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), LUSTRE_OBD_VERSION,
188 OST_GETATTR, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
215 OST_GETATTR, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
261 OST_SETATTR, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
307 OST_CREATE, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
385 OST_PUNCH, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
432 OST_SYNC, 1, &size, NULL);
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), LUSTRE_OBD_VERSION,
479 OST_DESTROY, 1, &size, NULL);
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);
501 body = lustre_swab_repbuf(request, 0, sizeof(*body),
502 lustre_swab_ost_body);
504 CERROR ("Can't unpack body\n");
505 GOTO (out, rc = -EPROTO);
508 memcpy(oa, &body->oa, sizeof(*oa));
512 ptlrpc_req_finished(request);
516 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
519 obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
521 LASSERT(!(oa->o_valid & bits));
524 spin_lock(&cli->cl_loi_list_lock);
525 oa->o_dirty = cli->cl_dirty;
526 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
527 oa->o_grant = cli->cl_avail_grant;
528 oa->o_dropped = cli->cl_lost_grant;
529 cli->cl_lost_grant = 0;
530 spin_unlock(&cli->cl_loi_list_lock);
531 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
532 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
535 /* caller must hold loi_list_lock */
536 static void osc_consume_write_grant(struct client_obd *cli,
537 struct osc_async_page *oap)
539 cli->cl_dirty += PAGE_SIZE;
540 cli->cl_avail_grant -= PAGE_SIZE;
541 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
542 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
543 LASSERT(cli->cl_avail_grant >= 0);
546 static unsigned long rpcs_in_flight(struct client_obd *cli)
548 return cli->cl_r_in_flight + cli->cl_w_in_flight;
551 /* caller must hold loi_list_lock */
552 void osc_wake_cache_waiters(struct client_obd *cli)
554 struct list_head *l, *tmp;
555 struct osc_cache_waiter *ocw;
557 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
558 /* if we can't dirty more, we must wait until some is written */
559 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
560 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
561 cli->cl_dirty, cli->cl_dirty_max);
565 /* if still dirty cache but no grant wait for pending RPCs that
566 * may yet return us some grant before doing sync writes */
567 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
568 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
569 cli->cl_w_in_flight);
571 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
572 list_del_init(&ocw->ocw_entry);
573 if (cli->cl_avail_grant < PAGE_SIZE) {
574 /* no more RPCs in flight to return grant, do sync IO */
575 ocw->ocw_rc = -EDQUOT;
576 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
578 osc_consume_write_grant(cli, ocw->ocw_oap);
581 wake_up(&ocw->ocw_waitq);
587 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
589 spin_lock(&cli->cl_loi_list_lock);
590 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
591 cli->cl_avail_grant += body->oa.o_grant;
592 /* waiters are woken in brw_interpret_oap */
593 spin_unlock(&cli->cl_loi_list_lock);
596 /* We assume that the reason this OSC got a short read is because it read
597 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
598 * via the LOV, and it _knows_ it's reading inside the file, it's just that
599 * this stripe never got written at or beyond this stripe offset yet. */
600 static void handle_short_read(int nob_read, obd_count page_count,
601 struct brw_page *pga)
605 /* skip bytes read OK */
606 while (nob_read > 0) {
607 LASSERT (page_count > 0);
609 if (pga->count > nob_read) {
610 /* EOF inside this page */
611 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
612 memset(ptr + nob_read, 0, pga->count - nob_read);
619 nob_read -= pga->count;
624 /* zero remaining pages */
625 while (page_count-- > 0) {
626 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
627 memset(ptr, 0, pga->count);
633 static int check_write_rcs(struct ptlrpc_request *request,
634 int requested_nob, int niocount,
635 obd_count page_count, struct brw_page *pga)
639 /* return error if any niobuf was in error */
640 remote_rcs = lustre_swab_repbuf(request, 1,
641 sizeof(*remote_rcs) * niocount, NULL);
642 if (remote_rcs == NULL) {
643 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
646 if (lustre_msg_swabbed(request->rq_repmsg))
647 for (i = 0; i < niocount; i++)
648 __swab32s(&remote_rcs[i]);
650 for (i = 0; i < niocount; i++) {
651 if (remote_rcs[i] < 0)
652 return(remote_rcs[i]);
654 if (remote_rcs[i] != 0) {
655 CERROR("rc[%d] invalid (%d) req %p\n",
656 i, remote_rcs[i], request);
661 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
662 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
663 requested_nob, request->rq_bulk->bd_nob_transferred);
670 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
672 if (p1->flag != p2->flag) {
673 unsigned mask = ~OBD_BRW_FROM_GRANT;
675 /* warn if we try to combine flags that we don't know to be
677 if ((p1->flag & mask) != (p2->flag & mask))
678 CERROR("is it ok to have flags 0x%x and 0x%x in the "
679 "same brw?\n", p1->flag, p2->flag);
683 return (p1->disk_offset + p1->count == p2->disk_offset);
687 static obd_count cksum_pages(int nob, obd_count page_count,
688 struct brw_page *pga)
694 LASSERT (page_count > 0);
697 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
698 pga->count > nob ? nob : pga->count);
710 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
711 struct lov_stripe_md *lsm, obd_count page_count,
712 struct brw_page *pga, int *requested_nobp,
713 int *niocountp, struct ptlrpc_request **reqp)
715 struct ptlrpc_request *req;
716 struct ptlrpc_bulk_desc *desc;
717 struct client_obd *cli = &imp->imp_obd->u.cli;
718 struct ost_body *body;
719 struct obd_ioobj *ioobj;
720 struct niobuf_remote *niobuf;
728 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
730 for (niocount = i = 1; i < page_count; i++)
731 if (!can_merge_pages(&pga[i - 1], &pga[i]))
734 size[0] = sizeof(*body);
735 size[1] = sizeof(*ioobj);
736 size[2] = niocount * sizeof(*niobuf);
738 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, 3, size, NULL);
742 if (opc == OST_WRITE)
743 desc = ptlrpc_prep_bulk_imp (req, page_count,
744 BULK_GET_SOURCE, OST_BULK_PORTAL);
746 desc = ptlrpc_prep_bulk_imp (req, page_count,
747 BULK_PUT_SINK, OST_BULK_PORTAL);
749 GOTO(out, rc = -ENOMEM);
750 /* NB request now owns desc and will free it when it gets freed */
752 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
753 ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
754 niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
756 memcpy(&body->oa, oa, sizeof(*oa));
758 obdo_to_ioobj(oa, ioobj);
759 ioobj->ioo_bufcnt = niocount;
761 LASSERT (page_count > 0);
763 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
764 struct brw_page *pg = &pga[i];
765 struct brw_page *pg_prev = pg - 1;
767 LASSERT(pg->count > 0);
768 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
769 "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
770 pg->page_offset, pg->count);
771 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
772 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
773 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
775 pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
776 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
777 pg_prev->disk_offset);
779 ptlrpc_prep_bulk_page(desc, pg->pg,
780 pg->page_offset & ~PAGE_MASK, pg->count);
781 requested_nob += pg->count;
783 if (i > 0 && can_merge_pages(pg_prev, pg)) {
785 niobuf->len += pg->count;
787 niobuf->offset = pg->disk_offset;
788 niobuf->len = pg->count;
789 niobuf->flags = pg->flag;
793 LASSERT((void *)(niobuf - niocount) ==
794 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
795 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
797 /* size[0] still sizeof (*body) */
798 if (opc == OST_WRITE) {
800 body->oa.o_valid |= OBD_MD_FLCKSUM;
801 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
803 /* 1 RC per niobuf */
804 size[1] = sizeof(__u32) * niocount;
805 req->rq_replen = lustre_msg_size(2, size);
807 /* 1 RC for the whole I/O */
808 req->rq_replen = lustre_msg_size(1, size);
811 *niocountp = niocount;
812 *requested_nobp = requested_nob;
817 ptlrpc_req_finished (req);
821 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
822 int requested_nob, int niocount,
823 obd_count page_count, struct brw_page *pga,
826 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
827 struct ost_body *body;
833 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
835 CERROR ("Can't unpack body\n");
839 osc_update_grant(cli, body);
840 memcpy(oa, &body->oa, sizeof(*oa));
842 if (req->rq_reqmsg->opc == OST_WRITE) {
844 CERROR ("Unexpected +ve rc %d\n", rc);
847 LASSERT (req->rq_bulk->bd_nob == requested_nob);
849 RETURN(check_write_rcs(req, requested_nob, niocount,
853 if (rc > requested_nob) {
854 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
858 if (rc != req->rq_bulk->bd_nob_transferred) {
859 CERROR ("Unexpected rc %d (%d transferred)\n",
860 rc, req->rq_bulk->bd_nob_transferred);
864 if (rc < requested_nob)
865 handle_short_read(rc, page_count, pga);
868 if (oa->o_valid & OBD_MD_FLCKSUM) {
869 const struct ptlrpc_peer *peer =
870 &req->rq_import->imp_connection->c_peer;
871 static int cksum_counter;
872 obd_count server_cksum = oa->o_cksum;
873 obd_count cksum = cksum_pages(rc, page_count, pga);
874 char str[PTL_NALFMT_SIZE];
876 ptlrpc_peernid2str(peer, str);
879 if (server_cksum != cksum) {
880 CERROR("Bad checksum: server %x, client %x, server NID "
881 LPX64" (%s)\n", server_cksum, cksum,
882 peer->peer_id.nid, str);
885 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
886 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
887 cksum_counter, peer->peer_id.nid, str, cksum);
890 static int cksum_missed;
893 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
894 CERROR("Request checksum %u from "LPX64", no reply\n",
896 req->rq_import->imp_connection->c_peer.peer_id.nid);
902 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
903 struct lov_stripe_md *lsm,
904 obd_count page_count, struct brw_page *pga)
908 struct ptlrpc_request *request;
913 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
914 page_count, pga, &requested_nob, &niocount,
919 rc = ptlrpc_queue_wait(request);
921 if (rc == -ETIMEDOUT && request->rq_resend) {
922 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
923 ptlrpc_req_finished(request);
927 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
928 page_count, pga, rc);
930 ptlrpc_req_finished(request);
934 static int brw_interpret(struct ptlrpc_request *request,
935 struct osc_brw_async_args *aa, int rc)
937 struct obdo *oa = aa->aa_oa;
938 int requested_nob = aa->aa_requested_nob;
939 int niocount = aa->aa_nio_count;
940 obd_count page_count = aa->aa_page_count;
941 struct brw_page *pga = aa->aa_pga;
944 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
945 page_count, pga, rc);
949 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
950 struct lov_stripe_md *lsm, obd_count page_count,
951 struct brw_page *pga, struct ptlrpc_request_set *set)
953 struct ptlrpc_request *request;
956 struct osc_brw_async_args *aa;
960 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
961 page_count, pga, &requested_nob, &nio_count,
964 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
965 aa = (struct osc_brw_async_args *)&request->rq_async_args;
967 aa->aa_requested_nob = requested_nob;
968 aa->aa_nio_count = nio_count;
969 aa->aa_page_count = page_count;
972 request->rq_interpret_reply = brw_interpret;
973 ptlrpc_set_add_req(set, request);
979 #define min_t(type,x,y) \
980 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
984 * ugh, we want disk allocation on the target to happen in offset order. we'll
985 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
986 * fine for our small page arrays and doesn't require allocation. its an
987 * insertion sort that swaps elements that are strides apart, shrinking the
988 * stride down until its '1' and the array is sorted.
990 static void sort_brw_pages(struct brw_page *array, int num)
997 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1002 for (i = stride ; i < num ; i++) {
1005 while (j >= stride && array[j - stride].disk_offset >
1007 array[j] = array[j - stride];
1012 } while (stride > 1);
1015 /* make sure we the regions we're passing to elan don't violate its '4
1016 * fragments' constraint. portal headers are a fragment, all full
1017 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1018 * counts as a fragment. I think. see bug 934. */
1019 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1022 int saw_whole_frag = 0;
1025 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1026 if (pg->count == PAGE_SIZE) {
1027 if (!saw_whole_frag) {
1038 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1039 struct lov_stripe_md *md, obd_count page_count,
1040 struct brw_page *pga, struct obd_trans_info *oti)
1044 if (cmd == OBD_BRW_CHECK) {
1045 /* The caller just wants to know if there's a chance that this
1046 * I/O can succeed */
1047 struct obd_import *imp = class_exp2cliimp(exp);
1049 if (imp == NULL || imp->imp_invalid)
1054 while (page_count) {
1055 obd_count pages_per_brw;
1058 if (page_count > PTLRPC_MAX_BRW_PAGES)
1059 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1061 pages_per_brw = page_count;
1063 sort_brw_pages(pga, pages_per_brw);
1064 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1066 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1071 page_count -= pages_per_brw;
1072 pga += pages_per_brw;
1077 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1078 struct lov_stripe_md *md, obd_count page_count,
1079 struct brw_page *pga, struct ptlrpc_request_set *set,
1080 struct obd_trans_info *oti)
1084 if (cmd == OBD_BRW_CHECK) {
1085 /* The caller just wants to know if there's a chance that this
1086 * I/O can succeed */
1087 struct obd_import *imp = class_exp2cliimp(exp);
1089 if (imp == NULL || imp->imp_invalid)
1094 while (page_count) {
1095 obd_count pages_per_brw;
1098 if (page_count > PTLRPC_MAX_BRW_PAGES)
1099 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1101 pages_per_brw = page_count;
1103 sort_brw_pages(pga, pages_per_brw);
1104 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1106 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1111 page_count -= pages_per_brw;
1112 pga += pages_per_brw;
1117 static void osc_check_rpcs(struct client_obd *cli);
1118 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1120 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1121 static void lop_update_pending(struct client_obd *cli,
1122 struct loi_oap_pages *lop, int cmd, int delta);
1124 /* this is called when a sync waiter receives an interruption. Its job is to
1125 * get the caller woken as soon as possible. If its page hasn't been put in an
1126 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1127 * desiring interruption which will forcefully complete the rpc once the rpc
1129 static void osc_occ_interrupted(struct oig_callback_context *occ)
1131 struct osc_async_page *oap;
1132 struct loi_oap_pages *lop;
1133 struct lov_oinfo *loi;
1136 /* XXX member_of() */
1137 oap = list_entry(occ, struct osc_async_page, oap_occ);
1139 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1141 oap->oap_interrupted = 1;
1143 /* ok, it's been put in an rpc. */
1144 if (oap->oap_request != NULL) {
1145 ptlrpc_mark_interrupted(oap->oap_request);
1146 ptlrpcd_wake(oap->oap_request);
1150 /* we don't get interruption callbacks until osc_trigger_sync_io()
1151 * has been called and put the sync oaps in the pending/urgent lists.*/
1152 if (!list_empty(&oap->oap_pending_item)) {
1153 list_del_init(&oap->oap_pending_item);
1154 if (oap->oap_async_flags & ASYNC_URGENT)
1155 list_del_init(&oap->oap_urgent_item);
1158 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1159 &loi->loi_write_lop : &loi->loi_read_lop;
1160 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1161 loi_list_maint(oap->oap_cli, oap->oap_loi);
1163 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1164 oap->oap_oig = NULL;
1168 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1171 /* this must be called holding the loi list lock to give coverage to exit_cache,
1172 * async_flag maintenance, and oap_request */
1173 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1174 struct osc_async_page *oap, int sent, int rc)
1176 osc_exit_cache(cli, oap, sent);
1177 oap->oap_async_flags = 0;
1178 oap->oap_interrupted = 0;
1180 if (oap->oap_request != NULL) {
1181 ptlrpc_req_finished(oap->oap_request);
1182 oap->oap_request = NULL;
1185 if (rc == 0 && oa != NULL)
1186 oap->oap_loi->loi_blocks = oa->o_blocks;
1189 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1190 oap->oap_oig = NULL;
1195 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1199 static int brw_interpret_oap(struct ptlrpc_request *request,
1200 struct osc_brw_async_args *aa, int rc)
1202 struct osc_async_page *oap;
1203 struct client_obd *cli;
1204 struct list_head *pos, *n;
1208 do_gettimeofday(&now);
1209 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1210 aa->aa_nio_count, aa->aa_page_count,
1213 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1216 /* in failout recovery we ignore writeback failure and want
1217 * to just tell llite to unlock the page and continue */
1218 if (request->rq_reqmsg->opc == OST_WRITE &&
1219 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1220 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1222 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1226 spin_lock(&cli->cl_loi_list_lock);
1228 if (request->rq_reqmsg->opc == OST_WRITE)
1229 lprocfs_stime_record(&cli->cl_write_stime, &now,
1230 &request->rq_rpcd_start);
1232 lprocfs_stime_record(&cli->cl_read_stime, &now,
1233 &request->rq_rpcd_start);
1237 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1238 * is called so we know whether to go to sync BRWs or wait for more
1239 * RPCs to complete */
1240 if (request->rq_reqmsg->opc == OST_WRITE)
1241 cli->cl_w_in_flight--;
1243 cli->cl_r_in_flight--;
1245 /* the caller may re-use the oap after the completion call so
1246 * we need to clean it up a little */
1247 list_for_each_safe(pos, n, &aa->aa_oaps) {
1248 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1250 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1251 //oap->oap_page, oap->oap_page->index, oap);
1253 list_del_init(&oap->oap_rpc_item);
1254 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1257 osc_wake_cache_waiters(cli);
1258 osc_check_rpcs(cli);
1260 spin_unlock(&cli->cl_loi_list_lock);
1262 obdo_free(aa->aa_oa);
1263 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1268 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1269 struct list_head *rpc_list,
1270 int page_count, int cmd)
1272 struct ptlrpc_request *req;
1273 struct brw_page *pga = NULL;
1274 int requested_nob, nio_count;
1275 struct osc_brw_async_args *aa;
1276 struct obdo *oa = NULL;
1277 struct obd_async_page_ops *ops = NULL;
1278 void *caller_data = NULL;
1279 struct list_head *pos;
1282 LASSERT(!list_empty(rpc_list));
1284 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1286 RETURN(ERR_PTR(-ENOMEM));
1290 GOTO(out, req = ERR_PTR(-ENOMEM));
1293 list_for_each(pos, rpc_list) {
1294 struct osc_async_page *oap;
1296 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1298 ops = oap->oap_caller_ops;
1299 caller_data = oap->oap_caller_data;
1301 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1302 pga[i].page_offset = pga[i].disk_offset;
1303 pga[i].pg = oap->oap_page;
1304 pga[i].count = oap->oap_count;
1305 pga[i].flag = oap->oap_brw_flags;
1306 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1307 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1311 /* always get the data for the obdo for the rpc */
1312 LASSERT(ops != NULL);
1313 ops->ap_fill_obdo(caller_data, cmd, oa);
1315 sort_brw_pages(pga, page_count);
1316 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1317 pga, &requested_nob, &nio_count, &req);
1319 CERROR("prep_req failed: %d\n", rc);
1320 GOTO(out, req = ERR_PTR(rc));
1323 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1324 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1326 aa->aa_requested_nob = requested_nob;
1327 aa->aa_nio_count = nio_count;
1328 aa->aa_page_count = page_count;
1337 OBD_FREE(pga, sizeof(*pga) * page_count);
1342 static void lop_update_pending(struct client_obd *cli,
1343 struct loi_oap_pages *lop, int cmd, int delta)
1345 lop->lop_num_pending += delta;
1346 if (cmd == OBD_BRW_WRITE)
1347 cli->cl_pending_w_pages += delta;
1349 cli->cl_pending_r_pages += delta;
1352 /* the loi lock is held across this function but it's allowed to release
1353 * and reacquire it during its work */
1354 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1355 int cmd, struct loi_oap_pages *lop)
1357 struct ptlrpc_request *request;
1358 obd_count page_count = 0;
1359 struct list_head *tmp, *pos;
1360 struct osc_async_page *oap = NULL;
1361 struct osc_brw_async_args *aa;
1362 struct obd_async_page_ops *ops;
1363 LIST_HEAD(rpc_list);
1366 /* first we find the pages we're allowed to work with */
1367 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1368 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1369 ops = oap->oap_caller_ops;
1371 LASSERT(oap->oap_magic == OAP_MAGIC);
1373 /* in llite being 'ready' equates to the page being locked
1374 * until completion unlocks it. commit_write submits a page
1375 * as not ready because its unlock will happen unconditionally
1376 * as the call returns. if we race with commit_write giving
1377 * us that page we dont' want to create a hole in the page
1378 * stream, so we stop and leave the rpc to be fired by
1379 * another dirtier or kupdated interval (the not ready page
1380 * will still be on the dirty list). we could call in
1381 * at the end of ll_file_write to process the queue again. */
1382 if (!(oap->oap_async_flags & ASYNC_READY)) {
1383 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1385 CDEBUG(D_INODE, "oap %p page %p returned %d "
1386 "instead of ready\n", oap,
1390 /* llite is telling us that the page is still
1391 * in commit_write and that we should try
1392 * and put it in an rpc again later. we
1393 * break out of the loop so we don't create
1394 * a hole in the sequence of pages in the rpc
1399 /* the io isn't needed.. tell the checks
1400 * below to complete the rpc with EINTR */
1401 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1402 oap->oap_count = -EINTR;
1405 oap->oap_async_flags |= ASYNC_READY;
1408 LASSERTF(0, "oap %p page %p returned %d "
1409 "from make_ready\n", oap,
1417 /* take the page out of our book-keeping */
1418 list_del_init(&oap->oap_pending_item);
1419 lop_update_pending(cli, lop, cmd, -1);
1420 list_del_init(&oap->oap_urgent_item);
1422 /* ask the caller for the size of the io as the rpc leaves. */
1423 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1425 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1426 if (oap->oap_count <= 0) {
1427 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1429 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1433 /* now put the page back in our accounting */
1434 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1435 if (++page_count >= cli->cl_max_pages_per_rpc)
1439 osc_wake_cache_waiters(cli);
1441 if (page_count == 0)
1444 loi_list_maint(cli, loi);
1445 spin_unlock(&cli->cl_loi_list_lock);
1447 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1448 if (IS_ERR(request)) {
1449 /* this should happen rarely and is pretty bad, it makes the
1450 * pending list not follow the dirty order */
1451 spin_lock(&cli->cl_loi_list_lock);
1452 list_for_each_safe(pos, tmp, &rpc_list) {
1453 oap = list_entry(pos, struct osc_async_page,
1455 list_del_init(&oap->oap_rpc_item);
1457 /* queued sync pages can be torn down while the pages
1458 * were between the pending list and the rpc */
1459 if (oap->oap_interrupted) {
1460 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1461 osc_ap_completion(cli, NULL, oap, 0,
1466 /* put the page back in the loi/lop lists */
1467 list_add_tail(&oap->oap_pending_item,
1469 lop_update_pending(cli, lop, cmd, 1);
1470 if (oap->oap_async_flags & ASYNC_URGENT)
1471 list_add(&oap->oap_urgent_item,
1474 loi_list_maint(cli, loi);
1475 RETURN(PTR_ERR(request));
1478 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1479 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1480 INIT_LIST_HEAD(&aa->aa_oaps);
1481 list_splice(&rpc_list, &aa->aa_oaps);
1482 INIT_LIST_HEAD(&rpc_list);
1485 if (cmd == OBD_BRW_READ) {
1486 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1487 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1489 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1490 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1491 cli->cl_w_in_flight);
1495 spin_lock(&cli->cl_loi_list_lock);
1497 if (cmd == OBD_BRW_READ)
1498 cli->cl_r_in_flight++;
1500 cli->cl_w_in_flight++;
1501 /* queued sync pages can be torn down while the pages
1502 * were between the pending list and the rpc */
1503 list_for_each(pos, &aa->aa_oaps) {
1504 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1505 if (oap->oap_interrupted) {
1506 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1508 ptlrpc_mark_interrupted(request);
1513 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %dr/%dw in flight\n",
1514 request, page_count, aa, cli->cl_r_in_flight,
1515 cli->cl_w_in_flight);
1517 oap->oap_request = ptlrpc_request_addref(request);
1518 request->rq_interpret_reply = brw_interpret_oap;
1519 ptlrpcd_add_req(request);
1523 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1529 if (lop->lop_num_pending == 0)
1532 /* if we have an invalid import we want to drain the queued pages
1533 * by forcing them through rpcs that immediately fail and complete
1534 * the pages. recovery relies on this to empty the queued pages
1535 * before canceling the locks and evicting down the llite pages */
1536 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1539 /* stream rpcs in queue order as long as as there is an urgent page
1540 * queued. this is our cheap solution for good batching in the case
1541 * where writepage marks some random page in the middle of the file as
1542 * urgent because of, say, memory pressure */
1543 if (!list_empty(&lop->lop_urgent))
1546 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1547 optimal = cli->cl_max_pages_per_rpc;
1548 if (cmd == OBD_BRW_WRITE) {
1549 /* trigger a write rpc stream as long as there are dirtiers
1550 * waiting for space. as they're waiting, they're not going to
1551 * create more pages to coallesce with what's waiting.. */
1552 if (!list_empty(&cli->cl_cache_waiters))
1555 /* *2 to avoid triggering rpcs that would want to include pages
1556 * that are being queued but which can't be made ready until
1557 * the queuer finishes with the page. this is a wart for
1558 * llite::commit_write() */
1561 if (lop->lop_num_pending >= optimal)
1567 static void on_list(struct list_head *item, struct list_head *list,
1570 if (list_empty(item) && should_be_on)
1571 list_add_tail(item, list);
1572 else if (!list_empty(item) && !should_be_on)
1573 list_del_init(item);
1576 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1577 * can find pages to build into rpcs quickly */
1578 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1580 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1581 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1582 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1584 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1585 loi->loi_write_lop.lop_num_pending);
1587 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1588 loi->loi_read_lop.lop_num_pending);
1591 #define LOI_DEBUG(LOI, STR, args...) \
1592 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1593 !list_empty(&(LOI)->loi_cli_item), \
1594 (LOI)->loi_write_lop.lop_num_pending, \
1595 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1596 (LOI)->loi_read_lop.lop_num_pending, \
1597 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1600 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1603 /* first return all objects which we already know to have
1604 * pages ready to be stuffed into rpcs */
1605 if (!list_empty(&cli->cl_loi_ready_list))
1606 RETURN(list_entry(cli->cl_loi_ready_list.next,
1607 struct lov_oinfo, loi_cli_item));
1609 /* then if we have cache waiters, return all objects with queued
1610 * writes. This is especially important when many small files
1611 * have filled up the cache and not been fired into rpcs because
1612 * they don't pass the nr_pending/object threshhold */
1613 if (!list_empty(&cli->cl_cache_waiters) &&
1614 !list_empty(&cli->cl_loi_write_list))
1615 RETURN(list_entry(cli->cl_loi_write_list.next,
1616 struct lov_oinfo, loi_write_item));
1618 /* then return all queued objects when we have an invalid import
1619 * so that they get flushed */
1620 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1621 if (!list_empty(&cli->cl_loi_write_list))
1622 RETURN(list_entry(cli->cl_loi_write_list.next,
1623 struct lov_oinfo, loi_write_item));
1624 if (!list_empty(&cli->cl_loi_read_list))
1625 RETURN(list_entry(cli->cl_loi_read_list.next,
1626 struct lov_oinfo, loi_read_item));
1631 /* called with the loi list lock held */
1632 static void osc_check_rpcs(struct client_obd *cli)
1634 struct lov_oinfo *loi;
1635 int rc = 0, race_counter = 0;
1638 while ((loi = osc_next_loi(cli)) != NULL) {
1639 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1641 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1644 /* attempt some read/write balancing by alternating between
1645 * reads and writes in an object. The makes_rpc checks here
1646 * would be redundant if we were getting read/write work items
1647 * instead of objects. we don't want send_oap_rpc to drain a
1648 * partial read pending queue when we're given this object to
1649 * do io on writes while there are cache waiters */
1650 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1651 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1652 &loi->loi_write_lop);
1660 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1661 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1662 &loi->loi_read_lop);
1671 /* attempt some inter-object balancing by issueing rpcs
1672 * for each object in turn */
1673 if (!list_empty(&loi->loi_cli_item))
1674 list_del_init(&loi->loi_cli_item);
1675 if (!list_empty(&loi->loi_write_item))
1676 list_del_init(&loi->loi_write_item);
1677 if (!list_empty(&loi->loi_read_item))
1678 list_del_init(&loi->loi_read_item);
1680 loi_list_maint(cli, loi);
1682 /* send_oap_rpc fails with 0 when make_ready tells it to
1683 * back off. llite's make_ready does this when it tries
1684 * to lock a page queued for write that is already locked.
1685 * we want to try sending rpcs from many objects, but we
1686 * don't want to spin failing with 0. */
1687 if (race_counter == 10)
1693 /* we're trying to queue a page in the osc so we're subject to the
1694 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1695 * If the osc's queued pages are already at that limit, then we want to sleep
1696 * until there is space in the osc's queue for us. We also may be waiting for
1697 * write credits from the OST if there are RPCs in flight that may return some
1698 * before we fall back to sync writes.
1700 * We need this know our allocation was granted in the presence of signals */
1701 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1705 spin_lock(&cli->cl_loi_list_lock);
1706 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1707 spin_unlock(&cli->cl_loi_list_lock);
1711 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1712 * grant or cache space. */
1713 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1714 struct osc_async_page *oap)
1716 struct osc_cache_waiter ocw;
1717 struct l_wait_info lwi = { 0 };
1718 struct timeval start, stop;
1720 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1721 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1722 cli->cl_avail_grant);
1724 if (cli->cl_dirty_max < PAGE_SIZE)
1727 /* Hopefully normal case - cache space and write credits available */
1728 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1729 cli->cl_avail_grant >= PAGE_SIZE) {
1730 /* account for ourselves */
1731 osc_consume_write_grant(cli, oap);
1735 /* Make sure that there are write rpcs in flight to wait for. This
1736 * is a little silly as this object may not have any pending but
1737 * other objects sure might. */
1738 if (cli->cl_w_in_flight) {
1739 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1740 init_waitqueue_head(&ocw.ocw_waitq);
1744 loi_list_maint(cli, loi);
1745 osc_check_rpcs(cli);
1746 spin_unlock(&cli->cl_loi_list_lock);
1748 CDEBUG(0, "sleeping for cache space\n");
1749 do_gettimeofday(&start);
1750 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1751 do_gettimeofday(&stop);
1752 spin_lock(&cli->cl_loi_list_lock);
1753 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1754 if (!list_empty(&ocw.ocw_entry)) {
1755 list_del(&ocw.ocw_entry);
1764 /* the companion to enter_cache, called when an oap is no longer part of the
1765 * dirty accounting.. so writeback completes or truncate happens before writing
1766 * starts. must be called with the loi lock held. */
1767 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1772 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1777 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1778 cli->cl_dirty -= PAGE_SIZE;
1780 cli->cl_lost_grant += PAGE_SIZE;
1781 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1782 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1788 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1789 struct lov_oinfo *loi, struct page *page,
1790 obd_off offset, struct obd_async_page_ops *ops,
1791 void *data, void **res)
1793 struct osc_async_page *oap;
1796 OBD_ALLOC(oap, sizeof(*oap));
1800 oap->oap_magic = OAP_MAGIC;
1801 oap->oap_cli = &exp->exp_obd->u.cli;
1804 oap->oap_caller_ops = ops;
1805 oap->oap_caller_data = data;
1807 oap->oap_page = page;
1808 oap->oap_obj_off = offset;
1810 INIT_LIST_HEAD(&oap->oap_pending_item);
1811 INIT_LIST_HEAD(&oap->oap_urgent_item);
1812 INIT_LIST_HEAD(&oap->oap_rpc_item);
1814 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1816 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1821 struct osc_async_page *oap_from_cookie(void *cookie)
1823 struct osc_async_page *oap = cookie;
1824 if (oap->oap_magic != OAP_MAGIC)
1825 return ERR_PTR(-EINVAL);
1829 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1830 struct lov_oinfo *loi, void *cookie,
1831 int cmd, obd_off off, int count,
1832 obd_flags brw_flags, enum async_flags async_flags)
1834 struct client_obd *cli = &exp->exp_obd->u.cli;
1835 struct osc_async_page *oap;
1836 struct loi_oap_pages *lop;
1840 oap = oap_from_cookie(cookie);
1842 RETURN(PTR_ERR(oap));
1844 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1847 if (!list_empty(&oap->oap_pending_item) ||
1848 !list_empty(&oap->oap_urgent_item) ||
1849 !list_empty(&oap->oap_rpc_item))
1853 loi = &lsm->lsm_oinfo[0];
1855 spin_lock(&cli->cl_loi_list_lock);
1858 oap->oap_async_flags = async_flags;
1859 oap->oap_page_off = off;
1860 oap->oap_count = count;
1861 oap->oap_brw_flags = brw_flags;
1863 if (cmd == OBD_BRW_WRITE) {
1864 rc = osc_enter_cache(cli, loi, oap);
1866 spin_unlock(&cli->cl_loi_list_lock);
1869 lop = &loi->loi_write_lop;
1871 lop = &loi->loi_read_lop;
1874 if (oap->oap_async_flags & ASYNC_URGENT)
1875 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1876 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1877 lop_update_pending(cli, lop, cmd, 1);
1879 loi_list_maint(cli, loi);
1881 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1884 osc_check_rpcs(cli);
1885 spin_unlock(&cli->cl_loi_list_lock);
1890 /* aka (~was & now & flag), but this is more clear :) */
1891 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1893 static int osc_set_async_flags(struct obd_export *exp,
1894 struct lov_stripe_md *lsm,
1895 struct lov_oinfo *loi, void *cookie,
1896 obd_flags async_flags)
1898 struct client_obd *cli = &exp->exp_obd->u.cli;
1899 struct loi_oap_pages *lop;
1900 struct osc_async_page *oap;
1904 oap = oap_from_cookie(cookie);
1906 RETURN(PTR_ERR(oap));
1908 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1912 loi = &lsm->lsm_oinfo[0];
1914 if (oap->oap_cmd == OBD_BRW_WRITE) {
1915 lop = &loi->loi_write_lop;
1917 lop = &loi->loi_read_lop;
1920 spin_lock(&cli->cl_loi_list_lock);
1922 if (list_empty(&oap->oap_pending_item))
1923 GOTO(out, rc = -EINVAL);
1925 if ((oap->oap_async_flags & async_flags) == async_flags)
1928 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1929 oap->oap_async_flags |= ASYNC_READY;
1931 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1932 if (list_empty(&oap->oap_rpc_item)) {
1933 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1934 loi_list_maint(cli, loi);
1938 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1939 oap->oap_async_flags);
1941 osc_check_rpcs(cli);
1942 spin_unlock(&cli->cl_loi_list_lock);
1946 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1947 struct lov_oinfo *loi,
1948 struct obd_io_group *oig, void *cookie,
1949 int cmd, obd_off off, int count,
1950 obd_flags brw_flags,
1951 obd_flags async_flags)
1953 struct client_obd *cli = &exp->exp_obd->u.cli;
1954 struct osc_async_page *oap;
1955 struct loi_oap_pages *lop;
1958 oap = oap_from_cookie(cookie);
1960 RETURN(PTR_ERR(oap));
1962 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1965 if (!list_empty(&oap->oap_pending_item) ||
1966 !list_empty(&oap->oap_urgent_item) ||
1967 !list_empty(&oap->oap_rpc_item))
1971 loi = &lsm->lsm_oinfo[0];
1973 spin_lock(&cli->cl_loi_list_lock);
1976 oap->oap_page_off = off;
1977 oap->oap_count = count;
1978 oap->oap_brw_flags = brw_flags;
1979 oap->oap_async_flags = async_flags;
1981 if (cmd == OBD_BRW_WRITE)
1982 lop = &loi->loi_write_lop;
1984 lop = &loi->loi_read_lop;
1986 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
1987 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
1989 oig_add_one(oig, &oap->oap_occ);
1992 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
1994 spin_unlock(&cli->cl_loi_list_lock);
1999 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2000 struct loi_oap_pages *lop, int cmd)
2002 struct list_head *pos, *tmp;
2003 struct osc_async_page *oap;
2005 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2006 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2007 list_del(&oap->oap_pending_item);
2008 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2009 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2010 lop_update_pending(cli, lop, cmd, 1);
2012 loi_list_maint(cli, loi);
2015 static int osc_trigger_group_io(struct obd_export *exp,
2016 struct lov_stripe_md *lsm,
2017 struct lov_oinfo *loi,
2018 struct obd_io_group *oig)
2020 struct client_obd *cli = &exp->exp_obd->u.cli;
2024 loi = &lsm->lsm_oinfo[0];
2026 spin_lock(&cli->cl_loi_list_lock);
2028 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2029 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2031 osc_check_rpcs(cli);
2032 spin_unlock(&cli->cl_loi_list_lock);
2037 static int osc_teardown_async_page(struct obd_export *exp,
2038 struct lov_stripe_md *lsm,
2039 struct lov_oinfo *loi, void *cookie)
2041 struct client_obd *cli = &exp->exp_obd->u.cli;
2042 struct loi_oap_pages *lop;
2043 struct osc_async_page *oap;
2047 oap = oap_from_cookie(cookie);
2049 RETURN(PTR_ERR(oap));
2052 loi = &lsm->lsm_oinfo[0];
2054 if (oap->oap_cmd == OBD_BRW_WRITE) {
2055 lop = &loi->loi_write_lop;
2057 lop = &loi->loi_read_lop;
2060 spin_lock(&cli->cl_loi_list_lock);
2062 if (!list_empty(&oap->oap_rpc_item))
2063 GOTO(out, rc = -EBUSY);
2065 osc_exit_cache(cli, oap, 0);
2066 osc_wake_cache_waiters(cli);
2068 if (!list_empty(&oap->oap_urgent_item)) {
2069 list_del_init(&oap->oap_urgent_item);
2070 oap->oap_async_flags &= ~ASYNC_URGENT;
2072 if (!list_empty(&oap->oap_pending_item)) {
2073 list_del_init(&oap->oap_pending_item);
2074 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2076 loi_list_maint(cli, loi);
2078 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2080 spin_unlock(&cli->cl_loi_list_lock);
2082 OBD_FREE(oap, sizeof(*oap));
2087 /* Note: caller will lock/unlock, and set uptodate on the pages */
2088 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2089 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2090 struct lov_stripe_md *lsm, obd_count page_count,
2091 struct brw_page *pga)
2093 struct ptlrpc_request *request = NULL;
2094 struct ost_body *body;
2095 struct niobuf_remote *nioptr;
2096 struct obd_ioobj *iooptr;
2097 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2101 /* XXX does not handle 'new' brw protocol */
2103 size[1] = sizeof(struct obd_ioobj);
2104 size[2] = page_count * sizeof(*nioptr);
2106 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2107 OST_SAN_READ, 3, size, NULL);
2111 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2112 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2113 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2114 sizeof(*nioptr) * page_count);
2116 memcpy(&body->oa, oa, sizeof(body->oa));
2118 obdo_to_ioobj(oa, iooptr);
2119 iooptr->ioo_bufcnt = page_count;
2121 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2122 LASSERT(PageLocked(pga[mapped].pg));
2123 LASSERT(mapped == 0 ||
2124 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2126 nioptr->offset = pga[mapped].disk_offset;
2127 nioptr->len = pga[mapped].count;
2128 nioptr->flags = pga[mapped].flag;
2131 size[1] = page_count * sizeof(*nioptr);
2132 request->rq_replen = lustre_msg_size(2, size);
2134 rc = ptlrpc_queue_wait(request);
2138 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2139 lustre_swab_ost_body);
2141 CERROR("Can't unpack body\n");
2142 GOTO(out_req, rc = -EPROTO);
2145 memcpy(oa, &body->oa, sizeof(*oa));
2147 swab = lustre_msg_swabbed(request->rq_repmsg);
2148 LASSERT_REPSWAB(request, 1);
2149 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2151 /* nioptr missing or short */
2152 GOTO(out_req, rc = -EPROTO);
2156 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2157 struct page *page = pga[mapped].pg;
2158 struct buffer_head *bh;
2162 lustre_swab_niobuf_remote (nioptr);
2164 /* got san device associated */
2165 LASSERT(exp->exp_obd != NULL);
2166 dev = exp->exp_obd->u.cli.cl_sandev;
2169 if (!nioptr->offset) {
2170 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2171 page->mapping->host->i_ino,
2173 memset(page_address(page), 0, PAGE_SIZE);
2177 if (!page->buffers) {
2178 create_empty_buffers(page, dev, PAGE_SIZE);
2181 clear_bit(BH_New, &bh->b_state);
2182 set_bit(BH_Mapped, &bh->b_state);
2183 bh->b_blocknr = (unsigned long)nioptr->offset;
2185 clear_bit(BH_Uptodate, &bh->b_state);
2187 ll_rw_block(READ, 1, &bh);
2191 /* if buffer already existed, it must be the
2192 * one we mapped before, check it */
2193 LASSERT(!test_bit(BH_New, &bh->b_state));
2194 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2195 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2197 /* wait it's io completion */
2198 if (test_bit(BH_Lock, &bh->b_state))
2201 if (!test_bit(BH_Uptodate, &bh->b_state))
2202 ll_rw_block(READ, 1, &bh);
2206 /* must do syncronous write here */
2208 if (!buffer_uptodate(bh)) {
2216 ptlrpc_req_finished(request);
2220 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2221 struct lov_stripe_md *lsm, obd_count page_count,
2222 struct brw_page *pga)
2224 struct ptlrpc_request *request = NULL;
2225 struct ost_body *body;
2226 struct niobuf_remote *nioptr;
2227 struct obd_ioobj *iooptr;
2228 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2232 size[1] = sizeof(struct obd_ioobj);
2233 size[2] = page_count * sizeof(*nioptr);
2235 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2236 OST_SAN_WRITE, 3, size, NULL);
2240 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2241 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2242 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2243 sizeof (*nioptr) * page_count);
2245 memcpy(&body->oa, oa, sizeof(body->oa));
2247 obdo_to_ioobj(oa, iooptr);
2248 iooptr->ioo_bufcnt = page_count;
2251 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2252 LASSERT(PageLocked(pga[mapped].pg));
2253 LASSERT(mapped == 0 ||
2254 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2256 nioptr->offset = pga[mapped].disk_offset;
2257 nioptr->len = pga[mapped].count;
2258 nioptr->flags = pga[mapped].flag;
2261 size[1] = page_count * sizeof(*nioptr);
2262 request->rq_replen = lustre_msg_size(2, size);
2264 rc = ptlrpc_queue_wait(request);
2268 swab = lustre_msg_swabbed (request->rq_repmsg);
2269 LASSERT_REPSWAB (request, 1);
2270 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2272 CERROR("absent/short niobuf array\n");
2273 GOTO(out_req, rc = -EPROTO);
2277 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2278 struct page *page = pga[mapped].pg;
2279 struct buffer_head *bh;
2283 lustre_swab_niobuf_remote (nioptr);
2285 /* got san device associated */
2286 LASSERT(exp->exp_obd != NULL);
2287 dev = exp->exp_obd->u.cli.cl_sandev;
2289 if (!page->buffers) {
2290 create_empty_buffers(page, dev, PAGE_SIZE);
2293 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2294 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2295 LASSERT(page->buffers->b_blocknr ==
2296 (unsigned long)nioptr->offset);
2302 /* if buffer locked, wait it's io completion */
2303 if (test_bit(BH_Lock, &bh->b_state))
2306 clear_bit(BH_New, &bh->b_state);
2307 set_bit(BH_Mapped, &bh->b_state);
2309 /* override the block nr */
2310 bh->b_blocknr = (unsigned long)nioptr->offset;
2312 /* we are about to write it, so set it
2314 * page lock should garentee no race condition here */
2315 set_bit(BH_Uptodate, &bh->b_state);
2316 set_bit(BH_Dirty, &bh->b_state);
2318 ll_rw_block(WRITE, 1, &bh);
2320 /* must do syncronous write here */
2322 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2330 ptlrpc_req_finished(request);
2334 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2335 struct lov_stripe_md *lsm, obd_count page_count,
2336 struct brw_page *pga, struct obd_trans_info *oti)
2340 while (page_count) {
2341 obd_count pages_per_brw;
2344 if (page_count > PTLRPC_MAX_BRW_PAGES)
2345 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2347 pages_per_brw = page_count;
2349 if (cmd & OBD_BRW_WRITE)
2350 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2352 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2357 page_count -= pages_per_brw;
2358 pga += pages_per_brw;
2365 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2367 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2370 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2374 l_lock(&lock->l_resource->lr_namespace->ns_lock);
2376 if (lock->l_ast_data && lock->l_ast_data != data) {
2377 struct inode *new_inode = data;
2378 struct inode *old_inode = lock->l_ast_data;
2379 LASSERTF(old_inode->i_state & I_FREEING,
2380 "Found existing inode %p/%lu/%u state %lu in lock: "
2381 "setting data to %p/%lu/%u\n", old_inode,
2382 old_inode->i_ino, old_inode->i_generation,
2384 new_inode, new_inode->i_ino, new_inode->i_generation);
2387 lock->l_ast_data = data;
2388 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2389 LDLM_LOCK_PUT(lock);
2392 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2393 ldlm_iterator_t replace, void *data)
2395 struct ldlm_res_id res_id = { .name = {0} };
2396 struct obd_device *obd = class_exp2obd(exp);
2398 res_id.name[0] = lsm->lsm_object_id;
2399 res_id.name[2] = lsm->lsm_object_gr;
2400 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2404 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2405 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2406 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2407 void *data, __u32 lvb_len, void *lvb_swabber,
2408 struct lustre_handle *lockh)
2410 struct obd_device *obd = exp->exp_obd;
2411 struct ldlm_res_id res_id = { .name = {0} };
2413 struct ldlm_reply *rep;
2414 struct ptlrpc_request *req = NULL;
2418 res_id.name[0] = lsm->lsm_object_id;
2419 res_id.name[2] = lsm->lsm_object_gr;
2421 /* Filesystem lock extents are extended to page boundaries so that
2422 * dealing with the page cache is a little smoother. */
2423 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2424 policy->l_extent.end |= ~PAGE_MASK;
2426 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2429 /* Next, search for already existing extent locks that will cover us */
2430 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2433 osc_set_data_with_check(lockh, data);
2434 if (*flags & LDLM_FL_HAS_INTENT) {
2435 /* I would like to be able to ASSERT here that rss <=
2436 * kms, but I can't, for reasons which are explained in
2439 /* We already have a lock, and it's referenced */
2443 /* If we're trying to read, we also search for an existing PW lock. The
2444 * VFS and page cache already protect us locally, so lots of readers/
2445 * writers can share a single PW lock.
2447 * There are problems with conversion deadlocks, so instead of
2448 * converting a read lock to a write lock, we'll just enqueue a new
2451 * At some point we should cancel the read lock instead of making them
2452 * send us a blocking callback, but there are problems with canceling
2453 * locks out from other users right now, too. */
2455 if (mode == LCK_PR) {
2456 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2457 policy, LCK_PW, lockh);
2459 /* FIXME: This is not incredibly elegant, but it might
2460 * be more elegant than adding another parameter to
2461 * lock_match. I want a second opinion. */
2462 ldlm_lock_addref(lockh, LCK_PR);
2463 ldlm_lock_decref(lockh, LCK_PW);
2464 osc_set_data_with_check(lockh, data);
2468 if (mode == LCK_PW) {
2469 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2470 policy, LCK_PR, lockh);
2472 rc = ldlm_cli_convert(lockh, mode, flags);
2474 /* Update readers/writers accounting */
2475 ldlm_lock_addref(lockh, LCK_PW);
2476 ldlm_lock_decref(lockh, LCK_PR);
2477 osc_set_data_with_check(lockh, data);
2480 /* If the conversion failed, we need to drop refcount
2481 on matched lock before we get new one */
2482 /* XXX Won't it save us some efforts if we cancel PR
2483 lock here? We are going to take PW lock anyway and it
2484 will invalidate PR lock */
2485 ldlm_lock_decref(lockh, LCK_PR);
2486 if (rc != EDEADLOCK) {
2493 if (*flags & LDLM_FL_HAS_INTENT) {
2494 int size[2] = {0, sizeof(struct ldlm_request)};
2496 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2497 LDLM_ENQUEUE, 2, size, NULL);
2501 size[0] = sizeof(*rep);
2502 size[1] = sizeof(lvb);
2503 req->rq_replen = lustre_msg_size(2, size);
2505 rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2506 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2507 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2509 if (rc == ELDLM_LOCK_ABORTED) {
2510 /* swabbed by ldlm_cli_enqueue() */
2511 LASSERT_REPSWABBED(req, 0);
2512 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2513 LASSERT(rep != NULL);
2514 if (rep->lock_policy_res1)
2515 rc = rep->lock_policy_res1;
2517 ptlrpc_req_finished(req);
2520 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2521 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2522 lvb.lvb_size, lvb.lvb_blocks);
2523 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2524 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2530 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2531 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2532 int *flags, void *data, struct lustre_handle *lockh)
2534 struct ldlm_res_id res_id = { .name = {0} };
2535 struct obd_device *obd = exp->exp_obd;
2539 res_id.name[0] = lsm->lsm_object_id;
2540 res_id.name[2] = lsm->lsm_object_gr;
2542 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2544 /* Filesystem lock extents are extended to page boundaries so that
2545 * dealing with the page cache is a little smoother */
2546 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2547 policy->l_extent.end |= ~PAGE_MASK;
2549 /* Next, search for already existing extent locks that will cover us */
2550 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2551 policy, mode, lockh);
2553 // if (!(*flags & LDLM_FL_TEST_LOCK))
2554 osc_set_data_with_check(lockh, data);
2557 /* If we're trying to read, we also search for an existing PW lock. The
2558 * VFS and page cache already protect us locally, so lots of readers/
2559 * writers can share a single PW lock. */
2560 if (mode == LCK_PR) {
2561 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2562 policy, LCK_PW, lockh);
2563 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2564 /* FIXME: This is not incredibly elegant, but it might
2565 * be more elegant than adding another parameter to
2566 * lock_match. I want a second opinion. */
2567 osc_set_data_with_check(lockh, data);
2568 ldlm_lock_addref(lockh, LCK_PR);
2569 ldlm_lock_decref(lockh, LCK_PW);
2575 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2576 __u32 mode, struct lustre_handle *lockh)
2580 if (mode == LCK_GROUP)
2581 ldlm_lock_decref_and_cancel(lockh, mode);
2583 ldlm_lock_decref(lockh, mode);
2588 static int osc_cancel_unused(struct obd_export *exp,
2589 struct lov_stripe_md *lsm, int flags, void *opaque)
2591 struct obd_device *obd = class_exp2obd(exp);
2592 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2595 res_id.name[0] = lsm->lsm_object_id;
2596 res_id.name[2] = lsm->lsm_object_gr;
2600 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2603 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2604 unsigned long max_age)
2606 struct obd_statfs *msfs;
2607 struct ptlrpc_request *request;
2608 int rc, size = sizeof(*osfs);
2611 /* We could possibly pass max_age in the request (as an absolute
2612 * timestamp or a "seconds.usec ago") so the target can avoid doing
2613 * extra calls into the filesystem if that isn't necessary (e.g.
2614 * during mount that would help a bit). Having relative timestamps
2615 * is not so great if request processing is slow, while absolute
2616 * timestamps are not ideal because they need time synchronization. */
2617 request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2618 OST_STATFS, 0, NULL, NULL);
2622 request->rq_replen = lustre_msg_size(1, &size);
2623 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2625 rc = ptlrpc_queue_wait(request);
2629 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2630 lustre_swab_obd_statfs);
2632 CERROR("Can't unpack obd_statfs\n");
2633 GOTO(out, rc = -EPROTO);
2636 memcpy(osfs, msfs, sizeof(*osfs));
2640 ptlrpc_req_finished(request);
2644 /* Retrieve object striping information.
2646 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2647 * the maximum number of OST indices which will fit in the user buffer.
2648 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2650 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2652 struct lov_user_md lum, *lumk;
2659 rc = copy_from_user(&lum, lump, sizeof(lum));
2663 if (lum.lmm_magic != LOV_USER_MAGIC)
2666 if (lum.lmm_stripe_count > 0) {
2667 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2668 OBD_ALLOC(lumk, lum_size);
2672 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2673 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2675 lum_size = sizeof(lum);
2679 lumk->lmm_object_id = lsm->lsm_object_id;
2680 lumk->lmm_object_gr = lsm->lsm_object_gr;
2681 lumk->lmm_stripe_count = 1;
2683 if (copy_to_user(lump, lumk, lum_size))
2687 OBD_FREE(lumk, lum_size);
2692 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2693 void *karg, void *uarg)
2695 struct obd_device *obd = exp->exp_obd;
2696 struct obd_ioctl_data *data = karg;
2700 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2703 if (!try_module_get(THIS_MODULE)) {
2704 CERROR("Can't get module. Is it alive?");
2709 case OBD_IOC_LOV_GET_CONFIG: {
2711 struct lov_desc *desc;
2712 struct obd_uuid uuid;
2716 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2717 GOTO(out, err = -EINVAL);
2719 data = (struct obd_ioctl_data *)buf;
2721 if (sizeof(*desc) > data->ioc_inllen1) {
2723 GOTO(out, err = -EINVAL);
2726 if (data->ioc_inllen2 < sizeof(uuid)) {
2728 GOTO(out, err = -EINVAL);
2731 if (data->ioc_inllen3 < sizeof(__u32)) {
2733 GOTO(out, err = -EINVAL);
2736 desc = (struct lov_desc *)data->ioc_inlbuf1;
2737 desc->ld_tgt_count = 1;
2738 desc->ld_active_tgt_count = 1;
2739 desc->ld_default_stripe_count = 1;
2740 desc->ld_default_stripe_size = 0;
2741 desc->ld_default_stripe_offset = 0;
2742 desc->ld_pattern = 0;
2743 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2744 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2745 *((__u32 *)data->ioc_inlbuf3) = 1;
2747 err = copy_to_user((void *)uarg, buf, len);
2750 obd_ioctl_freedata(buf, len);
2753 case LL_IOC_LOV_SETSTRIPE:
2754 err = obd_alloc_memmd(exp, karg);
2758 case LL_IOC_LOV_GETSTRIPE:
2759 err = osc_getstripe(karg, uarg);
2761 case OBD_IOC_CLIENT_RECOVER:
2762 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2767 case IOC_OSC_SET_ACTIVE:
2768 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2772 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2773 GOTO(out, err = -ENOTTY);
2776 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2779 module_put(THIS_MODULE);
2784 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2785 void *key, __u32 *vallen, void *val)
2788 if (!vallen || !val)
2791 if (keylen > strlen("lock_to_stripe") &&
2792 strcmp(key, "lock_to_stripe") == 0) {
2793 __u32 *stripe = val;
2794 *vallen = sizeof(*stripe);
2797 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2798 struct ptlrpc_request *req;
2800 char *bufs[1] = {key};
2802 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2803 OST_GET_INFO, 1, &keylen, bufs);
2807 req->rq_replen = lustre_msg_size(1, vallen);
2808 rc = ptlrpc_queue_wait(req);
2812 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2813 lustre_swab_ost_last_id);
2814 if (reply == NULL) {
2815 CERROR("Can't unpack OST last ID\n");
2816 GOTO(out, rc = -EPROTO);
2818 *((obd_id *)val) = *reply;
2820 ptlrpc_req_finished(req);
2826 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2827 void *key, obd_count vallen, void *val)
2829 struct obd_device *obd = exp->exp_obd;
2830 struct obd_import *imp = class_exp2cliimp(exp);
2831 struct llog_ctxt *ctxt;
2835 if (keylen == strlen("next_id") &&
2836 memcmp(key, "next_id", strlen("next_id")) == 0) {
2837 if (vallen != sizeof(obd_id))
2839 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
2840 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
2841 exp->exp_obd->obd_name,
2842 obd->u.cli.cl_oscc.oscc_next_id);
2847 if (keylen == strlen("growth_count") &&
2848 memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2849 if (vallen != sizeof(int))
2851 obd->u.cli.cl_oscc.oscc_max_grow_count = *((int*)val);
2855 if (keylen == strlen("unlinked") &&
2856 memcmp(key, "unlinked", keylen) == 0) {
2857 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2858 spin_lock(&oscc->oscc_lock);
2859 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2860 spin_unlock(&oscc->oscc_lock);
2863 if (keylen == strlen("unrecovery") &&
2864 memcmp(key, "unrecovery", keylen) == 0) {
2865 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2866 spin_lock(&oscc->oscc_lock);
2867 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
2868 spin_unlock(&oscc->oscc_lock);
2871 if (keylen == strlen("initial_recov") &&
2872 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2873 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2874 if (vallen != sizeof(int))
2876 imp->imp_initial_recov = *(int *)val;
2877 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2878 exp->exp_obd->obd_name,
2879 imp->imp_initial_recov);
2883 if (keylen < strlen("mds_conn") ||
2884 memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2887 ctxt = llog_get_context(&exp->exp_obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2890 rc = llog_initiator_connect(ctxt);
2892 CERROR("cannot establish the connect for ctxt %p: %d\n",
2896 imp->imp_server_timeout = 1;
2897 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2898 imp->imp_pingable = 1;
2904 static struct llog_operations osc_size_repl_logops = {
2905 lop_cancel: llog_obd_repl_cancel
2908 static struct llog_operations osc_unlink_orig_logops;
2909 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2910 struct obd_device *tgt, int count,
2911 struct llog_catid *catid)
2916 osc_unlink_orig_logops = llog_lvfs_ops;
2917 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2918 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
2919 osc_unlink_orig_logops.lop_add = llog_catalog_add;
2920 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2922 rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
2923 &catid->lci_logid, &osc_unlink_orig_logops);
2927 rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2928 &osc_size_repl_logops);
2932 static int osc_llog_finish(struct obd_device *obd,
2933 struct obd_llogs *llogs, int count)
2938 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
2942 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
2947 static int osc_connect(struct lustre_handle *exph,
2948 struct obd_device *obd, struct obd_uuid *cluuid,
2949 unsigned long connect_flags)
2953 rc = client_connect_import(exph, obd, cluuid, connect_flags);
2957 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
2959 struct obd_device *obd = class_exp2obd(exp);
2960 struct llog_ctxt *ctxt;
2964 ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
2965 if (obd->u.cli.cl_conn_count == 1)
2966 /* flush any remaining cancel messages out to the target */
2967 llog_sync(ctxt, exp);
2969 rc = client_disconnect_export(exp, flags);
2973 static int osc_import_event(struct obd_device *obd,
2974 struct obd_import *imp,
2975 enum obd_import_event event)
2977 struct client_obd *cli;
2980 LASSERT(imp->imp_obd == obd);
2983 case IMP_EVENT_DISCON: {
2984 /* Only do this on the MDS OSC's */
2985 if (imp->imp_server_timeout) {
2986 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2988 spin_lock(&oscc->oscc_lock);
2989 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
2990 spin_unlock(&oscc->oscc_lock);
2994 case IMP_EVENT_INACTIVE: {
2995 if (obd->obd_observer)
2996 rc = obd_notify(obd->obd_observer, obd, 0, 0);
2999 case IMP_EVENT_INVALIDATE: {
3000 struct ldlm_namespace *ns = obd->obd_namespace;
3004 spin_lock(&cli->cl_loi_list_lock);
3005 cli->cl_avail_grant = 0;
3006 cli->cl_lost_grant = 0;
3007 /* all pages go to failing rpcs due to the invalid import */
3008 osc_check_rpcs(cli);
3009 spin_unlock(&cli->cl_loi_list_lock);
3011 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3015 case IMP_EVENT_ACTIVE: {
3016 /* Only do this on the MDS OSC's */
3017 if (imp->imp_server_timeout) {
3018 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3020 spin_lock(&oscc->oscc_lock);
3021 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3022 spin_unlock(&oscc->oscc_lock);
3025 if (obd->obd_observer)
3026 rc = obd_notify(obd->obd_observer, obd, 1, 0);
3030 CERROR("Unknown import event %d\n", event);
3036 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3038 struct lprocfs_static_vars lvars;
3042 lprocfs_init_vars(osc,&lvars);
3043 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3047 rc = lproc_osc_attach_seqstat(dev);
3049 lprocfs_obd_detach(dev);
3053 ptlrpc_lprocfs_register_obd(dev);
3057 static int osc_detach(struct obd_device *dev)
3059 ptlrpc_lprocfs_unregister_obd(dev);
3060 return lprocfs_obd_detach(dev);
3063 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3067 rc = ptlrpcd_addref();
3071 rc = client_obd_setup(obd, len, buf);
3080 static int osc_cleanup(struct obd_device *obd, int flags)
3084 rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3085 LDLM_FL_CONFIG_CHANGE, NULL);
3089 rc = client_obd_cleanup(obd, flags);
3094 struct obd_ops osc_obd_ops = {
3095 .o_owner = THIS_MODULE,
3096 .o_attach = osc_attach,
3097 .o_detach = osc_detach,
3098 .o_setup = osc_setup,
3099 .o_cleanup = osc_cleanup,
3100 .o_add_conn = client_import_add_conn,
3101 .o_del_conn = client_import_del_conn,
3102 .o_connect = osc_connect,
3103 .o_disconnect = osc_disconnect,
3104 .o_statfs = osc_statfs,
3105 .o_packmd = osc_packmd,
3106 .o_unpackmd = osc_unpackmd,
3107 .o_create = osc_create,
3108 .o_destroy = osc_destroy,
3109 .o_getattr = osc_getattr,
3110 .o_getattr_async = osc_getattr_async,
3111 .o_setattr = osc_setattr,
3113 .o_brw_async = osc_brw_async,
3114 .o_prep_async_page = osc_prep_async_page,
3115 .o_queue_async_io = osc_queue_async_io,
3116 .o_set_async_flags = osc_set_async_flags,
3117 .o_queue_group_io = osc_queue_group_io,
3118 .o_trigger_group_io = osc_trigger_group_io,
3119 .o_teardown_async_page = osc_teardown_async_page,
3120 .o_punch = osc_punch,
3122 .o_enqueue = osc_enqueue,
3123 .o_match = osc_match,
3124 .o_change_cbdata = osc_change_cbdata,
3125 .o_cancel = osc_cancel,
3126 .o_cancel_unused = osc_cancel_unused,
3127 .o_iocontrol = osc_iocontrol,
3128 .o_get_info = osc_get_info,
3129 .o_set_info = osc_set_info,
3130 .o_import_event = osc_import_event,
3131 .o_llog_init = osc_llog_init,
3132 .o_llog_finish = osc_llog_finish,
3135 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3136 struct obd_ops sanosc_obd_ops = {
3137 .o_owner = THIS_MODULE,
3138 .o_attach = osc_attach,
3139 .o_detach = osc_detach,
3140 .o_cleanup = client_obd_cleanup,
3141 .o_add_conn = client_import_add_conn,
3142 .o_del_conn = client_import_del_conn,
3143 .o_connect = osc_connect,
3144 .o_disconnect = client_disconnect_export,
3145 .o_statfs = osc_statfs,
3146 .o_packmd = osc_packmd,
3147 .o_unpackmd = osc_unpackmd,
3148 .o_create = osc_real_create,
3149 .o_destroy = osc_destroy,
3150 .o_getattr = osc_getattr,
3151 .o_getattr_async = osc_getattr_async,
3152 .o_setattr = osc_setattr,
3153 .o_setup = client_sanobd_setup,
3154 .o_brw = sanosc_brw,
3155 .o_punch = osc_punch,
3157 .o_enqueue = osc_enqueue,
3158 .o_match = osc_match,
3159 .o_change_cbdata = osc_change_cbdata,
3160 .o_cancel = osc_cancel,
3161 .o_cancel_unused = osc_cancel_unused,
3162 .o_iocontrol = osc_iocontrol,
3163 .o_import_event = osc_import_event,
3164 .o_llog_init = osc_llog_init,
3165 .o_llog_finish = osc_llog_finish,
3169 int __init osc_init(void)
3171 struct lprocfs_static_vars lvars;
3172 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3173 struct lprocfs_static_vars sanlvars;
3178 lprocfs_init_vars(osc, &lvars);
3179 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3180 lprocfs_init_vars(osc, &sanlvars);
3183 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3188 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3189 rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3190 LUSTRE_SANOSC_NAME);
3192 class_unregister_type(LUSTRE_OSC_NAME);
3199 static void /*__exit*/ osc_exit(void)
3201 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3202 class_unregister_type(LUSTRE_SANOSC_NAME);
3204 class_unregister_type(LUSTRE_OSC_NAME);
3207 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3208 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3209 MODULE_LICENSE("GPL");
3211 module_init(osc_init);
3212 module_exit(osc_exit);