");
2626 case OBD_IOC_LOV_GET_CONFIG: {
2628 struct lov_desc *desc;
2629 struct obd_uuid uuid;
2633 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2634 GOTO(out, err = -EINVAL);
2636 data = (struct obd_ioctl_data *)buf;
2638 if (sizeof(*desc) > data->ioc_inllen1) {
2640 GOTO(out, err = -EINVAL);
2643 if (data->ioc_inllen2 < sizeof(uuid)) {
2645 GOTO(out, err = -EINVAL);
2648 if (data->ioc_inllen3 < sizeof(__u32)) {
2650 GOTO(out, err = -EINVAL);
2653 desc = (struct lov_desc *)data->ioc_inlbuf1;
2654 desc->ld_tgt_count = 1;
2655 desc->ld_active_tgt_count = 1;
2656 desc->ld_default_stripe_count = 1;
2657 desc->ld_default_stripe_size = 0;
2658 desc->ld_default_stripe_offset = 0;
2659 desc->ld_pattern = 0;
2660 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2661 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2662 *((__u32 *)data->ioc_inlbuf3) = 1;
2664 err = copy_to_user((void *)uarg, buf, len);
2667 obd_ioctl_freedata(buf, len);
2670 case LL_IOC_LOV_SETSTRIPE:
2671 err = obd_alloc_memmd(exp, karg);
2675 case LL_IOC_LOV_GETSTRIPE:
2676 err = osc_getstripe(karg, uarg);
2678 case OBD_IOC_CLIENT_RECOVER:
2679 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2684 case IOC_OSC_SET_ACTIVE:
2685 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2689 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2690 GOTO(out, err = -ENOTTY);
2693 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2696 module_put(THIS_MODULE);
2701 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2702 void *key, __u32 *vallen, void *val)
2705 if (!vallen || !val)
2708 if (keylen > strlen("lock_to_stripe") &&
2709 strcmp(key, "lock_to_stripe") == 0) {
2710 __u32 *stripe = val;
2711 *vallen = sizeof(*stripe);
2714 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2715 struct ptlrpc_request *req;
2717 char *bufs[1] = {key};
2719 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2724 req->rq_replen = lustre_msg_size(1, vallen);
2725 rc = ptlrpc_queue_wait(req);
2729 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2730 lustre_swab_ost_last_id);
2731 if (reply == NULL) {
2732 CERROR("Can't unpack OST last ID\n");
2733 GOTO(out, rc = -EPROTO);
2735 *((obd_id *)val) = *reply;
2737 ptlrpc_req_finished(req);
2743 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2744 void *key, obd_count vallen, void *val)
2746 struct obd_device *obd = exp->exp_obd;
2747 struct obd_import *imp = class_exp2cliimp(exp);
2748 struct llog_ctxt *ctxt;
2752 if (keylen == strlen("next_id") &&
2753 memcmp(key, "next_id", strlen("next_id")) == 0) {
2754 if (vallen != sizeof(obd_id))
2756 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
2757 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
2758 exp->exp_obd->obd_name,
2759 obd->u.cli.cl_oscc.oscc_next_id);
2764 if (keylen == strlen("growth_count") &&
2765 memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2766 if (vallen != sizeof(int))
2768 obd->u.cli.cl_oscc.oscc_max_grow_count = *((int*)val);
2772 if (keylen == strlen("unlinked") &&
2773 memcmp(key, "unlinked", keylen) == 0) {
2774 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2775 spin_lock(&oscc->oscc_lock);
2776 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2777 spin_unlock(&oscc->oscc_lock);
2780 if (keylen == strlen("unrecovery") &&
2781 memcmp(key, "unrecovery", keylen) == 0) {
2782 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2783 spin_lock(&oscc->oscc_lock);
2784 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
2785 spin_unlock(&oscc->oscc_lock);
2788 if (keylen == strlen("initial_recov") &&
2789 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2790 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2791 if (vallen != sizeof(int))
2793 imp->imp_initial_recov = *(int *)val;
2794 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2795 exp->exp_obd->obd_name,
2796 imp->imp_initial_recov);
2800 if (keylen < strlen("mds_conn") ||
2801 memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2804 ctxt = llog_get_context(&exp->exp_obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2806 rc = llog_initiator_connect(ctxt);
2811 imp->imp_server_timeout = 1;
2812 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2813 imp->imp_pingable = 1;
2819 static struct llog_operations osc_size_repl_logops = {
2820 lop_cancel: llog_obd_repl_cancel
2823 static struct llog_operations osc_unlink_orig_logops;
2824 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2825 struct obd_device *tgt, int count,
2826 struct llog_catid *catid)
2831 osc_unlink_orig_logops = llog_lvfs_ops;
2832 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2833 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
2834 osc_unlink_orig_logops.lop_add = llog_catalog_add;
2835 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2837 rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
2838 &catid->lci_logid, &osc_unlink_orig_logops);
2842 rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2843 &osc_size_repl_logops);
2847 static int osc_llog_finish(struct obd_device *obd,
2848 struct obd_llogs *llogs, int count)
2853 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
2857 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
2862 static int osc_connect(struct lustre_handle *exph,
2863 struct obd_device *obd, struct obd_uuid *cluuid,
2864 unsigned long connect_flags)
2868 rc = client_connect_import(exph, obd, cluuid, connect_flags);
2873 static int osc_disconnect(struct obd_export *exp, int flags)
2875 struct obd_device *obd = class_exp2obd(exp);
2876 struct llog_ctxt *ctxt;
2879 ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
2880 if (obd->u.cli.cl_conn_count == 1)
2881 /* flush any remaining cancel messages out to the target */
2882 llog_sync(ctxt, exp);
2884 rc = client_disconnect_export(exp, flags);
2888 static int osc_import_event(struct obd_device *obd,
2889 struct obd_import *imp,
2890 enum obd_import_event event)
2892 struct client_obd *cli;
2895 LASSERT(imp->imp_obd == obd);
2898 case IMP_EVENT_DISCON: {
2899 /* Only do this on the MDS OSC's */
2900 if (imp->imp_server_timeout) {
2901 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2903 spin_lock(&oscc->oscc_lock);
2904 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
2905 spin_unlock(&oscc->oscc_lock);
2909 case IMP_EVENT_INACTIVE: {
2910 if (obd->obd_observer)
2911 rc = obd_notify(obd->obd_observer, obd, 0, 0);
2914 case IMP_EVENT_INVALIDATE: {
2915 struct ldlm_namespace *ns = obd->obd_namespace;
2919 spin_lock(&cli->cl_loi_list_lock);
2920 cli->cl_avail_grant = 0;
2921 cli->cl_lost_grant = 0;
2922 /* all pages go to failing rpcs due to the invalid import */
2923 osc_check_rpcs(cli);
2924 spin_unlock(&cli->cl_loi_list_lock);
2926 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2930 case IMP_EVENT_ACTIVE: {
2931 if (obd->obd_observer)
2932 rc = obd_notify(obd->obd_observer, obd, 1, 0);
2936 CERROR("Unknown import event %d\n", event);
2942 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
2944 struct lprocfs_static_vars lvars;
2948 lprocfs_init_vars(osc,&lvars);
2949 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
2953 rc = lproc_osc_attach_seqstat(dev);
2955 lprocfs_obd_detach(dev);
2959 ptlrpc_lprocfs_register_obd(dev);
2963 static int osc_detach(struct obd_device *dev)
2965 ptlrpc_lprocfs_unregister_obd(dev);
2966 return lprocfs_obd_detach(dev);
2969 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2973 rc = ptlrpcd_addref();
2977 rc = client_obd_setup(obd, len, buf);
2986 static int osc_cleanup(struct obd_device *obd, int flags)
2990 rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
2991 LDLM_FL_CONFIG_CHANGE, NULL);
2995 rc = client_obd_cleanup(obd, flags);
3000 struct obd_ops osc_obd_ops = {
3001 .o_owner = THIS_MODULE,
3002 .o_attach = osc_attach,
3003 .o_detach = osc_detach,
3004 .o_setup = osc_setup,
3005 .o_cleanup = osc_cleanup,
3006 .o_connect = osc_connect,
3007 .o_disconnect = osc_disconnect,
3008 .o_statfs = osc_statfs,
3009 .o_packmd = osc_packmd,
3010 .o_unpackmd = osc_unpackmd,
3011 .o_create = osc_create,
3012 .o_destroy = osc_destroy,
3013 .o_getattr = osc_getattr,
3014 .o_getattr_async = osc_getattr_async,
3015 .o_setattr = osc_setattr,
3017 .o_brw_async = osc_brw_async,
3018 .o_prep_async_page = osc_prep_async_page,
3019 .o_queue_async_io = osc_queue_async_io,
3020 .o_set_async_flags = osc_set_async_flags,
3021 .o_queue_group_io = osc_queue_group_io,
3022 .o_trigger_group_io = osc_trigger_group_io,
3023 .o_teardown_async_page = osc_teardown_async_page,
3024 .o_punch = osc_punch,
3026 .o_enqueue = osc_enqueue,
3027 .o_match = osc_match,
3028 .o_change_cbdata = osc_change_cbdata,
3029 .o_cancel = osc_cancel,
3030 .o_cancel_unused = osc_cancel_unused,
3031 .o_iocontrol = osc_iocontrol,
3032 .o_get_info = osc_get_info,
3033 .o_set_info = osc_set_info,
3034 .o_import_event = osc_import_event,
3035 .o_llog_init = osc_llog_init,
3036 .o_llog_finish = osc_llog_finish,
3039 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3040 struct obd_ops sanosc_obd_ops = {
3041 .o_owner = THIS_MODULE,
3042 .o_attach = osc_attach,
3043 .o_detach = osc_detach,
3044 .o_cleanup = client_obd_cleanup,
3045 .o_connect = osc_connect,
3046 .o_disconnect = client_disconnect_export,
3047 .o_statfs = osc_statfs,
3048 .o_packmd = osc_packmd,
3049 .o_unpackmd = osc_unpackmd,
3050 .o_create = osc_real_create,
3051 .o_destroy = osc_destroy,
3052 .o_getattr = osc_getattr,
3053 .o_getattr_async = osc_getattr_async,
3054 .o_setattr = osc_setattr,
3055 .o_setup = client_sanobd_setup,
3056 .o_brw = sanosc_brw,
3057 .o_punch = osc_punch,
3059 .o_enqueue = osc_enqueue,
3060 .o_match = osc_match,
3061 .o_change_cbdata = osc_change_cbdata,
3062 .o_cancel = osc_cancel,
3063 .o_cancel_unused = osc_cancel_unused,
3064 .o_iocontrol = osc_iocontrol,
3065 .o_import_event = osc_import_event,
3066 .o_llog_init = osc_llog_init,
3067 .o_llog_finish = osc_llog_finish,
3071 int __init osc_init(void)
3073 struct lprocfs_static_vars lvars;
3074 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3075 struct lprocfs_static_vars sanlvars;
3080 lprocfs_init_vars(osc, &lvars);
3081 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3082 lprocfs_init_vars(osc, &sanlvars);
3085 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3090 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3091 rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3092 LUSTRE_SANOSC_NAME);
3094 class_unregister_type(LUSTRE_OSC_NAME);
3101 static void /*__exit*/ osc_exit(void)
3103 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3104 class_unregister_type(LUSTRE_SANOSC_NAME);
3106 class_unregister_type(LUSTRE_OSC_NAME);
3109 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3110 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3111 MODULE_LICENSE("GPL");
3113 module_init(osc_init);
3114 module_exit(osc_exit);