On Wed, Dec 29, 2010 at 11:12 AM, Mulyadi Santosa <mulyadi.santosa@gmail.com> wrote:

On Wed, Dec 29, 2010 at 20:06, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

...

Hello All,

ZFS file system has a property called devices. If turned off, ZFS would not allow access to the device files (block/character) present on the file system. I want to implement the same behavior on the a Linux File System.

I don't know about ZFS, so could you please elaborate on what you mean by "ZFS could disallow access"?

IMHO, (untested), you could simply do it using usual Linux file/directory permission up to SELinux/AppArmor....so, is that what you mean?

-- regards,

Mulyadi Santosa

Mulyadi, My guess is that it is more complex than that. Some filesystems have issues if the raw drive is read while the filesystem is mounted. I think it is caused by inconsistencies in the various cache's. ie. iirc, At least in the 2.4 kernel there was not a single unified cache for block layer and filesystems. So doing raw reads of underlying device while it was mounted could cause the caches to get out of sync. I don't recall the details, but either the kernel would oops or the filesystem would become corrupt. I don't know if any 2.6 filesystems still have that issue. Anyway ZFS must have a similar issue. So a ZFS filesystem developer knowing this was a conflict could add a check in the /dev/sda open() that would fail the open if there was a mounted filesystem of type ZFS on the drive. And the mount should fail if /dev/sda is already open. I'm not aware of the 2.6.x linux kernel offering any infrastructure to help with that issue. Greg

On Wednesday 29 Dec 2010 10:31:37 pm Greg Freemyer wrote:

On Wed, Dec 29, 2010 at 11:12 AM, Mulyadi Santosa

<mulyadi.santosa@gmail.com> wrote:

...

On Wed, Dec 29, 2010 at 20:06, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

...

Hello All,

ZFS file system has a property called devices. If turned off, ZFS would not allow access to the device files (block/character) present on the file system. I want to implement the same behavior on the a Linux File System.

I don't know about ZFS, so could you please elaborate on what you mean by "ZFS could disallow access"?

IMHO, (untested), you could simply do it using usual Linux file/directory permission up to SELinux/AppArmor....so, is that what you mean?

-- regards,

Mulyadi Santosa

Mulyadi,

My guess is that it is more complex than that.

Some filesystems have issues if the raw drive is read while the filesystem is mounted. I think it is caused by inconsistencies in the various cache's. ie. iirc, At least in the 2.4 kernel there was not a single unified cache for block layer and filesystems. So doing raw reads of underlying device while it was mounted could cause the caches to get out of sync.

I don't recall the details, but either the kernel would oops or the filesystem would become corrupt. I don't know if any 2.6 filesystems still have that issue. Anyway ZFS must have a similar issue.

So a ZFS filesystem developer knowing this was a conflict could add a check in the /dev/sda open() that would fail the open if there was a mounted filesystem of type ZFS on the drive.

And the mount should fail if /dev/sda is already open.

I'm not aware of the 2.6.x linux kernel offering any infrastructure to help with that issue.

Greg

_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies

Greg,

So doing raw reads of underlying device while it was mounted could cause the caches to get out of sync.

So doing a 'dd' would cause the kernel to oops? Thanks.

On Wed, Dec 29, 2010 at 1:54 PM, mindentropy <mindentropy@gmail.com> wrote:

On Wednesday 29 Dec 2010 10:31:37 pm Greg Freemyer wrote:

...

On Wed, Dec 29, 2010 at 11:12 AM, Mulyadi Santosa

<mulyadi.santosa@gmail.com> wrote:

...

On Wed, Dec 29, 2010 at 20:06, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

...

Hello All,

ZFS file system has a property called devices. If turned off, ZFS would not allow access to the device files (block/character) present on the file system. I want to implement the same behavior on the a Linux File System.

I don't know about ZFS, so could you please elaborate on what you mean by "ZFS could disallow access"?

IMHO, (untested), you could simply do it using usual Linux file/directory permission up to SELinux/AppArmor....so, is that what you mean?

-- regards,

Mulyadi Santosa

Mulyadi,

My guess is that it is more complex than that.

Some filesystems have issues if the raw drive is read while the filesystem is mounted.  I think it is caused by inconsistencies in the various cache's.  ie. iirc, At least in the 2.4 kernel there was not a single unified cache for block layer and filesystems.  So doing raw reads of underlying device while it was mounted could cause the caches to get out of sync.

I don't recall the details, but either the kernel would oops or the filesystem would become corrupt.  I don't know if any 2.6 filesystems still have that issue.  Anyway ZFS must have a similar issue.

So a ZFS filesystem developer knowing this was a conflict could add a check in the /dev/sda open() that would fail the open if there was a mounted filesystem of type ZFS on the drive.

And the mount should fail if /dev/sda is already open.

I'm not aware of the 2.6.x linux kernel offering any infrastructure to help with that issue.

Greg

_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies

Greg,

...

So doing raw reads of underlying device while it was mounted could cause the caches to get out of sync.

   So doing a 'dd' would cause the kernel to oops?

Thanks.

I really don't remember what the problem was. It was about 2003 or 2004 I think that I came across the issue. I don't recall if it was XFS or ext2/3 or what. I just know I came to the conclusion that dd if=/dev/sda of=/dev/null was not always safe if the drive had a mounted filesystem. Hopefully its been fixed by now. Greg

On Wed, Dec 29, 2010 at 2:07 PM, Mulyadi Santosa <mulyadi.santosa@gmail.com> wrote:

Hi Greg...

On Thu, Dec 30, 2010 at 02:02, Greg Freemyer <greg.freemyer@gmail.com> wrote:

...

I really don't remember what the problem was.  It was about 2003 or 2004 I think that I came across the issue.  I don't recall if it was XFS or ext2/3 or what.

I just know I came to the conclusion that dd if=/dev/sda of=/dev/null was not always safe if the drive had a mounted filesystem.  Hopefully its been fixed by now.

2.4.x? If yes, I am not so surprised since buffer cache and page cache aren't unified yet, unlike in 2.6.x. I highly guess that was indeed synchronization issue....

yes, 2.4.x was what I am talking about.

On Wed, Dec 29, 2010 at 6:32 PM, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

On Wed, Dec 29, 2010 at 4:12 PM, Mulyadi Santosa <mulyadi.santosa@gmail.com> wrote:

...

On Wed, Dec 29, 2010 at 20:06, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

...

Hello All,

ZFS file system has a property called devices. If turned off, ZFS would not allow access to the device files (block/character) present on the file system. I want to implement the same behavior on the a Linux File System.

I don't know about ZFS, so could you please elaborate on what you mean by "ZFS could disallow access"?

I am really sorry that I was not clear with the first mail. Thanks a lot for all mail replies and for sharing important information. By not disallowing access to device files I ment

root@prasad-laptop:~# mount disk -o loop arm/

root@prasad-laptop:~/arm# mount -t ext3 /dev/loop0 on /home/prasad/arm type ext3 (rw)

############# CREATING A DEVICE FILE ON THE FILE SYSTEM root@prasad-laptop:~/arm# mknod zero c 1 5

root@prasad-laptop:~/arm# ls lost+found  zero

root@prasad-laptop:~/arm# ls -l total 12 drwx------ 2 root root 12288 2010-12-23 11:28 lost+found crw-r--r-- 1 root root  1, 5 2010-12-23 11:28 zero

root@prasad-laptop:~/arm# dd if=zero of=disk bs=10K count=10K dd: writing `disk': No space left on device 9313+0 records in 9312+0 records out 95354880 bytes (95 MB) copied, 1.00106 s, 95.3 MB/s

root@prasad-laptop:~/arm# ls -l total 93499 -rw-r--r-- 1 root root 95354880 2010-12-23 11:28 disk drwx------ 2 root root    12288 2010-12-23 11:28 lost+found crw-r--r-- 1 root root     1, 5 2010-12-23 11:28 zero

Here the file system allowed access to the device file named zero. The requirement is to turn off the access to all of the device files present on the mounted file system. ie. considering the above case access (open/read/write) to/from device zero should not be allowed (even by root user). I don't know why would one create a device file on a file system other than /dev.

I could modify the open code to check if the file the file being opened is device file then return either EPERM or EACCESS (not sure which one). But before modifying the code I thought of checking mount flags, could not find one, hence thought of asking on mailing list.

Thanks a lot for wonderful replies and sharing valuable information. Hope the example above has made the requirement clear.

Thanks and Regards, Prasad

Well that's totally different than I thought. Why not stop the mknod from working instead of the open? mknod is not be part of the "hot path", where open could easily be for some apps. ie. If your FS does not support device files, then simply don't allow them to be created. Thanks Greg

Henry is right, nodev is the option for you, it passes down MS_NODEV flag to kernel which in turn sets MNT_NODEV flag in mount object, which open system call checks and return -EACCES if device it is a device node, snippet from linux/fs/namei.c: case S_IFBLK: case S_IFCHR: if (path->mnt->mnt_flags & MNT_NODEV) return -EACCES; The case which Greg was mentioning, I think that is solved through exclusive open call to block device in exclusive mode with open_bdev_exclusive. This allows only one opener of block device at a time, i.e. Filesystem is the exclusive opener of block device which mounting the block device, e.g. ext2_get_sb() -> get_sb_bdev() -> open_bdev_exclusive(); Rajat On Thu, Dec 30, 2010 at 5:54 AM, Henry Gebhardt <hsggebhardt@googlemail.com> wrote:

On Wed, Dec 29, 2010 at 11:32:18PM +0000, Prasad Joshi wrote:

...

On Wed, Dec 29, 2010 at 4:12 PM, Mulyadi Santosa <mulyadi.santosa@gmail.com> wrote:

...

On Wed, Dec 29, 2010 at 20:06, Prasad Joshi <prasadjoshi124@gmail.com> wrote:

...

Hello All,

ZFS file system has a property called devices. If turned off, ZFS would not allow access to the device files (block/character) present on the file system. I want to implement the same behavior on the a Linux File System.

I don't know about ZFS, so could you please elaborate on what you mean by "ZFS could disallow access"?

I am really sorry that I was not clear with the first mail. Thanks a lot for all mail replies and for sharing important information. By not disallowing access to device files I ment

root@prasad-laptop:~# mount disk -o loop arm/

root@prasad-laptop:~/arm# mount -t ext3 /dev/loop0 on /home/prasad/arm type ext3 (rw)

############# CREATING A DEVICE FILE ON THE FILE SYSTEM root@prasad-laptop:~/arm# mknod zero c 1 5

root@prasad-laptop:~/arm# ls lost+found  zero

root@prasad-laptop:~/arm# ls -l total 12 drwx------ 2 root root 12288 2010-12-23 11:28 lost+found crw-r--r-- 1 root root  1, 5 2010-12-23 11:28 zero

root@prasad-laptop:~/arm# dd if=zero of=disk bs=10K count=10K dd: writing `disk': No space left on device 9313+0 records in 9312+0 records out 95354880 bytes (95 MB) copied, 1.00106 s, 95.3 MB/s

root@prasad-laptop:~/arm# ls -l total 93499 -rw-r--r-- 1 root root 95354880 2010-12-23 11:28 disk drwx------ 2 root root    12288 2010-12-23 11:28 lost+found crw-r--r-- 1 root root     1, 5 2010-12-23 11:28 zero

Here the file system allowed access to the device file named zero. The requirement is to turn off the access to all of the device files present on the mounted file system. ie. considering the above case access (open/read/write) to/from device zero should not be allowed (even by root user). I don't know why would one create a device file on a file system other than /dev.

I could modify the open code to check if the file the file being opened is device file then return either EPERM or EACCESS (not sure which one). But before modifying the code I thought of checking mount flags, could not find one, hence thought of asking on mailing list.

Thanks a lot for wonderful replies and sharing valuable information. Hope the example above has made the requirement clear.

Maybe I, too, am completely misunderstanding you, but does the nodev option do what you want?  From the mount manpage:

   nodev - Do not interpret character or block special devices on the    file system.

Use like so:

   $ mount disk -o loop,nodev arm/

You can still create device special files, you just can't access them.

Greetings, Henry

_______________________________________________ Kernelnewbies mailing list Kernelnewbies@kernelnewbies.org http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies