On Mon, Jun 11, 2018 at 11:44:05AM -0700, Linus Torvalds wrote:

On Mon, Jun 11, 2018 at 11:32 AM Linus Torvalds
[off-list ref] wrote:

quoted

So now we apparently do a GFP_KERNEL allocation insider the mm zone
lock, which is broken on just _so_ many levels.

Oh, I see the comment about how it doesn't actually do an allocation
at all because it's a single-entry.

Still too damn ugly to live, and much too fragile. No way in hell do
we even _hint_ at a GFP_KERNEL when we're inside a context that can't
do any memory allocation at all.

Plus I'm not convinced it's a "no allocation" path even despite that
comment, because it also does a "dma_map_page()" etc, which can cause
allocations to do the dma mapping thing afaik. No?

Well no because DMA is triggered by the IOMMU flag and
that is always off for the balloon. But I hear what you
are saying about it being fragile.

Maybe there's some reason why that doesn't happen either, but
basically this whole callchain looks *way* to complicated to be used
under a core VM spinlock.

                Linus

Maybe it will help to have GFP_NONE which will make any allocation
fail if attempted. Linus, would this address your comment?
-- 
MST