lkml.org 
[lkml]   [2016]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] radix-tree: fix radix_tree_iter_retry() for tagged iterators.
On Fri, Jul 15, 2016 at 11:52:58AM +0300, Andrey Ryabinin wrote:
> On 07/15/2016 01:25 AM, Ross Zwisler wrote:
> > On Thu, Jul 14, 2016 at 02:19:56PM +0300, Andrey Ryabinin wrote:
> >> radix_tree_iter_retry() resets slot to NULL, but it doesn't reset tags.
> >> Then NULL slot and non-zero iter.tags passed to radix_tree_next_slot()
> >> leading to crash:
> >>
> >> RIP: [< inline >] radix_tree_next_slot include/linux/radix-tree.h:473
> >> [<ffffffff816951a4>] find_get_pages_tag+0x334/0x930 mm/filemap.c:1452
> >> ....
> >> Call Trace:
> >> [<ffffffff816cd91a>] pagevec_lookup_tag+0x3a/0x80 mm/swap.c:960
> >> [<ffffffff81ab4231>] mpage_prepare_extent_to_map+0x321/0xa90 fs/ext4/inode.c:2516
> >> [<ffffffff81ac883e>] ext4_writepages+0x10be/0x2b20 fs/ext4/inode.c:2736
> >> [<ffffffff816c99c7>] do_writepages+0x97/0x100 mm/page-writeback.c:2364
> >> [<ffffffff8169bee8>] __filemap_fdatawrite_range+0x248/0x2e0 mm/filemap.c:300
> >> [<ffffffff8169c371>] filemap_write_and_wait_range+0x121/0x1b0 mm/filemap.c:490
> >> [<ffffffff81aa584d>] ext4_sync_file+0x34d/0xdb0 fs/ext4/fsync.c:115
> >> [<ffffffff818b667a>] vfs_fsync_range+0x10a/0x250 fs/sync.c:195
> >> [< inline >] vfs_fsync fs/sync.c:209
> >> [<ffffffff818b6832>] do_fsync+0x42/0x70 fs/sync.c:219
> >> [< inline >] SYSC_fdatasync fs/sync.c:232
> >> [<ffffffff818b6f89>] SyS_fdatasync+0x19/0x20 fs/sync.c:230
> >> [<ffffffff86a94e00>] entry_SYSCALL_64_fastpath+0x23/0xc1 arch/x86/entry/entry_64.S:207
> >>
> >> We must reset iterator's tags to bail out from radix_tree_next_slot() and
> >> go to the slow-path in radix_tree_next_chunk().
> >
> > This analysis doesn't make sense to me. In find_get_pages_tag(), when we call
> > radix_tree_iter_retry(), this sets the local 'slot' variable to NULL, then
> > does a 'continue'. This will hop to the next iteration of the
> > radix_tree_for_each_tagged() loop, which will very check the exit condition of
> > the for() loop:
> >
> > #define radix_tree_for_each_tagged(slot, root, iter, start, tag) \
> > for (slot = radix_tree_iter_init(iter, start) ; \
> > slot || (slot = radix_tree_next_chunk(root, iter, \
> > RADIX_TREE_ITER_TAGGED | tag)) ; \
> > slot = radix_tree_next_slot(slot, iter, \
> > RADIX_TREE_ITER_TAGGED))
> >
> > So, we'll run the
> > slot || (slot = radix_tree_next_chunk(root, iter, \
> > RADIX_TREE_ITER_TAGGED | tag)) ; \
> >
> > bit first.
>
> This is not the way how the for() loop works. slot = radix_tree_next_slot() executed first
> and only after that goes the condition statement.

Right...*sigh*... Thanks for the sanity check. :)

> > 'slot' is NULL, so we'll set it via radix_tree_next_chunk(). At
> > this point radix_tree_next_slot() hasn't been called.
> >
> > radix_tree_next_chunk() will set up the iter->index, iter->next_index and
> > iter->tags before it returns. The next iteration of the loop in
> > find_get_pages_tag() will use the non-NULL slot provided by
> > radix_tree_next_chunk(), and only after that iteration will we call
> > radix_tree_next_slot() again. By then iter->tags should be up to date.
> >
> > Do you have a test setup that reliably fails without this code but passes when
> > you zero out iter->tags?
> >
>
>
> Yup, I run Dmitry's reproducer in a parallel loop:
> $ while true; do ./a.out & done
>
> Usually it takes just couple minutes maximum.

Cool - I was able to get this to work on my system as well by upping the
thread count.

In looking at this more, I agree that your patch fixes this particular bug,
but I think that ultimately we might want something more general.

IIUC, the real issue is that we shouldn't be running through
radix_tree_next_slot() with a NULL 'slot' parameter. In the end I think it's
fine to zero out iter->tags in radix_tree_iter_retry(), but really we want to
guarantee that we just bail out of radix_tree_next_slot() if we have a NULL
'slot'.

I've run this patch in my test setup, and it fixes the reproducer provided by
Dmitry. I've also run xfstests against it with out any failures.

--- 8< ---
From 533beefac12f61f467aeb72e2d2c46685247b9bc Mon Sep 17 00:00:00 2001
From: Ross Zwisler <ross.zwisler@linux.intel.com>
Date: Fri, 15 Jul 2016 12:46:38 -0600
Subject: [PATCH] radix-tree: 'slot' can be NULL in radix_tree_next_slot()

There are four cases I can see where we could end up with a NULL 'slot' in
radix_tree_next_slot() (there might be more):

1) radix_tree_iter_retry() via a non-tagged iteration like
radix_tree_for_each_slot(). In this case we currently aren't seeing a bug
because radix_tree_iter_retry() sets

iter->next_index = iter->index;

which means that in in the else case in radix_tree_next_slot(), 'count' is
zero, so we skip over the while() loop and effectively just return NULL
without ever dereferencing 'slot'.

2) radix_tree_iter_retry() via tagged iteration like
radix_tree_for_each_tagged(). With the current code this case is
unhandled and we have seen it result in a kernel crash when we dereference
'slot'.

3) radix_tree_iter_next() via via a non-tagged iteration like
radix_tree_for_each_slot(). This currently happens in shmem_tag_pins()
and shmem_partial_swap_usage().

I think that this case is currently unhandled. Unlike with
radix_tree_iter_retry() case (#1 above) we can't rely on 'count' in the else
case of radix_tree_next_slot() to be zero, so I think it's possible we'll end
up executing code in the while() loop in radix_tree_next_slot() that assumes
'slot' is valid.

I haven't actually seen this crash on a test setup, but I don't think the
current code is safe.

4) radix_tree_iter_next() via tagged iteration like
radix_tree_for_each_tagged(). This happens in shmem_wait_for_pins().

radix_tree_iter_next() zeros out iter->tags, so we end up exiting
radix_tree_next_slot() here:

if (flags & RADIX_TREE_ITER_TAGGED) {
void *canon = slot;

iter->tags >>= 1;
if (unlikely(!iter->tags))
return NULL;

Really we want to guarantee that we just bail out of
radix_tree_next_slot() if we have a NULL 'slot'. This is a more explicit
way of handling all the 4 above cases.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
---
include/linux/radix-tree.h | 3 +++
1 file changed, 3 insertions(+)

diff --git a/include/linux/radix-tree.h b/include/linux/radix-tree.h
index cb4b7e8..840308d 100644
--- a/include/linux/radix-tree.h
+++ b/include/linux/radix-tree.h
@@ -463,6 +463,9 @@ static inline struct radix_tree_node *entry_to_node(void *ptr)
static __always_inline void **
radix_tree_next_slot(void **slot, struct radix_tree_iter *iter, unsigned flags)
{
+ if (unlikely(!slot))
+ return NULL;
+
if (flags & RADIX_TREE_ITER_TAGGED) {
void *canon = slot;

--
2.9.0
\
 
 \ /
  Last update: 2016-07-15 21:41    [W:0.084 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site