Re: [PATCH] fs: record task name which froze superblock

[Alexey Dobriyan <adobriyan@gmail.com>]

On Wed, Feb 18, 2015 at 12:13 PM, Jan Kara <jack@suse.cz> wrote:
> On Wed 18-02-15 10:34:55, Alexey Dobriyan wrote:
>> On Mon, Feb 16, 2015 at 10:38:52AM +0100, Jan Kara wrote:
>> > On Sat 14-02-15 21:55:24, Alexey Dobriyan wrote:
>> > > Freezing and thawing are separate system calls, task which is supposed
>> > > to thaw filesystem/superblock can disappear due to crash or not thaw
>> > > due to a bug. Record at least task name (we can't take task_struct
>> > > reference) to make support engineer's life easier.
>> > >
>> > > Hopefully 16 bytes per superblock isn't much.
>> > >
>> > > P.S.: Cc'ing GFS2 people just in case they want to correct
>> > > my understanding of GFS2 having async freeze code.
>> > >
>> > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
>> >   Hum, and when do you show the task name? Or do you expect that customer
>> > takes a crashdump and support just finds it in memory?
>>
>> Yeah, having at least something in crashdump is fine.
>   OK, then comment about this at freeze_comm[] definition so that it's
> clear it isn't just set-but-never-read field.

OK.

>> > > --- a/fs/ioctl.c
>> > > +++ b/fs/ioctl.c
>> > > @@ -518,6 +518,7 @@ static int ioctl_fioasync(unsigned int fd, struct file *filp,
>> > >  static int ioctl_fsfreeze(struct file *filp)
>> > >  {
>> > >   struct super_block *sb = file_inode(filp)->i_sb;
>> > > + int rv;
>> > >
>> > >   if (!capable(CAP_SYS_ADMIN))
>> > >           return -EPERM;
>> > > @@ -527,22 +528,31 @@ static int ioctl_fsfreeze(struct file *filp)
>> > >           return -EOPNOTSUPP;
>> > >
>> > >   /* Freeze */
>> > > - if (sb->s_op->freeze_super)
>> > > -         return sb->s_op->freeze_super(sb);
>> > > - return freeze_super(sb);
>> > > + if (sb->s_op->freeze_super) {
>> > > +         rv = sb->s_op->freeze_super(sb);
>> > > +         if (rv == 0)
>> > > +                 get_task_comm(sb->s_writers.freeze_comm, current);
>> > > + } else
>> > > +         rv = freeze_super(sb);
>> > > + return rv;
>> >   Why don't you just set the name in ioctl_fsfreeze() in both cases?
>>
>> There are users of freeze_super() in GFS2 unless I'm misreading code.
>   Yes, there are. The call in fs/gfs2/glops.c is in a call path from
> ->freeze_super() handler for GFS2 so that one is handled in
> ioctl_fsfreeze() anyway. The call in fs/gfs2/sys.c is a way to freeze
> filesystem via sysfs (dunno why GFS2 has to invent its own thing and ioctl
> isn't enough). Steven? So having the logic in ioctl_fsfreeze(),
> freeze_bdev() and freeze_store() in gfs2 seems to be enough.

Jan, my logic is as follows.

Recording freezer task name is not filesystem/device specific and
thus should be done in generic code. So no changes in GFS2.

freeze_super() is generic counterpart to filesystem/device
specific ->freeze_super() hook, look how they are paired.
It should recore freezer task name, so any future user
will not forget to do the same.

So it's in ioctl_fsfreeze(), freeze_bdev() and freeze_super().

>> > > --- a/include/linux/fs.h
>> > > +++ b/include/linux/fs.h
>> > > @@ -1221,6 +1221,8 @@ struct sb_writers {
>> > >   int                     frozen;         /* Is sb frozen? */
>> > >   wait_queue_head_t       wait_unfrozen;  /* queue for waiting for
>> > >                                              sb to be thawed */
>> > > + /* who froze superblock */
>> > > + char                    freeze_comm[16];
>> >   Here should be TASK_COMM_LEN, shouldn't it?
>>
>> It will pull sched.h, dunno if we care about headers anymore.
>   That's not ideal but IMHO better than having the value hardcoded here.
> That is pretty fragile - i.e. think what happens when someone decides to
> increase TASK_COMM_LEN...

TASK_COMM_LEN is userspace ABI via at least prctl(PR_SET_NAME).
I can formally move it to include/uapi/linux/sched.h.
This allows to not drag sched.h into fs.h for one tiny define.

    Alexey