On 2016/6/21 23:29, Kirill A. Shutemov wrote: > On Tue, Jun 21, 2016 at 11:19:07PM +0800, zhong jiang wrote: >> On 2016/6/21 22:37, Kirill A. Shutemov wrote: >>> On Tue, Jun 21, 2016 at 10:05:56PM +0800, zhongjiang wrote: >>>> From: zhong jiang <zhongjiang@xxxxxxxxxx> >>>> >>>> with great pressure, I run some test cases. As a result, I found >>>> that the THP is not freed, it is detected by check_mm(). >>>> >>>> BUG: Bad rss-counter state mm:ffff8827edb70000 idx:1 val:512 >>>> >>>> Consider the following race : >>>> >>>> CPU0 CPU1 >>>> __handle_mm_fault() >>>> wp_huge_pmd() >>>> do_huge_pmd_wp_page() >>>> pmdp_huge_clear_flush_notify() >>>> (pmd_none = true) >>>> exit_mmap() >>>> unmap_vmas() >>>> zap_pmd_range() >>>> pmd_none_or_trans_huge_or_clear_bad() >>>> (result in memory leak) >>>> set_pmd_at() >>>> >>>> because of CPU0 have allocated huge page before pmdp_huge_clear_notify, >>>> and it make the pmd entry to be null. Therefore, The memory leak can occur. >>>> >>>> The patch fix the scenario that the pmd entry can lead to be null. >>> I don't think the scenario is possible. >>> >>> exit_mmap() called when all mm users have gone, so no parallel threads >>> exist. >>> >> Forget this patch. It 's my fault , it indeed don not exist. >> But I hit the following problem. we can see the memory leak when the process exit. >> >> >> Any suggestion will be apprecaited. > Could you try this: > > http://lkml.kernel.org/r/20160621150433.GA7536@xxxxxxxxxxxxxxxxxx The patch I have seen , but I don not think this patch can fix so problem . if that race occur, pmd entry points to the huge page will be changed , and freeze_page spilt pmd will fail. subsequent vm_bug_on() will fired. freeze_page() try_to_unmap() split_huge_pmd_address() (return fail) result in page_mapcount is not zero vm_bug_on() -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>