@cryptotaxi247 / netdata-1 / commits / b5e92cb48

Protect v2 journal populate walk from SIGBUS (#22514)

* fix(dbengine): protect v2 journal populate walk from SIGBUS journalfile_v2_populate_retention_to_mrg() walked the mmap'd v2 journal in two phases: an optional CRC check (when JOURNALFILE_FLAG_METRIC_CRC_CHECK is set -- the default cheap-load path on every startup) followed by the mrg update. Only the mrg update ran inside PROTECTED_ACCESS_SETUP. If a backing page could not be paged in (file truncated, sparse hole, transient I/O error) the CRC walk hit SIGBUS and the process aborted with no recovery handler armed. Move PROTECTED_ACCESS_SETUP above the CRC check so one region covers both phases. On signal recovery, clear JOURNALFILE_FLAG_IS_AVAILABLE so the next pass rebuilds the journal -- same effect as a CRC failure. Drop the bare early-return on CRC failure, which leaked the acquire refcount, and fall through to the single journalfile_v2_data_release at the end. Mirrors the pattern in journalfile_v2_validate() at the load site. * fix(dbengine): ensure clean teardown of v2 journal on failure Add immediate unmap of fd/mmap and cleanup of njfv2idx entry when v2 journals fail during critical paths. Prevents leaving journals marked unavailable but partially initialized, which could lead to fd/mmap overwrite and duplicate index entries during rebuilds. * fix(dbengine): validate v2 journal header offsets to prevent over-read and ensure clean failure handling Add bounds checks for header-controlled offsets against the mmap size to prevent corruption-driven over-reads. Centralize teardown logic by clearing IS_AVAILABLE under a spinlock and unmapping permanently, avoiding dangling index entries or fd/mmap leaks during rebuild. * fix(dbengine): improve v2 header validation to prevent underflow and size wrapping Refine bounds checks for metric offsets and metric count in v2 journal headers. Avoid underflow and prevent size_t wrapping on 32-bit builds by reordering checks and using division instead of multiplication for capacity validation. * fix(dbengine): enhance v2 journal validation and clarify teardown logic Add stricter checks for trailer offsets to detect malformed v2 journal headers and prevent corrupted metric lists. Adjust header-controlled size bounds in CRC checks for future-proofing. Refactor IS_AVAILABLE clearing and refcount release sequence for robust teardown, avoiding dangling mappings or index entries. * protect(dbengine): safeguard v2 journal retention updates from SIGBUS Add PROTECTED_ACCESS_SETUP to safely handle mmap walks during metric retention updates. Ensure clean failure handling and resource release in case of journal access errors. * fix(dbengine): mark volatile variables in v2 journal access to ensure safety after recovery Declare uuid_first_entry_list, count, and added as volatile to maintain well-defined behavior across setjmp/longjmp in PROTECTED_ACCESS_SETUP regions. * protect(dbengine): safeguard v2 journal size statistics walk from SIGBUS Add PROTECTED_ACCESS_SETUP to handle mmap walks during the collection of size statistics. Ensure clean failure handling and resource release in case of journal access errors. * fix(dbengine): initialize `has_references` to false in v2 journal setup Ensure `has_references` is explicitly set to false during v2 journal initialization for consistent reference tracking. * protect(dbengine): validate v2 journal bounds during size stats walk Add detailed bounds checks for extent and metric lists in v2 journals to prevent over-reads and ensure safe traversal. Enhance failure handling with consistent resource release on invalid mappings. * fix(dbengine): use uint8_t* for v2 journal byte arithmetic populate_v2_statistics() declared data_start as void* and walked the mmap'd v2 journal with `(void *)(data_start + offset)` pointer arithmetic. Pointer arithmetic on void* is a GCC extension, undefined under standards-strict semantics, and CodeQL's cpp/suspicious-pointer-scaling-void rule flags every such site (alerts 3315, 3317-3320 on this PR). Declaring data_start as uint8_t* makes the arithmetic well-defined (1 byte per step, exactly what the GCC void* extension produced) and silences the 5 alerts at no semantic cost. The existing (void *) casts on the rvalues compile unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(dbengine): mark `journal_access_failed` as volatile and extend bounds checks Declare `journal_access_failed` as volatile to ensure safety in recovery scenarios and prevent future invariance breaks. Add checks to avoid overflow in uuid entry list size computations during v2 journal setup. * protect(dbengine): isolate UUID stack copies to prevent SIGBUS during mmap access Safeguard `journalfile` and `rrdengine` operations by copying UUIDs to the stack before invoking mrg functions. Ensures clean recovery in protected regions on unreadable backing pages and avoids lock state corruption during SIGBUS errors. * fix(dbengine): improve v2 header validation and clarify static analysis intent Enhance error logging for out-of-range v2 journal header offsets, enabling clearer diagnostics during rebuild scenarios. Add a comment clarifying tautological bounds for `descr->type` to assist static analyzers. * protect(dbengine): scope protected frame tightly in v2 journal retention updates Reduce nesting depth and mask of unrelated faults by tightly scoping the PROTECTED_ACCESS_SETUP frame to mmap walks only. Ensure proper cleanup and avoid inflated nesting in subsequent operations. --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Stelios Fragkakis committed Jun 9, 2026 at 13:20 UTC b5e92cb4862c5034c57e84888294d14f71b70f56

3 files changed +301 -90

src/database/engine/journalfile.c

+112 -34

@@ -497,6 +497,7 @@ static void journalfile_v2_data_unmap_permanently(struct rrdengine_journalfile *

497 journalfile->v2.first_time_s = 0;

498 journalfile->v2.last_time_s = 0;

499 journalfile->v2.flags = 0;

500 + has_references = false;

501 }

502 else {

503 has_references = true;

@@ -1052,52 +1053,129 @@ void journalfile_v2_populate_retention_to_mrg(struct rrdengine_instance *ctx, st

1053

1054 uint8_t *data_start = (uint8_t *)j2_header;

1055

1055 - if (journalfile->v2.flags & JOURNALFILE_FLAG_METRIC_CRC_CHECK) {

1056 - journalfile->v2.flags &= ~JOURNALFILE_FLAG_METRIC_CRC_CHECK;

1057 - if (journalfile_check_v2_metric_list(data_start, j2_header->journal_v2_file_size)) {

1058 - journalfile->v2.flags &= ~JOURNALFILE_FLAG_IS_AVAILABLE;

1059 - // needs rebuild

1060 - return;

1061 - }

1062 - }

1063 -

1056 char path_v2[RRDENG_PATH_MAX];

1057 journalfile_v2_generate_path(journalfile->datafile, path_v2, sizeof(path_v2));

1066 - time_t global_first_time_s;

1058 + time_t global_first_time_s = 0;

1059 bool failed = false;

1068 - uint32_t entries;

1060 + uint32_t entries = 0;

1061 // Calculate number of samples here and update once the file is loaded

1062 uint64_t journal_samples = 0;

1071 - PROTECTED_ACCESS_SETUP(data_start, journalfile->mmap.size, path_v2, "mrg-load");

1063 +

1064 + // Protect the whole walk -- both the optional CRC check and the mrg update

1065 + // read into the mmap'd v2 journal. If a backing page cannot be paged in

1066 + // (file truncated, sparse hole, transient I/O error) the kernel raises

1067 + // SIGBUS; without a protected region active the process aborts. Same

1068 + // pattern as journalfile_v2_validate() in journalfile_v2_load().

1069 + PROTECTED_ACCESS_SETUP(journalfile->mmap.data, journalfile->mmap.size, path_v2, "mrg-load");

1070 if(no_signal_received) {

1073 - entries = j2_header->metric_count;

1074 - struct journal_metric_list *metric = (struct journal_metric_list *) (data_start + j2_header->metric_offset);

1075 - time_t header_start_time_s = (time_t) (j2_header->start_time_ut / USEC_PER_SEC);

1076 - global_first_time_s = header_start_time_s;

1077 - time_t now_s = max_acceptable_collected_time();

1078 - for (size_t i=0; i < entries; i++) {

1079 - time_t start_time_s = header_start_time_s + metric->delta_start_s;

1080 - time_t end_time_s = header_start_time_s + metric->delta_end_s;

1081 -

1082 - mrg_update_metric_retention_and_granularity_by_uuid(

1083 - main_mrg,

1084 - (Word_t)ctx,

1085 - &metric->uuid,

1086 - start_time_s,

1087 - end_time_s,

1088 - metric->update_every_s,

1089 - now_s,

1090 - &journal_samples);

1091 -

1092 - metric++;

1071 + // Validate header-controlled offsets against the actual mapping size

1072 + // BEFORE reading through them. PROTECTED_ACCESS_SETUP only catches

1073 + // faults whose address falls within [data_start, data_start+mmap.size);

1074 + // an over-read driven by a corrupted offset past the registered range

1075 + // would not be recovered and the process would still abort. Out-of-

1076 + // range offsets are treated as a rebuild trigger (same outcome as a

1077 + // CRC failure).

1078 + size_t mmap_size = journalfile->mmap.size;

1079 + // Order matters: short-circuit on metric_offset > mmap_size first so the

1080 + // subtraction below cannot underflow, then compare metric_count against

1081 + // the available slot capacity via division rather than multiplying out

1082 + // metric_count * sizeof(...), which would wrap size_t on 32-bit builds

1083 + // and silently pass a malformed header. Only after those two checks is

1084 + // metric_count * sizeof(...) known to fit in size_t, which lets the

1085 + // trailer-ordering check below compute the expected trailer offset

1086 + // safely. The on-disk layout (see journalfile_migrate_to_v2_callback)

1087 + // places the metric-list trailer immediately after the metric list, so

1088 + // any deviation indicates a corrupted header that would otherwise let

1089 + // the CRC compare against bytes inside the metric list itself.

1090 + if ((size_t)j2_header->metric_offset > mmap_size ||

1091 + (size_t)j2_header->metric_count > (mmap_size - (size_t)j2_header->metric_offset) / sizeof(struct journal_metric_list) ||

1092 + (size_t)j2_header->metric_trailer_offset > mmap_size ||

1093 + mmap_size - (size_t)j2_header->metric_trailer_offset < sizeof(struct journal_v2_block_trailer) ||

1094 + (size_t)j2_header->metric_trailer_offset != (size_t)j2_header->metric_offset + (size_t)j2_header->metric_count * sizeof(struct journal_metric_list)) {

1095 + // header offsets out of range -- needs rebuild

1096 + nd_log_daemon(NDLP_ERR,

1097 + "DBENGINE: journal v2 \"%s\" has out-of-range header offsets "

1098 + "(metric_offset=%u, metric_count=%u, metric_trailer_offset=%u, mmap_size=%zu); "

1099 + "marking unavailable for rebuild",

1100 + path_v2,

1101 + j2_header->metric_offset,

1102 + j2_header->metric_count,

1103 + j2_header->metric_trailer_offset,

1104 + mmap_size);

1105 + failed = true;

1106 + }

1107 + else if (journalfile->v2.flags & JOURNALFILE_FLAG_METRIC_CRC_CHECK) {

1108 + journalfile->v2.flags &= ~JOURNALFILE_FLAG_METRIC_CRC_CHECK;

1109 + // Pass the verified mmap_size, not the header-controlled

1110 + // j2_header->journal_v2_file_size; the helper currently ignores the

1111 + // size argument (UNUSED) but the value at the call site should

1112 + // still reflect the trusted bound for clarity and future-proofing.

1113 + if (journalfile_check_v2_metric_list(data_start, mmap_size)) {

1114 + // needs rebuild

1115 + failed = true;

1116 + }

1117 + }

1118 +

1119 + if (!failed) {

1120 + entries = j2_header->metric_count;

1121 + struct journal_metric_list *metric = (struct journal_metric_list *) (data_start + j2_header->metric_offset);

1122 + time_t header_start_time_s = (time_t) (j2_header->start_time_ut / USEC_PER_SEC);

1123 + global_first_time_s = header_start_time_s;

1124 + time_t now_s = max_acceptable_collected_time();

1125 + for (size_t i=0; i < entries; i++) {

1126 + // Copy uuid out of the mmap onto the stack BEFORE calling mrg.

1127 + // If a backing page is unreadable, uuid_copy SIGBUSes here and

1128 + // the protected region recovers cleanly; the mrg call then

1129 + // never executes. If we passed &metric->uuid into mrg, a

1130 + // SIGBUS could fire INSIDE mrg while it holds internal locks,

1131 + // and siglongjmp would skip mrg's unlock paths.

1132 + nd_uuid_t local_uuid;

1133 + uuid_copy(local_uuid, metric->uuid);

1134 + time_t start_time_s = header_start_time_s + metric->delta_start_s;

1135 + time_t end_time_s = header_start_time_s + metric->delta_end_s;

1136 + uint32_t update_every_s = metric->update_every_s;

1137 +

1138 + mrg_update_metric_retention_and_granularity_by_uuid(

1139 + main_mrg,

1140 + (Word_t)ctx,

1141 + &local_uuid,

1142 + start_time_s,

1143 + end_time_s,

1144 + update_every_s,

1145 + now_s,

1146 + &journal_samples);

1147 +

1148 + metric++;

1149 + }

1150 }

1094 - } else

1151 + }

1152 + else {

1153 + // SIGBUS/SIGSEGV inside the mmap walk. The PROTECTED_ACCESS_SETUP

1154 + // macro already rate-limits the error log.

1155 failed = true;

1156 + }

1157 +

1158 + if (unlikely(failed)) {

1159 + // Clear IS_AVAILABLE under the data spinlock BEFORE releasing our

1160 + // refcount, so no concurrent journalfile_v2_data_acquire_with_hint()

1161 + // can succeed and walk the (potentially corrupted) mmap between here

1162 + // and the permanent unmap. The bounds, CRC, and signal-recovery paths

1163 + // all converge through this single transition; without it, teardown

1164 + // (journalfile_close / journalfile_destroy_unsafe) would gate on

1165 + // IS_AVAILABLE and skip journalfile_v2_data_unmap_permanently(),

1166 + // leaving a dangling njfv2idx entry, an unclosed fd, and an unmapped

1167 + // region.

1168 + spinlock_lock(&journalfile->data_spinlock);

1169 + journalfile->v2.flags &= ~JOURNALFILE_FLAG_IS_AVAILABLE;

1170 + spinlock_unlock(&journalfile->data_spinlock);

1171 + }

1172

1173 journalfile_v2_data_release(journalfile);

1174

1099 - if (unlikely(failed))

1175 + if (unlikely(failed)) {

1176 + journalfile_v2_data_unmap_permanently(journalfile);

1177 return;

1178 + }

1179

1180 __atomic_add_fetch(&ctx->atomic.samples, journal_samples, __ATOMIC_RELAXED);

1181

src/database/engine/rrdengine.c

+95 -15

@@ -1387,24 +1387,96 @@ static void update_metrics_first_time_s(struct rrdengine_instance *ctx, struct r

1387

1388 __atomic_add_fetch(&rrdeng_cache_efficiency_stats.metrics_retention_started, 1, __ATOMIC_RELAXED);

1389

1390 - struct journal_metric_list *uuid_list = (struct journal_metric_list *)((uint8_t *) j2_header + j2_header->metric_offset);

1390 + char file_path[RRDENG_PATH_MAX];

1391 + journalfile_v2_generate_path(datafile_to_delete, file_path, sizeof(file_path));

1392

1392 - size_t count = j2_header->metric_count;

1393 struct uuid_first_time_s *uuid_first_t_entry;

1394 - struct uuid_first_time_s *uuid_first_entry_list = callocz(count, sizeof(struct uuid_first_time_s));

1395 -

1396 - size_t added = 0;

1397 - for (size_t index = 0; index < count; ++index) {

1398 - METRIC *metric = mrg_metric_get_and_acquire_by_uuid(main_mrg, &uuid_list[index].uuid, (Word_t)ctx);

1399 - if (!metric)

1400 - continue;

1394 + // PROTECTED_ACCESS_SETUP below uses sigsetjmp/siglongjmp (see

1395 + // src/daemon/protected-access.h). Per C11 7.13.2.1, non-volatile locals

1396 + // that are modified between setjmp and longjmp have indeterminate values

1397 + // on the recovery path. uuid_first_entry_list / count / added are all

1398 + // mutated inside the protected region and then read afterwards (the

1399 + // unconditional log line and the journal_access_failed cleanup loop), so

1400 + // they must be volatile to keep the recovery path well-defined.

1401 + // journal_access_failed is also marked volatile defensively: although it

1402 + // is only assigned on a path that does not subsequently SIGBUS, future

1403 + // edits could break that invariant, and the bool is read once on cleanup.

1404 + struct uuid_first_time_s * volatile uuid_first_entry_list = NULL;

1405 + volatile size_t count = 0;

1406 + volatile size_t added = 0;

1407 + volatile bool journal_access_failed = false;

1408 +

1409 + // Protect the mmap walk: reading j2_header->metric_offset, metric_count,

1410 + // and the per-metric uuid_list[] entries can SIGBUS if the underlying v2

1411 + // file has any unreadable page (truncated, sparse hole, transient I/O

1412 + // error). Without a protected region active the process aborts. Same

1413 + // pattern as find_uuid_first_time() added by commit 26b26ac25a (#22310);

1414 + // this caller was missed at the time.

1415 + // Scope the protected frame tightly to the mmap walk only. The

1416 + // PROTECTED_ACCESS_AUTO_CLEANUP() guard inside PROTECTED_ACCESS_SETUP

1417 + // declares a __attribute__((cleanup)) local; when this inner block exits

1418 + // (normally or via the SIGBUS recovery else-branch falling through), the

1419 + // cleanup runs and the protected-access depth drops back to its prior

1420 + // value. Without this scoping, the frame would stay live through the

1421 + // post-walk log, the data_release, the cleanup loop, the nested

1422 + // find_uuid_first_time() (which registers its own frame), and the final

1423 + // cleanup -- masking unrelated faults that might land in the mmap range

1424 + // and inflating nesting depth unnecessarily.

1425 + {

1426 + PROTECTED_ACCESS_SETUP(journalfile->mmap.data, journalfile->mmap.size, file_path, "mrg-retention");

1427 + if(no_signal_received) {

1428 + size_t journal_v2_file_size = journalfile->mmap.size;

1429 + size_t metric_offset = j2_header->metric_offset;

1430 + count = j2_header->metric_count;

1431 + size_t metric_list_size;

1432 + size_t entry_list_size;

1433 + // Also check count * sizeof(uuid_first_time_s) -- the allocation below

1434 + // sizes the working array by count, and on 32-bit builds count can

1435 + // pass the metric_list_size bound while still overflowing the entry

1436 + // list multiplication (struct uuid_first_time_s is larger than

1437 + // struct journal_metric_list).

1438 + if (__builtin_mul_overflow(count, sizeof(struct journal_metric_list), &metric_list_size) ||

1439 + __builtin_mul_overflow(count, sizeof(struct uuid_first_time_s), &entry_list_size) ||

1440 + metric_offset > journal_v2_file_size ||

1441 + metric_list_size > journal_v2_file_size - metric_offset) {

1442 + nd_log_daemon(NDLP_ERR,

1443 + "DBENGINE: metric list exceeds journal file size in journalfile \"%s\" "

1444 + "(metric_offset=%zu, list_size=%zu, file_size=%zu), skipping retention update",

1445 + file_path, metric_offset, metric_list_size, journal_v2_file_size);

1446 + journal_access_failed = true;

1447 + }

1448 + else {

1449 + struct journal_metric_list *uuid_list = (struct journal_metric_list *)((uint8_t *) j2_header + metric_offset);

1450 + uuid_first_entry_list = callocz(count, sizeof(struct uuid_first_time_s));

1451 +

1452 + for (size_t index = 0; index < count; ++index) {

1453 + // Copy uuid out of the mmap onto the stack BEFORE calling mrg.

1454 + // If a backing page is unreadable, uuid_copy SIGBUSes here and

1455 + // the protected region recovers cleanly; the mrg call then

1456 + // never executes. If we passed &uuid_list[index].uuid into

1457 + // mrg, a SIGBUS could fire INSIDE mrg while it holds internal

1458 + // locks, and siglongjmp would skip mrg's unlock paths.

1459 + nd_uuid_t local_uuid;

1460 + uuid_copy(local_uuid, uuid_list[index].uuid);

1461 +

1462 + METRIC *metric = mrg_metric_get_and_acquire_by_uuid(main_mrg, &local_uuid, (Word_t)ctx);

1463 + if (!metric)

1464 + continue;

1465

1402 - uuid_first_entry_list[added].metric = metric;

1403 - uuid_first_entry_list[added].first_time_s = LONG_MAX;

1404 - uuid_first_entry_list[added].df_matched = 0;

1405 - uuid_first_entry_list[added].df_index_oldest = 0;

1406 - uuid_first_entry_list[added].uuid = mrg_metric_uuid(main_mrg, metric);

1407 - added++;

1466 + uuid_first_entry_list[added].metric = metric;

1467 + uuid_first_entry_list[added].first_time_s = LONG_MAX;

1468 + uuid_first_entry_list[added].df_matched = 0;

1469 + uuid_first_entry_list[added].df_index_oldest = 0;

1470 + uuid_first_entry_list[added].uuid = mrg_metric_uuid(main_mrg, metric);

1471 + added++;

1472 + }

1473 + }

1474 + }

1475 + else {

1476 + // SIGBUS/SIGSEGV inside the mmap walk -- bail cleanly. The

1477 + // PROTECTED_ACCESS_SETUP macro already rate-limits the error log.

1478 + journal_access_failed = true;

1479 + }

1480 }

1481

1482 netdata_log_info(

@@ -1415,6 +1487,14 @@ static void update_metrics_first_time_s(struct rrdengine_instance *ctx, struct r

1487

1488 journalfile_v2_data_release(journalfile);

1489

1490 + if (unlikely(journal_access_failed)) {

1491 + // Release any partially-acquired metrics; uuid_first_entry_list may

1492 + // be NULL (signal received before callocz).

1493 + for (size_t index = 0; index < added; ++index)

1494 + mrg_metric_release(main_mrg, uuid_first_entry_list[index].metric);

1495 + goto done;

1496 + }

1497 +

1498 // Update the first time / last time for all metrics we plan to delete

1499

1500 if(worker)

src/database/engine/rrdengineapi.c

+94 -41

@@ -1293,68 +1293,121 @@ void rrdeng_quiesce(struct rrdengine_instance *ctx)

1293 static void populate_v2_statistics(struct rrdengine_datafile *datafile, RRDENG_SIZE_STATS *stats)

1294 {

1295 struct journal_v2_header *j2_header = journalfile_v2_data_acquire(datafile->journalfile, NULL, 0, 0);

1296 - void *data_start = (void *)j2_header;

1296 + uint8_t *data_start = (uint8_t *)j2_header;

1297

1298 if(unlikely(!j2_header))

1299 return;

1300

1301 - stats->extents += j2_header->extent_count;

1301 + char file_path[RRDENG_PATH_MAX];

1302 + journalfile_v2_generate_path(datafile, file_path, sizeof(file_path));

1303 +

1304 + // Protect the mmap walk: every j2_header->*, extent_list, metric, and

1305 + // descr access reads the mmap'd v2 journal. If the underlying file has

1306 + // any unreadable page (truncated, sparse hole, transient I/O error), the

1307 + // walk SIGBUSes and the process aborts. Same pattern as the sister sites

1308 + // (populate_retention_to_mrg, find_uuid_first_time, update_metrics_first_time_s).

1309 + // Partial accumulator increments on signal recovery are acceptable:

1310 + // size statistics are best-effort across datafiles, the caller continues

1311 + // to the next datafile in rrdeng_size_statistics().

1312 + PROTECTED_ACCESS_SETUP(datafile->journalfile->mmap.data, datafile->journalfile->mmap.size, file_path, "size-stats");

1313 + if(no_signal_received) {

1314 + size_t mmap_size = datafile->journalfile->mmap.size;

1315 +

1316 + // Bounds-check the extent list array against the mapping size before

1317 + // walking through header-controlled offsets. PROTECTED_ACCESS_SETUP

1318 + // only catches faults in [data_start, data_start+mmap_size); a

1319 + // corrupted offset that points past mmap_size would over-read into

1320 + // unrelated memory and abort. Same idiom as the bounds check in

1321 + // journalfile_v2_populate_retention_to_mrg.

1322 + bool extents_in_bounds = (size_t)j2_header->extent_offset <= mmap_size &&

1323 + (size_t)j2_header->extent_count <= (mmap_size - (size_t)j2_header->extent_offset) / sizeof(struct journal_extent_list);

1324 + if (extents_in_bounds) {

1325 + stats->extents += j2_header->extent_count;

1326 +

1327 + struct journal_extent_list *extent_list = (void *) (data_start + j2_header->extent_offset);

1328 + for (unsigned entries = 0; entries < j2_header->extent_count; entries++) {

1329 + stats->extents_compressed_bytes += extent_list->datafile_size;

1330 + stats->extents_pages += extent_list->pages;

1331 + extent_list++;

1332 + }

1333 + }

1334

1303 - unsigned entries;

1304 - struct journal_extent_list *extent_list = (void *) (data_start + j2_header->extent_offset);

1305 - for (entries = 0; entries < j2_header->extent_count; entries++) {

1306 - stats->extents_compressed_bytes += extent_list->datafile_size;

1307 - stats->extents_pages += extent_list->pages;

1308 - extent_list++;

1309 - }

1335 + bool metrics_in_bounds = (size_t)j2_header->metric_offset <= mmap_size &&

1336 + (size_t)j2_header->metric_count <= (mmap_size - (size_t)j2_header->metric_offset) / sizeof(struct journal_metric_list);

1337 + if (!metrics_in_bounds)

1338 + goto release;

1339

1311 - struct journal_metric_list *metric = (void *) (data_start + j2_header->metric_offset);

1312 - time_t journal_start_time_s = (time_t) (j2_header->start_time_ut / USEC_PER_SEC);

1340 + struct journal_metric_list *metric = (void *) (data_start + j2_header->metric_offset);

1341 + time_t journal_start_time_s = (time_t) (j2_header->start_time_ut / USEC_PER_SEC);

1342

1314 - stats->metrics += j2_header->metric_count;

1315 - for (entries = 0; entries < j2_header->metric_count; entries++) {

1343 + stats->metrics += j2_header->metric_count;

1344 + for (unsigned entries = 0; entries < j2_header->metric_count; entries++) {

1345

1317 - struct journal_page_header *metric_list_header = (void *) (data_start + metric->page_offset);

1318 - stats->metrics_pages += metric_list_header->entries;

1319 - struct journal_page_list *descr = (void *) (data_start + metric->page_offset + sizeof(struct journal_page_header));

1320 - for (uint32_t idx=0; idx < metric_list_header->entries; idx++) {

1346 + // Per-metric: page_offset is header-controlled. Validate it points

1347 + // into the mapping and that there is room for the page_header AND

1348 + // its trailing page_list[] before dereferencing through it.

1349 + if ((size_t)metric->page_offset > mmap_size ||

1350 + mmap_size - (size_t)metric->page_offset < sizeof(struct journal_page_header)) {

1351 + metric++;

1352 + continue;

1353 + }

1354

1322 - time_t update_every_s;

1355 + struct journal_page_header *metric_list_header = (void *) (data_start + metric->page_offset);

1356

1324 - size_t points = descr->page_length / CTX_POINT_SIZE_BYTES(datafile_ctx(datafile));

1357 + size_t page_list_room = mmap_size - (size_t)metric->page_offset - sizeof(struct journal_page_header);

1358 + if ((size_t)metric_list_header->entries > page_list_room / sizeof(struct journal_page_list)) {

1359 + metric++;

1360 + continue;

1361 + }

1362

1326 - time_t start_time_s = journal_start_time_s + descr->delta_start_s;

1327 - time_t end_time_s = journal_start_time_s + descr->delta_end_s;

1363 + stats->metrics_pages += metric_list_header->entries;

1364 + struct journal_page_list *descr = (void *) (data_start + metric->page_offset + sizeof(struct journal_page_header));

1365 + for (uint32_t idx=0; idx < metric_list_header->entries; idx++) {

1366

1329 - if(likely(points > 1))

1330 - update_every_s = (time_t) ((end_time_s - start_time_s) / (points - 1));

1331 - else {

1332 - update_every_s = (time_t) (nd_profile.update_every * get_tier_grouping(datafile_ctx(datafile)->config.tier));

1333 - stats->single_point_pages++;

1334 - }

1367 + time_t update_every_s;

1368

1336 - time_t duration_s = (time_t)((end_time_s - start_time_s + update_every_s));

1369 + size_t points = descr->page_length / CTX_POINT_SIZE_BYTES(datafile_ctx(datafile));

1370

1338 - stats->pages_uncompressed_bytes += descr->page_length;

1339 - stats->pages_duration_secs += duration_s;

1340 - stats->points += points;

1371 + time_t start_time_s = journal_start_time_s + descr->delta_start_s;

1372 + time_t end_time_s = journal_start_time_s + descr->delta_end_s;

1373

1342 - stats->page_types[descr->type].pages++;

1343 - stats->page_types[descr->type].pages_uncompressed_bytes += descr->page_length;

1344 - stats->page_types[descr->type].pages_duration_secs += duration_s;

1345 - stats->page_types[descr->type].points += points;

1374 + if(likely(points > 1))

1375 + update_every_s = (time_t) ((end_time_s - start_time_s) / (points - 1));

1376 + else {

1377 + update_every_s = (time_t) (nd_profile.update_every * get_tier_grouping(datafile_ctx(datafile)->config.tier));

1378 + stats->single_point_pages++;

1379 + }

1380 +

1381 + time_t duration_s = (time_t)((end_time_s - start_time_s + update_every_s));

1382

1347 - if(!stats->first_time_s || (start_time_s - update_every_s) < stats->first_time_s)

1348 - stats->first_time_s = (start_time_s - update_every_s);

1383 + stats->pages_uncompressed_bytes += descr->page_length;

1384 + stats->pages_duration_secs += duration_s;

1385 + stats->points += points;

1386

1350 - if(!stats->last_time_s || end_time_s > stats->last_time_s)

1351 - stats->last_time_s = end_time_s;

1387 + // descr->type is uint8_t (range [0, 255]); page_types is sized

1388 + // [256]. The index is bounded by the type width, so no runtime

1389 + // check is needed -- a `descr->type < 256` guard would be a

1390 + // tautology. Note for static analyzers flagging this site.

1391 + stats->page_types[descr->type].pages++;

1392 + stats->page_types[descr->type].pages_uncompressed_bytes += descr->page_length;

1393 + stats->page_types[descr->type].pages_duration_secs += duration_s;

1394 + stats->page_types[descr->type].points += points;

1395

1353 - descr++;

1396 + if(!stats->first_time_s || (start_time_s - update_every_s) < stats->first_time_s)

1397 + stats->first_time_s = (start_time_s - update_every_s);

1398 +

1399 + if(!stats->last_time_s || end_time_s > stats->last_time_s)

1400 + stats->last_time_s = end_time_s;

1401 +

1402 + descr++;

1403 + }

1404 + metric++;

1405 }

1355 - metric++;

1406 }

1407 + // On SIGBUS/SIGSEGV the PROTECTED_ACCESS_SETUP macro already

1408 + // rate-limits the error log; fall through to release the journal.

1409

1410 +release:

1411 journalfile_v2_data_release(datafile->journalfile);

1412 }

1413