Re: [RFC PATCH] mm/slub: fix a deadlock due to incomplete patching of cpusets_enabled()

[Dima Zavin]

On Wed, Jul 26, 2017 at 10:02 AM, Christopher Lameter <cl@xxxxxxxxx> wrote:
> On Wed, 26 Jul 2017, Dima Zavin wrote:
>
>> The fix is to cache the value that's returned by cpusets_enabled() at the
>> top of the loop, and only operate on the seqlock (both begin and retry) if
>> it was true.
>
> I think the proper fix would be to ensure that the calls to
> read_mems_allowed_{begin,retry} cannot cause the deadlock. Otherwise you
> have to fix this in multiple places.
>
> Maybe read_mems_allowed_* can do some form of synchronization or *_retry
> can implictly rely on the results of cpusets_enabled() by *_begin?
>

(res-ending because gmail hates me, sorry).

Thanks for the quick reply!

I can turn the cookie into a uint64, put the sequence into the low
order 32 bits and put the enabled state into bit 33 (or 63 :) ). Then
retry will not query cpusets_enabled() and will just look at the
enabled bit. This means that *_retry will always have a conditional
jump (i.e. lose the whole static_branch optimization) but maybe that's
ok since that's pretty rare and the *_begin() will still benefit from
it?