Re: the new VMt

[Andrea Arcangeli <andrea@suse.de>]

On Mon, Sep 25, 2000 at 11:51:39AM -0600, yodaiken@fsmlabs.com wrote:
> It should probably be GFP_ATOMIC, if I understand the mm right. 

poll_wait is called from the f_op->poll callback from select just before
a sleep and since it's allowed to sleep too it should be a GFP_KERNEL
(not ATOMIC). Using GFP_ATOMIC where GFP_KERNEL can be used is a bug
and it can lead to failed allocations even while there's huge amount
of freeable/recyclable cache.

The reason it isn't GFP_USER but it's a GFP_KERNEL is because the memory
isn't allocated in userspace.

On a solid VM the only difference between GFP_USER and GFP_KERNEL happens to be
when the machine runs truly out of memory. In 2.4.x GFP_KERNEL should probably
be changed not to short the PF_MEMALLOC atomic queue when memory balancing
fails (then they would be equal).

> The algorithm for requesting a collection of reources and freeing all of them
>  on failure is simple, fast, and robust. 

Yes, I tend to like that style too because it's obviously safe and it obviously
can't dealdock during oom.

Andrea
--
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.eu.org/Linux-MM/