Hi,

when trying to create a directory path with 'mkdir -p' with the root
being a mount point of a writeable filesystem that exists inside a
read-only file system, mkdir return -EROFS.

This was tested with the latest git-tree from:
git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6

The last kernel I know of this has been working was 2.6.22.

E.g.:
|root@triton1:/# mount
|rootfs on / type rootfs (rw)
|/dev/root on / type jffs2 (ro)
|proc on /proc type proc (rw)
|sysfs on /sys type sysfs (rw)
|/dev/root on /dev/.static/dev type jffs2 (ro)
|tmpfs on /dev type tmpfs (rw,size=10240k,mode=755)
|tmpfs on /var type tmpfs (rw)
|tmpfs on /tmp type tmpfs (rw)
|devpts on /dev/pts type devpts (rw,gid=5,mode=620)
|tmpfs on /dev/shm type tmpfs (rw)
|root@triton1:/# mkdir -p /dev/.udev/db
|mkdir: Cannot create directory `/dev/': Read-only file system

While:
|root@triton1:/# cd /dev
|root@triton1:/# mkdir -p .udev/db
works as expected...

Has anyone else encountered this problem already?
I've searched the archives but could not find anything related to this
problem.



Lothar Wa

(y'know that ~4000 people read your email and hit 'n'?  Except for me?

I had a shot at scripting this, but for me, everything works OK.

#!/bin/sh
dd if=/dev/zero of=file1 bs=1M count=32
dd if=/dev/zero of=file2 bs=1M count=32
yes | mkfs -t ext2 file1
yes | mkfs -t ext2 file2
mkdir -p mountpoint
mount -o loop file1 mountpoint
mkdir mountpoint/dir
umount mountpoint
mount -o loop,ro file1 mountpoint
mount -o loop file2 mountpoint/dir
mkdir -p $(/bin/pwd)/mountpoint/dir/subdir


are you able to modify the above so that it demonstrates the failure?

Thanks.
--

Hmmm.  The thing you are doing differently is that the reporter's root 
filesystem is mounted read only but in your case (and the quick try I 
had at it) the read only file system is mounted below root.

However, I tried reproducing that scenario with a chroot and I still 
couldn't get it to happen.  No time to try a real root mount ro this 
morning but perhaps others can.

Cheers,
Dan

-- 
                     /--------------- - -  -  -   -   -
                     |  Dan Noé
                     |  http://isomerica.net/~dpn/
--

My guess is that you have an old/buggy version of mkdir.

What does
    mkdir --version

say?
Mine says:
mkdir (GNU coreutils) 6.10
Copyright (C) 2008 Free Software Foundation, Inc.
etc...

What does

   strace mkdir -p /dev/.udev/db

show ???

NeilBrown


--

I would bet this is the problem for the reporter.

I tried the same here, and cannot exhibit the problem. Just like the
reporter, my / is ro and /dev is rw :

# mkdir --version
mkdir (coreutils) 5.2.1
...

# uname -a
Linux alix-2C3 2.6.25-wt4-alix #3 Thu Apr 24 08:30:05 CEST 2008 i586 i586 i386 GNU/Linux

(this is essentially 2.6.25+squashfs+some geode patches)

# mount
/dev/sda4 on / type ext3 (ro,errors=continue,data=ordered)
/dev on /dev type tmpfs (rw,size=0k,nr_inodes=4096,mode=755)

# mkdir -p /dev/a/b/c
# ls -lad /dev/a/b/c
drwxr-xr-x  2 root root 40 Jan  3 00:49 /dev/a/b/c/

# rmdir -p /dev/a/b/c
rmdir: `/dev': Read-only file system
# strace mkdir -p /dev/a/b/c

stat64("/dev/a/b/c", 0xbf9175dc)        = -1 ENOENT (No such file or directory)
umask(0)                                = 022
open(".", O_RDONLY|O_LARGEFILE|O_DIRECTORY) = 3
chdir("/")                              = 0
mkdir("dev", 0755)                      = -1 EEXIST (File exists)
stat64("dev", {st_mode=S_IFDIR|0755, st_size=17080, ...}) = 0
chdir("dev")                            = 0
mkdir("a", 0755)                        = 0
chdir("a")                              = 0
mkdir("b", 0755)                        = 0
chdir("b")                              = 0
umask(022)                              = 0
mkdir("c", 0777)                        = 0
fchdir(3)                               = 0
close(3)                                = 0

Regards,
Willy

--

Hi,

This is what strace shows in my case:
mkdir("/", 0777)                        = -1 EEXIST (File exists)
stat64("/", {st_mode=S_IFDIR|0755, st_size=0, ...}) = 0
mkdir("/dev/", 0777)                    = -1 EROFS (Read-only file system)
umask(022)                              = 022
write(2, "mkdir: ", 7mkdir: )                  = 7
write(2, "Cannot create directory `/dev/\'", 31Cannot create directory `/dev/') = 31
write(2, ": Read-only file system\n", 24: Read-only file system
) = 24
io_submit(0x1, 0x1, 0xfbad2088 <unfinished ... exit status 1>


Lothar Wa

So you should send this trace to the busybox maintainer. It
obviously exhibits a small but easily fixable bug in the
program ; it should check whether the directory exists in
case of error. It seems it only checks for EEXIST.

What puzzles me is why you have this problem only with a new kernel.
I suspect that previous ones did report EEXIST and now we report
EROFS for an existing directory on an ro fs.

Willy

--

Hi,

                                                   ^^^^^^
Interestingly your strace output showed a return value of EEXIST where
mine had EROFS.


Lothar Wa

Oh you're right! I did not notice. There's something odd here.
And I can reproduce it right now (2.6.25.1). I really have
EEXIST.

Willy

--

Hi,

As I noted in my first message I have this problem with the kernel
from git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
which is currently 2.6.26-rc1 and still exhibits this behaviour.

So, this problem obviously does not exist in the 2.6.25 release kernel
but would probably hit mainline in the 2.6.26 release.


Lothar Wa

Aye.  That's ro-bind fallout we'd not noticed; will fix...

And a fine example of the reasons why
	* relying on which particular errno value you get when several
are applicable is a Bad Idea(tm) and
	* changing the choice of the errno value in such situation might
be legal, but is to be done with care; sometimes it really has impact...
--