Re: [PATCH 1/1 v2] kernel/fork.c: avoid division by zero

[Guenter Roeck <linux@roeck-us.net>]

On Wed, Feb 18, 2015 at 08:50:21PM +0100, Heinrich Schuchardt wrote:
> On 18.02.2015 20:38, Guenter Roeck wrote:
> > On Tue, Feb 17, 2015 at 03:15:11PM -0800, Andrew Morton wrote:
> >> On Tue, 17 Feb 2015 20:01:38 +0100 Heinrich Schuchardt <xypron.glpk@gmx.de> wrote:
> >>
> >>> PAGE_SIZE is not guaranteed to be equal to or less than 8 times the
> >>> THREAD_SIZE.
> >>>
> >>> E.g. architecture hexagon may have page size 1M and thread size 4096.
> >>>
> >>> This would lead to a division by zero.
> >>>
> >>> The futex implementation assumes that tids fit into the FUTEX_TID_MASK.
> >>> This limits the number of allowable threads.
> >>>
> >>> ...
> >>>
> >>> --- a/kernel/fork.c
> >>> +++ b/kernel/fork.c
> >>> @@ -74,6 +74,7 @@
> >>>  #include <linux/uprobes.h>
> >>>  #include <linux/aio.h>
> >>>  #include <linux/compiler.h>
> >>> +#include <linux/math64.h>
> >>>  
> >>>  #include <asm/pgtable.h>
> >>>  #include <asm/pgalloc.h>
> >>> @@ -255,6 +256,8 @@ void __init __weak arch_task_cache_init(void) { }
> >>>  
> >>>  void __init fork_init(unsigned long mempages)
> >>>  {
> >>> +	u64 temp;
> >>
> >> That's a really poor name.  We should always try to make names
> >> meaningful.  Here, something like "threads" would be better.
> >>
> >>>  #ifndef CONFIG_ARCH_TASK_STRUCT_ALLOCATOR
> >>>  #ifndef ARCH_MIN_TASKALIGN
> >>>  #define ARCH_MIN_TASKALIGN	L1_CACHE_BYTES
> >>> @@ -273,7 +276,16 @@ void __init fork_init(unsigned long mempages)
> >>>  	 * value: the thread structures can take up at most half
> >>>  	 * of memory.
> >>>  	 */
> >>> -	max_threads = mempages / (8 * THREAD_SIZE / PAGE_SIZE);
> >>> +	temp = div64_u64((u64) mempages * (u64) PAGE_SIZE,
> >>> +			 (u64) THREAD_SIZE * 8UL);
> >>> +
> >>> +	/*
> >>> +	 * The futex code assumes that tids fit into the FUTEX_TID_MASK.
> >>> +	 */
> >>> +	if (temp < FUTEX_TID_MASK)
> >>> +		max_threads = temp;
> >>> +	else
> >>> +		max_threads = FUTEX_TID_MASK;
> >>
> >> Seems rather complicated.  How about
> >>
> >> 	max_threads = mempages / (8 * THREAD_SIZE);
> > 
> > Apparently it is possible that
> > 	mempages < 8 * THREAD_SIZE
> > which would result in max_threads == 0.
> > 
> > How about the following ?
> > 
> > 	max_threads = mempages / DIV_ROUND_UP(8 * THREAD_SIZE, PAGE_SIZE);
> 
> For PAGE_SIZE 2MB and THREAD_SIZE 4kB this is wrong by factor 64.
> 
That really depends on the definition of "wrong", doesn't it ?
It would ensure that max_threads is never larger than mempages,
which is more defensive.

One could also use
	max_threads = mempages / (max(1, 8 * THREAD_SIZE / PAGE_SIZE));
to solve the original problem without introducing 64 bit operations
and without affecting the result for more common architectures.

Guenter