Message-ID: <3b0611ff-8d8d-a95b-d37f-f43a8d2bdb82@gmx.com>
Date:   Fri, 24 Jun 2022 21:50:58 +0800
From:   Qu Wenruo <quwenruo.btrfs@....com>
To:     dsterba@...e.cz, Christoph Hellwig <hch@....de>, clm@...com,
        josef@...icpanda.com, dsterba@...e.com,
        linux-btrfs@...r.kernel.org,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux Memory Management List <linux-mm@...ck.org>
Subject: Re: [PATCH] btrfs: remove btrfs_writepage_cow_fixup



On 2022/6/24 21:27, David Sterba wrote:
> On Fri, Jun 24, 2022 at 09:12:44PM +0800, Qu Wenruo wrote:
>> On 2022/6/24 20:49, David Sterba wrote:
>>> On Fri, Jun 24, 2022 at 02:23:34PM +0200, Christoph Hellwig wrote:
>>>> Since the page_mkwrite address space operation was added, starting with
>>>> commit 9637a5efd4fb ("[PATCH] add page_mkwrite() vm_operations method")
>>>> in 2006, the kernel does not just dirty random pages without telling
>>>> the file system.
>>>
>>> It does and there's a history behind the fixup worker. tl;dr it can't be
>>> removed, though every now and then somebody comes and tries to.
>>>
>>> On s390 the page status is tracked in two places, hw and in memory and
>>> this needs to be synchronized manually.
>>>
>>> On x86_64 it's not a simple reason but it happens as well in some edge
>>> case where the mappings get removed and dirty page is set deep in the
>>> arch mm code.  We've been chasing it long time ago, I don't recall exact
>>> details and it's been a painful experience.
>>>
>>> If there's been any change on the s390 side or in arch/x86/mm code I
>>> don't know but to be on the safe side, I strongly assume the fixup code
>>> is needed unless proven otherwise.
>>
>> I'd say, if this can be a problem to btrfs, then all fs supporting COW
>> should also be affected, and should have similar workaround.
>
> Probably yes.
>
>> Furthermore, this means we can get a page dirtied without us knowing.
>
> This should not happen because we do have the detection of the page and
> extent state mismatch and the fixup worker makes things right again.
>
>> This is a super big surprise to any fs, and should be properly
>> documented, not just leaving some seemly dead and special code in some
>> random fs.
>
> You seem to be a non-believer that the bug is real and calling the code
> dead.

Nope, as Jan mentioned RDMA, it immediately light the bulb in my head.

Now I'm totally convinced we can have page marked dirty without proper
notification to fs.

So now I think this is real bug.

But unfortunately the code itself has no concrete reasons on which cases
this can happen, just mentioned kernel can mark page dirty (seemly
randomly).

Thus it "looks like" a dead code.


> Each filesystem should validate the implementation agains the
> platform where it is and btrfs once found the hard way that there are
> some corner cases where structures get out of sync.

In fact, from the fs point of view, there are quite some expectation on
its interfaces, if there is a surprise and such problem is no long
really specific to btrfs, then it should be addressed more generically.

>
>> Furthermore, I'm not sure even if handling this in a fs level is correct.
>> This looks like more a MM problem to me then.
>>
>>
>> I totally understand it's a pain to debug such lowlevel bug, but
>> shouldn't we have a proper regression for it then?
>
> The regression test is generic/208 and it was not reliable at all, it
> fired randomly once a week or month, there used to be a BUG() in the
> fixup worker callback.

And it doesn't have any comment even related to this unexpected dirty pages.

>
>> Instead of just keeping what we know works, I really want to handle this
>> old case/bug in a more modern way.
>
> As long as the guarantees stay the same, then fine. We need to be able
> to detect the unexpected dirty bit and have a way to react to it.
>
> f4b1363cae43 ("btrfs: do not do delalloc reservation under page lock")
> 25f3c5021985 ("Btrfs: keep pages dirty when using btrfs_writepage_fixup_worker")
> 1d53c9e67230 ("Btrfs: only associate the locked page with one async_chunk struct")
>
> And the commit that fixed it:
>
> 87826df0ec36 ("btrfs: delalloc for page dirtied out-of-band in fixup worker")
>
> You can find several reports in the mailing list archives (search term
> btrfs_writepage_fixup_worker):

To me, a proper and modern solution is not to rely on super old reports
(although they are definitely helpful as a record), but proper explanation.

Thanks to Jan, RDMA would be a very direct example for this.

Although personally speaking, I still think we should limit on who can
set a page from page cache dirty.
(AKA, ensuring fs receives notification on every dirtied page)

Thanks,
Qu

>
> https://lore.kernel.org/linux-btrfs/1295053074.15265.6.camel@mercury.localdomain
>
> https://lore.kernel.org/linux-btrfs/20110701174436.GA8352@yahoo.fr
>
> https://lore.kernel.org/linux-btrfs/j0k65i$29a$1@dough.gmane.org
>
> https://lore.kernel.org/linux-btrfs/CAO47_--H0+6bu4qQ2QA9gZcHvGVWO4QUGCAb3+9a5Kg3+23UiQ@mail.gmail.com
>
> https://lore.kernel.org/linux-btrfs/vqfmv8-9ch.ln1@hurikhan.ath.cx

Powered by blists - more mailing lists