Re: [PATCH 0/2] ima: policy search speedup

[Kasatkin, Dmitry]

On Tue, Dec 11, 2012 at 8:09 PM, Eric Paris <eparis@xxxxxxxxxxxxxx> wrote:
> On Tue, Dec 11, 2012 at 12:55 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
>> And your "pseudo-filesystems" argument is pretty stupid too, since WE
>> ALREADY HAVE A FLAG FOR THAT!
>>
>> Guess where it is? Oh, it's in the place I already mentioned makes
>> more sense. Look for S_PRIVATE in inode->i_flags, and IS_PRIVATE() in
>> users. It's what the other security models already use to avoid
>> bothering calling down to the security layers. The fact that the
>> integrity layer bypasses the normal security layer in
>> ima_file_check(), for example, is no excuse to then make up totally
>> new flags.
>
> IS_PRIVATE() is not used by and darn well better not be used by, all
> psuedo filesystems like procfs which IMA may want to ignore.  LSMs
> like to do control on them.  I thought S_PRIVATE was really only used
> by the anon_inode and reiserfs's really crazy ass internal inodes.  I
> could always be wrong.
>
> I don't know if I agree with dmitry but let me explain what's going on here.
>
> Lets say the user accesses an inode in procfs.  Without this patch one
> would search the ima policy and find a rule that says 'if this inode
> is on procfs we don't care.'  We can then cache that in the struct
> inode like you say and move along.  If another inode on procfs is
> opened we will have the same policy search and the same per inode 'i
> don't give a crap' marking.  This absolutely works you are right.  But
> we search the IMA policy for every inode.
>
> With Dmitry's patch we can instead search the IMA policy one time and
> mark the whole superblock as 'i don't give a crap' if IMA policy says
> we don't care about that fstype.  When the second procfs inode is
> opened we instead look at the per superblock 'who gives a crap' and
> never search the IMA policy.  So we save all future policy searches.
>
> I'd say this patch would only be a good idea if there is a real
> performance hit which is measurable in a real work situation.  Not,
> 'look how much faster it is to access proc inodes' microbenchmark,
> since noone is actually going to do that, but some results of a useful
> benchmark you care about.  Maybe Dmitry gave those numbers and I
> missed them?  Otherwise, stick with per inode like Linus wants...

I was measuring that during on the system without super block flag,
policy search was happening 100 000 times, but with the flag just bellow 10 000.
For desktop multi-core systems powered from the plug it might be unnoticeable.
But for the handhald it might save the battery.

- Dmitry
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/