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-util.h"
24 #include "coredump-vacuum.h"
25 #include "dirent-util.h"
32 #include "journal-importer.h"
33 #include "journal-send.h"
36 #include "main-func.h"
37 #include "memory-util.h"
38 #include "mkdir-label.h"
39 #include "parse-util.h"
40 #include "process-util.h"
41 #include "signal-util.h"
42 #include "socket-util.h"
44 #include "stat-util.h"
45 #include "string-table.h"
46 #include "string-util.h"
48 #include "sync-util.h"
49 #include "tmpfile-util.h"
50 #include "uid-alloc-range.h"
51 #include "user-util.h"
53 /* The maximum size up to which we process coredumps. We use 1G on 32bit systems, and 32G on 64bit systems */
54 #if __SIZEOF_POINTER__ == 4
55 #define PROCESS_SIZE_MAX ((uint64_t) (1LLU*1024LLU*1024LLU*1024LLU))
56 #elif __SIZEOF_POINTER__ == 8
57 #define PROCESS_SIZE_MAX ((uint64_t) (32LLU*1024LLU*1024LLU*1024LLU))
59 #error "Unexpected pointer size"
62 /* The maximum size up to which we leave the coredump around on disk */
63 #define EXTERNAL_SIZE_MAX PROCESS_SIZE_MAX
65 /* The maximum size up to which we store the coredump in the journal */
66 #ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
67 #define JOURNAL_SIZE_MAX ((size_t) (767LU*1024LU*1024LU))
69 /* oss-fuzz limits memory usage. */
70 #define JOURNAL_SIZE_MAX ((size_t) (10LU*1024LU*1024LU))
73 /* When checking for available memory and setting lower limits, don't
74 * go below 4MB for writing core files to storage. */
75 #define PROCESS_SIZE_MIN (4U*1024U*1024U)
77 /* Make sure to not make this larger than the maximum journal entry
78 * size. See DATA_SIZE_MAX in journal-importer.h. */
79 assert_cc(JOURNAL_SIZE_MAX
<= DATA_SIZE_MAX
);
82 /* We use these as array indexes for our process metadata cache.
84 * The first indices of the cache stores the same metadata as the ones passed by
85 * the kernel via argv[], ie the strings array passed by the kernel according to
86 * our pattern defined in /proc/sys/kernel/core_pattern (see man:core(5)). */
88 META_ARGV_PID
, /* %P: as seen in the initial pid namespace */
89 META_ARGV_UID
, /* %u: as seen in the initial user namespace */
90 META_ARGV_GID
, /* %g: as seen in the initial user namespace */
91 META_ARGV_SIGNAL
, /* %s: number of signal causing dump */
92 META_ARGV_TIMESTAMP
, /* %t: time of dump, expressed as seconds since the Epoch (we expand this to µs granularity) */
93 META_ARGV_RLIMIT
, /* %c: core file size soft resource limit */
94 META_ARGV_HOSTNAME
, /* %h: hostname */
97 /* The following indexes are cached for a couple of special fields we use (and
98 * thereby need to be retrieved quickly) for naming coredump files, and attaching
99 * xattrs. Unlike the previous ones they are retrieved from the runtime
102 META_COMM
= _META_ARGV_MAX
,
105 /* The rest are similar to the previous ones except that we won't fail if one of
106 * them is missing. */
108 META_EXE
= _META_MANDATORY_MAX
,
113 static const char * const meta_field_names
[_META_MAX
] = {
114 [META_ARGV_PID
] = "COREDUMP_PID=",
115 [META_ARGV_UID
] = "COREDUMP_UID=",
116 [META_ARGV_GID
] = "COREDUMP_GID=",
117 [META_ARGV_SIGNAL
] = "COREDUMP_SIGNAL=",
118 [META_ARGV_TIMESTAMP
] = "COREDUMP_TIMESTAMP=",
119 [META_ARGV_RLIMIT
] = "COREDUMP_RLIMIT=",
120 [META_ARGV_HOSTNAME
] = "COREDUMP_HOSTNAME=",
121 [META_COMM
] = "COREDUMP_COMM=",
122 [META_EXE
] = "COREDUMP_EXE=",
123 [META_UNIT
] = "COREDUMP_UNIT=",
126 typedef struct Context
{
127 const char *meta
[_META_MAX
];
133 typedef enum CoredumpStorage
{
134 COREDUMP_STORAGE_NONE
,
135 COREDUMP_STORAGE_EXTERNAL
,
136 COREDUMP_STORAGE_JOURNAL
,
137 _COREDUMP_STORAGE_MAX
,
138 _COREDUMP_STORAGE_INVALID
= -EINVAL
,
141 static const char* const coredump_storage_table
[_COREDUMP_STORAGE_MAX
] = {
142 [COREDUMP_STORAGE_NONE
] = "none",
143 [COREDUMP_STORAGE_EXTERNAL
] = "external",
144 [COREDUMP_STORAGE_JOURNAL
] = "journal",
147 DEFINE_PRIVATE_STRING_TABLE_LOOKUP(coredump_storage
, CoredumpStorage
);
148 static DEFINE_CONFIG_PARSE_ENUM(config_parse_coredump_storage
, coredump_storage
, CoredumpStorage
, "Failed to parse storage setting");
150 static CoredumpStorage arg_storage
= COREDUMP_STORAGE_EXTERNAL
;
151 static bool arg_compress
= true;
152 static uint64_t arg_process_size_max
= PROCESS_SIZE_MAX
;
153 static uint64_t arg_external_size_max
= EXTERNAL_SIZE_MAX
;
154 static uint64_t arg_journal_size_max
= JOURNAL_SIZE_MAX
;
155 static uint64_t arg_keep_free
= UINT64_MAX
;
156 static uint64_t arg_max_use
= UINT64_MAX
;
158 static int parse_config(void) {
159 static const ConfigTableItem items
[] = {
160 { "Coredump", "Storage", config_parse_coredump_storage
, 0, &arg_storage
},
161 { "Coredump", "Compress", config_parse_bool
, 0, &arg_compress
},
162 { "Coredump", "ProcessSizeMax", config_parse_iec_uint64
, 0, &arg_process_size_max
},
163 { "Coredump", "ExternalSizeMax", config_parse_iec_uint64_infinity
, 0, &arg_external_size_max
},
164 { "Coredump", "JournalSizeMax", config_parse_iec_size
, 0, &arg_journal_size_max
},
165 { "Coredump", "KeepFree", config_parse_iec_uint64
, 0, &arg_keep_free
},
166 { "Coredump", "MaxUse", config_parse_iec_uint64
, 0, &arg_max_use
},
170 return config_parse_many_nulstr(
171 PKGSYSCONFDIR
"/coredump.conf",
172 CONF_PATHS_NULSTR("systemd/coredump.conf.d"),
174 config_item_table_lookup
, items
,
180 static uint64_t storage_size_max(void) {
181 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
182 return arg_external_size_max
;
183 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
)
184 return arg_journal_size_max
;
185 assert(arg_storage
== COREDUMP_STORAGE_NONE
);
189 static int fix_acl(int fd
, uid_t uid
) {
195 assert(uid_is_valid(uid
));
197 if (uid_is_system(uid
) || uid_is_dynamic(uid
) || uid
== UID_NOBODY
)
200 /* Make sure normal users can read (but not write or delete) their own coredumps */
201 r
= fd_add_uid_acl_permission(fd
, uid
, ACL_READ
);
203 return log_error_errno(r
, "Failed to adjust ACL of the coredump: %m");
209 static int fix_xattr(int fd
, const Context
*context
) {
211 static const char * const xattrs
[_META_MAX
] = {
212 [META_ARGV_PID
] = "user.coredump.pid",
213 [META_ARGV_UID
] = "user.coredump.uid",
214 [META_ARGV_GID
] = "user.coredump.gid",
215 [META_ARGV_SIGNAL
] = "user.coredump.signal",
216 [META_ARGV_TIMESTAMP
] = "user.coredump.timestamp",
217 [META_ARGV_RLIMIT
] = "user.coredump.rlimit",
218 [META_ARGV_HOSTNAME
] = "user.coredump.hostname",
219 [META_COMM
] = "user.coredump.comm",
220 [META_EXE
] = "user.coredump.exe",
227 /* Attach some metadata to coredumps via extended
228 * attributes. Just because we can. */
230 for (unsigned i
= 0; i
< _META_MAX
; i
++) {
233 if (isempty(context
->meta
[i
]) || !xattrs
[i
])
236 k
= fsetxattr(fd
, xattrs
[i
], context
->meta
[i
], strlen(context
->meta
[i
]), XATTR_CREATE
);
244 #define filename_escape(s) xescape((s), "./ ")
246 static const char *coredump_tmpfile_name(const char *s
) {
247 return s
? s
: "(unnamed temporary file)";
250 static int fix_permissions(
252 const char *filename
,
254 const Context
*context
,
263 /* Ignore errors on these */
264 (void) fchmod(fd
, 0640);
265 (void) fix_acl(fd
, uid
);
266 (void) fix_xattr(fd
, context
);
270 return log_error_errno(r
, "Failed to sync coredump %s: %m", coredump_tmpfile_name(filename
));
272 r
= link_tmpfile(fd
, filename
, target
);
274 return log_error_errno(r
, "Failed to move coredump %s into place: %m", target
);
279 static int maybe_remove_external_coredump(const char *filename
, uint64_t size
) {
281 /* Returns 1 if might remove, 0 if will not remove, < 0 on error. */
283 if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
&&
284 size
<= arg_external_size_max
)
290 if (unlink(filename
) < 0 && errno
!= ENOENT
)
291 return log_error_errno(errno
, "Failed to unlink %s: %m", filename
);
296 static int make_filename(const Context
*context
, char **ret
) {
297 _cleanup_free_
char *c
= NULL
, *u
= NULL
, *p
= NULL
, *t
= NULL
;
298 sd_id128_t boot
= {};
303 c
= filename_escape(context
->meta
[META_COMM
]);
307 u
= filename_escape(context
->meta
[META_ARGV_UID
]);
311 r
= sd_id128_get_boot(&boot
);
315 p
= filename_escape(context
->meta
[META_ARGV_PID
]);
319 t
= filename_escape(context
->meta
[META_ARGV_TIMESTAMP
]);
324 "/var/lib/systemd/coredump/core.%s.%s." SD_ID128_FORMAT_STR
".%s.%s",
327 SD_ID128_FORMAT_VAL(boot
),
335 static int save_external_coredump(
336 const Context
*context
,
342 uint64_t *ret_compressed_size
,
343 bool *ret_truncated
) {
345 _cleanup_(unlink_and_freep
) char *tmp
= NULL
;
346 _cleanup_free_
char *fn
= NULL
;
347 _cleanup_close_
int fd
= -EBADF
;
348 uint64_t rlimit
, process_limit
, max_size
;
349 bool truncated
, storage_on_tmpfs
;
355 assert(ret_filename
);
359 assert(ret_compressed_size
);
360 assert(ret_truncated
);
362 r
= parse_uid(context
->meta
[META_ARGV_UID
], &uid
);
364 return log_error_errno(r
, "Failed to parse UID: %m");
366 r
= safe_atou64(context
->meta
[META_ARGV_RLIMIT
], &rlimit
);
368 return log_error_errno(r
, "Failed to parse resource limit '%s': %m",
369 context
->meta
[META_ARGV_RLIMIT
]);
370 if (rlimit
< page_size())
371 /* Is coredumping disabled? Then don't bother saving/processing the
372 * coredump. Anything below PAGE_SIZE cannot give a readable coredump
373 * (the kernel uses ELF_EXEC_PAGESIZE which is not easily accessible, but
374 * is usually the same as PAGE_SIZE. */
375 return log_info_errno(SYNTHETIC_ERRNO(EBADSLT
),
376 "Resource limits disable core dumping for process %s (%s).",
377 context
->meta
[META_ARGV_PID
], context
->meta
[META_COMM
]);
379 process_limit
= MAX(arg_process_size_max
, storage_size_max());
380 if (process_limit
== 0)
381 return log_debug_errno(SYNTHETIC_ERRNO(EBADSLT
),
382 "Limits for coredump processing and storage are both 0, not dumping core.");
384 /* Never store more than the process configured, or than we actually shall keep or process */
385 max_size
= MIN(rlimit
, process_limit
);
387 r
= make_filename(context
, &fn
);
389 return log_error_errno(r
, "Failed to determine coredump file name: %m");
391 (void) mkdir_parents_label(fn
, 0755);
393 fd
= open_tmpfile_linkable(fn
, O_RDWR
|O_CLOEXEC
, &tmp
);
395 return log_error_errno(fd
, "Failed to create temporary file for coredump %s: %m", fn
);
397 /* If storage is on tmpfs, the kernel oomd might kill us if there's MemoryMax set on
398 * the service or the slice it belongs to. This is common on low-resources systems,
399 * to avoid crashing processes to take away too many system resources.
400 * Check the cgroup settings, and set max_size to a bit less than half of the
401 * available memory left to the process.
402 * Then, attempt to write the core file uncompressed first - if the write gets
403 * interrupted, we know we won't be able to write it all, so instead compress what
404 * was written so far, delete the uncompressed truncated core, and then continue
405 * compressing from STDIN. Given the compressed core cannot be larger than the
406 * uncompressed one, and 1KB for metadata is accounted for in the calculation, we
407 * should be able to at least store the full compressed core file. */
409 storage_on_tmpfs
= fd_is_temporary_fs(fd
) > 0;
410 if (storage_on_tmpfs
&& arg_compress
) {
411 _cleanup_(sd_bus_flush_close_unrefp
) sd_bus
*bus
= NULL
;
412 uint64_t cgroup_limit
= UINT64_MAX
;
415 /* If we can't get the cgroup limit, just ignore it, but don't fail,
416 * try anyway with the config settings. */
417 r
= sd_bus_default_system(&bus
);
419 log_info_errno(r
, "Failed to connect to system bus, skipping MemoryAvailable check: %m");
421 _cleanup_(sd_bus_error_free
) sd_bus_error error
= SD_BUS_ERROR_NULL
;
423 r
= sd_bus_get_property_trivial(
425 "org.freedesktop.systemd1",
426 "/org/freedesktop/systemd1/unit/self",
427 "org.freedesktop.systemd1.Service",
433 "Failed to query MemoryAvailable for current unit, "
434 "falling back to static config settings: %s",
435 bus_error_message(&error
, r
));
438 max_size
= MIN(cgroup_limit
, max_size
);
439 max_size
= LESS_BY(max_size
, 1024U) / 2; /* Account for 1KB metadata overhead for compressing */
440 max_size
= MAX(PROCESS_SIZE_MIN
, max_size
); /* Impose a lower minimum */
442 /* tmpfs might get full quickly, so check the available space too.
443 * But don't worry about errors here, failing to access the storage
444 * location will be better logged when writing to it. */
445 if (statvfs("/var/lib/systemd/coredump/", &sv
) >= 0)
446 max_size
= MIN((uint64_t)sv
.f_frsize
* (uint64_t)sv
.f_bfree
, max_size
);
448 log_debug("Limiting core file size to %" PRIu64
" bytes due to cgroup memory limits.", max_size
);
451 r
= copy_bytes(input_fd
, fd
, max_size
, 0);
453 return log_error_errno(r
, "Cannot store coredump of %s (%s): %m",
454 context
->meta
[META_ARGV_PID
], context
->meta
[META_COMM
]);
459 _cleanup_(unlink_and_freep
) char *tmp_compressed
= NULL
;
460 _cleanup_free_
char *fn_compressed
= NULL
;
461 _cleanup_close_
int fd_compressed
= -EBADF
;
462 uint64_t uncompressed_size
= 0;
464 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
465 return log_error_errno(errno
, "Failed to seek on coredump %s: %m", fn
);
467 fn_compressed
= strjoin(fn
, default_compression_extension());
471 fd_compressed
= open_tmpfile_linkable(fn_compressed
, O_RDWR
|O_CLOEXEC
, &tmp_compressed
);
472 if (fd_compressed
< 0)
473 return log_error_errno(fd_compressed
, "Failed to create temporary file for coredump %s: %m", fn_compressed
);
475 r
= compress_stream(fd
, fd_compressed
, max_size
, &uncompressed_size
);
477 return log_error_errno(r
, "Failed to compress %s: %m", coredump_tmpfile_name(tmp_compressed
));
479 if (truncated
&& storage_on_tmpfs
) {
480 uint64_t partial_uncompressed_size
= 0;
482 /* Uncompressed write was truncated and we are writing to tmpfs: delete
483 * the uncompressed core, and compress the remaining part from STDIN. */
485 tmp
= unlink_and_free(tmp
);
488 r
= compress_stream(input_fd
, fd_compressed
, max_size
, &partial_uncompressed_size
);
490 return log_error_errno(r
, "Failed to compress %s: %m", coredump_tmpfile_name(tmp_compressed
));
491 uncompressed_size
+= partial_uncompressed_size
;
494 r
= fix_permissions(fd_compressed
, tmp_compressed
, fn_compressed
, context
, uid
);
498 if (fstat(fd_compressed
, &st
) < 0)
499 return log_error_errno(errno
,
500 "Failed to fstat core file %s: %m",
501 coredump_tmpfile_name(tmp_compressed
));
503 *ret_filename
= TAKE_PTR(fn_compressed
); /* compressed */
504 *ret_node_fd
= TAKE_FD(fd_compressed
); /* compressed */
505 *ret_compressed_size
= (uint64_t) st
.st_size
; /* compressed */
506 *ret_data_fd
= TAKE_FD(fd
);
507 *ret_size
= uncompressed_size
;
508 *ret_truncated
= truncated
;
509 tmp_compressed
= mfree(tmp_compressed
);
517 LOG_MESSAGE("Core file was truncated to %"PRIu64
" bytes.", max_size
),
518 "SIZE_LIMIT=%"PRIu64
, max_size
,
519 "MESSAGE_ID=" SD_MESSAGE_TRUNCATED_CORE_STR
);
521 r
= fix_permissions(fd
, tmp
, fn
, context
, uid
);
523 return log_error_errno(r
, "Failed to fix permissions and finalize coredump %s into %s: %m", coredump_tmpfile_name(tmp
), fn
);
525 if (fstat(fd
, &st
) < 0)
526 return log_error_errno(errno
, "Failed to fstat core file %s: %m", coredump_tmpfile_name(tmp
));
528 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
529 return log_error_errno(errno
, "Failed to seek on coredump %s: %m", fn
);
531 *ret_filename
= TAKE_PTR(fn
);
532 *ret_data_fd
= TAKE_FD(fd
);
533 *ret_size
= (uint64_t) st
.st_size
;
534 *ret_truncated
= truncated
;
539 static int allocate_journal_field(int fd
, size_t size
, char **ret
, size_t *ret_size
) {
540 _cleanup_free_
char *field
= NULL
;
547 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
548 return log_warning_errno(errno
, "Failed to seek: %m");
550 field
= malloc(9 + size
);
552 log_warning("Failed to allocate memory for coredump, coredump will not be stored.");
556 memcpy(field
, "COREDUMP=", 9);
558 n
= read(fd
, field
+ 9, size
);
560 return log_error_errno((int) n
, "Failed to read core data: %m");
561 if ((size_t) n
< size
)
562 return log_error_errno(SYNTHETIC_ERRNO(EIO
),
563 "Core data too short.");
565 *ret
= TAKE_PTR(field
);
566 *ret_size
= size
+ 9;
571 /* Joins /proc/[pid]/fd/ and /proc/[pid]/fdinfo/ into the following lines:
585 static int compose_open_fds(pid_t pid
, char **open_fds
) {
586 _cleanup_closedir_
DIR *proc_fd_dir
= NULL
;
587 _cleanup_close_
int proc_fdinfo_fd
= -EBADF
;
588 _cleanup_free_
char *buffer
= NULL
;
589 _cleanup_fclose_
FILE *stream
= NULL
;
590 const char *fddelim
= "", *path
;
595 assert(open_fds
!= NULL
);
597 path
= procfs_file_alloca(pid
, "fd");
598 proc_fd_dir
= opendir(path
);
602 proc_fdinfo_fd
= openat(dirfd(proc_fd_dir
), "../fdinfo", O_DIRECTORY
|O_NOFOLLOW
|O_CLOEXEC
|O_PATH
);
603 if (proc_fdinfo_fd
< 0)
606 stream
= open_memstream_unlocked(&buffer
, &size
);
610 FOREACH_DIRENT(de
, proc_fd_dir
, return -errno
) {
611 _cleanup_fclose_
FILE *fdinfo
= NULL
;
612 _cleanup_free_
char *fdname
= NULL
;
613 _cleanup_close_
int fd
= -EBADF
;
615 r
= readlinkat_malloc(dirfd(proc_fd_dir
), de
->d_name
, &fdname
);
619 fprintf(stream
, "%s%s:%s\n", fddelim
, de
->d_name
, fdname
);
622 /* Use the directory entry from /proc/[pid]/fd with /proc/[pid]/fdinfo */
623 fd
= openat(proc_fdinfo_fd
, de
->d_name
, O_NOFOLLOW
|O_CLOEXEC
|O_RDONLY
);
627 fdinfo
= take_fdopen(&fd
, "r");
632 _cleanup_free_
char *line
= NULL
;
634 r
= read_line(fdinfo
, LONG_LINE_MAX
, &line
);
646 stream
= safe_fclose(stream
);
651 *open_fds
= TAKE_PTR(buffer
);
656 static int get_process_ns(pid_t pid
, const char *namespace, ino_t
*ns
) {
659 _cleanup_close_
int proc_ns_dir_fd
= -EBADF
;
661 p
= procfs_file_alloca(pid
, "ns");
663 proc_ns_dir_fd
= open(p
, O_DIRECTORY
| O_CLOEXEC
| O_RDONLY
);
664 if (proc_ns_dir_fd
< 0)
667 if (fstatat(proc_ns_dir_fd
, namespace, &stbuf
, /* flags */0) < 0)
674 static int get_mount_namespace_leader(pid_t pid
, pid_t
*ret
) {
678 r
= get_process_ns(pid
, "mnt", &proc_mntns
);
686 r
= get_process_ppid(pid
, &ppid
);
687 if (r
== -EADDRNOTAVAIL
) /* Reached the top (i.e. typically PID 1, but could also be a process
688 * whose parent is not in our pidns) */
693 r
= get_process_ns(ppid
, "mnt", &parent_mntns
);
697 if (proc_mntns
!= parent_mntns
) {
706 /* Returns 1 if the parent was found.
707 * Returns 0 if there is not a process we can call the pid's
708 * container parent (the pid's process isn't 'containerized').
709 * Returns a negative number on errors.
711 static int get_process_container_parent_cmdline(pid_t pid
, char** cmdline
) {
713 const char *proc_root_path
;
714 struct stat root_stat
, proc_root_stat
;
717 /* To compare inodes of / and /proc/[pid]/root */
718 if (stat("/", &root_stat
) < 0)
721 proc_root_path
= procfs_file_alloca(pid
, "root");
722 if (stat(proc_root_path
, &proc_root_stat
) < 0)
725 /* The process uses system root. */
726 if (stat_inode_same(&proc_root_stat
, &root_stat
)) {
731 r
= get_mount_namespace_leader(pid
, &container_pid
);
735 r
= get_process_cmdline(container_pid
, SIZE_MAX
, PROCESS_CMDLINE_QUOTE_POSIX
, cmdline
);
742 static int change_uid_gid(const Context
*context
) {
747 r
= parse_uid(context
->meta
[META_ARGV_UID
], &uid
);
751 if (uid_is_system(uid
)) {
752 const char *user
= "systemd-coredump";
754 r
= get_user_creds(&user
, &uid
, &gid
, NULL
, NULL
, 0);
756 log_warning_errno(r
, "Cannot resolve %s user. Proceeding to dump core as root: %m", user
);
760 r
= parse_gid(context
->meta
[META_ARGV_GID
], &gid
);
765 return drop_privileges(uid
, gid
, 0);
768 static int submit_coredump(
770 struct iovec_wrapper
*iovw
,
773 _cleanup_(json_variant_unrefp
) JsonVariant
*json_metadata
= NULL
;
774 _cleanup_close_
int coredump_fd
= -EBADF
, coredump_node_fd
= -EBADF
;
775 _cleanup_free_
char *filename
= NULL
, *coredump_data
= NULL
;
776 _cleanup_free_
char *stacktrace
= NULL
;
778 const char *module_name
;
779 uint64_t coredump_size
= UINT64_MAX
, coredump_compressed_size
= UINT64_MAX
;
780 bool truncated
= false;
781 JsonVariant
*module_json
;
786 assert(input_fd
>= 0);
788 /* Vacuum before we write anything again */
789 (void) coredump_vacuum(-1, arg_keep_free
, arg_max_use
);
791 /* Always stream the coredump to disk, if that's possible */
792 r
= save_external_coredump(context
, input_fd
,
793 &filename
, &coredump_node_fd
, &coredump_fd
,
794 &coredump_size
, &coredump_compressed_size
, &truncated
);
796 /* Skip whole core dumping part */
799 /* If we don't want to keep the coredump on disk, remove it now, as later on we
800 * will lack the privileges for it. However, we keep the fd to it, so that we can
801 * still process it and log it. */
802 r
= maybe_remove_external_coredump(filename
, coredump_node_fd
>= 0 ? coredump_compressed_size
: coredump_size
);
806 (void) iovw_put_string_field(iovw
, "COREDUMP_FILENAME=", filename
);
807 else if (arg_storage
== COREDUMP_STORAGE_EXTERNAL
)
808 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
809 coredump_node_fd
>= 0 ? coredump_compressed_size
: coredump_size
, arg_external_size_max
);
811 /* Vacuum again, but exclude the coredump we just created */
812 (void) coredump_vacuum(coredump_node_fd
>= 0 ? coredump_node_fd
: coredump_fd
, arg_keep_free
, arg_max_use
);
814 /* Now, let's drop privileges to become the user who owns the segfaulted process
815 * and allocate the coredump memory under the user's uid. This also ensures that
816 * the credentials journald will see are the ones of the coredumping user, thus
817 * making sure the user gets access to the core dump. Let's also get rid of all
818 * capabilities, if we run as root, we won't need them anymore. */
819 r
= change_uid_gid(context
);
821 return log_error_errno(r
, "Failed to drop privileges: %m");
823 /* Try to get a stack trace if we can */
824 if (coredump_size
> arg_process_size_max
)
825 log_debug("Not generating stack trace: core size %"PRIu64
" is greater "
826 "than %"PRIu64
" (the configured maximum)",
827 coredump_size
, arg_process_size_max
);
828 else if (coredump_fd
>= 0) {
829 bool skip
= startswith(context
->meta
[META_COMM
], "systemd-coredum"); /* COMM is 16 bytes usually */
831 (void) parse_elf_object(coredump_fd
,
832 context
->meta
[META_EXE
],
833 /* fork_disable_dump= */ skip
, /* avoid loops */
839 core_message
= strjoina("Process ", context
->meta
[META_ARGV_PID
],
840 " (", context
->meta
[META_COMM
], ") of user ",
841 context
->meta
[META_ARGV_UID
], " dumped core.",
842 context
->is_journald
&& filename
? "\nCoredump diverted to " : NULL
,
843 context
->is_journald
&& filename
? filename
: NULL
);
845 core_message
= strjoina(core_message
, stacktrace
? "\n\n" : NULL
, stacktrace
);
847 if (context
->is_journald
)
848 /* We might not be able to log to the journal, so let's always print the message to another
849 * log target. The target was set previously to something safe. */
850 log_dispatch(LOG_ERR
, 0, core_message
);
852 (void) iovw_put_string_field(iovw
, "MESSAGE=", core_message
);
855 (void) iovw_put_string_field(iovw
, "COREDUMP_TRUNCATED=", "1");
857 /* If we managed to parse any ELF metadata (build-id, ELF package meta),
858 * attach it as journal metadata. */
860 _cleanup_free_
char *formatted_json
= NULL
;
862 r
= json_variant_format(json_metadata
, 0, &formatted_json
);
864 return log_error_errno(r
, "Failed to format JSON package metadata: %m");
866 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_JSON=", formatted_json
);
869 /* In the unlikely scenario that context->meta[META_EXE] is not available,
870 * let's avoid guessing the module name and skip the loop. */
871 if (context
->meta
[META_EXE
])
872 JSON_VARIANT_OBJECT_FOREACH(module_name
, module_json
, json_metadata
) {
875 /* We only add structured fields for the 'main' ELF module, and only if we can identify it. */
876 if (!path_equal_filename(module_name
, context
->meta
[META_EXE
]))
879 t
= json_variant_by_key(module_json
, "name");
881 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_NAME=", json_variant_string(t
));
883 t
= json_variant_by_key(module_json
, "version");
885 (void) iovw_put_string_field(iovw
, "COREDUMP_PACKAGE_VERSION=", json_variant_string(t
));
888 /* Optionally store the entire coredump in the journal */
889 if (arg_storage
== COREDUMP_STORAGE_JOURNAL
&& coredump_fd
>= 0) {
890 if (coredump_size
<= arg_journal_size_max
) {
893 /* Store the coredump itself in the journal */
895 r
= allocate_journal_field(coredump_fd
, (size_t) coredump_size
, &coredump_data
, &sz
);
897 if (iovw_put(iovw
, coredump_data
, sz
) >= 0)
898 TAKE_PTR(coredump_data
);
900 log_warning_errno(r
, "Failed to attach the core to the journal entry: %m");
902 log_info("The core will not be stored: size %"PRIu64
" is greater than %"PRIu64
" (the configured maximum)",
903 coredump_size
, arg_journal_size_max
);
906 /* If journald is coredumping, we have to be careful that we don't deadlock when trying to write the
907 * coredump to the journal, so we put the journal socket in nonblocking mode before trying to write
908 * the coredump to the socket. */
910 if (context
->is_journald
) {
911 r
= journal_fd_nonblock(true);
913 return log_error_errno(r
, "Failed to make journal socket non-blocking: %m");
916 r
= sd_journal_sendv(iovw
->iovec
, iovw
->count
);
918 if (context
->is_journald
) {
921 k
= journal_fd_nonblock(false);
923 return log_error_errno(k
, "Failed to make journal socket blocking: %m");
926 if (r
== -EAGAIN
&& context
->is_journald
)
927 log_warning_errno(r
, "Failed to log journal coredump, ignoring: %m");
929 return log_error_errno(r
, "Failed to log coredump: %m");
934 static int save_context(Context
*context
, const struct iovec_wrapper
*iovw
) {
940 assert(iovw
->count
>= _META_ARGV_MAX
);
942 /* The context does not allocate any memory on its own */
944 for (size_t n
= 0; n
< iovw
->count
; n
++) {
945 struct iovec
*iovec
= iovw
->iovec
+ n
;
947 for (size_t i
= 0; i
< ELEMENTSOF(meta_field_names
); i
++) {
950 /* Note that these strings are NUL terminated, because we made sure that a
951 * trailing NUL byte is in the buffer, though not included in the iov_len
952 * count (see process_socket() and gather_pid_metadata_*()) */
953 assert(((char*) iovec
->iov_base
)[iovec
->iov_len
] == 0);
955 p
= startswith(iovec
->iov_base
, meta_field_names
[i
]);
957 context
->meta
[i
] = p
;
963 if (!context
->meta
[META_ARGV_PID
])
964 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
965 "Failed to find the PID of crashing process");
967 r
= parse_pid(context
->meta
[META_ARGV_PID
], &context
->pid
);
969 return log_error_errno(r
, "Failed to parse PID \"%s\": %m", context
->meta
[META_ARGV_PID
]);
971 unit
= context
->meta
[META_UNIT
];
972 context
->is_pid1
= streq(context
->meta
[META_ARGV_PID
], "1") || streq_ptr(unit
, SPECIAL_INIT_SCOPE
);
973 context
->is_journald
= streq_ptr(unit
, SPECIAL_JOURNALD_SERVICE
);
978 static int process_socket(int fd
) {
979 _cleanup_close_
int input_fd
= -EBADF
;
980 Context context
= {};
981 struct iovec_wrapper iovw
= {};
989 log_debug("Processing coredump received on stdin...");
992 CMSG_BUFFER_TYPE(CMSG_SPACE(sizeof(int))) control
;
994 .msg_control
= &control
,
995 .msg_controllen
= sizeof(control
),
1001 l
= next_datagram_size_fd(fd
);
1003 r
= log_error_errno(l
, "Failed to determine datagram size to read: %m");
1008 iovec
.iov_base
= malloc(l
+ 1);
1009 if (!iovec
.iov_base
) {
1014 mh
.msg_iov
= &iovec
;
1016 n
= recvmsg_safe(fd
, &mh
, MSG_CMSG_CLOEXEC
);
1018 free(iovec
.iov_base
);
1019 r
= log_error_errno(n
, "Failed to receive datagram: %m");
1023 /* The final zero-length datagram carries the file descriptor and tells us
1024 * that we're done. */
1026 struct cmsghdr
*found
;
1028 free(iovec
.iov_base
);
1030 found
= cmsg_find(&mh
, SOL_SOCKET
, SCM_RIGHTS
, CMSG_LEN(sizeof(int)));
1032 cmsg_close_all(&mh
);
1033 r
= log_error_errno(SYNTHETIC_ERRNO(EBADMSG
),
1034 "Coredump file descriptor missing.");
1038 assert(input_fd
< 0);
1039 input_fd
= *(int*) CMSG_DATA(found
);
1042 cmsg_close_all(&mh
);
1044 /* Add trailing NUL byte, in case these are strings */
1045 ((char*) iovec
.iov_base
)[n
] = 0;
1046 iovec
.iov_len
= (size_t) n
;
1048 r
= iovw_put(&iovw
, iovec
.iov_base
, iovec
.iov_len
);
1053 /* Make sure we got all data we really need */
1054 assert(input_fd
>= 0);
1056 r
= save_context(&context
, &iovw
);
1060 /* Make sure we received at least all fields we need. */
1061 for (int i
= 0; i
< _META_MANDATORY_MAX
; i
++)
1062 if (!context
.meta
[i
]) {
1063 r
= log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1064 "A mandatory argument (%i) has not been sent, aborting.",
1069 r
= submit_coredump(&context
, &iovw
, input_fd
);
1072 iovw_free_contents(&iovw
, true);
1076 static int send_iovec(const struct iovec_wrapper
*iovw
, int input_fd
) {
1077 _cleanup_close_
int fd
= -EBADF
;
1081 assert(input_fd
>= 0);
1083 fd
= socket(AF_UNIX
, SOCK_SEQPACKET
|SOCK_CLOEXEC
, 0);
1085 return log_error_errno(errno
, "Failed to create coredump socket: %m");
1087 r
= connect_unix_path(fd
, AT_FDCWD
, "/run/systemd/coredump");
1089 return log_error_errno(r
, "Failed to connect to coredump service: %m");
1091 for (size_t i
= 0; i
< iovw
->count
; i
++) {
1092 struct msghdr mh
= {
1093 .msg_iov
= iovw
->iovec
+ i
,
1096 struct iovec copy
[2];
1099 if (sendmsg(fd
, &mh
, MSG_NOSIGNAL
) >= 0)
1102 if (errno
== EMSGSIZE
&& mh
.msg_iov
[0].iov_len
> 0) {
1103 /* This field didn't fit? That's a pity. Given that this is
1104 * just metadata, let's truncate the field at half, and try
1105 * again. We append three dots, in order to show that this is
1108 if (mh
.msg_iov
!= copy
) {
1109 /* We don't want to modify the caller's iovec, hence
1110 * let's create our own array, consisting of two new
1111 * iovecs, where the first is a (truncated) copy of
1112 * what we want to send, and the second one contains
1113 * the trailing dots. */
1114 copy
[0] = iovw
->iovec
[i
];
1115 copy
[1] = IOVEC_MAKE(((char[]){'.', '.', '.'}), 3);
1121 copy
[0].iov_len
/= 2; /* halve it, and try again */
1125 return log_error_errno(errno
, "Failed to send coredump datagram: %m");
1129 r
= send_one_fd(fd
, input_fd
, 0);
1131 return log_error_errno(r
, "Failed to send coredump fd: %m");
1136 static int gather_pid_metadata_from_argv(
1137 struct iovec_wrapper
*iovw
,
1139 int argc
, char **argv
) {
1141 _cleanup_free_
char *free_timestamp
= NULL
;
1145 /* We gather all metadata that were passed via argv[] into an array of iovecs that
1146 * we'll forward to the socket unit */
1148 if (argc
< _META_ARGV_MAX
)
1149 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1150 "Not enough arguments passed by the kernel (%i, expected %i).",
1151 argc
, _META_ARGV_MAX
);
1153 for (int i
= 0; i
< _META_ARGV_MAX
; i
++) {
1159 case META_ARGV_TIMESTAMP
:
1160 /* The journal fields contain the timestamp padded with six
1161 * zeroes, so that the kernel-supplied 1s granularity timestamps
1162 * becomes 1µs granularity, i.e. the granularity systemd usually
1164 t
= free_timestamp
= strjoin(argv
[i
], "000000");
1169 case META_ARGV_SIGNAL
:
1170 /* For signal, record its pretty name too */
1171 if (safe_atoi(argv
[i
], &signo
) >= 0 && SIGNAL_VALID(signo
))
1172 (void) iovw_put_string_field(iovw
, "COREDUMP_SIGNAL_NAME=SIG",
1173 signal_to_string(signo
));
1180 r
= iovw_put_string_field(iovw
, meta_field_names
[i
], t
);
1185 /* Cache some of the process metadata we collected so far and that we'll need to
1187 return save_context(context
, iovw
);
1190 static int gather_pid_metadata(struct iovec_wrapper
*iovw
, Context
*context
) {
1197 /* Note that if we fail on oom later on, we do not roll-back changes to the iovec
1198 * structure. (It remains valid, with the first iovec fields initialized.) */
1202 /* The following is mandatory */
1203 r
= get_process_comm(pid
, &t
);
1205 return log_error_errno(r
, "Failed to get COMM: %m");
1207 r
= iovw_put_string_field_free(iovw
, "COREDUMP_COMM=", t
);
1211 /* The following are optional, but we use them if present. */
1212 r
= get_process_exe(pid
, &t
);
1214 r
= iovw_put_string_field_free(iovw
, "COREDUMP_EXE=", t
);
1216 log_warning_errno(r
, "Failed to get EXE, ignoring: %m");
1218 if (cg_pid_get_unit(pid
, &t
) >= 0)
1219 (void) iovw_put_string_field_free(iovw
, "COREDUMP_UNIT=", t
);
1221 if (cg_pid_get_user_unit(pid
, &t
) >= 0)
1222 (void) iovw_put_string_field_free(iovw
, "COREDUMP_USER_UNIT=", t
);
1224 if (sd_pid_get_session(pid
, &t
) >= 0)
1225 (void) iovw_put_string_field_free(iovw
, "COREDUMP_SESSION=", t
);
1227 if (sd_pid_get_owner_uid(pid
, &owner_uid
) >= 0) {
1228 r
= asprintf(&t
, UID_FMT
, owner_uid
);
1230 (void) iovw_put_string_field_free(iovw
, "COREDUMP_OWNER_UID=", t
);
1233 if (sd_pid_get_slice(pid
, &t
) >= 0)
1234 (void) iovw_put_string_field_free(iovw
, "COREDUMP_SLICE=", t
);
1236 if (get_process_cmdline(pid
, SIZE_MAX
, PROCESS_CMDLINE_QUOTE_POSIX
, &t
) >= 0)
1237 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CMDLINE=", t
);
1239 if (cg_pid_get_path_shifted(pid
, NULL
, &t
) >= 0)
1240 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CGROUP=", t
);
1242 if (compose_open_fds(pid
, &t
) >= 0)
1243 (void) iovw_put_string_field_free(iovw
, "COREDUMP_OPEN_FDS=", t
);
1245 p
= procfs_file_alloca(pid
, "status");
1246 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1247 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_STATUS=", t
);
1249 p
= procfs_file_alloca(pid
, "maps");
1250 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1251 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_MAPS=", t
);
1253 p
= procfs_file_alloca(pid
, "limits");
1254 if (read_full_virtual_file(p
, &t
, NULL
) >= 0)
1255 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_LIMITS=", t
);
1257 p
= procfs_file_alloca(pid
, "cgroup");
1258 if (read_full_virtual_file(p
, &t
, NULL
) >=0)
1259 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_CGROUP=", t
);
1261 p
= procfs_file_alloca(pid
, "mountinfo");
1262 if (read_full_virtual_file(p
, &t
, NULL
) >=0)
1263 (void) iovw_put_string_field_free(iovw
, "COREDUMP_PROC_MOUNTINFO=", t
);
1265 if (get_process_cwd(pid
, &t
) >= 0)
1266 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CWD=", t
);
1268 if (get_process_root(pid
, &t
) >= 0) {
1269 bool proc_self_root_is_slash
;
1271 proc_self_root_is_slash
= strcmp(t
, "/") == 0;
1273 (void) iovw_put_string_field_free(iovw
, "COREDUMP_ROOT=", t
);
1275 /* If the process' root is "/", then there is a chance it has
1276 * mounted own root and hence being containerized. */
1277 if (proc_self_root_is_slash
&& get_process_container_parent_cmdline(pid
, &t
) > 0)
1278 (void) iovw_put_string_field_free(iovw
, "COREDUMP_CONTAINER_CMDLINE=", t
);
1281 if (get_process_environ(pid
, &t
) >= 0)
1282 (void) iovw_put_string_field_free(iovw
, "COREDUMP_ENVIRON=", t
);
1284 /* we successfully acquired all metadata */
1285 return save_context(context
, iovw
);
1288 static int process_kernel(int argc
, char* argv
[]) {
1289 Context context
= {};
1290 struct iovec_wrapper
*iovw
;
1293 /* When we're invoked by the kernel, stdout/stderr are closed which is dangerous because the fds
1294 * could get reallocated. To avoid hard to debug issues, let's instead bind stdout/stderr to
1296 r
= rearrange_stdio(STDIN_FILENO
, -1, -1);
1298 return log_error_errno(r
, "Failed to connect stdout/stderr to /dev/null: %m");
1300 log_debug("Processing coredump received from the kernel...");
1306 (void) iovw_put_string_field(iovw
, "MESSAGE_ID=", SD_MESSAGE_COREDUMP_STR
);
1307 (void) iovw_put_string_field(iovw
, "PRIORITY=", STRINGIFY(LOG_CRIT
));
1309 /* Collect all process metadata passed by the kernel through argv[] */
1310 r
= gather_pid_metadata_from_argv(iovw
, &context
, argc
- 1, argv
+ 1);
1314 /* Collect the rest of the process metadata retrieved from the runtime */
1315 r
= gather_pid_metadata(iovw
, &context
);
1319 if (!context
.is_journald
) {
1320 /* OK, now we know it's not the journal, hence we can make use of it now. */
1321 log_set_target(LOG_TARGET_JOURNAL_OR_KMSG
);
1325 /* If this is PID 1 disable coredump collection, we'll unlikely be able to process
1328 * FIXME: maybe we should disable coredumps generation from the beginning and
1329 * re-enable it only when we know it's either safe (ie we're not running OOM) or
1330 * it's not pid1 ? */
1331 if (context
.is_pid1
) {
1332 log_notice("Due to PID 1 having crashed coredump collection will now be turned off.");
1333 disable_coredumps();
1336 if (context
.is_journald
|| context
.is_pid1
)
1337 r
= submit_coredump(&context
, iovw
, STDIN_FILENO
);
1339 r
= send_iovec(iovw
, STDIN_FILENO
);
1342 iovw
= iovw_free_free(iovw
);
1346 static int process_backtrace(int argc
, char *argv
[]) {
1347 Context context
= {};
1348 struct iovec_wrapper
*iovw
;
1351 _cleanup_(journal_importer_cleanup
) JournalImporter importer
= JOURNAL_IMPORTER_INIT(STDIN_FILENO
);
1353 log_debug("Processing backtrace on stdin...");
1359 (void) iovw_put_string_field(iovw
, "MESSAGE_ID=", SD_MESSAGE_BACKTRACE_STR
);
1360 (void) iovw_put_string_field(iovw
, "PRIORITY=", STRINGIFY(LOG_CRIT
));
1362 /* Collect all process metadata from argv[] by making sure to skip the
1363 * '--backtrace' option */
1364 r
= gather_pid_metadata_from_argv(iovw
, &context
, argc
- 2, argv
+ 2);
1368 /* Collect the rest of the process metadata retrieved from the runtime */
1369 r
= gather_pid_metadata(iovw
, &context
);
1374 r
= journal_importer_process_data(&importer
);
1376 log_error_errno(r
, "Failed to parse journal entry on stdin: %m");
1379 if (r
== 1 || /* complete entry */
1380 journal_importer_eof(&importer
)) /* end of data */
1384 if (journal_importer_eof(&importer
)) {
1385 log_warning("Did not receive a full journal entry on stdin, ignoring message sent by reporter");
1387 message
= strjoina("Process ", context
.meta
[META_ARGV_PID
],
1388 " (", context
.meta
[META_COMM
], ")"
1389 " of user ", context
.meta
[META_ARGV_UID
],
1390 " failed with ", context
.meta
[META_ARGV_SIGNAL
]);
1392 r
= iovw_put_string_field(iovw
, "MESSAGE=", message
);
1396 /* The imported iovecs are not supposed to be freed by us so let's store
1397 * them at the end of the array so we can skip them while freeing the
1399 for (size_t i
= 0; i
< importer
.iovw
.count
; i
++) {
1400 struct iovec
*iovec
= importer
.iovw
.iovec
+ i
;
1402 iovw_put(iovw
, iovec
->iov_base
, iovec
->iov_len
);
1406 r
= sd_journal_sendv(iovw
->iovec
, iovw
->count
);
1408 log_error_errno(r
, "Failed to log backtrace: %m");
1411 iovw
->count
-= importer
.iovw
.count
;
1412 iovw
= iovw_free_free(iovw
);
1416 static int run(int argc
, char *argv
[]) {
1419 /* First, log to a safe place, since we don't know what crashed and it might
1420 * be journald which we'd rather not log to then. */
1422 log_set_target(LOG_TARGET_KMSG
);
1425 /* Make sure we never enter a loop */
1426 (void) prctl(PR_SET_DUMPABLE
, 0);
1428 /* Ignore all parse errors */
1429 (void) parse_config();
1431 log_debug("Selected storage '%s'.", coredump_storage_to_string(arg_storage
));
1432 log_debug("Selected compression %s.", yes_no(arg_compress
));
1434 r
= sd_listen_fds(false);
1436 return log_error_errno(r
, "Failed to determine the number of file descriptors: %m");
1438 /* If we got an fd passed, we are running in coredumpd mode. Otherwise we
1439 * are invoked from the kernel as coredump handler. */
1441 if (streq_ptr(argv
[1], "--backtrace"))
1442 return process_backtrace(argc
, argv
);
1444 return process_kernel(argc
, argv
);
1446 return process_socket(SD_LISTEN_FDS_START
);
1448 return log_error_errno(SYNTHETIC_ERRNO(EINVAL
),
1449 "Received unexpected number of file descriptors.");
1452 DEFINE_MAIN_FUNCTION(run
);