Re: [PATCH v4 3/3] locking/mutex: Ensure forward progress of waiter-spinner
From: Waiman Long
Date: Tue Aug 09 2016 - 14:32:39 EST
On 08/08/2016 01:37 PM, Peter Zijlstra wrote:
On Mon, Jul 18, 2016 at 04:39:26PM -0400, Waiman Long wrote:
As both an optimistic spinner and a waiter-spinner (a woken task from
the wait queue spinning) can be spinning on the lock at the same time,
we cannot ensure forward progress for the waiter-spinner. Therefore,
it is possible for the waiter-spinner to be starved of getting the
lock, though not likely.
Right; yet your previous two changelogs/comments implied otherwise.
This patch adds a flag to indicate that a waiter-spinner is
spinning and hence has priority over the acquisition of the lock. A
waiter-spinner sets this flag while spinning. An optimistic spinner
will check this flag and yield if set. This essentially makes the
waiter-spinner jump to the head of the optimistic spinning queue to
acquire the lock.
There will be no increase in size for the mutex structure for 64-bit
architectures. For 32-bit architectures, there will be a size increase
of 4 bytes.
Alternative might be to use the LSB of mutex::owner, but that's going to
be somewhat icky too.
I was thinking about doing that. However, the owner field is used in
quite a number of places. It may be a bit risky to change all of them.
I'm not sure the 32bit platforms are going to be excited about growing
struct mutex...
Or we can make this a 64-bit architecture specific change if the
increase in mutex size is a real concern. Actually, we don't need to use
a list_head structure for wait_list. It can be just a pointer to
mutex_waiter that has the list_head structure. This can save a pointer
from the structure.
Cheers,
Longman