[PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[Maxim Patlasov]

Changing size of a file on server and local update (fuse_write_update_size)
should be always protected by inode->i_mutex. Otherwise a race like this is
possible:

1. Process 'A' calls fallocate(2) to extend file (~FALLOC_FL_KEEP_SIZE).
fuse_file_fallocate() sends FUSE_FALLOCATE request to the server.
2. Process 'B' performs ordinary buffered write(2) with a length big enough
to extend the file beyond "offset + length" of fallocate call.
3. Process 'A' resumes execution of fuse_file_fallocate() and calls
fuse_write_update_size(inode, offset + length). But 'offset + length' was
obsoleted by write from previous step.

Signed-off-by: Maxim V. Patlasov <MPatlasov@parallels.com>
---
 fs/fuse/file.c |   10 ++++++----
 1 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/fs/fuse/file.c b/fs/fuse/file.c
index e570081..8dfbf7d 100644
--- a/fs/fuse/file.c
+++ b/fs/fuse/file.c
@@ -2470,14 +2470,16 @@ static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
 		.mode = mode
 	};
 	int err;
+	bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
+			   (mode & FALLOC_FL_PUNCH_HOLE);
 
 	if (fc->no_fallocate)
 		return -EOPNOTSUPP;
 
-	if (mode & FALLOC_FL_PUNCH_HOLE) {
+	if (lock_inode)
 		mutex_lock(&inode->i_mutex);
+	if (mode & FALLOC_FL_PUNCH_HOLE)
 		fuse_set_nowrite(inode);
-	}
 
 	req = fuse_get_req_nopages(fc);
 	if (IS_ERR(req)) {
@@ -2511,10 +2513,10 @@ static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
 	fuse_invalidate_attr(inode);
 
 out:
-	if (mode & FALLOC_FL_PUNCH_HOLE) {
+	if (mode & FALLOC_FL_PUNCH_HOLE)
 		fuse_release_nowrite(inode);
+	if (lock_inode)
 		mutex_unlock(&inode->i_mutex);
-	}
 
 	return err;
 }

Re: [PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[Brian Foster]

On 06/11/2013 06:59 AM, Maxim Patlasov wrote:
> Changing size of a file on server and local update (fuse_write_update_size)
> should be always protected by inode->i_mutex. Otherwise a race like this is
> possible:
> 
> 1. Process 'A' calls fallocate(2) to extend file (~FALLOC_FL_KEEP_SIZE).
> fuse_file_fallocate() sends FUSE_FALLOCATE request to the server.
> 2. Process 'B' performs ordinary buffered write(2) with a length big enough
> to extend the file beyond "offset + length" of fallocate call.
> 3. Process 'A' resumes execution of fuse_file_fallocate() and calls
> fuse_write_update_size(inode, offset + length). But 'offset + length' was
> obsoleted by write from previous step.
> 

Hi Maxim,

Doesn't fuse_write_update_size() already handle this particular case by
only ever extending the size?

Brian

> Signed-off-by: Maxim V. Patlasov <MPatlasov@parallels.com>
> ---
>  fs/fuse/file.c |   10 ++++++----
>  1 files changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> index e570081..8dfbf7d 100644
> --- a/fs/fuse/file.c
> +++ b/fs/fuse/file.c
> @@ -2470,14 +2470,16 @@ static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
>  		.mode = mode
>  	};
>  	int err;
> +	bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
> +			   (mode & FALLOC_FL_PUNCH_HOLE);
>  
>  	if (fc->no_fallocate)
>  		return -EOPNOTSUPP;
>  
> -	if (mode & FALLOC_FL_PUNCH_HOLE) {
> +	if (lock_inode)
>  		mutex_lock(&inode->i_mutex);
> +	if (mode & FALLOC_FL_PUNCH_HOLE)
>  		fuse_set_nowrite(inode);
> -	}
>  
>  	req = fuse_get_req_nopages(fc);
>  	if (IS_ERR(req)) {
> @@ -2511,10 +2513,10 @@ static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
>  	fuse_invalidate_attr(inode);
>  
>  out:
> -	if (mode & FALLOC_FL_PUNCH_HOLE) {
> +	if (mode & FALLOC_FL_PUNCH_HOLE)
>  		fuse_release_nowrite(inode);
> +	if (lock_inode)
>  		mutex_unlock(&inode->i_mutex);
> -	}
>  
>  	return err;
>  }
> 

Re: [PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[Anand Avati]

On 6/12/13 4:40 AM, Brian Foster wrote:
> On 06/11/2013 06:59 AM, Maxim Patlasov wrote:
>> Changing size of a file on server and local update (fuse_write_update_size)
>> should be always protected by inode->i_mutex. Otherwise a race like this is
>> possible:
>>
>> 1. Process 'A' calls fallocate(2) to extend file (~FALLOC_FL_KEEP_SIZE).
>> fuse_file_fallocate() sends FUSE_FALLOCATE request to the server.
>> 2. Process 'B' performs ordinary buffered write(2) with a length big enough
>> to extend the file beyond "offset + length" of fallocate call.
>> 3. Process 'A' resumes execution of fuse_file_fallocate() and calls
>> fuse_write_update_size(inode, offset + length). But 'offset + length' was
>> obsoleted by write from previous step.
>>
>
> Hi Maxim,
>
> Doesn't fuse_write_update_size() already handle this particular case by
> only ever extending the size?
>


As you say, fuse_write_update_size() does seem to protect against the 
case Maxim writes in the commit log.

However, there is still an issue with with truncate(shrinking_offset) 
and fallocate(growing_offset,~FALLOC_FL_KEEP_SIZE) racing, and changing 
inode size in opposing order between file server and in core ->i_size. 
Therefore, grabbing i_mutex is making fallocate and truncate atomic 
against each other.

I guess we just need an updated commit log, and same code change?

Avati

Re: [PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[Anand Avati]

On 6/11/13 3:59 AM, Maxim Patlasov wrote:

> -	if (mode & FALLOC_FL_PUNCH_HOLE) {
> +	if (lock_inode)
>   		mutex_lock(&inode->i_mutex);

> +	if (mode & FALLOC_FL_PUNCH_HOLE)
>   		fuse_set_nowrite(inode);
> -	}

Just for clarity, can you make the condition to check 
FALLOC_FL_PUNCH_HOLE and call to fuse_set_nowrite() nested within the 
larger if (lock_inode) { .. } block? fuse_set_nowrite() should not be 
called without i_mutex acquired. The current style of calling 
mutex_lock() and fuse_set_nowrite() in separate conditions can 
potentially cause bugs in the future if they were to get re-ordered due 
to some unrelated patch. Nesting them makes the relation more explicit 
and clear.

Thanks,
Avati

Re: [PATCH] fuse: hold i_mutex in fuse_file_fallocate()

[Maxim Patlasov]

Anand, Brian,

06/12/2013 11:04 PM, Anand Avati пишет:
> On 6/11/13 3:59 AM, Maxim Patlasov wrote:
>
>> -    if (mode & FALLOC_FL_PUNCH_HOLE) {
>> +    if (lock_inode)
>>           mutex_lock(&inode->i_mutex);
>
>> +    if (mode & FALLOC_FL_PUNCH_HOLE)
>>           fuse_set_nowrite(inode);
>> -    }
>
> Just for clarity, can you make the condition to check 
> FALLOC_FL_PUNCH_HOLE and call to fuse_set_nowrite() nested within the 
> larger if (lock_inode) { .. } block? fuse_set_nowrite() should not be 
> called without i_mutex acquired. The current style of calling 
> mutex_lock() and fuse_set_nowrite() in separate conditions can 
> potentially cause bugs in the future if they were to get re-ordered 
> due to some unrelated patch. Nesting them makes the relation more 
> explicit and clear.

Thanks a lot for review. I'll post updated patch soon.

Thanks,
Maxim