diff mbox series

[6/7] cifs: Fix creating of SFU fifo and socket special files

Message ID 20240912120548.15877-7-pali@kernel.org
State New
Headers show
Series cifs: Improve client SFU support for special files | expand

Commit Message

Pali Rohár Sept. 12, 2024, 12:05 p.m. UTC
SFU-style fifo is empty file with system attribute set. This format is used
by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
(which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).

SFU-style socket is file which has system attribute set and file content is
one zero byte. This format was introduced in Interix 3.0 subsystem, as part
of the Microsoft SFU 3.0 and is used also by all later versions. Previous
versions had no UNIX domain socket support.

Currently when sfu mount option is specified then CIFS creates fifo and
socket special files with some strange LnxSOCK or LnxFIFO content which is
not compatible with Microsoft SFU and neither recognized by other SMB
implementations which have some SFU support, including older Linux cifs
implementations.

So when sfu mount option is specified, create all fifo and socket special
files compatible with SFU format to achieve SFU interop, as it is expected
by sfu mount option.

Signed-off-by: Pali Rohár <pali@kernel.org>
---
 fs/smb/client/cifssmb.c |  8 ++++----
 fs/smb/client/smb1ops.c |  2 +-
 fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
 3 files changed, 24 insertions(+), 15 deletions(-)

Comments

Pali Rohár Sept. 13, 2024, 8:07 p.m. UTC | #1
On Thursday 12 September 2024 14:05:47 Pali Rohár wrote:
> SFU-style fifo is empty file with system attribute set. This format is used
> by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
> (which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).
> 
> SFU-style socket is file which has system attribute set and file content is
> one zero byte. This format was introduced in Interix 3.0 subsystem, as part
> of the Microsoft SFU 3.0 and is used also by all later versions. Previous
> versions had no UNIX domain socket support.
> 
> Currently when sfu mount option is specified then CIFS creates fifo and
> socket special files with some strange LnxSOCK or LnxFIFO content which is
> not compatible with Microsoft SFU and neither recognized by other SMB
> implementations which have some SFU support, including older Linux cifs
> implementations.
> 
> So when sfu mount option is specified, create all fifo and socket special
> files compatible with SFU format to achieve SFU interop, as it is expected
> by sfu mount option.
> 
> Signed-off-by: Pali Rohár <pali@kernel.org>

Fixes: 72bc63f5e23a ("smb3: fix creating FIFOs when mounting with "sfu" mount option")
Fixes: 518549c120e6 ("cifs: fix creating sockets when using sfu mount options")

I located commits which introduced those strange LnxSOCK or LnxFIFO
types which are not compatible with SFU. I would suggest to add those
two Fixes: tags into commit message for reference.

> ---
>  fs/smb/client/cifssmb.c |  8 ++++----
>  fs/smb/client/smb1ops.c |  2 +-
>  fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
>  3 files changed, 24 insertions(+), 15 deletions(-)
> 
> diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
> index cfae2e918209..0ffc45aa5e2c 100644
> --- a/fs/smb/client/cifssmb.c
> +++ b/fs/smb/client/cifssmb.c
> @@ -1076,8 +1076,8 @@ SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
>  	pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
>  	pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
>  	pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
> -	/* set file as system file if special file such
> -	   as fifo and server expecting SFU style and
> +	/* set file as system file if special file such as fifo,
> +	 * socket, char or block and server expecting SFU style and
>  	   no Unix extensions */
>  
>  	if (create_options & CREATE_OPTION_SPECIAL)
> @@ -1193,8 +1193,8 @@ CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
>  	req->AllocationSize = 0;
>  
>  	/*
> -	 * Set file as system file if special file such as fifo and server
> -	 * expecting SFU style and no Unix extensions.
> +	 * Set file as system file if special file such as fifo, socket, char
> +	 * or block and server expecting SFU style and no Unix extensions.
>  	 */
>  	if (create_options & CREATE_OPTION_SPECIAL)
>  		req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
> diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
> index e1f2feb56f45..e03c91a49650 100644
> --- a/fs/smb/client/smb1ops.c
> +++ b/fs/smb/client/smb1ops.c
> @@ -1078,7 +1078,7 @@ cifs_make_node(unsigned int xid, struct inode *inode,
>  	/*
>  	 * Check if mounted with mount parm 'sfu' mount parm.
>  	 * SFU emulation should work with all servers, but only
> -	 * supports block and char device (no socket & fifo),
> +	 * supports block and char device, socket & fifo,
>  	 * and was used by default in earlier versions of Windows
>  	 */
>  	if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
> diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
> index 9c2d065d3cc4..9e90672caf60 100644
> --- a/fs/smb/client/smb2ops.c
> +++ b/fs/smb/client/smb2ops.c
> @@ -5066,26 +5066,32 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
>  	struct cifs_fid fid;
>  	unsigned int bytes_written;
>  	struct win_dev pdev = {};
> +	size_t pdev_len = 0;
>  	struct kvec iov[2];
>  	__u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
>  	int rc;
>  
>  	switch (mode & S_IFMT) {
>  	case S_IFCHR:
> +		pdev_len = sizeof(pdev);
>  		memcpy(pdev.type, "IntxCHR\0", 8);
>  		pdev.major = cpu_to_le64(MAJOR(dev));
>  		pdev.minor = cpu_to_le64(MINOR(dev));
>  		break;
>  	case S_IFBLK:
> +		pdev_len = sizeof(pdev);
>  		memcpy(pdev.type, "IntxBLK\0", 8);
>  		pdev.major = cpu_to_le64(MAJOR(dev));
>  		pdev.minor = cpu_to_le64(MINOR(dev));
>  		break;
>  	case S_IFSOCK:
> -		strscpy(pdev.type, "LnxSOCK");
> +		/* SFU socket is system file with one zero byte */
> +		pdev_len = 1;
> +		pdev.type[0] = '\0';
>  		break;
>  	case S_IFIFO:
> -		strscpy(pdev.type, "LnxFIFO");
> +		/* SFU fifo is system file which is empty */
> +		pdev_len = 0;
>  		break;
>  	default:
>  		return -EPERM;
> @@ -5100,14 +5106,17 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
>  	if (rc)
>  		return rc;
>  
> -	io_parms.pid = current->tgid;
> -	io_parms.tcon = tcon;
> -	io_parms.length = sizeof(pdev);
> -	iov[1].iov_base = &pdev;
> -	iov[1].iov_len = sizeof(pdev);
> +	if (pdev_len > 0) {
> +		io_parms.pid = current->tgid;
> +		io_parms.tcon = tcon;
> +		io_parms.length = pdev_len;
> +		iov[1].iov_base = &pdev;
> +		iov[1].iov_len = pdev_len;
> +
> +		rc = server->ops->sync_write(xid, &fid, &io_parms,
> +					     &bytes_written, iov, 1);
> +	}
>  
> -	rc = server->ops->sync_write(xid, &fid, &io_parms,
> -				     &bytes_written, iov, 1);
>  	server->ops->close(xid, tcon, &fid);
>  	return rc;
>  }
> @@ -5149,7 +5158,7 @@ static int smb2_make_node(unsigned int xid, struct inode *inode,
>  	/*
>  	 * Check if mounted with mount parm 'sfu' mount parm.
>  	 * SFU emulation should work with all servers, but only
> -	 * supports block and char device (no socket & fifo),
> +	 * supports block and char device, socket & fifo,
>  	 * and was used by default in earlier versions of Windows
>  	 */
>  	if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {
> -- 
> 2.20.1
>
Steve French Sept. 13, 2024, 10:14 p.m. UTC | #2
How did you find the format of the FIFO and SOCK file types?  I
couldn't find any references to those so added two new types to allow
current Linux to be able to create these (especially since they are
opaque to the server and thus low risk).

> +     case S_IFSOCK:
> -             strscpy(pdev.type, "LnxSOCK");
> +             /* SFU socket is system file with one zero byte */
> +             pdev_len = 1;
> +             pdev.type[0] = '\0';
>               break;
>       case S_IFIFO:
> -             strscpy(pdev.type, "LnxFIFO");
> +             /* SFU fifo is system file which is empty */
> +             pdev_len = 0;

My worry about the suggested change above is that it is possible that
we could accidentally match to an empty file. We intentionally added
the two new dev.type fields for these to avoid collisions with other
things (and since they are Linux specific).  It seems risky to have an
empty file with the system attribute marked as a FIFO, and similarly a
file with one byte null as Socket.   Since this is for only the Linux
client to recognize, I wanted to do something safe for those file
types less likely to be confusing (ie strings for Socket and FIFO that
were similar in length and format to the other special files seemed
intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
file to reduce confusion ie the tags for those two start with "Lnx" -
ie "something used for Linux client" not related to the original
Interix (those begin with "Intx").

Note that in the long run we hope to use reparse points by default in
more servers to store special files like this but there are a few
cases for unusual workloads that need special file support that would
have to use sfu still.  The newer reparse tags that Windows uses "WSL"
have the advantage that they require fewer roundtrips to query (since
the file type is in the reparse tag).

Also noticed an interesting problem when mounted with "sfu" -
"smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
that expected for a FIFO?

On Fri, Sep 13, 2024 at 3:07 PM Pali Rohár <pali@kernel.org> wrote:
>
> On Thursday 12 September 2024 14:05:47 Pali Rohár wrote:
> > SFU-style fifo is empty file with system attribute set. This format is used
> > by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
> > (which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).
> >
> > SFU-style socket is file which has system attribute set and file content is
> > one zero byte. This format was introduced in Interix 3.0 subsystem, as part
> > of the Microsoft SFU 3.0 and is used also by all later versions. Previous
> > versions had no UNIX domain socket support.
> >
> > Currently when sfu mount option is specified then CIFS creates fifo and
> > socket special files with some strange LnxSOCK or LnxFIFO content which is
> > not compatible with Microsoft SFU and neither recognized by other SMB
> > implementations which have some SFU support, including older Linux cifs
> > implementations.
> >
> > So when sfu mount option is specified, create all fifo and socket special
> > files compatible with SFU format to achieve SFU interop, as it is expected
> > by sfu mount option.
> >
> > Signed-off-by: Pali Rohár <pali@kernel.org>
>
> Fixes: 72bc63f5e23a ("smb3: fix creating FIFOs when mounting with "sfu" mount option")
> Fixes: 518549c120e6 ("cifs: fix creating sockets when using sfu mount options")
>
> I located commits which introduced those strange LnxSOCK or LnxFIFO
> types which are not compatible with SFU. I would suggest to add those
> two Fixes: tags into commit message for reference.
>
> > ---
> >  fs/smb/client/cifssmb.c |  8 ++++----
> >  fs/smb/client/smb1ops.c |  2 +-
> >  fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
> >  3 files changed, 24 insertions(+), 15 deletions(-)
> >
> > diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
> > index cfae2e918209..0ffc45aa5e2c 100644
> > --- a/fs/smb/client/cifssmb.c
> > +++ b/fs/smb/client/cifssmb.c
> > @@ -1076,8 +1076,8 @@ SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
> >       pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
> >       pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
> >       pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
> > -     /* set file as system file if special file such
> > -        as fifo and server expecting SFU style and
> > +     /* set file as system file if special file such as fifo,
> > +      * socket, char or block and server expecting SFU style and
> >          no Unix extensions */
> >
> >       if (create_options & CREATE_OPTION_SPECIAL)
> > @@ -1193,8 +1193,8 @@ CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
> >       req->AllocationSize = 0;
> >
> >       /*
> > -      * Set file as system file if special file such as fifo and server
> > -      * expecting SFU style and no Unix extensions.
> > +      * Set file as system file if special file such as fifo, socket, char
> > +      * or block and server expecting SFU style and no Unix extensions.
> >        */
> >       if (create_options & CREATE_OPTION_SPECIAL)
> >               req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
> > diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
> > index e1f2feb56f45..e03c91a49650 100644
> > --- a/fs/smb/client/smb1ops.c
> > +++ b/fs/smb/client/smb1ops.c
> > @@ -1078,7 +1078,7 @@ cifs_make_node(unsigned int xid, struct inode *inode,
> >       /*
> >        * Check if mounted with mount parm 'sfu' mount parm.
> >        * SFU emulation should work with all servers, but only
> > -      * supports block and char device (no socket & fifo),
> > +      * supports block and char device, socket & fifo,
> >        * and was used by default in earlier versions of Windows
> >        */
> >       if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
> > diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
> > index 9c2d065d3cc4..9e90672caf60 100644
> > --- a/fs/smb/client/smb2ops.c
> > +++ b/fs/smb/client/smb2ops.c
> > @@ -5066,26 +5066,32 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> >       struct cifs_fid fid;
> >       unsigned int bytes_written;
> >       struct win_dev pdev = {};
> > +     size_t pdev_len = 0;
> >       struct kvec iov[2];
> >       __u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
> >       int rc;
> >
> >       switch (mode & S_IFMT) {
> >       case S_IFCHR:
> > +             pdev_len = sizeof(pdev);
> >               memcpy(pdev.type, "IntxCHR\0", 8);
> >               pdev.major = cpu_to_le64(MAJOR(dev));
> >               pdev.minor = cpu_to_le64(MINOR(dev));
> >               break;
> >       case S_IFBLK:
> > +             pdev_len = sizeof(pdev);
> >               memcpy(pdev.type, "IntxBLK\0", 8);
> >               pdev.major = cpu_to_le64(MAJOR(dev));
> >               pdev.minor = cpu_to_le64(MINOR(dev));
> >               break;
> >       case S_IFSOCK:
> > -             strscpy(pdev.type, "LnxSOCK");
> > +             /* SFU socket is system file with one zero byte */
> > +             pdev_len = 1;
> > +             pdev.type[0] = '\0';
> >               break;
> >       case S_IFIFO:
> > -             strscpy(pdev.type, "LnxFIFO");
> > +             /* SFU fifo is system file which is empty */
> > +             pdev_len = 0;
> >               break;
> >       default:
> >               return -EPERM;
> > @@ -5100,14 +5106,17 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> >       if (rc)
> >               return rc;
> >
> > -     io_parms.pid = current->tgid;
> > -     io_parms.tcon = tcon;
> > -     io_parms.length = sizeof(pdev);
> > -     iov[1].iov_base = &pdev;
> > -     iov[1].iov_len = sizeof(pdev);
> > +     if (pdev_len > 0) {
> > +             io_parms.pid = current->tgid;
> > +             io_parms.tcon = tcon;
> > +             io_parms.length = pdev_len;
> > +             iov[1].iov_base = &pdev;
> > +             iov[1].iov_len = pdev_len;
> > +
> > +             rc = server->ops->sync_write(xid, &fid, &io_parms,
> > +                                          &bytes_written, iov, 1);
> > +     }
> >
> > -     rc = server->ops->sync_write(xid, &fid, &io_parms,
> > -                                  &bytes_written, iov, 1);
> >       server->ops->close(xid, tcon, &fid);
> >       return rc;
> >  }
> > @@ -5149,7 +5158,7 @@ static int smb2_make_node(unsigned int xid, struct inode *inode,
> >       /*
> >        * Check if mounted with mount parm 'sfu' mount parm.
> >        * SFU emulation should work with all servers, but only
> > -      * supports block and char device (no socket & fifo),
> > +      * supports block and char device, socket & fifo,
> >        * and was used by default in earlier versions of Windows
> >        */
> >       if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {
> > --
> > 2.20.1
> >
>
Steve French Sept. 13, 2024, 10:33 p.m. UTC | #3
One other thing which worried me a bit is that I confirmed that an
empty file with the system attribute set was matched as a FIFO file
type (when you mount with "sfu") - this seems risky.

On Fri, Sep 13, 2024 at 5:14 PM Steve French <smfrench@gmail.com> wrote:
>
> How did you find the format of the FIFO and SOCK file types?  I
> couldn't find any references to those so added two new types to allow
> current Linux to be able to create these (especially since they are
> opaque to the server and thus low risk).
>
> > +     case S_IFSOCK:
> > -             strscpy(pdev.type, "LnxSOCK");
> > +             /* SFU socket is system file with one zero byte */
> > +             pdev_len = 1;
> > +             pdev.type[0] = '\0';
> >               break;
> >       case S_IFIFO:
> > -             strscpy(pdev.type, "LnxFIFO");
> > +             /* SFU fifo is system file which is empty */
> > +             pdev_len = 0;
>
> My worry about the suggested change above is that it is possible that
> we could accidentally match to an empty file. We intentionally added
> the two new dev.type fields for these to avoid collisions with other
> things (and since they are Linux specific).  It seems risky to have an
> empty file with the system attribute marked as a FIFO, and similarly a
> file with one byte null as Socket.   Since this is for only the Linux
> client to recognize, I wanted to do something safe for those file
> types less likely to be confusing (ie strings for Socket and FIFO that
> were similar in length and format to the other special files seemed
> intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> file to reduce confusion ie the tags for those two start with "Lnx" -
> ie "something used for Linux client" not related to the original
> Interix (those begin with "Intx").
>
> Note that in the long run we hope to use reparse points by default in
> more servers to store special files like this but there are a few
> cases for unusual workloads that need special file support that would
> have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> have the advantage that they require fewer roundtrips to query (since
> the file type is in the reparse tag).
>
> Also noticed an interesting problem when mounted with "sfu" -
> "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> that expected for a FIFO?
>
> On Fri, Sep 13, 2024 at 3:07 PM Pali Rohár <pali@kernel.org> wrote:
> >
> > On Thursday 12 September 2024 14:05:47 Pali Rohár wrote:
> > > SFU-style fifo is empty file with system attribute set. This format is used
> > > by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
> > > (which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).
> > >
> > > SFU-style socket is file which has system attribute set and file content is
> > > one zero byte. This format was introduced in Interix 3.0 subsystem, as part
> > > of the Microsoft SFU 3.0 and is used also by all later versions. Previous
> > > versions had no UNIX domain socket support.
> > >
> > > Currently when sfu mount option is specified then CIFS creates fifo and
> > > socket special files with some strange LnxSOCK or LnxFIFO content which is
> > > not compatible with Microsoft SFU and neither recognized by other SMB
> > > implementations which have some SFU support, including older Linux cifs
> > > implementations.
> > >
> > > So when sfu mount option is specified, create all fifo and socket special
> > > files compatible with SFU format to achieve SFU interop, as it is expected
> > > by sfu mount option.
> > >
> > > Signed-off-by: Pali Rohár <pali@kernel.org>
> >
> > Fixes: 72bc63f5e23a ("smb3: fix creating FIFOs when mounting with "sfu" mount option")
> > Fixes: 518549c120e6 ("cifs: fix creating sockets when using sfu mount options")
> >
> > I located commits which introduced those strange LnxSOCK or LnxFIFO
> > types which are not compatible with SFU. I would suggest to add those
> > two Fixes: tags into commit message for reference.
> >
> > > ---
> > >  fs/smb/client/cifssmb.c |  8 ++++----
> > >  fs/smb/client/smb1ops.c |  2 +-
> > >  fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
> > >  3 files changed, 24 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
> > > index cfae2e918209..0ffc45aa5e2c 100644
> > > --- a/fs/smb/client/cifssmb.c
> > > +++ b/fs/smb/client/cifssmb.c
> > > @@ -1076,8 +1076,8 @@ SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
> > >       pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
> > >       pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
> > >       pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
> > > -     /* set file as system file if special file such
> > > -        as fifo and server expecting SFU style and
> > > +     /* set file as system file if special file such as fifo,
> > > +      * socket, char or block and server expecting SFU style and
> > >          no Unix extensions */
> > >
> > >       if (create_options & CREATE_OPTION_SPECIAL)
> > > @@ -1193,8 +1193,8 @@ CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
> > >       req->AllocationSize = 0;
> > >
> > >       /*
> > > -      * Set file as system file if special file such as fifo and server
> > > -      * expecting SFU style and no Unix extensions.
> > > +      * Set file as system file if special file such as fifo, socket, char
> > > +      * or block and server expecting SFU style and no Unix extensions.
> > >        */
> > >       if (create_options & CREATE_OPTION_SPECIAL)
> > >               req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
> > > diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
> > > index e1f2feb56f45..e03c91a49650 100644
> > > --- a/fs/smb/client/smb1ops.c
> > > +++ b/fs/smb/client/smb1ops.c
> > > @@ -1078,7 +1078,7 @@ cifs_make_node(unsigned int xid, struct inode *inode,
> > >       /*
> > >        * Check if mounted with mount parm 'sfu' mount parm.
> > >        * SFU emulation should work with all servers, but only
> > > -      * supports block and char device (no socket & fifo),
> > > +      * supports block and char device, socket & fifo,
> > >        * and was used by default in earlier versions of Windows
> > >        */
> > >       if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
> > > diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
> > > index 9c2d065d3cc4..9e90672caf60 100644
> > > --- a/fs/smb/client/smb2ops.c
> > > +++ b/fs/smb/client/smb2ops.c
> > > @@ -5066,26 +5066,32 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > >       struct cifs_fid fid;
> > >       unsigned int bytes_written;
> > >       struct win_dev pdev = {};
> > > +     size_t pdev_len = 0;
> > >       struct kvec iov[2];
> > >       __u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
> > >       int rc;
> > >
> > >       switch (mode & S_IFMT) {
> > >       case S_IFCHR:
> > > +             pdev_len = sizeof(pdev);
> > >               memcpy(pdev.type, "IntxCHR\0", 8);
> > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > >               break;
> > >       case S_IFBLK:
> > > +             pdev_len = sizeof(pdev);
> > >               memcpy(pdev.type, "IntxBLK\0", 8);
> > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > >               break;
> > >       case S_IFSOCK:
> > > -             strscpy(pdev.type, "LnxSOCK");
> > > +             /* SFU socket is system file with one zero byte */
> > > +             pdev_len = 1;
> > > +             pdev.type[0] = '\0';
> > >               break;
> > >       case S_IFIFO:
> > > -             strscpy(pdev.type, "LnxFIFO");
> > > +             /* SFU fifo is system file which is empty */
> > > +             pdev_len = 0;
> > >               break;
> > >       default:
> > >               return -EPERM;
> > > @@ -5100,14 +5106,17 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > >       if (rc)
> > >               return rc;
> > >
> > > -     io_parms.pid = current->tgid;
> > > -     io_parms.tcon = tcon;
> > > -     io_parms.length = sizeof(pdev);
> > > -     iov[1].iov_base = &pdev;
> > > -     iov[1].iov_len = sizeof(pdev);
> > > +     if (pdev_len > 0) {
> > > +             io_parms.pid = current->tgid;
> > > +             io_parms.tcon = tcon;
> > > +             io_parms.length = pdev_len;
> > > +             iov[1].iov_base = &pdev;
> > > +             iov[1].iov_len = pdev_len;
> > > +
> > > +             rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > +                                          &bytes_written, iov, 1);
> > > +     }
> > >
> > > -     rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > -                                  &bytes_written, iov, 1);
> > >       server->ops->close(xid, tcon, &fid);
> > >       return rc;
> > >  }
> > > @@ -5149,7 +5158,7 @@ static int smb2_make_node(unsigned int xid, struct inode *inode,
> > >       /*
> > >        * Check if mounted with mount parm 'sfu' mount parm.
> > >        * SFU emulation should work with all servers, but only
> > > -      * supports block and char device (no socket & fifo),
> > > +      * supports block and char device, socket & fifo,
> > >        * and was used by default in earlier versions of Windows
> > >        */
> > >       if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {
> > > --
> > > 2.20.1
> > >
> >
>
>
> --
> Thanks,
>
> Steve
Pali Rohár Sept. 13, 2024, 10:42 p.m. UTC | #4
On Friday 13 September 2024 17:14:22 Steve French wrote:
> How did you find the format of the FIFO and SOCK file types?  I

For fifo there are multiple sources on internet, but none of them is
normative. Everything is just what people have tried. For example this
old email on samba list:
https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html

Format of the socket I have figured out by creating it in Interix
subsystem and then dumped content of the file from Win32 subsystem.
Then I checked that it has also same format over older MS NFS server.
It was easier than trying to search for some documentation (which I have
not found).

> couldn't find any references to those so added two new types to allow
> current Linux to be able to create these (especially since they are
> opaque to the server and thus low risk).

I was searching over internet again and now I have found patent
https://patents.google.com/patent/US20090049459 which describe symlink
content:

#define NFS_SPECFILE_LNK_V1  0x014b4e4c78746e49 /* “IntxLNK” */

But does not describe other types.

> > +     case S_IFSOCK:
> > -             strscpy(pdev.type, "LnxSOCK");
> > +             /* SFU socket is system file with one zero byte */
> > +             pdev_len = 1;
> > +             pdev.type[0] = '\0';
> >               break;
> >       case S_IFIFO:
> > -             strscpy(pdev.type, "LnxFIFO");
> > +             /* SFU fifo is system file which is empty */
> > +             pdev_len = 0;
> 
> My worry about the suggested change above is that it is possible that
> we could accidentally match to an empty file.

I fully understand your concerns, but code in this patch is for creating
new fifos. Not recognizing existing fifos.

Code for recognizing existing fifos (=empty file with system attribute)
was not changed and is in Linux cifs client for a very long time.

And if nobody complained yet then this is probably not an issue.

You can create manually on server empty file, mark it as system and then
Linux cifs client even without these changes would recognize that file
as fifo.

> We intentionally added
> the two new dev.type fields for these to avoid collisions with other
> things (and since they are Linux specific).  It seems risky to have an
> empty file with the system attribute marked as a FIFO, and similarly a
> file with one byte null as Socket.   Since this is for only the Linux
> client to recognize, I wanted to do something safe for those file
> types less likely to be confusing (ie strings for Socket and FIFO that
> were similar in length and format to the other special files seemed
> intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> file to reduce confusion ie the tags for those two start with "Lnx" -
> ie "something used for Linux client" not related to the original
> Interix (those begin with "Intx").

I see. Now I understand what are those types (as I have not seen them
before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
functionality is provided by SFU option, but is incompatible with MS SFU
and also with MS NFS server. And is also incompatible with older Linux
cifs clients (as they do not understand those Lnx types).

> Note that in the long run we hope to use reparse points by default in
> more servers to store special files like this but there are a few
> cases for unusual workloads that need special file support that would
> have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> have the advantage that they require fewer roundtrips to query (since
> the file type is in the reparse tag).

Yes, new WSL tags seems to be better. Also SFU mount option is not
activated by default.

> Also noticed an interesting problem when mounted with "sfu" -
> "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> that expected for a FIFO?

Reading from fifo sleep reading process until some other process write
data to fifo. This is how fifos are working. You can try it on local
filesystem (e.g. ext4 or tmpfs).

$ mkfifo fifo
$ strace cat fifo

You will see in strace output of cat something like:
...
openat(AT_FDCWD, "fifo", O_RDONLY

And once you write to that fifo by other process (e.g. echo abc > fifo)
then cat will unfreeze and continue:

                                        = 3
fstat(3, {st_mode=S_IFIFO|0644, st_size=0, ...}) = 0
fadvise64(3, 0, 0, POSIX_FADV_SEQUENTIAL) = -1 ESPIPE (Neprípustné nastavenie pozície)
mmap(NULL, 139264, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f17a34bb000
read(3, "abc\n", 131072)                = 4
...


> On Fri, Sep 13, 2024 at 3:07 PM Pali Rohár <pali@kernel.org> wrote:
> >
> > On Thursday 12 September 2024 14:05:47 Pali Rohár wrote:
> > > SFU-style fifo is empty file with system attribute set. This format is used
> > > by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
> > > (which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).
> > >
> > > SFU-style socket is file which has system attribute set and file content is
> > > one zero byte. This format was introduced in Interix 3.0 subsystem, as part
> > > of the Microsoft SFU 3.0 and is used also by all later versions. Previous
> > > versions had no UNIX domain socket support.
> > >
> > > Currently when sfu mount option is specified then CIFS creates fifo and
> > > socket special files with some strange LnxSOCK or LnxFIFO content which is
> > > not compatible with Microsoft SFU and neither recognized by other SMB
> > > implementations which have some SFU support, including older Linux cifs
> > > implementations.
> > >
> > > So when sfu mount option is specified, create all fifo and socket special
> > > files compatible with SFU format to achieve SFU interop, as it is expected
> > > by sfu mount option.
> > >
> > > Signed-off-by: Pali Rohár <pali@kernel.org>
> >
> > Fixes: 72bc63f5e23a ("smb3: fix creating FIFOs when mounting with "sfu" mount option")
> > Fixes: 518549c120e6 ("cifs: fix creating sockets when using sfu mount options")
> >
> > I located commits which introduced those strange LnxSOCK or LnxFIFO
> > types which are not compatible with SFU. I would suggest to add those
> > two Fixes: tags into commit message for reference.
> >
> > > ---
> > >  fs/smb/client/cifssmb.c |  8 ++++----
> > >  fs/smb/client/smb1ops.c |  2 +-
> > >  fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
> > >  3 files changed, 24 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
> > > index cfae2e918209..0ffc45aa5e2c 100644
> > > --- a/fs/smb/client/cifssmb.c
> > > +++ b/fs/smb/client/cifssmb.c
> > > @@ -1076,8 +1076,8 @@ SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
> > >       pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
> > >       pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
> > >       pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
> > > -     /* set file as system file if special file such
> > > -        as fifo and server expecting SFU style and
> > > +     /* set file as system file if special file such as fifo,
> > > +      * socket, char or block and server expecting SFU style and
> > >          no Unix extensions */
> > >
> > >       if (create_options & CREATE_OPTION_SPECIAL)
> > > @@ -1193,8 +1193,8 @@ CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
> > >       req->AllocationSize = 0;
> > >
> > >       /*
> > > -      * Set file as system file if special file such as fifo and server
> > > -      * expecting SFU style and no Unix extensions.
> > > +      * Set file as system file if special file such as fifo, socket, char
> > > +      * or block and server expecting SFU style and no Unix extensions.
> > >        */
> > >       if (create_options & CREATE_OPTION_SPECIAL)
> > >               req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
> > > diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
> > > index e1f2feb56f45..e03c91a49650 100644
> > > --- a/fs/smb/client/smb1ops.c
> > > +++ b/fs/smb/client/smb1ops.c
> > > @@ -1078,7 +1078,7 @@ cifs_make_node(unsigned int xid, struct inode *inode,
> > >       /*
> > >        * Check if mounted with mount parm 'sfu' mount parm.
> > >        * SFU emulation should work with all servers, but only
> > > -      * supports block and char device (no socket & fifo),
> > > +      * supports block and char device, socket & fifo,
> > >        * and was used by default in earlier versions of Windows
> > >        */
> > >       if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
> > > diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
> > > index 9c2d065d3cc4..9e90672caf60 100644
> > > --- a/fs/smb/client/smb2ops.c
> > > +++ b/fs/smb/client/smb2ops.c
> > > @@ -5066,26 +5066,32 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > >       struct cifs_fid fid;
> > >       unsigned int bytes_written;
> > >       struct win_dev pdev = {};
> > > +     size_t pdev_len = 0;
> > >       struct kvec iov[2];
> > >       __u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
> > >       int rc;
> > >
> > >       switch (mode & S_IFMT) {
> > >       case S_IFCHR:
> > > +             pdev_len = sizeof(pdev);
> > >               memcpy(pdev.type, "IntxCHR\0", 8);
> > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > >               break;
> > >       case S_IFBLK:
> > > +             pdev_len = sizeof(pdev);
> > >               memcpy(pdev.type, "IntxBLK\0", 8);
> > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > >               break;
> > >       case S_IFSOCK:
> > > -             strscpy(pdev.type, "LnxSOCK");
> > > +             /* SFU socket is system file with one zero byte */
> > > +             pdev_len = 1;
> > > +             pdev.type[0] = '\0';
> > >               break;
> > >       case S_IFIFO:
> > > -             strscpy(pdev.type, "LnxFIFO");
> > > +             /* SFU fifo is system file which is empty */
> > > +             pdev_len = 0;
> > >               break;
> > >       default:
> > >               return -EPERM;
> > > @@ -5100,14 +5106,17 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > >       if (rc)
> > >               return rc;
> > >
> > > -     io_parms.pid = current->tgid;
> > > -     io_parms.tcon = tcon;
> > > -     io_parms.length = sizeof(pdev);
> > > -     iov[1].iov_base = &pdev;
> > > -     iov[1].iov_len = sizeof(pdev);
> > > +     if (pdev_len > 0) {
> > > +             io_parms.pid = current->tgid;
> > > +             io_parms.tcon = tcon;
> > > +             io_parms.length = pdev_len;
> > > +             iov[1].iov_base = &pdev;
> > > +             iov[1].iov_len = pdev_len;
> > > +
> > > +             rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > +                                          &bytes_written, iov, 1);
> > > +     }
> > >
> > > -     rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > -                                  &bytes_written, iov, 1);
> > >       server->ops->close(xid, tcon, &fid);
> > >       return rc;
> > >  }
> > > @@ -5149,7 +5158,7 @@ static int smb2_make_node(unsigned int xid, struct inode *inode,
> > >       /*
> > >        * Check if mounted with mount parm 'sfu' mount parm.
> > >        * SFU emulation should work with all servers, but only
> > > -      * supports block and char device (no socket & fifo),
> > > +      * supports block and char device, socket & fifo,
> > >        * and was used by default in earlier versions of Windows
> > >        */
> > >       if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {
> > > --
> > > 2.20.1
> > >
> >
> 
> 
> -- 
> Thanks,
> 
> Steve
Pali Rohár Sept. 13, 2024, 10:45 p.m. UTC | #5
Yes, but it was enabled also before this patch series.

Now I have checked that even old Linux 4.19 cifs kernel client detects
empty file with system attribute as fifo. This functionality is there
for a long time. Probably since first cifs/sfu version.

On Friday 13 September 2024 17:33:22 Steve French wrote:
> One other thing which worried me a bit is that I confirmed that an
> empty file with the system attribute set was matched as a FIFO file
> type (when you mount with "sfu") - this seems risky.
> 
> On Fri, Sep 13, 2024 at 5:14 PM Steve French <smfrench@gmail.com> wrote:
> >
> > How did you find the format of the FIFO and SOCK file types?  I
> > couldn't find any references to those so added two new types to allow
> > current Linux to be able to create these (especially since they are
> > opaque to the server and thus low risk).
> >
> > > +     case S_IFSOCK:
> > > -             strscpy(pdev.type, "LnxSOCK");
> > > +             /* SFU socket is system file with one zero byte */
> > > +             pdev_len = 1;
> > > +             pdev.type[0] = '\0';
> > >               break;
> > >       case S_IFIFO:
> > > -             strscpy(pdev.type, "LnxFIFO");
> > > +             /* SFU fifo is system file which is empty */
> > > +             pdev_len = 0;
> >
> > My worry about the suggested change above is that it is possible that
> > we could accidentally match to an empty file. We intentionally added
> > the two new dev.type fields for these to avoid collisions with other
> > things (and since they are Linux specific).  It seems risky to have an
> > empty file with the system attribute marked as a FIFO, and similarly a
> > file with one byte null as Socket.   Since this is for only the Linux
> > client to recognize, I wanted to do something safe for those file
> > types less likely to be confusing (ie strings for Socket and FIFO that
> > were similar in length and format to the other special files seemed
> > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > file to reduce confusion ie the tags for those two start with "Lnx" -
> > ie "something used for Linux client" not related to the original
> > Interix (those begin with "Intx").
> >
> > Note that in the long run we hope to use reparse points by default in
> > more servers to store special files like this but there are a few
> > cases for unusual workloads that need special file support that would
> > have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> > have the advantage that they require fewer roundtrips to query (since
> > the file type is in the reparse tag).
> >
> > Also noticed an interesting problem when mounted with "sfu" -
> > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > that expected for a FIFO?
> >
> > On Fri, Sep 13, 2024 at 3:07 PM Pali Rohár <pali@kernel.org> wrote:
> > >
> > > On Thursday 12 September 2024 14:05:47 Pali Rohár wrote:
> > > > SFU-style fifo is empty file with system attribute set. This format is used
> > > > by old Microsoft POSIX subsystem and later also by OpenNT/Interix subsystem
> > > > (which replaced Microsoft POSIX subsystem and is part of Microsoft SFU).
> > > >
> > > > SFU-style socket is file which has system attribute set and file content is
> > > > one zero byte. This format was introduced in Interix 3.0 subsystem, as part
> > > > of the Microsoft SFU 3.0 and is used also by all later versions. Previous
> > > > versions had no UNIX domain socket support.
> > > >
> > > > Currently when sfu mount option is specified then CIFS creates fifo and
> > > > socket special files with some strange LnxSOCK or LnxFIFO content which is
> > > > not compatible with Microsoft SFU and neither recognized by other SMB
> > > > implementations which have some SFU support, including older Linux cifs
> > > > implementations.
> > > >
> > > > So when sfu mount option is specified, create all fifo and socket special
> > > > files compatible with SFU format to achieve SFU interop, as it is expected
> > > > by sfu mount option.
> > > >
> > > > Signed-off-by: Pali Rohár <pali@kernel.org>
> > >
> > > Fixes: 72bc63f5e23a ("smb3: fix creating FIFOs when mounting with "sfu" mount option")
> > > Fixes: 518549c120e6 ("cifs: fix creating sockets when using sfu mount options")
> > >
> > > I located commits which introduced those strange LnxSOCK or LnxFIFO
> > > types which are not compatible with SFU. I would suggest to add those
> > > two Fixes: tags into commit message for reference.
> > >
> > > > ---
> > > >  fs/smb/client/cifssmb.c |  8 ++++----
> > > >  fs/smb/client/smb1ops.c |  2 +-
> > > >  fs/smb/client/smb2ops.c | 29 +++++++++++++++++++----------
> > > >  3 files changed, 24 insertions(+), 15 deletions(-)
> > > >
> > > > diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
> > > > index cfae2e918209..0ffc45aa5e2c 100644
> > > > --- a/fs/smb/client/cifssmb.c
> > > > +++ b/fs/smb/client/cifssmb.c
> > > > @@ -1076,8 +1076,8 @@ SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
> > > >       pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
> > > >       pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
> > > >       pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
> > > > -     /* set file as system file if special file such
> > > > -        as fifo and server expecting SFU style and
> > > > +     /* set file as system file if special file such as fifo,
> > > > +      * socket, char or block and server expecting SFU style and
> > > >          no Unix extensions */
> > > >
> > > >       if (create_options & CREATE_OPTION_SPECIAL)
> > > > @@ -1193,8 +1193,8 @@ CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
> > > >       req->AllocationSize = 0;
> > > >
> > > >       /*
> > > > -      * Set file as system file if special file such as fifo and server
> > > > -      * expecting SFU style and no Unix extensions.
> > > > +      * Set file as system file if special file such as fifo, socket, char
> > > > +      * or block and server expecting SFU style and no Unix extensions.
> > > >        */
> > > >       if (create_options & CREATE_OPTION_SPECIAL)
> > > >               req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
> > > > diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
> > > > index e1f2feb56f45..e03c91a49650 100644
> > > > --- a/fs/smb/client/smb1ops.c
> > > > +++ b/fs/smb/client/smb1ops.c
> > > > @@ -1078,7 +1078,7 @@ cifs_make_node(unsigned int xid, struct inode *inode,
> > > >       /*
> > > >        * Check if mounted with mount parm 'sfu' mount parm.
> > > >        * SFU emulation should work with all servers, but only
> > > > -      * supports block and char device (no socket & fifo),
> > > > +      * supports block and char device, socket & fifo,
> > > >        * and was used by default in earlier versions of Windows
> > > >        */
> > > >       if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
> > > > diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
> > > > index 9c2d065d3cc4..9e90672caf60 100644
> > > > --- a/fs/smb/client/smb2ops.c
> > > > +++ b/fs/smb/client/smb2ops.c
> > > > @@ -5066,26 +5066,32 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > > >       struct cifs_fid fid;
> > > >       unsigned int bytes_written;
> > > >       struct win_dev pdev = {};
> > > > +     size_t pdev_len = 0;
> > > >       struct kvec iov[2];
> > > >       __u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
> > > >       int rc;
> > > >
> > > >       switch (mode & S_IFMT) {
> > > >       case S_IFCHR:
> > > > +             pdev_len = sizeof(pdev);
> > > >               memcpy(pdev.type, "IntxCHR\0", 8);
> > > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > > >               break;
> > > >       case S_IFBLK:
> > > > +             pdev_len = sizeof(pdev);
> > > >               memcpy(pdev.type, "IntxBLK\0", 8);
> > > >               pdev.major = cpu_to_le64(MAJOR(dev));
> > > >               pdev.minor = cpu_to_le64(MINOR(dev));
> > > >               break;
> > > >       case S_IFSOCK:
> > > > -             strscpy(pdev.type, "LnxSOCK");
> > > > +             /* SFU socket is system file with one zero byte */
> > > > +             pdev_len = 1;
> > > > +             pdev.type[0] = '\0';
> > > >               break;
> > > >       case S_IFIFO:
> > > > -             strscpy(pdev.type, "LnxFIFO");
> > > > +             /* SFU fifo is system file which is empty */
> > > > +             pdev_len = 0;
> > > >               break;
> > > >       default:
> > > >               return -EPERM;
> > > > @@ -5100,14 +5106,17 @@ static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
> > > >       if (rc)
> > > >               return rc;
> > > >
> > > > -     io_parms.pid = current->tgid;
> > > > -     io_parms.tcon = tcon;
> > > > -     io_parms.length = sizeof(pdev);
> > > > -     iov[1].iov_base = &pdev;
> > > > -     iov[1].iov_len = sizeof(pdev);
> > > > +     if (pdev_len > 0) {
> > > > +             io_parms.pid = current->tgid;
> > > > +             io_parms.tcon = tcon;
> > > > +             io_parms.length = pdev_len;
> > > > +             iov[1].iov_base = &pdev;
> > > > +             iov[1].iov_len = pdev_len;
> > > > +
> > > > +             rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > > +                                          &bytes_written, iov, 1);
> > > > +     }
> > > >
> > > > -     rc = server->ops->sync_write(xid, &fid, &io_parms,
> > > > -                                  &bytes_written, iov, 1);
> > > >       server->ops->close(xid, tcon, &fid);
> > > >       return rc;
> > > >  }
> > > > @@ -5149,7 +5158,7 @@ static int smb2_make_node(unsigned int xid, struct inode *inode,
> > > >       /*
> > > >        * Check if mounted with mount parm 'sfu' mount parm.
> > > >        * SFU emulation should work with all servers, but only
> > > > -      * supports block and char device (no socket & fifo),
> > > > +      * supports block and char device, socket & fifo,
> > > >        * and was used by default in earlier versions of Windows
> > > >        */
> > > >       if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {
> > > > --
> > > > 2.20.1
> > > >
> > >
> >
> >
> > --
> > Thanks,
> >
> > Steve
> 
> 
> 
> -- 
> Thanks,
> 
> Steve
Steve French Sept. 14, 2024, 6:21 a.m. UTC | #6
On Fri, Sep 13, 2024 at 5:42 PM Pali Rohár <pali@kernel.org> wrote:
>
> On Friday 13 September 2024 17:14:22 Steve French wrote:
> > How did you find the format of the FIFO and SOCK file types?  I
>
> For fifo there are multiple sources on internet, but none of them is
> normative. Everything is just what people have tried. For example this
> old email on samba list:
> https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html
>
> Format of the socket I have figured out by creating it in Interix
> subsystem and then dumped content of the file from Win32 subsystem.
> Then I checked that it has also same format over older MS NFS server.
> It was easier than trying to search for some documentation (which I have
> not found).
>
> > couldn't find any references to those so added two new types to allow
> > current Linux to be able to create these (especially since they are
> > opaque to the server and thus low risk).
>
> I was searching over internet again and now I have found patent
> https://patents.google.com/patent/US20090049459 which describe symlink
> content:
>
> #define NFS_SPECFILE_LNK_V1  0x014b4e4c78746e49 /* “IntxLNK” */
>
> But does not describe other types.
>
> > > +     case S_IFSOCK:
> > > -             strscpy(pdev.type, "LnxSOCK");
> > > +             /* SFU socket is system file with one zero byte */
> > > +             pdev_len = 1;
> > > +             pdev.type[0] = '\0';
> > >               break;
> > >       case S_IFIFO:
> > > -             strscpy(pdev.type, "LnxFIFO");
> > > +             /* SFU fifo is system file which is empty */
> > > +             pdev_len = 0;
> >
> > My worry about the suggested change above is that it is possible that
> > we could accidentally match to an empty file.
>
> I fully understand your concerns, but code in this patch is for creating
> new fifos. Not recognizing existing fifos.
>
> Code for recognizing existing fifos (=empty file with system attribute)
> was not changed and is in Linux cifs client for a very long time.
<>
> > We intentionally added
> > the two new dev.type fields for these to avoid collisions with other
> > things (and since they are Linux specific).  It seems risky to have an
> > empty file with the system attribute marked as a FIFO, and similarly a
> > file with one byte null as Socket.   Since this is for only the Linux
> > client to recognize, I wanted to do something safe for those file
> > types less likely to be confusing (ie strings for Socket and FIFO that
> > were similar in length and format to the other special files seemed
> > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > file to reduce confusion ie the tags for those two start with "Lnx" -
> > ie "something used for Linux client" not related to the original
> > Interix (those begin with "Intx").
>
> I see. Now I understand what are those types (as I have not seen them
> before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
> functionality is provided by SFU option, but is incompatible with MS SFU
> and also with MS NFS server. And is also incompatible with older Linux
> cifs clients (as they do not understand those Lnx types).

I am not as worried about FIFO and SOCK type being recognized by
older servers (since almost every use case for them would be for them
to be seen (only) by the client - e.g. for mounts to servers that
don't implement reparse points yet), and since they are less
common file types I am willing to let them be unrecognized by
old clients (we can tag them for stable if older distros don't have
them),  but I am concerned about allowing "current clients" to
create empty files for an unusual purpose which could be
confusing/non-intuitive.

And since this change (at least the one to allow FIFOs to be created with "sfu"
has been in mainline for a year and also since it uses a more intuitive tag
("LnxFIFO") than the empty one used by very old Windows) the only
clients who would have created these would be already using this newer tag
(older Linux clients couldn't have created such files - there seems more
risk of regression with reverting the change than with continuing with
the Linux client specific tag (at least to the one for FIFOs
since that has been in much longer than the socket one which is recent)

Will discuss with others - opinions welcome.

There is an upcoming SMB3.1.1 test event coming up next week (and the annual
Storage Developer Conference too) so I can see if others have opinions one
way or another on whether to move to empty (or 1 byte) files for
creating fifos/sockets

> > Note that in the long run we hope to use reparse points by default in
> > more servers to store special files like this but there are a few
> > cases for unusual workloads that need special file support that would
> > have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> > have the advantage that they require fewer roundtrips to query (since
> > the file type is in the reparse tag).
>
> Yes, new WSL tags seems to be better. Also SFU mount option is not
> activated by default.
>
> > Also noticed an interesting problem when mounted with "sfu" -
> > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > that expected for a FIFO?
>
> Reading from fifo sleep reading process until some other process write
> data to fifo. This is how fifos are working. You can try it on local
> filesystem (e.g. ext4 or tmpfs).

makes sense - thx
Pali Rohár Sept. 14, 2024, 8:17 a.m. UTC | #7
On Saturday 14 September 2024 01:21:17 Steve French wrote:
> On Fri, Sep 13, 2024 at 5:42 PM Pali Rohár <pali@kernel.org> wrote:
> >
> > On Friday 13 September 2024 17:14:22 Steve French wrote:
> > > How did you find the format of the FIFO and SOCK file types?  I
> >
> > For fifo there are multiple sources on internet, but none of them is
> > normative. Everything is just what people have tried. For example this
> > old email on samba list:
> > https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html
> >
> > Format of the socket I have figured out by creating it in Interix
> > subsystem and then dumped content of the file from Win32 subsystem.
> > Then I checked that it has also same format over older MS NFS server.
> > It was easier than trying to search for some documentation (which I have
> > not found).
> >
> > > couldn't find any references to those so added two new types to allow
> > > current Linux to be able to create these (especially since they are
> > > opaque to the server and thus low risk).
> >
> > I was searching over internet again and now I have found patent
> > https://patents.google.com/patent/US20090049459 which describe symlink
> > content:
> >
> > #define NFS_SPECFILE_LNK_V1  0x014b4e4c78746e49 /* “IntxLNK” */
> >
> > But does not describe other types.
> >
> > > > +     case S_IFSOCK:
> > > > -             strscpy(pdev.type, "LnxSOCK");
> > > > +             /* SFU socket is system file with one zero byte */
> > > > +             pdev_len = 1;
> > > > +             pdev.type[0] = '\0';
> > > >               break;
> > > >       case S_IFIFO:
> > > > -             strscpy(pdev.type, "LnxFIFO");
> > > > +             /* SFU fifo is system file which is empty */
> > > > +             pdev_len = 0;
> > >
> > > My worry about the suggested change above is that it is possible that
> > > we could accidentally match to an empty file.
> >
> > I fully understand your concerns, but code in this patch is for creating
> > new fifos. Not recognizing existing fifos.
> >
> > Code for recognizing existing fifos (=empty file with system attribute)
> > was not changed and is in Linux cifs client for a very long time.
> <>
> > > We intentionally added
> > > the two new dev.type fields for these to avoid collisions with other
> > > things (and since they are Linux specific).  It seems risky to have an
> > > empty file with the system attribute marked as a FIFO, and similarly a
> > > file with one byte null as Socket.   Since this is for only the Linux
> > > client to recognize, I wanted to do something safe for those file
> > > types less likely to be confusing (ie strings for Socket and FIFO that
> > > were similar in length and format to the other special files seemed
> > > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > > file to reduce confusion ie the tags for those two start with "Lnx" -
> > > ie "something used for Linux client" not related to the original
> > > Interix (those begin with "Intx").
> >
> > I see. Now I understand what are those types (as I have not seen them
> > before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
> > functionality is provided by SFU option, but is incompatible with MS SFU
> > and also with MS NFS server. And is also incompatible with older Linux
> > cifs clients (as they do not understand those Lnx types).
> 
> I am not as worried about FIFO and SOCK type being recognized by
> older servers (since almost every use case for them would be for them
> to be seen (only) by the client - e.g. for mounts to servers that
> don't implement reparse points yet), and since they are less
> common file types I am willing to let them be unrecognized by
> old clients (we can tag them for stable if older distros don't have
> them),

This is quite pity for old clients, to break existing interoperability.
At least I see sfu as an compatibility option either for ecosystem with
old clients, or option where server itself does not support reparse
points.

> but I am concerned about allowing "current clients" to
> create empty files for an unusual purpose which could be
> confusing/non-intuitive.

I understand this concern. I thought that this should not be an issue
because files are created with system attribute which is not common for
normal/ordinary usage (system attribute could be less confusing) and
also because this format, at least for fifo is used and understood by
many SW for about 30 years.

> And since this change (at least the one to allow FIFOs to be created with "sfu"
> has been in mainline for a year and also since it uses a more intuitive tag
> ("LnxFIFO") than the empty one used by very old Windows) the only
> clients who would have created these would be already using this newer tag
> (older Linux clients couldn't have created such files - there seems more
> risk of regression with reverting the change than with continuing with
> the Linux client specific tag (at least to the one for FIFOs
> since that has been in much longer than the socket one which is recent)

This kind of stuff is lot of times used on LTS/stable linux
distributions and new kernel to these users/admins do not have to be
delivered yet. Mostly it takes 2-3 years after release. Look for example
at RHEL cycles.

I'm looking on this from opposite perspective. I see this an regression
in -o sfu option that after upgrading from previous LTS version to new,
-o sfu stopped to be compatible with SFU-style fifos.

But your point is valid. But maybe it is not an issue because users
do not have updated yet to new version?

> Will discuss with others - opinions welcome.
> 
> There is an upcoming SMB3.1.1 test event coming up next week (and the annual
> Storage Developer Conference too) so I can see if others have opinions one
> way or another on whether to move to empty (or 1 byte) files for
> creating fifos/sockets

Ok, perfect, let me know then about the result.

> > > Note that in the long run we hope to use reparse points by default in
> > > more servers to store special files like this but there are a few
> > > cases for unusual workloads that need special file support that would
> > > have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> > > have the advantage that they require fewer roundtrips to query (since
> > > the file type is in the reparse tag).
> >
> > Yes, new WSL tags seems to be better. Also SFU mount option is not
> > activated by default.
> >
> > > Also noticed an interesting problem when mounted with "sfu" -
> > > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > > that expected for a FIFO?
> >
> > Reading from fifo sleep reading process until some other process write
> > data to fifo. This is how fifos are working. You can try it on local
> > filesystem (e.g. ext4 or tmpfs).
> 
> makes sense - thx
Pali Rohár Sept. 15, 2024, 5:41 p.m. UTC | #8
Hello Steve,

I have another argument for the empty system file as fifo over the
file with "LnxFIFO" content.

Now I have figured out that even the latest Windows Server 2022 version
provides interoperability of FIFOs in SFU format with Windows NFS 4.1
Server. So if you configure on Windows Server 2022 one share which is
exported over SMB and also NFS at the same time, and over SMB you create
SFU-style fifo, then Windows NFS4.1 server recognize it and properly
reports nfs4type as NFS4FIFO for NFSv4.1 client.

So this SFU FIFO style is not only for old clients and servers, but
still relevant and useful even for latest Windows Server.

And same applies for named sockets.

For testing this scenario it is enough to use just trial version of
latest Windows Server from:
https://go.microsoft.com/fwlink/p/?linkid=2208182
https://go.microsoft.com/fwlink/p/?Linkid=2195280

For setting NFS4.1 and SMB share named "test" I used these cmd commands:

  dism /Online /Enable-Feature /All /FeatureName:ServerForNFS-Infrastructure
  md C:\test
  icacls C:\test /grant Everyone:(OI)(CI)F /T
  nfsshare test=C:\test -o rw unmapped=yes
  net share test=C:\test /grant:Everyone,FULL

(for everyone who is going to reproduce this scenario, beware that new
Windows servers use by default powershell, so first launch cmd.exe then
copy+paste those commands)

Pali

On Saturday 14 September 2024 10:17:42 Pali Rohár wrote:
> On Saturday 14 September 2024 01:21:17 Steve French wrote:
> > On Fri, Sep 13, 2024 at 5:42 PM Pali Rohár <pali@kernel.org> wrote:
> > >
> > > On Friday 13 September 2024 17:14:22 Steve French wrote:
> > > > How did you find the format of the FIFO and SOCK file types?  I
> > >
> > > For fifo there are multiple sources on internet, but none of them is
> > > normative. Everything is just what people have tried. For example this
> > > old email on samba list:
> > > https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html
> > >
> > > Format of the socket I have figured out by creating it in Interix
> > > subsystem and then dumped content of the file from Win32 subsystem.
> > > Then I checked that it has also same format over older MS NFS server.
> > > It was easier than trying to search for some documentation (which I have
> > > not found).
> > >
> > > > couldn't find any references to those so added two new types to allow
> > > > current Linux to be able to create these (especially since they are
> > > > opaque to the server and thus low risk).
> > >
> > > I was searching over internet again and now I have found patent
> > > https://patents.google.com/patent/US20090049459 which describe symlink
> > > content:
> > >
> > > #define NFS_SPECFILE_LNK_V1  0x014b4e4c78746e49 /* “IntxLNK” */
> > >
> > > But does not describe other types.
> > >
> > > > > +     case S_IFSOCK:
> > > > > -             strscpy(pdev.type, "LnxSOCK");
> > > > > +             /* SFU socket is system file with one zero byte */
> > > > > +             pdev_len = 1;
> > > > > +             pdev.type[0] = '\0';
> > > > >               break;
> > > > >       case S_IFIFO:
> > > > > -             strscpy(pdev.type, "LnxFIFO");
> > > > > +             /* SFU fifo is system file which is empty */
> > > > > +             pdev_len = 0;
> > > >
> > > > My worry about the suggested change above is that it is possible that
> > > > we could accidentally match to an empty file.
> > >
> > > I fully understand your concerns, but code in this patch is for creating
> > > new fifos. Not recognizing existing fifos.
> > >
> > > Code for recognizing existing fifos (=empty file with system attribute)
> > > was not changed and is in Linux cifs client for a very long time.
> > <>
> > > > We intentionally added
> > > > the two new dev.type fields for these to avoid collisions with other
> > > > things (and since they are Linux specific).  It seems risky to have an
> > > > empty file with the system attribute marked as a FIFO, and similarly a
> > > > file with one byte null as Socket.   Since this is for only the Linux
> > > > client to recognize, I wanted to do something safe for those file
> > > > types less likely to be confusing (ie strings for Socket and FIFO that
> > > > were similar in length and format to the other special files seemed
> > > > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > > > file to reduce confusion ie the tags for those two start with "Lnx" -
> > > > ie "something used for Linux client" not related to the original
> > > > Interix (those begin with "Intx").
> > >
> > > I see. Now I understand what are those types (as I have not seen them
> > > before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
> > > functionality is provided by SFU option, but is incompatible with MS SFU
> > > and also with MS NFS server. And is also incompatible with older Linux
> > > cifs clients (as they do not understand those Lnx types).
> > 
> > I am not as worried about FIFO and SOCK type being recognized by
> > older servers (since almost every use case for them would be for them
> > to be seen (only) by the client - e.g. for mounts to servers that
> > don't implement reparse points yet), and since they are less
> > common file types I am willing to let them be unrecognized by
> > old clients (we can tag them for stable if older distros don't have
> > them),
> 
> This is quite pity for old clients, to break existing interoperability.
> At least I see sfu as an compatibility option either for ecosystem with
> old clients, or option where server itself does not support reparse
> points.
> 
> > but I am concerned about allowing "current clients" to
> > create empty files for an unusual purpose which could be
> > confusing/non-intuitive.
> 
> I understand this concern. I thought that this should not be an issue
> because files are created with system attribute which is not common for
> normal/ordinary usage (system attribute could be less confusing) and
> also because this format, at least for fifo is used and understood by
> many SW for about 30 years.
> 
> > And since this change (at least the one to allow FIFOs to be created with "sfu"
> > has been in mainline for a year and also since it uses a more intuitive tag
> > ("LnxFIFO") than the empty one used by very old Windows) the only
> > clients who would have created these would be already using this newer tag
> > (older Linux clients couldn't have created such files - there seems more
> > risk of regression with reverting the change than with continuing with
> > the Linux client specific tag (at least to the one for FIFOs
> > since that has been in much longer than the socket one which is recent)
> 
> This kind of stuff is lot of times used on LTS/stable linux
> distributions and new kernel to these users/admins do not have to be
> delivered yet. Mostly it takes 2-3 years after release. Look for example
> at RHEL cycles.
> 
> I'm looking on this from opposite perspective. I see this an regression
> in -o sfu option that after upgrading from previous LTS version to new,
> -o sfu stopped to be compatible with SFU-style fifos.
> 
> But your point is valid. But maybe it is not an issue because users
> do not have updated yet to new version?
> 
> > Will discuss with others - opinions welcome.
> > 
> > There is an upcoming SMB3.1.1 test event coming up next week (and the annual
> > Storage Developer Conference too) so I can see if others have opinions one
> > way or another on whether to move to empty (or 1 byte) files for
> > creating fifos/sockets
> 
> Ok, perfect, let me know then about the result.
> 
> > > > Note that in the long run we hope to use reparse points by default in
> > > > more servers to store special files like this but there are a few
> > > > cases for unusual workloads that need special file support that would
> > > > have to use sfu still.  The newer reparse tags that Windows uses "WSL"
> > > > have the advantage that they require fewer roundtrips to query (since
> > > > the file type is in the reparse tag).
> > >
> > > Yes, new WSL tags seems to be better. Also SFU mount option is not
> > > activated by default.
> > >
> > > > Also noticed an interesting problem when mounted with "sfu" -
> > > > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > > > that expected for a FIFO?
> > >
> > > Reading from fifo sleep reading process until some other process write
> > > data to fifo. This is how fifos are working. You can try it on local
> > > filesystem (e.g. ext4 or tmpfs).
> > 
> > makes sense - thx
Pali Rohár Sept. 15, 2024, 5:48 p.m. UTC | #9
It works for this new 2022 version and also for old 2012 version. So my
personal guess is yes, that it would work also for 2019 (I do not see
reason why something would not work for version in-the-middle). If
needed, I may try to do this test against 2019 version during next or
another weekend.

On Sunday 15 September 2024 12:42:27 Steve French wrote:
> Do you know if this is the case for windows server 2019
> 
> On Sun, Sep 15, 2024, 12:41 PM Pali Rohár <pali@kernel.org> wrote:
> 
> > Hello Steve,
> >
> > I have another argument for the empty system file as fifo over the
> > file with "LnxFIFO" content.
> >
> > Now I have figured out that even the latest Windows Server 2022 version
> > provides interoperability of FIFOs in SFU format with Windows NFS 4.1
> > Server. So if you configure on Windows Server 2022 one share which is
> > exported over SMB and also NFS at the same time, and over SMB you create
> > SFU-style fifo, then Windows NFS4.1 server recognize it and properly
> > reports nfs4type as NFS4FIFO for NFSv4.1 client.
> >
> > So this SFU FIFO style is not only for old clients and servers, but
> > still relevant and useful even for latest Windows Server.
> >
> > And same applies for named sockets.
> >
> > For testing this scenario it is enough to use just trial version of
> > latest Windows Server from:
> > https://go.microsoft.com/fwlink/p/?linkid=2208182
> > https://go.microsoft.com/fwlink/p/?Linkid=2195280
> >
> > For setting NFS4.1 and SMB share named "test" I used these cmd commands:
> >
> >   dism /Online /Enable-Feature /All
> > /FeatureName:ServerForNFS-Infrastructure
> >   md C:\test
> >   icacls C:\test /grant Everyone:(OI)(CI)F /T
> >   nfsshare test=C:\test -o rw unmapped=yes
> >   net share test=C:\test /grant:Everyone,FULL
> >
> > (for everyone who is going to reproduce this scenario, beware that new
> > Windows servers use by default powershell, so first launch cmd.exe then
> > copy+paste those commands)
> >
> > Pali
> >
> > On Saturday 14 September 2024 10:17:42 Pali Rohár wrote:
> > > On Saturday 14 September 2024 01:21:17 Steve French wrote:
> > > > On Fri, Sep 13, 2024 at 5:42 PM Pali Rohár <pali@kernel.org> wrote:
> > > > >
> > > > > On Friday 13 September 2024 17:14:22 Steve French wrote:
> > > > > > How did you find the format of the FIFO and SOCK file types?  I
> > > > >
> > > > > For fifo there are multiple sources on internet, but none of them is
> > > > > normative. Everything is just what people have tried. For example
> > this
> > > > > old email on samba list:
> > > > >
> > https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html
> > > > >
> > > > > Format of the socket I have figured out by creating it in Interix
> > > > > subsystem and then dumped content of the file from Win32 subsystem.
> > > > > Then I checked that it has also same format over older MS NFS server.
> > > > > It was easier than trying to search for some documentation (which I
> > have
> > > > > not found).
> > > > >
> > > > > > couldn't find any references to those so added two new types to
> > allow
> > > > > > current Linux to be able to create these (especially since they are
> > > > > > opaque to the server and thus low risk).
> > > > >
> > > > > I was searching over internet again and now I have found patent
> > > > > https://patents.google.com/patent/US20090049459 which describe
> > symlink
> > > > > content:
> > > > >
> > > > > #define NFS_SPECFILE_LNK_V1 0x014b4e4c78746e49 /* “IntxLNK” */
> > > > >
> > > > > But does not describe other types.
> > > > >
> > > > > > > +     case S_IFSOCK:
> > > > > > > -             strscpy(pdev.type, "LnxSOCK");
> > > > > > > +             /* SFU socket is system file with one zero byte */
> > > > > > > +             pdev_len = 1;
> > > > > > > +             pdev.type[0] = '\0';
> > > > > > >               break;
> > > > > > >       case S_IFIFO:
> > > > > > > -             strscpy(pdev.type, "LnxFIFO");
> > > > > > > +             /* SFU fifo is system file which is empty */
> > > > > > > +             pdev_len = 0;
> > > > > >
> > > > > > My worry about the suggested change above is that it is possible
> > that
> > > > > > we could accidentally match to an empty file.
> > > > >
> > > > > I fully understand your concerns, but code in this patch is for
> > creating
> > > > > new fifos. Not recognizing existing fifos.
> > > > >
> > > > > Code for recognizing existing fifos (=empty file with system
> > attribute)
> > > > > was not changed and is in Linux cifs client for a very long time.
> > > > <>
> > > > > > We intentionally added
> > > > > > the two new dev.type fields for these to avoid collisions with
> > other
> > > > > > things (and since they are Linux specific).  It seems risky to
> > have an
> > > > > > empty file with the system attribute marked as a FIFO, and
> > similarly a
> > > > > > file with one byte null as Socket.   Since this is for only the
> > Linux
> > > > > > client to recognize, I wanted to do something safe for those file
> > > > > > types less likely to be confusing (ie strings for Socket and FIFO
> > that
> > > > > > were similar in length and format to the other special files seemed
> > > > > > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > > > > > file to reduce confusion ie the tags for those two start with
> > "Lnx" -
> > > > > > ie "something used for Linux client" not related to the original
> > > > > > Interix (those begin with "Intx").
> > > > >
> > > > > I see. Now I understand what are those types (as I have not seen them
> > > > > before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
> > > > > functionality is provided by SFU option, but is incompatible with MS
> > SFU
> > > > > and also with MS NFS server. And is also incompatible with older
> > Linux
> > > > > cifs clients (as they do not understand those Lnx types).
> > > >
> > > > I am not as worried about FIFO and SOCK type being recognized by
> > > > older servers (since almost every use case for them would be for them
> > > > to be seen (only) by the client - e.g. for mounts to servers that
> > > > don't implement reparse points yet), and since they are less
> > > > common file types I am willing to let them be unrecognized by
> > > > old clients (we can tag them for stable if older distros don't have
> > > > them),
> > >
> > > This is quite pity for old clients, to break existing interoperability.
> > > At least I see sfu as an compatibility option either for ecosystem with
> > > old clients, or option where server itself does not support reparse
> > > points.
> > >
> > > > but I am concerned about allowing "current clients" to
> > > > create empty files for an unusual purpose which could be
> > > > confusing/non-intuitive.
> > >
> > > I understand this concern. I thought that this should not be an issue
> > > because files are created with system attribute which is not common for
> > > normal/ordinary usage (system attribute could be less confusing) and
> > > also because this format, at least for fifo is used and understood by
> > > many SW for about 30 years.
> > >
> > > > And since this change (at least the one to allow FIFOs to be created
> > with "sfu"
> > > > has been in mainline for a year and also since it uses a more
> > intuitive tag
> > > > ("LnxFIFO") than the empty one used by very old Windows) the only
> > > > clients who would have created these would be already using this newer
> > tag
> > > > (older Linux clients couldn't have created such files - there seems
> > more
> > > > risk of regression with reverting the change than with continuing with
> > > > the Linux client specific tag (at least to the one for FIFOs
> > > > since that has been in much longer than the socket one which is recent)
> > >
> > > This kind of stuff is lot of times used on LTS/stable linux
> > > distributions and new kernel to these users/admins do not have to be
> > > delivered yet. Mostly it takes 2-3 years after release. Look for example
> > > at RHEL cycles.
> > >
> > > I'm looking on this from opposite perspective. I see this an regression
> > > in -o sfu option that after upgrading from previous LTS version to new,
> > > -o sfu stopped to be compatible with SFU-style fifos.
> > >
> > > But your point is valid. But maybe it is not an issue because users
> > > do not have updated yet to new version?
> > >
> > > > Will discuss with others - opinions welcome.
> > > >
> > > > There is an upcoming SMB3.1.1 test event coming up next week (and the
> > annual
> > > > Storage Developer Conference too) so I can see if others have opinions
> > one
> > > > way or another on whether to move to empty (or 1 byte) files for
> > > > creating fifos/sockets
> > >
> > > Ok, perfect, let me know then about the result.
> > >
> > > > > > Note that in the long run we hope to use reparse points by default
> > in
> > > > > > more servers to store special files like this but there are a few
> > > > > > cases for unusual workloads that need special file support that
> > would
> > > > > > have to use sfu still.  The newer reparse tags that Windows uses
> > "WSL"
> > > > > > have the advantage that they require fewer roundtrips to query
> > (since
> > > > > > the file type is in the reparse tag).
> > > > >
> > > > > Yes, new WSL tags seems to be better. Also SFU mount option is not
> > > > > activated by default.
> > > > >
> > > > > > Also noticed an interesting problem when mounted with "sfu" -
> > > > > > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > > > > > that expected for a FIFO?
> > > > >
> > > > > Reading from fifo sleep reading process until some other process
> > write
> > > > > data to fifo. This is how fifos are working. You can try it on local
> > > > > filesystem (e.g. ext4 or tmpfs).
> > > >
> > > > makes sense - thx
> >
Volker Lendecke Sept. 16, 2024, 12:51 p.m. UTC | #10
Hello Pali,

Am Sun, Sep 15, 2024 at 07:41:26PM +0200 schrieb Pali Rohár:
> Now I have figured out that even the latest Windows Server 2022 version
> provides interoperability of FIFOs in SFU format with Windows NFS 4.1
> Server. So if you configure on Windows Server 2022 one share which is
> exported over SMB and also NFS at the same time, and over SMB you create
> SFU-style fifo, then Windows NFS4.1 server recognize it and properly
> reports nfs4type as NFS4FIFO for NFSv4.1 client.

it might work in this direction. However, in my tests if you create a
fifo over NFS you end up with a NFS-style reparse point, which is much
less prone to misinterpretation. So my suggestion would be to use the
way that Microsoft properly documented in

https://learn.microsoft.com/en-us/openspecs/windows_protocols/ms-fscc/ff4df658-7f27-476a-8025-4074c0121eec

for anything that is newly created on all servers that can do reparse
points. For servers that don't support reparse points, WSL might be an
alternative. For me it would be highly confusing to add yet another
representation of unix special files from new code.

Reading existing special files -- sure, where strictly required for
compat reasons. Creating them? No. Too many options make the test
matrix explode.

Volker
Pali Rohár Sept. 16, 2024, 4:23 p.m. UTC | #11
Now I have tested this behavior against one Windows Server 2019 instance
and it works too. So confirmed, it is the case also for 2019 version.

On Sunday 15 September 2024 19:48:24 Pali Rohár wrote:
> It works for this new 2022 version and also for old 2012 version. So my
> personal guess is yes, that it would work also for 2019 (I do not see
> reason why something would not work for version in-the-middle). If
> needed, I may try to do this test against 2019 version during next or
> another weekend.
> 
> On Sunday 15 September 2024 12:42:27 Steve French wrote:
> > Do you know if this is the case for windows server 2019
> > 
> > On Sun, Sep 15, 2024, 12:41 PM Pali Rohár <pali@kernel.org> wrote:
> > 
> > > Hello Steve,
> > >
> > > I have another argument for the empty system file as fifo over the
> > > file with "LnxFIFO" content.
> > >
> > > Now I have figured out that even the latest Windows Server 2022 version
> > > provides interoperability of FIFOs in SFU format with Windows NFS 4.1
> > > Server. So if you configure on Windows Server 2022 one share which is
> > > exported over SMB and also NFS at the same time, and over SMB you create
> > > SFU-style fifo, then Windows NFS4.1 server recognize it and properly
> > > reports nfs4type as NFS4FIFO for NFSv4.1 client.
> > >
> > > So this SFU FIFO style is not only for old clients and servers, but
> > > still relevant and useful even for latest Windows Server.
> > >
> > > And same applies for named sockets.
> > >
> > > For testing this scenario it is enough to use just trial version of
> > > latest Windows Server from:
> > > https://go.microsoft.com/fwlink/p/?linkid=2208182
> > > https://go.microsoft.com/fwlink/p/?Linkid=2195280
> > >
> > > For setting NFS4.1 and SMB share named "test" I used these cmd commands:
> > >
> > >   dism /Online /Enable-Feature /All
> > > /FeatureName:ServerForNFS-Infrastructure
> > >   md C:\test
> > >   icacls C:\test /grant Everyone:(OI)(CI)F /T
> > >   nfsshare test=C:\test -o rw unmapped=yes
> > >   net share test=C:\test /grant:Everyone,FULL
> > >
> > > (for everyone who is going to reproduce this scenario, beware that new
> > > Windows servers use by default powershell, so first launch cmd.exe then
> > > copy+paste those commands)
> > >
> > > Pali
> > >
> > > On Saturday 14 September 2024 10:17:42 Pali Rohár wrote:
> > > > On Saturday 14 September 2024 01:21:17 Steve French wrote:
> > > > > On Fri, Sep 13, 2024 at 5:42 PM Pali Rohár <pali@kernel.org> wrote:
> > > > > >
> > > > > > On Friday 13 September 2024 17:14:22 Steve French wrote:
> > > > > > > How did you find the format of the FIFO and SOCK file types?  I
> > > > > >
> > > > > > For fifo there are multiple sources on internet, but none of them is
> > > > > > normative. Everything is just what people have tried. For example
> > > this
> > > > > > old email on samba list:
> > > > > >
> > > https://lists.samba.org/archive/linux-cifs-client/2005-May/000856.html
> > > > > >
> > > > > > Format of the socket I have figured out by creating it in Interix
> > > > > > subsystem and then dumped content of the file from Win32 subsystem.
> > > > > > Then I checked that it has also same format over older MS NFS server.
> > > > > > It was easier than trying to search for some documentation (which I
> > > have
> > > > > > not found).
> > > > > >
> > > > > > > couldn't find any references to those so added two new types to
> > > allow
> > > > > > > current Linux to be able to create these (especially since they are
> > > > > > > opaque to the server and thus low risk).
> > > > > >
> > > > > > I was searching over internet again and now I have found patent
> > > > > > https://patents.google.com/patent/US20090049459 which describe
> > > symlink
> > > > > > content:
> > > > > >
> > > > > > #define NFS_SPECFILE_LNK_V1 0x014b4e4c78746e49 /* “IntxLNK” */
> > > > > >
> > > > > > But does not describe other types.
> > > > > >
> > > > > > > > +     case S_IFSOCK:
> > > > > > > > -             strscpy(pdev.type, "LnxSOCK");
> > > > > > > > +             /* SFU socket is system file with one zero byte */
> > > > > > > > +             pdev_len = 1;
> > > > > > > > +             pdev.type[0] = '\0';
> > > > > > > >               break;
> > > > > > > >       case S_IFIFO:
> > > > > > > > -             strscpy(pdev.type, "LnxFIFO");
> > > > > > > > +             /* SFU fifo is system file which is empty */
> > > > > > > > +             pdev_len = 0;
> > > > > > >
> > > > > > > My worry about the suggested change above is that it is possible
> > > that
> > > > > > > we could accidentally match to an empty file.
> > > > > >
> > > > > > I fully understand your concerns, but code in this patch is for
> > > creating
> > > > > > new fifos. Not recognizing existing fifos.
> > > > > >
> > > > > > Code for recognizing existing fifos (=empty file with system
> > > attribute)
> > > > > > was not changed and is in Linux cifs client for a very long time.
> > > > > <>
> > > > > > > We intentionally added
> > > > > > > the two new dev.type fields for these to avoid collisions with
> > > other
> > > > > > > things (and since they are Linux specific).  It seems risky to
> > > have an
> > > > > > > empty file with the system attribute marked as a FIFO, and
> > > similarly a
> > > > > > > file with one byte null as Socket.   Since this is for only the
> > > Linux
> > > > > > > client to recognize, I wanted to do something safe for those file
> > > > > > > types less likely to be confusing (ie strings for Socket and FIFO
> > > that
> > > > > > > were similar in length and format to the other special files seemed
> > > > > > > intuitive) and "LnxFIFO" and LnxSOCK" were used as the tags in the
> > > > > > > file to reduce confusion ie the tags for those two start with
> > > "Lnx" -
> > > > > > > ie "something used for Linux client" not related to the original
> > > > > > > Interix (those begin with "Intx").
> > > > > >
> > > > > > I see. Now I understand what are those types (as I have not seen them
> > > > > > before). It is somehow misleading if such "LnxFIFO" and LnxSOCK"
> > > > > > functionality is provided by SFU option, but is incompatible with MS
> > > SFU
> > > > > > and also with MS NFS server. And is also incompatible with older
> > > Linux
> > > > > > cifs clients (as they do not understand those Lnx types).
> > > > >
> > > > > I am not as worried about FIFO and SOCK type being recognized by
> > > > > older servers (since almost every use case for them would be for them
> > > > > to be seen (only) by the client - e.g. for mounts to servers that
> > > > > don't implement reparse points yet), and since they are less
> > > > > common file types I am willing to let them be unrecognized by
> > > > > old clients (we can tag them for stable if older distros don't have
> > > > > them),
> > > >
> > > > This is quite pity for old clients, to break existing interoperability.
> > > > At least I see sfu as an compatibility option either for ecosystem with
> > > > old clients, or option where server itself does not support reparse
> > > > points.
> > > >
> > > > > but I am concerned about allowing "current clients" to
> > > > > create empty files for an unusual purpose which could be
> > > > > confusing/non-intuitive.
> > > >
> > > > I understand this concern. I thought that this should not be an issue
> > > > because files are created with system attribute which is not common for
> > > > normal/ordinary usage (system attribute could be less confusing) and
> > > > also because this format, at least for fifo is used and understood by
> > > > many SW for about 30 years.
> > > >
> > > > > And since this change (at least the one to allow FIFOs to be created
> > > with "sfu"
> > > > > has been in mainline for a year and also since it uses a more
> > > intuitive tag
> > > > > ("LnxFIFO") than the empty one used by very old Windows) the only
> > > > > clients who would have created these would be already using this newer
> > > tag
> > > > > (older Linux clients couldn't have created such files - there seems
> > > more
> > > > > risk of regression with reverting the change than with continuing with
> > > > > the Linux client specific tag (at least to the one for FIFOs
> > > > > since that has been in much longer than the socket one which is recent)
> > > >
> > > > This kind of stuff is lot of times used on LTS/stable linux
> > > > distributions and new kernel to these users/admins do not have to be
> > > > delivered yet. Mostly it takes 2-3 years after release. Look for example
> > > > at RHEL cycles.
> > > >
> > > > I'm looking on this from opposite perspective. I see this an regression
> > > > in -o sfu option that after upgrading from previous LTS version to new,
> > > > -o sfu stopped to be compatible with SFU-style fifos.
> > > >
> > > > But your point is valid. But maybe it is not an issue because users
> > > > do not have updated yet to new version?
> > > >
> > > > > Will discuss with others - opinions welcome.
> > > > >
> > > > > There is an upcoming SMB3.1.1 test event coming up next week (and the
> > > annual
> > > > > Storage Developer Conference too) so I can see if others have opinions
> > > one
> > > > > way or another on whether to move to empty (or 1 byte) files for
> > > > > creating fifos/sockets
> > > >
> > > > Ok, perfect, let me know then about the result.
> > > >
> > > > > > > Note that in the long run we hope to use reparse points by default
> > > in
> > > > > > > more servers to store special files like this but there are a few
> > > > > > > cases for unusual workloads that need special file support that
> > > would
> > > > > > > have to use sfu still.  The newer reparse tags that Windows uses
> > > "WSL"
> > > > > > > have the advantage that they require fewer roundtrips to query
> > > (since
> > > > > > > the file type is in the reparse tag).
> > > > > >
> > > > > > Yes, new WSL tags seems to be better. Also SFU mount option is not
> > > > > > activated by default.
> > > > > >
> > > > > > > Also noticed an interesting problem when mounted with "sfu" -
> > > > > > > "smbgetinfo filebasicinfo /mnt/fifo1" will hang (in sys_open).  Is
> > > > > > > that expected for a FIFO?
> > > > > >
> > > > > > Reading from fifo sleep reading process until some other process
> > > write
> > > > > > data to fifo. This is how fifos are working. You can try it on local
> > > > > > filesystem (e.g. ext4 or tmpfs).
> > > > >
> > > > > makes sense - thx
> > >
Pali Rohár Sept. 16, 2024, 4:31 p.m. UTC | #12
Hello Volker,

On Monday 16 September 2024 14:51:49 Volker Lendecke wrote:
> Hello Pali,
> 
> Am Sun, Sep 15, 2024 at 07:41:26PM +0200 schrieb Pali Rohár:
> > Now I have figured out that even the latest Windows Server 2022 version
> > provides interoperability of FIFOs in SFU format with Windows NFS 4.1
> > Server. So if you configure on Windows Server 2022 one share which is
> > exported over SMB and also NFS at the same time, and over SMB you create
> > SFU-style fifo, then Windows NFS4.1 server recognize it and properly
> > reports nfs4type as NFS4FIFO for NFSv4.1 client.
> 
> it might work in this direction. However, in my tests if you create a
> fifo over NFS you end up with a NFS-style reparse point, which is much
> less prone to misinterpretation.

Yes, that is truth. New NFS server versions by default create reparse
points with NFS tags for special files. I just pointed out that even new
NFS server can recognize old SFU-style of special files, which is useful
to know in some situation (e.g. for mix of old and cifs clients together
with nfs clients).

> So my suggestion would be to use the
> way that Microsoft properly documented in
> 
> https://learn.microsoft.com/en-us/openspecs/windows_protocols/ms-fscc/ff4df658-7f27-476a-8025-4074c0121eec
> 
> for anything that is newly created on all servers that can do reparse
> points. For servers that don't support reparse points, WSL might be an
> alternative. For me it would be highly confusing to add yet another
> representation of unix special files from new code.
> 
> Reading existing special files -- sure, where strictly required for
> compat reasons. Creating them? No. Too many options make the test
> matrix explode.
> 
> Volker

Important is that all my changes and whole discussion is about the
non-default cifs mount option '-o sfu' which disables creating reparse
points and uses only SFU-style of special files. So this is not a new
code but special compatibility case used in scenarios where reparse
points are either not supported or cifs client want to use normal files
for storing fifos/sockets/... in SFU-style.

I fully agree with you that in the default conditions new reparse points
should be used and created. And this behavior is not changed or affected
by these my patches.

But in this compatibility mode, which has to be manually activated by
user at cifs mount time by -o sfu option, we should follow what SFU is.

Note that when this -o sfu option is not specified then these special
system files are not recognized (or created) at all.
diff mbox series

Patch

diff --git a/fs/smb/client/cifssmb.c b/fs/smb/client/cifssmb.c
index cfae2e918209..0ffc45aa5e2c 100644
--- a/fs/smb/client/cifssmb.c
+++ b/fs/smb/client/cifssmb.c
@@ -1076,8 +1076,8 @@  SMBLegacyOpen(const unsigned int xid, struct cifs_tcon *tcon,
 	pSMB->OpenFlags |= cpu_to_le16(REQ_MORE_INFO);
 	pSMB->Mode = cpu_to_le16(access_flags_to_smbopen_mode(access_flags));
 	pSMB->Mode |= cpu_to_le16(0x40); /* deny none */
-	/* set file as system file if special file such
-	   as fifo and server expecting SFU style and
+	/* set file as system file if special file such as fifo,
+	 * socket, char or block and server expecting SFU style and
 	   no Unix extensions */
 
 	if (create_options & CREATE_OPTION_SPECIAL)
@@ -1193,8 +1193,8 @@  CIFS_open(const unsigned int xid, struct cifs_open_parms *oparms, int *oplock,
 	req->AllocationSize = 0;
 
 	/*
-	 * Set file as system file if special file such as fifo and server
-	 * expecting SFU style and no Unix extensions.
+	 * Set file as system file if special file such as fifo, socket, char
+	 * or block and server expecting SFU style and no Unix extensions.
 	 */
 	if (create_options & CREATE_OPTION_SPECIAL)
 		req->FileAttributes = cpu_to_le32(ATTR_SYSTEM);
diff --git a/fs/smb/client/smb1ops.c b/fs/smb/client/smb1ops.c
index e1f2feb56f45..e03c91a49650 100644
--- a/fs/smb/client/smb1ops.c
+++ b/fs/smb/client/smb1ops.c
@@ -1078,7 +1078,7 @@  cifs_make_node(unsigned int xid, struct inode *inode,
 	/*
 	 * Check if mounted with mount parm 'sfu' mount parm.
 	 * SFU emulation should work with all servers, but only
-	 * supports block and char device (no socket & fifo),
+	 * supports block and char device, socket & fifo,
 	 * and was used by default in earlier versions of Windows
 	 */
 	if (!(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL))
diff --git a/fs/smb/client/smb2ops.c b/fs/smb/client/smb2ops.c
index 9c2d065d3cc4..9e90672caf60 100644
--- a/fs/smb/client/smb2ops.c
+++ b/fs/smb/client/smb2ops.c
@@ -5066,26 +5066,32 @@  static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
 	struct cifs_fid fid;
 	unsigned int bytes_written;
 	struct win_dev pdev = {};
+	size_t pdev_len = 0;
 	struct kvec iov[2];
 	__u32 oplock = server->oplocks ? REQ_OPLOCK : 0;
 	int rc;
 
 	switch (mode & S_IFMT) {
 	case S_IFCHR:
+		pdev_len = sizeof(pdev);
 		memcpy(pdev.type, "IntxCHR\0", 8);
 		pdev.major = cpu_to_le64(MAJOR(dev));
 		pdev.minor = cpu_to_le64(MINOR(dev));
 		break;
 	case S_IFBLK:
+		pdev_len = sizeof(pdev);
 		memcpy(pdev.type, "IntxBLK\0", 8);
 		pdev.major = cpu_to_le64(MAJOR(dev));
 		pdev.minor = cpu_to_le64(MINOR(dev));
 		break;
 	case S_IFSOCK:
-		strscpy(pdev.type, "LnxSOCK");
+		/* SFU socket is system file with one zero byte */
+		pdev_len = 1;
+		pdev.type[0] = '\0';
 		break;
 	case S_IFIFO:
-		strscpy(pdev.type, "LnxFIFO");
+		/* SFU fifo is system file which is empty */
+		pdev_len = 0;
 		break;
 	default:
 		return -EPERM;
@@ -5100,14 +5106,17 @@  static int __cifs_sfu_make_node(unsigned int xid, struct inode *inode,
 	if (rc)
 		return rc;
 
-	io_parms.pid = current->tgid;
-	io_parms.tcon = tcon;
-	io_parms.length = sizeof(pdev);
-	iov[1].iov_base = &pdev;
-	iov[1].iov_len = sizeof(pdev);
+	if (pdev_len > 0) {
+		io_parms.pid = current->tgid;
+		io_parms.tcon = tcon;
+		io_parms.length = pdev_len;
+		iov[1].iov_base = &pdev;
+		iov[1].iov_len = pdev_len;
+
+		rc = server->ops->sync_write(xid, &fid, &io_parms,
+					     &bytes_written, iov, 1);
+	}
 
-	rc = server->ops->sync_write(xid, &fid, &io_parms,
-				     &bytes_written, iov, 1);
 	server->ops->close(xid, tcon, &fid);
 	return rc;
 }
@@ -5149,7 +5158,7 @@  static int smb2_make_node(unsigned int xid, struct inode *inode,
 	/*
 	 * Check if mounted with mount parm 'sfu' mount parm.
 	 * SFU emulation should work with all servers, but only
-	 * supports block and char device (no socket & fifo),
+	 * supports block and char device, socket & fifo,
 	 * and was used by default in earlier versions of Windows
 	 */
 	if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_UNX_EMUL) {