Progress:
* UM-2 [QEMU upstream maintainership]
- Joint Linaro/Arm meeting on the Realm Management Extension (RME).
Mostly this was "how the software stack is going to work"; some
useful discussion on how this intersects with QEMU, both as a
potential emulation platform for software development and also
as the VMM for a realm-using KVM virtual machine.
- Kicked off a discussion thread on getting rid of the last parts
of our "ad-hoc CI" this release cycle. The main thing that's
not yet handled by Gitlab CI is FreeBSD and NetBSD hosts.
We might be able to do this if we can find an x86 Linux machine
to act as a gitlab runner (it needs to be able to run guest
BSD VMs).
* QEMU-422 [QEMU Arm Neoverse V1 vCPU for TCG]
- Several small features have now made it upstream so those
subtasks have been closed out
- Looking at FEAT_RASv1p1 and FEAT_DoubleFault. The first of these
is easy because we only model the minimal "no error records"
RAS: I think we can simply bump the ID register fields to
indicate support. The second is a little bit more interesting
because the new SCR_EL3.EASE bit means we suddenly have a
source of physical SErrors (from synchronous external aborts)
when we previously could not have any...
-- PMM
Hello,
# [GNU-767] Support changing SVE vector length in remote debugging
* Continued working on the testsuite regressions introduced by my code.
- Found out that the crash I was investigating (which happens when
gdbserver thinks that the kernel doesn't support SVE and crashes)
only occurs with kernel v5.4 from Ubuntu 20.04 and not with
Ubuntu 22.04's v5.15. In the former version,
ptrace(PTRACE_GETREGSET, …, NT_ARM_SVE, …) often returns -1 (but not
always) and gdbserver interprets this as meaning that SVE isn't
supported. The newer kernel always returns successfully from this
call. gdbserver still acts weird so there's more digging to do, but
at least it doesn't crash.
# Misc
- Took one day of vacation.
--
Thiago
Progress:
* UM-2 [QEMU upstream maintainership]
+ had a look at adding an option to allow semihosting from
userspace (handy for some test case purposes); have a working
prototype which I used to test the FEAT_IDST patch, but
probably won't pursue further until some refactoring of the
whole semihosting code has landed, to avoid conflicts
+ reading through the Realm Management Extension documentation:
there's a Linaro Kernel Working Group sprint in Cambridge
next week, and possible QEMU RME support is on the agenda
+ back before QEMU 6.0 we tried to fix a bug where QEMU always
provided 4 PMU counters, not the number the emulated CPU really
has. We had to revert that change before the 6.0 release, but
then forgot to go back and revisit it afterwards. Alex
reminded me of it this week, so I rebased it, fixed the problem
that meant it got reverted, and sent it out for review.
* QEMU-422 [QEMU Arm Neoverse V1 vCPU for TCG]
+ sent patch implementing FEAT_IDST
+ next up: FEAT_DoubleFault (likely a no-op for us, as we don't
ever have physical SErrors)
-- PMM
Hello,
# [GNU-767] Support changing SVE vector length in remote debugging
* Continued working on the testsuite regressions introduced by my code.
- Currently investigating a crash that happens when for some reason
gdbserver thinks that the kernel doesn't support SVE and crashes
while building the internal representation of the regular vector
registers.
# Misc
- Attended Google's Fuchsia boot camp.
--
Thiago
Progress:
* UM-2 [QEMU upstream maintainership]
+ tracked down the UEFI crash with KVM on AArch64 to a combination of
(a) host system was heterogenous and user hadn't restricted QEMU to
only running on one set of cores and (b) QEMU silently throws away
the error it gets from KVM in this situation, resulting in most
things seeming to work except that vcpu power on/off state isn't
written to KVM and so the guest starts with all vcpus running
instead of just the primary. We should really improve the error
handling, but I worry that that might break previously functional
setups...
+ tidied up some patches I wrote while doing the GICv4 work that
fix an odd inconsistency in our GIC emulation where we correctly
implement the right number of virtual priority bits for the CPU
but always provide 8 bits of physical priority regardless of what
the real CPU implementation should have, and sent them out for review
+ usual code review and pullreq management
* QEMU-422 [QEMU Arm Neoverse V1 vCPU for TCG]
+ Implemented support for FEAT_S2FWB and sent patches to the list
+ Next up is probably FEAT_IDST (ID space trap handling)
-- PMM
Hello,
# [GNU-767] Support changing SVE vector length in remote debugging
* Continued working on the testsuite regressions introduced by my code.
- Realised that gdbserver uses “thread->tdesc != NULL” as a proxy of
whether it has already attached to the inferior process. My change to
track the target description per-thread instead of per-process
invalidated this equivalence. Fixed a few places to use “proc->attached”
instead, which is the direct way of testing that condition.
- Now looking into an issue with gdbserver loading libthread_db.so.
# Misc
- Reviewed Tom Tromey's “[PATCH 00/36] C++-ify breakpoints” patch series¹.
Spotted one bug.
--
Thiago
¹ https://sourceware.org/pipermail/gdb-patches/2022-January/185256.html
msgid:20220118194007.2853108-1-tom@tromey.com
Progress:
* UM-2 [QEMU upstream maintainership]
+ investigating a bug report about the guest UEFI crashing on
startup when using KVM on AArch64 with an SMP guest (sadly I haven't
been able to repro it myself thus far)
* QEMU-422 [QEMU Arm Neoverse V1 vCPU for TCG]
+ audited code to check we can enable FEAT_TTL in QEMU (ie we ignore the
new TTL hint field in TLB invalidate operations)
+ audited code to check we can enable FEAT_BBM level 2 in QEMU
(our TLB implementation never allows multiple clashing TLB entries)
+ audited code to check we can enable SMMUv3.2-BBML2 in our SMMUv3
(our SMMU TLB does allow multiple entries but will always select
the one for the lowest level and ignore the others)
+ sent patches to advertise FEAT_TTL, FEAT_BBM, SMMUv3.2-BBML2 in
CPU and SMMU ID registers
-- PMM