Skip to content

Commit 36a31a7

Browse files
committed
NVIDIA: add fresh H20 GPU and TDX binding validation
Signed-off-by: Jiale Zhang <zhangjiale@linux.alibaba.com>
1 parent 76c207f commit 36a31a7

2 files changed

Lines changed: 172 additions & 3 deletions

File tree

‎deps/verifier/src/nvidia.rs‎

Lines changed: 115 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -543,4 +543,119 @@ mod tests {
543543
.unwrap_err();
544544
assert!(error.to_string().contains("exit 12"), "{error}");
545545
}
546+
547+
// Files stay outside the repository: fresh reports identify the test GPU.
548+
#[cfg(not(target_arch = "wasm32"))]
549+
fn hardware_evidence() -> (Value, Vec<u8>) {
550+
let evidence = std::fs::read(
551+
std::env::var("TRUSTEE_NVIDIA_HARDWARE_EVIDENCE")
552+
.expect("Set TRUSTEE_NVIDIA_HARDWARE_EVIDENCE to collected evidence JSON"),
553+
)
554+
.unwrap();
555+
let report_data =
556+
std::fs::read(std::env::var("TRUSTEE_NVIDIA_HARDWARE_REPORT_DATA").expect(
557+
"Set TRUSTEE_NVIDIA_HARDWARE_REPORT_DATA to the original 64-byte challenge",
558+
))
559+
.unwrap();
560+
assert_eq!(report_data.len(), 64);
561+
(serde_json::from_slice(&evidence).unwrap(), report_data)
562+
}
563+
564+
#[cfg(not(target_arch = "wasm32"))]
565+
#[tokio::test]
566+
#[ignore = "Requires fresh GPU evidence and live NVIDIA services; see nvidia/README.md"]
567+
async fn live_hardware_gpu_evidence() {
568+
let (evidence, report_data) = hardware_evidence();
569+
let mut config = crate::nvidia_config::NvidiaVerifierConfig::default()
570+
.resolve()
571+
.unwrap();
572+
let verifier = match config.mode {
573+
crate::nvidia_config::NvidiaVerificationMode::Rim => {
574+
config.nras_url = "https://127.0.0.1:1/unused".into();
575+
config.nras_jwks_url = "https://127.0.0.1:1/unused-jwks".into();
576+
NvidiaVerifier::Rim(config)
577+
}
578+
crate::nvidia_config::NvidiaVerificationMode::Nras => {
579+
NvidiaVerifier::Nras(NrasVerifier::new(&config).await.unwrap())
580+
}
581+
};
582+
let nonce = crate::tdx::gpu::gpu_nonce(&report_data);
583+
let gpus = evidence["gpu_evidence"]["evidence_list"]
584+
.as_array()
585+
.unwrap();
586+
assert!(!gpus.is_empty());
587+
for gpu in gpus {
588+
let arch = gpu["arch"].as_str().unwrap();
589+
let report = gpu["attestation_report"].as_str().unwrap();
590+
let certificate = gpu["certificate"].as_str().unwrap();
591+
let claims = verifier
592+
.verify(arch, report, certificate, &nonce)
593+
.await
594+
.unwrap();
595+
assert_eq!(claims["x-nvidia-gpu-driver-version"], gpu["driver_version"]);
596+
let mut wrong_nonce = nonce;
597+
wrong_nonce[0] ^= 1;
598+
assert!(verifier
599+
.verify(arch, report, certificate, &wrong_nonce)
600+
.await
601+
.is_err());
602+
let mut tampered = STANDARD.decode(report).unwrap();
603+
*tampered.last_mut().unwrap() ^= 1;
604+
assert!(verifier
605+
.verify(arch, &STANDARD.encode(tampered), certificate, &nonce)
606+
.await
607+
.is_err());
608+
let wrong_certificate = STANDARD.encode(include_bytes!(
609+
"../test_data/nvidia/blackwell_coco_certs.txt"
610+
));
611+
assert!(verifier
612+
.verify(arch, report, &wrong_certificate, &nonce)
613+
.await
614+
.is_err());
615+
println!("GPU hardware verification passed: mode={}, arch={}, driver={}; wrong nonce, signature and certificate rejected",
616+
claims["verifier"], arch, claims["x-nvidia-gpu-driver-version"]);
617+
}
618+
}
619+
620+
#[cfg(not(target_arch = "wasm32"))]
621+
#[tokio::test]
622+
#[ignore = "Requires fresh combined TDX/GPU evidence, PCCS and NVIDIA services"]
623+
async fn live_hardware_tdx_gpu_binding() {
624+
use crate::{InitDataHash, ReportData, Verifier};
625+
let (evidence, report_data) = hardware_evidence();
626+
let verifier = crate::tdx::Tdx::default();
627+
let expected_report_data = ReportData::Value(&report_data);
628+
let init_data_hash = InitDataHash::NotProvided;
629+
let verify = |evidence| verifier.evaluate(evidence, &expected_report_data, &init_data_hash);
630+
let (claims, _) = verify(evidence.clone()).await.unwrap();
631+
let count = evidence["gpu_evidence"]["evidence_list"]
632+
.as_array()
633+
.unwrap()
634+
.len();
635+
assert!(count > 0);
636+
assert_eq!(claims["nvidia_gpu_count"], count);
637+
let mut wrong_data = report_data.clone();
638+
wrong_data[0] ^= 1;
639+
assert!(verifier
640+
.evaluate(
641+
evidence.clone(),
642+
&ReportData::Value(&wrong_data),
643+
&InitDataHash::NotProvided
644+
)
645+
.await
646+
.is_err());
647+
let mut removed = evidence.clone();
648+
removed.as_object_mut().unwrap().remove("gpu_evidence");
649+
assert!(verify(removed).await.is_err());
650+
let mut changed = evidence.clone();
651+
let original = changed["gpu_evidence"]["evidence_list"][0]["attestation_report"]
652+
.as_str()
653+
.unwrap();
654+
let mut bytes = STANDARD.decode(original).unwrap();
655+
*bytes.last_mut().unwrap() ^= 1;
656+
changed["gpu_evidence"]["evidence_list"][0]["attestation_report"] =
657+
json!(STANDARD.encode(bytes));
658+
assert!(verify(changed).await.is_err());
659+
println!("TDX + {count} GPU(s) verification passed; changed challenge, removed GPU evidence and changed GPU report rejected");
660+
}
546661
}

‎deps/verifier/src/nvidia/README.md‎

Lines changed: 57 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -147,6 +147,60 @@ cargo +1.88.0 test -p verifier --no-default-features --features tdx-dcap-rust \
147147
On 2026-11-17, Blackwell report verification passed in both modes; wrong nonce,
148148
changed signature and a substituted certificate were rejected. The old Hopper
149149
report was rejected because its Driver RIM signing certificate expired on
150-
2026-06-01. This validates hardware-report replay and live verifier services;
151-
fresh device collection and a combined TDX+GPU hardware run remain separate
152-
validation steps.
150+
2026-06-01. These fixed-challenge tests validate hardware-report replay and
151+
live verifier services; fresh collection is covered separately below.
152+
153+
For a fresh hardware run, build the paired guest-components `evidence_getter`
154+
with `tdx-attester,bin` (Rust) or `-tags gpu` (Go). On the TDX/GPU guest,
155+
generate a random 64-byte challenge and collect evidence:
156+
157+
```sh
158+
head -c 64 /dev/urandom > report-data.bin
159+
./evidence_getter file report-data.bin > evidence.json
160+
```
161+
162+
Transfer these two files to the verifier over an authenticated channel; keep
163+
device-specific evidence outside the source tree. Then run each mode explicitly:
164+
165+
```sh
166+
export TRUSTEE_NVIDIA_HARDWARE_EVIDENCE=/absolute/path/evidence.json
167+
export TRUSTEE_NVIDIA_HARDWARE_REPORT_DATA=/absolute/path/report-data.bin
168+
for mode in rim nras; do
169+
TRUSTEE_NVIDIA_VERIFIER_MODE=$mode cargo +1.88.0 test -p verifier \
170+
--no-default-features --features tdx-dcap-rust --lib \
171+
nvidia::tests::live_hardware -- --ignored --nocapture --test-threads=1
172+
done
173+
```
174+
175+
The GPU test verifies the supplied report and rejects a wrong nonce, changed
176+
signature and substituted certificate. RIM mode uses unreachable NRAS URLs.
177+
The combined test additionally checks the TDX signature, event log, CPU/GPU
178+
binding, and rejection of a changed challenge or removed GPU evidence. It
179+
requires a working quote-generation service and PCCS connectivity.
180+
181+
## H20 hardware validation (2026-11-17)
182+
183+
Fresh evidence was collected on an Alibaba Cloud `ecs.gn8v-tee.4xlarge` in
184+
`cn-beijing-l`: one NVIDIA H20, GPU CC ON, driver `595.71.05`, VBIOS
185+
`96.00.CF.00.05`, and ALinux kernel `5.10.134-19.8.al8.x86_64`. The guest used
186+
an 8 GiB SWIOTLB and NVIDIA persistence daemon as described in the
187+
[provider instructions](https://help.aliyun.com/en/ecs/user-guide/build-a-heterogeneous-confidential-computing-environment).
188+
189+
| Collector | Local RIM/OCSP | NRAS | Full TDX/GPU binding |
190+
| --- | --- | --- | --- |
191+
| Rust, paired guest-components PR | Passed | Passed | Passed in both modes |
192+
| Go, built with `-tags gpu` | Passed | Passed | Passed in both modes |
193+
194+
Each collector returned a 5006-byte TDX quote, a 4129-byte GPU report and a
195+
4826-byte attestation certificate chain for a random 64-byte challenge. The
196+
GPU tests rejected a wrong nonce, modified signature and substituted chain.
197+
The combined verifier rejected a changed challenge, removed GPU evidence and
198+
modified GPU report. Local tests used unreachable NRAS URLs.
199+
200+
Hardware testing also reproduced NVML returning zero devices after GPU
201+
initialization failed. The paired collectors now reject this PCI/NVML
202+
inventory mismatch. Rust additionally selects the GPU attestation chain
203+
instead of the empty CEC chain and slices both NVML buffers by their reported
204+
lengths. Both fixes were validated on the same H20 before releasing the
205+
temporary instance. This run covers fresh collection and combined evidence
206+
verification; it does not exercise KBS secret delivery.

0 commit comments

Comments
 (0)