vmspawn: add --ephemeral (#40505)

This PR implements --ephemeral for vmspawn in the same way as nspawn.

Note that this does not support snapshots for --extra-drives.

man/systemd-vmspawn.xml

@@ -125,6 +125,17 @@
           <xi:include href="version-info.xml" xpointer="v260"/></listitem>
         </varlistentry>
 
+        <varlistentry>
+          <term><option>-x</option></term>
+          <term><option>--ephemeral</option></term>
+
+          <listitem><para>If specified, the VM is run with a temporary snapshot of its file system that is removed
+          immediately when the VM terminates. Only works with <option>--image=</option> currently.
+
+          Note that <option>--ephemeral</option> will not work with <option>--extra-drive=</option>.</para>
+
+          <xi:include href="version-info.xml" xpointer="v260"/></listitem>
+        </varlistentry>
       </variablelist>
     </refsect2>
 

src/nspawn/nspawn.c

@@ -107,6 +107,7 @@
 #include "shift-uid.h"
 #include "signal-util.h"
 #include "siphash24.h"
+#include "snapshot-util.h"
 #include "socket-util.h"
 #include "stat-util.h"
 #include "stdio-util.h"
@@ -5915,7 +5916,7 @@ static int do_cleanup(void) {
 }
 
 static int run(int argc, char *argv[]) {
-        bool remove_directory = false, remove_image = false, veth_created = false;
+        bool remove_image = false, veth_created = false;
         _cleanup_close_ int master = -EBADF, userns_fd = -EBADF, mount_fd = -EBADF;
         _cleanup_fdset_free_ FDSet *fds = NULL;
         int r, ret = EXIT_SUCCESS;
@@ -5923,6 +5924,7 @@ static int run(int argc, char *argv[]) {
         struct ExposeArgs expose_args = {};
         _cleanup_(release_lock_file) LockFile tree_global_lock = LOCK_FILE_INIT, tree_local_lock = LOCK_FILE_INIT;
         _cleanup_(rmdir_and_freep) char *rootdir = NULL;
+        _cleanup_(rm_rf_subvolume_and_freep) char *snapshot_dir = NULL;
         _cleanup_(loop_device_unrefp) LoopDevice *loop = NULL;
         _cleanup_(dissected_image_unrefp) DissectedImage *dissected_image = NULL;
         _cleanup_(sd_netlink_unrefp) sd_netlink *nfnl = NULL;
@@ -6064,63 +6066,27 @@ static int run(int argc, char *argv[]) {
                 }
 
                 if (arg_ephemeral) {
-                        _cleanup_free_ char *np = NULL;
-
                         r = chase_and_update(&arg_directory, 0);
                         if (r < 0)
                                 goto finish;
 
-                        /* If the specified path is a mount point we generate the new snapshot immediately
+                        r = create_ephemeral_snapshot(
-                         * inside it under a random name. However if the specified is not a mount point we
+                                        arg_directory,
-                         * create the new snapshot in the parent directory, just next to it. */
-                        r = path_is_mount_point(arg_directory);
-                        if (r < 0) {
-                                log_error_errno(r, "Failed to determine whether directory %s is mount point: %m", arg_directory);
-                                goto finish;
-                        }
-                        if (r > 0)
-                                r = tempfn_random_child(arg_directory, "machine.", &np);
-                        else
-                                r = tempfn_random(arg_directory, "machine.", &np);
-                        if (r < 0) {
-                                log_error_errno(r, "Failed to generate name for directory snapshot: %m");
-                                goto finish;
-                        }
-
-                        /* We take an exclusive lock on this image, since it's our private, ephemeral copy
-                         * only owned by us and no one else. */
-                        r = image_path_lock(
                                         arg_privileged ? RUNTIME_SCOPE_SYSTEM : RUNTIME_SCOPE_USER,
-                                        np,
+                                        arg_read_only,
-                                        LOCK_EX|LOCK_NB,
+                                        &tree_global_lock,
-                                        arg_privileged ? &tree_global_lock : NULL,
+                                        &tree_local_lock,
-                                        &tree_local_lock);
+                                        &snapshot_dir);
                         if (r < 0) {
-                                log_error_errno(r, "Failed to lock %s: %m", np);
+                                log_error_errno(r, "Failed to create ephemeral snapshot: %m");
                                 goto finish;
                         }
 
-                        {
+                        r = free_and_strdup(&arg_directory, snapshot_dir);
-                                BLOCK_SIGNALS(SIGINT);
-                                r = btrfs_subvol_snapshot_at(AT_FDCWD, arg_directory, AT_FDCWD, np,
-                                                             (arg_read_only ? BTRFS_SNAPSHOT_READ_ONLY : 0) |
-                                                             BTRFS_SNAPSHOT_FALLBACK_COPY |
-                                                             BTRFS_SNAPSHOT_FALLBACK_DIRECTORY |
-                                                             BTRFS_SNAPSHOT_RECURSIVE |
-                                                             BTRFS_SNAPSHOT_QUOTA |
-                                                             BTRFS_SNAPSHOT_SIGINT);
-                        }
-                        if (r == -EINTR) {
-                                log_error_errno(r, "Interrupted while copying file system tree to %s, removed again.", np);
-                                goto finish;
-                        }
                         if (r < 0) {
-                                log_error_errno(r, "Failed to create snapshot %s from %s: %m", np, arg_directory);
+                                log_oom();
                                 goto finish;
                         }
-
-                        free_and_replace(arg_directory, np);
-                        remove_directory = true;
                 } else {
                         r = chase_and_update(&arg_directory, arg_template ? CHASE_NONEXISTENT : 0);
                         if (r < 0)
@@ -6471,14 +6437,6 @@ finish:
 
         pager_close();
 
-        if (remove_directory && arg_directory) {
-                int k;
-
-                k = rm_rf(arg_directory, REMOVE_ROOT|REMOVE_PHYSICAL|REMOVE_SUBVOLUME);
-                if (k < 0)
-                        log_warning_errno(k, "Cannot remove '%s', ignoring: %m", arg_directory);
-        }
-
         if (remove_image && arg_image) {
                 if (unlink(arg_image) < 0)
                         log_warning_errno(errno, "Can't remove image file '%s', ignoring: %m", arg_image);

src/shared/meson.build

@@ -177,6 +177,7 @@ shared_sources = files(
         'sleep-config.c',
         'smack-util.c',
         'smbios11.c',
+        'snapshot-util.c',
         'socket-label.c',
         'socket-netlink.c',
         'specifier.c',

src/shared/snapshot-util.c

@@ -0,0 +1,65 @@
+/* SPDX-License-Identifier: LGPL-2.1-or-later */
+
+#include <sys/file.h>
+
+#include "alloc-util.h"
+#include "btrfs-util.h"
+#include "discover-image.h"
+#include "log.h"
+#include "mountpoint-util.h"
+#include "snapshot-util.h"
+#include "signal-util.h"
+#include "tmpfile-util.h"
+
+int create_ephemeral_snapshot(
+                const char *directory,
+                RuntimeScope scope,
+                bool read_only,
+                LockFile *tree_global_lock,
+                LockFile *tree_local_lock,
+                char **ret_new_path) {
+
+        _cleanup_free_ char *np = NULL;
+        int r;
+
+        /* If the specified path is a mount point we generate the new snapshot immediately
+         * inside it under a random name. However if the specified is not a mount point we
+         * create the new snapshot in the parent directory, just next to it. */
+        r = path_is_mount_point(directory);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to determine whether directory %s is mount point: %m", directory);
+        if (r > 0)
+                r = tempfn_random_child(directory, "snapshot.", &np);
+        else
+                r = tempfn_random(directory, "snapshot.", &np);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to generate name for directory snapshot: %m");
+
+        /* We take an exclusive lock on this image, since it's our private, ephemeral copy
+         * only owned by us and no one else. */
+        r = image_path_lock(
+                        scope,
+                        np,
+                        LOCK_EX|LOCK_NB,
+                        scope == RUNTIME_SCOPE_SYSTEM ? tree_global_lock : NULL,
+                        tree_local_lock);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to lock %s: %m", np);
+
+        {
+                BLOCK_SIGNALS(SIGINT);
+                r = btrfs_subvol_snapshot_at(AT_FDCWD, directory, AT_FDCWD, np,
+                                             (read_only ? BTRFS_SNAPSHOT_READ_ONLY : 0) |
+                                             BTRFS_SNAPSHOT_FALLBACK_COPY |
+                                             BTRFS_SNAPSHOT_FALLBACK_DIRECTORY |
+                                             BTRFS_SNAPSHOT_RECURSIVE |
+                                             BTRFS_SNAPSHOT_QUOTA |
+                                             BTRFS_SNAPSHOT_SIGINT);
+        }
+        if (r < 0)
+                return log_debug_errno(r, "Failed to create snapshot %s from %s: %m", np, directory);
+
+        *ret_new_path = TAKE_PTR(np);
+
+        return 0;
+}

src/shared/snapshot-util.h

@@ -0,0 +1,21 @@
+/* SPDX-License-Identifier: LGPL-2.1-or-later */
+#pragma once
+
+#include "runtime-scope.h"
+
+/* create_ephemeral_snapshot - create a snapshot of the given directory.
+ *
+ * It will use a btrfs snapshot when available with fallback to traditional dir copy. It will set the global
+ * and local lock files based on the passed runtime scope. On success the new directory path is returned via
+ * `ret_new_path`.
+ *
+ * The caller is responsible for the cleanup of the directory, using `_cleanup_(rm_rf_subvolume_and_freep)`
+ * is recommended.
+ */
+int create_ephemeral_snapshot(
+                const char *directory,
+                RuntimeScope scope,
+                bool read_only,
+                LockFile *tree_global_lock,
+                LockFile *tree_local_lock,
+                char **ret_new_path);

src/vmspawn/vmspawn-settings.h

@@ -35,6 +35,7 @@ typedef enum SettingsMask {
         SETTING_START_MODE        = UINT64_C(1) << 0,
         SETTING_MACHINE_ID        = UINT64_C(1) << 6,
         SETTING_BIND_MOUNTS       = UINT64_C(1) << 11,
+        SETTING_EPHEMERAL         = UINT64_C(1) << 24,
         SETTING_DIRECTORY         = UINT64_C(1) << 26,
         SETTING_CREDENTIALS       = UINT64_C(1) << 30,
         _SETTING_FORCE_ENUM_WIDTH = UINT64_MAX

src/vmspawn/vmspawn.c

@@ -62,6 +62,7 @@
 #include "random-util.h"
 #include "rm-rf.h"
 #include "signal-util.h"
+#include "snapshot-util.h"
 #include "socket-util.h"
 #include "stat-util.h"
 #include "stdio-util.h"
@@ -145,6 +146,7 @@ static char **arg_bind_user = NULL;
 static char *arg_bind_user_shell = NULL;
 static bool arg_bind_user_shell_copy = false;
 static char **arg_bind_user_groups = NULL;
+static bool arg_ephemeral = false;
 static RuntimeScope arg_runtime_scope = _RUNTIME_SCOPE_INVALID;
 
 STATIC_DESTRUCTOR_REGISTER(arg_directory, freep);
@@ -190,6 +192,7 @@ static int help(void) {
                "     --system              Interact with system manager\n"
                "\n%3$sImage:%4$s\n"
                "  -D --directory=PATH      Root directory for the VM\n"
+               "  -x --ephemeral           Run VM with snapshot of the disk or directory\n"
                "  -i --image=FILE|DEVICE   Root file system disk image or device for the VM\n"
                "     --image-format=FORMAT Specify disk image format (raw, qcow2; default: raw)\n"
                "\n%3$sHost Configuration:%4$s\n"
@@ -327,6 +330,7 @@ static int parse_argv(int argc, char *argv[]) {
                 { "no-pager",          no_argument,       NULL, ARG_NO_PAGER          },
                 { "image",             required_argument, NULL, 'i'                   },
                 { "image-format",      required_argument, NULL, ARG_IMAGE_FORMAT      },
+                { "ephemeral",         no_argument,       NULL, 'x'                   },
                 { "directory",         required_argument, NULL, 'D'                   },
                 { "machine",           required_argument, NULL, 'M'                   },
                 { "slice",             required_argument, NULL, 'S'                   },
@@ -382,7 +386,7 @@ static int parse_argv(int argc, char *argv[]) {
         assert(argv);
 
         optind = 0;
-        while ((c = getopt_long(argc, argv, "+hD:i:M:nqs:G:S:", options, NULL)) >= 0)
+        while ((c = getopt_long(argc, argv, "+hD:i:xM:nqs:G:S:", options, NULL)) >= 0)
                 switch (c) {
                 case 'h':
                         return help();
@@ -429,6 +433,10 @@ static int parse_argv(int argc, char *argv[]) {
                         }
                         break;
 
+                case 'x':
+                        arg_ephemeral = true;
+                        break;
+
                 case ARG_NO_PAGER:
                         arg_pager_flags |= PAGER_DISABLE;
                         break;
@@ -794,6 +802,9 @@ static int parse_argv(int argc, char *argv[]) {
         if (!strv_isempty(arg_bind_user_groups) && strv_isempty(arg_bind_user))
                 return log_error_errno(SYNTHETIC_ERRNO(EINVAL), "Cannot use --bind-user-group= without --bind-user=");
 
+        if (arg_ephemeral && arg_extra_drives.n_drives > 0)
+                return log_error_errno(SYNTHETIC_ERRNO(EINVAL), "Cannot use --ephemeral with --extra-drive=");
+
         if (argc > optind) {
                 arg_kernel_cmdline_extra = strv_copy(argv + optind);
                 if (!arg_kernel_cmdline_extra)
@@ -1843,6 +1854,8 @@ static int run_virtual_machine(int kvm_device_fd, int vhost_device_fd) {
         _cleanup_(ovmf_config_freep) OvmfConfig *ovmf_config = NULL;
         _cleanup_free_ char *qemu_binary = NULL, *mem = NULL, *kernel = NULL;
         _cleanup_(rm_rf_physical_and_freep) char *ssh_private_key_path = NULL, *ssh_public_key_path = NULL;
+        _cleanup_(rm_rf_subvolume_and_freep) char *snapshot_directory = NULL;
+        _cleanup_(release_lock_file) LockFile tree_global_lock = LOCK_FILE_INIT, tree_local_lock = LOCK_FILE_INIT;
         _cleanup_close_ int notify_sock_fd = -EBADF;
         _cleanup_strv_free_ char **cmdline = NULL;
         _cleanup_free_ int *pass_fds = NULL;
@@ -2319,7 +2332,8 @@ static int run_virtual_machine(int kvm_device_fd, int vhost_device_fd) {
                 if (!escaped_image)
                         return log_oom();
 
-                if (strv_extendf(&cmdline, "if=none,id=vmspawn,file=%s,format=%s,discard=%s", escaped_image, image_format_to_string(arg_image_format), on_off(arg_discard_disk)) < 0)
+                if (strv_extendf(&cmdline, "if=none,id=vmspawn,file=%s,format=%s,discard=%s,snapshot=%s",
+                                 escaped_image, image_format_to_string(arg_image_format), on_off(arg_discard_disk), on_off(arg_ephemeral)) < 0)
                         return log_oom();
 
                 _cleanup_free_ char *image_fn = NULL;
@@ -2366,6 +2380,21 @@ static int run_virtual_machine(int kvm_device_fd, int vhost_device_fd) {
                 if (!GREEDY_REALLOC(children, n_children + 1))
                         return log_oom();
 
+                if (arg_ephemeral) {
+                        r = create_ephemeral_snapshot(arg_directory,
+                                                      arg_runtime_scope,
+                                                      /* read-only */ false,
+                                                      &tree_global_lock,
+                                                      &tree_local_lock,
+                                                      &snapshot_directory);
+                        if (r < 0)
+                                return r;
+
+                        arg_directory = strdup(snapshot_directory);
+                        if (!arg_directory)
+                                return log_oom();
+                }
+
                 r = start_virtiofsd(
                                 unit,
                                 arg_directory,
@@ -3065,6 +3094,11 @@ static int determine_names(void) {
                         if (r < 0)
                                 return log_error_errno(r, "Failed to extract file name from '%s': %m", arg_directory);
                 }
+                /* Add a random suffix when this is an ephemeral machine, so that we can run many
+                 * instances at once without manually having to specify -M each time. */
+                if (arg_ephemeral)
+                        if (strextendf(&arg_machine, "-%016" PRIx64, random_u64()) < 0)
+                                return log_oom();
 
                 hostname_cleanup(arg_machine);
                 if (!hostname_is_valid(arg_machine, 0))