From: yebin <yebin@huaweicloud.com>
To: Jan Kara <jack@suse.cz>, Christian Brauner <brauner@kernel.org>
Cc: viro@zeniv.linux.org.uk, linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH] fs: hold freeze protection across writable regular file ->open()
Date: Sat, 21 Nov 2026 09:19:18 +0800 [thread overview]
Message-ID: <6B00F196.20601@huaweicloud.com> (raw)
In-Reply-To: <4czsdzcffhzpkcdf7zffpj2ow2oa2skb64f4brrbf6z2jqmbik@zkeuyyuyvcdp>
On 2026/11/20 20:56, Jan Kara wrote:
> On Fri 20-11-26 12:19:03, Christian Brauner wrote:
>> On Thu, Nov 12, 2026 at 09:52:45PM +0800, Ye Bin wrote:
>>> From: Ye Bin <yebin10@huawei.com>
>>>
>>> A writable open of a regular file reaches the filesystem's ->open()
>>> callback without holding any superblock write (freeze) protection.
>>> ->open() can nevertheless write to the filesystem: ext4_file_open() ->
>>> dquot_file_open() -> dquot_initialize() may acquire a dquot and write
>>> the quota file, which starts a jbd2 transaction. If the filesystem
>>> reaches SB_FREEZE_COMPLETE in that window, ext4_journal_check_start()
>>> hits
>>>
>>> WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
>>>
>>> and the open then sleeps in jbd2 until the filesystem is thawed.
>>>
>>> The hole was created over time:
>>>
>>> - Commit 907f4554e252 ("dquot: move dquot initialization responsibility
>>> into the filesystem") moved quota initialization into ->open() and
>>> introduced dquot_file_open(), making ->open() a potentially writing
>>> operation for writable opens.
>>>
>>> - Commit eb04c28288bb ("fs: Add freezing handling to mnt_want_write() /
>>> mnt_drop_write()") added sb_start_write() to mnt_want_write() but
>>> deliberately kept the write reference held by an open file on the
>>> non-freezing __mnt_want_write() variant, so the file's own write
>>> reference does not protect its ->open().
>>>
>>> - Commit 64894cf84327 ("simplify lookup_open()/atomic_open() - do the
>>> temporary mnt_want_write() early") moved the temporary write count
>>> into do_last() and dropped it before finish_open(), so a writable
>>> open that is not O_TRUNC reaches ->open() with no sb write count at
>>> all.
>>>
>>> Since then only O_TRUNC opens take mnt_want_write() in do_open().
>>>
>>> Fix this by taking mnt_want_write() for any writable open of a regular
>>> file and holding it across vfs_open()/->open(). Regular files are
>>> checked explicitly so that opening device nodes on a read-only mount is
>>> not newly rejected. Read-only opens are unaffected.
>>>
>>> Fixes: eb04c28288bb ("fs: Add freezing handling to mnt_want_write() / mnt_drop_write()")
>>> Signed-off-by: Ye Bin <yebin10@huawei.com>
>>> ---
>>> fs/namei.c | 16 ++++++++++++++--
>>> 1 file changed, 14 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/namei.c b/fs/namei.c
>>> index f573fd0aacaf..0ee4cf282cf5 100644
>>> --- a/fs/namei.c
>>> +++ b/fs/namei.c
>>> @@ -4906,6 +4906,7 @@ static int do_open(struct nameidata *nd,
>>> const struct mnt_idmap *idmap;
>>> int open_flag = op->open_flag;
>>> bool do_truncate;
>>> + bool got_write;
>>> int acc_mode;
>>> int error;
>>>
>>> @@ -4935,12 +4936,23 @@ static int do_open(struct nameidata *nd,
>>> return -ENOTDIR;
>>>
>>> do_truncate = false;
>>> + got_write = false;
>>> acc_mode = op->acc_mode;
>>> if (file->f_mode & FMODE_CREATED) {
>>> /* Don't check for write permission, don't truncate */
>>> open_flag &= ~O_TRUNC;
>>> acc_mode = 0;
>>> } else if (d_is_reg(nd->path.dentry) && open_flag & O_TRUNC) {
>>> + do_truncate = true;
>>> + }
>>> +
>>> + /*
>>> + * ->open() can write to the filesystem, e.g. when it initializes
>>> + * quota, so writable opens of regular files need freeze protection
>>> + * as well. O_TRUNC opens already need it for handle_truncate().
>>> + */
>>> + if (do_truncate ||
>>> + (d_is_reg(nd->path.dentry) && (file->f_mode & FMODE_WRITE))) {
>>> error = mnt_want_write(nd->path.mnt);
>>
>> No, I don't think that's the right thing to do. Every
>> open(O_WRONLY|O_RDWR) would now block for a regular file on a frozen
>> filesystem even if it never writes. That's not acceptable. Log files,
>> editors, backup tools, they all would hang rather than fail on write.
>> It's a drastic uapi change.
>>
>> The write in ->open() is from the quota stuff. So that should be a quota
>> fix and not penalize the whole VFS layer.
>
> Conceptually I agree with you but we cannot easily grab freeze protection
> from ->open() where we do quota stuff because VFS may but need not be already
> holding it.
>
> Somewhat paradoxical is that quota code will actually write anything only
> if we are creating quota structures for a new user but that never happens
> from ->open (it would happen already during vfs_create() in
> dentry_create()). It is just ext4 that unconditionally starts a transaction
> because it doesn't know the context... But ocfs2 actually does write if we
> are touching the first inode of this user so there the problem is real.
>
> Hrm, but I think we could just use sb_start_intwrite() in filesystem's
> .acquire_dquot methods and be done with that. That won't deadlock and it
> will silence the warnings / fix ocfs2 breakage. AFAICS only ext4 & ocfs2
> actually need anything done. Ye?
>
I checked the code of ocfs2 and it should be fine, because sb_start_intwrite()
is called when starting a transaction.
ocfs2_acquire_dquot
ocfs2_start_trans
sb_start_intwrite(osb->sb);
So, it should only require fixing the EXT4 file system. That should be enough.
In the ext4_evict_inode() process, calling dquot_initialize() and freezing the
file system concurrently may also trigger a warning.
ext4_evict_inode
dquot_initialize(inode);
if (!ext4_journal_current_handle())
sb_start_intwrite(inode->i_sb);
ext4_journal_start()
The same processing logic in ext4_acquire_dquot() and ext4_evict_inode() should
be fine.
> Honza
>
prev parent reply other threads:[~2026-11-21 1:19 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-12 13:52 [PATCH] fs: hold freeze protection across writable regular file ->open() Ye Bin
2026-11-12 17:56 ` Josef Bacik
2026-11-16 7:11 ` yebin
2026-11-20 11:19 ` Christian Brauner
2026-11-20 12:56 ` Jan Kara
2026-11-21 1:19 ` yebin [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=6B00F196.20601@huaweicloud.com \
--to=yebin@huaweicloud.com \
--cc=brauner@kernel.org \
--cc=jack@suse.cz \
--cc=linux-fsdevel@vger.kernel.org \
--cc=viro@zeniv.linux.org.uk \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox