[lkml]   [2008]   [Oct]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] regression: vmalloc easily fail.
On Tue, Oct 28, 2008 at 11:03:22PM +0200, Avi Kivity wrote:
> Glauber Costa wrote:
>> Commit db64fe02258f1507e13fe5212a989922323685ce broke
>> KVM (the symptom) for me. The cause is that vmalloc
>> allocations fail, despite of the fact that /proc/meminfo
>> shows plenty of vmalloc space available.
>> After some investigation, it seems to me that the current
>> way to compute the next addr in the rb-tree transversal
>> leaves a spare page between each allocation. After a few
>> allocations, regardless of their size, we run out of vmalloc
>> space.
>> while (addr + size >= first->va_start && addr + size <= vend) {
>> - addr = ALIGN(first->va_end + PAGE_SIZE, align);
>> + addr = ALIGN(first->va_end, align);
>> n = rb_next(&first->rb_node);
>> if (n)
> I'm guessing that the missing comment explains that this is intentional,
> to trap buffer overflows?
> (okay that was a cheap shot. I don't comment nearly enough either)
> Even if you leave a page between allocations, I don't see how you can
> fail a one page allocation, unless you've allocated at least N/2 pages
> (where N is the size of the vmalloc space in pages).

I'm hoping Nick will comment on it. I might well be wrong.
but it nicely fixes the problem for me, and actually, you don't need
"at least N/2 pages". The size of the allocations hardly matters, just
the amount of allocations we did. Since kvm does some small
vmalloc allocations, that may be the reason for we triggering it.

 \ /
  Last update: 2008-10-28 22:11    [W:0.117 / U:0.608 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site