Nouveau - Sep 2014 - [RFC PATCH 7/7] drm/prime: Support explicit fence on export

Hi guys,


I'd like to start a new thread about explicit fence synchronization.  This
time
with a Nouveau twist. :-)

First, let me define what I understand by implicit/explicit sync:

Implicit synchronization
* Fences are attached to buffers
* Kernel manages fences automatically based on buffer read/write access

Explicit synchronization
* Fences are passed around independently
* Kernel takes and emits fences to/from user space when submitting work

Implicit synchronization is already implemented in open source drivers, and
works well for most use cases.  I don't seek to change any of that.  My
proposal aims at allowing some drm drivers to operate in explicit sync mode to
get maximal performance, while still remaining fully compatible with the
implicit paradigm.

I will try to explain why I think we should support the explicit model as well.


1. Bindless graphics

Bindless graphics is a central concept when trying to reduce the OpenGL driver
overhead.  The idea is that the application can bind a large set of buffers to
the working set up front using extensions such as GL_ARB_bindless_texture, and
they remain resident until the application releases them (note that compute
APIs have typically similar semantics).  These working sets can be huge,
hundreds or even thousands of buffers, so we would like to opt out from the
per-submit overhead of acquiring locks, waiting for fences, and storing fences.
Automatically synchronizing these working sets in kernel will also prevent
parallelism between channels that are sharing the working set (in fact sharing
just one buffer from the working set will cause the jobs of the two channels to
be serialized).

2. Evolution of graphics APIs

The graphics API evolution seems to be going to a direction where game engine
and middleware vendors demand more control over work submission and
synchronization.  We expect that this trend will continue, and more and more
synchronization decisions will be pushed to the API level.  OpenGL and EGL
already provide good explicit command stream level synchronization primitives:
glFenceSync and EGL_KHR_wait_sync.  Their use is also encouraged - for example
EGL_KHR_image_base spec clearly states that the application is responsible for
synchronizing accesses to EGLImages.  If the API that is exposed to developers
gives the control over synchronization to the developer, then implicit waits
that are inserted by the kernel are unnecessary and unexpected, and can
severely hurt performance.  It also makes it easy for the developer to write
code that happens to work on Linux because of implicit sync, but will fail on
other platforms.

3. Suballocation

Using user space suballocation can help reduce the overhead when a large number
of small textures are used.  Synchronizing suballocated surfaces implicitly in
kernel doesn't make sense - many channels should be able to access the same
kernel-level buffer object simultaneously.

4. Buffer sharing complications

This is not really an argument for explicit sync as such, but I'd like to
point
out that sharing buffers across SoC engines is often much more complex than
just exporting and importing a dma-buf and waiting for the dma-buf fences.
Sometimes we need to do color format or tiling layout conversion.  Sometimes,
at least on Tegra, we need to decompress buffers when we pass them from the GPU
to an engine that doesn't support framebuffer compression.  These things are
not uncommon, particularly when we have SoC's that combine licensed IP
blocks
from different vendors.  My point is that user space is already heavily
involved when sharing buffers between drivers, and giving it some more control
over synchronization is not adding that much complexity.


Because of the above arguments, I think it makes sense to let some user space
drm drivers opt out from implicit synchronization, while allowing them to still
remain fully compatible with the rest of the drm world that uses implicit
synchronization.  In practice, this would require three things:

(1) Support passing fences (that are not tied to buffer objects) between kernel
    and user space.

(2) Stop automatically storing fences to the buffers that user space wants to
    synchronize explicitly.

(3) Allow user space to attach an explicit fence to dma-buf when exporting to
    another driver that uses implicit sync.

There are still some open issues beyond these.  For example, can we skip
acquiring the ww mutex for explicitly synchronized buffers?  I think we could
eventually, at least on unified memory systems where we don't need to
migrate
between heaps (our downstream Tegra GPU driver does not lock any buffers at
submit, it just grabs refcounts for hw).  Another quirk is that now Nouveau
waits on the buffer fences when closing the gem object to ensure that it
doesn't unmap too early.  We need to rework that for explicit sync, but that
shouldn't be difficult.

I have written a prototype that demonstrates (1) by adding explicit sync fd
support to Nouveau.  It's not a lot of code, because I only use a relatively
small subset of the android sync driver functionality.  Thanks to Maarten's
rewrite, all I need to do is to allow creating a sync_fence from a drm fence in
order to pass it to user space.  I don't need to use sync_pt or
sync_timeline,
or fill in sync_timeline_ops.

I can see why the upstream has been reluctant to de-stage the android sync
driver in its current form, since (even though it now builds on struct fence)
it still duplicates some of the drm fence concepts.  I'd like to think that
my
patches only use the parts of the android sync driver that genuinely are
missing from the drm fence model: allowing user space to operate on fence
objects that are independent of buffer objects.

The last two patches are mocks that show how (2) and (3) might work out.  I
haven't done any testing with them yet.  Before going any further, I'd
like to
get your feedback.  Can you see the benefits of explicit sync as an alternative
synchronization model?  Do you think we could use the android sync_fence for
passing fences between user space?  Or did you have something else in mind for
explicit sync in the drm world?


Thanks,
Lauri


Lauri Peltonen (7):
  android: Support creating sync fence from drm fences
  drm/nouveau: Split nouveau_fence_sync
  drm/nouveau: Add fence fd helpers
  drm/nouveau: Support fence fd's at kickoff
  libdrm: nouveau: Support fence fds
  drm/nouveau: Support marking buffers for explicit sync
  drm/prime: Support explicit fence on export

-- 
1.8.1.5

Modify sync_fence_create to accept an array of 'struct fence' objects.
This will allow drm drivers to create sync_fence objects and pass sync
fd's between user space with minimal modifications, without ever creating
sync_timeline or sync_pt objects, and without implementing the
sync_timeline_ops interface.

Modify the sync driver debug code to not assume that every 'struct
fence'
(that is associated with a 'struct sync_fence') is embedded within a
'struct sync_pt'.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drivers/staging/android/sw_sync.c    |  3 ++-
 drivers/staging/android/sync.c       | 34 ++++++++++++++++++---------------
 drivers/staging/android/sync.h       | 11 ++++++-----
 drivers/staging/android/sync_debug.c | 37 +++++++++++++++++-------------------
 4 files changed, 44 insertions(+), 41 deletions(-)

diff --git a/drivers/staging/android/sw_sync.c
b/drivers/staging/android/sw_sync.c
index a76db3f..6949812 100644
--- a/drivers/staging/android/sw_sync.c
+++ b/drivers/staging/android/sw_sync.c
@@ -184,7 +184,8 @@ static long sw_sync_ioctl_create_fence(struct
sw_sync_timeline *obj,
 	}
 
 	data.name[sizeof(data.name) - 1] = '\0';
-	fence = sync_fence_create(data.name, pt);
+	fence = sync_fence_create(data.name,
+				  (struct fence *[]){ &pt->base }, 1);
 	if (fence == NULL) {
 		sync_pt_free(pt);
 		err = -ENOMEM;
diff --git a/drivers/staging/android/sync.c b/drivers/staging/android/sync.c
index e7b2e02..1d0d968 100644
--- a/drivers/staging/android/sync.c
+++ b/drivers/staging/android/sync.c
@@ -187,28 +187,32 @@ static void fence_check_cb_func(struct fence *f, struct
fence_cb *cb)
 		wake_up_all(&fence->wq);
 }
 
-/* TODO: implement a create which takes more that one sync_pt */
-struct sync_fence *sync_fence_create(const char *name, struct sync_pt *pt)
+struct sync_fence *sync_fence_create(const char *name,
+				     struct fence **fences, int num_fences)
 {
-	struct sync_fence *fence;
+	struct sync_fence *sync_fence;
+	int size = offsetof(struct sync_fence, cbs[num_fences]);
+	int i;
 
-	fence = sync_fence_alloc(offsetof(struct sync_fence, cbs[1]), name);
-	if (fence == NULL)
+	sync_fence = sync_fence_alloc(size, name);
+	if (sync_fence == NULL)
 		return NULL;
 
-	fence->num_fences = 1;
-	atomic_set(&fence->status, 1);
+	sync_fence->num_fences = num_fences;
+	atomic_set(&sync_fence->status, 0);
 
-	fence_get(&pt->base);
-	fence->cbs[0].sync_pt = &pt->base;
-	fence->cbs[0].fence = fence;
-	if (fence_add_callback(&pt->base, &fence->cbs[0].cb,
-			       fence_check_cb_func))
-		atomic_dec(&fence->status);
+	for (i = 0; i < num_fences; i++) {
+		struct fence *f = fences[i];
+		struct sync_fence_cb *cb = &sync_fence->cbs[i];
 
-	sync_fence_debug_add(fence);
+		cb->sync_pt = fence_get(f);
+		cb->fence = sync_fence;
+		if (!fence_add_callback(f, &cb->cb, fence_check_cb_func))
+			atomic_inc(&sync_fence->status);
+	}
+	sync_fence_debug_add(sync_fence);
 
-	return fence;
+	return sync_fence;
 }
 EXPORT_SYMBOL(sync_fence_create);
 
diff --git a/drivers/staging/android/sync.h b/drivers/staging/android/sync.h
index 66b0f43..b8ad72c 100644
--- a/drivers/staging/android/sync.h
+++ b/drivers/staging/android/sync.h
@@ -246,13 +246,14 @@ void sync_pt_free(struct sync_pt *pt);
 
 /**
  * sync_fence_create() - creates a sync fence
- * @name:	name of fence to create
- * @pt:		sync_pt to add to the fence
+ * @name:	name of the sync fence to create
+ * @fences:	fences to add to the sync fence
+ * @num_fences:	the number of fences in the @fences array
  *
- * Creates a fence containg @pt.  Once this is called, the fence takes
- * ownership of @pt.
+ * Creates a sync fence from an array of drm fences. Takes refs to @fences.
  */
-struct sync_fence *sync_fence_create(const char *name, struct sync_pt *pt);
+struct sync_fence *sync_fence_create(const char *name,
+				     struct fence **fences, int num_fences);
 
 /*
  * API for sync_fence consumers
diff --git a/drivers/staging/android/sync_debug.c
b/drivers/staging/android/sync_debug.c
index 257fc91..2d8873e 100644
--- a/drivers/staging/android/sync_debug.c
+++ b/drivers/staging/android/sync_debug.c
@@ -81,33 +81,33 @@ static const char *sync_status_str(int status)
 	return "error";
 }
 
-static void sync_print_pt(struct seq_file *s, struct sync_pt *pt, bool fence)
+static void sync_print_pt(struct seq_file *s, struct fence *pt, bool fence)
 {
 	int status = 1;
-	struct sync_timeline *parent = sync_pt_parent(pt);
 
-	if (fence_is_signaled_locked(&pt->base))
-		status = pt->base.status;
+	if (fence_is_signaled_locked(pt))
+		status = pt->status;
 
-	seq_printf(s, "  %s%spt %s",
-		   fence ? parent->name : "",
-		   fence ? "_" : "",
-		   sync_status_str(status));
+	if (fence)
+		seq_printf(s, "  %d_pt %s", pt->context,
+			   sync_status_str(status));
+	else
+		seq_printf(s, "  pt %s", sync_status_str(status));
 
 	if (status <= 0) {
-		struct timeval tv = ktime_to_timeval(pt->base.timestamp);
+		struct timeval tv = ktime_to_timeval(pt->timestamp);
 
 		seq_printf(s, "@%ld.%06ld", tv.tv_sec, tv.tv_usec);
 	}
 
-	if (parent->ops->timeline_value_str &&
-	    parent->ops->pt_value_str) {
+	if (pt->ops->timeline_value_str &&
+	    pt->ops->fence_value_str) {
 		char value[64];
 
-		parent->ops->pt_value_str(pt, value, sizeof(value));
+		pt->ops->fence_value_str(pt, value, sizeof(value));
 		seq_printf(s, ": %s", value);
 		if (fence) {
-			parent->ops->timeline_value_str(parent, value,
+			pt->ops->timeline_value_str(pt, value,
 						    sizeof(value));
 			seq_printf(s, " / %s", value);
 		}
@@ -121,7 +121,8 @@ static void sync_print_obj(struct seq_file *s, struct
sync_timeline *obj)
 	struct list_head *pos;
 	unsigned long flags;
 
-	seq_printf(s, "%s %s", obj->name, obj->ops->driver_name);
+	seq_printf(s, "%d %s %s", obj->context, obj->name,
+		   obj->ops->driver_name);
 
 	if (obj->ops->timeline_value_str) {
 		char value[64];
@@ -136,7 +137,7 @@ static void sync_print_obj(struct seq_file *s, struct
sync_timeline *obj)
 	list_for_each(pos, &obj->child_list_head) {
 		struct sync_pt *pt  			container_of(pos, struct sync_pt, child_list);
-		sync_print_pt(s, pt, false);
+		sync_print_pt(s, &pt->base, false);
 	}
 	spin_unlock_irqrestore(&obj->child_list_lock, flags);
 }
@@ -151,11 +152,7 @@ static void sync_print_fence(struct seq_file *s, struct
sync_fence *fence)
 		   sync_status_str(atomic_read(&fence->status)));
 
 	for (i = 0; i < fence->num_fences; ++i) {
-		struct sync_pt *pt -			container_of(fence->cbs[i].sync_pt,
-				     struct sync_pt, base);
-
-		sync_print_pt(s, pt, true);
+		sync_print_pt(s, fence->cbs[i].sync_pt, true);
 	}
 
 	spin_lock_irqsave(&fence->wq.lock, flags);
-- 
1.8.1.5

Split nouveau_fence_sync to two functions:

 * nouveau_fence_sync, which only adds a fence wait to the channel
   command stream, and
 * nouveau_bo_sync, which gets the fences from the reservation object
   and passes them to nouveau_fence_sync.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drm/nouveau_bo.c      | 38 +++++++++++++++++++++++++++++++++++-
 drm/nouveau_bo.h      |  2 ++
 drm/nouveau_display.c |  4 ++--
 drm/nouveau_fence.c   | 54 ++++++++-------------------------------------------
 drm/nouveau_fence.h   |  2 +-
 drm/nouveau_gem.c     |  2 +-
 6 files changed, 51 insertions(+), 51 deletions(-)

diff --git a/drm/nouveau_bo.c b/drm/nouveau_bo.c
index 70c3cb5..534734a 100644
--- a/drm/nouveau_bo.c
+++ b/drm/nouveau_bo.c
@@ -463,6 +463,42 @@ nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
 }
 
 int
+nouveau_bo_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool
exclusive)
+{
+	struct fence *fence;
+	struct reservation_object *resv = nvbo->bo.resv;
+	struct reservation_object_list *fobj;
+	int ret = 0, i;
+
+	if (!exclusive) {
+		ret = reservation_object_reserve_shared(resv);
+
+		if (ret)
+			return ret;
+	}
+
+	fobj = reservation_object_get_list(resv);
+	fence = reservation_object_get_excl(resv);
+
+	if (fence && (!exclusive || !fobj || !fobj->shared_count))
+		return nouveau_fence_sync(fence, chan);
+
+	if (!exclusive || !fobj)
+		return ret;
+
+	for (i = 0; i < fobj->shared_count && !ret; ++i) {
+		fence = rcu_dereference_protected(fobj->shared[i],
+						reservation_object_held(resv));
+		ret = nouveau_fence_sync(fence, chan);
+
+		if (ret)
+			break;
+	}
+
+	return ret;
+}
+
+int
 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
 		    bool no_wait_gpu)
 {
@@ -1070,7 +1106,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int
evict, bool intr,
 	}
 
 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
-	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true);
+	ret = nouveau_bo_sync(nouveau_bo(bo), chan, true);
 	if (ret == 0) {
 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
 		if (ret == 0) {
diff --git a/drm/nouveau_bo.h b/drm/nouveau_bo.h
index 8ab877b..f97bc26 100644
--- a/drm/nouveau_bo.h
+++ b/drm/nouveau_bo.h
@@ -84,6 +84,8 @@ int  nouveau_bo_validate(struct nouveau_bo *, bool
interruptible,
 			 bool no_wait_gpu);
 void nouveau_bo_sync_for_device(struct nouveau_bo *nvbo);
 void nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo);
+int  nouveau_bo_sync(struct nouveau_bo *, struct nouveau_channel *,
+		     bool exclusive);
 
 struct nouveau_vma *
 nouveau_bo_vma_find(struct nouveau_bo *, struct nouveau_vm *);
diff --git a/drm/nouveau_display.c b/drm/nouveau_display.c
index 6d0a3cd..41b130c 100644
--- a/drm/nouveau_display.c
+++ b/drm/nouveau_display.c
@@ -658,7 +658,7 @@ nouveau_page_flip_emit(struct nouveau_channel *chan,
 	spin_unlock_irqrestore(&dev->event_lock, flags);
 
 	/* Synchronize with the old framebuffer */
-	ret = nouveau_fence_sync(old_bo, chan, false);
+	ret = nouveau_bo_sync(old_bo, chan, false);
 	if (ret)
 		goto fail;
 
@@ -722,7 +722,7 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct
drm_framebuffer *fb,
 		goto fail_unpin;
 
 	/* synchronise rendering channel with the kernel's channel */
-	ret = nouveau_fence_sync(new_bo, chan, false);
+	ret = nouveau_bo_sync(new_bo, chan, false);
 	if (ret) {
 		ttm_bo_unreserve(&new_bo->bo);
 		goto fail_unpin;
diff --git a/drm/nouveau_fence.c b/drm/nouveau_fence.c
index decfe6c..39b5436 100644
--- a/drm/nouveau_fence.c
+++ b/drm/nouveau_fence.c
@@ -342,57 +342,19 @@ nouveau_fence_wait(struct nouveau_fence *fence, bool lazy,
bool intr)
 }
 
 int
-nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool
exclusive)
+nouveau_fence_sync(struct fence *fence, struct nouveau_channel *chan)
 {
 	struct nouveau_fence_chan *fctx = chan->fence;
-	struct fence *fence;
-	struct reservation_object *resv = nvbo->bo.resv;
-	struct reservation_object_list *fobj;
+	struct nouveau_channel *prev = NULL;
 	struct nouveau_fence *f;
-	int ret = 0, i;
-
-	if (!exclusive) {
-		ret = reservation_object_reserve_shared(resv);
-
-		if (ret)
-			return ret;
-	}
-
-	fobj = reservation_object_get_list(resv);
-	fence = reservation_object_get_excl(resv);
-
-	if (fence && (!exclusive || !fobj || !fobj->shared_count)) {
-		struct nouveau_channel *prev = NULL;
-
-		f = nouveau_local_fence(fence, chan->drm);
-		if (f)
-			prev = f->channel;
-
-		if (!prev || (prev != chan && (ret = fctx->sync(f, prev, chan))))
-			ret = fence_wait(fence, true);
-
-		return ret;
-	}
-
-	if (!exclusive || !fobj)
-		return ret;
-
-	for (i = 0; i < fobj->shared_count && !ret; ++i) {
-		struct nouveau_channel *prev = NULL;
-
-		fence = rcu_dereference_protected(fobj->shared[i],
-						reservation_object_held(resv));
-
-		f = nouveau_local_fence(fence, chan->drm);
-		if (f)
-			prev = f->channel;
+	int ret = 0;
 
-		if (!prev || (ret = fctx->sync(f, prev, chan)))
-			ret = fence_wait(fence, true);
+	f = nouveau_local_fence(fence, chan->drm);
+	if (f)
+		prev = f->channel;
 
-		if (ret)
-			break;
-	}
+	if (!prev || (prev != chan && (ret = fctx->sync(f, prev, chan))))
+		ret = fence_wait(fence, true);
 
 	return ret;
 }
diff --git a/drm/nouveau_fence.h b/drm/nouveau_fence.h
index 986c813..8b70166 100644
--- a/drm/nouveau_fence.h
+++ b/drm/nouveau_fence.h
@@ -26,7 +26,7 @@ int  nouveau_fence_emit(struct nouveau_fence *, struct
nouveau_channel *);
 bool nouveau_fence_done(struct nouveau_fence *);
 void nouveau_fence_work(struct fence *, void (*)(void *), void *);
 int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
-int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool
exclusive);
+int  nouveau_fence_sync(struct fence *, struct nouveau_channel *);
 
 struct nouveau_fence_chan {
 	spinlock_t lock;
diff --git a/drm/nouveau_gem.c b/drm/nouveau_gem.c
index a43af30..78398d4 100644
--- a/drm/nouveau_gem.c
+++ b/drm/nouveau_gem.c
@@ -459,7 +459,7 @@ validate_list(struct nouveau_channel *chan, struct
nouveau_cli *cli,
 			return ret;
 		}
 
-		ret = nouveau_fence_sync(nvbo, chan, !!b->write_domains);
+		ret = nouveau_bo_sync(nvbo, chan, !!b->write_domains);
 		if (unlikely(ret)) {
 			if (ret != -ERESTARTSYS)
 				NV_PRINTK(error, cli, "fail post-validate sync\n");
-- 
1.8.1.5

Add nouveau_fence_install, which installs a drm fence as a file
descriptor that can be returned to user space.

Add nouveau_fence_sync_fd, which pushes semaphore wait commands for
each Nouveau fence contained within the sync fd.  If the sync fd
contains non-Nouveau fences, those are waited on the CPU.

Add missing fence_value_str and timeline_value_str callbacks to
nouveau fence_ops.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drm/nouveau_fence.c | 71 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 drm/nouveau_fence.h |  2 ++
 2 files changed, 72 insertions(+), 1 deletion(-)

diff --git a/drm/nouveau_fence.c b/drm/nouveau_fence.c
index 39b5436..e67d467 100644
--- a/drm/nouveau_fence.c
+++ b/drm/nouveau_fence.c
@@ -37,6 +37,8 @@
 #include "nouveau_dma.h"
 #include "nouveau_fence.h"
 
+#include "../drivers/staging/android/sync.h"
+
 static const struct fence_ops nouveau_fence_ops_uevent;
 static const struct fence_ops nouveau_fence_ops_legacy;
 
@@ -471,11 +473,78 @@ static bool nouveau_fence_enable_signaling(struct fence
*f)
 	return ret;
 }
 
+static void nouveau_fence_timeline_value_str(struct fence *fence,
+					     char *str, int size)
+{
+	struct nouveau_fence *f = from_fence(fence);
+	struct nouveau_fence_chan *fctx = nouveau_fctx(f);
+	u32 cur;
+
+	cur = f->channel ? fctx->read(f->channel) : 0;
+	snprintf(str, size, "%d", cur);
+}
+
+static void
+nouveau_fence_value_str(struct fence *fence, char *str, int size)
+{
+	snprintf(str, size, "%d", fence->seqno);
+}
+
 static const struct fence_ops nouveau_fence_ops_uevent = {
 	.get_driver_name = nouveau_fence_get_get_driver_name,
 	.get_timeline_name = nouveau_fence_get_timeline_name,
 	.enable_signaling = nouveau_fence_enable_signaling,
 	.signaled = nouveau_fence_is_signaled,
 	.wait = fence_default_wait,
-	.release = NULL
+	.release = NULL,
+	.fence_value_str = nouveau_fence_value_str,
+	.timeline_value_str = nouveau_fence_timeline_value_str,
 };
+
+int
+nouveau_fence_install(struct fence *fence, const char *name, int *fd_out)
+{
+#ifdef CONFIG_SYNC
+	struct sync_fence *f;
+	int fd;
+
+	f = sync_fence_create(name, &fence, 1);
+	if (!f)
+		return -ENOMEM;
+
+	fd = get_unused_fd();
+	if (fd < 0) {
+		sync_fence_put(f);
+		return fd;
+	}
+	sync_fence_install(f, fd);
+	*fd_out = fd;
+	return 0;
+#else
+	return -ENODEV;
+#endif
+}
+
+int
+nouveau_fence_sync_fd(int fence_fd, struct nouveau_channel *chan)
+{
+#ifdef CONFIG_SYNC
+	int i, ret = 0;
+	struct sync_fence *fence;
+
+	fence = sync_fence_fdget(fence_fd);
+	if (!fence)
+		return -EINVAL;
+
+	for (i = 0; i < fence->num_fences; ++i) {
+		struct fence *pt = fence->cbs[i].sync_pt;
+
+		ret = nouveau_fence_sync(pt, chan);
+		if (ret)
+			break;
+	}
+	sync_fence_put(fence);
+	return ret;
+#else
+	return -ENODEV;
+#endif
+}
diff --git a/drm/nouveau_fence.h b/drm/nouveau_fence.h
index 8b70166..76342ea 100644
--- a/drm/nouveau_fence.h
+++ b/drm/nouveau_fence.h
@@ -27,6 +27,8 @@ bool nouveau_fence_done(struct nouveau_fence *);
 void nouveau_fence_work(struct fence *, void (*)(void *), void *);
 int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
 int  nouveau_fence_sync(struct fence *, struct nouveau_channel *);
+int  nouveau_fence_sync_fd(int, struct nouveau_channel *);
+int  nouveau_fence_install(struct fence *, const char *name, int *);
 
 struct nouveau_fence_chan {
 	spinlock_t lock;
-- 
1.8.1.5

Add a new NOUVEAU_GEM_PUSHBUF_2 ioctl that accepts and emits a sync
fence fd from/to user space if the user space requests it by passing
corresponding flags.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drm/nouveau_drm.c          |  1 +
 drm/nouveau_gem.c          | 46 ++++++++++++++++++++++++++++++++++++++++++----
 drm/nouveau_gem.h          |  2 ++
 drm/uapi/drm/nouveau_drm.h | 11 +++++++++++
 4 files changed, 56 insertions(+), 4 deletions(-)

diff --git a/drm/nouveau_drm.c b/drm/nouveau_drm.c
index 244d78f..74d5ac6 100644
--- a/drm/nouveau_drm.c
+++ b/drm/nouveau_drm.c
@@ -812,6 +812,7 @@ nouveau_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GPUOBJ_FREE, nouveau_abi16_ioctl_gpuobj_free,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_NEW, nouveau_gem_ioctl_new,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_PUSHBUF, nouveau_gem_ioctl_pushbuf,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_PUSHBUF_2, nouveau_gem_ioctl_pushbuf_2,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_PREP, nouveau_gem_ioctl_cpu_prep,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_FINI, nouveau_gem_ioctl_cpu_fini,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_INFO, nouveau_gem_ioctl_info,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
diff --git a/drm/nouveau_gem.c b/drm/nouveau_gem.c
index 78398d4..ee5782c 100644
--- a/drm/nouveau_gem.c
+++ b/drm/nouveau_gem.c
@@ -636,15 +636,16 @@ nouveau_gem_pushbuf_reloc_apply(struct nouveau_cli *cli,
 	return ret;
 }
 
-int
-nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
-			  struct drm_file *file_priv)
+static int
+__nouveau_gem_ioctl_pushbuf(struct drm_device *dev,
+			    struct drm_nouveau_gem_pushbuf *req,
+			    struct drm_nouveau_gem_pushbuf_2 *req_2,
+			    struct drm_file *file_priv)
 {
 	struct nouveau_abi16 *abi16 = nouveau_abi16_get(file_priv, dev);
 	struct nouveau_cli *cli = nouveau_cli(file_priv);
 	struct nouveau_abi16_chan *temp;
 	struct nouveau_drm *drm = nouveau_drm(dev);
-	struct drm_nouveau_gem_pushbuf *req = data;
 	struct drm_nouveau_gem_pushbuf_push *push;
 	struct drm_nouveau_gem_pushbuf_bo *bo;
 	struct nouveau_channel *chan = NULL;
@@ -725,6 +726,14 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void
*data,
 		}
 	}
 
+	if (req_2 && (req_2->flags & NOUVEAU_GEM_PUSHBUF_2_FENCE_WAIT))
{
+		ret = nouveau_fence_sync_fd(req_2->fence, chan);
+		if (ret) {
+			NV_PRINTK(error, cli, "fence wait: %d\n", ret);
+			goto out;
+		}
+	}
+
 	if (chan->dma.ib_max) {
 		ret = nouveau_dma_wait(chan, req->nr_push + 1, 16);
 		if (ret) {
@@ -800,6 +809,16 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void
*data,
 		goto out;
 	}
 
+	if (req_2 && (req_2->flags & NOUVEAU_GEM_PUSHBUF_2_FENCE_EMIT))
{
+		ret = nouveau_fence_install(&fence->base, "nv-pushbuf",
+					    &req_2->fence);
+		if (ret) {
+			NV_PRINTK(error, cli, "fence install: %d\n", ret);
+			WIND_RING(chan);
+			goto out;
+		}
+	}
+
 out:
 	validate_fini(&op, fence, bo);
 	nouveau_fence_unref(&fence);
@@ -825,6 +844,25 @@ out_next:
 	return nouveau_abi16_put(abi16, ret);
 }
 
+int
+nouveau_gem_ioctl_pushbuf_2(struct drm_device *dev, void *data,
+			    struct drm_file *file_priv)
+{
+	struct drm_nouveau_gem_pushbuf_2 *req_2 = data;
+	struct drm_nouveau_gem_pushbuf *req = &req_2->base;
+
+	return __nouveau_gem_ioctl_pushbuf(dev, req, req_2, file_priv);
+}
+
+int
+nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
+			  struct drm_file *file_priv)
+{
+	struct drm_nouveau_gem_pushbuf *req = data;
+
+	return __nouveau_gem_ioctl_pushbuf(dev, req, NULL, file_priv);
+}
+
 static inline uint32_t
 domain_to_ttm(struct nouveau_bo *nvbo, uint32_t domain)
 {
diff --git a/drm/nouveau_gem.h b/drm/nouveau_gem.h
index ddab762..7454dea 100644
--- a/drm/nouveau_gem.h
+++ b/drm/nouveau_gem.h
@@ -27,6 +27,8 @@ extern int nouveau_gem_ioctl_new(struct drm_device *, void *,
 				 struct drm_file *);
 extern int nouveau_gem_ioctl_pushbuf(struct drm_device *, void *,
 				     struct drm_file *);
+extern int nouveau_gem_ioctl_pushbuf_2(struct drm_device *, void *,
+				       struct drm_file *);
 extern int nouveau_gem_ioctl_cpu_prep(struct drm_device *, void *,
 				      struct drm_file *);
 extern int nouveau_gem_ioctl_cpu_fini(struct drm_device *, void *,
diff --git a/drm/uapi/drm/nouveau_drm.h b/drm/uapi/drm/nouveau_drm.h
index 0d7608d..394cd94 100644
--- a/drm/uapi/drm/nouveau_drm.h
+++ b/drm/uapi/drm/nouveau_drm.h
@@ -115,6 +115,15 @@ struct drm_nouveau_gem_pushbuf {
 	uint64_t gart_available;
 };
 
+#define NOUVEAU_GEM_PUSHBUF_2_FENCE_WAIT                             0x00000001
+#define NOUVEAU_GEM_PUSHBUF_2_FENCE_EMIT                             0x00000002
+struct drm_nouveau_gem_pushbuf_2 {
+	struct drm_nouveau_gem_pushbuf base;
+	uint32_t flags;
+	int32_t  fence;
+	uint64_t reserved;
+};
+
 #define NOUVEAU_GEM_CPU_PREP_NOWAIT                                  0x00000001
 #define NOUVEAU_GEM_CPU_PREP_WRITE                                   0x00000004
 struct drm_nouveau_gem_cpu_prep {
@@ -139,9 +148,11 @@ struct drm_nouveau_gem_cpu_fini {
 #define DRM_NOUVEAU_GEM_CPU_PREP       0x42
 #define DRM_NOUVEAU_GEM_CPU_FINI       0x43
 #define DRM_NOUVEAU_GEM_INFO           0x44
+#define DRM_NOUVEAU_GEM_PUSHBUF_2      0x45
 
 #define DRM_IOCTL_NOUVEAU_GEM_NEW            DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_NEW, struct drm_nouveau_gem_new)
 #define DRM_IOCTL_NOUVEAU_GEM_PUSHBUF        DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_PUSHBUF, struct drm_nouveau_gem_pushbuf)
+#define DRM_IOCTL_NOUVEAU_GEM_PUSHBUF_2      DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_PUSHBUF_2, struct drm_nouveau_gem_pushbuf_2)
 #define DRM_IOCTL_NOUVEAU_GEM_CPU_PREP       DRM_IOW (DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_CPU_PREP, struct drm_nouveau_gem_cpu_prep)
 #define DRM_IOCTL_NOUVEAU_GEM_CPU_FINI       DRM_IOW (DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_CPU_FINI, struct drm_nouveau_gem_cpu_fini)
 #define DRM_IOCTL_NOUVEAU_GEM_INFO           DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_INFO, struct drm_nouveau_gem_info)
-- 
1.8.1.5

Add a new nouveau_pushbuf_kick_fence function that takes and emits
a sync fence fd.  The fence fd can be waited on, or merged with
other fence fd's, or passed back to kernel as a prerquisite for a
subsequent hw operation.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 include/drm/nouveau_drm.h | 10 +++++
 nouveau/nouveau.h         |  2 +
 nouveau/pushbuf.c         | 93 ++++++++++++++++++++++++++++++++---------------
 3 files changed, 76 insertions(+), 29 deletions(-)

diff --git a/include/drm/nouveau_drm.h b/include/drm/nouveau_drm.h
index b18cad0..3e40210 100644
--- a/include/drm/nouveau_drm.h
+++ b/include/drm/nouveau_drm.h
@@ -171,6 +171,15 @@ struct drm_nouveau_gem_pushbuf {
 	uint64_t gart_available;
 };
 
+#define NOUVEAU_GEM_PUSHBUF_2_FENCE_WAIT                             0x00000001
+#define NOUVEAU_GEM_PUSHBUF_2_FENCE_EMIT                             0x00000002
+struct drm_nouveau_gem_pushbuf_2 {
+	struct drm_nouveau_gem_pushbuf base;
+	uint32_t flags;
+	int32_t  fence;
+	uint64_t reserved;
+};
+
 #define NOUVEAU_GEM_CPU_PREP_NOWAIT                                  0x00000001
 #define NOUVEAU_GEM_CPU_PREP_NOBLOCK                                 0x00000002
 #define NOUVEAU_GEM_CPU_PREP_WRITE                                   0x00000004
@@ -204,5 +213,6 @@ struct drm_nouveau_sarea {
 #define DRM_NOUVEAU_GEM_CPU_PREP       0x42
 #define DRM_NOUVEAU_GEM_CPU_FINI       0x43
 #define DRM_NOUVEAU_GEM_INFO           0x44
+#define DRM_NOUVEAU_GEM_PUSHBUF_2      0x45
 
 #endif /* __NOUVEAU_DRM_H__ */
diff --git a/nouveau/nouveau.h b/nouveau/nouveau.h
index a55e2b0..281420f 100644
--- a/nouveau/nouveau.h
+++ b/nouveau/nouveau.h
@@ -225,6 +225,8 @@ void nouveau_pushbuf_reloc(struct nouveau_pushbuf *, struct
nouveau_bo *,
 int  nouveau_pushbuf_validate(struct nouveau_pushbuf *);
 uint32_t nouveau_pushbuf_refd(struct nouveau_pushbuf *, struct nouveau_bo *);
 int  nouveau_pushbuf_kick(struct nouveau_pushbuf *, struct nouveau_object
*channel);
+int  nouveau_pushbuf_kick_fence(struct nouveau_pushbuf *,
+				struct nouveau_object *channel, int *fence);
 struct nouveau_bufctx *
 nouveau_pushbuf_bufctx(struct nouveau_pushbuf *, struct nouveau_bufctx *);
 
diff --git a/nouveau/pushbuf.c b/nouveau/pushbuf.c
index 6e703a4..8667d05 100644
--- a/nouveau/pushbuf.c
+++ b/nouveau/pushbuf.c
@@ -33,6 +33,7 @@
 #include <string.h>
 #include <assert.h>
 #include <errno.h>
+#include <unistd.h>
 
 #include <xf86drm.h>
 #include <xf86atomic.h>
@@ -77,7 +78,7 @@ nouveau_pushbuf(struct nouveau_pushbuf *push)
 }
 
 static int pushbuf_validate(struct nouveau_pushbuf *, bool);
-static int pushbuf_flush(struct nouveau_pushbuf *);
+static int pushbuf_flush(struct nouveau_pushbuf *, int *);
 
 static bool
 pushbuf_kref_fits(struct nouveau_pushbuf *push, struct nouveau_bo *bo,
@@ -172,7 +173,7 @@ pushbuf_kref(struct nouveau_pushbuf *push, struct nouveau_bo
*bo,
 	 */
 	fpush = cli_push_get(push->client, bo);
 	if (fpush && fpush != push)
-		pushbuf_flush(fpush);
+		pushbuf_flush(fpush, NULL);
 
 	kref = cli_kref_get(push->client, bo);
 	if (kref) {
@@ -305,18 +306,21 @@ pushbuf_dump(struct nouveau_pushbuf_krec *krec, int
krec_id, int chid)
 }
 
 static int
-pushbuf_submit(struct nouveau_pushbuf *push, struct nouveau_object *chan)
+pushbuf_submit(struct nouveau_pushbuf *push, struct nouveau_object *chan,
+	       int *fence)
 {
 	struct nouveau_pushbuf_priv *nvpb = nouveau_pushbuf(push);
 	struct nouveau_pushbuf_krec *krec = nvpb->list;
 	struct nouveau_device *dev = push->client->device;
 	struct drm_nouveau_gem_pushbuf_bo_presumed *info;
 	struct drm_nouveau_gem_pushbuf_bo *kref;
-	struct drm_nouveau_gem_pushbuf req;
+	struct drm_nouveau_gem_pushbuf_2 req_2;
+	struct drm_nouveau_gem_pushbuf *req = &req_2.base;
 	struct nouveau_fifo *fifo = chan->data;
 	struct nouveau_bo *bo;
 	int krec_id = 0;
 	int ret = 0, i;
+	int fence_out = -1;
 
 	if (chan->oclass != NOUVEAU_FIFO_CHANNEL_CLASS)
 		return -EINVAL;
@@ -326,30 +330,42 @@ pushbuf_submit(struct nouveau_pushbuf *push, struct
nouveau_object *chan)
 
 	nouveau_pushbuf_data(push, NULL, 0, 0);
 
+	/* TODO: If fence is requested, force kickoff. */
 	while (krec && krec->nr_push) {
-		req.channel = fifo->channel;
-		req.nr_buffers = krec->nr_buffer;
-		req.buffers = (uint64_t)(unsigned long)krec->buffer;
-		req.nr_relocs = krec->nr_reloc;
-		req.nr_push = krec->nr_push;
-		req.relocs = (uint64_t)(unsigned long)krec->reloc;
-		req.push = (uint64_t)(unsigned long)krec->push;
-		req.suffix0 = nvpb->suffix0;
-		req.suffix1 = nvpb->suffix1;
-		req.vram_available = 0; /* for valgrind */
-		req.gart_available = 0;
+		req->channel = fifo->channel;
+		req->nr_buffers = krec->nr_buffer;
+		req->buffers = (uint64_t)(unsigned long)krec->buffer;
+		req->nr_relocs = krec->nr_reloc;
+		req->nr_push = krec->nr_push;
+		req->relocs = (uint64_t)(unsigned long)krec->reloc;
+		req->push = (uint64_t)(unsigned long)krec->push;
+		req->suffix0 = nvpb->suffix0;
+		req->suffix1 = nvpb->suffix1;
+		req->vram_available = 0; /* for valgrind */
+		req->gart_available = 0;
+		if (fence) {
+			req_2.flags = NOUVEAU_GEM_PUSHBUF_2_FENCE_EMIT;
+			if (*fence >= 0)
+				req_2.flags |= NOUVEAU_GEM_PUSHBUF_2_FENCE_WAIT;
+			req_2.fence = *fence;
+			req_2.reserved = 0;
+		}
 
 		if (dbg_on(0))
 			pushbuf_dump(krec, krec_id++, fifo->channel);
 
 #ifndef SIMULATE
-		ret = drmCommandWriteRead(dev->fd, DRM_NOUVEAU_GEM_PUSHBUF,
-					  &req, sizeof(req));
-		nvpb->suffix0 = req.suffix0;
-		nvpb->suffix1 = req.suffix1;
-		dev->vram_limit = (req.vram_available *
+		if (fence)
+			ret = drmCommandWriteRead(dev->fd, DRM_NOUVEAU_GEM_PUSHBUF_2,
+						  &req_2, sizeof(req_2));
+		else
+			ret = drmCommandWriteRead(dev->fd, DRM_NOUVEAU_GEM_PUSHBUF,
+						  req, sizeof(*req));
+		nvpb->suffix0 = req->suffix0;
+		nvpb->suffix1 = req->suffix1;
+		dev->vram_limit = (req->vram_available *
 				nouveau_device(dev)->vram_limit_percent) / 100;
-		dev->gart_limit = (req.gart_available *
+		dev->gart_limit = (req->gart_available *
 				nouveau_device(dev)->gart_limit_percent) / 100;
 #else
 		if (dbg_on(31))
@@ -362,6 +378,12 @@ pushbuf_submit(struct nouveau_pushbuf *push, struct
nouveau_object *chan)
 			break;
 		}
 
+		if (fence) {
+			if (fence_out >= 0)
+				close(fence_out);
+			fence_out = req_2.fence;
+		}
+
 		kref = krec->buffer;
 		for (i = 0; i < krec->nr_buffer; i++, kref++) {
 			bo = (void *)(unsigned long)kref->user_priv;
@@ -385,11 +407,17 @@ pushbuf_submit(struct nouveau_pushbuf *push, struct
nouveau_object *chan)
 		krec = krec->next;
 	}
 
+	if (!ret && fence) {
+		if (*fence >= 0)
+			close(*fence);
+		*fence = fence_out;
+	}
+
 	return ret;
 }
 
 static int
-pushbuf_flush(struct nouveau_pushbuf *push)
+pushbuf_flush(struct nouveau_pushbuf *push, int *fence)
 {
 	struct nouveau_pushbuf_priv *nvpb = nouveau_pushbuf(push);
 	struct nouveau_pushbuf_krec *krec = nvpb->krec;
@@ -399,7 +427,7 @@ pushbuf_flush(struct nouveau_pushbuf *push)
 	int ret = 0, i;
 
 	if (push->channel) {
-		ret = pushbuf_submit(push, push->channel);
+		ret = pushbuf_submit(push, push->channel, fence);
 	} else {
 		nouveau_pushbuf_data(push, NULL, 0, 0);
 		krec->next = malloc(sizeof(*krec));
@@ -469,7 +497,7 @@ pushbuf_refn(struct nouveau_pushbuf *push, bool retry,
 	if (ret) {
 		pushbuf_refn_fail(push, sref, krec->nr_reloc);
 		if (retry) {
-			pushbuf_flush(push);
+			pushbuf_flush(push, NULL);
 			nouveau_pushbuf_space(push, 0, 0, 0);
 			return pushbuf_refn(push, false, refs, nr);
 		}
@@ -521,7 +549,7 @@ pushbuf_validate(struct nouveau_pushbuf *push, bool retry)
 	if (ret) {
 		pushbuf_refn_fail(push, sref, srel);
 		if (retry) {
-			pushbuf_flush(push);
+			pushbuf_flush(push, NULL);
 			return pushbuf_validate(push, false);
 		}
 	}
@@ -673,7 +701,7 @@ nouveau_pushbuf_space(struct nouveau_pushbuf *push,
 	    krec->nr_reloc + relocs >= NOUVEAU_GEM_MAX_RELOCS ||
 	    krec->nr_push + pushes >= NOUVEAU_GEM_MAX_PUSH) {
 		if (nvpb->bo && krec->nr_buffer)
-			pushbuf_flush(push);
+			pushbuf_flush(push, NULL);
 		flushed = true;
 	}
 
@@ -767,10 +795,17 @@ nouveau_pushbuf_refd(struct nouveau_pushbuf *push, struct
nouveau_bo *bo)
 }
 
 drm_public int
-nouveau_pushbuf_kick(struct nouveau_pushbuf *push, struct nouveau_object *chan)
+nouveau_pushbuf_kick_fence(struct nouveau_pushbuf *push,
+			   struct nouveau_object *chan, int *fence)
 {
 	if (!push->channel)
-		return pushbuf_submit(push, chan);
-	pushbuf_flush(push);
+		return pushbuf_submit(push, chan, fence);
+	pushbuf_flush(push, fence);
 	return pushbuf_validate(push, false);
 }
+
+drm_public int
+nouveau_pushbuf_kick(struct nouveau_pushbuf *push, struct nouveau_object *chan)
+{
+	return nouveau_pushbuf_kick_fence(push, chan, NULL);
+}
-- 
1.8.1.5

Do not attach fences automatically to buffers that are marked for
explicit synchronization.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drm/nouveau_bo.c           |  8 ++++----
 drm/nouveau_bo.h           |  4 ++--
 drm/nouveau_drm.c          |  1 +
 drm/nouveau_gem.c          | 47 +++++++++++++++++++++++++++++++++++++++-------
 drm/nouveau_gem.h          |  6 ++++--
 drm/nouveau_ttm.c          |  8 ++++----
 drm/nv50_display.c         |  2 +-
 drm/uapi/drm/nouveau_drm.h |  5 ++++-
 8 files changed, 60 insertions(+), 21 deletions(-)

diff --git a/drm/nouveau_bo.c b/drm/nouveau_bo.c
index 534734a..68b7bdd 100644
--- a/drm/nouveau_bo.c
+++ b/drm/nouveau_bo.c
@@ -180,7 +180,7 @@ nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
 
 int
 nouveau_bo_new(struct drm_device *dev, int size, int align,
-	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
+	       uint32_t flags, uint32_t tile_mode, uint32_t bo_flags,
 	       struct sg_table *sg,
 	       struct nouveau_bo **pnvbo)
 {
@@ -211,7 +211,7 @@ nouveau_bo_new(struct drm_device *dev, int size, int align,
 	INIT_LIST_HEAD(&nvbo->entry);
 	INIT_LIST_HEAD(&nvbo->vma_list);
 	nvbo->tile_mode = tile_mode;
-	nvbo->tile_flags = tile_flags;
+	nvbo->bo_flags = bo_flags;
 	nvbo->bo.bdev = &drm->ttm.bdev;
 
 	if (!nv_device_is_cpu_coherent(nvkm_device(&drm->device)))
@@ -272,7 +272,7 @@ set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
 		 * speed up when alpha-blending and depth-test are enabled
 		 * at the same time.
 		 */
-		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
+		if (nvbo->bo_flags & NOUVEAU_GEM_TILE_ZETA) {
 			fpfn = vram_pages / 2;
 			lpfn = ~0;
 		} else {
@@ -1291,7 +1291,7 @@ nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct
ttm_mem_reg *new_mem,
 	if (drm->device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
 		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
 						nvbo->tile_mode,
-						nvbo->tile_flags);
+						nvbo->bo_flags);
 	}
 
 	return 0;
diff --git a/drm/nouveau_bo.h b/drm/nouveau_bo.h
index f97bc26..ff1edba 100644
--- a/drm/nouveau_bo.h
+++ b/drm/nouveau_bo.h
@@ -25,7 +25,7 @@ struct nouveau_bo {
 	unsigned page_shift;
 
 	u32 tile_mode;
-	u32 tile_flags;
+	u32 bo_flags;
 	struct nouveau_drm_tile *tile;
 
 	/* Only valid if allocated via nouveau_gem_new() and iff you hold a
@@ -68,7 +68,7 @@ extern struct ttm_bo_driver nouveau_bo_driver;
 
 void nouveau_bo_move_init(struct nouveau_drm *);
 int  nouveau_bo_new(struct drm_device *, int size, int align, u32 flags,
-		    u32 tile_mode, u32 tile_flags, struct sg_table *sg,
+		    u32 tile_mode, u32 bo_flags, struct sg_table *sg,
 		    struct nouveau_bo **);
 int  nouveau_bo_pin(struct nouveau_bo *, u32 flags);
 int  nouveau_bo_unpin(struct nouveau_bo *);
diff --git a/drm/nouveau_drm.c b/drm/nouveau_drm.c
index 74d5ac6..3d84f3a 100644
--- a/drm/nouveau_drm.c
+++ b/drm/nouveau_drm.c
@@ -816,6 +816,7 @@ nouveau_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_PREP, nouveau_gem_ioctl_cpu_prep,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_FINI, nouveau_gem_ioctl_cpu_fini,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_INFO, nouveau_gem_ioctl_info,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_SET_INFO, nouveau_gem_ioctl_set_info,
DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW),
 };
 
 long
diff --git a/drm/nouveau_gem.c b/drm/nouveau_gem.c
index ee5782c..bb19507 100644
--- a/drm/nouveau_gem.c
+++ b/drm/nouveau_gem.c
@@ -149,7 +149,7 @@ nouveau_gem_object_close(struct drm_gem_object *gem, struct
drm_file *file_priv)
 
 int
 nouveau_gem_new(struct drm_device *dev, int size, int align, uint32_t domain,
-		uint32_t tile_mode, uint32_t tile_flags,
+		uint32_t tile_mode, uint32_t bo_flags,
 		struct nouveau_bo **pnvbo)
 {
 	struct nouveau_drm *drm = nouveau_drm(dev);
@@ -165,7 +165,7 @@ nouveau_gem_new(struct drm_device *dev, int size, int align,
uint32_t domain,
 		flags |= TTM_PL_FLAG_SYSTEM;
 
 	ret = nouveau_bo_new(dev, size, align, flags, tile_mode,
-			     tile_flags, NULL, pnvbo);
+			     bo_flags, NULL, pnvbo);
 	if (ret)
 		return ret;
 	nvbo = *pnvbo;
@@ -216,7 +216,21 @@ nouveau_gem_info(struct drm_file *file_priv, struct
drm_gem_object *gem,
 	rep->size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
 	rep->map_handle = drm_vma_node_offset_addr(&nvbo->bo.vma_node);
 	rep->tile_mode = nvbo->tile_mode;
-	rep->tile_flags = nvbo->tile_flags;
+	rep->bo_flags = nvbo->bo_flags;
+	return 0;
+}
+
+static int
+nouveau_gem_set_info(struct drm_file *file_priv, struct drm_gem_object *gem,
+		     struct drm_nouveau_gem_info *req)
+{
+	struct nouveau_bo *nvbo = nouveau_gem_object(gem);
+
+	if (req->bo_flags & NOUVEAU_GEM_EXPLICIT_SYNC)
+		nvbo->bo_flags |= NOUVEAU_GEM_EXPLICIT_SYNC;
+	else
+		nvbo->bo_flags &= ~NOUVEAU_GEM_EXPLICIT_SYNC;
+
 	return 0;
 }
 
@@ -231,14 +245,15 @@ nouveau_gem_ioctl_new(struct drm_device *dev, void *data,
 	struct nouveau_bo *nvbo = NULL;
 	int ret = 0;
 
-	if (!pfb->memtype_valid(pfb, req->info.tile_flags)) {
-		NV_PRINTK(error, cli, "bad page flags: 0x%08x\n",
req->info.tile_flags);
+	if (!pfb->memtype_valid(pfb, req->info.bo_flags)) {
+		NV_PRINTK(error, cli, "bad page flags: 0x%08x\n",
+			  req->info.bo_flags);
 		return -EINVAL;
 	}
 
 	ret = nouveau_gem_new(dev, req->info.size, req->align,
 			      req->info.domain, req->info.tile_mode,
-			      req->info.tile_flags, &nvbo);
+			      req->info.bo_flags, &nvbo);
 	if (ret)
 		return ret;
 
@@ -303,12 +318,14 @@ validate_fini_no_ticket(struct validate_op *op, struct
nouveau_fence *fence,
 {
 	struct nouveau_bo *nvbo;
 	struct drm_nouveau_gem_pushbuf_bo *b;
+	bool explicit_sync;
 
 	while (!list_empty(&op->list)) {
 		nvbo = list_entry(op->list.next, struct nouveau_bo, entry);
 		b = &pbbo[nvbo->pbbo_index];
+		explicit_sync = !!(nvbo->bo_flags & NOUVEAU_GEM_EXPLICIT_SYNC);
 
-		if (likely(fence))
+		if (likely(fence) && !explicit_sync)
 			nouveau_bo_fence(nvbo, fence, !!b->write_domains);
 
 		if (unlikely(nvbo->validate_mapped)) {
@@ -947,3 +964,19 @@ nouveau_gem_ioctl_info(struct drm_device *dev, void *data,
 	return ret;
 }
 
+int
+nouveau_gem_ioctl_set_info(struct drm_device *dev, void *data,
+			   struct drm_file *file_priv)
+{
+	struct drm_nouveau_gem_info *req = data;
+	struct drm_gem_object *gem;
+	int ret;
+
+	gem = drm_gem_object_lookup(dev, file_priv, req->handle);
+	if (!gem)
+		return -ENOENT;
+
+	ret = nouveau_gem_set_info(file_priv, gem, req);
+	drm_gem_object_unreference_unlocked(gem);
+	return ret;
+}
diff --git a/drm/nouveau_gem.h b/drm/nouveau_gem.h
index 7454dea..abac606 100644
--- a/drm/nouveau_gem.h
+++ b/drm/nouveau_gem.h
@@ -7,7 +7,7 @@
 #include "nouveau_bo.h"
 
 #define nouveau_bo_tile_layout(nvbo)				\
-	((nvbo)->tile_flags & NOUVEAU_GEM_TILE_LAYOUT_MASK)
+	((nvbo)->bo_flags & NOUVEAU_GEM_TILE_LAYOUT_MASK)
 
 static inline struct nouveau_bo *
 nouveau_gem_object(struct drm_gem_object *gem)
@@ -18,7 +18,7 @@ nouveau_gem_object(struct drm_gem_object *gem)
 /* nouveau_gem.c */
 extern int nouveau_gem_new(struct drm_device *, int size, int align,
 			   uint32_t domain, uint32_t tile_mode,
-			   uint32_t tile_flags, struct nouveau_bo **);
+			   uint32_t bo_flags, struct nouveau_bo **);
 extern void nouveau_gem_object_del(struct drm_gem_object *);
 extern int nouveau_gem_object_open(struct drm_gem_object *, struct drm_file *);
 extern void nouveau_gem_object_close(struct drm_gem_object *,
@@ -35,6 +35,8 @@ extern int nouveau_gem_ioctl_cpu_fini(struct drm_device *,
void *,
 				      struct drm_file *);
 extern int nouveau_gem_ioctl_info(struct drm_device *, void *,
 				  struct drm_file *);
+extern int nouveau_gem_ioctl_set_info(struct drm_device *, void *,
+				      struct drm_file *);
 
 extern int nouveau_gem_prime_pin(struct drm_gem_object *);
 struct reservation_object *nouveau_gem_prime_res_obj(struct drm_gem_object *);
diff --git a/drm/nouveau_ttm.c b/drm/nouveau_ttm.c
index 8058013..a224820 100644
--- a/drm/nouveau_ttm.c
+++ b/drm/nouveau_ttm.c
@@ -81,12 +81,12 @@ nouveau_vram_manager_new(struct ttm_mem_type_manager *man,
 	u32 size_nc = 0;
 	int ret;
 
-	if (nvbo->tile_flags & NOUVEAU_GEM_TILE_NONCONTIG)
+	if (nvbo->bo_flags & NOUVEAU_GEM_TILE_NONCONTIG)
 		size_nc = 1 << nvbo->page_shift;
 
 	ret = pfb->ram->get(pfb, mem->num_pages << PAGE_SHIFT,
 			   mem->page_alignment << PAGE_SHIFT, size_nc,
-			   (nvbo->tile_flags >> 8) & 0x3ff, &node);
+			   (nvbo->bo_flags >> 8) & 0x3ff, &node);
 	if (ret) {
 		mem->mm_node = NULL;
 		return (ret == -ENOSPC) ? 0 : ret;
@@ -174,11 +174,11 @@ nouveau_gart_manager_new(struct ttm_mem_type_manager *man,
 	switch (drm->device.info.family) {
 	case NV_DEVICE_INFO_V0_TESLA:
 		if (drm->device.info.chipset != 0x50)
-			node->memtype = (nvbo->tile_flags & 0x7f00) >> 8;
+			node->memtype = (nvbo->bo_flags & 0x7f00) >> 8;
 		break;
 	case NV_DEVICE_INFO_V0_FERMI:
 	case NV_DEVICE_INFO_V0_KEPLER:
-		node->memtype = (nvbo->tile_flags & 0xff00) >> 8;
+		node->memtype = (nvbo->bo_flags & 0xff00) >> 8;
 		break;
 	default:
 		break;
diff --git a/drm/nv50_display.c b/drm/nv50_display.c
index fdb3e1a..ce0df41 100644
--- a/drm/nv50_display.c
+++ b/drm/nv50_display.c
@@ -2352,7 +2352,7 @@ nv50_fb_ctor(struct drm_framebuffer *fb)
 	u8 kind = nouveau_bo_tile_layout(nvbo) >> 8;
 	u8 tile = nvbo->tile_mode;
 
-	if (nvbo->tile_flags & NOUVEAU_GEM_TILE_NONCONTIG) {
+	if (nvbo->bo_flags & NOUVEAU_GEM_TILE_NONCONTIG) {
 		NV_ERROR(drm, "framebuffer requires contiguous bo\n");
 		return -EINVAL;
 	}
diff --git a/drm/uapi/drm/nouveau_drm.h b/drm/uapi/drm/nouveau_drm.h
index 394cd94..49fc749 100644
--- a/drm/uapi/drm/nouveau_drm.h
+++ b/drm/uapi/drm/nouveau_drm.h
@@ -46,6 +46,7 @@
 #define NOUVEAU_GEM_TILE_32BPP       0x00000002
 #define NOUVEAU_GEM_TILE_ZETA        0x00000004
 #define NOUVEAU_GEM_TILE_NONCONTIG   0x00000008
+#define NOUVEAU_GEM_EXPLICIT_SYNC    0x00040000
 
 struct drm_nouveau_gem_info {
 	uint32_t handle;
@@ -54,7 +55,7 @@ struct drm_nouveau_gem_info {
 	uint64_t offset;
 	uint64_t map_handle;
 	uint32_t tile_mode;
-	uint32_t tile_flags;
+	uint32_t bo_flags;
 };
 
 struct drm_nouveau_gem_new {
@@ -149,6 +150,7 @@ struct drm_nouveau_gem_cpu_fini {
 #define DRM_NOUVEAU_GEM_CPU_FINI       0x43
 #define DRM_NOUVEAU_GEM_INFO           0x44
 #define DRM_NOUVEAU_GEM_PUSHBUF_2      0x45
+#define DRM_NOUVEAU_GEM_SET_INFO       0x46
 
 #define DRM_IOCTL_NOUVEAU_GEM_NEW            DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_NEW, struct drm_nouveau_gem_new)
 #define DRM_IOCTL_NOUVEAU_GEM_PUSHBUF        DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_PUSHBUF, struct drm_nouveau_gem_pushbuf)
@@ -156,5 +158,6 @@ struct drm_nouveau_gem_cpu_fini {
 #define DRM_IOCTL_NOUVEAU_GEM_CPU_PREP       DRM_IOW (DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_CPU_PREP, struct drm_nouveau_gem_cpu_prep)
 #define DRM_IOCTL_NOUVEAU_GEM_CPU_FINI       DRM_IOW (DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_CPU_FINI, struct drm_nouveau_gem_cpu_fini)
 #define DRM_IOCTL_NOUVEAU_GEM_INFO           DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_INFO, struct drm_nouveau_gem_info)
+#define DRM_IOCTL_NOUVEAU_GEM_SET_INFO       DRM_IOWR(DRM_COMMAND_BASE +
DRM_NOUVEAU_GEM_SET_INFO, struct drm_nouveau_gem_info)
 
 #endif /* __NOUVEAU_DRM_H__ */
-- 
1.8.1.5

Allow user space to provide an explicit sync fence fd when exporting
a dma-buf from gem handle.  The fence will be stored as the explicit
fence to the reservation object.

Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
---
 drivers/gpu/drm/drm_prime.c | 41 +++++++++++++++++++++++++++++++++--------
 include/uapi/drm/drm.h      |  9 ++++++++-
 2 files changed, 41 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
index 2807a77..c69df2e 100644
--- a/drivers/gpu/drm/drm_prime.c
+++ b/drivers/gpu/drm/drm_prime.c
@@ -28,8 +28,10 @@
 
 #include <linux/export.h>
 #include <linux/dma-buf.h>
+#include <linux/reservation.h>
 #include <drm/drmP.h>
 #include "drm_internal.h"
+#include "../../staging/android/sync.h"
 
 /*
  * DMA-BUF/GEM Object references and lifetime overview:
@@ -329,7 +331,7 @@ static const struct dma_buf_ops drm_gem_prime_dmabuf_ops = 
{
  * drm_gem_prime_export - helper library implemention of the export callback
  * @dev: drm_device to export from
  * @obj: GEM object to export
- * @flags: flags like DRM_CLOEXEC
+ * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
  *
  * This is the implementation of the gem_prime_export functions for GEM drivers
  * using the PRIME helpers.
@@ -385,7 +387,7 @@ static struct dma_buf *export_and_register_object(struct
drm_device *dev,
  * @dev: dev to export the buffer from
  * @file_priv: drm file-private structure
  * @handle: buffer handle to export
- * @flags: flags like DRM_CLOEXEC
+ * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
  * @prime_fd: pointer to storage for the fd id of the create dma-buf
  *
  * This is the PRIME export function which must be used mandatorily by GEM
@@ -401,6 +403,24 @@ int drm_gem_prime_handle_to_fd(struct drm_device *dev,
 	struct drm_gem_object *obj;
 	int ret = 0;
 	struct dma_buf *dmabuf;
+	struct fence *fence = NULL;
+
+	if (flags & DRM_SYNC_FD) {
+#ifdef CONFIG_SYNC
+		struct sync_fence *sf = sync_fence_fdget(*prime_fd);
+		if (!sf)
+			return -ENOENT;
+		if (sf->num_fences != 1) {
+			sync_fence_put(sf);
+			return -EINVAL;
+		}
+		fence = fence_get(sf->cbs[0].sync_pt);
+		sync_fence_put(sf);
+		flags &= ~DRM_SYNC_FD;
+#else
+		return -ENODEV;
+#endif
+	}
 
 	mutex_lock(&file_priv->prime.lock);
 	obj = drm_gem_object_lookup(dev, file_priv, handle);
@@ -453,6 +473,14 @@ out_have_obj:
 		goto fail_put_dmabuf;
 
 out_have_handle:
+	if (fence) {
+		if (!dmabuf->resv) {
+			ret = -ENODEV;
+			goto fail_put_dmabuf;
+		}
+		reservation_object_add_excl_fence(dmabuf->resv, fence);
+	}
+
 	ret = dma_buf_fd(dmabuf, flags);
 	/*
 	 * We must _not_ remove the buffer from the handle cache since the newly
@@ -475,6 +503,7 @@ out:
 	drm_gem_object_unreference_unlocked(obj);
 out_unlock:
 	mutex_unlock(&file_priv->prime.lock);
+	fence_put(fence);
 
 	return ret;
 }
@@ -624,7 +653,6 @@ int drm_prime_handle_to_fd_ioctl(struct drm_device *dev,
void *data,
 				 struct drm_file *file_priv)
 {
 	struct drm_prime_handle *args = data;
-	uint32_t flags;
 
 	if (!drm_core_check_feature(dev, DRIVER_PRIME))
 		return -EINVAL;
@@ -633,14 +661,11 @@ int drm_prime_handle_to_fd_ioctl(struct drm_device *dev,
void *data,
 		return -ENOSYS;
 
 	/* check flags are valid */
-	if (args->flags & ~DRM_CLOEXEC)
+	if (args->flags & ~(DRM_CLOEXEC | DRM_SYNC_FD))
 		return -EINVAL;
 
-	/* we only want to pass DRM_CLOEXEC which is == O_CLOEXEC */
-	flags = args->flags & DRM_CLOEXEC;
-
 	return dev->driver->prime_handle_to_fd(dev, file_priv,
-			args->handle, flags, &args->fd);
+			args->handle, args->flags, &args->fd);
 }
 
 int drm_prime_fd_to_handle_ioctl(struct drm_device *dev, void *data,
diff --git a/include/uapi/drm/drm.h b/include/uapi/drm/drm.h
index b0b8556..a11b893 100644
--- a/include/uapi/drm/drm.h
+++ b/include/uapi/drm/drm.h
@@ -661,13 +661,20 @@ struct drm_set_client_cap {
 };
 
 #define DRM_CLOEXEC O_CLOEXEC
+#define DRM_SYNC_FD O_DSYNC
 struct drm_prime_handle {
 	__u32 handle;
 
 	/** Flags.. only applicable for handle->fd */
 	__u32 flags;
 
-	/** Returned dmabuf file descriptor */
+	/**
+	 * DRM_IOCTL_PRIME_FD_TO_HANDLE:
+	 *   in: dma-buf fd
+	 * DRM_IOCTL_PRIME_HANDLE_TO_FD:
+	 *   in: sync fence fd if DRM_SYNC_FD flag is passed
+	 *   out: dma-buf fd
+	 */
 	__s32 fd;
 };
 
-- 
1.8.1.5

On 26-09-14 12:00, Lauri Peltonen wrote:
> Allow user space to provide an explicit sync fence fd when exporting
> a dma-buf from gem handle.  The fence will be stored as the explicit
> fence to the reservation object.
> 
> Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
> ---
>  drivers/gpu/drm/drm_prime.c | 41 +++++++++++++++++++++++++++++++++--------
>  include/uapi/drm/drm.h      |  9 ++++++++-
>  2 files changed, 41 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
> index 2807a77..c69df2e 100644
> --- a/drivers/gpu/drm/drm_prime.c
> +++ b/drivers/gpu/drm/drm_prime.c
> @@ -28,8 +28,10 @@
>  
>  #include <linux/export.h>
>  #include <linux/dma-buf.h>
> +#include <linux/reservation.h>
>  #include <drm/drmP.h>
>  #include "drm_internal.h"
> +#include "../../staging/android/sync.h"
>  
>  /*
>   * DMA-BUF/GEM Object references and lifetime overview:
> @@ -329,7 +331,7 @@ static const struct dma_buf_ops
drm_gem_prime_dmabuf_ops =  {
>   * drm_gem_prime_export - helper library implemention of the export
callback
>   * @dev: drm_device to export from
>   * @obj: GEM object to export
> - * @flags: flags like DRM_CLOEXEC
> + * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
>   *
>   * This is the implementation of the gem_prime_export functions for GEM
drivers
>   * using the PRIME helpers.
> @@ -385,7 +387,7 @@ static struct dma_buf
*export_and_register_object(struct drm_device *dev,
>   * @dev: dev to export the buffer from
>   * @file_priv: drm file-private structure
>   * @handle: buffer handle to export
> - * @flags: flags like DRM_CLOEXEC
> + * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
>   * @prime_fd: pointer to storage for the fd id of the create dma-buf
>   *
>   * This is the PRIME export function which must be used mandatorily by GEM
> @@ -401,6 +403,24 @@ int drm_gem_prime_handle_to_fd(struct drm_device *dev,
>  	struct drm_gem_object *obj;
>  	int ret = 0;
>  	struct dma_buf *dmabuf;
> +	struct fence *fence = NULL;
> +
> +	if (flags & DRM_SYNC_FD) {
> +#ifdef CONFIG_SYNC
> +		struct sync_fence *sf = sync_fence_fdget(*prime_fd);
> +		if (!sf)
> +			return -ENOENT;
> +		if (sf->num_fences != 1) {
> +			sync_fence_put(sf);
> +			return -EINVAL;
> +		}
> +		fence = fence_get(sf->cbs[0].sync_pt);
> +		sync_fence_put(sf);
> +		flags &= ~DRM_SYNC_FD;
> +#else
> +		return -ENODEV;
> +#endif
> +	}
>  
>  	mutex_lock(&file_priv->prime.lock);
>  	obj = drm_gem_object_lookup(dev, file_priv, handle);
> @@ -453,6 +473,14 @@ out_have_obj:
>  		goto fail_put_dmabuf;
>  
>  out_have_handle:
> +	if (fence) {
> +		if (!dmabuf->resv) {
> +			ret = -ENODEV;
> +			goto fail_put_dmabuf;
> +		}
This is never true. A default resv gets allocated, see dma_buf's create
function.

~Maarten

~Maarten

Hey,

On 26-09-14 12:00, Lauri Peltonen wrote:
> Do not attach fences automatically to buffers that are marked for
> explicit synchronization.
> 
> Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
> ---
>  drm/nouveau_bo.c           |  8 ++++----
>  drm/nouveau_bo.h           |  4 ++--
>  drm/nouveau_drm.c          |  1 +
>  drm/nouveau_gem.c          | 47
+++++++++++++++++++++++++++++++++++++++-------
>  drm/nouveau_gem.h          |  6 ++++--
>  drm/nouveau_ttm.c          |  8 ++++----
>  drm/nv50_display.c         |  2 +-
>  drm/uapi/drm/nouveau_drm.h |  5 ++++-
>  8 files changed, 60 insertions(+), 21 deletions(-)
Considering this will break on eviction, is it really needed?
If a fence is waited on before starting an operation synchronization will be a
noop.

~Maarten

Hey,

On 26-09-14 12:00, Lauri Peltonen wrote:
> Modify sync_fence_create to accept an array of 'struct fence'
objects.
> This will allow drm drivers to create sync_fence objects and pass sync
> fd's between user space with minimal modifications, without ever
creating
> sync_timeline or sync_pt objects, and without implementing the
> sync_timeline_ops interface.
> 
> Modify the sync driver debug code to not assume that every 'struct
fence'
> (that is associated with a 'struct sync_fence') is embedded within
a
> 'struct sync_pt'.
> 
> Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
> ---
>  drivers/staging/android/sw_sync.c    |  3 ++-
>  drivers/staging/android/sync.c       | 34
++++++++++++++++++---------------
>  drivers/staging/android/sync.h       | 11 ++++++-----
>  drivers/staging/android/sync_debug.c | 37
+++++++++++++++++-------------------
>  4 files changed, 44 insertions(+), 41 deletions(-)
> 
> diff --git a/drivers/staging/android/sw_sync.c
b/drivers/staging/android/sw_sync.c
> index a76db3f..6949812 100644
> --- a/drivers/staging/android/sw_sync.c
> +++ b/drivers/staging/android/sw_sync.c
> @@ -184,7 +184,8 @@ static long sw_sync_ioctl_create_fence(struct
sw_sync_timeline *obj,
>  	}
>  
>  	data.name[sizeof(data.name) - 1] = '\0';
> -	fence = sync_fence_create(data.name, pt);
> +	fence = sync_fence_create(data.name,
> +				  (struct fence *[]){ &pt->base }, 1);
>  	if (fence == NULL) {
>  		sync_pt_free(pt);
>  		err = -ENOMEM;
> diff --git a/drivers/staging/android/sync.c
b/drivers/staging/android/sync.c
> index e7b2e02..1d0d968 100644
> --- a/drivers/staging/android/sync.c
> +++ b/drivers/staging/android/sync.c
> @@ -187,28 +187,32 @@ static void fence_check_cb_func(struct fence *f,
struct fence_cb *cb)
>  		wake_up_all(&fence->wq);
>  }
>  
> -/* TODO: implement a create which takes more that one sync_pt */
> -struct sync_fence *sync_fence_create(const char *name, struct sync_pt *pt)
> +struct sync_fence *sync_fence_create(const char *name,
> +				     struct fence **fences, int num_fences)
>  {
> -	struct sync_fence *fence;
> +	struct sync_fence *sync_fence;
> +	int size = offsetof(struct sync_fence, cbs[num_fences]);
> +	int i;
>  
> -	fence = sync_fence_alloc(offsetof(struct sync_fence, cbs[1]), name);
> -	if (fence == NULL)
> +	sync_fence = sync_fence_alloc(size, name);
> +	if (sync_fence == NULL)
>  		return NULL;
>  
> -	fence->num_fences = 1;
> -	atomic_set(&fence->status, 1);
> +	sync_fence->num_fences = num_fences;
> +	atomic_set(&sync_fence->status, 0);
>  
> -	fence_get(&pt->base);
> -	fence->cbs[0].sync_pt = &pt->base;
> -	fence->cbs[0].fence = fence;
> -	if (fence_add_callback(&pt->base, &fence->cbs[0].cb,
> -			       fence_check_cb_func))
> -		atomic_dec(&fence->status);
> +	for (i = 0; i < num_fences; i++) {
> +		struct fence *f = fences[i];
> +		struct sync_fence_cb *cb = &sync_fence->cbs[i];
>  
> -	sync_fence_debug_add(fence);
> +		cb->sync_pt = fence_get(f);
> +		cb->fence = sync_fence;
> +		if (!fence_add_callback(f, &cb->cb, fence_check_cb_func))
> +			atomic_inc(&sync_fence->status);
> +	}
> +	sync_fence_debug_add(sync_fence);
>  
> -	return fence;
> +	return sync_fence;
>  }
>  EXPORT_SYMBOL(sync_fence_create);
sync_fence_merge currently depends on the list of fences to be sorted to remove
duplicates efficiently.
Feeding it a unsorted list will probably result in hard to debug bugs. :-)
> diff --git a/drivers/staging/android/sync.h
b/drivers/staging/android/sync.h
> index 66b0f43..b8ad72c 100644
> --- a/drivers/staging/android/sync.h
> +++ b/drivers/staging/android/sync.h
> @@ -246,13 +246,14 @@ void sync_pt_free(struct sync_pt *pt);
>  
>  /**
>   * sync_fence_create() - creates a sync fence
> - * @name:	name of fence to create
> - * @pt:		sync_pt to add to the fence
> + * @name:	name of the sync fence to create
> + * @fences:	fences to add to the sync fence
> + * @num_fences:	the number of fences in the @fences array
>   *
> - * Creates a fence containg @pt.  Once this is called, the fence takes
> - * ownership of @pt.
> + * Creates a sync fence from an array of drm fences. Takes refs to
@fences.
>   */
> -struct sync_fence *sync_fence_create(const char *name, struct sync_pt
*pt);
> +struct sync_fence *sync_fence_create(const char *name,
> +				     struct fence **fences, int num_fences);
>  
>  /*
>   * API for sync_fence consumers
> diff --git a/drivers/staging/android/sync_debug.c
b/drivers/staging/android/sync_debug.c
> index 257fc91..2d8873e 100644
> --- a/drivers/staging/android/sync_debug.c
> +++ b/drivers/staging/android/sync_debug.c
> @@ -81,33 +81,33 @@ static const char *sync_status_str(int status)
>  	return "error";
>  }
>  
> -static void sync_print_pt(struct seq_file *s, struct sync_pt *pt, bool
fence)
> +static void sync_print_pt(struct seq_file *s, struct fence *pt, bool
fence)
>  {
>  	int status = 1;
> -	struct sync_timeline *parent = sync_pt_parent(pt);
>  
> -	if (fence_is_signaled_locked(&pt->base))
> -		status = pt->base.status;
> +	if (fence_is_signaled_locked(pt))
> +		status = pt->status;
>  
> -	seq_printf(s, "  %s%spt %s",
> -		   fence ? parent->name : "",
> -		   fence ? "_" : "",
> -		   sync_status_str(status));
> +	if (fence)
> +		seq_printf(s, "  %d_pt %s", pt->context,
> +			   sync_status_str(status));
> +	else
> +		seq_printf(s, "  pt %s", sync_status_str(status));
>  
>  	if (status <= 0) {
> -		struct timeval tv = ktime_to_timeval(pt->base.timestamp);
> +		struct timeval tv = ktime_to_timeval(pt->timestamp);
>  
>  		seq_printf(s, "@%ld.%06ld", tv.tv_sec, tv.tv_usec);
>  	}
>  
> -	if (parent->ops->timeline_value_str &&
> -	    parent->ops->pt_value_str) {
> +	if (pt->ops->timeline_value_str &&
> +	    pt->ops->fence_value_str) {
>  		char value[64];
>  
> -		parent->ops->pt_value_str(pt, value, sizeof(value));
> +		pt->ops->fence_value_str(pt, value, sizeof(value));
>  		seq_printf(s, ": %s", value);
>  		if (fence) {
> -			parent->ops->timeline_value_str(parent, value,
> +			pt->ops->timeline_value_str(pt, value,
>  						    sizeof(value));
>  			seq_printf(s, " / %s", value);
>  		}
> @@ -121,7 +121,8 @@ static void sync_print_obj(struct seq_file *s, struct
sync_timeline *obj)
>  	struct list_head *pos;
>  	unsigned long flags;
>  
> -	seq_printf(s, "%s %s", obj->name,
obj->ops->driver_name);
> +	seq_printf(s, "%d %s %s", obj->context, obj->name,
> +		   obj->ops->driver_name);
>  
>  	if (obj->ops->timeline_value_str) {
>  		char value[64];
> @@ -136,7 +137,7 @@ static void sync_print_obj(struct seq_file *s, struct
sync_timeline *obj)
>  	list_for_each(pos, &obj->child_list_head) {
>  		struct sync_pt *pt >  			container_of(pos, struct sync_pt,
child_list);
> -		sync_print_pt(s, pt, false);
> +		sync_print_pt(s, &pt->base, false);
>  	}
>  	spin_unlock_irqrestore(&obj->child_list_lock, flags);
>  }
> @@ -151,11 +152,7 @@ static void sync_print_fence(struct seq_file *s,
struct sync_fence *fence)
>  		   sync_status_str(atomic_read(&fence->status)));
>  
>  	for (i = 0; i < fence->num_fences; ++i) {
> -		struct sync_pt *pt > -			container_of(fence->cbs[i].sync_pt,
> -				     struct sync_pt, base);
> -
> -		sync_print_pt(s, pt, true);
> +		sync_print_pt(s, fence->cbs[i].sync_pt, true);
>  	}
>  
>  	spin_lock_irqsave(&fence->wq.lock, flags);
>

On Fri, Sep 26, 2014 at 01:00:05PM +0300, Lauri Peltonen
wrote:
> 
> Hi guys,
> 
> 
> I'd like to start a new thread about explicit fence synchronization. 
This time
> with a Nouveau twist. :-)
> 
> First, let me define what I understand by implicit/explicit sync:
> 
> Implicit synchronization
> * Fences are attached to buffers
> * Kernel manages fences automatically based on buffer read/write access
> 
> Explicit synchronization
> * Fences are passed around independently
> * Kernel takes and emits fences to/from user space when submitting work
> 
> Implicit synchronization is already implemented in open source drivers, and
> works well for most use cases.  I don't seek to change any of that.  My
> proposal aims at allowing some drm drivers to operate in explicit sync mode
to
> get maximal performance, while still remaining fully compatible with the
> implicit paradigm.
Yeah, pretty much what we have in mind on the i915 side too. I didn't look
too closely at your patches, so just a few high level comments on your rfc
here.
> I will try to explain why I think we should support the explicit model as
well.
> 
> 
> 1. Bindless graphics
> 
> Bindless graphics is a central concept when trying to reduce the OpenGL
driver
> overhead.  The idea is that the application can bind a large set of buffers
to
> the working set up front using extensions such as GL_ARB_bindless_texture,
and
> they remain resident until the application releases them (note that compute
> APIs have typically similar semantics).  These working sets can be huge,
> hundreds or even thousands of buffers, so we would like to opt out from the
> per-submit overhead of acquiring locks, waiting for fences, and storing
fences.
> Automatically synchronizing these working sets in kernel will also prevent
> parallelism between channels that are sharing the working set (in fact
sharing
> just one buffer from the working set will cause the jobs of the two
channels to
> be serialized).
> 
> 2. Evolution of graphics APIs
> 
> The graphics API evolution seems to be going to a direction where game
engine
> and middleware vendors demand more control over work submission and
> synchronization.  We expect that this trend will continue, and more and
more
> synchronization decisions will be pushed to the API level.  OpenGL and EGL
> already provide good explicit command stream level synchronization
primitives:
> glFenceSync and EGL_KHR_wait_sync.  Their use is also encouraged - for
example
> EGL_KHR_image_base spec clearly states that the application is responsible
for
> synchronizing accesses to EGLImages.  If the API that is exposed to
developers
> gives the control over synchronization to the developer, then implicit
waits
> that are inserted by the kernel are unnecessary and unexpected, and can
> severely hurt performance.  It also makes it easy for the developer to
write
> code that happens to work on Linux because of implicit sync, but will fail
on
> other platforms.
> 
> 3. Suballocation
> 
> Using user space suballocation can help reduce the overhead when a large
number
> of small textures are used.  Synchronizing suballocated surfaces implicitly
in
> kernel doesn't make sense - many channels should be able to access the
same
> kernel-level buffer object simultaneously.
> 
> 4. Buffer sharing complications
> 
> This is not really an argument for explicit sync as such, but I'd like
to point
> out that sharing buffers across SoC engines is often much more complex than
> just exporting and importing a dma-buf and waiting for the dma-buf fences.
> Sometimes we need to do color format or tiling layout conversion. 
Sometimes,
> at least on Tegra, we need to decompress buffers when we pass them from the
GPU
> to an engine that doesn't support framebuffer compression.  These
things are
> not uncommon, particularly when we have SoC's that combine licensed IP
blocks
> from different vendors.  My point is that user space is already heavily
> involved when sharing buffers between drivers, and giving it some more
control
> over synchronization is not adding that much complexity.
> 
> 
> Because of the above arguments, I think it makes sense to let some user
space
> drm drivers opt out from implicit synchronization, while allowing them to
still
> remain fully compatible with the rest of the drm world that uses implicit
> synchronization.  In practice, this would require three things:
> 
> (1) Support passing fences (that are not tied to buffer objects) between
kernel
>     and user space.
> 
> (2) Stop automatically storing fences to the buffers that user space wants
to
>     synchronize explicitly.
The problem with this approach is that you then need hw faulting to make
sure the memory is there. Implicit fences aren't just used for syncing,
but also to make sure that the gpu still has access to the buffer as long
as it needs it. So you need at least a non-exclusive fence attached for
each command submission.

Of course on Android you don't have swap (would kill the puny mmc within
seconds) and you don't care for letting userspace pin most of memory for
gfx. So you'll get away with no fences at all. But for upstream I don't
see a good solution unfortunately. Ideas very much welcome.
> (3) Allow user space to attach an explicit fence to dma-buf when exporting
to
>     another driver that uses implicit sync.
> 
> There are still some open issues beyond these.  For example, can we skip
> acquiring the ww mutex for explicitly synchronized buffers?  I think we
could
> eventually, at least on unified memory systems where we don't need to
migrate
> between heaps (our downstream Tegra GPU driver does not lock any buffers at
> submit, it just grabs refcounts for hw).  Another quirk is that now Nouveau
> waits on the buffer fences when closing the gem object to ensure that it
> doesn't unmap too early.  We need to rework that for explicit sync, but
that
> shouldn't be difficult.
See above, but you can't avoid to attach fences as long as we still use a
buffer-object based gfx memory management model. At least afaics. Which
means you need the ordering guarantees imposed by ww mutexes to ensure
that the oddball implicit ordered client can't deadlock the kernel's
memory management code.
> I have written a prototype that demonstrates (1) by adding explicit sync fd
> support to Nouveau.  It's not a lot of code, because I only use a
relatively
> small subset of the android sync driver functionality.  Thanks to
Maarten's
> rewrite, all I need to do is to allow creating a sync_fence from a drm
fence in
> order to pass it to user space.  I don't need to use sync_pt or
sync_timeline,
> or fill in sync_timeline_ops.
> 
> I can see why the upstream has been reluctant to de-stage the android sync
> driver in its current form, since (even though it now builds on struct
fence)
> it still duplicates some of the drm fence concepts.  I'd like to think
that my
> patches only use the parts of the android sync driver that genuinely are
> missing from the drm fence model: allowing user space to operate on fence
> objects that are independent of buffer objects.
Imo de-staging the android syncpt stuff needs to happen first, before
drivers can use it. Since non-staging stuff really shouldn't depend upon
code from staging.
> The last two patches are mocks that show how (2) and (3) might work out.  I
> haven't done any testing with them yet.  Before going any further,
I'd like to
> get your feedback.  Can you see the benefits of explicit sync as an
alternative
> synchronization model?  Do you think we could use the android sync_fence
for
> passing fences between user space?  Or did you have something else in mind
for
> explicit sync in the drm world?
I'm all for adding explicit syncing. Our plans are roughly.
- Add both an in and and out fence to execbuf to sync with other rendering
  and give userspace a fence back. Needs to different flags probably.

- Maybe add an ioctl to dma-bufs to get at the current implicit fences
  attached to them (both an exclusive and non-exclusive version). This
  should help with making explicit and implicit sync work together nicely.

- Add fence support to kms. Probably only worth it together with the new
  atomic stuff. Again we need an in fence to wait for (one for each
  buffer) and an out fence. The later can easily be implemented by
  extending struct drm_event, which means not a single driver code line
  needs to be changed for this.

- For de-staging android syncpts we need to de-clutter the internal
  interfaces and also review all the ioctls exposed. Like you say it
  should be just the userspace interface for struct drm_fence. Also, it
  needs testcases and preferrably manpages.

Unfortunately it looks like Intel won't do this all for you due to a bunch
of hilarious internal reasons :( At least not anytime soon.

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

On Fri, Sep 26, 2014 at 01:00:12PM +0300, Lauri Peltonen
wrote:
> Allow user space to provide an explicit sync fence fd when exporting
> a dma-buf from gem handle.  The fence will be stored as the explicit
> fence to the reservation object.
> 
> Signed-off-by: Lauri Peltonen <lpeltonen at nvidia.com>
All existing userspace treats dma_bufs as long-lived objects. Well, all
the userspace that expects implicit syncing, afaik Android shovels lots of
dma-bufs around all the time (since ion is using them as it's native
buffer handles).

So adding an exclusive fence once at export time isn't going to be
terribly useful.

So I think a better approach would be to add this as ioctls to the dma-buf
fd itself. Then you can also add a "give me the fence(s) for this
dma_buf"
ioctl, which is useful for interop the other way round (i.e. implicit ->
explicit).
-Daniel
> ---
>  drivers/gpu/drm/drm_prime.c | 41 +++++++++++++++++++++++++++++++++--------
>  include/uapi/drm/drm.h      |  9 ++++++++-
>  2 files changed, 41 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
> index 2807a77..c69df2e 100644
> --- a/drivers/gpu/drm/drm_prime.c
> +++ b/drivers/gpu/drm/drm_prime.c
> @@ -28,8 +28,10 @@
>  
>  #include <linux/export.h>
>  #include <linux/dma-buf.h>
> +#include <linux/reservation.h>
>  #include <drm/drmP.h>
>  #include "drm_internal.h"
> +#include "../../staging/android/sync.h"
>  
>  /*
>   * DMA-BUF/GEM Object references and lifetime overview:
> @@ -329,7 +331,7 @@ static const struct dma_buf_ops
drm_gem_prime_dmabuf_ops =  {
>   * drm_gem_prime_export - helper library implemention of the export
callback
>   * @dev: drm_device to export from
>   * @obj: GEM object to export
> - * @flags: flags like DRM_CLOEXEC
> + * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
>   *
>   * This is the implementation of the gem_prime_export functions for GEM
drivers
>   * using the PRIME helpers.
> @@ -385,7 +387,7 @@ static struct dma_buf
*export_and_register_object(struct drm_device *dev,
>   * @dev: dev to export the buffer from
>   * @file_priv: drm file-private structure
>   * @handle: buffer handle to export
> - * @flags: flags like DRM_CLOEXEC
> + * @flags: flags like DRM_CLOEXEC or DRM_SYNC_FD
>   * @prime_fd: pointer to storage for the fd id of the create dma-buf
>   *
>   * This is the PRIME export function which must be used mandatorily by GEM
> @@ -401,6 +403,24 @@ int drm_gem_prime_handle_to_fd(struct drm_device *dev,
>  	struct drm_gem_object *obj;
>  	int ret = 0;
>  	struct dma_buf *dmabuf;
> +	struct fence *fence = NULL;
> +
> +	if (flags & DRM_SYNC_FD) {
> +#ifdef CONFIG_SYNC
> +		struct sync_fence *sf = sync_fence_fdget(*prime_fd);
> +		if (!sf)
> +			return -ENOENT;
> +		if (sf->num_fences != 1) {
> +			sync_fence_put(sf);
> +			return -EINVAL;
> +		}
> +		fence = fence_get(sf->cbs[0].sync_pt);
> +		sync_fence_put(sf);
> +		flags &= ~DRM_SYNC_FD;
> +#else
> +		return -ENODEV;
> +#endif
> +	}
>  
>  	mutex_lock(&file_priv->prime.lock);
>  	obj = drm_gem_object_lookup(dev, file_priv, handle);
> @@ -453,6 +473,14 @@ out_have_obj:
>  		goto fail_put_dmabuf;
>  
>  out_have_handle:
> +	if (fence) {
> +		if (!dmabuf->resv) {
> +			ret = -ENODEV;
> +			goto fail_put_dmabuf;
> +		}
> +		reservation_object_add_excl_fence(dmabuf->resv, fence);
> +	}
> +
>  	ret = dma_buf_fd(dmabuf, flags);
>  	/*
>  	 * We must _not_ remove the buffer from the handle cache since the newly
> @@ -475,6 +503,7 @@ out:
>  	drm_gem_object_unreference_unlocked(obj);
>  out_unlock:
>  	mutex_unlock(&file_priv->prime.lock);
> +	fence_put(fence);
>  
>  	return ret;
>  }
> @@ -624,7 +653,6 @@ int drm_prime_handle_to_fd_ioctl(struct drm_device
*dev, void *data,
>  				 struct drm_file *file_priv)
>  {
>  	struct drm_prime_handle *args = data;
> -	uint32_t flags;
>  
>  	if (!drm_core_check_feature(dev, DRIVER_PRIME))
>  		return -EINVAL;
> @@ -633,14 +661,11 @@ int drm_prime_handle_to_fd_ioctl(struct drm_device
*dev, void *data,
>  		return -ENOSYS;
>  
>  	/* check flags are valid */
> -	if (args->flags & ~DRM_CLOEXEC)
> +	if (args->flags & ~(DRM_CLOEXEC | DRM_SYNC_FD))
>  		return -EINVAL;
>  
> -	/* we only want to pass DRM_CLOEXEC which is == O_CLOEXEC */
> -	flags = args->flags & DRM_CLOEXEC;
> -
>  	return dev->driver->prime_handle_to_fd(dev, file_priv,
> -			args->handle, flags, &args->fd);
> +			args->handle, args->flags, &args->fd);
>  }
>  
>  int drm_prime_fd_to_handle_ioctl(struct drm_device *dev, void *data,
> diff --git a/include/uapi/drm/drm.h b/include/uapi/drm/drm.h
> index b0b8556..a11b893 100644
> --- a/include/uapi/drm/drm.h
> +++ b/include/uapi/drm/drm.h
> @@ -661,13 +661,20 @@ struct drm_set_client_cap {
>  };
>  
>  #define DRM_CLOEXEC O_CLOEXEC
> +#define DRM_SYNC_FD O_DSYNC
>  struct drm_prime_handle {
>  	__u32 handle;
>  
>  	/** Flags.. only applicable for handle->fd */
>  	__u32 flags;
>  
> -	/** Returned dmabuf file descriptor */
> +	/**
> +	 * DRM_IOCTL_PRIME_FD_TO_HANDLE:
> +	 *   in: dma-buf fd
> +	 * DRM_IOCTL_PRIME_HANDLE_TO_FD:
> +	 *   in: sync fence fd if DRM_SYNC_FD flag is passed
> +	 *   out: dma-buf fd
> +	 */
>  	__s32 fd;
>  };
>  
> -- 
> 1.8.1.5
> 
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

On Mon, Sep 29, 2014 at 09:43:02AM +0200, Daniel Vetter
wrote:
> On Fri, Sep 26, 2014 at 01:00:05PM +0300, Lauri Peltonen wrote:
> > 
> > Hi guys,
> > 
> > 
> > I'd like to start a new thread about explicit fence
synchronization.  This time
> > with a Nouveau twist. :-)
> > 
> > First, let me define what I understand by implicit/explicit sync:
> > 
> > Implicit synchronization
> > * Fences are attached to buffers
> > * Kernel manages fences automatically based on buffer read/write
access
> > 
> > Explicit synchronization
> > * Fences are passed around independently
> > * Kernel takes and emits fences to/from user space when submitting
work
> > 
> > Implicit synchronization is already implemented in open source
drivers, and
> > works well for most use cases.  I don't seek to change any of
that.  My
> > proposal aims at allowing some drm drivers to operate in explicit sync
mode to
> > get maximal performance, while still remaining fully compatible with
the
> > implicit paradigm.
> 
> Yeah, pretty much what we have in mind on the i915 side too. I didn't
look
> too closely at your patches, so just a few high level comments on your rfc
> here.
> 
> > I will try to explain why I think we should support the explicit model
as well.
> > 
> > 
> > 1. Bindless graphics
> > 
> > Bindless graphics is a central concept when trying to reduce the
OpenGL driver
> > overhead.  The idea is that the application can bind a large set of
buffers to
> > the working set up front using extensions such as
GL_ARB_bindless_texture, and
> > they remain resident until the application releases them (note that
compute
> > APIs have typically similar semantics).  These working sets can be
huge,
> > hundreds or even thousands of buffers, so we would like to opt out
from the
> > per-submit overhead of acquiring locks, waiting for fences, and
storing fences.
> > Automatically synchronizing these working sets in kernel will also
prevent
> > parallelism between channels that are sharing the working set (in fact
sharing
> > just one buffer from the working set will cause the jobs of the two
channels to
> > be serialized).
> > 
> > 2. Evolution of graphics APIs
> > 
> > The graphics API evolution seems to be going to a direction where game
engine
> > and middleware vendors demand more control over work submission and
> > synchronization.  We expect that this trend will continue, and more
and more
> > synchronization decisions will be pushed to the API level.  OpenGL and
EGL
> > already provide good explicit command stream level synchronization
primitives:
> > glFenceSync and EGL_KHR_wait_sync.  Their use is also encouraged - for
example
> > EGL_KHR_image_base spec clearly states that the application is
responsible for
> > synchronizing accesses to EGLImages.  If the API that is exposed to
developers
> > gives the control over synchronization to the developer, then implicit
waits
> > that are inserted by the kernel are unnecessary and unexpected, and
can
> > severely hurt performance.  It also makes it easy for the developer to
write
> > code that happens to work on Linux because of implicit sync, but will
fail on
> > other platforms.
> > 
> > 3. Suballocation
> > 
> > Using user space suballocation can help reduce the overhead when a
large number
> > of small textures are used.  Synchronizing suballocated surfaces
implicitly in
> > kernel doesn't make sense - many channels should be able to access
the same
> > kernel-level buffer object simultaneously.
> > 
> > 4. Buffer sharing complications
> > 
> > This is not really an argument for explicit sync as such, but I'd
like to point
> > out that sharing buffers across SoC engines is often much more complex
than
> > just exporting and importing a dma-buf and waiting for the dma-buf
fences.
> > Sometimes we need to do color format or tiling layout conversion. 
Sometimes,
> > at least on Tegra, we need to decompress buffers when we pass them
from the GPU
> > to an engine that doesn't support framebuffer compression.  These
things are
> > not uncommon, particularly when we have SoC's that combine
licensed IP blocks
> > from different vendors.  My point is that user space is already
heavily
> > involved when sharing buffers between drivers, and giving it some more
control
> > over synchronization is not adding that much complexity.
> > 
> > 
> > Because of the above arguments, I think it makes sense to let some
user space
> > drm drivers opt out from implicit synchronization, while allowing them
to still
> > remain fully compatible with the rest of the drm world that uses
implicit
> > synchronization.  In practice, this would require three things:
> > 
> > (1) Support passing fences (that are not tied to buffer objects)
between kernel
> >     and user space.
> > 
> > (2) Stop automatically storing fences to the buffers that user space
wants to
> >     synchronize explicitly.
> 
> The problem with this approach is that you then need hw faulting to make
> sure the memory is there. Implicit fences aren't just used for syncing,
> but also to make sure that the gpu still has access to the buffer as long
> as it needs it. So you need at least a non-exclusive fence attached for
> each command submission.
> 
> Of course on Android you don't have swap (would kill the puny mmc
within
> seconds) and you don't care for letting userspace pin most of memory
for
> gfx. So you'll get away with no fences at all. But for upstream I
don't
> see a good solution unfortunately. Ideas very much welcome.
Well i am gonna repeat myself. But yes you can do explicit without associating
fence (at least no struct alloc) just associate a unique per command stream
number and have it be global ie irrespective of different execution pipeline
your hw have.

For non scheduling GPU, today generation roughly, you keep buffer on lru and
you know you can not evict buffer to swap for those that have an active id
(hw did not yet write the sequence number back). You update the lru with each
command stream ioctl.

For binding to GPU GART you can do that as a preamble to the command stream
which most hardware (AMD, Intel, NVidia) should be able to do.

For VRAM you have several choice that depends on how you want to manage VRAM.
For instance you might want to use it more like a cache and have each command
stream preamble with a bunch of copy to VRAM and posibly bunch of post copy
back to RAM. Or you can hold to current scheme but buffer move now becomes
preamble to command stream (ie buffer move are scheduled as a preamble to
command stream).


So i do not see what you would consider rocket science about this ?

Cheers,
J?r?me
> 
> > (3) Allow user space to attach an explicit fence to dma-buf when
exporting to
> >     another driver that uses implicit sync.
> > 
> > There are still some open issues beyond these.  For example, can we
skip
> > acquiring the ww mutex for explicitly synchronized buffers?  I think
we could
> > eventually, at least on unified memory systems where we don't need
to migrate
> > between heaps (our downstream Tegra GPU driver does not lock any
buffers at
> > submit, it just grabs refcounts for hw).  Another quirk is that now
Nouveau
> > waits on the buffer fences when closing the gem object to ensure that
it
> > doesn't unmap too early.  We need to rework that for explicit
sync, but that
> > shouldn't be difficult.
> 
> See above, but you can't avoid to attach fences as long as we still use
a
> buffer-object based gfx memory management model. At least afaics. Which
> means you need the ordering guarantees imposed by ww mutexes to ensure
> that the oddball implicit ordered client can't deadlock the
kernel's
> memory management code.
> 
> > I have written a prototype that demonstrates (1) by adding explicit
sync fd
> > support to Nouveau.  It's not a lot of code, because I only use a
relatively
> > small subset of the android sync driver functionality.  Thanks to
Maarten's
> > rewrite, all I need to do is to allow creating a sync_fence from a drm
fence in
> > order to pass it to user space.  I don't need to use sync_pt or
sync_timeline,
> > or fill in sync_timeline_ops.
> > 
> > I can see why the upstream has been reluctant to de-stage the android
sync
> > driver in its current form, since (even though it now builds on struct
fence)
> > it still duplicates some of the drm fence concepts.  I'd like to
think that my
> > patches only use the parts of the android sync driver that genuinely
are
> > missing from the drm fence model: allowing user space to operate on
fence
> > objects that are independent of buffer objects.
> 
> Imo de-staging the android syncpt stuff needs to happen first, before
> drivers can use it. Since non-staging stuff really shouldn't depend
upon
> code from staging.
> 
> > The last two patches are mocks that show how (2) and (3) might work
out.  I
> > haven't done any testing with them yet.  Before going any further,
I'd like to
> > get your feedback.  Can you see the benefits of explicit sync as an
alternative
> > synchronization model?  Do you think we could use the android
sync_fence for
> > passing fences between user space?  Or did you have something else in
mind for
> > explicit sync in the drm world?
> 
> I'm all for adding explicit syncing. Our plans are roughly.
> - Add both an in and and out fence to execbuf to sync with other rendering
>   and give userspace a fence back. Needs to different flags probably.
> 
> - Maybe add an ioctl to dma-bufs to get at the current implicit fences
>   attached to them (both an exclusive and non-exclusive version). This
>   should help with making explicit and implicit sync work together nicely.
> 
> - Add fence support to kms. Probably only worth it together with the new
>   atomic stuff. Again we need an in fence to wait for (one for each
>   buffer) and an out fence. The later can easily be implemented by
>   extending struct drm_event, which means not a single driver code line
>   needs to be changed for this.
> 
> - For de-staging android syncpts we need to de-clutter the internal
>   interfaces and also review all the ioctls exposed. Like you say it
>   should be just the userspace interface for struct drm_fence. Also, it
>   needs testcases and preferrably manpages.
> 
> Unfortunately it looks like Intel won't do this all for you due to a
bunch
> of hilarious internal reasons :( At least not anytime soon.
> 
> Cheers, Daniel
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

Thanks Daniel for your input!

On Mon, Sep 29, 2014 at 09:43:02AM +0200, Daniel Vetter
wrote:
> On Fri, Sep 26, 2014 at 01:00:05PM +0300, Lauri Peltonen wrote:
> > (2) Stop automatically storing fences to the buffers that user space
wants to
> >     synchronize explicitly.
> 
> The problem with this approach is that you then need hw faulting to make
> sure the memory is there. Implicit fences aren't just used for syncing,
> but also to make sure that the gpu still has access to the buffer as long
> as it needs it. So you need at least a non-exclusive fence attached for
> each command submission.
> 
> Of course on Android you don't have swap (would kill the puny mmc
within
> seconds) and you don't care for letting userspace pin most of memory
for
> gfx. So you'll get away with no fences at all. But for upstream I
don't
> see a good solution unfortunately. Ideas very much welcome.
> 
> > (3) Allow user space to attach an explicit fence to dma-buf when
exporting to
> >     another driver that uses implicit sync.
> > 
> > There are still some open issues beyond these.  For example, can we
skip
> > acquiring the ww mutex for explicitly synchronized buffers?  I think
we could
> > eventually, at least on unified memory systems where we don't need
to migrate
> > between heaps (our downstream Tegra GPU driver does not lock any
buffers at
> > submit, it just grabs refcounts for hw).  Another quirk is that now
Nouveau
> > waits on the buffer fences when closing the gem object to ensure that
it
> > doesn't unmap too early.  We need to rework that for explicit
sync, but that
> > shouldn't be difficult.
> 
> See above, but you can't avoid to attach fences as long as we still use
a
> buffer-object based gfx memory management model. At least afaics. Which
> means you need the ordering guarantees imposed by ww mutexes to ensure
> that the oddball implicit ordered client can't deadlock the
kernel's
> memory management code.
Implicit fences attached to individual buffers are one way for residency
management.  Do you think a working set based model could work in the DRM
framework?  For example, something like this:

- Allow user space to create "working set objects" and associate
buffers with
  them.  If the user space doesn't want to manage working sets explicitly,
it
  could also use an implicit default working set that contains all buffers that
  are mapped to the channel vm (on Android we could always use the default
  working set since we don't need to manage residency).  The working sets
are
  initially marked as dirty.
- User space tells which working sets are referenced by each work submission.
  Kernel locks these working sets, pins all buffers in dirty working sets, and
  resets the dirty bits.  After kicking off work, kernel stores the fence to
  the _working sets_, and then releases the locks (if an implicit default
  working set is used, then this would be roughly equivalent to storing a fence
  to channel vm that tells "this is the last hw operation that might have
  touched buffers in this address space").
- If swapping doesn't happen, then we just need to check the working set
dirty
  bits at each submit.
- When a buffer is swapped out, all working sets that refer to it need to be
  marked as dirty.
- When a buffer is swapped out or unmapped, we need to wait for the fences from
  all working sets that refer to the buffer.

Initially one might think of working sets as a mere optimization - we now need
to process a few working sets at every submit instead of many individual
buffers.  However, it makes a huge difference because of fences: fences that
are attached to buffers are used for implicitly synchronizing work across
different channels and engines.  They are in the performance critical path, and
we want to carefully manage them (that's the idea of explicit
synchronization).
The working set fences, on the other hand, would only be used to guarantee that
we don't swap out or unmap something that the GPU might be accessing.  We
never
need to wait for those fences (except when swapping or unmapping), so we can be
conservative without hurting performance.

> Imo de-staging the android syncpt stuff needs to happen first, before
drivers
> can use it. Since non-staging stuff really shouldn't depend upon code
from
> staging.
Fully agree.  I thought the best way towards that would be to show some driver
code that _would_ use it. :)

> I'm all for adding explicit syncing. Our plans are roughly.  - Add both
an in
> and and out fence to execbuf to sync with other rendering and give
userspace
> a fence back. Needs to different flags probably.
> 
> - Maybe add an ioctl to dma-bufs to get at the current implicit fences
>   attached to them (both an exclusive and non-exclusive version). This
>   should help with making explicit and implicit sync work together nicely.
> 
> - Add fence support to kms. Probably only worth it together with the new
>   atomic stuff. Again we need an in fence to wait for (one for each
>   buffer) and an out fence. The later can easily be implemented by
>   extending struct drm_event, which means not a single driver code line
>   needs to be changed for this.
> 
> - For de-staging android syncpts we need to de-clutter the internal
>   interfaces and also review all the ioctls exposed. Like you say it
>   should be just the userspace interface for struct drm_fence. Also, it
>   needs testcases and preferrably manpages.
This all sounds very similar to what we'd like to do!  Maybe we can move
forward with these parts, and continue to attach fences at submit until we have
a satisfactory solution for the pinning problem?

I'd like to understand what are the concrete steps to de-stage struct
sync_fence, since that's the first thing that needs to be done.  For
example,
what do you mean by "de-cluttering the internal interfaces"?  Just
that we'd
move the sync_fence parts from drivers/staging/android/sync.c to, say,
drivers/dma-buf/sync-fence.c ?  Would we still leave a copy of the existing
full driver to staging/android?


Thanks,
Lauri