The IFramebuffer::map() function is not thread-safe, which is why the
threaded implementation of draw_test_pattern_impl() maps all planes
before starting to draw. A typo slipped in the code, resulting in only
plane 0 being mapped. This didn't result in an immediate segfault, as
drawing operations in the worker threads map the remaining planes.
However, due to the implementation of DumbFramebuffer::map(), this can
result in the same plane being mapped multiple times, with only one of
the mapping recorded in the mapping cache. The other mappings are then
leaked, leading not only to extra memory consumption, but also to the
DRM device never being released even after the destruction of the Card
object.
Fix this.
Fixes: 40d96062a37c ("Revert "testpat: remove threaded drawing"")
Signed-off-by: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
---
I've noticed this issue when running the DU KMS tests, when the tests
run right after multi-planar format tests would sometimes fail, while
running fine in isolation. It was "fun" to debug, for some definition of
"fun".
---
kms++util/src/testpat.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kms++util/src/testpat.cpp b/kms++util/src/testpat.cpp
index 78c9d19a5ad7..1102588cc9f7 100644
--- a/kms++util/src/testpat.cpp
+++ b/kms++util/src/testpat.cpp
@@ -173,7 +173,7 @@ static void draw_test_pattern_impl(IFramebuffer& fb, YUVType yuvt)
// Create the mmaps before starting the threads
for (unsigned i = 0; i < fb.num_planes(); ++i)
- fb.map(0);
+ fb.map(i);
unsigned num_threads = thread::hardware_concurrency();
vector<thread> workers;
base-commit: c23e7548ee317c043660f9b992388257e99f1776
--
Regards,
Laurent Pinchart
