Re: [PATCH v3 1/2] kvm: sev: Add SEV-SNP guest request throttling

[Sean Christopherson]

AMD folks, unless y'all object to the concept itself, can this be tacked onto the
SNP series?  Responsibility for responding to feedback and making changes can still
be punted to Dionna (or whoever), I just get briefly confused every time this series
is posted because I think it's addressing a problem that needs attention _now_.

On Thu, Jan 19, 2023, Dionna Glaze wrote:
> The AMD-SP is a precious resource that doesn't have a scheduler other
> than a mutex lock queue. To avoid customers from causing a DoS, a
> module_param-set rate limit is added with a default of 2 requests
> per 2 seconds.
> 
> These defaults were chosen empirically with a the assumption that
> current server-grade SEV-SNP machines will rarely exceed 128 VMs under
> usual circumstance.
> 
> The throttling code is 2 << 32 given that invalid length is 1 and 2 is
> the next available code. This was suggested by Tom Lendacky, and will
> be included in a new revision of the GHCB specification.

Why does throttling just punt back to the guest?  E.g. why not exit to userspace
and let userspace stall the vCPU?  Is the guest expected to schedule out the task
that's trying to make the request?

> @@ -158,6 +158,7 @@ struct snp_psc_desc {
>  
>  /* Guest message request error code */
>  #define SNP_GUEST_REQ_INVALID_LEN	BIT_ULL(32)
> +#define SNP_GUEST_REQ_THROTTLED		(((u64)2) << 32)

Someone please add macros to define the shift and generate error codes, the above
is way too hard to read for such a simple concept.  E.g. I want to see something
like

 #define SNP_GUEST_REQ_INVALID_LEN	SNP_GUEST_REQ_ERROR_CODE(1)

>  #define GHCB_MSR_TERM_REQ		0x100
>  #define GHCB_MSR_TERM_REASON_SET_POS	12
> diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
> index d0e58cffd1ed..cd9372ce6fc2 100644
> --- a/arch/x86/kvm/svm/sev.c
> +++ b/arch/x86/kvm/svm/sev.c
> @@ -58,6 +58,14 @@ module_param_named(sev_es, sev_es_enabled, bool, 0444);
>  /* enable/disable SEV-SNP support */
>  static bool sev_snp_enabled = true;
>  module_param_named(sev_snp, sev_snp_enabled, bool, 0444);
> +
> +/* Throttle guest requests to a burst # per this many seconds */
> +unsigned int guest_request_throttle_s = 2;

"seconds" seems like to coarse of a granularity, e.g. if userspace wants to allow
one request every half-second.  Why not go with milliseconds?  As a bonus, the (IMO)
odd "s" gets replaced with the more intuitive "ms".

That said, I wonder if this should be a per-VM capability, not a module param.
Since the throttling is per VM and not per user, making it a module param doesn't
prevent a malicious user or even a compromised VMM from spamming the PSP, e.g. just
spin up a big pile o' VMs.

> +module_param(guest_request_throttle_s, int, 0444);
> +
> +/* Throttle guest requests to this many per the above many seconds */
> +unsigned int guest_request_throttle_burst = 2;
> +module_param(guest_request_throttle_burst, int, 0444);
>  #else
>  #define sev_enabled false
>  #define sev_es_enabled false
> @@ -333,6 +341,9 @@ static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
>  			goto e_free;
>  
>  		mutex_init(&sev->guest_req_lock);
> +		ratelimit_state_init(&sev->snp_guest_msg_rs,
> +				guest_request_throttle_s * HZ,
> +				guest_request_throttle_burst);
>  		ret = sev_snp_init(&argp->error, false);
>  	} else {
>  		ret = sev_platform_init(&argp->error);
> @@ -3595,6 +3606,14 @@ static void snp_cleanup_guest_buf(struct sev_data_snp_guest_request *data, unsig
>  		*rc = SEV_RET_INVALID_ADDRESS;
>  }
>  
> +static bool snp_throttle_guest_request(struct kvm_sev_info *sev) {

Curly brace goes on its own line for functions.

> +	if (__ratelimit(&sev->snp_guest_msg_rs))
> +		return false;
> +
> +	pr_info_ratelimited("svm: too many guest message requests\n");

Drop the printk, it doesn't help understand _which_ guest is being throttled.
If userspace really wants to get notified, then KVM should exit to userspace
instead of throttling manually.