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 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
98 /* Unpack OSC object metadata from disk storage (LE byte order). */
99 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
100 struct lov_mds_md *lmm, int lmm_bytes)
106 if (lmm_bytes < sizeof (*lmm)) {
107 CERROR("lov_mds_md too small: %d, need %d\n",
108 lmm_bytes, (int)sizeof(*lmm));
111 /* XXX LOV_MAGIC etc check? */
113 if (lmm->lmm_object_id == 0) {
114 CERROR("lov_mds_md: zero lmm_object_id\n");
119 lsm_size = lov_stripe_md_size(1);
123 if (*lsmp != NULL && lmm == NULL) {
124 OBD_FREE(*lsmp, lsm_size);
130 OBD_ALLOC(*lsmp, lsm_size);
133 loi_init((*lsmp)->lsm_oinfo);
137 /* XXX zero *lsmp? */
138 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
139 LASSERT((*lsmp)->lsm_object_id);
142 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
147 static int osc_getattr_interpret(struct ptlrpc_request *req,
148 struct osc_getattr_async_args *aa, int rc)
150 struct ost_body *body;
156 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
158 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
159 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
161 /* This should really be sent by the OST */
162 aa->aa_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
163 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
165 CERROR("can't unpack ost_body\n");
167 aa->aa_oa->o_valid = 0;
173 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
174 struct lov_stripe_md *md,
175 struct ptlrpc_request_set *set)
177 struct ptlrpc_request *request;
178 struct ost_body *body;
179 int size = sizeof(*body);
180 struct osc_getattr_async_args *aa;
183 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
188 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
189 memcpy(&body->oa, oa, sizeof(*oa));
191 request->rq_replen = lustre_msg_size(1, &size);
192 request->rq_interpret_reply = osc_getattr_interpret;
194 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
195 aa = (struct osc_getattr_async_args *)&request->rq_async_args;
198 ptlrpc_set_add_req (set, request);
202 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
203 struct lov_stripe_md *md)
205 struct ptlrpc_request *request;
206 struct ost_body *body;
207 int rc, size = sizeof(*body);
210 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
215 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
216 memcpy(&body->oa, oa, sizeof(*oa));
218 request->rq_replen = lustre_msg_size(1, &size);
220 rc = ptlrpc_queue_wait(request);
222 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
226 body = lustre_swab_repbuf(request, 0, sizeof (*body),
227 lustre_swab_ost_body);
229 CERROR ("can't unpack ost_body\n");
230 GOTO (out, rc = -EPROTO);
233 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
234 memcpy(oa, &body->oa, sizeof(*oa));
236 /* This should really be sent by the OST */
237 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
238 oa->o_valid |= OBD_MD_FLBLKSZ;
242 ptlrpc_req_finished(request);
246 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
247 struct lov_stripe_md *md, struct obd_trans_info *oti)
249 struct ptlrpc_request *request;
250 struct ost_body *body;
251 int rc, size = sizeof(*body);
254 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SETATTR, 1, &size,
259 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
260 memcpy(&body->oa, oa, sizeof(*oa));
262 request->rq_replen = lustre_msg_size(1, &size);
264 rc = ptlrpc_queue_wait(request);
268 body = lustre_swab_repbuf(request, 0, sizeof(*body),
269 lustre_swab_ost_body);
271 GOTO(out, rc = -EPROTO);
273 memcpy(oa, &body->oa, sizeof(*oa));
277 ptlrpc_req_finished(request);
281 int osc_real_create(struct obd_export *exp, struct obdo *oa,
282 struct lov_stripe_md **ea, struct obd_trans_info *oti)
284 struct ptlrpc_request *request;
285 struct ost_body *body;
286 struct lov_stripe_md *lsm;
287 int rc, size = sizeof(*body);
295 rc = obd_alloc_memmd(exp, &lsm);
300 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_CREATE, 1, &size,
303 GOTO(out, rc = -ENOMEM);
305 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
306 memcpy(&body->oa, oa, sizeof(body->oa));
308 request->rq_replen = lustre_msg_size(1, &size);
309 if (oa->o_valid & OBD_MD_FLINLINE) {
310 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
311 oa->o_flags == OBD_FL_DELORPHAN);
312 DEBUG_REQ(D_HA, request,
313 "delorphan from OST integration");
316 rc = ptlrpc_queue_wait(request);
320 body = lustre_swab_repbuf(request, 0, sizeof(*body),
321 lustre_swab_ost_body);
323 CERROR ("can't unpack ost_body\n");
324 GOTO (out_req, rc = -EPROTO);
327 memcpy(oa, &body->oa, sizeof(*oa));
329 /* This should really be sent by the OST */
330 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
331 oa->o_valid |= OBD_MD_FLBLKSZ;
333 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
334 * have valid lsm_oinfo data structs, so don't go touching that.
335 * This needs to be fixed in a big way.
337 lsm->lsm_object_id = oa->o_id;
341 oti->oti_transno = request->rq_repmsg->transno;
343 if (oa->o_valid & OBD_MD_FLCOOKIE) {
344 if (!oti->oti_logcookies)
345 oti_alloc_cookies(oti, 1);
346 memcpy(oti->oti_logcookies, obdo_logcookie(oa),
347 sizeof(oti->oti_onecookie));
351 CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
354 ptlrpc_req_finished(request);
357 obd_free_memmd(exp, &lsm);
361 static int osc_punch(struct obd_export *exp, struct obdo *oa,
362 struct lov_stripe_md *md, obd_size start,
363 obd_size end, struct obd_trans_info *oti)
365 struct ptlrpc_request *request;
366 struct ost_body *body;
367 int rc, size = sizeof(*body);
375 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_PUNCH, 1, &size,
380 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
381 memcpy(&body->oa, oa, sizeof(*oa));
383 /* overload the size and blocks fields in the oa with start/end */
384 body->oa.o_size = start;
385 body->oa.o_blocks = end;
386 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
388 request->rq_replen = lustre_msg_size(1, &size);
390 rc = ptlrpc_queue_wait(request);
394 body = lustre_swab_repbuf (request, 0, sizeof (*body),
395 lustre_swab_ost_body);
397 CERROR ("can't unpack ost_body\n");
398 GOTO (out, rc = -EPROTO);
401 memcpy(oa, &body->oa, sizeof(*oa));
405 ptlrpc_req_finished(request);
409 static int osc_sync(struct obd_export *exp, struct obdo *oa,
410 struct lov_stripe_md *md, obd_size start, obd_size end)
412 struct ptlrpc_request *request;
413 struct ost_body *body;
414 int rc, size = sizeof(*body);
422 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SYNC, 1, &size,
427 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
428 memcpy(&body->oa, oa, sizeof(*oa));
430 /* overload the size and blocks fields in the oa with start/end */
431 body->oa.o_size = start;
432 body->oa.o_blocks = end;
433 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
435 request->rq_replen = lustre_msg_size(1, &size);
437 rc = ptlrpc_queue_wait(request);
441 body = lustre_swab_repbuf(request, 0, sizeof(*body),
442 lustre_swab_ost_body);
444 CERROR ("can't unpack ost_body\n");
445 GOTO (out, rc = -EPROTO);
448 memcpy(oa, &body->oa, sizeof(*oa));
452 ptlrpc_req_finished(request);
456 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
457 struct lov_stripe_md *ea, struct obd_trans_info *oti)
459 struct ptlrpc_request *request;
460 struct ost_body *body;
461 int rc, size = sizeof(*body);
469 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_DESTROY, 1,
474 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
476 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
477 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
478 sizeof(*oti->oti_logcookies));
479 oti->oti_logcookies++;
482 memcpy(&body->oa, oa, sizeof(*oa));
483 request->rq_replen = lustre_msg_size(1, &size);
485 rc = ptlrpc_queue_wait(request);
489 body = lustre_swab_repbuf(request, 0, sizeof(*body),
490 lustre_swab_ost_body);
492 CERROR ("Can't unpack body\n");
493 GOTO (out, rc = -EPROTO);
496 memcpy(oa, &body->oa, sizeof(*oa));
500 ptlrpc_req_finished(request);
504 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
507 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
509 LASSERT(!(oa->o_valid & bits));
512 spin_lock(&cli->cl_loi_list_lock);
513 oa->o_dirty = cli->cl_dirty;
514 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
515 oa->o_grant = cli->cl_avail_grant;
516 oa->o_dropped = cli->cl_lost_grant;
517 cli->cl_lost_grant = 0;
518 spin_unlock(&cli->cl_loi_list_lock);
519 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
520 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
523 /* caller must hold loi_list_lock */
524 static void osc_consume_write_grant(struct client_obd *cli,
525 struct osc_async_page *oap)
527 cli->cl_dirty += PAGE_SIZE;
528 cli->cl_avail_grant -= PAGE_SIZE;
529 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
530 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
531 LASSERT(cli->cl_avail_grant >= 0);
534 /* caller must hold loi_list_lock */
535 void osc_wake_cache_waiters(struct client_obd *cli)
537 struct list_head *l, *tmp;
538 struct osc_cache_waiter *ocw;
540 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
541 /* if we can't dirty more, we must wait until some is written */
542 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
543 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
544 cli->cl_dirty, cli->cl_dirty_max);
548 /* if still dirty cache but no grant wait for pending RPCs that
549 * may yet return us some grant before doing sync writes */
550 if (cli->cl_brw_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
551 CDEBUG(D_CACHE, "%d BRWs in flight, no grant\n",
552 cli->cl_brw_in_flight);
556 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
557 list_del_init(&ocw->ocw_entry);
558 if (cli->cl_avail_grant < PAGE_SIZE) {
559 /* no more RPCs in flight to return grant, do sync IO */
560 ocw->ocw_rc = -EDQUOT;
561 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
563 osc_consume_write_grant(cli, ocw->ocw_oap);
566 wake_up(&ocw->ocw_waitq);
572 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
574 spin_lock(&cli->cl_loi_list_lock);
575 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
576 cli->cl_avail_grant += body->oa.o_grant;
577 /* waiters are woken in brw_interpret_oap */
578 spin_unlock(&cli->cl_loi_list_lock);
581 /* We assume that the reason this OSC got a short read is because it read
582 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
583 * via the LOV, and it _knows_ it's reading inside the file, it's just that
584 * this stripe never got written at or beyond this stripe offset yet. */
585 static void handle_short_read(int nob_read, obd_count page_count,
586 struct brw_page *pga)
590 /* skip bytes read OK */
591 while (nob_read > 0) {
592 LASSERT (page_count > 0);
594 if (pga->count > nob_read) {
595 /* EOF inside this page */
596 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
597 memset(ptr + nob_read, 0, pga->count - nob_read);
604 nob_read -= pga->count;
609 /* zero remaining pages */
610 while (page_count-- > 0) {
611 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
612 memset(ptr, 0, pga->count);
618 static int check_write_rcs(struct ptlrpc_request *request,
619 int requested_nob, int niocount,
620 obd_count page_count, struct brw_page *pga)
624 /* return error if any niobuf was in error */
625 remote_rcs = lustre_swab_repbuf(request, 1,
626 sizeof(*remote_rcs) * niocount, NULL);
627 if (remote_rcs == NULL) {
628 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
631 if (lustre_msg_swabbed(request->rq_repmsg))
632 for (i = 0; i < niocount; i++)
633 __swab32s(&remote_rcs[i]);
635 for (i = 0; i < niocount; i++) {
636 if (remote_rcs[i] < 0)
637 return(remote_rcs[i]);
639 if (remote_rcs[i] != 0) {
640 CERROR("rc[%d] invalid (%d) req %p\n",
641 i, remote_rcs[i], request);
646 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
647 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
648 requested_nob, request->rq_bulk->bd_nob_transferred);
655 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
657 if (p1->flag != p2->flag) {
658 unsigned mask = ~OBD_BRW_FROM_GRANT;
660 /* warn if we try to combine flags that we don't know to be
662 if ((p1->flag & mask) != (p2->flag & mask))
663 CERROR("is it ok to have flags 0x%x and 0x%x in the "
664 "same brw?\n", p1->flag, p2->flag);
668 return (p1->off + p1->count == p2->off);
672 static obd_count cksum_pages(int nob, obd_count page_count,
673 struct brw_page *pga)
679 LASSERT (page_count > 0);
682 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
683 pga->count > nob ? nob : pga->count);
695 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
696 struct lov_stripe_md *lsm, obd_count page_count,
697 struct brw_page *pga, int *requested_nobp,
698 int *niocountp, struct ptlrpc_request **reqp)
700 struct ptlrpc_request *req;
701 struct ptlrpc_bulk_desc *desc;
702 struct client_obd *cli = &imp->imp_obd->u.cli;
703 struct ost_body *body;
704 struct obd_ioobj *ioobj;
705 struct niobuf_remote *niobuf;
714 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
716 for (niocount = i = 1; i < page_count; i++)
717 if (!can_merge_pages(&pga[i - 1], &pga[i]))
720 size[0] = sizeof(*body);
721 size[1] = sizeof(*ioobj);
722 size[2] = niocount * sizeof(*niobuf);
724 req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
728 if (opc == OST_WRITE)
729 desc = ptlrpc_prep_bulk_imp (req, page_count,
730 BULK_GET_SOURCE, OST_BULK_PORTAL);
732 desc = ptlrpc_prep_bulk_imp (req, page_count,
733 BULK_PUT_SINK, OST_BULK_PORTAL);
735 GOTO(out, rc = -ENOMEM);
736 /* NB request now owns desc and will free it when it gets freed */
738 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
739 ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
740 niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
742 memcpy(&body->oa, oa, sizeof(*oa));
744 obdo_to_ioobj(oa, ioobj);
745 ioobj->ioo_bufcnt = niocount;
747 LASSERT (page_count > 0);
748 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
749 struct brw_page *pg = &pga[i];
750 struct brw_page *pg_prev = pg - 1;
752 LASSERT(pg->count > 0);
753 LASSERT((pg->off & ~PAGE_MASK) + pg->count <= PAGE_SIZE);
754 LASSERTF(i == 0 || pg->off > pg_prev->off,
755 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
756 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
758 pg->pg, pg->pg->private, pg->pg->index, pg->off,
759 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
762 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~PAGE_MASK,
764 requested_nob += pg->count;
766 if (i > 0 && can_merge_pages(pg_prev, pg)) {
768 niobuf->len += pg->count;
770 niobuf->offset = pg->off;
771 niobuf->len = pg->count;
772 niobuf->flags = pg->flag;
776 LASSERT((void *)(niobuf - niocount) ==
777 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
778 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
779 spin_lock_irqsave(&req->rq_lock, flags);
780 req->rq_no_resend = 1;
781 spin_unlock_irqrestore(&req->rq_lock, flags);
783 /* size[0] still sizeof (*body) */
784 if (opc == OST_WRITE) {
786 body->oa.o_valid |= OBD_MD_FLCKSUM;
787 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
789 /* 1 RC per niobuf */
790 size[1] = sizeof(__u32) * niocount;
791 req->rq_replen = lustre_msg_size(2, size);
793 /* 1 RC for the whole I/O */
794 req->rq_replen = lustre_msg_size(1, size);
797 *niocountp = niocount;
798 *requested_nobp = requested_nob;
803 ptlrpc_req_finished (req);
807 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
808 int requested_nob, int niocount,
809 obd_count page_count, struct brw_page *pga,
812 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
813 struct ost_body *body;
819 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
821 CERROR ("Can't unpack body\n");
825 osc_update_grant(cli, body);
826 memcpy(oa, &body->oa, sizeof(*oa));
828 if (req->rq_reqmsg->opc == OST_WRITE) {
830 CERROR ("Unexpected +ve rc %d\n", rc);
833 LASSERT (req->rq_bulk->bd_nob == requested_nob);
835 RETURN(check_write_rcs(req, requested_nob, niocount,
839 if (rc > requested_nob) {
840 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
844 if (rc != req->rq_bulk->bd_nob_transferred) {
845 CERROR ("Unexpected rc %d (%d transferred)\n",
846 rc, req->rq_bulk->bd_nob_transferred);
850 if (rc < requested_nob)
851 handle_short_read(rc, page_count, pga);
854 if (oa->o_valid & OBD_MD_FLCKSUM) {
855 const struct ptlrpc_peer *peer =
856 &req->rq_import->imp_connection->c_peer;
857 static int cksum_counter;
858 obd_count server_cksum = oa->o_cksum;
859 obd_count cksum = cksum_pages(rc, page_count, pga);
860 char str[PTL_NALFMT_SIZE];
862 ptlrpc_peernid2str(peer, str);
865 if (server_cksum != cksum) {
866 CERROR("Bad checksum: server %x, client %x, server NID "
867 LPX64" (%s)\n", server_cksum, cksum,
868 peer->peer_nid, str);
871 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
872 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
873 cksum_counter, peer->peer_nid, str, cksum);
876 static int cksum_missed;
879 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
880 CERROR("Request checksum %u from "LPX64", no reply\n",
882 req->rq_import->imp_connection->c_peer.peer_nid);
888 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
889 struct lov_stripe_md *lsm,
890 obd_count page_count, struct brw_page *pga)
894 struct ptlrpc_request *request;
899 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
900 page_count, pga, &requested_nob, &niocount,
902 /* NB ^ sets rq_no_resend */
907 rc = ptlrpc_queue_wait(request);
909 if (rc == -ETIMEDOUT && request->rq_resend) {
910 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
911 ptlrpc_req_finished(request);
915 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
916 page_count, pga, rc);
918 ptlrpc_req_finished(request);
922 static int brw_interpret(struct ptlrpc_request *request,
923 struct osc_brw_async_args *aa, int rc)
925 struct obdo *oa = aa->aa_oa;
926 int requested_nob = aa->aa_requested_nob;
927 int niocount = aa->aa_nio_count;
928 obd_count page_count = aa->aa_page_count;
929 struct brw_page *pga = aa->aa_pga;
932 /* XXX bug 937 here */
933 if (rc == -ETIMEDOUT && request->rq_resend) {
934 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
935 LBUG(); /* re-send. later. */
939 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
940 page_count, pga, rc);
944 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
945 struct lov_stripe_md *lsm, obd_count page_count,
946 struct brw_page *pga, struct ptlrpc_request_set *set)
948 struct ptlrpc_request *request;
951 struct osc_brw_async_args *aa;
955 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
956 page_count, pga, &requested_nob, &nio_count,
958 /* NB ^ sets rq_no_resend */
961 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
962 aa = (struct osc_brw_async_args *)&request->rq_async_args;
964 aa->aa_requested_nob = requested_nob;
965 aa->aa_nio_count = nio_count;
966 aa->aa_page_count = page_count;
969 request->rq_interpret_reply = brw_interpret;
970 ptlrpc_set_add_req(set, request);
976 #define min_t(type,x,y) \
977 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
981 * ugh, we want disk allocation on the target to happen in offset order. we'll
982 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
983 * fine for our small page arrays and doesn't require allocation. its an
984 * insertion sort that swaps elements that are strides apart, shrinking the
985 * stride down until its '1' and the array is sorted.
987 static void sort_brw_pages(struct brw_page *array, int num)
994 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
999 for (i = stride ; i < num ; i++) {
1002 while (j >= stride && array[j - stride].off > tmp.off) {
1003 array[j] = array[j - stride];
1008 } while (stride > 1);
1011 /* make sure we the regions we're passing to elan don't violate its '4
1012 * fragments' constraint. portal headers are a fragment, all full
1013 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1014 * counts as a fragment. I think. see bug 934. */
1015 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1018 int saw_whole_frag = 0;
1021 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1022 if (pg->count == PAGE_SIZE) {
1023 if (!saw_whole_frag) {
1034 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1035 struct lov_stripe_md *md, obd_count page_count,
1036 struct brw_page *pga, struct obd_trans_info *oti)
1040 if (cmd == OBD_BRW_CHECK) {
1041 /* The caller just wants to know if there's a chance that this
1042 * I/O can succeed */
1043 struct obd_import *imp = class_exp2cliimp(exp);
1045 if (imp == NULL || imp->imp_invalid)
1050 while (page_count) {
1051 obd_count pages_per_brw;
1054 if (page_count > PTLRPC_MAX_BRW_PAGES)
1055 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1057 pages_per_brw = page_count;
1059 sort_brw_pages(pga, pages_per_brw);
1060 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1062 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1067 page_count -= pages_per_brw;
1068 pga += pages_per_brw;
1073 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1074 struct lov_stripe_md *md, obd_count page_count,
1075 struct brw_page *pga, struct ptlrpc_request_set *set,
1076 struct obd_trans_info *oti)
1080 if (cmd == OBD_BRW_CHECK) {
1081 /* The caller just wants to know if there's a chance that this
1082 * I/O can succeed */
1083 struct obd_import *imp = class_exp2cliimp(exp);
1085 if (imp == NULL || imp->imp_invalid)
1090 while (page_count) {
1091 obd_count pages_per_brw;
1094 if (page_count > PTLRPC_MAX_BRW_PAGES)
1095 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1097 pages_per_brw = page_count;
1099 sort_brw_pages(pga, pages_per_brw);
1100 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1102 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1107 page_count -= pages_per_brw;
1108 pga += pages_per_brw;
1113 static void osc_check_rpcs(struct client_obd *cli);
1114 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1116 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1117 static void lop_update_pending(struct client_obd *cli,
1118 struct loi_oap_pages *lop, int cmd, int delta);
1120 /* this is called when a sync waiter receives an interruption. Its job is to
1121 * get the caller woken as soon as possible. If its page hasn't been put in an
1122 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1123 * desiring interruption which will forcefully complete the rpc once the rpc
1125 static void osc_occ_interrupted(struct oig_callback_context *occ)
1127 struct osc_async_page *oap;
1128 struct loi_oap_pages *lop;
1129 struct lov_oinfo *loi;
1132 /* XXX member_of() */
1133 oap = list_entry(occ, struct osc_async_page, oap_occ);
1135 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1137 oap->oap_interrupted = 1;
1139 /* ok, it's been put in an rpc. */
1140 if (oap->oap_request != NULL) {
1141 ptlrpc_mark_interrupted(oap->oap_request);
1142 ptlrpcd_wake(oap->oap_request);
1146 /* we don't get interruption callbacks until osc_trigger_sync_io()
1147 * has been called and put the sync oaps in the pending/urgent lists.*/
1148 if (!list_empty(&oap->oap_pending_item)) {
1149 list_del_init(&oap->oap_pending_item);
1150 if (oap->oap_async_flags & ASYNC_URGENT)
1151 list_del_init(&oap->oap_urgent_item);
1154 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1155 &loi->loi_write_lop : &loi->loi_read_lop;
1156 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1157 loi_list_maint(oap->oap_cli, oap->oap_loi);
1159 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1160 oap->oap_oig = NULL;
1164 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1167 /* this must be called holding the loi list lock to give coverage to exit_cache,
1168 * async_flag maintenance, and oap_request */
1169 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1170 struct osc_async_page *oap, int sent, int rc)
1172 osc_exit_cache(cli, oap, sent);
1173 oap->oap_async_flags = 0;
1174 oap->oap_interrupted = 0;
1176 if (oap->oap_request != NULL) {
1177 ptlrpc_req_finished(oap->oap_request);
1178 oap->oap_request = NULL;
1181 if (rc == 0 && oa != NULL)
1182 oap->oap_loi->loi_blocks = oa->o_blocks;
1185 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1186 oap->oap_oig = NULL;
1191 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1195 static int brw_interpret_oap(struct ptlrpc_request *request,
1196 struct osc_brw_async_args *aa, int rc)
1198 struct osc_async_page *oap;
1199 struct client_obd *cli;
1200 struct list_head *pos, *n;
1204 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1205 aa->aa_nio_count, aa->aa_page_count,
1208 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1211 /* in failout recovery we ignore writeback failure and want
1212 * to just tell llite to unlock the page and continue */
1213 if (request->rq_reqmsg->opc == OST_WRITE &&
1214 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1215 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1217 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1221 spin_lock(&cli->cl_loi_list_lock);
1223 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1224 * is called so we know whether to go to sync BRWs or wait for more
1225 * RPCs to complete */
1226 cli->cl_brw_in_flight--;
1228 /* the caller may re-use the oap after the completion call so
1229 * we need to clean it up a little */
1230 list_for_each_safe(pos, n, &aa->aa_oaps) {
1231 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1233 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1234 //oap->oap_page, oap->oap_page->index, oap);
1236 list_del_init(&oap->oap_rpc_item);
1237 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1240 osc_wake_cache_waiters(cli);
1241 osc_check_rpcs(cli);
1243 spin_unlock(&cli->cl_loi_list_lock);
1245 obdo_free(aa->aa_oa);
1246 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1251 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1252 struct list_head *rpc_list,
1253 int page_count, int cmd)
1255 struct ptlrpc_request *req;
1256 struct brw_page *pga = NULL;
1257 int requested_nob, nio_count;
1258 struct osc_brw_async_args *aa;
1259 struct obdo *oa = NULL;
1260 struct obd_async_page_ops *ops = NULL;
1261 void *caller_data = NULL;
1262 struct list_head *pos;
1265 LASSERT(!list_empty(rpc_list));
1267 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1269 RETURN(ERR_PTR(-ENOMEM));
1273 GOTO(out, req = ERR_PTR(-ENOMEM));
1276 list_for_each(pos, rpc_list) {
1277 struct osc_async_page *oap;
1279 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1281 ops = oap->oap_caller_ops;
1282 caller_data = oap->oap_caller_data;
1284 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1285 pga[i].pg = oap->oap_page;
1286 pga[i].count = oap->oap_count;
1287 pga[i].flag = oap->oap_brw_flags;
1288 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1289 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1293 /* always get the data for the obdo for the rpc */
1294 LASSERT(ops != NULL);
1295 ops->ap_fill_obdo(caller_data, cmd, oa);
1297 sort_brw_pages(pga, page_count);
1298 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1299 pga, &requested_nob, &nio_count, &req);
1301 CERROR("prep_req failed: %d\n", rc);
1302 GOTO(out, req = ERR_PTR(rc));
1305 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1306 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1308 aa->aa_requested_nob = requested_nob;
1309 aa->aa_nio_count = nio_count;
1310 aa->aa_page_count = page_count;
1319 OBD_FREE(pga, sizeof(*pga) * page_count);
1324 static void lop_update_pending(struct client_obd *cli,
1325 struct loi_oap_pages *lop, int cmd, int delta)
1327 lop->lop_num_pending += delta;
1328 if (cmd == OBD_BRW_WRITE)
1329 cli->cl_pending_w_pages += delta;
1331 cli->cl_pending_r_pages += delta;
1334 /* the loi lock is held across this function but it's allowed to release
1335 * and reacquire it during its work */
1336 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1337 int cmd, struct loi_oap_pages *lop)
1339 struct ptlrpc_request *request;
1340 obd_count page_count = 0;
1341 struct list_head *tmp, *pos;
1342 struct osc_async_page *oap = NULL;
1343 struct osc_brw_async_args *aa;
1344 struct obd_async_page_ops *ops;
1345 LIST_HEAD(rpc_list);
1348 /* first we find the pages we're allowed to work with */
1349 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1350 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1351 ops = oap->oap_caller_ops;
1353 LASSERT(oap->oap_magic == OAP_MAGIC);
1355 /* in llite being 'ready' equates to the page being locked
1356 * until completion unlocks it. commit_write submits a page
1357 * as not ready because its unlock will happen unconditionally
1358 * as the call returns. if we race with commit_write giving
1359 * us that page we dont' want to create a hole in the page
1360 * stream, so we stop and leave the rpc to be fired by
1361 * another dirtier or kupdated interval (the not ready page
1362 * will still be on the dirty list). we could call in
1363 * at the end of ll_file_write to process the queue again. */
1364 if (!(oap->oap_async_flags & ASYNC_READY)) {
1365 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1367 CDEBUG(D_INODE, "oap %p page %p returned %d "
1368 "instead of ready\n", oap,
1372 /* llite is telling us that the page is still
1373 * in commit_write and that we should try
1374 * and put it in an rpc again later. we
1375 * break out of the loop so we don't create
1376 * a hole in the sequence of pages in the rpc
1381 /* the io isn't needed.. tell the checks
1382 * below to complete the rpc with EINTR */
1383 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1384 oap->oap_count = -EINTR;
1387 oap->oap_async_flags |= ASYNC_READY;
1390 LASSERTF(0, "oap %p page %p returned %d "
1391 "from make_ready\n", oap,
1399 /* take the page out of our book-keeping */
1400 list_del_init(&oap->oap_pending_item);
1401 lop_update_pending(cli, lop, cmd, -1);
1402 if (!list_empty(&oap->oap_urgent_item))
1403 list_del_init(&oap->oap_urgent_item);
1405 /* ask the caller for the size of the io as the rpc leaves. */
1406 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1408 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1409 if (oap->oap_count <= 0) {
1410 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1412 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1416 /* now put the page back in our accounting */
1417 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1418 if (++page_count >= cli->cl_max_pages_per_rpc)
1422 osc_wake_cache_waiters(cli);
1424 if (page_count == 0)
1427 loi_list_maint(cli, loi);
1428 spin_unlock(&cli->cl_loi_list_lock);
1430 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1431 if (IS_ERR(request)) {
1432 /* this should happen rarely and is pretty bad, it makes the
1433 * pending list not follow the dirty order */
1434 spin_lock(&cli->cl_loi_list_lock);
1435 list_for_each_safe(pos, tmp, &rpc_list) {
1436 oap = list_entry(pos, struct osc_async_page,
1438 list_del_init(&oap->oap_rpc_item);
1440 /* queued sync pages can be torn down while the pages
1441 * were between the pending list and the rpc */
1442 if (oap->oap_interrupted) {
1443 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1444 osc_ap_completion(cli, NULL, oap, 0,
1449 /* put the page back in the loi/lop lists */
1450 list_add_tail(&oap->oap_pending_item,
1452 lop_update_pending(cli, lop, cmd, 1);
1453 if (oap->oap_async_flags & ASYNC_URGENT)
1454 list_add(&oap->oap_urgent_item,
1457 loi_list_maint(cli, loi);
1458 RETURN(PTR_ERR(request));
1461 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1462 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1463 INIT_LIST_HEAD(&aa->aa_oaps);
1464 list_splice(&rpc_list, &aa->aa_oaps);
1465 INIT_LIST_HEAD(&rpc_list);
1468 if (cmd == OBD_BRW_READ) {
1469 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1470 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1472 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1473 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1474 cli->cl_brw_in_flight);
1478 spin_lock(&cli->cl_loi_list_lock);
1480 cli->cl_brw_in_flight++;
1481 /* queued sync pages can be torn down while the pages
1482 * were between the pending list and the rpc */
1483 list_for_each(pos, &aa->aa_oaps) {
1484 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1485 if (oap->oap_interrupted) {
1486 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1488 ptlrpc_mark_interrupted(request);
1493 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %d in flight\n", request,
1494 page_count, aa, cli->cl_brw_in_flight);
1496 oap->oap_request = ptlrpc_request_addref(request);
1497 request->rq_interpret_reply = brw_interpret_oap;
1498 ptlrpcd_add_req(request);
1502 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1508 if (lop->lop_num_pending == 0)
1511 /* if we have an invalid import we want to drain the queued pages
1512 * by forcing them through rpcs that immediately fail and complete
1513 * the pages. recovery relies on this to empty the queued pages
1514 * before canceling the locks and evicting down the llite pages */
1515 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1518 /* stream rpcs in queue order as long as as there is an urgent page
1519 * queued. this is our cheap solution for good batching in the case
1520 * where writepage marks some random page in the middle of the file as
1521 * urgent because of, say, memory pressure */
1522 if (!list_empty(&lop->lop_urgent))
1525 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1526 optimal = cli->cl_max_pages_per_rpc;
1527 if (cmd == OBD_BRW_WRITE) {
1528 /* trigger a write rpc stream as long as there are dirtiers
1529 * waiting for space. as they're waiting, they're not going to
1530 * create more pages to coallesce with what's waiting.. */
1531 if (!list_empty(&cli->cl_cache_waiters))
1534 /* *2 to avoid triggering rpcs that would want to include pages
1535 * that are being queued but which can't be made ready until
1536 * the queuer finishes with the page. this is a wart for
1537 * llite::commit_write() */
1540 if (lop->lop_num_pending >= optimal)
1546 static void on_list(struct list_head *item, struct list_head *list,
1549 if (list_empty(item) && should_be_on)
1550 list_add_tail(item, list);
1551 else if (!list_empty(item) && !should_be_on)
1552 list_del_init(item);
1555 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1556 * can find pages to build into rpcs quickly */
1557 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1559 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1560 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1561 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1563 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1564 loi->loi_write_lop.lop_num_pending);
1566 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1567 loi->loi_read_lop.lop_num_pending);
1570 #define LOI_DEBUG(LOI, STR, args...) \
1571 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1572 !list_empty(&(LOI)->loi_cli_item), \
1573 (LOI)->loi_write_lop.lop_num_pending, \
1574 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1575 (LOI)->loi_read_lop.lop_num_pending, \
1576 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1579 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1582 /* first return all objects which we already know to have
1583 * pages ready to be stuffed into rpcs */
1584 if (!list_empty(&cli->cl_loi_ready_list))
1585 RETURN(list_entry(cli->cl_loi_ready_list.next,
1586 struct lov_oinfo, loi_cli_item));
1588 /* then if we have cache waiters, return all objects with queued
1589 * writes. This is especially important when many small files
1590 * have filled up the cache and not been fired into rpcs because
1591 * they don't pass the nr_pending/object threshhold */
1592 if (!list_empty(&cli->cl_cache_waiters) &&
1593 !list_empty(&cli->cl_loi_write_list))
1594 RETURN(list_entry(cli->cl_loi_write_list.next,
1595 struct lov_oinfo, loi_write_item));
1597 /* then return all queued objects when we have an invalid import
1598 * so that they get flushed */
1599 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1600 if (!list_empty(&cli->cl_loi_write_list))
1601 RETURN(list_entry(cli->cl_loi_write_list.next,
1602 struct lov_oinfo, loi_write_item));
1603 if (!list_empty(&cli->cl_loi_read_list))
1604 RETURN(list_entry(cli->cl_loi_read_list.next,
1605 struct lov_oinfo, loi_read_item));
1610 /* called with the loi list lock held */
1611 static void osc_check_rpcs(struct client_obd *cli)
1613 struct lov_oinfo *loi;
1614 int rc = 0, race_counter = 0;
1617 while ((loi = osc_next_loi(cli)) != NULL) {
1618 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1620 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1623 /* attempt some read/write balancing by alternating between
1624 * reads and writes in an object. The makes_rpc checks here
1625 * would be redundant if we were getting read/write work items
1626 * instead of objects. we don't want send_oap_rpc to drain a
1627 * partial read pending queue when we're given this object to
1628 * do io on writes while there are cache waiters */
1629 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1630 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1631 &loi->loi_write_lop);
1639 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1640 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1641 &loi->loi_read_lop);
1650 /* attempt some inter-object balancing by issueing rpcs
1651 * for each object in turn */
1652 if (!list_empty(&loi->loi_cli_item))
1653 list_del_init(&loi->loi_cli_item);
1654 if (!list_empty(&loi->loi_write_item))
1655 list_del_init(&loi->loi_write_item);
1656 if (!list_empty(&loi->loi_read_item))
1657 list_del_init(&loi->loi_read_item);
1659 loi_list_maint(cli, loi);
1661 /* send_oap_rpc fails with 0 when make_ready tells it to
1662 * back off. llite's make_ready does this when it tries
1663 * to lock a page queued for write that is already locked.
1664 * we want to try sending rpcs from many objects, but we
1665 * don't want to spin failing with 0. */
1666 if (race_counter == 10)
1672 /* we're trying to queue a page in the osc so we're subject to the
1673 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1674 * If the osc's queued pages are already at that limit, then we want to sleep
1675 * until there is space in the osc's queue for us. We also may be waiting for
1676 * write credits from the OST if there are RPCs in flight that may return some
1677 * before we fall back to sync writes.
1679 * We need this know our allocation was granted in the presence of signals */
1680 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1684 spin_lock(&cli->cl_loi_list_lock);
1685 rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1686 spin_unlock(&cli->cl_loi_list_lock);
1690 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1691 * grant or cache space. */
1692 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1693 struct osc_async_page *oap)
1695 struct osc_cache_waiter ocw;
1696 struct l_wait_info lwi = { 0 };
1698 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1699 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1700 cli->cl_avail_grant);
1702 if (cli->cl_dirty_max < PAGE_SIZE)
1705 /* Hopefully normal case - cache space and write credits available */
1706 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1707 cli->cl_avail_grant >= PAGE_SIZE) {
1708 /* account for ourselves */
1709 osc_consume_write_grant(cli, oap);
1713 /* Make sure that there are write rpcs in flight to wait for. This
1714 * is a little silly as this object may not have any pending but
1715 * other objects sure might. */
1716 if (cli->cl_brw_in_flight) {
1717 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1718 init_waitqueue_head(&ocw.ocw_waitq);
1722 loi_list_maint(cli, loi);
1723 osc_check_rpcs(cli);
1724 spin_unlock(&cli->cl_loi_list_lock);
1726 CDEBUG(0, "sleeping for cache space\n");
1727 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1729 spin_lock(&cli->cl_loi_list_lock);
1730 if (!list_empty(&ocw.ocw_entry)) {
1731 list_del(&ocw.ocw_entry);
1740 /* the companion to enter_cache, called when an oap is no longer part of the
1741 * dirty accounting.. so writeback completes or truncate happens before writing
1742 * starts. must be called with the loi lock held. */
1743 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1748 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1753 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1754 cli->cl_dirty -= PAGE_SIZE;
1756 cli->cl_lost_grant += PAGE_SIZE;
1757 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1758 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1764 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1765 struct lov_oinfo *loi, struct page *page,
1766 obd_off offset, struct obd_async_page_ops *ops,
1767 void *data, void **res)
1769 struct osc_async_page *oap;
1772 OBD_ALLOC(oap, sizeof(*oap));
1776 oap->oap_magic = OAP_MAGIC;
1777 oap->oap_cli = &exp->exp_obd->u.cli;
1780 oap->oap_caller_ops = ops;
1781 oap->oap_caller_data = data;
1783 oap->oap_page = page;
1784 oap->oap_obj_off = offset;
1786 INIT_LIST_HEAD(&oap->oap_pending_item);
1787 INIT_LIST_HEAD(&oap->oap_urgent_item);
1788 INIT_LIST_HEAD(&oap->oap_rpc_item);
1790 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1792 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1797 struct osc_async_page *oap_from_cookie(void *cookie)
1799 struct osc_async_page *oap = cookie;
1800 if (oap->oap_magic != OAP_MAGIC)
1801 return ERR_PTR(-EINVAL);
1805 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1806 struct lov_oinfo *loi, void *cookie,
1807 int cmd, obd_off off, int count,
1808 obd_flag brw_flags, enum async_flags async_flags)
1810 struct client_obd *cli = &exp->exp_obd->u.cli;
1811 struct osc_async_page *oap;
1812 struct loi_oap_pages *lop;
1816 oap = oap_from_cookie(cookie);
1818 RETURN(PTR_ERR(oap));
1820 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1823 if (!list_empty(&oap->oap_pending_item) ||
1824 !list_empty(&oap->oap_urgent_item) ||
1825 !list_empty(&oap->oap_rpc_item))
1829 loi = &lsm->lsm_oinfo[0];
1831 spin_lock(&cli->cl_loi_list_lock);
1834 oap->oap_async_flags = async_flags;
1835 oap->oap_page_off = off;
1836 oap->oap_count = count;
1837 oap->oap_brw_flags = brw_flags;
1839 if (cmd == OBD_BRW_WRITE) {
1840 rc = osc_enter_cache(cli, loi, oap);
1842 spin_unlock(&cli->cl_loi_list_lock);
1845 lop = &loi->loi_write_lop;
1847 lop = &loi->loi_read_lop;
1850 if (oap->oap_async_flags & ASYNC_URGENT)
1851 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1852 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1853 lop_update_pending(cli, lop, cmd, 1);
1855 loi_list_maint(cli, loi);
1857 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1860 osc_check_rpcs(cli);
1861 spin_unlock(&cli->cl_loi_list_lock);
1866 /* aka (~was & now & flag), but this is more clear :) */
1867 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1869 static int osc_set_async_flags(struct obd_export *exp,
1870 struct lov_stripe_md *lsm,
1871 struct lov_oinfo *loi, void *cookie,
1872 obd_flag async_flags)
1874 struct client_obd *cli = &exp->exp_obd->u.cli;
1875 struct loi_oap_pages *lop;
1876 struct osc_async_page *oap;
1880 oap = oap_from_cookie(cookie);
1882 RETURN(PTR_ERR(oap));
1884 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1888 loi = &lsm->lsm_oinfo[0];
1890 if (oap->oap_cmd == OBD_BRW_WRITE) {
1891 lop = &loi->loi_write_lop;
1893 lop = &loi->loi_read_lop;
1896 spin_lock(&cli->cl_loi_list_lock);
1898 if (list_empty(&oap->oap_pending_item))
1899 GOTO(out, rc = -EINVAL);
1901 if ((oap->oap_async_flags & async_flags) == async_flags)
1904 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1905 oap->oap_async_flags |= ASYNC_READY;
1907 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1908 if (list_empty(&oap->oap_rpc_item)) {
1909 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1910 loi_list_maint(cli, loi);
1914 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1915 oap->oap_async_flags);
1917 osc_check_rpcs(cli);
1918 spin_unlock(&cli->cl_loi_list_lock);
1922 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1923 struct lov_oinfo *loi,
1924 struct obd_io_group *oig, void *cookie,
1925 int cmd, obd_off off, int count,
1927 obd_flag async_flags)
1929 struct client_obd *cli = &exp->exp_obd->u.cli;
1930 struct osc_async_page *oap;
1931 struct loi_oap_pages *lop;
1934 oap = oap_from_cookie(cookie);
1936 RETURN(PTR_ERR(oap));
1938 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1941 if (!list_empty(&oap->oap_pending_item) ||
1942 !list_empty(&oap->oap_urgent_item) ||
1943 !list_empty(&oap->oap_rpc_item))
1947 loi = &lsm->lsm_oinfo[0];
1949 spin_lock(&cli->cl_loi_list_lock);
1952 oap->oap_page_off = off;
1953 oap->oap_count = count;
1954 oap->oap_brw_flags = brw_flags;
1955 oap->oap_async_flags = async_flags;
1957 if (cmd == OBD_BRW_WRITE)
1958 lop = &loi->loi_write_lop;
1960 lop = &loi->loi_read_lop;
1962 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
1963 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
1965 oig_add_one(oig, &oap->oap_occ);
1968 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
1970 spin_unlock(&cli->cl_loi_list_lock);
1975 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1976 struct loi_oap_pages *lop, int cmd)
1978 struct list_head *pos, *tmp;
1979 struct osc_async_page *oap;
1981 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
1982 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1983 list_del(&oap->oap_pending_item);
1984 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1985 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1986 lop_update_pending(cli, lop, cmd, 1);
1988 loi_list_maint(cli, loi);
1991 static int osc_trigger_group_io(struct obd_export *exp,
1992 struct lov_stripe_md *lsm,
1993 struct lov_oinfo *loi,
1994 struct obd_io_group *oig)
1996 struct client_obd *cli = &exp->exp_obd->u.cli;
2000 loi = &lsm->lsm_oinfo[0];
2002 spin_lock(&cli->cl_loi_list_lock);
2004 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2005 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2007 osc_check_rpcs(cli);
2008 spin_unlock(&cli->cl_loi_list_lock);
2013 static int osc_teardown_async_page(struct obd_export *exp,
2014 struct lov_stripe_md *lsm,
2015 struct lov_oinfo *loi, void *cookie)
2017 struct client_obd *cli = &exp->exp_obd->u.cli;
2018 struct loi_oap_pages *lop;
2019 struct osc_async_page *oap;
2023 oap = oap_from_cookie(cookie);
2025 RETURN(PTR_ERR(oap));
2028 loi = &lsm->lsm_oinfo[0];
2030 if (oap->oap_cmd == OBD_BRW_WRITE) {
2031 lop = &loi->loi_write_lop;
2033 lop = &loi->loi_read_lop;
2036 spin_lock(&cli->cl_loi_list_lock);
2038 if (!list_empty(&oap->oap_rpc_item))
2039 GOTO(out, rc = -EBUSY);
2041 osc_exit_cache(cli, oap, 0);
2042 osc_wake_cache_waiters(cli);
2044 if (!list_empty(&oap->oap_urgent_item)) {
2045 list_del_init(&oap->oap_urgent_item);
2046 oap->oap_async_flags &= ~ASYNC_URGENT;
2048 if (!list_empty(&oap->oap_pending_item)) {
2049 list_del_init(&oap->oap_pending_item);
2050 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2052 loi_list_maint(cli, loi);
2054 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2056 spin_unlock(&cli->cl_loi_list_lock);
2058 OBD_FREE(oap, sizeof(*oap));
2063 /* Note: caller will lock/unlock, and set uptodate on the pages */
2064 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2065 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2066 struct lov_stripe_md *lsm, obd_count page_count,
2067 struct brw_page *pga)
2069 struct ptlrpc_request *request = NULL;
2070 struct ost_body *body;
2071 struct niobuf_remote *nioptr;
2072 struct obd_ioobj *iooptr;
2073 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2077 /* XXX does not handle 'new' brw protocol */
2079 size[1] = sizeof(struct obd_ioobj);
2080 size[2] = page_count * sizeof(*nioptr);
2082 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2087 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2088 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2089 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2090 sizeof(*nioptr) * page_count);
2092 memcpy(&body->oa, oa, sizeof(body->oa));
2094 obdo_to_ioobj(oa, iooptr);
2095 iooptr->ioo_bufcnt = page_count;
2097 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2098 LASSERT(PageLocked(pga[mapped].pg));
2099 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2101 nioptr->offset = pga[mapped].off;
2102 nioptr->len = pga[mapped].count;
2103 nioptr->flags = pga[mapped].flag;
2106 size[1] = page_count * sizeof(*nioptr);
2107 request->rq_replen = lustre_msg_size(2, size);
2109 rc = ptlrpc_queue_wait(request);
2113 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2114 lustre_swab_ost_body);
2116 CERROR("Can't unpack body\n");
2117 GOTO(out_req, rc = -EPROTO);
2120 memcpy(oa, &body->oa, sizeof(*oa));
2122 swab = lustre_msg_swabbed(request->rq_repmsg);
2123 LASSERT_REPSWAB(request, 1);
2124 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2126 /* nioptr missing or short */
2127 GOTO(out_req, rc = -EPROTO);
2131 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2132 struct page *page = pga[mapped].pg;
2133 struct buffer_head *bh;
2137 lustre_swab_niobuf_remote (nioptr);
2139 /* got san device associated */
2140 LASSERT(exp->exp_obd != NULL);
2141 dev = exp->exp_obd->u.cli.cl_sandev;
2144 if (!nioptr->offset) {
2145 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2146 page->mapping->host->i_ino,
2148 memset(page_address(page), 0, PAGE_SIZE);
2152 if (!page->buffers) {
2153 create_empty_buffers(page, dev, PAGE_SIZE);
2156 clear_bit(BH_New, &bh->b_state);
2157 set_bit(BH_Mapped, &bh->b_state);
2158 bh->b_blocknr = (unsigned long)nioptr->offset;
2160 clear_bit(BH_Uptodate, &bh->b_state);
2162 ll_rw_block(READ, 1, &bh);
2166 /* if buffer already existed, it must be the
2167 * one we mapped before, check it */
2168 LASSERT(!test_bit(BH_New, &bh->b_state));
2169 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2170 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2172 /* wait it's io completion */
2173 if (test_bit(BH_Lock, &bh->b_state))
2176 if (!test_bit(BH_Uptodate, &bh->b_state))
2177 ll_rw_block(READ, 1, &bh);
2181 /* must do syncronous write here */
2183 if (!buffer_uptodate(bh)) {
2191 ptlrpc_req_finished(request);
2195 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2196 struct lov_stripe_md *lsm, obd_count page_count,
2197 struct brw_page *pga)
2199 struct ptlrpc_request *request = NULL;
2200 struct ost_body *body;
2201 struct niobuf_remote *nioptr;
2202 struct obd_ioobj *iooptr;
2203 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2207 size[1] = sizeof(struct obd_ioobj);
2208 size[2] = page_count * sizeof(*nioptr);
2210 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2215 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2216 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2217 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2218 sizeof (*nioptr) * page_count);
2220 memcpy(&body->oa, oa, sizeof(body->oa));
2222 obdo_to_ioobj(oa, iooptr);
2223 iooptr->ioo_bufcnt = page_count;
2226 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2227 LASSERT(PageLocked(pga[mapped].pg));
2228 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2230 nioptr->offset = pga[mapped].off;
2231 nioptr->len = pga[mapped].count;
2232 nioptr->flags = pga[mapped].flag;
2235 size[1] = page_count * sizeof(*nioptr);
2236 request->rq_replen = lustre_msg_size(2, size);
2238 rc = ptlrpc_queue_wait(request);
2242 swab = lustre_msg_swabbed (request->rq_repmsg);
2243 LASSERT_REPSWAB (request, 1);
2244 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2246 CERROR("absent/short niobuf array\n");
2247 GOTO(out_req, rc = -EPROTO);
2251 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2252 struct page *page = pga[mapped].pg;
2253 struct buffer_head *bh;
2257 lustre_swab_niobuf_remote (nioptr);
2259 /* got san device associated */
2260 LASSERT(exp->exp_obd != NULL);
2261 dev = exp->exp_obd->u.cli.cl_sandev;
2263 if (!page->buffers) {
2264 create_empty_buffers(page, dev, PAGE_SIZE);
2267 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2268 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2269 LASSERT(page->buffers->b_blocknr ==
2270 (unsigned long)nioptr->offset);
2276 /* if buffer locked, wait it's io completion */
2277 if (test_bit(BH_Lock, &bh->b_state))
2280 clear_bit(BH_New, &bh->b_state);
2281 set_bit(BH_Mapped, &bh->b_state);
2283 /* override the block nr */
2284 bh->b_blocknr = (unsigned long)nioptr->offset;
2286 /* we are about to write it, so set it
2288 * page lock should garentee no race condition here */
2289 set_bit(BH_Uptodate, &bh->b_state);
2290 set_bit(BH_Dirty, &bh->b_state);
2292 ll_rw_block(WRITE, 1, &bh);
2294 /* must do syncronous write here */
2296 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2304 ptlrpc_req_finished(request);
2308 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2309 struct lov_stripe_md *lsm, obd_count page_count,
2310 struct brw_page *pga, struct obd_trans_info *oti)
2314 while (page_count) {
2315 obd_count pages_per_brw;
2318 if (page_count > PTLRPC_MAX_BRW_PAGES)
2319 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2321 pages_per_brw = page_count;
2323 if (cmd & OBD_BRW_WRITE)
2324 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2326 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2331 page_count -= pages_per_brw;
2332 pga += pages_per_brw;
2339 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2341 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2343 LASSERT(lock != NULL);
2344 l_lock(&lock->l_resource->lr_namespace->ns_lock);
2346 if (lock->l_ast_data && lock->l_ast_data != data) {
2347 struct inode *new_inode = data;
2348 struct inode *old_inode = lock->l_ast_data;
2349 LASSERTF(old_inode->i_state & I_FREEING,
2350 "Found existing inode %p/%lu/%u state %lu in lock: "
2351 "setting data to %p/%lu/%u\n", old_inode,
2352 old_inode->i_ino, old_inode->i_generation,
2354 new_inode, new_inode->i_ino, new_inode->i_generation);
2357 lock->l_ast_data = data;
2358 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2359 LDLM_LOCK_PUT(lock);
2362 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2363 ldlm_iterator_t replace, void *data)
2365 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2366 struct obd_device *obd = class_exp2obd(exp);
2368 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2372 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2373 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2374 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2375 void *data, __u32 lvb_len, void *lvb_swabber,
2376 struct lustre_handle *lockh)
2378 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2379 struct obd_device *obd = exp->exp_obd;
2384 /* Filesystem lock extents are extended to page boundaries so that
2385 * dealing with the page cache is a little smoother. */
2386 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2387 policy->l_extent.end |= ~PAGE_MASK;
2389 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2392 /* Next, search for already existing extent locks that will cover us */
2393 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2396 osc_set_data_with_check(lockh, data);
2397 if (*flags & LDLM_FL_HAS_INTENT) {
2398 /* I would like to be able to ASSERT here that rss <=
2399 * kms, but I can't, for reasons which are explained in
2402 /* We already have a lock, and it's referenced */
2406 /* If we're trying to read, we also search for an existing PW lock. The
2407 * VFS and page cache already protect us locally, so lots of readers/
2408 * writers can share a single PW lock.
2410 * There are problems with conversion deadlocks, so instead of
2411 * converting a read lock to a write lock, we'll just enqueue a new
2414 * At some point we should cancel the read lock instead of making them
2415 * send us a blocking callback, but there are problems with canceling
2416 * locks out from other users right now, too. */
2418 if (mode == LCK_PR) {
2419 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2420 policy, LCK_PW, lockh);
2422 /* FIXME: This is not incredibly elegant, but it might
2423 * be more elegant than adding another parameter to
2424 * lock_match. I want a second opinion. */
2425 ldlm_lock_addref(lockh, LCK_PR);
2426 ldlm_lock_decref(lockh, LCK_PW);
2427 osc_set_data_with_check(lockh, data);
2433 rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, res_id, type,
2434 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2435 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2437 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2438 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2439 lvb.lvb_size, lvb.lvb_blocks);
2440 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2441 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2447 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2448 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2449 int *flags, void *data, struct lustre_handle *lockh)
2451 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2452 struct obd_device *obd = exp->exp_obd;
2456 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2458 /* Filesystem lock extents are extended to page boundaries so that
2459 * dealing with the page cache is a little smoother */
2460 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2461 policy->l_extent.end |= ~PAGE_MASK;
2463 /* Next, search for already existing extent locks that will cover us */
2464 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2465 policy, mode, lockh);
2467 osc_set_data_with_check(lockh, data);
2470 /* If we're trying to read, we also search for an existing PW lock. The
2471 * VFS and page cache already protect us locally, so lots of readers/
2472 * writers can share a single PW lock. */
2473 if (mode == LCK_PR) {
2474 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2475 policy, LCK_PW, lockh);
2476 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2477 /* FIXME: This is not incredibly elegant, but it might
2478 * be more elegant than adding another parameter to
2479 * lock_match. I want a second opinion. */
2480 osc_set_data_with_check(lockh, data);
2481 ldlm_lock_addref(lockh, LCK_PR);
2482 ldlm_lock_decref(lockh, LCK_PW);
2488 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2489 __u32 mode, struct lustre_handle *lockh)
2493 if (mode == LCK_GROUP)
2494 ldlm_lock_decref_and_cancel(lockh, mode);
2496 ldlm_lock_decref(lockh, mode);
2501 static int osc_cancel_unused(struct obd_export *exp,
2502 struct lov_stripe_md *lsm, int flags, void *opaque)
2504 struct obd_device *obd = class_exp2obd(exp);
2505 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2507 return ldlm_cli_cancel_unused(obd->obd_namespace, &res_id, flags,
2511 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2512 unsigned long max_age)
2514 struct obd_statfs *msfs;
2515 struct ptlrpc_request *request;
2516 int rc, size = sizeof(*osfs);
2519 /* We could possibly pass max_age in the request (as an absolute
2520 * timestamp or a "seconds.usec ago") so the target can avoid doing
2521 * extra calls into the filesystem if that isn't necessary (e.g.
2522 * during mount that would help a bit). Having relative timestamps
2523 * is not so great if request processing is slow, while absolute
2524 * timestamps are not ideal because they need time synchronization. */
2525 request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2529 request->rq_replen = lustre_msg_size(1, &size);
2530 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2532 rc = ptlrpc_queue_wait(request);
2536 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2537 lustre_swab_obd_statfs);
2539 CERROR("Can't unpack obd_statfs\n");
2540 GOTO(out, rc = -EPROTO);
2543 memcpy(osfs, msfs, sizeof(*osfs));
2547 ptlrpc_req_finished(request);
2551 /* Retrieve object striping information.
2553 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2554 * the maximum number of OST indices which will fit in the user buffer.
2555 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2557 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2559 struct lov_user_md lum, *lumk;
2566 rc = copy_from_user(&lum, lump, sizeof(lum));
2570 if (lum.lmm_magic != LOV_USER_MAGIC)
2573 if (lum.lmm_stripe_count > 0) {
2574 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2575 OBD_ALLOC(lumk, lum_size);
2579 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2581 lum_size = sizeof(lum);
2585 lumk->lmm_object_id = lsm->lsm_object_id;
2586 lumk->lmm_stripe_count = 1;
2588 if (copy_to_user(lump, lumk, lum_size))
2592 OBD_FREE(lumk, lum_size);
2597 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2598 void *karg, void *uarg)
2600 struct obd_device *obd = exp->exp_obd;
2601 struct obd_ioctl_data *data = karg;
2605 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2608 if (!try_module_get(THIS_MODULE)) {
2609 CERROR("Can't get module. Is it alive?");
2614 case OBD_IOC_LOV_GET_CONFIG: {
2616 struct lov_desc *desc;
2617 struct obd_uuid uuid;
2621 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2622 GOTO(out, err = -EINVAL);
2624 data = (struct obd_ioctl_data *)buf;
2626 if (sizeof(*desc) > data->ioc_inllen1) {
2628 GOTO(out, err = -EINVAL);
2631 if (data->ioc_inllen2 < sizeof(uuid)) {
2633 GOTO(out, err = -EINVAL);
2636 desc = (struct lov_desc *)data->ioc_inlbuf1;
2637 desc->ld_tgt_count = 1;
2638 desc->ld_active_tgt_count = 1;
2639 desc->ld_default_stripe_count = 1;
2640 desc->ld_default_stripe_size = 0;
2641 desc->ld_default_stripe_offset = 0;
2642 desc->ld_pattern = 0;
2643 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2645 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2647 err = copy_to_user((void *)uarg, buf, len);
2650 obd_ioctl_freedata(buf, len);
2653 case LL_IOC_LOV_SETSTRIPE:
2654 err = obd_alloc_memmd(exp, karg);
2658 case LL_IOC_LOV_GETSTRIPE:
2659 err = osc_getstripe(karg, uarg);
2661 case OBD_IOC_CLIENT_RECOVER:
2662 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2667 case IOC_OSC_SET_ACTIVE:
2668 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2672 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2673 GOTO(out, err = -ENOTTY);
2676 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2679 module_put(THIS_MODULE);
2684 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2685 void *key, __u32 *vallen, void *val)
2688 if (!vallen || !val)
2691 if (keylen > strlen("lock_to_stripe") &&
2692 strcmp(key, "lock_to_stripe") == 0) {
2693 __u32 *stripe = val;
2694 *vallen = sizeof(*stripe);
2697 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2698 struct ptlrpc_request *req;
2700 char *bufs[1] = {key};
2702 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2707 req->rq_replen = lustre_msg_size(1, vallen);
2708 rc = ptlrpc_queue_wait(req);
2712 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2713 lustre_swab_ost_last_id);
2714 if (reply == NULL) {
2715 CERROR("Can't unpack OST last ID\n");
2716 GOTO(out, rc = -EPROTO);
2718 *((obd_id *)val) = *reply;
2720 ptlrpc_req_finished(req);
2726 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2727 void *key, obd_count vallen, void *val)
2729 struct ptlrpc_request *req;
2730 struct obd_device *obd = exp->exp_obd;
2731 struct obd_import *imp = class_exp2cliimp(exp);
2732 struct llog_ctxt *ctxt;
2733 int rc, size = keylen;
2734 char *bufs[1] = {key};
2737 if (keylen == strlen("next_id") &&
2738 memcmp(key, "next_id", strlen("next_id")) == 0) {
2739 if (vallen != sizeof(obd_id))
2741 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
2742 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
2743 exp->exp_obd->obd_name,
2744 obd->u.cli.cl_oscc.oscc_next_id);
2749 if (keylen == strlen("growth_count") &&
2750 memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2751 if (vallen != sizeof(int))
2753 obd->u.cli.cl_oscc.oscc_grow_count = *((int*)val);
2757 if (keylen == strlen("unlinked") &&
2758 memcmp(key, "unlinked", keylen) == 0) {
2759 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2760 spin_lock(&oscc->oscc_lock);
2761 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2762 spin_unlock(&oscc->oscc_lock);
2767 if (keylen == strlen("initial_recov") &&
2768 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2769 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2770 if (vallen != sizeof(int))
2772 imp->imp_initial_recov = *(int *)val;
2773 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2774 exp->exp_obd->obd_name,
2775 imp->imp_initial_recov);
2779 if (keylen < strlen("mds_conn") ||
2780 memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2784 req = ptlrpc_prep_req(imp, OST_SET_INFO, 1, &size, bufs);
2788 req->rq_replen = lustre_msg_size(0, NULL);
2789 rc = ptlrpc_queue_wait(req);
2790 ptlrpc_req_finished(req);
2792 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_ORIG_CTXT);
2794 rc = llog_initiator_connect(ctxt);
2799 imp->imp_server_timeout = 1;
2800 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2801 imp->imp_pingable = 1;
2807 static struct llog_operations osc_size_repl_logops = {
2808 lop_cancel: llog_obd_repl_cancel
2811 static struct llog_operations osc_unlink_orig_logops;
2812 static int osc_llog_init(struct obd_device *obd, struct obd_device *tgt,
2813 int count, struct llog_catid *catid)
2818 osc_unlink_orig_logops = llog_lvfs_ops;
2819 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2820 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
2821 osc_unlink_orig_logops.lop_add = llog_catalog_add;
2822 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2824 rc = obd_llog_setup(obd, LLOG_UNLINK_ORIG_CTXT, tgt, count,
2825 &catid->lci_logid, &osc_unlink_orig_logops);
2829 rc = obd_llog_setup(obd, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2830 &osc_size_repl_logops);
2834 static int osc_llog_finish(struct obd_device *obd, int count)
2839 rc = obd_llog_cleanup(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT));
2843 rc = obd_llog_cleanup(llog_get_context(obd, LLOG_SIZE_REPL_CTXT));
2848 static int osc_connect(struct lustre_handle *exph,
2849 struct obd_device *obd, struct obd_uuid *cluuid)
2853 rc = client_connect_import(exph, obd, cluuid);
2858 static int osc_disconnect(struct obd_export *exp, int flags)
2860 struct obd_device *obd = class_exp2obd(exp);
2861 struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2864 if (obd->u.cli.cl_conn_count == 1)
2865 /* flush any remaining cancel messages out to the target */
2866 llog_sync(ctxt, exp);
2868 rc = client_disconnect_export(exp, flags);
2872 static int osc_import_event(struct obd_device *obd,
2873 struct obd_import *imp,
2874 enum obd_import_event event)
2876 struct client_obd *cli;
2879 LASSERT(imp->imp_obd == obd);
2882 case IMP_EVENT_DISCON: {
2883 /* Only do this on the MDS OSC's */
2884 if (imp->imp_server_timeout) {
2885 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2887 spin_lock(&oscc->oscc_lock);
2888 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
2889 spin_unlock(&oscc->oscc_lock);
2893 case IMP_EVENT_INACTIVE: {
2894 if (obd->obd_observer)
2895 rc = obd_notify(obd->obd_observer, obd, 0);
2898 case IMP_EVENT_INVALIDATE: {
2899 struct ldlm_namespace *ns = obd->obd_namespace;
2903 spin_lock(&cli->cl_loi_list_lock);
2904 cli->cl_avail_grant = 0;
2905 cli->cl_lost_grant = 0;
2906 /* all pages go to failing rpcs due to the invalid import */
2907 osc_check_rpcs(cli);
2908 spin_unlock(&cli->cl_loi_list_lock);
2910 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2914 case IMP_EVENT_ACTIVE: {
2915 if (obd->obd_observer)
2916 rc = obd_notify(obd->obd_observer, obd, 1);
2920 CERROR("Unknown import event %d\n", event);
2926 int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2930 rc = ptlrpcd_addref();
2934 rc = client_obd_setup(obd, len, buf);
2938 struct lprocfs_static_vars lvars;
2940 lprocfs_init_vars(osc, &lvars);
2941 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
2942 lproc_osc_attach_seqstat(obd);
2943 ptlrpc_lprocfs_register_obd(obd);
2952 int osc_cleanup(struct obd_device *obd, int flags)
2956 ptlrpc_lprocfs_unregister_obd(obd);
2957 lprocfs_obd_cleanup(obd);
2959 rc = client_obd_cleanup(obd, flags);
2965 struct obd_ops osc_obd_ops = {
2966 .o_owner = THIS_MODULE,
2967 .o_setup = osc_setup,
2968 .o_cleanup = osc_cleanup,
2969 .o_connect = osc_connect,
2970 .o_disconnect = osc_disconnect,
2971 .o_statfs = osc_statfs,
2972 .o_packmd = osc_packmd,
2973 .o_unpackmd = osc_unpackmd,
2974 .o_create = osc_create,
2975 .o_destroy = osc_destroy,
2976 .o_getattr = osc_getattr,
2977 .o_getattr_async = osc_getattr_async,
2978 .o_setattr = osc_setattr,
2980 .o_brw_async = osc_brw_async,
2981 .o_prep_async_page = osc_prep_async_page,
2982 .o_queue_async_io = osc_queue_async_io,
2983 .o_set_async_flags = osc_set_async_flags,
2984 .o_queue_group_io = osc_queue_group_io,
2985 .o_trigger_group_io = osc_trigger_group_io,
2986 .o_teardown_async_page = osc_teardown_async_page,
2987 .o_punch = osc_punch,
2989 .o_enqueue = osc_enqueue,
2990 .o_match = osc_match,
2991 .o_change_cbdata = osc_change_cbdata,
2992 .o_cancel = osc_cancel,
2993 .o_cancel_unused = osc_cancel_unused,
2994 .o_iocontrol = osc_iocontrol,
2995 .o_get_info = osc_get_info,
2996 .o_set_info = osc_set_info,
2997 .o_import_event = osc_import_event,
2998 .o_llog_init = osc_llog_init,
2999 .o_llog_finish = osc_llog_finish,
3002 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3003 struct obd_ops sanosc_obd_ops = {
3004 .o_owner = THIS_MODULE,
3005 .o_cleanup = client_obd_cleanup,
3006 .o_connect = osc_connect,
3007 .o_disconnect = client_disconnect_export,
3008 .o_statfs = osc_statfs,
3009 .o_packmd = osc_packmd,
3010 .o_unpackmd = osc_unpackmd,
3011 .o_create = osc_real_create,
3012 .o_destroy = osc_destroy,
3013 .o_getattr = osc_getattr,
3014 .o_getattr_async = osc_getattr_async,
3015 .o_setattr = osc_setattr,
3016 .o_setup = client_sanobd_setup,
3017 .o_brw = sanosc_brw,
3018 .o_punch = osc_punch,
3020 .o_enqueue = osc_enqueue,
3021 .o_match = osc_match,
3022 .o_change_cbdata = osc_change_cbdata,
3023 .o_cancel = osc_cancel,
3024 .o_cancel_unused = osc_cancel_unused,
3025 .o_iocontrol = osc_iocontrol,
3026 .o_import_event = osc_import_event,
3027 .o_llog_init = osc_llog_init,
3028 .o_llog_finish = osc_llog_finish,
3032 int __init osc_init(void)
3034 struct lprocfs_static_vars lvars;
3035 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3036 struct lprocfs_static_vars sanlvars;
3041 lprocfs_init_vars(osc, &lvars);
3042 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3043 lprocfs_init_vars(osc, &sanlvars);
3046 rc = class_register_type(&osc_obd_ops, lvars.module_vars,
3051 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3052 rc = class_register_type(&sanosc_obd_ops, sanlvars.module_vars,
3053 LUSTRE_SANOSC_NAME);
3055 class_unregister_type(LUSTRE_OSC_NAME);
3062 static void /*__exit*/ osc_exit(void)
3064 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3065 class_unregister_type(LUSTRE_SANOSC_NAME);
3067 class_unregister_type(LUSTRE_OSC_NAME);
3070 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3071 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3072 MODULE_LICENSE("GPL");
3074 module_init(osc_init);
3075 module_exit(osc_exit);