On 11/10/18 11:59 AM, Catalin Marinas wrote: > On Sat, Nov 10, 2018 at 10:08:10AM -0500, Qian Cai wrote: >> On Nov 8, 2018, at 4:23 PM, Qian Cai <c...@gmx.us> wrote: >>> The maximum value for DEBUG_KMEMLEAK_EARLY_LOG_SIZE is only 40000, so it >>> disables kmemleak every time on this aarch64 server running the latest >>> mainline >>> (b00d209). >>> >>> # echo scan > /sys/kernel/debug/kmemleak >>> -bash: echo: write error: Device or resource busy >>> >>> Any idea on how to enable kmemleak there? >> >> I have managed to hard-code DEBUG_KMEMLEAK_EARLY_LOG_SIZE to 600000, > > That's quite a high number, I wouldn't have thought it is needed. > Basically the early log buffer is only used until the slub allocator > gets initialised and kmemleak_init() is called from start_kernel(). I > don't know what allocates that much memory so early. > It turned out that kmemleak does not play well with KASAN on those aarch64 (HPE Apollo 70 and Huawei TaiShan 2280) servers. After calling start_kernel()->setup_arch()->kasan_init(), kmemleak early log buffer went from something like from 280 to 260000. The multitude of kmemleak_alloc() calls is, for_each_memblock(memory, reg) x \ while (pgdp++, addr = next, addr != end) x \ while (ptep++, addr = next, addr != end && \ pte_none(READ_ONCE(*ptep))) Is this expected?