udev loops eating 100 % CPU [mmotm 2010-05-21-16-05]

udev loops eating 100 % CPU [mmotm 2010-05-21-16-05]

[Jiri Slaby]

Hi,

>From next/master:
commit a7cf4145bb86aaf85d4d4d29a69b50b688e2e49d
Author: Eric Paris <eparis@redhat.com>
Date:   Fri May 14 11:44:10 2010 -0400

    anon_inode: set S_IFREG on the anon_inode

    anon_inode_mkinode() sets inode->i_mode = S_IRUSR | S_IWUSR;  This means
    that (inode->i_mode & S_IFMT) == 0.  This trips up some SELinux code
that
    needs to determine if a given inode is a regular file, a directory, etc.
    The easiest solution is to just make sure that the anon_inode also sets
    S_IFREG.

    Signed-off-by: Eric Paris <eparis@redhat.com>
    Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>





causes my udev and psi both eating 100 % CPU. When I revert that on the
top of mmotm 2010-05-21-16-05, the problem disappears.

These are red-numbered lines from perf:
    13.96%    udevd  [kernel.kallsyms]      [k] _raw_spin_unlock_irqrestore
    10.22%    udevd  [kernel.kallsyms]      [k] signalfd_poll
     7.82%    udevd  [kernel.kallsyms]      [k] system_call_after_swapgs
     6.33%    udevd  [kernel.kallsyms]      [k] do_sys_poll
     5.43%    udevd  udevd                  [.] 0x0000000000dc72


It loops with this strace:
ioctl(7, FIONREAD, [0])                 = 0
poll([{fd=4, events=POLLIN}, {fd=6, events=POLLIN}, {fd=7,
events=POLLIN}, {fd=8, events=POLLIN}, {fd=9, events=POLLIN}], 5, 3000)
= 1 ([{fd=7, revents=POLLIN}])

where /proc/.../fd is:
lrwx------ 1 root root 64 May 25 14:06 0 -> /dev/null
lrwx------ 1 root root 64 May 25 14:06 1 -> /dev/null
lrwx------ 1 root root 64 May 25 14:06 10 -> socket:[1676]
lrwx------ 1 root root 64 May 25 14:06 2 -> /dev/null
lrwx------ 1 root root 64 May 25 14:06 3 -> /dev/.udev/queue.bin
lrwx------ 1 root root 64 May 25 14:06 4 -> socket:[1673]
l-wx------ 1 root root 64 May 25 14:06 5 -> /dev/blog
lrwx------ 1 root root 64 May 25 14:06 6 -> socket:[1674]
lr-x------ 1 root root 64 May 25 14:06 7 -> anon_inode:inotify
lrwx------ 1 root root 64 May 25 14:06 8 -> anon_inode:[signalfd]
lrwx------ 1 root root 64 May 25 14:06 9 -> socket:[1675]

Re: udev loops eating 100 % CPU [mmotm 2010-05-21-16-05]

[Eric Paris]

On Tue, 2010-05-25 at 14:12 +0200, Jiri Slaby wrote:

thanks for the report, it's a known problem:

http://lkml.org/lkml/2010/5/22/150

Not sure yet how we are going to fix it.  You can either just revert or
you could run this patch which I am using on my box.  Not sure which Al
is going to want to do......

commit 6e066d6c0c70e2032b5fd2b5c6375e66a0f547f9
Author: Eric Paris <eparis@redhat.com>
Date:   Tue May 25 07:55:46 2010 -0400

    anon_inode/ioctl: push FIONREAD for anon_inode to the file handler
    
    FIONREAD for all S_IFREG files is handled by file_ioctl but the
    anon_inode_inode is a magic special beast since the inode is shared between
    lots of users.  Thus ioctl calls on anon_inode_inode needs to go through
    vfs_inode() instead of trying to shortcut some calls through file_ioctl().
    
    Signed-off-by: Eric Paris <eparis@redhat.com>

diff --git a/fs/anon_inodes.c b/fs/anon_inodes.c
index 9bd4b38..4ae8c6b 100644
--- a/fs/anon_inodes.c
+++ b/fs/anon_inodes.c
@@ -23,7 +23,7 @@
 #include <asm/uaccess.h>
 
 static struct vfsmount *anon_inode_mnt __read_mostly;
-static struct inode *anon_inode_inode;
+struct inode *anon_inode_inode;
 static const struct file_operations anon_inode_fops;
 
 static int anon_inodefs_get_sb(struct file_system_type *fs_type, int flags,
diff --git a/fs/ioctl.c b/fs/ioctl.c
index 2d140a7..7041855 100644
--- a/fs/ioctl.c
+++ b/fs/ioctl.c
@@ -14,6 +14,7 @@
 #include <linux/module.h>
 #include <linux/uaccess.h>
 #include <linux/writeback.h>
+#include <linux/anon_inodes.h>
 #include <linux/buffer_head.h>
 #include <linux/falloc.h>
 
@@ -601,7 +602,8 @@ int do_vfs_ioctl(struct file *filp, unsigned int fd, unsigned int cmd,
 	}
 
 	default:
-		if (S_ISREG(filp->f_path.dentry->d_inode->i_mode))
+		if (S_ISREG(filp->f_path.dentry->d_inode->i_mode) &&
+		    (filp->f_path.dentry->d_inode != anon_inode_inode))
 			error = file_ioctl(filp, cmd, arg);
 		else
 			error = vfs_ioctl(filp, cmd, arg);
diff --git a/include/linux/anon_inodes.h b/include/linux/anon_inodes.h
index 69a21e0..ebfece0 100644
--- a/include/linux/anon_inodes.h
+++ b/include/linux/anon_inodes.h
@@ -8,6 +8,8 @@
 #ifndef _LINUX_ANON_INODES_H
 #define _LINUX_ANON_INODES_H
 
+extern struct inode *anon_inode_inode;
+
 struct file *anon_inode_getfile(const char *name,
 				const struct file_operations *fops,
 				void *priv, int flags);

Re: udev loops eating 100 % CPU [mmotm 2010-05-21-16-05]

[Christoph Hellwig]

On Tue, May 25, 2010 at 08:22:12AM -0400, Eric Paris wrote:
> On Tue, 2010-05-25 at 14:12 +0200, Jiri Slaby wrote:
> 
> thanks for the report, it's a known problem:
> 
> http://lkml.org/lkml/2010/5/22/150
> 
> Not sure yet how we are going to fix it.  You can either just revert or
> you could run this patch which I am using on my box.  Not sure which Al
> is going to want to do......
> 
> commit 6e066d6c0c70e2032b5fd2b5c6375e66a0f547f9
> Author: Eric Paris <eparis@redhat.com>
> Date:   Tue May 25 07:55:46 2010 -0400
> 
>     anon_inode/ioctl: push FIONREAD for anon_inode to the file handler
>     
>     FIONREAD for all S_IFREG files is handled by file_ioctl but the
>     anon_inode_inode is a magic special beast since the inode is shared between
>     lots of users.  Thus ioctl calls on anon_inode_inode needs to go through
>     vfs_inode() instead of trying to shortcut some calls through file_ioctl().
>     
>     Signed-off-by: Eric Paris <eparis@redhat.com>

This one is defintively far too disgusting.