1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 Copyright 2012 Lennart Poettering
10 #include <sys/xattr.h>
15 #include <elfutils/libdwfl.h>
18 #include "sd-daemon.h"
19 #include "sd-journal.h"
21 #include "sd-messages.h"
24 #include "alloc-util.h"
25 #include "capability-util.h"
26 #include "cgroup-util.h"
28 #include "conf-parser.h"
30 #include "coredump-vacuum.h"
31 #include "dirent-util.h"
37 #include "journal-importer.h"
42 #include "parse-util.h"
43 #include "process-util.h"
44 #include "signal-util.h"
45 #include "socket-util.h"
47 #include "stacktrace.h"
48 #include "string-table.h"
49 #include "string-util.h"
51 #include "user-util.h"
54 /* The maximum size up to which we process coredumps */
55 #define PROCESS_SIZE_MAX ((uint64_t) (2LLU*1024LLU*1024LLU*1024LLU))
57 /* The maximum size up to which we leave the coredump around on disk */
58 #define EXTERNAL_SIZE_MAX PROCESS_SIZE_MAX
60 /* The maximum size up to which we store the coredump in the journal */
61 #define JOURNAL_SIZE_MAX ((size_t) (767LU*1024LU*1024LU))
63 /* Make sure to not make this larger than the maximum journal entry
64 * size. See DATA_SIZE_MAX in journald-native.c. */
65 assert_cc(JOURNAL_SIZE_MAX
<= DATA_SIZE_MAX
);
68 /* We use this as array indexes for a couple of special fields we use for
69 * naming coredump files, and attaching xattrs, and for indexing argv[].
71 * Our pattern for man:systectl(1) kernel.core_pattern is such that the
72 * kernel passes fields until CONTEXT_RLIMIT as arguments in argv[]. After
73 * that it gets complicated: the kernel passes "comm" as one or more fields
74 * starting at index CONTEXT_COMM (in other words, full "comm" is under index
75 * CONTEXT_COMM when it does not contain spaces, which is the common
76 * case). This mapping is not reversible, so we prefer to retrieve "comm"
77 * from /proc. We only fall back to argv[CONTEXT_COMM...] when that fails.
79 * In the internal context[] array, fields before CONTEXT_COMM are the
80 * strings from argv[], so they should not be freed. The strings at indices
81 * CONTEXT_COMM and higher are allocated by us and should be freed at the
97 typedef enum CoredumpStorage
{
98 COREDUMP_STORAGE_NONE
,
99 COREDUMP_STORAGE_EXTERNAL
,
100 COREDUMP_STORAGE_JOURNAL
,
101 _COREDUMP_STORAGE_MAX
,
102 _COREDUMP_STORAGE_INVALID
= -1
105 static const char* const coredump_storage_table
[_COREDUMP_STORAGE_MAX
] = {
106 [COREDUMP_STORAGE_NONE
] = "none",
107 [COREDUMP_STORAGE_EXTERNAL
] = "external",
108 [COREDUMP_STORAGE_JOURNAL
] = "journal",
111 DEFINE_PRIVATE_STRING_TABLE_LOOKUP(coredump_storage
, CoredumpStorage
);
112 static DEFINE_CONFIG_PARSE_ENUM(config_parse_coredump_storage
, coredump_storage
, CoredumpStorage
, "Failed to parse storage setting");
114 static CoredumpStorage arg_storage
= COREDUMP_STORAGE_EXTERNAL
;
115 static bool arg_compress
= true;
116 static uint64_t arg_process_size_max
= PROCESS_SIZE_MAX
;
117 static uint64_t arg_external_size_max
= EXTERNAL_SIZE_MAX
;
118 static uint64_t arg_journal_size_max
= JOURNAL_SIZE_MAX
;
119 static uint64_t arg_keep_free
= (uint64_t) -1;
120 static uint64_t arg_max_use
= (uint64_t) -1;
122 static int parse_config(void) {
123 static const ConfigTableItem items
[] = {
124 { "Coredump", "Storage", config_parse_coredump_storage
, 0, &arg_storage
},
125 { "Coredump", "Compress", config_parse_bool
, 0, &arg_compress
},
126 { "Coredump", "ProcessSizeMax", config_parse_iec_uint64
, 0, &arg_process_size_max
},
127 { "Coredump", "ExternalSizeMax", config_parse_iec_uint64
, 0, &arg_external_size_max
},
128 { "Coredump", "JournalSizeMax", config_parse_iec_size
, 0, &arg_journal_size_max
},
129 { "Coredump", "KeepFree", config_parse_iec_uint64
, 0, &arg_keep_free
},
130 { "Coredump", "MaxUse", config_parse_iec_uint64
, 0, &arg_max_use
},
134 return config_parse_many_nulstr(PKGSYSCONFDIR
"/coredump.conf",
135 CONF_PATHS_NULSTR("systemd/coredump.conf.d"),
137 config_item_table_lookup
, items
,
138 CONFIG_PARSE_WARN
, NULL
);
141 static inline uint64_t storage_size_max(void) {
142 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
143 return arg_external_size_max
;
144 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
)
145 return arg_journal_size_max
;
146 assert(arg_storage
== COREDUMP_STORAGE_NONE
);
150 static int fix_acl(int fd
, uid_t uid
) {
153 _cleanup_(acl_freep
) acl_t acl
= NULL
;
155 acl_permset_t permset
;
160 if (uid_is_system(uid
) || uid_is_dynamic(uid
) || uid
== UID_NOBODY
)
163 /* Make sure normal users can read (but not write or delete)
164 * their own coredumps */
166 acl
= acl_get_fd(fd
);
168 return log_error_errno(errno
, "Failed to get ACL: %m");
170 if (acl_create_entry(&acl
, &entry
) < 0 ||
171 acl_set_tag_type(entry
, ACL_USER
) < 0 ||
172 acl_set_qualifier(entry
, &uid
) < 0)
173 return log_error_errno(errno
, "Failed to patch ACL: %m");
175 if (acl_get_permset(entry
, &permset
) < 0 ||
176 acl_add_perm(permset
, ACL_READ
) < 0)
177 return log_warning_errno(errno
, "Failed to patch ACL: %m");
179 r
= calc_acl_mask_if_needed(&acl
);
181 return log_warning_errno(r
, "Failed to patch ACL: %m");
183 if (acl_set_fd(fd
, acl
) < 0)
184 return log_error_errno(errno
, "Failed to apply ACL: %m");
190 static int fix_xattr(int fd
, const char *context
[_CONTEXT_MAX
]) {
192 static const char * const xattrs
[_CONTEXT_MAX
] = {
193 [CONTEXT_PID
] = "user.coredump.pid",
194 [CONTEXT_UID
] = "user.coredump.uid",
195 [CONTEXT_GID
] = "user.coredump.gid",
196 [CONTEXT_SIGNAL
] = "user.coredump.signal",
197 [CONTEXT_TIMESTAMP
] = "user.coredump.timestamp",
198 [CONTEXT_RLIMIT
] = "user.coredump.rlimit",
199 [CONTEXT_HOSTNAME
] = "user.coredump.hostname",
200 [CONTEXT_COMM
] = "user.coredump.comm",
201 [CONTEXT_EXE
] = "user.coredump.exe",
209 /* Attach some metadata to coredumps via extended
210 * attributes. Just because we can. */
212 for (i
= 0; i
< _CONTEXT_MAX
; i
++) {
215 if (isempty(context
[i
]) || !xattrs
[i
])
218 k
= fsetxattr(fd
, xattrs
[i
], context
[i
], strlen(context
[i
]), XATTR_CREATE
);
226 #define filename_escape(s) xescape((s), "./ ")
228 static inline const char *coredump_tmpfile_name(const char *s
) {
229 return s
? s
: "(unnamed temporary file)";
232 static int fix_permissions(
234 const char *filename
,
236 const char *context
[_CONTEXT_MAX
],
245 /* Ignore errors on these */
246 (void) fchmod(fd
, 0640);
247 (void) fix_acl(fd
, uid
);
248 (void) fix_xattr(fd
, context
);
251 return log_error_errno(errno
, "Failed to sync coredump %s: %m", coredump_tmpfile_name(filename
));
253 (void) fsync_directory_of_file(fd
);
255 r
= link_tmpfile(fd
, filename
, target
);
257 return log_error_errno(r
, "Failed to move coredump %s into place: %m", target
);
262 static int maybe_remove_external_coredump(const char *filename
, uint64_t size
) {
264 /* Returns 1 if might remove, 0 if will not remove, < 0 on error. */
266 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
&&
267 size
<= arg_external_size_max
)
273 if (unlink(filename
) < 0 && errno
!= ENOENT
)
274 return log_error_errno(errno
, "Failed to unlink %s: %m", filename
);
279 static int make_filename(const char *context
[_CONTEXT_MAX
], char **ret
) {
280 _cleanup_free_
char *c
= NULL
, *u
= NULL
, *p
= NULL
, *t
= NULL
;
281 sd_id128_t boot
= {};
286 c
= filename_escape(context
[CONTEXT_COMM
]);
290 u
= filename_escape(context
[CONTEXT_UID
]);
294 r
= sd_id128_get_boot(&boot
);
298 p
= filename_escape(context
[CONTEXT_PID
]);
302 t
= filename_escape(context
[CONTEXT_TIMESTAMP
]);
307 "/var/lib/systemd/coredump/core.%s.%s." SD_ID128_FORMAT_STR
".%s.%s000000",
310 SD_ID128_FORMAT_VAL(boot
),
318 static int save_external_coredump(
319 const char *context
[_CONTEXT_MAX
],
325 bool *ret_truncated
) {
327 _cleanup_free_
char *fn
= NULL
, *tmp
= NULL
;
328 _cleanup_close_
int fd
= -1;
329 uint64_t rlimit
, process_limit
, max_size
;
335 assert(ret_filename
);
340 r
= parse_uid(context
[CONTEXT_UID
], &uid
);
342 return log_error_errno(r
, "Failed to parse UID: %m");
344 r
= safe_atou64(context
[CONTEXT_RLIMIT
], &rlimit
);
346 return log_error_errno(r
, "Failed to parse resource limit: %s", context
[CONTEXT_RLIMIT
]);
347 if (rlimit
< page_size()) {
348 /* Is coredumping disabled? Then don't bother saving/processing the coredump.
349 * Anything below PAGE_SIZE cannot give a readable coredump (the kernel uses
350 * ELF_EXEC_PAGESIZE which is not easily accessible, but is usually the same as PAGE_SIZE. */
351 log_info("Resource limits disable core dumping for process %s (%s).",
352 context
[CONTEXT_PID
], context
[CONTEXT_COMM
]);
356 process_limit
= MAX(arg_process_size_max
, storage_size_max());
357 if (process_limit
== 0) {
358 log_debug("Limits for coredump processing and storage are both 0, not dumping core.");
362 /* Never store more than the process configured, or than we actually shall keep or process */
363 max_size
= MIN(rlimit
, process_limit
);
365 r
= make_filename(context
, &fn
);
367 return log_error_errno(r
, "Failed to determine coredump file name: %m");
369 mkdir_p_label("/var/lib/systemd/coredump", 0755);
371 fd
= open_tmpfile_linkable(fn
, O_RDWR
|O_CLOEXEC
, &tmp
);
373 return log_error_errno(fd
, "Failed to create temporary file for coredump %s: %m", fn
);
375 r
= copy_bytes(input_fd
, fd
, max_size
, 0);
377 log_error_errno(r
, "Cannot store coredump of %s (%s): %m", context
[CONTEXT_PID
], context
[CONTEXT_COMM
]);
380 *ret_truncated
= r
== 1;
383 LOG_MESSAGE("Core file was truncated to %zu bytes.", max_size
),
384 "SIZE_LIMIT=%zu", max_size
,
385 "MESSAGE_ID=" SD_MESSAGE_TRUNCATED_CORE_STR
);
387 if (fstat(fd
, &st
) < 0) {
388 log_error_errno(errno
, "Failed to fstat core file %s: %m", coredump_tmpfile_name(tmp
));
392 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1) {
393 log_error_errno(errno
, "Failed to seek on %s: %m", coredump_tmpfile_name(tmp
));
397 #if HAVE_XZ || HAVE_LZ4
398 /* If we will remove the coredump anyway, do not compress. */
399 if (arg_compress
&& !maybe_remove_external_coredump(NULL
, st
.st_size
)) {
401 _cleanup_free_
char *fn_compressed
= NULL
, *tmp_compressed
= NULL
;
402 _cleanup_close_
int fd_compressed
= -1;
404 fn_compressed
= strappend(fn
, COMPRESSED_EXT
);
405 if (!fn_compressed
) {
410 fd_compressed
= open_tmpfile_linkable(fn_compressed
, O_RDWR
|O_CLOEXEC
, &tmp_compressed
);
411 if (fd_compressed
< 0) {
412 log_error_errno(fd_compressed
, "Failed to create temporary file for coredump %s: %m", fn_compressed
);
416 r
= compress_stream(fd
, fd_compressed
, -1);
418 log_error_errno(r
, "Failed to compress %s: %m", coredump_tmpfile_name(tmp_compressed
));
419 goto fail_compressed
;
422 r
= fix_permissions(fd_compressed
, tmp_compressed
, fn_compressed
, context
, uid
);
424 goto fail_compressed
;
426 /* OK, this worked, we can get rid of the uncompressed version now */
430 *ret_filename
= TAKE_PTR(fn_compressed
); /* compressed */
431 *ret_node_fd
= TAKE_FD(fd_compressed
); /* compressed */
432 *ret_data_fd
= TAKE_FD(fd
); /* uncompressed */
433 *ret_size
= (uint64_t) st
.st_size
; /* uncompressed */
439 (void) unlink(tmp_compressed
);
445 r
= fix_permissions(fd
, tmp
, fn
, context
, uid
);
449 *ret_filename
= TAKE_PTR(fn
);
450 *ret_data_fd
= TAKE_FD(fd
);
452 *ret_size
= (uint64_t) st
.st_size
;
462 static int allocate_journal_field(int fd
, size_t size
, char **ret
, size_t *ret_size
) {
463 _cleanup_free_
char *field
= NULL
;
470 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
471 return log_warning_errno(errno
, "Failed to seek: %m");
473 field
= malloc(9 + size
);
475 log_warning("Failed to allocate memory for coredump, coredump will not be stored.");
479 memcpy(field
, "COREDUMP=", 9);
481 n
= read(fd
, field
+ 9, size
);
483 return log_error_errno((int) n
, "Failed to read core data: %m");
484 if ((size_t) n
< size
) {
485 log_error("Core data too short.");
489 *ret
= TAKE_PTR(field
);
490 *ret_size
= size
+ 9;
495 /* Joins /proc/[pid]/fd/ and /proc/[pid]/fdinfo/ into the following lines:
509 static int compose_open_fds(pid_t pid
, char **open_fds
) {
510 _cleanup_closedir_
DIR *proc_fd_dir
= NULL
;
511 _cleanup_close_
int proc_fdinfo_fd
= -1;
512 _cleanup_free_
char *buffer
= NULL
;
513 _cleanup_fclose_
FILE *stream
= NULL
;
514 const char *fddelim
= "", *path
;
515 struct dirent
*dent
= NULL
;
520 assert(open_fds
!= NULL
);
522 path
= procfs_file_alloca(pid
, "fd");
523 proc_fd_dir
= opendir(path
);
527 proc_fdinfo_fd
= openat(dirfd(proc_fd_dir
), "../fdinfo", O_DIRECTORY
|O_NOFOLLOW
|O_CLOEXEC
|O_PATH
);
528 if (proc_fdinfo_fd
< 0)
531 stream
= open_memstream(&buffer
, &size
);
535 (void) __fsetlocking(stream
, FSETLOCKING_BYCALLER
);
537 FOREACH_DIRENT(dent
, proc_fd_dir
, return -errno
) {
538 _cleanup_fclose_
FILE *fdinfo
= NULL
;
539 _cleanup_free_
char *fdname
= NULL
;
543 r
= readlinkat_malloc(dirfd(proc_fd_dir
), dent
->d_name
, &fdname
);
547 fprintf(stream
, "%s%s:%s\n", fddelim
, dent
->d_name
, fdname
);
550 /* Use the directory entry from /proc/[pid]/fd with /proc/[pid]/fdinfo */
551 fd
= openat(proc_fdinfo_fd
, dent
->d_name
, O_NOFOLLOW
|O_CLOEXEC
|O_RDONLY
);
555 fdinfo
= fdopen(fd
, "re");
561 FOREACH_LINE(line
, fdinfo
, break) {
563 if (!endswith(line
, "\n"))
569 stream
= safe_fclose(stream
);
574 *open_fds
= TAKE_PTR(buffer
);
579 static int get_process_ns(pid_t pid
, const char *namespace, ino_t
*ns
) {
582 _cleanup_close_
int proc_ns_dir_fd
;
584 p
= procfs_file_alloca(pid
, "ns");
586 proc_ns_dir_fd
= open(p
, O_DIRECTORY
| O_CLOEXEC
| O_RDONLY
);
587 if (proc_ns_dir_fd
< 0)
590 if (fstatat(proc_ns_dir_fd
, namespace, &stbuf
, /* flags */0) < 0)
597 static int get_mount_namespace_leader(pid_t pid
, pid_t
*container_pid
) {
598 pid_t cpid
= pid
, ppid
= 0;
602 r
= get_process_ns(pid
, "mnt", &proc_mntns
);
609 r
= get_process_ppid(cpid
, &ppid
);
613 r
= get_process_ns(ppid
, "mnt", &parent_mntns
);
617 if (proc_mntns
!= parent_mntns
)
626 *container_pid
= ppid
;
630 /* Returns 1 if the parent was found.
631 * Returns 0 if there is not a process we can call the pid's
632 * container parent (the pid's process isn't 'containerized').
633 * Returns a negative number on errors.
635 static int get_process_container_parent_cmdline(pid_t pid
, char** cmdline
) {
638 const char *proc_root_path
;
639 struct stat root_stat
, proc_root_stat
;
641 /* To compare inodes of / and /proc/[pid]/root */
642 if (stat("/", &root_stat
) < 0)
645 proc_root_path
= procfs_file_alloca(pid
, "root");
646 if (stat(proc_root_path
, &proc_root_stat
) < 0)
649 /* The process uses system root. */
650 if (proc_root_stat
.st_ino
== root_stat
.st_ino
) {
655 r
= get_mount_namespace_leader(pid
, &container_pid
);
659 r
= get_process_cmdline(container_pid
, 0, false, cmdline
);
666 static int change_uid_gid(const char *context
[]) {
671 r
= parse_uid(context
[CONTEXT_UID
], &uid
);
675 if (uid
<= SYSTEM_UID_MAX
) {
676 const char *user
= "systemd-coredump";
678 r
= get_user_creds(&user
, &uid
, &gid
, NULL
, NULL
);
680 log_warning_errno(r
, "Cannot resolve %s user. Proceeding to dump core as root: %m", user
);
684 r
= parse_gid(context
[CONTEXT_GID
], &gid
);
689 return drop_privileges(uid
, gid
, 0);
692 static bool is_journald_crash(const char *context
[_CONTEXT_MAX
]) {
695 return streq_ptr(context
[CONTEXT_UNIT
], SPECIAL_JOURNALD_SERVICE
);
698 static bool is_pid1_crash(const char *context
[_CONTEXT_MAX
]) {
701 return streq_ptr(context
[CONTEXT_UNIT
], SPECIAL_INIT_SCOPE
) ||
702 streq_ptr(context
[CONTEXT_PID
], "1");
705 #define SUBMIT_COREDUMP_FIELDS 4
707 static int submit_coredump(
708 const char *context
[_CONTEXT_MAX
],
710 size_t n_iovec_allocated
,
714 _cleanup_close_
int coredump_fd
= -1, coredump_node_fd
= -1;
715 _cleanup_free_
char *core_message
= NULL
, *filename
= NULL
, *coredump_data
= NULL
;
716 uint64_t coredump_size
= UINT64_MAX
;
717 bool truncated
= false, journald_crash
;
722 assert(n_iovec_allocated
>= n_iovec
+ SUBMIT_COREDUMP_FIELDS
);
723 assert(input_fd
>= 0);
725 journald_crash
= is_journald_crash(context
);
727 /* Vacuum before we write anything again */
728 (void) coredump_vacuum(-1, arg_keep_free
, arg_max_use
);
730 /* Always stream the coredump to disk, if that's possible */
731 r
= save_external_coredump(context
, input_fd
,
732 &filename
, &coredump_node_fd
, &coredump_fd
, &coredump_size
, &truncated
);
734 /* Skip whole core dumping part */
737 /* If we don't want to keep the coredump on disk, remove it now, as later on we will lack the privileges for
738 * it. However, we keep the fd to it, so that we can still process it and log it. */
739 r
= maybe_remove_external_coredump(filename
, coredump_size
);
743 const char *coredump_filename
;
745 coredump_filename
= strjoina("COREDUMP_FILENAME=", filename
);
746 iovec
[n_iovec
++] = IOVEC_MAKE_STRING(coredump_filename
);
747 } else if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
748 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
749 coredump_size
, arg_external_size_max
);
751 /* Vacuum again, but exclude the coredump we just created */
752 (void) coredump_vacuum(coredump_node_fd
>= 0 ? coredump_node_fd
: coredump_fd
, arg_keep_free
, arg_max_use
);
754 /* Now, let's drop privileges to become the user who owns the segfaulted process and allocate the coredump
755 * memory under the user's uid. This also ensures that the credentials journald will see are the ones of the
756 * coredumping user, thus making sure the user gets access to the core dump. Let's also get rid of all
757 * capabilities, if we run as root, we won't need them anymore. */
758 r
= change_uid_gid(context
);
760 return log_error_errno(r
, "Failed to drop privileges: %m");
763 /* Try to get a strack trace if we can */
764 if (coredump_size
<= arg_process_size_max
) {
765 _cleanup_free_
char *stacktrace
= NULL
;
767 r
= coredump_make_stack_trace(coredump_fd
, context
[CONTEXT_EXE
], &stacktrace
);
769 core_message
= strjoin("MESSAGE=Process ", context
[CONTEXT_PID
],
770 " (", context
[CONTEXT_COMM
], ") of user ",
771 context
[CONTEXT_UID
], " dumped core.",
772 journald_crash
? "\nCoredump diverted to " : "",
773 journald_crash
? filename
: "",
775 else if (r
== -EINVAL
)
776 log_warning("Failed to generate stack trace: %s", dwfl_errmsg(dwfl_errno()));
778 log_warning_errno(r
, "Failed to generate stack trace: %m");
780 log_debug("Not generating stack trace: core size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
781 coredump_size
, arg_process_size_max
);
786 core_message
= strjoin("MESSAGE=Process ", context
[CONTEXT_PID
],
787 " (", context
[CONTEXT_COMM
], ") of user ",
788 context
[CONTEXT_UID
], " dumped core.",
789 journald_crash
? "\nCoredump diverted to " : NULL
,
790 journald_crash
? filename
: NULL
);
794 if (journald_crash
) {
795 /* We cannot log to the journal, so just print the MESSAGE.
796 * The target was set previously to something safe. */
797 log_dispatch(LOG_ERR
, 0, core_message
);
801 iovec
[n_iovec
++] = IOVEC_MAKE_STRING(core_message
);
804 iovec
[n_iovec
++] = IOVEC_MAKE_STRING("COREDUMP_TRUNCATED=1");
806 /* Optionally store the entire coredump in the journal */
807 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
) {
808 if (coredump_size
<= arg_journal_size_max
) {
811 /* Store the coredump itself in the journal */
813 r
= allocate_journal_field(coredump_fd
, (size_t) coredump_size
, &coredump_data
, &sz
);
815 iovec
[n_iovec
++] = IOVEC_MAKE(coredump_data
, sz
);
817 log_warning_errno(r
, "Failed to attach the core to the journal entry: %m");
819 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
820 coredump_size
, arg_journal_size_max
);
823 assert(n_iovec
<= n_iovec_allocated
);
825 r
= sd_journal_sendv(iovec
, n_iovec
);
827 return log_error_errno(r
, "Failed to log coredump: %m");
832 static void map_context_fields(const struct iovec
*iovec
, const char* context
[]) {
834 static const char * const context_field_names
[] = {
835 [CONTEXT_PID
] = "COREDUMP_PID=",
836 [CONTEXT_UID
] = "COREDUMP_UID=",
837 [CONTEXT_GID
] = "COREDUMP_GID=",
838 [CONTEXT_SIGNAL
] = "COREDUMP_SIGNAL=",
839 [CONTEXT_TIMESTAMP
] = "COREDUMP_TIMESTAMP=",
840 [CONTEXT_RLIMIT
] = "COREDUMP_RLIMIT=",
841 [CONTEXT_HOSTNAME
] = "COREDUMP_HOSTNAME=",
842 [CONTEXT_COMM
] = "COREDUMP_COMM=",
843 [CONTEXT_EXE
] = "COREDUMP_EXE=",
851 for (i
= 0; i
< ELEMENTSOF(context_field_names
); i
++) {
854 if (!context_field_names
[i
])
857 p
= memory_startswith(iovec
->iov_base
, iovec
->iov_len
, context_field_names
[i
]);
861 /* Note that these strings are NUL terminated, because we made sure that a trailing NUL byte is in the
862 * buffer, though not included in the iov_len count. (see below) */
868 static int process_socket(int fd
) {
869 _cleanup_close_
int coredump_fd
= -1;
870 struct iovec
*iovec
= NULL
;
871 size_t n_iovec
= 0, n_allocated
= 0, i
, k
;
872 const char *context
[_CONTEXT_MAX
] = {};
877 log_set_target(LOG_TARGET_AUTO
);
878 log_parse_environment();
881 log_debug("Processing coredump received on stdin...");
885 struct cmsghdr cmsghdr
;
886 uint8_t buf
[CMSG_SPACE(sizeof(int))];
889 .msg_control
= &control
,
890 .msg_controllen
= sizeof(control
),
896 if (!GREEDY_REALLOC(iovec
, n_allocated
, n_iovec
+ SUBMIT_COREDUMP_FIELDS
)) {
901 l
= next_datagram_size_fd(fd
);
903 r
= log_error_errno(l
, "Failed to determine datagram size to read: %m");
909 iovec
[n_iovec
].iov_len
= l
;
910 iovec
[n_iovec
].iov_base
= malloc(l
+ 1);
911 if (!iovec
[n_iovec
].iov_base
) {
916 mh
.msg_iov
= iovec
+ n_iovec
;
918 n
= recvmsg(fd
, &mh
, MSG_NOSIGNAL
|MSG_CMSG_CLOEXEC
);
920 free(iovec
[n_iovec
].iov_base
);
921 r
= log_error_errno(errno
, "Failed to receive datagram: %m");
926 struct cmsghdr
*cmsg
, *found
= NULL
;
927 /* The final zero-length datagram carries the file descriptor and tells us that we're done. */
929 free(iovec
[n_iovec
].iov_base
);
931 CMSG_FOREACH(cmsg
, &mh
) {
932 if (cmsg
->cmsg_level
== SOL_SOCKET
&&
933 cmsg
->cmsg_type
== SCM_RIGHTS
&&
934 cmsg
->cmsg_len
== CMSG_LEN(sizeof(int))) {
941 log_error("Coredump file descriptor missing.");
946 assert(coredump_fd
< 0);
947 coredump_fd
= *(int*) CMSG_DATA(found
);
951 /* Add trailing NUL byte, in case these are strings */
952 ((char*) iovec
[n_iovec
].iov_base
)[n
] = 0;
953 iovec
[n_iovec
].iov_len
= (size_t) n
;
956 map_context_fields(iovec
+ n_iovec
, context
);
960 if (!GREEDY_REALLOC(iovec
, n_allocated
, n_iovec
+ SUBMIT_COREDUMP_FIELDS
)) {
965 /* Make sure we got all data we really need */
966 assert(context
[CONTEXT_PID
]);
967 assert(context
[CONTEXT_UID
]);
968 assert(context
[CONTEXT_GID
]);
969 assert(context
[CONTEXT_SIGNAL
]);
970 assert(context
[CONTEXT_TIMESTAMP
]);
971 assert(context
[CONTEXT_RLIMIT
]);
972 assert(context
[CONTEXT_HOSTNAME
]);
973 assert(context
[CONTEXT_COMM
]);
974 assert(coredump_fd
>= 0);
976 /* Small quirk: the journal fields contain the timestamp padded with six zeroes, so that the kernel-supplied 1s
977 * granularity timestamps becomes 1µs granularity, i.e. the granularity systemd usually operates in. Since we
978 * are reconstructing the original kernel context, we chop this off again, here. */
979 k
= strlen(context
[CONTEXT_TIMESTAMP
]);
981 context
[CONTEXT_TIMESTAMP
] = strndupa(context
[CONTEXT_TIMESTAMP
], k
- 6);
983 r
= submit_coredump(context
, iovec
, n_allocated
, n_iovec
, coredump_fd
);
986 for (i
= 0; i
< n_iovec
; i
++)
987 free(iovec
[i
].iov_base
);
993 static int send_iovec(const struct iovec iovec
[], size_t n_iovec
, int input_fd
) {
995 static const union sockaddr_union sa
= {
996 .un
.sun_family
= AF_UNIX
,
997 .un
.sun_path
= "/run/systemd/coredump",
999 _cleanup_close_
int fd
= -1;
1003 assert(iovec
|| n_iovec
<= 0);
1004 assert(input_fd
>= 0);
1006 fd
= socket(AF_UNIX
, SOCK_SEQPACKET
|SOCK_CLOEXEC
, 0);
1008 return log_error_errno(errno
, "Failed to create coredump socket: %m");
1010 if (connect(fd
, &sa
.sa
, SOCKADDR_UN_LEN(sa
.un
)) < 0)
1011 return log_error_errno(errno
, "Failed to connect to coredump service: %m");
1013 for (i
= 0; i
< n_iovec
; i
++) {
1014 struct msghdr mh
= {
1015 .msg_iov
= (struct iovec
*) iovec
+ i
,
1018 struct iovec copy
[2];
1021 if (sendmsg(fd
, &mh
, MSG_NOSIGNAL
) >= 0)
1024 if (errno
== EMSGSIZE
&& mh
.msg_iov
[0].iov_len
> 0) {
1025 /* This field didn't fit? That's a pity. Given that this is just metadata,
1026 * let's truncate the field at half, and try again. We append three dots, in
1027 * order to show that this is truncated. */
1029 if (mh
.msg_iov
!= copy
) {
1030 /* We don't want to modify the caller's iovec, hence let's create our
1031 * own array, consisting of two new iovecs, where the first is a
1032 * (truncated) copy of what we want to send, and the second one
1033 * contains the trailing dots. */
1035 copy
[1] = (struct iovec
) {
1036 .iov_base
= (char[]) { '.', '.', '.' },
1044 copy
[0].iov_len
/= 2; /* halve it, and try again */
1048 return log_error_errno(errno
, "Failed to send coredump datagram: %m");
1052 r
= send_one_fd(fd
, input_fd
, 0);
1054 return log_error_errno(r
, "Failed to send coredump fd: %m");
1059 static char* set_iovec_field(struct iovec
*iovec
, size_t *n_iovec
, const char *field
, const char *value
) {
1062 x
= strappend(field
, value
);
1064 iovec
[(*n_iovec
)++] = IOVEC_MAKE_STRING(x
);
1068 static char* set_iovec_field_free(struct iovec
*iovec
, size_t *n_iovec
, const char *field
, char *value
) {
1071 x
= set_iovec_field(iovec
, n_iovec
, field
, value
);
1076 static int gather_pid_metadata(
1077 char* context
[_CONTEXT_MAX
],
1078 char **comm_fallback
,
1079 struct iovec
*iovec
, size_t *n_iovec
) {
1081 /* We need 27 empty slots in iovec!
1083 * Note that if we fail on oom later on, we do not roll-back changes to the iovec structure. (It remains valid,
1084 * with the first n_iovec fields initialized.) */
1092 r
= parse_pid(context
[CONTEXT_PID
], &pid
);
1094 return log_error_errno(r
, "Failed to parse PID \"%s\": %m", context
[CONTEXT_PID
]);
1096 r
= get_process_comm(pid
, &context
[CONTEXT_COMM
]);
1098 log_warning_errno(r
, "Failed to get COMM, falling back to the command line: %m");
1099 context
[CONTEXT_COMM
] = strv_join(comm_fallback
, " ");
1100 if (!context
[CONTEXT_COMM
])
1104 r
= get_process_exe(pid
, &context
[CONTEXT_EXE
]);
1106 log_warning_errno(r
, "Failed to get EXE, ignoring: %m");
1108 if (cg_pid_get_unit(pid
, &context
[CONTEXT_UNIT
]) >= 0) {
1109 if (!is_journald_crash((const char**) context
)) {
1110 /* OK, now we know it's not the journal, hence we can make use of it now. */
1111 log_set_target(LOG_TARGET_JOURNAL_OR_KMSG
);
1115 /* If this is PID 1 disable coredump collection, we'll unlikely be able to process it later on. */
1116 if (is_pid1_crash((const char**) context
)) {
1117 log_notice("Due to PID 1 having crashed coredump collection will now be turned off.");
1118 disable_coredumps();
1121 set_iovec_field(iovec
, n_iovec
, "COREDUMP_UNIT=", context
[CONTEXT_UNIT
]);
1124 if (cg_pid_get_user_unit(pid
, &t
) >= 0)
1125 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_USER_UNIT=", t
);
1127 /* The next few are mandatory */
1128 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_PID=", context
[CONTEXT_PID
]))
1131 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_UID=", context
[CONTEXT_UID
]))
1134 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_GID=", context
[CONTEXT_GID
]))
1137 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_SIGNAL=", context
[CONTEXT_SIGNAL
]))
1140 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_RLIMIT=", context
[CONTEXT_RLIMIT
]))
1143 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_HOSTNAME=", context
[CONTEXT_HOSTNAME
]))
1146 if (!set_iovec_field(iovec
, n_iovec
, "COREDUMP_COMM=", context
[CONTEXT_COMM
]))
1149 if (context
[CONTEXT_EXE
] &&
1150 !set_iovec_field(iovec
, n_iovec
, "COREDUMP_EXE=", context
[CONTEXT_EXE
]))
1153 if (sd_pid_get_session(pid
, &t
) >= 0)
1154 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_SESSION=", t
);
1156 if (sd_pid_get_owner_uid(pid
, &owner_uid
) >= 0) {
1157 r
= asprintf(&t
, "COREDUMP_OWNER_UID=" UID_FMT
, owner_uid
);
1159 iovec
[(*n_iovec
)++] = IOVEC_MAKE_STRING(t
);
1162 if (sd_pid_get_slice(pid
, &t
) >= 0)
1163 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_SLICE=", t
);
1165 if (get_process_cmdline(pid
, 0, false, &t
) >= 0)
1166 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_CMDLINE=", t
);
1168 if (cg_pid_get_path_shifted(pid
, NULL
, &t
) >= 0)
1169 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_CGROUP=", t
);
1171 if (compose_open_fds(pid
, &t
) >= 0)
1172 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_OPEN_FDS=", t
);
1174 p
= procfs_file_alloca(pid
, "status");
1175 if (read_full_file(p
, &t
, NULL
) >= 0)
1176 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_PROC_STATUS=", t
);
1178 p
= procfs_file_alloca(pid
, "maps");
1179 if (read_full_file(p
, &t
, NULL
) >= 0)
1180 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_PROC_MAPS=", t
);
1182 p
= procfs_file_alloca(pid
, "limits");
1183 if (read_full_file(p
, &t
, NULL
) >= 0)
1184 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_PROC_LIMITS=", t
);
1186 p
= procfs_file_alloca(pid
, "cgroup");
1187 if (read_full_file(p
, &t
, NULL
) >=0)
1188 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_PROC_CGROUP=", t
);
1190 p
= procfs_file_alloca(pid
, "mountinfo");
1191 if (read_full_file(p
, &t
, NULL
) >=0)
1192 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_PROC_MOUNTINFO=", t
);
1194 if (get_process_cwd(pid
, &t
) >= 0)
1195 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_CWD=", t
);
1197 if (get_process_root(pid
, &t
) >= 0) {
1198 bool proc_self_root_is_slash
;
1200 proc_self_root_is_slash
= strcmp(t
, "/") == 0;
1202 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_ROOT=", t
);
1204 /* If the process' root is "/", then there is a chance it has
1205 * mounted own root and hence being containerized. */
1206 if (proc_self_root_is_slash
&& get_process_container_parent_cmdline(pid
, &t
) > 0)
1207 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_CONTAINER_CMDLINE=", t
);
1210 if (get_process_environ(pid
, &t
) >= 0)
1211 set_iovec_field_free(iovec
, n_iovec
, "COREDUMP_ENVIRON=", t
);
1213 t
= strjoin("COREDUMP_TIMESTAMP=", context
[CONTEXT_TIMESTAMP
], "000000");
1215 iovec
[(*n_iovec
)++] = IOVEC_MAKE_STRING(t
);
1217 if (safe_atoi(context
[CONTEXT_SIGNAL
], &signo
) >= 0 && SIGNAL_VALID(signo
))
1218 set_iovec_field(iovec
, n_iovec
, "COREDUMP_SIGNAL_NAME=SIG", signal_to_string(signo
));
1220 return 0; /* we successfully acquired all metadata */
1223 static int process_kernel(int argc
, char* argv
[]) {
1225 char* context
[_CONTEXT_MAX
] = {};
1226 struct iovec iovec
[29 + SUBMIT_COREDUMP_FIELDS
];
1227 size_t i
, n_iovec
, n_to_free
= 0;
1230 log_debug("Processing coredump received from the kernel...");
1232 if (argc
< CONTEXT_COMM
+ 1) {
1233 log_error("Not enough arguments passed by the kernel (%i, expected %i).", argc
- 1, CONTEXT_COMM
+ 1 - 1);
1237 context
[CONTEXT_PID
] = argv
[1 + CONTEXT_PID
];
1238 context
[CONTEXT_UID
] = argv
[1 + CONTEXT_UID
];
1239 context
[CONTEXT_GID
] = argv
[1 + CONTEXT_GID
];
1240 context
[CONTEXT_SIGNAL
] = argv
[1 + CONTEXT_SIGNAL
];
1241 context
[CONTEXT_TIMESTAMP
] = argv
[1 + CONTEXT_TIMESTAMP
];
1242 context
[CONTEXT_RLIMIT
] = argv
[1 + CONTEXT_RLIMIT
];
1243 context
[CONTEXT_HOSTNAME
] = argv
[1 + CONTEXT_HOSTNAME
];
1245 r
= gather_pid_metadata(context
, argv
+ 1 + CONTEXT_COMM
, iovec
, &n_to_free
);
1249 n_iovec
= n_to_free
;
1251 iovec
[n_iovec
++] = IOVEC_MAKE_STRING("MESSAGE_ID=" SD_MESSAGE_COREDUMP_STR
);
1253 assert_cc(2 == LOG_CRIT
);
1254 iovec
[n_iovec
++] = IOVEC_MAKE_STRING("PRIORITY=2");
1256 assert(n_iovec
<= ELEMENTSOF(iovec
));
1258 if (is_journald_crash((const char**) context
) || is_pid1_crash((const char**) context
))
1259 r
= submit_coredump((const char**) context
,
1260 iovec
, ELEMENTSOF(iovec
), n_iovec
,
1263 r
= send_iovec(iovec
, n_iovec
, STDIN_FILENO
);
1266 for (i
= 0; i
< n_to_free
; i
++)
1267 free(iovec
[i
].iov_base
);
1269 /* Those fields are allocated by gather_pid_metadata */
1270 free(context
[CONTEXT_COMM
]);
1271 free(context
[CONTEXT_EXE
]);
1272 free(context
[CONTEXT_UNIT
]);
1277 static int process_backtrace(int argc
, char *argv
[]) {
1278 char *context
[_CONTEXT_MAX
] = {};
1279 _cleanup_free_
char *message
= NULL
;
1280 _cleanup_free_
struct iovec
*iovec
= NULL
;
1281 size_t n_iovec
, n_allocated
, n_to_free
= 0, i
;
1283 JournalImporter importer
= {
1287 log_debug("Processing backtrace on stdin...");
1289 if (argc
< CONTEXT_COMM
+ 1) {
1290 log_error("Not enough arguments passed (%i, expected %i).", argc
- 1, CONTEXT_COMM
+ 1 - 1);
1294 context
[CONTEXT_PID
] = argv
[2 + CONTEXT_PID
];
1295 context
[CONTEXT_UID
] = argv
[2 + CONTEXT_UID
];
1296 context
[CONTEXT_GID
] = argv
[2 + CONTEXT_GID
];
1297 context
[CONTEXT_SIGNAL
] = argv
[2 + CONTEXT_SIGNAL
];
1298 context
[CONTEXT_TIMESTAMP
] = argv
[2 + CONTEXT_TIMESTAMP
];
1299 context
[CONTEXT_RLIMIT
] = argv
[2 + CONTEXT_RLIMIT
];
1300 context
[CONTEXT_HOSTNAME
] = argv
[2 + CONTEXT_HOSTNAME
];
1302 n_allocated
= 34 + COREDUMP_STORAGE_EXTERNAL
;
1303 /* 26 metadata, 2 static, +unknown input, 4 storage, rounded up */
1304 iovec
= new(struct iovec
, n_allocated
);
1308 r
= gather_pid_metadata(context
, argv
+ 2 + CONTEXT_COMM
, iovec
, &n_to_free
);
1312 /* This was a special crash, and has already been processed. */
1316 n_iovec
= n_to_free
;
1319 r
= journal_importer_process_data(&importer
);
1321 log_error_errno(r
, "Failed to parse journal entry on stdin: %m");
1324 if (r
== 1 || /* complete entry */
1325 journal_importer_eof(&importer
)) /* end of data */
1329 if (!GREEDY_REALLOC(iovec
, n_allocated
, n_iovec
+ importer
.iovw
.count
+ 2))
1332 if (journal_importer_eof(&importer
)) {
1333 log_warning("Did not receive a full journal entry on stdin, ignoring message sent by reporter");
1335 message
= strjoin("MESSAGE=Process ", context
[CONTEXT_PID
],
1336 " (", context
[CONTEXT_COMM
], ")"
1337 " of user ", context
[CONTEXT_UID
],
1338 " failed with ", context
[CONTEXT_SIGNAL
]);
1343 iovec
[n_iovec
++] = IOVEC_MAKE_STRING(message
);
1345 for (i
= 0; i
< importer
.iovw
.count
; i
++)
1346 iovec
[n_iovec
++] = importer
.iovw
.iovec
[i
];
1349 iovec
[n_iovec
++] = IOVEC_MAKE_STRING("MESSAGE_ID=" SD_MESSAGE_BACKTRACE_STR
);
1350 assert_cc(2 == LOG_CRIT
);
1351 iovec
[n_iovec
++] = IOVEC_MAKE_STRING("PRIORITY=2");
1353 assert(n_iovec
<= n_allocated
);
1355 r
= sd_journal_sendv(iovec
, n_iovec
);
1357 log_error_errno(r
, "Failed to log backtrace: %m");
1360 for (i
= 0; i
< n_to_free
; i
++)
1361 free(iovec
[i
].iov_base
);
1363 /* Those fields are allocated by gather_pid_metadata */
1364 free(context
[CONTEXT_COMM
]);
1365 free(context
[CONTEXT_EXE
]);
1366 free(context
[CONTEXT_UNIT
]);
1371 int main(int argc
, char *argv
[]) {
1374 /* First, log to a safe place, since we don't know what crashed and it might
1375 * be journald which we'd rather not log to then. */
1377 log_set_target(LOG_TARGET_KMSG
);
1380 /* Make sure we never enter a loop */
1381 (void) prctl(PR_SET_DUMPABLE
, 0);
1383 /* Ignore all parse errors */
1384 (void) parse_config();
1386 log_debug("Selected storage '%s'.", coredump_storage_to_string(arg_storage
));
1387 log_debug("Selected compression %s.", yes_no(arg_compress
));
1389 r
= sd_listen_fds(false);
1391 log_error_errno(r
, "Failed to determine number of file descriptor: %m");
1395 /* If we got an fd passed, we are running in coredumpd mode. Otherwise we
1396 * are invoked from the kernel as coredump handler. */
1398 if (streq_ptr(argv
[1], "--backtrace"))
1399 r
= process_backtrace(argc
, argv
);
1401 r
= process_kernel(argc
, argv
);
1403 r
= process_socket(SD_LISTEN_FDS_START
);
1405 log_error("Received unexpected number of file descriptors.");
1410 return r
< 0 ? EXIT_FAILURE
: EXIT_SUCCESS
;