Skip to contentSkip to conversation

drm/i915/gvt: Dma-buf support for GVT-g

PATCH Superseded Tina Zhang 6 participants 31 messages Aug 18, 2017 – Oct 7, 2017

Subsystems (5)

v14 of 19 Superseded Latest is v19 →

lore
Share & tools
lore

Tina Zhang

author
#0 lore raw

v14 cover drm/i915/gvt: Dma-buf support for GVT-g

v13->v14:
1) add PROBE, DMABUF and REGION flags. (Alex)
2) return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)
3) add some details about the float pixel format. (Daniel)
4) add F suffix to the defined name. (Daniel)
5) refine pixel format table. (Zhenyu)

v12->v13:
1) add comments to GEM proxy. (Chris)
2) don't ban GEM proxy in i915_gem_sw_finish_ioctl. (Chris)
3) check GEM proxy bar after finishing i915_gem_object_wait. (Chris)
4) remove GEM proxy bar in i915_gem_madvise_ioctl.

v11->v12:
1) add drm_format_mod back. (Gerd and Zhenyu)
2) add region_index. (Gerd)
3) refine the lifecycle of dmabuf.
4) send to [email protected]. (Ville)

v10->v11:
1) rename plane_type to drm_plane_type. (Gerd)
2) move fields of vfio_device_query_gfx_plane to
vfio_device_gfx_plane_info. (Gerd)
3) remove drm_format_mod, start fields. (Daniel)
4) remove plane_id.

v9->v10:
1) remove dma-buf management
2) refine the ABI API VFIO_DEVICE_QUERY_GFX_PLANE
3) track the dma-buf create and release in kernel mode

v8->v9:
1) refine the dma-buf ioctl definition
2) add a lock to protect the dmabuf list
3) move drm format change to a separate patch
4) codes cleanup

v7->v8:
1) refine framebuffer decoder code
2) fix a bug in decoding primary plane

v6->v7:
1) release dma-buf related allocations in dma-buf's associated release function.
2) refine ioctl interface for querying plane info or create dma-buf
3) refine framebuffer decoder code
4) the patch series is based on 4.12.0-rc1

v5->v6:
1) align the dma-buf life cycle with the vfio device.
2) add the dma-buf related operations in a separate patch.
3) i915 releated changes.

v4->v5:
1) fix bug while checking whether the gem obj is gvt's dma-buf when user change caching mode or domains. Add a helper function to do it.
2) add definition for the query plane and create dma-buf.

v3->v4:
1) fix bug while checking whether the gem obj is gvt's dma-buf when set caching mode or doamins.

v2->v3:
1) add a field gvt_plane_info in the drm_i915_gem_obj structure to save the decoded plane information to avoid look up while need the plane info.
2) declare a new flag I915_GEM_OBJECT_IS_GVT_DMABUF in drm_i915_gem_object to represent the gem obj for gvt's dma-buf. The tiling mode, caching mode and domains can not be changed for this kind of gem object.
3) change dma-buf related information to be more generic. So other vendor can use the same interface.

v1->v2:
1) create a management fd for dma-buf operations.
2) alloc gem object's backing storage in gem obj's get_pages() callback.

This patch set adds the dma-buf support for intel GVT-g. dma-buf is a uniform mechanism to share DMA buffers across different devices and sub-systems.
dma-buf for intel GVT-g is mainly used to share the vgpu's framebuffer to other users or sub-systems so they can use the dma-buf to show the desktop of a VM which uses intel vgpu.

The main idea is we create a gem object and set vgpu's framebuffer as the backing storage of this gem object. And associate this gem obj to a dma-buf object then export this dma-buf at the meantime generate a file descriptor for this dma-buf. Finally deliver this file descriptor to user space. And user can use this dma-buf fd to do rendering or other operations.

This patch set can be tried with the following example:
https://github.com/01org/Igvtg-qemu branch: topic/dmabuf

Tina Zhang (7):
drm/i915/gvt: Add framebuffer decoder support
drm: Introduce RGB 64-bit 16:16:16:16 float format drm/i915/gvt: Add RGB 64-bit 16:16:16:16 float format drm/i915/gvt: Add opregion support
vfio: ABI for mdev display dma-buf operation drm/i915: Introduce GEM proxy
drm/i915/gvt: Dmabuf support for GVT-g

 drivers/gpu/drm/i915/gvt/Makefile      |   3 +-
 drivers/gpu/drm/i915/gvt/display.c     |   2 +-
 drivers/gpu/drm/i915/gvt/display.h     |   2 +
 drivers/gpu/drm/i915/gvt/dmabuf.c      | 387 ++++++++++++++++++++++++
 drivers/gpu/drm/i915/gvt/dmabuf.h      |  58 ++++
 drivers/gpu/drm/i915/gvt/fb_decoder.c  | 521 +++++++++++++++++++++++++++++++++
 drivers/gpu/drm/i915/gvt/fb_decoder.h  | 169 +++++++++++
 drivers/gpu/drm/i915/gvt/gvt.c         |   1 +
 drivers/gpu/drm/i915/gvt/gvt.h         |  10 +
 drivers/gpu/drm/i915/gvt/hypercall.h   |   3 +
 drivers/gpu/drm/i915/gvt/kvmgt.c       | 151 +++++++++-
 drivers/gpu/drm/i915/gvt/mpt.h         |  45 +++
 drivers/gpu/drm/i915/gvt/opregion.c    |  26 +-
 drivers/gpu/drm/i915/gvt/vgpu.c        |   6 +-
 drivers/gpu/drm/i915/i915_gem.c        |  24 +-
 drivers/gpu/drm/i915/i915_gem_object.h |   9 +
 drivers/gpu/drm/i915/i915_gem_tiling.c |   8 +
 include/uapi/drm/drm_fourcc.h          |   4 +
 include/uapi/linux/vfio.h              |  43 +++
 19 files changed, 1459 insertions(+), 13 deletions(-)
 create mode 100644 drivers/gpu/drm/i915/gvt/dmabuf.c
 create mode 100644 drivers/gpu/drm/i915/gvt/dmabuf.h
 create mode 100644 drivers/gpu/drm/i915/gvt/fb_decoder.c
 create mode 100644 drivers/gpu/drm/i915/gvt/fb_decoder.h
Signature · 6 lines
2.7.4

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

PATCH v14 1/7 drm/i915/gvt: Add framebuffer decoder support

This patch is to introduce the framebuffer decoder which can decode guest OS's framebuffer information, including primary, cursor and sprite plane.

v14:
- refine pixel format table. (Zhenyu)

v9:
- move drm format change to a separate patch. (Xiaoguang)

v8:
- fix a bug in decoding primary plane. (Tina)

v7:
- refine framebuffer decoder code. (Zhenyu)

drivers/gpu/drm/i915/gvt/Makefile +2 -1
index f5486cb9..019d596 100644
@@ -1,7 +1,8 @@
 GVT_DIR := gvt
 GVT_SOURCE := gvt.o aperture_gm.o handlers.o vgpu.o trace_points.o firmware.o \
 	interrupt.o gtt.o cfg_space.o opregion.o mmio.o display.o edid.o \
-	execlist.o scheduler.o sched_policy.o render.o cmd_parser.o
+	execlist.o scheduler.o sched_policy.o render.o cmd_parser.o \
+	fb_decoder.o
 
 ccflags-y				+= -I$(src) -I$(src)/$(GVT_DIR)
 i915-y					+= $(addprefix $(GVT_DIR)/, $(GVT_SOURCE))
drivers/gpu/drm/i915/gvt/display.c +1 -1
index 3c31843..fb7fdba 100644
@@ -67,7 +67,7 @@ static int edp_pipe_is_enabled(struct intel_vgpu *vgpu)
 	return 1;
 }
 
-static int pipe_is_enabled(struct intel_vgpu *vgpu, int pipe)
+int pipe_is_enabled(struct intel_vgpu *vgpu, int pipe)
 {
 	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 
drivers/gpu/drm/i915/gvt/display.h +2 -0
index d73de22..b46b868 100644
@@ -179,4 +179,6 @@ int intel_vgpu_init_display(struct intel_vgpu *vgpu, u64 resolution);
 void intel_vgpu_reset_display(struct intel_vgpu *vgpu);
 void intel_vgpu_clean_display(struct intel_vgpu *vgpu);
 
+int pipe_is_enabled(struct intel_vgpu *vgpu, int pipe);
+
 #endif
drivers/gpu/drm/i915/gvt/fb_decoder.c new file +507 -0

507 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/fb_decoder.h new file +169 -0

169 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/gvt.h +1 -0
index 4c1a442..0f562ee 100644
@@ -46,6 +46,7 @@
 #include "sched_policy.h"
 #include "render.h"
 #include "cmd_parser.h"
+#include "fb_decoder.h"
 
 #define GVT_MAX_VGPU 8
 
Signature · 6 lines
2.7.4

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

Re: [PATCH v14 1/7] drm/i915/gvt: Add framebuffer decoder support

Quoted 41 lines · Tina Zhang

This patch is to introduce the framebuffer decoder which can decode guest OS's framebuffer information, including primary, cursor and sprite plane.

v14:
- refine pixel format table. (Zhenyu)

v9:
- move drm format change to a separate patch. (Xiaoguang)

v8:
- fix a bug in decoding primary plane. (Tina)

v7:
- refine framebuffer decoder code. (Zhenyu)

drivers/gpu/drm/i915/gvt/fb_decoder.c new file +22 -0
+static struct pixel_format bdw_pixel_formats[] = {
+	{DRM_FORMAT_C8, 8, "8-bit Indexed"},
+	{DRM_FORMAT_RGB565, 16, "16-bit BGRX (5:6:5 MSB-R:G:B)"},
+	{DRM_FORMAT_XRGB8888, 32, "32-bit BGRX (8:8:8:8 MSB-X:R:G:B)"},
+	{DRM_FORMAT_XBGR2101010, 32, "32-bit RGBX (2:10:10:10 MSB-X:B:G:R)"},
+
+	{DRM_FORMAT_XRGB2101010, 32, "32-bit BGRX (2:10:10:10 MSB-X:R:G:B)"},
+	{DRM_FORMAT_XBGR8888, 32, "32-bit RGBX (8:8:8:8 MSB-X:B:G:R)"},
+
+	/* non-supported format has bpp default to 0 */
+	{0, 0, NULL},
+};
+
+static struct pixel_format skl_pixel_formats[] = {
+	{DRM_FORMAT_YUYV, 16, "16-bit packed YUYV (8:8:8:8 MSB-V:Y2:U:Y1)"},
+	{DRM_FORMAT_UYVY, 16, "16-bit packed UYVY (8:8:8:8 MSB-Y2:V:Y1:U)"},
+	{DRM_FORMAT_YVYU, 16, "16-bit packed YVYU (8:8:8:8 MSB-U:Y2:V:Y1)"},
+	{DRM_FORMAT_VYUY, 16, "16-bit packed VYUY (8:8:8:8 MSB-Y2:U:Y1:V)"},
+
+	{DRM_FORMAT_C8, 8, "8-bit Indexed"},
+	{DRM_FORMAT_RGB565, 16, "16-bit BGRX (5:6:5 MSB-R:G:B)"},
+	{DRM_FORMAT_ABGR8888, 32, "32-bit RGBA (8:8:8:8 MSB-A:B:G:R)"},

drivers/gpu/drm/i915/gvt/fb_decoder.c line 78

Quoted 10 lines
drivers/gpu/drm/i915/gvt/fb_decoder.c new file +10 -0
+	{DRM_FORMAT_XBGR8888, 32, "32-bit RGBX (8:8:8:8 MSB-X:B:G:R)"},
+
+	{DRM_FORMAT_ARGB8888, 32, "32-bit BGRA (8:8:8:8 MSB-A:R:G:B)"},
+	{DRM_FORMAT_XRGB8888, 32, "32-bit BGRX (8:8:8:8 MSB-X:R:G:B)"},
+	{DRM_FORMAT_XBGR2101010, 32, "32-bit RGBX (2:10:10:10 MSB-X:B:G:R)"},
+	{DRM_FORMAT_XRGB2101010, 32, "32-bit BGRX (2:10:10:10 MSB-X:R:G:B)"},
+
+	/* non-supported format has bpp default to 0 */
+	{0, 0, NULL},
+};

What about KBL support of this? If not fully supported now, need to set proper state at "probe" ioctl time.

Signature · 3 lines
Open Source Technology Center, Intel ltd.

$gpg --keyserver wwwkeys.pgp.net --recv-keys 4D781827

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

author
#3 lore raw

PATCH v14 2/7 drm: Introduce RGB 64-bit 16:16:16:16 float format

In reply to Tina Zhang

The RGB 64-bit 16:16:16:16 float pixel format is needed by some Apps in windows. The float format in each component is 1:5:10 MSb-sign:exponent:
fraction.

This patch is to introduce the format to drm, so that the windows guest's framebuffer in this kind of format can be recognized and used by linux host.

v14:
- add some details about the float pixel format. (Daniel)
- add F suffix to the defined name. (Daniel)

v12:
- send to dri-devel at lists.freedesktop.org. (Ville)

v9:
- separated from framebuffer decoder patch. (Zhenyu) (Xiaoguang)

Cc Ville Syrjälä <[email protected]>
Cc Dave Airlie <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Joonas Lahtinen <[email protected]>
include/uapi/drm/drm_fourcc.h +4 -0
index 76c9101..575014f 100644
@@ -113,6 +113,10 @@ extern "C" {
 
 #define DRM_FORMAT_AYUV		fourcc_code('A', 'Y', 'U', 'V') /* [31:0] A:Y:Cb:Cr 8:8:8:8 little endian */
 
+/* 64 bpp RGB 16:16:16:16 Floating Point */
+#define DRM_FORMAT_XRGB161616F  fourcc_code('X', 'R', '3', 'F') /* [63:0] x:R:G:B 16:16:16:16 little endian */
+#define DRM_FORMAT_XBGR161616F  fourcc_code('X', 'B', '3', 'F') /* [63:0] x:B:G:R 16:16:16:16 little endian */
+
 /*
  * 2 plane RGB + A
  * index 0 = RGB plane, same format as the corresponding non _A8 format has
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#4 lore raw

PATCH v14 3/7 drm/i915/gvt: Add RGB 64-bit 16:16:16:16 float format

In reply to Tina Zhang

The RGB 64-bit 16:16:16:16 float pixel format is needed by windows 10 guest VM. This patch is to add this pixel format support to gvt device model. Without this patch, some Apps, e.g. "DXGIGammaVM.exe", will crash and make guest screen black.

drivers/gpu/drm/i915/gvt/fb_decoder.c +21 -7
index d380ab8..49053f4 100644
@@ -52,6 +52,8 @@ static struct pixel_format bdw_pixel_formats[] = {
 
 	{DRM_FORMAT_XRGB2101010, 32, "32-bit BGRX (2:10:10:10 MSB-X:R:G:B)"},
 	{DRM_FORMAT_XBGR8888, 32, "32-bit RGBX (8:8:8:8 MSB-X:B:G:R)"},
+	{DRM_FORMAT_XRGB161616F, 64,
+		"64-bit XRGB Floating Point (16:16:16:16 MSB-X:R:G:B)"},
 
 	/* non-supported format has bpp default to 0 */
 	{0, 0, NULL},
@@ -73,13 +75,18 @@ static struct pixel_format skl_pixel_formats[] = {
 	{DRM_FORMAT_XBGR2101010, 32, "32-bit RGBX (2:10:10:10 MSB-X:B:G:R)"},
 	{DRM_FORMAT_XRGB2101010, 32, "32-bit BGRX (2:10:10:10 MSB-X:R:G:B)"},
 
+	{DRM_FORMAT_XRGB161616F, 64,
+		"64-bit XRGB Floating Point (16:16:16:16 MSB-X:R:G:B)"},
+	{DRM_FORMAT_XBGR161616F, 64,
+		"64-bit XBGR Floating Point (16:16:16:16 MSB-X:B:G:R)"},
+
 	/* non-supported format has bpp default to 0 */
 	{0, 0, NULL},
 };
 
 static int bdw_format_to_drm(int format)
 {
-	int bdw_pixel_formats_index = 6;
+	int bdw_pixel_formats_index = 7;
 
 	switch (format) {
 	case DISPPLANE_8BPP:
@@ -100,7 +107,9 @@ static int bdw_format_to_drm(int format)
 	case DISPPLANE_RGBX888:
 		bdw_pixel_formats_index = 5;
 		break;
-
+	case DISPPLANE_RGBX161616:
+		bdw_pixel_formats_index = 6;
+		break;
 	default:
 		break;
 	}
@@ -111,7 +120,7 @@ static int bdw_format_to_drm(int format)
 static int skl_format_to_drm(int format, bool rgb_order, bool alpha,
 	int yuv_order)
 {
-	int skl_pixel_formats_index = 12;
+	int skl_pixel_formats_index = 14;
 
 	switch (format) {
 	case PLANE_CTL_FORMAT_INDEXED:
@@ -129,6 +138,9 @@ static int skl_format_to_drm(int format, bool rgb_order, bool alpha,
 	case PLANE_CTL_FORMAT_XRGB_2101010:
 		skl_pixel_formats_index = rgb_order ? 10 : 11;
 		break;
+	case PLANE_CTL_FORMAT_XRGB_16161616F:
+		skl_pixel_formats_index = rgb_order ? 12 : 13;
+		break;
 	case PLANE_CTL_FORMAT_YUV422:
 		skl_pixel_formats_index = yuv_order >> 16;
 		if (skl_pixel_formats_index > 3)
@@ -389,11 +401,13 @@ int intel_vgpu_decode_cursor_plane(struct intel_vgpu *vgpu,
 #define SPRITE_FORMAT_NUM	(1 << 3)
 
 static struct pixel_format sprite_pixel_formats[SPRITE_FORMAT_NUM] = {
-	[0x0] = {DRM_FORMAT_YUV422, 16, "YUV 16-bit 4:2:2 packed"},
-	[0x1] = {DRM_FORMAT_XRGB2101010, 32, "RGB 32-bit 2:10:10:10"},
-	[0x2] = {DRM_FORMAT_XRGB8888, 32, "RGB 32-bit 8:8:8:8"},
+	[0x0] = {DRM_FORMAT_YUV422, 16, "16-bit YUV 4:2:2 packed"},
+	[0x1] = {DRM_FORMAT_XRGB2101010, 32, "32-bit BGRX (2:10:10:10 MSB-X:R:G:B)"},
+	[0x2] = {DRM_FORMAT_XRGB8888, 32, "32-bit RGBX (8:8:8:8 MSB-X:B:G:R)"},
+	[0x3]  = {DRM_FORMAT_XRGB161616F, 64,
+		    "64-bit XRGB Floating Point (16:16:16:16 MSB-X:R:G:B)"},
 	[0x4] = {DRM_FORMAT_AYUV, 32,
-		"YUV 32-bit 4:4:4 packed (8:8:8:8 MSB-X:Y:U:V)"},
+		"32-bit YUV 4:4:4 packed (8:8:8:8 MSB-X:Y:U:V)"},
 };
 
 /**
Signature · 6 lines
2.7.4

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

PATCH v14 4/7 drm/i915/gvt: Add opregion support

In reply to Tina Zhang

Windows guest driver needs vbt in opregion, to configure the setting for display. Without opregion support, the display registers won't be set and this blocks display model to get the correct information of the guest display plane.

This patch is to provide a virtual opregion for guest. Current implementation is to fill the virtual opregion with the content in host's opregion. The original author of this patch is Xiaoguang Chen.

drivers/gpu/drm/i915/gvt/hypercall.h +1 -0
index df7f33a..32c345c 100644
@@ -55,6 +55,7 @@ struct intel_gvt_mpt {
 			      unsigned long mfn, unsigned int nr, bool map);
 	int (*set_trap_area)(unsigned long handle, u64 start, u64 end,
 			     bool map);
+	int (*set_opregion)(void *vgpu);
 };
 
 extern struct intel_gvt_mpt xengt_mpt;
drivers/gpu/drm/i915/gvt/kvmgt.c +106 -3

157 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/mpt.h +15 -0
index f0e5487..9e73f2e 100644
@@ -292,4 +292,19 @@ static inline int intel_gvt_hypervisor_set_trap_area(
 	return intel_gvt_host.mpt->set_trap_area(vgpu->handle, start, end, map);
 }
 
+/**
+ * intel_gvt_hypervisor_set_opregion - Set opregion for guest
+ * @vgpu: a vGPU
+ *
+ * Returns:
+ * Zero on success, negative error code if failed.
+ */
+static inline int intel_gvt_hypervisor_set_opregion(struct intel_vgpu *vgpu)
+{
+	if (!intel_gvt_host.mpt->set_opregion)
+		return 0;
+
+	return intel_gvt_host.mpt->set_opregion(vgpu);
+}
+
 #endif /* _GVT_MPT_H_ */
drivers/gpu/drm/i915/gvt/opregion.c +20 -6
index 3117991..04c0452 100644
@@ -113,23 +113,37 @@ void intel_vgpu_clean_opregion(struct intel_vgpu *vgpu)
  */
 int intel_vgpu_init_opregion(struct intel_vgpu *vgpu, u32 gpa)
 {
-	int ret;
+	int ret = 0;
+	unsigned long pfn;
 
 	gvt_dbg_core("vgpu%d: init vgpu opregion\n", vgpu->id);
 
-	if (intel_gvt_host.hypervisor_type == INTEL_GVT_HYPERVISOR_XEN) {
+	switch (intel_gvt_host.hypervisor_type) {
+	case INTEL_GVT_HYPERVISOR_KVM:
+		pfn = intel_gvt_hypervisor_gfn_to_mfn(vgpu, gpa >> PAGE_SHIFT);
+		vgpu_opregion(vgpu)->va = memremap(pfn << PAGE_SHIFT,
+						INTEL_GVT_OPREGION_SIZE,
+						MEMREMAP_WB);
+		if (!vgpu_opregion(vgpu)->va) {
+			gvt_vgpu_err("failed to map guest opregion\n");
+			ret = -EFAULT;
+		}
+		break;
+	case INTEL_GVT_HYPERVISOR_XEN:
 		gvt_dbg_core("emulate opregion from kernel\n");
 
 		ret = init_vgpu_opregion(vgpu, gpa);
 		if (ret)
-			return ret;
+			break;
 
 		ret = map_vgpu_opregion(vgpu, true);
-		if (ret)
-			return ret;
+		break;
+	default:
+		ret = -EINVAL;
+		gvt_vgpu_err("not supported hypervisor\n");
 	}
 
-	return 0;
+	return ret;
 }
 
 /**
drivers/gpu/drm/i915/gvt/vgpu.c +4 -0
index 3deadcb..bcbf535 100644
@@ -383,6 +383,10 @@ static struct intel_vgpu *__intel_gvt_create_vgpu(struct intel_gvt *gvt,
 	if (ret)
 		goto out_clean_shadow_ctx;
 
+	ret = intel_gvt_hypervisor_set_opregion(vgpu);
+	if (ret)
+		goto out_clean_shadow_ctx;
+
 	mutex_unlock(&gvt->lock);
 
 	return vgpu;
Signature · 6 lines
2.7.4

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

PATCH v14 5/7 vfio: ABI for mdev display dma-buf operation

In reply to Tina Zhang

Add VFIO_DEVICE_QUERY_GFX_PLANE ioctl command to let user mode query and get the plan and its related information. This ioctl can be invoked with:
1) either flag DMABUF or REGION is set. Vendor driver returns success and the plane_info only when the specific kind of buffer is supported.
2) flag PROBE is set with either DMABUF or REGION. Vendor driver returns success only when the specific kind of buffer is supported.

The dma-buf's life cycle is handled by user mode and tracked by kernel. The returned fd in struct vfio_device_query_gfx_plane can be a new fd or an old fd of a re-exported dma-buf. Host user mode can check the value of fd and to see if it needs to create new resource according to the new fd or just use the existed resource related to the old fd.

v14:
- add PROBE, DMABUF and REGION flags. (Alex)

v12:
- add drm_format_mod back. (Gerd and Zhenyu)
- add region_index. (Gerd)

v11:
- rename plane_type to drm_plane_type. (Gerd)
- move fields of vfio_device_query_gfx_plane to vfio_device_gfx_plane_info. (Gerd)
- remove drm_format_mod, start fields. (Daniel)
- remove plane_id.

v10:
- refine the ABI API VFIO_DEVICE_QUERY_GFX_PLANE. (Alex) (Gerd)

v3:
- add a field gvt_plane_info in the drm_i915_gem_obj structure to save the decoded plane information to avoid look up while need the plane info. (Gerd)

Cc Gerd Hoffmann <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Daniel Vetter <[email protected]>
include/uapi/linux/vfio.h +43 -0
index ae46105..fccc87f 100644
@@ -502,6 +502,49 @@ struct vfio_pci_hot_reset {
 
 #define VFIO_DEVICE_PCI_HOT_RESET	_IO(VFIO_TYPE, VFIO_BASE + 13)
 
+/**
+ * VFIO_DEVICE_QUERY_GFX_PLANE - _IOW(VFIO_TYPE, VFIO_BASE + 14, struct vfio_device_query_gfx_plane)
+ *
+ * Set the drm_plane_type and flags, then retrieve information about the gfx plane.
+ *
+ * flags:
+ *	VFIO_GFX_PLANE_TYPE_PROBE and VFIO_GFX_PLANE_TYPE_DMABUF are set to ask if the mdev
+ * 	    supports dma-buf. 0 on support, -EINVAL on no support for dma-buf.
+ *      VFIO_GFX_PLANE_TYPE_PROBE and VFIO_GFX_PLANE_TYPE_REGION are set to ask if the mdev
+ *          supports region. 0 on support, -EINVAL on no support for region.
+ *      VFIO_GFX_PLANE_TYPE_DMABUF or VFIO_GFX_PLANE_TYPE_REGION is set with each call to
+ *          query the plane info.
+ *      Others are invalid and return -EINVAL.
+ *
+ * Return: 0 on success, -ENODEV with all out fields zero on mdev device initialization, -errno on
+ * other failure.
+ */
+struct vfio_device_gfx_plane_info {
+	__u32 argsz;
+	__u32 flags;
+#define VFIO_GFX_PLANE_TYPE_PROBE (1 << 0)
+#define VFIO_GFX_PLANE_TYPE_DMABUF (1 << 1)
+#define VFIO_GFX_PLANE_TYPE_REGION (1 << 2)
+	/* in */
+	__u32 drm_plane_type;	/* type of plane: DRM_PLANE_TYPE_* */
+	/* out */
+	__u32 drm_format;	/* drm format of plane */
+	__u64 drm_format_mod;   /* tiled mode */
+	__u32 width;	/* width of plane */
+	__u32 height;	/* height of plane */
+	__u32 stride;	/* stride of plane */
+	__u32 size;	/* size of plane in bytes, align on page*/
+	__u32 x_pos;	/* horizontal position of cursor plane, upper left corner in pixels */
+	__u32 y_pos;	/* vertical position of cursor plane, upper left corner in lines*/
+	union {
+		__u32 region_index;	/* region index */
+		__s32 fd;	/* dma-buf fd */
+	};
+};
+
+#define VFIO_DEVICE_QUERY_GFX_PLANE _IO(VFIO_TYPE, VFIO_BASE + 14)
+
+
 /* -------- API for Type1 VFIO IOMMU -------- */
 
 /**
Signature · 6 lines
2.7.4

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

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 15 lines · Tina Zhang

+/**
+ * VFIO_DEVICE_QUERY_GFX_PLANE - _IOW(VFIO_TYPE, VFIO_BASE + 14, struct vfio_device_query_gfx_plane)
+ *
+ * Set the drm_plane_type and flags, then retrieve information about the gfx plane.
+ *
+ * flags:
+ *     VFIO_GFX_PLANE_TYPE_PROBE and VFIO_GFX_PLANE_TYPE_DMABUF are set to ask if the mdev
+ *         supports dma-buf. 0 on support, -EINVAL on no support for dma-buf.
+ *      VFIO_GFX_PLANE_TYPE_PROBE and VFIO_GFX_PLANE_TYPE_REGION are set to ask if the mdev
+ *          supports region. 0 on support, -EINVAL on no support for

Quoted 8 lines

region.
+ *      VFIO_GFX_PLANE_TYPE_DMABUF or VFIO_GFX_PLANE_TYPE_REGION is set with each call to
+ *          query the plane info.
+ *      Others are invalid and return -EINVAL.
+ *
+ * Return: 0 on success, -ENODEV with all out fields zero on mdev device initialization,

Do you mean PROBE calls here?

The behavior in case the guest driver hasn't yet initialized the plane should be documented here too.

cheers,
Gerd

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

#8 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

In reply to Tina Zhang

Hi,

[ bringing a private discussion back to the list ]

Quoted 8 lines

The dma-buf's life cycle is handled by user mode and tracked by kernel.
The returned fd in struct vfio_device_query_gfx_plane can be a new fd or an old fd of a re-exported dma-buf. Host user mode can check the
value of fd and to see if it needs to create new resource according to
the new fd or just use the existed resource related to the old fd.

Ok, this idea has a fundamental flaw: The life cycle of the dma-buf and the file handle are not identical. The dma-buf can exist longer than the file handle in case other references to the dma-buf exist. So when trying to use the file handle as identifier for the dma-buf you'll end up with all sorts of strange effects.

So, I'd suggest to use a id instead, and add a ioctl to get a dmabuf for a given id (incremental patch):

include/uapi/linux/vfio.h +0 -0
@@ -538,12 +538,22 @@ struct vfio_device_gfx_plane_info {

        __u32 y_pos;    /* vertical position of cursor plane, upper left corner in lines*/
        union {
                __u32 region_index;     /* region index */ -               __s32 fd;       /* dma-buf fd */
+               __u32 dmabuf_id;        /* dma-buf id */ };
 };

 #define VFIO_DEVICE_QUERY_GFX_PLANE _IO(VFIO_TYPE, VFIO_BASE + 14)
+struct vfio_device_gfx_dmabuf_fd {
+       __u32 argsz;
+       __u32 flags;
+       /* in */
+       __u32 dmabuf_id;
+       /* out */
+       __s32 dmabuf_fd;
+};
+
+#define VFIO_DEVICE_GET_GFX_DMABUF _IO(VFIO_TYPE, VFIO_BASE + 15)
 /* -------- API for Type1 VFIO IOMMU -------- */

[ no changes for a region-based display ]

git branch, kernel, with updated dmabuf patch:
https://www.kraxel.org/cgit/linux/log/?h=gvt-dmabuf-v14

qemu branch:
https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu

cheers,
Gerd

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

#9 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

It's really tricky to handle the cached dmabuf_obj's life-cycle without touching other kernel modules (e.g. i915 or dmabuf).The proposed two ioctls will be helpful.

So, there is a problem about the releasing cached dmabuf_obj. We cannot rely on the drm_i915_gem_object_ops.release() to release the cached dmabuf_obj, as this release operation is running in another thread, which leads to a racing condition and tricky to be solved without touching other modules.

So, in order to solve that kind of problem, I’d like to add one more ioctl, which is used for user mode to close the dmabuf_obj.

Including the proposed two ioctls, the incremental patch is like this:

linux-headers/linux/vfio.h +21 -1
index bf40f7b..6aa6860 100644
@@ -538,12 +538,33 @@ struct vfio_device_gfx_plane_info {
        __u32 y_pos;    /* vertical position of cursor plane, upper left corner in lines*/
        union {
                __u32 region_index;     /* region index */
-               __s32 fd;       /* dma-buf fd */
+               __s32 dmabuf_id;        /* dma-buf fd */
        };
 };
 
 #define VFIO_DEVICE_QUERY_GFX_PLANE _IO(VFIO_TYPE, VFIO_BASE + 14)
 
+struct vfio_device_gfx_dmabuf_fd {
+         __u32 argsz;
+         __u32 flags;
+         /* in */
+        __u32 dmabuf_id;
+        /* out */
+        __s32 dmabuf_fd;
+};
+
+#define VFIO_DEVICE_GET_GFX_DMABUF _IO(VFIO_TYPE, VFIO_BASE + 15)
+
+
+struct vfio_device_gfx_buffer {
+        __u32 argsz;
+        __u32 flags;
+        /* in */
+        __u32 id;
+};
+
+#define VFIO_DEVICE_CLOSE_BUF _IO(VFIO_TYPE, VFIO_BASE + 16)
 

And here are some details:

1. VFIO_DEVICE_QUERY_GFX_PLANE: is used for user mode to ask kernel part to create a buffer (in dmabuf case is dmabuf_obj in kernel) and return the buffer info.

2. VFIO_DEVICE_GET_GFX_DMABUF: is used for user mode to ask kernel part which interface it likes the buffer to be wrapped. Actually, I think there could be a bunch of types, including DMBUF type.
So, maybe we can change the IOCTL's name to some generic name and use flags field to distinguish them.

In dmabuf case, a new dmabuf will be created each time with this ioctl invoked, and installed with a new fd. The dmabuf is just a wrapper of kernel dmabuf_obj distinguished by dmabuf_id. So there could be more than one dmabuf as the wrappers of the same dmabuf_obj, if this ioctl was invoked with the same dmabuf_id many times. The dmabuf and its fd is fully controlled by user mode. And the VFIO_DEVICE_CLOSE_BUF can guarantee the referenced dmabuf_obj will be released at last, after all the dmabufs are released.

3. VFIO_DEVICE_CLOSE_BUF: is used for user mode to tell kernel part to release that buffer.
In dmabuf case, this ioctl won't release the dmabuf_obj at once. It just decreases the reference count of the dmabuf_obj and make sure this dmabuf_obj won't be reused through VFIO_DEVICE_QUERY_GFX_PLANE again. At last, after all the referencing dmabuf are released by user mode, the dmabuf_obj will be released by kernel.

Thanks,

BR,
Tina

Quoted 57 lines

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Tuesday, September 26, 2017 3:12 PM
To: Zhang, Tina <[email protected]>; [email protected]
Cc: [email protected]; [email protected]; Alex Williamson <[email protected]>; Daniel Vetter <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

[ bringing a private discussion back to the list ]

Quoted 6 lines

The dma-buf's life cycle is handled by user mode and tracked by kernel.
The returned fd in struct vfio_device_query_gfx_plane can be a new fd or an old fd of a re-exported dma-buf. Host user mode can check the value of fd and to see if it needs to create new resource according to the new fd or just use the existed resource related to the old fd.

Ok, this idea has a fundamental flaw: The life cycle of the dma-buf and the file handle are not identical. The dma-buf can exist longer than the file handle in case other references to the dma-buf exist. So when trying to use the file handle as identifier for the dma-buf you'll end up with all sorts of strange effects.

So, I'd suggest to use a id instead, and add a ioctl to get a dmabuf for a given id (incremental patch):

include/uapi/linux/vfio.h +0 -0
@@ -538,12 +538,22 @@ struct vfio_device_gfx_plane_info {

        __u32 y_pos;    /* vertical position of cursor plane, upper left corner in lines*/
        union {
                __u32 region_index;     /* region index */
-               __s32 fd;       /* dma-buf fd */
+               __u32 dmabuf_id;        /* dma-buf id */
        };
 };

 #define VFIO_DEVICE_QUERY_GFX_PLANE _IO(VFIO_TYPE, VFIO_BASE + 14)

include/uapi/linux/vfio.h +10 -0
+struct vfio_device_gfx_dmabuf_fd {
+       __u32 argsz;
+       __u32 flags;
+       /* in */
+       __u32 dmabuf_id;
+       /* out */
+       __s32 dmabuf_fd;
+};
+
+#define VFIO_DEVICE_GET_GFX_DMABUF _IO(VFIO_TYPE, VFIO_BASE + 15)

 /* -------- API for Type1 VFIO IOMMU -------- */

[ no changes for a region-based display ]

Quoted 8 lines

git branch, kernel, with updated dmabuf patch:
https://www.kraxel.org/cgit/linux/log/?h=gvt-dmabuf-v14

qemu branch:
https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu

cheers,
Gerd

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

#10 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 6 lines

So, there is a problem about the releasing cached dmabuf_obj. We cannot rely on the drm_i915_gem_object_ops.release() to release the cached dmabuf_obj,
as this release operation is running in another thread, which leads to a racing condition and tricky to be solved without touching other modules.

PLANE_INFO just creates a intel_vgpu_dmabuf_obj.

GET_DMABUF creates a fresh proxy gem object and dmabuf.

proxy gem object references intel_vgpu_dmabuf_obj but not the other way around. Then you can simply refcount intel_vgpu_dmabuf_obj and be done with it.

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=350a0e834
971e6f53d7235d8b6167bed4dccf074

Note: Patch renamed intel_vgpu_dmabuf_obj to intel_vgpu_fb_obj, because it doesn't refer to dmabufs any more. It basically carries the guest plane/framebuffer information and the ID associated with it.

Quoted 2 lines

So, in order to solve that kind of problem, I’d like to add one more ioctl, which is used for user mode to close the dmabuf_obj.

Depending on userspace notifying the kernel for that kind of cleanups is a bad idea. What happens in case userspace crashes? Do you leak dmabufs then?

cheers,
Gerd

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

#11 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Thanks for the patch. Actually, I did the same thing in my local repo and also, I have a patch for the local Qemu repo to test it. I will send them out later.

The reason why I want to propose the close IOCTL is because that the current lock (fb_obj_list_lock), cannot sync the intel_vgpu_fb_info releasing and reusing. You see, the intel_vgpu_fb_info reusing and releasing are in different threads. There is a case that intel_vgpu_find_dmabuf can return a intel_vgpu_fb_obj, while the intel_vgpu_fb_obj is on the way to be released. That's the problem.

The invoker of the close IOCTL is only Qemu. So, if the Qemu crashes, the whole vGPU's resource is going to be released. We can handle our dmabuf_obj to be released there.

Thanks.

BR,
Tina

Quoted 40 lines

-----Original Message-----
From: intel-gvt-dev [mailto:[email protected]] On Behalf Of Gerd Hoffmann
Sent: Wednesday, September 27, 2017 6:11 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Alex Williamson <[email protected]>; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 5 lines

So, there is a problem about the releasing cached dmabuf_obj. We cannot rely on the drm_i915_gem_object_ops.release() to release the cached dmabuf_obj, as this release operation is running in another thread, which leads to a racing condition and tricky to be solved without touching other modules.

PLANE_INFO just creates a intel_vgpu_dmabuf_obj.

GET_DMABUF creates a fresh proxy gem object and dmabuf.

proxy gem object references intel_vgpu_dmabuf_obj but not the other way around. Then you can simply refcount intel_vgpu_dmabuf_obj and be done with it.

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=350a0e834
971e6f53d7235d8b6167bed4dccf074

Note: Patch renamed intel_vgpu_dmabuf_obj to intel_vgpu_fb_obj, because it doesn't refer to dmabufs any more. It basically carries the guest plane/framebuffer information and the ID associated with it.

Quoted 2 lines

So, in order to solve that kind of problem, I’d like to add one more ioctl, which is used for user mode to close the dmabuf_obj.

Depending on userspace notifying the kernel for that kind of cleanups is a bad idea. What happens in case userspace crashes? Do you leak dmabufs then?

Quoted 7 lines

cheers,
Gerd

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

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

#12 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 7 lines

The reason why I want to propose the close IOCTL is because that the current lock (fb_obj_list_lock), cannot sync the intel_vgpu_fb_info releasing and reusing.
You see, the intel_vgpu_fb_info reusing and releasing are in different threads. There is a case that intel_vgpu_find_dmabuf can return a intel_vgpu_fb_obj, while the intel_vgpu_fb_obj is on the way to be released. That's the problem.

Oh, right. But that race is fixable. We need to move the locks one level up, so we don't only cover list operations (add/lookup/delete) but also the kref_{get,put} operations for the list elements.

Patch against my tree, only build-tested so far.

cheers,
Gerd

Quoted 1 line

From 3e8c30a857d98d36357e8d9bb04b7ccb72264543 Mon Sep 17 00:00:00 2001

From: Gerd Hoffmann <[email protected]>
Date: Fri, 29 Sep 2017 08:59:34 +0200
Subject: [PATCH] fix locking

---
 drivers/gpu/drm/i915/gvt/dmabuf.c | 26 +++++++++++++-------------
 1 file changed, 13 insertions(+), 13 deletions(-)
drivers/gpu/drm/i915/gvt/dmabuf.c +13 -13
index 2fb3247eff..06ff7bb04e 100644
@@ -84,24 +84,25 @@ static void intel_vgpu_fb_obj_release(struct kref *kref)
 {
 	struct intel_vgpu_fb_obj *fb_obj =
 		container_of(kref, struct intel_vgpu_fb_obj, kref);
-	struct intel_vgpu *vgpu;
 
-	vgpu = fb_obj->vgpu;
-	mutex_lock(&vgpu->fb_obj_list_lock);
 	list_del(&fb_obj->list);
-	mutex_unlock(&vgpu->fb_obj_list_lock);
 	kfree(fb_obj);
 }
 
 static void intel_vgpu_gem_release(struct drm_i915_gem_object *obj)
 {
+	struct intel_vgpu *vgpu;
+
 	if (WARN_ON(!obj->gvt || !obj->gvt->vgpu)) {
 		gvt_err("gvt info is invalid\n");
 		return;
 	}
 
-	intel_gvt_hypervisor_put_vfio_device(obj->gvt->vgpu);
+	vgpu = obj->gvt->vgpu;
+	intel_gvt_hypervisor_put_vfio_device(vgpu);
+	mutex_lock(&vgpu->fb_obj_list_lock);
 	kref_put(&obj->gvt->kref, intel_vgpu_fb_obj_release);
+	mutex_unlock(&vgpu->fb_obj_list_lock);
 	obj->gvt = NULL;
 }
 
@@ -239,7 +240,6 @@ intel_vgpu_pick_exposed_dmabuf(struct intel_vgpu *vgpu,
 	struct list_head *pos;
 	struct intel_vgpu_fb_obj *fb_obj;
 
-	mutex_lock(&vgpu->fb_obj_list_lock);
 	list_for_each(pos, &vgpu->fb_obj_list_head) {
 		fb_obj = container_of(pos, struct intel_vgpu_fb_obj,
 					  list);
@@ -251,11 +251,9 @@ intel_vgpu_pick_exposed_dmabuf(struct intel_vgpu *vgpu,
 		    (fb_obj->fb.width == latest_info->width) &&
 		    (fb_obj->fb.height == latest_info->height) &&
 		    (fb_obj->fb.stride == latest_info->stride)) {
-			mutex_unlock(&vgpu->fb_obj_list_lock);
 			return fb_obj;
 		}
 	}
-	mutex_unlock(&vgpu->fb_obj_list_lock);
 	return NULL;
 }
 
@@ -265,16 +263,13 @@ intel_vgpu_find_dmabuf(struct intel_vgpu *vgpu, u32 dmabuf_id)
 	struct list_head *pos;
 	struct intel_vgpu_fb_obj *fb_obj;
 
-	mutex_lock(&vgpu->fb_obj_list_lock);
 	list_for_each(pos, &vgpu->fb_obj_list_head) {
 		fb_obj = container_of(pos, struct intel_vgpu_fb_obj,
 					  list);
 		if (fb_obj->dmabuf_id == dmabuf_id) {
-			mutex_unlock(&vgpu->fb_obj_list_lock);
 			return fb_obj;
 		}
 	}
-	mutex_unlock(&vgpu->fb_obj_list_lock);
 	return NULL;
 }
 
@@ -327,8 +322,10 @@ int intel_vgpu_query_plane(struct intel_vgpu *vgpu, void *args)
 		return ret;
 
 	/* If exists, pick up the exposed dmabuf fd */
+	mutex_lock(&vgpu->fb_obj_list_lock);
 	fb_obj = intel_vgpu_pick_exposed_dmabuf(vgpu, &fb_info);
 	if (fb_obj != NULL) {
+		mutex_unlock(&vgpu->fb_obj_list_lock);
 		update_fb_info(gvt_dmabuf, fb_obj);
 		return 0;
 	}
@@ -345,7 +342,6 @@ int intel_vgpu_query_plane(struct intel_vgpu *vgpu, void *args)
 	fb_obj->fb = fb_info;
 	fb_obj->dmabuf_id = id++;
 
-	mutex_lock(&vgpu->fb_obj_list_lock);
 	list_add_tail(&fb_obj->list, &vgpu->fb_obj_list_head);
 	mutex_unlock(&vgpu->fb_obj_list_lock);
 	update_fb_info(gvt_dmabuf, fb_obj);
@@ -362,11 +358,15 @@ int intel_vgpu_get_dmabuf(struct intel_vgpu *vgpu, void *args)
 	struct dma_buf *dmabuf;
 	int ret;
 
+	mutex_lock(&vgpu->fb_obj_list_lock);
 	fb_obj = intel_vgpu_find_dmabuf(vgpu, gvt_dmabuf->dmabuf_id);
-	if (NULL == fb_obj)
+	if (NULL == fb_obj) {
+		mutex_unlock(&vgpu->fb_obj_list_lock);
 		return -EINVAL;
+	}
 
 	obj = intel_vgpu_create_gem(dev, fb_obj);
+	mutex_unlock(&vgpu->fb_obj_list_lock);
 	if (obj == NULL) {
 		gvt_vgpu_err("create gvt gem obj failed:%d\n", vgpu->id);
 		return -ENOMEM;
Signature · 1 line
2.9.3

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

#13 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

In reply to Gerd Hoffmann

So, there won't be dmabuf leaking problem, as we release all the dmabuf_obj in the release ops when user space crashing.

Can we just stop considering the way to fix the dmabuf life-cycle issue and try to just consider the generic way to handle buffer exposing? Does the generic way need the close ioctl?
In my opinion, it's like to build up a producer-consumer way to expose the buffer:

Create buffer and return its info
Mdev devices -----------------------------------------------------> User space

                           <----------------------------------------------------- 

(producer) Close it (consumer)

Alex and Gerd, can you share your thoughts?
Thanks.

BR,
Tina

Quoted 68 lines

-----Original Message-----
From: Zhang, Tina
Sent: Friday, September 29, 2017 7:43 AM
To: 'Gerd Hoffmann' <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Alex Williamson <[email protected]>; Lv, Zhiyuan <[email protected]>
Subject: RE: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Thanks for the patch. Actually, I did the same thing in my local repo and also, I have a patch for the local Qemu repo to test it. I will send them out later.

The reason why I want to propose the close IOCTL is because that the current lock (fb_obj_list_lock), cannot sync the intel_vgpu_fb_info releasing and reusing.
You see, the intel_vgpu_fb_info reusing and releasing are in different threads. There is a case that intel_vgpu_find_dmabuf can return a intel_vgpu_fb_obj, while the intel_vgpu_fb_obj is on the way to be released. That's the problem.

The invoker of the close IOCTL is only Qemu. So, if the Qemu crashes, the whole vGPU's resource is going to be released. We can handle our dmabuf_obj to be released there.

Thanks.

BR,
Tina

Quoted 24 lines

-----Original Message-----
From: intel-gvt-dev [mailto:[email protected]] On Behalf Of Gerd Hoffmann
Sent: Wednesday, September 27, 2017 6:11 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Alex Williamson <[email protected]>; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 5 lines

So, there is a problem about the releasing cached dmabuf_obj. We cannot rely on the drm_i915_gem_object_ops.release() to release the cached dmabuf_obj, as this release operation is running in another thread, which leads to a racing condition and tricky to be solved without touching other modules.

PLANE_INFO just creates a intel_vgpu_dmabuf_obj.

Quoted 7 lines

GET_DMABUF creates a fresh proxy gem object and dmabuf.

proxy gem object references intel_vgpu_dmabuf_obj but not the other way around. Then you can simply refcount intel_vgpu_dmabuf_obj and be done with it.

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-

v14&id=350a0e83

Quoted 7 lines

4
971e6f53d7235d8b6167bed4dccf074

Note: Patch renamed intel_vgpu_dmabuf_obj to intel_vgpu_fb_obj, because it doesn't refer to dmabufs any more. It basically carries the guest plane/framebuffer information and the ID associated with it.

Quoted 14 lines
Quoted 5 lines
Quoted 2 lines

So, in order to solve that kind of problem, I’d like to add one more ioctl, which is used for user mode to close the dmabuf_obj.

Depending on userspace notifying the kernel for that kind of cleanups is a bad idea. What happens in case userspace crashes? Do you leak dmabufs

then?

Quoted 8 lines

cheers,
Gerd

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

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

#14 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 6 lines · Tina

So, there won't be dmabuf leaking problem, as we release all the dmabuf_obj in the release ops when user space crashing.

Can we just stop considering the way to fix the dmabuf life-cycle issue and try to just consider the generic way to handle buffer exposing?

Can you describe in more detail what you have in mind?

Quoted 1 line

Does the generic way need the close ioctl?

I think we don't need a close ioctl anyway.

cheers,
Gerd

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

#15 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 20 lines

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, September 29, 2017 3:29 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; Alex Williamson <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 6 lines · Tina

So, there won't be dmabuf leaking problem, as we release all the dmabuf_obj in the release ops when user space crashing.

Can we just stop considering the way to fix the dmabuf life-cycle issue and try to just consider the generic way to handle buffer exposing?

Quoted 5 lines

Can you describe in more detail what you have in mind?

Quoted 1 line

Does the generic way need the close ioctl?

I think we don't need a close ioctl anyway.

Can you share your thoughts?
Do you think the fd interface is enough for all kinds of buffer exposed by Mdev? Thanks.

BR,
Tina

Quoted 4 lines

cheers,
Gerd

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

#16 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Does the generic way need the close ioctl?

I think we don't need a close ioctl anyway.

Can you share your thoughts?

See other mail. I think the race can be fixed by changing the locking, so a explicit close ioctl isn't needed.

Quoted 2 lines

Do you think the fd interface is enough for all kinds of buffer exposed by Mdev?

What kind of buffers do you have in mind which might not be covered?

cheers,
Gerd

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

#17 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 14 lines

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, September 29, 2017 4:03 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; Alex Williamson <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Does the generic way need the close ioctl?

I think we don't need a close ioctl anyway.

Can you share your thoughts?

See other mail. I think the race can be fixed by changing the locking, so a explicit close ioctl isn't needed.

Yeah, I understand your idea. But unfortunately, it cannot solve the current issue.
There will still be a racing condition between releasing dmabuf_obj and reusing it.
For example, if the old reused dmabuf_obj is released just after query ioctl return it, the next get_fd ioctl would return error as the dmabuf_obj has already been closed.

Quoted 5 lines
Quoted 3 lines

Do you think the fd interface is enough for all kinds of buffer exposed by Mdev?

What kind of buffers do you have in mind which might not be covered?

I thinking about the case that would like to postpone the buffers releasing operation, after user space has closed all the fd. Later these buffers may be used to expose to other kinds of fd to user space.
Thanks.

BR,
Tina

Quoted 3 lines

cheers,
Gerd

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

#18 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 3 lines

For example, if the old reused dmabuf_obj is released just after query ioctl return it,  the next get_fd ioctl would return error as the dmabuf_obj has already been closed.

My branch already grabs an extra reference when creating a new dmabuf_obj, which will be dropped on GET_DMABUF ioctl, exactly to avoid the dmabuf_obj disappear between QUERY_PLANE and GET_DMABUF ioctls.

Can easily be extended to handle the reuse case too.

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=9959109ae
52cf15e119715a6b7de080fb849e3d2

While being at it also cleanup properly on close (so we don't leak structs in case userspace never calls GET_DMABUF for a plane).

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=c0b0c407e
33904e749dec1ef44ec01099c16d39f

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

Do you think the fd interface is enough for all kinds of buffer exposed by Mdev?

What kind of buffers do you have in mind which might not be covered?

I thinking about the case that would like to postpone the buffers releasing operation, after user space has closed all the fd.

Work fine. qemu can import the dma-buf as opengl texture, which creates a extra reference. Then close the fd. dma-buf continues to exist as long as the texture referencing it exists.

Quoted 2 lines

Later these buffers may be used to expose to other kinds of fd to user space.

Sorry, I don't understand that sentence.

cheers,
Gerd

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

#19 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 14 lines

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, September 29, 2017 6:21 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; Alex Williamson <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 7 lines
Quoted 3 lines

For example, if the old reused dmabuf_obj is released just after query ioctl return it,  the next get_fd ioctl would return error as the dmabuf_obj has already been closed.

My branch already grabs an extra reference when creating a new dmabuf_obj, which will be dropped on GET_DMABUF ioctl, exactly to avoid the dmabuf_obj disappear between QUERY_PLANE and GET_DMABUF ioctls.

Yeah, that could solve the problem. But I'm not sure if it could be acceptable. Zhenyu, can you share your comments?
Thanks.

BR,
Tina

Quoted 24 lines

Can easily be extended to handle the reuse case too.

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=9959109ae
52cf15e119715a6b7de080fb849e3d2

While being at it also cleanup properly on close (so we don't leak structs in case userspace never calls GET_DMABUF for a plane).

https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=c0b0c407e
33904e749dec1ef44ec01099c16d39f

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Do you think the fd interface is enough for all kinds of buffer exposed by Mdev?

What kind of buffers do you have in mind which might not be covered?

I thinking about the case that would like to postpone the buffers releasing operation, after user space has closed all the fd.

Work fine. qemu can import the dma-buf as opengl texture, which creates a extra reference. Then close the fd. dma-buf continues to exist as long as the texture referencing it exists.

Quoted 7 lines
Quoted 2 lines

Later these buffers may be used to expose to other kinds of fd to user space.

Sorry, I don't understand that sentence.

cheers,
Gerd

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

#20 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Hi,

Quoted 3 lines

Yeah, that could solve the problem. But I'm not sure if it could be acceptable.
Zhenyu, can you share your comments?

Ping? Any progress here? We are at 4.14-rc3 already. v15 is needed really soon now otherwise the 4.15 merge window will be missed.

cheers,
Gerd

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

#21 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 11 lines

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, October 6, 2017 8:13 PM
To: Zhang, Tina <[email protected]>; [email protected]; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; Alex Williamson <[email protected]>
Cc: Daniel Vetter <[email protected]>; [email protected]; [email protected]; Lv, Zhiyuan <[email protected]>
Subject: Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

Quoted 8 lines

Hi,

Quoted 3 lines

Yeah, that could solve the problem. But I'm not sure if it could be acceptable.
Zhenyu, can you share your comments?

Ping? Any progress here? We are at 4.14-rc3 already. v15 is needed really soon now otherwise the 4.15 merge window will be missed.

V15 will be submitted next week, after discussion with maintainers. Thanks.

BR,
Tina

Quoted 3 lines

cheers,
Gerd

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

#22 lore raw

Re: [PATCH v14 5/7] vfio: ABI for mdev display dma-buf operation

In reply to Tina Zhang

Hi,

Quoted 4 lines
(quoted diff) +1 -0
+	__u32 x_pos;	/* horizontal position of cursor plane,

upper left corner in pixels */

(quoted diff) +1 -0
+	__u32 y_pos;	/* vertical position of cursor plane,

upper left corner in lines*/

Completely separate question: Does the host know the cursor hotspot?

cheers,
Gerd

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

author
#23 lore raw

PATCH v14 6/7 drm/i915: Introduce GEM proxy

In reply to Tina Zhang

GEM proxy is a kind of GEM, whose backing physical memory is pinned and produced by guest VM and is used by host as read only. With GEM proxy, host is able to access guest physical memory through GEM object interface. As GEM proxy is such a special kind of GEM, a new flag I915_GEM_OBJECT_IS_PROXY is introduced to ban host from changing the backing storage of GEM proxy.

v14:
- return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)

v13:
- add comments to GEM proxy. (Chris)
- don't ban GEM proxy in i915_gem_sw_finish_ioctl. (Chris)
- check GEM proxy bar after finishing i915_gem_object_wait. (Chris)
- remove GEM proxy bar in i915_gem_madvise_ioctl.

v6:
- add gem proxy barrier in the following ioctls. (Chris) i915_gem_set_caching_ioctl i915_gem_set_domain_ioctl i915_gem_sw_finish_ioctl i915_gem_set_tiling_ioctl i915_gem_madvise_ioctl

Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>
drivers/gpu/drm/i915/i915_gem.c +23 -1
index 000a764..7f1e7ab 100644
@@ -1584,6 +1584,16 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
 	if (err)
 		goto out;
 
+	/* Proxy objects do not control access to the backing storage, ergo
+	 * they cannot be used as a means to manipulate the cache domain
+	 * tracking for that backing storage. The proxy object is always
+	 * considered to be outside of any cache domain.
+	 */
+	if (i915_gem_object_is_proxy(obj)) {
+		err = -ENXIO;
+		goto out;
+	}
+
 	/* Flush and acquire obj->pages so that we are coherent through
 	 * direct access in memory with previous cached writes through
 	 * shmemfs and that our cache domain tracking remains valid.
@@ -1640,6 +1650,10 @@ i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
 	if (!obj)
 		return -ENOENT;
 
+	/* Proxy objects are barred from CPU access, so there is no
+	 * need to ban sw_finish as it is a nop.
+	 */
+
 	/* Pinned buffers may be scanout, so flush the cache */
 	i915_gem_object_flush_if_display(obj);
 	i915_gem_object_put(obj);
@@ -1690,7 +1704,7 @@ i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
 	 */
 	if (!obj->base.filp) {
 		i915_gem_object_put(obj);
-		return -EINVAL;
+		return -ENXIO;
 	}
 
 	addr = vm_mmap(obj->base.filp, 0, args->size,
@@ -3749,6 +3763,14 @@ int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
 	if (!obj)
 		return -ENOENT;
 
+	/* The caching mode of proxy object is handled by its generator, and not
+	 * expected to be changed by user mode.
+	 */
+	if (i915_gem_object_is_proxy(obj)) {
+		ret = -ENXIO;
+		goto out;
+	}
+
 	if (obj->cache_level == level)
 		goto out;
 
drivers/gpu/drm/i915/i915_gem_object.h +7 -0
index 5b19a49..f3b382a 100644
@@ -39,6 +39,7 @@ struct drm_i915_gem_object_ops {
 	unsigned int flags;
 #define I915_GEM_OBJECT_HAS_STRUCT_PAGE BIT(0)
 #define I915_GEM_OBJECT_IS_SHRINKABLE   BIT(1)
+#define I915_GEM_OBJECT_IS_PROXY	BIT(2)
 
 	/* Interface between the GEM object and its backing storage.
 	 * get_pages() is called once prior to the use of the associated set
@@ -300,6 +301,12 @@ i915_gem_object_is_shrinkable(const struct drm_i915_gem_object *obj)
 }
 
 static inline bool
+i915_gem_object_is_proxy(const struct drm_i915_gem_object *obj)
+{
+	return obj->ops->flags & I915_GEM_OBJECT_IS_PROXY;
+}
+
+static inline bool
 i915_gem_object_is_active(const struct drm_i915_gem_object *obj)
 {
 	return obj->active_count;
drivers/gpu/drm/i915/i915_gem_tiling.c +8 -0
index fb5231f..f617012 100644
@@ -345,6 +345,14 @@ i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
 	if (!obj)
 		return -ENOENT;
 
+	/* The tiling mode of proxy objects is handled by its generator, and not
+	 * expected to be changed by user mode.
+	 */
+	if (i915_gem_object_is_proxy(obj)) {
+		err = -ENXIO;
+		goto err;
+	}
+
 	if (!i915_tiling_ok(obj, args->tiling_mode, args->stride)) {
 		err = -EINVAL;
 		goto err;
Signature · 6 lines
2.7.4

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer Reviewed-by on v14 6/7
#24 lore raw

Re: [PATCH v14 6/7] drm/i915: Introduce GEM proxy

Quoted 25 lines · Tina Zhang

GEM proxy is a kind of GEM, whose backing physical memory is pinned and produced by guest VM and is used by host as read only. With GEM proxy, host is able to access guest physical memory through GEM object interface. As GEM proxy is such a special kind of GEM, a new flag I915_GEM_OBJECT_IS_PROXY is introduced to ban host from changing the backing storage of GEM proxy.

v14:
- return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)

v13:
- add comments to GEM proxy. (Chris)
- don't ban GEM proxy in i915_gem_sw_finish_ioctl. (Chris)
- check GEM proxy bar after finishing i915_gem_object_wait. (Chris)
- remove GEM proxy bar in i915_gem_madvise_ioctl.

v6:
- add gem proxy barrier in the following ioctls. (Chris) i915_gem_set_caching_ioctl i915_gem_set_domain_ioctl i915_gem_sw_finish_ioctl i915_gem_set_tiling_ioctl i915_gem_madvise_ioctl

Quoted 4 lines
Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>

This is now as agreed.

Regards, Joonas

PS. I think you have --no-stat option turned on in git format-patch. Please turn it back off, it'll make it far easier for anybody reading the patches to get an overview of which files the patch touches.

Signature · 7 lines
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
author
#25 lore raw

PATCH v14 7/7 drm/i915/gvt: Dmabuf support for GVT-g

In reply to Tina Zhang

This patch introduces a guest's framebuffer sharing mechanism based on dma-buf subsystem. With this sharing mechanism, guest's framebuffer can be shared between guest VM and host.

v14:
- add PROBE, DMABUF and REGION flags. (Alex)

v12:
- refine the lifecycle of dmabuf.

v9:
- remove dma-buf management. (Alex)
- track the dma-buf create and release in kernel mode. (Gerd) (Daniel)

v8:
- refine the dma-buf ioctl definition.(Alex)
- add a lock to protect the dmabuf list. (Alex)

v7:
- release dma-buf related allocations in dma-buf's associated release function. (Alex)
- refine ioctl interface for querying plane info or create dma-buf. (Alex)

v6:
- align the dma-buf life cycle with the vfio device. (Alex)
- add the dma-buf related operations in a separate patch. (Gerd)
- i915 related changes. (Chris)

v5:
- fix bug while checking whether the gem obj is gvt's dma-buf when user change caching mode or domains. Add a helper function to do it. (Xiaoguang)
- add definition for the query plane and create dma-buf. (Xiaoguang)

v4:
- fix bug while checking whether the gem obj is gvt's dma-buf when set caching mode or doamins. (Xiaoguang)

v3:
- declare a new flag I915_GEM_OBJECT_IS_GVT_DMABUF in drm_i915_gem_object to represent the gem obj for gvt's dma-buf. The tiling mode, caching mode and domains can not be changed for this kind of gem object. (Alex)
- change dma-buf related information to be more generic. So other vendor can use the same interface. (Alex)

v2:
- create a management fd for dma-buf operations. (Alex)
- alloc gem object's backing storage in gem obj's get_pages() callback. (Chris)

Cc Alex Williamson <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Gerd Hoffmann <[email protected]>
drivers/gpu/drm/i915/gvt/Makefile +1 -1
index 019d596..18f43cb 100644
@@ -2,7 +2,7 @@ GVT_DIR := gvt
 GVT_SOURCE := gvt.o aperture_gm.o handlers.o vgpu.o trace_points.o firmware.o \
 	interrupt.o gtt.o cfg_space.o opregion.o mmio.o display.o edid.o \
 	execlist.o scheduler.o sched_policy.o render.o cmd_parser.o \
-	fb_decoder.o
+	fb_decoder.o dmabuf.o
 
 ccflags-y				+= -I$(src) -I$(src)/$(GVT_DIR)
 i915-y					+= $(addprefix $(GVT_DIR)/, $(GVT_SOURCE))
drivers/gpu/drm/i915/gvt/dmabuf.c new file +387 -0

387 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/dmabuf.h new file +58 -0

58 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/gvt.c +1 -0
index c27c683..f77770b 100644
@@ -54,6 +54,7 @@ static const struct intel_gvt_ops intel_gvt_ops = {
 	.vgpu_reset = intel_gvt_reset_vgpu,
 	.vgpu_activate = intel_gvt_activate_vgpu,
 	.vgpu_deactivate = intel_gvt_deactivate_vgpu,
+	.vgpu_query_plane = intel_vgpu_query_plane,
 };
 
 /**
drivers/gpu/drm/i915/gvt/gvt.h +9 -0
index 0f562ee..0f313ee 100644
@@ -47,6 +47,7 @@
 #include "render.h"
 #include "cmd_parser.h"
 #include "fb_decoder.h"
+#include "dmabuf.h"
 
 #define GVT_MAX_VGPU 8
 
@@ -183,8 +184,15 @@ struct intel_vgpu {
 		struct kvm *kvm;
 		struct work_struct release_work;
 		atomic_t released;
+		struct vfio_device *vfio_device;
 	} vdev;
 #endif
+
+	struct list_head dmabuf_obj_list_head;
+	struct mutex dmabuf_list_lock;
+
+	struct completion vblank_done;
+
 };
 
 struct intel_gvt_gm {
@@ -500,6 +508,7 @@ struct intel_gvt_ops {
 	void (*vgpu_reset)(struct intel_vgpu *);
 	void (*vgpu_activate)(struct intel_vgpu *);
 	void (*vgpu_deactivate)(struct intel_vgpu *);
+	int (*vgpu_query_plane)(struct intel_vgpu *vgpu, void *);
 };
 
 
drivers/gpu/drm/i915/gvt/hypercall.h +2 -0
index 32c345c..a1bd82f 100644
@@ -56,6 +56,8 @@ struct intel_gvt_mpt {
 	int (*set_trap_area)(unsigned long handle, u64 start, u64 end,
 			     bool map);
 	int (*set_opregion)(void *vgpu);
+	int (*get_vfio_device)(void *vgpu);
+	void (*put_vfio_device)(void *vgpu);
 };
 
 extern struct intel_gvt_mpt xengt_mpt;
drivers/gpu/drm/i915/gvt/kvmgt.c +42 -0

76 lines not on this page · raw message

drivers/gpu/drm/i915/gvt/mpt.h +30 -0
index 9e73f2e..86443ff 100644
@@ -307,4 +307,34 @@ static inline int intel_gvt_hypervisor_set_opregion(struct intel_vgpu *vgpu)
 	return intel_gvt_host.mpt->set_opregion(vgpu);
 }
 
+/**
+ * intel_gvt_hypervisor_get_vfio_device - increase vfio device ref count
+ * @vgpu: a vGPU
+ *
+ * Returns:
+ * Zero on success, negative error code if failed.
+ */
+static inline int intel_gvt_hypervisor_get_vfio_device(struct intel_vgpu *vgpu)
+{
+	if (!intel_gvt_host.mpt->get_vfio_device)
+		return 0;
+
+	return intel_gvt_host.mpt->get_vfio_device(vgpu);
+}
+
+/**
+ * intel_gvt_hypervisor_put_vfio_device - decrease vfio device ref count
+ * @vgpu: a vGPU
+ *
+ * Returns:
+ * Zero on success, negative error code if failed.
+ */
+static inline void intel_gvt_hypervisor_put_vfio_device(struct intel_vgpu *vgpu)
+{
+	if (!intel_gvt_host.mpt->put_vfio_device)
+		return;
+
+	intel_gvt_host.mpt->put_vfio_device(vgpu);
+}
+
 #endif /* _GVT_MPT_H_ */
drivers/gpu/drm/i915/gvt/vgpu.c +1 -1
index bcbf535..19cbbbb 100644
@@ -346,7 +346,7 @@ static struct intel_vgpu *__intel_gvt_create_vgpu(struct intel_gvt *gvt,
 	vgpu->gvt = gvt;
 	vgpu->sched_ctl.weight = param->weight;
 	bitmap_zero(vgpu->tlb_handle_pending, I915_NUM_ENGINES);
-
+	INIT_LIST_HEAD(&vgpu->dmabuf_obj_list_head);
 	intel_vgpu_init_cfg_space(vgpu, param->primary);
 
 	ret = intel_vgpu_init_mmio(vgpu);
drivers/gpu/drm/i915/i915_gem_object.h +2 -0
index f3b382a..c91e336 100644
@@ -199,6 +199,8 @@ struct drm_i915_gem_object {
 		} userptr;
 
 		unsigned long scratch;
+
+		void *gvt_info;
 	};
 
 	/** for phys allocated objects */
Signature · 6 lines
2.7.4

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

✓ Fi.CI.BAT: success for drm/i915/gvt: Dma-buf support for GVT-g

In reply to Tina Zhang
== Series Details ==

== Series Details ==

Series: drm/i915/gvt: Dma-buf support for GVT-g
URL : https://patchwork.freedesktop.org/series/28980/
State : success

== Summary ==

Series 28980v1 drm/i915/gvt: Dma-buf support for GVT-g
https://patchwork.freedesktop.org/api/1.0/series/28980/revisions/1/mbox/

Test gem_exec_flush:

        Subgroup basic-batch-kernel-default-uc:
                pass       -> FAIL       (fi-snb-2600) fdo#100007
Test gem_ringfill:
        Subgroup basic-default:
                pass       -> SKIP       (fi-bsw-n3050) fdo#101915

fdo#100007 https://bugs.freedesktop.org/show_bug.cgi?id=100007
fdo#101915 https://bugs.freedesktop.org/show_bug.cgi?id=101915

fi-bdw-5557u     total:279  pass:268  dwarn:0   dfail:0   fail:0   skip:11  time:451s
fi-bdw-gvtdvm    total:279  pass:265  dwarn:0   dfail:0   fail:0   skip:14  time:441s
fi-blb-e6850     total:279  pass:224  dwarn:1   dfail:0   fail:0   skip:54  time:362s
fi-bsw-n3050     total:279  pass:242  dwarn:0   dfail:0   fail:0   skip:37  time:560s
fi-bxt-j4205     total:279  pass:260  dwarn:0   dfail:0   fail:0   skip:19  time:521s
fi-byt-j1900     total:279  pass:254  dwarn:1   dfail:0   fail:0   skip:24  time:528s
fi-byt-n2820     total:279  pass:250  dwarn:1   dfail:0   fail:0   skip:28  time:511s
fi-glk-2a        total:279  pass:260  dwarn:0   dfail:0   fail:0   skip:19  time:609s
fi-hsw-4770      total:279  pass:263  dwarn:0   dfail:0   fail:0   skip:16  time:443s
fi-hsw-4770r     total:279  pass:263  dwarn:0   dfail:0   fail:0   skip:16  time:423s
fi-ilk-650       total:279  pass:229  dwarn:0   dfail:0   fail:0   skip:50  time:425s
fi-ivb-3520m     total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:510s
fi-ivb-3770      total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:478s
fi-kbl-7500u     total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:476s
fi-kbl-7560u     total:279  pass:269  dwarn:0   dfail:0   fail:0   skip:10  time:596s
fi-kbl-r         total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:597s
fi-pnv-d510      total:279  pass:223  dwarn:1   dfail:0   fail:0   skip:55  time:533s
fi-skl-6260u     total:279  pass:269  dwarn:0   dfail:0   fail:0   skip:10  time:473s
fi-skl-6700k     total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:483s
fi-skl-6770hq    total:279  pass:269  dwarn:0   dfail:0   fail:0   skip:10  time:488s
fi-skl-gvtdvm    total:279  pass:266  dwarn:0   dfail:0   fail:0   skip:13  time:446s
fi-skl-x1585l    total:279  pass:268  dwarn:0   dfail:0   fail:0   skip:11  time:484s
fi-snb-2520m     total:279  pass:251  dwarn:0   dfail:0   fail:0   skip:28  time:546s
fi-snb-2600      total:279  pass:249  dwarn:0   dfail:0   fail:1   skip:29  time:404s

8183095ca4abce6ec2ad43c1e36e877792c140f1 drm-tip: 2017y-08m-18d-09h-08m-28s UTC integration manifest 06591ef22442 drm/i915/gvt: Dmabuf support for GVT-g
69d9112b7ab5 drm/i915: Introduce GEM proxy
d905c36647f8 vfio: ABI for mdev display dma-buf operation
bd594b6c0125 drm/i915/gvt: Add opregion support
83284f2a7ef6 drm/i915/gvt: Add RGB 64-bit 16:16:16:16 float format
09d19af42ee3 drm: Introduce RGB 64-bit 16:16:16:16 float format
3fc810c1addb drm/i915/gvt: Add framebuffer decoder support

== Logs ==

For more details see: https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_5437/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#27 lore raw
In reply to Tina Zhang
Quoted 7 lines · Tina Zhang

v13->v14:
1) add PROBE, DMABUF and REGION flags. (Alex)
2) return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)
3) add some details about the float pixel format. (Daniel)
4) add F suffix to the defined name. (Daniel)
5) refine pixel format table. (Zhenyu)

Ok, patch series applies cleanly to 4.13-rc6. The new flags are working fine.

However, I see VFIO_DEVICE_QUERY_GFX_PLANE failures which I think should not be there. When the guest didn't define a plane yet I get "No such device" errors instead of a plane_info struct with fields (drm_format, width, height, size) set to zero. I also see "Bad address" errors now and then with no obvious cause.

Cursor support isn't working too.

I'm testing with "-display egl-headless -spice gl=off,port=...". In that configuration qemu will import the dma-bufs as textures and reads them using ReadPixels for display.

qemu branch: https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu

The qemu branch has support for both dmabufs and regions. The region- based display code is totaly untested though.

cheers,
Gerd

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

#28 lore raw

Hi Gerd:
Thanks for the testing. We will find out the problem and refresh the whole patch series.

Thanks,
Zhi.

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, August 25, 2017 2:47 PM
To: Zhang, Tina <[email protected]>; [email protected]; Lv, Zhiyuan <[email protected]>; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]
Subject: Re: [PATCH v14 0/7] drm/i915/gvt: Dma-buf support for GVT-g

Quoted 7 lines · Tina Zhang

v13->v14:
1) add PROBE, DMABUF and REGION flags. (Alex)
2) return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)
3) add some details about the float pixel format. (Daniel)
4) add F suffix to the defined name. (Daniel)
5) refine pixel format table. (Zhenyu)

Ok, patch series applies cleanly to 4.13-rc6. The new flags are working fine.

However, I see VFIO_DEVICE_QUERY_GFX_PLANE failures which I think should not be there. When the guest didn't define a plane yet I get "No such device" errors instead of a plane_info struct with fields (drm_format, width, height, size) set to zero. I also see "Bad address" errors now and then with no obvious cause.

Cursor support isn't working too.

I'm testing with "-display egl-headless -spice gl=off,port=...". In that configuration qemu will import the dma-bufs as textures and reads them using ReadPixels for display.

qemu branch: https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu

The qemu branch has support for both dmabufs and regions. The region- based display code is totaly untested though.

cheers,
Gerd

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

#29 lore raw

Thanks Zhi and Gerd for testing v14 patch-set.

Quoted 38 lines

-----Original Message-----
From: Wang, Zhi A
Sent: Friday, August 25, 2017 8:53 PM
To: Gerd Hoffmann <[email protected]>; Zhang, Tina <[email protected]>; [email protected]; Lv, Zhiyuan <[email protected]>; Tian, Kevin <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]
Subject: RE: [PATCH v14 0/7] drm/i915/gvt: Dma-buf support for GVT-g

Hi Gerd:
Thanks for the testing. We will find out the problem and refresh the whole
patch series.

Thanks,
Zhi.

-----Original Message-----
From: Gerd Hoffmann [mailto:[email protected]]
Sent: Friday, August 25, 2017 2:47 PM
To: Zhang, Tina <[email protected]>; [email protected]; Lv, Zhiyuan <[email protected]>; Wang, Zhi A <[email protected]>; Tian, Kevin <[email protected]>; [email protected]; chris@chris- wilson.co.uk; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]
Subject: Re: [PATCH v14 0/7] drm/i915/gvt: Dma-buf support for GVT-g

Quoted 7 lines · Tina Zhang

v13->v14:
1) add PROBE, DMABUF and REGION flags. (Alex)
2) return -ENXIO when gem proxy object is banned by ioctl. (Chris) (Daniel)
3) add some details about the float pixel format. (Daniel)
4) add F suffix to the defined name. (Daniel)
5) refine pixel format table. (Zhenyu)

Quoted 6 lines

Ok, patch series applies cleanly to 4.13-rc6. The new flags are working fine.

However, I see VFIO_DEVICE_QUERY_GFX_PLANE failures which I think should not be there. When the guest didn't define a plane yet I get "No such device" errors instead of a plane_info struct with fields (drm_format, width, height, size) set to zero. I also see "Bad address" errors now and then with no obvious cause.

The idea is to return "No such device" error with fields set to zero.
There are two cases, in which the "No such device" error is returned: one is the guest IGD driver has not finished the initialization and the other is the plane is disabled by guest IGD driver. If we prefer to return success in these two situations with fields set to zero, I can add it in the next version.

We didn't meet the "Bad address" errors before. I will try your qemu repo to see whether I can meet the issue. Thanks.

Quoted 1 line

Cursor support isn't working too.

It seems there is some issue in i915 of version 4.13-rc6, which blocks the cursor on native platform. I just tried the rc7 (the latest staging), there is on such issue.
Thanks.

Tina

Quoted 5 lines

I'm testing with "-display egl-headless -spice gl=off,port=...". In that configuration qemu will import the dma-bufs as textures and reads them using ReadPixels for display.

Quoted 7 lines

qemu branch: https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu

The qemu branch has support for both dmabufs and regions. The region- based display code is totaly untested though.

cheers,
Gerd

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

#30 lore raw

Hi,

Quoted 9 lines
Quoted 8 lines

However, I see VFIO_DEVICE_QUERY_GFX_PLANE failures which I think should
not be there.  When the guest didn't define a plane yet I get "No such device"
errors instead of a plane_info struct with fields (drm_format, width, height, size)
set to zero.  I also see "Bad address" errors now and then with no obvious cause.

Quoted 8 lines

The idea is to return "No such device" error with fields set to zero. There are two cases, in which the "No such device" error is returned:
one is the guest IGD driver
has not finished the initialization and the other is the plane is disabled by guest IGD driver.
If we prefer to return success in these two situations with fields set to zero, I can add it in the
next version.

Yes, success should be returned in those cases. Querying the plane information worked and the struct is filled by the driver. Userspace can figure that there is no plane defined by the guest by looking at the struct fields.

Quoted 6 lines
Quoted 1 line

Cursor support isn't working too.

It seems there is some issue in i915 of version 4.13-rc6, which blocks the cursor on native platform.
I just tried the rc7 (the latest staging), there is on such issue. Thanks.

OK, I'll go re-test with the just-released 4.13 kernel then.

thanks,
Gerd

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