Skip to contentSkip to conversation

intel: Add fence fd support to execbuf

PATCH New Chad Versace 1 participant 4 messages Jan 13, 2017 DRM drivers

lore
Share & tools
lore
#0 lore raw

cover intel: Add fence fd support to execbuf

Chris wrote these patches, and the kernel patches too. I wrote the Mesa patches that use the new feature. I'm submitting these patches myself to get things moving.

This series depends on fence fd support in I915_GEM_EXECBUFFER2, which isn't upstream in the kernel yet. I tested this with kmscube on Skylake, and everything looked good to me.

I pushed tags for this series as well as all the code I tested with:

    mesa: http://git.kiwitree.net/cgit/~chadv/mesa/tag/?h=chadv/review/i965-exec-fence-v03
    libdrm: http://git.kiwitree.net/cgit/~chadv/libdrm/tag/?h=chadv/review/intel-exec-fence-v01
    linux: http://git.kiwitree.net/cgit/~chadv/linux/tag/?h=chadv/test/i915-exec-fence-v03
    kmscube: http://git.kiwitree.net/cgit/~chadv/kmscube/tag/?h=chadv/test/fences-v02

I also sent Mesa patches to mesa-dev. Someone else should submit the kernel patches, as I tested them but don't grok them.

Chad Versace (1):
WAIT: headers: Update drm_i915.h

Chris Wilson (2):
intel: Allow the client to control implicit synchronisation
intel: Support passing of explicit fencing from execbuf

 include/drm/i915_drm.h   | 277 +++++++++++++++++++++++++++++++++++++++++++++--
 intel/intel_bufmgr.h     |   8 ++
 intel/intel_bufmgr_gem.c |  64 +++++++++--
 3 files changed, 330 insertions(+), 19 deletions(-)
Signature · 6 lines
2.11.0.21.ga274e0a

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH 1/3 WAIT: headers: Update drm_i915.h

Generated with make headers_install
from tag 'chadv/test/i915-exec-fence-v03'
of <git://git.kiwitree.net/~chadv/linux>.

TODO: Wait until header updates are in upstream kernel.
References: http://git.kiwitree.net/cgit/~chadv/linux/tag/?h=chadv/test/i915-exec-fence-v03

---
 include/drm/i915_drm.h | 277 +++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 267 insertions(+), 10 deletions(-)
include/drm/i915_drm.h +267 -10
index eb611a7a..48e023fa 100644
@@ -29,6 +29,10 @@
 
 #include "drm.h"
 
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
 /* Please note that modifications to all structs defined here are
  * subject to backwards-compatibility constraints.
  */
@@ -58,6 +62,30 @@
 #define I915_ERROR_UEVENT		"ERROR"
 #define I915_RESET_UEVENT		"RESET"
 
+/*
+ * MOCS indexes used for GPU surfaces, defining the cacheability of the
+ * surface data and the coherency for this data wrt. CPU vs. GPU accesses.
+ */
+enum i915_mocs_table_index {
+	/*
+	 * Not cached anywhere, coherency between CPU and GPU accesses is
+	 * guaranteed.
+	 */
+	I915_MOCS_UNCACHED,
+	/*
+	 * Cacheability and coherency controlled by the kernel automatically
+	 * based on the DRM_I915_GEM_SET_CACHING IOCTL setting and the current
+	 * usage of the surface (used for display scanout or not).
+	 */
+	I915_MOCS_PTE,
+	/*
+	 * Cached in all GPU caches available on the platform.
+	 * Coherency between CPU and GPU accesses to the surface is not
+	 * guaranteed without extra synchronization.
+	 */
+	I915_MOCS_CACHED,
+};
+
 /* Each region is a minimum of 16k, and there are at most 255 of them.
  */
 #define I915_NR_TEX_REGIONS 255	/* table size 2k - maximum due to use
@@ -218,6 +246,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_I915_OVERLAY_PUT_IMAGE	0x27
 #define DRM_I915_OVERLAY_ATTRS	0x28
 #define DRM_I915_GEM_EXECBUFFER2	0x29
+#define DRM_I915_GEM_EXECBUFFER2_WR	DRM_I915_GEM_EXECBUFFER2
 #define DRM_I915_GET_SPRITE_COLORKEY	0x2a
 #define DRM_I915_SET_SPRITE_COLORKEY	0x2b
 #define DRM_I915_GEM_WAIT	0x2c
@@ -230,6 +259,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_I915_GEM_USERPTR		0x33
 #define DRM_I915_GEM_CONTEXT_GETPARAM	0x34
 #define DRM_I915_GEM_CONTEXT_SETPARAM	0x35
+#define DRM_I915_PERF_OPEN		0x36
 
 #define DRM_IOCTL_I915_INIT		DRM_IOW( DRM_COMMAND_BASE + DRM_I915_INIT, drm_i915_init_t)
 #define DRM_IOCTL_I915_FLUSH		DRM_IO ( DRM_COMMAND_BASE + DRM_I915_FLUSH)
@@ -251,6 +281,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_IOCTL_I915_GEM_INIT		DRM_IOW(DRM_COMMAND_BASE + DRM_I915_GEM_INIT, struct drm_i915_gem_init)
 #define DRM_IOCTL_I915_GEM_EXECBUFFER	DRM_IOW(DRM_COMMAND_BASE + DRM_I915_GEM_EXECBUFFER, struct drm_i915_gem_execbuffer)
 #define DRM_IOCTL_I915_GEM_EXECBUFFER2	DRM_IOW(DRM_COMMAND_BASE + DRM_I915_GEM_EXECBUFFER2, struct drm_i915_gem_execbuffer2)
+#define DRM_IOCTL_I915_GEM_EXECBUFFER2_WR	DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_EXECBUFFER2_WR, struct drm_i915_gem_execbuffer2)
 #define DRM_IOCTL_I915_GEM_PIN		DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_PIN, struct drm_i915_gem_pin)
 #define DRM_IOCTL_I915_GEM_UNPIN	DRM_IOW(DRM_COMMAND_BASE + DRM_I915_GEM_UNPIN, struct drm_i915_gem_unpin)
 #define DRM_IOCTL_I915_GEM_BUSY		DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_BUSY, struct drm_i915_gem_busy)
@@ -283,6 +314,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_IOCTL_I915_GEM_USERPTR			DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_USERPTR, struct drm_i915_gem_userptr)
 #define DRM_IOCTL_I915_GEM_CONTEXT_GETPARAM	DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_GETPARAM, struct drm_i915_gem_context_param)
 #define DRM_IOCTL_I915_GEM_CONTEXT_SETPARAM	DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_SETPARAM, struct drm_i915_gem_context_param)
+#define DRM_IOCTL_I915_PERF_OPEN	DRM_IOW(DRM_COMMAND_BASE + DRM_I915_PERF_OPEN, struct drm_i915_perf_open_param)
 
 /* Allow drivers to submit batchbuffers directly to hardware, relying
  * on the security mechanisms provided by hardware.
@@ -357,8 +389,28 @@ typedef struct drm_i915_irq_wait {
 #define I915_PARAM_HAS_GPU_RESET	 35
 #define I915_PARAM_HAS_RESOURCE_STREAMER 36
 #define I915_PARAM_HAS_EXEC_SOFTPIN	 37
-#define I915_PARAM_HAS_POOLED_EU         38
-#define I915_PARAM_MIN_EU_IN_POOL        39
+#define I915_PARAM_HAS_POOLED_EU	 38
+#define I915_PARAM_MIN_EU_IN_POOL	 39
+#define I915_PARAM_MMAP_GTT_VERSION	 40
+
+/* Query whether DRM_I915_GEM_EXECBUFFER2 supports user defined execution
+ * priorities and the driver will attempt to execute batches in priority order.
+ * The initial priority for each batch is supplied by the context and is
+ * controlled via I915_CONTEXT_PARAM_PRIORITY.
+ */
+#define I915_PARAM_HAS_SCHEDULER	 41
+
+/* Query whether DRM_I915_GEM_EXECBUFFER2 supports the ability to opt-out of
+ * synchronisation with implicit fencing on individual objects.
+ */
+#define I915_PARAM_HAS_EXEC_ASYNC	 42
+
+/* Query whether DRM_I915_GEM_EXECBUFFER2 supports explicit fence support -
+ * both being able to pass in a sync_file fd to wait upon before executing,
+ * and being able to return a new sync_file fd that is signaled when the
+ * current request is complete.
+ */
+#define I915_PARAM_HAS_EXEC_FENCE	 43
 
 typedef struct drm_i915_getparam {
 	__s32 param;
@@ -694,15 +746,41 @@ struct drm_i915_gem_exec_object2 {
 	 */
 	__u64 offset;
 
-#define EXEC_OBJECT_NEEDS_FENCE (1<<0)
-#define EXEC_OBJECT_NEEDS_GTT	(1<<1)
-#define EXEC_OBJECT_WRITE	(1<<2)
+#define EXEC_OBJECT_NEEDS_FENCE		 (1<<0)
+#define EXEC_OBJECT_NEEDS_GTT		 (1<<1)
+#define EXEC_OBJECT_WRITE		 (1<<2)
 #define EXEC_OBJECT_SUPPORTS_48B_ADDRESS (1<<3)
-#define EXEC_OBJECT_PINNED	(1<<4)
-#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_PINNED<<1)
+#define EXEC_OBJECT_PINNED		 (1<<4)
+#define EXEC_OBJECT_PAD_TO_SIZE		 (1<<5)
+/* The kernel implicitly tracks GPU activity on all GEM objects, and
+ * synchronises operations with outstanding rendering. This includes
+ * rendering on other devices if exported via dma-buf. However, sometimes
+ * this tracking is too coarse and the user knows better. For example,
+ * if the object is split into non-overlapping ranges shared between different
+ * clients or engines (i.e. suballocating objects), the implicit tracking
+ * by kernel assumes that each operation affects the whole object rather
+ * than an individual range, causing needless synchronisation between clients.
+ * The kernel will also forgo any CPU cache flushes prior to rendering from
+ * the object as the client is expected to be also handling such domain
+ * tracking.
+ *
+ * The kernel maintains the implicit tracking in order to manage resources
+ * used by the GPU - this flag only disables the synchronisation prior to
+ * rendering with this object in this execbuf.
+ *
+ * Opting out of implicit synhronisation requires the user to do its own
+ * explicit tracking to avoid rendering corruption. See, for example,
+ * I915_PARAM_HAS_EXEC_FENCE to order execbufs and execute them asynchronously.
+ */
+#define EXEC_OBJECT_ASYNC		(1<<6)
+/* All remaining bits are MBZ and RESERVED FOR FUTURE USE */
+#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_ASYNC<<1)
 	__u64 flags;
 
-	__u64 rsvd1;
+	union {
+		__u64 rsvd1;
+		__u64 pad_to_size;
+	};
 	__u64 rsvd2;
 };
 
@@ -786,7 +864,32 @@ struct drm_i915_gem_execbuffer2 {
  */
 #define I915_EXEC_RESOURCE_STREAMER     (1<<15)
 
-#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_RESOURCE_STREAMER<<1)
+/* Setting I915_EXEC_FENCE_IN implies that lower_32_bits(rsvd2) represent
+ * a sync_file fd to wait upon (in a nonblocking manner) prior to executing
+ * the batch.
+ *
+ * Returns -EINVAL if the sync_file fd cannot be found.
+ */
+#define I915_EXEC_FENCE_IN		(1<<16)
+
+/* Setting I915_EXEC_FENCE_OUT causes the ioctl to return a sync_file fd
+ * in the upper_32_bits(rsvd2) upon success. Ownership of the fd is given
+ * to the caller, and it should be close() after use. (The fd is a regular
+ * file descriptor and will be cleaned up on process termination. It holds
+ * a reference to the request, but nothing else.)
+ *
+ * The sync_file fd can be combined with other sync_file and passed either
+ * to execbuf using I915_EXEC_FENCE_IN, to atomic KMS ioctls (so that a flip
+ * will only occur after this request completes), or to other devices.
+ *
+ * Using I915_EXEC_FENCE_OUT requires use of
+ * DRM_IOCTL_I915_GEM_EXECBUFFER2_WR ioctl so that the result is written
+ * back to userspace. Failure to do so will cause the out-fence to always
+ * be reported as zero, and the real fence fd to be leaked.
+ */
+#define I915_EXEC_FENCE_OUT		(1<<17)
+
+#define __I915_EXEC_UNKNOWN_FLAGS (-(I915_EXEC_FENCE_OUT<<1))
 
 #define I915_EXEC_CONTEXT_ID_MASK	(0xffffffff)
 #define i915_execbuffer2_set_context_id(eb2, context) \
@@ -822,7 +925,16 @@ struct drm_i915_gem_busy {
 	 * having flushed any pending activity), and a non-zero return that
 	 * the object is still in-flight on the GPU. (The GPU has not yet
 	 * signaled completion for all pending requests that reference the
-	 * object.)
+	 * object.) An object is guaranteed to become idle eventually (so
+	 * long as no new GPU commands are executed upon it). Due to the
+	 * asynchronous nature of the hardware, an object reported
+	 * as busy may become idle before the ioctl is completed.
+	 *
+	 * Furthermore, if the object is busy, which engine is busy is only
+	 * provided as a guide. There are race conditions which prevent the
+	 * report of which engines are busy from being always accurate.
+	 * However, the converse is not true. If the object is idle, the
+	 * result of the ioctl, that all engines are idle, is accurate.
 	 *
 	 * The returned dword is split into two fields to indicate both
 	 * the engines on which the object is being read, and the
@@ -845,6 +957,11 @@ struct drm_i915_gem_busy {
 	 * execution engines, e.g. multiple media engines, which are
 	 * mapped to the same identifier in the EXECBUFFER2 ioctl and
 	 * so are not separately reported for busyness.
+	 *
+	 * Caveat emptor:
+	 * Only the boolean result of this query is reliable; that is whether
+	 * the object is idle or busy. The report of which engines are busy
+	 * should be only used as a heuristic.
 	 */
 	__u32 busy;
 };
@@ -893,6 +1010,7 @@ struct drm_i915_gem_caching {
 #define I915_TILING_NONE	0
 #define I915_TILING_X		1
 #define I915_TILING_Y		2
+#define I915_TILING_LAST	I915_TILING_Y
 
 #define I915_BIT_6_SWIZZLE_NONE		0
 #define I915_BIT_6_SWIZZLE_9		1
@@ -1169,7 +1287,146 @@ struct drm_i915_gem_context_param {
 #define I915_CONTEXT_PARAM_BAN_PERIOD	0x1
 #define I915_CONTEXT_PARAM_NO_ZEROMAP	0x2
 #define I915_CONTEXT_PARAM_GTT_SIZE	0x3
+#define I915_CONTEXT_PARAM_NO_ERROR_CAPTURE	0x4
+#define I915_CONTEXT_PARAM_BANNABLE	0x5
+#define I915_CONTEXT_PARAM_PRIORITY	0x6
 	__u64 value;
 };
 
+enum drm_i915_oa_format {
+	I915_OA_FORMAT_A13 = 1,
+	I915_OA_FORMAT_A29,
+	I915_OA_FORMAT_A13_B8_C8,
+	I915_OA_FORMAT_B4_C8,
+	I915_OA_FORMAT_A45_B8_C8,
+	I915_OA_FORMAT_B4_C8_A16,
+	I915_OA_FORMAT_C4_B8,
+
+	I915_OA_FORMAT_MAX	    /* non-ABI */
+};
+
+enum drm_i915_perf_property_id {
+	/**
+	 * Open the stream for a specific context handle (as used with
+	 * execbuffer2). A stream opened for a specific context this way
+	 * won't typically require root privileges.
+	 */
+	DRM_I915_PERF_PROP_CTX_HANDLE = 1,
+
+	/**
+	 * A value of 1 requests the inclusion of raw OA unit reports as
+	 * part of stream samples.
+	 */
+	DRM_I915_PERF_PROP_SAMPLE_OA,
+
+	/**
+	 * The value specifies which set of OA unit metrics should be
+	 * be configured, defining the contents of any OA unit reports.
+	 */
+	DRM_I915_PERF_PROP_OA_METRICS_SET,
+
+	/**
+	 * The value specifies the size and layout of OA unit reports.
+	 */
+	DRM_I915_PERF_PROP_OA_FORMAT,
+
+	/**
+	 * Specifying this property implicitly requests periodic OA unit
+	 * sampling and (at least on Haswell) the sampling frequency is derived
+	 * from this exponent as follows:
+	 *
+	 *   80ns * 2^(period_exponent + 1)
+	 */
+	DRM_I915_PERF_PROP_OA_EXPONENT,
+
+	DRM_I915_PERF_PROP_MAX /* non-ABI */
+};
+
+struct drm_i915_perf_open_param {
+	__u32 flags;
+#define I915_PERF_FLAG_FD_CLOEXEC	(1<<0)
+#define I915_PERF_FLAG_FD_NONBLOCK	(1<<1)
+#define I915_PERF_FLAG_DISABLED		(1<<2)
+
+	/** The number of u64 (id, value) pairs */
+	__u32 num_properties;
+
+	/**
+	 * Pointer to array of u64 (id, value) pairs configuring the stream
+	 * to open.
+	 */
+	__u64 properties_ptr;
+};
+
+/**
+ * Enable data capture for a stream that was either opened in a disabled state
+ * via I915_PERF_FLAG_DISABLED or was later disabled via
+ * I915_PERF_IOCTL_DISABLE.
+ *
+ * It is intended to be cheaper to disable and enable a stream than it may be
+ * to close and re-open a stream with the same configuration.
+ *
+ * It's undefined whether any pending data for the stream will be lost.
+ */
+#define I915_PERF_IOCTL_ENABLE	_IO('i', 0x0)
+
+/**
+ * Disable data capture for a stream.
+ *
+ * It is an error to try and read a stream that is disabled.
+ */
+#define I915_PERF_IOCTL_DISABLE	_IO('i', 0x1)
+
+/**
+ * Common to all i915 perf records
+ */
+struct drm_i915_perf_record_header {
+	__u32 type;
+	__u16 pad;
+	__u16 size;
+};
+
+enum drm_i915_perf_record_type {
+
+	/**
+	 * Samples are the work horse record type whose contents are extensible
+	 * and defined when opening an i915 perf stream based on the given
+	 * properties.
+	 *
+	 * Boolean properties following the naming convention
+	 * DRM_I915_PERF_SAMPLE_xyz_PROP request the inclusion of 'xyz' data in
+	 * every sample.
+	 *
+	 * The order of these sample properties given by userspace has no
+	 * affect on the ordering of data within a sample. The order is
+	 * documented here.
+	 *
+	 * struct {
+	 *     struct drm_i915_perf_record_header header;
+	 *
+	 *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
+	 * };
+	 */
+	DRM_I915_PERF_RECORD_SAMPLE = 1,
+
+	/*
+	 * Indicates that one or more OA reports were not written by the
+	 * hardware. This can happen for example if an MI_REPORT_PERF_COUNT
+	 * command collides with periodic sampling - which would be more likely
+	 * at higher sampling frequencies.
+	 */
+	DRM_I915_PERF_RECORD_OA_REPORT_LOST = 2,
+
+	/**
+	 * An error occurred that resulted in all pending OA reports being lost.
+	 */
+	DRM_I915_PERF_RECORD_OA_BUFFER_LOST = 3,
+
+	DRM_I915_PERF_RECORD_MAX /* non-ABI */
+};
+
+#if defined(__cplusplus)
+}
+#endif
+
 #endif /* _I915_DRM_H_ */
Signature · 6 lines
2.11.0.21.ga274e0a

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

PATCH 2/3 intel: Allow the client to control implicit synchronisation

In reply to Chad Versace

From: Chris Wilson <[email protected]>

The kernel allows implicit synchronisation to be disabled on individual buffers. Use at your own risk.

---
 intel/intel_bufmgr.h     |  2 ++
 intel/intel_bufmgr_gem.c | 32 ++++++++++++++++++++++++++++----
 2 files changed, 30 insertions(+), 4 deletions(-)
intel/intel_bufmgr.h +2 -0
index 85e4ff78..6f128679 100644
@@ -184,6 +184,8 @@ int drm_intel_gem_bo_map_unsynchronized(drm_intel_bo *bo);
 int drm_intel_gem_bo_map_gtt(drm_intel_bo *bo);
 int drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo);
 
+void drm_intel_gem_bo_disable_implicit_sync(drm_intel_bo *bo);
+
 void *drm_intel_gem_bo_map__cpu(drm_intel_bo *bo);
 void *drm_intel_gem_bo_map__gtt(drm_intel_bo *bo);
 void *drm_intel_gem_bo_map__wc(drm_intel_bo *bo);
intel/intel_bufmgr_gem.c +28 -4
index 75949b9f..c5fc1b36 100644
@@ -195,8 +195,11 @@ struct _drm_intel_bo_gem {
 	uint32_t swizzle_mode;
 	unsigned long stride;
 
+	unsigned long kflags;
+
 	time_t free_time;
 
+
 	/** Array passed to the DRM containing relocation information. */
 	struct drm_i915_gem_relocation_entry *relocs;
 	/**
@@ -575,12 +578,11 @@ drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence)
 	bufmgr_gem->exec2_objects[index].relocation_count = bo_gem->reloc_count;
 	bufmgr_gem->exec2_objects[index].relocs_ptr = (uintptr_t)bo_gem->relocs;
 	bufmgr_gem->exec2_objects[index].alignment = bo->align;
-	bufmgr_gem->exec2_objects[index].offset = bo_gem->is_softpin ?
-		bo->offset64 : 0;
-	bufmgr_gem->exec_bos[index] = bo;
-	bufmgr_gem->exec2_objects[index].flags = flags;
+	bufmgr_gem->exec2_objects[index].offset = bo->offset64;
+	bufmgr_gem->exec2_objects[index].flags = flags | bo_gem->kflags;
 	bufmgr_gem->exec2_objects[index].rsvd1 = 0;
 	bufmgr_gem->exec2_objects[index].rsvd2 = 0;
+	bufmgr_gem->exec_bos[index] = bo;
 	bufmgr_gem->exec_count++;
 }
 
@@ -1368,6 +1370,7 @@ drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)
 	for (i = 0; i < bo_gem->softpin_target_count; i++)
 		drm_intel_gem_bo_unreference_locked_timed(bo_gem->softpin_target[i],
 								  time);
+	bo_gem->kflags = 0;
 	bo_gem->reloc_count = 0;
 	bo_gem->used_as_reloc_target = false;
 	bo_gem->softpin_target_count = 0;
@@ -2765,6 +2768,27 @@ drm_intel_bufmgr_gem_enable_reuse(drm_intel_bufmgr *bufmgr)
 }
 
 /**
+ * Disables implicit synchronisation before executing the bo
+ *
+ * This will cause rendering corruption unless you correctly manage explicit
+ * fences for all rendering involving this buffer - including use by others.
+ * Disabling the implicit serialisation is only required if that serialisation
+ * is too coarse (for example, you have split the buffer into many
+ * non-overlapping regions and are sharing the whole buffer between concurrent
+ * independent command streams).
+ *
+ * Note the kernel must advertise support via I915_PARAM_HAS_EXEC_ASYNC or
+ * subsequent execbufs involving the bo will generate EINVAL.
+ */
+void
+drm_intel_gem_bo_disable_implicit_sync(drm_intel_bo *bo)
+{
+	drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+	bo_gem->kflags |= EXEC_OBJECT_ASYNC;
+}
+
+/**
  * Enable use of fenced reloc type.
  *
  * New code should enable this to avoid unnecessary fence register
Signature · 6 lines
2.11.0.21.ga274e0a

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

PATCH 3/3 intel: Support passing of explicit fencing from execbuf

In reply to Chad Versace

From: Chris Wilson <[email protected]>

Allow the caller to pass in an fd to an array of fences to control serialisation of the execbuf in the kernel and on the GPU, and in return allow creation of a fence fd for signaling the completion (and flushing) of the batch. When the returned fence is signaled, all writes to the buffers inside the batch will be complete and coherent from the cpu, or other consumers. The return fence is a sync_file object and can be passed to other users (such as atomic modesetting, or other drivers).

Fixes from rantogno:
- Fix the in/out fence flags on execbuf.

Squashed-fixes-from: Rafael Antognolli <[email protected]>

---
 intel/intel_bufmgr.h     |  6 ++++++
 intel/intel_bufmgr_gem.c | 32 +++++++++++++++++++++++++++-----
 2 files changed, 33 insertions(+), 5 deletions(-)
intel/intel_bufmgr.h +6 -0
index 6f128679..49cabf67 100644
@@ -219,6 +219,12 @@ int drm_intel_gem_context_get_id(drm_intel_context *ctx,
 void drm_intel_gem_context_destroy(drm_intel_context *ctx);
 int drm_intel_gem_bo_context_exec(drm_intel_bo *bo, drm_intel_context *ctx,
 				  int used, unsigned int flags);
+int drm_intel_gem_bo_fence_exec(drm_intel_bo *bo,
+				drm_intel_context *ctx,
+				int used,
+				int in_fence,
+				int *out_fence,
+				unsigned int flags);
 
 int drm_intel_bo_gem_export_to_prime(drm_intel_bo *bo, int *prime_fd);
 drm_intel_bo *drm_intel_bo_gem_create_from_prime(drm_intel_bufmgr *bufmgr,
intel/intel_bufmgr_gem.c +27 -5
index c5fc1b36..da993002 100644
@@ -2376,6 +2376,7 @@ drm_intel_gem_bo_exec(drm_intel_bo *bo, int used,
 static int
 do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx,
 	 drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
+	 int in_fence, int *out_fence,
 	 unsigned int flags)
 {
 	drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
@@ -2429,13 +2430,20 @@ do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx,
 		i915_execbuffer2_set_context_id(execbuf, 0);
 	else
 		i915_execbuffer2_set_context_id(execbuf, ctx->ctx_id);
-	execbuf.rsvd2 = 0;
+	if (in_fence != -1) {
+		execbuf.rsvd2 = in_fence;
+		execbuf.flags |= I915_EXEC_FENCE_IN;
+	}
+	if (out_fence != NULL) {
+		*out_fence = -1;
+		execbuf.flags |= I915_EXEC_FENCE_OUT;
+	}
 
 	if (bufmgr_gem->no_exec)
 		goto skip_execution;
 
 	ret = drmIoctl(bufmgr_gem->fd,
-		       DRM_IOCTL_I915_GEM_EXECBUFFER2,
+		       DRM_IOCTL_I915_GEM_EXECBUFFER2_WR,
 		       &execbuf);
 	if (ret != 0) {
 		ret = -errno;
@@ -2451,6 +2459,9 @@ do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx,
 	}
 	drm_intel_update_buffer_offsets2(bufmgr_gem);
 
+	if (out_fence != NULL)
+		*out_fence = execbuf.rsvd2 >> 32;
+
 skip_execution:
 	if (bufmgr_gem->bufmgr.debug)
 		drm_intel_gem_dump_validation_list(bufmgr_gem);
@@ -2476,7 +2487,7 @@ drm_intel_gem_bo_exec2(drm_intel_bo *bo, int used,
 		       int DR4)
 {
 	return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
-			I915_EXEC_RENDER);
+			-1, NULL, I915_EXEC_RENDER);
 }
 
 static int
@@ -2485,14 +2496,25 @@ drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used,
 			unsigned int flags)
 {
 	return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
-			flags);
+			-1, NULL, flags);
 }
 
 int
 drm_intel_gem_bo_context_exec(drm_intel_bo *bo, drm_intel_context *ctx,
 			      int used, unsigned int flags)
 {
-	return do_exec2(bo, used, ctx, NULL, 0, 0, flags);
+	return do_exec2(bo, used, ctx, NULL, 0, 0, -1, NULL, flags);
+}
+
+int
+drm_intel_gem_bo_fence_exec(drm_intel_bo *bo,
+			    drm_intel_context *ctx,
+			    int used,
+			    int in_fence,
+			    int *out_fence,
+			    unsigned int flags)
+{
+	return do_exec2(bo, used, ctx, NULL, 0, 0, in_fence, out_fence, flags);
 }
 
 static int
Signature · 6 lines
2.11.0.21.ga274e0a

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