1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
6 #include <sys/statvfs.h>
10 #include "sd-daemon.h"
11 #include "sd-journal.h"
13 #include "sd-messages.h"
16 #include "alloc-util.h"
17 #include "bus-error.h"
18 #include "capability-util.h"
19 #include "cgroup-util.h"
21 #include "conf-parser.h"
23 #include "coredump-vacuum.h"
24 #include "dirent-util.h"
31 #include "journal-importer.h"
34 #include "main-func.h"
35 #include "memory-util.h"
36 #include "mkdir-label.h"
37 #include "parse-util.h"
38 #include "process-util.h"
39 #include "signal-util.h"
40 #include "socket-util.h"
42 #include "stat-util.h"
43 #include "string-table.h"
44 #include "string-util.h"
46 #include "sync-util.h"
47 #include "tmpfile-util.h"
48 #include "uid-alloc-range.h"
49 #include "user-util.h"
51 /* The maximum size up to which we process coredumps */
52 #define PROCESS_SIZE_MAX ((uint64_t) (2LLU*1024LLU*1024LLU*1024LLU))
54 /* The maximum size up to which we leave the coredump around on disk */
55 #define EXTERNAL_SIZE_MAX PROCESS_SIZE_MAX
57 /* The maximum size up to which we store the coredump in the journal */
58 #ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
59 #define JOURNAL_SIZE_MAX ((size_t) (767LU*1024LU*1024LU))
61 /* oss-fuzz limits memory usage. */
62 #define JOURNAL_SIZE_MAX ((size_t) (10LU*1024LU*1024LU))
65 /* When checking for available memory and setting lower limits, don't
66 * go below 4MB for writing core files to storage. */
67 #define PROCESS_SIZE_MIN (4U*1024U*1024U)
69 /* Make sure to not make this larger than the maximum journal entry
70 * size. See DATA_SIZE_MAX in journal-importer.h. */
71 assert_cc(JOURNAL_SIZE_MAX
<= DATA_SIZE_MAX
);
74 /* We use these as array indexes for our process metadata cache.
76 * The first indices of the cache stores the same metadata as the ones passed by
77 * the kernel via argv[], ie the strings array passed by the kernel according to
78 * our pattern defined in /proc/sys/kernel/core_pattern (see man:core(5)). */
80 META_ARGV_PID
, /* %P: as seen in the initial pid namespace */
81 META_ARGV_UID
, /* %u: as seen in the initial user namespace */
82 META_ARGV_GID
, /* %g: as seen in the initial user namespace */
83 META_ARGV_SIGNAL
, /* %s: number of signal causing dump */
84 META_ARGV_TIMESTAMP
, /* %t: time of dump, expressed as seconds since the Epoch (we expand this to µs granularity) */
85 META_ARGV_RLIMIT
, /* %c: core file size soft resource limit */
86 META_ARGV_HOSTNAME
, /* %h: hostname */
89 /* The following indexes are cached for a couple of special fields we use (and
90 * thereby need to be retrieved quickly) for naming coredump files, and attaching
91 * xattrs. Unlike the previous ones they are retrieved from the runtime
94 META_COMM
= _META_ARGV_MAX
,
97 /* The rest are similar to the previous ones except that we won't fail if one of
100 META_EXE
= _META_MANDATORY_MAX
,
105 static const char * const meta_field_names
[_META_MAX
] = {
106 [META_ARGV_PID
] = "COREDUMP_PID=",
107 [META_ARGV_UID
] = "COREDUMP_UID=",
108 [META_ARGV_GID
] = "COREDUMP_GID=",
109 [META_ARGV_SIGNAL
] = "COREDUMP_SIGNAL=",
110 [META_ARGV_TIMESTAMP
] = "COREDUMP_TIMESTAMP=",
111 [META_ARGV_RLIMIT
] = "COREDUMP_RLIMIT=",
112 [META_ARGV_HOSTNAME
] = "COREDUMP_HOSTNAME=",
113 [META_COMM
] = "COREDUMP_COMM=",
114 [META_EXE
] = "COREDUMP_EXE=",
115 [META_UNIT
] = "COREDUMP_UNIT=",
118 typedef struct Context
{
119 const char *meta
[_META_MAX
];
125 typedef enum CoredumpStorage
{
126 COREDUMP_STORAGE_NONE
,
127 COREDUMP_STORAGE_EXTERNAL
,
128 COREDUMP_STORAGE_JOURNAL
,
129 _COREDUMP_STORAGE_MAX
,
130 _COREDUMP_STORAGE_INVALID
= -EINVAL
,
133 static const char* const coredump_storage_table
[_COREDUMP_STORAGE_MAX
] = {
134 [COREDUMP_STORAGE_NONE
] = "none",
135 [COREDUMP_STORAGE_EXTERNAL
] = "external",
136 [COREDUMP_STORAGE_JOURNAL
] = "journal",
139 DEFINE_PRIVATE_STRING_TABLE_LOOKUP(coredump_storage
, CoredumpStorage
);
140 static DEFINE_CONFIG_PARSE_ENUM(config_parse_coredump_storage
, coredump_storage
, CoredumpStorage
, "Failed to parse storage setting");
142 static CoredumpStorage arg_storage
= COREDUMP_STORAGE_EXTERNAL
;
143 static bool arg_compress
= true;
144 static uint64_t arg_process_size_max
= PROCESS_SIZE_MAX
;
145 static uint64_t arg_external_size_max
= EXTERNAL_SIZE_MAX
;
146 static uint64_t arg_journal_size_max
= JOURNAL_SIZE_MAX
;
147 static uint64_t arg_keep_free
= UINT64_MAX
;
148 static uint64_t arg_max_use
= UINT64_MAX
;
150 static int parse_config(void) {
151 static const ConfigTableItem items
[] = {
152 { "Coredump", "Storage", config_parse_coredump_storage
, 0, &arg_storage
},
153 { "Coredump", "Compress", config_parse_bool
, 0, &arg_compress
},
154 { "Coredump", "ProcessSizeMax", config_parse_iec_uint64
, 0, &arg_process_size_max
},
155 { "Coredump", "ExternalSizeMax", config_parse_iec_uint64_infinity
, 0, &arg_external_size_max
},
156 { "Coredump", "JournalSizeMax", config_parse_iec_size
, 0, &arg_journal_size_max
},
157 { "Coredump", "KeepFree", config_parse_iec_uint64
, 0, &arg_keep_free
},
158 { "Coredump", "MaxUse", config_parse_iec_uint64
, 0, &arg_max_use
},
162 return config_parse_many_nulstr(
163 PKGSYSCONFDIR
"/coredump.conf",
164 CONF_PATHS_NULSTR("systemd/coredump.conf.d"),
166 config_item_table_lookup
, items
,
172 static uint64_t storage_size_max(void) {
173 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
174 return arg_external_size_max
;
175 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
)
176 return arg_journal_size_max
;
177 assert(arg_storage
== COREDUMP_STORAGE_NONE
);
181 static int fix_acl(int fd
, uid_t uid
) {
187 assert(uid_is_valid(uid
));
189 if (uid_is_system(uid
) || uid_is_dynamic(uid
) || uid
== UID_NOBODY
)
192 /* Make sure normal users can read (but not write or delete) their own coredumps */
193 r
= fd_add_uid_acl_permission(fd
, uid
, ACL_READ
);
195 return log_error_errno(r
, "Failed to adjust ACL of the coredump: %m");
201 static int fix_xattr(int fd
, const Context
*context
) {
203 static const char * const xattrs
[_META_MAX
] = {
204 [META_ARGV_PID
] = "user.coredump.pid",
205 [META_ARGV_UID
] = "user.coredump.uid",
206 [META_ARGV_GID
] = "user.coredump.gid",
207 [META_ARGV_SIGNAL
] = "user.coredump.signal",
208 [META_ARGV_TIMESTAMP
] = "user.coredump.timestamp",
209 [META_ARGV_RLIMIT
] = "user.coredump.rlimit",
210 [META_ARGV_HOSTNAME
] = "user.coredump.hostname",
211 [META_COMM
] = "user.coredump.comm",
212 [META_EXE
] = "user.coredump.exe",
219 /* Attach some metadata to coredumps via extended
220 * attributes. Just because we can. */
222 for (unsigned i
= 0; i
< _META_MAX
; i
++) {
225 if (isempty(context
->meta
[i
]) || !xattrs
[i
])
228 k
= fsetxattr(fd
, xattrs
[i
], context
->meta
[i
], strlen(context
->meta
[i
]), XATTR_CREATE
);
236 #define filename_escape(s) xescape((s), "./ ")
238 static const char *coredump_tmpfile_name(const char *s
) {
239 return s
? s
: "(unnamed temporary file)";
242 static int fix_permissions(
244 const char *filename
,
246 const Context
*context
,
255 /* Ignore errors on these */
256 (void) fchmod(fd
, 0640);
257 (void) fix_acl(fd
, uid
);
258 (void) fix_xattr(fd
, context
);
262 return log_error_errno(r
, "Failed to sync coredump %s: %m", coredump_tmpfile_name(filename
));
264 r
= link_tmpfile(fd
, filename
, target
);
266 return log_error_errno(r
, "Failed to move coredump %s into place: %m", target
);
271 static int maybe_remove_external_coredump(const char *filename
, uint64_t size
) {
273 /* Returns 1 if might remove, 0 if will not remove, < 0 on error. */
275 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
&&
276 size
<= arg_external_size_max
)
282 if (unlink(filename
) < 0 && errno
!= ENOENT
)
283 return log_error_errno(errno
, "Failed to unlink %s: %m", filename
);
288 static int make_filename(const Context
*context
, char **ret
) {
289 _cleanup_free_
char *c
= NULL
, *u
= NULL
, *p
= NULL
, *t
= NULL
;
290 sd_id128_t boot
= {};
295 c
= filename_escape(context
->meta
[META_COMM
]);
299 u
= filename_escape(context
->meta
[META_ARGV_UID
]);
303 r
= sd_id128_get_boot(&boot
);
307 p
= filename_escape(context
->meta
[META_ARGV_PID
]);
311 t
= filename_escape(context
->meta
[META_ARGV_TIMESTAMP
]);
316 "/var/lib/systemd/coredump/core.%s.%s." SD_ID128_FORMAT_STR
".%s.%s",
319 SD_ID128_FORMAT_VAL(boot
),
327 static int save_external_coredump(
328 const Context
*context
,
334 uint64_t *ret_compressed_size
,
335 bool *ret_truncated
) {
337 _cleanup_(unlink_and_freep
) char *tmp
= NULL
;
338 _cleanup_free_
char *fn
= NULL
;
339 _cleanup_close_
int fd
= -1;
340 uint64_t rlimit
, process_limit
, max_size
;
341 bool truncated
, storage_on_tmpfs
;
347 assert(ret_filename
);
351 assert(ret_compressed_size
);
352 assert(ret_truncated
);
354 r
= parse_uid(context
->meta
[META_ARGV_UID
], &uid
);
356 return log_error_errno(r
, "Failed to parse UID: %m");
358 r
= safe_atou64(context
->meta
[META_ARGV_RLIMIT
], &rlimit
);
360 return log_error_errno(r
, "Failed to parse resource limit '%s': %m",
361 context
->meta
[META_ARGV_RLIMIT
]);
362 if (rlimit
< page_size())
363 /* Is coredumping disabled? Then don't bother saving/processing the
364 * coredump. Anything below PAGE_SIZE cannot give a readable coredump
365 * (the kernel uses ELF_EXEC_PAGESIZE which is not easily accessible, but
366 * is usually the same as PAGE_SIZE. */
367 return log_info_errno(SYNTHETIC_ERRNO(EBADSLT
),
368 "Resource limits disable core dumping for process %s (%s).",
369 context
->meta
[META_ARGV_PID
], context
->meta
[META_COMM
]);
371 process_limit
= MAX(arg_process_size_max
, storage_size_max());
372 if (process_limit
== 0)
373 return log_debug_errno(SYNTHETIC_ERRNO(EBADSLT
),
374 "Limits for coredump processing and storage are both 0, not dumping core.");
376 /* Never store more than the process configured, or than we actually shall keep or process */
377 max_size
= MIN(rlimit
, process_limit
);
379 r
= make_filename(context
, &fn
);
381 return log_error_errno(r
, "Failed to determine coredump file name: %m");
383 (void) mkdir_parents_label(fn
, 0755);
385 fd
= open_tmpfile_linkable(fn
, O_RDWR
|O_CLOEXEC
, &tmp
);
387 return log_error_errno(fd
, "Failed to create temporary file for coredump %s: %m", fn
);
389 /* If storage is on tmpfs, the kernel oomd might kill us if there's MemoryMax set on
390 * the service or the slice it belongs to. This is common on low-resources systems,
391 * to avoid crashing processes to take away too many system resources.
392 * Check the cgroup settings, and set max_size to a bit less than half of the
393 * available memory left to the process.
394 * Then, attempt to write the core file uncompressed first - if the write gets
395 * interrupted, we know we won't be able to write it all, so instead compress what
396 * was written so far, delete the uncompressed truncated core, and then continue
397 * compressing from STDIN. Given the compressed core cannot be larger than the
398 * uncompressed one, and 1KB for metadata is accounted for in the calculation, we
399 * should be able to at least store the full compressed core file. */
401 storage_on_tmpfs
= fd_is_temporary_fs(fd
) > 0;
402 if (storage_on_tmpfs
&& arg_compress
) {
403 _cleanup_(sd_bus_flush_close_unrefp
) sd_bus
*bus
= NULL
;
404 uint64_t cgroup_limit
= UINT64_MAX
;
407 /* If we can't get the cgroup limit, just ignore it, but don't fail,
408 * try anyway with the config settings. */
409 r
= sd_bus_default_system(&bus
);
411 log_info_errno(r
, "Failed to connect to system bus, skipping MemoryAvailable check: %m");
413 _cleanup_(sd_bus_error_free
) sd_bus_error error
= SD_BUS_ERROR_NULL
;
415 r
= sd_bus_get_property_trivial(
417 "org.freedesktop.systemd1",
418 "/org/freedesktop/systemd1/unit/self",
419 "org.freedesktop.systemd1.Service",
425 "Failed to query MemoryAvailable for current unit, "
426 "falling back to static config settings: %s",
427 bus_error_message(&error
, r
));
430 max_size
= MIN(cgroup_limit
, max_size
);
431 max_size
= LESS_BY(max_size
, 1024U) / 2; /* Account for 1KB metadata overhead for compressing */
432 max_size
= MAX(PROCESS_SIZE_MIN
, max_size
); /* Impose a lower minimum */
434 /* tmpfs might get full quickly, so check the available space too.
435 * But don't worry about errors here, failing to access the storage
436 * location will be better logged when writing to it. */
437 if (statvfs("/var/lib/systemd/coredump/", &sv
) >= 0)
438 max_size
= MIN((uint64_t)sv
.f_frsize
* (uint64_t)sv
.f_bfree
, max_size
);
440 log_debug("Limiting core file size to %" PRIu64
" bytes due to cgroup memory limits.", max_size
);
443 r
= copy_bytes(input_fd
, fd
, max_size
, 0);
445 return log_error_errno(r
, "Cannot store coredump of %s (%s): %m",
446 context
->meta
[META_ARGV_PID
], context
->meta
[META_COMM
]);
451 _cleanup_(unlink_and_freep
) char *tmp_compressed
= NULL
;
452 _cleanup_free_
char *fn_compressed
= NULL
;
453 _cleanup_close_
int fd_compressed
= -1;
454 uint64_t uncompressed_size
= 0;
456 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
457 return log_error_errno(errno
, "Failed to seek on coredump %s: %m", fn
);
459 fn_compressed
= strjoin(fn
, COMPRESSED_EXT
);
463 fd_compressed
= open_tmpfile_linkable(fn_compressed
, O_RDWR
|O_CLOEXEC
, &tmp_compressed
);
464 if (fd_compressed
< 0)
465 return log_error_errno(fd_compressed
, "Failed to create temporary file for coredump %s: %m", fn_compressed
);
467 r
= compress_stream(fd
, fd_compressed
, max_size
, &uncompressed_size
);
469 return log_error_errno(r
, "Failed to compress %s: %m", coredump_tmpfile_name(tmp_compressed
));
471 if (truncated
&& storage_on_tmpfs
) {
472 uint64_t partial_uncompressed_size
= 0;
474 /* Uncompressed write was truncated and we are writing to tmpfs: delete
475 * the uncompressed core, and compress the remaining part from STDIN. */
477 tmp
= unlink_and_free(tmp
);
480 r
= compress_stream(input_fd
, fd_compressed
, max_size
, &partial_uncompressed_size
);
482 return log_error_errno(r
, "Failed to compress %s: %m", coredump_tmpfile_name(tmp_compressed
));
483 uncompressed_size
+= partial_uncompressed_size
;
486 r
= fix_permissions(fd_compressed
, tmp_compressed
, fn_compressed
, context
, uid
);
490 if (fstat(fd_compressed
, &st
) < 0)
491 return log_error_errno(errno
,
492 "Failed to fstat core file %s: %m",
493 coredump_tmpfile_name(tmp_compressed
));
495 *ret_filename
= TAKE_PTR(fn_compressed
); /* compressed */
496 *ret_node_fd
= TAKE_FD(fd_compressed
); /* compressed */
497 *ret_compressed_size
= (uint64_t) st
.st_size
; /* compressed */
498 *ret_data_fd
= TAKE_FD(fd
);
499 *ret_size
= uncompressed_size
;
500 *ret_truncated
= truncated
;
501 tmp_compressed
= mfree(tmp_compressed
);
509 LOG_MESSAGE("Core file was truncated to %zu bytes.", max_size
),
510 "SIZE_LIMIT=%zu", max_size
,
511 "MESSAGE_ID=" SD_MESSAGE_TRUNCATED_CORE_STR
);
513 r
= fix_permissions(fd
, tmp
, fn
, context
, uid
);
515 return log_error_errno(r
, "Failed to fix permissions and finalize coredump %s into %s: %m", coredump_tmpfile_name(tmp
), fn
);
517 if (fstat(fd
, &st
) < 0)
518 return log_error_errno(errno
, "Failed to fstat core file %s: %m", coredump_tmpfile_name(tmp
));
520 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
521 return log_error_errno(errno
, "Failed to seek on coredump %s: %m", fn
);
523 *ret_filename
= TAKE_PTR(fn
);
524 *ret_data_fd
= TAKE_FD(fd
);
525 *ret_size
= (uint64_t) st
.st_size
;
526 *ret_truncated
= truncated
;
531 static int allocate_journal_field(int fd
, size_t size
, char **ret
, size_t *ret_size
) {
532 _cleanup_free_
char *field
= NULL
;
539 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
540 return log_warning_errno(errno
, "Failed to seek: %m");
542 field
= malloc(9 + size
);
544 log_warning("Failed to allocate memory for coredump, coredump will not be stored.");
548 memcpy(field
, "COREDUMP=", 9);
550 n
= read(fd
, field
+ 9, size
);
552 return log_error_errno((int) n
, "Failed to read core data: %m");
553 if ((size_t) n
< size
)
554 return log_error_errno(SYNTHETIC_ERRNO(EIO
),
555 "Core data too short.");
557 *ret
= TAKE_PTR(field
);
558 *ret_size
= size
+ 9;
563 /* Joins /proc/[pid]/fd/ and /proc/[pid]/fdinfo/ into the following lines:
577 static int compose_open_fds(pid_t pid
, char **open_fds
) {
578 _cleanup_closedir_
DIR *proc_fd_dir
= NULL
;
579 _cleanup_close_
int proc_fdinfo_fd
= -1;
580 _cleanup_free_
char *buffer
= NULL
;
581 _cleanup_fclose_
FILE *stream
= NULL
;
582 const char *fddelim
= "", *path
;
587 assert(open_fds
!= NULL
);
589 path
= procfs_file_alloca(pid
, "fd");
590 proc_fd_dir
= opendir(path
);
594 proc_fdinfo_fd
= openat(dirfd(proc_fd_dir
), "../fdinfo", O_DIRECTORY
|O_NOFOLLOW
|O_CLOEXEC
|O_PATH
);
595 if (proc_fdinfo_fd
< 0)
598 stream
= open_memstream_unlocked(&buffer
, &size
);
602 FOREACH_DIRENT(de
, proc_fd_dir
, return -errno
) {
603 _cleanup_fclose_
FILE *fdinfo
= NULL
;
604 _cleanup_free_
char *fdname
= NULL
;
605 _cleanup_close_
int fd
= -1;
607 r
= readlinkat_malloc(dirfd(proc_fd_dir
), de
->d_name
, &fdname
);
611 fprintf(stream
, "%s%s:%s\n", fddelim
, de
->d_name
, fdname
);
614 /* Use the directory entry from /proc/[pid]/fd with /proc/[pid]/fdinfo */
615 fd
= openat(proc_fdinfo_fd
, de
->d_name
, O_NOFOLLOW
|O_CLOEXEC
|O_RDONLY
);
619 fdinfo
= take_fdopen(&fd
, "r");
624 _cleanup_free_
char *line
= NULL
;
626 r
= read_line(fdinfo
, LONG_LINE_MAX
, &line
);
638 stream
= safe_fclose(stream
);
643 *open_fds
= TAKE_PTR(buffer
);
648 static int get_process_ns(pid_t pid
, const char *namespace, ino_t
*ns
) {
651 _cleanup_close_
int proc_ns_dir_fd
= -1;
653 p
= procfs_file_alloca(pid
, "ns");
655 proc_ns_dir_fd
= open(p
, O_DIRECTORY
| O_CLOEXEC
| O_RDONLY
);
656 if (proc_ns_dir_fd
< 0)
659 if (fstatat(proc_ns_dir_fd
, namespace, &stbuf
, /* flags */0) < 0)
666 static int get_mount_namespace_leader(pid_t pid
, pid_t
*ret
) {
670 r
= get_process_ns(pid
, "mnt", &proc_mntns
);
678 r
= get_process_ppid(pid
, &ppid
);
679 if (r
== -EADDRNOTAVAIL
) /* Reached the top (i.e. typically PID 1, but could also be a process
680 * whose parent is not in our pidns) */
685 r
= get_process_ns(ppid
, "mnt", &parent_mntns
);
689 if (proc_mntns
!= parent_mntns
) {
698 /* Returns 1 if the parent was found.
699 * Returns 0 if there is not a process we can call the pid's
700 * container parent (the pid's process isn't 'containerized').
701 * Returns a negative number on errors.
703 static int get_process_container_parent_cmdline(pid_t pid
, char** cmdline
) {
705 const char *proc_root_path
;
706 struct stat root_stat
, proc_root_stat
;
709 /* To compare inodes of / and /proc/[pid]/root */
710 if (stat("/", &root_stat
) < 0)
713 proc_root_path
= procfs_file_alloca(pid
, "root");
714 if (stat(proc_root_path
, &proc_root_stat
) < 0)
717 /* The process uses system root. */
718 if (proc_root_stat
.st_ino
== root_stat
.st_ino
) {
723 r
= get_mount_namespace_leader(pid
, &container_pid
);
727 r
= get_process_cmdline(container_pid
, SIZE_MAX
, PROCESS_CMDLINE_QUOTE_POSIX
, cmdline
);
734 static int change_uid_gid(const Context
*context
) {
739 r
= parse_uid(context
->meta
[META_ARGV_UID
], &uid
);
743 if (uid_is_system(uid
)) {
744 const char *user
= "systemd-coredump";
746 r
= get_user_creds(&user
, &uid
, &gid
, NULL
, NULL
, 0);
748 log_warning_errno(r
, "Cannot resolve %s user. Proceeding to dump core as root: %m", user
);
752 r
= parse_gid(context
->meta
[META_ARGV_GID
], &gid
);
757 return drop_privileges(uid
, gid
, 0);
760 static int submit_coredump(
762 struct iovec_wrapper
*iovw
,
765 _cleanup_(json_variant_unrefp
) JsonVariant
*json_metadata
= NULL
;
766 _cleanup_close_
int coredump_fd
= -1, coredump_node_fd
= -1;
767 _cleanup_free_
char *filename
= NULL
, *coredump_data
= NULL
;
768 _cleanup_free_
char *stacktrace
= NULL
;
770 const char *module_name
;
771 uint64_t coredump_size
= UINT64_MAX
, coredump_compressed_size
= UINT64_MAX
;
772 bool truncated
= false;
773 JsonVariant
*module_json
;
778 assert(input_fd
>= 0);
780 /* Vacuum before we write anything again */
781 (void) coredump_vacuum(-1, arg_keep_free
, arg_max_use
);
783 /* Always stream the coredump to disk, if that's possible */
784 r
= save_external_coredump(context
, input_fd
,
785 &filename
, &coredump_node_fd
, &coredump_fd
,
786 &coredump_size
, &coredump_compressed_size
, &truncated
);
788 /* Skip whole core dumping part */
791 /* If we don't want to keep the coredump on disk, remove it now, as later on we
792 * will lack the privileges for it. However, we keep the fd to it, so that we can
793 * still process it and log it. */
794 r
= maybe_remove_external_coredump(filename
, coredump_node_fd
>= 0 ? coredump_compressed_size
: coredump_size
);
798 (void) iovw_put_string_field(iovw
, "COREDUMP_FILENAME=", filename
);
799 else if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
800 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
801 coredump_node_fd
>= 0 ? coredump_compressed_size
: coredump_size
, arg_external_size_max
);
803 /* Vacuum again, but exclude the coredump we just created */
804 (void) coredump_vacuum(coredump_node_fd
>= 0 ? coredump_node_fd
: coredump_fd
, arg_keep_free
, arg_max_use
);
806 /* Now, let's drop privileges to become the user who owns the segfaulted process
807 * and allocate the coredump memory under the user's uid. This also ensures that
808 * the credentials journald will see are the ones of the coredumping user, thus
809 * making sure the user gets access to the core dump. Let's also get rid of all
810 * capabilities, if we run as root, we won't need them anymore. */
811 r
= change_uid_gid(context
);
813 return log_error_errno(r
, "Failed to drop privileges: %m");
815 /* Try to get a stack trace if we can */
816 if (coredump_size
> arg_process_size_max
)
817 log_debug("Not generating stack trace: core size %"PRIu64
" is greater "
818 "than %"PRIu64
" (the configured maximum)",
819 coredump_size
, arg_process_size_max
);
820 else if (coredump_fd
>= 0) {
821 bool skip
= startswith(context
->meta
[META_COMM
], "systemd-coredum"); /* COMM is 16 bytes usually */
823 (void) parse_elf_object(coredump_fd
,
824 context
->meta
[META_EXE
],
825 /* fork_disable_dump= */ skip
, /* avoid loops */
831 core_message
= strjoina("Process ", context
->meta
[META_ARGV_PID
],
832 " (", context
->meta
[META_COMM
], ") of user ",
833 context
->meta
[META_ARGV_UID
], " dumped core.",
834 context
->is_journald
&& filename
? "\nCoredump diverted to " : NULL
,
835 context
->is_journald
&& filename
? filename
: NULL
);
837 core_message
= strjoina(core_message
, stacktrace
? "\n\n" : NULL
, stacktrace
);
839 if (context
->is_journald
) {
840 /* We cannot log to the journal, so just print the message.
841 * The target was set previously to something safe. */
842 log_dispatch(LOG_ERR
, 0, core_message
);
846 (void) iovw_put_string_field(iovw
, "MESSAGE=", core_message
);
849 (void) iovw_put_string_field(iovw
, "COREDUMP_TRUNCATED=", "1");
851 /* If we managed to parse any ELF metadata (build-id, ELF package meta),
852 * attach it as journal metadata. */
854 _cleanup_free_
char *formatted_json
= NULL
;
856 r
= json_variant_format(json_metadata
, 0, &formatted_json
);
858 return log_error_errno(r
, "Failed to format JSON package metadata: %m");
860 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_JSON=", formatted_json
);
863 /* In the unlikely scenario that context->meta[META_EXE] is not available,
864 * let's avoid guessing the module name and skip the loop. */
865 if (context
->meta
[META_EXE
])
866 JSON_VARIANT_OBJECT_FOREACH(module_name
, module_json
, json_metadata
) {
869 /* We only add structured fields for the 'main' ELF module, and only if we can identify it. */
870 if (!path_equal_filename(module_name
, context
->meta
[META_EXE
]))
873 t
= json_variant_by_key(module_json
, "name");
875 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_NAME=", json_variant_string(t
));
877 t
= json_variant_by_key(module_json
, "version");
879 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_VERSION=", json_variant_string(t
));
882 /* Optionally store the entire coredump in the journal */
883 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
&& coredump_fd
>= 0) {
884 if (coredump_size
<= arg_journal_size_max
) {
887 /* Store the coredump itself in the journal */
889 r
= allocate_journal_field(coredump_fd
, (size_t) coredump_size
, &coredump_data
, &sz
);
891 if (iovw_put(iovw
, coredump_data
, sz
) >= 0)
892 TAKE_PTR(coredump_data
);
894 log_warning_errno(r
, "Failed to attach the core to the journal entry: %m");
896 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
897 coredump_size
, arg_journal_size_max
);
900 r
= sd_journal_sendv(iovw
->iovec
, iovw
->count
);
902 return log_error_errno(r
, "Failed to log coredump: %m");
907 static int save_context(Context
*context
, const struct iovec_wrapper
*iovw
) {
914 assert(iovw
->count
>= _META_ARGV_MAX
);
916 /* The context does not allocate any memory on its own */
918 for (size_t n
= 0; n
< iovw
->count
; n
++) {
919 struct iovec
*iovec
= iovw
->iovec
+ n
;
921 for (size_t i
= 0; i
< ELEMENTSOF(meta_field_names
); i
++) {
924 /* Note that these strings are NUL terminated, because we made sure that a
925 * trailing NUL byte is in the buffer, though not included in the iov_len
926 * count (see process_socket() and gather_pid_metadata_*()) */
927 assert(((char*) iovec
->iov_base
)[iovec
->iov_len
] == 0);
929 p
= startswith(iovec
->iov_base
, meta_field_names
[i
]);
931 context
->meta
[i
] = p
;
938 if (!context
->meta
[META_ARGV_PID
])
939 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
940 "Failed to find the PID of crashing process");
942 r
= parse_pid(context
->meta
[META_ARGV_PID
], &context
->pid
);
944 return log_error_errno(r
, "Failed to parse PID \"%s\": %m", context
->meta
[META_ARGV_PID
]);
946 unit
= context
->meta
[META_UNIT
];
947 context
->is_pid1
= streq(context
->meta
[META_ARGV_PID
], "1") || streq_ptr(unit
, SPECIAL_INIT_SCOPE
);
948 context
->is_journald
= streq_ptr(unit
, SPECIAL_JOURNALD_SERVICE
);
953 static int process_socket(int fd
) {
954 _cleanup_close_
int input_fd
= -1;
955 Context context
= {};
956 struct iovec_wrapper iovw
= {};
964 log_debug("Processing coredump received on stdin...");
967 CMSG_BUFFER_TYPE(CMSG_SPACE(sizeof(int))) control
;
969 .msg_control
= &control
,
970 .msg_controllen
= sizeof(control
),
976 l
= next_datagram_size_fd(fd
);
978 r
= log_error_errno(l
, "Failed to determine datagram size to read: %m");
983 iovec
.iov_base
= malloc(l
+ 1);
984 if (!iovec
.iov_base
) {
991 n
= recvmsg_safe(fd
, &mh
, MSG_CMSG_CLOEXEC
);
993 free(iovec
.iov_base
);
994 r
= log_error_errno(n
, "Failed to receive datagram: %m");
998 /* The final zero-length datagram carries the file descriptor and tells us
999 * that we're done. */
1001 struct cmsghdr
*found
;
1003 free(iovec
.iov_base
);
1005 found
= cmsg_find(&mh
, SOL_SOCKET
, SCM_RIGHTS
, CMSG_LEN(sizeof(int)));
1007 cmsg_close_all(&mh
);
1008 r
= log_error_errno(SYNTHETIC_ERRNO(EBADMSG
),
1009 "Coredump file descriptor missing.");
1013 assert(input_fd
< 0);
1014 input_fd
= *(int*) CMSG_DATA(found
);
1017 cmsg_close_all(&mh
);
1019 /* Add trailing NUL byte, in case these are strings */
1020 ((char*) iovec
.iov_base
)[n
] = 0;
1021 iovec
.iov_len
= (size_t) n
;
1023 r
= iovw_put(&iovw
, iovec
.iov_base
, iovec
.iov_len
);
1028 /* Make sure we got all data we really need */
1029 assert(input_fd
>= 0);
1031 r
= save_context(&context
, &iovw
);
1035 /* Make sure we received at least all fields we need. */
1036 for (int i
= 0; i
< _META_MANDATORY_MAX
; i
++)
1037 if (!context
.meta
[i
]) {
1038 r
= log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1039 "A mandatory argument (%i) has not been sent, aborting.",
1044 r
= submit_coredump(&context
, &iovw
, input_fd
);
1047 iovw_free_contents(&iovw
, true);
1051 static int send_iovec(const struct iovec_wrapper
*iovw
, int input_fd
) {
1053 static const union sockaddr_union sa
= {
1054 .un
.sun_family
= AF_UNIX
,
1055 .un
.sun_path
= "/run/systemd/coredump",
1057 _cleanup_close_
int fd
= -1;
1061 assert(input_fd
>= 0);
1063 fd
= socket(AF_UNIX
, SOCK_SEQPACKET
|SOCK_CLOEXEC
, 0);
1065 return log_error_errno(errno
, "Failed to create coredump socket: %m");
1067 if (connect(fd
, &sa
.sa
, SOCKADDR_UN_LEN(sa
.un
)) < 0)
1068 return log_error_errno(errno
, "Failed to connect to coredump service: %m");
1070 for (size_t i
= 0; i
< iovw
->count
; i
++) {
1071 struct msghdr mh
= {
1072 .msg_iov
= iovw
->iovec
+ i
,
1075 struct iovec copy
[2];
1078 if (sendmsg(fd
, &mh
, MSG_NOSIGNAL
) >= 0)
1081 if (errno
== EMSGSIZE
&& mh
.msg_iov
[0].iov_len
> 0) {
1082 /* This field didn't fit? That's a pity. Given that this is
1083 * just metadata, let's truncate the field at half, and try
1084 * again. We append three dots, in order to show that this is
1087 if (mh
.msg_iov
!= copy
) {
1088 /* We don't want to modify the caller's iovec, hence
1089 * let's create our own array, consisting of two new
1090 * iovecs, where the first is a (truncated) copy of
1091 * what we want to send, and the second one contains
1092 * the trailing dots. */
1093 copy
[0] = iovw
->iovec
[i
];
1094 copy
[1] = IOVEC_MAKE(((char[]){'.', '.', '.'}), 3);
1100 copy
[0].iov_len
/= 2; /* halve it, and try again */
1104 return log_error_errno(errno
, "Failed to send coredump datagram: %m");
1108 r
= send_one_fd(fd
, input_fd
, 0);
1110 return log_error_errno(r
, "Failed to send coredump fd: %m");
1115 static int gather_pid_metadata_from_argv(
1116 struct iovec_wrapper
*iovw
,
1118 int argc
, char **argv
) {
1120 _cleanup_free_
char *free_timestamp
= NULL
;
1124 /* We gather all metadata that were passed via argv[] into an array of iovecs that
1125 * we'll forward to the socket unit */
1127 if (argc
< _META_ARGV_MAX
)
1128 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1129 "Not enough arguments passed by the kernel (%i, expected %i).",
1130 argc
, _META_ARGV_MAX
);
1132 for (int i
= 0; i
< _META_ARGV_MAX
; i
++) {
1138 case META_ARGV_TIMESTAMP
:
1139 /* The journal fields contain the timestamp padded with six
1140 * zeroes, so that the kernel-supplied 1s granularity timestamps
1141 * becomes 1µs granularity, i.e. the granularity systemd usually
1143 t
= free_timestamp
= strjoin(argv
[i
], "000000");
1148 case META_ARGV_SIGNAL
:
1149 /* For signal, record its pretty name too */
1150 if (safe_atoi(argv
[i
], &signo
) >= 0 && SIGNAL_VALID(signo
))
1151 (void) iovw_put_string_field(iovw
, "COREDUMP_SIGNAL_NAME=SIG",
1152 signal_to_string(signo
));
1159 r
= iovw_put_string_field(iovw
, meta_field_names
[i
], t
);
1164 /* Cache some of the process metadata we collected so far and that we'll need to
1166 return save_context(context
, iovw
);
1169 static int gather_pid_metadata(struct iovec_wrapper
*iovw
, Context
*context
) {
1176 /* Note that if we fail on oom later on, we do not roll-back changes to the iovec
1177 * structure. (It remains valid, with the first iovec fields initialized.) */
1181 /* The following is mandatory */
1182 r
= get_process_comm(pid
, &t
);
1184 return log_error_errno(r
, "Failed to get COMM: %m");
1186 r
= iovw_put_string_field_free(iovw
, "COREDUMP_COMM=", t
);
1190 /* The following are optional, but we use them if present. */
1191 r
= get_process_exe(pid
, &t
);
1193 r
= iovw_put_string_field_free(iovw
, "COREDUMP_EXE=", t
);
1195 log_warning_errno(r
, "Failed to get EXE, ignoring: %m");
1197 if (cg_pid_get_unit(pid
, &t
) >= 0)
1198 (void) iovw_put_string_field_free(iovw
, "COREDUMP_UNIT=", t
);
1200 if (cg_pid_get_user_unit(pid
, &t
) >= 0)
1201 (void) iovw_put_string_field_free(iovw
, "COREDUMP_USER_UNIT=", t
);
1203 if (sd_pid_get_session(pid
, &t
) >= 0)
1204 (void) iovw_put_string_field_free(iovw
, "COREDUMP_SESSION=", t
);
1206 if (sd_pid_get_owner_uid(pid
, &owner_uid
) >= 0) {
1207 r
= asprintf(&t
, UID_FMT
, owner_uid
);
1209 (void) iovw_put_string_field_free(iovw
, "COREDUMP_OWNER_UID=", t
);
1212 if (sd_pid_get_slice(pid
, &t
) >= 0)
1213 (void) iovw_put_string_field_free(iovw
, "COREDUMP_SLICE=", t
);
1215 if (get_process_cmdline(pid
, SIZE_MAX
, PROCESS_CMDLINE_QUOTE_POSIX
, &t
) >= 0)
1216 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CMDLINE=", t
);
1218 if (cg_pid_get_path_shifted(pid
, NULL
, &t
) >= 0)
1219 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CGROUP=", t
);
1221 if (compose_open_fds(pid
, &t
) >= 0)
1222 (void) iovw_put_string_field_free(iovw
, "COREDUMP_OPEN_FDS=", t
);
1224 p
= procfs_file_alloca(pid
, "status");
1225 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1226 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_STATUS=", t
);
1228 p
= procfs_file_alloca(pid
, "maps");
1229 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1230 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_MAPS=", t
);
1232 p
= procfs_file_alloca(pid
, "limits");
1233 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1234 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_LIMITS=", t
);
1236 p
= procfs_file_alloca(pid
, "cgroup");
1237 if (read_full_virtual_file(p
, &t
, NULL
) >=0)
1238 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_CGROUP=", t
);
1240 p
= procfs_file_alloca(pid
, "mountinfo");
1241 if (read_full_virtual_file(p
, &t
, NULL
) >=0)
1242 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_MOUNTINFO=", t
);
1244 if (get_process_cwd(pid
, &t
) >= 0)
1245 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CWD=", t
);
1247 if (get_process_root(pid
, &t
) >= 0) {
1248 bool proc_self_root_is_slash
;
1250 proc_self_root_is_slash
= strcmp(t
, "/") == 0;
1252 (void) iovw_put_string_field_free(iovw
, "COREDUMP_ROOT=", t
);
1254 /* If the process' root is "/", then there is a chance it has
1255 * mounted own root and hence being containerized. */
1256 if (proc_self_root_is_slash
&& get_process_container_parent_cmdline(pid
, &t
) > 0)
1257 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CONTAINER_CMDLINE=", t
);
1260 if (get_process_environ(pid
, &t
) >= 0)
1261 (void) iovw_put_string_field_free(iovw
, "COREDUMP_ENVIRON=", t
);
1263 /* we successfully acquired all metadata */
1264 return save_context(context
, iovw
);
1267 static int process_kernel(int argc
, char* argv
[]) {
1268 Context context
= {};
1269 struct iovec_wrapper
*iovw
;
1272 log_debug("Processing coredump received from the kernel...");
1278 (void) iovw_put_string_field(iovw
, "MESSAGE_ID=", SD_MESSAGE_COREDUMP_STR
);
1279 (void) iovw_put_string_field(iovw
, "PRIORITY=", STRINGIFY(LOG_CRIT
));
1281 /* Collect all process metadata passed by the kernel through argv[] */
1282 r
= gather_pid_metadata_from_argv(iovw
, &context
, argc
- 1, argv
+ 1);
1286 /* Collect the rest of the process metadata retrieved from the runtime */
1287 r
= gather_pid_metadata(iovw
, &context
);
1291 if (!context
.is_journald
) {
1292 /* OK, now we know it's not the journal, hence we can make use of it now. */
1293 log_set_target(LOG_TARGET_JOURNAL_OR_KMSG
);
1297 /* If this is PID 1 disable coredump collection, we'll unlikely be able to process
1300 * FIXME: maybe we should disable coredumps generation from the beginning and
1301 * re-enable it only when we know it's either safe (ie we're not running OOM) or
1302 * it's not pid1 ? */
1303 if (context
.is_pid1
) {
1304 log_notice("Due to PID 1 having crashed coredump collection will now be turned off.");
1305 disable_coredumps();
1308 if (context
.is_journald
|| context
.is_pid1
)
1309 r
= submit_coredump(&context
, iovw
, STDIN_FILENO
);
1311 r
= send_iovec(iovw
, STDIN_FILENO
);
1314 iovw
= iovw_free_free(iovw
);
1318 static int process_backtrace(int argc
, char *argv
[]) {
1319 Context context
= {};
1320 struct iovec_wrapper
*iovw
;
1323 _cleanup_(journal_importer_cleanup
) JournalImporter importer
= JOURNAL_IMPORTER_INIT(STDIN_FILENO
);
1325 log_debug("Processing backtrace on stdin...");
1331 (void) iovw_put_string_field(iovw
, "MESSAGE_ID=", SD_MESSAGE_BACKTRACE_STR
);
1332 (void) iovw_put_string_field(iovw
, "PRIORITY=", STRINGIFY(LOG_CRIT
));
1334 /* Collect all process metadata from argv[] by making sure to skip the
1335 * '--backtrace' option */
1336 r
= gather_pid_metadata_from_argv(iovw
, &context
, argc
- 2, argv
+ 2);
1340 /* Collect the rest of the process metadata retrieved from the runtime */
1341 r
= gather_pid_metadata(iovw
, &context
);
1346 r
= journal_importer_process_data(&importer
);
1348 log_error_errno(r
, "Failed to parse journal entry on stdin: %m");
1351 if (r
== 1 || /* complete entry */
1352 journal_importer_eof(&importer
)) /* end of data */
1356 if (journal_importer_eof(&importer
)) {
1357 log_warning("Did not receive a full journal entry on stdin, ignoring message sent by reporter");
1359 message
= strjoina("Process ", context
.meta
[META_ARGV_PID
],
1360 " (", context
.meta
[META_COMM
], ")"
1361 " of user ", context
.meta
[META_ARGV_UID
],
1362 " failed with ", context
.meta
[META_ARGV_SIGNAL
]);
1364 r
= iovw_put_string_field(iovw
, "MESSAGE=", message
);
1368 /* The imported iovecs are not supposed to be freed by us so let's store
1369 * them at the end of the array so we can skip them while freeing the
1371 for (size_t i
= 0; i
< importer
.iovw
.count
; i
++) {
1372 struct iovec
*iovec
= importer
.iovw
.iovec
+ i
;
1374 iovw_put(iovw
, iovec
->iov_base
, iovec
->iov_len
);
1378 r
= sd_journal_sendv(iovw
->iovec
, iovw
->count
);
1380 log_error_errno(r
, "Failed to log backtrace: %m");
1383 iovw
->count
-= importer
.iovw
.count
;
1384 iovw
= iovw_free_free(iovw
);
1388 static int run(int argc
, char *argv
[]) {
1391 /* First, log to a safe place, since we don't know what crashed and it might
1392 * be journald which we'd rather not log to then. */
1394 log_set_target(LOG_TARGET_KMSG
);
1397 /* Make sure we never enter a loop */
1398 (void) prctl(PR_SET_DUMPABLE
, 0);
1400 /* Ignore all parse errors */
1401 (void) parse_config();
1403 log_debug("Selected storage '%s'.", coredump_storage_to_string(arg_storage
));
1404 log_debug("Selected compression %s.", yes_no(arg_compress
));
1406 r
= sd_listen_fds(false);
1408 return log_error_errno(r
, "Failed to determine the number of file descriptors: %m");
1410 /* If we got an fd passed, we are running in coredumpd mode. Otherwise we
1411 * are invoked from the kernel as coredump handler. */
1413 if (streq_ptr(argv
[1], "--backtrace"))
1414 return process_backtrace(argc
, argv
);
1416 return process_kernel(argc
, argv
);
1418 return process_socket(SD_LISTEN_FDS_START
);
1420 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1421 "Received unexpected number of file descriptors.");
1424 DEFINE_MAIN_FUNCTION(run
);