diff --git a/src/Enclave.edl b/src/Enclave.edl
index 50b9f63b..30b43994 100644
--- a/src/Enclave.edl
+++ b/src/Enclave.edl
@@ -61,6 +61,23 @@ enclave {
          *      ESRCH - Cannot find the LibOS thread.
          */
         public int occlum_ecall_exec_thread(int libos_tid, int host_tid);
+
+        /*
+         * Send a signal to one or multiple LibOS processes.
+         *
+         * If pid == -1, send the signal to all processes. If pid > 0, send
+         * the signal to the specific process. For the purpose of security,
+         * the only allowed signals for now are SIGKILL and SIGTERM.
+         *
+         * @retval On success, return 0. On error, return -errno.
+         *
+         * The possible values of errno are
+         *      EAGAIN - The LibOS is not initialized.
+         *      EINVAL - The value of an argument are invalid.
+         *      ESRCH - Cannot find the process specified by pid.
+         *      EPERM - No permission to send the signal or to the process.
+         */
+        public int occlum_ecall_kill(int pid, int sig);
     };
 
     untrusted {
diff --git a/src/libos/src/entry.rs b/src/libos/src/entry.rs
index a6a0cdf1..db48cf1c 100644
--- a/src/libos/src/entry.rs
+++ b/src/libos/src/entry.rs
@@ -1,14 +1,16 @@
-use super::*;
-use exception::*;
-use fs::HostStdioFds;
-use process::pid_t;
 use std::ffi::{CStr, CString, OsString};
 use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Once;
-use util::log::LevelFilter;
-use util::mem_util::from_untrusted::*;
-use util::sgx::allow_debug as sgx_allow_debug;
+
+use super::*;
+use crate::exception::*;
+use crate::fs::HostStdioFds;
+use crate::process::ProcessFilter;
+use crate::signal::SigNum;
+use crate::util::log::LevelFilter;
+use crate::util::mem_util::from_untrusted::*;
+use crate::util::sgx::allow_debug as sgx_allow_debug;
 use sgx_tse::*;
 
 pub static mut INSTANCE_DIR: String = String::new();
@@ -53,7 +55,7 @@ pub extern "C" fn occlum_ecall_init(log_level: *const c_char, instance_dir: *con
     INIT_ONCE.call_once(|| {
         // Init the log infrastructure first so that log messages will be printed afterwards
         util::log::init(log_level);
- 
+
         // Init MPX for SFI if MPX is available
         let report = rsgx_self_report();
         if (report.body.attributes.xfrm & SGX_XFRM_MPX != 0) {
@@ -129,6 +131,25 @@ pub extern "C" fn occlum_ecall_exec_thread(libos_pid: i32, host_tid: i32) -> i32
     .unwrap_or(ecall_errno!(EFAULT))
 }
 
+#[no_mangle]
+pub extern "C" fn occlum_ecall_kill(pid: i32, sig: i32) -> i32 {
+    if HAS_INIT.load(Ordering::SeqCst) == false {
+        return ecall_errno!(EAGAIN);
+    }
+
+    let _ = unsafe { backtrace::enable_backtrace(&ENCLAVE_PATH, PrintFormat::Short) };
+    panic::catch_unwind(|| {
+        backtrace::__rust_begin_short_backtrace(|| match do_kill(pid, sig) {
+            Ok(()) => 0,
+            Err(e) => {
+                eprintln!("failed to kill: {}", e.backtrace());
+                ecall_errno!(e.errno())
+            }
+        })
+    })
+    .unwrap_or(ecall_errno!(EFAULT))
+}
+
 fn parse_log_level(level_chars: *const c_char) -> Result<LevelFilter> {
     const DEFAULT_LEVEL: LevelFilter = LevelFilter::Off;
 
@@ -255,3 +276,23 @@ fn validate_program_path(target_path: &PathBuf) -> Result<()> {
     }
     Ok(())
 }
+
+fn do_kill(pid: i32, sig: i32) -> Result<()> {
+    let filter = if pid > 0 {
+        ProcessFilter::WithPid(pid as pid_t)
+    } else if pid == -1 {
+        ProcessFilter::WithAnyPid
+    } else if pid < 0 {
+        return_errno!(EINVAL, "Invalid pid");
+    } else {
+        // pid == 0
+        return_errno!(EPERM, "Process 0 cannot be killed");
+    };
+    let signum = {
+        if sig < 0 {
+            return_errno!(EINVAL, "invalid arguments");
+        }
+        SigNum::from_u8(sig as u8)?
+    };
+    crate::signal::do_kill_from_outside_enclave(filter, signum)
+}
diff --git a/src/libos/src/signal/do_kill.rs b/src/libos/src/signal/do_kill.rs
index 8cfe8d35..3b7b0c4e 100644
--- a/src/libos/src/signal/do_kill.rs
+++ b/src/libos/src/signal/do_kill.rs
@@ -1,4 +1,5 @@
-use super::signals::{UserSignal, UserSignalKind};
+use super::constants::*;
+use super::signals::{KernelSignal, UserSignal, UserSignalKind};
 use super::{SigNum, Signal};
 use crate::prelude::*;
 use crate::process::{table, ProcessFilter, ProcessRef, ProcessStatus, ThreadRef, ThreadStatus};
@@ -21,6 +22,32 @@ pub fn do_kill(filter: ProcessFilter, signum: SigNum) -> Result<()> {
     Ok(())
 }
 
+/// Send a signal from the outside the enclave.
+///
+/// Such a call must be performed very carefully. The obvious reason
+/// is that the call is not trusted. And there is a less obvious reason:
+/// the function is not executed during a normal syscall. Thus, current!() does
+/// not refer to a valid LibOS thread. So let's implement this function with
+/// these two insights in mind.
+pub fn do_kill_from_outside_enclave(filter: ProcessFilter, signum: SigNum) -> Result<()> {
+    let signal = {
+        if signum != SIGKILL && signum != SIGTERM {
+            return_errno!(EPERM, "The signal is not allowed");
+        }
+        Box::new(KernelSignal::new(signum))
+    };
+    let processes = get_processes(&filter)?;
+    for process in processes {
+        if process.status() == ProcessStatus::Zombie {
+            continue;
+        }
+
+        let mut sig_queues = process.sig_queues().lock().unwrap();
+        sig_queues.enqueue(signal.clone());
+    }
+    Ok(())
+}
+
 fn get_processes(filter: &ProcessFilter) -> Result<Vec<ProcessRef>> {
     let processes = match filter {
         ProcessFilter::WithAnyPid => table::get_all_processes(),
diff --git a/src/libos/src/signal/mod.rs b/src/libos/src/signal/mod.rs
index 2dcd29f9..f119fb61 100644
--- a/src/libos/src/signal/mod.rs
+++ b/src/libos/src/signal/mod.rs
@@ -6,6 +6,7 @@ use sig_action::{SigAction, SigActionFlags, SigDefaultAction};
 
 pub use self::c_types::{sigaction_t, sigset_t};
 pub use self::constants::*;
+pub use self::do_kill::do_kill_from_outside_enclave;
 pub use self::do_sigreturn::{deliver_signal, force_signal};
 pub use self::sig_dispositions::SigDispositions;
 pub use self::sig_num::SigNum;
diff --git a/src/pal/include/occlum_pal_api.h b/src/pal/include/occlum_pal_api.h
index 394e34f9..26679b1e 100644
--- a/src/pal/include/occlum_pal_api.h
+++ b/src/pal/include/occlum_pal_api.h
@@ -73,6 +73,18 @@ int occlum_pal_exec(const char* cmd_path,
                     const struct occlum_stdio_fds* io_fds,
                     int* exit_status);
 
+/*
+ * @brief Send a signal to one or multiple LibOS processes
+ *
+ * @param pid   If pid > 0, send the signal to the process with the
+ *              pid; if pid == -1, send the signal to all processes.
+ * @param sig   The signal number. For the purpose of security, the
+ *              only allowed signals for now are SIGKILL and SIGTERM.
+ *
+ * @retval If 0, then success; otherwise, check errno for the exact error type.
+ */
+int occlum_pal_kill(int pid, int sig);
+
 /*
  * @brief Destroy teh Occlum enclave
  *
diff --git a/src/pal/src/pal_api.c b/src/pal/src/pal_api.c
index cb28a189..1199f720 100644
--- a/src/pal/src/pal_api.c
+++ b/src/pal/src/pal_api.c
@@ -94,6 +94,32 @@ int occlum_pal_exec(const char* cmd_path,
     return 0;
 }
 
+int occlum_pal_kill(int pid, int sig) {
+    errno = 0;
+
+    sgx_enclave_id_t eid = pal_get_enclave_id();
+    if (eid == SGX_INVALID_ENCLAVE_ID) {
+        errno = ENOENT;
+        PAL_ERROR("Enclave is not initialized yet.");
+        return -1;
+    }
+
+    int ecall_ret = 0;
+    sgx_status_t ecall_status = occlum_ecall_kill(eid, &ecall_ret, pid, sig);
+    if (ecall_status != SGX_SUCCESS) {
+        const char* sgx_err = pal_get_sgx_error_msg(ecall_status);
+        PAL_ERROR("Failed to do ECall: %s", sgx_err);
+        return -1;
+    }
+    if (ecall_ret < 0) {
+        errno = -ecall_ret;
+        PAL_ERROR("Failed to occlum_ecall_kill: %s", errno2str(errno));
+        return -1;
+    }
+
+    return 0;
+}
+
 int occlum_pal_destroy(void) {
     errno = 0;