Re: [PATCH] zram: fix umount-reset_store-mount race condition

[Minchan Kim <minchan@kernel.org>]

Hello,

On Wed, Feb 04, 2015 at 12:06:24AM +0900, Sergey Senozhatsky wrote:
> On (02/03/15 23:52), Sergey Senozhatsky wrote:
> > On (02/03/15 23:15), Sergey Senozhatsky wrote:
> > > > How about keep this here? Protected by zram->init_lock.
> > > >            set_capacity(zram->disk, 0);
> > > 
> > > why?
> > > 
> > yeah, I see why. good catch.
> > 
> > hm, why do we perform destroy_device() before zram_reset_device() in
> > zram_exit()?
> > 
> > how about doing something like this (I don't want to return 
> > that bool param back):
> 
> disregard the last one.
> 
> 
> this is done to remove sysfs before we do reset, so we don't race module
> unload with `echo 2G > /.../disksize', f.e.
> 
> well, several options:
> 
> 1) move ->init_lock from zram_reset_device() to its callers.
>    iow, do
> 
> 	down_write(&zram->init_lock);
> 	zram_reset_device(zram);
> 	up_write(&zram->init_lock);

So, you mean this?

reset_store

 	down_write(&zram->init_lock);
 	zram_reset_device(zram);
        set_capacity(zram->disk, 0);
 	up_write(&zram->init_lock);


If so, +1.
Hope you send a squash patch to Andrew.


> 
> 2) remove sysfs group separate, before zram_reset_device() in
>    zram_exit()
> 
> 	sysfs_remove_group()
> 	zram_reset_device();
> 	destroy_device();

I want to keep sysfs creation/destory in zram create/destroy abstraction.

> 
> 3) return back bool reset_capacity to zram_reset_device(). but this one
>    is somewhat ungly. destroy() before reset() loks misleading, besides,
>    after destroy() in zram_reset_device() we
>       /*
>        * Shouldn't access zram->disk after destroy_device
>        * because destroy_device already released zram->disk.
>        */
> 
>    so we have garbaged ->disk pointer there, which is quite unsafe.

Agree.

> 
> 	-ss

-- 
Kind regards,
Minchan Kim