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 ExecSharedRuntime ExecSharedRuntime;
   8 typedef struct DynamicCreds DynamicCreds;
   9 typedef struct ExecRuntime ExecRuntime;
  10 typedef struct ExecParameters ExecParameters;
  11 typedef struct Manager Manager;
  12
  13 #include <sched.h>
  14 #include <stdbool.h>
  15 #include <stdio.h>
  16 #include <sys/capability.h>
  17
  18 #include "cgroup-util.h"
  19 #include "coredump-util.h"
  20 #include "cpu-set-util.h"
  21 #include "exec-util.h"
  22 #include "fdset.h"
  23 #include "list.h"
  24 #include "missing_resource.h"
  25 #include "namespace.h"
  26 #include "nsflags.h"
  27 #include "numa-util.h"
  28 #include "open-file.h"
  29 #include "path-util.h"
  30 #include "set.h"
  31 #include "time-util.h"
  32
  33 #define EXEC_STDIN_DATA_MAX (64U*1024U*1024U)
  34
  35 typedef enum ExecUtmpMode {
  36         EXEC_UTMP_INIT,
  37         EXEC_UTMP_LOGIN,
  38         EXEC_UTMP_USER,
  39         _EXEC_UTMP_MODE_MAX,
  40         _EXEC_UTMP_MODE_INVALID = -EINVAL,
  41 } ExecUtmpMode;
  42
  43 typedef enum ExecInput {
  44         EXEC_INPUT_NULL,
  45         EXEC_INPUT_TTY,
  46         EXEC_INPUT_TTY_FORCE,
  47         EXEC_INPUT_TTY_FAIL,
  48         EXEC_INPUT_SOCKET,
  49         EXEC_INPUT_NAMED_FD,
  50         EXEC_INPUT_DATA,
  51         EXEC_INPUT_FILE,
  52         _EXEC_INPUT_MAX,
  53         _EXEC_INPUT_INVALID = -EINVAL,
  54 } ExecInput;
  55
  56 typedef enum ExecOutput {
  57         EXEC_OUTPUT_INHERIT,
  58         EXEC_OUTPUT_NULL,
  59         EXEC_OUTPUT_TTY,
  60         EXEC_OUTPUT_KMSG,
  61         EXEC_OUTPUT_KMSG_AND_CONSOLE,
  62         EXEC_OUTPUT_JOURNAL,
  63         EXEC_OUTPUT_JOURNAL_AND_CONSOLE,
  64         EXEC_OUTPUT_SOCKET,
  65         EXEC_OUTPUT_NAMED_FD,
  66         EXEC_OUTPUT_FILE,
  67         EXEC_OUTPUT_FILE_APPEND,
  68         EXEC_OUTPUT_FILE_TRUNCATE,
  69         _EXEC_OUTPUT_MAX,
  70         _EXEC_OUTPUT_INVALID = -EINVAL,
  71 } ExecOutput;
  72
  73 typedef enum ExecPreserveMode {
  74         EXEC_PRESERVE_NO,
  75         EXEC_PRESERVE_YES,
  76         EXEC_PRESERVE_RESTART,
  77         _EXEC_PRESERVE_MODE_MAX,
  78         _EXEC_PRESERVE_MODE_INVALID = -EINVAL,
  79 } ExecPreserveMode;
  80
  81 typedef enum ExecKeyringMode {
  82         EXEC_KEYRING_INHERIT,
  83         EXEC_KEYRING_PRIVATE,
  84         EXEC_KEYRING_SHARED,
  85         _EXEC_KEYRING_MODE_MAX,
  86         _EXEC_KEYRING_MODE_INVALID = -EINVAL,
  87 } ExecKeyringMode;
  88
  89 /* Contains start and exit information about an executed command.  */
  90 struct ExecStatus {
  91         dual_timestamp start_timestamp;
  92         dual_timestamp exit_timestamp;
  93         pid_t pid;
  94         int code;     /* as in siginfo_t::si_code */
  95         int status;   /* as in siginfo_t::si_status */
  96 };
  97
  98 /* Stores information about commands we execute. Covers both configuration settings as well as runtime data. */
  99 struct ExecCommand {
 100         char *path;
 101         char **argv;
 102         ExecStatus exec_status;
 103         ExecCommandFlags flags;
 104         LIST_FIELDS(ExecCommand, command); /* useful for chaining commands */
 105 };
 106
 107 /* Encapsulates certain aspects of the runtime environment that is to be shared between multiple otherwise separate
 108  * invocations of commands. Specifically, this allows sharing of /tmp and /var/tmp data as well as network namespaces
 109  * between invocations of commands. This is a reference counted object, with one reference taken by each currently
 110  * active command invocation that wants to share this runtime. */
 111 struct ExecSharedRuntime {
 112         unsigned n_ref;
 113
 114         Manager *manager;
 115
 116         char *id; /* Unit id of the owner */
 117
 118         char *tmp_dir;
 119         char *var_tmp_dir;
 120
 121         /* An AF_UNIX socket pair, that contains a datagram containing a file descriptor referring to the network
 122          * namespace. */
 123         int netns_storage_socket[2];
 124
 125         /* Like netns_storage_socket, but the file descriptor is referring to the IPC namespace. */
 126         int ipcns_storage_socket[2];
 127 };
 128
 129 struct ExecRuntime {
 130         ExecSharedRuntime *shared;
 131         DynamicCreds *dynamic_creds;
 132 };
 133
 134 typedef enum ExecDirectoryType {
 135         EXEC_DIRECTORY_RUNTIME = 0,
 136         EXEC_DIRECTORY_STATE,
 137         EXEC_DIRECTORY_CACHE,
 138         EXEC_DIRECTORY_LOGS,
 139         EXEC_DIRECTORY_CONFIGURATION,
 140         _EXEC_DIRECTORY_TYPE_MAX,
 141         _EXEC_DIRECTORY_TYPE_INVALID = -EINVAL,
 142 } ExecDirectoryType;
 143
 144 typedef struct ExecDirectoryItem {
 145         char *path;
 146         char **symlinks;
 147         bool only_create;
 148 } ExecDirectoryItem;
 149
 150 typedef struct ExecDirectory {
 151         mode_t mode;
 152         size_t n_items;
 153         ExecDirectoryItem *items;
 154 } ExecDirectory;
 155
 156 typedef enum ExecCleanMask {
 157         /* In case you wonder why the bitmask below doesn't use "directory" in its name: we want to keep this
 158          * generic so that .timer timestamp files can nicely be covered by this too, and similar. */
 159         EXEC_CLEAN_RUNTIME       = 1U << EXEC_DIRECTORY_RUNTIME,
 160         EXEC_CLEAN_STATE         = 1U << EXEC_DIRECTORY_STATE,
 161         EXEC_CLEAN_CACHE         = 1U << EXEC_DIRECTORY_CACHE,
 162         EXEC_CLEAN_LOGS          = 1U << EXEC_DIRECTORY_LOGS,
 163         EXEC_CLEAN_CONFIGURATION = 1U << EXEC_DIRECTORY_CONFIGURATION,
 164         EXEC_CLEAN_FDSTORE       = 1U << _EXEC_DIRECTORY_TYPE_MAX,
 165         EXEC_CLEAN_NONE          = 0,
 166         EXEC_CLEAN_ALL           = (1U << (_EXEC_DIRECTORY_TYPE_MAX+1)) - 1,
 167         _EXEC_CLEAN_MASK_INVALID = -EINVAL,
 168 } ExecCleanMask;
 169
 170 /* A credential configured with LoadCredential= */
 171 typedef struct ExecLoadCredential {
 172         char *id, *path;
 173         bool encrypted;
 174 } ExecLoadCredential;
 175
 176 /* A credential configured with SetCredential= */
 177 typedef struct ExecSetCredential {
 178         char *id;
 179         bool encrypted;
 180         void *data;
 181         size_t size;
 182 } ExecSetCredential;
 183
 184 /* Encodes configuration parameters applied to invoked commands. Does not carry runtime data, but only configuration
 185  * changes sourced from unit files and suchlike. ExecContext objects are usually embedded into Unit objects, and do not
 186  * change after being loaded. */
 187 struct ExecContext {
 188         char **environment;
 189         char **environment_files;
 190         char **pass_environment;
 191         char **unset_environment;
 192
 193         struct rlimit *rlimit[_RLIMIT_MAX];
 194         char *working_directory, *root_directory, *root_image, *root_verity, *root_hash_path, *root_hash_sig_path;
 195         void *root_hash, *root_hash_sig;
 196         size_t root_hash_size, root_hash_sig_size;
 197         LIST_HEAD(MountOptions, root_image_options);
 198         bool working_directory_missing_ok:1;
 199         bool working_directory_home:1;
 200
 201         bool oom_score_adjust_set:1;
 202         bool coredump_filter_set:1;
 203         bool nice_set:1;
 204         bool ioprio_set:1;
 205         bool cpu_sched_set:1;
 206         bool mount_apivfs_set:1;
 207
 208         /* This is not exposed to the user but available internally. We need it to make sure that whenever we
 209          * spawn /usr/bin/mount it is run in the same process group as us so that the autofs logic detects
 210          * that it belongs to us and we don't enter a trigger loop. */
 211         bool same_pgrp;
 212
 213         bool cpu_sched_reset_on_fork;
 214         bool non_blocking;
 215
 216         mode_t umask;
 217         int oom_score_adjust;
 218         int nice;
 219         int ioprio;
 220         int cpu_sched_policy;
 221         int cpu_sched_priority;
 222         uint64_t coredump_filter;
 223
 224         CPUSet cpu_set;
 225         NUMAPolicy numa_policy;
 226         bool cpu_affinity_from_numa;
 227
 228         ExecInput std_input;
 229         ExecOutput std_output;
 230         ExecOutput std_error;
 231         bool stdio_as_fds;
 232         char *stdio_fdname[3];
 233         char *stdio_file[3];
 234
 235         void *stdin_data;
 236         size_t stdin_data_size;
 237
 238         nsec_t timer_slack_nsec;
 239
 240         char *tty_path;
 241
 242         bool tty_reset;
 243         bool tty_vhangup;
 244         bool tty_vt_disallocate;
 245
 246         unsigned tty_rows;
 247         unsigned tty_cols;
 248
 249         bool ignore_sigpipe;
 250
 251         ExecKeyringMode keyring_mode;
 252
 253         /* Since resolving these names might involve socket
 254          * connections and we don't want to deadlock ourselves these
 255          * names are resolved on execution only and in the child
 256          * process. */
 257         char *user;
 258         char *group;
 259         char **supplementary_groups;
 260
 261         char *pam_name;
 262
 263         char *utmp_id;
 264         ExecUtmpMode utmp_mode;
 265
 266         bool no_new_privileges;
 267
 268         bool selinux_context_ignore;
 269         bool apparmor_profile_ignore;
 270         bool smack_process_label_ignore;
 271
 272         char *selinux_context;
 273         char *apparmor_profile;
 274         char *smack_process_label;
 275
 276         char **read_write_paths, **read_only_paths, **inaccessible_paths, **exec_paths, **no_exec_paths;
 277         char **exec_search_path;
 278         unsigned long mount_propagation_flag;
 279         BindMount *bind_mounts;
 280         size_t n_bind_mounts;
 281         TemporaryFileSystem *temporary_filesystems;
 282         size_t n_temporary_filesystems;
 283         MountImage *mount_images;
 284         size_t n_mount_images;
 285         MountImage *extension_images;
 286         size_t n_extension_images;
 287         char **extension_directories;
 288
 289         uint64_t capability_bounding_set;
 290         uint64_t capability_ambient_set;
 291         int secure_bits;
 292
 293         int syslog_priority;
 294         bool syslog_level_prefix;
 295         char *syslog_identifier;
 296
 297         struct iovec* log_extra_fields;
 298         size_t n_log_extra_fields;
 299         Set *log_filter_allowed_patterns;
 300         Set *log_filter_denied_patterns;
 301
 302         usec_t log_ratelimit_interval_usec;
 303         unsigned log_ratelimit_burst;
 304
 305         int log_level_max;
 306
 307         char *log_namespace;
 308
 309         ProtectProc protect_proc;  /* hidepid= */
 310         ProcSubset proc_subset;    /* subset= */
 311
 312         int private_mounts;
 313         bool private_tmp;
 314         bool private_network;
 315         bool private_devices;
 316         bool private_users;
 317         bool private_ipc;
 318         bool protect_kernel_tunables;
 319         bool protect_kernel_modules;
 320         bool protect_kernel_logs;
 321         bool protect_clock;
 322         bool protect_control_groups;
 323         ProtectSystem protect_system;
 324         ProtectHome protect_home;
 325         bool protect_hostname;
 326         bool mount_apivfs;
 327
 328         bool dynamic_user;
 329         bool remove_ipc;
 330
 331         bool memory_deny_write_execute;
 332         bool restrict_realtime;
 333         bool restrict_suid_sgid;
 334
 335         bool lock_personality;
 336         unsigned long personality;
 337
 338         unsigned long restrict_namespaces; /* The CLONE_NEWxyz flags permitted to the unit's processes */
 339
 340         Set *restrict_filesystems;
 341         bool restrict_filesystems_allow_list:1;
 342
 343         Hashmap *syscall_filter;
 344         Set *syscall_archs;
 345         int syscall_errno;
 346         bool syscall_allow_list:1;
 347
 348         Hashmap *syscall_log;
 349         bool syscall_log_allow_list:1; /* Log listed system calls */
 350
 351         bool address_families_allow_list:1;
 352         Set *address_families;
 353
 354         char *network_namespace_path;
 355         char *ipc_namespace_path;
 356
 357         ExecDirectory directories[_EXEC_DIRECTORY_TYPE_MAX];
 358         ExecPreserveMode runtime_directory_preserve_mode;
 359         usec_t timeout_clean_usec;
 360
 361         Hashmap *set_credentials; /* output id → ExecSetCredential */
 362         Hashmap *load_credentials; /* output id → ExecLoadCredential */
 363
 364         ImagePolicy *root_image_policy, *mount_image_policy, *extension_image_policy;
 365 };
 366
 367 static inline bool exec_context_restrict_namespaces_set(const ExecContext *c) {
 368         assert(c);
 369
 370         return (c->restrict_namespaces & NAMESPACE_FLAGS_ALL) != NAMESPACE_FLAGS_ALL;
 371 }
 372
 373 static inline bool exec_context_restrict_filesystems_set(const ExecContext *c) {
 374         assert(c);
 375
 376         return c->restrict_filesystems_allow_list ||
 377           !set_isempty(c->restrict_filesystems);
 378 }
 379
 380 static inline bool exec_context_with_rootfs(const ExecContext *c) {
 381         assert(c);
 382
 383         /* Checks if RootDirectory= or RootImage= are used */
 384
 385         return !empty_or_root(c->root_directory) || c->root_image;
 386 }
 387
 388 typedef enum ExecFlags {
 389         EXEC_APPLY_SANDBOXING      = 1 << 0,
 390         EXEC_APPLY_CHROOT          = 1 << 1,
 391         EXEC_APPLY_TTY_STDIN       = 1 << 2,
 392         EXEC_PASS_LOG_UNIT         = 1 << 3, /* Whether to pass the unit name to the service's journal stream connection */
 393         EXEC_CHOWN_DIRECTORIES     = 1 << 4, /* chown() the runtime/state/cache/log directories to the user we run as, under all conditions */
 394         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 */
 395         EXEC_CGROUP_DELEGATE       = 1 << 6,
 396         EXEC_IS_CONTROL            = 1 << 7,
 397         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 */
 398         EXEC_WRITE_CREDENTIALS     = 1 << 9, /* Set up the credential store logic */
 399
 400         /* The following are not used by execute.c, but by consumers internally */
 401         EXEC_PASS_FDS              = 1 << 10,
 402         EXEC_SETENV_RESULT         = 1 << 11,
 403         EXEC_SET_WATCHDOG          = 1 << 12,
 404         EXEC_SETENV_MONITOR_RESULT = 1 << 13, /* Pass exit status to OnFailure= and OnSuccess= dependencies. */
 405 } ExecFlags;
 406
 407 /* Parameters for a specific invocation of a command. This structure is put together right before a command is
 408  * executed. */
 409 struct ExecParameters {
 410         char **environment;
 411
 412         int *fds;
 413         char **fd_names;
 414         size_t n_socket_fds;
 415         size_t n_storage_fds;
 416
 417         ExecFlags flags;
 418         bool selinux_context_net:1;
 419
 420         CGroupMask cgroup_supported;
 421         const char *cgroup_path;
 422
 423         char **prefix;
 424         const char *received_credentials_directory;
 425         const char *received_encrypted_credentials_directory;
 426
 427         const char *confirm_spawn;
 428
 429         usec_t watchdog_usec;
 430
 431         int *idle_pipe;
 432
 433         int stdin_fd;
 434         int stdout_fd;
 435         int stderr_fd;
 436
 437         /* An fd that is closed by the execve(), and thus will result in EOF when the execve() is done */
 438         int exec_fd;
 439
 440         const char *notify_socket;
 441
 442         LIST_HEAD(OpenFile, open_files);
 443 };
 444
 445 #include "unit.h"
 446 #include "dynamic-user.h"
 447
 448 int exec_spawn(Unit *unit,
 449                ExecCommand *command,
 450                const ExecContext *context,
 451                const ExecParameters *exec_params,
 452                ExecRuntime *runtime,
 453                const CGroupContext *cgroup_context,
 454                pid_t *ret);
 455
 456 void exec_command_done_array(ExecCommand *c, size_t n);
 457 ExecCommand* exec_command_free_list(ExecCommand *c);
 458 void exec_command_free_array(ExecCommand **c, size_t n);
 459 void exec_command_reset_status_array(ExecCommand *c, size_t n);
 460 void exec_command_reset_status_list_array(ExecCommand **c, size_t n);
 461 void exec_command_dump_list(ExecCommand *c, FILE *f, const char *prefix);
 462 void exec_command_append_list(ExecCommand **l, ExecCommand *e);
 463 int exec_command_set(ExecCommand *c, const char *path, ...) _sentinel_;
 464 int exec_command_append(ExecCommand *c, const char *path, ...) _sentinel_;
 465
 466 void exec_context_init(ExecContext *c);
 467 void exec_context_done(ExecContext *c);
 468 void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix);
 469
 470 int exec_context_destroy_runtime_directory(const ExecContext *c, const char *runtime_root);
 471 int exec_context_destroy_credentials(const ExecContext *c, const char *runtime_root, const char *unit);
 472 int exec_context_destroy_mount_ns_dir(Unit *u);
 473
 474 const char* exec_context_fdname(const ExecContext *c, int fd_index);
 475
 476 bool exec_context_may_touch_console(const ExecContext *c);
 477 bool exec_context_maintains_privileges(const ExecContext *c);
 478 bool exec_context_has_encrypted_credentials(ExecContext *c);
 479 bool exec_context_has_credentials(const ExecContext *context);
 480
 481 int exec_context_get_effective_ioprio(const ExecContext *c);
 482 bool exec_context_get_effective_mount_apivfs(const ExecContext *c);
 483
 484 void exec_context_free_log_extra_fields(ExecContext *c);
 485
 486 void exec_context_revert_tty(ExecContext *c);
 487
 488 int exec_context_get_clean_directories(ExecContext *c, char **prefix, ExecCleanMask mask, char ***ret);
 489 int exec_context_get_clean_mask(ExecContext *c, ExecCleanMask *ret);
 490 int exec_context_add_default_dependencies(Unit *u, const ExecContext *c);
 491
 492 void exec_status_start(ExecStatus *s, pid_t pid);
 493 void exec_status_exit(ExecStatus *s, const ExecContext *context, pid_t pid, int code, int status);
 494 void exec_status_dump(const ExecStatus *s, FILE *f, const char *prefix);
 495 void exec_status_reset(ExecStatus *s);
 496
 497 int exec_shared_runtime_acquire(Manager *m, const ExecContext *c, const char *name, bool create, ExecSharedRuntime **ret);
 498 ExecSharedRuntime *exec_shared_runtime_destroy(ExecSharedRuntime *r);
 499 ExecSharedRuntime *exec_shared_runtime_unref(ExecSharedRuntime *r);
 500 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSharedRuntime*, exec_shared_runtime_unref);
 501
 502 int exec_shared_runtime_serialize(const Manager *m, FILE *f, FDSet *fds);
 503 int exec_shared_runtime_deserialize_compat(Unit *u, const char *key, const char *value, FDSet *fds);
 504 int exec_shared_runtime_deserialize_one(Manager *m, const char *value, FDSet *fds);
 505 void exec_shared_runtime_vacuum(Manager *m);
 506
 507 int exec_runtime_make(ExecSharedRuntime *shared, DynamicCreds *creds, ExecRuntime **ret);
 508 ExecRuntime* exec_runtime_free(ExecRuntime *rt);
 509 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecRuntime*, exec_runtime_free);
 510 ExecRuntime* exec_runtime_destroy(ExecRuntime *rt);
 511
 512 void exec_params_clear(ExecParameters *p);
 513
 514 bool exec_context_get_cpu_affinity_from_numa(const ExecContext *c);
 515
 516 ExecSetCredential *exec_set_credential_free(ExecSetCredential *sc);
 517 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSetCredential*, exec_set_credential_free);
 518
 519 ExecLoadCredential *exec_load_credential_free(ExecLoadCredential *lc);
 520 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecLoadCredential*, exec_load_credential_free);
 521
 522 void exec_directory_done(ExecDirectory *d);
 523 int exec_directory_add(ExecDirectory *d, const char *path, const char *symlink);
 524 void exec_directory_sort(ExecDirectory *d);
 525
 526 ExecCleanMask exec_clean_mask_from_string(const char *s);
 527
 528 extern const struct hash_ops exec_set_credential_hash_ops;
 529 extern const struct hash_ops exec_load_credential_hash_ops;
 530
 531 const char* exec_output_to_string(ExecOutput i) _const_;
 532 ExecOutput exec_output_from_string(const char *s) _pure_;
 533
 534 const char* exec_input_to_string(ExecInput i) _const_;
 535 ExecInput exec_input_from_string(const char *s) _pure_;
 536
 537 const char* exec_utmp_mode_to_string(ExecUtmpMode i) _const_;
 538 ExecUtmpMode exec_utmp_mode_from_string(const char *s) _pure_;
 539
 540 const char* exec_preserve_mode_to_string(ExecPreserveMode i) _const_;
 541 ExecPreserveMode exec_preserve_mode_from_string(const char *s) _pure_;
 542
 543 const char* exec_keyring_mode_to_string(ExecKeyringMode i) _const_;
 544 ExecKeyringMode exec_keyring_mode_from_string(const char *s) _pure_;
 545
 546 const char* exec_directory_type_to_string(ExecDirectoryType i) _const_;
 547 ExecDirectoryType exec_directory_type_from_string(const char *s) _pure_;
 548
 549 const char* exec_directory_type_symlink_to_string(ExecDirectoryType i) _const_;
 550 ExecDirectoryType exec_directory_type_symlink_from_string(const char *s) _pure_;
 551
 552 const char* exec_resource_type_to_string(ExecDirectoryType i) _const_;
 553 ExecDirectoryType exec_resource_type_from_string(const char *s) _pure_;
 554
 555 bool exec_needs_mount_namespace(const ExecContext *context, const ExecParameters *params, const ExecRuntime *runtime);
 556 bool exec_needs_network_namespace(const ExecContext *context);