-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 21, 2017 3:39 AM
To: Li, Samuel <samuel...@amd.com>; amd-gfx@lists.freedesktop.org
Subject: Re: [PATCH v2 1/1] drm/amdgpu: Add gem_prime_mmap support
Am 20.09.2017 um 22:31 schrieb Li, Samuel:
The waiting done here is only for the shared fence to switch from
explicitly to implicitly synchronization for correct interaction with the Intel
driver.
Actually here it is going to wait for all fences,
94 ret = reservation_object_wait_timeout_rcu(bo->tbo.resv, true...
That is actually a bug, we only need to wait here when prime_shared_count
== 0.
ttm_bo_validate() is an asynchronous operation to enable GPU access
to the BO, it isn't related at all to possible CPU access and can
actually prevent it a number of cases.
amdgpu_ttm_bind() in turn binds the BO to the GART space and isn't
related
to CPU access either.
ttm_bo_validate() can move buffer is necessary,
No, that is a very common misconception.
ttm_bo_validate() schedules the buffer to be moved at some point in the
future.
The only thing which saves us here is the fact that TTM waits for the move to
be completed in it's page fault handler.
and amdgpu_ttm_bind() will flush hdp, which we have discussed before.
And as concluded before that is unnecessary.
What you most likely need to do here is to use
reservation_object_wait_timeout_rcu() to wait for all GPU operations to
end.
Right, that is called by amdgpu_gem_prime_pin().
Please just completely drop the begin_cpu_access() and end_cpu_access()
callbacks.
Checking other drivers the only one actually implementing something special
is the i915 driver which needs to remove the BO from the GTT domain for
cache coherency I think.
The omap driver implements those callbacks to grab an extra reference to
the pages while they are access, we don't need to do this since TTM grabs an
extra reference during the mmap anyway.
All other drivers implementing mmap (exynos, rockchip, tilcdc, arcpgu, msm,
meson, etnaviv, sti, qxl, vc4, rcar-du, atmel-hlcdc, imx, mediatek, mali, vgem,
fsl-dcu, zte, hisilicon, sun4i, mxsfb) don't have a begin/end callback.
Pinning the BO can actually cause problem with the display code when the
BO needs to be scanned out, so we should avoid that as much as possible.
Regards,
Christian.
Sam
-----Original Message-----
From: Christian König [mailto:ckoenig.leichtzumer...@gmail.com]
Sent: Wednesday, September 20, 2017 2:15 PM
To: Li, Samuel <samuel...@amd.com>; Koenig, Christian
<christian.koe...@amd.com>; amd-gfx@lists.freedesktop.org
Subject: Re: [PATCH v2 1/1] drm/amdgpu: Add gem_prime_mmap
support
Am 20.09.2017 um 19:47 schrieb Li, Samuel:
No that isn't correct. Pinning just notes that the BO shouldn't be
moved any more. It doesn't wait for anything.
It does. The following is from amdgpu_gem_prime_pin(),
91 * Wait for all shared fences to complete before we switch to
future
92 * use of exclusive fence on this prime shared bo.
93 */
94 ret = reservation_object_wait_timeout_rcu(bo->tbo.resv, true,
false,
95 MAX_SCHEDULE_TIMEOUT);
96 if (unlikely(ret < 0)) {
97 DRM_DEBUG_PRIME("Fence wait failed: %li\n", ret);
98 amdgpu_bo_unreserve(bo);
99 return ret;
100 }
Besides, pinning process prepares all the stuff before and after
moving buffer(ttm_bo_validate, amdgpu_ttm_bind),
No, you misunderstood what this is good for.
The waiting done here is only for the shared fence to switch from
explicitly to implicitly synchronization for correct interaction with the Intel
driver.
As soon the the BO is exported that shouldn't wait for anything any more.
I think if a buffer can be moved, it is probably also in a good
condition to be
accessed.
That is complete nonsense.
ttm_bo_validate() is an asynchronous operation to enable GPU access
to the BO, it isn't related at all to possible CPU access and can
actually prevent it a number of cases.
amdgpu_ttm_bind() in turn binds the BO to the GART space and isn't
related to CPU access either.
What you most likely need to do here is to use
reservation_object_wait_timeout_rcu() to wait for all GPU operations to
end.
Regards,
Christian.
Sam
-----Original Message-----
From: Koenig, Christian
Sent: Wednesday, September 20, 2017 1:38 PM
To: Li, Samuel <samuel...@amd.com>; amd-gfx@lists.freedesktop.org
Subject: Re: [PATCH v2 1/1] drm/amdgpu: Add gem_prime_mmap
support
Am 20.09.2017 um 19:34 schrieb Li, Samuel:
If that is what this callback should do then this implementation
would be incorrect. Pinning doesn't wait for any GPU operation to
finish.
During pining, it will all the fences to finish. That shall be OK.
No that isn't correct. Pinning just notes that the BO shouldn't be
moved any more.
It doesn't wait for anything.
Christian.
Sam
-----Original Message-----
From: Koenig, Christian
Sent: Wednesday, September 20, 2017 12:21 PM
To: Li, Samuel <samuel...@amd.com>; amd-
g...@lists.freedesktop.org
Subject: Re: [PATCH v2 1/1] drm/amdgpu: Add gem_prime_mmap
support
Am 20.09.2017 um 17:44 schrieb Li, Samuel:
What happens when another thread is using
amdgpu_dmabuf_ops
to
call
begin_cpu_access/end_cpu_access when you are fixing it up
again?
Right, that is an issue.
A simple "if (!amdgpu_dmabuf_ops.begin_cpu_access)" should be
able
to
deal with that.
I would just completely drop the two callbacks, pinning is not
necessary for CPU access and thinking more about it it actually
has some unwanted side effects.
CPU access needs synchronization anyway, so the two callbacks
cannot be
dropped (other drivers implemented too), so I would like to keep
it there for now.
Wait a second what do you mean with "CPU access needs
synchronization"?
At least i915 makes the memory GPU inaccessible when you start to
use it with the CPU.
If that is what this callback should do then this implementation
would be incorrect. Pinning doesn't wait for any GPU operation to
finish.
Regards,
Christian.
Sam
-----Original Message-----
From: Koenig, Christian
Sent: Wednesday, September 20, 2017 2:58 AM
To: Li, Samuel <samuel...@amd.com>; amd-
g...@lists.freedesktop.org
Subject: Re: [PATCH v2 1/1] drm/amdgpu: Add
gem_prime_mmap
support
What do you mean "This isn't race free"?
Take a look at the code again:
+ dma_buf = drm_gem_prime_export(dev, gobj, flags);
+ amdgpu_dmabuf_ops = *(dma_buf->ops);
+ amdgpu_dmabuf_ops.begin_cpu_access =
amdgpu_gem_begin_cpu_access;
+ amdgpu_dmabuf_ops.end_cpu_access =
amdgpu_gem_end_cpu_access;
+ dma_buf->ops = &amdgpu_dmabuf_ops;
What happens when another thread is using
amdgpu_dmabuf_ops
to
call
begin_cpu_access/end_cpu_access when you are fixing it up
again?
I would just completely drop the two callbacks, pinning is not
necessary for CPU access and thinking more about it it actually
has some unwanted side effects.
Regards,
Christian.
Am 19.09.2017 um 23:22 schrieb Samuel Li:
+ vma->vm_pgoff = amdgpu_bo_mmap_offset(bo) >>
PAGE_SHIFT;
Maybe better use "vma->vm_pgoff +=
amdgpu_bo_mmap_offset(bo) >>
PAGE_SHIFT;", but I'm not sure.
How other drivers handle this?
This is a good catch. Looks like pgoff is honored during prime
mmap, not a
fake offset here.
+ dma_buf->ops = &amdgpu_dmabuf_ops;
This isn't race free and needs to be fixed.
Better add callbacks to drm_prime.c similar to
drm_gem_dmabuf_mmap().
What do you mean "This isn't race free"?
Regards,
Sam
On 2017-09-15 11:05 AM, Christian König wrote:
Am 14.09.2017 um 00:39 schrieb Samuel Li:
v2: drop hdp invalidate/flush.
Signed-off-by: Samuel Li <samuel...@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 ++
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c | 77
++++++++++++++++++++++++++++++-
3 files changed, 81 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index d2aaad7..188b705 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -395,11 +395,14 @@
amdgpu_gem_prime_import_sg_table(struct
drm_device *dev,
struct dma_buf *amdgpu_gem_prime_export(struct
drm_device
*dev,
struct drm_gem_object *gobj,
int flags);
+struct drm_gem_object
*amdgpu_gem_prime_import(struct
drm_device *dev,
+ struct dma_buf *dma_buf);
int amdgpu_gem_prime_pin(struct drm_gem_object
*obj);
void amdgpu_gem_prime_unpin(struct drm_gem_object
*obj);
struct reservation_object
*amdgpu_gem_prime_res_obj(struct
drm_gem_object *);
void *amdgpu_gem_prime_vmap(struct drm_gem_object
*obj);
void amdgpu_gem_prime_vunmap(struct
drm_gem_object
*obj,
void
*vaddr);
+int amdgpu_gem_prime_mmap(struct drm_gem_object
*obj,
struct
+vm_area_struct *vma);
int amdgpu_gem_debugfs_init(struct amdgpu_device
*adev);
/* sub-allocation manager, it has to be protected
by another
lock.
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 2cdf844..9b63ac5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -835,7 +835,7 @@ static struct drm_driver kms_driver = {
.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
.gem_prime_export = amdgpu_gem_prime_export,
- .gem_prime_import = drm_gem_prime_import,
+ .gem_prime_import = amdgpu_gem_prime_import,
.gem_prime_pin = amdgpu_gem_prime_pin,
.gem_prime_unpin = amdgpu_gem_prime_unpin,
.gem_prime_res_obj = amdgpu_gem_prime_res_obj,
@@
-843,6
+843,7 @@ static struct drm_driver kms_driver = {
.gem_prime_import_sg_table =
amdgpu_gem_prime_import_sg_table,
.gem_prime_vmap = amdgpu_gem_prime_vmap,
.gem_prime_vunmap = amdgpu_gem_prime_vunmap,
+ .gem_prime_mmap = amdgpu_gem_prime_mmap,
.name = DRIVER_NAME,
.desc = DRIVER_DESC, diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
index 5b3f928..13c977a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
@@ -57,6 +57,40 @@ void
amdgpu_gem_prime_vunmap(struct
drm_gem_object *obj, void *vaddr)
ttm_bo_kunmap(&bo->dma_buf_vmap);
}
+int amdgpu_gem_prime_mmap(struct drm_gem_object
*obj,
struct
vm_area_struct *vma)
+{
+ struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+ struct amdgpu_device *adev = amdgpu_ttm_adev(bo-
tbo.bdev);
+ unsigned asize = amdgpu_bo_size(bo);
+ int ret;
+
+ if (!vma->vm_file)
+ return -ENODEV;
+
+ if (adev == NULL)
+ return -ENODEV;
+
+ /* Check for valid size. */
+ if (asize < vma->vm_end - vma->vm_start)
+ return -EINVAL;
+
+ if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
+ (bo->flags &
AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
+ return -EPERM;
+ }
+ vma->vm_pgoff = amdgpu_bo_mmap_offset(bo) >>
PAGE_SHIFT;
Maybe better use "vma->vm_pgoff +=
amdgpu_bo_mmap_offset(bo) >>
PAGE_SHIFT;", but I'm not sure.
How other drivers handle this?
+
+ /* prime mmap does not need to check access, so allow
+ here
*/
+ ret = drm_vma_node_allow(&obj->vma_node, vma-
vm_file-
private_data);
+ if (ret)
+ return ret;
+
+ ret = ttm_bo_mmap(vma->vm_file, vma, &adev-
mman.bdev);
+ drm_vma_node_revoke(&obj->vma_node,
+ vma->vm_file->private_data);
+
+ return ret;
+}
+
struct drm_gem_object *
amdgpu_gem_prime_import_sg_table(struct drm_device
*dev,
struct dma_buf_attachment *attach,
@@
-130,14
+164,55 @@ struct reservation_object
*amdgpu_gem_prime_res_obj(struct drm_gem_object *obj)
return bo->tbo.resv;
}
+static int amdgpu_gem_begin_cpu_access(struct
dma_buf *dma_buf, enum dma_data_direction direction)
+{
+ return amdgpu_gem_prime_pin(dma_buf->priv);
+}
+
+static int amdgpu_gem_end_cpu_access(struct dma_buf
*dma_buf,
enum
+dma_data_direction direction) {
+ amdgpu_gem_prime_unpin(dma_buf->priv);
+
+ return 0;
+}
+
+static struct dma_buf_ops amdgpu_dmabuf_ops;
+
struct dma_buf *amdgpu_gem_prime_export(struct
drm_device
*dev,
struct drm_gem_object *gobj,
int flags)
{
struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
+ struct dma_buf *dma_buf;
if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
return ERR_PTR(-EPERM);
- return drm_gem_prime_export(dev, gobj, flags);
+ dma_buf = drm_gem_prime_export(dev, gobj, flags);
+ amdgpu_dmabuf_ops = *(dma_buf->ops);
+ amdgpu_dmabuf_ops.begin_cpu_access =
amdgpu_gem_begin_cpu_access;
+ amdgpu_dmabuf_ops.end_cpu_access =
amdgpu_gem_end_cpu_access;
+ dma_buf->ops = &amdgpu_dmabuf_ops;
This isn't race free and needs to be fixed.
Better add callbacks to drm_prime.c similar to
drm_gem_dmabuf_mmap().
Alternative you could just completely drop
amdgpu_gem_begin_cpu_access() and
amdgpu_gem_end_cpu_access()
as
well.
When the buffer is shared between device it is pinned anyway
and when
it isn't shared ttm_bo_mmap() is able to handle VRAM mappings
as
well.
Regards,
Christian.
+
+ return dma_buf;
+}
+
+struct drm_gem_object
*amdgpu_gem_prime_import(struct
drm_device *dev,
+ struct dma_buf *dma_buf) {
+ struct drm_gem_object *obj;
+
+ if (dma_buf->ops == &amdgpu_dmabuf_ops) {
+ obj = dma_buf->priv;
+ if (obj->dev == dev) {
+ /*
+ * Importing dmabuf exported from out own gem
increases
+ * refcount on gem itself instead of f_count of dmabuf.
+ */
+ drm_gem_object_get(obj);
+ return obj;
+ }
+ }
+
+ return drm_gem_prime_import(dev, dma_buf);
}
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
