[Mesa-dev] [PATCH] st/dri: Fix dangling pointer to a destroyed dri_drawable

[Johan Helsing]

Emil: Your alternative patch won't work because dri_make_current is not necessarily called with NULL after a buffer has been destroyed.


The problematic sequence is a pattern we use in QtWayland:


//create temporary context

surface1 = eglCreateWindowSurface() <-- dri_drawable pointer is malloced

eglMakeCurrent(surface1) <-- ctx->dPriv is set


// ... (Get some information about available GL extensions etc)


eglDestroySurface(surface1) <-- pointer is freed, ctx->dPriv is now dangling

surface2 = eglCreateWindowSurface() <-- Creating a new surface. Sometimes it's address will be the same as the free'd pointer.

eglMakeCurrent(surface2) <-- In dri_make_current, ctx->dPriv == driReadPriv may return true because the pointers may be equal

      => The drawable info is not updated. Width and height for the drawable is not set from the wl_egl_window on the first frame.


Marek: How exactly does it crash? Are you sure firefox didn't previously access free'd memory through the dangling pointer and that it was just exposed now that the pointer is NULL?


Johan



________________________________
From: Marek Olšák <maraeo at gmail.com>
Sent: Tuesday, April 24, 2018 6:08:16 AM
To: Johan Helsing
Cc: ML Mesa-dev; pekka.paalanen at collabora.co.uk; Daniel Stone; Charmaine Lee; Thomas Hellstrom; Michel Dänzer
Subject: Re: [PATCH] st/dri: Fix dangling pointer to a destroyed dri_drawable

FYI, the commit was causing crashes of qtcreator and firefox, so I reverted it.

Marek

On Fri, Apr 20, 2018 at 6:29 AM, Johan Klokkhammer Helsing <johan.helsing at qt.io<mailto:johan.helsing at qt.io>> wrote:
If an EGLSurface is created, made current and destroyed, and then a second
EGLSurface is created. Then the second malloc in driCreateNewDrawable may
return the same pointer address the first surface's drawable had.
Consequently, when dri_make_current later tries to determine if it should
update the texture_stamp it compares the surface's drawable pointer against
the drawable in the last call to dri_make_current and assumes it's the same
surface (which it isn't).

When texture_stamp is left unset, then dri_st_framebuffer_validate thinks
it has already called update_drawable_info for that drawable, leaving it
unvalidated and this is when bad things starts to happen. In my case it
manifested itself by the width and height of the surface being unset.

This is fixed this by setting the pointer to NULL before freeing the
surface.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=106126
Signed-off-by: Johan Klokkhammer Helsing <johan.helsing at qt.io<mailto:johan.helsing at qt.io>>
---
 src/gallium/state_trackers/dri/dri_drawable.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/gallium/state_trackers/dri/dri_drawable.c b/src/gallium/state_trackers/dri/dri_drawable.c
index e5a7537e47..02328acd98 100644
--- a/src/gallium/state_trackers/dri/dri_drawable.c
+++ b/src/gallium/state_trackers/dri/dri_drawable.c
@@ -185,6 +185,7 @@ fail:
 void
 dri_destroy_buffer(__DRIdrawable * dPriv)
 {
+   struct dri_context *ctx = dri_context(dPriv->driContextPriv);
    struct dri_drawable *drawable = dri_drawable(dPriv);
    struct dri_screen *screen = drawable->screen;
    struct st_api *stapi = screen->st_api;
@@ -202,6 +203,9 @@ dri_destroy_buffer(__DRIdrawable * dPriv)
    /* Notify the st manager that this drawable is no longer valid */
    stapi->destroy_drawable(stapi, &drawable->base);

+   if (ctx && ctx->dPriv == dPriv)
+      ctx->dPriv = NULL;
+
    FREE(drawable);
 }

--
2.17.0


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180424/e8ed5c9b/attachment.html>