Skip to contentSkip to conversation

drm/hibmc: Fix plane helpers and convert to gem-shmem

PATCH Superseded Thomas Zimmermann 2 participants 9 messages Jun 1 – Jun 11

Subsystems (4)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Thomas Zimmermann

author maintainer
#0 lore raw

v3 cover drm/hibmc: Fix plane helpers and convert to gem-shmem

Convert hibmc's memory management from gem-vram ot gem-shmem. Fix a number of bugs in preparation.

Patches 1 to 3 fix various bugs in the plane code.

Patch 4 replaces gem-vram with gem-shmem. The rational is that gem-vram has corner cases on low-end devices and is limited in functionality. The commit messages state a list of possible issues.

Switching to gem-shmem avoids these problems. It also enables buffer sharing via dma_buf, which improves the hibmc driver's integration with the overall graphics ecosystem.

We've done these changes on drivers for similar hardware, such as ast and mgag200. Hibmc is the last of these drivers to run on gem-shmem. Gem-vram instead is deprecated and can hopefully be removed in a later release.

Tested on an RH1288 v3 system.

v3:
- fix coding style
- rebase on display-detection fixes
v2:
- extend commit message of patch 2 (Yongbang)
- so not select TTM any longer

Thomas Zimmermann (4):
drm/hibmc: Use drm_atomic_helper_check_plane_state() drm/hibmc: Fix list of formats on the primary plane drm/hibmc: Do not use cpp from struct drm_format_info drm/hibmc: Use gem-shmem with shadow-plane helpers for memory
management

 drivers/gpu/drm/drm_gem_shmem_helper.c        |  22 +++-
 drivers/gpu/drm/hisilicon/hibmc/Kconfig       |   4 +-
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    | 110 +++++++++---------
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c   |  62 +++++++---
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h   |   5 +
 include/drm/drm_gem_shmem_helper.h            |   4 +
 6 files changed, 129 insertions(+), 78 deletions(-)

base-commit: 4f554688dffcacf48630c14f9fb77a9f60394c1c

Signature · 1 line
2.54.0
author maintainer
#1 lore raw

PATCH v3 1/4 drm/hibmc: Use drm_atomic_helper_check_plane_state()

Call drm_atomic_helper_check_plane_state() from the primary plane's atomic-check helper and replace the custom implementation.

All plane's implementations of atomic_check should call the shared _check_plane_state() helper first. It adjusts the plane state for correct positioning, rotation and scaling of the plane. Do this even if the plane's CRTC has been disabled by setting the parameter can_update_disabled. The original code returned early in this case, but it's safe to so and cleaner to have all plane state initialized.

As we don't set can_position, drm_atomic_helper_check_plane_state()'s visibility check tests if the plane covers all of the CRTC. This is a small change from the original code, which tested if the plane is exactly the size of the CRTC. With the new test, the plane still has to cover all of the CRTC, but can be larger than the CRTC's size. A later patch can fully implement this feature in hibmc.

If the plane is disabled, the helper clears the visibility flag in the plane state. On errors or if the plane is not visible, the atomic-check helper can return early. Implement all this in hibmc and drop the custom code that does some of it.

v2:
- extend the commit description (Yongbang)

Fixes da52605eea8f ("drm/hisilicon/hibmc: Add support for display engine")
Cc Rongrong Zou <[email protected]>
Cc Sean Paul <[email protected]>
Cc Xinliang Liu <[email protected]>
Cc Dmitry Baryshkov <[email protected]>
Cc Baihan Li <[email protected]>
Cc Yongbang Shi <[email protected]>
Cc <[email protected]> # v4.10+
---
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    | 46 ++++++-------------
 1 file changed, 14 insertions(+), 32 deletions(-)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +14 -32
index 3066dc9ebc64..7c0b88c774b5 100644
@@ -72,46 +72,28 @@ static int hibmc_get_best_clock_idx(const struct drm_display_mode *mode)
 static int hibmc_plane_atomic_check(struct drm_plane *plane,
 				    struct drm_atomic_commit *state)
 {
-	struct drm_plane_state *new_plane_state = drm_atomic_get_new_plane_state(state,
-										 plane);
-	struct drm_framebuffer *fb = new_plane_state->fb;
-	struct drm_crtc *crtc = new_plane_state->crtc;
-	struct drm_crtc_state *crtc_state;
-	u32 src_w = new_plane_state->src_w >> 16;
-	u32 src_h = new_plane_state->src_h >> 16;
-
-	if (!crtc || !fb)
-		return 0;
+	struct drm_plane_state *new_plane_state =
+		drm_atomic_get_new_plane_state(state, plane);
+	struct drm_crtc_state *new_crtc_state = NULL;
+	int ret;
 
-	crtc_state = drm_atomic_get_crtc_state(state, crtc);
-	if (IS_ERR(crtc_state))
-		return PTR_ERR(crtc_state);
+	if (new_plane_state->crtc)
+		new_crtc_state = drm_atomic_get_new_crtc_state(state, new_plane_state->crtc);
 
-	if (src_w != new_plane_state->crtc_w || src_h != new_plane_state->crtc_h) {
-		drm_dbg_atomic(plane->dev, "scale not support\n");
-		return -EINVAL;
-	}
-
-	if (new_plane_state->crtc_x < 0 || new_plane_state->crtc_y < 0) {
-		drm_dbg_atomic(plane->dev, "crtc_x/y of drm_plane state is invalid\n");
-		return -EINVAL;
-	}
-
-	if (!crtc_state->enable)
+	ret = drm_atomic_helper_check_plane_state(new_plane_state, new_crtc_state,
+						  DRM_PLANE_NO_SCALING,
+						  DRM_PLANE_NO_SCALING,
+						  false, true);
+	if (ret)
+		return ret;
+	else if (!new_plane_state->visible)
 		return 0;
 
-	if (new_plane_state->crtc_x + new_plane_state->crtc_w >
-	    crtc_state->adjusted_mode.hdisplay ||
-	    new_plane_state->crtc_y + new_plane_state->crtc_h >
-	    crtc_state->adjusted_mode.vdisplay) {
-		drm_dbg_atomic(plane->dev, "visible portion of plane is invalid\n");
-		return -EINVAL;
-	}
-
 	if (new_plane_state->fb->pitches[0] % 128 != 0) {
 		drm_dbg_atomic(plane->dev, "wrong stride with 128-byte aligned\n");
 		return -EINVAL;
 	}
+
 	return 0;
 }
 
Signature · 1 line
2.54.0
author maintainer
#2 lore raw

PATCH v3 2/4 drm/hibmc: Fix list of formats on the primary plane

In reply to Thomas Zimmermann

Remove all formats from the primary plane that are unsupported for various reasons.

* Formats with alpha channel: planes should not announce alpha channels unless they support transparency. There's no transparency support in the primary plane's implementation.

* Formats with BGR order. The common format is in RGB channel order. There's no BGR support in the primary plane's implementation.

* RGB888: atomic_update programs the format from cpp[0] * 8 / 16. For RGB888's cpp value of 3 this returns 1.5; rounded to 1. Programming the value of 1 to HIBMC_CRT_DISP_CTL_FORMAT sets up RGB565. Hence, the output is distorted. This can be tested by booting with video=1024x768-24.

Removing all unsupported formats leaves XRGB8888 and RGB565. Both of which are supported and work correctly.

Fixes da52605eea8f ("drm/hisilicon/hibmc: Add support for display engine")
Cc Rongrong Zou <[email protected]>
Cc Sean Paul <[email protected]>
Cc Xinliang Liu <[email protected]>
Cc Dmitry Baryshkov <[email protected]>
Cc Yongbang Shi <[email protected]>
Cc Baihan Li <[email protected]>
Cc <[email protected]> # v4.10+
---
 drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +2 -4
index 7c0b88c774b5..2e6e189bec1a 100644
@@ -135,10 +135,8 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 }
 
 static const u32 channel_formats1[] = {
-	DRM_FORMAT_RGB565, DRM_FORMAT_BGR565, DRM_FORMAT_RGB888,
-	DRM_FORMAT_BGR888, DRM_FORMAT_XRGB8888, DRM_FORMAT_XBGR8888,
-	DRM_FORMAT_RGBA8888, DRM_FORMAT_BGRA8888, DRM_FORMAT_ARGB8888,
-	DRM_FORMAT_ABGR8888
+	DRM_FORMAT_XRGB8888,
+	DRM_FORMAT_RGB565,
 };
 
 static const struct drm_plane_funcs hibmc_plane_funcs = {
Signature · 1 line
2.54.0
author maintainer
#3 lore raw

PATCH v3 3/4 drm/hibmc: Do not use cpp from struct drm_format_info

In reply to Thomas Zimmermann

Replace uses of struct drm_format_info's cpp with appropriate interfaces. The cpp field contains the characters per pixel. It is deprecated and should be avoided.

Calculate the line width in bytes with drm_format_info_min_pitch(). This is the preferred way of getting pixel and line sizes.

Program HIB_CRT_DISP_CTL_FORMAT from the format's 4CC code instead of calculating the field's value from the cpp.

---
 drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +11 -5
index 2e6e189bec1a..79c33c778d2c 100644
@@ -100,8 +100,8 @@ static int hibmc_plane_atomic_check(struct drm_plane *plane,
 static void hibmc_plane_atomic_update(struct drm_plane *plane,
 				      struct drm_atomic_commit *state)
 {
-	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state,
-									   plane);
+	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state, plane);
+	struct drm_framebuffer *fb = new_state->fb;
 	u32 reg;
 	s64 gpu_addr = 0;
 	u32 line_l;
@@ -119,7 +119,7 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 
 	writel(gpu_addr, priv->mmio + HIBMC_CRT_FB_ADDRESS);
 
-	reg = new_state->fb->width * (new_state->fb->format->cpp[0]);
+	reg = drm_format_info_min_pitch(fb->format, 0, fb->width);
 
 	line_l = new_state->fb->pitches[0];
 	writel(HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_WIDTH, reg) |
@@ -129,8 +129,14 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 	/* SET PIXEL FORMAT */
 	reg = readl(priv->mmio + HIBMC_CRT_DISP_CTL);
 	reg &= ~HIBMC_CRT_DISP_CTL_FORMAT_MASK;
-	reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT,
-			   new_state->fb->format->cpp[0] * 8 / 16);
+	switch (fb->format->format) {
+	case DRM_FORMAT_XRGB8888:
+		reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT, 2);
+		break;
+	case DRM_FORMAT_RGB565:
+		reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT, 1);
+		break;
+	}
 	writel(reg, priv->mmio + HIBMC_CRT_DISP_CTL);
 }
 
Signature · 1 line
2.54.0
maintainer
#4 lore raw

Re: [PATCH v3 3/4] drm/hibmc: Do not use cpp from struct drm_format_info

Hi, Thomas.

Quoted 31 lines

Replace uses of struct drm_format_info's cpp with appropriate interfaces. The cpp field contains the characters per pixel. It is deprecated and should be avoided.

Calculate the line width in bytes with drm_format_info_min_pitch(). This is the preferred way of getting pixel and line sizes.

Program HIB_CRT_DISP_CTL_FORMAT from the format's 4CC code instead of calculating the field's value from the cpp.

---
 drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +2 -2
index 2e6e189bec1a..79c33c778d2c 100644
@@ -100,8 +100,8 @@ static int hibmc_plane_atomic_check(struct drm_plane *plane,
 static void hibmc_plane_atomic_update(struct drm_plane *plane,
 				      struct drm_atomic_commit *state)
 {
-	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state,
-									   plane);
+	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state, plane);
+	struct drm_framebuffer *fb = new_state->fb;
 	u32 reg;
 	s64 gpu_addr = 0;
 	u32 line_l;

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c static void hibmc_plane_atomic_update(struct drm_plane *plane, · line 124

Quoted 7 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +1 -1
@@ -119,7 +119,7 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 
 	writel(gpu_addr, priv->mmio + HIBMC_CRT_FB_ADDRESS);
 
-	reg = new_state->fb->width * (new_state->fb->format->cpp[0]);
+	reg = drm_format_info_min_pitch(fb->format, 0, fb->width);
 
 	line_l = new_state->fb->pitches[0];

Since the variable fb was defined earlier, it can be used directly here without needing new_state->fb.

This function uses new_state->fb in two other places:
https://gitlab.freedesktop.org/drm/misc/kernel/-/blob/drm-misc-next/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c?ref_type=heads#L129
https://gitlab.freedesktop.org/drm/misc/kernel/-/blob/drm-misc-next/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c?ref_type=heads#L132

Thanks,
Yongbang.

Quoted 18 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +8 -2
 	writel(HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_WIDTH, reg) |
@@ -129,8 +129,14 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 	/* SET PIXEL FORMAT */
 	reg = readl(priv->mmio + HIBMC_CRT_DISP_CTL);
 	reg &= ~HIBMC_CRT_DISP_CTL_FORMAT_MASK;
-	reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT,
-			   new_state->fb->format->cpp[0] * 8 / 16);
+	switch (fb->format->format) {
+	case DRM_FORMAT_XRGB8888:
+		reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT, 2);
+		break;
+	case DRM_FORMAT_RGB565:
+		reg |= HIBMC_FIELD(HIBMC_CRT_DISP_CTL_FORMAT, 1);
+		break;
+	}
 	writel(reg, priv->mmio + HIBMC_CRT_DISP_CTL);
 }
 
author maintainer
#5 lore raw

PATCH v3 4/4 drm/hibmc: Use gem-shmem with shadow-plane helpers for memory management

In reply to Thomas Zimmermann

Replace the gem-vram memory manager with gem-shmem. Makes the driver more robust and enables dma-buf sharing with other hardware.

Gem-vram was created from various drivers that used TTM for their memory management. All these drivers have meanwhile been converted to gem-shmem. Using gem-vram is deprecated because it has several problems.

* TTM requires significant overcommitment of video memory for reliable page flips. There needs to be 3 times the size of the largest possible framebuffer available or page flips can fail. This leaves the display dark without further warning. Hibmc hardware with 32 MiB and a maximum framebuffer size of 1920x2000 is at the limit.

* No dma-buf sharing without GTT support. Neither gem-vram nor hibmc hardware support a GTT address space. This is required to share buffers with other devices via dma-buf interfaces.

* TTM requires hardware-accelerated rendering into video memory for optimal results. As hibmc hardware cannot do this, hibmc renders in system memory and copies the result to video memory. This can be more effectively implemented with gem-shmem and DRM's shadow-plane helpers.

Converting hibmc to gem-shmem and shadow-plane helpers.

* Replace gem-vram entry points in struct drm_driver with gem-shmem equivalents. This makes the driver allocate struct drm_gem_shmem_object for its buffers.

* Use DRM_GEM_SHADOW_*_PLANE for its plane funcs and plane-helper funcs. The shadow-plane helpers map a plane's gem buffer objects into kernel address space during a page flip, so that atomic_update can copy them to video memory.

* Handle framebuffer damage in hibmc_plane_atomic_update(). This updates video memory from the plane's framebuffer. It automatically synchronizes shared buffers with other devices. Create the framebuffer with drm_gem_fb_create_with_dirty() to trigger the update on each page flip.

* Initialize the plane with drm_plane_enable_fb_damage_clips() to limit the damage updates to the framebuffer areas that changed. We don't want to do a full-buffer memcpy if only a small area has changed.

* Test display modes against the available video memory in hibmc_mode_config_mode_valid(). We only want to announce display modes that fit into display memory.

* Map the display memory itself into kernel address space.

* Do not set drm_mode_config.prefer_shadow. This would advise user space to install a shadow buffer. But with gem-shmem, the gem buffer object already acts as a shadow buffer for video memory.

We use these patterns in many other drivers with similar limitation as hibmc and its hardware. With these changes in place, hibmc is more robust and better integrated into the overall DRM framework.

v3:
- fix coding style
v2:
- do not select TTM symbols

---
 drivers/gpu/drm/drm_gem_shmem_helper.c        | 22 +++++--
 drivers/gpu/drm/hisilicon/hibmc/Kconfig       |  4 +-
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    | 42 ++++++++-----
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c   | 62 ++++++++++++++-----
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h   |  5 ++
 include/drm/drm_gem_shmem_helper.h            |  4 ++
 6 files changed, 102 insertions(+), 37 deletions(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +18 -4
index 545933c7f712..d4ace2a1c47d 100644
@@ -453,10 +453,23 @@ void drm_gem_shmem_vunmap_locked(struct drm_gem_shmem_object *shmem,
 }
 EXPORT_SYMBOL_GPL(drm_gem_shmem_vunmap_locked);
 
-static int
-drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
-				 struct drm_device *dev, size_t size,
-				 uint32_t *handle)
+/**
+ * drm_gem_shmem_create_with_handle - Allocate an object with the given size and
+ *	returns a GEM handle
+ * @file_priv: DRM file structure to create the dumb buffer for
+ * @dev: DRM device
+ * @size: Size of the object to allocate
+ * @handle: Returns the GEM handle on success
+ *
+ * Allocates an shmem GEM buffer using drm_gem_shmem_create() and returns
+ * a GEM handle to it.
+ *
+ * Returns:
+ * Zero on success, or an error code otherwise.
+ */
+int drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
+				     struct drm_device *dev, size_t size,
+				     uint32_t *handle)
 {
 	struct drm_gem_shmem_object *shmem;
 	int ret;
@@ -475,6 +488,7 @@ drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
 
 	return ret;
 }
+EXPORT_SYMBOL_GPL(drm_gem_shmem_create_with_handle);
 
 /* Update madvise status, returns true if not purged, else
  * false or -errno.
drivers/gpu/drm/hisilicon/hibmc/Kconfig +1 -3
index d1f3f5793f34..adf4516bf8f6 100644
@@ -5,10 +5,8 @@ config DRM_HISI_HIBMC
 	select DRM_CLIENT_SELECTION
 	select DRM_DISPLAY_HELPER
 	select DRM_DISPLAY_DP_HELPER
+	select DRM_GEM_SHMEM_HELPER
 	select DRM_KMS_HELPER
-	select DRM_VRAM_HELPER
-	select DRM_TTM
-	select DRM_TTM_HELPER
 	select I2C
 	select I2C_ALGOBIT
 	help
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +28 -14

83 lines not on this page · raw message

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +46 -16

108 lines not on this page · raw message

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h +5 -0
index cd3a3fca1fe6..dce8572bf63e 100644
@@ -38,6 +38,11 @@ struct hibmc_drm_private {
 	/* hw */
 	void __iomem   *mmio;
 
+	/* vram */
+	void __iomem *vram;
+	resource_size_t	vram_base;
+	resource_size_t	vram_size;
+
 	/* drm */
 	struct drm_device dev;
 	struct drm_plane primary_plane;
include/drm/drm_gem_shmem_helper.h +4 -0
index 5ccdae21b94a..86b967174b60 100644
@@ -141,6 +141,10 @@ struct sg_table *drm_gem_shmem_get_pages_sgt(struct drm_gem_shmem_object *shmem)
 void drm_gem_shmem_print_info(const struct drm_gem_shmem_object *shmem,
 			      struct drm_printer *p, unsigned int indent);
 
+int drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
+				     struct drm_device *dev, size_t size,
+				     uint32_t *handle);
+
 extern const struct vm_operations_struct drm_gem_shmem_vm_ops;
 
 /*
Signature · 1 line
2.54.0
maintainer
#6 lore raw

Re: [PATCH v3 4/4] drm/hibmc: Use gem-shmem with shadow-plane helpers for memory management

Hi Thomas.

Quoted 33 lines

#include "hibmc_drm_drv.h"

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +14 -5
@@ -100,28 +102,41 @@ static int hibmc_plane_atomic_check(struct drm_plane *plane,
 static void hibmc_plane_atomic_update(struct drm_plane *plane,
 				      struct drm_atomic_commit *state)
 {
+	struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
 	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state, plane);
+	struct drm_shadow_plane_state *shadow_plane_state = to_drm_shadow_plane_state(new_state);
 	struct drm_framebuffer *fb = new_state->fb;
+	struct drm_plane_state *old_state = drm_atomic_get_old_plane_state(state, plane);
+	u32 gpu_addr = 0;
 	u32 reg;
-	s64 gpu_addr = 0;
 	u32 line_l;
-	struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
-	struct drm_gem_vram_object *gbo;
 
-	if (!new_state->fb)
+	if (!fb)
 		return;
 
-	gbo = drm_gem_vram_of_gem(new_state->fb->obj[0]);
+	if (drm_gem_fb_begin_cpu_access(fb, DMA_FROM_DEVICE) == 0) {
+		struct drm_rect damage;
+		struct drm_atomic_helper_damage_iter iter;
+
+		drm_atomic_helper_damage_iter_init(&iter, old_state, new_state);
+		drm_atomic_for_each_plane_damage(&iter, &damage) {
+			struct iosys_map dst[DRM_FORMAT_MAX_PLANES] = {
+				IOSYS_MAP_INIT_VADDR_IOMEM(priv->vram + gpu_addr),
+			};
 

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c static int hibmc_plane_atomic_check(struct drm_plane *plane, · line 133

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +7 -3
-	gpu_addr = drm_gem_vram_offset(gbo);
-	if (WARN_ON_ONCE(gpu_addr < 0))
-		return; /* Bug: we didn't pin the BO to VRAM in prepare_fb. */
+			iosys_map_incr(&dst[0],
+				       drm_fb_clip_offset(fb->pitches[0], fb->format, &damage));
+			drm_fb_memcpy(dst, fb->pitches, shadow_plane_state->data, fb, &damage);
+		}
+
+		drm_gem_fb_end_cpu_access(fb, DMA_FROM_DEVICE);
+	}

This implementation uses shadow planes for damage rect updates. If the entire plane is updated frequently, could this lead to a performance drop or frame drops?

Originally, the window system(X11/Wayland) drew directly to the hardware framebuffer, but now there is an extra copy in atomic_update.

Quoted 92 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +4 -5
 	writel(gpu_addr, priv->mmio + HIBMC_CRT_FB_ADDRESS);
 
 	reg = drm_format_info_min_pitch(fb->format, 0, fb->width);
 
-	line_l = new_state->fb->pitches[0];
+	line_l = fb->pitches[0];
 	writel(HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_WIDTH, reg) |
 	       HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_OFFS, line_l),
 	       priv->mmio + HIBMC_CRT_FB_WIDTH);
@@ -149,13 +164,11 @@ static const struct drm_plane_funcs hibmc_plane_funcs = {
 	.update_plane	= drm_atomic_helper_update_plane,
 	.disable_plane	= drm_atomic_helper_disable_plane,
 	.destroy = drm_plane_cleanup,
-	.reset = drm_atomic_helper_plane_reset,
-	.atomic_duplicate_state = drm_atomic_helper_plane_duplicate_state,
-	.atomic_destroy_state = drm_atomic_helper_plane_destroy_state,
+	DRM_GEM_SHADOW_PLANE_FUNCS,
 };
 
 static const struct drm_plane_helper_funcs hibmc_plane_helper_funcs = {
-	DRM_GEM_VRAM_PLANE_HELPER_FUNCS,
+	DRM_GEM_SHADOW_PLANE_HELPER_FUNCS,
 	.atomic_check = hibmc_plane_atomic_check,
 	.atomic_update = hibmc_plane_atomic_update,
 };
@@ -515,6 +528,7 @@ int hibmc_de_init(struct hibmc_drm_private *priv)
 	}
 
 	drm_plane_helper_add(plane, &hibmc_plane_helper_funcs);
+	drm_plane_enable_fb_damage_clips(plane);
 
 	ret = drm_crtc_init_with_planes(dev, crtc, plane,
 					NULL, &hibmc_crtc_funcs, NULL);
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +21 -7
index 99b36de1fe13..19d3193e2f76 100644
@@ -18,9 +18,10 @@
 #include <drm/clients/drm_client_setup.h>
 #include <drm/drm_atomic_helper.h>
 #include <drm/drm_drv.h>
-#include <drm/drm_fbdev_ttm.h>
+#include <drm/drm_dumb_buffers.h>
+#include <drm/drm_fbdev_shmem.h>
 #include <drm/drm_gem_framebuffer_helper.h>
-#include <drm/drm_gem_vram_helper.h>
+#include <drm/drm_gem_shmem_helper.h>
 #include <drm/drm_managed.h>
 #include <drm/drm_module.h>
 #include <drm/drm_vblank.h>
@@ -71,7 +72,13 @@ static irqreturn_t hibmc_dp_interrupt(int irq, void *arg)
 static int hibmc_dumb_create(struct drm_file *file, struct drm_device *dev,
 			     struct drm_mode_create_dumb *args)
 {
-	return drm_gem_vram_fill_create_dumb(file, dev, 0, 128, args);
+	int ret;
+
+	ret = drm_mode_size_dumb(dev, args, SZ_128, 0);
+	if (ret)
+		return ret;
+
+	return drm_gem_shmem_create_with_handle(file, dev, args->size, &args->handle);
 }
 
 static const struct drm_driver hibmc_driver = {
@@ -81,10 +88,9 @@ static const struct drm_driver hibmc_driver = {
 	.desc			= "hibmc drm driver",
 	.major			= 1,
 	.minor			= 0,
-	.debugfs_init		= drm_vram_mm_debugfs_init,
-	.dumb_create            = hibmc_dumb_create,
-	.dumb_map_offset        = drm_gem_ttm_dumb_map_offset,
-	DRM_FBDEV_TTM_DRIVER_OPS,
+	.gem_prime_import	= drm_gem_shmem_prime_import_no_map,
+	.dumb_create		= hibmc_dumb_create,
+	DRM_FBDEV_SHMEM_DRIVER_OPS,
 };
 
 static int __maybe_unused hibmc_pm_suspend(struct device *dev)
@@ -106,11 +112,32 @@ static const struct dev_pm_ops hibmc_pm_ops = {
 				hibmc_pm_resume)
 };
 
+static enum drm_mode_status hibmc_mode_config_mode_valid(struct drm_device *dev,
+							 const struct drm_display_mode *mode)
+{
+	const struct drm_format_info *info =
+		drm_get_format_info(dev, DRM_FORMAT_XRGB8888, DRM_FORMAT_MOD_LINEAR);
+	struct hibmc_drm_private *priv = to_hibmc_drm_private(dev);
+	unsigned long max_fb_size = priv->vram_size;
+	u64 pitch;

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c static const struct dev_pm_ops hibmc_pm_ops = { · line 132

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +10 -0
+
+	if (drm_WARN_ON_ONCE(dev, !info))
+		return MODE_ERROR; /* driver bug */
+
+	pitch = drm_format_info_min_pitch(info, 0, mode->hdisplay);
+	if (!pitch)
+		return MODE_BAD_WIDTH;
+	else if (pitch > max_fb_size / mode->vdisplay)
+		return MODE_MEM;
+

I reviewed the AST and Cirrus implementations, and they include checks for the maximum size supported by the hardware.

The hardware FB width register has only 14 bits, and the HIBMC_CRT_FB_WIDTH_WIDTH_MASK macro defines the maximum size as 0x3FFF. Would adding the following check make the code more rigorous?

Also, the pitch requires 128-byte alignment. Should we add a check here, or does the buffer created by hibmc_dumb_create already ensure that the FB start address is aligned to 128 bytes via SZ_128? I’m not entirely sure about this.

Of course, the original code used drm_vram_helper_mode_valid without strictly checking these hardware limitations. Thank you for your modification.

Thanks,
Yongbang.

Quoted 81 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +15 -9
+	return MODE_OK;
+}
+
 static const struct drm_mode_config_funcs hibmc_mode_funcs = {
-	.mode_valid = drm_vram_helper_mode_valid,
+	.mode_valid = hibmc_mode_config_mode_valid,
 	.atomic_check = drm_atomic_helper_check,
 	.atomic_commit = drm_atomic_helper_commit,
-	.fb_create = drm_gem_fb_create,
+	.fb_create = drm_gem_fb_create_with_dirty,
 };
 
 static int hibmc_kms_init(struct hibmc_drm_private *priv)
@@ -130,7 +157,6 @@ static int hibmc_kms_init(struct hibmc_drm_private *priv)
 	dev->mode_config.max_height = 1200;
 
 	dev->mode_config.preferred_depth = 24;
-	dev->mode_config.prefer_shadow = 1;
 
 	dev->mode_config.funcs = (void *)&hibmc_mode_funcs;
 
@@ -335,18 +361,22 @@ static int hibmc_load(struct drm_device *dev)
 {
 	struct pci_dev *pdev = to_pci_dev(dev->dev);
 	struct hibmc_drm_private *priv = to_hibmc_drm_private(dev);
+	resource_size_t vram_base, vram_size;
 	int ret;
 
 	ret = hibmc_hw_init(priv);
 	if (ret)
 		return ret;
 
-	ret = drmm_vram_helper_init(dev, pci_resource_start(pdev, 0),
-				    pci_resource_len(pdev, 0));
-	if (ret) {
-		drm_err(dev, "Error initializing VRAM MM; %d\n", ret);
-		return ret;
-	}
+	vram_base = pci_resource_start(pdev, 0);
+	vram_size = pci_resource_len(pdev, 0);
+
+	priv->vram = devm_ioremap_wc(dev->dev, vram_base, vram_size);
+	if (!priv->vram)
+		return -ENOMEM;
+
+	priv->vram_base = vram_base;
+	priv->vram_size = vram_size;
 
 	ret = hibmc_kms_init(priv);
 	if (ret)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h +5 -0
index cd3a3fca1fe6..dce8572bf63e 100644
@@ -38,6 +38,11 @@ struct hibmc_drm_private {
 	/* hw */
 	void __iomem   *mmio;
 
+	/* vram */
+	void __iomem *vram;
+	resource_size_t	vram_base;
+	resource_size_t	vram_size;
+
 	/* drm */
 	struct drm_device dev;
 	struct drm_plane primary_plane;
include/drm/drm_gem_shmem_helper.h +4 -0
index 5ccdae21b94a..86b967174b60 100644
@@ -141,6 +141,10 @@ struct sg_table *drm_gem_shmem_get_pages_sgt(struct drm_gem_shmem_object *shmem)
 void drm_gem_shmem_print_info(const struct drm_gem_shmem_object *shmem,
 			      struct drm_printer *p, unsigned int indent);
 
+int drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
+				     struct drm_device *dev, size_t size,
+				     uint32_t *handle);
+
 extern const struct vm_operations_struct drm_gem_shmem_vm_ops;
 
 /*
author maintainer
#7 lore raw

Re: [PATCH v3 4/4] drm/hibmc: Use gem-shmem with shadow-plane helpers for memory management

Hi

Am 08.06.26 um 15:42 schrieb Yongbang Shi:

Quoted 48 lines

Hi Thomas.

Quoted 33 lines
  
  #include "hibmc_drm_drv.h"
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +14 -5
@@ -100,28 +102,41 @@ static int hibmc_plane_atomic_check(struct drm_plane *plane,
  static void hibmc_plane_atomic_update(struct drm_plane *plane,
  				      struct drm_atomic_commit *state)
  {
+	struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
  	struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state, plane);
+	struct drm_shadow_plane_state *shadow_plane_state = to_drm_shadow_plane_state(new_state);
  	struct drm_framebuffer *fb = new_state->fb;
+	struct drm_plane_state *old_state = drm_atomic_get_old_plane_state(state, plane);
+	u32 gpu_addr = 0;
  	u32 reg;
-	s64 gpu_addr = 0;
  	u32 line_l;
-	struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
-	struct drm_gem_vram_object *gbo;
 
-	if (!new_state->fb)
+	if (!fb)
  		return;
 
-	gbo = drm_gem_vram_of_gem(new_state->fb->obj[0]);
+	if (drm_gem_fb_begin_cpu_access(fb, DMA_FROM_DEVICE) == 0) {
+		struct drm_rect damage;
+		struct drm_atomic_helper_damage_iter iter;
+
+		drm_atomic_helper_damage_iter_init(&iter, old_state, new_state);
+		drm_atomic_for_each_plane_damage(&iter, &damage) {
+			struct iosys_map dst[DRM_FORMAT_MAX_PLANES] = {
+				IOSYS_MAP_INIT_VADDR_IOMEM(priv->vram + gpu_addr),
+			};
 

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c static int hibmc_plane_atomic_check(struct drm_plane *plane, · line 133

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +7 -3
-	gpu_addr = drm_gem_vram_offset(gbo);
-	if (WARN_ON_ONCE(gpu_addr < 0))
-		return; /* Bug: we didn't pin the BO to VRAM in prepare_fb. */
+			iosys_map_incr(&dst[0],
+				       drm_fb_clip_offset(fb->pitches[0], fb->format, &damage));
+			drm_fb_memcpy(dst, fb->pitches, shadow_plane_state->data, fb, &damage);
+		}
+
+		drm_gem_fb_end_cpu_access(fb, DMA_FROM_DEVICE);
+	}

This implementation uses shadow planes for damage rect updates. If the entire plane is updated frequently, could this lead to a performance drop or frame drops?

Quoted 2 lines

Originally, the window system(X11/Wayland) drew directly to the hardware framebuffer, but now there is an extra copy in atomic_update.

Generally speaking, shadow buffering is necessary for acceptable performance of the software renderer, because PCI access is rather slow (when compared to system memory).

The current hibmc driver sets drm_mode_config.prefer_shadow, [1] which tells compositors to maintain a shadow buffer.  IIRC Xorg announces the by mentioning shadowfb in its log files. It then writes the fully composed shadow buffer into video memory. Wayland compositors do the same AFAIK. Related code from Gnome's windows manager is a [2].

With the patch applied, prefer_shadow is no longer required. The kernel driver now maintains the shadow buffer. If there's a performance penalty, I'm not aware of it.

All this also applies to the kernel's framebuffer console. It currently uses an internal shadow buffer [3], but that is no longer needed with gem-shmem.

[1] https://elixir.bootlin.com/linux/v7.0.12/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c#L132
[2] https://gitlab.gnome.org/GNOME/mutter/-/blob/50.2/src/backends/native/meta-kms-impl-device.c?ref_type=tags#L587
[3] https://elixir.bootlin.com/linux/v7.0.12/source/drivers/gpu/drm/drm_fbdev_ttm.c#L200

Quoted 105 lines
Quoted 93 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +4 -5
  	writel(gpu_addr, priv->mmio + HIBMC_CRT_FB_ADDRESS);
 
  	reg = drm_format_info_min_pitch(fb->format, 0, fb->width);
 
-	line_l = new_state->fb->pitches[0];
+	line_l = fb->pitches[0];
  	writel(HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_WIDTH, reg) |
  	       HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_OFFS, line_l),
  	       priv->mmio + HIBMC_CRT_FB_WIDTH);
@@ -149,13 +164,11 @@ static const struct drm_plane_funcs hibmc_plane_funcs = {
  	.update_plane	= drm_atomic_helper_update_plane,
  	.disable_plane	= drm_atomic_helper_disable_plane,
  	.destroy = drm_plane_cleanup,
-	.reset = drm_atomic_helper_plane_reset,
-	.atomic_duplicate_state = drm_atomic_helper_plane_duplicate_state,
-	.atomic_destroy_state = drm_atomic_helper_plane_destroy_state,
+	DRM_GEM_SHADOW_PLANE_FUNCS,
  };
 
  static const struct drm_plane_helper_funcs hibmc_plane_helper_funcs = {
-	DRM_GEM_VRAM_PLANE_HELPER_FUNCS,
+	DRM_GEM_SHADOW_PLANE_HELPER_FUNCS,
  	.atomic_check = hibmc_plane_atomic_check,
  	.atomic_update = hibmc_plane_atomic_update,
  };
@@ -515,6 +528,7 @@ int hibmc_de_init(struct hibmc_drm_private *priv)
  	}
 
  	drm_plane_helper_add(plane, &hibmc_plane_helper_funcs);
+	drm_plane_enable_fb_damage_clips(plane);
 
  	ret = drm_crtc_init_with_planes(dev, crtc, plane,
  					NULL, &hibmc_crtc_funcs, NULL);
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +21 -7
index 99b36de1fe13..19d3193e2f76 100644
@@ -18,9 +18,10 @@
  #include <drm/clients/drm_client_setup.h>
  #include <drm/drm_atomic_helper.h>
  #include <drm/drm_drv.h>
-#include <drm/drm_fbdev_ttm.h>
+#include <drm/drm_dumb_buffers.h>
+#include <drm/drm_fbdev_shmem.h>
  #include <drm/drm_gem_framebuffer_helper.h>
-#include <drm/drm_gem_vram_helper.h>
+#include <drm/drm_gem_shmem_helper.h>
  #include <drm/drm_managed.h>
  #include <drm/drm_module.h>
  #include <drm/drm_vblank.h>
@@ -71,7 +72,13 @@ static irqreturn_t hibmc_dp_interrupt(int irq, void *arg)
  static int hibmc_dumb_create(struct drm_file *file, struct drm_device *dev,
  			     struct drm_mode_create_dumb *args)
  {
-	return drm_gem_vram_fill_create_dumb(file, dev, 0, 128, args);
+	int ret;
+
+	ret = drm_mode_size_dumb(dev, args, SZ_128, 0);
+	if (ret)
+		return ret;
+
+	return drm_gem_shmem_create_with_handle(file, dev, args->size, &args->handle);
  }
 
  static const struct drm_driver hibmc_driver = {
@@ -81,10 +88,9 @@ static const struct drm_driver hibmc_driver = {
  	.desc			= "hibmc drm driver",
  	.major			= 1,
  	.minor			= 0,
-	.debugfs_init		= drm_vram_mm_debugfs_init,
-	.dumb_create            = hibmc_dumb_create,
-	.dumb_map_offset        = drm_gem_ttm_dumb_map_offset,
-	DRM_FBDEV_TTM_DRIVER_OPS,
+	.gem_prime_import	= drm_gem_shmem_prime_import_no_map,
+	.dumb_create		= hibmc_dumb_create,
+	DRM_FBDEV_SHMEM_DRIVER_OPS,
  };
 
  static int __maybe_unused hibmc_pm_suspend(struct device *dev)
@@ -106,11 +112,32 @@ static const struct dev_pm_ops hibmc_pm_ops = {
  				hibmc_pm_resume)
  };
 
+static enum drm_mode_status hibmc_mode_config_mode_valid(struct drm_device *dev,
+							 const struct drm_display_mode *mode)
+{
+	const struct drm_format_info *info =
+		drm_get_format_info(dev, DRM_FORMAT_XRGB8888, DRM_FORMAT_MOD_LINEAR);
+	struct hibmc_drm_private *priv = to_hibmc_drm_private(dev);
+	unsigned long max_fb_size = priv->vram_size;
+	u64 pitch;

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c static const struct dev_pm_ops hibmc_pm_ops = { · line 132

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +10 -0
+
+	if (drm_WARN_ON_ONCE(dev, !info))
+		return MODE_ERROR; /* driver bug */
+
+	pitch = drm_format_info_min_pitch(info, 0, mode->hdisplay);
+	if (!pitch)
+		return MODE_BAD_WIDTH;
+	else if (pitch > max_fb_size / mode->vdisplay)
+		return MODE_MEM;
+

I reviewed the AST and Cirrus implementations, and they include checks for the maximum size supported by the hardware.

Quoted 2 lines

The hardware FB width register has only 14 bits, and the HIBMC_CRT_FB_WIDTH_WIDTH_MASK macro defines the maximum size as 0x3FFF. Would adding the following check make the code more rigorous?

The helper here is the general mode_valid callback from drm_mode_config_funcs. It should only test against devide-wide limits, such as the available video memory.

The framebuffer size is apparently a limit of the CRTC. There's already a check in the CRTC's mode_valid for the supported resolutions. [4] I think this should be fine.

[4] https://elixir.bootlin.com/linux/v7.1-rc7/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c#L223

Quoted 4 lines

Also, the pitch requires 128-byte alignment. Should we add a check here, or does the buffer created by hibmc_dumb_create already ensure that the FB start address is aligned to 128 bytes via SZ_128? I’m not entirely sure about this.

Yes, using SZ_128 aligns each scanline in the framebuffer to a multiple of 128 bytes. The underlying GEM buffer is sized accordingly. And the GEM buffer itself always starts at a page boundary. Any GEM buffer allocated via the dumb-buffer interface should be correct and the ioctl also returns the correct pitch back into user space.

The driver later programs the hardware pitch to the value stored in the framebuffer (in pitches[0]). If it then does a memcpy from the GEM buffer to video memory, it all works out.

A corner case would be that user space allocates via the dumb-buffers ioctl, but ignores the returned pitch. If we want to do additional checks for the pitch, we could add a wrapper function around fb_create. [5]  This would be the right place to verify that user space does not allocate unaligned framebuffers.

[5] https://elixir.bootlin.com/linux/v7.1-rc7/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c#L112

Best regards
Thomas

Quoted 89 lines

Of course, the original code used drm_vram_helper_mode_valid without strictly checking these hardware limitations. Thank you for your modification.

Thanks,
Yongbang.

Quoted 81 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +15 -9
+	return MODE_OK;
+}
+
  static const struct drm_mode_config_funcs hibmc_mode_funcs = {
-	.mode_valid = drm_vram_helper_mode_valid,
+	.mode_valid = hibmc_mode_config_mode_valid,
  	.atomic_check = drm_atomic_helper_check,
  	.atomic_commit = drm_atomic_helper_commit,
-	.fb_create = drm_gem_fb_create,
+	.fb_create = drm_gem_fb_create_with_dirty,
  };
 
  static int hibmc_kms_init(struct hibmc_drm_private *priv)
@@ -130,7 +157,6 @@ static int hibmc_kms_init(struct hibmc_drm_private *priv)
  	dev->mode_config.max_height = 1200;
 
  	dev->mode_config.preferred_depth = 24;
-	dev->mode_config.prefer_shadow = 1;
 
  	dev->mode_config.funcs = (void *)&hibmc_mode_funcs;
 
@@ -335,18 +361,22 @@ static int hibmc_load(struct drm_device *dev)
  {
  	struct pci_dev *pdev = to_pci_dev(dev->dev);
  	struct hibmc_drm_private *priv = to_hibmc_drm_private(dev);
+	resource_size_t vram_base, vram_size;
  	int ret;
 
  	ret = hibmc_hw_init(priv);
  	if (ret)
  		return ret;
 
-	ret = drmm_vram_helper_init(dev, pci_resource_start(pdev, 0),
-				    pci_resource_len(pdev, 0));
-	if (ret) {
-		drm_err(dev, "Error initializing VRAM MM; %d\n", ret);
-		return ret;
-	}
+	vram_base = pci_resource_start(pdev, 0);
+	vram_size = pci_resource_len(pdev, 0);
+
+	priv->vram = devm_ioremap_wc(dev->dev, vram_base, vram_size);
+	if (!priv->vram)
+		return -ENOMEM;
+
+	priv->vram_base = vram_base;
+	priv->vram_size = vram_size;
 
  	ret = hibmc_kms_init(priv);
  	if (ret)
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h +5 -0
index cd3a3fca1fe6..dce8572bf63e 100644
@@ -38,6 +38,11 @@ struct hibmc_drm_private {
  	/* hw */
  	void __iomem   *mmio;
 
+	/* vram */
+	void __iomem *vram;
+	resource_size_t	vram_base;
+	resource_size_t	vram_size;
+
  	/* drm */
  	struct drm_device dev;
  	struct drm_plane primary_plane;
include/drm/drm_gem_shmem_helper.h +4 -0
index 5ccdae21b94a..86b967174b60 100644
@@ -141,6 +141,10 @@ struct sg_table *drm_gem_shmem_get_pages_sgt(struct drm_gem_shmem_object *shmem)
  void drm_gem_shmem_print_info(const struct drm_gem_shmem_object *shmem,
  			      struct drm_printer *p, unsigned int indent);
 
+int drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
+				     struct drm_device *dev, size_t size,
+				     uint32_t *handle);
+
  extern const struct vm_operations_struct drm_gem_shmem_vm_ops;
 
  /*
Signature · 6 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
maintainer
#8 lore raw

Re: [PATCH v3 4/4] drm/hibmc: Use gem-shmem with shadow-plane helpers for memory management

Quoted 60 lines

Hi

Am 08.06.26 um 15:42 schrieb Yongbang Shi:

Quoted 44 lines

Hi Thomas.

Quoted 29 lines

    #include "hibmc_drm_drv.h"

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +14 -3
@@ -100,28 +102,41 @@ static int hibmc_plane_atomic_check(struct drm_plane *plane,
   static void hibmc_plane_atomic_update(struct drm_plane *plane,
                         struct drm_atomic_commit *state)
   {
+    struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
       struct drm_plane_state *new_state = drm_atomic_get_new_plane_state(state, plane);
+    struct drm_shadow_plane_state *shadow_plane_state = to_drm_shadow_plane_state(new_state);
       struct drm_framebuffer *fb = new_state->fb;
+    struct drm_plane_state *old_state = drm_atomic_get_old_plane_state(state, plane);
+    u32 gpu_addr = 0;
       u32 reg;
-    s64 gpu_addr = 0;
       u32 line_l;
-    struct hibmc_drm_private *priv = to_hibmc_drm_private(plane->dev);
-    struct drm_gem_vram_object *gbo;
   -    if (!new_state->fb)
+    if (!fb)
           return;
   -    gbo = drm_gem_vram_of_gem(new_state->fb->obj[0]);
+    if (drm_gem_fb_begin_cpu_access(fb, DMA_FROM_DEVICE) == 0) {
+        struct drm_rect damage;
+        struct drm_atomic_helper_damage_iter iter;
+
+        drm_atomic_helper_damage_iter_init(&iter, old_state, new_state);
+        drm_atomic_for_each_plane_damage(&iter, &damage) {
+            struct iosys_map dst[DRM_FORMAT_MAX_PLANES] = {
+                IOSYS_MAP_INIT_VADDR_IOMEM(priv->vram + gpu_addr),
+            };

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c static int hibmc_plane_atomic_check(struct drm_plane *plane, · line 133

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +7 -2
   -    gpu_addr = drm_gem_vram_offset(gbo);
-    if (WARN_ON_ONCE(gpu_addr < 0))
-        return; /* Bug: we didn't pin the BO to VRAM in prepare_fb. */
+            iosys_map_incr(&dst[0],
+                       drm_fb_clip_offset(fb->pitches[0], fb->format, &damage));
+            drm_fb_memcpy(dst, fb->pitches, shadow_plane_state->data, fb, &damage);
+        }
+
+        drm_gem_fb_end_cpu_access(fb, DMA_FROM_DEVICE);
+    }

This implementation uses shadow planes for damage rect updates. If the entire plane is updated frequently, could this lead to a performance drop or frame drops?

Quoted 2 lines

Originally, the window system(X11/Wayland) drew directly to the hardware framebuffer, but now there is an extra copy in atomic_update.

Generally speaking, shadow buffering is necessary for acceptable performance of the software renderer, because PCI access is rather slow (when compared to system memory).

The current hibmc driver sets drm_mode_config.prefer_shadow, [1] which tells compositors to maintain a shadow buffer. IIRC Xorg announces the by mentioning shadowfb in its log files. It then writes the fully composed shadow buffer into video memory. Wayland compositors do the same AFAIK. Related code from Gnome's windows manager is a [2].

With the patch applied, prefer_shadow is no longer required. The kernel driver now maintains the shadow buffer. If there's a performance penalty, I'm not aware of it.

Quoted 7 lines

All this also applies to the kernel's framebuffer console. It currently uses an internal shadow buffer [3], but that is no longer needed with gem-shmem.

[1] https://elixir.bootlin.com/linux/v7.0.12/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c#L132
[2] https://gitlab.gnome.org/GNOME/mutter/-/blob/50.2/src/backends/native/meta-kms-impl-device.c?ref_type=tags#L587
[3] https://elixir.bootlin.com/linux/v7.0.12/source/drivers/gpu/drm/drm_fbdev_ttm.c#L200

Okay, I get it. Thanks for explaining.

Quoted 96 lines
Quoted 96 lines
Quoted 84 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +1 -0
         writel(gpu_addr, priv->mmio + HIBMC_CRT_FB_ADDRESS);
         reg = drm_format_info_min_pitch(fb->format, 0, fb->width);
   -    line_l = new_state->fb->pitches[0];
+    line_l = fb->pitches[0];

      writel(HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_WIDTH, reg) | HIBMC_FIELD(HIBMC_CRT_FB_WIDTH_OFFS, line_l), priv->mmio + HIBMC_CRT_FB_WIDTH);

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +3 -4
@@ -149,13 +164,11 @@ static const struct drm_plane_funcs hibmc_plane_funcs = {
       .update_plane    = drm_atomic_helper_update_plane,
       .disable_plane    = drm_atomic_helper_disable_plane,
       .destroy = drm_plane_cleanup,
-    .reset = drm_atomic_helper_plane_reset,
-    .atomic_duplicate_state = drm_atomic_helper_plane_duplicate_state,
-    .atomic_destroy_state = drm_atomic_helper_plane_destroy_state,
+    DRM_GEM_SHADOW_PLANE_FUNCS,
   };
     static const struct drm_plane_helper_funcs hibmc_plane_helper_funcs = {
-    DRM_GEM_VRAM_PLANE_HELPER_FUNCS,
+    DRM_GEM_SHADOW_PLANE_HELPER_FUNCS,
       .atomic_check = hibmc_plane_atomic_check,
       .atomic_update = hibmc_plane_atomic_update,
   };
@@ -515,6 +528,7 @@ int hibmc_de_init(struct hibmc_drm_private *priv)
       }
         drm_plane_helper_add(plane, &hibmc_plane_helper_funcs);
+    drm_plane_enable_fb_damage_clips(plane);
         ret = drm_crtc_init_with_planes(dev, crtc, plane,
                       NULL, &hibmc_crtc_funcs, NULL);
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +20 -7
index 99b36de1fe13..19d3193e2f76 100644
@@ -18,9 +18,10 @@
   #include <drm/clients/drm_client_setup.h>
   #include <drm/drm_atomic_helper.h>
   #include <drm/drm_drv.h>
-#include <drm/drm_fbdev_ttm.h>
+#include <drm/drm_dumb_buffers.h>
+#include <drm/drm_fbdev_shmem.h>
   #include <drm/drm_gem_framebuffer_helper.h>
-#include <drm/drm_gem_vram_helper.h>
+#include <drm/drm_gem_shmem_helper.h>
   #include <drm/drm_managed.h>
   #include <drm/drm_module.h>
   #include <drm/drm_vblank.h>
@@ -71,7 +72,13 @@ static irqreturn_t hibmc_dp_interrupt(int irq, void *arg)
   static int hibmc_dumb_create(struct drm_file *file, struct drm_device *dev,
                    struct drm_mode_create_dumb *args)
   {
-    return drm_gem_vram_fill_create_dumb(file, dev, 0, 128, args);
+    int ret;
+
+    ret = drm_mode_size_dumb(dev, args, SZ_128, 0);
+    if (ret)
+        return ret;
+
+    return drm_gem_shmem_create_with_handle(file, dev, args->size, &args->handle);
   }
     static const struct drm_driver hibmc_driver = {
@@ -81,10 +88,9 @@ static const struct drm_driver hibmc_driver = {
       .desc            = "hibmc drm driver",
       .major            = 1,
       .minor            = 0,
-    .debugfs_init        = drm_vram_mm_debugfs_init,
-    .dumb_create            = hibmc_dumb_create,
-    .dumb_map_offset        = drm_gem_ttm_dumb_map_offset,
-    DRM_FBDEV_TTM_DRIVER_OPS,
+    .gem_prime_import    = drm_gem_shmem_prime_import_no_map,
+    .dumb_create        = hibmc_dumb_create,
+    DRM_FBDEV_SHMEM_DRIVER_OPS,
   };
     static int __maybe_unused hibmc_pm_suspend(struct device *dev)
@@ -106,11 +112,32 @@ static const struct dev_pm_ops hibmc_pm_ops = {
                   hibmc_pm_resume)
   };
   +static enum drm_mode_status hibmc_mode_config_mode_valid(struct drm_device *dev,
+                             const struct drm_display_mode *mode)
+{
+    const struct drm_format_info *info =
+        drm_get_format_info(dev, DRM_FORMAT_XRGB8888, DRM_FORMAT_MOD_LINEAR);
+    struct hibmc_drm_private *priv = to_hibmc_drm_private(dev);
+    unsigned long max_fb_size = priv->vram_size;
+    u64 pitch;

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c static const struct dev_pm_ops hibmc_pm_ops = { · line 131

Quoted 10 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +10 -0
+
+    if (drm_WARN_ON_ONCE(dev, !info))
+        return MODE_ERROR; /* driver bug */
+
+    pitch = drm_format_info_min_pitch(info, 0, mode->hdisplay);
+    if (!pitch)
+        return MODE_BAD_WIDTH;
+    else if (pitch > max_fb_size / mode->vdisplay)
+        return MODE_MEM;
+

I reviewed the AST and Cirrus implementations, and they include checks for the maximum size supported by the hardware.

Quoted 8 lines
Quoted 2 lines

The hardware FB width register has only 14 bits, and the HIBMC_CRT_FB_WIDTH_WIDTH_MASK macro defines the maximum size as 0x3FFF. Would adding the following check make the code more rigorous?

The helper here is the general mode_valid callback from drm_mode_config_funcs. It should only test against devide-wide limits, such as the available video memory.

The framebuffer size is apparently a limit of the CRTC. There's already a check in the CRTC's mode_valid for the supported resolutions. [4] I think this should be fine.

Yes, the CRTC's mode_valid function ensures that the resolution falls within the valid fb width range; adding this check wouldn't really provide any additional practical benefit.

Quoted 12 lines

[4] https://elixir.bootlin.com/linux/v7.1-rc7/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c#L223

Quoted 4 lines

Also, the pitch requires 128-byte alignment. Should we add a check here, or does the buffer created by hibmc_dumb_create already ensure that the FB start address is aligned to 128 bytes via SZ_128? I’m not entirely sure about this.

Yes, using SZ_128 aligns each scanline in the framebuffer to a multiple of 128 bytes. The underlying GEM buffer is sized accordingly. And the GEM buffer itself always starts at a page boundary. Any GEM buffer allocated via the dumb-buffer interface should be correct and the ioctl also returns the correct pitch back into user space.

Quoted 7 lines

The driver later programs the hardware pitch to the value stored in the framebuffer (in pitches[0]). If it then does a memcpy from the GEM buffer to video memory, it all works out.

A corner case would be that user space allocates via the dumb-buffers ioctl, but ignores the returned pitch. If we want to do additional checks for the pitch, we could add a wrapper function around fb_create. [5]  This would be the right place to verify that user space does not allocate unaligned framebuffers.

Thanks for the suggestion. Yes, since the kernel does not trust user space, a check could be added in fb_create.

For windowing systems like Xorg or Wayland, I believe the pitch is typically handled correctly. However, for some simple or buggy DRM clients, if the driver does not check the pitch, it may result in misaligned or distorted screen display.

Thanks,
Yongbang.

Quoted 85 lines

[5] https://elixir.bootlin.com/linux/v7.1-rc7/source/drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c#L112

Best regards
Thomas

Quoted 79 lines

Of course, the original code used drm_vram_helper_mode_valid without strictly checking these hardware limitations. Thank you for your modification.

Thanks,
Yongbang.

Quoted 71 lines
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +5 -3
+    return MODE_OK;
+}
+
   static const struct drm_mode_config_funcs hibmc_mode_funcs = {
-    .mode_valid = drm_vram_helper_mode_valid,
+    .mode_valid = hibmc_mode_config_mode_valid,
       .atomic_check = drm_atomic_helper_check,
       .atomic_commit = drm_atomic_helper_commit,
-    .fb_create = drm_gem_fb_create,
+    .fb_create = drm_gem_fb_create_with_dirty,
   };
     static int hibmc_kms_init(struct hibmc_drm_private *priv)
@@ -130,7 +157,6 @@ static int hibmc_kms_init(struct hibmc_drm_private *priv)
       dev->mode_config.max_height = 1200;
         dev->mode_config.preferred_depth = 24;
-    dev->mode_config.prefer_shadow = 1;
         dev->mode_config.funcs = (void *)&hibmc_mode_funcs;
   @@ -335,18 +361,22 @@ static int hibmc_load(struct drm_device *dev)
   {
       struct pci_dev *pdev = to_pci_dev(dev->dev);

      struct hibmc_drm_private *priv = to_hibmc_drm_private(dev); +    resource_size_t vram_base, vram_size;
      int ret;
        ret = hibmc_hw_init(priv);
      if (ret)
          return ret;
  -    ret = drmm_vram_helper_init(dev, pci_resource_start(pdev, 0), -                    pci_resource_len(pdev, 0));
-    if (ret) {
-        drm_err(dev, "Error initializing VRAM MM; %d\n", ret); -        return ret;
-    }
+    vram_base = pci_resource_start(pdev, 0);
+    vram_size = pci_resource_len(pdev, 0);

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.c +7 -0
+
+    priv->vram = devm_ioremap_wc(dev->dev, vram_base, vram_size);
+    if (!priv->vram)
+        return -ENOMEM;
+
+    priv->vram_base = vram_base;
+    priv->vram_size = vram_size;

        ret = hibmc_kms_init(priv);
      if (ret)

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_drv.h +0 -0
index cd3a3fca1fe6..dce8572bf63e 100644
@@ -38,6 +38,11 @@ struct hibmc_drm_private {
       /* hw */
       void __iomem   *mmio;

  +    /* vram */
+    void __iomem *vram;
+    resource_size_t    vram_base;
+    resource_size_t    vram_size;
+
      /* drm */
      struct drm_device dev;
      struct drm_plane primary_plane;

include/drm/drm_gem_shmem_helper.h +3 -0
index 5ccdae21b94a..86b967174b60 100644
@@ -141,6 +141,10 @@ struct sg_table *drm_gem_shmem_get_pages_sgt(struct drm_gem_shmem_object *shmem)
   void drm_gem_shmem_print_info(const struct drm_gem_shmem_object *shmem,
                     struct drm_printer *p, unsigned int indent);
   +int drm_gem_shmem_create_with_handle(struct drm_file *file_priv,
+                     struct drm_device *dev, size_t size,
+                     uint32_t *handle);
+
   extern const struct vm_operations_struct drm_gem_shmem_vm_ops;
     /*