lib: log vCPU diagnostics on triple fault and for some unhandled exit types

bin/propolis-server/src/lib/migrate/source.rs

@@ -4,7 +4,7 @@
 
 use bitvec::prelude::{BitSlice, Lsb0};
 use futures::{SinkExt, StreamExt};
-use propolis::common::{GuestAddr, PAGE_SIZE};
+use propolis::common::{GuestAddr, GuestData, PAGE_SIZE};
 use propolis::migrate::{
     MigrateCtx, MigrateStateError, Migrator, PayloadOutputs,
 };
@@ -657,9 +657,12 @@ impl<'vm, T: MigrateConn> RonV0Runner<'vm, T> {
         info!(self.log(), "ram_push: xfer RAM between {start:#x} and {end:#x}",);
         self.send_msg(memx::make_mem_xfer(start, end, bits)).await?;
         for addr in PageIter::new(start, end, bits) {
-            let mut bytes = [0u8; PAGE_SIZE];
-            self.read_guest_mem(GuestAddr(addr), &mut bytes).await?;
-            self.send_msg(codec::Message::Page(bytes.into())).await?;
+            let mut byte_buffer = [0u8; PAGE_SIZE];
+            {
+                let mut bytes = GuestData::from(byte_buffer.as_mut_slice());
+                self.read_guest_mem(GuestAddr(addr), &mut bytes).await?;
+            }
+            self.send_msg(codec::Message::Page(byte_buffer.into())).await?;
             probes::migrate_xfer_ram_page!(|| (addr, PAGE_SIZE as u64));
         }
         Ok(())
@@ -919,7 +922,7 @@ impl<'vm, T: MigrateConn> RonV0Runner<'vm, T> {
     async fn read_guest_mem(
         &mut self,
         addr: GuestAddr,
-        buf: &mut [u8],
+        buf: &mut GuestData<&mut [u8]>,
     ) -> Result<(), MigrateError> {
         let objects = self.vm.lock_shared().await;
         let memctx = objects.access_mem().unwrap();

bin/propolis-server/src/lib/vcpu_tasks.rs

@@ -159,6 +159,7 @@ impl VcpuTasks {
                         VmEntry::Run
                     }
                     VmExitKind::Suspended(SuspendDetail { kind, when }) => {
+                        use propolis::vcpu::Diagnostics;
                         match kind {
                             exits::Suspend::Halt => {
                                 event_handler.suspend_halt_event(when);
@@ -167,6 +168,13 @@ impl VcpuTasks {
                                 event_handler.suspend_reset_event(when);
                             }
                             exits::Suspend::TripleFault(vcpuid) => {
+                                slog::info!(
+                                    &log,
+                                    "triple fault on vcpu {}",
+                                    vcpu.id;
+                                    "state" => %Diagnostics::capture(vcpu)
+                                );
+
                                 if vcpuid == -1 || vcpuid == vcpu.id {
                                     event_handler
                                         .suspend_triple_fault_event(vcpu.id, when);
@@ -187,7 +195,76 @@ impl VcpuTasks {
                         task.force_hold();
                         VmEntry::Run
                     }
-                    _ => {
+                    VmExitKind::InstEmul(inst) => {
+                        let diag = propolis::vcpu::Diagnostics::capture(vcpu);
+                        error!(log,
+                               "instruction emulation exit on vCPU {}",
+                               vcpu.id;
+                               "context" => ?inst,
+                               "vcpu_state" => %diag);
+
+                        event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
+                        VmEntry::Run
+                    }
+                    VmExitKind::Unknown(code) => {
+                        error!(log,
+                               "unrecognized exit code on vCPU {}",
+                               vcpu.id;
+                               "code" => code);
+
+                        event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
+                        VmEntry::Run
+                    }
+                    // Bhyve emits the `Bogus` exit kind when there is no actual
+                    // guest exit for user space to handle, but circumstances
+                    // nevertheless dictate that the kernel VMM should exit to
+                    // user space (e.g. a caller requested that all vCPUs be
+                    // forced to exit to user space so their threads can
+                    // rendezvous there).
+                    //
+                    // `process_vmexit` should always successfully handle this
+                    // exit, since it never entails any work that could fail to
+                    // be completed.
+                    VmExitKind::Bogus => {
+                        unreachable!(
+                            "propolis-lib always handles VmExitKind::Bogus"
+                        );
+                    }
+                    VmExitKind::Debug => {
+                        error!(log,
+                               "lib returned debug exit from vCPU {}",
+                               vcpu.id);
+
+                        event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
+                        VmEntry::Run
+                    }
+                    VmExitKind::VmxError(detail) => {
+                        error!(log,
+                               "unclassified VMX exit on vCPU {}",
+                               vcpu.id;
+                               "detail" => ?detail);
+
+                        event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
+                        VmEntry::Run
+                    }
+                    VmExitKind::SvmError(detail) => {
+                        error!(log,
+                               "unclassified SVM exit on vCPU {}",
+                               vcpu.id;
+                               "detail" => ?detail);
+
+                        event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
+                        VmEntry::Run
+                    }
+                    VmExitKind::Paging(gpa, fault_type) => {
+                        let diag = propolis::vcpu::Diagnostics::capture(vcpu);
+                        error!(log,
+                               "unhandled paging exit on vCPU {}",
+                               vcpu.id;
+                               "gpa" => gpa,
+                               "fault_type" => fault_type,
+                               "vcpu_state" => %diag);
+
                         event_handler.unhandled_vm_exit(vcpu.id, exit.kind);
                         VmEntry::Run
                     }

bin/propolis-server/src/main.rs

@@ -136,6 +136,13 @@ fn run_server(
         Err(e).context("API version checks")?;
     }
 
+    // If this is a development image being run outside of an Omicron zone,
+    // enable the display (in logs, panic messages, and the like) of diagnostic
+    // data that may have originated in the guest.
+    #[cfg(not(feature = "omicron-build"))]
+    propolis::common::DISPLAY_GUEST_DATA
+        .store(true, std::sync::atomic::Ordering::SeqCst);
+
     let use_reservoir = config::reservoir_decide(&log);
 
     let context = server::DropshotEndpointContext::new(

bin/propolis-standalone/src/main.rs

@@ -1421,6 +1421,9 @@ fn main() -> anyhow::Result<ExitCode> {
     // Check that vmm and viona device version match what we expect
     api_version_checks(&log).context("API version checks")?;
 
+    propolis::common::DISPLAY_GUEST_DATA
+        .store(true, std::sync::atomic::Ordering::SeqCst);
+
     // Load/parse the config first, since it's required to size the tokio runtime
     // used to run the instance.
     let config = if restore {

lib/propolis/src/common.rs

@@ -5,9 +5,88 @@
 use std::ops::{Add, BitAnd};
 use std::ops::{Bound::*, RangeBounds};
 use std::slice::SliceIndex;
+use std::sync::atomic::{AtomicBool, Ordering};
 
 use crate::vmm::SubMapping;
 
+/// Controls whether items wrapped in a [`GuestData`] are displayed or redacted
+/// when the wrappers are printed via their `Display` or `Debug` impls.
+//
+// The Propolis server binary should only link the Propolis lib once (any
+// structure that links the lib multiple times means something is very odd about
+// its dependency graph), so there should never be any ambiguity about what
+// `DISPLAY_GUEST_DATA` refers to when linking. But to be maximally cautious,
+// label this static as `no_mangle` so that pulling in multiple Propolis
+// libraries will break the build instead of possibly resolving ambiguously.
+#[no_mangle]
+pub static DISPLAY_GUEST_DATA: AtomicBool = AtomicBool::new(false);
+
+/// A wrapper type denoting that the contained `T` was obtained from the guest
+/// (e.g. by reading the guest's memory). This type implements various traits
+/// (`Deref`, `DerefMut`, and `Borrow`) that allow it to be treated in most
+/// cases as just another instance of a `T`. The main difference is that this
+/// wrapper has custom `Display` and `Debug` implementations that redact the
+/// wrapped value unless the program has set the [`DISPLAY_GUEST_DATA`] flag.
+///
+/// NOTE: This wrapper type is not airtight: owners of a wrapper can always
+/// dereference it and invoke the Display/Debug impls directly on the resulting
+/// reference to the wrapped value. If `T` is `Clone`, they can also clone the
+/// dereferenced value and display the clone. (This comes with the territory
+/// here: users need to be able to get at the wrapped value to be able to do
+/// anything useful with it!)
+///
+/// NOTE: This type does not provide any other security guarantees (e.g. it does
+/// not ensure that the wrapped memory will be zeroed on drop).
+#[derive(Clone, Copy)]
+#[repr(transparent)]
+pub struct GuestData<T: ?Sized>(T);
+
+impl<T: std::fmt::Display> std::fmt::Display for GuestData<T> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        if DISPLAY_GUEST_DATA.load(Ordering::Relaxed) {
+            write!(f, "{}", self.0)
+        } else {
+            write!(f, "<guest data redacted>")
+        }
+    }
+}
+
+impl<T: std::fmt::Debug> std::fmt::Debug for GuestData<T> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        if DISPLAY_GUEST_DATA.load(Ordering::Relaxed) {
+            write!(f, "{:?}", self.0)
+        } else {
+            write!(f, "<guest data redacted>")
+        }
+    }
+}
+
+impl<T> std::ops::Deref for GuestData<T> {
+    type Target = T;
+
+    fn deref(&self) -> &Self::Target {
+        &self.0
+    }
+}
+
+impl<T> std::ops::DerefMut for GuestData<T> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.0
+    }
+}
+
+impl<T> From<T> for GuestData<T> {
+    fn from(value: T) -> Self {
+        Self(value)
+    }
+}
+
+impl<T> std::borrow::Borrow<T> for GuestData<T> {
+    fn borrow(&self) -> &T {
+        &self.0
+    }
+}
+
 fn numeric_bounds(
     bound: impl RangeBounds<usize>,
     len: usize,

lib/propolis/src/exits.rs

@@ -12,6 +12,8 @@ use bhyve_api::{
     vm_suspend_how,
 };
 
+use crate::common::GuestData;
+
 /// Describes the reason for exiting execution of a vCPU.
 pub struct VmExit {
     /// The instruction pointer of the guest at the time of exit.
@@ -95,17 +97,19 @@ impl From<&bhyve_api::vm_exit_vmx> for VmxDetail {
 
 #[derive(Copy, Clone, Debug)]
 pub struct InstEmul {
-    pub inst_data: [u8; 15],
+    pub inst_data: GuestData<[u8; 15]>,
     pub len: u8,
 }
+
 impl InstEmul {
     pub fn bytes(&self) -> &[u8] {
         &self.inst_data[..usize::min(self.inst_data.len(), self.len as usize)]
     }
 }
 impl From<&bhyve_api::vm_inst_emul> for InstEmul {
     fn from(raw: &bhyve_api::vm_inst_emul) -> Self {
-        let mut res = Self { inst_data: [0u8; 15], len: raw.num_valid };
+        let mut res =
+            Self { inst_data: GuestData::from([0u8; 15]), len: raw.num_valid };
         assert!(res.len as usize <= res.inst_data.len());
         res.inst_data.copy_from_slice(&raw.inst[..]);
 

lib/propolis/src/hw/nvme/cmds.rs

@@ -54,12 +54,14 @@ pub enum AdminCmd {
     /// Asynchronous Event Request Command
     AsyncEventReq,
     /// An unknown admin command
-    Unknown(#[allow(dead_code)] SubmissionQueueEntry),
+    Unknown(#[allow(dead_code)] GuestData<SubmissionQueueEntry>),
 }
 
 impl AdminCmd {
     /// Try to parse an `AdminCmd` out of a raw Submission Entry.
-    pub fn parse(raw: SubmissionQueueEntry) -> Result<Self, ParseErr> {
+    pub fn parse(
+        raw: GuestData<SubmissionQueueEntry>,
+    ) -> Result<Self, ParseErr> {
         let cmd = match raw.opcode() {
             bits::ADMIN_OPC_DELETE_IO_SQ => {
                 AdminCmd::DeleteIOSubQ(raw.cdw10 as u16)
@@ -645,12 +647,14 @@ pub enum NvmCmd {
     /// Read data and metadata
     Read(ReadCmd),
     /// An unknown NVM command
-    Unknown(SubmissionQueueEntry),
+    Unknown(GuestData<SubmissionQueueEntry>),
 }
 
 impl NvmCmd {
     /// Try to parse an `NvmCmd` out of a raw Submission Entry.
-    pub fn parse(raw: SubmissionQueueEntry) -> Result<Self, ParseErr> {
+    pub fn parse(
+        raw: GuestData<SubmissionQueueEntry>,
+    ) -> Result<Self, ParseErr> {
         let _fuse = match (raw.cdw0 >> 8) & 0b11 {
             0b00 => Ok(()),               // Normal (non-fused) operation
             0b01 => Err(ParseErr::Fused), // First fused op
@@ -882,29 +886,29 @@ impl PrpIter<'_> {
             PrpNext::List(base, idx) => {
                 assert!(idx <= PRP_LIST_MAX);
                 let entry_addr = base + u64::from(idx) * 8;
-                let entry: u64 = self
+                let entry: GuestData<u64> = self
                     .mem
                     .read(GuestAddr(entry_addr))
                     .ok_or_else(|| "Unable to read PRP list entry")?;
                 probes::nvme_prp_entry!(
                     || (self as *const Self as u64, entry,)
                 );
 
-                if entry & PAGE_OFFSET as u64 != 0 {
+                if *entry & PAGE_OFFSET as u64 != 0 {
                     return Err("Inappropriate PRP list entry offset");
                 }
 
                 if self.remain <= PAGE_SIZE as u64 {
-                    (entry, self.remain, PrpNext::Done)
+                    (*entry, self.remain, PrpNext::Done)
                 } else if idx != PRP_LIST_MAX {
-                    (entry, PAGE_SIZE as u64, PrpNext::List(base, idx + 1))
+                    (*entry, PAGE_SIZE as u64, PrpNext::List(base, idx + 1))
                 } else {
                     // The last PRP in this list chains to another
                     // (page-aligned) list with the next PRP.
-                    self.next = PrpNext::List(entry, 0);
+                    self.next = PrpNext::List(*entry, 0);
                     probes::nvme_prp_list!(|| (
                         self as *const Self as u64,
-                        entry,
+                        *entry,
                         0,
                     ));
                     return self.get_next();

lib/propolis/src/hw/nvme/queue.rs

@@ -429,7 +429,7 @@ impl SubQueue {
     pub fn pop(
         self: &Arc<SubQueue>,
         mem: &MemCtx,
-    ) -> Option<(SubmissionQueueEntry, Permit, u16)> {
+    ) -> Option<(GuestData<SubmissionQueueEntry>, Permit, u16)> {
         // Attempt to reserve an entry on the Completion Queue
         let permit = self.cq.reserve_entry(&self)?;
         let mut state = self.state.lock();

lib/propolis/src/hw/qemu/fwcfg.rs

@@ -599,7 +599,7 @@ impl FwCfg {
 
         let mem_guard = self.acc_mem.access().expect("usable mem accessor");
         let mem = mem_guard.deref();
-        let req: FwCfgDmaAccess =
+        let req: GuestData<FwCfgDmaAccess> =
             mem.read(GuestAddr(addr)).ok_or(FwCfgErr::BadAddr)?;
 
         fn dma_write_result(
@@ -632,7 +632,7 @@ impl FwCfg {
     fn dma_operation(
         &self,
         state: &mut MutexGuard<State>,
-        req: FwCfgDmaAccess,
+        req: GuestData<FwCfgDmaAccess>,
         mem: &MemCtx,
     ) -> Result<(), FwCfgErr> {
         let opts = FwCfgDmaCtrl::from_bits_truncate(req.ctrl.get());
@@ -995,9 +995,9 @@ mod test {
         submit_dma_req(&dev, req_addr);
 
         // DMA should have successfully completed now
-        assert_eq!(mem.read::<u32>(GuestAddr(req_addr)).unwrap(), 0);
+        assert_eq!(*mem.read::<u32>(GuestAddr(req_addr)).unwrap(), 0);
         let data = mem.read::<[u8; 4]>(GuestAddr(dma_addr)).unwrap();
-        assert_eq!(&data, "QEMU".as_bytes());
+        assert_eq!(&*data, "QEMU".as_bytes());
     }
 
     #[test]
@@ -1019,9 +1019,9 @@ mod test {
         submit_dma_req(&dev, req_addr);
 
         // DMA should have successfully completed now
-        assert_eq!(mem.read::<u32>(GuestAddr(req_addr)).unwrap(), 0);
+        assert_eq!(*mem.read::<u32>(GuestAddr(req_addr)).unwrap(), 0);
         let data = mem.read::<[u8; 4]>(GuestAddr(dma_addr)).unwrap();
-        assert_eq!(data, [0u8; 4]);
+        assert_eq!(*data, [0u8; 4]);
     }
 
     #[test]