Re: [RFC PATCH] fs/pipe.c: implement minimum pipe size for arg==0

[Josh Poimboeuf]

On Mon, Sep 11, 2017 at 04:15:49PM -0700, Randy Dunlap wrote:
> From: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> 
> Shankara reports that running Syskaller with UBSAN causes this message:
>   UBSAN: Undefined behaviour in ./include/linux/log2.h:57:13
> 
> Syzkaller is trying to set the pipe size to 0UL. The call chain is:
> 	pipe_set_size(pipe, 0UL)
> 	...
> 	size = round_pipe_size(arg); // arg == 0UL
> which does
> 	nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; // = 0UL
> 	return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
> which is undefined when the argument is 0... and which calls
> 	fls_long(-1) // == 64
> and then returns 1UL << 64. This is where UBSAN kicks in.
> 
> The fcntl() man page [http://man7.org/linux/man-pages/man2/fcntl.2.html]
> says that:
> 	Attempts to set the pipe capacity below the page size are
> 	silently rounded up to the page size.
> 
> We could try to fix the basic low-level functions to handle 0 (where
> <linux/log2.h> says the result is undefined when n == 0), but the safest
> path for now is probably just to patch fs/pipe.c to make the documented
> default happen when arg is 0.
> 
> Reported-by: Shankara Pailoor <sp3485@xxxxxxxxxxxx>
> Signed-off-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> ---
>  fs/pipe.c |    2 ++
>  1 file changed, 2 insertions(+)
> 
> We could just return -EINVAL when arg == 0, but we don't know how that might
> adversely affect some programs.
> 
> 
> --- lnx-413.orig/fs/pipe.c
> +++ lnx-413/fs/pipe.c
> @@ -1038,6 +1038,8 @@ static long pipe_set_size(struct pipe_in
>  	unsigned long user_bufs;
>  	long ret = 0;
>  
> +	if (!arg)
> +		arg = PAGE_SIZE;
>  	size = round_pipe_size(arg);
>  	nr_pages = size >> PAGE_SHIFT;

Searching through lkml, it looks like there have been four attempts to
fix this issue in the last year or so, but none of them have been
merged.  It would be nice to finally get a fix in.

I agree with the patch, though I would argue that the fix belongs in
round_pipe_size() so that other callers don't get the same bug.

-- 
Josh