Re: 2.4.10pre11 vm rewrite fixes for mainline inclusion and testing

[Andrea Arcangeli <andrea@suse.de>]

On Tue, Sep 18, 2001 at 04:39:34PM -0300, Marcelo Tosatti wrote:
> 
> 
> On Tue, 18 Sep 2001, Marcelo Tosatti wrote:
> 
> > 
> > 
> > On Tue, 18 Sep 2001, Andrea Arcangeli wrote:
> > 
> > (testing now)
> 
> Well, unfortunately:

I'm not so pessimistic, I was pretty sure it would have worked out.

> 
> Sep 18 17:59:01 matrix PAM_pwdb[842]: (sshd) session opened for user
> marcelo by (uid=0)
> Sep 18 17:59:27 matrix kernel: __alloc_pages: 0-order allocation failed
> (gfp=0x20/0) from c012c5fe
> Sep 18 17:59:28 matrix sshd[860]: Accepted password for marcelo from
> 10.0.17.22 port 1020
> Sep 18 17:59:29 matrix PAM_pwdb[860]: (sshd) session opened for user
> marcelo by (uid=0)
> Sep 18 17:59:42 matrix sshd[873]: Accepted password for marcelo from
> 10.0.17.22 port 1021
> Sep 18 17:59:43 matrix PAM_pwdb[873]: (sshd) session opened for user
> marcelo by (uid=0)
> Sep 18 17:59:48 matrix PAM_pwdb[875]: (su) session opened for user root by
> marcelo(uid=719)
> Sep 18 17:59:55 matrix kernel: __alloc_pages: 0-order allocation failed
> (gfp=0x20/0) from c012c5fe

That's just a GREAT log: those are GFP_ATOMIC allocations, all is _more_
than _perfect_ in the above log.

> I really think we need the "fail: loop again: try_to_free_pages()" logic,

What do you want to do with GFP_ATOMIC? we cannot do anything at the
moment unless you want to make an atomic list that we can free from
atomic context, I thought about something like that but I guess we can
do it after 2.4.11 (possibly in 2.4.x, it doesn't sound a showstopper).

> Andrea: If there is not enough memory we HAVE to block in the page
> reclaiming work.

we obviously cannot block with GFP_ATOMIC.

Andrea