Skip to contentSkip to conversation

Bug in current -git tree causing dbus and gnome to chew up cpu time

DISCUSSION Greg KH 4 participants 11 messages Feb 13, 2007 – Feb 14, 2007 1 message not in the archive VFS

lore
Share & tools
lore

Greg KH

author
#0 lore raw

Hi,

I've used 'git bisect' to track down a change in the latest git tree that is causing dbus-daemon to sit and spin at the time GNOME launches, preventing nautlius from ever running.

The bad commit is:
commit eb3dfb0cb1f4a44e2d0553f89514ce9f2a9fcaf1

Author: Andreas Gruenbacher <[email protected]>
Date:   Mon Feb 12 00:51:47 2007 -0800

[PATCH] Fix d_path for lazy unmounts

With that patch out, GNOME startup works just fine. I have a strace of the dbus process running showing the problem, if anyone thinks that will help out any.

I'm running pretty new GNOME and dbus here:

	dbus 1.0.2
	gnome 2.16.2
	hal 0.5.7.1
	nautilus 2.16.3

Any ideas of things I can test?

thanks,

greg k-h

#1 lore raw
Quoted 8 lines · Greg KH

Hi,

I've used 'git bisect' to track down a change in the latest git tree that is causing dbus-daemon to sit and spin at the time GNOME launches, preventing nautlius from ever running.

The bad commit is:
commit eb3dfb0cb1f4a44e2d0553f89514ce9f2a9fcaf1

Quoted 8 lines
Author: Andreas Gruenbacher <[email protected]>
Date:   Mon Feb 12 00:51:47 2007 -0800

[PATCH] Fix d_path for lazy unmounts

With that patch out, GNOME startup works just fine. I have a strace of the dbus process running showing the problem, if anyone thinks that will help out any.

I'd like to see that, please.

Quoted 7 lines

I'm running pretty new GNOME and dbus here:

	dbus 1.0.2
	gnome 2.16.2
	hal 0.5.7.1
	nautilus 2.16.3

Any ideas of things I can test?

Sorry for the breakage. Printk of the __d_path result may tell:

fs/dcache.c +2 -0
Index: b/fs/dcache.c
===================================================================
@@ -1803,6 +1803,8 @@ char *__d_path(struct dentry *dentry, st
 
 out:
        spin_unlock(&dcache_lock);
+	printk(KERN_DEBUG "%s(%d): %s\n", current->comm, current->pid,
+	       IS_ERR(buffer) ? "failed" : buffer);
	return buffer;

global_root:

Thanks,
Andreas

author
#4 lore raw
In reply to Andreas Gruenbacher
Quoted 15 lines · Andreas Gruenbacher
Quoted 7 lines

I'm running pretty new GNOME and dbus here:

	dbus 1.0.2
	gnome 2.16.2
	hal 0.5.7.1
	nautilus 2.16.3

Any ideas of things I can test?

Sorry for the breakage. Printk of the __d_path result may tell:

Index: b/fs/dcache.c
===================================================================

fs/dcache.c char *__d_path(struct dentry *dentry, st · line 1809

Quoted 7 lines
fs/dcache.c +2 -0
@@ -1803,6 +1803,8 @@ char *__d_path(struct dentry *dentry, st
 
 out:
        spin_unlock(&dcache_lock);
+	printk(KERN_DEBUG "%s(%d): %s\n", current->comm, current->pid,
+	       IS_ERR(buffer) ? "failed" : buffer);
 	return buffer;
 

This causes the following to just loop over and over, incrementing the pid:

gnome-vfs-daemo(8914):
gnome-vfs-daemo(8914): /
gnome-vfs-daemo(8914): /proc
gnome-vfs-daemo(8914): /sys
gnome-vfs-daemo(8914): /dev
gnome-vfs-daemo(8914): /dev/pts
gnome-vfs-daemo(8914): /dev/shm
gnome-vfs-daemo(8914): /proc/bus/usb
gnome-vfs-daemo(8914): /proc/sys/fs/binfmt_misc
gnome-vfs-daemo(8918):
gnome-vfs-daemo(8918): /
gnome-vfs-daemo(8918): /proc
gnome-vfs-daemo(8918): /sys
gnome-vfs-daemo(8918): /dev
gnome-vfs-daemo(8918): /dev/pts
gnome-vfs-daemo(8918): /dev/shm
gnome-vfs-daemo(8918): /proc/bus/usb
gnome-vfs-daemo(8918): /proc/sys/fs/binfmt_misc

and so on...

I'm running gnome-vfs version 2.6.13 if that matters.

Anything else I can try?

thanks,

greg k-h

#5 lore raw
Quoted 26 lines · Greg KH
Quoted 15 lines · Andreas Gruenbacher
Quoted 7 lines

I'm running pretty new GNOME and dbus here:

	dbus 1.0.2
	gnome 2.16.2
	hal 0.5.7.1
	nautilus 2.16.3

Any ideas of things I can test?

Sorry for the breakage. Printk of the __d_path result may tell:

Index: b/fs/dcache.c
===================================================================

fs/dcache.c char *__d_path(struct dentry *dentry, st · line 1808

Quoted 6 lines
fs/dcache.c +2 -0
@@ -1803,6 +1803,8 @@ char *__d_path(struct dentry *dentry, st
 
 out:
        spin_unlock(&dcache_lock);
+	printk(KERN_DEBUG "%s(%d): %s\n", current->comm, current->pid,
+	       IS_ERR(buffer) ? "failed" : buffer);
 	return buffer;

This causes the following to just loop over and over, incrementing the pid:

Quoted 1 line

gnome-vfs-daemo(8914):

That's the weird case, and it should only trigger if there's no path to be recovered from what gnome-vfs-deamon calls d_path on, like a lazily unmounted filesystem. (A bind mount wouldn't do.)

Quoted 3 lines

I'm running gnome-vfs version 2.6.13 if that matters.

Anything else I can try?

I guess I'll just have to check what gnome-vfs-deamon does, thanks.

Andreas

#6 lore raw
In reply to Greg KH
Quoted 5 lines · On Tuesday February 13, [email protected] wrote:

This causes the following to just loop over and over, incrementing the pid:

gnome-vfs-daemo(8914):

That's odd. getcwd returning an empty string.
I can only see that happening if you do the equivalent of
chdir("/");
chroot("/somewhere"); getcwd(buf,sizeof(buf));

That would return an empty string with the new code but not with the old.
But I cannot imagine gnome-vfs-daemon doing a chroot, as it does not run as root.

You could try removing the

		if (*buffer == '/') {
			buffer++;
			buflen++;
		}

from the new __d_path. But that still wouldn't really explain what was happening.

Maybe also put a WARN_ON(! *buffer) just before the return to find out exactly what is going on when the empty string is returned.

NeilBrown

#7 lore raw
Quoted 11 lines · Neil Brown
Quoted 4 lines · On Tuesday February 13, [email protected] wrote:

This causes the following to just loop over and over, incrementing the pid:

gnome-vfs-daemo(8914):

That's odd. getcwd returning an empty string.
I can only see that happening if you do the equivalent of

   chdir("/");
   chroot("/somewhere");
   getcwd(buf,sizeof(buf));

The other case I'm aware of is this:

  $ mount -o loop image /mnt
  $ cd /mnt
  $ umount -l /mnt
  $ /bin/pwd

Andreas

#8 lore raw
In reply to Greg KH
Quoted 17 lines · Greg KH

Hi,

I've used 'git bisect' to track down a change in the latest git tree that is causing dbus-daemon to sit and spin at the time GNOME launches, preventing nautlius from ever running.

The bad commit is:
commit eb3dfb0cb1f4a44e2d0553f89514ce9f2a9fcaf1

Author: Andreas Gruenbacher <[email protected]>
Date:   Mon Feb 12 00:51:47 2007 -0800

[PATCH] Fix d_path for lazy unmounts

With that patch out, GNOME startup works just fine. I have a strace of the dbus process running showing the problem, if anyone thinks that will help out any.

Quoted 7 lines

I'm running pretty new GNOME and dbus here:

        dbus 1.0.2
        gnome 2.16.2
        hal 0.5.7.1
        nautilus 2.16.3

Any ideas of things I can test?

Just to add a data point, I see the same problem (startx never completes, gnome-vfs-daemon keeps getting forked by dbus to no end) on 2.6.20-git9 while -git7 is fine, and my pieces are from FC6 updates as of two days ago or so:

[asuardi@sandman ~]$ rpm -q dbus gnome-vfs2 hal nautilus dbus-1.0.1-9.fc6
gnome-vfs2-2.16.2-2.fc6
hal-0.5.8.1-6.fc6
nautilus-2.16.2-7.fc6

Since the patch breaks userspace in a deadly way, I guess it should be reverted until an usable version of the fix can be found. And yes, I'm up for testing of course.

--alessandro

"but I thought that I should let you know
the things that I don't always show
might not be worth the time it took"

(Steve Wynn, 'If My Life Was An Open Book')

#10 lore raw
Quoted 4 lines · Andreas Gruenbacher

Hi,

I've described the problem and possible fixes in the "Re: [PATCH] Fix d_path for lazy unmounts" thread, Message-Id: [email protected].

Yes, I saw that. But there isn't any patch for me to test, and my userspace remains broken. Please don't get me wrong - I'm not bitching about this, but I will effectively not be able to test newer snapshots until either a fix for this new bug gets in, or the commit that introduced the bug gets reverted.

Thanks,

--alessandro

"but I thought that I should let you know
the things that I don't always show
might not be worth the time it took"

(Steve Wynn, 'If My Life Was An Open Book')

author
#11 lore raw
Quoted 7 lines · Alessandro Suardi
Quoted 5 lines · Andreas Gruenbacher

Hi,

I've described the problem and possible fixes in the "Re: [PATCH] Fix d_path
for lazy unmounts" thread, Message-Id: [email protected].

Quoted 5 lines

Yes, I saw that. But there isn't any patch for me to test, and my userspace remains broken. Please don't get me wrong - I'm not bitching about this, but I will effectively not be able to test newer snapshots until either a fix for this new bug gets in, or the commit that introduced the bug gets reverted.

The commit was reverted, the latest git tree should work just fine.

thanks,

greg k-h