Hello,
here is the sixth iteration of my patches to improve filesystem freezing.
The change since last iteration is that filesystem can be frozen with open but
unlinked files. After some thinking, I've decided that the best way to handle
this is to block removal inside ->evict_inode() of each filesystem and use
fs-internal level of freeze protection for that (usually I've instrumented
filesystem's transaction system to use freeze protection). Handling
inside VFS would be less work but the only level of freeze protection that
has a chance of not causing deadlocks is the one used for page faults and even
there it's not clear lock ordering would be correct wrt some fs-specific locks.
I've converted ext2, ext4, btrfs, xfs, nilfs2, ocfs2, gfs2 and also checked
that ext3, reiserfs, jfs should work as well (they have their internal freeze
protection mechanisms, possibly they could be replaced by a generic one but
given these are mostly aging filesystems, it's not a real priority IHMO).
So finally I'm not aware of any pending issue with this patch set so if you
have some concern, please speak up!
Introductory text to first time readers:
Filesystem freezing is currently racy and thus we can end up with dirty data on
frozen filesystem (see changelog patch 13 for detailed race description). This
patch series aims at fixing this.
To be able to block all places where inodes get dirtied, I've moved filesystem
file_update_time() call to ->page_mkwrite callback (patches 01-07) and put
freeze handling in mnt_want_write() / mnt_drop_write(). That however required
some code shuffling and changes to kern_path_create() (see patches 09-12). I
think the result is OK but opinions may differ ;). The advantage of this change
also is that all filesystems get freeze protection almost for free - even ext2
can handle freezing well now.
Another potential contention point might be patch 19. In that patch we make
freeze_super() refuse to freeze the filesystem when there are open but unlinked
files which may be impractical in some cases. The main reason for this is the
problem with handling of file deletion from fput() called with mmap_sem held
(e.g. from munmap(2)), and then there's the fact that we cannot really force
such filesystem into a consistent state... But if people think that freezing
with open but unlinked files should happen, then I have some possible
solutions in mind (maybe as a separate patchset since this is large enough).
I'm not able to hit any deadlocks, lockdep warnings, or dirty data on frozen
filesystem despite beating it with fsstress and bash-shared-mapping while
freezing and unfreezing for several hours (using ext4 and xfs) so I'm
reasonably confident this could finally be the right solution.
Changes since v5:
* handle unlinked & open files on frozen filesystem
* lockdep keys for freeze protection are now per filesystem type
* taught lockdep that freeze protection at lower level does not create
dependency when we already hold freeze protection at higher level
* rebased on 3.5-rc1-ish
Changes since v4:
* added a couple of Acked-by's
* added some comments & doc update
* added patches from series "Push file_update_time() into .page_mkwrite"
since it doesn't make much sense to keep them separate anymore
* rebased on top of 3.4-rc2
Changes since v3:
* added third level of freezing for fs internal purposes - hooked some
filesystems to use it (XFS, nilfs2)
* removed racy i_size check from filemap_mkwrite()
Changes since v2:
* completely rewritten
* freezing is now blocked at VFS entry points
* two stage freezing to handle both mmapped writes and other IO
The biggest changes since v1:
* have two counters to provide safe state transitions for SB_FREEZE_WRITE
and SB_FREEZE_TRANS states
* use percpu counters instead of own percpu structure
* added documentation fixes from the old fs freezing series
* converted XFS to use SB_FREEZE_TRANS counter instead of its private
m_active_trans counter
Honza
CC: Alex Elder <[email protected]>
CC: Anton Altaparmakov <[email protected]>
CC: Ben Myers <[email protected]>
CC: Chris Mason <[email protected]>
CC: [email protected]
CC: "David S. Miller" <[email protected]>
CC: [email protected]
CC: "J. Bruce Fields" <[email protected]>
CC: Joel Becker <[email protected]>
CC: KONISHI Ryusuke <[email protected]>
CC: [email protected]
CC: [email protected]
CC: [email protected]
CC: [email protected]
CC: [email protected]
CC: Mark Fasheh <[email protected]>
CC: Miklos Szeredi <[email protected]>
CC: [email protected]
CC: OGAWA Hirofumi <[email protected]>
CC: Steven Whitehouse <[email protected]>
CC: "Theodore Ts'o" <[email protected]>
CC: [email protected]
We remove most of frozen checks since upper layer takes care of blocking all
writes. We have to handle protection in ext4_page_mkwrite() in a special way
because we cannot use generic block_page_mkwrite(). Also we add a freeze
protection to ext4_evict_inode() so that iput() of unlinked inode cannot modify
a frozen filesystem (we cannot easily instrument ext4_journal_start() /
ext4_journal_stop() with freeze protection because we are missing the
superblock pointer in ext4_journal_stop() in nojournal mode).
CC: [email protected]
CC: "Theodore Ts'o" <[email protected]>
BugLink: https://bugs.launchpad.net/bugs/897421
Tested-by: Kamal Mostafa <[email protected]>
Tested-by: Peter M. Petrakis <[email protected]>
Tested-by: Dann Frazier <[email protected]>
Tested-by: Massimo Morana <[email protected]>
Signed-off-by: Jan Kara <[email protected]>
---
fs/ext4/inode.c | 15 ++++++++++-----
fs/ext4/mmp.c | 14 ++++++++++----
fs/ext4/super.c | 31 +++++++------------------------
3 files changed, 27 insertions(+), 33 deletions(-)
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 07eaf56..4884127 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -166,6 +166,11 @@ void ext4_evict_inode(struct inode *inode)
if (is_bad_inode(inode))
goto no_delete;
+ /*
+ * Protect us against freezing - iput() caller didn't have to have any
+ * protection against it
+ */
+ sb_start_intwrite(inode->i_sb);
handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
if (IS_ERR(handle)) {
ext4_std_error(inode->i_sb, PTR_ERR(handle));
@@ -175,6 +180,7 @@ void ext4_evict_inode(struct inode *inode)
* cleaned up.
*/
ext4_orphan_del(NULL, inode);
+ sb_end_intwrite(inode->i_sb);
goto no_delete;
}
@@ -206,6 +212,7 @@ void ext4_evict_inode(struct inode *inode)
stop_handle:
ext4_journal_stop(handle);
ext4_orphan_del(NULL, inode);
+ sb_end_intwrite(inode->i_sb);
goto no_delete;
}
}
@@ -234,6 +241,7 @@ void ext4_evict_inode(struct inode *inode)
else
ext4_free_inode(handle, inode);
ext4_journal_stop(handle);
+ sb_end_intwrite(inode->i_sb);
return;
no_delete:
ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
@@ -4606,11 +4614,7 @@ int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
get_block_t *get_block;
int retries = 0;
- /*
- * This check is racy but catches the common case. We rely on
- * __block_page_mkwrite() to do a reliable check.
- */
- vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
+ sb_start_pagefault(inode->i_sb);
/* Delalloc case is easy... */
if (test_opt(inode->i_sb, DELALLOC) &&
!ext4_should_journal_data(inode) &&
@@ -4678,5 +4682,6 @@ retry_alloc:
out_ret:
ret = block_page_mkwrite_return(ret);
out:
+ sb_end_pagefault(inode->i_sb);
return ret;
}
diff --git a/fs/ext4/mmp.c b/fs/ext4/mmp.c
index ed6548d..4f63f90 100644
--- a/fs/ext4/mmp.c
+++ b/fs/ext4/mmp.c
@@ -10,14 +10,20 @@
* Write the MMP block using WRITE_SYNC to try to get the block on-disk
* faster.
*/
-static int write_mmp_block(struct buffer_head *bh)
+static int write_mmp_block(struct super_block *sb, struct buffer_head *bh)
{
+ /*
+ * We protect against freezing so that we don't create dirty buffers
+ * on frozen filesystem.
+ */
+ sb_start_write(sb);
mark_buffer_dirty(bh);
lock_buffer(bh);
bh->b_end_io = end_buffer_write_sync;
get_bh(bh);
submit_bh(WRITE_SYNC, bh);
wait_on_buffer(bh);
+ sb_end_write(sb);
if (unlikely(!buffer_uptodate(bh)))
return 1;
@@ -120,7 +126,7 @@ static int kmmpd(void *data)
mmp->mmp_time = cpu_to_le64(get_seconds());
last_update_time = jiffies;
- retval = write_mmp_block(bh);
+ retval = write_mmp_block(sb, bh);
/*
* Don't spew too many error messages. Print one every
* (s_mmp_update_interval * 60) seconds.
@@ -200,7 +206,7 @@ static int kmmpd(void *data)
mmp->mmp_seq = cpu_to_le32(EXT4_MMP_SEQ_CLEAN);
mmp->mmp_time = cpu_to_le64(get_seconds());
- retval = write_mmp_block(bh);
+ retval = write_mmp_block(sb, bh);
failed:
kfree(data);
@@ -299,7 +305,7 @@ skip:
seq = mmp_new_seq();
mmp->mmp_seq = cpu_to_le32(seq);
- retval = write_mmp_block(bh);
+ retval = write_mmp_block(sb, bh);
if (retval)
goto failed;
diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index 35b5954..cd6a516 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -290,33 +290,17 @@ static void ext4_put_nojournal(handle_t *handle)
* journal_end calls result in the superblock being marked dirty, so
* that sync() will call the filesystem's write_super callback if
* appropriate.
- *
- * To avoid j_barrier hold in userspace when a user calls freeze(),
- * ext4 prevents a new handle from being started by s_frozen, which
- * is in an upper layer.
*/
handle_t *ext4_journal_start_sb(struct super_block *sb, int nblocks)
{
journal_t *journal;
- handle_t *handle;
trace_ext4_journal_start(sb, nblocks, _RET_IP_);
if (sb->s_flags & MS_RDONLY)
return ERR_PTR(-EROFS);
+ WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
journal = EXT4_SB(sb)->s_journal;
- handle = ext4_journal_current_handle();
-
- /*
- * If a handle has been started, it should be allowed to
- * finish, otherwise deadlock could happen between freeze
- * and others(e.g. truncate) due to the restart of the
- * journal handle if the filesystem is forzen and active
- * handles are not stopped.
- */
- if (!handle)
- vfs_check_frozen(sb, SB_FREEZE_TRANS);
-
if (!journal)
return ext4_get_nojournal();
/*
@@ -2633,6 +2617,7 @@ static int ext4_run_li_request(struct ext4_li_request *elr)
sb = elr->lr_super;
ngroups = EXT4_SB(sb)->s_groups_count;
+ sb_start_write(sb);
for (group = elr->lr_next_group; group < ngroups; group++) {
gdp = ext4_get_group_desc(sb, group, NULL);
if (!gdp) {
@@ -2659,6 +2644,7 @@ static int ext4_run_li_request(struct ext4_li_request *elr)
elr->lr_next_sched = jiffies + elr->lr_timeout;
elr->lr_next_group = group + 1;
}
+ sb_end_write(sb);
return ret;
}
@@ -4135,10 +4121,8 @@ int ext4_force_commit(struct super_block *sb)
return 0;
journal = EXT4_SB(sb)->s_journal;
- if (journal) {
- vfs_check_frozen(sb, SB_FREEZE_TRANS);
+ if (journal)
ret = ext4_journal_force_commit(journal);
- }
return ret;
}
@@ -4170,9 +4154,8 @@ static int ext4_sync_fs(struct super_block *sb, int wait)
* gives us a chance to flush the journal completely and mark the fs clean.
*
* Note that only this function cannot bring a filesystem to be in a clean
- * state independently, because ext4 prevents a new handle from being started
- * by @sb->s_frozen, which stays in an upper layer. It thus needs help from
- * the upper layer.
+ * state independently. It relies on upper layer to stop all data & metadata
+ * modifications.
*/
static int ext4_freeze(struct super_block *sb)
{
@@ -4199,7 +4182,7 @@ static int ext4_freeze(struct super_block *sb)
EXT4_CLEAR_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_RECOVER);
error = ext4_commit_super(sb, 1);
out:
- /* we rely on s_frozen to stop further updates */
+ /* we rely on upper layer to stop further updates */
jbd2_journal_unlock_updates(EXT4_SB(sb)->s_journal);
return error;
}
--
1.7.1
On Sat, Jun 02, 2012 at 12:30:14AM +0200, Jan Kara wrote:
> Hello,
>
> here is the sixth iteration of my patches to improve filesystem freezing.
> The change since last iteration is that filesystem can be frozen with open but
> unlinked files. After some thinking, I've decided that the best way to handle
> this is to block removal inside ->evict_inode() of each filesystem and use
> fs-internal level of freeze protection for that (usually I've instrumented
> filesystem's transaction system to use freeze protection). Handling
> inside VFS would be less work but the only level of freeze protection that
> has a chance of not causing deadlocks is the one used for page faults and even
> there it's not clear lock ordering would be correct wrt some fs-specific locks.
> I've converted ext2, ext4, btrfs, xfs, nilfs2, ocfs2, gfs2 and also checked
> that ext3, reiserfs, jfs should work as well (they have their internal freeze
> protection mechanisms, possibly they could be replaced by a generic one but
> given these are mostly aging filesystems, it's not a real priority IHMO).
> So finally I'm not aware of any pending issue with this patch set so if you
> have some concern, please speak up!
Could you rebase on top of e.g. -rc2 and repost?
On Sat, Jun 02, 2012 at 12:30:31AM +0200, Jan Kara wrote:
> We remove most of frozen checks since upper layer takes care of blocking all
> writes. We have to handle protection in ext4_page_mkwrite() in a special way
> because we cannot use generic block_page_mkwrite(). Also we add a freeze
> protection to ext4_evict_inode() so that iput() of unlinked inode cannot modify
> a frozen filesystem (we cannot easily instrument ext4_journal_start() /
> ext4_journal_stop() with freeze protection because we are missing the
> superblock pointer in ext4_journal_stop() in nojournal mode).
>
> CC: [email protected]
> CC: "Theodore Ts'o" <[email protected]>
> BugLink: https://bugs.launchpad.net/bugs/897421
> Tested-by: Kamal Mostafa <[email protected]>
> Tested-by: Peter M. Petrakis <[email protected]>
> Tested-by: Dann Frazier <[email protected]>
> Tested-by: Massimo Morana <[email protected]>
> Signed-off-by: Jan Kara <[email protected]>
Acked-by: "Theodore Ts'o" <[email protected]>
- Ted