[PATCH v11 00/25] mm/gup: track dma-pinned pages: FOLL_PIN

[John Hubbard]

On 12/21/19 2:08 AM, Leon Romanovsky wrote:
> On Fri, Dec 20, 2019 at 03:54:55PM -0800, John Hubbard wrote:
>> On 12/20/19 10:29 AM, Leon Romanovsky wrote:
>> ...
>>>> $ ./build.sh
>>>> $ build/bin/run_tests.py
>>>>
>>>> If you get things that far I think Leon can get a reproduction for you
>>>
>>> I'm not so optimistic about that.
>>>
>>
>> OK, I'm going to proceed for now on the assumption that I've got an overflow
>> problem that happens when huge pages are pinned. If I can get more information,
>> great, otherwise it's probably enough.
>>
>> One thing: for your repro, if you know the huge page size, and the system
>> page size for that case, that would really help. Also the number of pins per
>> page, more or less, that you'd expect. Because Jason says that only 2M huge
>> pages are used...
>>
>> Because the other possibility is that the refcount really is going negative,
>> likely due to a mismatched pin/unpin somehow.
>>
>> If there's not an obvious repro case available, but you do have one (is it easy
>> to repro, though?), then *if* you have the time, I could point you to a github
>> branch that reduces GUP_PIN_COUNTING_BIAS by, say, 4x, by applying this:
> 
> I'll see what I can do this Sunday.
> 

The other data point that might shed light on whether it's a mismatch (this only
works if the system is not actually crashing, though), is checking the new
vmstat items, like this:

$ grep foll_pin /proc/vmstat
nr_foll_pin_requested 16288188
nr_foll_pin_returned 16288188

...but OTOH, if you've got long-term pins, then those are *supposed* to be
mismatched, so it only really helps in between tests.

thanks,
-- 
John Hubbard
NVIDIA