src/core/execute.h

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2 #pragma once
   3
   4 typedef struct ExecStatus ExecStatus;
   5 typedef struct ExecCommand ExecCommand;
   6 typedef struct ExecContext ExecContext;
   7 typedef struct ExecRuntime ExecRuntime;
   8 typedef struct ExecParameters ExecParameters;
   9 typedef struct Manager Manager;
  10
  11 #include <sched.h>
  12 #include <stdbool.h>
  13 #include <stdio.h>
  14 #include <sys/capability.h>
  15
  16 #include "cgroup-util.h"
  17 #include "coredump-util.h"
  18 #include "cpu-set-util.h"
  19 #include "exec-util.h"
  20 #include "fdset.h"
  21 #include "list.h"
  22 #include "missing_resource.h"
  23 #include "namespace.h"
  24 #include "nsflags.h"
  25 #include "numa-util.h"
  26 #include "path-util.h"
  27 #include "time-util.h"
  28
  29 #define EXEC_STDIN_DATA_MAX (64U*1024U*1024U)
  30
  31 typedef enum ExecUtmpMode {
  32         EXEC_UTMP_INIT,
  33         EXEC_UTMP_LOGIN,
  34         EXEC_UTMP_USER,
  35         _EXEC_UTMP_MODE_MAX,
  36         _EXEC_UTMP_MODE_INVALID = -EINVAL,
  37 } ExecUtmpMode;
  38
  39 typedef enum ExecInput {
  40         EXEC_INPUT_NULL,
  41         EXEC_INPUT_TTY,
  42         EXEC_INPUT_TTY_FORCE,
  43         EXEC_INPUT_TTY_FAIL,
  44         EXEC_INPUT_SOCKET,
  45         EXEC_INPUT_NAMED_FD,
  46         EXEC_INPUT_DATA,
  47         EXEC_INPUT_FILE,
  48         _EXEC_INPUT_MAX,
  49         _EXEC_INPUT_INVALID = -EINVAL,
  50 } ExecInput;
  51
  52 typedef enum ExecOutput {
  53         EXEC_OUTPUT_INHERIT,
  54         EXEC_OUTPUT_NULL,
  55         EXEC_OUTPUT_TTY,
  56         EXEC_OUTPUT_KMSG,
  57         EXEC_OUTPUT_KMSG_AND_CONSOLE,
  58         EXEC_OUTPUT_JOURNAL,
  59         EXEC_OUTPUT_JOURNAL_AND_CONSOLE,
  60         EXEC_OUTPUT_SOCKET,
  61         EXEC_OUTPUT_NAMED_FD,
  62         EXEC_OUTPUT_FILE,
  63         EXEC_OUTPUT_FILE_APPEND,
  64         EXEC_OUTPUT_FILE_TRUNCATE,
  65         _EXEC_OUTPUT_MAX,
  66         _EXEC_OUTPUT_INVALID = -EINVAL,
  67 } ExecOutput;
  68
  69 typedef enum ExecPreserveMode {
  70         EXEC_PRESERVE_NO,
  71         EXEC_PRESERVE_YES,
  72         EXEC_PRESERVE_RESTART,
  73         _EXEC_PRESERVE_MODE_MAX,
  74         _EXEC_PRESERVE_MODE_INVALID = -EINVAL,
  75 } ExecPreserveMode;
  76
  77 typedef enum ExecKeyringMode {
  78         EXEC_KEYRING_INHERIT,
  79         EXEC_KEYRING_PRIVATE,
  80         EXEC_KEYRING_SHARED,
  81         _EXEC_KEYRING_MODE_MAX,
  82         _EXEC_KEYRING_MODE_INVALID = -EINVAL,
  83 } ExecKeyringMode;
  84
  85 /* Contains start and exit information about an executed command.  */
  86 struct ExecStatus {
  87         dual_timestamp start_timestamp;
  88         dual_timestamp exit_timestamp;
  89         pid_t pid;
  90         int code;     /* as in siginfo_t::si_code */
  91         int status;   /* as in siginfo_t::si_status */
  92 };
  93
  94 /* Stores information about commands we execute. Covers both configuration settings as well as runtime data. */
  95 struct ExecCommand {
  96         char *path;
  97         char **argv;
  98         ExecStatus exec_status;
  99         ExecCommandFlags flags;
 100         LIST_FIELDS(ExecCommand, command); /* useful for chaining commands */
 101 };
 102
 103 /* Encapsulates certain aspects of the runtime environment that is to be shared between multiple otherwise separate
 104  * invocations of commands. Specifically, this allows sharing of /tmp and /var/tmp data as well as network namespaces
 105  * between invocations of commands. This is a reference counted object, with one reference taken by each currently
 106  * active command invocation that wants to share this runtime. */
 107 struct ExecRuntime {
 108         unsigned n_ref;
 109
 110         Manager *manager;
 111
 112         char *id; /* Unit id of the owner */
 113
 114         char *tmp_dir;
 115         char *var_tmp_dir;
 116
 117         /* An AF_UNIX socket pair, that contains a datagram containing a file descriptor referring to the network
 118          * namespace. */
 119         int netns_storage_socket[2];
 120
 121         /* Like netns_storage_socket, but the file descriptor is referring to the IPC namespace. */
 122         int ipcns_storage_socket[2];
 123 };
 124
 125 typedef enum ExecDirectoryType {
 126         EXEC_DIRECTORY_RUNTIME = 0,
 127         EXEC_DIRECTORY_STATE,
 128         EXEC_DIRECTORY_CACHE,
 129         EXEC_DIRECTORY_LOGS,
 130         EXEC_DIRECTORY_CONFIGURATION,
 131         _EXEC_DIRECTORY_TYPE_MAX,
 132         _EXEC_DIRECTORY_TYPE_INVALID = -EINVAL,
 133 } ExecDirectoryType;
 134
 135 typedef struct ExecDirectoryItem {
 136         char *path;
 137         char **symlinks;
 138         bool only_create;
 139 } ExecDirectoryItem;
 140
 141 typedef struct ExecDirectory {
 142         mode_t mode;
 143         size_t n_items;
 144         ExecDirectoryItem *items;
 145 } ExecDirectory;
 146
 147 typedef enum ExecCleanMask {
 148         /* In case you wonder why the bitmask below doesn't use "directory" in its name: we want to keep this
 149          * generic so that .timer timestamp files can nicely be covered by this too, and similar. */
 150         EXEC_CLEAN_RUNTIME       = 1U << EXEC_DIRECTORY_RUNTIME,
 151         EXEC_CLEAN_STATE         = 1U << EXEC_DIRECTORY_STATE,
 152         EXEC_CLEAN_CACHE         = 1U << EXEC_DIRECTORY_CACHE,
 153         EXEC_CLEAN_LOGS          = 1U << EXEC_DIRECTORY_LOGS,
 154         EXEC_CLEAN_CONFIGURATION = 1U << EXEC_DIRECTORY_CONFIGURATION,
 155         EXEC_CLEAN_NONE          = 0,
 156         EXEC_CLEAN_ALL           = (1U << _EXEC_DIRECTORY_TYPE_MAX) - 1,
 157         _EXEC_CLEAN_MASK_INVALID = -EINVAL,
 158 } ExecCleanMask;
 159
 160 /* A credential configured with LoadCredential= */
 161 typedef struct ExecLoadCredential {
 162         char *id, *path;
 163         bool encrypted;
 164 } ExecLoadCredential;
 165
 166 /* A credential configured with SetCredential= */
 167 typedef struct ExecSetCredential {
 168         char *id;
 169         bool encrypted;
 170         void *data;
 171         size_t size;
 172 } ExecSetCredential;
 173
 174 /* Encodes configuration parameters applied to invoked commands. Does not carry runtime data, but only configuration
 175  * changes sourced from unit files and suchlike. ExecContext objects are usually embedded into Unit objects, and do not
 176  * change after being loaded. */
 177 struct ExecContext {
 178         char **environment;
 179         char **environment_files;
 180         char **pass_environment;
 181         char **unset_environment;
 182
 183         struct rlimit *rlimit[_RLIMIT_MAX];
 184         char *working_directory, *root_directory, *root_image, *root_verity, *root_hash_path, *root_hash_sig_path;
 185         void *root_hash, *root_hash_sig;
 186         size_t root_hash_size, root_hash_sig_size;
 187         LIST_HEAD(MountOptions, root_image_options);
 188         bool working_directory_missing_ok:1;
 189         bool working_directory_home:1;
 190
 191         bool oom_score_adjust_set:1;
 192         bool coredump_filter_set:1;
 193         bool nice_set:1;
 194         bool ioprio_set:1;
 195         bool cpu_sched_set:1;
 196         bool mount_apivfs_set:1;
 197
 198         /* This is not exposed to the user but available internally. We need it to make sure that whenever we
 199          * spawn /usr/bin/mount it is run in the same process group as us so that the autofs logic detects
 200          * that it belongs to us and we don't enter a trigger loop. */
 201         bool same_pgrp;
 202
 203         bool cpu_sched_reset_on_fork;
 204         bool non_blocking;
 205
 206         mode_t umask;
 207         int oom_score_adjust;
 208         int nice;
 209         int ioprio;
 210         int cpu_sched_policy;
 211         int cpu_sched_priority;
 212         uint64_t coredump_filter;
 213
 214         CPUSet cpu_set;
 215         NUMAPolicy numa_policy;
 216         bool cpu_affinity_from_numa;
 217
 218         ExecInput std_input;
 219         ExecOutput std_output;
 220         ExecOutput std_error;
 221         bool stdio_as_fds;
 222         char *stdio_fdname[3];
 223         char *stdio_file[3];
 224
 225         void *stdin_data;
 226         size_t stdin_data_size;
 227
 228         nsec_t timer_slack_nsec;
 229
 230         char *tty_path;
 231
 232         bool tty_reset;
 233         bool tty_vhangup;
 234         bool tty_vt_disallocate;
 235
 236         unsigned tty_rows;
 237         unsigned tty_cols;
 238
 239         bool ignore_sigpipe;
 240
 241         ExecKeyringMode keyring_mode;
 242
 243         /* Since resolving these names might involve socket
 244          * connections and we don't want to deadlock ourselves these
 245          * names are resolved on execution only and in the child
 246          * process. */
 247         char *user;
 248         char *group;
 249         char **supplementary_groups;
 250
 251         char *pam_name;
 252
 253         char *utmp_id;
 254         ExecUtmpMode utmp_mode;
 255
 256         bool no_new_privileges;
 257
 258         bool selinux_context_ignore;
 259         bool apparmor_profile_ignore;
 260         bool smack_process_label_ignore;
 261
 262         char *selinux_context;
 263         char *apparmor_profile;
 264         char *smack_process_label;
 265
 266         char **read_write_paths, **read_only_paths, **inaccessible_paths, **exec_paths, **no_exec_paths;
 267         char **exec_search_path;
 268         unsigned long mount_flags;
 269         BindMount *bind_mounts;
 270         size_t n_bind_mounts;
 271         TemporaryFileSystem *temporary_filesystems;
 272         size_t n_temporary_filesystems;
 273         MountImage *mount_images;
 274         size_t n_mount_images;
 275         MountImage *extension_images;
 276         size_t n_extension_images;
 277         char **extension_directories;
 278
 279         uint64_t capability_bounding_set;
 280         uint64_t capability_ambient_set;
 281         int secure_bits;
 282
 283         int syslog_priority;
 284         bool syslog_level_prefix;
 285         char *syslog_identifier;
 286
 287         struct iovec* log_extra_fields;
 288         size_t n_log_extra_fields;
 289
 290         usec_t log_ratelimit_interval_usec;
 291         unsigned log_ratelimit_burst;
 292
 293         int log_level_max;
 294
 295         char *log_namespace;
 296
 297         ProtectProc protect_proc;  /* hidepid= */
 298         ProcSubset proc_subset;    /* subset= */
 299
 300         bool private_tmp;
 301         bool private_network;
 302         bool private_devices;
 303         bool private_users;
 304         bool private_mounts;
 305         bool private_ipc;
 306         bool protect_kernel_tunables;
 307         bool protect_kernel_modules;
 308         bool protect_kernel_logs;
 309         bool protect_clock;
 310         bool protect_control_groups;
 311         ProtectSystem protect_system;
 312         ProtectHome protect_home;
 313         bool protect_hostname;
 314         bool mount_apivfs;
 315
 316         bool dynamic_user;
 317         bool remove_ipc;
 318
 319         bool memory_deny_write_execute;
 320         bool restrict_realtime;
 321         bool restrict_suid_sgid;
 322
 323         bool lock_personality;
 324         unsigned long personality;
 325
 326         unsigned long restrict_namespaces; /* The CLONE_NEWxyz flags permitted to the unit's processes */
 327
 328         Set *restrict_filesystems;
 329         bool restrict_filesystems_allow_list:1;
 330
 331         Hashmap *syscall_filter;
 332         Set *syscall_archs;
 333         int syscall_errno;
 334         bool syscall_allow_list:1;
 335
 336         Hashmap *syscall_log;
 337         bool syscall_log_allow_list:1; /* Log listed system calls */
 338
 339         bool address_families_allow_list:1;
 340         Set *address_families;
 341
 342         char *network_namespace_path;
 343         char *ipc_namespace_path;
 344
 345         ExecDirectory directories[_EXEC_DIRECTORY_TYPE_MAX];
 346         ExecPreserveMode runtime_directory_preserve_mode;
 347         usec_t timeout_clean_usec;
 348
 349         Hashmap *set_credentials; /* output id → ExecSetCredential */
 350         Hashmap *load_credentials; /* output id → ExecLoadCredential */
 351 };
 352
 353 static inline bool exec_context_restrict_namespaces_set(const ExecContext *c) {
 354         assert(c);
 355
 356         return (c->restrict_namespaces & NAMESPACE_FLAGS_ALL) != NAMESPACE_FLAGS_ALL;
 357 }
 358
 359 static inline bool exec_context_restrict_filesystems_set(const ExecContext *c) {
 360         assert(c);
 361
 362         return c->restrict_filesystems_allow_list ||
 363           !set_isempty(c->restrict_filesystems);
 364 }
 365
 366 static inline bool exec_context_with_rootfs(const ExecContext *c) {
 367         assert(c);
 368
 369         /* Checks if RootDirectory= or RootImage= are used */
 370
 371         return !empty_or_root(c->root_directory) || c->root_image;
 372 }
 373
 374 typedef enum ExecFlags {
 375         EXEC_APPLY_SANDBOXING      = 1 << 0,
 376         EXEC_APPLY_CHROOT          = 1 << 1,
 377         EXEC_APPLY_TTY_STDIN       = 1 << 2,
 378         EXEC_PASS_LOG_UNIT         = 1 << 3, /* Whether to pass the unit name to the service's journal stream connection */
 379         EXEC_CHOWN_DIRECTORIES     = 1 << 4, /* chown() the runtime/state/cache/log directories to the user we run as, under all conditions */
 380         EXEC_NSS_DYNAMIC_BYPASS    = 1 << 5, /* Set the SYSTEMD_NSS_DYNAMIC_BYPASS environment variable, to disable nss-systemd blocking on PID 1, for use by dbus-daemon */
 381         EXEC_CGROUP_DELEGATE       = 1 << 6,
 382         EXEC_IS_CONTROL            = 1 << 7,
 383         EXEC_CONTROL_CGROUP        = 1 << 8, /* Place the process not in the indicated cgroup but in a subcgroup '/.control', but only EXEC_CGROUP_DELEGATE and EXEC_IS_CONTROL is set, too */
 384         EXEC_WRITE_CREDENTIALS     = 1 << 9, /* Set up the credential store logic */
 385
 386         /* The following are not used by execute.c, but by consumers internally */
 387         EXEC_PASS_FDS              = 1 << 10,
 388         EXEC_SETENV_RESULT         = 1 << 11,
 389         EXEC_SET_WATCHDOG          = 1 << 12,
 390         EXEC_SETENV_MONITOR_RESULT = 1 << 13, /* Pass exit status to OnFailure= and OnSuccess= dependencies. */
 391 } ExecFlags;
 392
 393 /* Parameters for a specific invocation of a command. This structure is put together right before a command is
 394  * executed. */
 395 struct ExecParameters {
 396         char **environment;
 397
 398         int *fds;
 399         char **fd_names;
 400         size_t n_socket_fds;
 401         size_t n_storage_fds;
 402
 403         ExecFlags flags;
 404         bool selinux_context_net:1;
 405
 406         CGroupMask cgroup_supported;
 407         const char *cgroup_path;
 408
 409         char **prefix;
 410         const char *received_credentials_directory;
 411         const char *received_encrypted_credentials_directory;
 412
 413         const char *confirm_spawn;
 414
 415         usec_t watchdog_usec;
 416
 417         int *idle_pipe;
 418
 419         int stdin_fd;
 420         int stdout_fd;
 421         int stderr_fd;
 422
 423         /* An fd that is closed by the execve(), and thus will result in EOF when the execve() is done */
 424         int exec_fd;
 425
 426         const char *notify_socket;
 427 };
 428
 429 #include "unit.h"
 430 #include "dynamic-user.h"
 431
 432 int exec_spawn(Unit *unit,
 433                ExecCommand *command,
 434                const ExecContext *context,
 435                const ExecParameters *exec_params,
 436                ExecRuntime *runtime,
 437                DynamicCreds *dynamic_creds,
 438                pid_t *ret);
 439
 440 void exec_command_done_array(ExecCommand *c, size_t n);
 441 ExecCommand* exec_command_free_list(ExecCommand *c);
 442 void exec_command_free_array(ExecCommand **c, size_t n);
 443 void exec_command_reset_status_array(ExecCommand *c, size_t n);
 444 void exec_command_reset_status_list_array(ExecCommand **c, size_t n);
 445 void exec_command_dump_list(ExecCommand *c, FILE *f, const char *prefix);
 446 void exec_command_append_list(ExecCommand **l, ExecCommand *e);
 447 int exec_command_set(ExecCommand *c, const char *path, ...) _sentinel_;
 448 int exec_command_append(ExecCommand *c, const char *path, ...) _sentinel_;
 449
 450 void exec_context_init(ExecContext *c);
 451 void exec_context_done(ExecContext *c);
 452 void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix);
 453
 454 int exec_context_destroy_runtime_directory(const ExecContext *c, const char *runtime_root);
 455 int exec_context_destroy_credentials(const ExecContext *c, const char *runtime_root, const char *unit);
 456
 457 const char* exec_context_fdname(const ExecContext *c, int fd_index);
 458
 459 bool exec_context_may_touch_console(const ExecContext *c);
 460 bool exec_context_maintains_privileges(const ExecContext *c);
 461
 462 int exec_context_get_effective_ioprio(const ExecContext *c);
 463 bool exec_context_get_effective_mount_apivfs(const ExecContext *c);
 464
 465 void exec_context_free_log_extra_fields(ExecContext *c);
 466
 467 void exec_context_revert_tty(ExecContext *c);
 468
 469 int exec_context_get_clean_directories(ExecContext *c, char **prefix, ExecCleanMask mask, char ***ret);
 470 int exec_context_get_clean_mask(ExecContext *c, ExecCleanMask *ret);
 471
 472 void exec_status_start(ExecStatus *s, pid_t pid);
 473 void exec_status_exit(ExecStatus *s, const ExecContext *context, pid_t pid, int code, int status);
 474 void exec_status_dump(const ExecStatus *s, FILE *f, const char *prefix);
 475 void exec_status_reset(ExecStatus *s);
 476
 477 int exec_runtime_acquire(Manager *m, const ExecContext *c, const char *name, bool create, ExecRuntime **ret);
 478 ExecRuntime *exec_runtime_unref(ExecRuntime *r, bool destroy);
 479
 480 int exec_runtime_serialize(const Manager *m, FILE *f, FDSet *fds);
 481 int exec_runtime_deserialize_compat(Unit *u, const char *key, const char *value, FDSet *fds);
 482 int exec_runtime_deserialize_one(Manager *m, const char *value, FDSet *fds);
 483 void exec_runtime_vacuum(Manager *m);
 484
 485 void exec_params_clear(ExecParameters *p);
 486
 487 bool exec_context_get_cpu_affinity_from_numa(const ExecContext *c);
 488
 489 ExecSetCredential *exec_set_credential_free(ExecSetCredential *sc);
 490 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSetCredential*, exec_set_credential_free);
 491
 492 ExecLoadCredential *exec_load_credential_free(ExecLoadCredential *lc);
 493 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecLoadCredential*, exec_load_credential_free);
 494
 495 void exec_directory_done(ExecDirectory *d);
 496 int exec_directory_add(ExecDirectory *d, const char *path, const char *symlink);
 497 void exec_directory_sort(ExecDirectory *d);
 498
 499 extern const struct hash_ops exec_set_credential_hash_ops;
 500 extern const struct hash_ops exec_load_credential_hash_ops;
 501
 502 const char* exec_output_to_string(ExecOutput i) _const_;
 503 ExecOutput exec_output_from_string(const char *s) _pure_;
 504
 505 const char* exec_input_to_string(ExecInput i) _const_;
 506 ExecInput exec_input_from_string(const char *s) _pure_;
 507
 508 const char* exec_utmp_mode_to_string(ExecUtmpMode i) _const_;
 509 ExecUtmpMode exec_utmp_mode_from_string(const char *s) _pure_;
 510
 511 const char* exec_preserve_mode_to_string(ExecPreserveMode i) _const_;
 512 ExecPreserveMode exec_preserve_mode_from_string(const char *s) _pure_;
 513
 514 const char* exec_keyring_mode_to_string(ExecKeyringMode i) _const_;
 515 ExecKeyringMode exec_keyring_mode_from_string(const char *s) _pure_;
 516
 517 const char* exec_directory_type_to_string(ExecDirectoryType i) _const_;
 518 ExecDirectoryType exec_directory_type_from_string(const char *s) _pure_;
 519
 520 const char* exec_directory_type_symlink_to_string(ExecDirectoryType i) _const_;
 521 ExecDirectoryType exec_directory_type_symlink_from_string(const char *s) _pure_;
 522
 523 const char* exec_resource_type_to_string(ExecDirectoryType i) _const_;
 524 ExecDirectoryType exec_resource_type_from_string(const char *s) _pure_;
 525
 526 bool exec_needs_mount_namespace(const ExecContext *context, const ExecParameters *params, const ExecRuntime *runtime);