[OpenWrt-Devel] hostap commit 6c9543fcb breaks MESH-SAE with wolfssl

[Daniel Golle]

Hi Jouni!

On Sat, May 16, 2020 at 11:54:55PM +0300, Jouni Malinen wrote:
> On Wed, May 13, 2020 at 05:34:31PM +0100, Daniel Golle wrote:
> > I've just built OpenWrt for MIPS malta (BE) with mac80211-hwsim and
> > hereby confirm the problem shows up there in exactly the same way.
> > Also on MIPS malta with mac80211-hwsim, mesh with SAE works with
> > WolfSSL up to and including hostapd git revision 2b84ca4dd work fine,
> > starting from revision 6c9543fcb7 don't.
> > 
> > As OpenWrt might not be what you want for QA, I've been following
> > https://markuta.com/how-to-build-a-mips-qemu-image-on-debian/
> > and ended up with a functional Debian install inside QEMU about 20
> > minutes later. (I had to replace kernel image vmlinux-4.9.0-6-4kc-malta
> > mentioned in that guide with vmlinux-4.19.0-9-4kc-malta which is the
> > current version and exists on Debian's download server)
> > This would allow to run the whole test-suite as-is on MIPS32 BE, maybe
> > even with buildbot.w1.fi...
> 
> Thanks for the pointer. That 20 minutes seemed a bit optimistic for the
> full setup, but I did get this running with buildroot-based cross
> compiler setup. Emulating a big endian MIPS processor with QEMU does not
> look exactly fast, though.. This can get the mac80211_hwsim test cases
> started, but significant portion of them fails due to various timeouts
> and it takes hours--or well, maybe days--to run through the full test
> set (but to be fair, I could run multiple VM instances in parallel to
> speed this up). Anyway, this can be quite useful to have available for
> manually testing some specific implementation details..

I was intrigued by how fast Debian installer went through and after
only a few minutes it booted with a working system, then installing
build-essentials, automake, autoconf, libtool and git.
The actual compile of wolfssl and wpa_supplicant then took several
hours...


> 
> As far as this issue with SAE is concerned, there is actually nothing
> wrong with the calculation results, i.e., all the values are correct and
> I was able to get SAE completed with the current snapshot.. However, it
> is really slow. To the point of taking close to a minute to complete
> authentication. I'd assume you are seeing timeouts from this or just
> giving up on waiting before the operation is completed.
> 
> While the current standard version of SAE is inconveniently slow on low
> end processors, it was not supposed to be this slow. It turned out that
> the real issue here is in not exactly ideal implementation of the
> wolfssl wrapper function crypto_bignum_rand(). This function is expected
> to return a random value between 0 and m-1. That is what the function
> did, but it did that by finding a random _prime_ double the maximum
> length of that range and then scaling to to the range which is
> significantly harder calculation (took around 40 times longer than
> needed in some of my tests) and completely unnecessary for this
> function. This commit fixes that issue:
> https://w1.fi/cgit/hostap/commit/?id=eb595b3e3ab531645a5bde71cf6385335b7a4b95

Thank you for finding the cause of the issue, fixing it and also for
the detailed explaination!
I can confirm that your commit does fix the issue on the QCA MIPS
devices which were previously affected, they now reliably connect
instantly now.


Cheers


Daniel

_______________________________________________
openwrt-devel mailing list
openwrt-devel at lists.openwrt.org
https://lists.openwrt.org/mailman/listinfo/openwrt-devel