Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> writes:
> On Mon, Dec 02, 2013 at 04:47:46PM +0200, Mika Kuoppala wrote:
>> Fork and create another filedesc to wait access to already
>> hung GPU and then kill it. This triggers use after free of the
>> request->batch_obj.
>>
>> Signed-off-by: Mika Kuoppala <mika.kuoppala@xxxxxxxxx>
>> ---
>> +static int __test_open_any(void)
>> +{
>> + char *name;
>> + int ret, fd;
>> +
>> + ret = asprintf(&name, "/dev/dri/card%d", drm_get_card());
>> + if (ret == -1)
>> + return -1;
>> +
>> + fd = open(name, O_RDWR);
>> + free(name);
>> +
>> + return fd;
>> +}
>> +
>> +static void test_close_pending_fork(void)
>> +{
>> + int pid;
>> + int fd, fd2, h;
>> +
>> + fd = drm_open_any();
>> + igt_assert(fd >= 0);
>
> [snip]
>
>> + close(fd);
>> +
>> + fd = drm_open_any();
>
> This breaks the testsuite since drm_open_any() will now return the
> closed fd. Hence your __test_open_any() above.
>
> However this does not justify kernel doing anything other than terminating
> your process with extreme prejudice.
I discussed this with Chris in IRC. There seems to be no problem
as drm_open_any doesn't cache file descriptors.
The reason i use __test_open_any is that I needed a way to bypass
testsuites exit handling.
--Mika
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
