On Mon, Jul 31, 2017 at 10:37:24AM +0200, Michal Hocko wrote:

On Mon 31-07-17 16:23:26, ZhenweiPi wrote:

quoted

On 07/31/2017 03:51 PM, Michal Hocko wrote:

quoted

On Mon 31-07-17 15:41:49, Wei Wang wrote:

quoted

quoted

On 07/31/2017 02:55 PM, Michal Hocko wrote:

quoted

quoted

On Mon 31-07-17 12:13:33, Wei Wang wrote:

quoted

quoted

Ballooned pages will be marked as MADV_DONTNEED by the hypervisor and
shouldn't be given to the host ksmd to scan.

Could you point me where this MADV_DONTNEED is done, please?

Sure. It's done in the hypervisor when the balloon pages are received.

Please see line 40 at
https://github.com/qemu/qemu/blob/master/hw/virtio/virtio-balloon.c

And one more thing. I am not familiar with ksm much. But how is
MADV_DONTNEED even helping? This madvise is not sticky - aka it will
unmap the range without leaving any note behind. AFAICS the only way
to have vma scanned is to have VM_MERGEABLE and that is an opt in:
See Documentation/vm/ksm.txt
"
KSM only operates on those areas of address space which an application
has advised to be likely candidates for merging, by using the madvise(2)
system call: int madvise(addr, length, MADV_MERGEABLE).
"

So what exactly is going on here? The original patch looks highly
suspicious as well. If somebody wants to make that memory mergable then
the user of that memory should zero them out.

Kernel starts a kthread named "ksmd". ksmd scans the VM_MERGEABLE
memory, and merge the same pages.(same page means memcmp(page1,
page2, PAGESIZE) == 0).

Guest can not use ballooned pages, and these pages will not be accessed
in a long time. Kswapd on host will swap these pages out and get more
free memory.

Rather than swapping, KSM has better performence.  Presently pages in
the balloon device have random value,  they usually cannot be merged.
So enqueue zero pages will resolve this problem.

Because MADV_DONTNEED depends on host os capability and hypervisor capability,
I prefer to enqueue zero pages to balloon device and made this patch.

I think you should have hypervisor zero them out if it wants to then. Seems cleaner.

So why exactly are we zeroying pages (and pay some cost for that) in
guest when we do not know what host actually does with them?

I suspect this is some special hypervisor that somehow benefits from
this patch. It should just use a feature bit for its special needs
I think.

Michal is also exactly right that patches like this should come
with some performance numbers.
I'll post a patch adding virtio lists for mm/balloon_compaction.c
so that we notice when people tweak it like that.

-- 
Michal Hocko
SUSE Labs