On 23/08/2021 16:48, Bertrand Marquis wrote:
Hi Julien,
Hi Bertrand,
[Keep only arm maintainers in the CC list]
On 23 Aug 2021, at 13:10, Julien Grall <[email protected]> wrote:
Hi Bertrand,
On 23/08/2021 11:32, Bertrand Marquis wrote:
On arm architecture we might have heterogeneous platforms with different
types of cores. As a guest can potentialy run on any of those cores we
have to present them cpu features which are compatible with all cores
and discard the features which are only available on some cores.
Sanitizing the CPU info is important for Xen so it can correctly size the P2M,
flush the cache... However, I don't think this is going to be sufficient to be
able to move a vCPU between different type of pCPU.
The main hurdle I can see so far is errata handling. Not all the errata can be
fully handled in Xen so some of them are left to the guest to mitigate.
I agree this is something to work on and a problem with the current serie.
The errata are usually detected based on the MIDR while the OS is booting. IOW,
a guest will not be able to know that it needs to handle an errata for pCPU B
if it only runs on pCPU A.
Ack.
I don't know yet how this can be solved, but errata are not that uncommon on
Arm. So until this addressed, we will still need to make sure that vCPUs are
not migrated between pCPUs with at least a different MIDR.
This prevention can be either done manually by pinning the vCPUs or
implementing the proposal that Dario sent a few years ago (see [1]).
My current proposal would be the following:
- add a command line option to allow to use all cores on a heterogeneous
platform (different MIDR)
We already have a command line for heterogenous platform. How about
re-using it?
- taint Xen on this case
- keep the feature sanitize as it is as on this case it will create a safer
setup (apart from the errata potential problem)
- keep current behaviour if command line option is not passed
Having a solution to enable all cores (even if it is unsafe) could still be a
good improvement for development on big.LITTLE
platforms or for people knowing how to properly configure the system to prevent
the problems by using properly cpupools so
I still think this serie with the proposed changes is still making a lot of
sense.
+1.
I will start looking at a long term solution, maybe automatically create a
cpupools on boot or investigate on the design you provided.
IIRC, all vCPUs of a domain needs to be in the same cpupool. If your
plan is to expose big.LITTLE to the guest, then you would need to look
at the design proposal from Dario.
Cheers,
--
Julien Grall
