lkml.org 
[lkml]   [2019]   [Dec]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2] mm/hugetlb: defer free_huge_page() to a workqueue
From
Date
On 12/16/19 8:17 AM, Davidlohr Bueso wrote:
> On Mon, 16 Dec 2019, Michal Hocko wrote:
>> I am afraid that work_struct is too large to be stuffed into the struct
>> page array (because of the lockdep part).
>
> Yeah, this needs to be done without touching struct page.
>
> Which is why I had done the stack allocated way in this patch, but we
> cannot wait for it to complete in irq, so that's out the window. Andi
> had suggested percpu allocated work items, but having played with the
> idea over the weekend, I don't see how we can prevent another page being
> freed on the same cpu before previous work on the same cpu is complete
> (cpu0 wants to free pageA, schedules the work, in the mean time cpu0
> wants to free pageB and workerfn for pageA still hasn't been called).
>
>> I think that it would be just safer to make hugetlb_lock irq safe. Are
>> there any other locks that would require the same?
>
> It would be simpler. Any performance issues that arise would probably
> be only seen in microbenchmarks, assuming we want to have full irq safety.
> If we don't need to worry about hardirq, then even better.
>
> The subpool lock would also need to be irq safe.

I do think we need to worry about hardirq. There are no restruictions that
put_page can not be called from hardirq context.

I am concerned about the latency of making hugetlb_lock (and potentially
subpool lock) hardirq safe. When these locks were introduced (before my
time) the concept of making them irq safe was not considered. Recently,
I learned that the hugetlb_lock is held for a linear scan of ALL hugetlb
pages during a cgroup reparentling operation. That is just too long.

If there is no viable work queue solution, then I think we would like to
restructure the hugetlb locking before a change to just make hugetlb_lock
irq safe. The idea would be to split the scope of what is done under
hugetlb_lock. Most of it would never be executed in irq context. Then
have a small/limited set of functionality that really needs to be irq
safe protected by an irq safe lock.

--
Mike Kravetz

\
 
 \ /
  Last update: 2019-12-16 20:10    [W:0.048 / U:12.376 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site