[Linaro-mm-sig] thoughts of looking at android fences

[Rom Lemarchand]

I ran some benchmarks and things seem to be running about the same.
No one on our graphics team seemed concerned about the change.

The only concern I heard was about the increased complexity of the new sync
code as opposed to the old sync framework which tried to keep things
straightforward.


On Fri, Nov 8, 2013 at 3:43 AM, Maarten Lankhorst <
maarten.lankhorst at canonical.com> wrote:

> op 07-11-13 22:11, Rom Lemarchand schreef:
> > Hi Maarten, I tested your changes and needed the attached patch: behavior
> > now seems equivalent as android sync. I haven't tested performance.
> >
> > The issue resolved by this patch happens when i_b < b->num_fences and i_a
> >> = a->num_fences (or vice versa). Then, pt_a is invalid and so
> > dereferencing pt_a->context causes a crash.
> >
> Yeah, I pushed my original fix. I intended to keep android userspace
> behavior the same, and I tried to keep the kernelspace the api same as much
> as I could. If peformance is the same, or not noticeably worse, would there
> be any objections on the android side about renaming dma-fence to
> syncpoint, and getting it in mainline?
>
> ~Maarten
>
-- next part --
An HTML attachment was scrubbed...
URL: 

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 07-11-13 22:11, Rom Lemarchand schreef:
> Hi Maarten, I tested your changes and needed the attached patch: behavior
> now seems equivalent as android sync. I haven't tested performance.
>
> The issue resolved by this patch happens when i_b < b->num_fences and i_a
>> = a->num_fences (or vice versa). Then, pt_a is invalid and so
> dereferencing pt_a->context causes a crash.
>
Yeah, I pushed my original fix. I intended to keep android userspace behavior 
the same, and I tried to keep the kernelspace the api same as much as I could. 
If peformance is the same, or not noticeably worse, would there be any 
objections on the android side about renaming dma-fence to syncpoint, and 
getting it in mainline?

~Maarten

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 07-11-13 22:11, Rom Lemarchand schreef:
> Hi Maarten, I tested your changes and needed the attached patch: behavior
> now seems equivalent as android sync. I haven't tested performance.
>
> The issue resolved by this patch happens when i_b < b->num_fences and
> i_a >= a->num_fences (or vice versa). Then, pt_a is invalid and so
> dereferencing pt_a->context causes a crash.

Oops, thinko. :) Originally I had it correct by doing this:

+   /*
+* Assume sync_fence a and b are both ordered and have no
+* duplicates with the same context.
+*
+* If a sync_fence can only be created with sync_fence_merge
+* and sync_fence_create, this is a reasonable assumption.
+*/
+   for (i = i_a = i_b = 0; i_a < a->num_fences && i_b < b->num_fences; ) {
+   struct fence *pt_a = a->cbs[i_a].sync_pt;
+   struct fence *pt_b = b->cbs[i_b].sync_pt;
+
+   if (pt_a->context < pt_b->context) {
+   sync_fence_add_pt(fence, , pt_a);
+
+   i_a++;
+   } else if (pt_a->context > pt_b->context) {
+   sync_fence_add_pt(fence, , pt_b);
+
+   i_b++;
+   } else {
+   if (pt_a->seqno - pt_b->seqno <= INT_MAX)
+   sync_fence_add_pt(fence, , pt_a);
+   else
+   sync_fence_add_pt(fence, , pt_b);
+
+   i_a++;
+   i_b++;
+   }
+}
+
+/* Add remaining fences from a or b*/
+for (; i_a < a->num_fences; i_a++)
+   sync_fence_add_pt(fence, , a->cbs[i_a].sync_pt);
+
+for (; i_b < b->num_fences; i_b++)
+   sync_fence_add_pt(fence, , b->cbs[i_b].sync_pt);

Then I thought I could clean it up by merging it, but that ended up being
more unreadable and crashing... so I guess I'll revert back to this version. :)

[Linaro-mm-sig] thoughts of looking at android fences

[Rom Lemarchand]

Let me run some benchmarks today, talk to people internally, and I'll let
you know.
On Nov 8, 2013 3:43 AM, "Maarten Lankhorst" 
wrote:

> op 07-11-13 22:11, Rom Lemarchand schreef:
> > Hi Maarten, I tested your changes and needed the attached patch: behavior
> > now seems equivalent as android sync. I haven't tested performance.
> >
> > The issue resolved by this patch happens when i_b < b->num_fences and i_a
> >> = a->num_fences (or vice versa). Then, pt_a is invalid and so
> > dereferencing pt_a->context causes a crash.
> >
> Yeah, I pushed my original fix. I intended to keep android userspace
> behavior the same, and I tried to keep the kernelspace the api same as much
> as I could. If peformance is the same, or not noticeably worse, would there
> be any objections on the android side about renaming dma-fence to
> syncpoint, and getting it in mainline?
>
> ~Maarten
>
-- next part --
An HTML attachment was scrubbed...
URL: 

[Linaro-mm-sig] thoughts of looking at android fences

[Rom Lemarchand]

Hi Maarten, I tested your changes and needed the attached patch: behavior
now seems equivalent as android sync. I haven't tested performance.

The issue resolved by this patch happens when i_b < b->num_fences and i_a
>= a->num_fences (or vice versa). Then, pt_a is invalid and so
dereferencing pt_a->context causes a crash.


On Mon, Nov 4, 2013 at 2:31 AM, Maarten Lankhorst <
maarten.lankhorst at canonical.com> wrote:

> op 02-11-13 22:36, Colin Cross schreef:
> > On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst
> >  wrote:
> >> op 24-10-13 14:13, Maarten Lankhorst schreef:
> >>> So I actually tried to implement it now. I killed all the deprecated
> members and assumed a linear timeline.
> >>> This means that syncpoints can only be added at the end, not in
> between. In particular it means sw_sync
> >>> might be slightly broken.
> >>>
> >>> I only tested it with a simple program I wrote called ufence.c, it's
> in drivers/staging/android/ufence.c in the following tree:
> >>>
> >>> http://cgit.freedesktop.org/~mlankhorst/linux
> >>>
> >>> the "rfc: convert android to fence api" has all the changes from my
> dma-fence proposal to what android would need,
> >>> it also converts the userspace fence api to use the dma-fence api.
> >>>
> >>> sync_pt is implemented as fence too. This meant not having to convert
> all of android right away, though I did make some changes.
> >>> I killed the deprecated members and made all the fence calls forward
> to the sync_timeline_ops. dup and compare are no longer used.
> >>>
> >>> I haven't given this a spin on a full android kernel, only with the
> components that are in mainline kernel under staging and my dumb test
> program.
> >>>
> >>> ~Maarten
> >>>
> >>> PS: The nomenclature is very confusing. I want to rename dma-fence to
> syncpoint, but I want some feedback from the android devs first. :)
> >>>
> >> Come on, any feedback? I want to move the discussion forward.
> >>
> >> ~Maarten
> > I experimented with it a little on a device that uses sync and came
> > across a few bugs:
> > 1.  sync_timeline_signal needs to call __fence_signal on all signaled
> > points on the timeline, not just the first
> > 2.  fence_add_callback doesn't always initialize cb.node
> > 3.  sync_fence_wait should take ms
> > 4.  sync_print_pt status printing was incorrect
> > 5.  there is a deadlock:
> >sync_print_obj takes obj->child_list_lock
> >sync_print_pt
> >fence_is_signaled
> >fence_signal takes fence->lock == obj->child_list_lock
> > 6.  freeing a timeline before all the fences holding points on that
> > timeline have timed out results in a crash
> >
> > With the attached patch to fix these issues, our libsync and sync_test
> > give the same results as with our sync code.  I haven't tested against
> > the full Android framework yet.
> >
> > The compare op and timeline ordering is critical to the efficiency of
> > sync points on Android.  The compare op is used when merging fences to
> > drop all but the latest point on the same timeline.  This is necessary
> > for example when the same buffer is submitted to the display on
> > multiple frames, like when there is a live wallpaper in the background
> > updating at 60 fps and a static screen of widgets on top of it.  The
> > static widget buffer is submitted on every frame, returning a new
> > fence each time.  The compositor merges the new fence with the fence
> > for the previous buffer, and because they are on the same timeline it
> > merges down to a single point.  I experimented with disabling the
> > merge optimization on a Nexus 10, and found that leaving the screen on
> > running a live wallpaper eventually resulted in 100k outstanding sync
> > points.
>
> Well, here I did the same for dma-fence, can you take a look?
>
> ---
>
> diff --git a/drivers/staging/android/sync.c
> b/drivers/staging/android/sync.c
> index 2c7fd3f2ab23..d1d89f1f8553 100644
> --- a/drivers/staging/android/sync.c
> +++ b/drivers/staging/android/sync.c
> @@ -232,39 +232,62 @@ void sync_fence_install(struct sync_fence *fence,
> int fd)
>  }
>  EXPORT_SYMBOL(sync_fence_install);
>
> +static void sync_fence_add_pt(struct sync_fence *fence, int *i, struct
> fence *pt) {
> +   fence->cbs[*i].sync_pt = pt;
> +   fence->cbs[*i].fence = fence;
> +
> +   if (!fence_add_callback(pt, >cbs[*i].cb,
> fence_check_cb_func)) {
> +   fence_get(pt);
> +   (*i)++;
> +   }
> +}
> +
>  struct sync_fence *sync_fence_merge(const char *name,
> struct sync_fence *a, struct
> sync_fence *b)
>  {
> int num_fences = a->num_fences + b->num_fences;
> struct sync_fence *fence;
> -   int i;
> +   int i, i_a, i_b;
>
> fence = sync_fence_alloc(offsetof(struct sync_fence,
> cbs[num_fences]), name);
> if (fence == NULL)
> return NULL;
>
> -   fence->num_fences = num_fences;
> atomic_set(>status, num_fences);
>
> -   for (i = 0; i < 

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 02-11-13 22:36, Colin Cross schreef:
> On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst
>  wrote:
>> op 24-10-13 14:13, Maarten Lankhorst schreef:
>>> So I actually tried to implement it now. I killed all the deprecated 
>>> members and assumed a linear timeline.
>>> This means that syncpoints can only be added at the end, not in between. In 
>>> particular it means sw_sync
>>> might be slightly broken.
>>>
>>> I only tested it with a simple program I wrote called ufence.c, it's in 
>>> drivers/staging/android/ufence.c in the following tree:
>>>
>>> http://cgit.freedesktop.org/~mlankhorst/linux
>>>
>>> the "rfc: convert android to fence api" has all the changes from my 
>>> dma-fence proposal to what android would need,
>>> it also converts the userspace fence api to use the dma-fence api.
>>>
>>> sync_pt is implemented as fence too. This meant not having to convert all 
>>> of android right away, though I did make some changes.
>>> I killed the deprecated members and made all the fence calls forward to the 
>>> sync_timeline_ops. dup and compare are no longer used.
>>>
>>> I haven't given this a spin on a full android kernel, only with the 
>>> components that are in mainline kernel under staging and my dumb test 
>>> program.
>>>
>>> ~Maarten
>>>
>>> PS: The nomenclature is very confusing. I want to rename dma-fence to 
>>> syncpoint, but I want some feedback from the android devs first. :)
>>>
>> Come on, any feedback? I want to move the discussion forward.
>>
>> ~Maarten
> I experimented with it a little on a device that uses sync and came
> across a few bugs:
> 1.  sync_timeline_signal needs to call __fence_signal on all signaled
> points on the timeline, not just the first
> 2.  fence_add_callback doesn't always initialize cb.node
> 3.  sync_fence_wait should take ms
> 4.  sync_print_pt status printing was incorrect
> 5.  there is a deadlock:
>sync_print_obj takes obj->child_list_lock
>sync_print_pt
>fence_is_signaled
>fence_signal takes fence->lock == obj->child_list_lock
> 6.  freeing a timeline before all the fences holding points on that
> timeline have timed out results in a crash
>
> With the attached patch to fix these issues, our libsync and sync_test
> give the same results as with our sync code.  I haven't tested against
> the full Android framework yet.
>
> The compare op and timeline ordering is critical to the efficiency of
> sync points on Android.  The compare op is used when merging fences to
> drop all but the latest point on the same timeline.  This is necessary
> for example when the same buffer is submitted to the display on
> multiple frames, like when there is a live wallpaper in the background
> updating at 60 fps and a static screen of widgets on top of it.  The
> static widget buffer is submitted on every frame, returning a new
> fence each time.  The compositor merges the new fence with the fence
> for the previous buffer, and because they are on the same timeline it
> merges down to a single point.  I experimented with disabling the
> merge optimization on a Nexus 10, and found that leaving the screen on
> running a live wallpaper eventually resulted in 100k outstanding sync
> points.

Well, here I did the same for dma-fence, can you take a look?

---

diff --git a/drivers/staging/android/sync.c b/drivers/staging/android/sync.c
index 2c7fd3f2ab23..d1d89f1f8553 100644
--- a/drivers/staging/android/sync.c
+++ b/drivers/staging/android/sync.c
@@ -232,39 +232,62 @@ void sync_fence_install(struct sync_fence *fence, int fd)
 }
 EXPORT_SYMBOL(sync_fence_install);

+static void sync_fence_add_pt(struct sync_fence *fence, int *i, struct fence 
*pt) {
+   fence->cbs[*i].sync_pt = pt;
+   fence->cbs[*i].fence = fence;
+
+   if (!fence_add_callback(pt, >cbs[*i].cb, fence_check_cb_func)) {
+   fence_get(pt);
+   (*i)++;
+   }
+}
+
 struct sync_fence *sync_fence_merge(const char *name,
struct sync_fence *a, struct sync_fence *b)
 {
int num_fences = a->num_fences + b->num_fences;
struct sync_fence *fence;
-   int i;
+   int i, i_a, i_b;

fence = sync_fence_alloc(offsetof(struct sync_fence, cbs[num_fences]), 
name);
if (fence == NULL)
return NULL;

-   fence->num_fences = num_fences;
atomic_set(>status, num_fences);

-   for (i = 0; i < a->num_fences; ++i) {
-   struct fence *pt = a->cbs[i].sync_pt;
-
-   fence_get(pt);
-   fence->cbs[i].sync_pt = pt;
-   fence->cbs[i].fence = fence;
-   if (fence_add_callback(pt, >cbs[i].cb, 
fence_check_cb_func))
-   atomic_dec(>status);
+   /*
+* Assume sync_fence a and b are both ordered and have no
+* duplicates with the same context.
+*
+* If a sync_fence can only be created with sync_fence_merge
+* and sync_fence_create, this is a reasonable assumption.
+   

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 02-11-13 22:36, Colin Cross schreef:
> On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst
>  wrote:
>> op 24-10-13 14:13, Maarten Lankhorst schreef:
>>> So I actually tried to implement it now. I killed all the deprecated 
>>> members and assumed a linear timeline.
>>> This means that syncpoints can only be added at the end, not in between. In 
>>> particular it means sw_sync
>>> might be slightly broken.
>>>
>>> I only tested it with a simple program I wrote called ufence.c, it's in 
>>> drivers/staging/android/ufence.c in the following tree:
>>>
>>> http://cgit.freedesktop.org/~mlankhorst/linux
>>>
>>> the "rfc: convert android to fence api" has all the changes from my 
>>> dma-fence proposal to what android would need,
>>> it also converts the userspace fence api to use the dma-fence api.
>>>
>>> sync_pt is implemented as fence too. This meant not having to convert all 
>>> of android right away, though I did make some changes.
>>> I killed the deprecated members and made all the fence calls forward to the 
>>> sync_timeline_ops. dup and compare are no longer used.
>>>
>>> I haven't given this a spin on a full android kernel, only with the 
>>> components that are in mainline kernel under staging and my dumb test 
>>> program.
>>>
>>> ~Maarten
>>>
>>> PS: The nomenclature is very confusing. I want to rename dma-fence to 
>>> syncpoint, but I want some feedback from the android devs first. :)
>>>
>> Come on, any feedback? I want to move the discussion forward.
>>
>> ~Maarten
> I experimented with it a little on a device that uses sync and came
> across a few bugs:
> 1.  sync_timeline_signal needs to call __fence_signal on all signaled
> points on the timeline, not just the first
> 2.  fence_add_callback doesn't always initialize cb.node
> 3.  sync_fence_wait should take ms
> 4.  sync_print_pt status printing was incorrect
> 5.  there is a deadlock:
>sync_print_obj takes obj->child_list_lock
>sync_print_pt
>fence_is_signaled
>fence_signal takes fence->lock == obj->child_list_lock
> 6.  freeing a timeline before all the fences holding points on that
> timeline have timed out results in a crash
>
> With the attached patch to fix these issues, our libsync and sync_test
> give the same results as with our sync code.  I haven't tested against
> the full Android framework yet.
>
> The compare op and timeline ordering is critical to the efficiency of
> sync points on Android.  The compare op is used when merging fences to
> drop all but the latest point on the same timeline.  This is necessary
> for example when the same buffer is submitted to the display on
> multiple frames, like when there is a live wallpaper in the background
> updating at 60 fps and a static screen of widgets on top of it.  The
> static widget buffer is submitted on every frame, returning a new
> fence each time.  The compositor merges the new fence with the fence
> for the previous buffer, and because they are on the same timeline it
> merges down to a single point.  I experimented with disabling the
> merge optimization on a Nexus 10, and found that leaving the screen on
> running a live wallpaper eventually resulted in 100k outstanding sync
> points.

Hey,


fence_add_callback will now always initialize cb->node, even on failure.
I added __fence_is_signaled, to be used with the lock held.
sync_print_pt didn't work when the fence was signaled with an error, I fixed 
that.

So I reworked patch below, no merge optimization yet. It will be done as a 
separate patch. :)

---
diff --git a/drivers/base/fence.c b/drivers/base/fence.c
index 89c89ae19f58..9e7a63c4b07f 100644
--- a/drivers/base/fence.c
+++ b/drivers/base/fence.c
@@ -185,8 +185,10 @@ int fence_add_callback(struct fence *fence, struct 
fence_cb *cb,
if (WARN_ON(!fence || !func))
return -EINVAL;

-   if (test_bit(FENCE_FLAG_SIGNALED_BIT, >flags))
+   if (test_bit(FENCE_FLAG_SIGNALED_BIT, >flags)) {
+   LIST_INIT_HEAD(>node);
return -ENOENT;
+   }

spin_lock_irqsave(fence->lock, flags);

@@ -202,7 +204,8 @@ int fence_add_callback(struct fence *fence, struct fence_cb 
*cb,
if (!ret) {
cb->func = func;
list_add_tail(>node, >cb_list);
-   }
+   } else
+   LIST_INIT_HEAD(>node);
spin_unlock_irqrestore(fence->lock, flags);

return ret;
diff --git a/drivers/staging/android/sync.c b/drivers/staging/android/sync.c
index 110a9e99cb71..2c7fd3f2ab23 100644
--- a/drivers/staging/android/sync.c
+++ b/drivers/staging/android/sync.c
@@ -74,6 +74,16 @@ static void sync_timeline_free(struct kref *kref)
kfree(obj);
 }

+static void sync_timeline_get(struct sync_timeline *obj)
+{
+   kref_get(>kref);
+}
+
+static void sync_timeline_put(struct sync_timeline *obj)
+{
+   kref_put(>kref, sync_timeline_free);
+}
+
 void sync_timeline_destroy(struct sync_timeline *obj)
 {
obj->destroyed = true;
@@ -83,8 +93,8 @@ void 

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 02-11-13 22:36, Colin Cross schreef:
> On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst
>  wrote:
>> op 24-10-13 14:13, Maarten Lankhorst schreef:
>>> So I actually tried to implement it now. I killed all the deprecated 
>>> members and assumed a linear timeline.
>>> This means that syncpoints can only be added at the end, not in between. In 
>>> particular it means sw_sync
>>> might be slightly broken.
>>>
>>> I only tested it with a simple program I wrote called ufence.c, it's in 
>>> drivers/staging/android/ufence.c in the following tree:
>>>
>>> http://cgit.freedesktop.org/~mlankhorst/linux
>>>
>>> the "rfc: convert android to fence api" has all the changes from my 
>>> dma-fence proposal to what android would need,
>>> it also converts the userspace fence api to use the dma-fence api.
>>>
>>> sync_pt is implemented as fence too. This meant not having to convert all 
>>> of android right away, though I did make some changes.
>>> I killed the deprecated members and made all the fence calls forward to the 
>>> sync_timeline_ops. dup and compare are no longer used.
>>>
>>> I haven't given this a spin on a full android kernel, only with the 
>>> components that are in mainline kernel under staging and my dumb test 
>>> program.
>>>
>>> ~Maarten
>>>
>>> PS: The nomenclature is very confusing. I want to rename dma-fence to 
>>> syncpoint, but I want some feedback from the android devs first. :)
>>>
>> Come on, any feedback? I want to move the discussion forward.
>>
>> ~Maarten
> I experimented with it a little on a device that uses sync and came
> across a few bugs:
> 1.  sync_timeline_signal needs to call __fence_signal on all signaled
> points on the timeline, not just the first
> 2.  fence_add_callback doesn't always initialize cb.node
> 3.  sync_fence_wait should take ms
> 4.  sync_print_pt status printing was incorrect
Well, in the normal case status may not be set, but fence may still be 
signaled. Any status should be set before signaling. In cases where non-android 
dma-fences are used.
I deliberately tried not to depend on any android stuff in the android fences, 
this way it could be used outside android too.

I'm trying to only set status on error. If there is a race with 2 threads 
calling fence_signal, one with error, one without, then the error will always 
be visible in fence->status that way.
But I guess this might be a corner case we wouldn't worry about normally...
> 5.  there is a deadlock:
>sync_print_obj takes obj->child_list_lock
>sync_print_pt
>fence_is_signaled
>fence_signal takes fence->lock == obj->child_list_lock
> 6.  freeing a timeline before all the fences holding points on that
> timeline have timed out results in a crash
To be honest, I was surprised this code mostly worked for the sw_fence test 
code I wrote originally. :)
I think I'll have to add a __fence_is_signaled for 5, which can be called with 
fence->lock held.


>
> With the attached patch to fix these issues, our libsync and sync_test
> give the same results as with our sync code.  I haven't tested against
> the full Android framework yet.
>
> The compare op and timeline ordering is critical to the efficiency of
> sync points on Android.  The compare op is used when merging fences to
> drop all but the latest point on the same timeline.  This is necessary
> for example when the same buffer is submitted to the display on
> multiple frames, like when there is a live wallpaper in the background
> updating at 60 fps and a static screen of widgets on top of it.  The
> static widget buffer is submitted on every frame, returning a new
> fence each time.  The compositor merges the new fence with the fence
> for the previous buffer, and because they are on the same timeline it
> merges down to a single point.  I experimented with disabling the
> merge optimization on a Nexus 10, and found that leaving the screen on
> running a live wallpaper eventually resulted in 100k outstanding sync
> points.
Yeah I've been looking into how to do that. It's very easy to optimize 
actually. The dma-fence code requires a context,
which is a number (even if it may become a pointer later, it could be seen as a 
number). If we order all fences based on number,
fences could have an order based on context number. Merging fences would simply 
become adding 2 ordered lists, and dropping
any duplicates and signaled points.

I left it out for the RFC I wanted to keep things as readable as possible. :)

~Maarten

[Linaro-mm-sig] thoughts of looking at android fences

[Colin Cross]

On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst
 wrote:
> op 24-10-13 14:13, Maarten Lankhorst schreef:
>> So I actually tried to implement it now. I killed all the deprecated members 
>> and assumed a linear timeline.
>> This means that syncpoints can only be added at the end, not in between. In 
>> particular it means sw_sync
>> might be slightly broken.
>>
>> I only tested it with a simple program I wrote called ufence.c, it's in 
>> drivers/staging/android/ufence.c in the following tree:
>>
>> http://cgit.freedesktop.org/~mlankhorst/linux
>>
>> the "rfc: convert android to fence api" has all the changes from my 
>> dma-fence proposal to what android would need,
>> it also converts the userspace fence api to use the dma-fence api.
>>
>> sync_pt is implemented as fence too. This meant not having to convert all of 
>> android right away, though I did make some changes.
>> I killed the deprecated members and made all the fence calls forward to the 
>> sync_timeline_ops. dup and compare are no longer used.
>>
>> I haven't given this a spin on a full android kernel, only with the 
>> components that are in mainline kernel under staging and my dumb test 
>> program.
>>
>> ~Maarten
>>
>> PS: The nomenclature is very confusing. I want to rename dma-fence to 
>> syncpoint, but I want some feedback from the android devs first. :)
>>
> Come on, any feedback? I want to move the discussion forward.
>
> ~Maarten

I experimented with it a little on a device that uses sync and came
across a few bugs:
1.  sync_timeline_signal needs to call __fence_signal on all signaled
points on the timeline, not just the first
2.  fence_add_callback doesn't always initialize cb.node
3.  sync_fence_wait should take ms
4.  sync_print_pt status printing was incorrect
5.  there is a deadlock:
   sync_print_obj takes obj->child_list_lock
   sync_print_pt
   fence_is_signaled
   fence_signal takes fence->lock == obj->child_list_lock
6.  freeing a timeline before all the fences holding points on that
timeline have timed out results in a crash

With the attached patch to fix these issues, our libsync and sync_test
give the same results as with our sync code.  I haven't tested against
the full Android framework yet.

The compare op and timeline ordering is critical to the efficiency of
sync points on Android.  The compare op is used when merging fences to
drop all but the latest point on the same timeline.  This is necessary
for example when the same buffer is submitted to the display on
multiple frames, like when there is a live wallpaper in the background
updating at 60 fps and a static screen of widgets on top of it.  The
static widget buffer is submitted on every frame, returning a new
fence each time.  The compositor merges the new fence with the fence
for the previous buffer, and because they are on the same timeline it
merges down to a single point.  I experimented with disabling the
merge optimization on a Nexus 10, and found that leaving the screen on
running a live wallpaper eventually resulted in 100k outstanding sync
points.
-- next part --
A non-text attachment was scrubbed...
Name: 0001-dma-fence-fixes.patch
Type: text/x-patch
Size: 4962 bytes
Desc: not available
URL: 

[Linaro-mm-sig] thoughts of looking at android fences

[Rom Lemarchand]

Sorry about the delay.
Hopefully other people from Android will also chime in.
We need the ability to merge sync fences and keep the sync_pt ordered: the
idea behind sync timelines is that we promise an ordering of operations.

Our reference device is Nexus 10: we need to make sure that any new
implementation satisfies the same requirements.

You can find sample use-cases here, we also use it in our hardware composer
libraries:
https://android.googlesource.com/platform/system/core/+/refs/heads/master/libsync/
https://android.googlesource.com/platform/frameworks/native/+/master/libs/ui/Fence.cpp



On Wed, Oct 30, 2013 at 5:17 AM, Maarten Lankhorst <
maarten.lankhorst at canonical.com> wrote:

> op 24-10-13 14:13, Maarten Lankhorst schreef:
> > op 09-10-13 16:39, Maarten Lankhorst schreef:
> >> Hey,
> >>
> >>  op 08-10-13 19:37, John Stultz schreef:
> >>> On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling 
> wrote:
>  On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
>   wrote:
> > Depending on feedback I'll try reflashing my nexus 7 to stock
> android, and work on trying to convert android
> > syncpoints to dma-fence, which I'll probably rename to syncpoints.
>  I thought the plan decided at plumbers was to investigate backing
>  dma_buf with the android sync solution not the other way around.  It
>  doesn't make sense to me to take a working, tested, end-to-end
>  solution with a released compositing system built around it, throw it
>  out, and replace it with new un-tested code to
>  support a system which is not yet built.
> >>> Hey Erik,
> >>>   Thanks for the clarifying points in your email, your insights and
> >>> feedback are critical, and I think having you and Maarten continue to
> >>> work out the details here will make this productive.
> >>>
> >>> My recollection from the discussion was that Rob was ok with trying to
> >>> pipe the sync arguments through the various interfaces in order to
> >>> support the explicit sync, but I think he did suggest having it backed
> >>> by the dma-buf fences underneath.
> >>>
> >>> I know this can be frustrating to watch things be reimplemented when
> >>> you have a pre-baked solution, but some compromise will be needed to
> >>> get things merged (and Maarten is taking the initiative here), but its
> >>> important to keep discussing this so the *right* compromises are made
> >>> that don't hurt performance, etc.
> >>>
> >>> My hope is Maarten's approach of getting the dma-fence core
> >>> integrated, and then moving the existing Android sync interface over
> >>> to the shared back-end, will allow for proper apples-to-apples
> >>> comparisons of the same interface. And if the functionality isn't
> >>> sufficient we can hold off on merging the sync interface conversion
> >>> until that gets resolved.
> >>>
> >> Yeah, I'm trying to understand the android side too. I think a unified
> interface would benefit both. I'm
> >> toying a bit with the sw_sync driver in staging because it's the
> easiest to try out on my desktop.
> >>
> >> The timeline stuff looks like it could be simplified. The main
> difference that there seems to be is that
> >> I didn't want to create a separate timeline struct for synchronization
> but let the drivers handle it.
> >>
> >> If you use rcu for reference lifetime management of timeline, the kref
> can be dropped. Signalling all
> >> syncpts on timeline destroy to a new destroyed state would kill the
> need for a destroyed member.
> >> The active list is unneeded and can be killed if only a linear
> progression of child_list is allowed.
> >>
> >> Which probably leaves this nice structure:
> >> struct sync_timeline {
> >> const struct sync_timeline_ops*ops;
> >> charname[32];
> >>
> >> struct list_headchild_list_head;
> >> spinlock_tchild_list_lock;
> >>
> >> struct list_headsync_timeline_list;
> >> };
> >>
> >> Where name, and sync_timeline_list are nice for debugging, but I guess
> not necessarily required. so that
> >> could be split out into a separate debugfs thing if required. I've
> moved the pointer to ops to the fence
> >> for dma-fence, which leaves this..
> >>
> >> struct sync_timeline {
> >> struct list_headchild_list_head;
> >> spinlock_tchild_list_lock;
> >>
> >> struct  sync_timeline_debug {
> >> struct list_headsync_timeline_list;
> >> char name[32];
> >> };
> >> };
> >>
> >> Hm, this looks familiar, the drm drivers had some structure for
> protecting the active fence list that has
> >> an identical definition, but with a slightly different list offset..
> >>
> >> struct __wait_queue_head {
> >> spinlock_t lock;
> >> struct list_head task_list;
> >> };
> >>
> >> typedef struct __wait_queue_head wait_queue_head_t;
> >>
> >> This is nicer to convert the existing drm drivers, which already
> implement synchronous wait with a waitqueue.
> >> The default wait op is in fact
> >>
> >> Ok 

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 24-10-13 14:13, Maarten Lankhorst schreef:
> op 09-10-13 16:39, Maarten Lankhorst schreef:
>> Hey,
>>
>>  op 08-10-13 19:37, John Stultz schreef:
>>> On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling  
>>> wrote:
 On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
  wrote:
> Depending on feedback I'll try reflashing my nexus 7 to stock android, 
> and work on trying to convert android
> syncpoints to dma-fence, which I'll probably rename to syncpoints.
 I thought the plan decided at plumbers was to investigate backing
 dma_buf with the android sync solution not the other way around.  It
 doesn't make sense to me to take a working, tested, end-to-end
 solution with a released compositing system built around it, throw it
 out, and replace it with new un-tested code to
 support a system which is not yet built.
>>> Hey Erik,
>>>   Thanks for the clarifying points in your email, your insights and
>>> feedback are critical, and I think having you and Maarten continue to
>>> work out the details here will make this productive.
>>>
>>> My recollection from the discussion was that Rob was ok with trying to
>>> pipe the sync arguments through the various interfaces in order to
>>> support the explicit sync, but I think he did suggest having it backed
>>> by the dma-buf fences underneath.
>>>
>>> I know this can be frustrating to watch things be reimplemented when
>>> you have a pre-baked solution, but some compromise will be needed to
>>> get things merged (and Maarten is taking the initiative here), but its
>>> important to keep discussing this so the *right* compromises are made
>>> that don't hurt performance, etc.
>>>
>>> My hope is Maarten's approach of getting the dma-fence core
>>> integrated, and then moving the existing Android sync interface over
>>> to the shared back-end, will allow for proper apples-to-apples
>>> comparisons of the same interface. And if the functionality isn't
>>> sufficient we can hold off on merging the sync interface conversion
>>> until that gets resolved.
>>>
>> Yeah, I'm trying to understand the android side too. I think a unified 
>> interface would benefit both. I'm
>> toying a bit with the sw_sync driver in staging because it's the easiest to 
>> try out on my desktop.
>>
>> The timeline stuff looks like it could be simplified. The main difference 
>> that there seems to be is that
>> I didn't want to create a separate timeline struct for synchronization but 
>> let the drivers handle it.
>>
>> If you use rcu for reference lifetime management of timeline, the kref can 
>> be dropped. Signalling all
>> syncpts on timeline destroy to a new destroyed state would kill the need for 
>> a destroyed member.
>> The active list is unneeded and can be killed if only a linear progression 
>> of child_list is allowed.
>>
>> Which probably leaves this nice structure:
>> struct sync_timeline {
>> const struct sync_timeline_ops*ops;
>> charname[32];
>>
>> struct list_headchild_list_head;
>> spinlock_tchild_list_lock;
>>
>> struct list_headsync_timeline_list;
>> };
>>
>> Where name, and sync_timeline_list are nice for debugging, but I guess not 
>> necessarily required. so that
>> could be split out into a separate debugfs thing if required. I've moved the 
>> pointer to ops to the fence
>> for dma-fence, which leaves this..
>>
>> struct sync_timeline {
>> struct list_headchild_list_head;
>> spinlock_tchild_list_lock;
>>
>> struct  sync_timeline_debug {
>> struct list_headsync_timeline_list;
>> char name[32];
>> };
>> };
>>
>> Hm, this looks familiar, the drm drivers had some structure for protecting 
>> the active fence list that has
>> an identical definition, but with a slightly different list offset..
>>
>> struct __wait_queue_head {
>> spinlock_t lock;
>> struct list_head task_list;
>> };
>>
>> typedef struct __wait_queue_head wait_queue_head_t;
>>
>> This is nicer to convert the existing drm drivers, which already implement 
>> synchronous wait with a waitqueue.
>> The default wait op is in fact
>>
>> Ok enough of this little excercise. I just wanted to see how different the 2 
>> are. I think even if the
>> fence interface will end up being incompatible it wouldn't be too hard to 
>> convert android..
>>
>> Main difference is the ops, android has a lot more than what I used for 
>> dma-fence:
>>
>> struct fence_ops {
>>  bool (*enable_signaling)(struct fence *fence); // required, callback 
>> called with fence->lock held,
>>  // fence->lock is a pointer passed to __fence_init. Callback should 
>> make sure that the fence will
>>  // be signaled asap.
>>  bool (*signaled)(struct fence *fence); // optional, but if set to NULL 
>> fence_is_signaled is not
>>  // required to ever return true, unless enable_signaling is called, 
>> similar to has_signaled
>>  long (*wait)(struct fence *fence, bool intr, signed long 

[Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

op 09-10-13 16:39, Maarten Lankhorst schreef:
> Hey,
>
>  op 08-10-13 19:37, John Stultz schreef:
>> On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling  wrote:
>>> On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
>>>  wrote:
 Depending on feedback I'll try reflashing my nexus 7 to stock android, and 
 work on trying to convert android
 syncpoints to dma-fence, which I'll probably rename to syncpoints.
>>> I thought the plan decided at plumbers was to investigate backing
>>> dma_buf with the android sync solution not the other way around.  It
>>> doesn't make sense to me to take a working, tested, end-to-end
>>> solution with a released compositing system built around it, throw it
>>> out, and replace it with new un-tested code to
>>> support a system which is not yet built.
>> Hey Erik,
>>   Thanks for the clarifying points in your email, your insights and
>> feedback are critical, and I think having you and Maarten continue to
>> work out the details here will make this productive.
>>
>> My recollection from the discussion was that Rob was ok with trying to
>> pipe the sync arguments through the various interfaces in order to
>> support the explicit sync, but I think he did suggest having it backed
>> by the dma-buf fences underneath.
>>
>> I know this can be frustrating to watch things be reimplemented when
>> you have a pre-baked solution, but some compromise will be needed to
>> get things merged (and Maarten is taking the initiative here), but its
>> important to keep discussing this so the *right* compromises are made
>> that don't hurt performance, etc.
>>
>> My hope is Maarten's approach of getting the dma-fence core
>> integrated, and then moving the existing Android sync interface over
>> to the shared back-end, will allow for proper apples-to-apples
>> comparisons of the same interface. And if the functionality isn't
>> sufficient we can hold off on merging the sync interface conversion
>> until that gets resolved.
>>
> Yeah, I'm trying to understand the android side too. I think a unified 
> interface would benefit both. I'm
> toying a bit with the sw_sync driver in staging because it's the easiest to 
> try out on my desktop.
>
> The timeline stuff looks like it could be simplified. The main difference 
> that there seems to be is that
> I didn't want to create a separate timeline struct for synchronization but 
> let the drivers handle it.
>
> If you use rcu for reference lifetime management of timeline, the kref can be 
> dropped. Signalling all
> syncpts on timeline destroy to a new destroyed state would kill the need for 
> a destroyed member.
> The active list is unneeded and can be killed if only a linear progression of 
> child_list is allowed.
>
> Which probably leaves this nice structure:
> struct sync_timeline {
> const struct sync_timeline_ops*ops;
> charname[32];
>
> struct list_headchild_list_head;
> spinlock_tchild_list_lock;
>
> struct list_headsync_timeline_list;
> };
>
> Where name, and sync_timeline_list are nice for debugging, but I guess not 
> necessarily required. so that
> could be split out into a separate debugfs thing if required. I've moved the 
> pointer to ops to the fence
> for dma-fence, which leaves this..
>
> struct sync_timeline {
> struct list_headchild_list_head;
> spinlock_tchild_list_lock;
>
> struct  sync_timeline_debug {
> struct list_headsync_timeline_list;
> char name[32];
> };
> };
>
> Hm, this looks familiar, the drm drivers had some structure for protecting 
> the active fence list that has
> an identical definition, but with a slightly different list offset..
>
> struct __wait_queue_head {
> spinlock_t lock;
> struct list_head task_list;
> };
>
> typedef struct __wait_queue_head wait_queue_head_t;
>
> This is nicer to convert the existing drm drivers, which already implement 
> synchronous wait with a waitqueue.
> The default wait op is in fact
>
> Ok enough of this little excercise. I just wanted to see how different the 2 
> are. I think even if the
> fence interface will end up being incompatible it wouldn't be too hard to 
> convert android..
>
> Main difference is the ops, android has a lot more than what I used for 
> dma-fence:
>
> struct fence_ops {
>   bool (*enable_signaling)(struct fence *fence); // required, callback 
> called with fence->lock held,
>   // fence->lock is a pointer passed to __fence_init. Callback should 
> make sure that the fence will
>   // be signaled asap.
>   bool (*signaled)(struct fence *fence); // optional, but if set to NULL 
> fence_is_signaled is not
>   // required to ever return true, unless enable_signaling is called, 
> similar to has_signaled
>   long (*wait)(struct fence *fence, bool intr, signed long timeout); // 
> required, but it can be set
>   // to the default fence_default_wait implementation which is 
> recommended. It calls enable_signaling
>   // 

Re: [Linaro-mm-sig] thoughts of looking at android fences

[Maarten Lankhorst]

Hey,

 op 08-10-13 19:37, John Stultz schreef:
 On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling konk...@android.com wrote:
 On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
 maarten.lankho...@canonical.com wrote:
 Depending on feedback I'll try reflashing my nexus 7 to stock android, and 
 work on trying to convert android
 syncpoints to dma-fence, which I'll probably rename to syncpoints.
 I thought the plan decided at plumbers was to investigate backing
 dma_buf with the android sync solution not the other way around.  It
 doesn't make sense to me to take a working, tested, end-to-end
 solution with a released compositing system built around it, throw it
 out, and replace it with new un-tested code to
 support a system which is not yet built.
 Hey Erik,
   Thanks for the clarifying points in your email, your insights and
 feedback are critical, and I think having you and Maarten continue to
 work out the details here will make this productive.

 My recollection from the discussion was that Rob was ok with trying to
 pipe the sync arguments through the various interfaces in order to
 support the explicit sync, but I think he did suggest having it backed
 by the dma-buf fences underneath.

 I know this can be frustrating to watch things be reimplemented when
 you have a pre-baked solution, but some compromise will be needed to
 get things merged (and Maarten is taking the initiative here), but its
 important to keep discussing this so the *right* compromises are made
 that don't hurt performance, etc.

 My hope is Maarten's approach of getting the dma-fence core
 integrated, and then moving the existing Android sync interface over
 to the shared back-end, will allow for proper apples-to-apples
 comparisons of the same interface. And if the functionality isn't
 sufficient we can hold off on merging the sync interface conversion
 until that gets resolved.

Yeah, I'm trying to understand the android side too. I think a unified 
interface would benefit both. I'm
toying a bit with the sw_sync driver in staging because it's the easiest to try 
out on my desktop.

The timeline stuff looks like it could be simplified. The main difference that 
there seems to be is that
I didn't want to create a separate timeline struct for synchronization but let 
the drivers handle it.

If you use rcu for reference lifetime management of timeline, the kref can be 
dropped. Signalling all
syncpts on timeline destroy to a new destroyed state would kill the need for a 
destroyed member.
The active list is unneeded and can be killed if only a linear progression of 
child_list is allowed.

Which probably leaves this nice structure:

struct sync_timeline {
const struct sync_timeline_ops*ops;
charname[32];

struct list_headchild_list_head;
spinlock_tchild_list_lock;

struct list_headsync_timeline_list;
};

Where name, and sync_timeline_list are nice for debugging, but I guess not 
necessarily required. so that
could be split out into a separate debugfs thing if required. I've moved the 
pointer to ops to the fence
for dma-fence, which leaves this..

struct sync_timeline {
struct list_headchild_list_head;
spinlock_tchild_list_lock;

struct  sync_timeline_debug {
struct list_headsync_timeline_list;
char name[32];
};
};

Hm, this looks familiar, the drm drivers had some structure for protecting the 
active fence list that has
an identical definition, but with a slightly different list offset..

struct __wait_queue_head {
spinlock_t lock;
struct list_head task_list;
};

typedef struct __wait_queue_head wait_queue_head_t;

This is nicer to convert the existing drm drivers, which already implement 
synchronous wait with a waitqueue.
The default wait op is in fact

Ok enough of this little excercise. I just wanted to see how different the 2 
are. I think even if the
fence interface will end up being incompatible it wouldn't be too hard to 
convert android..

Main difference is the ops, android has a lot more than what I used for 
dma-fence:

struct fence_ops {
bool (*enable_signaling)(struct fence *fence); // required, callback 
called with fence-lock held,
// fence-lock is a pointer passed to __fence_init. Callback should 
make sure that the fence will
// be signaled asap.
bool (*signaled)(struct fence *fence); // optional, but if set to NULL 
fence_is_signaled is not
// required to ever return true, unless enable_signaling is called, 
similar to has_signaled
long (*wait)(struct fence *fence, bool intr, signed long timeout); // 
required, but it can be set
// to the default fence_default_wait implementation which is 
recommended. It calls enable_signaling
// and appends itself to async callback list. Identical semantics to 
wait_event_interruptible_timeout.
void (*release)(struct fence *fence); // free_pt
};

Because every fence has a stamp, there is no need for a 

Re: [Linaro-mm-sig] thoughts of looking at android fences

[John Stultz]

On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling konk...@android.com wrote:
 On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
 maarten.lankho...@canonical.com wrote:
 Depending on feedback I'll try reflashing my nexus 7 to stock android, and 
 work on trying to convert android
 syncpoints to dma-fence, which I'll probably rename to syncpoints.

 I thought the plan decided at plumbers was to investigate backing
 dma_buf with the android sync solution not the other way around.  It
 doesn't make sense to me to take a working, tested, end-to-end
 solution with a released compositing system built around it, throw it
 out, and replace it with new un-tested code to
 support a system which is not yet built.

Hey Erik,
  Thanks for the clarifying points in your email, your insights and
feedback are critical, and I think having you and Maarten continue to
work out the details here will make this productive.

My recollection from the discussion was that Rob was ok with trying to
pipe the sync arguments through the various interfaces in order to
support the explicit sync, but I think he did suggest having it backed
by the dma-buf fences underneath.

I know this can be frustrating to watch things be reimplemented when
you have a pre-baked solution, but some compromise will be needed to
get things merged (and Maarten is taking the initiative here), but its
important to keep discussing this so the *right* compromises are made
that don't hurt performance, etc.

My hope is Maarten's approach of getting the dma-fence core
integrated, and then moving the existing Android sync interface over
to the shared back-end, will allow for proper apples-to-apples
comparisons of the same interface. And if the functionality isn't
sufficient we can hold off on merging the sync interface conversion
until that gets resolved.

Does that sound reasonable?

thanks
-john
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [Linaro-mm-sig] thoughts of looking at android fences

[Rob Clark]

On Wed, Oct 2, 2013 at 2:13 PM, Erik Gilling konk...@android.com wrote:
 The lock is provided by the driver, which makes adding support for old 
 hardware that has no reliable way of notifying completion of events easier.

 I'm a bit confused here how it's possible to implement sync on
 hardware with no reliable way of notifying completion of events.  That
 seems like a non-starter to me.

I suspect Maarten meant no reliable way of notifying completion *to
the cpu*.. which isn't strictly needed for gpu-gpu sharing.

BR,
-R
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [Linaro-mm-sig] thoughts of looking at android fences

[Rob Clark]

On Tue, Oct 8, 2013 at 1:37 PM, John Stultz john.stu...@linaro.org wrote:
 On Wed, Oct 2, 2013 at 11:13 AM, Erik Gilling konk...@android.com wrote:
 On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
 maarten.lankho...@canonical.com wrote:
 Depending on feedback I'll try reflashing my nexus 7 to stock android, and 
 work on trying to convert android
 syncpoints to dma-fence, which I'll probably rename to syncpoints.

 I thought the plan decided at plumbers was to investigate backing
 dma_buf with the android sync solution not the other way around.  It
 doesn't make sense to me to take a working, tested, end-to-end
 solution with a released compositing system built around it, throw it
 out, and replace it with new un-tested code to
 support a system which is not yet built.

 Hey Erik,
   Thanks for the clarifying points in your email, your insights and
 feedback are critical, and I think having you and Maarten continue to
 work out the details here will make this productive.

 My recollection from the discussion was that Rob was ok with trying to
 pipe the sync arguments through the various interfaces in order to
 support the explicit sync, but I think he did suggest having it backed
 by the dma-buf fences underneath.

Yeah, my comment was mainly about userspace API for different driver
subsystems.  I'd rather add some extra parameter(s?) to drm and v4l
ioctls, even if they are unused by linux userspace, vs having
different ABI for android kernel vs linux kernel.

We probably do however need the zero value to indicate unusued.. at
least for adding new parameters to existing drm ioctls since
drm_ioctl() will be zero'ing stuff out to deal w/ new userspace / old
kernel, or old userspace / new kernel combos.  For new ioctls (like
'atomic') we don't have this constraint.

BR,
-R

 I know this can be frustrating to watch things be reimplemented when
 you have a pre-baked solution, but some compromise will be needed to
 get things merged (and Maarten is taking the initiative here), but its
 important to keep discussing this so the *right* compromises are made
 that don't hurt performance, etc.

 My hope is Maarten's approach of getting the dma-fence core
 integrated, and then moving the existing Android sync interface over
 to the shared back-end, will allow for proper apples-to-apples
 comparisons of the same interface. And if the functionality isn't
 sufficient we can hold off on merging the sync interface conversion
 until that gets resolved.

 Does that sound reasonable?

 thanks
 -john
 ___
 dri-devel mailing list
 dri-devel@lists.freedesktop.org
 http://lists.freedesktop.org/mailman/listinfo/dri-devel
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [Linaro-mm-sig] thoughts of looking at android fences

[Erik Gilling]

On Wed, Oct 2, 2013 at 12:35 AM, Maarten Lankhorst
maarten.lankho...@canonical.com wrote:
 The timeline is similar to what I called a fence context. Each command stream 
 on a gpu can have a context. Because
 nvidia hardware can have 4095 separate timelines, I didn't want to keep the 
 bookkeeping for each timeline, although
 I guess that it's already done. Maybe it could be done in a unified way for 
 each driver, making a transition to
 timelines that can be used by android easier.

 I did not have an explicit syncpoint addition, but I think that sync points + 
 sync_fence were similar to what I did with
 my dma-fence stuff, except slightly different.
 In my approach the dma-fence is signaled after all sync_points are done AND 
 the queued commands are executed.
 In effect the dma-fence becomes the next syncpoint, depending on all previous 
 dma-fence syncpoints.

What makes queued command completion different than any other sync point?

 An important thing to note is that dma-fence is kernelspace only, so it might 
 be better to rename it to syncpoint,
 and use fence for the userspace interface.

 A big difference is locking, I assume in my code that most fences emitted are 
 not waited on, so the fastpath
 fence_signal is a test_and_set_bit plus test_bit. A single lock is used for 
 the waitqueue and callbacks,
 with the waitqueue being implemented internally as an asynchronous callback.

I assume very little lock contention so the performance impact is
negligible.  Also, because sync_pts on a timeline are strictly
ordered, it's necessary to check all active pts on a timeline signal.
A future optimization could involve keeping active pts in a sorted
list or other data structure so that you only need to iterate over the
pts that are about to signal.  So far we've not seen any bottlenecks
here so I've kept it simple.

 The lock is provided by the driver, which makes adding support for old 
 hardware that has no reliable way of notifying completion of events easier.

I'm a bit confused here how it's possible to implement sync on
hardware with no reliable way of notifying completion of events.  That
seems like a non-starter to me.

 I avoided using global locks, but I think for debugfs support I may end up 
 having to add some.

As did I, except for debugfs support.

 One thing though: is it really required to merge fences? It seems to me that 
 if I add a poll callback userspace
 could simply do a poll on a list of fences. This would give userspace all the 
 information it needs about each
 individual fence.

This is very important.  It greatly simplifies they way the userspace
deals with fences.  It means that it only has to track one fd per
buffer and both the kernel API and userspace RPC apis don't have to
take a variable number of fds per buffer.  FWIW the android sync
driver already implements poll.

 Depending on feedback I'll try reflashing my nexus 7 to stock android, and 
 work on trying to convert android
 syncpoints to dma-fence, which I'll probably rename to syncpoints.

I thought the plan decided at plumbers was to investigate backing
dma_buf with the android sync solution not the other way around.  It
doesn't make sense to me to take a working, tested, end-to-end
solution with a released compositing system built around it, throw it
out, and replace it with new un-tested code to
support a system which is not yet built.

Cheers,
   Erik
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel