Merge pull request #4624 from chu11/issue4612_broker_crash

job-info: return error on invalid eventlog append

Author: mergify[bot]

src/modules/job-info/watch.c

@@ -246,13 +246,24 @@ static void watch_continuation (flux_future_t *f, void *arg)
     const char *input;
     const char *tok;
     size_t toklen;
+    const char *errmsg = NULL;
 
     if (flux_kvs_lookup_get (f, &s) < 0) {
         if (errno != ENOENT && errno != ENODATA && errno != ENOTSUP)
             flux_log_error (ctx->h, "%s: flux_kvs_lookup_get", __FUNCTION__);
         goto error;
     }
 
+    /* Issue #4612 - zero length append illegal for an eventlog.  This
+     * most likely occurred through an illegal overwrite of the whole
+     * eventlog.
+     */
+    if (!s) {
+        errmsg = "illegal append of zero bytes";
+        errno = EINVAL;
+        goto error;
+    }
+
     if (w->cancel) {
         errno = ENODATA;
         goto error;
@@ -310,7 +321,7 @@ static void watch_continuation (flux_future_t *f, void *arg)
     }
 
 error:
-    if (flux_respond_error (ctx->h, w->msg, errno, NULL) < 0)
+    if (flux_respond_error (ctx->h, w->msg, errno, errmsg) < 0)
         flux_log_error (ctx->h, "%s: flux_respond_error", __FUNCTION__);
 
     /* flux future destroyed in watch_ctx_destroy, which is called

t/Makefile.am

@@ -312,6 +312,7 @@ dist_check_SCRIPTS = \
 	issues/t4413-empty-eventlog.sh \
 	issues/t4465-job-list-use-after-free.sh \
 	issues/t4482-flush-list-corruption.sh \
+	issues/t4612-eventlog-overwrite-crash.sh \
 	python/__init__.py \
 	python/subflux.py \
 	python/tap \

t/issues/t4612-eventlog-overwrite-crash.sh

@@ -0,0 +1,23 @@
+#!/bin/bash -e
+
+waitfile=${SHARNESS_TEST_SRCDIR}/scripts/waitfile.lua
+
+jobid=$(flux mini submit --wait-event=start sh -c "echo foo; sleep 300")
+
+kvsdir=$(flux job id --to=kvs $jobid)
+
+# issue a command that will watch / monitor the job's eventlog
+flux job attach $jobid > t4612.out &
+
+# ensure backgrounded process has started to monitor eventlog
+$waitfile --count=1 --timeout=30 --pattern=foo t4612.out
+
+# now overwrite the eventlog without changing its length
+flux kvs get --raw ${kvsdir}.eventlog \
+	| sed -e s/submit/foobar/ \
+	| flux kvs put --raw ${kvsdir}.eventlog=-
+
+wait
+
+# if flux broker segfaulted, this won't work
+flux mini run hostname