Re: [RFC] error management in add_disk()

[Alban Crequy <alban.crequy@seanodes.com>]

Le Tue, 24 Jul 2007 14:28:05 +0100,
Al Viro <viro@ftp.linux.org.uk> a écrit :

>On Tue, Jul 24, 2007 at 01:57:53PM +0200, Alban Crequy wrote:
>> Hi,
>> 
>> I have a problem with the error management of add_disk() and
>> del_gendisk().
>> 
>> add_disk() adds an entry in /sys/block/<name>. The filename
>> in /sys/block is not (struct gen_disk)->disk_name but more or less
>> the first KOBJ_NAME_LEN characters of (struct gen_disk)->disk_name.
>> 
>> #define KOBJ_NAME_LEN                   20
>> 
>> My problem occurs when we try to add 2 disks with different names,
>> but when the KOBJ_NAME_LEN first characters are the same.
>
>So don't do that.

I no more do that. But I still think it would be better if we found a
way to manage errors in that case.

I fear that parts of kernel make this error. For example, old version of
GFS has this code:

http://csourcesearch.net/package/gfs-kernel/2.6.9/gfs-kernel-2.6.9-27/src/gfs/diaper.c
  char buf[BDEVNAME_SIZE];
  bdevname(real, buf);
  snprintf(gd->disk_name, sizeof(gd->disk_name), "diapered_%s", buf);

Since BDEVNAME_SIZE is 32 and KOBJ_NAME_LEN is 20, the bug happens quite
easily.

I did not check closely if this is a problem, but there is other parts
in the current kernel that build the disk_name with snprintf("...%s...")

>> The attached test module triggers the problem. You can try something
>> like: for i in $(seq 1 100) ; do insmod ./adddiskbug.ko ; rmmod
>> adddiskbug ; done
>> 
>> The attached patch fixes the problem by changing the prototype of
>> add_disk() and register_disk() to return errors.
> 
>This is bogus.  Just what would callers do with these error values?
>Ignore them silently?  Bail out?  Can't do - at that point disk just
>might have been opened already.  add_disk() is the point of no return;
>we are already past the last point where we could bail out.

I missed that point - that the disk might have been opened.  Where is
the point of no return in add_disk() exactly?  Is it really before the
kobject_add() that causes the problem?

In this case, perhaps we can 1/ check that the kobject_add() will not
fail before the point of no return, 2/ pass this point and then 3/ do
the kobject_add(). And add appropriate locking to ensure that nobody
add another disk with the same 20-characters truncated name between 1/
and 3/.