diff options
author | Jeremy Allison <jra@samba.org> | 2004-09-09 10:02:44 +0000 |
---|---|---|
committer | Gerald (Jerry) Carter <jerry@samba.org> | 2007-10-10 10:52:38 -0500 |
commit | 39a9c5dcc6e8c155bb102c7659460f2e738a6e13 (patch) | |
tree | 5dfcdba484875887a9aea935b9569db3d36d0970 /Roadmap | |
parent | 516753896a6eb713ccaac53e261a833d082b79e8 (diff) | |
download | samba-39a9c5dcc6e8c155bb102c7659460f2e738a6e13.tar.gz samba-39a9c5dcc6e8c155bb102c7659460f2e738a6e13.tar.bz2 samba-39a9c5dcc6e8c155bb102c7659460f2e738a6e13.zip |
r2265: Volkers change to HEAD looks very good. Commit message
from HEAD follows :
While torturing winbind a bit I found the following unfortunate behaviour:
Sending multiple requests at a high rate for a slow operation exposed that no
response comes back until the last request in the queue has been
processed. This is an unfortunate result of serially going through all sockets> that have shown to be readable or writable. All client sockets become readable> at the same time, none of them is writable. We go through them, read the
request, process the complete request. Before we enter the select system call
the next time all requests have to have completed.
This patch optimizes this by first looking at the sockets for writability. A
write on a socket that came back from select does not block, so this
additional loop might have a non-zero cost, but it can't prevent other
operations from proceeding.
After a possibly long-running winbindd_process() we directly start select()
again. To avoid starvation the currently processed client is demoted to be the> last one in the list of clients.
Jeremy.
(This used to be commit bfdeb22c69d09eb73305b6034fa6d0ec67275789)
Diffstat (limited to 'Roadmap')
0 files changed, 0 insertions, 0 deletions