Merge pull request #5254 from garlick/issue#5245

Ignore errors when starting flux from a restart dump containing giant blobs

Author: mergify[bot]

doc/man1/flux-restore.rst

@@ -53,6 +53,8 @@ OPTIONS
    Bypass the broker content cache and interact directly with the backing
    store.  Performance will vary depending on the content of the archive.
 
+**--size-limit**\ =\ *SIZE*
+   Skip restoring keys that exceed SIZE bytes (default: no limit).
 
 RESOURCES
 =========

etc/rc1

@@ -47,7 +47,7 @@ if test $RANK -eq 0; then
             flux module load ${backingmod} truncate
         fi
         echo "restoring content from ${dumpfile}"
-        flux restore --quiet --checkpoint ${dumpfile}
+        flux restore --quiet --checkpoint --size-limit=104857600 ${dumpfile}
         if test -n "${dumplink}"; then
             rm -f ${dumplink}
         fi

src/cmd/builtin/restore.c

@@ -37,6 +37,7 @@ static int content_flags;
 static time_t restore_timestamp;
 static int blobcount;
 static int keycount;
+static int blob_size_limit;
 
 static void progress (int delta_blob, int delta_keys)
 {
@@ -270,6 +271,17 @@ static json_t *restore_snapshot (struct archive *ar, flux_t *h)
         else if (type == AE_IFREG) {
             int size = archive_entry_size (entry);
 
+            if (blob_size_limit > 0 && size > blob_size_limit) {
+                fprintf (stderr,
+                         "%s%s size %d exceeds %d limit, skipping\n",
+                         (!quiet && !verbose) ? "\r" : "",
+                         path,
+                         size,
+                         blob_size_limit);
+                // N.B.  archive_read_next_header() skips unconsumed data
+                //   automatically so it is safe to "continue" here.
+                continue;
+            }
             if (size > bufsize) {
                 void *newbuf;
                 if (!(newbuf = realloc (buf, size)))
@@ -354,6 +366,7 @@ static int cmd_restore (optparse_t *p, int ac, char *av[])
         content_flags |= CONTENT_FLAG_CACHE_BYPASS;
         kvs_checkpoint_flags |= KVS_CHECKPOINT_FLAG_CACHE_BYPASS;
     }
+    blob_size_limit = optparse_get_int (p, "size-limit", 0);
 
     h = builtin_get_flux_handle (p);
     ar = restore_create (infile);
@@ -458,6 +471,9 @@ static struct optparse_option restore_opts[] = {
     { .name = "no-cache", .has_arg = 0,
       .usage = "Bypass the broker content cache",
     },
+    { .name = "size-limit", .has_arg = 1, .arginfo = "SIZE",
+      .usage = "Do not restore blobs greater than SIZE bytes",
+    },
     OPTPARSE_TABLE_END
 };
 

t/t2807-dump-cmd.t

@@ -227,4 +227,34 @@ test_expect_success 'restart flux instance and try to run a job' '
 		flux run /bin/true
 '
 
+# Cover --size-limit
+
+test_expect_success 'create bigdump.tar with a 12M blob in it' '
+	mkdir -p big &&
+	dd if=/dev/zero of=big/tinyblob bs=1048576 count=1 &&
+	dd if=/dev/zero of=big/bigblob bs=1048576 count=12 &&
+	dd if=/dev/zero of=big/smallblob bs=1048576 count=3 &&
+	dd if=/dev/zero of=big/medblob bs=1048576 count=6 &&
+	dd if=/dev/zero of=big/med2blob bs=1048576 count=6 &&
+	tar cvf bigdump.tar big
+'
+test_expect_success 'restore bigdump.tar and verify blob count' '
+	flux start flux restore \
+		--key=foo bigdump.tar 2>bigdump.err &&
+	grep "restored 5 keys (7 blobs)" bigdump.err
+'
+test_expect_success 'restore bigdump.tar with size limit' '
+	flux start flux restore --size-limit=10485760 \
+		--key=foo bigdump.tar 2>bigdump2.err &&
+	grep "exceeds" bigdump2.err &&
+	grep "restored 4 keys (6 blobs)" bigdump2.err
+'
+test_expect_success 'rc1 skips blob that exceeds 100M limit' '
+	dd if=/dev/zero of=big/hugeblob bs=1048576 count=120 &&
+	tar cvf bigdump2.tar big &&
+	flux start -o,-Scontent.restore=bigdump2.tar \
+		/bin/true 2>bigdump3.err &&
+	grep "exceeds" bigdump3.err
+'
+
 test_done