Message ID | 20230119213426.379312-2-dionnaglaze@google.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | kvm: sev: Add SNP guest request throttling | expand |
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.
diff --git a/arch/x86/include/asm/sev-common.h b/arch/x86/include/asm/sev-common.h index 1b111cde8c82..e3a6b039480d 100644 --- a/arch/x86/include/asm/sev-common.h +++ b/arch/x86/include/asm/sev-common.h @@ -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) #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; +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) { + if (__ratelimit(&sev->snp_guest_msg_rs)) + return false; + + pr_info_ratelimited("svm: too many guest message requests\n"); + return true; +} + static void snp_handle_guest_request(struct vcpu_svm *svm, gpa_t req_gpa, gpa_t resp_gpa) { struct sev_data_snp_guest_request data = {0}; @@ -3611,6 +3630,11 @@ static void snp_handle_guest_request(struct vcpu_svm *svm, gpa_t req_gpa, gpa_t sev = &to_kvm_svm(kvm)->sev_info; + if (snp_throttle_guest_request(sev)) { + rc = SNP_GUEST_REQ_THROTTLED; + goto e_fail; + } + mutex_lock(&sev->guest_req_lock); rc = snp_setup_guest_buf(svm, &data, req_gpa, resp_gpa); @@ -3648,6 +3672,11 @@ static void snp_handle_ext_guest_request(struct vcpu_svm *svm, gpa_t req_gpa, gp sev = &to_kvm_svm(kvm)->sev_info; + if (snp_throttle_guest_request(sev)) { + rc = SNP_GUEST_REQ_THROTTLED; + goto e_fail; + } + data_gpa = vcpu->arch.regs[VCPU_REGS_RAX]; data_npages = vcpu->arch.regs[VCPU_REGS_RBX]; diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h index 8d1ba66860a4..7048f817efb0 100644 --- a/arch/x86/kvm/svm/svm.h +++ b/arch/x86/kvm/svm/svm.h @@ -18,6 +18,7 @@ #include <linux/kvm_types.h> #include <linux/kvm_host.h> #include <linux/bits.h> +#include <linux/ratelimit.h> #include <asm/svm.h> #include <asm/sev-common.h> @@ -105,6 +106,8 @@ struct kvm_sev_info { unsigned int snp_certs_len; /* Size of instance override for certs */ struct mutex guest_req_lock; + struct ratelimit_state snp_guest_msg_rs; /* Limit guest requests */ + u64 sev_features; /* Features set at VMSA creation */ }; diff --git a/include/uapi/linux/in.h b/include/uapi/linux/in.h index f243ce665f74..07a4cb149305 100644 --- a/include/uapi/linux/in.h +++ b/include/uapi/linux/in.h @@ -20,6 +20,7 @@ #define _UAPI_LINUX_IN_H #include <linux/types.h> +#include <linux/stddef.h> #include <linux/libc-compat.h> #include <linux/socket.h>
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 2 burst per 2 seconds means on average 1 request every second. We allow 2 requests back to back to allow for the guest to query the certificate length in an extended guest request without a pause. The 1 second average is our target for quality of service since empirical tests show that 64 VMs can concurrently request an attestation report with a maximum latency of 1 second. We don't anticipate more concurrency than that for a seldom used request for a majority well- behaved set of VMs. The majority point is decided as >64 VMs given the assumed 128 VM count for "extreme load". 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. Cc: Thomas Lendacky <Thomas.Lendacky@amd.com> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Joerg Roedel <jroedel@suse.de> Cc: Peter Gonda <pgonda@google.com> Cc: Borislav Petkov <bp@alien8.de> Signed-off-by: Dionna Glaze <dionnaglaze@google.com> --- arch/x86/include/asm/sev-common.h | 1 + arch/x86/kvm/svm/sev.c | 29 +++++++++++++++++++++++++++++ arch/x86/kvm/svm/svm.h | 3 +++ include/uapi/linux/in.h | 1 + 4 files changed, 34 insertions(+)