On 24.08.22 14:03, Brian Foster wrote: > Hi all, > > When running a fuzzer workload to test an unrelated patch[1], I've been > reproducing the VM_BUG_ON() splat below[2] on s390x. I've narrowed the > problem down to a deterministic reproducer. The code for that is also > appended below[3]. > > The splat occurs because during fork() we end up down in > copy_present_pte() -> page_try_dup_anon_rmap() -> > page_needs_cow_for_dma() for a !is_cow mapping, so copy_page_range() > didn't acquire the ->write_protect_seq seqlock as expected. After > digging into this a bit, I _think_ this boils down to a bug in the s390 > arch fault code dealing with a write fault to a !VM_WRITE mapping.. > > The sequence of events implemented by the reproducer that leads to this: > > 1. Create a shmem segment and attach it SHM_RDONLY. This causes > do_mmap() to set up a !VM_WRITE mapping, but also clear > (VM_MAYWRITE|VM_SHARED) on the mapping because the backing shmem file is > read-only. > > 2. Take a write fault on the mapping in kernel mode (via getrlimit() in > this case). The write fault ultimately causes getrlimit() to fail with a > bad access error, but not before the generic fault code creates an > anon_vma mapping for the page. > > This occurs because first do_dat_exception() calls handle_mm_fault() > with FAULT_FLAG_WRITE via the following logic: > > access = VM_ACCESS_FLAGS; > ... > if (access == VM_WRITE || is_write) > flags |= FAULT_FLAG_WRITE; > ... > if (unlikely(!(vma->vm_flags & access))) > goto out_up; > ... > fault = handle_mm_fault(...); > > So the FAULT_FLAG_WRITE fault proceeds because is_write is true and > ->vm_flags has read or exec permission (but not VM_WRITE). This > eventually gets down into do_cow_fault() -> finish_fault() -> > do_set_pte(), the latter of which calls page_add_new_anon_rmap() because > this is a write fault to a !shared mapping. > > Note this is immediately followed by a do_protection_exception() that > uses access = VM_WRITE and thus fails the above check and returns with > VM_FAULT_BADACCESS. So I think this ultimately DTRT wrt to failing the > syscall to userspace, but the do_dat_exception() handling sets up an > unexpected situation for fork().. > > 3. fork() -> dup_mm() comes across this mapping with ->anon_vma set (so > vma_needs_copy() returns true), but is_cow_mapping() returns false > because VM_MAYWRITE is cleared. From there we fall down into the page > table copying path described by the BUG splat. > > This problem doesn't occur on x86 seemingly because we don't call into > handle_mm_fault() for a write fault to a !VM_WRITE mapping, which is > specifically checked in access_error(). Therefore, something like the > patch below[4] seems to prevent the problem on s390. However, the access > checking logic looks wonky enough to me that I wonder whether it > warrants a closer look from s390 experts. For example, does this code > need to care about any flags/context beyond write or read faults vs. > write or !write mappings? I don't have enough context to reason about > it. Could somebody more familiar with these two s390 exception variants > chime in? > > Finally, note that so far I've only really tested the patch against the > reproducer. I'm happy to try and form it into a proper patch and test > further after some feedback... thanks. > > Brian > > [1] https://lore.kernel.org/linux-s390/20220816155407.537372-1-bfoster@xxxxxxxxxx > [2] BUG splat: > > kernel BUG at include/linux/mm.h:1529! > monitor event: 0040 ilc:2 [#1] SMP > Modules linked in: rfkill sunrpc ghash_s390 prng xts aes_s390 des_s390 libdes sha3_512_s390 sha3_256_s390 vfio_ccw sha512_s390 mdev vfio_iommu_type1 vfio xfs libcrc32c virtio_blk virtio_net net_failover failover dm_mirror dm_region_hash dm_log dm_mod pkey zcrypt > CPU: 1 PID: 1401 Comm: shmem-fork-test Not tainted 6.0.0-rc2+ #20 > Hardware name: IBM 8561 LT1 400 (KVM/Linux) > Krnl PSW : 0704c00180000000 0000000014928240 (copy_pte_range+0xa40/0xe58) > R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3 > Krnl GPRS: 000003ff85b80000 000000000000000c 0000000000000000 000003ff85b80000 > 0000000091c5f31f 0000000087d70640 000000008160e800 00000372024717c0 > 000003ff85b80000 0000000000000000 00000000831c9c00 0000000091c5f31f > 00000000823ada00 0000000087d70640 00000000149279c2 0000038000773880 > Krnl Code: 0000000014928232: c0e5fffff48f brasl %r14,0000000014926b50 > 0000000014928238: a7f4fd43 brc 15,0000000014927cbe > #000000001492823c: af000000 mc 0,0 > >0000000014928240: b904005b lgr %r5,%r11 > 0000000014928244: a7f4ffde brc 15,0000000014928200 > 0000000014928248: e310f0e80004 lg %r1,232(%r15) > 000000001492824e: a7f4ff17 brc 15,000000001492807c > 0000000014928252: ec3800091c7c cgij %r3,28,8,0000000014928264 > Call Trace: > [<0000000014928240>] copy_pte_range+0xa40/0xe58 > ([<00000000149279c2>] copy_pte_range+0x1c2/0xe58) > [<000000001492e258>] copy_page_range+0x510/0x770 > [<00000000146f3896>] dup_mmap+0x47e/0x6c0 > [<00000000146f3b52>] dup_mm+0x7a/0x278 > [<00000000146f5a48>] copy_process+0x1298/0x1a48 > [<00000000146f62fe>] kernel_clone+0x5e/0x3c0 > [<00000000146f6742>] __do_sys_clone+0x5a/0x68 > [<00000000146f67e0>] __s390x_sys_clone+0x40/0x50 > [<0000000014f68dac>] __do_syscall+0x1d4/0x200 > [<0000000014f78c22>] system_call+0x82/0xb0 > Last Breaking-Event-Address: > [<0000000014927a46>] copy_pte_range+0x246/0xe58 > Kernel panic - not syncing: Fatal exception: panic_on_oops > > [3] minimal reproducer: > > #include <unistd.h> > #include <sys/shm.h> > #include <sys/resource.h> > > int main() > { > int id; > void *p; > > id = shmget(IPC_PRIVATE, 4096, IPC_CREAT); > p = shmat(id, NULL, SHM_RDONLY); > getrlimit(RLIMIT_AS, p); > fork(); > return 0; > } > > [4] RFC patch: > > diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c > index 13449941516c..c12722da1558 100644 > --- a/arch/s390/mm/fault.c > +++ b/arch/s390/mm/fault.c > @@ -418,6 +418,8 @@ static inline vm_fault_t do_exception(struct pt_regs *regs, int access) > fault = VM_FAULT_BADACCESS; > if (unlikely(!(vma->vm_flags & access))) > goto out_up; > + if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_WRITE)) > + goto out_up; > > if (is_vm_hugetlb_page(vma)) > address &= HPAGE_MASK; > Heh, we might have identified this independently just recently: https://lore.kernel.org/all/20220816113359.33843f54@thinkpad/T/#u Can you take a look if that proposed small change also fixes the issue? -- Thanks, David / dhildenb
