mrbpix

Memory leak in fglrx.ko (regression in Catalyst 11.2)

Discussion created by mrbpix on Apr 23, 2011
Latest reply on Jul 22, 2011 by mrbpix

While upgrading a cluster of 64-bit Linux machines from Catalyst 10.12 to 11.3, all running my CAL (not OpenCL) whitepixel application, I found a regression that I later tracked down was introduced in the intermediary Catalyst 11.2 version: fglrx.ko leaks memory at a rate of about 10-30 MB per minute as reported by free(1). I know it is the fglrx.ko module because the resident memory used by X11 and my app is pretty much constant as seen in top(1).

My machines have 2GB RAM. The leak causes the amount of memory used by fglrx.ko to grow up to about 1.5GB after a few hours. The behavior is then erratic: either the application SIGSEGV, or hangs, or more surprisingly some machines appear to continue running fine after plateau'ing at 1.5GB (but the compute shaders could be behaving incorrectly: threads returning prematurely, etc - I need to check that).

AMD engineers: it should be simple for you to reproduce. whitepixel is an open source app. Download it and run it as per the README. The app compiles the CAL kernel once, and then calls calCtxRunProgramGrid() and calIsEventDone() indefinitely in a loop to process work items. Please advise. In the mean time I have reverted to 11.1, which was the last non-leaky driver.

Outcomes