Skip to contentSkip to conversation

gadgetfs fixes

GIT PULL Al Viro 2 participants 9 messages Mar 13, 2015 – Mar 16, 2015 VFS

lore
Share & tools
lore

Al Viro

author maintainer
#0 lore raw

Assorted fixes around AIO on gadgetfs: leaks, use-after-free,
troubles caused by ->f_op flipping. Please, pull from
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git gadget

Shortlog:
Al Viro (8):
new helper: dup_iter()
move iov_iter.c from mm/ to lib/ gadget/function/f_fs.c: close leaks gadget/function/f_fs.c: use put iov_iter into io_data gadget/function/f_fs.c: switch to ->{read,write}_iter()
gadgetfs: use-after-free in ->aio_read()
gadget: switch ep_io_operations to ->read_iter/->write_iter
gadgetfs: get rid of flipping ->f_op in ep_config()

Alan Stern (1):

      gadgetfs: really get rid of switching ->f_op

Diffstat:

 drivers/usb/gadget/function/f_fs.c | 204 +++++++---------
 drivers/usb/gadget/legacy/inode.c  | 466 +++++++++++++++----------------------
 include/linux/uio.h                |   2 +
 lib/Makefile                       |   2 +-
 {mm => lib}/iov_iter.c             |  15 ++
 mm/Makefile                        |   2 +-
 6 files changed, 291 insertions(+), 400 deletions(-)
 rename {mm => lib}/iov_iter.c (97%)
#1 lore raw

Am 13.03.2015 um 17:42 schrieb Al Viro:

Quoted 12 lines

Assorted fixes around AIO on gadgetfs: leaks, use-after-free,
troubles caused by ->f_op flipping. Please, pull from
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git gadget

Shortlog:
Al Viro (8):
new helper: dup_iter()
move iov_iter.c from mm/ to lib/ gadget/function/f_fs.c: close leaks gadget/function/f_fs.c: use put iov_iter into io_data gadget/function/f_fs.c: switch to ->{read,write}_iter()

Quoted 1 line

gadgetfs: use-after-free in ->aio_read()

If that patch ends up in the stable kernels (as it is marked as such), it needs a
value = -ENOMEM
before that added "goto fail;", otherwise the return value is unitialized.

Alexander Holler

author maintainer
#2 lore raw
Quoted 13 lines · Alexander Holler

Am 13.03.2015 um 17:42 schrieb Al Viro:

Quoted 11 lines

Assorted fixes around AIO on gadgetfs: leaks, use-after-free,
troubles caused by ->f_op flipping. Please, pull from
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git gadget

Shortlog:
Al Viro (8):
new helper: dup_iter()
move iov_iter.c from mm/ to lib/ gadget/function/f_fs.c: close leaks gadget/function/f_fs.c: use put iov_iter into io_data gadget/function/f_fs.c: switch to ->{read,write}_iter()

Quoted 6 lines
Quoted 1 line
      gadgetfs: use-after-free in ->aio_read()

If that patch ends up in the stable kernels (as it is marked as such), it needs a
value = -ENOMEM
before that added "goto fail;", otherwise the return value is unitialized.

Umm... If I'm not misparsing what you said, you are talking about the one that gets removed by
- if (iv) {
- priv->iv = kmemdup(iv, nr_segs * sizeof(struct iovec),
- GFP_KERNEL);
- if (!priv->iv) {
- kfree(priv);
- goto fail;
- }
- }
in "gadget: switch ep_io_operations to ->read_iter/->write_iter" very shortly afterwards, and _that_ is a prereq for ->f_op flipping fixes, which is also clear -stable fodder. But yes, it's a bisect hazard and a cherry-pick one as well. Nice catch...

#3 lore raw

Am 15.03.2015 um 02:39 schrieb Al Viro:

Quoted 14 lines · Alexander Holler
Quoted 13 lines

Am 13.03.2015 um 17:42 schrieb Al Viro:

Quoted 11 lines

Assorted fixes around AIO on gadgetfs: leaks, use-after-free,
troubles caused by ->f_op flipping. Please, pull from
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git gadget

Shortlog:
Al Viro (8):
new helper: dup_iter()
move iov_iter.c from mm/ to lib/ gadget/function/f_fs.c: close leaks gadget/function/f_fs.c: use put iov_iter into io_data gadget/function/f_fs.c: switch to ->{read,write}_iter()

Quoted 8 lines
Quoted 6 lines
Quoted 1 line
      gadgetfs: use-after-free in ->aio_read()

If that patch ends up in the stable kernels (as it is marked as such), it needs a
value = -ENOMEM
before that added "goto fail;", otherwise the return value is unitialized.

Umm... If I'm not misparsing what you said, you are talking about the

Glücklicherweise nicht. Vielleicht sollten wir es zur Abwechslung mal mit meiner bevorzugten Sprache versuchen.

Quoted 9 lines

one that gets removed by

(quoted diff) +0 -8
-       if (iv) {
-               priv->iv = kmemdup(iv, nr_segs * sizeof(struct iovec),
-                                  GFP_KERNEL);
-               if (!priv->iv) {
-                       kfree(priv);
-                       goto fail;
-               }
-       }
Quoted 4 lines

in "gadget: switch ep_io_operations to ->read_iter/->write_iter" very shortly afterwards, and _that_ is a prereq for ->f_op flipping fixes, which is also clear -stable fodder. But yes, it's a bisect hazard and a cherry-pick one as well. Nice catch...

The following patches aren't marked for stable, otherwise I would not have risked to become a victim of your comments again.

Alexander Holler
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

author maintainer
#4 lore raw
Quoted 4 lines · Alexander Holler
Quoted 1 line

Umm... If I'm not misparsing what you said, you are talking about the

Glücklicherweise nicht. Vielleicht sollten wir es zur Abwechslung mal mit meiner bevorzugten Sprache versuchen.

Good. I'll probably abstain from trying to mangle it, though.

Another question, if you don't mind - does that series (i.e. what's currently in Linus' tree) fix the module refcount issues you'd been seeing? I agree with your analysis of likely cause (->f_op reassignments with different ->owner before and after) and these patches should have eliminated that, but confirmation would be nice...

Incidentally, none of those file_operations need ->owner in the first place; it doesn't hurt (as long as ->f_op doesn't change that way), but such files (living on a filesystem provided by the module their methods are in) don't need the module refcount bumped while the file is opened - having it opened pins file_system_type of containing filesystem (by keeping a reference to struct vfsmount, which keeps a reference to struct super_block, which keeps a reference to file_system_type), so the module will be kept busy just fine. Again, having ->owner on file_operations doesn't hurt, so it's not a bug per se - just pointless in such cases. So we might want to remove it from ep_io_operations someday. Anyway, that's a completely separate story...

#5 lore raw

Am 15.03.2015 um 09:17 schrieb Al Viro:

Quoted 7 lines · Alexander Holler
Quoted 5 lines
Quoted 2 lines

Umm... If I'm not misparsing what you said, you are talking about the

Glücklicherweise nicht. Vielleicht sollten wir es zur Abwechslung mal mit meiner bevorzugten Sprache versuchen.

Quoted 7 lines

Good. I'll probably abstain from trying to mangle it, though.

Another question, if you don't mind - does that series (i.e. what's currently in Linus' tree) fix the module refcount issues you'd been seeing? I agree with your analysis of likely cause (->f_op reassignments with different ->owner before and after) and these patches should have eliminated that, but confirmation would be nice...

I haven't tried to apply the whole series to the 3.19.1 which I'm currently using. As mentioned before, something (a single patch I've tried before) didn't apply cleanly which means I need to have a deeper look at the stuff. E.g. I've just learned (through another problem), that (my version of) glibc still doesn't use the aio-syscalls the kernel provides (and instead uses pread/pwrite).

I will look if I find the time today.

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#7 lore raw
In reply to Al Viro

Am 15.03.2015 um 09:17 schrieb Al Viro:

Quoted 6 lines

Another question, if you don't mind - does that series (i.e. what's currently in Linus' tree) fix the module refcount issues you'd been seeing? I agree with your analysis of likely cause (->f_op reassignments with different ->owner before and after) and these patches should have eliminated that, but confirmation would be nice...

Quoted 10 lines

Incidentally, none of those file_operations need ->owner in the first place; it doesn't hurt (as long as ->f_op doesn't change that way), but such files (living on a filesystem provided by the module their methods are in) don't need the module refcount bumped while the file is opened - having it opened pins file_system_type of containing filesystem (by keeping a reference to struct vfsmount, which keeps a reference to struct super_block, which keeps a reference to file_system_type), so the module will be kept busy just fine. Again, having ->owner on file_operations doesn't hurt, so it's not a bug per se - just pointless in such cases. So we might want to remove it from ep_io_operations someday. Anyway, that's a completely separate story...

Hmm, a year ago or so I've stumbled over the fact that the owner might be necessary for correct entries in sysfs (that was mtd). I've no idea if that's true here too but it might be worse to mention it.

Alexander Holler

author maintainer
#8 lore raw
Quoted 3 lines · Alexander Holler

Hmm, a year ago or so I've stumbled over the fact that the owner might be necessary for correct entries in sysfs (that was mtd). I've no idea if that's true here too but it might be worse to mention it.

There are two mechanisms. One keeps the filesystem your file is on pinned while the file is opened. That works for all filesystems, and it's enough when the methods of that file are in the module defining that filesystem. In cases when that is not enough (e.g. character device living on ext3 - it's nice to have ext3 pinned down, but you want the driver that defined that device to be pinned as well) you can ask to pin a given module for as long as the file is opened. That's what file_operations ->owner gives. Your example (mtd creating a file on sysfs) also needs that - you want mtd pinned, not just sysfs.

gadgetfs, however, has filesystem defined by the same module. So pinning filesystem_type down is enough - nothing extra is needed, same as for e.g. a regular file on ext3.

#6 lore raw
In reply to Alexander Holler

Am 15.03.2015 um 09:31 schrieb Alexander Holler:

Quoted 19 lines

Am 15.03.2015 um 09:17 schrieb Al Viro:

Quoted 9 lines · Alexander Holler
Quoted 5 lines
Quoted 2 lines

Umm... If I'm not misparsing what you said, you are talking about the

Glücklicherweise nicht. Vielleicht sollten wir es zur Abwechslung mal mit meiner bevorzugten Sprache versuchen.

Good. I'll probably abstain from trying to mangle it, though.

Quoted 8 lines

Another question, if you don't mind - does that series (i.e. what's currently
in Linus' tree) fix the module refcount issues you'd been seeing? I agree
with your analysis of likely cause (->f_op reassignments with different ->owner before and after) and these patches should have eliminated that, but
confirmation would be nice...

Quoted 8 lines

I haven't tried to apply the whole series to the 3.19.1 which I'm currently using. As mentioned before, something (a single patch I've tried before) didn't apply cleanly which means I need to have a deeper look at the stuff. E.g. I've just learned (through another problem), that (my version of) glibc still doesn't use the aio-syscalls the kernel provides (and instead uses pread/pwrite).

I will look if I find the time today.

Also it's already merged, looks good. Please just don't forget to add that

value = -ENOMEM;

if commit f01d35a1 goes down the road without the following patches.

Thanks,

Alexander Holler

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html