Re: [PATCH] mm: fix possible cause of a page_mapped BUG

[Robert ÅwiÄcki]

On Sat, Mar 19, 2011 at 6:34 AM, Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> On Thu, 17 Mar 2011, Robert Swiecki wrote:
>> On Tue, Mar 1, 2011 at 12:35 AM, Robert Swiecki <robert@xxxxxxxxxxx> wrote:
>>
>> So, I compiled 2.6.38 and started fuzzing it. I'm bumping into other
>> problems, and never seen anything about mremap in 2.6.38 (yet),
>
> Thanks a lot for getting back to this, Robert, and thanks for the update.
> I won't be celebrating, but this sounds like good news for my mremap patch.
>
>> as it had been happening in 2.6.37-rc2. The output goes to
>> http://alt.swiecki.net/linux_kernel/ - I'm still trying.
>
> A problem in sys_mlock: I've Cc'ed Michel who is the current expert.
>
> A problem in sys_munlock: Michel again, except vma_prio_tree_add is
> implicated, and I used to be involved with that. ÂI've appended below
> a debug patch which I wrote years ago, and have largely forgotten, but
> Andrew keeps it around in mmotm: we might learn more if you add that
> into your kernel build.

Hey, I'll apply your patch and check it out. In the meantime I
triggered another Oops (NULL-ptr deref via sys_mprotect).

The oops is here:

http://alt.swiecki.net/linux_kernel/sys_mprotect-2.6.38.txt

> A problem in next_pidmap from find_ge_pid from ... proc_pid_readdir.
> I did spend a while looking into that when you first reported it.
> I'm pretty sure, from the register values, that it's a result of
> a pid number (in some places signed int, in some places unsigned)
> getting unexpectedly sign-extended to negative, so indexing before
> the beginning of an array; but I never tracked down the root of the
> problem, and failed to reproduce it with odd lseeks on the directory.
>
> Ah, the one you report now comes from compat_sys_getdents,
> whereas the original one came from compat_sys_old_readdir: okay,
> I had been wondering whether it was peculiar to the old_readdir case,
> but no, it's reproduced with getdents too. ÂMight be peculiar to compat.
>
> Anyway, I've Cc'ed Eric who will be the best for that one.
>
> And a couple of watchdog problems: I haven't even glanced at
> those, hope someone else can suggest a good way forward on them.
>
> Hugh
>
>>
>> > Btw, the fuzzer is here: http://code.google.com/p/iknowthis/
>> >
>> > I think i was trying it with this revision:
>> > http://code.google.com/p/iknowthis/source/detail?r=11 (i386 mode,
>> > newest 'iknowthis' supports x86-64 natively), so feel free to try it.
>> >
>> > It used to crash the machine (it's BUG_ON but the system became
>> > unusable) in matter of hours. Btw, when I was testing it for the last
>> > time it Ooopsed much more frequently in proc_readdir (I sent report in
>> > one of earliet e-mails).
>
> From: Hugh Dickins <hughd@xxxxxxxxxx>
>
> Jayson Santos has sighted mm/prio_tree.c:78,79 BUGs (kernel bugzilla 8446),
> and one was sighted a couple of years ago. ÂNo reason yet to suppose
> they're prio_tree bugs, but we can't tell much about them without seeing
> the vmas.
>
> So dump vma and the one it's supposed to resemble: I had expected to use
> print_hex_dump(), but that's designed for u8 dumps, whereas almost every
> field of vm_area_struct is either a pointer or an unsigned long - which
> look nonsense dumped as u8s.
>
> Replace the two BUG_ONs by a single WARN_ON; and if it fires, just keep
> this vma out of the tree (truncation and swapout won't be able to find it).
> ÂHow safe this is depends on what the error really is; but we hold a file's
> i_mmap_lock here, so it may be impossible to recover from BUG_ON.
>
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: Jayson Santos <jaysonsantos2003@xxxxxxxxxxxx>
> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> ---
>
> Âmm/prio_tree.c | Â 33 ++++++++++++++++++++++++++++-----
> Â1 file changed, 28 insertions(+), 5 deletions(-)
>
> diff -puN mm/prio_tree.c~prio_tree-debugging-patch mm/prio_tree.c
> --- a/mm/prio_tree.c~prio_tree-debugging-patch
> +++ a/mm/prio_tree.c
> @@ -67,6 +67,20 @@
> Â* Â Â vma->shared.vm_set.head == NULL ==> a list node
> Â*/
>
> +static void dump_vma(struct vm_area_struct *vma)
> +{
> + Â Â Â void **ptr = (void **) vma;
> + Â Â Â int i;
> +
> + Â Â Â printk("vm_area_struct at %p:", ptr);
> + Â Â Â for (i = 0; i < sizeof(*vma)/sizeof(*ptr); i++, ptr++) {
> + Â Â Â Â Â Â Â if (!(i & 3))
> + Â Â Â Â Â Â Â Â Â Â Â printk("\n");
> + Â Â Â Â Â Â Â printk(" %p", *ptr);
> + Â Â Â }
> + Â Â Â printk("\n");
> +}
> +
> Â/*
> Â* Add a new vma known to map the same set of pages as the old vma:
> Â* useful for fork's dup_mmap as well as vma_prio_tree_insert below.
> @@ -74,14 +88,23 @@
> Â*/
> Âvoid vma_prio_tree_add(struct vm_area_struct *vma, struct vm_area_struct *old)
> Â{
> - Â Â Â /* Leave these BUG_ONs till prio_tree patch stabilizes */
> - Â Â Â BUG_ON(RADIX_INDEX(vma) != RADIX_INDEX(old));
> - Â Â Â BUG_ON(HEAP_INDEX(vma) != HEAP_INDEX(old));
> -
> Â Â Â Âvma->shared.vm_set.head = NULL;
> Â Â Â Âvma->shared.vm_set.parent = NULL;
>
> - Â Â Â if (!old->shared.vm_set.parent)
> + Â Â Â if (WARN_ON(RADIX_INDEX(vma) != RADIX_INDEX(old) ||
> + Â Â Â Â Â Â Â Â Â HEAP_INDEX(vma) Â!= HEAP_INDEX(old))) {
> + Â Â Â Â Â Â Â /*
> + Â Â Â Â Â Â Â Â* This should never happen, yet it has been seen a few times:
> + Â Â Â Â Â Â Â Â* we cannot say much about it without seeing the vma contents.
> + Â Â Â Â Â Â Â Â*/
> + Â Â Â Â Â Â Â dump_vma(vma);
> + Â Â Â Â Â Â Â dump_vma(old);
> + Â Â Â Â Â Â Â /*
> + Â Â Â Â Â Â Â Â* Don't try to link this (corrupt?) vma into the (corrupt?)
> + Â Â Â Â Â Â Â Â* prio_tree, but arrange for its removal to succeed later.
> + Â Â Â Â Â Â Â Â*/
> + Â Â Â Â Â Â Â INIT_LIST_HEAD(&vma->shared.vm_set.list);
> + Â Â Â } else if (!old->shared.vm_set.parent)
> Â Â Â Â Â Â Â Âlist_add(&vma->shared.vm_set.list,
> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â&old->shared.vm_set.list);
> Â Â Â Âelse if (old->shared.vm_set.head)
>



-- 
Robert ÅwiÄcki
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/