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_NONE          = 0,
 165         EXEC_CLEAN_ALL           = (1U << _EXEC_DIRECTORY_TYPE_MAX) - 1,
 166         _EXEC_CLEAN_MASK_INVALID = -EINVAL,
 167 } ExecCleanMask;
 168
 169 /* A credential configured with LoadCredential= */
 170 typedef struct ExecLoadCredential {
 171         char *id, *path;
 172         bool encrypted;
 173 } ExecLoadCredential;
 174
 175 /* A credential configured with SetCredential= */
 176 typedef struct ExecSetCredential {
 177         char *id;
 178         bool encrypted;
 179         void *data;
 180         size_t size;
 181 } ExecSetCredential;
 182
 183 /* Encodes configuration parameters applied to invoked commands. Does not carry runtime data, but only configuration
 184  * changes sourced from unit files and suchlike. ExecContext objects are usually embedded into Unit objects, and do not
 185  * change after being loaded. */
 186 struct ExecContext {
 187         char **environment;
 188         char **environment_files;
 189         char **pass_environment;
 190         char **unset_environment;
 191
 192         struct rlimit *rlimit[_RLIMIT_MAX];
 193         char *working_directory, *root_directory, *root_image, *root_verity, *root_hash_path, *root_hash_sig_path;
 194         void *root_hash, *root_hash_sig;
 195         size_t root_hash_size, root_hash_sig_size;
 196         LIST_HEAD(MountOptions, root_image_options);
 197         bool working_directory_missing_ok:1;
 198         bool working_directory_home:1;
 199
 200         bool oom_score_adjust_set:1;
 201         bool coredump_filter_set:1;
 202         bool nice_set:1;
 203         bool ioprio_set:1;
 204         bool cpu_sched_set:1;
 205         bool mount_apivfs_set:1;
 206
 207         /* This is not exposed to the user but available internally. We need it to make sure that whenever we
 208          * spawn /usr/bin/mount it is run in the same process group as us so that the autofs logic detects
 209          * that it belongs to us and we don't enter a trigger loop. */
 210         bool same_pgrp;
 211
 212         bool cpu_sched_reset_on_fork;
 213         bool non_blocking;
 214
 215         mode_t umask;
 216         int oom_score_adjust;
 217         int nice;
 218         int ioprio;
 219         int cpu_sched_policy;
 220         int cpu_sched_priority;
 221         uint64_t coredump_filter;
 222
 223         CPUSet cpu_set;
 224         NUMAPolicy numa_policy;
 225         bool cpu_affinity_from_numa;
 226
 227         ExecInput std_input;
 228         ExecOutput std_output;
 229         ExecOutput std_error;
 230         bool stdio_as_fds;
 231         char *stdio_fdname[3];
 232         char *stdio_file[3];
 233
 234         void *stdin_data;
 235         size_t stdin_data_size;
 236
 237         nsec_t timer_slack_nsec;
 238
 239         char *tty_path;
 240
 241         bool tty_reset;
 242         bool tty_vhangup;
 243         bool tty_vt_disallocate;
 244
 245         unsigned tty_rows;
 246         unsigned tty_cols;
 247
 248         bool ignore_sigpipe;
 249
 250         ExecKeyringMode keyring_mode;
 251
 252         /* Since resolving these names might involve socket
 253          * connections and we don't want to deadlock ourselves these
 254          * names are resolved on execution only and in the child
 255          * process. */
 256         char *user;
 257         char *group;
 258         char **supplementary_groups;
 259
 260         char *pam_name;
 261
 262         char *utmp_id;
 263         ExecUtmpMode utmp_mode;
 264
 265         bool no_new_privileges;
 266
 267         bool selinux_context_ignore;
 268         bool apparmor_profile_ignore;
 269         bool smack_process_label_ignore;
 270
 271         char *selinux_context;
 272         char *apparmor_profile;
 273         char *smack_process_label;
 274
 275         char **read_write_paths, **read_only_paths, **inaccessible_paths, **exec_paths, **no_exec_paths;
 276         char **exec_search_path;
 277         unsigned long mount_propagation_flag;
 278         BindMount *bind_mounts;
 279         size_t n_bind_mounts;
 280         TemporaryFileSystem *temporary_filesystems;
 281         size_t n_temporary_filesystems;
 282         MountImage *mount_images;
 283         size_t n_mount_images;
 284         MountImage *extension_images;
 285         size_t n_extension_images;
 286         char **extension_directories;
 287
 288         uint64_t capability_bounding_set;
 289         uint64_t capability_ambient_set;
 290         int secure_bits;
 291
 292         int syslog_priority;
 293         bool syslog_level_prefix;
 294         char *syslog_identifier;
 295
 296         struct iovec* log_extra_fields;
 297         size_t n_log_extra_fields;
 298         Set *log_filter_allowed_patterns;
 299         Set *log_filter_denied_patterns;
 300
 301         usec_t log_ratelimit_interval_usec;
 302         unsigned log_ratelimit_burst;
 303
 304         int log_level_max;
 305
 306         char *log_namespace;
 307
 308         ProtectProc protect_proc;  /* hidepid= */
 309         ProcSubset proc_subset;    /* subset= */
 310
 311         int private_mounts;
 312         bool private_tmp;
 313         bool private_network;
 314         bool private_devices;
 315         bool private_users;
 316         bool private_ipc;
 317         bool protect_kernel_tunables;
 318         bool protect_kernel_modules;
 319         bool protect_kernel_logs;
 320         bool protect_clock;
 321         bool protect_control_groups;
 322         ProtectSystem protect_system;
 323         ProtectHome protect_home;
 324         bool protect_hostname;
 325         bool mount_apivfs;
 326
 327         bool dynamic_user;
 328         bool remove_ipc;
 329
 330         bool memory_deny_write_execute;
 331         bool restrict_realtime;
 332         bool restrict_suid_sgid;
 333
 334         bool lock_personality;
 335         unsigned long personality;
 336
 337         unsigned long restrict_namespaces; /* The CLONE_NEWxyz flags permitted to the unit's processes */
 338
 339         Set *restrict_filesystems;
 340         bool restrict_filesystems_allow_list:1;
 341
 342         Hashmap *syscall_filter;
 343         Set *syscall_archs;
 344         int syscall_errno;
 345         bool syscall_allow_list:1;
 346
 347         Hashmap *syscall_log;
 348         bool syscall_log_allow_list:1; /* Log listed system calls */
 349
 350         bool address_families_allow_list:1;
 351         Set *address_families;
 352
 353         char *network_namespace_path;
 354         char *ipc_namespace_path;
 355
 356         ExecDirectory directories[_EXEC_DIRECTORY_TYPE_MAX];
 357         ExecPreserveMode runtime_directory_preserve_mode;
 358         usec_t timeout_clean_usec;
 359
 360         Hashmap *set_credentials; /* output id → ExecSetCredential */
 361         Hashmap *load_credentials; /* output id → ExecLoadCredential */
 362
 363         ImagePolicy *root_image_policy, *mount_image_policy, *extension_image_policy;
 364 };
 365
 366 static inline bool exec_context_restrict_namespaces_set(const ExecContext *c) {
 367         assert(c);
 368
 369         return (c->restrict_namespaces & NAMESPACE_FLAGS_ALL) != NAMESPACE_FLAGS_ALL;
 370 }
 371
 372 static inline bool exec_context_restrict_filesystems_set(const ExecContext *c) {
 373         assert(c);
 374
 375         return c->restrict_filesystems_allow_list ||
 376           !set_isempty(c->restrict_filesystems);
 377 }
 378
 379 static inline bool exec_context_with_rootfs(const ExecContext *c) {
 380         assert(c);
 381
 382         /* Checks if RootDirectory= or RootImage= are used */
 383
 384         return !empty_or_root(c->root_directory) || c->root_image;
 385 }
 386
 387 typedef enum ExecFlags {
 388         EXEC_APPLY_SANDBOXING      = 1 << 0,
 389         EXEC_APPLY_CHROOT          = 1 << 1,
 390         EXEC_APPLY_TTY_STDIN       = 1 << 2,
 391         EXEC_PASS_LOG_UNIT         = 1 << 3, /* Whether to pass the unit name to the service's journal stream connection */
 392         EXEC_CHOWN_DIRECTORIES     = 1 << 4, /* chown() the runtime/state/cache/log directories to the user we run as, under all conditions */
 393         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 */
 394         EXEC_CGROUP_DELEGATE       = 1 << 6,
 395         EXEC_IS_CONTROL            = 1 << 7,
 396         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 */
 397         EXEC_WRITE_CREDENTIALS     = 1 << 9, /* Set up the credential store logic */
 398
 399         /* The following are not used by execute.c, but by consumers internally */
 400         EXEC_PASS_FDS              = 1 << 10,
 401         EXEC_SETENV_RESULT         = 1 << 11,
 402         EXEC_SET_WATCHDOG          = 1 << 12,
 403         EXEC_SETENV_MONITOR_RESULT = 1 << 13, /* Pass exit status to OnFailure= and OnSuccess= dependencies. */
 404 } ExecFlags;
 405
 406 /* Parameters for a specific invocation of a command. This structure is put together right before a command is
 407  * executed. */
 408 struct ExecParameters {
 409         char **environment;
 410
 411         int *fds;
 412         char **fd_names;
 413         size_t n_socket_fds;
 414         size_t n_storage_fds;
 415
 416         ExecFlags flags;
 417         bool selinux_context_net:1;
 418
 419         CGroupMask cgroup_supported;
 420         const char *cgroup_path;
 421
 422         char **prefix;
 423         const char *received_credentials_directory;
 424         const char *received_encrypted_credentials_directory;
 425
 426         const char *confirm_spawn;
 427
 428         usec_t watchdog_usec;
 429
 430         int *idle_pipe;
 431
 432         int stdin_fd;
 433         int stdout_fd;
 434         int stderr_fd;
 435
 436         /* An fd that is closed by the execve(), and thus will result in EOF when the execve() is done */
 437         int exec_fd;
 438
 439         const char *notify_socket;
 440
 441         LIST_HEAD(OpenFile, open_files);
 442 };
 443
 444 #include "unit.h"
 445 #include "dynamic-user.h"
 446
 447 int exec_spawn(Unit *unit,
 448                ExecCommand *command,
 449                const ExecContext *context,
 450                const ExecParameters *exec_params,
 451                ExecRuntime *runtime,
 452                const CGroupContext *cgroup_context,
 453                pid_t *ret);
 454
 455 void exec_command_done_array(ExecCommand *c, size_t n);
 456 ExecCommand* exec_command_free_list(ExecCommand *c);
 457 void exec_command_free_array(ExecCommand **c, size_t n);
 458 void exec_command_reset_status_array(ExecCommand *c, size_t n);
 459 void exec_command_reset_status_list_array(ExecCommand **c, size_t n);
 460 void exec_command_dump_list(ExecCommand *c, FILE *f, const char *prefix);
 461 void exec_command_append_list(ExecCommand **l, ExecCommand *e);
 462 int exec_command_set(ExecCommand *c, const char *path, ...) _sentinel_;
 463 int exec_command_append(ExecCommand *c, const char *path, ...) _sentinel_;
 464
 465 void exec_context_init(ExecContext *c);
 466 void exec_context_done(ExecContext *c);
 467 void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix);
 468
 469 int exec_context_destroy_runtime_directory(const ExecContext *c, const char *runtime_root);
 470 int exec_context_destroy_credentials(const ExecContext *c, const char *runtime_root, const char *unit);
 471 int exec_context_destroy_mount_ns_dir(Unit *u);
 472
 473 const char* exec_context_fdname(const ExecContext *c, int fd_index);
 474
 475 bool exec_context_may_touch_console(const ExecContext *c);
 476 bool exec_context_maintains_privileges(const ExecContext *c);
 477 bool exec_context_has_encrypted_credentials(ExecContext *c);
 478
 479 int exec_context_get_effective_ioprio(const ExecContext *c);
 480 bool exec_context_get_effective_mount_apivfs(const ExecContext *c);
 481
 482 void exec_context_free_log_extra_fields(ExecContext *c);
 483
 484 void exec_context_revert_tty(ExecContext *c);
 485
 486 int exec_context_get_clean_directories(ExecContext *c, char **prefix, ExecCleanMask mask, char ***ret);
 487 int exec_context_get_clean_mask(ExecContext *c, ExecCleanMask *ret);
 488
 489 void exec_status_start(ExecStatus *s, pid_t pid);
 490 void exec_status_exit(ExecStatus *s, const ExecContext *context, pid_t pid, int code, int status);
 491 void exec_status_dump(const ExecStatus *s, FILE *f, const char *prefix);
 492 void exec_status_reset(ExecStatus *s);
 493
 494 int exec_shared_runtime_acquire(Manager *m, const ExecContext *c, const char *name, bool create, ExecSharedRuntime **ret);
 495 ExecSharedRuntime *exec_shared_runtime_destroy(ExecSharedRuntime *r);
 496 ExecSharedRuntime *exec_shared_runtime_unref(ExecSharedRuntime *r);
 497 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSharedRuntime*, exec_shared_runtime_unref);
 498
 499 int exec_shared_runtime_serialize(const Manager *m, FILE *f, FDSet *fds);
 500 int exec_shared_runtime_deserialize_compat(Unit *u, const char *key, const char *value, FDSet *fds);
 501 int exec_shared_runtime_deserialize_one(Manager *m, const char *value, FDSet *fds);
 502 void exec_shared_runtime_vacuum(Manager *m);
 503
 504 int exec_runtime_make(ExecSharedRuntime *shared, DynamicCreds *creds, ExecRuntime **ret);
 505 ExecRuntime* exec_runtime_free(ExecRuntime *rt);
 506 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecRuntime*, exec_runtime_free);
 507 ExecRuntime* exec_runtime_destroy(ExecRuntime *rt);
 508
 509 void exec_params_clear(ExecParameters *p);
 510
 511 bool exec_context_get_cpu_affinity_from_numa(const ExecContext *c);
 512
 513 ExecSetCredential *exec_set_credential_free(ExecSetCredential *sc);
 514 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecSetCredential*, exec_set_credential_free);
 515
 516 ExecLoadCredential *exec_load_credential_free(ExecLoadCredential *lc);
 517 DEFINE_TRIVIAL_CLEANUP_FUNC(ExecLoadCredential*, exec_load_credential_free);
 518
 519 void exec_directory_done(ExecDirectory *d);
 520 int exec_directory_add(ExecDirectory *d, const char *path, const char *symlink);
 521 void exec_directory_sort(ExecDirectory *d);
 522
 523 extern const struct hash_ops exec_set_credential_hash_ops;
 524 extern const struct hash_ops exec_load_credential_hash_ops;
 525
 526 const char* exec_output_to_string(ExecOutput i) _const_;
 527 ExecOutput exec_output_from_string(const char *s) _pure_;
 528
 529 const char* exec_input_to_string(ExecInput i) _const_;
 530 ExecInput exec_input_from_string(const char *s) _pure_;
 531
 532 const char* exec_utmp_mode_to_string(ExecUtmpMode i) _const_;
 533 ExecUtmpMode exec_utmp_mode_from_string(const char *s) _pure_;
 534
 535 const char* exec_preserve_mode_to_string(ExecPreserveMode i) _const_;
 536 ExecPreserveMode exec_preserve_mode_from_string(const char *s) _pure_;
 537
 538 const char* exec_keyring_mode_to_string(ExecKeyringMode i) _const_;
 539 ExecKeyringMode exec_keyring_mode_from_string(const char *s) _pure_;
 540
 541 const char* exec_directory_type_to_string(ExecDirectoryType i) _const_;
 542 ExecDirectoryType exec_directory_type_from_string(const char *s) _pure_;
 543
 544 const char* exec_directory_type_symlink_to_string(ExecDirectoryType i) _const_;
 545 ExecDirectoryType exec_directory_type_symlink_from_string(const char *s) _pure_;
 546
 547 const char* exec_resource_type_to_string(ExecDirectoryType i) _const_;
 548 ExecDirectoryType exec_resource_type_from_string(const char *s) _pure_;
 549
 550 bool exec_needs_mount_namespace(const ExecContext *context, const ExecParameters *params, const ExecRuntime *runtime);
 551 bool exec_needs_network_namespace(const ExecContext *context);