Re: [RFC 6/7] mm: extend process_madvise syscall to support vector arrary

On Wed 29-05-19 03:08:32, Daniel Colascione wrote:

On Mon, May 27, 2019 at 12:49 AM Minchan Kim [off-list ref] wrote:

quoted

On Tue, May 21, 2019 at 12:37:26PM +0200, Michal Hocko wrote:

quoted

On Tue 21-05-19 19:26:13, Minchan Kim wrote:

quoted

On Tue, May 21, 2019 at 08:24:21AM +0200, Michal Hocko wrote:

quoted

On Tue 21-05-19 11:48:20, Minchan Kim wrote:

quoted

On Mon, May 20, 2019 at 11:22:58AM +0200, Michal Hocko wrote:

quoted

[Cc linux-api]

On Mon 20-05-19 12:52:53, Minchan Kim wrote:

quoted

Currently, process_madvise syscall works for only one address range
so user should call the syscall several times to give hints to
multiple address range.

Is that a problem? How big of a problem? Any numbers?

We easily have 2000+ vma so it's not trivial overhead. I will come up
with number in the description at respin.

Does this really have to be a fast operation? I would expect the monitor
is by no means a fast path. The system call overhead is not what it used
to be, sigh, but still for something that is not a hot path it should be
tolerable, especially when the whole operation is quite expensive on its
own (wrt. the syscall entry/exit).

What's different with process_vm_[readv|writev] and vmsplice?
If the range needed to be covered is a lot, vector operation makes senese
to me.

I am not saying that the vector API is wrong. All I am trying to say is
that the benefit is not really clear so far. If you want to push it
through then you should better get some supporting data.

I measured 1000 madvise syscall vs. a vector range syscall with 1000
ranges on ARM64 mordern device. Even though I saw 15% improvement but
absoluate gain is just 1ms so I don't think it's worth to support.
I will drop vector support at next revision.

Please do keep the vector support. Absolute timing is misleading,
since in a tight loop, you're not going to contend on mmap_sem. We've
seen tons of improvements in things like camera start come from
coalescing mprotect calls, with the gains coming from taking and
releasing various locks a lot less often and bouncing around less on
the contended lock paths. Raw throughput doesn't tell the whole story,
especially on mobile.

This will always be a double edge sword. Taking a lock for longer can
improve a throughput of a single call but it would make a latency for
anybody contending on the lock much worse.

Besides that, please do not overcomplicate the thing from the early
beginning please. Let's start with a simple and well defined remote
madvise alternative first and build a vector API on top with some
numbers based on _real_ workloads.
-- 
Michal Hocko
SUSE Labs