]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/journal/journald-context.c
8253a45128c6a29e61424153700f75d1ce99b0a5
1 /* SPDX-License-Identifier: LGPL-2.1+ */
4 #include <selinux/selinux.h>
7 #include "alloc-util.h"
8 #include "audit-util.h"
9 #include "cgroup-util.h"
14 #include "journal-util.h"
15 #include "journald-context.h"
16 #include "parse-util.h"
17 #include "path-util.h"
18 #include "process-util.h"
19 #include "string-util.h"
20 #include "syslog-util.h"
21 #include "unaligned.h"
22 #include "user-util.h"
24 /* This implements a metadata cache for clients, which are identified by their PID. Requesting metadata through /proc
25 * is expensive, hence let's cache the data if we can. Note that this means the metadata might be out-of-date when we
26 * store it, but it might already be anyway, as we request the data asynchronously from /proc at a different time the
27 * log entry was originally created. We hence just increase the "window of inaccuracy" a bit.
29 * The cache is indexed by the PID. Entries may be "pinned" in the cache, in which case the entries are not removed
30 * until they are unpinned. Unpinned entries are kept around until cache pressure is seen. Cache entries older than 5s
31 * are never used (a sad attempt to deal with the UNIX weakness of PIDs reuse), cache entries older than 1s are
32 * refreshed in an incremental way (meaning: data is reread from /proc, but any old data we can't refresh is not
33 * flushed out). Data newer than 1s is used immediately without refresh.
35 * Log stream clients (i.e. all clients using the AF_UNIX/SOCK_STREAM stdout/stderr transport) will pin a cache entry
36 * as long as their socket is connected. Note that cache entries are shared between different transports. That means a
37 * cache entry pinned for the stream connection logic may be reused for the syslog or native protocols.
39 * Caching metadata like this has two major benefits:
41 * 1. Reading metadata is expensive, and we can thus substantially speed up log processing under flood.
43 * 2. Because metadata caching is shared between stream and datagram transports and stream connections pin a cache
44 * entry there's a good chance we can properly map a substantial set of datagram log messages to their originating
45 * service, as all services (unless explicitly configured otherwise) will have their stdout/stderr connected to a
46 * stream connection. This should improve cases where a service process logs immediately before exiting and we
47 * previously had trouble associating the log message with the service.
49 * NB: With and without the metadata cache: the implicitly added entry metadata in the journal (with the exception of
50 * UID/PID/GID and SELinux label) must be understood as possibly slightly out of sync (i.e. sometimes slighly older
51 * and sometimes slightly newer than what was current at the log event).
54 /* We refresh every 1s */
55 #define REFRESH_USEC (1*USEC_PER_SEC)
57 /* Data older than 5s we flush out */
58 #define MAX_USEC (5*USEC_PER_SEC)
60 /* Keep at most 16K entries in the cache. (Note though that this limit may be violated if enough streams pin entries in
61 * the cache, in which case we *do* permit this limit to be breached. That's safe however, as the number of stream
62 * clients itself is limited.) */
63 #define CACHE_MAX (16*1024)
65 static int client_context_compare(const void *a
, const void *b
) {
66 const ClientContext
*x
= a
, *y
= b
;
69 r
= CMP(x
->timestamp
, y
->timestamp
);
73 return CMP(x
->pid
, y
->pid
);
76 static int client_context_new(Server
*s
, pid_t pid
, ClientContext
**ret
) {
81 assert(pid_is_valid(pid
));
84 r
= hashmap_ensure_allocated(&s
->client_contexts
, NULL
);
88 r
= prioq_ensure_allocated(&s
->client_contexts_lru
, client_context_compare
);
92 c
= new0(ClientContext
, 1);
100 c
->auditid
= AUDIT_SESSION_INVALID
;
101 c
->loginuid
= UID_INVALID
;
102 c
->owner_uid
= UID_INVALID
;
103 c
->lru_index
= PRIOQ_IDX_NULL
;
104 c
->timestamp
= USEC_INFINITY
;
105 c
->extra_fields_mtime
= NSEC_INFINITY
;
106 c
->log_level_max
= -1;
107 c
->log_rate_limit_interval
= s
->rate_limit_interval
;
108 c
->log_rate_limit_burst
= s
->rate_limit_burst
;
110 r
= hashmap_put(s
->client_contexts
, PID_TO_PTR(pid
), c
);
120 static void client_context_reset(Server
*s
, ClientContext
*c
) {
124 c
->timestamp
= USEC_INFINITY
;
126 c
->uid
= UID_INVALID
;
127 c
->gid
= GID_INVALID
;
129 c
->comm
= mfree(c
->comm
);
130 c
->exe
= mfree(c
->exe
);
131 c
->cmdline
= mfree(c
->cmdline
);
132 c
->capeff
= mfree(c
->capeff
);
134 c
->auditid
= AUDIT_SESSION_INVALID
;
135 c
->loginuid
= UID_INVALID
;
137 c
->cgroup
= mfree(c
->cgroup
);
138 c
->session
= mfree(c
->session
);
139 c
->owner_uid
= UID_INVALID
;
140 c
->unit
= mfree(c
->unit
);
141 c
->user_unit
= mfree(c
->user_unit
);
142 c
->slice
= mfree(c
->slice
);
143 c
->user_slice
= mfree(c
->user_slice
);
145 c
->invocation_id
= SD_ID128_NULL
;
147 c
->label
= mfree(c
->label
);
150 c
->extra_fields_iovec
= mfree(c
->extra_fields_iovec
);
151 c
->extra_fields_n_iovec
= 0;
152 c
->extra_fields_data
= mfree(c
->extra_fields_data
);
153 c
->extra_fields_mtime
= NSEC_INFINITY
;
155 c
->log_level_max
= -1;
157 c
->log_rate_limit_interval
= s
->rate_limit_interval
;
158 c
->log_rate_limit_burst
= s
->rate_limit_burst
;
161 static ClientContext
* client_context_free(Server
*s
, ClientContext
*c
) {
167 assert_se(hashmap_remove(s
->client_contexts
, PID_TO_PTR(c
->pid
)) == c
);
170 assert_se(prioq_remove(s
->client_contexts_lru
, c
, &c
->lru_index
) >= 0);
172 client_context_reset(s
, c
);
177 static void client_context_read_uid_gid(ClientContext
*c
, const struct ucred
*ucred
) {
179 assert(pid_is_valid(c
->pid
));
181 /* The ucred data passed in is always the most current and accurate, if we have any. Use it. */
182 if (ucred
&& uid_is_valid(ucred
->uid
))
185 (void) get_process_uid(c
->pid
, &c
->uid
);
187 if (ucred
&& gid_is_valid(ucred
->gid
))
190 (void) get_process_gid(c
->pid
, &c
->gid
);
193 static void client_context_read_basic(ClientContext
*c
) {
197 assert(pid_is_valid(c
->pid
));
199 if (get_process_comm(c
->pid
, &t
) >= 0)
200 free_and_replace(c
->comm
, t
);
202 if (get_process_exe(c
->pid
, &t
) >= 0)
203 free_and_replace(c
->exe
, t
);
205 if (get_process_cmdline(c
->pid
, 0, false, &t
) >= 0)
206 free_and_replace(c
->cmdline
, t
);
208 if (get_process_capeff(c
->pid
, &t
) >= 0)
209 free_and_replace(c
->capeff
, t
);
212 static int client_context_read_label(
214 const char *label
, size_t label_size
) {
217 assert(pid_is_valid(c
->pid
));
218 assert(label_size
== 0 || label
);
220 if (label_size
> 0) {
223 /* If we got an SELinux label passed in it counts. */
225 l
= newdup_suffix0(char, label
, label_size
);
229 free_and_replace(c
->label
, l
);
230 c
->label_size
= label_size
;
236 /* If we got no SELinux label passed in, let's try to acquire one */
238 if (getpidcon(c
->pid
, &con
) >= 0) {
239 free_and_replace(c
->label
, con
);
240 c
->label_size
= strlen(c
->label
);
248 static int client_context_read_cgroup(Server
*s
, ClientContext
*c
, const char *unit_id
) {
254 /* Try to acquire the current cgroup path */
255 r
= cg_pid_get_path_shifted(c
->pid
, s
->cgroup_root
, &t
);
256 if (r
< 0 || empty_or_root(t
)) {
258 /* We use the unit ID passed in as fallback if we have nothing cached yet and cg_pid_get_path_shifted()
259 * failed or process is running in a root cgroup. Zombie processes are automatically migrated to root cgroup
260 * on cgroupsv1 and we want to be able to map log messages from them too. */
261 if (unit_id
&& !c
->unit
) {
262 c
->unit
= strdup(unit_id
);
270 /* Let's shortcut this if the cgroup path didn't change */
271 if (streq_ptr(c
->cgroup
, t
)) {
276 free_and_replace(c
->cgroup
, t
);
278 (void) cg_path_get_session(c
->cgroup
, &t
);
279 free_and_replace(c
->session
, t
);
281 if (cg_path_get_owner_uid(c
->cgroup
, &c
->owner_uid
) < 0)
282 c
->owner_uid
= UID_INVALID
;
284 (void) cg_path_get_unit(c
->cgroup
, &t
);
285 free_and_replace(c
->unit
, t
);
287 (void) cg_path_get_user_unit(c
->cgroup
, &t
);
288 free_and_replace(c
->user_unit
, t
);
290 (void) cg_path_get_slice(c
->cgroup
, &t
);
291 free_and_replace(c
->slice
, t
);
293 (void) cg_path_get_user_slice(c
->cgroup
, &t
);
294 free_and_replace(c
->user_slice
, t
);
299 static int client_context_read_invocation_id(
303 _cleanup_free_
char *value
= NULL
;
310 /* Read the invocation ID of a unit off a unit. PID 1 stores it in a per-unit symlink in /run/systemd/units/ */
315 p
= strjoina("/run/systemd/units/invocation:", c
->unit
);
316 r
= readlink_malloc(p
, &value
);
320 return sd_id128_from_string(value
, &c
->invocation_id
);
323 static int client_context_read_log_level_max(
327 _cleanup_free_
char *value
= NULL
;
334 p
= strjoina("/run/systemd/units/log-level-max:", c
->unit
);
335 r
= readlink_malloc(p
, &value
);
339 ll
= log_level_from_string(value
);
343 c
->log_level_max
= ll
;
347 static int client_context_read_extra_fields(
351 size_t size
= 0, n_iovec
= 0, n_allocated
= 0, left
;
352 _cleanup_free_
struct iovec
*iovec
= NULL
;
353 _cleanup_free_
void *data
= NULL
;
354 _cleanup_fclose_
FILE *f
= NULL
;
363 p
= strjoina("/run/systemd/units/log-extra-fields:", c
->unit
);
365 if (c
->extra_fields_mtime
!= NSEC_INFINITY
) {
366 if (stat(p
, &st
) < 0) {
373 if (timespec_load_nsec(&st
.st_mtim
) == c
->extra_fields_mtime
)
385 if (fstat(fileno(f
), &st
) < 0) /* The file might have been replaced since the stat() above, let's get a new
386 * one, that matches the stuff we are reading */
389 r
= read_full_stream(f
, (char**) &data
, &size
);
393 q
= data
, left
= size
;
398 if (left
< sizeof(uint64_t))
401 v
= unaligned_read_le64(q
);
405 n
= sizeof(uint64_t) + v
;
409 field
= q
+ sizeof(uint64_t);
411 eq
= memchr(field
, '=', v
);
415 if (!journal_field_valid((const char *) field
, eq
- field
, false))
418 if (!GREEDY_REALLOC(iovec
, n_allocated
, n_iovec
+1))
421 iovec
[n_iovec
++] = IOVEC_MAKE(field
, v
);
426 free(c
->extra_fields_iovec
);
427 free(c
->extra_fields_data
);
429 c
->extra_fields_iovec
= TAKE_PTR(iovec
);
430 c
->extra_fields_n_iovec
= n_iovec
;
431 c
->extra_fields_data
= TAKE_PTR(data
);
432 c
->extra_fields_mtime
= timespec_load_nsec(&st
.st_mtim
);
437 static int client_context_read_log_rate_limit_interval(ClientContext
*c
) {
438 _cleanup_free_
char *value
= NULL
;
447 p
= strjoina("/run/systemd/units/log-rate-limit-interval:", c
->unit
);
448 r
= readlink_malloc(p
, &value
);
452 return safe_atou64(value
, &c
->log_rate_limit_interval
);
455 static int client_context_read_log_rate_limit_burst(ClientContext
*c
) {
456 _cleanup_free_
char *value
= NULL
;
465 p
= strjoina("/run/systemd/units/log-rate-limit-burst:", c
->unit
);
466 r
= readlink_malloc(p
, &value
);
470 return safe_atou(value
, &c
->log_rate_limit_burst
);
473 static void client_context_really_refresh(
476 const struct ucred
*ucred
,
477 const char *label
, size_t label_size
,
483 assert(pid_is_valid(c
->pid
));
485 if (timestamp
== USEC_INFINITY
)
486 timestamp
= now(CLOCK_MONOTONIC
);
488 client_context_read_uid_gid(c
, ucred
);
489 client_context_read_basic(c
);
490 (void) client_context_read_label(c
, label
, label_size
);
492 (void) audit_session_from_pid(c
->pid
, &c
->auditid
);
493 (void) audit_loginuid_from_pid(c
->pid
, &c
->loginuid
);
495 (void) client_context_read_cgroup(s
, c
, unit_id
);
496 (void) client_context_read_invocation_id(s
, c
);
497 (void) client_context_read_log_level_max(s
, c
);
498 (void) client_context_read_extra_fields(s
, c
);
499 (void) client_context_read_log_rate_limit_interval(c
);
500 (void) client_context_read_log_rate_limit_burst(c
);
502 c
->timestamp
= timestamp
;
505 assert(c
->n_ref
== 0);
506 assert_se(prioq_reshuffle(s
->client_contexts_lru
, c
, &c
->lru_index
) >= 0);
510 void client_context_maybe_refresh(
513 const struct ucred
*ucred
,
514 const char *label
, size_t label_size
,
521 if (timestamp
== USEC_INFINITY
)
522 timestamp
= now(CLOCK_MONOTONIC
);
524 /* No cached data so far? Let's fill it up */
525 if (c
->timestamp
== USEC_INFINITY
)
528 /* If the data isn't pinned and if the cashed data is older than the upper limit, we flush it out
529 * entirely. This follows the logic that as long as an entry is pinned the PID reuse is unlikely. */
530 if (c
->n_ref
== 0 && c
->timestamp
+ MAX_USEC
< timestamp
) {
531 client_context_reset(s
, c
);
535 /* If the data is older than the lower limit, we refresh, but keep the old data for all we can't update */
536 if (c
->timestamp
+ REFRESH_USEC
< timestamp
)
539 /* If the data passed along doesn't match the cached data we also do a refresh */
540 if (ucred
&& uid_is_valid(ucred
->uid
) && c
->uid
!= ucred
->uid
)
543 if (ucred
&& gid_is_valid(ucred
->gid
) && c
->gid
!= ucred
->gid
)
546 if (label_size
> 0 && (label_size
!= c
->label_size
|| memcmp(label
, c
->label
, label_size
) != 0))
552 client_context_really_refresh(s
, c
, ucred
, label
, label_size
, unit_id
, timestamp
);
555 static void client_context_try_shrink_to(Server
*s
, size_t limit
) {
558 /* Bring the number of cache entries below the indicated limit, so that we can create a new entry without
559 * breaching the limit. Note that we only flush out entries that aren't pinned here. This means the number of
560 * cache entries may very well grow beyond the limit, if all entries stored remain pinned. */
562 while (hashmap_size(s
->client_contexts
) > limit
) {
565 c
= prioq_pop(s
->client_contexts_lru
);
567 break; /* All remaining entries are pinned, give up */
570 assert(c
->n_ref
== 0);
574 client_context_free(s
, c
);
578 void client_context_flush_all(Server
*s
) {
581 /* Flush out all remaining entries. This assumes all references are already dropped. */
583 s
->my_context
= client_context_release(s
, s
->my_context
);
584 s
->pid1_context
= client_context_release(s
, s
->pid1_context
);
586 client_context_try_shrink_to(s
, 0);
588 assert(prioq_size(s
->client_contexts_lru
) == 0);
589 assert(hashmap_size(s
->client_contexts
) == 0);
591 s
->client_contexts_lru
= prioq_free(s
->client_contexts_lru
);
592 s
->client_contexts
= hashmap_free(s
->client_contexts
);
595 static int client_context_get_internal(
598 const struct ucred
*ucred
,
599 const char *label
, size_t label_len
,
602 ClientContext
**ret
) {
610 if (!pid_is_valid(pid
))
613 c
= hashmap_get(s
->client_contexts
, PID_TO_PTR(pid
));
618 /* The entry wasn't pinned so far, let's remove it from the LRU list then */
619 assert(c
->n_ref
== 0);
620 assert_se(prioq_remove(s
->client_contexts_lru
, c
, &c
->lru_index
) >= 0);
627 client_context_maybe_refresh(s
, c
, ucred
, label
, label_len
, unit_id
, USEC_INFINITY
);
633 client_context_try_shrink_to(s
, CACHE_MAX
-1);
635 r
= client_context_new(s
, pid
, &c
);
642 r
= prioq_put(s
->client_contexts_lru
, c
, &c
->lru_index
);
644 client_context_free(s
, c
);
651 client_context_really_refresh(s
, c
, ucred
, label
, label_len
, unit_id
, USEC_INFINITY
);
657 int client_context_get(
660 const struct ucred
*ucred
,
661 const char *label
, size_t label_len
,
663 ClientContext
**ret
) {
665 return client_context_get_internal(s
, pid
, ucred
, label
, label_len
, unit_id
, false, ret
);
668 int client_context_acquire(
671 const struct ucred
*ucred
,
672 const char *label
, size_t label_len
,
674 ClientContext
**ret
) {
676 return client_context_get_internal(s
, pid
, ucred
, label
, label_len
, unit_id
, true, ret
);
679 ClientContext
*client_context_release(Server
*s
, ClientContext
*c
) {
685 assert(c
->n_ref
> 0);
692 /* The entry is not pinned anymore, let's add it to the LRU prioq if we can. If we can't we'll drop it
695 if (prioq_put(s
->client_contexts_lru
, c
, &c
->lru_index
) < 0)
696 client_context_free(s
, c
);
703 void client_context_acquire_default(Server
*s
) {
708 /* Ensure that our own and PID1's contexts are always pinned. Our own context is particularly useful to
709 * generate driver messages. */
711 if (!s
->my_context
) {
712 struct ucred ucred
= {
713 .pid
= getpid_cached(),
718 r
= client_context_acquire(s
, ucred
.pid
, &ucred
, NULL
, 0, NULL
, &s
->my_context
);
720 log_warning_errno(r
, "Failed to acquire our own context, ignoring: %m");
723 if (!s
->pid1_context
) {
725 r
= client_context_acquire(s
, 1, NULL
, NULL
, 0, NULL
, &s
->pid1_context
);
727 log_warning_errno(r
, "Failed to acquire PID1's context, ignoring: %m");