Re: [PATCH v4 3/3] squashfs: implement readahead

[Phillip Lougher <phillip@squashfs.org.uk>]

On 03/06/2022 16:58, Marek Szyprowski wrote:
> Hi Matthew,
> 
> On 03.06.2022 17:29, Matthew Wilcox wrote:
>> On Fri, Jun 03, 2022 at 10:55:01PM +0800, Hsin-Yi Wang wrote:
>>> On Fri, Jun 3, 2022 at 10:10 PM Marek Szyprowski
>>> <m.szyprowski@samsung.com> wrote:
>>>> Hi Matthew,
>>>>
>>>> On 03.06.2022 14:59, Matthew Wilcox wrote:
>>>>> On Fri, Jun 03, 2022 at 02:54:21PM +0200, Marek Szyprowski wrote:
>>>>>> On 01.06.2022 12:39, Hsin-Yi Wang wrote:
>>>>>>> Implement readahead callback for squashfs. It will read datablocks
>>>>>>> which cover pages in readahead request. For a few cases it will
>>>>>>> not mark page as uptodate, including:
>>>>>>> - file end is 0.
>>>>>>> - zero filled blocks.
>>>>>>> - current batch of pages isn't in the same datablock or not enough in a
>>>>>>>       datablock.
>>>>>>> - decompressor error.
>>>>>>> Otherwise pages will be marked as uptodate. The unhandled pages will be
>>>>>>> updated by readpage later.
>>>>>>>
>>>>>>> Suggested-by: Matthew Wilcox <willy@infradead.org>
>>>>>>> Signed-off-by: Hsin-Yi Wang <hsinyi@chromium.org>
>>>>>>> Reported-by: Matthew Wilcox <willy@infradead.org>
>>>>>>> Reported-by: Phillip Lougher <phillip@squashfs.org.uk>
>>>>>>> Reported-by: Xiongwei Song <Xiongwei.Song@windriver.com>
>>>>>>> ---
>>>>>> This patch landed recently in linux-next as commit 95f7a26191de
>>>>>> ("squashfs: implement readahead"). I've noticed that it causes serious
>>>>>> issues on my test systems (various ARM 32bit and 64bit based boards).
>>>>>> The easiest way to observe is udev timeout 'waiting for /dev to be fully
>>>>>> populated' and prolonged booting time. I'm using squashfs for deploying
>>>>>> kernel modules via initrd. Reverting aeefca9dfae7 & 95f7a26191deon on
>>>>>> top of the next-20220603 fixes the issue.
>>>>> How large are these files?  Just a few kilobytes?
>>>> Yes, they are small, most of them are smaller than 16KB, some about
>>>> 128KB and a few about 256KB. I've sent a detailed list in private mail.
>>>>
>>> Hi Marek,
>>>
>>> Are there any obvious squashfs errors in dmesg? Did you enable
>>> CONFIG_SQUASHFS_FILE_DIRECT or CONFIG_SQUASHFS_FILE_CACHE?
>> I don't think it's an error problem.  I think it's a short file problem.
>>
>> As I understand the current code (and apologies for not keeping up
>> to date with how the patch is progressing), if the file is less than
>> msblk->block_size bytes, we'll leave all the pages as !uptodate, leaving
>> them to be brough uptodate by squashfs_read_folio().  So Marek is hitting
>> the worst case scenario where we re-read the entire block for each page
>> in it.  I think we have to handle this tail case in ->readahead().
> 
> I'm not sure if this is related to reading of small files. There are
> only 50 modules being loaded from squashfs volume. I did a quick test of
> reading the files.
> 
> Simple file read with this patch:
> 
> root@target:~# time find /initrd/ -type f | while read f; do cat $f
>   >/dev/null; done
> 
> real    0m5.865s
> user    0m2.362s
> sys     0m3.844s
> 
> Without:
> 
> root@target:~# time find /initrd/ -type f | while read f; do cat $f
>   >/dev/null; done
> 
> real    0m6.619s
> user    0m2.112s
> sys     0m4.827s
> 

It has been a four day holiday in the UK (Queen's Platinum Jubilee),
hence the delay in responding.

The above read use-case is sequential (only one thread/process),
whereas the use-case where the slow-down is observed may be
parallel (multiple threads/processes entering Squashfs).

The above sequential use-case if the small files are held in
fragments, will be exhibiting caching behaviour that will
ameliorate the case where the same block is being repeatedly
re-read for each page in it.  Because each time
Squashfs is re-entered handling only a single page, the
decompressed block will be found in the fragment
cache, eliminating a block decompression for each page.

In a parallel use-case the decompressed fragment block
may be being eliminated from the cache (by other reading
processes), hence forcing the block to be repeatedly
decompressed.

Hence the slow-down will be much more noticable with a
parallel use-case than a sequential use-case.  It also may
be why this slipped through testing, if the test cases
are purely sequential in nature.

So Matthew's previous comment is still the most likely
explanation for the slow-down.

Phillip

> Best regards