Re: [PATCH v2] btrfs: do not wait for short bulk allocation

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

在 2024/3/29 02:27, David Sterba 写道:
> On Tue, Mar 26, 2024 at 09:16:46AM +1030, Qu Wenruo wrote:
>> [BUG]
>> There is a recent report that when memory pressure is high (including
>> cached pages), btrfs can spend most of its time on memory allocation in
>> btrfs_alloc_page_array() for compressed read/write.
>>
>> [CAUSE]
>> For btrfs_alloc_page_array() we always go alloc_pages_bulk_array(), and
>> even if the bulk allocation failed (fell back to single page
>> allocation) we still retry but with extra memalloc_retry_wait().
>>
>> If the bulk alloc only returned one page a time, we would spend a lot of
>> time on the retry wait.
>>
>> The behavior was introduced in commit 395cb57e8560 ("btrfs: wait between
>> incomplete batch memory allocations").
>>
>> [FIX]
>> Although the commit mentioned that other filesystems do the wait, it's
>> not the case at least nowadays.
>>
>> All the mainlined filesystems only call memalloc_retry_wait() if they
>> failed to allocate any page (not only for bulk allocation).
>> If there is any progress, they won't call memalloc_retry_wait() at all.
>>
>> For example, xfs_buf_alloc_pages() would only call memalloc_retry_wait()
>> if there is no allocation progress at all, and the call is not for
>> metadata readahead.
>>
>> So I don't believe we should call memalloc_retry_wait() unconditionally
>> for short allocation.
>>
>> This patch would only call memalloc_retry_wait() if failed to allocate
>> any page for tree block allocation (which goes with __GFP_NOFAIL and may
>> not need the special handling anyway), and reduce the latency for
>> btrfs_alloc_page_array().
>
> Is this saying that it can fail with GFP_NOFAIL?

I'd say no, but never say no to memory allocation failure.

>
>> Reported-by: Julian Taylor <julian.taylor@1und1.de>
>> Tested-by: Julian Taylor <julian.taylor@1und1.de>
>> Link: https://lore.kernel.org/all/8966c095-cbe7-4d22-9784-a647d1bf27c3@1und1.de/
>> Fixes: 395cb57e8560 ("btrfs: wait between incomplete batch memory allocations")
>> Reviewed-by: Filipe Manana <fdmanana@suse.com>
>> Signed-off-by: Qu Wenruo <wqu@suse.com>
>> ---
>> Changelog:
>> v2:
>> - Still use bulk allocation function
>>    Since alloc_pages_bulk_array() would fall back to single page
>>    allocation by itself, there is no need to go alloc_page() manually.
>>
>> - Update the commit message to indicate other fses do not call
>>    memalloc_retry_wait() unconditionally
>>    In fact, they only call it when they need to retry hard and can not
>>    really fail.
>> ---
>>   fs/btrfs/extent_io.c | 22 +++++++++-------------
>>   1 file changed, 9 insertions(+), 13 deletions(-)
>>
>> diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
>> index 7441245b1ceb..c96089b6f388 100644
>> --- a/fs/btrfs/extent_io.c
>> +++ b/fs/btrfs/extent_io.c
>> @@ -681,31 +681,27 @@ static void end_bbio_data_read(struct btrfs_bio *bbio)
>>   int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
>>   			   gfp_t extra_gfp)
>>   {
>> +	const gfp_t gfp = GFP_NOFS | extra_gfp;
>>   	unsigned int allocated;
>>
>>   	for (allocated = 0; allocated < nr_pages;) {
>>   		unsigned int last = allocated;
>>
>> -		allocated = alloc_pages_bulk_array(GFP_NOFS | extra_gfp,
>> -						   nr_pages, page_array);
>> +		allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
>> +		if (unlikely(allocated == last)) {
>> +			/* Can not fail, wait and retry. */
>> +			if (extra_gfp & __GFP_NOFAIL) {
>> +				memalloc_retry_wait(GFP_NOFS);
>
> Can this happen? alloc_pages_bulk_array() should not fail when
> GFP_NOFAIL is passed, there are two allocation phases in
> __alloc_pages_bulk() and if it falls back to __alloc_pages() + NOFAIL it
> will not fail ... so what's the point of the retry?

Yeah, that's also one of my concern.

Unlike other fses, btrfs utilizes NOFAIL for metadata memory allocation,
meanwhile others do not.
E.g. XFS always do the retry wait even the allocation does not got a
page allocated. (aka, another kind of NOFAIL).

If needed, I can drop the retry part completely.

>
> Anyway the whole thing with NOFAIL flag that's passed only from
> alloc_extent_buffer() could be made a bit more straightforward. The gfp
> flags replaced by a bool with 'nofail' semantics or 2 helpers one that
> is for normal use an the one for the special purpose.
>

Sure I can do extra cleanups on this.

Thanks,
Qu