On Wed, Feb 11, 2015 at 3:22 PM, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > On Tue, Feb 10, 2015 at 10:42:31PM +0400, Konstantin Khlebnikov wrote: >> On Mon, Dec 22, 2014 at 10:14 PM, Kirill A. Shutemov >> <kirill@xxxxxxxxxxxxx> wrote: >> > On Mon, Dec 22, 2014 at 01:05:13PM -0500, Sasha Levin wrote: >> >> On 12/22/2014 01:01 PM, Kirill A. Shutemov wrote: >> >> > On Mon, Dec 22, 2014 at 10:04:02AM -0500, Sasha Levin wrote: >> >> >> > Hi all, >> >> >> > >> >> >> > While fuzzing with trinity inside a KVM tools guest running the latest -next >> >> >> > kernel, I've stumbled on the following spew: >> >> >> > >> >> >> > [ 432.376425] BUG: unable to handle kernel NULL pointer dereference at 0000000000000038 >> >> >> > [ 432.378876] IP: down_write (./arch/x86/include/asm/rwsem.h:105 ./arch/x86/include/asm/rwsem.h:121 kernel/locking/rwsem.c:71) >> >> > Looks like vma->vm_file->mapping is NULL. Somebody freed ->vm_file from >> >> > under us? >> >> > >> >> > I suspect Davidlohr's patchset on i_mmap_lock, but I cannot find any code >> >> > path which could lead to the crash. >> >> >> >> I've reported a different issue which that patchset: https://lkml.org/lkml/2014/12/9/741 >> >> >> >> I guess it could be related? >> > >> > Maybe. >> > >> > Other thing: >> > >> > unmap_mapping_range() >> > i_mmap_lock_read(mapping); >> > unmap_mapping_range_tree() >> > unmap_mapping_range_vma() >> > zap_page_range_single() >> > unmap_single_vma() >> > untrack_pfn() >> > vma->vm_flags &= ~VM_PAT; >> > >> > It seems we modify ->vm_flags without mmap_sem taken, means we can corrupt >> > them. >> > >> > Sasha could you check if you hit untrack_pfn()? >> > >> > The problem probably was hidden by exclusive i_mmap_lock on >> > unmap_mapping_range(), but it's not exclusive anymore afrer Dave's >> > patchset. >> > >> > Konstantin, you've modified untrack_pfn() back in 2012 to change >> > ->vm_flags. Any coments? >> >> Hmm. I don't really understand how >> unmap_mapping_range() could be used for VM_PFNMAP mappings >> except unmap() or exit_mmap() where mm is locked anyway. >> Somebody truncates memory mapped device and unmaps mapped PFNs? > > Hm. Probably not. But it's not obvious to me what would stop this. > Should we at least have assert on mmap_sem locked in untrack_pfn()? exit_mmap() runs without mmap_sem thus this should be something like thus: WARN_ON_ONCE(atomic_read(&mm->mm_users) && !rwsem_is_locked(&mm->mmap_sem)); Clearing VM_MAYSHARE in __unmap_hugepage_range_final() has the same problem, it's called from unmap_single_vma() as well as untrack_pfn(). Also clear_soft_dirty_pmd() clears VM_SOFTDIRTY under mmap_sem locked for read. > >> If it's a problem then I think VM_PAT could be tuned into hint which >> means PAT tracking was here and we pat should check internal structure >> for details and take actions if pat tracking is still here. As I see >> pat tracking probably also have problems if somebody unmaps that vma >> partially. > > IIUC, we only mark a vma with VM_PAT if whole vma is subject for > remap_pfn_range(). I don't see a point in cleaning VM_PAT -- just let it > die with vma. Or do I miss something? Yep, for now track/untrack works only for whole vma and cannot handle vma split or partial munmap. > > -- > Kirill A. Shutemov -- 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>