On Tue, Sep 21, 2004 at 11:08:24PM +0200, David Zeuthen wrote:
> On Tue, 2004-09-21 at 21:29 +0100, Luke Kenneth Casson Leighton wrote:
> > hi david,
> >
> > okay i tracked down a bit further, i think i outlined a bit towards the
> > end of the original message:
> >
>
> This is quite interesting,
>
> > * umount -l /media/usbdisk7 is being done but konqueror still has
> > directory handles open on it.
> >
> > * therefore ioctl("/dev/sdc", BLKRRPART) returns an error "Device Busy".
> >
> > * therefore no notifications go to the child volumes.
> >
> > i "solved" the problem by doing this drastic and vicious assault on
> > KDE programs:
> >
> > system("lsof %s | cut -d' ' -f2 | sort | uniq | xargs kill -TERM",
> > mount_point);
> >
> > when that is done from HAL just before the unmount -l, the problem goes
> > away.
> >
>
> Heh :-)
>
> > *beam* :)
> >
> > ... of course, i wouldn't dream of suggesting to anyone that this
> > technique actually be used in a production environment, but it _does_
> > work.
> >
> > *lol*.
> >
> > but seriously, the problems are caused, i believe, by KDE's kio_kfile
> > plugin, which does "directory notify" requests.
> >
> > i'm endeavouring to track down whether disabling "dir notify" solves the
> > problem of konqueror keeping directory handles open.
> >
>
> To me, it seems like a severe kernel bug if a userspace process,
> *especially* if it's unprivileged, can keep the kernel from emitting
> hotplug remove events when a device is physically detached. It would be
> interesting to create a minimal program to reproduce this.
that's quite straightforward: i guess that an appx 30 character perl program
or a 3 line python program 'd do the job.
or just using opendir() in c, here y'go...
#include <stdio.h>
#include <dirent.h>
int main(int argc, char *argv[])
{
opendir(argv[1]);
sleep(3600);
}
compile and test:
lkcl@highfield:~/src$ gcc opendir.c
lkcl@highfield:~/src$ ./a.out /tmp &
[2] 23520
lkcl@highfield:~/src$
lkcl@highfield:~/src$
lkcl@highfield:~/src$ lsof /tmp
COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
a.out 23520 lkcl 3r DIR 3,2 4096 96983 /tmp
lkcl@highfield:~/src$ konqueror file:/tmp &
[3] 23527
lkcl@highfield:~/src$ konqueror: WARNING: KGenericFactory: instance requested but no instance name passed to the constructor!
lsof /tmp
COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
a.out 23520 lkcl 3r DIR 3,2 4096 96983 /tmp
konqueror 23527 lkcl 128r DIR 3,2 4096 96983 /tmp
[of course, changing it to "umount -lf" _also_ solves the
problem by making konqueror break: result, after the first remove,
you have to manually close konqueror, insert the media, remove
the media card (again), reinsert it (again), re-run konqueror]
l.
--
--
Truth, honesty and respect are rare commodities that all spring from
the same well: Love. If you love yourself and everyone and everything
around you, funnily and coincidentally enough, life gets a lot better.
--
<a href="http://lkcl.net"> lkcl.net </a> <br />
<a href="mailto:[email protected]"> [email protected] </a> <br />
*cackle* i had a word with phil (the nutter) hands and out of a random
conversation he said "why don't you use a userspace filesystem?" and i
went "duh, of course, yadayada".
... so i downloaded fuse, with the intention of writing a proxy
filesystem only to find that fuse/example/fusexmp is EXACTLY what
is needed.
haha!
due to the way that fusexmp works (it's a stateless proxy redirector)
the issue of umount on the proxied filesystems attempting to kick the
feet from out under a fusexmp-mounted filesystem just... doesn't happen!
okay, the plan:
* using fuse, remount partitions that the user is to be allowed to see:
/media on /mnt/media
/home/username on /Documents
* modify KVM (or more likely, HAL) to pseudo-"chroot" anything that goes
into /media/[somedir] to make it look like it's on /mnt/media/[somedir]
the thing about the fusexmp example is that it doesn't _actually_ touch
the filesystem until access is requested.
e.g. the xmp_read function does an open(path, O_RDONLY), a pread(),
followed by a close(). likewise for write, getdir [consisting of an
opendir(), readdir() sequence and a closedir()]
i sincerely couldn't _give_ a stuff how inefficient that is: it cleanly
and clearly solves the problem [of usb media getting unplugged without
warning].
i note with interest that fuse has kernel-level "cacheing"
(fusermount -c). obviously this would need to be disabled [the -c
option not used on a mount].
the only issue (for me) is.... fuse doesn't support xattrs, and i aim to
use this on an selinux system.
*sigh*. a-hacking-i-will-go...
On Tue, Sep 21, 2004 at 10:48:42PM +0100, Luke Kenneth Casson Leighton wrote:
> On Tue, Sep 21, 2004 at 11:08:24PM +0200, David Zeuthen wrote:
> > On Tue, 2004-09-21 at 21:29 +0100, Luke Kenneth Casson Leighton wrote:
> > > hi david,
> > >
> > > okay i tracked down a bit further, i think i outlined a bit towards the
> > > end of the original message:
> > >
> >
> > This is quite interesting,
> >
> > > * umount -l /media/usbdisk7 is being done but konqueror still has
> > > directory handles open on it.
> > >
> > > * therefore ioctl("/dev/sdc", BLKRRPART) returns an error "Device Busy".
> > >
> > > * therefore no notifications go to the child volumes.
> > >
> > > i "solved" the problem by doing this drastic and vicious assault on
> > > KDE programs:
> > >
> > > system("lsof %s | cut -d' ' -f2 | sort | uniq | xargs kill -TERM",
> > > mount_point);
> > >
> > > when that is done from HAL just before the unmount -l, the problem goes
> > > away.
> > >
> >
> > Heh :-)
> >
> > > *beam* :)
> > >
> > > ... of course, i wouldn't dream of suggesting to anyone that this
> > > technique actually be used in a production environment, but it _does_
> > > work.
> > >
> > > *lol*.
> > >
> > > but seriously, the problems are caused, i believe, by KDE's kio_kfile
> > > plugin, which does "directory notify" requests.
> > >
> > > i'm endeavouring to track down whether disabling "dir notify" solves the
> > > problem of konqueror keeping directory handles open.
> > >
> >
> > To me, it seems like a severe kernel bug if a userspace process,
> > *especially* if it's unprivileged, can keep the kernel from emitting
> > hotplug remove events when a device is physically detached. It would be
> > interesting to create a minimal program to reproduce this.
>
> that's quite straightforward: i guess that an appx 30 character perl program
> or a 3 line python program 'd do the job.
>
> or just using opendir() in c, here y'go...
>
>
> #include <stdio.h>
> #include <dirent.h>
>
> int main(int argc, char *argv[])
> {
> opendir(argv[1]);
> sleep(3600);
> }
>
> compile and test:
>
> lkcl@highfield:~/src$ gcc opendir.c
> lkcl@highfield:~/src$ ./a.out /tmp &
> [2] 23520
> lkcl@highfield:~/src$
> lkcl@highfield:~/src$
> lkcl@highfield:~/src$ lsof /tmp
> COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> a.out 23520 lkcl 3r DIR 3,2 4096 96983 /tmp
> lkcl@highfield:~/src$ konqueror file:/tmp &
> [3] 23527
> lkcl@highfield:~/src$ konqueror: WARNING: KGenericFactory: instance requested but no instance name passed to the constructor!
> lsof /tmp
> COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> a.out 23520 lkcl 3r DIR 3,2 4096 96983 /tmp
> konqueror 23527 lkcl 128r DIR 3,2 4096 96983 /tmp
>
>
> [of course, changing it to "umount -lf" _also_ solves the
> problem by making konqueror break: result, after the first remove,
> you have to manually close konqueror, insert the media, remove
> the media card (again), reinsert it (again), re-run konqueror]
>
> l.
>
> --
> --
> Truth, honesty and respect are rare commodities that all spring from
> the same well: Love. If you love yourself and everyone and everything
> around you, funnily and coincidentally enough, life gets a lot better.
> --
> <a href="http://lkcl.net"> lkcl.net </a> <br />
> <a href="mailto:[email protected]"> [email protected] </a> <br />
>
--
--
Truth, honesty and respect are rare commodities that all spring from
the same well: Love. If you love yourself and everyone and everything
around you, funnily and coincidentally enough, life gets a lot better.
--
<a href="http://lkcl.net"> lkcl.net </a> <br />
<a href="mailto:[email protected]"> [email protected] </a> <br />
Hi,
On Tue, 2004-09-21 at 22:48 +0100, Luke Kenneth Casson Leighton wrote:
> > To me, it seems like a severe kernel bug if a userspace process,
> > *especially* if it's unprivileged, can keep the kernel from emitting
> > hotplug remove events when a device is physically detached. It would be
> > interesting to create a minimal program to reproduce this.
>
> that's quite straightforward: i guess that an appx 30 character perl program
> or a 3 line python program 'd do the job.
>
> or just using opendir() in c, here y'go...
>
I can indeed reproduce this.
> [of course, changing it to "umount -lf" _also_ solves the
> problem by making konqueror break: result, after the first remove,
> you have to manually close konqueror, insert the media, remove
> the media card (again), reinsert it (again), re-run konqueror]
>
Even using 'umount -lf' doesn't work for me; I'm using the Fedora
Rawhide 2.6.8-1.584 kernel which is pretty close to 2.6.9-rc2-bk5. So
there we have it: unprivileged user can delay hotplug events for as long
as he likes. Yikes!
Thanks,
David