On 2/12/2025 8:18 AM, Sean Christopherson wrote:
On Tue, Feb 11, 2025, Binbin Wu wrote:
+static int tdx_report_fatal_error(struct kvm_vcpu *vcpu)
+{
+ struct vcpu_tdx *tdx = to_tdx(vcpu);
+ u64 reg_mask = tdx->vp_enter_args.rcx;
+ u64 *opt_regs;
+
+ /*
+ * Skip sanity checks and let userspace decide what to do if sanity
+ * checks fail.
+ */
+ vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
+ vcpu->run->system_event.type = KVM_SYSTEM_EVENT_TDX_FATAL;
+ /* Error codes. */
+ vcpu->run->system_event.data[0] = tdx->vp_enter_args.r12;
+ /* GPA of additional information page. */
+ vcpu->run->system_event.data[1] = tdx->vp_enter_args.r13;
+ /* Information passed via registers (up to 64 bytes). */
+ opt_regs = &vcpu->run->system_event.data[2];
+
+#define COPY_REG(REG, MASK) \
+ do { \
+ if (reg_mask & MASK) { \
Based on past experience with conditionally filling kvm_run fields, I think KVM
should copy all registers and let userspace sort out the reg_mask. Unless the
guest passes an ASCII byte stream exactly as the GHCI suggests,
Yea, GHCI doesn't enforce it to be ASCII byte stream.
the information
is quite useless because userspace doesn't have reg_mask and so can't know what's
in data[4], data[5], etc... And I won't be the least bit surprised if guests
deviate from the GHCI.
But it also confuses userspace if guests uses special protocol to pass
information other than ASCII byte stream.
Anyway, dumping all registers to userspace and let userspace to have all
the information passed from guest for parsing is definitely workable.
+ *opt_regs = tdx->vp_enter_args.REG; \
+ opt_regs++; \
+ } \
+ } while (0)
+
+ /* The order is defined in GHCI. */
Assuming I haven't missed something, to hell with the GCHI, just dump *all*
registers, sorted by their index (ascending). Including RAX (TDCALL), RBP, and
RSP.
