Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Andreas Mohr]

Hi,

[no properly binding reference via In-Reply-To: available thus manually re-creating, sorry]

> > But initerim I guess we could set our own owner field and check that
> > to keep the duct-tape from getting off completely.
> > -Daniel
> 
> Another alternative is to provide a standard mutex API that returns the
> owner of the lock if there is a real need for this capability. Peeking
> into lock internal is not a good practice.

>From personal experience here I would suggest that
the core issue here is that
this would create an inherently race-window-tainted API,
which clearly is something to be avoided:

The point is that the lock *owner* value is *volatile*
whenever it is *not* our own context instance
that is currently holding the lock
while querying this API
(i.e., thus not guaranteeing that the owner value will *not* be changed interim!),
since in such a situation
(not-privately-locked case!!)
lock ownership may change at any point from under
our just-observed result value.

Returning such inherently racy information from a publicly offered mutex API
is, errrrr, not so good, to put it rather mildly.

So, it seems the most we could provide
which would offer a reliable, non-racy API protocol
is something like:

static bool mutex_is_locked_by_us(struct mutex *mutex)

since during execution of this processing it would be guaranteed that:
- either we do have the lock, thus *we* *RELIABLY* are and will be "the owner"
- or we simply do not have it, thus *we* *RELIABLY* are and will be "not the owner"


[but note that in that case
this mutex API implementation code would have
a potentially unholy dependency on task stuff such as "current",
but which it probably already has anyway]

HTH,

Andreas Mohr

Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Peter Zijlstra]

On Fri, Aug 26, 2016 at 05:25:09AM +0200, Andreas Mohr wrote:
> > Another alternative is to provide a standard mutex API that returns the
> > owner of the lock if there is a real need for this capability. Peeking
> > into lock internal is not a good practice.


> So, it seems the most we could provide which would offer a reliable,
> non-racy API protocol is something like:
> 
> static bool mutex_is_locked_by_us(struct mutex *mutex)
> 
> since during execution of this processing it would be guaranteed that:
> - either we do have the lock, thus *we* *RELIABLY* are and will be "the owner"
> - or we simply do not have it, thus *we* *RELIABLY* are and will be "not the owner"

Right, and that is exactly what they attempted and need. And the new
mutex implementation could actually do this much better than the old
one.

But yes, such an interface should be part of the mutex implementation
proper, not something hacked on in random places.

Fwiw, the build bot seems to have found another instance of this thing
:/ drivers/gpu/drm/msm/msm_gem_shrinker.c includes an exact copy.

Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Waiman Long]

On 08/26/2016 05:10 AM, Peter Zijlstra wrote:
> On Fri, Aug 26, 2016 at 05:25:09AM +0200, Andreas Mohr wrote:
>>> Another alternative is to provide a standard mutex API that returns the
>>> owner of the lock if there is a real need for this capability. Peeking
>>> into lock internal is not a good practice.
>
>> So, it seems the most we could provide which would offer a reliable,
>> non-racy API protocol is something like:
>>
>> static bool mutex_is_locked_by_us(struct mutex *mutex)
>>
>> since during execution of this processing it would be guaranteed that:
>> - either we do have the lock, thus *we* *RELIABLY* are and will be "the owner"
>> - or we simply do not have it, thus *we* *RELIABLY* are and will be "not the owner"
> Right, and that is exactly what they attempted and need. And the new
> mutex implementation could actually do this much better than the old
> one.
>
> But yes, such an interface should be part of the mutex implementation
> proper, not something hacked on in random places.

It is what exactly I have in mind. The actual API implemented is subject 
to negotiation. The important thing is that it has to be within the core 
mutex code.

> Fwiw, the build bot seems to have found another instance of this thing
> :/ drivers/gpu/drm/msm/msm_gem_shrinker.c includes an exact copy.

This seems to be a new file that was introduced since 4.8.

Cheers,
Longman

Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Waiman Long]

On 08/25/2016 03:36 PM, Daniel Vetter wrote:
> On Thu, Aug 25, 2016 at 8:37 PM, Peter Zijlstra<peterz@infradead.org>  wrote:
>> Poking at lock internals is not cool. Since I'm going to change the
>> implementation this will break, take it out.
>>
>> Cc: Chris Wilson<chris@chris-wilson.co.uk>
>> Cc: Daniel Vetter<daniel.vetter@ffwll.ch>
>> Signed-off-by: Peter Zijlstra (Intel)<peterz@infradead.org>
> It's horrible, but we die without this in spurious oom. And we haven't
> made much progress in recent years to throw out the locking scheme and
> replace it by something else that doesn't need a recursive mutex.
>
> But initerim I guess we could set our own owner field and check that
> to keep the duct-tape from getting off completely.
> -Daniel

Another alternative is to provide a standard mutex API that returns the 
owner of the lock if there is a real need for this capability. Peeking 
into lock internal is not a good practice.

Cheers,
Longman

Re: [RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Daniel Vetter]

On Thu, Aug 25, 2016 at 8:37 PM, Peter Zijlstra <peterz@infradead.org> wrote:
> Poking at lock internals is not cool. Since I'm going to change the
> implementation this will break, take it out.
>
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>

It's horrible, but we die without this in spurious oom. And we haven't
made much progress in recent years to throw out the locking scheme and
replace it by something else that doesn't need a recursive mutex.

But initerim I guess we could set our own owner field and check that
to keep the duct-tape from getting off completely.
-Daniel

> ---
>  drivers/gpu/drm/i915/i915_gem_shrinker.c |   26 +++-----------------------
>  1 file changed, 3 insertions(+), 23 deletions(-)
>
> --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
> +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> @@ -35,19 +35,6 @@
>  #include "i915_drv.h"
>  #include "i915_trace.h"
>
> -static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
> -{
> -       if (!mutex_is_locked(mutex))
> -               return false;
> -
> -#if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
> -       return mutex->owner == task;
> -#else
> -       /* Since UP may be pre-empted, we cannot assume that we own the lock */
> -       return false;
> -#endif
> -}
> -
>  static int num_vma_bound(struct drm_i915_gem_object *obj)
>  {
>         struct i915_vma *vma;
> @@ -238,17 +225,10 @@ unsigned long i915_gem_shrink_all(struct
>
>  static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
>  {
> -       if (!mutex_trylock(&dev->struct_mutex)) {
> -               if (!mutex_is_locked_by(&dev->struct_mutex, current))
> -                       return false;
> -
> -               if (to_i915(dev)->mm.shrinker_no_lock_stealing)
> -                       return false;
> -
> -               *unlock = false;
> -       } else
> -               *unlock = true;
> +       if (!mutex_trylock(&dev->struct_mutex))
> +               return false;
>
> +       *unlock = true;
>         return true;
>  }
>
>
>



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

[RFC][PATCH -v2 1/4] locking/drm/i915: Kill mutex trickery

[Peter Zijlstra]

[-- Attachment #1: peterz-revert-drm-i915-brainmelt.patch --]
[-- Type: text/plain, Size: 1470 bytes --]

Poking at lock internals is not cool. Since I'm going to change the
implementation this will break, take it out.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 drivers/gpu/drm/i915/i915_gem_shrinker.c |   26 +++-----------------------
 1 file changed, 3 insertions(+), 23 deletions(-)

--- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
@@ -35,19 +35,6 @@
 #include "i915_drv.h"
 #include "i915_trace.h"
 
-static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
-{
-	if (!mutex_is_locked(mutex))
-		return false;
-
-#if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
-	return mutex->owner == task;
-#else
-	/* Since UP may be pre-empted, we cannot assume that we own the lock */
-	return false;
-#endif
-}
-
 static int num_vma_bound(struct drm_i915_gem_object *obj)
 {
 	struct i915_vma *vma;
@@ -238,17 +225,10 @@ unsigned long i915_gem_shrink_all(struct
 
 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 {
-	if (!mutex_trylock(&dev->struct_mutex)) {
-		if (!mutex_is_locked_by(&dev->struct_mutex, current))
-			return false;
-
-		if (to_i915(dev)->mm.shrinker_no_lock_stealing)
-			return false;
-
-		*unlock = false;
-	} else
-		*unlock = true;
+	if (!mutex_trylock(&dev->struct_mutex))
+		return false;
 
+	*unlock = true;
 	return true;
 }