Do not loop on ERESTARTSYS using interruptible waits

This problem happens using KMS surfaces and QXL driver.
To easy reproduce use KDE Plasma (which use surfaces a lot) and assure you are using KMS surfaces (QXL driver on Fedora/RedHat has a patch to stop using them). Open some complex application like LibreOffice and after a while your machine get stuck using 100% CPU on Xorg. The problem occurs as creating new surfaces not interruptible wait are used however instead of returning ERESTARTSYS back to userspace you try to loop but wait routines always keep returning ERESTARTSYS once the signal is marked.
On out of memory conditions TTM module try to move objects to system memory and QXL assure surface is updated before the move. The fix handle differently this case using no interruptible wait so wait functions will wait instead of returning ERESTARTSYS. Note the when the loop occurs driver will send a lot of update requests causing more CPU usage on Qemu side too.

Signed-off-by Frediano Ziglio <[email protected]>

---
 qxl/qxl_cmd.c   | 12 +++---------
 qxl/qxl_drv.h   |  2 +-
 qxl/qxl_ioctl.c |  2 +-
 3 files changed, 5 insertions(+), 11 deletions(-)

drivers/gpu/drm/qxl/qxl_cmd.c +0 -0

index 9782364..bd5404e 100644

drivers/gpu/drm/qxl/qxl_cmd.c +2 -2

 }
 
 int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,
-			const struct qxl_rect *area)
+			const struct qxl_rect *area, bool intr)
 {
 	int surface_id;
 	uint32_t surface_width, surface_height;

@@ -350,7 +347,7 @@ int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,
 	mutex_lock(&qdev->update_area_mutex);
 	qdev->ram_header->update_area = *area;
 	qdev->ram_header->update_surface = surface_id;
-	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, true);
+	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, intr);
 	mutex_unlock(&qdev->update_area_mutex);
 	return ret;
 }

Hey,

Quoted 32 lines · Frediano Ziglio

This problem happens using KMS surfaces and QXL driver.
To easy reproduce use KDE Plasma (which use surfaces a lot) and assure you are using KMS surfaces (QXL driver on Fedora/RedHat has a patch to stop using them). Open some complex application like LibreOffice and after a while your machine get stuck using 100% CPU on Xorg. The problem occurs as creating new surfaces not interruptible wait are used however instead of returning ERESTARTSYS back to userspace you try to loop but wait routines always keep returning ERESTARTSYS once the signal is marked.
On out of memory conditions TTM module try to move objects to system memory and QXL assure surface is updated before the move. The fix handle differently this case using no interruptible wait so wait functions will wait instead of returning ERESTARTSYS. Note the when the loop occurs driver will send a lot of update requests causing more CPU usage on Qemu side too.

Signed-off-by Frediano Ziglio <[email protected]>

---
 qxl/qxl_cmd.c   | 12 +++---------
 qxl/qxl_drv.h   |  2 +-
 qxl/qxl_ioctl.c |  2 +-
 3 files changed, 5 insertions(+), 11 deletions(-)

drivers/gpu/drm/qxl/qxl_cmd.c +0 -0

index 9782364..bd5404e 100644
@@ -317,14 +317,11 @@ static void wait_for_io_cmd(struct qxl_device *qdev,
 uint8_t val, long port)
 {
 	int ret;
 

drivers/gpu/drm/qxl/qxl_cmd.c static void wait_for_io_cmd(struct qxl_device *qdev, uint8_t val, long port) · line 323

Quoted 4 lines

drivers/gpu/drm/qxl/qxl_cmd.c +0 -3

-restart:
 	ret = wait_for_io_cmd_user(qdev, val, port, false);
-	if (ret == -ERESTARTSYS)
-		goto restart;

I think this one is not directly related to the fix, but can be removed because wait_for_io_cmd_user(qdev, val, port, false); will call wait_event_timeout() which cannot return ERESTARTSYS? Or was this loop causing issues too?

Quoted 19 lines

drivers/gpu/drm/qxl/qxl_cmd.c +1 -1

 }
 
 int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,
-			const struct qxl_rect *area)
+			const struct qxl_rect *area, bool intr)
 {
 	int surface_id;

drivers/gpu/drm/qxl/qxl_cmd.c +1 -1

 	uint32_t surface_width, surface_height;

@@ -350,7 +347,7 @@ int qxl_io_update_area(struct qxl_device *qdev, struct
 qxl_bo *surf,
 	mutex_lock(&qdev->update_area_mutex);
 	qdev->ram_header->update_area = *area;
 	qdev->ram_header->update_surface = surface_id;
-	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, true);
+	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, intr);
 	mutex_unlock(&qdev->update_area_mutex);
 	return ret;

}

drivers/gpu/drm/qxl/qxl_cmd.c int qxl_update_surface(struct qxl_device *qdev, struct · line 589

Quoted 9 lines

drivers/gpu/drm/qxl/qxl_cmd.c +1 -4

@@ -588,10 +585,7 @@ int qxl_update_surface(struct qxl_device *qdev, struct
 qxl_bo *surf)
 	rect.right = surf->surf.width;
 	rect.top = 0;
 	rect.bottom = surf->surf.height;
-retry:
-	ret = qxl_io_update_area(qdev, surf, &rect);
-	if (ret == -ERESTARTSYS)
-		goto retry;
+	ret = qxl_io_update_area(qdev, surf, &rect, false);

My understanding is that the fix is this hunk? If so, this could be made more obvious with an intermediate commit adding the 'bool intr' arg to qxl_io_update_area and only calling it with 'true' in the appropriate places.
This code path is only triggered from qxl_surface_evict() which I assume is not necessarily easily interruptible, so this change makes sense to me. However it would be much better to get a review from Dave Airlie ;)

Christophe

Quoted 37 lines

Quoted 19 lines

}

int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,

drivers/gpu/drm/qxl/qxl_cmd.c +2 -2

-			const struct qxl_rect *area)
+			const struct qxl_rect *area, bool intr)
 {
 	int surface_id;
 	uint32_t surface_width, surface_height;

@@ -350,7 +347,7 @@ int qxl_io_update_area(struct qxl_device *qdev, struct
 qxl_bo *surf,
 	mutex_lock(&qdev->update_area_mutex);
 	qdev->ram_header->update_area = *area;
 	qdev->ram_header->update_surface = surface_id;
-	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, true);
+	ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, intr);
 	mutex_unlock(&qdev->update_area_mutex);
 	return ret;

}

drivers/gpu/drm/qxl/qxl_cmd.c int qxl_update_surface(struct qxl_device *qdev, struct · line 589

Quoted 9 lines

drivers/gpu/drm/qxl/qxl_cmd.c +1 -4

@@ -588,10 +585,7 @@ int qxl_update_surface(struct qxl_device *qdev, struct
 qxl_bo *surf)
 	rect.right = surf->surf.width;
 	rect.top = 0;
 	rect.bottom = surf->surf.height;
-retry:
-	ret = qxl_io_update_area(qdev, surf, &rect);
-	if (ret == -ERESTARTSYS)
-		goto retry;
+	ret = qxl_io_update_area(qdev, surf, &rect, false);

My understanding is that the fix is this hunk? If so, this could be made more obvious with an intermediate commit adding the 'bool intr' arg to qxl_io_update_area and only calling it with 'true' in the appropriate places.
This code path is only triggered from qxl_surface_evict() which I assume is not necessarily easily interruptible, so this change makes sense to me. However it would be much better to get a review from Dave Airlie ;)

Quoted 2 lines

Christophe

Are you asking if just removing the loop would fix the issue? So you are proposing a first patch that add the argument always passing true and another that change some calls to false? It make sense but still the loop should be removed.

This problem happens using KMS surfaces and QXL driver.
To easy reproduce use KDE Plasma (which use surfaces a lot) and assure you are using KMS surfaces (QXL driver on Fedora/RedHat has a patch to stop using them). Open some complex application like LibreOffice and after a while your machine get stuck using 100% CPU on Xorg. The problem occurs as creating new surfaces not interruptible wait are used however instead of returning ERESTARTSYS back to userspace

you try to loop but wait routines always keep returning ERESTARTSYS once the signal is marked.
On out of memory conditions TTM module try to move objects to system memory and QXL assure surface is updated before the move. The fix handle differently this case using no interruptible wait so wait functions will wait instead of returning ERESTARTSYS. Note the when the loop occurs driver will send a lot of update requests causing more CPU usage on Qemu side too.

Signed-off-by Frediano Ziglio <[email protected]>

---
 qxl/qxl_cmd.c   | 12 +++---------
 qxl/qxl_drv.h   |  2 +-
 qxl/qxl_ioctl.c |  2 +-
 3 files changed, 5 insertions(+), 11 deletions(-)

drivers/gpu/drm/qxl/qxl_cmd.c +3 -9

index 9782364..bd5404e 100644
@@ -317,14 +317,11 @@ static void wait_for_io_cmd(struct qxl_device *qdev, uint8_t val, long port)
 {
        int ret;
 
-restart:
        ret = wait_for_io_cmd_user(qdev, val, port, false);
-       if (ret == -ERESTARTSYS)
-               goto restart;
 }
 
 int qxl_io_update_area(struct qxl_ddevice *qdev, struct qxl_bo *surf,
-                       const struct qxl_rect *area)
+                       const struct qxl_rect *area, bool intr)
 {
        int surface_id;
        uint32_t surface_width, surface_height;

@@ -350,7 +347,7 @@ int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,
        mutex_lock(&qdev->update_area_mutex);
        qdev->ram_header->update_area = *area;
        qdev->ram_header->update_surface = surface_id;
-       ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, true);
+       ret = wait_for_io_cmd_user(qdev, 0, QXL_IO_UPDATE_AREA_ASYNC, intr);
        mutex_unlock(&qdev->update_area_mutex);
        return ret;
 }

@@ -588,10 +585,7 @@ int qxl_update_surface(struct qxl_device *qdev, struct qxl_bo *surf)
        rect.right = surf->surf.width;
        rect.top = 0;
        rect.bottom = surf->surf.height;
-retry:
-       ret = qxl_io_update_area(qdev, surf, &rect);
-       if (ret == -ERESTARTSYS)
-               goto retry;
+       ret = qxl_io_update_area(qdev, surf, &rect, false);
        return ret;
 }
 

drivers/gpu/drm/qxl/qxl_drv.h +1 -1

index 7c6cafe..6745c44 100644
@@ -462,7 +462,7 @@ void qxl_io_memslot_add(struct qxl_device *qdev, uint8_t id);
 void qxl_io_notify_oom(struct qxl_device *qdev);
 
 int qxl_io_update_area(struct qxl_device *qdev, struct qxl_bo *surf,
-                      const struct qxl_rect *area);
+                      const struct qxl_rect *area, bool intr);
 
 void qxl_io_reset(struct qxl_device *qdev);
 void qxl_io_monitors_config(struct qxl_device *qdev);

drivers/gpu/drm/qxl/qxl_ioctl.c +1 -2

index b110883..afd7297 100644
@@ -348,7 +348,7 @@ static int qxl_update_area_ioctl(struct drm_device *dev, void *data,
                goto out2;
        if (!qobj->surface_id)
                DRM_ERROR("got update area for surface with no id %d\n", update_area->handle);
-       ret = qxl_io_update_area(qdev, qobj, &area);
+       ret = qxl_io_update_area(qdev, qobj, &area, true);
 
 out2:
        qxl_bo_unreserve(qobj);
--

2.1.0
_______________________________________________
Spice-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/spice-devel