examples: Have UFFD handler kill Firecracker should it die

If the UFFD handler exits abnormaly for some reason, have it take down
Firecracker as well by SIGKILL-ing it from a panic hook. For this,
reintroduce the "get peer creds" logic. We have to use SIGKILL because
Firecracker could be inside the handler for a KVM-originated page fault
that is not marked as interruptible, in which case all signals but
SIGKILL are ignored (happens for example during KVM_SET_MSRS when it
triggers the initialization of a gfn_to_pfn_cache for the kvm-clock
page, which uses GUP without FOLL_INTERRUPTIBLE).

While we're at it, add a hint to the generic "process not found" error
message to indicate that potentially Firecracker died, and that the
cause of this could be the UFFD handler crashing (for example, in firecracker-microvm#4601
the cause of the mystery hang is the UFFD handler crashing, but we were
stumped by what's going on for over half a year. Let's avoid that going
forward).

We can't enable this by default because it interferes with unittests,
and also the "malicious_handler", so expose a function on `Runtime` to
enable it only in valid_handler and fault_all_handler.

Signed-off-by: Patrick Roy <[email protected]>

src/firecracker/examples/uffd/fault_all_handler.rs

@@ -24,6 +24,7 @@ fn main() {
     let (stream, _) = listener.accept().expect("Cannot listen on UDS socket");
 
     let mut runtime = Runtime::new(stream, file);
+    runtime.install_panic_hook();
     runtime.run(|uffd_handler: &mut UffdHandler| {
         // Read an event from the userfaultfd.
         let event = uffd_handler

src/firecracker/examples/uffd/uffd_utils.rs

@@ -208,6 +208,43 @@ impl Runtime {
         }
     }
 
+    fn peer_process_credentials(&self) -> libc::ucred {
+        let mut creds: libc::ucred = libc::ucred {
+            pid: 0,
+            gid: 0,
+            uid: 0,
+        };
+        let mut creds_size = size_of::<libc::ucred>() as u32;
+        let ret = unsafe {
+            libc::getsockopt(
+                self.stream.as_raw_fd(),
+                libc::SOL_SOCKET,
+                libc::SO_PEERCRED,
+                &mut creds as *mut _ as *mut _,
+                &mut creds_size as *mut libc::socklen_t,
+            )
+        };
+        if ret != 0 {
+            panic!("Failed to get peer process credentials");
+        }
+        creds
+    }
+
+    pub fn install_panic_hook(&self) {
+        let peer_creds = self.peer_process_credentials();
+
+        let default_panic_hook = std::panic::take_hook();
+        std::panic::set_hook(Box::new(move |panic_info| {
+            let r = unsafe { libc::kill(peer_creds.pid, libc::SIGKILL) };
+
+            if r != 0 {
+                eprintln!("Failed to kill Firecracker process from panic hook");
+            }
+
+            default_panic_hook(panic_info);
+        }));
+    }
+
     /// Polls the `UnixStream` and UFFD fds in a loop.
     /// When stream is polled, new uffd is retrieved.
     /// When uffd is polled, page fault is handled by

src/firecracker/examples/uffd/valid_handler.rs

@@ -24,6 +24,7 @@ fn main() {
     let (stream, _) = listener.accept().expect("Cannot listen on UDS socket");
 
     let mut runtime = Runtime::new(stream, file);
+    runtime.install_panic_hook();
     runtime.run(|uffd_handler: &mut UffdHandler| {
         // Read an event from the userfaultfd.
         let event = uffd_handler

tests/framework/microvm.py

@@ -310,6 +310,9 @@ def kill(self):
             if self.screen_pid:
                 os.kill(self.screen_pid, signal.SIGKILL)
         except:
+            LOG.error(
+                "Failed to kill Firecracker Process. Did it already die (or did the UFFD handler process die and take it down)?"
+            )
             LOG.error(self.log_data)
             raise