xf86drm: warn about GEM handle reference counting

Users need to be careful when using drmPrimeHandleToFD or
drmPrimeFDToHandle directly. Mention GBM as a solution.

See [1] for an example mistake.

[1]: https://gitlab.freedesktop.org/drm/nouveau/-/issues/43#note_772661

Signed-off-by: Simon Ser <contact@emersion.fr>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

1 files changed, 18 insertions, 0 deletions

diff --git a/xf86drm.h b/xf86drm.h
index 7b85079a..9fc06ab8 100644
--- a/xf86drm.h
+++ b/xf86drm.h
@@ -813,6 +813,24 @@ extern char *drmGetDeviceNameFromFd(int fd);
 extern char *drmGetDeviceNameFromFd2(int fd);
 extern int drmGetNodeTypeFromFd(int fd);
 
+/* Convert between GEM handles and DMA-BUF file descriptors.
+ *
+ * Warning: since GEM handles are not reference-counted and are unique per
+ * DRM file description, the caller is expected to perform its own reference
+ * counting. drmPrimeFDToHandle is guaranteed to return the same handle for
+ * different FDs if they reference the same underlying buffer object. This
+ * could even be a buffer object originally created on the same DRM FD.
+ *
+ * When sharing a DRM FD with an API such as EGL or GBM, the caller must not
+ * use drmPrimeHandleToFD nor drmPrimeFDToHandle. A single user-space
+ * reference-counting implementation is necessary to avoid double-closing GEM
+ * handles.
+ *
+ * Two processes can't share the same DRM FD and both use it to create or
+ * import GEM handles, even when using a single user-space reference-counting
+ * implementation like GBM, because GBM doesn't share its state between
+ * processes.
+ */
 extern int drmPrimeHandleToFD(int fd, uint32_t handle, uint32_t flags, int *prime_fd);
 extern int drmPrimeFDToHandle(int fd, int prime_fd, uint32_t *handle);