On 11/27/2017 09:42 AM, Victor Kaplansky wrote:
----- Original Message -----
From: "Maxime Coquelin" <maxime.coque...@redhat.com>
To: "Victor Kaplansky" <vkapl...@redhat.com>
Cc: dev@dpdk.org, y...@fridaylinux.org, "tiwei bie" <tiwei....@intel.com>, "jianfeng
tan" <jianfeng....@intel.com>,
sta...@dpdk.org, jfrei...@redhat.com
Sent: Monday, November 27, 2017 10:27:22 AM
Subject: Re: [PATCH v2 2/3] vhost: protect dirty logging against logging base
change
Hi Victor,
On 11/27/2017 09:16 AM, Victor Kaplansky wrote:
Hi,
While I agree that taking full fledged lock by rte_rwlock_read_lock()
solves the race condition,
I'm afraid that it would be too expensive in case when logging is off,
since it introduces
acquiring and releasing lock into the main flow of ring updates.
Actually my v2 fixes the performance penalty when logging is off. The
lock is now taken after the logging feature check.
But still, I agree logging on case will suffer from a performance
penalty.
Yes, checking of logging feature is better than nothing, but VHOST_F_LOG_ALL
marks only whether logging is supported by the device and not if
logging is in the action. Thus, any guest will hit the performance
degradation even not during migration.
My understanding is that VHOST_USER_SET_FEATURES is called again with
VHOST_F_LOG_ALL on on migration start and with VHOST_F_LOG_ALL off on
migration stop.
It is OK for now, as it fixes the bug, but we need to perform more careful
performance measurements,
and see whether the performance degradation is not too prohibitive.
As alternative, we may consider using more light weighted busy looping.
I think it will end up almost being the same, as both threads will need
to busy loop. PMD thread to be sure the protocol thread isn't being
unmapping the region before doing the logging, and protocol thread to be
sure the PMD thread is not doing logging before handling the set log
base.
I'm not fully aware how rte_rwlock_read_lock() is implemented, but
theoretically busy looping should be much cheaper in cases when
taking lock by one side is very rare.
we could improve it by only taking the lock once per burst instead of
per page logging, as we don't care the protocol thread waits a bit more
when it wants to remap the area.
Maybe you have something else in mind?
Also, lets fix by this series the __sync_fetch_and_or_8 ->
__sync_fetch_and_or,
as it may improve the performance slightly.
Sure, this can be done, but it would need to be benchmarked first.
Agree.
Regards,
Maxime
