| Message ID | 1361446379-7970-1-git-send-email-jlayton@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Feb 21, 2013 at 6:32 AM, Jeff Layton <jlayton@redhat.com> wrote: > > ...as promised for 3.9. > > Signed-off-by: Jeff Layton <jlayton@redhat.com> > --- > fs/cifs/connect.c | 16 +--------------- > 1 file changed, 1 insertion(+), 15 deletions(-) > > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c > index d997737..8609c42 100644 > --- a/fs/cifs/connect.c > +++ b/fs/cifs/connect.c > @@ -97,7 +97,7 @@ enum { > Opt_user, Opt_pass, Opt_ip, > Opt_unc, Opt_domain, > Opt_srcaddr, Opt_prefixpath, > - Opt_iocharset, Opt_sockopt, > + Opt_iocharset, > Opt_netbiosname, Opt_servern, > Opt_ver, Opt_vers, Opt_sec, Opt_cache, > > @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = > { > { Opt_srcaddr, "srcaddr=%s" }, > { Opt_prefixpath, "prefixpath=%s" }, > { Opt_iocharset, "iocharset=%s" }, > - { Opt_sockopt, "sockopt=%s" }, > { Opt_netbiosname, "netbiosname=%s" }, > { Opt_servern, "servern=%s" }, > { Opt_ver, "ver=%s" }, > @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, > const char *devname, > */ > cFYI(1, "iocharset set to %s", string); > break; > - case Opt_sockopt: > - string = match_strdup(args); > - if (string == NULL) > - goto out_nomem; > - > - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { > - printk(KERN_WARNING "CIFS: the " > - "sockopt=TCP_NODELAY option has > been " > - "deprecated and will be removed " > - "in 3.9\n"); > - vol->sockopt_tcp_nodelay = 1; > - } > - break; > case Opt_netbiosname: > string = match_strdup(args); > if (string == NULL) Gar. Why was TCP_NODELAY pulled? I only ask because I'm updating the socket options section of the smb.conf man page. Supposedly NODELAY can have great results in some network configurations[citation needed]. -- Peace and Blessings, -Scott. -- To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
We should do some quick performance runs to verify this - but with corking and uncorking the socket explicitly presumably TCP_NODELAY should not be needed. On Thu, Feb 21, 2013 at 9:37 AM, Scott Lovenberg <scott.lovenberg@gmail.com> wrote: > On Thu, Feb 21, 2013 at 6:32 AM, Jeff Layton <jlayton@redhat.com> wrote: >> >> ...as promised for 3.9. >> >> Signed-off-by: Jeff Layton <jlayton@redhat.com> >> --- >> fs/cifs/connect.c | 16 +--------------- >> 1 file changed, 1 insertion(+), 15 deletions(-) >> >> diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c >> index d997737..8609c42 100644 >> --- a/fs/cifs/connect.c >> +++ b/fs/cifs/connect.c >> @@ -97,7 +97,7 @@ enum { >> Opt_user, Opt_pass, Opt_ip, >> Opt_unc, Opt_domain, >> Opt_srcaddr, Opt_prefixpath, >> - Opt_iocharset, Opt_sockopt, >> + Opt_iocharset, >> Opt_netbiosname, Opt_servern, >> Opt_ver, Opt_vers, Opt_sec, Opt_cache, >> >> @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = >> { >> { Opt_srcaddr, "srcaddr=%s" }, >> { Opt_prefixpath, "prefixpath=%s" }, >> { Opt_iocharset, "iocharset=%s" }, >> - { Opt_sockopt, "sockopt=%s" }, >> { Opt_netbiosname, "netbiosname=%s" }, >> { Opt_servern, "servern=%s" }, >> { Opt_ver, "ver=%s" }, >> @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, >> const char *devname, >> */ >> cFYI(1, "iocharset set to %s", string); >> break; >> - case Opt_sockopt: >> - string = match_strdup(args); >> - if (string == NULL) >> - goto out_nomem; >> - >> - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { >> - printk(KERN_WARNING "CIFS: the " >> - "sockopt=TCP_NODELAY option has >> been " >> - "deprecated and will be removed " >> - "in 3.9\n"); >> - vol->sockopt_tcp_nodelay = 1; >> - } >> - break; >> case Opt_netbiosname: >> string = match_strdup(args); >> if (string == NULL) > > Gar. Why was TCP_NODELAY pulled? I only ask because I'm updating the > socket options section of the smb.conf man page. Supposedly NODELAY > can have great results in some network configurations[citation > needed]. > > -- > Peace and Blessings, > -Scott.
On Thu, 21 Feb 2013 10:37:21 -0500 Scott Lovenberg <scott.lovenberg@gmail.com> wrote: > On Thu, Feb 21, 2013 at 6:32 AM, Jeff Layton <jlayton@redhat.com> wrote: > > > > ...as promised for 3.9. > > > > Signed-off-by: Jeff Layton <jlayton@redhat.com> > > --- > > fs/cifs/connect.c | 16 +--------------- > > 1 file changed, 1 insertion(+), 15 deletions(-) > > > > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c > > index d997737..8609c42 100644 > > --- a/fs/cifs/connect.c > > +++ b/fs/cifs/connect.c > > @@ -97,7 +97,7 @@ enum { > > Opt_user, Opt_pass, Opt_ip, > > Opt_unc, Opt_domain, > > Opt_srcaddr, Opt_prefixpath, > > - Opt_iocharset, Opt_sockopt, > > + Opt_iocharset, > > Opt_netbiosname, Opt_servern, > > Opt_ver, Opt_vers, Opt_sec, Opt_cache, > > > > @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = > > { > > { Opt_srcaddr, "srcaddr=%s" }, > > { Opt_prefixpath, "prefixpath=%s" }, > > { Opt_iocharset, "iocharset=%s" }, > > - { Opt_sockopt, "sockopt=%s" }, > > { Opt_netbiosname, "netbiosname=%s" }, > > { Opt_servern, "servern=%s" }, > > { Opt_ver, "ver=%s" }, > > @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, > > const char *devname, > > */ > > cFYI(1, "iocharset set to %s", string); > > break; > > - case Opt_sockopt: > > - string = match_strdup(args); > > - if (string == NULL) > > - goto out_nomem; > > - > > - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { > > - printk(KERN_WARNING "CIFS: the " > > - "sockopt=TCP_NODELAY option has > > been " > > - "deprecated and will be removed " > > - "in 3.9\n"); > > - vol->sockopt_tcp_nodelay = 1; > > - } > > - break; > > case Opt_netbiosname: > > string = match_strdup(args); > > if (string == NULL) > > Gar. Why was TCP_NODELAY pulled? I only ask because I'm updating the > socket options section of the smb.conf man page. Supposedly NODELAY > can have great results in some network configurations[citation > needed]. > commit b8eed28375a43e1c9aaa9d15af2a052aae0d0725 Author: Jeff Layton <jlayton@redhat.com> Date: Tue Sep 18 16:20:35 2012 -0700 cifs: cork the socket before a send and uncork it afterward 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@samba.org> Signed-off-by: Jeff Layton <jlayton@redhat.com> Signed-off-by: Steve French <smfrench@gmail.com>
On Thu, 21 Feb 2013 10:00:07 -0600 Steve French <smfrench@gmail.com> wrote: > We should do some quick performance runs to verify this - but with > corking and uncorking the socket explicitly presumably TCP_NODELAY > should not be needed. > I would think time for that sort of investigation was before you merged the patch that said we were going to remove it. I assumed that since you took that patch, you were OK with it. In any case, I'll turn this question around... TCP_NODELAY says: Send frames as soon as possible after a sendmsg (or kernel_sendmsg in this case) TCP_CORK says: hold sending until I uncork the socket If the socket is CORK'ed then that overrides TCP_NODELAY (at least in my reading of the code). Since we always cork/uncork the socket now, what possible benefit could there be to keeping the sockopt=TCP_NODELAY option around? > On Thu, Feb 21, 2013 at 9:37 AM, Scott Lovenberg > <scott.lovenberg@gmail.com> wrote: > > On Thu, Feb 21, 2013 at 6:32 AM, Jeff Layton <jlayton@redhat.com> wrote: > >> > >> ...as promised for 3.9. > >> > >> Signed-off-by: Jeff Layton <jlayton@redhat.com> > >> --- > >> fs/cifs/connect.c | 16 +--------------- > >> 1 file changed, 1 insertion(+), 15 deletions(-) > >> > >> diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c > >> index d997737..8609c42 100644 > >> --- a/fs/cifs/connect.c > >> +++ b/fs/cifs/connect.c > >> @@ -97,7 +97,7 @@ enum { > >> Opt_user, Opt_pass, Opt_ip, > >> Opt_unc, Opt_domain, > >> Opt_srcaddr, Opt_prefixpath, > >> - Opt_iocharset, Opt_sockopt, > >> + Opt_iocharset, > >> Opt_netbiosname, Opt_servern, > >> Opt_ver, Opt_vers, Opt_sec, Opt_cache, > >> > >> @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = > >> { > >> { Opt_srcaddr, "srcaddr=%s" }, > >> { Opt_prefixpath, "prefixpath=%s" }, > >> { Opt_iocharset, "iocharset=%s" }, > >> - { Opt_sockopt, "sockopt=%s" }, > >> { Opt_netbiosname, "netbiosname=%s" }, > >> { Opt_servern, "servern=%s" }, > >> { Opt_ver, "ver=%s" }, > >> @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, > >> const char *devname, > >> */ > >> cFYI(1, "iocharset set to %s", string); > >> break; > >> - case Opt_sockopt: > >> - string = match_strdup(args); > >> - if (string == NULL) > >> - goto out_nomem; > >> - > >> - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { > >> - printk(KERN_WARNING "CIFS: the " > >> - "sockopt=TCP_NODELAY option has > >> been " > >> - "deprecated and will be removed " > >> - "in 3.9\n"); > >> - vol->sockopt_tcp_nodelay = 1; > >> - } > >> - break; > >> case Opt_netbiosname: > >> string = match_strdup(args); > >> if (string == NULL) > > > > Gar. Why was TCP_NODELAY pulled? I only ask because I'm updating the > > socket options section of the smb.conf man page. Supposedly NODELAY > > can have great results in some network configurations[citation > > needed]. > > > > -- > > Peace and Blessings, > > -Scott. > > >
On 2/21/2013 11:00 AM, Steve French wrote: > We should do some quick performance runs to verify this - but with > corking and uncorking the socket explicitly presumably TCP_NODELAY > should not be needed. I got that info from the old Samba HOWTO and Reference Guide, but it's very, very old info. I'm guessing that when matched with an older kernel and possibly older TCP/IP congestion algorithm, it might have made more of a difference than it does now. -- To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 21, 2013 at 10:27 AM, Jeff Layton <jlayton@redhat.com> wrote: > On Thu, 21 Feb 2013 10:00:07 -0600 > Steve French <smfrench@gmail.com> wrote: > >> We should do some quick performance runs to verify this - but with >> corking and uncorking the socket explicitly presumably TCP_NODELAY >> should not be needed. >> > > I would think time for that sort of investigation was before you merged > the patch that said we were going to remove it. I assumed that since > you took that patch, you were OK with it. Yes - I am ok with it and it didn't seem to hurt performance. Corking/uncorking explicitly makes more sense if the complex network layers work as expected - but the problem is that they change so best to be safe and at least try some quick tests. > In any case, I'll turn this question around... > > TCP_NODELAY says: Send frames as soon as possible after a sendmsg (or > kernel_sendmsg in this case) > > TCP_CORK says: hold sending until I uncork the socket > > If the socket is CORK'ed then that overrides TCP_NODELAY (at least in > my reading of the code). Since we always cork/uncork the socket now, > what possible benefit could there be to keeping the sockopt=TCP_NODELAY > option around? None if corking/uncorking works as one would intuitively think - but the tcp network layers under us are very complicated. (I am not objecting to the patch to remove sockopt setting - just want to do the simple checks to make sure we don't miss something)
On Thu, 21 Feb 2013 10:31:24 -0600 Steve French <smfrench@gmail.com> wrote: > On Thu, Feb 21, 2013 at 10:27 AM, Jeff Layton <jlayton@redhat.com> wrote: > > On Thu, 21 Feb 2013 10:00:07 -0600 > > Steve French <smfrench@gmail.com> wrote: > > > >> We should do some quick performance runs to verify this - but with > >> corking and uncorking the socket explicitly presumably TCP_NODELAY > >> should not be needed. > >> > > > > I would think time for that sort of investigation was before you merged > > the patch that said we were going to remove it. I assumed that since > > you took that patch, you were OK with it. > > Yes - I am ok with it and it didn't seem to hurt performance. > Corking/uncorking explicitly makes more sense if the complex network > layers work as expected - but the problem is that they change so best > to be safe and at least try some quick tests. > > > In any case, I'll turn this question around... > > > > TCP_NODELAY says: Send frames as soon as possible after a sendmsg (or > > kernel_sendmsg in this case) > > > > TCP_CORK says: hold sending until I uncork the socket > > > > If the socket is CORK'ed then that overrides TCP_NODELAY (at least in > > my reading of the code). Since we always cork/uncork the socket now, > > what possible benefit could there be to keeping the sockopt=TCP_NODELAY > > option around? > > None if corking/uncorking works as one would intuitively think - but > the tcp network layers under us are very complicated. > > (I am not objecting to the patch to remove sockopt setting - just want > to do the simple checks to make sure we don't miss something) > Fair enough, let me know what you find. Thanks,
Resending patch to a slightly broader list for last minute check if anyone objects. Although setting this particular socket option (TCP_NODELAY) may not be as useful when corking/uncorking explicitly, I want to doublecheck before removing them because there has been some utility to the server allowing override of various sockopt options. Samba server has long supported at least the following set of settable socket options (although I don't know if the defaults are frequently overridden now, by setting sockopts in smb.conf as used to be common for the server). SO_KEEPALIVE SO_REUSEADDR SO_BROADCAST TCP_NODELAY IPTOS_LOWDELAY IPTOS_THROUGHPUT SO_SNDBUF * SO_RCVBUF * SO_SNDLOWAT * SO_RCVLOWAT * * takes an integer argument rather than a boolean on/off Any objections to removing the ability to set socket options explicitly for the cifs network file system client? On Thu, Feb 21, 2013 at 5:32 AM, Jeff Layton <jlayton@redhat.com> wrote: > > ...as promised for 3.9. > > Signed-off-by: Jeff Layton <jlayton@redhat.com> > --- > fs/cifs/connect.c | 16 +--------------- > 1 file changed, 1 insertion(+), 15 deletions(-) > > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c > index d997737..8609c42 100644 > --- a/fs/cifs/connect.c > +++ b/fs/cifs/connect.c > @@ -97,7 +97,7 @@ enum { > Opt_user, Opt_pass, Opt_ip, > Opt_unc, Opt_domain, > Opt_srcaddr, Opt_prefixpath, > - Opt_iocharset, Opt_sockopt, > + Opt_iocharset, > Opt_netbiosname, Opt_servern, > Opt_ver, Opt_vers, Opt_sec, Opt_cache, > > @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = > { > { Opt_srcaddr, "srcaddr=%s" }, > { Opt_prefixpath, "prefixpath=%s" }, > { Opt_iocharset, "iocharset=%s" }, > - { Opt_sockopt, "sockopt=%s" }, > { Opt_netbiosname, "netbiosname=%s" }, > { Opt_servern, "servern=%s" }, > { Opt_ver, "ver=%s" }, > @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, > const char *devname, > */ > cFYI(1, "iocharset set to %s", string); > break; > - case Opt_sockopt: > - string = match_strdup(args); > - if (string == NULL) > - goto out_nomem; > - > - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { > - printk(KERN_WARNING "CIFS: the " > - "sockopt=TCP_NODELAY option has > been " > - "deprecated and will be removed " > - "in 3.9\n"); > - vol->sockopt_tcp_nodelay = 1; > - } > - break; > case Opt_netbiosname: > string = match_strdup(args); > if (string == NULL) > -- > 1.7.11.7 > -- Thanks, Steve -- To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, 4 Mar 2013 16:08:30 -0600 Steve French <smfrench@gmail.com> wrote: > Resending patch to a slightly broader list for last minute check if > anyone objects. Although setting this particular socket option > (TCP_NODELAY) may not be as useful when corking/uncorking explicitly, > I want to doublecheck before removing them because there has been some > utility to the server allowing override of various sockopt options. > Samba server has long supported at least the following set of settable > socket options (although I don't know if the defaults are frequently > overridden now, by setting sockopts in smb.conf as used to be common > for the server). > > SO_KEEPALIVE > SO_REUSEADDR > SO_BROADCAST > TCP_NODELAY > IPTOS_LOWDELAY > IPTOS_THROUGHPUT > SO_SNDBUF * > SO_RCVBUF * > SO_SNDLOWAT * > SO_RCVLOWAT * > > * takes an integer argument rather than a boolean on/off > > Any objections to removing the ability to set socket options > explicitly for the cifs network file system client? > A couple of points... The sockopt= option was never documented in the mount.cifs manpage and the only value it ever accepted was TCP_NODELAY. Now that we're explicitly corking the socket, TCP_NODELAY has no effect. I don't think there's any value in leaving in a "placeholder" socket= option.
On Wed, Mar 6, 2013 at 8:40 AM, Jeff Layton <jlayton@redhat.com> wrote: > On Mon, 4 Mar 2013 16:08:30 -0600 > Steve French <smfrench@gmail.com> wrote: > >> Resending patch to a slightly broader list for last minute check if >> anyone objects. Although setting this particular socket option >> (TCP_NODELAY) may not be as useful when corking/uncorking explicitly, >> I want to doublecheck before removing them because there has been some >> utility to the server allowing override of various sockopt options. >> Samba server has long supported at least the following set of settable >> socket options (although I don't know if the defaults are frequently >> overridden now, by setting sockopts in smb.conf as used to be common >> for the server). >> >> SO_KEEPALIVE >> SO_REUSEADDR >> SO_BROADCAST >> TCP_NODELAY >> IPTOS_LOWDELAY >> IPTOS_THROUGHPUT >> SO_SNDBUF * >> SO_RCVBUF * >> SO_SNDLOWAT * >> SO_RCVLOWAT * >> >> * takes an integer argument rather than a boolean on/off >> >> Any objections to removing the ability to set socket options >> explicitly for the cifs network file system client? >> > > A couple of points... > > The sockopt= option was never documented in the mount.cifs manpage and > the only value it ever accepted was TCP_NODELAY. Now that we're > explicitly corking the socket, TCP_NODELAY has no effect. I don't think > there's any value in leaving in a "placeholder" socket= option. > After doing some research, I agree. I've been updating the socket options section of the Samba smb.conf man page and the performance section of the Samba HOWTO. The more research I do the more convinced that there is almost no reason to set most socket options on a modern Linux kernel or in Samba (server or client side). One thing pointed out to me in another thread is that the smbd server sets TCP_NODELAY by default. AIUI this was because the TCP_NODELAY option used to almost double the speed of SMB in some cases due to the way Microsoft's old TCP/IP stack sent ACKs. This is all to say that I think the "sockopt" option can go altogether since it's unlikely that any socket options will be needed in the future and the only one that was supported is now useless.
diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c index d997737..8609c42 100644 --- a/fs/cifs/connect.c +++ b/fs/cifs/connect.c @@ -97,7 +97,7 @@ enum { Opt_user, Opt_pass, Opt_ip, Opt_unc, Opt_domain, Opt_srcaddr, Opt_prefixpath, - Opt_iocharset, Opt_sockopt, + Opt_iocharset, Opt_netbiosname, Opt_servern, Opt_ver, Opt_vers, Opt_sec, Opt_cache, @@ -202,7 +202,6 @@ static const match_table_t cifs_mount_option_tokens = { { Opt_srcaddr, "srcaddr=%s" }, { Opt_prefixpath, "prefixpath=%s" }, { Opt_iocharset, "iocharset=%s" }, - { Opt_sockopt, "sockopt=%s" }, { Opt_netbiosname, "netbiosname=%s" }, { Opt_servern, "servern=%s" }, { Opt_ver, "ver=%s" }, @@ -1722,19 +1721,6 @@ cifs_parse_mount_options(const char *mountdata, const char *devname, */ cFYI(1, "iocharset set to %s", string); break; - case Opt_sockopt: - string = match_strdup(args); - if (string == NULL) - goto out_nomem; - - if (strnicmp(string, "TCP_NODELAY", 11) == 0) { - printk(KERN_WARNING "CIFS: the " - "sockopt=TCP_NODELAY option has been " - "deprecated and will be removed " - "in 3.9\n"); - vol->sockopt_tcp_nodelay = 1; - } - break; case Opt_netbiosname: string = match_strdup(args); if (string == NULL)
...as promised for 3.9. Signed-off-by: Jeff Layton <jlayton@redhat.com> --- fs/cifs/connect.c | 16 +--------------- 1 file changed, 1 insertion(+), 15 deletions(-)