Re: possible deadlock in __do_page_fault

[Tetsuo Handa]

On 2019/01/24 22:46, Joel Fernandes wrote:
> On Thu, Jan 24, 2019 at 10:52:30AM +0900, Tetsuo Handa wrote:
>> Then, I'm tempted to eliminate shrinker and LRU list (like a draft patch shown
>> below). I think this is not equivalent to current code because this shrinks
>> upon only range_alloc() time and I don't know whether it is OK to temporarily
>> release ashmem_mutex during range_alloc() at "Case #4" of ashmem_pin(), but
>> can't we go this direction? 
> 
> No, the point of the shrinker is to do a lazy free. We cannot free things
> during unpin since it can be pinned again and we need to find that range by
> going through the list. We also cannot get rid of any lists. Since if
> something is re-pinned, we need to find it and find out if it was purged. We
> also need the list for knowing what was unpinned so the shrinker works.
> 
> By the way, all this may be going away quite soon (the whole driver) as I
> said, so just give it a little bit of time.
> 
> I am happy to fix it soon if that's not the case (which I should know soon -
> like a couple of weeks) but I'd like to hold off till then.
> 
>> By the way, why not to check range_alloc() failure before calling range_shrink() ?
> 
> That would be a nice thing to do. Send a patch?

OK. Here is a patch. I chose __GFP_NOFAIL rather than adding error handling,
for small GFP_KERNEL allocation won't fail unless current thread was killed by
the OOM killer or memory allocation fault injection forces it fail, and
range_alloc() will not be called for multiple times from one syscall.

But note that doing GFP_KERNEL allocation with ashmem_mutex held has a risk of
needlessly invoking the OOM killer because "the point of the shrinker is to do
a lazy free" counts on ashmem_mutex not held by GFP_KERNEL allocating thread.
Although other shrinkers likely make forward progress by releasing memory,
technically you should avoid doing GFP_KERNEL allocation with ashmem_mutex held
if shrinker depends on ashmem_mutex not held.