Add support for execve

RFC: https://github.com/occlum/occlum/issues/429

src/libos/src/process/do_exec.rs

@@ -0,0 +1,96 @@
+use std::ffi::{CStr, CString};
+use std::path::Path;
+
+use super::do_exit::exit_old_process_for_execve;
+use super::do_spawn::new_process_for_exec;
+use super::process::ProcessFilter;
+use super::term_status::TermStatus;
+use super::wait::Waiter;
+use super::{do_exit, do_exit_group};
+use super::{table, ProcessRef, ProcessStatus};
+use super::{task, ThreadRef};
+use crate::interrupt::broadcast_interrupts;
+use crate::prelude::*;
+
+// FIXME: `occlum exec` command will return early if the application calls execve successfully.
+// Because the "execved"-ed application will run on a new thread and the current thread will exit.
+// `occlum run` will not have this problem.
+
+pub fn do_exec(
+    path: &str,
+    argv: &[CString],
+    envp: &[CString],
+    current_ref: &ThreadRef,
+) -> Result<isize> {
+    trace!(
+        "exec current process pid = {:?}",
+        current_ref.process().pid()
+    );
+
+    // Construct new process structure but with same parent, pid, tid
+    let current = current!();
+    let new_process_ref = super::do_spawn::new_process_for_exec(path, argv, envp, current_ref);
+
+    if let Ok(new_process_ref) = new_process_ref {
+        let new_main_thread = new_process_ref
+            .main_thread()
+            .expect("the main thread is just created; it must exist");
+
+        // Force exit all child threads of current process
+        let term_status = TermStatus::Exited(0 as u8);
+        current.process().force_exit(term_status);
+
+        // Don't hesitate. Interrupt all threads right now (except the calling thread).
+        broadcast_interrupts();
+
+        // Wait for all threads (except calling thread) to exit
+        wait_for_other_threads_to_exit(current);
+
+        // Exit current thread and let new process to adopt current's child process
+        exit_old_process_for_execve(term_status, new_process_ref.clone());
+
+        // Update process and thread in global table
+        table::replace_process(new_process_ref.pid(), new_process_ref.clone());
+        table::replace_thread(
+            new_process_ref.pid(),
+            new_process_ref.main_thread().unwrap(),
+        );
+
+        // Finally, enqueue the new thread for execution
+        task::enqueue_and_exec(new_main_thread);
+        return Ok(0);
+    } else {
+        // There is something wrong when constructing new process. Just return the error.
+        let error = new_process_ref.unwrap_err();
+        return Err(error);
+    }
+}
+
+// Blocking wait until there is only one thread in the calling process
+fn wait_for_other_threads_to_exit(current_ref: ThreadRef) {
+    use super::do_futex::{self, FutexTimeout};
+    use crate::time::{timespec_t, ClockID};
+    use core::time::Duration;
+
+    // Set timeout to 50ms
+    let timeout = FutexTimeout::new(
+        ClockID::CLOCK_MONOTONIC,
+        timespec_t::from(Duration::from_millis(50)),
+    );
+    // Use calling process's pointer as futex value
+    let futex_val = Arc::as_ptr(&current_ref.process()) as *const i32;
+    loop {
+        let thread_num = current_ref.process().threads().len();
+        if thread_num == 1 {
+            return;
+        }
+        // Blocking wait here. When a thread exit, it will notify us.
+        unsafe {
+            do_futex::futex_wait(
+                Arc::as_ptr(&current_ref.process()) as *const i32,
+                *futex_val,
+                &Some(timeout),
+            )
+        };
+    }
+}

src/libos/src/process/do_exit.rs

@@ -3,7 +3,7 @@ use std::intrinsics::atomic_store;
 
 use super::do_futex::futex_wake;
 use super::process::{Process, ProcessFilter};
-use super::{table, TermStatus, ThreadRef, ThreadStatus};
+use super::{table, ProcessRef, TermStatus, ThreadRef, ThreadStatus};
 use crate::prelude::*;
 use crate::signal::{KernelSignal, SigNum};
 
@@ -56,6 +56,10 @@ fn exit_thread(term_status: TermStatus) {
     if num_remaining_threads == 0 {
         exit_process(&thread, term_status);
     }
+
+    // Notify a thread, if any, that wait on this thread to exit.
+    // E.g. In execve, the new thread should wait for old process's all thread to exit
+    futex_wake(Arc::as_ptr(&thread.process()) as *const i32, 1);
 }
 
 fn exit_process(thread: &ThreadRef, term_status: TermStatus) {
@@ -133,3 +137,68 @@ fn send_sigchld_to(parent: &Arc<Process>) {
     let mut sig_queues = parent.sig_queues().write().unwrap();
     sig_queues.enqueue(signal);
 }
+
+pub fn exit_old_process_for_execve(term_status: TermStatus, new_parent_ref: ProcessRef) {
+    let thread = current!();
+
+    // Exit current thread
+    let num_remaining_threads = thread.exit(term_status);
+    if thread.tid() != thread.process().pid() {
+        // Keep the main thread's tid available as long as the process is not destroyed.
+        // Main thread doesn't need to delete here. It will be repalced later.
+        table::del_thread(thread.tid()).expect("tid must be in the table");
+    }
+
+    debug_assert!(num_remaining_threads == 0);
+    exit_process_for_execve(&thread, new_parent_ref, term_status);
+}
+
+// Let new parent process to adopt current process' children
+fn exit_process_for_execve(
+    thread: &ThreadRef,
+    new_parent_ref: ProcessRef,
+    term_status: TermStatus,
+) {
+    let process = thread.process();
+
+    // Deadlock note: always lock parent first, then child.
+    // Lock the idle process since it may adopt new children.
+    let idle_ref = super::IDLE.process().clone();
+    let mut idle_inner = idle_ref.inner();
+
+    // Lock the parent process as we want to prevent race conditions between
+    // current's exit() and parent's wait5().
+    let mut parent;
+    let mut parent_inner = loop {
+        parent = process.parent();
+        if parent.pid() == 0 {
+            // If the parent is the idle process, don't need to lock again
+            break None;
+        }
+
+        let parent_inner = parent.inner();
+        // To prevent the race condition that parent is changed after `parent()`,
+        // but before `parent().innner()`, we need to check again here.
+        if parent.pid() != process.parent().pid() {
+            continue;
+        }
+        break Some(parent_inner);
+    };
+
+    // Lock the current process
+    let mut process_inner = process.inner();
+    let mut new_parent_inner = new_parent_ref.inner();
+    let pid = process.pid();
+
+    // Let new_process to adopt the children of current process
+    process_inner.exit(term_status, &new_parent_ref, &mut new_parent_inner);
+
+    // Remove current process from parent process' zombie list.
+    if parent_inner.is_none() {
+        debug_assert!(parent.pid() == 0);
+        idle_inner.remove_zombie_child(pid);
+        return;
+    }
+
+    parent_inner.unwrap().remove_zombie_child(pid);
+}

src/libos/src/process/do_spawn/gdb_hook_load_elf.c

@@ -18,4 +18,4 @@ void __attribute__((optimize("O0"))) occlum_gdb_hook_load_elf(
     uint64_t elf_base,
     const char *elf_path,
     uint64_t elf_path_len) {
-}
+}

src/libos/src/process/do_spawn/mod.rs

@@ -7,7 +7,7 @@ use super::elf_file::{ElfFile, ElfHeader, ProgramHeaderExt};
 use super::process::ProcessBuilder;
 use super::spawn_attribute::SpawnAttr;
 use super::task::Task;
-use super::thread::ThreadName;
+use super::thread::{ThreadId, ThreadName};
 use super::{table, task, ProcessRef, ThreadRef};
 use crate::fs::{
     CreationFlags, File, FileDesc, FileTable, FsView, HostStdioFds, StdinFile, StdoutFile,
@@ -107,6 +107,56 @@ fn new_process(
     spawn_attributes: Option<SpawnAttr>,
     host_stdio_fds: Option<&HostStdioFds>,
     current_ref: &ThreadRef,
+) -> Result<ProcessRef> {
+    let new_process_ref = new_process_common(
+        file_path,
+        argv,
+        envp,
+        file_actions,
+        spawn_attributes,
+        host_stdio_fds,
+        current_ref,
+        None,
+    )?;
+    table::add_process(new_process_ref.clone());
+    table::add_thread(new_process_ref.main_thread().unwrap());
+
+    Ok(new_process_ref)
+}
+
+/// Create a new process for execve which will use same parent, pid, tid
+pub fn new_process_for_exec(
+    file_path: &str,
+    argv: &[CString],
+    envp: &[CString],
+    current_ref: &ThreadRef,
+) -> Result<ProcessRef> {
+    let tid = ThreadId {
+        tid: current_ref.process().pid() as u32,
+    };
+    let new_process_ref = new_process_common(
+        file_path,
+        argv,
+        envp,
+        &Vec::new(),
+        None,
+        None,
+        current_ref,
+        Some(tid),
+    )?;
+
+    Ok(new_process_ref)
+}
+
+fn new_process_common(
+    file_path: &str,
+    argv: &[CString],
+    envp: &[CString],
+    file_actions: &[FileAction],
+    spawn_attributes: Option<SpawnAttr>,
+    host_stdio_fds: Option<&HostStdioFds>,
+    current_ref: &ThreadRef,
+    reuse_tid: Option<ThreadId>,
 ) -> Result<ProcessRef> {
     let mut argv = argv.clone().to_vec();
     let (is_script, elf_inode, mut elf_buf, elf_header) =
@@ -223,10 +273,19 @@ fn new_process(
         let elf_name = elf_path.rsplit('/').collect::<Vec<&str>>()[0];
         let thread_name = ThreadName::new(elf_name);
 
-        ProcessBuilder::new()
+        let mut parent;
+        let mut process_builder = ProcessBuilder::new();
+        if reuse_tid.is_some() {
+            process_builder = process_builder.tid(reuse_tid.unwrap());
+            parent = current!().process().parent();
+        } else {
+            parent = process_ref;
+        }
+
+        process_builder
             .vm(vm_ref)
             .exec_path(&elf_path)
-            .parent(process_ref)
+            .parent(parent)
             .task(task)
             .sched(sched_ref)
             .rlimits(rlimit_ref)
@@ -238,9 +297,6 @@ fn new_process(
             .build()?
     };
 
-    table::add_process(new_process_ref.clone());
-    table::add_thread(new_process_ref.main_thread().unwrap());
-
     info!(
         "Process created: elf = {}, pid = {}",
         elf_path,

src/libos/src/process/mod.rs

@@ -32,6 +32,7 @@ pub use self::thread::{Thread, ThreadStatus};
 
 mod do_arch_prctl;
 mod do_clone;
+mod do_exec;
 mod do_exit;
 mod do_futex;
 mod do_getpid;

src/libos/src/process/syscalls.rs

@@ -1,5 +1,6 @@
 use super::do_arch_prctl::ArchPrctlCode;
 use super::do_clone::CloneFlags;
+use super::do_exec::do_exec;
 use super::do_futex::{FutexFlags, FutexOp, FutexTimeout};
 use super::do_spawn::FileAction;
 use super::do_wait4::WaitOptions;
@@ -429,3 +430,16 @@ pub fn do_getgroups(size: isize, buf_ptr: *mut u32) -> Result<isize> {
         Ok(1)
     }
 }
+
+pub fn do_execve(path: *const i8, argv: *const *const i8, envp: *const *const i8) -> Result<isize> {
+    let path = clone_cstring_safely(path)?.to_string_lossy().into_owned();
+    let argv = clone_cstrings_safely(argv)?;
+    let envp = clone_cstrings_safely(envp)?;
+    let current = current!();
+    debug!(
+        "execve: path: {:?}, argv: {:?}, envp: {:?}",
+        path, argv, envp
+    );
+
+    do_exec(&path, &argv, &envp, &current)
+}

src/libos/src/process/table.rs

@@ -31,6 +31,11 @@ pub(super) fn del_process(pid: pid_t) -> Result<ProcessRef> {
     PROCESS_TABLE.lock().unwrap().del(pid)
 }
 
+pub fn replace_process(pid: pid_t, new_process: ProcessRef) -> Result<()> {
+    del_process(pid);
+    add_process(new_process)
+}
+
 pub fn get_thread(tid: pid_t) -> Result<ThreadRef> {
     THREAD_TABLE.lock().unwrap().get(tid)
 }
@@ -43,6 +48,11 @@ pub(super) fn del_thread(tid: pid_t) -> Result<ThreadRef> {
     THREAD_TABLE.lock().unwrap().del(tid)
 }
 
+pub(super) fn replace_thread(tid: pid_t, new_thread: ThreadRef) -> Result<()> {
+    del_thread(tid);
+    add_thread(new_thread)
+}
+
 pub fn debug() {
     println!("process table = {:#?}", PROCESS_TABLE.lock().unwrap());
     println!("thread table = {:#?}", THREAD_TABLE.lock().unwrap());

src/libos/src/process/thread/id.rs

@@ -8,7 +8,7 @@ use crate::prelude::*;
 /// And when a ThreadID instance is freed, its ID is automatically freed too.
 #[derive(Debug, PartialEq)]
 pub struct ThreadId {
-    tid: u32,
+    pub tid: u32,
 }
 
 impl ThreadId {
@@ -41,7 +41,7 @@ impl Drop for ThreadId {
         }
 
         let mut alloc = THREAD_ID_ALLOC.lock().unwrap();
-        alloc.free(self.tid).expect("tid must has been allocated");
+        alloc.free(self.tid);
     }
 }
 
@@ -90,7 +90,9 @@ impl IdAlloc {
     }
 
     pub fn free(&mut self, id: u32) -> Option<u32> {
-        debug_assert!(self.used_ids.contains(&id));
+        // Note: When enableing "execve", there is situation that the ThreadId is reused.
+        // And thus when exit, it may free twice.
+        // debug_assert!(self.used_ids.contains(&id));
         if self.used_ids.remove(&id) {
             Some(id)
         } else {

src/libos/src/syscall/mod.rs

@@ -40,8 +40,8 @@ use crate::net::{
     do_shutdown, do_socket, do_socketpair, mmsghdr, msghdr, msghdr_mut,
 };
 use crate::process::{
-    do_arch_prctl, do_clone, do_exit, do_exit_group, do_futex, do_getegid, do_geteuid, do_getgid,
-    do_getgroups, do_getpgid, do_getpid, do_getppid, do_gettid, do_getuid, do_prctl,
+    do_arch_prctl, do_clone, do_execve, do_exit, do_exit_group, do_futex, do_getegid, do_geteuid,
+    do_getgid, do_getgroups, do_getpgid, do_getpid, do_getppid, do_gettid, do_getuid, do_prctl,
     do_set_tid_address, do_spawn_for_glibc, do_spawn_for_musl, do_wait4, pid_t, posix_spawnattr_t,
     FdOp, SpawnFileActions, ThreadStatus,
 };
@@ -144,7 +144,7 @@ macro_rules! process_syscall_table_with_callback {
             (Clone = 56) => do_clone(flags: u32, stack_addr: usize, ptid: *mut pid_t, ctid: *mut pid_t, new_tls: usize),
             (Fork = 57) => handle_unsupported(),
             (Vfork = 58) => handle_unsupported(),
-            (Execve = 59) => handle_unsupported(),
+            (Execve = 59) => do_execve(path: *const i8, argv: *const *const i8, envp: *const *const i8),
             (Exit = 60) => do_exit(exit_status: i32),
             (Wait4 = 61) => do_wait4(pid: i32, _exit_status: *mut i32, options: u32),
             (Kill = 62) => do_kill(pid: i32, sig: c_int),

test/Makefile

@@ -19,7 +19,7 @@ TESTS ?= env empty hello_world malloc mmap file fs_perms getpid spawn sched pipe
 	truncate readdir mkdir open stat link symlink chmod chown tls pthread system_info rlimit \
 	server server_epoll unix_socket cout hostfs cpuid rdtsc device sleep exit_group \
 	ioctl fcntl eventfd emulate_syscall access signal sysinfo prctl rename procfs wait \
-	spawn_attribute statfs
+	spawn_attribute exec statfs
 # Benchmarks: need to be compiled and run by bench-% target
 BENCHES := spawn_and_exit_latency pipe_throughput unix_socket_throughput
 

test/exec/Makefile

@@ -0,0 +1,5 @@
+include ../test_common.mk
+
+EXTRA_C_FLAGS := -g -lpthread
+EXTRA_LINK_FLAGS :=
+BIN_ARGS :=

test/exec/main.c

@@ -0,0 +1,92 @@
+#include <unistd.h>
+#include <string.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <spawn.h>
+#include <sys/wait.h>
+#include <pthread.h>
+#include <stdbool.h>
+#include "test.h"
+
+static void *just_sleep(void *_arg) {
+    bool should_exit_by_execve = *(bool *)_arg;
+    sleep(3);
+
+    // If should_exit_by_execve is true, execve should be done before sleep returns.
+    if (should_exit_by_execve) {
+        printf("This should never be reached");
+        exit(-1);
+    } else {
+        printf("sleep is done\n");
+    }
+    return NULL;
+}
+
+int test_execve_no_return(void) {
+    bool should_exit_by_execve = true;
+    pthread_t child_thread;
+    if (pthread_create(&child_thread, NULL, just_sleep, (void *)&should_exit_by_execve) < 0) {
+        THROW_ERROR("pthread_create failed");
+    }
+
+    char *args[] = {"spawn", NULL};
+    execve("/bin/spawn", args, NULL);
+
+    THROW_ERROR("The program shouldn't reach here.");
+    return -1;
+}
+
+int test_execve_error_return(void) {
+    // execve will fail in this case and thus the child thread will not exit until finish
+    bool should_exit_by_execve = false;
+    pthread_t child_thread;
+    if (pthread_create(&child_thread, NULL, just_sleep, (void *)&should_exit_by_execve) < 0) {
+        THROW_ERROR("pthread_create failed");
+    }
+
+    // during the time, try execve a non-exit process
+    int ret = execve("/bin/joke", NULL, NULL);
+    if (ret != -1 || errno != ENOENT) {
+        THROW_ERROR("execve error code wrong");
+    }
+
+    pthread_join(child_thread, NULL);
+    return 0;
+}
+
+int test_execve_on_child_thread(void) {
+    int ret, child_pid, status;
+
+    // construct child process args
+    int child_argc = 3; // ./nauty_child -t execve_thread
+    char **child_argv = calloc(1, sizeof(char *) * (child_argc + 1));
+    child_argv[0] = strdup("naughty_child");
+    child_argv[1] = strdup("-t");
+    child_argv[2] = strdup("execve_thread");
+
+    ret = posix_spawn(&child_pid, "/bin/naughty_child", NULL, NULL, child_argv, NULL);
+    if (ret != 0) {
+        THROW_ERROR("failed to spawn a child process");
+    }
+
+    ret = waitpid(child_pid, &status, 0);
+    if (ret < 0) {
+        THROW_ERROR("failed to wait4 the child process");
+    }
+    printf("child process %d exit status = %d\n", child_pid, status);
+    if (status != 0) {
+        THROW_ERROR("child process exit with error");
+    }
+
+    return 0;
+}
+
+static test_case_t test_cases[] = {
+    TEST_CASE(test_execve_on_child_thread),
+    TEST_CASE(test_execve_error_return),
+    TEST_CASE(test_execve_no_return),
+};
+
+int main() {
+    return test_suite_run(test_cases, ARRAY_SIZE(test_cases));
+}

test/getpid/main.c

@@ -1,8 +1,11 @@
+#define _GNU_SOURCE
 #include <sys/types.h>
 #include <unistd.h>
 #include <stdio.h>
+#include <sys/syscall.h>
 
 int main(int argc, const char *argv[]) {
     printf("Run a new process with pid = %d and ppid = %d\n", getpid(), getppid());
+    printf("tid = %ld\n", syscall(SYS_gettid));
     return 0;
 }

test/naughty_child/Makefile

@@ -1,5 +1,5 @@
 include ../test_common.mk
 
-EXTRA_C_FLAGS := -g
+EXTRA_C_FLAGS := -g -lpthread
 EXTRA_LINK_FLAGS :=
 BIN_ARGS :=

test/naughty_child/main.c

@@ -5,6 +5,8 @@
 #include <stdlib.h>
 #include <features.h>
 #include <sys/stat.h>
+#include <pthread.h>
+#include <unistd.h>
 #include "test.h"
 
 char **g_argv;
@@ -100,6 +102,94 @@ int test_ioctl_fioclex() {
     return 0;
 }
 
+// This child process will first create multiple threads which are waiting on a condition
+// and then a child thread will call execve and all threads should be destroyed except the
+// main thread. Use "pthread" test case as reference.
+#define NTHREADS                (5)
+#define WAIT_ROUND          (100000)
+
+struct thread_cond_arg {
+    int                         ti;
+    volatile unsigned int      *val;
+    volatile int               *exit_thread_count;
+    pthread_cond_t             *cond_val;
+    pthread_mutex_t            *mutex;
+};
+
+static void *thread_cond_wait(void *_arg) {
+    struct thread_cond_arg *arg = _arg;
+    printf("Thread #%d: start to wait on condition variable.\n", arg->ti);
+    for (unsigned int i = 0; i < WAIT_ROUND; ++i) {
+        pthread_mutex_lock(arg->mutex);
+        // execve on a child thread with mutex
+        if (arg->ti == NTHREADS - 4) {
+            char *args[] = {"/bin/getpid", NULL};
+            if (execve("/bin/getpid", args, NULL) < 0) {
+                printf("execve failed with errno: %d", errno);
+                exit(errno);
+            }
+        }
+        while (*(arg->val) == 0) {
+            pthread_cond_wait(arg->cond_val, arg->mutex);
+        }
+        pthread_mutex_unlock(arg->mutex);
+    }
+    (*arg->exit_thread_count)++;
+    printf("Thread #%d: exited.\n", arg->ti);
+    return NULL;
+}
+
+int test_execve_child_thread() {
+    volatile unsigned int val = 0;
+    volatile int exit_thread_count = 0;
+    pthread_t threads[NTHREADS];
+    struct thread_cond_arg thread_args[NTHREADS];
+    pthread_cond_t cond_val = PTHREAD_COND_INITIALIZER;
+    pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
+
+    /*
+     * Start the threads waiting on the condition variable
+     */
+    for (int ti = 0; ti < NTHREADS; ti++) {
+        struct thread_cond_arg *thread_arg = &thread_args[ti];
+        thread_arg->ti = ti;
+        thread_arg->val = &val;
+        thread_arg->exit_thread_count = &exit_thread_count;
+        thread_arg->cond_val = &cond_val;
+        thread_arg->mutex = &mutex;
+
+        if (pthread_create(&threads[ti], NULL, thread_cond_wait, thread_arg) < 0) {
+            printf("ERROR: pthread_create failed (ti = %d)\n", ti);
+            return -1;
+        }
+    }
+
+    /*
+    * Unblock all threads currently waiting on the condition variable
+    */
+    while (exit_thread_count < NTHREADS) {
+        pthread_mutex_lock(&mutex);
+        val = 1;
+        pthread_cond_broadcast(&cond_val);
+        pthread_mutex_unlock(&mutex);
+
+        pthread_mutex_lock(&mutex);
+        val = 0;
+        pthread_mutex_unlock(&mutex);
+    }
+
+    // wait for all threads to finish
+    for (int ti = 0; ti < NTHREADS; ti++) {
+        if (pthread_join(threads[ti], NULL) < 0) {
+            printf("ERROR: pthread_join failed (ti = %d)\n", ti);
+            return -1;
+        }
+    }
+
+    THROW_ERROR("This should never be reached!");
+    return -1;
+}
+
 // ============================================================================
 // Test suite
 // ============================================================================
@@ -113,6 +203,8 @@ int start_test(const char *test_name) {
         return test_spawn_attribute_sigdef();
     } else if (strcmp(test_name, "fioclex") == 0) {
         return test_ioctl_fioclex();
+    } else if (strcmp(test_name, "execve_thread") == 0) {
+        return test_execve_child_thread();
     } else {
         fprintf(stderr, "[child] test case not found\n");
         return -1;
@@ -120,8 +212,8 @@ int start_test(const char *test_name) {
 }
 
 void print_usage() {
-    fprintf(stderr, "Usage:\n nauty_child [-t testcase1] [-t testcase2] ...\n\n");
-    fprintf(stderr, " Now support testcase: <sigmask, sigdef, fioclex>\n");
+    fprintf(stderr, "Usage:\n naughty_child [-t testcase1] [-t testcase2] ...\n\n");
+    fprintf(stderr, " Now support testcase: <sigmask, sigdef, fioclex, execve_thread>\n");
 }
 
 int main(int argc, char *argv[]) {