Comment # 8 on bug 947825 from Petr Tesa����k

(In reply to Petr Tesa����k from comment #7)
> In my test VM, I saw additional 6M used by kmalloc caches. I managed to
> track it down to the qxl framebuffer driver, which may allocate delayed work
> if called in atomic context. In my case the length of the queue was 20k+
> operations. Since bitblt operations also include the bitmap data, it adds
> up...

Turns out this was a result of multiple chained OOM backtraces, which should
never happen. Under normal circumstances, there may be a spurious backtrace in
atomic context, but that needs only approx. 1M backlog buffer, which is covered
by general reserve margins.

In short, no special adjustments are necessary for qxl.