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 "open-file.h"
  27 #include "path-util.h"
  28 #include "set.h"
  29 #include "time-util.h"
  30
  31 #define EXEC_STDIN_DATA_MAX (64U*1024U*1024U)
  32
  33 typedef enum ExecUtmpMode {
  34         EXEC_UTMP_INIT,
  35         EXEC_UTMP_LOGIN,
  36         EXEC_UTMP_USER,
  37         _EXEC_UTMP_MODE_MAX,
  38         _EXEC_UTMP_MODE_INVALID = -EINVAL,
  39 } ExecUtmpMode;
  40
  41 typedef enum ExecInput {
  42         EXEC_INPUT_NULL,
  43         EXEC_INPUT_TTY,
  44         EXEC_INPUT_TTY_FORCE,
  45         EXEC_INPUT_TTY_FAIL,
  46         EXEC_INPUT_SOCKET,
  47         EXEC_INPUT_NAMED_FD,
  48         EXEC_INPUT_DATA,
  49         EXEC_INPUT_FILE,
  50         _EXEC_INPUT_MAX,
  51         _EXEC_INPUT_INVALID = -EINVAL,
  52 } ExecInput;
  53
  54 typedef enum ExecOutput {
  55         EXEC_OUTPUT_INHERIT,
  56         EXEC_OUTPUT_NULL,
  57         EXEC_OUTPUT_TTY,
  58         EXEC_OUTPUT_KMSG,
  59         EXEC_OUTPUT_KMSG_AND_CONSOLE,
  60         EXEC_OUTPUT_JOURNAL,
  61         EXEC_OUTPUT_JOURNAL_AND_CONSOLE,
  62         EXEC_OUTPUT_SOCKET,
  63         EXEC_OUTPUT_NAMED_FD,
  64         EXEC_OUTPUT_FILE,
  65         EXEC_OUTPUT_FILE_APPEND,
  66         EXEC_OUTPUT_FILE_TRUNCATE,
  67         _EXEC_OUTPUT_MAX,
  68         _EXEC_OUTPUT_INVALID = -EINVAL,
  69 } ExecOutput;
  70
  71 typedef enum ExecPreserveMode {
  72         EXEC_PRESERVE_NO,
  73         EXEC_PRESERVE_YES,
  74         EXEC_PRESERVE_RESTART,
  75         _EXEC_PRESERVE_MODE_MAX,
  76         _EXEC_PRESERVE_MODE_INVALID = -EINVAL,
  77 } ExecPreserveMode;
  78
  79 typedef enum ExecKeyringMode {
  80         EXEC_KEYRING_INHERIT,
  81         EXEC_KEYRING_PRIVATE,
  82         EXEC_KEYRING_SHARED,
  83         _EXEC_KEYRING_MODE_MAX,
  84         _EXEC_KEYRING_MODE_INVALID = -EINVAL,
  85 } ExecKeyringMode;
  86
  87 /* Contains start and exit information about an executed command.  */
  88 struct ExecStatus {
  89         dual_timestamp start_timestamp;
  90         dual_timestamp exit_timestamp;
  91         pid_t pid;
  92         int code;     /* as in siginfo_t::si_code */
  93         int status;   /* as in siginfo_t::si_status */
  94 };
  95
  96 /* Stores information about commands we execute. Covers both configuration settings as well as runtime data. */
  97 struct ExecCommand {
  98         char *path;
  99         char **argv;
 100         ExecStatus exec_status;
 101         ExecCommandFlags flags;
 102         LIST_FIELDS(ExecCommand, command); /* useful for chaining commands */
 103 };
 104
 105 /* Encapsulates certain aspects of the runtime environment that is to be shared between multiple otherwise separate
 106  * invocations of commands. Specifically, this allows sharing of /tmp and /var/tmp data as well as network namespaces
 107  * between invocations of commands. This is a reference counted object, with one reference taken by each currently
 108  * active command invocation that wants to share this runtime. */
 109 struct ExecRuntime {
 110         unsigned n_ref;
 111
 112         Manager *manager;
 113
 114         char *id; /* Unit id of the owner */
 115
 116         char *tmp_dir;
 117         char *var_tmp_dir;
 118
 119         /* An AF_UNIX socket pair, that contains a datagram containing a file descriptor referring to the network
 120          * namespace. */
 121         int netns_storage_socket[2];
 122
 123         /* Like netns_storage_socket, but the file descriptor is referring to the IPC namespace. */
 124         int ipcns_storage_socket[2];
 125 };
 126
 127 typedef enum ExecDirectoryType {
 128         EXEC_DIRECTORY_RUNTIME = 0,
 129         EXEC_DIRECTORY_STATE,
 130         EXEC_DIRECTORY_CACHE,
 131         EXEC_DIRECTORY_LOGS,
 132         EXEC_DIRECTORY_CONFIGURATION,
 133         _EXEC_DIRECTORY_TYPE_MAX,
 134         _EXEC_DIRECTORY_TYPE_INVALID = -EINVAL,
 135 } ExecDirectoryType;
 136
 137 typedef struct ExecDirectoryItem {
 138         char *path;
 139         char **symlinks;
 140         bool only_create;
 141 } ExecDirectoryItem;
 142
 143 typedef struct ExecDirectory {
 144         mode_t mode;
 145         size_t n_items;
 146         ExecDirectoryItem *items;
 147 } ExecDirectory;
 148
 149 typedef enum ExecCleanMask {
 150         /* In case you wonder why the bitmask below doesn't use "directory" in its name: we want to keep this
 151          * generic so that .timer timestamp files can nicely be covered by this too, and similar. */
 152         EXEC_CLEAN_RUNTIME       = 1U << EXEC_DIRECTORY_RUNTIME,
 153         EXEC_CLEAN_STATE         = 1U << EXEC_DIRECTORY_STATE,
 154         EXEC_CLEAN_CACHE         = 1U << EXEC_DIRECTORY_CACHE,
 155         EXEC_CLEAN_LOGS          = 1U << EXEC_DIRECTORY_LOGS,
 156         EXEC_CLEAN_CONFIGURATION = 1U << EXEC_DIRECTORY_CONFIGURATION,
 157         EXEC_CLEAN_NONE          = 0,
 158         EXEC_CLEAN_ALL           = (1U << _EXEC_DIRECTORY_TYPE_MAX) - 1,
 159         _EXEC_CLEAN_MASK_INVALID = -EINVAL,
 160 } ExecCleanMask;
 161
 162 /* A credential configured with LoadCredential= */
 163 typedef struct ExecLoadCredential {
 164         char *id, *path;
 165         bool encrypted;
 166 } ExecLoadCredential;
 167
 168 /* A credential configured with SetCredential= */
 169 typedef struct ExecSetCredential {
 170         char *id;
 171         bool encrypted;
 172         void *data;
 173         size_t size;
 174 } ExecSetCredential;
 175
 176 /* Encodes configuration parameters applied to invoked commands. Does not carry runtime data, but only configuration
 177  * changes sourced from unit files and suchlike. ExecContext objects are usually embedded into Unit objects, and do not
 178  * change after being loaded. */
 179 struct ExecContext {
 180         char **environment;
 181         char **environment_files;
 182         char **pass_environment;
 183         char **unset_environment;
 184
 185         struct rlimit *rlimit[_RLIMIT_MAX];
 186         char *working_directory, *root_directory, *root_image, *root_verity, *root_hash_path, *root_hash_sig_path;
 187         void *root_hash, *root_hash_sig;
 188         size_t root_hash_size, root_hash_sig_size;
 189         LIST_HEAD(MountOptions, root_image_options);
 190         bool working_directory_missing_ok:1;
 191         bool working_directory_home:1;
 192
 193         bool oom_score_adjust_set:1;
 194         bool coredump_filter_set:1;
 195         bool nice_set:1;
 196         bool ioprio_set:1;
 197         bool cpu_sched_set:1;
 198         bool mount_apivfs_set:1;
 199
 200         /* This is not exposed to the user but available internally. We need it to make sure that whenever we
 201          * spawn /usr/bin/mount it is run in the same process group as us so that the autofs logic detects
 202          * that it belongs to us and we don't enter a trigger loop. */
 203         bool same_pgrp;
 204
 205         bool cpu_sched_reset_on_fork;
 206         bool non_blocking;
 207
 208         mode_t umask;
 209         int oom_score_adjust;
 210         int nice;
 211         int ioprio;
 212         int cpu_sched_policy;
 213         int cpu_sched_priority;
 214         uint64_t coredump_filter;
 215
 216         CPUSet cpu_set;
 217         NUMAPolicy numa_policy;
 218         bool cpu_affinity_from_numa;
 219
 220         ExecInput std_input;
 221         ExecOutput std_output;
 222         ExecOutput std_error;
 223         bool stdio_as_fds;
 224         char *stdio_fdname[3];
 225         char *stdio_file[3];
 226
 227         void *stdin_data;
 228         size_t stdin_data_size;
 229
 230         nsec_t timer_slack_nsec;
 231
 232         char *tty_path;
 233
 234         bool tty_reset;
 235         bool tty_vhangup;
 236         bool tty_vt_disallocate;
 237
 238         unsigned tty_rows;
 239         unsigned tty_cols;
 240
 241         bool ignore_sigpipe;
 242
 243         ExecKeyringMode keyring_mode;
 244
 245         /* Since resolving these names might involve socket
 246          * connections and we don't want to deadlock ourselves these
 247          * names are resolved on execution only and in the child
 248          * process. */
 249         char *user;
 250         char *group;
 251         char **supplementary_groups;
 252
 253         char *pam_name;
 254
 255         char *utmp_id;
 256         ExecUtmpMode utmp_mode;
 257
 258         bool no_new_privileges;
 259
 260         bool selinux_context_ignore;
 261         bool apparmor_profile_ignore;
 262         bool smack_process_label_ignore;
 263
 264         char *selinux_context;
 265         char *apparmor_profile;
 266         char *smack_process_label;
 267
 268         char **read_write_paths, **read_only_paths, **inaccessible_paths, **exec_paths, **no_exec_paths;
 269         char **exec_search_path;
 270         unsigned long mount_propagation_flag;
 271         BindMount *bind_mounts;
 272         size_t n_bind_mounts;
 273         TemporaryFileSystem *temporary_filesystems;
 274         size_t n_temporary_filesystems;
 275         MountImage *mount_images;
 276         size_t n_mount_images;
 277         MountImage *extension_images;
 278         size_t n_extension_images;
 279         char **extension_directories;
 280
 281         uint64_t capability_bounding_set;
 282         uint64_t capability_ambient_set;
 283         int secure_bits;
 284
 285         int syslog_priority;
 286         bool syslog_level_prefix;
 287         char *syslog_identifier;
 288
 289         struct iovec* log_extra_fields;
 290         size_t n_log_extra_fields;
 291         Set *log_filter_allowed_patterns;
 292         Set *log_filter_denied_patterns;
 293
 294         usec_t log_ratelimit_interval_usec;
 295         unsigned log_ratelimit_burst;
 296
 297         int log_level_max;
 298
 299         char *log_namespace;
 300
 301         ProtectProc protect_proc;  /* hidepid= */
 302         ProcSubset proc_subset;    /* subset= */
 303
 304         int private_mounts;
 305         bool private_tmp;
 306         bool private_network;
 307         bool private_devices;
 308         bool private_users;
 309         bool private_ipc;
 310         bool protect_kernel_tunables;
 311         bool protect_kernel_modules;
 312         bool protect_kernel_logs;
 313         bool protect_clock;
 314         bool protect_control_groups;
 315         ProtectSystem protect_system;
 316         ProtectHome protect_home;
 317         bool protect_hostname;
 318         bool mount_apivfs;
 319
 320         bool dynamic_user;
 321         bool remove_ipc;
 322
 323         bool memory_deny_write_execute;
 324         bool restrict_realtime;
 325         bool restrict_suid_sgid;
 326
 327         bool lock_personality;
 328         unsigned long personality;
 329
 330         unsigned long restrict_namespaces; /* The CLONE_NEWxyz flags permitted to the unit's processes */
 331
 332         Set *restrict_filesystems;
 333         bool restrict_filesystems_allow_list:1;
 334
 335         Hashmap *syscall_filter;
 336         Set *syscall_archs;
 337         int syscall_errno;
 338         bool syscall_allow_list:1;
 339
 340         Hashmap *syscall_log;
 341         bool syscall_log_allow_list:1; /* Log listed system calls */
 342
 343         bool address_families_allow_list:1;
 344         Set *address_families;
 345
 346         char *network_namespace_path;
 347         char *ipc_namespace_path;
 348
 349         ExecDirectory directories[_EXEC_DIRECTORY_TYPE_MAX];
 350         ExecPreserveMode runtime_directory_preserve_mode;
 351         usec_t timeout_clean_usec;
 352
 353         Hashmap *set_credentials; /* output id → ExecSetCredential */
 354         Hashmap *load_credentials; /* output id → ExecLoadCredential */
 355 };
 356
 357 static inline bool exec_context_restrict_namespaces_set(const ExecContext *c) {
 358         assert(c);
 359
 360         return (c->restrict_namespaces & NAMESPACE_FLAGS_ALL) != NAMESPACE_FLAGS_ALL;
 361 }
 362
 363 static inline bool exec_context_restrict_filesystems_set(const ExecContext *c) {
 364         assert(c);
 365
 366         return c->restrict_filesystems_allow_list ||
 367           !set_isempty(c->restrict_filesystems);
 368 }
 369
 370 static inline bool exec_context_with_rootfs(const ExecContext *c) {
 371         assert(c);
 372
 373         /* Checks if RootDirectory= or RootImage= are used */
 374
 375         return !empty_or_root(c->root_directory) || c->root_image;
 376 }
 377
 378 typedef enum ExecFlags {
 379         EXEC_APPLY_SANDBOXING      = 1 << 0,
 380         EXEC_APPLY_CHROOT          = 1 << 1,
 381         EXEC_APPLY_TTY_STDIN       = 1 << 2,
 382         EXEC_PASS_LOG_UNIT         = 1 << 3, /* Whether to pass the unit name to the service's journal stream connection */
 383         EXEC_CHOWN_DIRECTORIES     = 1 << 4, /* chown() the runtime/state/cache/log directories to the user we run as, under all conditions */
 384         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 */
 385         EXEC_CGROUP_DELEGATE       = 1 << 6,
 386         EXEC_IS_CONTROL            = 1 << 7,
 387         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 */
 388         EXEC_WRITE_CREDENTIALS     = 1 << 9, /* Set up the credential store logic */
 389
 390         /* The following are not used by execute.c, but by consumers internally */
 391         EXEC_PASS_FDS              = 1 << 10,
 392         EXEC_SETENV_RESULT         = 1 << 11,
 393         EXEC_SET_WATCHDOG          = 1 << 12,
 394         EXEC_SETENV_MONITOR_RESULT = 1 << 13, /* Pass exit status to OnFailure= and OnSuccess= dependencies. */
 395 } ExecFlags;
 396
 397 /* Parameters for a specific invocation of a command. This structure is put together right before a command is
 398  * executed. */
 399 struct ExecParameters {
 400         char **environment;
 401
 402         int *fds;
 403         char **fd_names;
 404         size_t n_socket_fds;
 405         size_t n_storage_fds;
 406
 407         ExecFlags flags;
 408         bool selinux_context_net:1;
 409
 410         CGroupMask cgroup_supported;
 411         const char *cgroup_path;
 412
 413         char **prefix;
 414         const char *received_credentials_directory;
 415         const char *received_encrypted_credentials_directory;
 416
 417         const char *confirm_spawn;
 418
 419         usec_t watchdog_usec;
 420
 421         int *idle_pipe;
 422
 423         int stdin_fd;
 424         int stdout_fd;
 425         int stderr_fd;
 426
 427         /* An fd that is closed by the execve(), and thus will result in EOF when the execve() is done */
 428         int exec_fd;
 429
 430         const char *notify_socket;
 431
 432         LIST_HEAD(OpenFile, open_files);
 433 };
 434
 435 #include "unit.h"
 436 #include "dynamic-user.h"
 437
 438 int exec_spawn(Unit *unit,
 439                ExecCommand *command,
 440                const ExecContext *context,
 441                const ExecParameters *exec_params,
 442                ExecRuntime *runtime,
 443                DynamicCreds *dynamic_creds,
 444                const CGroupContext *cgroup_context,
 445                pid_t *ret);
 446
 447 void exec_command_done_array(ExecCommand *c, size_t n);
 448 ExecCommand* exec_command_free_list(ExecCommand *c);
 449 void exec_command_free_array(ExecCommand **c, size_t n);
 450 void exec_command_reset_status_array(ExecCommand *c, size_t n);
 451 void exec_command_reset_status_list_array(ExecCommand **c, size_t n);
 452 void exec_command_dump_list(ExecCommand *c, FILE *f, const char *prefix);
 453 void exec_command_append_list(ExecCommand **l, ExecCommand *e);
 454 int exec_command_set(ExecCommand *c, const char *path, ...) _sentinel_;
 455 int exec_command_append(ExecCommand *c, const char *path, ...) _sentinel_;
 456
 457 void exec_context_init(ExecContext *c);
 458 void exec_context_done(ExecContext *c);
 459 void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix);
 460
 461 int exec_context_destroy_runtime_directory(const ExecContext *c, const char *runtime_root);
 462 int exec_context_destroy_credentials(const ExecContext *c, const char *runtime_root, const char *unit);
 463 int exec_context_destroy_mount_ns_dir(Unit *u);
 464
 465 const char* exec_context_fdname(const ExecContext *c, int fd_index);
 466
 467 bool exec_context_may_touch_console(const ExecContext *c);
 468 bool exec_context_maintains_privileges(const ExecContext *c);
 469 bool exec_context_has_encrypted_credentials(ExecContext *c);
 470
 471 int exec_context_get_effective_ioprio(const ExecContext *c);
 472 bool exec_context_get_effective_mount_apivfs(const ExecContext *c);
 473
 474 void exec_context_free_log_extra_fields(ExecContext *c);
 475
 476 void exec_context_revert_tty(ExecContext *c);
 477
 478 int exec_context_get_clean_directories(ExecContext *c, char **prefix, ExecCleanMask mask, char ***ret);
 479 int exec_context_get_clean_mask(ExecContext *c, ExecCleanMask *ret);
 480
 481 void exec_status_start(ExecStatus *s, pid_t pid);
 482 void exec_status_exit(ExecStatus *s, const ExecContext *context, pid_t pid, int code, int status);
 483 void exec_status_dump(const ExecStatus *s, FILE *f, const char *prefix);
 484 void exec_status_reset(ExecStatus *s);
 485
 486 int exec_runtime_acquire(Manager *m, const ExecContext *c, const char *name, bool create, ExecRuntime **ret);
 487 ExecRuntime *exec_runtime_unref(ExecRuntime *r, bool destroy);
 488
 489 int exec_runtime_serialize(const Manager *m, FILE *f, FDSet *fds);
 490 int exec_runtime_deserialize_compat(Unit *u, const char *key, const char *value, FDSet *fds);
 491 int exec_runtime_deserialize_one(Manager *m, const char *value, FDSet *fds);
 492 void exec_runtime_vacuum(Manager *m);
 493
 494 void exec_params_clear(ExecParameters *p);
 495
 496 bool exec_context_get_cpu_affinity_from_numa(const ExecContext *c);
 497
 498 ExecSetCredential *exec_set_credential_free(ExecSetCredential *sc);
 499 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSetCredential*, exec_set_credential_free);
 500
 501 ExecLoadCredential *exec_load_credential_free(ExecLoadCredential *lc);
 502 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecLoadCredential*, exec_load_credential_free);
 503
 504 void exec_directory_done(ExecDirectory *d);
 505 int exec_directory_add(ExecDirectory *d, const char *path, const char *symlink);
 506 void exec_directory_sort(ExecDirectory *d);
 507
 508 extern const struct hash_ops exec_set_credential_hash_ops;
 509 extern const struct hash_ops exec_load_credential_hash_ops;
 510
 511 const char* exec_output_to_string(ExecOutput i) _const_;
 512 ExecOutput exec_output_from_string(const char *s) _pure_;
 513
 514 const char* exec_input_to_string(ExecInput i) _const_;
 515 ExecInput exec_input_from_string(const char *s) _pure_;
 516
 517 const char* exec_utmp_mode_to_string(ExecUtmpMode i) _const_;
 518 ExecUtmpMode exec_utmp_mode_from_string(const char *s) _pure_;
 519
 520 const char* exec_preserve_mode_to_string(ExecPreserveMode i) _const_;
 521 ExecPreserveMode exec_preserve_mode_from_string(const char *s) _pure_;
 522
 523 const char* exec_keyring_mode_to_string(ExecKeyringMode i) _const_;
 524 ExecKeyringMode exec_keyring_mode_from_string(const char *s) _pure_;
 525
 526 const char* exec_directory_type_to_string(ExecDirectoryType i) _const_;
 527 ExecDirectoryType exec_directory_type_from_string(const char *s) _pure_;
 528
 529 const char* exec_directory_type_symlink_to_string(ExecDirectoryType i) _const_;
 530 ExecDirectoryType exec_directory_type_symlink_from_string(const char *s) _pure_;
 531
 532 const char* exec_resource_type_to_string(ExecDirectoryType i) _const_;
 533 ExecDirectoryType exec_resource_type_from_string(const char *s) _pure_;
 534
 535 bool exec_needs_mount_namespace(const ExecContext *context, const ExecParameters *params, const ExecRuntime *runtime);
 536 bool exec_needs_network_namespace(const ExecContext *context);