Message-Id: <20200114222550.3BA08A4040@d06av23.portsmouth.uk.ibm.com>
Date:   Wed, 15 Jan 2020 03:55:49 +0530
From:   Ritesh Harjani <riteshh@...ux.ibm.com>
To:     Jan Kara <jack@...e.cz>
Cc:     linux-ext4@...r.kernel.org, tytso@....edu
Subject: Re: [RFC 2/2] ext4: Fix stale data read issue with DIO read &
 ext4_page_mkwrite path

Hello Jan,

On 1/14/20 3:17 PM, Jan Kara wrote:
> On Mon 13-01-20 16:34:22, Ritesh Harjani wrote:
>> Currently there is a small race window where ext4 tries to allocate
>> a written block for mapped files and if DIO read is in progress, then
>> this may result into stale data read exposure problem.
>>
>> This patch fixes the mentioned issue by:
>> 1. For non-delalloc path, page_mkwrite will use unwritten blocks by
>>     default for extent based files.
>>
>> 2. For delalloc path, we check if DIO is in progress during writeback.
>>     If yes, then we use unwritten blocks method to avoid this race.
>>
>> Signed-off-by: Ritesh Harjani <riteshh@...ux.ibm.com>
>> ---
>>   fs/ext4/inode.c | 45 ++++++++++++++++++++++++++++++++-------------
>>   1 file changed, 32 insertions(+), 13 deletions(-)
>>
>> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
>> index d035acab5b2a..07f66782335b 100644
>> --- a/fs/ext4/inode.c
>> +++ b/fs/ext4/inode.c
>> @@ -1529,6 +1529,7 @@ struct mpage_da_data {
>>   	struct ext4_map_blocks map;
>>   	struct ext4_io_submit io_submit;	/* IO submission data */
>>   	unsigned int do_map:1;
>> +	bool dio_in_progress:1;
>>   };
>>   
>>   static void mpage_release_unused_pages(struct mpage_da_data *mpd,
>> @@ -2359,7 +2360,7 @@ static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
>>   			   EXT4_GET_BLOCKS_METADATA_NOFAIL |
>>   			   EXT4_GET_BLOCKS_IO_SUBMIT;
>>   	dioread_nolock = ext4_should_dioread_nolock(inode);
>> -	if (dioread_nolock)
>> +	if (dioread_nolock || mpd->dio_in_progress)
>>   		get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
>>   	if (map->m_flags & (1 << BH_Delay))
>>   		get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
>> @@ -2367,7 +2368,8 @@ static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
>>   	err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
>>   	if (err < 0)
>>   		return err;
>> -	if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
>> +	if ((dioread_nolock || mpd->dio_in_progress) &&
>> +	    (map->m_flags & EXT4_MAP_UNWRITTEN)) {
>>   		if (!mpd->io_submit.io_end->handle &&
>>   		    ext4_handle_valid(handle)) {
>>   			mpd->io_submit.io_end->handle = handle->h_rsv_handle;
>> @@ -2626,6 +2628,7 @@ static int ext4_writepages(struct address_space *mapping,
>>   	bool done;
>>   	struct blk_plug plug;
>>   	bool give_up_on_write = false;
>> +	bool dio_in_progress = false;
>>   
>>   	if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
>>   		return -EIO;
>> @@ -2680,15 +2683,6 @@ static int ext4_writepages(struct address_space *mapping,
>>   		ext4_journal_stop(handle);
>>   	}
>>   
>> -	if (ext4_should_dioread_nolock(inode)) {
>> -		/*
>> -		 * We may need to convert up to one extent per block in
>> -		 * the page and we may dirty the inode.
>> -		 */
>> -		rsv_blocks = 1 + ext4_chunk_trans_blocks(inode,
>> -						PAGE_SIZE >> inode->i_blkbits);
>> -	}
>> -
>>   	if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
>>   		range_whole = 1;
>>   
>> @@ -2712,6 +2706,26 @@ static int ext4_writepages(struct address_space *mapping,
>>   	done = false;
>>   	blk_start_plug(&plug);
>>   
>> +	/*
>> +	 * If DIO is in progress, then we use unwritten blocks for allocation.
>> +	 * This is to avoid a small window of race (stale read) with
>> +	 * ext4_page_mkwrite path in delalloc case & with DIO read in parallel.
>> +	 *
>> +	 * Let's check for i_dio_count after we have tagged pages for writeback.
>> +	 */
>> +	smp_mb__before_atomic();
>> +	dio_in_progress = !!atomic_read(&inode->i_dio_count);
>> +	mpd.dio_in_progress = dio_in_progress;
> 
> Two problems here:
> 
> 1) smp_mb__before_atomic() does not work with atomic_read(). This kind of
> barrier works only with read-modify-write kinds of atomic operations like
> atomic_inc(). See Documentation/atomic_t.txt for more details.

Yes, I was not 100% sure on that part. But thanks for confirmation.


> 
> 2) Even if the barrier worked, this is still too early for the check.
> Consider the following race:
> 
> Task 1 - flusher		Task 2 - dio read	Task 3 - fault
> ext4_writepages()
>    atomic_read(&inode->i_dio_count) -> 0
>    ...
> 				iomap_dio_rw()
> 				  inode_dio_begin()
> 				  filemap_write_and_wait_range()
> 				  ...
> 							ext4_page_mkwrite()
> 							  fills hole at index I
>    ...
>    mpage_prepare_extent_to_map()
>      finds dirty page at index I - tagging
>      not in effect because this is WB_SYNC_NONE
>      writeback so we look for PAGECACHE_TAG_DIRTY
>      mpage_map_and_submit_extent()
>        - allocates block for page I
> 				  ext4_iomap_begin()
> 				    finds block under offset I
> 				  submit_bio()
> 				    - reads stale data

My bad. So a fault at hole may set the page dirty bit, while
ext4_writepages may be in progress. And if we check for i_dio_count very
early, then we still end up exposing stale data.

> 
> And what I wanted to use to stop this race is page lock / page writeback
> bit on page 'I' because filemap_write_and_wait_range() called from
> iomap_dio_rw() ends up waiting for both if the page is seen as dirty. For

Yes, agreed here. I guess earlier I was thinking of simplifying it
by checking for DIO early on so that we could use the same type of
extent type mapping (unwritten/written) for ext4_writepages.


> this to work, you need to check inode->i_dio_count after you hold the page
> locks for written range - i.e., after mpage_prepare_extent_to_map(). And

Yes, I am thinking we should add that check in 
mpage_map_and_submit_extent() before the do while{} loop.
I guess we should keep extent type mapping common for a given io_end
type, since we have to do unwritten to written handling at the end
of IO transfer.
This can be done if we check for DIO in progress early inside
mpage_map_and_submit_extent(), because by then, we already have the page
lock in place.

But let me also check more on this.

> that means you always have to have rsv_blocks set when starting a
> transaction because you don't know in advance whether you'll need them or
> not.

Yes. Thanks for pointing.


-ritesh

Powered by blists - more mailing lists