* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*/
/*
/* Grab the dirty and seen grant announcements from the incoming obdo.
* We will later calculate the clients new grant and return it.
* Caller must hold osfs lock */
-static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
+void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
{
struct filter_export_data *fed;
struct obd_device *obd = exp->exp_obd;
obd->u.filter.fo_tot_granted -= oa->o_dropped;
fed->fed_grant -= oa->o_dropped;
fed->fed_dirty = oa->o_dirty;
+
+ if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_SHRINK_GRANT) {
+ obd_size left_space = filter_grant_space_left(exp);
+ struct filter_obd *filter = &exp->exp_obd->u.filter;
+
+ /*Only if left_space < fo_tot_clients * 32M,
+ *then the grant space could be shrinked */
+ if (left_space < filter->fo_tot_granted_clients *
+ FILTER_GRANT_SHRINK_LIMIT) {
+ fed->fed_grant -= oa->o_grant;
+ filter->fo_tot_granted -= oa->o_grant;
+ CDEBUG(D_CACHE, "%s: cli %s/%p shrink "LPU64
+ "fed_grant %ld total "LPU64"\n",
+ obd->obd_name, exp->exp_client_uuid.uuid,
+ exp, oa->o_grant, fed->fed_grant,
+ filter->fo_tot_granted);
+ oa->o_grant = 0;
+ }
+ }
+
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);
+ cfs_spin_unlock(&obd->obd_osfs_lock);
LBUG();
}
EXIT;
LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock);
- if (cfs_time_before_64(obd->obd_osfs_age, cfs_time_current_64() - HZ)) {
+ if (cfs_time_before_64(obd->obd_osfs_age,
+ cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
restat:
rc = fsfilt_statfs(obd, obd->u.obt.obt_sb,
- cfs_time_current_64() + HZ);
+ cfs_time_shift_64(OBD_STATFS_CACHE_SECONDS));
if (rc) /* N.B. statfs can't really fail */
RETURN(0);
statfs_done = 1;
/* Calculate how much grant space to allocate to this client, based on how
* much space is currently free and how much of that is already granted.
*
+ * if @conservative != 0, we limit the maximum grant to FILTER_GRANT_CHUNK;
+ * otherwise we'll satisfy the requested amount as possible as we can, this
+ * is usually due to client reconnect.
+ *
* Caller must hold obd_osfs_lock. */
long filter_grant(struct obd_export *exp, obd_size current_grant,
- obd_size want, obd_size fs_space_left)
+ obd_size want, obd_size fs_space_left, int conservative)
{
struct obd_device *obd = exp->exp_obd;
struct filter_export_data *fed = &exp->exp_filter_data;
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;
+ grant = min(want + (1 << blockbits) - 1, fs_space_left / 8);
+ grant &= ~((1ULL << blockbits) - 1);
if (grant) {
- /* Allow >FILTER_GRANT_CHUNK size when clients
- * reconnect due to a server reboot.
- */
- if ((grant > FILTER_GRANT_CHUNK) &&
- (!obd->obd_recovering))
+ if (grant > FILTER_GRANT_CHUNK && conservative)
grant = FILTER_GRANT_CHUNK;
obd->u.filter.fo_tot_granted += grant;
"current"LPU64"\n",
obd->obd_name, exp->exp_client_uuid.uuid,
exp, fed->fed_grant, want,current_grant);
- spin_unlock(&obd->obd_osfs_lock);
+ cfs_spin_unlock(&obd->obd_osfs_lock);
LBUG();
}
}
/*
* the routine is used to request pages from pagecache
*
- * use GFP_NOFS not allowing to enter FS as the client can run on this node
- * and we might end waiting on a page he sent in the request we're serving.
- *
+ * use GFP_NOFS for requests from a local client not allowing to enter FS
+ * as we might end up waiting on a page he sent in the request we're serving.
* use __GFP_HIGHMEM so that the pages can use all of the available memory
* on 32-bit machines
+ * use more aggressive GFP_HIGHUSER flags from non-local clients to be able to
+ * generate more memory pressure.
+ *
+ * See Bug 19529 and Bug 19917 for details.
*/
static struct page *filter_get_page(struct obd_device *obd,
struct inode *inode,
- obd_off offset)
+ obd_off offset,
+ int localreq)
{
struct page *page;
page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
- GFP_NOFS | __GFP_HIGHMEM);
+ (localreq ? (GFP_NOFS | __GFP_HIGHMEM)
+ : GFP_HIGHUSER));
if (unlikely(page == NULL))
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_NO_PAGE, 1);
}
/*
- * the function is used to free all pages used for request
- * just to mimic cacheless OSS which don't occupy much memory
+ * the invalidate above doesn't work during read because lnet pins pages.
+ * The truncate is used here instead to drop pages from cache
*/
-void filter_invalidate_cache(struct obd_device *obd, struct obd_ioobj *obj,
- struct niobuf_remote *nb, struct inode *inode)
+void filter_release_cache(struct obd_device *obd, struct obd_ioobj *obj,
+ struct niobuf_remote *rnb, struct inode *inode)
{
- struct niobuf_remote *rnb;
int i;
LASSERT(inode != NULL);
-
- for (i = 0, rnb = nb; i < obj->ioo_bufcnt; i++, rnb++) {
- obd_off start;
- obd_off end;
-
- start = rnb->offset >> CFS_PAGE_SHIFT;
- end = (rnb->offset + rnb->len) >> CFS_PAGE_SHIFT;
- invalidate_mapping_pages(inode->i_mapping, start, end);
- /* just to avoid warnings */
- start = 0;
- end = 0;
+ for (i = 0; i < obj->ioo_bufcnt; i++, rnb++) {
+#ifdef HAVE_TRUNCATE_RANGE
+ /* remove pages in which range is fit */
+ truncate_inode_pages_range(inode->i_mapping,
+ rnb->offset & CFS_PAGE_MASK,
+ (rnb->offset + rnb->len - 1) |
+ ~CFS_PAGE_MASK);
+#else
+ /* use invalidate for old kernels */
+ invalidate_mapping_pages(inode->i_mapping,
+ rnb->offset >> CFS_PAGE_SHIFT,
+ (rnb->offset + rnb->len) >>
+ CFS_PAGE_SHIFT);
+#endif
}
}
int rc = 0, i, tot_bytes = 0;
unsigned long now = jiffies;
long timediff;
+ loff_t isize;
ENTRY;
/* We are currently not supporting multi-obj BRW_READ RPCS at all.
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,
+ rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
CAPA_OPC_OSS_READ);
if (rc)
RETURN(rc);
if (oa && oa->o_valid & OBD_MD_FLGRANT) {
- spin_lock(&obd->obd_osfs_lock);
+ cfs_spin_lock(&obd->obd_osfs_lock);
filter_grant_incoming(exp, oa);
- oa->o_grant = 0;
- spin_unlock(&obd->obd_osfs_lock);
+ if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
+ !(oa->o_flags & OBD_FL_SHRINK_GRANT))
+ oa->o_grant = 0;
+ cfs_spin_unlock(&obd->obd_osfs_lock);
}
iobuf = filter_iobuf_get(&obd->u.filter, oti);
RETURN(PTR_ERR(iobuf));
push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
- dentry = filter_oa2dentry(obd, oa);
+ dentry = filter_oa2dentry(obd, &oa->o_oi);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
dentry = NULL;
}
inode = dentry->d_inode;
+ isize = i_size_read(inode);
obdo_to_inode(inode, oa, OBD_MD_FLATIME);
fsfilt_check_slow(obd, now, "preprw_read setup");
/* find pages for all segments, fill array with them */
- do_gettimeofday(&start);
+ cfs_gettimeofday(&start);
for (i = 0, lnb = res; i < *npages; i++, lnb++) {
lnb->dentry = dentry;
- if (i_size_read(inode) <= lnb->offset)
+ if (isize <= lnb->offset)
/* If there's no more data, abort early. lnb->rc == 0,
* so it's easy to detect later. */
break;
- lnb->page = filter_get_page(obd, inode, lnb->offset);
+ lnb->page = filter_get_page(obd, inode, lnb->offset, 0);
if (lnb->page == NULL)
GOTO(cleanup, rc = -ENOMEM);
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS, 1);
- if (i_size_read(inode) < lnb->offset + lnb->len - 1)
- lnb->rc = i_size_read(inode) - lnb->offset;
+ if (isize < lnb->offset + lnb->len - 1)
+ lnb->rc = isize - lnb->offset;
else
lnb->rc = lnb->len;
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_CACHE_MISS, 1);
filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
}
- do_gettimeofday(&end);
+ cfs_gettimeofday(&end);
timediff = cfs_timeval_sub(&end, &start, NULL);
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
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;
+ int resend = 0;
+
+ if ((oa->o_valid & OBD_MD_FLFLAGS) &&
+ (oa->o_flags & OBD_FL_RECOV_RESEND)) {
+ resend = 1;
+ CDEBUG(D_CACHE, "Recoverable resend arrived, skipping "
+ "accounting\n");
+ }
LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock);
if ((lnb[n].flags & OBD_BRW_FROM_GRANT) &&
(oa->o_valid & OBD_MD_FLGRANT)) {
- if (fed->fed_grant < used + bytes) {
+ if (resend) {
+ /* this is a recoverable resent */
+ lnb[n].flags |= OBD_BRW_GRANTED;
+ rc = 0;
+ continue;
+ } else if (fed->fed_grant < used + bytes) {
CDEBUG(D_CACHE,
"%s: cli %s/%p claims %ld+%d "
"GRANT, real grant %lu idx %d\n",
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);
+ cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
LBUG();
}
return rc;
void *iobuf;
obd_size left;
unsigned long now = jiffies, timediff;
- int rc = 0, i, tot_bytes = 0, cleanup_phase = 0;
+ int rc = 0, i, tot_bytes = 0, cleanup_phase = 0, localreq = 0;
ENTRY;
LASSERT(objcount == 1);
LASSERT(obj->ioo_bufcnt > 0);
- rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa,
+ rc = filter_auth_capa(exp, NULL, oa->o_seq, capa,
CAPA_OPC_OSS_WRITE);
if (rc)
RETURN(rc);
+ if (exp->exp_connection &&
+ exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)
+ localreq = 1;
+
push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
iobuf = filter_iobuf_get(&obd->u.filter, oti);
if (IS_ERR(iobuf))
GOTO(cleanup, rc = PTR_ERR(iobuf));
cleanup_phase = 1;
- dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr,
+ dentry = filter_fid2dentry(obd, NULL, obj->ioo_seq,
obj->ioo_id);
if (IS_ERR(dentry))
GOTO(cleanup, rc = PTR_ERR(dentry));
cleanup_phase = 2;
if (dentry->d_inode == NULL) {
- CERROR("%s: trying to BRW to non-existent file "LPU64"\n",
- obd->obd_name, obj->ioo_id);
- GOTO(cleanup, rc = -ENOENT);
+ if (exp->exp_obd->obd_recovering) {
+ struct obdo *noa = oa;
+
+ if (oa == NULL) {
+ OBDO_ALLOC(noa);
+ if (noa == NULL)
+ GOTO(recreate_out, rc = -ENOMEM);
+ noa->o_id = obj->ioo_id;
+ noa->o_valid = OBD_MD_FLID;
+ }
+
+ if (filter_create(exp, noa, NULL, oti) == 0) {
+ f_dput(dentry);
+ dentry = filter_fid2dentry(exp->exp_obd, NULL,
+ obj->ioo_seq,
+ obj->ioo_id);
+ }
+ if (oa == NULL)
+ OBDO_FREE(noa);
+ }
+ recreate_out:
+ if (IS_ERR(dentry) || dentry->d_inode == NULL) {
+ CERROR("%s: BRW to missing obj "LPU64"/"LPU64":rc %d\n",
+ exp->exp_obd->obd_name,
+ obj->ioo_id, obj->ioo_seq,
+ IS_ERR(dentry) ? (int)PTR_ERR(dentry) : -ENOENT);
+ if (IS_ERR(dentry))
+ cleanup_phase = 1;
+ GOTO(cleanup, rc = -ENOENT);
+ }
}
if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID) &&
dentry->d_inode->i_mode & (S_ISUID | S_ISGID)) {
- rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
+ rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
if (rc)
GOTO(cleanup, rc);
}
fsfilt_check_slow(obd, now, "preprw_write setup");
+ /* Filter truncate first locks i_mutex then partially truncated
+ * page, filter write code first locks pages then take
+ * i_mutex. To avoid a deadlock in case of concurrent
+ * punch/write requests from one client, filter writes and
+ * filter truncates are serialized by i_alloc_sem, allowing
+ * multiple writes or single truncate. */
+ down_read(&dentry->d_inode->i_alloc_sem);
+ fsfilt_check_slow(obd, now, "i_alloc_sem");
+
/* 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);
+ fmd = filter_fmd_find(exp, obj->ioo_id, obj->ioo_seq);
LASSERT(oa != NULL);
- spin_lock(&obd->obd_osfs_lock);
+ cfs_spin_lock(&obd->obd_osfs_lock);
filter_grant_incoming(exp, oa);
if (fmd && fmd->fmd_mactime_xid > oti->oti_xid)
oa->o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLCTIME |
/* 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(&obd->obd_osfs_lock);
+ /* If OBD_FL_SHRINK_GRANT is set, the client just returned us some grant
+ * so no sense in allocating it some more. We either return the grant
+ * back to the client if we have plenty of space or we don't return
+ * anything if we are short. This was decided in filter_grant_incoming*/
+ if ((oa->o_valid & OBD_MD_FLGRANT) &&
+ (!(oa->o_valid & OBD_MD_FLFLAGS) ||
+ !(oa->o_flags & OBD_FL_SHRINK_GRANT)))
+ oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty,
+ left, 1);
+
+ cfs_spin_unlock(&obd->obd_osfs_lock);
filter_fmd_put(exp, fmd);
+ OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, (obd_timeout + 1) / 4);
+
if (rc)
GOTO(cleanup, rc);
+ cleanup_phase = 4;
- do_gettimeofday(&start);
+ cfs_gettimeofday(&start);
for (i = 0, lnb = res; i < *npages; i++, lnb++) {
/* We still set up for ungranted pages so that granted pages
* needs to keep the pages all aligned properly. */
lnb->dentry = dentry;
- lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset);
+ lnb->page = filter_get_page(obd, dentry->d_inode, lnb->offset,
+ localreq);
if (lnb->page == NULL)
GOTO(cleanup, rc = -ENOMEM);
- cleanup_phase = 4;
/* DLM locking protects us from write and truncate competing
* for same region, but truncate can leave dirty page in the
* be able to proceed in filter_commitrw_write(). thus let's
* just wait for writeout completion, should be rare enough.
* -bzzz */
- if (obd->u.filter.fo_writethrough_cache)
- wait_on_page_writeback(lnb->page);
+ wait_on_page_writeback(lnb->page);
BUG_ON(PageWriteback(lnb->page));
/* If the filter writes a partial page, then has the file
if (lnb->rc == 0)
tot_bytes += lnb->len;
}
- do_gettimeofday(&end);
+ cfs_gettimeofday(&end);
timediff = cfs_timeval_sub(&end, &start, NULL);
lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_GET_PAGE, timediff);
fsfilt_check_slow(obd, now, "start_page_write");
+ lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
+ tot_bytes);
+
if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats)
lprocfs_counter_add(exp->exp_nid_stats->nid_stats,
LPROC_FILTER_WRITE_BYTES, tot_bytes);
}
}
case 3:
+ if (rc)
+ up_read(&dentry->d_inode->i_alloc_sem);
+
filter_iobuf_put(&obd->u.filter, iobuf, oti);
case 2:
pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
case 1:
filter_iobuf_put(&obd->u.filter, iobuf, oti);
case 0:
- spin_lock(&obd->obd_osfs_lock);
+ cfs_spin_lock(&obd->obd_osfs_lock);
if (oa)
filter_grant_incoming(exp, oa);
- spin_unlock(&obd->obd_osfs_lock);
+ cfs_spin_unlock(&obd->obd_osfs_lock);
pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
break;
default:;
int i;
ENTRY;
- osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
+ osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
/* 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 (resource != NULL) {
LDLM_RESOURCE_ADDREF(resource);
- ns->ns_lvbo->lvbo_update(resource, NULL, 0, 1);
+ ns->ns_lvbo->lvbo_update(resource, NULL, 1);
LDLM_RESOURCE_DELREF(resource);
ldlm_resource_putref(resource);
}
lnb->page = NULL;
}
}
-
if (inode && (fo->fo_read_cache == 0 ||
- i_size_read(inode) > fo->fo_readcache_max_filesize))
- filter_invalidate_cache(exp->exp_obd, obj, rnb, inode);
+ i_size_read(inode) > fo->fo_readcache_max_filesize))
+ filter_release_cache(exp->exp_obd, obj, rnb, inode);
if (res->dentry != NULL)
f_dput(res->dentry);
unsigned long pending = 0;
int i;
- spin_lock(&exp->exp_obd->obd_osfs_lock);
+ cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
for (i = 0, lnb = res; i < niocount; i++, lnb++)
pending += lnb->lnb_grant_used;
filter->fo_tot_pending, pending);
filter->fo_tot_pending -= pending;
- spin_unlock(&exp->exp_obd->obd_osfs_lock);
+ cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
}
int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
lnb[i].page = pga[i].pg;
rnb[i].offset = pga[i].off;
rnb[i].len = pga[i].count;
+ lnb[i].flags = rnb[i].flags = pga[i].flag;
}
obdo_to_ioobj(oinfo->oi_oa, &ioo);