drm/msm: return fence_fd = -1 if gem_submit fails

Commit Message

Gustavo Padovan Dec. 12, 2016, 7:41 p.m. UTC

From: Gustavo Padovan <[e-mail]>

Previously we were returning garbage here, fix it by setting it to -1
before the first possible point of failure.

Signed-off-by: Gustavo Padovan <[e-mail]>
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Chris Wilson Dec. 12, 2016, 8:42 p.m. UTC | #1

On Mon, Dec 12, 2016 at 05:41:08PM -0200, Gustavo Padovan wrote:
> From: Gustavo Padovan <[e-mail]>
> 
> Previously we were returning garbage here, fix it by setting it to -1
> before the first possible point of failure.

The convention is that on error paths you do not modify user inputs. In
particular, consider EINTR where the usual pattern (e.g. drmIoctl) is

	do {
		err = ioctl(fd, SUBMIT, arg);
	} while (err == -EINTR);

If you modify the in fence before you consume it, you can't recreate it
after handling the signal.
-Chris

Gustavo Padovan Dec. 12, 2016, 9:23 p.m. UTC | #2

2016-12-12 Chris Wilson <[e-mail]>:

> On Mon, Dec 12, 2016 at 05:41:08PM -0200, Gustavo Padovan wrote:
> > From: Gustavo Padovan <[e-mail]>
> > 
> > Previously we were returning garbage here, fix it by setting it to -1
> > before the first possible point of failure.
> 
> The convention is that on error paths you do not modify user inputs. In
> particular, consider EINTR where the usual pattern (e.g. drmIoctl) is
> 
> 	do {
> 		err = ioctl(fd, SUBMIT, arg);
> 	} while (err == -EINTR);
> 
> If you modify the in fence before you consume it, you can't recreate it
> after handling the signal.

Right. I didn't know about that convention. So maybe we let it as is. :)

Gustavo

Patch

diff --git a/drivers/gpu/drm/msm/msm_gem_submit.c b/drivers/gpu/drm/msm/msm_gem_submit.c
index 166e84e..c102b55 100644
--- a/drivers/gpu/drm/msm/msm_gem_submit.c
+++ b/drivers/gpu/drm/msm/msm_gem_submit.c
@@ -383,10 +383,13 @@  int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 	struct msm_gpu *gpu = priv->gpu;
 	struct dma_fence *in_fence = NULL;
 	struct sync_file *sync_file = NULL;
+	int in_fence_fd = args->fence_fd;
 	int out_fence_fd = -1;
 	unsigned i;
 	int ret;
 
+	args->fence_fd = -1;
+
 	if (!gpu)
 		return -ENXIO;
 
@@ -427,7 +430,7 @@  int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 		goto out;
 
 	if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
-		in_fence = sync_file_get_fence(args->fence_fd);
+		in_fence = sync_file_get_fence(in_fence_fd);
 
 		if (!in_fence) {
 			ret = -EINVAL;