Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Date:   Fri, 22 Apr 2022 07:58:25 +0000
From:   Oliver Upton <oupton@google.com>
To:     Paolo Bonzini <pbonzini@redhat.com>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org, will@kernel.org,
        maz@kernel.org, apatel@ventanamicro.com, atishp@rivosinc.com,
        seanjc@google.com, pgonda@google.com
Subject: Re: [PATCH 0/4] KVM: fix KVM_EXIT_SYSTEM_EVENT mess
Message-ID: <YmJgIQe+5zGbrxoF@google.com>
References: <20220421180443.1465634-1-pbonzini@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220421180443.1465634-1-pbonzini@redhat.com>
Precedence: bulk

Hi Paolo,

On Thu, Apr 21, 2022 at 02:04:39PM -0400, Paolo Bonzini wrote:
> The KVM_SYSTEM_EVENT_NDATA_VALID mechanism that was introduced
> contextually with KVM_SYSTEM_EVENT_SEV_TERM is not a good match
> for ARM and RISC-V, which want to communicate information even
> for existing KVM_SYSTEM_EVENT_* constants.  Userspace is not ready
> to filter out bit 31 of type, and fails to process the
> KVM_EXIT_SYSTEM_EVENT exit.
> 
> Therefore, tie the availability of ndata to a system capability;
> if the capability is present, ndata is always valid, so patch 1
> makes x86 always initialize it.  Then patches 2 and 3 fix
> ARM and RISC-V compilation and patch 4 enables the capability.
> 
> Only compiled on x86, waiting for acks.
> 
> Paolo
> 
> Paolo Bonzini (4):
>   KVM: x86: always initialize system_event.ndata
>   KVM: ARM: replace system_event.flags with ndata and data[0]
>   KVM: RISC-V: replace system_event.flags with ndata and data[0]
>   KVM: tell userspace that system_event.ndata is valid

Is there any way we could clean this up in 5.18 and leave the whole
ndata/data pattern for 5.19?

IOW, for 5.18 go back and fix the padding:

	struct {
		__u32 type;
		__u32 pad;
		__u64 flags;
	} system_event;

Then for 5.19 circle back on the data business, except use a flag bit
for it:

	struct {
		__u32 type;
		__u32 pad;
	#define KVM_SYSTEM_EVENT_NDATA_VALID	(1u << 63)
		__u64 flags;
		__u64 ndata;
		__u64 data[16];
	} system_event;

Where we apply that bit to system_event::flags this time instead of
::type. Could also go the CAP route.

Wouldn't this be enough to preserve ABI with whatever userspace has been
spun up for system_event::flags already and also keep the SEV stuff
happy in 5.19?

It is a bit weird to churn existing UAPI usage in the very next kernel
cycle, but could be convinced otherwise :)

--
Thanks,
Oliver