criu/plugins/cuda/cuda_plugin.c

#include "criu-log.h"
#include "plugin.h"
#include "util.h"
#include "cr_options.h"
#include "pid.h"
#include "proc_parse.h"
#include "seize.h"
#include "fault-injection.h"

#include <common/list.h>
#include <compel/infect.h>

#include <ctype.h>
#include <fcntl.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/ptrace.h>
#include <sys/wait.h>

/* cuda-checkpoint binary should live in your PATH */
#define CUDA_CHECKPOINT "cuda-checkpoint"

/* cuda-checkpoint --action flags */
#define ACTION_LOCK	  "lock"
#define ACTION_CHECKPOINT "checkpoint"
#define ACTION_RESTORE	  "restore"
#define ACTION_UNLOCK	  "unlock"

#define CUDA_CKPT_BUF_SIZE (128)

#ifdef LOG_PREFIX
#undef LOG_PREFIX
#endif
#define LOG_PREFIX "cuda_plugin: "

/* Disable plugin functionality if cuda-checkpoint is not in $PATH or driver
 * version doesn't support --action flag
 */
bool plugin_disabled = false;

bool plugin_added_to_inventory = false;

struct pid_info {
	int pid;
	char checkpointed;
	struct list_head list;
};

/* Used to track which PID's we've paused CUDA operations on so far so we can
 * release them after we're done with the DUMP
 */
static LIST_HEAD(cuda_pids);

static void dealloc_pid_buffer(struct list_head *pid_buf)
{
	struct pid_info *info;
	struct pid_info *n;

	list_for_each_entry_safe(info, n, pid_buf, list) {
		list_del(&info->list);
		xfree(info);
	}
}

static int add_pid_to_buf(struct list_head *pid_buf, int pid)
{
	struct pid_info *new = xmalloc(sizeof(*new));

	if (new == NULL) {
		return -1;
	}

	new->pid = pid;
	new->checkpointed = 0;
	list_add_tail(&new->list, pid_buf);

	return 0;
}

static int update_checkpointed_pid(struct list_head *pid_buf, int pid)
{
	struct pid_info *info;

	list_for_each_entry(info, pid_buf, list) {
		if (info->pid == pid) {
			info->checkpointed = 1;
			return 0;
		}
	}

	return -1;
}

static int launch_cuda_checkpoint(const char **args, char *buf, int buf_size)
{
#define READ  0
#define WRITE 1
	int fd[2], buf_off;

	if (pipe(fd) != 0) {
		pr_err("Couldn't create pipes for reading cuda-checkpoint output\n");
		return -1;
	}

	buf[0] = '\0';

	int child_pid = fork();
	if (child_pid == -1) {
		pr_err("Failed to fork to exec cuda-checkpoint\n");
		close(fd[READ]);
		close(fd[WRITE]);
		return -1;
	}

	if (child_pid == 0) { // child
		if (dup2(fd[WRITE], STDOUT_FILENO) == -1) {
			pr_perror("unable to clone fd %d->%d", fd[WRITE], STDOUT_FILENO);
			_exit(EXIT_FAILURE);
		}
		if (dup2(fd[WRITE], STDERR_FILENO) == -1) {
			pr_perror("unable to clone fd %d->%d", fd[WRITE], STDERR_FILENO);
			_exit(EXIT_FAILURE);
		}
		close(fd[READ]);

		close_fds(STDERR_FILENO + 1);

		execvp(args[0], (char **)args);

		/* We can't use pr_error() as log file fd is closed. */
		fprintf(stderr, "execvp(\"%s\") failed: %s\n", args[0], strerror(errno));

		_exit(EXIT_FAILURE);
	}

	close(fd[WRITE]);
	buf_off = 0;
	/* Reserve one byte for the null charracter. */
	buf_size--;
	while (buf_off < buf_size) {
		int bytes_read;
		bytes_read = read(fd[READ], buf + buf_off, buf_size - buf_off);
		if (bytes_read == -1) {
			pr_perror("Unable to read output of cuda-checkpoint");
			goto err;
		}
		if (bytes_read == 0)
			break;
		buf_off += bytes_read;
	}
	buf[buf_off] = '\0';

	/* Clear out any of the remaining output in the pipe in case the buffer wasn't large enough */
	while (true) {
		char scratch[1024];
		int bytes_read;
		bytes_read = read(fd[READ], scratch, sizeof(scratch));
		if (bytes_read == -1) {
			pr_perror("Unable to read output of cuda-checkpoint");
			goto err;
		}
		if (bytes_read == 0)
			break;
	}
	close(fd[READ]);

	int status, exit_code = -1;
	if (waitpid(child_pid, &status, 0) == -1) {
		pr_perror("Unable to wait for the cuda-checkpoint process %d", child_pid);
		goto err;
	}
	if (WIFSIGNALED(status)) {
		int sig = WTERMSIG(status);

		pr_err("cuda-checkpoint unexpectedly signaled with %d: %s\n", sig, strsignal(sig));
	} else if (WIFEXITED(status)) {
		exit_code = WEXITSTATUS(status);
	} else {
		pr_err("cuda-checkpoint exited improperly: %u\n", status);
	}

	if (exit_code != EXIT_SUCCESS)
		pr_debug("cuda-checkpoint output ===>\n%s\n"
			 "<=== cuda-checkpoint output\n",
			 buf);

	return exit_code;
err:
	kill(child_pid, SIGKILL);
	waitpid(child_pid, NULL, 0);
	return -1;
}

/**
 * Checks if a given flag is supported by the cuda-checkpoint utility
 *
 * Returns:
 *  1 if the flag is supported,
 *  0 if the flag is not supported,
 *  -1 if there was an error launching the cuda-checkpoint utility.
 */
static int cuda_checkpoint_supports_flag(const char *flag)
{
	char msg_buf[2048];
	const char *args[] = { CUDA_CHECKPOINT, "-h", NULL };

	if (launch_cuda_checkpoint(args, msg_buf, sizeof(msg_buf)) != 0)
		return -1;

	if (strstr(msg_buf, flag) == NULL)
		return 0;

	return 1;
}

/* Retrieve the cuda restore thread TID from the root pid */
static int get_cuda_restore_tid(int root_pid)
{
	char pid_buf[16];
	char pid_out[CUDA_CKPT_BUF_SIZE];

	snprintf(pid_buf, sizeof(pid_buf), "%d", root_pid);

	const char *args[] = { CUDA_CHECKPOINT, "--get-restore-tid", "--pid", pid_buf, NULL };
	int ret = launch_cuda_checkpoint(args, pid_out, sizeof(pid_out));
	if (ret != 0) {
		pr_err("Failed to launch cuda-checkpoint to retrieve restore tid: %s\n", pid_out);
		return -1;
	}

	return atoi(pid_out);
}

static int cuda_process_checkpoint_action(int pid, const char *action, unsigned int timeout, char *msg_buf,
					  int buf_size)
{
	char pid_buf[16];
	char timeout_buf[16];

	snprintf(pid_buf, sizeof(pid_buf), "%d", pid);

	const char *args[] = { CUDA_CHECKPOINT, "--action", action, "--pid", pid_buf, NULL /* --timeout */,
			       NULL /* timeout_val */, NULL };
	if (timeout > 0) {
		snprintf(timeout_buf, sizeof(timeout_buf), "%d", timeout);
		args[5] = "--timeout";
		args[6] = timeout_buf;
	}

	return launch_cuda_checkpoint(args, msg_buf, buf_size);
}

static int interrupt_restore_thread(int restore_tid, k_rtsigset_t *restore_sigset)
{
	/* Since we resumed a thread that CRIU previously already froze we need to
	 * INTERRUPT it once again, task was already SEIZE'd so we don't need to do
	 * a compel_interrupt_task()
	 */
	if (ptrace(PTRACE_INTERRUPT, restore_tid, NULL, 0)) {
		pr_err("Could not interrupt cuda restore tid %d after checkpoint, process may be in strange state\n",
		       restore_tid);
		return -1;
	}

	struct proc_status_creds creds;
	if (compel_wait_task(restore_tid, -1, parse_pid_status, NULL, &creds.s, NULL) != COMPEL_TASK_ALIVE) {
		pr_err("compel_wait_task failed after interrupt\n");
		return -1;
	}

	if (ptrace(PTRACE_SETOPTIONS, restore_tid, NULL, PTRACE_O_SUSPEND_SECCOMP | PTRACE_O_TRACESYSGOOD)) {
		pr_err("Failed to set ptrace options on interrupt for restore tid %d\n", restore_tid);
		return -1;
	}

	if (ptrace(PTRACE_SETSIGMASK, restore_tid, sizeof(*restore_sigset), restore_sigset)) {
		pr_err("Unable to restore original sigmask to restore tid %d\n", restore_tid);
		return -1;
	}

	return 0;
}

static int resume_restore_thread(int restore_tid, k_rtsigset_t *save_sigset)
{
	k_rtsigset_t block;

	if (ptrace(PTRACE_GETSIGMASK, restore_tid, sizeof(*save_sigset), save_sigset)) {
		pr_err("Failed to get current sigmask for restore tid %d\n", restore_tid);
		return -1;
	}

	ksigfillset(&block);
	ksigdelset(&block, SIGTRAP);

	if (ptrace(PTRACE_SETSIGMASK, restore_tid, sizeof(block), &block)) {
		pr_err("Failed to block signals on restore tid %d\n", restore_tid);
		return -1;
	}

	// Clear out PTRACE_O_SUSPEND_SECCOMP when we resume the restore thread
	if (ptrace(PTRACE_SETOPTIONS, restore_tid, NULL, 0)) {
		pr_err("Could not clear ptrace options on restore tid %d\n", restore_tid);
		return -1;
	}

	if (ptrace(PTRACE_CONT, restore_tid, NULL, 0)) {
		pr_err("Could not resume cuda restore tid %d\n", restore_tid);
		return -1;
	}

	return 0;
}

int cuda_plugin_checkpoint_devices(int pid)
{
	int restore_tid;
	char msg_buf[CUDA_CKPT_BUF_SIZE];
	int int_ret;
	int status;
	k_rtsigset_t save_sigset;

	if (plugin_disabled) {
		return -ENOTSUP;
	}

	restore_tid = get_cuda_restore_tid(pid);

	/* We can possibly hit a race with cuInit() where we are past the point of
	 * locking the process but at lock time cuInit() hadn't completed in which
	 * case cuda-checkpoint will report that we're in an invalid state to
	 * checkpoint
	 */
	if (restore_tid == -1) {
		pr_info("No need to checkpoint devices on pid %d\n", pid);
		return 0;
	}

	pr_info("Checkpointing CUDA devices on pid %d restore_tid %d\n", pid, restore_tid);
	/* We need to resume the checkpoint thread to prepare the mappings for
	 * checkpointing
	 */
	if (resume_restore_thread(restore_tid, &save_sigset)) {
		return -1;
	}
	status = cuda_process_checkpoint_action(pid, ACTION_CHECKPOINT, 0, msg_buf, sizeof(msg_buf));
	if (status) {
		pr_err("CHECKPOINT_DEVICES failed with %s\n", msg_buf);
		goto interrupt;
	}
	status = update_checkpointed_pid(&cuda_pids, pid);
	if (status) {
		pr_err("Failed to track checkpointed pid %d\n", pid);
		status = cuda_process_checkpoint_action(pid, ACTION_RESTORE, 0, msg_buf, sizeof(msg_buf));
		if (status) {
			pr_err("Failed to restore process after error %s on pid %d\n", msg_buf, pid);
		}
	}

	if (!status && !plugin_added_to_inventory) {
		status = add_inventory_plugin(CR_PLUGIN_DESC.name);
		if (status)
			pr_err("Failed to add CUDA plugin to inventory image\n");
		else
			plugin_added_to_inventory = true;
	}

interrupt:
	int_ret = interrupt_restore_thread(restore_tid, &save_sigset);

	return status != 0 ? status : int_ret;
}
CR_PLUGIN_REGISTER_HOOK(CR_PLUGIN_HOOK__CHECKPOINT_DEVICES, cuda_plugin_checkpoint_devices);

int cuda_plugin_pause_devices(int pid)
{
	int restore_tid;
	char msg_buf[CUDA_CKPT_BUF_SIZE];

	if (plugin_disabled) {
		return -ENOTSUP;
	}

	restore_tid = get_cuda_restore_tid(pid);

	if (restore_tid == -1) {
		pr_info("no need to pause devices on pid %d\n", pid);
		return 0;
	}

	pr_info("pausing devices on pid %d\n", pid);
	int status = cuda_process_checkpoint_action(pid, ACTION_LOCK, opts.timeout * 1000, msg_buf, sizeof(msg_buf));
	if (status) {
		pr_err("PAUSE_DEVICES failed with %s\n", msg_buf);
		if (alarm_timeouted())
			goto unlock;
		return -1;
	}

	if (add_pid_to_buf(&cuda_pids, pid)) {
		pr_err("unable to track paused pid %d\n", pid);
		goto unlock;
	}

	return 0;
unlock:
	status = cuda_process_checkpoint_action(pid, ACTION_UNLOCK, 0, msg_buf, sizeof(msg_buf));
	if (status) {
		pr_err("Failed to unlock process status %s, pid %d may hang\n", msg_buf, pid);
	}
	return -1;
}
CR_PLUGIN_REGISTER_HOOK(CR_PLUGIN_HOOK__PAUSE_DEVICES, cuda_plugin_pause_devices)

int resume_device(int pid, int checkpointed)
{
	char msg_buf[CUDA_CKPT_BUF_SIZE];
	int status;
	int ret = 0;
	int int_ret;
	k_rtsigset_t save_sigset;

	int restore_tid = get_cuda_restore_tid(pid);
	if (restore_tid == -1) {
		pr_info("No need to resume devices on pid %d\n", pid);
		return 0;
	}

	pr_info("resuming devices on pid %d\n", pid);
	/* The resuming process has to stay frozen during this time otherwise
	 * attempting to access a UVM pointer will crash if we haven't restored the
	 * underlying mappings yet
	 */
	pr_debug("Restore thread pid %d found for real pid %d\n", restore_tid, pid);
	/* wakeup the restore thread so we can handle the restore for this pid,
	 * rseq_cs has to be restored before execution
	 */
	if (resume_restore_thread(restore_tid, &save_sigset)) {
		return -1;
	}

	if (checkpointed) {
		status = cuda_process_checkpoint_action(pid, ACTION_RESTORE, 0, msg_buf, sizeof(msg_buf));
		if (status) {
			pr_err("RESUME_DEVICES RESTORE failed with %s\n", msg_buf);
			ret = -1;
			goto interrupt;
		}
	}

	status = cuda_process_checkpoint_action(pid, ACTION_UNLOCK, 0, msg_buf, sizeof(msg_buf));
	if (status) {
		pr_err("RESUME_DEVICES UNLOCK failed with %s\n", msg_buf);
		ret = -1;
	}

interrupt:
	int_ret = interrupt_restore_thread(restore_tid, &save_sigset);

	return ret != 0 ? ret : int_ret;
}

int cuda_plugin_resume_devices_late(int pid)
{
	if (plugin_disabled) {
		return -ENOTSUP;
	}

	return resume_device(pid, 1);
}
CR_PLUGIN_REGISTER_HOOK(CR_PLUGIN_HOOK__RESUME_DEVICES_LATE, cuda_plugin_resume_devices_late)

int cuda_plugin_init(int stage)
{
	int ret;

	if (stage == CR_PLUGIN_STAGE__RESTORE) {
		if (!check_and_remove_inventory_plugin(CR_PLUGIN_DESC.name, strlen(CR_PLUGIN_DESC.name))) {
			plugin_disabled = true;
			return 0;
		}
	}

	if (!fault_injected(FI_PLUGIN_CUDA_FORCE_ENABLE) && access("/dev/nvidiactl", F_OK)) {
		pr_info("/dev/nvidiactl doesn't exist. The CUDA plugin is disabled.\n");
		plugin_disabled = true;
		return 0;
	}

	ret = cuda_checkpoint_supports_flag("--action");
	if (ret == -1) {
		pr_warn("check that %s is present in $PATH\n", CUDA_CHECKPOINT);
		plugin_disabled = true;
		return 0;
	}

	if (ret == 0) {
		pr_warn("cuda-checkpoint --action flag not supported, an r555 or higher version driver is required. Disabling CUDA plugin\n");
		plugin_disabled = true;
		return 0;
	}

	pr_info("initialized: %s stage %d\n", CR_PLUGIN_DESC.name, stage);

	/* In the DUMP stage track all the PID's we've paused CUDA operations on to
	 * release them when we're done if the user requested the leave-running option
	 */
	if (stage == CR_PLUGIN_STAGE__DUMP) {
		INIT_LIST_HEAD(&cuda_pids);
	}

	dont_use_freeze_cgroup();

	return 0;
}

void cuda_plugin_fini(int stage, int ret)
{
	if (plugin_disabled) {
		return;
	}

	pr_info("finished %s stage %d err %d\n", CR_PLUGIN_DESC.name, stage, ret);

	/* Release all the paused PID's at the end of the DUMP stage in case the
	 * user provides the -R (leave-running) flag or an error occurred
	 */
	if (stage == CR_PLUGIN_STAGE__DUMP && (opts.final_state == TASK_ALIVE || ret != 0)) {
		struct pid_info *info;
		list_for_each_entry(info, &cuda_pids, list) {
			resume_device(info->pid, info->checkpointed);
		}
	}
	if (stage == CR_PLUGIN_STAGE__DUMP) {
		dealloc_pid_buffer(&cuda_pids);
	}
}
CR_PLUGIN_REGISTER("cuda_plugin", cuda_plugin_init, cuda_plugin_fini)