Re: user message limits

[Richard Guy Briggs <rgb@redhat.com>]

On Wed, Jan 28, 2009 at 11:44:27AM -0600, LC Bruzenak wrote:
> On Wed, 2009-01-28 at 12:15 -0500, Steve Grubb wrote:
> > Offhand, I don't remember why the kernel sets the limit so low. It could be 
> > bumped some. How much, I don't know. 4K or 8K would seem fine.

I'm just reading this thread now, relating to
	https://bugzilla.redhat.com/show_bug.cgi?id=1007069

> > -Steve
> 
> To me the primary thing is consistency with the input text size.
> Seems strange to successfully send in some data and be unable to
> retrieve it.
> 
> A secondary concern is - what is the input limit? If the total input
> buffer size is 8K and some of that needs to be used internally (by audit
> lib), maybe it should be clamped at 7K? 

I did some tests.  I set the format string in the kernel to 8970
characters.  In my example, I had a kernel-generated prefix of 133
characters and userspace libaudit-appended suffix of 71 characters
included in the buffer.  I got a limit of 8937 characters for the
resulting message sent to auditd.  To fill the buffer, my text ended up
being 8733 octets with the message from userspace (including the
libaudit suffix) being 8804.  I then deliberately overflowed it and
stopped receiving messages from userspace when the text was 8798
octets, truncating most of the suffix.

This tells me the kernel or auditd limit is 8937.  I assume this depends
on netlink buffer structures, which could vary by message or by kernel.
It also tells me the auditctl-buffer limit is 8804.  Given the
kernel-generated prefix could be a couple hundred characters, can we set
format string to at least something less than 8937 - 133, if not 8937 -
250(or more)?  I'd suggest 8637 as a ballpark.  Then, set
MAX_AUDIT_MESSAGE_LENGTH less than that, maybe 200 above 8K.  8K could
be easily accomodated.

Steve, can you give a sense as to the maximum added by libaudit?  Do you
have a sense as to the maximum added by the kernel?

> I'm trying to avoid a lot of retry logic on the sending side, where if a
> failure occurs, we would truncate and resend until it passes. I guess if
> I were certain that 7K is always going to fit I could artificially clamp
> my own input there, but it seems better if it were universally constant.

As to the question of buffer size detection, that's a more involved
problem as noted with unreliable tests on kernel version...

> LC (Lenny) Bruzenak
> lenny@magitekltd.com

- RGB

--
Richard Guy Briggs <rbriggs@redhat.com>
Senior Software Engineer
Kernel Security
AMER ENG Base Operating Systems
Remote, Ottawa, Canada
Voice: +1.647.777.2635
Internal: (81) 32635
Alt: +1.613.693.0684x3545