tdb: add run-fcntl-deadlock test
authorStefan Metzmacher <metze@samba.org>
Tue, 11 Apr 2017 15:21:20 +0000 (17:21 +0200)
committerGarming Sam <garming@samba.org>
Wed, 14 Jun 2017 23:24:25 +0000 (01:24 +0200)
This verifies the F_RDLCK => F_WRLCK upgrade logic in the kernel
for conflicting locks.

This is a standalone test to check the traverse_read vs.
allrecord_lock/prepare_commit interaction.

This is based on the example from
https://lists.samba.org/archive/samba-technical/2017-April/119861.html
from Douglas Bagnall <douglas.bagnall@catalyst.net.nz> and Volker Lendecke <vl@samba.org>.

Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Garming Sam <garming@catalyst.net.nz>
lib/tdb/test/run-fcntl-deadlock.c [new file with mode: 0644]
lib/tdb/wscript

diff --git a/lib/tdb/test/run-fcntl-deadlock.c b/lib/tdb/test/run-fcntl-deadlock.c
new file mode 100644 (file)
index 0000000..0a328af
--- /dev/null
@@ -0,0 +1,202 @@
+#include "../common/tdb_private.h"
+#include "../common/io.c"
+#include "../common/tdb.c"
+#include "../common/lock.c"
+#include "../common/freelist.c"
+#include "../common/traverse.c"
+#include "../common/transaction.c"
+#include "../common/error.c"
+#include "../common/open.c"
+#include "../common/check.c"
+#include "../common/hash.c"
+#include "../common/mutex.c"
+#include "replace.h"
+#include "system/filesys.h"
+#include "system/time.h"
+#include <errno.h>
+#include "tap-interface.h"
+
+/*
+ * This tests the low level locking requirement
+ * for the allrecord lock/prepare_commit and traverse_read interaction.
+ *
+ * The pattern with the traverse_read and prepare_commit interaction is
+ * the following:
+ *
+ * 1. transaction_start got the allrecord lock with F_RDLCK.
+ *
+ * 2. the traverse_read code walks the database in a sequence like this
+ * (per chain):
+ *    2.1  chainlock(chainX, F_RDLCK)
+ *    2.2  recordlock(chainX.record1, F_RDLCK)
+ *    2.3  chainunlock(chainX, F_RDLCK)
+ *    2.4  callback(chainX.record1)
+ *    2.5  chainlock(chainX, F_RDLCK)
+ *    2.6  recordunlock(chainX.record1, F_RDLCK)
+ *    2.7  recordlock(chainX.record2, F_RDLCK)
+ *    2.8  chainunlock(chainX, F_RDLCK)
+ *    2.9  callback(chainX.record2)
+ *    2.10 chainlock(chainX, F_RDLCK)
+ *    2.11 recordunlock(chainX.record2, F_RDLCK)
+ *    2.12 chainunlock(chainX, F_RDLCK)
+ *    2.13 goto next chain
+ *
+ *    So it has always one record locked in F_RDLCK mode and tries to
+ *    get the 2nd one before it releases the first one.
+ *
+ * 3. prepare_commit tries to upgrade the allrecord lock to F_RWLCK
+ *    If that happens at the time of 2.4, the operation of
+ *    2.5 may deadlock with the allrecord lock upgrade.
+ *    On Linux step 2.5 works in order to make some progress with the
+ *    locking, but on solaris it might fail because the kernel
+ *    wants to satisfy the 1st lock requester before the 2nd one.
+ *
+ * I think the first step is a standalone test that does this:
+ *
+ * process1: F_RDLCK for ofs=0 len=2
+ * process2: F_RDLCK for ofs=0 len=1
+ * process1: upgrade ofs=0 len=2 to F_RWLCK (in blocking mode)
+ * process2: F_RDLCK for ofs=1 len=1
+ * process2: unlock ofs=0 len=2
+ * process1: should continue at that point
+ *
+ * Such a test follows here...
+ */
+
+static int raw_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag)
+{
+       struct flock fl;
+       int cmd;
+       fl.l_type = rw;
+       fl.l_whence = SEEK_SET;
+       fl.l_start = off;
+       fl.l_len = len;
+       fl.l_pid = 0;
+
+       cmd = waitflag ? F_SETLKW : F_SETLK;
+
+       return fcntl(fd, cmd, &fl);
+}
+
+static int raw_fcntl_unlock(int fd, off_t off, off_t len)
+{
+       struct flock fl;
+       fl.l_type = F_UNLCK;
+       fl.l_whence = SEEK_SET;
+       fl.l_start = off;
+       fl.l_len = len;
+       fl.l_pid = 0;
+
+       return fcntl(fd, F_SETLKW, &fl);
+}
+
+
+int pipe_r;
+int pipe_w;
+char buf[2];
+
+static void expect_char(char c)
+{
+       read(pipe_r, buf, 1);
+       if (*buf != c) {
+               fail("We were expecting %c, but got %c", c, buf[0]);
+       }
+}
+
+static void send_char(char c)
+{
+       write(pipe_w, &c, 1);
+}
+
+
+int main(int argc, char *argv[])
+{
+       int process;
+       int fd;
+       const char *filename = "run-fcntl-deadlock.lck";
+       int pid;
+       int pipes_1_2[2];
+       int pipes_2_1[2];
+       int ret;
+
+       pipe(pipes_1_2);
+       pipe(pipes_2_1);
+       fd = open(filename, O_RDWR | O_CREAT, 0755);
+
+       pid = fork();
+       if (pid == 0) {
+               pipe_r = pipes_1_2[0];
+               pipe_w = pipes_2_1[1];
+               process = 2;
+               alarm(15);
+       } else {
+               pipe_r = pipes_2_1[0];
+               pipe_w = pipes_1_2[1];
+               process = 1;
+               alarm(15);
+       }
+
+       /* a: process1: F_RDLCK for ofs=0 len=2 */
+       if (process == 1) {
+               ret = raw_fcntl_lock(fd, F_RDLCK, 0, 2, true);
+               ok(ret == 0,
+                  "process 1 lock ofs=0 len=2: %d - %s",
+                  ret, strerror(errno));
+               diag("process 1 took read lock on range 0,2");
+               send_char('a');
+       }
+
+       /* process2: F_RDLCK for ofs=0 len=1 */
+       if (process == 2) {
+               expect_char('a');
+               ret = raw_fcntl_lock(fd, F_RDLCK, 0, 1, true);
+               ok(ret == 0,
+                  "process 2 lock ofs=0 len=1: %d - %s",
+                  ret, strerror(errno));;
+               diag("process 2 took read lock on range 0,1");
+               send_char('b');
+       }
+
+       /* process1: upgrade ofs=0 len=2 to F_RWLCK (in blocking mode) */
+       if (process == 1) {
+               expect_char('b');
+               send_char('c');
+               diag("process 1 starts upgrade on range 0,2");
+               ret = raw_fcntl_lock(fd, F_WRLCK, 0, 2, true);
+               ok(ret == 0,
+                  "process 1 RW lock ofs=0 len=2: %d - %s",
+                  ret, strerror(errno));
+               diag("process 1 got read upgrade done");
+               /* at this point process 1 is blocked on 2 releasing the
+                  read lock */
+       }
+
+       /*
+        * process2: F_RDLCK for ofs=1 len=1
+        * process2: unlock ofs=0 len=2
+        */
+       if (process == 2) {
+               expect_char('c'); /* we know process 1 is *about* to lock */
+               sleep(1);
+               ret = raw_fcntl_lock(fd, F_RDLCK, 1, 1, true);
+               ok(ret == 0,
+                 "process 2 lock ofs=1 len=1: %d - %s",
+                 ret, strerror(errno));
+               diag("process 2 got read lock on 1,1\n");
+               ret = raw_fcntl_unlock(fd, 0, 2);
+               ok(ret == 0,
+                 "process 2 unlock ofs=0 len=2: %d - %s",
+                 ret, strerror(errno));
+               diag("process 2 released read lock on 0,2\n");
+               sleep(1);
+               send_char('d');
+       }
+
+       if (process == 1) {
+               expect_char('d');
+       }
+
+       diag("process %d has got to the end\n", process);
+
+       return 0;
+}
index 987d78c..09bc0a3 100644 (file)
@@ -41,6 +41,7 @@ tdb1_unit_tests = [
     'run-traverse-in-transaction',
     'run-wronghash-fail',
     'run-zero-append',
+    'run-fcntl-deadlock',
     'run-marklock-deadlock',
     'run-allrecord-traverse-deadlock',
     'run-mutex-openflags2',