Skip to contentSkip to conversation

amdgpu: implement context priority for amdgpu_cs_ctx_create2

PATCH New Andres Rodriguez 2 participants 4 messages Jan 4, 2017 3 messages not in the archive DRM drivers

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): 20170104015628.6472-1-andresx7-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#2 lore raw

PATCH 1/1 amdgpu: implement context priority for amdgpu_cs_ctx_create2 v2

Add a new context creation function that allows specifying the context priority.

A high priority context has the potential of starving lower priority contexts. The current kernel driver implementation only allows only the root user to allocate high priority contexts.

v2: corresponding changes for kernel patch v2

---
 amdgpu/amdgpu.h          | 17 +++++++++++++++--
 amdgpu/amdgpu_cs.c       | 17 +++++++++++++----
 configure.ac             |  2 +-
 include/drm/amdgpu_drm.h |  6 +++++-
 4 files changed, 34 insertions(+), 8 deletions(-)
amdgpu/amdgpu.h +15 -2
index 7b26a04..eef7a62 100644
@@ -794,8 +794,9 @@ int amdgpu_bo_list_update(amdgpu_bo_list_handle handle,
  * context will always be executed in order (first come, first serve).
  *
  *
- * \param   dev	    - \c [in] Device handle. See #amdgpu_device_initialize()
- * \param   context - \c [out] GPU Context handle
+ * \param   dev      - \c [in] Device handle. See #amdgpu_device_initialize()
+ * \param   priority - \c [in] Context creation flags. See AMDGPU_CTX_PRIORITY_*
+ * \param   context  - \c [out] GPU Context handle
  *
  * \return   0 on success\n
  *          <0 - Negative POSIX Error code
@@ -803,6 +804,18 @@ int amdgpu_bo_list_update(amdgpu_bo_list_handle handle,
  * \sa amdgpu_cs_ctx_free()
  *
 */
+int amdgpu_cs_ctx_create2(amdgpu_device_handle dev,
+			 uint32_t priority,
+			 amdgpu_context_handle *context);
+/**
+ * Create GPU execution Context
+ *
+ * Refer to amdgpu_cs_ctx_create2 for full documentation. This call
+ * is missing the priority parameter.
+ *
+ * \sa amdgpu_cs_ctx_create2()
+ *
+*/
 int amdgpu_cs_ctx_create(amdgpu_device_handle dev,
 			 amdgpu_context_handle *context);
 
amdgpu/amdgpu_cs.c +13 -4
index fb5b3a8..2f7669f 100644
@@ -46,13 +46,14 @@ static int amdgpu_cs_reset_sem(amdgpu_semaphore_handle sem);
 /**
  * Create command submission context
  *
- * \param   dev - \c [in] amdgpu device handle
- * \param   context - \c [out] amdgpu context handle
+ * \param   dev      - \c [in] Device handle. See #amdgpu_device_initialize()
+ * \param   priority - \c [in] Context creation flags. See AMDGPU_CTX_PRIORITY_*
+ * \param   context  - \c [out] GPU Context handle
  *
  * \return  0 on success otherwise POSIX Error code
 */
-int amdgpu_cs_ctx_create(amdgpu_device_handle dev,
-			 amdgpu_context_handle *context)
+int amdgpu_cs_ctx_create2(amdgpu_device_handle dev, uint32_t priority,
+							amdgpu_context_handle *context)
 {
 	struct amdgpu_context *gpu_context;
 	union drm_amdgpu_ctx args;
@@ -77,6 +78,8 @@ int amdgpu_cs_ctx_create(amdgpu_device_handle dev,
 	/* Create the context */
 	memset(&args, 0, sizeof(args));
 	args.in.op = AMDGPU_CTX_OP_ALLOC_CTX;
+	args.in.priority = priority;
+
 	r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_CTX, &args, sizeof(args));
 	if (r)
 		goto error;
@@ -96,6 +99,12 @@ error:
 	return r;
 }
 
+int amdgpu_cs_ctx_create(amdgpu_device_handle dev,
+			 amdgpu_context_handle *context)
+{
+	return amdgpu_cs_ctx_create2(dev, 0, context);
+}
+
 /**
  * Release command submission context
  *
configure.ac +1 -1
index e0597c3..682171b 100644
@@ -20,7 +20,7 @@
 
 AC_PREREQ([2.63])
 AC_INIT([libdrm],
-        [2.4.74],
+        [2.4.75],
         [https://bugs.freedesktop.org/enter_bug.cgi?product=DRI],
         [libdrm])
 
include/drm/amdgpu_drm.h +5 -1
index d8f2497..01cf3c9 100644
@@ -154,13 +154,17 @@ union drm_amdgpu_bo_list {
 /* unknown cause */
 #define AMDGPU_CTX_UNKNOWN_RESET	3
 
+/* Context priority level */
+#define AMDGPU_CTX_PRIORITY_HIGH	0
+#define AMDGPU_CTX_PRIORITY_NORMAL	1
+
 struct drm_amdgpu_ctx_in {
 	/** AMDGPU_CTX_OP_* */
 	uint32_t	op;
 	/** For future use, no flags defined so far */
 	uint32_t	flags;
 	uint32_t	ctx_id;
-	uint32_t	_pad;
+	uint32_t	priority;
 };
 
 union drm_amdgpu_ctx_out {
Signature · 6 lines
2.9.3

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 20170104015628.6472-2-andresx7-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#4 lore raw

Re: [PATCH] amdgpu: implement context priority for amdgpu_cs_ctx_create2 v2

Hi Andres,

Pardon for jumping in uninvited, there's a few comments below that you might find useful.

configure.ac line 23

Quoted 8 lines · Andres Rodriguez
configure.ac +1 -1
@@ -20,7 +20,7 @@
 
 AC_PREREQ([2.63])
 AC_INIT([libdrm],
-        [2.4.74],
+        [2.4.75],

This part should be separate and done only as one rolls a release. Currently we have more than a dozen of people who can do that, with 3+ from the AMD team.

include/drm/amdgpu_drm.h

Quoted 7 lines
configure.ac +0 -0
         [https://bugs.freedesktop.org/enter_bug.cgi?product=DRI],
         [libdrm])
 
include/drm/amdgpu_drm.h +0 -0
index d8f2497..01cf3c9 100644

Please sync this file as described in the README, in the same directory.

include/drm/amdgpu_drm.h union drm_amdgpu_bo_list { · line 159

Quoted 4 lines
include/drm/amdgpu_drm.h +3 -0
+/* Context priority level */
+#define AMDGPU_CTX_PRIORITY_HIGH       0
+#define AMDGPU_CTX_PRIORITY_NORMAL     1

You really want NORMAL to be 0 here, to preserve the behaviour. In other words, [sane] old userspace with new kernel will set HIGH prio, where NORMAL is used with older kernels.

include/drm/amdgpu_drm.h union drm_amdgpu_bo_list { · line 167

Quoted 9 lines
include/drm/amdgpu_drm.h +2 -1
+
 struct drm_amdgpu_ctx_in {
        /** AMDGPU_CTX_OP_* */
        uint32_t        op;
        /** For future use, no flags defined so far */
        uint32_t        flags;
        uint32_t        ctx_id;
-       uint32_t        _pad;
+       uint32_t        priority;

This part is a bit of meh:

Namely things can go crazy in either one of these scenarios:
- old kernel [validates _pad to be zero/other], new userspace feeds
"invalid" value.
- new kernel. old [nasty] userspace which leaves the value uninitialised.

Re-using _pad is ok, just make sure things don't explode all over the place.

Regards,
Emil
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CACvgo53X0TOGEYwO0vKx1m97+DH9P-mUCsEsPw+5t9o6Xqs+5w-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#6 lore raw

Re: [PATCH] amdgpu: implement context priority for amdgpu_cs_ctx_create2 v2

Thanks for the feedback Emil, it is very welcome :)

Quoted 11 lines · Emil Velikov

Hi Andres,

Pardon for jumping in uninvited, there's a few comments below that you might find useful.

Quoted 4 lines · Andres Rodriguez
configure.ac +0 -0
@@ -20,7 +20,7 @@
 
Quoted 7 lines

configure.ac line 23

Quoted 4 lines
configure.ac +1 -1
  AC_PREREQ([2.63])
  AC_INIT([libdrm],
-        [2.4.74],
+        [2.4.75],

This part should be separate and done only as one rolls a release. Currently we have more than a dozen of people who can do that, with 3+ from the AMD team.

I'll take this out of the patch.

Quoted 9 lines

include/drm/amdgpu_drm.h

Quoted 8 lines
configure.ac +0 -0
          [https://bugs.freedesktop.org/enter_bug.cgi?product=DRI],
          [libdrm])
 
include/drm/amdgpu_drm.h +0 -0
index d8f2497..01cf3c9 100644

Please sync this file as described in the README, in the same directory.

Thanks for pointing that out, I totally missed the README

Quoted 7 lines

include/drm/amdgpu_drm.h union drm_amdgpu_bo_list { · line 159

Quoted 4 lines
include/drm/amdgpu_drm.h +3 -0
+/* Context priority level */
+#define AMDGPU_CTX_PRIORITY_HIGH       0
+#define AMDGPU_CTX_PRIORITY_NORMAL     1

You really want NORMAL to be 0 here, to preserve the behaviour. In other words, [sane] old userspace with new kernel will set HIGH prio, where NORMAL is used with older kernels.

Will fix.

Quoted 17 lines

include/drm/amdgpu_drm.h union drm_amdgpu_bo_list { · line 167

Quoted 10 lines
include/drm/amdgpu_drm.h +2 -1
+
  struct drm_amdgpu_ctx_in {
         /** AMDGPU_CTX_OP_* */
         uint32_t        op;
         /** For future use, no flags defined so far */
         uint32_t        flags;
         uint32_t        ctx_id;
-       uint32_t        _pad;
+       uint32_t        priority;

This part is a bit of meh:

Namely things can go crazy in either one of these scenarios:
- old kernel [validates _pad to be zero/other], new userspace feeds
"invalid" value.
- new kernel. old [nasty] userspace which leaves the value uninitialised.

Quoted 4 lines

Re-using _pad is ok, just make sure things don't explode all over the place.

Regards,
Emil

Regards,
Andres

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx