046d91c83c
With the thread pool disabled, we add the requests in the queue to a GList, processing by iterating over there afterwards. For adding them, we're using "g_list_prepend()", which is more efficient but causes the requests to be processed in reverse order, breaking the read-ahead and request-merging optimizations in the host for sequential operations. According to the documentation, if you need to process the request in-order, using "g_list_prepend()" and then reversing the list with "g_list_reverse()" is more efficient than using "g_list_append()", so let's do it that way. Testing on a spinning disk (to boost the increase of read-ahead and request-merging) shows a 4x improvement on sequential write fio test: Test: fio --directory=/mnt/virtio-fs --filename=fio-file1 --runtime=20 --iodepth=16 --size=4G --direct=1 --blocksize=4K --ioengine libaio --rw write --name seqwrite-libaio Without "g_list_reverse()": ... Jobs: 1 (f=1): [W(1)][100.0%][w=22.4MiB/s][w=5735 IOPS][eta 00m:00s] seqwrite-libaio: (groupid=0, jobs=1): err= 0: pid=710: Tue Aug 24 12:58:16 2021 write: IOPS=5709, BW=22.3MiB/s (23.4MB/s)(446MiB/20002msec); 0 zone resets ... With "g_list_reverse()": ... Jobs: 1 (f=1): [W(1)][100.0%][w=84.0MiB/s][w=21.5k IOPS][eta 00m:00s] seqwrite-libaio: (groupid=0, jobs=1): err= 0: pid=716: Tue Aug 24 13:00:15 2021 write: IOPS=21.3k, BW=83.1MiB/s (87.2MB/s)(1663MiB/20001msec); 0 zone resets ... Signed-off-by: Sergio Lopez <slp@redhat.com> Message-Id: <20210824131158.39970-1-slp@redhat.com> Reviewed-by: Vivek Goyal <vgoyal@redhat.com> Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com> |
||
---|---|---|
.. | ||
50-qemu-virtiofsd.json.in | ||
buffer.c | ||
fuse_common.h | ||
fuse_i.h | ||
fuse_log.c | ||
fuse_log.h | ||
fuse_lowlevel.c | ||
fuse_lowlevel.h | ||
fuse_misc.h | ||
fuse_opt.c | ||
fuse_opt.h | ||
fuse_signals.c | ||
fuse_virtio.c | ||
fuse_virtio.h | ||
helper.c | ||
meson.build | ||
passthrough_helpers.h | ||
passthrough_ll.c | ||
passthrough_seccomp.c | ||
passthrough_seccomp.h |