On Sunday 06 July 2003 10.52, Henrik Nordstrom wrote:
> Hi,
>
> There obviously is a problem with the close of server-side
> connections, most visible on non-persistent connections. For some
> reason there is a long delay between comm_read detecting the close
> to this being signalled back.
Found it. Was not at all connection related, but a slight design error
in the io callback queue.
The poll/select loop did not account for the io callback queue,
causing large delays if events was queued for callback by another
callback.
poll -> filedescriptor ready
read() queued
queue invoked
callback queued
poll -> waiting until timeout
queue invoked
callback called
Fixed by adding a comm_iocallbackpending() query to comm_poll/select
loops, terminating the loop quickly if there is events pending, and
also by making the main loop invoke the queue twice to deal with
these indirectly queued events in a more efficient manner. These
indirect callbacks should be rather frequent in the current design
with queued I/O operations.
Regards
Henrik
