Eric Blake
2019-Jun-20 11:50 UTC
[Libguestfs] [libnbd PATCH] docs: Improve nbd_aio_get_direction documentation
Mention things to remember for avoiding the deadlock of polling for a POLLIN from a server that has no replies to send. Perhaps we could split the READY state into two - one for when there are no commands in flight (and get_direction returns 0 to state that polling is pointless, although a multi-threaded reader can still poll for POLLIN), and the other when there ARE commands in flight. Such a separation would also make it easier to know when nbd_aio_disconnect is not going to strand an in-flight command's response. But I'm not sure it is worth the effort. --- I'm pushing this one now, but sending the email because of the question it raises on whether we want the user to be able to distinguish a difference between any vs. no in-flight commands. generator/generator | 25 ++++++++++++++++++------- 1 file changed, 18 insertions(+), 7 deletions(-) diff --git a/generator/generator b/generator/generator index a289741..1d4db23 100755 --- a/generator/generator +++ b/generator/generator @@ -1719,6 +1719,18 @@ We are expected next to read from the server. If using L<poll(2)> you would set C<events = POLLIN>. If C<revents> returns C<POLLIN> you would then call C<nbd_aio_notify_read>. +Note that once libnbd reaches C<nbd_aio_is_ready>, this direction is +returned even before a command is issued via C<nbd_aio_pwrite> and +friends. In a single-threaded use of libnbd, it is not worth polling +until after issuing a command, as otherwise the server will never wake +up the poll. In a multi-threaded scenario, you can have one thread +begin a polling loop prior to any commands, but any other thread that +issues a command will need a way to kick the polling thread out of +poll in case issuing the command changes the needed polling +direction. Possible ways to do this include polling for activity on a +pipe-to-self, or using L<pthread_kill(3)> to send a signal that is +masked except during L<ppoll(2)>. + =item C<LIBNBD_AIO_DIRECTION_WRITE> = 2 We are expected next to write to the server. If using L<poll(2)> @@ -1727,13 +1739,12 @@ you would then call C<nbd_aio_notify_write>. =item C<LIBNBD_AIO_DIRECTION_BOTH> = 3 -We are expected next to either read or write to the server. -If using L<poll(2)> you would set C<events = POLLIN|POLLOUT>. -If one of C<POLLIN> or C<POLLOUT> is returned, then see above. -However note that you shouldn't call C<nbd_aio_notify_read> -and C<nbd_aio_notify_write> because calling the first one will -change the state of the connection, possibly invalidating the second -notification. +We are expected next to either read or write to the server. If using +L<poll(2)> you would set C<events = POLLIN|POLLOUT>. If only one of +C<POLLIN> or C<POLLOUT> is returned, then see above. However, if both +are returned, it is better to call only C<nbd_aio_notify_read>, as +processing the server's reply may change the state of the connection +and invalidate the need to write more commands. =back"; }; -- 2.20.1
Eric Blake
2019-Jun-20 21:33 UTC
Re: [Libguestfs] [libnbd PATCH] docs: Improve nbd_aio_get_direction documentation
On 6/20/19 6:50 AM, Eric Blake wrote:> Mention things to remember for avoiding the deadlock of polling for a > POLLIN from a server that has no replies to send. > > Perhaps we could split the READY state into two - one for when there > are no commands in flight (and get_direction returns 0 to state that > polling is pointless, although a multi-threaded reader can still poll > for POLLIN), and the other when there ARE commands in flight. Such a > separation would also make it easier to know when nbd_aio_disconnect > is not going to strand an in-flight command's response. But I'm not > sure it is worth the effort. > --- > > I'm pushing this one now, but sending the email because of the > question it raises on whether we want the user to be able to > distinguish a difference between any vs. no in-flight commands.Actually, I'm not pushing it yet. I noticed that we can also return 0 (such as on a just-created handle), and need to document that as well. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
Richard W.M. Jones
2019-Jun-25 09:08 UTC
Re: [Libguestfs] [libnbd PATCH] docs: Improve nbd_aio_get_direction documentation
On Thu, Jun 20, 2019 at 04:33:11PM -0500, Eric Blake wrote:> On 6/20/19 6:50 AM, Eric Blake wrote: > > Mention things to remember for avoiding the deadlock of polling for a > > POLLIN from a server that has no replies to send. > > > > Perhaps we could split the READY state into two - one for when there > > are no commands in flight (and get_direction returns 0 to state that > > polling is pointless, although a multi-threaded reader can still poll > > for POLLIN), and the other when there ARE commands in flight. Such a > > separation would also make it easier to know when nbd_aio_disconnect > > is not going to strand an in-flight command's response. But I'm not > > sure it is worth the effort. > > --- > > > > I'm pushing this one now, but sending the email because of the > > question it raises on whether we want the user to be able to > > distinguish a difference between any vs. no in-flight commands. > > Actually, I'm not pushing it yet. I noticed that we can also return 0 > (such as on a just-created handle), and need to document that as well.Patch is still good, with or without the extra documentation, so ACK anyway. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-builder quickly builds VMs from scratch http://libguestfs.org/virt-builder.1.html
Maybe Matching Threads
- [libnbd] Simultaneous read and write
- [PATCH libnbd v2 1/4] examples, tests: Remove want_to_send / ready logic, increase limit on cmds in flight.
- [libnbd] Simultaneous read and write
- [libnbd PATCH 2/2] poll: Improve our interface
- [nbdkit PATCH] Experiment: nbd: Use ppoll() instead of pipe-to-self