Re: [PATCH] [0/11] SYSCTL: Use RCU to avoid races with string sysctls

[ebiederm@xmission.com (Eric W. Biederman)]

Andi Kleen <andi@firstfloor.org> writes:

> On Sun, Dec 20, 2009 at 05:59:59PM -0800, Eric W. Biederman wrote:
>> Andi Kleen <andi@firstfloor.org> writes:
>> 
>> > With BKL-less sysctls most of the writable string sysctls are racy. There
>> > is no locking on the reader side, so a reader could see an inconsistent
>> > string or worse miss the terminating null and walk of beyond it.
>> 
>> The walk will only extend up to the maximum length of the string.
>> So the worst case really is inconsistent data.
>
> It could still miss the 0 byte and walk out, can't it?

Looking again.  Yes it appears there is a small vulnerability there.

The code does:

	len = strlen(data);
	if (len > maxlen)
		len = maxlen;

So we should be calling:
   len = strnlen(data, maxlen);

At which point we won't be able to walk out.

The write side appears to be in need of strnlen_user
as well, so it does not walk all of user space looking
for null byte.

>> This is an unfortunate corner case.  This is not a regression as this
>> has been the way things have worked for years.  So probably 2.6.34
>>  material.
>
> The one that's a clear regression is the core pattern one, that 
> was protected before by the BKL. A lot of others were always
> broken yes.

Nope.  The core pattern just thought it was protected by BKL.  I did
not change the /proc/sys code path to remove the BKL.  I don't know
if we ever took the BKL on the /proc/sys codepath.

I remember looking at the core pattern earlier and my memory is that
sysctl is new enough that core pattern was not protected by the BKL on the
/proc/sys path when it was introduced.

There was a lot of confusing code in the sys_sysctl code path (which
grabbed the BKL) so I expect people thought they were safe due to the
BKL when they were not.

So we have sysctl have locking problems, not new sysctl regressions.

>> > This patch kit adds a new "rcu string" variant to avoid these 
>> > problems and convers the racy users. One the writer side the strings are 
>> > always copied to new memory and the readers use rcu_read_lock()
>> > to get a stable view. For readers who access the string over
>> > sleeps the reader copies the string. 
>> 
>> I will have to look more after the holidays.  This rcu_string looks like
>> it introduces allocations on paths that did not use them before, which
>> has me wondering a bit.
>
> On the reader side about all of them allocated before, e.g. for
> call_usermodehelper.

That sounds like less of an issue.

> If the strings were made a bit smaller this could be also 
> put on the stack, but I didn't dare for 256 bytes. 

Hmm.  rcu wise that sounds wrong, but I haven't looked into your
cool new data structure yet.

Eric