Re: [RFC] vmalloc: add warning in __vmalloc

[David Rientjes <rientjes@google.com>]

On Fri, 27 Apr 2012, Minchan Kim wrote:

> Now there are several places to use __vmalloc with GFP_ATOMIC,
> GFP_NOIO, GFP_NOFS but unfortunately __vmalloc calls map_vm_area
> which calls alloc_pages with GFP_KERNEL to allocate page tables.
> It means it's possible to happen deadlock.
> I don't know why it doesn't have reported until now.
> 
> Firstly, I tried passing gfp_t to lower functions to support __vmalloc
> with such flags but other mm guys don't want and decided that
> all of caller should be fixed.
> 
> http://marc.info/?l=linux-kernel&m=133517143616544&w=2
> 
> To begin with, let's listen other's opinion whether they can fix it
> by other approach without calling __vmalloc with such flags.
> 
> So this patch adds warning to detect and to be fixed hopely.
> I Cced related maintainers.
> If I miss someone, please Cced them.
> 
> side-note:
>   I added WARN_ON instead of WARN_ONCE to detect all of callers
>   and each WARN_ON for each flag to detect to use any flag easily.
>   After we fix all of caller or reduce such caller, we can merge
>   a warning with WARN_ONCE.
> 

I disagree with this approach since it's going to violently spam an 
innocent kernel user's log with no ratelimiting and for a situation that 
actually may not be problematic.

Passing any of these bits (the difference between GFP_KERNEL and 
GFP_ATOMIC) only means anything when we're going to do reclaim.  And I'm 
suspecting we would have seen problems with this already since 
pte_alloc_kernel() does __GFP_REPEAT on most architectures meaning that it 
will loop infinitely in the page allocator until at least one page is 
freed (since its an order-0 allocation) which would hardly ever happen if 
__GFP_FS or __GFP_IO actually meant something in this context.

In other words, we would already have seen these deadlocks and it would 
have been diagnosed as a vmalloc(GFP_ATOMIC) problem.  Where are those bug 
reports?

At best, you'd need _some_ sort of ratelimiting like a static variable and 
only allowing 100 WARN_ON()s which could output dozens of lines for each 
call to vmalloc().

But the page allocator already has a might_sleep_if(gfp_mask & GFP_WAIT) 
which will dump the stack for CONFIG_DEBUG_ATOMIC_SLEEP.  So for this 
effect, just enable that config option and check your kernel log.

So I'm afraid this is complete overkill for something that we can't prove 
is a problem in the first place and will potentially fill the kernel logs 
for warnings where the allocation succeeds immediately.  If you want the 
bug reports, ask people to enable CONFIG_DEBUG_ATOMIC_SLEEP.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>