summaryrefslogtreecommitdiffstats
path: root/include/linux/sh_clk.h
diff options
context:
space:
mode:
authorTakuya Yoshikawa <takuya.yoshikawa@gmail.com>2012-03-03 14:21:48 +0900
committerAvi Kivity <avi@redhat.com>2012-04-08 12:50:00 +0300
commit60c34612b70711fb14a8dcbc6a79509902450d2e (patch)
treeaaf14cd91200a6b28cb6e720a61e8a4d31cec1d9 /include/linux/sh_clk.h
parent5dc99b2380d59b8aeafa98791f92b96400ed3187 (diff)
downloadblackbird-op-linux-60c34612b70711fb14a8dcbc6a79509902450d2e.tar.gz
blackbird-op-linux-60c34612b70711fb14a8dcbc6a79509902450d2e.zip
KVM: Switch to srcu-less get_dirty_log()
We have seen some problems of the current implementation of get_dirty_log() which uses synchronize_srcu_expedited() for updating dirty bitmaps; e.g. it is noticeable that this sometimes gives us ms order of latency when we use VGA displays. Furthermore the recent discussion on the following thread "srcu: Implement call_srcu()" http://lkml.org/lkml/2012/1/31/211 also motivated us to implement get_dirty_log() without SRCU. This patch achieves this goal without sacrificing the performance of both VGA and live migration: in practice the new code is much faster than the old one unless we have too many dirty pages. Implementation: The key part of the implementation is the use of xchg() operation for clearing dirty bits atomically. Since this allows us to update only BITS_PER_LONG pages at once, we need to iterate over the dirty bitmap until every dirty bit is cleared again for the next call. Although some people may worry about the problem of using the atomic memory instruction many times to the concurrently accessible bitmap, it is usually accessed with mmu_lock held and we rarely see concurrent accesses: so what we need to care about is the pure xchg() overheads. Another point to note is that we do not use for_each_set_bit() to check which ones in each BITS_PER_LONG pages are actually dirty. Instead we simply use __ffs() in a loop. This is much faster than repeatedly call find_next_bit(). Performance: The dirty-log-perf unit test showed nice improvements, some times faster than before, except for some extreme cases; for such cases the speed of getting dirty page information is much faster than we process it in the userspace. For real workloads, both VGA and live migration, we have observed pure improvements: when the guest was reading a file during live migration, we originally saw a few ms of latency, but with the new method the latency was less than 200us. Signed-off-by: Takuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp> Signed-off-by: Avi Kivity <avi@redhat.com>
Diffstat (limited to 'include/linux/sh_clk.h')
0 files changed, 0 insertions, 0 deletions
OpenPOWER on IntegriCloud