On 21 Jan 2021, at 12:24, Anna Schumaker wrote:
Hi Ben,
On Wed, Jan 20, 2021 at 12:04 PM Benjamin Coddington
<bcodding@xxxxxxxxxx> wrote:
It is now possible that a reader will resume a directory listing
after an
invalidation and fill the rest of the pages with the offset left over
from
the last partially-filled page. These pages will then be recycled
and
refilled by the next reader since their alignment is incorrect.
Add an index to the nfs_cache_array that will indicate where the next
entry
should be filled. This allows partially-filled pages to have the
best
alignment possible. They are more likely to be useful to readers
that
follow.
This optimization targets the case when there are multiple processes
listing the directory simultaneously. Often the processes will
collect and
block on the same page waiting for a READDIR call to fill the
pagecache.
If the pagecache is invalidated, a partially-filled page will usually
result. This partially-filled page can immediately be used by all
processes to emit entries rather than having to discard and refill it
for
every process.
The addition of another integer to struct nfs_cache_array increases
its
size to 24 bytes. We do not lose the original capacity of 127 entries
per
page.
This patch causes cthon basic test #6 to start failing with unexpected
dir entries across all NFS versions:
./test6: readdir
basic tests failed
./test6: (/mnt/test/anna/Connectathon/cthon04) unexpected dir
entry 'h'
Ah, yes -- because we /must/ revalidate before using any cached pages.
I
think this test unlinks files and that would be a breakage.
Luckily, the next patch seems to resolve this issue. Could they maybe
be squashed together?
Yes, they should be squashed - I'll do that in another pass. I'd like
to
get rid of uncached_readdir() altogether and then send another version.
Thanks for the test!
Ben
