On Wed, 2024-10-16 at 17:53 +0800, Herbert Xu wrote:
> Alright, so next I'm going to try to make TDMA entirely single-
> threaded and see if that fixes it.
>
> Thanks,
Hi, since this was marked as "not worth tracking*, here a summary of my understanding
- hardware: Turris Omnia, Marvell Armada 385 (same behaviour on 2 devices)
- the crypto self-tests on the hash algorithms provided by the Marvell CESA
driver fail randomly (1-5 failures in 90% of the boots, rarely without failure)
- this is likely a bug in the driver, which had been hidden for a long time
- it is now exposed by parallel invocation of self-tests, introduced in v6.12-rc1,
commit 37da5d0ffa ("crypto: api - Do not wait for tests during registration")
- to be safe, the algorithms in question (6 in total) have been set to priority 0 in
commit e845d2399a ("crypto: marvell/cesa - Disable hash algorithms")
- so, there should be no immediate harm by wrong hashes (at least in 6.12)
- there were backport patches floating around, but as of today they have not
been picked up
- no fix to the Marvell CESA driver has been found so far
Consequences
- I have decided to stop using the Marvell CESA driver on my production device
- I can easily run test patches on my spare Omnia, if that helps finding a fix
- I imagine more users will run into the failing self-tests, as soon as 6.12
hits OpenWRT's mvebu target
Best regards, Klaus
