Re: [TEST] page tables filling non-highmem

[Andrea Arcangeli <andrea@suse.de>]

On Fri, Feb 15, 2002 at 09:59:45AM +0100, Daniel Phillips wrote:
> On February 15, 2002 05:51 am, William Lee Irwin III wrote:
> > The following testcase brought down 2.4.17 mainline on an
> > 8-way P-III 700MHz machine with 12GB of RAM. The last thing
> > logged from it was a LowFree of 2MB with 9GB of highmem free
> > after something like 6-8 hours of pounding away, at which
> > time the machine stopped responding (IIRC it was given ~12
> > hours to echo another character).
> > 
> > This testcase is a blatant attempt to fill the direct-mapped
> > portion of the kernel virtual address space with process pagetables.
> > It was suspected such a thing was happening in another failure scenario
> > which is what motivated me to devise this testcase. I believe a fix
> > already exists (i.e. aa's ptes in highmem stuff) though I've not yet
> > verified its correct operation here.
> 
> As you described it to me on irc, this demonstration turns up a
> considerably worse problem than just having insufficient space for
> page tables - the system locks up hard instead of doing anything
> reasonable on page table-related oom.  It's wrong that the system
> should behave this way, it is after all, just an oom.
> 
> Now that basic stability issues seem to be under control, perhaps
> it's time to give the oom problem the attention it deserves?

My tree doesn't lock up hard even without pte-highmem applied.  The task
gets killed. backout pte-highmem, try the same testcase again on my tree
and you'll see. The oom handling in mainline is deadlock prone, I always
known this and that's why I always rejected it. Nobody but me
acklowledged this problem and I spent quite an amount of time convincing
mainline maintainers about those deadlock flaws of the mainline approch
but I failed so I giveup waiting for a report like this, just like with
all the other stuff that is now in my vm patch, 90% of it I tried to
push it separately into mainline before having to accumulate it.

Andrea