Skip to contentSkip to conversation

vfio: fix use-after-free in display

PATCH In review Gerd Hoffmann 3 participants 6 messages Jul 13, 2020 – Jul 15, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vfio: fix use-after-free in display

Calling ramfb_display_update() might replace the DisplaySurface with the boot display, which in turn will free the currently active DisplaySurface.

So clear our DisplaySurface pinter (dpy->region.surface pointer) to (a) avoid use-after-free and (b) force replacing the boot display with the real display when switching back.

---
 hw/vfio/display.c | 1 +
 1 file changed, 1 insertion(+)
hw/vfio/display.c +1 -0
index a57a22674d62..342054193b3c 100644
@@ -405,6 +405,7 @@ static void vfio_display_region_update(void *opaque)
     if (!plane.drm_format || !plane.size) {
         if (dpy->ramfb) {
             ramfb_display_update(dpy->con, dpy->ramfb);
+            dpy->region.surface = NULL;
         }
         return;
     }
Signature · 1 line
2.18.4
#1 lore raw
Quoted 19 lines · Gerd Hoffmann

Calling ramfb_display_update() might replace the DisplaySurface with the boot display, which in turn will free the currently active DisplaySurface.

So clear our DisplaySurface pinter (dpy->region.surface pointer) to (a) avoid use-after-free and (b) force replacing the boot display with the real display when switching back.

---
 hw/vfio/display.c | 1 +
 1 file changed, 1 insertion(+)
hw/vfio/display.c +0 -0
index a57a22674d62..342054193b3c 100644

hw/vfio/display.c static void vfio_display_region_update(void *opaque) · line 411

Quoted 7 lines
hw/vfio/display.c +1 -0
@@ -405,6 +405,7 @@ static void vfio_display_region_update(void *opaque)
     if (!plane.drm_format || !plane.size) {
         if (dpy->ramfb) {
             ramfb_display_update(dpy->con, dpy->ramfb);
+            dpy->region.surface = NULL;
         }
         return;
     }

More generic fix:

-- >8 --

ui/console.c +4 -4
@@ -1580,10 +1580,10 @@ void dpy_gfx_replace_surface(QemuConsole *con,
                              DisplaySurface *surface)
 {
     DisplayState *s = con->ds;
-    DisplaySurface *old_surface = con->surface;
+    QemuConsole *old_con = con;
     DisplayChangeListener *dcl;
 
-    assert(old_surface != surface || surface == NULL);
+    assert(con->surface != surface || surface == NULL);
 
     con->surface = surface;
     QLIST_FOREACH(dcl, &s->listeners, next) {
@@ -1594,7 +1594,8 @@ void dpy_gfx_replace_surface(QemuConsole *con,
             dcl->ops->dpy_gfx_switch(dcl, surface);
         }
     }
-    qemu_free_displaysurface(old_surface);
+    qemu_free_displaysurface(old_con->surface);
+    old_con->surface = NULL;
 }
---
 
#2 lore raw
Quoted 47 lines · Philippe Mathieu-Daudé
Quoted 19 lines · Gerd Hoffmann

Calling ramfb_display_update() might replace the DisplaySurface with the boot display, which in turn will free the currently active DisplaySurface.

So clear our DisplaySurface pinter (dpy->region.surface pointer) to (a) avoid use-after-free and (b) force replacing the boot display with the real display when switching back.

---
 hw/vfio/display.c | 1 +
 1 file changed, 1 insertion(+)
hw/vfio/display.c +0 -0
index a57a22674d62..342054193b3c 100644

hw/vfio/display.c static void vfio_display_region_update(void *opaque) · line 411

Quoted 7 lines
hw/vfio/display.c +1 -0
@@ -405,6 +405,7 @@ static void vfio_display_region_update(void *opaque)
     if (!plane.drm_format || !plane.size) {
         if (dpy->ramfb) {
             ramfb_display_update(dpy->con, dpy->ramfb);
+            dpy->region.surface = NULL;
         }
         return;
     }

More generic fix:

-- >8 --

ui/console.c +2 -2
@@ -1580,10 +1580,10 @@ void dpy_gfx_replace_surface(QemuConsole *con,
                              DisplaySurface *surface)
 {
     DisplayState *s = con->ds;
-    DisplaySurface *old_surface = con->surface;
+    QemuConsole *old_con = con;
     DisplayChangeListener *dcl;
 
-    assert(old_surface != surface || surface == NULL);
+    assert(con->surface != surface || surface == NULL);
 
     con->surface = surface;
     QLIST_FOREACH(dcl, &s->listeners, next) {

ui/console.c void dpy_gfx_replace_surface(QemuConsole *con, · line 1598

Quoted 6 lines
ui/console.c +2 -1
@@ -1594,7 +1594,8 @@ void dpy_gfx_replace_surface(QemuConsole *con,
             dcl->ops->dpy_gfx_switch(dcl, surface);
         }
     }
-    qemu_free_displaysurface(old_surface);
+    qemu_free_displaysurface(old_con->surface);
+    old_con->surface = NULL;

No.

That doesn't clear VFIODisplay->region.surface, but it sets QemuConsole->surface to NULL no matter what got passed to dpy_gfx_replace_surface().

Guesswork based just on the patch chunk doesn't always work, sometimes you have to consult the source code to see what the patch actually does ;)

take care,
Gerd

#3 lore raw
Quoted 57 lines · Gerd Hoffmann
Quoted 47 lines · Philippe Mathieu-Daudé
Quoted 19 lines · Gerd Hoffmann

Calling ramfb_display_update() might replace the DisplaySurface with the boot display, which in turn will free the currently active DisplaySurface.

So clear our DisplaySurface pinter (dpy->region.surface pointer) to (a) avoid use-after-free and (b) force replacing the boot display with the real display when switching back.

---
 hw/vfio/display.c | 1 +
 1 file changed, 1 insertion(+)
hw/vfio/display.c +0 -0
index a57a22674d62..342054193b3c 100644

hw/vfio/display.c static void vfio_display_region_update(void *opaque) · line 411

Quoted 7 lines
hw/vfio/display.c +1 -0
@@ -405,6 +405,7 @@ static void vfio_display_region_update(void *opaque)
     if (!plane.drm_format || !plane.size) {
         if (dpy->ramfb) {
             ramfb_display_update(dpy->con, dpy->ramfb);
+            dpy->region.surface = NULL;
         }
         return;
     }

More generic fix:

-- >8 --

ui/console.c +2 -2
@@ -1580,10 +1580,10 @@ void dpy_gfx_replace_surface(QemuConsole *con,
                              DisplaySurface *surface)
 {
     DisplayState *s = con->ds;
-    DisplaySurface *old_surface = con->surface;
+    QemuConsole *old_con = con;
     DisplayChangeListener *dcl;
 
-    assert(old_surface != surface || surface == NULL);
+    assert(con->surface != surface || surface == NULL);
 
     con->surface = surface;
     QLIST_FOREACH(dcl, &s->listeners, next) {

ui/console.c void dpy_gfx_replace_surface(QemuConsole *con, · line 1598

Quoted 6 lines
ui/console.c +2 -1
@@ -1594,7 +1594,8 @@ void dpy_gfx_replace_surface(QemuConsole *con,
             dcl->ops->dpy_gfx_switch(dcl, surface);
         }
     }
-    qemu_free_displaysurface(old_surface);
+    qemu_free_displaysurface(old_con->surface);
+    old_con->surface = NULL;

No.

Quoted 7 lines

That doesn't clear VFIODisplay->region.surface, but it sets QemuConsole->surface to NULL no matter what got passed to dpy_gfx_replace_surface().

Guesswork based just on the patch chunk doesn't always work, sometimes you have to consult the source code to see what the patch actually does ;)

I'm certainly not a source of trust producing perfect code :) I appreciate being reviewed on snippets, so I can learn from my mistakes and improve. I looked at the code (ui/console.c was not in the previous context) but I missed to look at the other call sites for dpy_gfx_replace_surface() and how it is used.

Thanks,

Phil.

Quoted 4 lines

take care,
Gerd

Acked-by Reviewed-by on v1 1/1
#4 lore raw
In reply to Gerd Hoffmann
Quoted 5 lines · On Mon, 13 Jul 2020 14:45:20 +0200 Gerd Hoffmann <[email protected]> wrote:

Calling ramfb_display_update() might replace the DisplaySurface with the boot display, which in turn will free the currently active DisplaySurface.

So clear our DisplaySurface pinter (dpy->region.surface pointer) to (a)

                               ^o
Quoted 13 lines

avoid use-after-free and (b) force replacing the boot display with the real display when switching back.

---
 hw/vfio/display.c | 1 +
 1 file changed, 1 insertion(+)
hw/vfio/display.c +0 -0
index a57a22674d62..342054193b3c 100644

hw/vfio/display.c static void vfio_display_region_update(void *opaque) · line 411

Quoted 7 lines
hw/vfio/display.c +1 -0
@@ -405,6 +405,7 @@ static void vfio_display_region_update(void *opaque)
     if (!plane.drm_format || !plane.size) {
         if (dpy->ramfb) {
             ramfb_display_update(dpy->con, dpy->ramfb);
+            dpy->region.surface = NULL;
         }
         return;
     }

Tricky, but I think I follow that dpy->region.surface is only ever allocated to replace dpy->con->surface, so when ramfb_display_update() then replaces and frees dpy->con->surface with dpy->ramfb->ds, that's where the object point to by dpy->region.surface was freed. Right?

If so, looks ok to me. If you're constructing a pull request, I'll give you an:

If you need me to send a pull, let me know. Thanks,

Alex

#5 lore raw

Hi,

Quoted 4 lines

Tricky, but I think I follow that dpy->region.surface is only ever allocated to replace dpy->con->surface, so when ramfb_display_update() then replaces and frees dpy->con->surface with dpy->ramfb->ds, that's where the object point to by dpy->region.surface was freed. Right?

Correct.

Quoted 7 lines

If so, looks ok to me. If you're constructing a pull request, I'll give you an:

If you need me to send a pull, let me know.

I'll go create a pull req, I have one or two other patches pending anyway.

take care,
Gerd