Yong,
On Sun, Jun 27, 2021 at 01:38:13PM +0800, huangy81@chinatelecom.cn wrote:
From: Hyman Huang(黄勇) <huangy81@chinatelecom.cn>
the dirtyrate measurement implemented by page-sampling originally, it
is not accurate in some scenarios, so we have introduced dirty-ring
based dirtyrate measurement(maybe it will be merged soon), it fix the
accuracy of page-sampling, and more importantly, it is at the
granualrity of vcpu.
dirty-ring method can be used when dirty-ring enable, as supplementary,
we introduce dirty-bitmap method to calculating dirtyrate when dirty log
enable, so that we can also get the accurate dirtyrate if needed in the
absence of dirty-ring.
three things has done to implement the measurement:
- introduce a fresh new dirty bits named DIRTY_MEMORY_DIRTY_RATE, which
is used to store dirty bitmap after fetching it from kvm. why we do
not reuse the existing DIRTY_MEMORY_MIGRATION dirty bits is we do not
want to interfere with migration of and let implementation clear, this
is also the reason why dirty_memory be split.
DIRTY_MEMORY_DIRTY_RATE dirty bits will be filled when
memory_global_dirty_log_sync executed if GLOBAL_DIRTY_DIRTY_RATE bit
be set in the global_dirty_tracking flag.
I'm not 100% sure this is needed.
Dirty rate measurements do not care about which page is dirtied, it looks like
an overkill to introduce a new bitmap for it.
- introduce kvm_get_manual_dirty_log_protect function so that we can
probe the protect caps of kvm when calculating.
- implement dirtyrate measurement with dirty bitmap with following step:
1. start the dirty log.
2. probe the protect cap, if KVM_DIRTY_LOG_INITIALLY_SET enable, skip
skip the 1'R and do the reset page protection manually, since kvm
file bitmap with 1 bits if this cap is enabled.
3. clear the DIRTY_MEMORY_DIRTY_RATE dirty bits, prepare to store
the dirty bitmap.
4. start memory_global_dirty_log_sync and fetch dirty bitmap from kvm
5. reap the DIRTY_MEMORY_DIRTY_RATE dirty bits and do the calculation.
this patchset rebases on the commit
"migration/dirtyrate: implement dirty-ring dirtyrate calculation",
since the above feature has not been merged, so we post this patch
for the sake of RFC. ideally, this patshset may be merged after it.
I gave it a shot with some setup dirty workload, it runs well so far and also I
do get accurate numbers (200MB/s measured as 201MB/s; 300MB/s measured as
301MB/s, and so on). Looks good to me in general.
But as I mentioned above I feel like the changeset can be shrinked quite a bit
if we can drop the extra bitmap; maybe it means we can drop half of the whole
series. But it's also possible I missed something, let's see.
It'll slightly differ from dirty ring in that same page written will always
only be counted once between two dirty map sync, but that's expected. Dirty
ring "sync" more frequently (either ring full, or current 1-sec timeout in the
reaper), so it re-protects more frequently too.
I still have some other small comments, I'll go into the patches.
Thanks,