Re: [PATCH] mm: add gfp_mask parameter to vm_map_ram()

[David Rientjes <rientjes@xxxxxxxxxx>]

On Thu, 14 Jun 2012, Dave Chinner wrote:

> Oh, please. I have been hearing this for years, and are we any
> closer to it? No, we are further away from ever being able to
> acheive this than ever. Face it, filesystems require memory
> allocation to write dirty data to disk, and the amount is almost
> impossible to define. Hence mempools can't be used because we can't
> give any guarantees of forward progress. And for vmalloc?
> 
> Filesystems widely use vmalloc/vm_map_ram because kmalloc fails on
> large contiguous allocations. This renders kmalloc unfit for
> purpose, so we have to fall back to single page allocation and
> vm_map_ram or vmalloc so that the filesystem can function properly.
> And to avoid deadlocks, all memory allocation must be able to
> specify GFP_NOFS to prevent the MM subsystem from recursing into the
> filesystem. Therefore, vmalloc needs to support GFP_NOFS.
> 
> I don't care how you make it happen, just fix it. Trying to place
> the blame on the filesystem folk for using vmalloc in GFP_NOFS
> contexts is a total and utter cop-out, because mm folk of all people
> should know that non-zero order kmalloc is not a reliable
> alternative....
> 

I'd actually like to see a demonstrated problem (i.e. not theoretical) 
where vmalloc() stalls indefinitely because its passed GFP_NOFS.  I've 
never seen one reported.

This is because the per-arch pte allocators have hardwired GFP_KERNEL 
flags, but then again they also have __GFP_REPEAT which would cause them 
to loop infinitely in the page allocator if a page was not reclaimed, 
which has little success without __GFP_FS.  But nobody has ever reported a 
livelock that was triaged back to passing !__GFP_FS to vmalloc().