diff options
Diffstat (limited to 'gitlab/issues_text/target_missing/host_missing/accel_missing/1595')
| -rw-r--r-- | gitlab/issues_text/target_missing/host_missing/accel_missing/1595 | 29 |
1 files changed, 29 insertions, 0 deletions
diff --git a/gitlab/issues_text/target_missing/host_missing/accel_missing/1595 b/gitlab/issues_text/target_missing/host_missing/accel_missing/1595 new file mode 100644 index 000000000..24b2547bb --- /dev/null +++ b/gitlab/issues_text/target_missing/host_missing/accel_missing/1595 @@ -0,0 +1,29 @@ +CPU boot sometimes fails on big.LITTLE CPUs with varying cache sizes +Description of problem: +The RK3588 SoC has three core clusters; one with A55 cores, and the other two have A76 cores. The big cores have more L2 cache than the little cores, so the value of `CCSIDR` depends on the core that it is read from. + +In `write_list_to_kvmstate`, QEMU attempts to use `KVM_SET_ONE_REG` with an ID for `KVM_REG_ARM_DEMUX_ID_CCSIDR`, trying to set `CCSIDR` to a previously read value. + +Normally, that works fine, but if the host kernel has moved QEMU from one core cluster to the other, then the value will be different and `demux_c15_set` will return `EINVAL`, causing the entire `arm_set_cpu_on` to fail, and the guest kernel to print an error. + +https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/kvm/sys_regs.c?h=v6.2#n2827 + +I tried changing the condition for the `ok = false` line in `write_list_to_kvmstate` to `ret && r.id >> 8 != 0x60200000001100`. This causes all CPUs to initialize correctly in the guest, but obviously that's a hack. + +I assume that `CCSIDR` not being uniform across all CPUs means that the guest's copy of `CCSIDR` may be wrong, and so cache maintenance operations may not act on the entire cache. I do not know whether that could actually cause problems. Will QEMU need to find the maximum cache size across all CPUs and present that to guests? +Steps to reproduce: +On a SoC where big and little cores have different cache sizes (e.g. RK3588): + +```text +$ qemu-system-aarch64 -M virt -accel kvm -cpu host -smp 4 -nographic -kernel arch/arm64/boot/Image -append quiet +[ 0.001399][ T1] psci: failed to boot CPU1 (-22) +[ 0.001407][ T1] CPU1: failed to boot: -22 +[ 0.001685][ T1] psci: failed to boot CPU2 (-22) +[ 0.001691][ T1] CPU2: failed to boot: -22 +[ 0.001809][ T1] psci: failed to boot CPU3 (-22) +[ 0.001814][ T1] CPU3: failed to boot: -22 +``` + +The error is not always printed, because it depends on which core cluster the processes are scheduled on. + +Using `taskset -c 0-3` or `taskset -c 4-7` to force QEMU to stick to the little or big cores respectively makes the bug not reproduce. |