On Mon, Jul 31, 2023 at 09:04:14AM -0300, Daniel Henrique Barboza wrote:
> Hi,
>
> This work was motivated by the changes made in QEMU in commit df817297d7
> ("target/riscv: update multi-letter extension KVM properties") where it
> was required to use EINVAL to try to assess whether a given extension is
> available in the host.
>
> It turns out that the RISC-V KVM module makes regular use of the EIVAL
> error code in all kvm_get_one_reg() and kvm_set_one_reg() callbacks,
> which is not ideal. For example, ENOENT is a better error code for the
> case where a given register does not exist. While at it I decided to
> take a look at other error codes from these callbacks, comparing them
> with how other archs (ARM) handles the same error types, and changed
> some of the EOPNOTSUPP instances to either ENOENT or EBUSY.
>
> I am aware that changing error codes can be problematic to existing
> userspaces. I took a look and no other major userspace (checked crosvm,
> rust-vmm, QEMU, kvmtool), aside from QEMU now checking for EIVAL (and we
> can't change that because of backwards compat for that particular case
> we're covering),
Merging this series just prior to some other API change, like adding
the get-reg-list ioctl[1], will allow QEMU to eventually drop the EINVAL
handling. This is because when QEMU sees the get-reg-list ioctl it will
know that get/set-reg-list has the updated API. At some point when QEMU
no longer wants to support a KVM which doesn't have get-reg-list, then
the EINVAL handling can be dropped.
And, once KVM has get-reg-list, then QEMU won't want to probe for
registers with get-one-reg anyway. Instead, it'll get the list once
and then check that each register it would have probed is in the list.
[1] https://lore.kernel.org/lkml/cover.1690273969.git.haibo1.xu@xxxxxxxxx/T/
> will be impacted by this kind of change since they're
> all checking for "return < 0 then ..." instead of doing specific error
> handling based on the error value. This means that we're still in good
> time to make this kind of change while we're still in the initial steps
> of the RISC-V KVM ecossystem support.
>
> Patch 3 happens to also change the behavior of the set_reg() for
> zicbom/zicboz block sizes. Instead of always erroring out we'll check if
> userspace is writing the same value that the host uses. In this case,
> allow the write to succeed (i.e. do nothing). This avoids the situation
> in which userspace reads cbom_block_size, find out that it's set to X,
> and then can't write the same value back to the register. It's a QOL
> improvement to allow userspace to be lazier when reading/writing regs.
Being able to write back the same value will greatly simplify
save/restore for QEMU, particularly when get-reg-list is available. QEMU
would then get the list of registers and, for save, it'll iterate the list
blindly, calling get-one-reg on each, storing each value. Then, for
restore, it'll blindly iterate again, writing each saved value back with
set-one-reg.
Thanks,
drew
> A
> similar change was also made in patch 4.
>
> Patches are based on top of riscv_kvm_queue.
>
> Daniel Henrique Barboza (6):
> RISC-V: KVM: return ENOENT in *_one_reg() when reg is unknown
> RISC-V: KVM: use ENOENT in *_one_reg() when extension is unavailable
> RISC-V: KVM: do not EOPNOTSUPP in set_one_reg() zicbo(m|z)
> RISC-V: KVM: do not EOPNOTSUPP in set KVM_REG_RISCV_TIMER_REG
> RISC-V: KVM: use EBUSY when !vcpu->arch.ran_atleast_once
> docs: kvm: riscv: document EBUSY in KVM_SET_ONE_REG
>
> Documentation/virt/kvm/api.rst | 2 ++
> arch/riscv/kvm/aia.c | 4 +--
> arch/riscv/kvm/vcpu_fp.c | 12 ++++----
> arch/riscv/kvm/vcpu_onereg.c | 52 ++++++++++++++++++++--------------
> arch/riscv/kvm/vcpu_sbi.c | 16 ++++++-----
> arch/riscv/kvm/vcpu_timer.c | 11 +++----
> 6 files changed, 55 insertions(+), 42 deletions(-)
>
> --
> 2.41.0
>
