| Message ID | a5234d11b8cc158352a2f97fc33aa9ad90bb287b.1551550112.git.jstancek@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] mm/memory.c: do_fault: avoid usage of stale vm_area_struct | expand |
On Sat, Mar 02, 2019 at 07:19:39PM +0100, Jan Stancek wrote: > static vm_fault_t do_fault(struct vm_fault *vmf) > { > struct vm_area_struct *vma = vmf->vma; > + struct mm_struct *vm_mm = READ_ONCE(vma->vm_mm); Would this not need a corresponding WRITE_ONCE() in vma_init() ?
Hello Jan, On Sat, Mar 02, 2019 at 07:19:39PM +0100, Jan Stancek wrote: > + struct mm_struct *vm_mm = READ_ONCE(vma->vm_mm); The vma->vm_mm cannot change under gcc there, so no need of READ_ONCE. The release of mmap_sem has release semantics so the vma->vm_mm access cannot be reordered after up_read(mmap_sem) either. Other than the above detail: Reviewed-by: Andrea Arcangeli <aarcange@redhat.com> Thanks, Andrea
----- Original Message ----- > Hello Jan, > > On Sat, Mar 02, 2019 at 07:19:39PM +0100, Jan Stancek wrote: > > + struct mm_struct *vm_mm = READ_ONCE(vma->vm_mm); > > The vma->vm_mm cannot change under gcc there, so no need of > READ_ONCE. The release of mmap_sem has release semantics so the > vma->vm_mm access cannot be reordered after up_read(mmap_sem) either. > > Other than the above detail: > > Reviewed-by: Andrea Arcangeli <aarcange@redhat.com> Thank you for review, I dropped READ_ONCE and sent v3 with your Reviewed-by included. I also successfully re-ran tests over-night. > Would this not need a corresponding WRITE_ONCE() in vma_init() ? There's at least 2 context switches between, so I think it wouldn't matter. My concern was gcc optimizing out vm_mm, and vma->vm_mm access happening only after do_read_fault().
diff --git a/mm/memory.c b/mm/memory.c index e11ca9dd823f..6c1afc1ece50 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -3517,10 +3517,13 @@ static vm_fault_t do_shared_fault(struct vm_fault *vmf) * but allow concurrent faults). * The mmap_sem may have been released depending on flags and our * return value. See filemap_fault() and __lock_page_or_retry(). + * If mmap_sem is released, vma may become invalid (for example + * by other thread calling munmap()). */ static vm_fault_t do_fault(struct vm_fault *vmf) { struct vm_area_struct *vma = vmf->vma; + struct mm_struct *vm_mm = READ_ONCE(vma->vm_mm); vm_fault_t ret; /* @@ -3561,7 +3564,7 @@ static vm_fault_t do_fault(struct vm_fault *vmf) /* preallocated pagetable is unused: free it */ if (vmf->prealloc_pte) { - pte_free(vma->vm_mm, vmf->prealloc_pte); + pte_free(vm_mm, vmf->prealloc_pte); vmf->prealloc_pte = NULL; } return ret;
LTP testcase mtest06 [1] can trigger a crash on s390x running 5.0.0-rc8. This is a stress test, where one thread mmaps/writes/munmaps memory area and other thread is trying to read from it: CPU: 0 PID: 2611 Comm: mmap1 Not tainted 5.0.0-rc8+ #51 Hardware name: IBM 2964 N63 400 (z/VM 6.4.0) Krnl PSW : 0404e00180000000 00000000001ac8d8 (__lock_acquire+0x7/0x7a8) Call Trace: ([<0000000000000000>] (null)) [<00000000001adae4>] lock_acquire+0xec/0x258 [<000000000080d1ac>] _raw_spin_lock_bh+0x5c/0x98 [<000000000012a780>] page_table_free+0x48/0x1a8 [<00000000002f6e54>] do_fault+0xdc/0x670 [<00000000002fadae>] __handle_mm_fault+0x416/0x5f0 [<00000000002fb138>] handle_mm_fault+0x1b0/0x320 [<00000000001248cc>] do_dat_exception+0x19c/0x2c8 [<000000000080e5ee>] pgm_check_handler+0x19e/0x200 page_table_free() is called with NULL mm parameter, but because "0" is a valid address on s390 (see S390_lowcore), it keeps going until it eventually crashes in lockdep's lock_acquire. This crash is reproducible at least since 4.14. Problem is that "vmf->vma" used in do_fault() can become stale. Because mmap_sem may be released, other threads can come in, call munmap() and cause "vma" be returned to kmem cache, and get zeroed/re-initialized and re-used: handle_mm_fault | __handle_mm_fault | do_fault | vma = vmf->vma | do_read_fault | __do_fault | vma->vm_ops->fault(vmf); | mmap_sem is released | | | do_munmap() | remove_vma_list() | remove_vma() | vm_area_free() | # vma is released | ... | # same vma is allocated | # from kmem cache | do_mmap() | vm_area_alloc() | memset(vma, 0, ...) | pte_free(vma->vm_mm, ...); | page_table_free | spin_lock_bh(&mm->context.lock);| <crash> | Cache mm_struct to avoid using potentially stale "vma". [1] https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/mem/mtest06/mmap1.c Signed-off-by: Jan Stancek <jstancek@redhat.com> --- mm/memory.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)