On 6/1/23 7:45 PM, Vivek Goyal wrote:
> On Thu, Jun 01, 2023 at 09:45:52AM +0800, Jingbo Xu wrote:
>>
>>
>> On 6/1/23 4:03 AM, Vivek Goyal wrote:
>>> On Mon, Apr 24, 2023 at 08:32:50PM +0800, Jingbo Xu wrote:
>>>> When range already got reclaimed by somebody else, return NULL so that
>>>> the caller could retry to allocate or reclaim another range, instead of
>>>> mistakenly returning the range already got reclaimed and reused by
>>>> others.
>>>>
>>>> Reported-by: Liu Jiang <gerry@xxxxxxxxxxxxxxxxx>
>>>> Fixes: 9a752d18c85a ("virtiofs: add logic to free up a memory range")
>>>> Signed-off-by: Jingbo Xu <jefflexu@xxxxxxxxxxxxxxxxx>
>>>
>>> Hi Jingbo,
>>>
>>> This patch looks correct to me.
>>>
>>> Are you able to reproduce the problem? Or you are fixing it based on
>>> code inspection?
>>
>> It's spotted by Liu Jiang during code review. Not tested yet.
>>
>>>
>>> How are you testing this? We don't have virtiofsd DAX implementation yet
>>> in rust virtiofsd yet.
>>>
>>> I am not sure how to test this chagne now. We had out of tree patches
>>> in qemu and now qemu has gotten rid of C version of virtiofsd so these
>>> patches might not even work now.
>>
>> Yeah this exception path may not be so easy to be tested as it is only
>> triggered in the race condition. I have the old branch (of qemu) with
>> support for DAX, and maybe I could try to reproduce the exception path
>> by configuring limited DAX window and heavy IO workload.
>
> That would be great. Please test it with really small DAX window size.
> Also put some pr_debug() statements to make sure you are hitting this
> particular path during testing.
I tried to reproduce it but failed. It seems the race is impossible
theoretically.
In theory, the race occurs when a freeable dmap is found in inode's
interval tree but found it is removed from the interval tree in the
second query.
However the above procedure is protected with
filemap_invalidate_lock(inode->i_mapping) held in
inode_inline_reclaim_one_dmap(). Given the dmap deletion operations
from inode's interval tree are all protected with
filemap_invalidate_lock(inode->i_mapping) held, e.g. inside
inode_inline_reclaim_one_dmap() and lookup_and_reclaim_dmap(), the above
race seems impossible then.
--
Thanks,
Jingbo
