Re: [PATCH v2 3/9] cifs: cork the socket before a send and uncork it afterward

[Jeff Layton <jlayton-eUNUBHrolfbYtjvyW6yDsg@public.gmane.org>]

On Mon, 30 Jul 2012 21:17:53 -0400
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:

> On Mon, 30 Jul 2012 23:11:10 +0200
> Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> 
> > 2012/7/29 Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>:
> > > On Fri, 27 Jul 2012 10:05:32 +0400
> > > Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> > >
> > >> 2012/7/27 Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>:
> > >> > On Fri, 27 Jul 2012 03:57:44 +0400
> > >> > Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> > >> >
> > >> >> 2012/7/25 Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>:
> > >> >> > We want to send SMBs as "atomically" as possible. Prior to sending any
> > >> >> > data on the socket, cork it to make sure that no non-full frames go
> > >> >> > out. Afterward, uncork it to make sure all of the data gets pushed out
> > >> >> > to the wire.
> > >> >> >
> > >> >> > Note that this more or less renders the socket=TCP_NODELAY mount option
> > >> >> > obsolete. When TCP_CORK and TCP_NODELAY are used on the same socket,
> > >> >> > TCP_NODELAY is essentially ignored.
> > >> >> >
> > >> >> > Acked-by: Pavel Shilovsky <pshilovsky-eUNUBHrolfbYtjvyW6yDsg@public.gmane.org>
> > >> >> > Signed-off-by: Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> > >> >> > ---
> > >> >> >  fs/cifs/connect.c   |  4 ++++
> > >> >> >  fs/cifs/transport.c | 12 ++++++++++++
> > >> >> >  2 files changed, 16 insertions(+)
> > >> >> >
> > >> >> > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
> > >> >> > index 6df6fa1..a828a8c 100644
> > >> >> > --- a/fs/cifs/connect.c
> > >> >> > +++ b/fs/cifs/connect.c
> > >> >> > @@ -1676,6 +1676,10 @@ cifs_parse_mount_options(const char *mountdata, const char *devname,
> > >> >> >                         if (string == NULL)
> > >> >> >                                 goto out_nomem;
> > >> >> >
> > >> >> > +                       /*
> > >> >> > +                        * FIXME: since we now cork/uncork the socket while
> > >> >> > +                        *        sending, should we deprecate this option?
> > >> >> > +                        */
> > >> >> >                         if (strnicmp(string, "TCP_NODELAY", 11) == 0)
> > >> >> >                                 vol->sockopt_tcp_nodelay = 1;
> > >> >> >                         break;
> > >> >> > diff --git a/fs/cifs/transport.c b/fs/cifs/transport.c
> > >> >> > index d93f15d..a3e58b2 100644
> > >> >> > --- a/fs/cifs/transport.c
> > >> >> > +++ b/fs/cifs/transport.c
> > >> >> > @@ -27,6 +27,7 @@
> > >> >> >  #include <linux/net.h>
> > >> >> >  #include <linux/delay.h>
> > >> >> >  #include <linux/freezer.h>
> > >> >> > +#include <linux/tcp.h>
> > >> >> >  #include <asm/uaccess.h>
> > >> >> >  #include <asm/processor.h>
> > >> >> >  #include <linux/mempool.h>
> > >> >> > @@ -247,12 +248,23 @@ smb_send_rqst(struct TCP_Server_Info *server, struct smb_rqst *rqst)
> > >> >> >         int n_vec = rqst->rq_nvec;
> > >> >> >         unsigned int smb_buf_length = get_rfc1002_length(iov[0].iov_base);
> > >> >> >         size_t total_len;
> > >> >> > +       struct socket *ssocket = server->ssocket;
> > >> >> > +       int val = 1;
> > >> >> >
> > >> >> >         cFYI(1, "Sending smb: smb_len=%u", smb_buf_length);
> > >> >> >         dump_smb(iov[0].iov_base, iov[0].iov_len);
> > >> >> >
> > >> >> > +       /* cork the socket */
> > >> >> > +       kernel_setsockopt(ssocket, SOL_TCP, TCP_CORK,
> > >> >> > +                               (char *)&val, sizeof(val));
> > >> >> > +
> > >> >> >         rc = smb_send_kvec(server, iov, n_vec, &total_len);
> > >> >> >
> > >> >> > +       /* uncork it */
> > >> >> > +       val = 0;
> > >> >> > +       kernel_setsockopt(ssocket, SOL_TCP, TCP_CORK,
> > >> >> > +                               (char *)&val, sizeof(val));
> > >> >> > +
> > >> >> >         if ((total_len > 0) && (total_len != smb_buf_length + 4)) {
> > >> >> >                 cFYI(1, "partial send (wanted=%u sent=%zu): terminating "
> > >> >> >                         "session", smb_buf_length + 4, total_len);
> > >> >> > --
> > >> >> > 1.7.11.2
> > >> >> >
> > >> >> > --
> > >> >> > To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
> > >> >> > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > >> >> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > >> >>
> > >> >> I tested it with SMB2 against Windows 7 server. When iosize is 64K
> > >> >> everything is ok but when we increase iosize to 1M (by using
> > >> >> multicredit requests) and the server loses the network connection and
> > >> >> only reboot helps.
> > >> >>
> > >> >> Also if I commented corking/uncorking the socket - everything is ok. I
> > >> >> think this change needs some more investigation (how does it deals
> > >> >> with 1M iosize on Samba, etc?)
> > >> >>
> > >> >
> > >> > Hmm, haven't seen that with a 1M iosize with smb1 against samba.
> > >> >
> > >> > I'll see if I can reproduce it.
> > >> >
> > >> > --
> > >> > Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> > >>
> > >> Forgot to mentioned how I reproduce it - dbench with 5 clients.
> > >>
> > >
> > > Ok, I've built a cifs.ko from your smb2-dev-cifs-3.6 branch. Here are my mount options:
> > >
> > > //win7.poochiereds.net/export /mnt/win7 cifs rw,relatime,vers=2.1,sec=ntlmsspi,cache=strict,unc=\\win7.poochiereds.net\export,username=testuser,domain=WIN7,uid=0,noforceuid,gid=0,noforcegid,addr=192.168.1.32,file_mode=0755,dir_mode=0755,nounix,serverino,rsize=1048576,wsize=1048576,actimeo=1 0 0
> > 
> > I don't set rsize and wsize explicitly but I don't think it's related.
> > On what connection did you test it? I use 100Mbit LAN.
> > 
> 
> The clients and servers are both KVM guests. I'll give it a go over a
> physical network tomorrow.
> 

Ok, ran it with client as a KVM guest and the server as win7 running on
bare metal over a gigE network. It still ran just fine.

So what are the symptoms that you see here? Does dbench just hang? If
so, could you collect /proc/<pid>/stack from the hung process(es)?
Maybe that would tell us what's going on...

-- 
Jeff Layton <jlayton-eUNUBHrolfbYtjvyW6yDsg@public.gmane.org>