* Author: Andreas Dilger <adilger@clusterfs.com>
* Author: Phil Schwan <phil@clusterfs.com>
*
- * This file is part of Lustre, http://www.lustre.org.
+ * This file is part of the Lustre file system, http://www.lustre.org
+ * Lustre is a trademark of Cluster File Systems, Inc.
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * You may have signed or agreed to another license before downloading
+ * this software. If so, you are bound by the terms and conditions
+ * of that agreement, and the following does not apply to you. See the
+ * LICENSE file included with this distribution for more information.
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * If you did not agree to a different license, then this copy of Lustre
+ * is open source software; you can redistribute it and/or modify it
+ * under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * In either case, Lustre is distributed in the hope that it will be
+ * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
+ * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * license text for more details.
*/
#define DEBUG_SUBSYSTEM S_FILTER
+#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
+#endif
#include <linux/module.h>
#include <linux/pagemap.h> // XXX kill me soon
#include <linux/version.h>
-#include <linux/obd_class.h>
-#include <linux/lustre_fsfilt.h>
-#include <linux/lustre_smfs.h>
-#include <linux/lustre_snap.h>
+#include <obd_class.h>
+#include <lustre_fsfilt.h>
#include "filter_internal.h"
+int *obdfilter_created_scratchpad;
+
static int filter_alloc_dio_page(struct obd_device *obd, struct inode *inode,
struct niobuf_local *lnb)
-
{
struct page *page;
- ENTRY;
-
- page = alloc_pages(GFP_HIGHUSER, 0);
- if (page == NULL) {
- CERROR("no memory for a temp page\n");
- lnb->rc = -ENOMEM;
- RETURN(-ENOMEM);
- }
+ LASSERT(lnb->page != NULL);
+
+ page = lnb->page;
#if 0
POISON_PAGE(page, 0xf1);
- if (lnb->len != PAGE_SIZE) {
- memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len);
+ if (lnb->len != CFS_PAGE_SIZE) {
+ memset(kmap(page) + lnb->len, 0, CFS_PAGE_SIZE - lnb->len);
kunmap(page);
}
#endif
- page->index = lnb->offset >> PAGE_SHIFT;
-
- lnb->page = page;
+ page->index = lnb->offset >> CFS_PAGE_SHIFT;
RETURN(0);
}
-void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
+static void filter_free_dio_pages(int objcount, struct obd_ioobj *obj,
int niocount, struct niobuf_local *res)
{
int i, j;
for (i = 0; i < objcount; i++, obj++) {
- for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++) {
- if (res->page != NULL) {
- __free_page(res->page);
+ for (j = 0 ; j < obj->ioo_bufcnt ; j++, res++)
res->page = NULL;
- }
- }
}
}
{
struct filter_export_data *fed;
struct obd_device *obd = exp->exp_obd;
- static unsigned long last_msg;
- static int last_count;
- int mask = D_CACHE;
ENTRY;
LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
fed = &exp->exp_filter_data;
- /* Don't print this to the console the first time it happens, since
- * it can happen legitimately on occasion, but only rarely. */
- if (time_after(jiffies, last_msg + 60 * HZ)) {
- last_count = 0;
- last_msg = jiffies;
- }
- if ((last_count & (-last_count)) == last_count)
- mask = D_WARNING;
- last_count++;
-
/* Add some margin, since there is a small race if other RPCs arrive
* out-or-order and have already consumed some grant. We want to
* leave this here in case there is a large error in accounting. */
- CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? mask:D_CACHE,
+ CDEBUG(D_CACHE,
"%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n",
obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant,
oa->o_dropped, fed->fed_grant);
/* Update our accounting now so that statfs takes it into account.
* Note that fed_dirty is only approximate and can become incorrect
* if RPCs arrive out-of-order. No important calculations depend
- * on fed_dirty however. */
+ * on fed_dirty however, but we must check sanity to not assert. */
+ if ((long long)oa->o_dirty < 0)
+ oa->o_dirty = 0;
+ else if (oa->o_dirty > fed->fed_grant + 4 * FILTER_GRANT_CHUNK)
+ oa->o_dirty = fed->fed_grant + 4 * FILTER_GRANT_CHUNK;
obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty;
if (fed->fed_grant < oa->o_dropped) {
- CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n",
+ CDEBUG(D_CACHE,"%s: cli %s/%p reports %u dropped > grant %lu\n",
obd->obd_name, exp->exp_client_uuid.uuid, exp,
oa->o_dropped, fed->fed_grant);
oa->o_dropped = 0;
obd->u.filter.fo_tot_granted -= oa->o_dropped;
fed->fed_grant -= oa->o_dropped;
fed->fed_dirty = oa->o_dirty;
+ if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
+ CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
+ obd->obd_name, exp->exp_client_uuid.uuid, exp,
+ fed->fed_dirty, fed->fed_pending, fed->fed_grant);
+ spin_unlock(&obd->obd_osfs_lock);
+ LBUG();
+ }
EXIT;
}
-#define GRANT_FOR_LLOG(obd) 16
-
/* Figure out how much space is available between what we've granted
* and what remains in the filesystem. Compensate for ext3 indirect
* block overhead when computing how much free space is left ungranted.
obd_size filter_grant_space_left(struct obd_export *exp)
{
struct obd_device *obd = exp->exp_obd;
- int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
+ int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0;
int rc, statfs_done = 0;
LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
- if (time_before(obd->obd_osfs_age, jiffies - HZ)) {
+ if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
restat:
- rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1);
+ rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
+ cfs_time_current_64() + HZ);
if (rc) /* N.B. statfs can't really fail */
RETURN(0);
statfs_done = 1;
if (left >= tot_granted) {
left -= tot_granted;
} else {
- static unsigned long next;
- if (left < tot_granted - obd->u.filter.fo_tot_pending &&
- time_after(jiffies, next)) {
- spin_unlock(&obd->obd_osfs_lock);
+ if (left < tot_granted - obd->u.filter.fo_tot_pending) {
CERROR("%s: cli %s/%p grant "LPU64" > available "
LPU64" and pending "LPU64"\n", obd->obd_name,
exp->exp_client_uuid.uuid, exp, tot_granted,
left, obd->u.filter.fo_tot_pending);
- if (next == 0)
- portals_debug_dumplog();
- next = jiffies + 20 * HZ;
- spin_lock(&obd->obd_osfs_lock);
}
left = 0;
}
{
struct obd_device *obd = exp->exp_obd;
struct filter_export_data *fed = &exp->exp_filter_data;
- int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
+ int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
__u64 grant = 0;
LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
* has and what we think it has, don't grant very much and let the
* client consume its grant first. Either it just has lots of RPCs
* in flight, or it was evicted and its grants will soon be used up. */
- if (current_grant < want &&
- current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
- grant = min((want >> blockbits) / 2,
+ if (want > 0x7fffffff) {
+ CERROR("%s: client %s/%p requesting > 2GB grant "LPU64"\n",
+ obd->obd_name, exp->exp_client_uuid.uuid, exp, want);
+ } else if (current_grant < want &&
+ current_grant < fed->fed_grant + FILTER_GRANT_CHUNK) {
+ grant = min((want >> blockbits),
(fs_space_left >> blockbits) / 8);
grant <<= blockbits;
if (grant) {
- if (grant > FILTER_GRANT_CHUNK)
+ /* Allow >FILTER_GRANT_CHUNK size when clients
+ * reconnect due to a server reboot.
+ */
+ if ((grant > FILTER_GRANT_CHUNK) &&
+ (!obd->obd_recovering))
grant = FILTER_GRANT_CHUNK;
obd->u.filter.fo_tot_granted += grant;
fed->fed_grant += grant;
+ if (fed->fed_grant < 0) {
+ CERROR("%s: cli %s/%p grant %ld want "LPU64
+ "current"LPU64"\n",
+ obd->obd_name, exp->exp_client_uuid.uuid,
+ exp, fed->fed_grant, want,current_grant);
+ spin_unlock(&obd->obd_osfs_lock);
+ LBUG();
+ }
}
}
- CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n",
- obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant);
+ CDEBUG(D_CACHE,
+ "%s: cli %s/%p wants: "LPU64" current grant "LPU64
+ " granting: "LPU64"\n", obd->obd_name, exp->exp_client_uuid.uuid,
+ exp, want, current_grant, grant);
CDEBUG(D_CACHE,
"%s: cli %s/%p tot cached:"LPU64" granted:"LPU64
" num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid,
return grant;
}
-
static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
int objcount, struct obd_ioobj *obj,
int niocount, struct niobuf_remote *nb,
struct niobuf_local *res,
- struct obd_trans_info *oti)
+ struct obd_trans_info *oti,
+ struct lustre_capa *capa)
{
struct obd_device *obd = exp->exp_obd;
struct lvfs_run_ctxt saved;
struct niobuf_local *lnb;
struct dentry *dentry = NULL;
struct inode *inode;
- void *iobuf = NULL;
+ void *iobuf = NULL;
int rc = 0, i, tot_bytes = 0;
unsigned long now = jiffies;
ENTRY;
/* We are currently not supporting multi-obj BRW_READ RPCS at all.
- * When we do this function's dentry cleanup will need to be fixed */
+ * When we do this function's dentry cleanup will need to be fixed.
+ * These values are verified in ost_brw_write() from the wire. */
LASSERTF(objcount == 1, "%d\n", objcount);
LASSERTF(obj->ioo_bufcnt > 0, "%d\n", obj->ioo_bufcnt);
+ rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
+ CAPA_OPC_OSS_READ);
+ if (rc)
+ RETURN(rc);
+
if (oa && oa->o_valid & OBD_MD_FLGRANT) {
spin_lock(&obd->obd_osfs_lock);
filter_grant_incoming(exp, oa);
oa->o_grant = 0;
-
spin_unlock(&obd->obd_osfs_lock);
}
- memset(res, 0, niocount * sizeof(*res));
-
- push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
- if (rc)
- GOTO(cleanup, rc);
+ iobuf = filter_iobuf_get(&obd->u.filter, oti);
+ if (IS_ERR(iobuf))
+ RETURN(PTR_ERR(iobuf));
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
dentry = filter_oa2dentry(obd, oa);
- if (IS_ERR(dentry))
- GOTO(cleanup, rc = PTR_ERR(dentry));
-
- if (dentry->d_inode == NULL) {
- CERROR("trying to BRW to non-existent file "LPU64"\n",
- obj->ioo_id);
- GOTO(cleanup, rc = -ENOENT);
+ if (IS_ERR(dentry)) {
+ rc = PTR_ERR(dentry);
+ dentry = NULL;
+ GOTO(cleanup, rc);
}
- inode = dentry->d_inode;
+ inode = dentry->d_inode;
- if (time_after(jiffies, now + 15 * HZ))
- CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ);
- else
- CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
- (jiffies - now));
+ obdo_to_inode(inode, oa, OBD_MD_FLATIME);
+ fsfilt_check_slow(obd, now, obd_timeout, "preprw_read setup");
for (i = 0, lnb = res, rnb = nb; i < obj->ioo_bufcnt;
i++, rnb++, lnb++) {
lnb->len = rnb->len;
lnb->flags = rnb->flags;
- if (inode->i_size <= rnb->offset)
- /* If there's no more data, abort early.
- * lnb->page == NULL and lnb->rc == 0, so it's
- * easy to detect later. */
+ /*
+ * ost_brw_write()->ost_nio_pages_get() already initialized
+ * lnb->page to point to the page from the per-thread page
+ * pool (bug 5137), initialize page.
+ */
+ LASSERT(lnb->page != NULL);
+
+ if (i_size_read(inode) <= rnb->offset)
+ /* If there's no more data, abort early. lnb->rc == 0,
+ * so it's easy to detect later. */
break;
else
- rc = filter_alloc_dio_page(obd, inode, lnb);
- if (rc) {
- CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
- "page err %u@"LPU64" %u/%u %p: rc %d\n",
- lnb->len, lnb->offset, i, obj->ioo_bufcnt,
- dentry, rc);
- GOTO(cleanup, rc);
- }
+ filter_alloc_dio_page(obd, inode, lnb);
- if (inode->i_size < lnb->offset + lnb->len - 1)
- lnb->rc = inode->i_size - lnb->offset;
+ if (i_size_read(inode) < lnb->offset + lnb->len - 1)
+ lnb->rc = i_size_read(inode) - lnb->offset;
else
lnb->rc = lnb->len;
filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
}
- if (time_after(jiffies, now + 15 * HZ))
- CERROR("slow start_page_read %lus\n", (jiffies - now) / HZ);
- else
- CDEBUG(D_INFO, "start_page_read: %lu jiffies\n",
- (jiffies - now));
+ fsfilt_check_slow(obd, now, obd_timeout, "start_page_read");
- rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
- NULL, NULL, NULL);
+ rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf,
+ exp, NULL, NULL, NULL);
if (rc)
GOTO(cleanup, rc);
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
- filter_tally_read(&exp->exp_obd->u.filter, res, niocount);
+ lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
+ tot_bytes);
EXIT;
-cleanup:
+ cleanup:
if (rc != 0) {
filter_free_dio_pages(objcount, obj, niocount, res);
if (dentry != NULL)
f_dput(dentry);
- else
- CERROR("NULL dentry in cleanup -- tell CFS\n");
}
- if (iobuf != NULL)
- filter_free_iobuf(iobuf);
+ filter_iobuf_put(&obd->u.filter, iobuf, oti);
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
if (rc)
CERROR("io error %d\n", rc);
+
return rc;
}
* right on through.
*
* Caller must hold obd_osfs_lock. */
-static int filter_grant_check(struct obd_export *exp, int objcount,
- struct fsfilt_objinfo *fso, int niocount,
- struct niobuf_remote *rnb,
+static int filter_grant_check(struct obd_export *exp, struct obdo *oa,
+ int objcount, struct fsfilt_objinfo *fso,
+ int niocount, struct niobuf_remote *rnb,
struct niobuf_local *lnb, obd_size *left,
struct inode *inode)
{
struct filter_export_data *fed = &exp->exp_filter_data;
- int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
+ int blocksize = exp->exp_obd->u.obt.obt_sb->s_blocksize;
unsigned long used = 0, ungranted = 0, using;
- int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE;
+ int i, rc = -ENOSPC, obj, n = 0;
LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
int tmp, bytes;
- /* FIXME: this is calculated with PAGE_SIZE on client */
+ /* should match the code in osc_exit_cache */
bytes = rnb[n].len;
bytes += rnb[n].offset & (blocksize - 1);
tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
if (tmp)
bytes += blocksize - tmp;
- if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
+ if ((rnb[n].flags & OBD_BRW_FROM_GRANT) &&
+ (oa->o_valid & OBD_MD_FLGRANT)) {
if (fed->fed_grant < used + bytes) {
CDEBUG(D_CACHE,
"%s: cli %s/%p claims %ld+%d "
exp->exp_obd->obd_name,
exp->exp_client_uuid.uuid, exp,
used, bytes, fed->fed_grant, n);
- mask = D_ERROR;
} else {
used += bytes;
rnb[n].flags |= OBD_BRW_GRANTED;
continue;
}
}
- if (*left > ungranted) {
+ if (*left > ungranted + bytes) {
/* if enough space, pretend it was granted */
ungranted += bytes;
rnb[n].flags |= OBD_BRW_GRANTED;
+ lnb[n].lnb_grant_used = bytes;
CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted);
rc = 0;
continue;
* marked BRW_GRANTED are already mapped and we can
* ignore this error. */
lnb[n].rc = -ENOSPC;
- rnb[n].flags &= OBD_BRW_GRANTED;
+ rnb[n].flags &= ~OBD_BRW_GRANTED;
CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n",
exp->exp_obd->obd_name,
exp->exp_client_uuid.uuid, exp, n, bytes);
* happens in filter_grant_commit() after the writes are done. */
*left -= ungranted;
fed->fed_grant -= used;
- fed->fed_pending += used;
- exp->exp_obd->u.filter.fo_tot_pending += used;
+ fed->fed_pending += used + ungranted;
+ exp->exp_obd->u.filter.fo_tot_granted += ungranted;
+ exp->exp_obd->u.filter.fo_tot_pending += used + ungranted;
- CDEBUG(mask,
+ CDEBUG(D_CACHE,
"%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n",
exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used,
ungranted, fed->fed_grant, fed->fed_dirty);
/* Rough calc in case we don't refresh cached statfs data */
using = (used + ungranted + 1 ) >>
- exp->exp_obd->u.filter.fo_sb->s_blocksize_bits;
+ exp->exp_obd->u.obt.obt_sb->s_blocksize_bits;
if (exp->exp_obd->obd_osfs.os_bavail > using)
exp->exp_obd->obd_osfs.os_bavail -= using;
else
exp->exp_obd->u.filter.fo_tot_dirty -= used;
fed->fed_dirty -= used;
+ if (fed->fed_dirty < 0 || fed->fed_grant < 0 || fed->fed_pending < 0) {
+ CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
+ exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
+ fed->fed_dirty, fed->fed_pending, fed->fed_grant);
+ spin_unlock(&exp->exp_obd->obd_osfs_lock);
+ LBUG();
+ }
return rc;
}
int objcount, struct obd_ioobj *obj,
int niocount, struct niobuf_remote *nb,
struct niobuf_local *res,
- struct obd_trans_info *oti)
+ struct obd_trans_info *oti,
+ struct lustre_capa *capa)
{
struct lvfs_run_ctxt saved;
struct niobuf_remote *rnb;
struct niobuf_local *lnb = res;
struct fsfilt_objinfo fso;
+ struct filter_mod_data *fmd;
struct dentry *dentry = NULL;
- void *iobuf;
+ void *iobuf;
obd_size left;
unsigned long now = jiffies;
int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
LASSERT(objcount == 1);
LASSERT(obj->ioo_bufcnt > 0);
- memset(res, 0, niocount * sizeof(*res));
-
- rc = filter_alloc_iobuf(OBD_BRW_READ, obj->ioo_bufcnt, &iobuf);
+ rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
+ CAPA_OPC_OSS_WRITE);
if (rc)
- GOTO(cleanup, rc);
- cleanup_phase = 1;
+ RETURN(rc);
push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
- obj->ioo_id);
+ iobuf = filter_iobuf_get(&exp->exp_obd->u.filter, oti);
+ if (IS_ERR(iobuf))
+ GOTO(cleanup, rc = PTR_ERR(iobuf));
+ cleanup_phase = 1;
+
+ dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
+ obj->ioo_id);
if (IS_ERR(dentry))
GOTO(cleanup, rc = PTR_ERR(dentry));
-
cleanup_phase = 2;
-
+
if (dentry->d_inode == NULL) {
- CERROR("trying to BRW to non-existent file "LPU64"\n",
- obj->ioo_id);
+ CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
+ exp->exp_obd->obd_name, obj->ioo_id);
GOTO(cleanup, rc = -ENOENT);
}
fso.fso_dentry = dentry;
fso.fso_bufcnt = obj->ioo_bufcnt;
- if (time_after(jiffies, now + 15 * HZ))
- CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
- else
- CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
- (jiffies - now));
+ fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "preprw_write setup");
+ /* Don't update inode timestamps if this write is older than a
+ * setattr which modifies the timestamps. b=10150 */
+ /* XXX when we start having persistent reservations this needs to
+ * be changed to filter_fmd_get() to create the fmd if it doesn't
+ * already exist so we can store the reservation handle there. */
+ fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_gr);
+
+ LASSERT(oa != NULL);
spin_lock(&exp->exp_obd->obd_osfs_lock);
- if (oa)
- filter_grant_incoming(exp, oa);
-
+ filter_grant_incoming(exp, oa);
+ if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
+ oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
+ OBD_MD_FLATIME);
+ else
+ obdo_to_inode(dentry->d_inode, oa, OBD_MD_FLATIME |
+ OBD_MD_FLMTIME | OBD_MD_FLCTIME);
cleanup_phase = 3;
left = filter_grant_space_left(exp);
- rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res,
+ rc = filter_grant_check(exp, oa, objcount, &fso, niocount, nb, res,
&left, dentry->d_inode);
- if (oa && oa->o_valid & OBD_MD_FLGRANT)
+
+ /* do not zero out oa->o_valid as it is used in filter_commitrw_write()
+ * for setting UID/GID and fid EA in first write time. */
+ if (oa->o_valid & OBD_MD_FLGRANT)
oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left);
spin_unlock(&exp->exp_obd->obd_osfs_lock);
+ filter_fmd_put(exp, fmd);
- if (rc)
+ if (rc)
GOTO(cleanup, rc);
for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
lnb->len = rnb->len;
lnb->flags = rnb->flags;
- rc = filter_alloc_dio_page(exp->exp_obd, dentry->d_inode,lnb);
- if (rc) {
- CERROR("page err %u@"LPU64" %u/%u %p: rc %d\n",
- lnb->len, lnb->offset,
- i, obj->ioo_bufcnt, dentry, rc);
- GOTO(cleanup, rc);
+ /*
+ * ost_brw_write()->ost_nio_pages_get() already initialized
+ * lnb->page to point to the page from the per-thread page
+ * pool (bug 5137), initialize page.
+ */
+ LASSERT(lnb->page != NULL);
+ if (lnb->len != CFS_PAGE_SIZE) {
+ memset(kmap(lnb->page) + lnb->len,
+ 0, CFS_PAGE_SIZE - lnb->len);
+ kunmap(lnb->page);
}
+ lnb->page->index = lnb->offset >> CFS_PAGE_SHIFT;
+
cleanup_phase = 4;
/* If the filter writes a partial page, then has the file
* page on disk. we do it by hand for partial extending
* writes, send_bio() is responsible for zeroing pages when
* asked to read unmapped blocks -- brw_kiovec() does this. */
- if (lnb->len != PAGE_SIZE) {
- if (lnb->offset + lnb->len < dentry->d_inode->i_size) {
+ if (lnb->len != CFS_PAGE_SIZE) {
+ __s64 maxidx;
+
+ maxidx = ((i_size_read(dentry->d_inode) +
+ CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
+ if (maxidx >= lnb->page->index) {
+ LL_CDEBUG_PAGE(D_PAGE, lnb->page, "write %u @ "
+ LPU64" flg %x before EOF %llu\n",
+ lnb->len, lnb->offset,lnb->flags,
+ i_size_read(dentry->d_inode));
filter_iobuf_add_page(exp->exp_obd, iobuf,
dentry->d_inode,
lnb->page);
} else {
- memset(kmap(lnb->page) + lnb->len, 0,
- PAGE_SIZE - lnb->len);
+ long off;
+ char *p = kmap(lnb->page);
+
+ off = lnb->offset & ~CFS_PAGE_MASK;
+ if (off)
+ memset(p, 0, off);
+ off = (lnb->offset + lnb->len) & ~CFS_PAGE_MASK;
+ if (off)
+ memset(p + off, 0, CFS_PAGE_SIZE - off);
kunmap(lnb->page);
}
}
rc = filter_direct_io(OBD_BRW_READ, dentry, iobuf, exp,
NULL, NULL, NULL);
-
- if (time_after(jiffies, now + 15 * HZ))
- CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
- else
- CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
- (jiffies - now));
- lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
+ fsfilt_check_slow(exp->exp_obd, now, obd_timeout, "start_page_write");
+
+ lprocfs_counter_add(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
tot_bytes);
EXIT;
cleanup:
switch(cleanup_phase) {
case 4:
- if (rc)
- filter_free_dio_pages(objcount, obj, niocount, res);
case 3:
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- filter_free_iobuf(iobuf);
+ filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
case 2:
+ pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
if (rc)
f_dput(dentry);
break;
case 1:
+ filter_iobuf_put(&exp->exp_obd->u.filter, iobuf, oti);
+ case 0:
spin_lock(&exp->exp_obd->obd_osfs_lock);
if (oa)
filter_grant_incoming(exp, oa);
spin_unlock(&exp->exp_obd->obd_osfs_lock);
pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- filter_free_iobuf(iobuf);
break;
default:;
-
}
- RETURN(rc);
+ return rc;
}
int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
int objcount, struct obd_ioobj *obj, int niocount,
struct niobuf_remote *nb, struct niobuf_local *res,
- struct obd_trans_info *oti)
+ struct obd_trans_info *oti, struct lustre_capa *capa)
{
if (cmd == OBD_BRW_WRITE)
return filter_preprw_write(cmd, exp, oa, objcount, obj,
- niocount, nb, res, oti);
-
+ niocount, nb, res, oti, capa);
if (cmd == OBD_BRW_READ)
return filter_preprw_read(cmd, exp, oa, objcount, obj,
- niocount, nb, res, oti);
-
+ niocount, nb, res, oti, capa);
LBUG();
return -EPROTO;
}
int drop = 0;
if (inode != NULL &&
- (inode->i_size > filter->fo_readcache_max_filesize))
+ (i_size_read(inode) > filter->fo_readcache_max_filesize))
drop = 1;
/* drop from cache like truncate_list_pages() */
struct obd_trans_info *oti, int rc)
{
struct inode *inode = NULL;
+ struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0,
+ obj->ioo_gr, 0} };
+ struct ldlm_resource *resource = NULL;
+ struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
ENTRY;
+ /* If oa != NULL then filter_preprw_read updated the inode atime
+ * and we should update the lvb so that other glimpses will also
+ * get the updated value. bug 5972 */
+ if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
+ resource = ldlm_resource_get(ns, NULL, &res_id, LDLM_EXTENT, 0);
+
+ if (resource != NULL) {
+ ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
+ ldlm_resource_putref(resource);
+ }
+ }
+
if (res->dentry != NULL)
inode = res->dentry->d_inode;
filter_free_dio_pages(objcount, obj, niocount, res);
-
+
if (res->dentry != NULL)
f_dput(res->dentry);
RETURN(rc);
spin_unlock(&exp->exp_obd->obd_osfs_lock);
}
-int filter_do_cow(struct obd_export *exp, struct obd_ioobj *obj,
- int nioo, struct niobuf_remote *rnb)
-{
- struct dentry *dentry;
- struct lvfs_run_ctxt saved;
- struct write_extents *extents = NULL;
- int j, rc = 0, numexts = 0, flags = 0;
-
- ENTRY;
-
- LASSERT(nioo == 1);
-
- push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
-
- dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
- obj->ioo_id);
- if (IS_ERR(dentry)) {
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- RETURN (PTR_ERR(dentry));
- }
-
- if (dentry->d_inode == NULL) {
- CERROR("trying to write extents to non-existent file "LPU64"\n",
- obj->ioo_id);
- GOTO(cleanup, rc = -ENOENT);
- }
-
- flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
- if (!(flags & SM_DO_COW)) {
- GOTO(cleanup, rc);
- }
- OBD_ALLOC(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
- if (!extents) {
- CERROR("No Memory\n");
- GOTO(cleanup, rc = -ENOMEM);
- }
- for (j = 0; j < obj->ioo_bufcnt; j++) {
- if (rnb[j].len != 0) {
- extents[numexts].w_count = rnb[j].len;
- extents[numexts].w_pos = rnb[j].offset;
- numexts++;
- }
- }
- rc = fsfilt_do_write_cow(exp->exp_obd, dentry, extents, numexts);
- if (rc) {
- CERROR("Do cow error id "LPU64" rc:%d \n",
- obj->ioo_id, rc);
- GOTO(cleanup, rc);
- }
-
-cleanup:
- if (extents) {
- OBD_FREE(extents, obj->ioo_bufcnt * sizeof(struct write_extents));
- }
- f_dput(dentry);
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- RETURN(rc);
-
-}
-int filter_write_extents(struct obd_export *exp, struct obd_ioobj *obj, int nobj,
- int niocount, struct niobuf_local *local, int rc)
-{
- struct lvfs_run_ctxt saved;
- struct dentry *dentry;
- struct niobuf_local *lnb;
- __u64 offset = 0;
- __u32 len = 0;
- int i, flags;
-
- ENTRY;
-
- LASSERT(nobj == 1);
-
- push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- dentry = filter_id2dentry(exp->exp_obd, NULL, obj->ioo_gr,
- obj->ioo_id);
- if (IS_ERR(dentry)) {
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- RETURN (PTR_ERR(dentry));
- }
-
- if (dentry->d_inode == NULL) {
- CERROR("trying to write extents to non-existent file "LPU64"\n",
- obj->ioo_id);
- GOTO(cleanup, rc = -ENOENT);
- }
-
- flags = fsfilt_get_fs_flags(exp->exp_obd, dentry);
- if (!(flags & SM_DO_REC)) {
- GOTO(cleanup, rc);
- }
-
- for (i = 0, lnb = local; i < obj->ioo_bufcnt; i++, lnb++) {
- if (len == 0) {
- offset = lnb->offset;
- len = lnb->len;
- } else if (lnb->offset == (offset + len)) {
- len += lnb->len;
- } else {
- rc = fsfilt_write_extents(exp->exp_obd, dentry,
- offset, len);
- if (rc) {
- CERROR("write exts off "LPU64" num %u rc:%d\n",
- offset, len, rc);
- GOTO(cleanup, rc);
- }
- offset = lnb->offset;
- len = lnb->len;
- }
- }
- if (len > 0) {
- rc = fsfilt_write_extents(exp->exp_obd, dentry,
- offset, len);
- if (rc) {
- CERROR("write exts off "LPU64" num %u rc:%d\n",
- offset, len, rc);
- GOTO(cleanup, rc);
- }
- }
-cleanup:
- f_dput(dentry);
- pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- RETURN(rc);
-}
int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
int objcount, struct obd_ioobj *obj, int niocount,
- struct niobuf_local *res, struct obd_trans_info *oti,int rc)
+ struct niobuf_local *res, struct obd_trans_info *oti,
+ int rc)
{
if (cmd == OBD_BRW_WRITE)
return filter_commitrw_write(exp, oa, objcount, obj, niocount,
return -EPROTO;
}
-int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa,
- struct lov_stripe_md *lsm, obd_count oa_bufs,
- struct brw_page *pga, struct obd_trans_info *oti)
+int filter_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
+ obd_count oa_bufs, struct brw_page *pga,
+ struct obd_trans_info *oti)
{
struct obd_ioobj ioo;
struct niobuf_local *lnb;
GOTO(out, ret = -ENOMEM);
for (i = 0; i < oa_bufs; i++) {
- rnb[i].offset = pga[i].disk_offset;
+ lnb[i].page = pga[i].pg;
+ rnb[i].offset = pga[i].off;
rnb[i].len = pga[i].count;
}
- obdo_to_ioobj(oa, &ioo);
+ obdo_to_ioobj(oinfo->oi_oa, &ioo);
ioo.ioo_bufcnt = oa_bufs;
- ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti);
+ ret = filter_preprw(cmd, exp, oinfo->oi_oa, 1, &ioo,
+ oa_bufs, rnb, lnb, oti, oinfo_capa(oinfo));
if (ret != 0)
GOTO(out, ret);
- for (i = 0; i < oa_bufs; i++) {
- void *virt;
- obd_off off;
- void *addr;
-
- if (lnb[i].page == NULL)
- break;
-
- off = pga[i].disk_offset & ~PAGE_MASK;
- virt = kmap(pga[i].pg);
- addr = kmap(lnb[i].page);
-
- /* 2 kmaps == vanishingly small deadlock opportunity */
-
- if (cmd & OBD_BRW_WRITE)
- memcpy(addr + off, virt + off, pga[i].count);
- else
- memcpy(virt + off, addr + off, pga[i].count);
-
- kunmap(lnb[i].page);
- kunmap(pga[i].pg);
- }
-
- ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret);
+ ret = filter_commitrw(cmd, exp, oinfo->oi_oa, 1, &ioo,
+ oa_bufs, lnb, oti, ret);
out:
if (lnb)