Merge pull request torvalds#130 from pscollins/free-all-mem

Free all memory allocated by LKL (except for undetached pthreads)

arch/lkl/include/uapi/asm/host_ops.h

@@ -36,6 +36,8 @@ typedef unsigned long lkl_thread_t;
  * @thread_detach - on POSIX systems, free up resources held by
  * pthreads. Noop on Win32.
  * @thread_exit - terminates the current thread
+ * @thread_join - wait for the given thread to terminate. Returns 0
+ * for success, -1 otherwise
  *
  * @tls_alloc - allocate a thread local storage key; returns 0 if succesful
  * @tls_free - frees a thread local storage key; returns 0 if succesful
@@ -58,7 +60,8 @@ typedef unsigned long lkl_thread_t;
  * @iomem_acess - reads or writes to and I/O memory region; addr must be in the
  * range returned by ioremap
  *
- * @gettid - returns the host thread id of the caller
+ * @gettid - returns the host thread id of the caller, which need not
+ * be the same as the handle returned by thread_create
  */
 struct lkl_host_operations {
 	const char *virtio_devices;
@@ -80,6 +83,7 @@ struct lkl_host_operations {
 	lkl_thread_t (*thread_create)(void (*f)(void *), void *arg);
 	void (*thread_detach)(void);
 	void (*thread_exit)(void);
+	int (*thread_join)(lkl_thread_t tid);
 
 	int (*tls_alloc)(unsigned int *key);
 	int (*tls_free)(unsigned int key);
@@ -113,7 +117,7 @@ struct lkl_host_operations {
  * generate the Linux kernel command line
  */
 int lkl_start_kernel(struct lkl_host_operations *lkl_ops,
-		     unsigned long mem_size,
-		     const char *cmd_line, ...);
+		    unsigned long mem_size,
+		    const char *cmd_line, ...);
 
 #endif

arch/lkl/kernel/setup.c

@@ -50,8 +50,8 @@ static void __init lkl_run_kernel(void *arg)
 }
 
 int __init lkl_start_kernel(struct lkl_host_operations *ops,
-			    unsigned long _mem_size,
-			    const char *fmt, ...)
+			unsigned long _mem_size,
+			const char *fmt, ...)
 {
 	va_list ap;
 	int ret;
@@ -117,6 +117,8 @@ void machine_restart(char *unused)
 	machine_halt();
 }
 
+extern int lkl_netdevs_remove(void);
+
 long lkl_sys_halt(void)
 {
 	long err;
@@ -139,6 +141,10 @@ long lkl_sys_halt(void)
 	lkl_ops->sem_free(init_sem);
 
 	free_initial_syscall_thread();
+	if (lkl_netdevs_remove() == 0)
+		/* We know that there is nothing else touching our
+		 * memory. */
+		free_mem();
 
 	return 0;
 }
@@ -147,10 +153,6 @@ void arch_cpu_idle(void)
 {
 	if (halt) {
 		threads_cleanup();
-		/* TODO(pscollins): If we free here, it causes a
-		 * segfault because the tx/rx threads are still
-		 * running in parallel. */
-		/* free_mem(); */
 
 		/* Shutdown the clockevents source. */
 		tick_suspend_local();

tools/lkl/include/lkl.h

@@ -209,12 +209,9 @@ int lkl_if_set_ipv4(int ifindex, unsigned int addr, unsigned int netmask_len);
 int lkl_set_ipv4_gateway(unsigned int addr);
 
 /**
- * lkl_netdev - host network device handle
+ * lkl_netdev - host network device handle, defined in lkl_host.h.
  */
-struct lkl_netdev {
-	struct lkl_dev_net_ops *ops;
-};
-
+struct lkl_netdev;
 
 /**
  * lkl_netdev_add - add a new network device
@@ -228,6 +225,17 @@ struct lkl_netdev {
  */
 int lkl_netdev_add(struct lkl_netdev *nd, void *mac);
 
+/**
+* lkl_netdevs_remove - destroy all network devices
+*
+* Attempts to release all resources held by network devices created
+* via lkl_netdev_add.
+*
+* @returns 0 if all devices are successfully removed, -1 if at least
+* one fails.
+*/
+int lkl_netdevs_remove(void);
+
 /**
  * lkl_netdev_get_ifindex - retrieve the interface index for a given network
  * device id

tools/lkl/include/lkl_host.h

@@ -47,12 +47,21 @@ struct lkl_dev_blk_ops {
 	int (*request)(union lkl_disk disk, struct lkl_blk_req *req);
 };
 
+struct lkl_netdev {
+	struct lkl_dev_net_ops *ops;
+	lkl_thread_t rx_tid, tx_tid;
+};
+
 struct lkl_dev_net_ops {
 	int (*tx)(struct lkl_netdev *nd, void *data, int len);
 	int (*rx)(struct lkl_netdev *nd, void *data, int *len);
 #define LKL_DEV_NET_POLL_RX		1
 #define LKL_DEV_NET_POLL_TX		2
 	int (*poll)(struct lkl_netdev *nd, int events);
+	/* Release all resources acquired --- in particular, kill the
+	 * polling threads and close any open handles. Not implemented
+	 * by all netdev types. 0 for success, -1 for failure. */
+	int (*close)(struct lkl_netdev *nd);
 };
 
 #ifdef __cplusplus

tools/lkl/lib/hijack/init.c

@@ -1,5 +1,3 @@
-
-
 /*
  * system calls hijack code
  * Copyright (c) 2015 Hajime Tazaki
@@ -27,6 +25,10 @@
 #include <lkl_host.h>
 
 #include "xlate.h"
+#include "../virtio_net_tap.h"
+
+#define __USE_GNU
+#include <dlfcn.h>
 
 /* Mount points are named after filesystem types so they should never
  * be longer than ~6 characters. */
@@ -139,6 +141,13 @@ static void mount_cmds_exec(char *_cmds, int (*callback)(char*))
 	free(cmds);
 }
 
+void fixup_netdev_tap_ops(void)
+{
+	/* It's okay if this is NULL, because then netdev close will
+	 * fall back onto an uncloseable implementation. */
+	lkl_netdev_tap_ops.eventfd = dlsym(RTLD_NEXT, "eventfd");
+}
+
 void __attribute__((constructor(102)))
 hijack_init(void)
 {
@@ -157,6 +166,9 @@ hijack_init(void)
 	char *mount = getenv("LKL_HIJACK_MOUNT");
 	struct lkl_netdev *nd = NULL;
 
+	/* Must be run before lkl_netdev_tap_create */
+	fixup_netdev_tap_ops();
+
 	if (tap) {
 		fprintf(stderr,
 			"WARN: variable LKL_HIJACK_NET_TAP is now obsoleted.\n"

tools/lkl/lib/nt-host.c

@@ -67,7 +67,7 @@ static lkl_thread_t thread_create(void (*fn)(void *), void *arg)
 {
 	DWORD WINAPI (*win_fn)(LPVOID arg) = (DWORD WINAPI (*)(LPVOID))fn;
 
-	return CreateThread(NULL, 0, win_fn, arg, 0, NULL);
+	return (lkl_thread_t)CreateThread(NULL, 0, win_fn, arg, 0, NULL);
 }
 
 static void thread_detach(void)
@@ -79,6 +79,13 @@ static void thread_exit(void)
 	ExitThread(0);
 }
 
+static int thread_join(lkl_thread_t tid)
+{
+	/* TODO: error handling */
+	WaitForSingleObject(tid, INFINITE);
+	return 0;
+}
+
 static int tls_alloc(unsigned int *key)
 {
 	*key = TlsAlloc();
@@ -202,6 +209,7 @@ struct lkl_host_operations lkl_host_ops = {
 	.thread_create = thread_create,
 	.thread_detach = thread_detach,
 	.thread_exit = thread_exit,
+	.thread_join = thread_join,
 	.sem_alloc = sem_alloc,
 	.sem_free = sem_free,
 	.sem_up = sem_up,
@@ -301,3 +309,10 @@ struct lkl_dev_blk_ops lkl_dev_blk_ops = {
 	.get_capacity = handle_get_capacity,
 	.request = blk_request,
 };
+
+/* Needed to resolve linker error on Win32. We don't really support
+ * any network IO on Windows, anyway, so there's no loss here. */
+int lkl_netdevs_remove(void)
+{
+	return -1;
+}

tools/lkl/lib/posix-host.c

@@ -199,6 +199,14 @@ static void thread_exit(void)
 	pthread_exit(NULL);
 }
 
+static int thread_join(lkl_thread_t tid)
+{
+	if (WARN_PTHREAD(pthread_join(tid, NULL)))
+		return -1;
+	else
+		return 0;
+}
+
 static int tls_alloc(unsigned int *key)
 {
 	return pthread_key_create((pthread_key_t*)key, NULL);
@@ -219,7 +227,6 @@ static void *tls_get(unsigned int key)
 	return pthread_getspecific(key);
 }
 
-
 static unsigned long long time_ns(void)
 {
 	struct timeval tv;
@@ -286,6 +293,7 @@ struct lkl_host_operations lkl_host_ops = {
 	.thread_create = thread_create,
 	.thread_detach = thread_detach,
 	.thread_exit = thread_exit,
+	.thread_join = thread_join,
 	.sem_alloc = sem_alloc,
 	.sem_free = sem_free,
 	.sem_up = sem_up,

tools/lkl/lib/virtio_net.c

@@ -14,6 +14,10 @@
 #define NUM_QUEUES (TX_QUEUE_IDX + 1)
 #define QUEUE_DEPTH 32
 
+/* In fact, we'll hit the limit on the devs string below long before
+ * we hit this, but it's good enough for now. */
+#define MAX_NET_DEVS 16
+
 #ifdef DEBUG
 #define bad_request(s) do {			\
 		lkl_printf("%s\n", s);		\
@@ -23,7 +27,6 @@
 #define bad_request(s) lkl_printf("virtio_net: %s\n", s);
 #endif /* DEBUG */
 
-
 struct virtio_net_poll {
 	struct virtio_net_dev *dev;
 	int event;
@@ -124,6 +127,22 @@ void poll_thread(void *arg)
 	}
 }
 
+struct virtio_net_dev *registered_devs[MAX_NET_DEVS];
+static int registered_dev_idx = 0;
+
+static int dev_register(struct virtio_net_dev *dev)
+{
+	if (registered_dev_idx == MAX_NET_DEVS) {
+		lkl_printf("Too many virtio_net devices!\n");
+		/* This error code is a little bit of a lie */
+		return -LKL_ENOMEM;
+	} else {
+		/* registered_dev_idx is incremented by the caller */
+		registered_devs[registered_dev_idx] = dev;
+		return 0;
+	}
+}
+
 static void free_queue_locks(struct lkl_mutex_t **queues, int num_queues)
 {
 	int i = 0;
@@ -158,7 +177,6 @@ static struct lkl_mutex_t **init_queue_locks(int num_queues)
 int lkl_netdev_add(struct lkl_netdev *nd, void *mac)
 {
 	struct virtio_net_dev *dev;
-	static int count;
 	int ret = -LKL_ENOMEM;
 
 	dev = lkl_host_ops.mem_alloc(sizeof(*dev));
@@ -198,15 +216,19 @@ int lkl_netdev_add(struct lkl_netdev *nd, void *mac)
 	if (ret)
 		goto out_free;
 
-	if (lkl_host_ops.thread_create(poll_thread, &dev->rx_poll) == 0)
+	nd->rx_tid = lkl_host_ops.thread_create(poll_thread, &dev->rx_poll);
+	if (nd->rx_tid == 0)
 		goto out_cleanup_dev;
 
-	if (lkl_host_ops.thread_create(poll_thread, &dev->tx_poll) == 0)
+	nd->tx_tid = lkl_host_ops.thread_create(poll_thread, &dev->tx_poll);
+	if (nd->tx_tid == 0)
 		goto out_cleanup_dev;
 
-	/* RX/TX thread polls will exit when the host netdev handle is closed */
+	ret = dev_register(dev);
+	if (ret < 0)
+		goto out_cleanup_dev;
 
-	return count++;
+	return registered_dev_idx++;
 
 out_cleanup_dev:
 	virtio_dev_cleanup(&dev->dev);
@@ -218,3 +240,40 @@ int lkl_netdev_add(struct lkl_netdev *nd, void *mac)
 
 	return ret;
 }
+
+/* Return 0 for success, -1 for failure. */
+static int lkl_netdev_remove(struct virtio_net_dev *dev)
+{
+	if (!dev->nd->ops->close)
+		/* Can't kill the poll threads, so we can't do
+		 * anything safely. */
+		return -1;
+
+	if (dev->nd->ops->close(dev->nd) < 0)
+		/* Something went wrong */
+		return -1;
+
+	virtio_dev_cleanup(&dev->dev);
+
+	lkl_host_ops.mem_free(dev->nd);
+	free_queue_locks(dev->queue_locks, NUM_QUEUES);
+	lkl_host_ops.mem_free(dev);
+
+	return 0;
+}
+
+int lkl_netdevs_remove(void)
+{
+	int i = 0, failure_count = 0;
+
+	for (; i < registered_dev_idx; i++)
+		failure_count -= lkl_netdev_remove(registered_devs[i]);
+
+	if (failure_count) {
+		lkl_printf("WARN: failed to free %d of %d netdevs.\n",
+			failure_count, registered_dev_idx);
+		return -1;
+	}
+
+	return 0;
+}