Re: 2.6.27-rc1: IP: iov_iter_advance+0x2e/0x90

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, 30 Jul 2008, Alexey Dobriyan wrote:
>
> Steps to reproduce:
> 
> 	# while true; do ./ftest03; done
> 
> ftest03 from LTP 20080603

Hmm. The oops disassembles to

 -12:	4c 8b 07             	mov    (%rdi),%r8
  -9:	48 8b 4f 10          	mov    0x10(%rdi),%rcx
  -5:	48 85 f6             	test   %rsi,%rsi
  -2:	75 17                	jne    0x42
   0:	49 83 78 08 00       	cmpq   $0x0,0x8(%r8)	<---
   5:	75 07                	jne    0xe
   7:	48 83 7f 18 00       	cmpq   $0x0,0x18(%rdi)
   c:	75 09                	jne    0x17

So it looks like we just overflowed %r8 to a new page and you presumably 
have DEBUG_PAGEALLOC on.

(And yes, I see in the oops that you do)

> RIP  [<ffffffff8026190e>] iov_iter_advance+0x2e/0x90
> Call Trace:
>  [<ffffffff80263452>] generic_file_buffered_write+0x1e2/0x710
>  [<ffffffff8040cfd0>] ? _spin_unlock+0x30/0x60
>  [<ffffffff80263e0f>] __generic_file_aio_write_nolock+0x29f/0x450
>  [<ffffffff80264026>] generic_file_aio_write+0x66/0xd0
>  [<ffffffff802c9506>] ext3_file_write+0x26/0xc0
>  [<ffffffff80264250>] ? generic_file_aio_read+0x0/0x670
>  [<ffffffff802c94e0>] ? ext3_file_write+0x0/0xc0
>  [<ffffffff8028921b>] do_sync_readv_writev+0xeb/0x130
>  [<ffffffff8025284d>] ? trace_hardirqs_on+0xd/0x10
>  [<ffffffff802449c0>] ? autoremove_wake_function+0x0/0x40
>  [<ffffffff80289055>] ? rw_copy_check_uvector+0x95/0x130
>  [<ffffffff80289953>] do_readv_writev+0xc3/0x120
>  [<ffffffff8025284d>] ? trace_hardirqs_on+0xd/0x10
>  [<ffffffff802527b5>] ? trace_hardirqs_on_caller+0xd5/0x160
>  [<ffffffff8025284d>] ? trace_hardirqs_on+0xd/0x10
>  [<ffffffff802899e9>] vfs_writev+0x39/0x60
>  [<ffffffff80289d60>] sys_writev+0x50/0x90
>  [<ffffffff8020b65b>] system_call_fastpath+0x16/0x1b
> 
> 0xffffffff8026190e is in iov_iter_advance (mm/filemap.c:1882).
> 1877
> 1878                    /*
> 1879                     * The !iov->iov_len check ensures we skip over unlikely
> 1880                     * zero-length segments (without overruning the iovec).
> 1881                     */
> 1882     ===>           while (bytes || unlikely(!iov->iov_len && i->count)) {

And yes, that oopsing op would be the one that loads 'iov->iov_len'.

So it very much looks like iov_iter_advance() advances past the end of the 
iov array. We've had issues like that before. And I bet it's due to a 
combination of Nick's commit f7009264c519603b8ec67c881bd368a56703cfc9 
("iov_iter_advance() fix") and 124d3b7041f9a0ca7c43a6293e1cae4576c32fd5 
("fix writev regression: pan hanging unkillable and un-straceable").

It's simply _not_ acceptable to look at iov->iov_len when 'bytes' has gone 
down to zero, because there may be no 'iov' left!

Nick?

That said, I do think that we have another issue with iovec's - I think we 
should strive to always pass in the number of iovec's when we pass a 
pointer to an iovec, in addition to the bytes. The sad part is that 
'iov_iter_advance' actually -has- the count, but it's the byte count 
remaining, not the iovec's remaining. 

In this particular case, the trivial fix _may_ be to just change the order 
of testing iov->iov_len && i->count, but I really think we should also 
count actual iov entries and pass them around (and keep them updated).

So does this (hacky, ugly) patch fix it for you?

			Linus

---
 mm/filemap.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 42bbc69..d97d1ad 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1879,7 +1879,7 @@ void iov_iter_advance(struct iov_iter *i, size_t bytes)
 		 * The !iov->iov_len check ensures we skip over unlikely
 		 * zero-length segments (without overruning the iovec).
 		 */
-		while (bytes || unlikely(!iov->iov_len && i->count)) {
+		while (bytes || unlikely(i->count && !iov->iov_len)) {
 			int copy;
 
 			copy = min(bytes, iov->iov_len - base);