Re: [PATCH] Fix do_move_pages_to_node() error handling

[Piotr Jaroszynski <pjaroszynski@nvidia.com>]

On 11/15/18 12:47 AM, Michal Hocko wrote:
> On Wed 14-11-18 17:04:37, Piotr Jaroszynski wrote:
>> On 11/14/18 1:22 PM, Michal Hocko wrote:
>>> On Wed 14-11-18 10:04:45, Piotr Jaroszynski wrote:
>>>> On 11/14/18 3:29 AM, Michal Hocko wrote:
>>>>> On Wed 14-11-18 08:34:15, Michal Hocko wrote:
>>>>>> On Tue 13-11-18 16:40:59, p.jaroszynski@gmail.com wrote:
>>>>>>> From: Piotr Jaroszynski <pjaroszynski@nvidia.com>
>>>>>>>
>>>>>>> migrate_pages() can return the number of pages that failed to migrate
>>>>>>> instead of 0 or an error code. If that happens, the positive return is
>>>>>>> treated as an error all the way up through the stack leading to the
>>>>>>> move_pages() syscall returning a positive number. I believe this
>>>>>>> regressed with commit a49bd4d71637 ("mm, numa: rework do_pages_move")
>>>>>>> that refactored a lot of this code.
>>>>>>
>>>>>> Yes this is correct.
>>>>>>
>>>>>>> Fix this by treating positive returns as success in
>>>>>>> do_move_pages_to_node() as that seems to most closely follow the
>>>>>>> previous code. This still leaves the question whether silently
>>>>>>> considering this case a success is the right thing to do as even the
>>>>>>> status of the pages will be set as if they were successfully migrated,
>>>>>>> but that seems to have been the case before as well.
>>>>>>
>>>>>> Yes, I believe the previous semantic was just wrong and we want to fix
>>>>>> it. Jan has already brought this up [1]. I believe we want to update the
>>>>>> documentation rather than restore the previous hazy semantic.
>>>>
>>>> That's probably fair although at least some code we have will have to be
>>>> updated as it just checks for non-zero returns from move_pages() and
>>>> assumes errno is set when that happens.
>>>
>>> Can you tell me more about your usecase plase? I definitely do not want
>>> to break any existing userspace. Making the syscall return code more
>>> reasonable is still attractive. So if this new semantic can work better
>>> for you it would be one argument more to keep it this way.
>>>  
>>
>> One of our APIs exposes a way to move a VA range to a GPU NUMA node or one of
>> the CPU NUMA nodes. The code keeps retrying move_pages() and relies on
>> the reported page status to decide whether each page is done, needs a
>> retry (EAGAIN or EBUSY) or possibly needs a fallback (EMEM).
>>
>> With the previous behaviour we would get a success, but the page status
>> would be reported incorrectly. That's bad as we skip the migration
>> without knowing about it.
> 
> Exactly.
> 
>> With the current code we get what we interpret as success as errno is 0,
>> but the page status is gargabe/untouched. That's also bad.
> 
> Agreed.
> 
>> The proposed solution adds a new case to handle, but it will just tell
>> us the page status is unusable and all we can do is just retry blindly.
>> If it was possible to plumb through the migration status for each page
>> accurately that would allow us to save redoing the call for pages that
>> actually worked. Perhaps we would need a special status for pages
>> skipped due to errors.
> 
> This would be possible but with this patch applied you should know how
> many pages to skip from the tail of the array.

At least in our case the node target is the same for all the pages so we
would just learn that all the pages failed to migrate as they would be
all batched together to the do_move_pages_to_node() call.

> 
>> But maybe this is all a tiny corner case short of the bug I hit (see
>> more below) and it's not worth thinking too much about.
>>
>>>>>> Just wondering, how have you found out? Is there any real application
>>>>>> failing because of the change or this is a result of some test?
>>>>
>>>> I have a test that creates a tmp file, mmaps it as shared, memsets the
>>>> memory and then attempts to move it to a different node. It used to
>>>> work, but now fails. I suspect the filesystem's migratepage() callback
>>>> regressed and will look into it next. So far I have only tested this on
>>>> powerpc with the xfs filesystem.
>>>
>>> I would be surprise if the rewor changed the migration behavior.
>>
>> It didn't, I tracked it down to the new fs/iomap.c code used by xfs not
>> being compatible with migrate_page_move_mapping() and prepared a perhaps
>> naive fix in [1].
> 
> I am not familiar with iomap code much TBH so I cannot really judge your
> fix.
> 

Christoph reviewed it already (thanks!) so it should be good after all.
But in its context, I wanted to ask about migrate_page_move_mapping()
page count checks that it was hitting. Is it true that the count checks
are to handle the case when a page might be temporarily pinned and hence
have the count too high temporarily? That would explain why it returns
EAGAIN in this case. But should having the count too low (what the bug
was hitting) be a fatal error with a WARN maybe? Or are there expected
cases where the count is too low temporarily too? I could send a patch
for that, but also just wanted to understand the expectations.