aboutsummaryrefslogtreecommitdiff
path: root/fs/ocfs2
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ocfs2')
-rw-r--r--fs/ocfs2/aops.c46
-rw-r--r--fs/ocfs2/aops.h4
-rw-r--r--fs/ocfs2/extent_map.c6
-rw-r--r--fs/ocfs2/file.c86
-rw-r--r--fs/ocfs2/journal.c8
-rw-r--r--fs/ocfs2/uptodate.c4
-rw-r--r--fs/ocfs2/vote.c6
7 files changed, 116 insertions, 44 deletions
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 0d858d0b25b..47152bf9a7f 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -276,13 +276,29 @@ static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
return ret;
}
+/* This can also be called from ocfs2_write_zero_page() which has done
+ * it's own cluster locking. */
+int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
+ unsigned from, unsigned to)
+{
+ int ret;
+
+ down_read(&OCFS2_I(inode)->ip_alloc_sem);
+
+ ret = block_prepare_write(page, from, to, ocfs2_get_block);
+
+ up_read(&OCFS2_I(inode)->ip_alloc_sem);
+
+ return ret;
+}
+
/*
* ocfs2_prepare_write() can be an outer-most ocfs2 call when it is called
* from loopback. It must be able to perform its own locking around
* ocfs2_get_block().
*/
-int ocfs2_prepare_write(struct file *file, struct page *page,
- unsigned from, unsigned to)
+static int ocfs2_prepare_write(struct file *file, struct page *page,
+ unsigned from, unsigned to)
{
struct inode *inode = page->mapping->host;
int ret;
@@ -295,11 +311,7 @@ int ocfs2_prepare_write(struct file *file, struct page *page,
goto out;
}
- down_read(&OCFS2_I(inode)->ip_alloc_sem);
-
- ret = block_prepare_write(page, from, to, ocfs2_get_block);
-
- up_read(&OCFS2_I(inode)->ip_alloc_sem);
+ ret = ocfs2_prepare_write_nolock(inode, page, from, to);
ocfs2_meta_unlock(inode, 0);
out:
@@ -625,11 +637,31 @@ static ssize_t ocfs2_direct_IO(int rw,
int ret;
mlog_entry_void();
+
+ /*
+ * We get PR data locks even for O_DIRECT. This allows
+ * concurrent O_DIRECT I/O but doesn't let O_DIRECT with
+ * extending and buffered zeroing writes race. If they did
+ * race then the buffered zeroing could be written back after
+ * the O_DIRECT I/O. It's one thing to tell people not to mix
+ * buffered and O_DIRECT writes, but expecting them to
+ * understand that file extension is also an implicit buffered
+ * write is too much. By getting the PR we force writeback of
+ * the buffered zeroing before proceeding.
+ */
+ ret = ocfs2_data_lock(inode, 0);
+ if (ret < 0) {
+ mlog_errno(ret);
+ goto out;
+ }
+ ocfs2_data_unlock(inode, 0);
+
ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
inode->i_sb->s_bdev, iov, offset,
nr_segs,
ocfs2_direct_IO_get_blocks,
ocfs2_dio_end_io);
+out:
mlog_exit(ret);
return ret;
}
diff --git a/fs/ocfs2/aops.h b/fs/ocfs2/aops.h
index d40456d509a..e88c3f0b8fa 100644
--- a/fs/ocfs2/aops.h
+++ b/fs/ocfs2/aops.h
@@ -22,8 +22,8 @@
#ifndef OCFS2_AOPS_H
#define OCFS2_AOPS_H
-int ocfs2_prepare_write(struct file *file, struct page *page,
- unsigned from, unsigned to);
+int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
+ unsigned from, unsigned to);
struct ocfs2_journal_handle *ocfs2_start_walk_page_trans(struct inode *inode,
struct page *page,
diff --git a/fs/ocfs2/extent_map.c b/fs/ocfs2/extent_map.c
index 4601fc256f1..1a5c69071df 100644
--- a/fs/ocfs2/extent_map.c
+++ b/fs/ocfs2/extent_map.c
@@ -569,7 +569,7 @@ static int ocfs2_extent_map_insert(struct inode *inode,
ret = -ENOMEM;
ctxt.new_ent = kmem_cache_alloc(ocfs2_em_ent_cachep,
- GFP_KERNEL);
+ GFP_NOFS);
if (!ctxt.new_ent) {
mlog_errno(ret);
return ret;
@@ -583,14 +583,14 @@ static int ocfs2_extent_map_insert(struct inode *inode,
if (ctxt.need_left && !ctxt.left_ent) {
ctxt.left_ent =
kmem_cache_alloc(ocfs2_em_ent_cachep,
- GFP_KERNEL);
+ GFP_NOFS);
if (!ctxt.left_ent)
break;
}
if (ctxt.need_right && !ctxt.right_ent) {
ctxt.right_ent =
kmem_cache_alloc(ocfs2_em_ent_cachep,
- GFP_KERNEL);
+ GFP_NOFS);
if (!ctxt.right_ent)
break;
}
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index 581eb451a41..a9559c87453 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -613,7 +613,8 @@ leave:
/* Some parts of this taken from generic_cont_expand, which turned out
* to be too fragile to do exactly what we need without us having to
- * worry about recursive locking in ->commit_write(). */
+ * worry about recursive locking in ->prepare_write() and
+ * ->commit_write(). */
static int ocfs2_write_zero_page(struct inode *inode,
u64 size)
{
@@ -641,7 +642,7 @@ static int ocfs2_write_zero_page(struct inode *inode,
goto out;
}
- ret = ocfs2_prepare_write(NULL, page, offset, offset);
+ ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
if (ret < 0) {
mlog_errno(ret);
goto out_unlock;
@@ -695,13 +696,26 @@ out:
return ret;
}
+/*
+ * A tail_to_skip value > 0 indicates that we're being called from
+ * ocfs2_file_aio_write(). This has the following implications:
+ *
+ * - we don't want to update i_size
+ * - di_bh will be NULL, which is fine because it's only used in the
+ * case where we want to update i_size.
+ * - ocfs2_zero_extend() will then only be filling the hole created
+ * between i_size and the start of the write.
+ */
static int ocfs2_extend_file(struct inode *inode,
struct buffer_head *di_bh,
- u64 new_i_size)
+ u64 new_i_size,
+ size_t tail_to_skip)
{
int ret = 0;
u32 clusters_to_add;
+ BUG_ON(!tail_to_skip && !di_bh);
+
/* setattr sometimes calls us like this. */
if (new_i_size == 0)
goto out;
@@ -714,27 +728,44 @@ static int ocfs2_extend_file(struct inode *inode,
OCFS2_I(inode)->ip_clusters;
if (clusters_to_add) {
- ret = ocfs2_extend_allocation(inode, clusters_to_add);
+ /*
+ * protect the pages that ocfs2_zero_extend is going to
+ * be pulling into the page cache.. we do this before the
+ * metadata extend so that we don't get into the situation
+ * where we've extended the metadata but can't get the data
+ * lock to zero.
+ */
+ ret = ocfs2_data_lock(inode, 1);
if (ret < 0) {
mlog_errno(ret);
goto out;
}
- ret = ocfs2_zero_extend(inode, new_i_size);
+ ret = ocfs2_extend_allocation(inode, clusters_to_add);
if (ret < 0) {
mlog_errno(ret);
- goto out;
+ goto out_unlock;
}
- }
- /* No allocation required, we just use this helper to
- * do a trivial update of i_size. */
- ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
- if (ret < 0) {
- mlog_errno(ret);
- goto out;
+ ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
+ if (ret < 0) {
+ mlog_errno(ret);
+ goto out_unlock;
+ }
+ }
+
+ if (!tail_to_skip) {
+ /* We're being called from ocfs2_setattr() which wants
+ * us to update i_size */
+ ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
+ if (ret < 0)
+ mlog_errno(ret);
}
+out_unlock:
+ if (clusters_to_add) /* this is the only case in which we lock */
+ ocfs2_data_unlock(inode, 1);
+
out:
return ret;
}
@@ -793,7 +824,7 @@ int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
if (i_size_read(inode) > attr->ia_size)
status = ocfs2_truncate_file(inode, bh, attr->ia_size);
else
- status = ocfs2_extend_file(inode, bh, attr->ia_size);
+ status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
if (status < 0) {
if (status != -ENOSPC)
mlog_errno(status);
@@ -1049,21 +1080,12 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
if (!clusters)
break;
- ret = ocfs2_extend_allocation(inode, clusters);
+ ret = ocfs2_extend_file(inode, NULL, newsize, count);
if (ret < 0) {
if (ret != -ENOSPC)
mlog_errno(ret);
goto out;
}
-
- /* Fill any holes which would've been created by this
- * write. If we're O_APPEND, this will wind up
- * (correctly) being a noop. */
- ret = ocfs2_zero_extend(inode, (u64) newsize - count);
- if (ret < 0) {
- mlog_errno(ret);
- goto out;
- }
break;
}
@@ -1146,6 +1168,22 @@ static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
ocfs2_iocb_set_rw_locked(iocb);
}
+ /*
+ * We're fine letting folks race truncates and extending
+ * writes with read across the cluster, just like they can
+ * locally. Hence no rw_lock during read.
+ *
+ * Take and drop the meta data lock to update inode fields
+ * like i_size. This allows the checks down below
+ * generic_file_aio_read() a chance of actually working.
+ */
+ ret = ocfs2_meta_lock(inode, NULL, NULL, 0);
+ if (ret < 0) {
+ mlog_errno(ret);
+ goto bail;
+ }
+ ocfs2_meta_unlock(inode, 0);
+
ret = generic_file_aio_read(iocb, buf, count, iocb->ki_pos);
if (ret == -EINVAL)
mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index 6a610ae5358..eebc3cfa6be 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -117,7 +117,7 @@ struct ocfs2_journal_handle *ocfs2_alloc_handle(struct ocfs2_super *osb)
{
struct ocfs2_journal_handle *retval = NULL;
- retval = kcalloc(1, sizeof(*retval), GFP_KERNEL);
+ retval = kcalloc(1, sizeof(*retval), GFP_NOFS);
if (!retval) {
mlog(ML_ERROR, "Failed to allocate memory for journal "
"handle!\n");
@@ -870,9 +870,11 @@ static int ocfs2_force_read_journal(struct inode *inode)
if (p_blocks > CONCURRENT_JOURNAL_FILL)
p_blocks = CONCURRENT_JOURNAL_FILL;
+ /* We are reading journal data which should not
+ * be put in the uptodate cache */
status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
p_blkno, p_blocks, bhs, 0,
- inode);
+ NULL);
if (status < 0) {
mlog_errno(status);
goto bail;
@@ -982,7 +984,7 @@ static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
{
struct ocfs2_la_recovery_item *item;
- item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_KERNEL);
+ item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
if (!item) {
/* Though we wish to avoid it, we are in fact safe in
* skipping local alloc cleanup as fsck.ocfs2 is more
diff --git a/fs/ocfs2/uptodate.c b/fs/ocfs2/uptodate.c
index 04a684dfdd9..b8a00a79332 100644
--- a/fs/ocfs2/uptodate.c
+++ b/fs/ocfs2/uptodate.c
@@ -337,7 +337,7 @@ static void __ocfs2_set_buffer_uptodate(struct ocfs2_inode_info *oi,
(unsigned long long)oi->ip_blkno,
(unsigned long long)block, expand_tree);
- new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_KERNEL);
+ new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
if (!new) {
mlog_errno(-ENOMEM);
return;
@@ -349,7 +349,7 @@ static void __ocfs2_set_buffer_uptodate(struct ocfs2_inode_info *oi,
* has no way of tracking that. */
for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++) {
tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
- GFP_KERNEL);
+ GFP_NOFS);
if (!tree[i]) {
mlog_errno(-ENOMEM);
goto out_free;
diff --git a/fs/ocfs2/vote.c b/fs/ocfs2/vote.c
index 53049a20419..ee42765a855 100644
--- a/fs/ocfs2/vote.c
+++ b/fs/ocfs2/vote.c
@@ -586,7 +586,7 @@ static struct ocfs2_net_wait_ctxt *ocfs2_new_net_wait_ctxt(unsigned int response
{
struct ocfs2_net_wait_ctxt *w;
- w = kcalloc(1, sizeof(*w), GFP_KERNEL);
+ w = kcalloc(1, sizeof(*w), GFP_NOFS);
if (!w) {
mlog_errno(-ENOMEM);
goto bail;
@@ -749,7 +749,7 @@ static struct ocfs2_vote_msg * ocfs2_new_vote_request(struct ocfs2_super *osb,
BUG_ON(!ocfs2_is_valid_vote_request(type));
- request = kcalloc(1, sizeof(*request), GFP_KERNEL);
+ request = kcalloc(1, sizeof(*request), GFP_NOFS);
if (!request) {
mlog_errno(-ENOMEM);
} else {
@@ -1129,7 +1129,7 @@ static int ocfs2_handle_vote_message(struct o2net_msg *msg,
struct ocfs2_super *osb = data;
struct ocfs2_vote_work *work;
- work = kmalloc(sizeof(struct ocfs2_vote_work), GFP_KERNEL);
+ work = kmalloc(sizeof(struct ocfs2_vote_work), GFP_NOFS);
if (!work) {
status = -ENOMEM;
mlog_errno(status);
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-28 15:11:32 +0000