K 10
svn:author
V 7
glebius
K 8
svn:date
V 27
2014-03-07T22:53:29.532395Z
K 7
svn:log
V 887
Provide a hack to workaround the following condition: two sendfile(2)
calls are issues on the same object, with ranges overlapping on at
least one page. One of them grabs all pages, schedules I/O and returns.
The second one encounters an overlapping page in a busy state. It
sleeps with 'pgrbwt' wait channel, and sleep time is effectively the
I/O time. Thus, second syscall is degenerating to the blocking on
disk I/O mode.

To avoid this, a sysctl kern.ipc.sendfile.pgrabnowait is provided
(default to off). When sysctl is on and userland supplies SF_NODISKIO
flag, then we are calling vm_page_grab(VM_ALLOC_NOWAIT). In case of
a failure we return EAGAIN, hinting userland that it should continue
monitor the socket via select/kevent/whatever. If the socket has zero
data to send, then we have to notify it immediately, to avoid stall.

Sponsored by:	Netflix
Sponsored by:	Nginx, Inc.

END