Re: [PATCH v3 1/2] io: Add support for MSG_PEEK for socket channel

[Peter Xu]

On Tue, Nov 22, 2022 at 10:12:25PM +0530, manish.mishra wrote:
> 
> On 22/11/22 10:03 pm, Peter Xu wrote:
> > On Tue, Nov 22, 2022 at 11:29:05AM -0500, Peter Xu wrote:
> > > On Tue, Nov 22, 2022 at 11:10:18AM -0500, Peter Xu wrote:
> > > > On Tue, Nov 22, 2022 at 09:01:59PM +0530, manish.mishra wrote:
> > > > > On 22/11/22 8:19 pm, Daniel P. Berrangé wrote:
> > > > > > On Tue, Nov 22, 2022 at 09:41:01AM -0500, Peter Xu wrote:
> > > > > > > On Tue, Nov 22, 2022 at 02:38:53PM +0530, manish.mishra wrote:
> > > > > > > > On 22/11/22 2:30 pm, Daniel P. Berrangé wrote:
> > > > > > > > > On Sat, Nov 19, 2022 at 09:36:14AM +0000, manish.mishra wrote:
> > > > > > > > > > MSG_PEEK reads from the peek of channel, The data is 
> > > > > > > > > > treated as
> > > > > > > > > > unread and the next read shall still return this data. This
> > > > > > > > > > support is currently added only for socket class. Extra 
> > > > > > > > > > parameter
> > > > > > > > > > 'flags' is added to io_readv calls to pass extra read flags 
> > > > > > > > > > like
> > > > > > > > > > MSG_PEEK.
> > > > > > > > > > 
> > > > > > > > > > Suggested-by: Daniel P. Berrangé <berrange@redhat.com
> > > > > > > > > > Signed-off-by: manish.mishra<manish.mishra@nutanix.com>
> > > > > > > > > > ---
> > > > > > > > > >     chardev/char-socket.c               |  4 +-
> > > > > > > > > >     include/io/channel.h                | 83 
> > > > > > > > > > +++++++++++++++++++++++++++++
> > > > > > > > > >     io/channel-buffer.c                 |  1 +
> > > > > > > > > >     io/channel-command.c                |  1 +
> > > > > > > > > >     io/channel-file.c                   |  1 +
> > > > > > > > > >     io/channel-null.c                   |  1 +
> > > > > > > > > >     io/channel-socket.c                 | 16 +++++-
> > > > > > > > > >     io/channel-tls.c                    |  1 +
> > > > > > > > > >     io/channel-websock.c                |  1 +
> > > > > > > > > >     io/channel.c                        | 73 
> > > > > > > > > > +++++++++++++++++++++++--
> > > > > > > > > >     migration/channel-block.c           |  1 +
> > > > > > > > > >     scsi/qemu-pr-helper.c               |  2 +-
> > > > > > > > > >     tests/qtest/tpm-emu.c               |  2 +-
> > > > > > > > > >     tests/unit/test-io-channel-socket.c |  1 +
> > > > > > > > > >     util/vhost-user-server.c            |  2 +-
> > > > > > > > > >     15 files changed, 179 insertions(+), 11 deletions(-)
> > > > > > > > > > diff --git a/io/channel-socket.c b/io/channel-socket.c
> > > > > > > > > > index b76dca9cc1..a06b24766d 100644
> > > > > > > > > > --- a/io/channel-socket.c
> > > > > > > > > > +++ b/io/channel-socket.c
> > > > > > > > > > @@ -406,6 +406,8 @@ 
> > > > > > > > > > qio_channel_socket_accept(QIOChannelSocket *ioc,
> > > > > > > > > >         }
> > > > > > > > > >     #endif /* WIN32 */
> > > > > > > > > > +    qio_channel_set_feature(QIO_CHANNEL(cioc), 
> > > > > > > > > > QIO_CHANNEL_FEATURE_READ_MSG_PEEK);
> > > > > > > > > > +
> > > > > > > > > This covers the incoming server side socket.
> > > > > > > > > 
> > > > > > > > > This also needs to be set in outgoing client side socket in
> > > > > > > > > qio_channel_socket_connect_async
> > > > > > > > Yes sorry, i considered only current use-case, but as it is 
> > > > > > > > generic one both should be there. Thanks will update it.
> > > > > > > > 
> > > > > > > > > > @@ -705,7 +718,6 @@ static ssize_t 
> > > > > > > > > > qio_channel_socket_writev(QIOChannel *ioc,
> > > > > > > > > >     }
> > > > > > > > > >     #endif /* WIN32 */
> > > > > > > > > > -
> > > > > > > > > >     #ifdef QEMU_MSG_ZEROCOPY
> > > > > > > > > >     static int qio_channel_socket_flush(QIOChannel *ioc,
> > > > > > > > > >                                         Error **errp)
> > > > > > > > > Please remove this unrelated whitespace change.
> > > > > > > > > 
> > > > > > > > > 
> > > > > > > > > > @@ -109,6 +117,37 @@ int 
> > > > > > > > > > qio_channel_readv_all_eof(QIOChannel *ioc,
> > > > > > > > > >         return qio_channel_readv_full_all_eof(ioc, iov, 
> > > > > > > > > > niov, NULL, NULL, errp);
> > > > > > > > > >     }
> > > > > > > > > > +int qio_channel_readv_peek_all_eof(QIOChannel *ioc,
> > > > > > > > > > +                                   const struct iovec *iov,
> > > > > > > > > > +                                   size_t niov,
> > > > > > > > > > +                                   Error **errp)
> > > > > > > > > > +{
> > > > > > > > > > +   ssize_t len = 0;
> > > > > > > > > > +   ssize_t total = iov_size(iov, niov);
> > > > > > > > > > +
> > > > > > > > > > +   while (len < total) {
> > > > > > > > > > +       len = qio_channel_readv_full(ioc, iov, niov, NULL,
> > > > > > > > > > +                                    NULL, 
> > > > > > > > > > QIO_CHANNEL_READ_FLAG_MSG_PEEK, errp);
> > > > > > > > > > +
> > > > > > > > > > +       if (len == QIO_CHANNEL_ERR_BLOCK) {
> > > > > > > > > > +            if (qemu_in_coroutine()) {
> > > > > > > > > > +                qio_channel_yield(ioc, G_IO_IN);
> > > > > > > > > > +            } else {
> > > > > > > > > > +                qio_channel_wait(ioc, G_IO_IN);
> > > > > > > > > > +            }
> > > > > > > > > > +            continue;
> > > > > > > > > > +       }
> > > > > > > > > > +       if (len == 0) {
> > > > > > > > > > +           return 0;
> > > > > > > > > > +       }
> > > > > > > > > > +       if (len < 0) {
> > > > > > > > > > +           return -1;
> > > > > > > > > > +       }
> > > > > > > > > > +   }
> > > > > > > > > This will busy wait burning CPU where there is a read > 0 and 
> > > > > > > > > < total.
> > > > > > > > > 
> > > > > > > > Daniel, i could use MSG_WAITALL too if that works but then we 
> > > > > > > > will lose opportunity to yield. Or if you have some other idea.
> > > > > > > How easy would this happen?
> > > > > > > 
> > > > > > > Another alternative is we could just return the partial len to 
> > > > > > > caller then
> > > > > > > we fallback to the original channel orders if it happens.  And 
> > > > > > > then if it
> > > > > > > mostly will never happen it'll behave merely the same as what we 
> > > > > > > want.
> > > > > > Well we're trying to deal with a bug where the slow and/or 
> > > > > > unreliable
> > > > > > network causes channels to arrive in unexpected order. Given we know
> > > > > > we're having network trouble, I wouldn't want to make more 
> > > > > > assumptions
> > > > > > about things happening correctly.
> > > > > > 
> > > > > > 
> > > > > > With regards,
> > > > > > Daniel
> > > > > 
> > > > > Peter, I have seen MSG_PEEK used in combination with MSG_WAITALL, but 
> > > > > looks like even though chances are less it can still return partial 
> > > > > data even with multiple retries for signal case, so is not full proof.
> > > > > 
> > > > > *MSG_WAITALL *(since Linux 2.2)
> > > > >                This flag requests that the operation block until the 
> > > > > full
> > > > >                request is satisfied.  However, the call may still 
> > > > > return
> > > > >                less data than requested if a signal is caught, an 
> > > > > error
> > > > >                or disconnect occurs, or the next data to be received 
> > > > > is
> > > > >                of a different type than that returned.  This flag has 
> > > > > no
> > > > >                effect for datagram sockets.
> > > > > 
> > > > > Actual read ahead will be little hackish, so just confirming we all 
> > > > > are in agreement to do actual read ahead and i can send patch? :)
> > > > Yet another option is the caller handles partial PEEK and then we can 
> > > > sleep
> > > > in the migration code before another PEEK attempt until it reaches the 
> > > > full
> > > > length.
> > > > 
> > > > Even with that explicit sleep code IMHO it is cleaner than the 
> > > > read-header
> > > > flag plus things like !tls check just to avoid the handshake dead lock
> > > > itself (and if to go with this route we'd better also have a full 
> > > > document
> > > > on why !tls, aka, how the dead lock can happen).
> > > Nah, I forgot we're in the same condition as in the main thread.. sorry.
> > > 
> > > Then how about using qemu_co_sleep_ns_wakeable() to replace
> > > qio_channel_yield() either above, or in the caller?
> > A better one is qemu_co_sleep_ns().  Off-topic: I'd even think we should
> > have one qemu_co_sleep_realtime_ns() because currently all callers of
> I am not aware of this :) , will check it.
> > qemu_co_sleep_ns() is for the rt clock.
> 
> 
> Yes that also works Peter. In that case, should i have a default time or take 
> it from upper layers. And for live migration does something like of scale 1ms 
> works?

Sounds good to me on migration side.  When making it formal we'd also want
to know how Juan/Dave think.

But let's also wait for Dan's input about this before going forward.  If
the io code wants an _eof() version of PEEK then maybe we'd better do the
timeout-yield there even if not as elegant as G_IO_IN.  IIUC it's a matter
of whether we want to allow the PEEK interface return partial len.

Thanks,

-- 
Peter Xu