server gpus: fix race between buffer evict and fbdev updates

server gpus: fix race between buffer evict and fbdev updates

[Dave Airlie]

This stores up dirty updates to the fbdev until the next update after
the buffer has finished moving, otherwise you'd see a lot of failed
to reserve bo during bootup with plymouth was starting.

Dave.

[PATCH 1/3] drm/mgag200: deal with bo reserve fail in dirty update path

[Dave Airlie]

From: Dave Airlie <airlied@redhat.com>

On F19 testing, it was noticed we get a lot of errors in dmesg
about being unable to reserve the buffer when plymouth starts,
this is due to the buffer being in the process of migrating,
so it makes sense we can't reserve it.

In order to deal with it, this adds delayed updates for the dirty
updates, when the bo is unreservable, in the normal console case
this shouldn't ever happen, its just when plymouth or X is
pushing the console bo to system memory.

Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
---
 drivers/gpu/drm/mgag200/mgag200_drv.h |  2 ++
 drivers/gpu/drm/mgag200/mgag200_fb.c  | 43 ++++++++++++++++++++++++++++++++---
 drivers/gpu/drm/mgag200/mgag200_ttm.c |  4 ++--
 3 files changed, 44 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/mgag200/mgag200_drv.h b/drivers/gpu/drm/mgag200/mgag200_drv.h
index dcfc973..bf29b2f 100644
--- a/drivers/gpu/drm/mgag200/mgag200_drv.h
+++ b/drivers/gpu/drm/mgag200/mgag200_drv.h
@@ -115,6 +115,8 @@ struct mga_fbdev {
 	void *sysram;
 	int size;
 	struct ttm_bo_kmap_obj mapping;
+	int x1, y1, x2, y2; /* dirty rect */
+	spinlock_t dirty_lock;
 };
 
 struct mga_crtc {
diff --git a/drivers/gpu/drm/mgag200/mgag200_fb.c b/drivers/gpu/drm/mgag200/mgag200_fb.c
index 421beab..5da824c 100644
--- a/drivers/gpu/drm/mgag200/mgag200_fb.c
+++ b/drivers/gpu/drm/mgag200/mgag200_fb.c
@@ -29,16 +29,52 @@ static void mga_dirty_update(struct mga_fbdev *mfbdev,
 	int bpp = (mfbdev->mfb.base.bits_per_pixel + 7)/8;
 	int ret;
 	bool unmap = false;
+	bool store_for_later = false;
+	int x2, y2;
+	unsigned long flags;
 
 	obj = mfbdev->mfb.obj;
 	bo = gem_to_mga_bo(obj);
 
+	/*
+	 * try and reserve the BO, if we fail with busy
+	 * then the BO is being moved and we should
+	 * store up the damage until later.
+	 */
 	ret = mgag200_bo_reserve(bo, true);
 	if (ret) {
-		DRM_ERROR("failed to reserve fb bo\n");
+		if (ret != -EBUSY)
+			return;
+
+		store_for_later = true;
+	}
+
+	x2 = x + width - 1;
+	y2 = y + height - 1;
+	spin_lock_irqsave(&mfbdev->dirty_lock, flags);
+
+	if (mfbdev->y1 < y)
+		y = mfbdev->y1;
+	if (mfbdev->y2 > y2)
+		y2 = mfbdev->y2;
+	if (mfbdev->x1 < x)
+		x = mfbdev->x1;
+	if (mfbdev->x2 > x2)
+		x2 = mfbdev->x2;
+
+	if (store_for_later) {
+		mfbdev->x1 = x;
+		mfbdev->x2 = x2;
+		mfbdev->y1 = y;
+		mfbdev->y2 = y2;
+		spin_unlock_irqrestore(&mfbdev->dirty_lock, flags);
 		return;
 	}
 
+	mfbdev->x1 = mfbdev->y1 = INT_MAX;
+	mfbdev->x2 = mfbdev->y2 = 0;
+	spin_unlock_irqrestore(&mfbdev->dirty_lock, flags);
+
 	if (!bo->kmap.virtual) {
 		ret = ttm_bo_kmap(&bo->bo, 0, bo->bo.num_pages, &bo->kmap);
 		if (ret) {
@@ -48,10 +84,10 @@ static void mga_dirty_update(struct mga_fbdev *mfbdev,
 		}
 		unmap = true;
 	}
-	for (i = y; i < y + height; i++) {
+	for (i = y; i <= y2; i++) {
 		/* assume equal stride for now */
 		src_offset = dst_offset = i * mfbdev->mfb.base.pitches[0] + (x * bpp);
-		memcpy_toio(bo->kmap.virtual + src_offset, mfbdev->sysram + src_offset, width * bpp);
+		memcpy_toio(bo->kmap.virtual + src_offset, mfbdev->sysram + src_offset, (x2 - x + 1) * bpp);
 
 	}
 	if (unmap)
@@ -252,6 +288,7 @@ int mgag200_fbdev_init(struct mga_device *mdev)
 
 	mdev->mfbdev = mfbdev;
 	mfbdev->helper.funcs = &mga_fb_helper_funcs;
+	spin_lock_init(&mfbdev->dirty_lock);
 
 	ret = drm_fb_helper_init(mdev->dev, &mfbdev->helper,
 				 mdev->num_crtc, MGAG200FB_CONN_LIMIT);
diff --git a/drivers/gpu/drm/mgag200/mgag200_ttm.c b/drivers/gpu/drm/mgag200/mgag200_ttm.c
index 8fc9d92..401c989 100644
--- a/drivers/gpu/drm/mgag200/mgag200_ttm.c
+++ b/drivers/gpu/drm/mgag200/mgag200_ttm.c
@@ -315,8 +315,8 @@ int mgag200_bo_reserve(struct mgag200_bo *bo, bool no_wait)
 
 	ret = ttm_bo_reserve(&bo->bo, true, no_wait, false, 0);
 	if (ret) {
-		if (ret != -ERESTARTSYS)
-			DRM_ERROR("reserve failed %p\n", bo);
+		if (ret != -ERESTARTSYS && ret != -EBUSY)
+			DRM_ERROR("reserve failed %p %d\n", bo, ret);
 		return ret;
 	}
 	return 0;
-- 
1.8.1.2

[PATCH 2/3] drm/ast: deal with bo reserve fail in dirty update path

[Dave Airlie]

From: Dave Airlie <airlied@redhat.com>

Port over the mgag200 fix to ast as it suffers the same issue.

    On F19 testing, it was noticed we get a lot of errors in dmesg
    about being unable to reserve the buffer when plymouth starts,
    this is due to the buffer being in the process of migrating,
    so it makes sense we can't reserve it.

    In order to deal with it, this adds delayed updates for the dirty
    updates, when the bo is unreservable, in the normal console case
    this shouldn't ever happen, its just when plymouth or X is
    pushing the console bo to system memory.

Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
---
 drivers/gpu/drm/ast/ast_drv.h |  2 ++
 drivers/gpu/drm/ast/ast_fb.c  | 43 ++++++++++++++++++++++++++++++++++++++++---
 drivers/gpu/drm/ast/ast_ttm.c |  2 +-
 3 files changed, 43 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/ast/ast_drv.h b/drivers/gpu/drm/ast/ast_drv.h
index 5284292..02e52d5 100644
--- a/drivers/gpu/drm/ast/ast_drv.h
+++ b/drivers/gpu/drm/ast/ast_drv.h
@@ -241,6 +241,8 @@ struct ast_fbdev {
 	void *sysram;
 	int size;
 	struct ttm_bo_kmap_obj mapping;
+	int x1, y1, x2, y2; /* dirty rect */
+	spinlock_t dirty_lock;
 };
 
 #define to_ast_crtc(x) container_of(x, struct ast_crtc, base)
diff --git a/drivers/gpu/drm/ast/ast_fb.c b/drivers/gpu/drm/ast/ast_fb.c
index 34931fe..fbc0823 100644
--- a/drivers/gpu/drm/ast/ast_fb.c
+++ b/drivers/gpu/drm/ast/ast_fb.c
@@ -53,16 +53,52 @@ static void ast_dirty_update(struct ast_fbdev *afbdev,
 	int bpp = (afbdev->afb.base.bits_per_pixel + 7)/8;
 	int ret;
 	bool unmap = false;
+	bool store_for_later = false;
+	int x2, y2;
+	unsigned long flags;
 
 	obj = afbdev->afb.obj;
 	bo = gem_to_ast_bo(obj);
 
+	/*
+	 * try and reserve the BO, if we fail with busy
+	 * then the BO is being moved and we should
+	 * store up the damage until later.
+	 */
 	ret = ast_bo_reserve(bo, true);
 	if (ret) {
-		DRM_ERROR("failed to reserve fb bo\n");
+		if (ret != -EBUSY)
+			return;
+
+		store_for_later = true;
+	}
+
+	x2 = x + width - 1;
+	y2 = y + height - 1;
+	spin_lock_irqsave(&afbdev->dirty_lock, flags);
+
+	if (afbdev->y1 < y)
+		y = afbdev->y1;
+	if (afbdev->y2 > y2)
+		y2 = afbdev->y2;
+	if (afbdev->x1 < x)
+		x = afbdev->x1;
+	if (afbdev->x2 > x2)
+		x2 = afbdev->x2;
+
+	if (store_for_later) {
+		afbdev->x1 = x;
+		afbdev->x2 = x2;
+		afbdev->y1 = y;
+		afbdev->y2 = y2;
+		spin_unlock_irqrestore(&afbdev->dirty_lock, flags);
 		return;
 	}
 
+	afbdev->x1 = afbdev->y1 = INT_MAX;
+	afbdev->x2 = afbdev->y2 = 0;
+	spin_unlock_irqrestore(&afbdev->dirty_lock, flags);
+
 	if (!bo->kmap.virtual) {
 		ret = ttm_bo_kmap(&bo->bo, 0, bo->bo.num_pages, &bo->kmap);
 		if (ret) {
@@ -72,10 +108,10 @@ static void ast_dirty_update(struct ast_fbdev *afbdev,
 		}
 		unmap = true;
 	}
-	for (i = y; i < y + height; i++) {
+	for (i = y; i <= y2; i++) {
 		/* assume equal stride for now */
 		src_offset = dst_offset = i * afbdev->afb.base.pitches[0] + (x * bpp);
-		memcpy_toio(bo->kmap.virtual + src_offset, afbdev->sysram + src_offset, width * bpp);
+		memcpy_toio(bo->kmap.virtual + src_offset, afbdev->sysram + src_offset, (x2 - x + 1) * bpp);
 
 	}
 	if (unmap)
@@ -292,6 +328,7 @@ int ast_fbdev_init(struct drm_device *dev)
 
 	ast->fbdev = afbdev;
 	afbdev->helper.funcs = &ast_fb_helper_funcs;
+	spin_lock_init(&afbdev->dirty_lock);
 	ret = drm_fb_helper_init(dev, &afbdev->helper,
 				 1, 1);
 	if (ret) {
diff --git a/drivers/gpu/drm/ast/ast_ttm.c b/drivers/gpu/drm/ast/ast_ttm.c
index 3602731..09da339 100644
--- a/drivers/gpu/drm/ast/ast_ttm.c
+++ b/drivers/gpu/drm/ast/ast_ttm.c
@@ -316,7 +316,7 @@ int ast_bo_reserve(struct ast_bo *bo, bool no_wait)
 
 	ret = ttm_bo_reserve(&bo->bo, true, no_wait, false, 0);
 	if (ret) {
-		if (ret != -ERESTARTSYS)
+		if (ret != -ERESTARTSYS && ret != -EBUSY)
 			DRM_ERROR("reserve failed %p\n", bo);
 		return ret;
 	}
-- 
1.8.1.2

[PATCH 3/3] drm/cirrus: deal with bo reserve fail in dirty update path

[Dave Airlie]

From: Dave Airlie <airlied@redhat.com>

Port over the mgag200 fix to cirrus as it suffers the same issue.

    On F19 testing, it was noticed we get a lot of errors in dmesg
    about being unable to reserve the buffer when plymouth starts,
    this is due to the buffer being in the process of migrating,
    so it makes sense we can't reserve it.

    In order to deal with it, this adds delayed updates for the dirty
    updates, when the bo is unreservable, in the normal console case
    this shouldn't ever happen, its just when plymouth or X is
    pushing the console bo to system memory.

Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
---
 drivers/gpu/drm/cirrus/cirrus_drv.h   |  2 ++
 drivers/gpu/drm/cirrus/cirrus_fbdev.c | 38 ++++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/cirrus/cirrus_ttm.c   |  2 +-
 3 files changed, 40 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/cirrus/cirrus_drv.h b/drivers/gpu/drm/cirrus/cirrus_drv.h
index 6e0cc72..7ca0595 100644
--- a/drivers/gpu/drm/cirrus/cirrus_drv.h
+++ b/drivers/gpu/drm/cirrus/cirrus_drv.h
@@ -154,6 +154,8 @@ struct cirrus_fbdev {
 	struct list_head fbdev_list;
 	void *sysram;
 	int size;
+	int x1, y1, x2, y2; /* dirty rect */
+	spinlock_t dirty_lock;
 };
 
 struct cirrus_bo {
diff --git a/drivers/gpu/drm/cirrus/cirrus_fbdev.c b/drivers/gpu/drm/cirrus/cirrus_fbdev.c
index e25afcc..3541b56 100644
--- a/drivers/gpu/drm/cirrus/cirrus_fbdev.c
+++ b/drivers/gpu/drm/cirrus/cirrus_fbdev.c
@@ -27,16 +27,51 @@ static void cirrus_dirty_update(struct cirrus_fbdev *afbdev,
 	int bpp = (afbdev->gfb.base.bits_per_pixel + 7)/8;
 	int ret;
 	bool unmap = false;
+	bool store_for_later = false;
+	int x2, y2;
+	unsigned long flags;
 
 	obj = afbdev->gfb.obj;
 	bo = gem_to_cirrus_bo(obj);
 
+	/*
+	 * try and reserve the BO, if we fail with busy
+	 * then the BO is being moved and we should
+	 * store up the damage until later.
+	 */
 	ret = cirrus_bo_reserve(bo, true);
 	if (ret) {
-		DRM_ERROR("failed to reserve fb bo\n");
+		if (ret != -EBUSY)
+			return;
+		store_for_later = true;
+	}
+
+	x2 = x + width - 1;
+	y2 = y + height - 1;
+	spin_lock_irqsave(&afbdev->dirty_lock, flags);
+
+	if (afbdev->y1 < y)
+		y = afbdev->y1;
+	if (afbdev->y2 > y2)
+		y2 = afbdev->y2;
+	if (afbdev->x1 < x)
+		x = afbdev->x1;
+	if (afbdev->x2 > x2)
+		x2 = afbdev->x2;
+
+	if (store_for_later) {
+		afbdev->x1 = x;
+		afbdev->x2 = x2;
+		afbdev->y1 = y;
+		afbdev->y2 = y2;
+		spin_unlock_irqrestore(&afbdev->dirty_lock, flags);
 		return;
 	}
 
+	afbdev->x1 = afbdev->y1 = INT_MAX;
+	afbdev->x2 = afbdev->y2 = 0;
+	spin_unlock_irqrestore(&afbdev->dirty_lock, flags);
+
 	if (!bo->kmap.virtual) {
 		ret = ttm_bo_kmap(&bo->bo, 0, bo->bo.num_pages, &bo->kmap);
 		if (ret) {
@@ -268,6 +303,7 @@ int cirrus_fbdev_init(struct cirrus_device *cdev)
 
 	cdev->mode_info.gfbdev = gfbdev;
 	gfbdev->helper.funcs = &cirrus_fb_helper_funcs;
+	spin_lock_init(&gfbdev->dirty_lock);
 
 	ret = drm_fb_helper_init(cdev->dev, &gfbdev->helper,
 				 cdev->num_crtc, CIRRUSFB_CONN_LIMIT);
diff --git a/drivers/gpu/drm/cirrus/cirrus_ttm.c b/drivers/gpu/drm/cirrus/cirrus_ttm.c
index 1413a26..2ed8cfc 100644
--- a/drivers/gpu/drm/cirrus/cirrus_ttm.c
+++ b/drivers/gpu/drm/cirrus/cirrus_ttm.c
@@ -321,7 +321,7 @@ int cirrus_bo_reserve(struct cirrus_bo *bo, bool no_wait)
 
 	ret = ttm_bo_reserve(&bo->bo, true, no_wait, false, 0);
 	if (ret) {
-		if (ret != -ERESTARTSYS)
+		if (ret != -ERESTARTSYS && ret != -EBUSY)
 			DRM_ERROR("reserve failed %p\n", bo);
 		return ret;
 	}
-- 
1.8.1.2