On Mon, 27 Jul 2015 21:35:20 +1200, Karl Tomlinson wrote: > Sometimes it would be nice to check in crashtests that use, or > attempt to use large memory allocations, but I'm concerned that > checking in these crashtests could disrupt subsequent tests > because there is then not enough memory to test what they want to > test.
Following up here, mainly just to report back on findings, for anyone considering the same in the future. I ended up backing out the test that I wanted to add [1] and I think the main blocker is that at least some platforms overcommit memory. Large allocations can succeed, but the lack of memory is not detected until the memory is actually used. When the memory is used, a process is killed, and AFAIK there is no guarantee that it is the process using the new memory. Even if we configured platforms to not overcommit, I'd still be uncomfortable adding a test that may force everything to swap to disk. It seems that fallible memory allocation is only a complete solution when constrained by address space, but perhaps we could run our tests with process data segment size constraints tuned for the hardware environment. The dummy unload event listener is helpful to avoid keeping things alive in BF cache, but AFAIK SimpleTest.forceGC/CC() (or similar DOMWindowUtils methods) and carefully removing all references is the only way to ensure that memory is released immediately. [1] https://bugzilla.mozilla.org/show_bug.cgi?id=999376#c12 _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
