[PATCH] memtx: cancel checkpoint thread at exit

[Vladimir Davydov]

On Thu, Apr 18, 2019 at 06:46:54PM +0300, Vladimir Davydov wrote:
> +static void
> +checkpoint_cancel(struct checkpoint *ckpt)
> +{
> +	/*
> +	 * Cancel the checkpoint thread if it's running and wait
> +	 * for it to terminate so as to eliminate the possibility
> +	 * of use-after-free.
> +	 */
> +	if (ckpt->cord.id != 0 &&
> +	    tt_pthread_cancel(ckpt->cord.id) != ESRCH)
> +		tt_pthread_join(ckpt->cord.id, NULL);
> +	checkpoint_delete(ckpt);
> +}

Oops, turned out this check isn't quite correct as we don't reset
cord.id after joining the thread. As a result, we might call
pthread_cancel on an already joined thread, which leads to a crash
(caught by Travis CI):

 | #0  0x557f6332e9a5 in print_backtrace+1b
 | #1  0x557f6310614f in _ZL12sig_fatal_cbiP9siginfo_tPv+451
 | #2  0x7f98504e10e0 in __restore_rt+0
 | #3  0x7f98504de499 in pthread_cancel+9
 | #4  0x557f6314aa48 in checkpoint_cancel+52
 | #5  0x557f6314866c in memtx_engine_shutdown+55
 | #6  0x557f63144fb1 in engine_shutdown+6f
 | #7  0x557f6326bb52 in box_free+98
 | #8  0x557f631083cf in _Z14tarantool_freev+a9
 | #9  0x557f63109ae2 in main+10cf
 | #10 0x7f984f24e2e1 in __libc_start_main+f1
 | #11 0x557f63104f6a in _start+2a
 | #12 (nil) in +2a

Actually, we don't need to check cord.id at all - we can use already
existing waiting_for_snap_thread flag instead. Fixed this as follows
(this is an incremental diff; for the full diff see below):

diff --git a/src/box/memtx_engine.c b/src/box/memtx_engine.c
index 2b9bef24..4d99910c 100644
--- a/src/box/memtx_engine.c
+++ b/src/box/memtx_engine.c
@@ -595,11 +595,6 @@ checkpoint_new(const char *snap_dirname, uint64_t snap_io_rate_limit)
 	xdir_create(&ckpt->dir, snap_dirname, SNAP, &INSTANCE_UUID, &opts);
 	vclock_create(&ckpt->vclock);
 	ckpt->touch = false;
-	/*
-	 * Reset the checkpoint thread id so that checkpoint_cancel()
-	 * doesn't attempt to cancel the thread if it isn't running.
-	 */
-	ckpt->cord.id = 0;
 	return ckpt;
 }
 
@@ -623,9 +618,10 @@ checkpoint_cancel(struct checkpoint *ckpt)
 	 * for it to terminate so as to eliminate the possibility
 	 * of use-after-free.
 	 */
-	if (ckpt->cord.id != 0 &&
-	    tt_pthread_cancel(ckpt->cord.id) != ESRCH)
+	if (ckpt->waiting_for_snap_thread) {
+		tt_pthread_cancel(ckpt->cord.id);
 		tt_pthread_join(ckpt->cord.id, NULL);
+	}
 	checkpoint_delete(ckpt);
 }
 

>From 66ff448a89864c3de7d34563a515295d62f7f944 Mon Sep 17 00:00:00 2001
From: Vladimir Davydov <vdavydov.dev at gmail.com>
Date: Thu, 18 Apr 2019 18:03:26 +0300
Subject: [PATCH] memtx: cancel checkpoint thread at exit

If a tarantool instance exits while checkpointing is in progress, the
memtx checkpoint thread, which writes the snap file, can access already
freed data resulting in a crash. Let's fix this the same way we did for
relay and vinyl threads - simply cancel the thread forcefully and wait
for it to terminate.

Closes #4170

diff --git a/src/box/memtx_engine.c b/src/box/memtx_engine.c
index 48f700a0..4d99910c 100644
--- a/src/box/memtx_engine.c
+++ b/src/box/memtx_engine.c
@@ -52,6 +52,9 @@
 /* sync snapshot every 16MB */
 #define SNAP_SYNC_INTERVAL	(1 << 24)
 
+static void
+checkpoint_cancel(struct checkpoint *ckpt);
+
 /*
  * Memtx yield-in-transaction trigger: roll back the effects
  * of the transaction and mark the transaction as aborted.
@@ -179,6 +182,8 @@ static void
 memtx_engine_shutdown(struct engine *engine)
 {
 	struct memtx_engine *memtx = (struct memtx_engine *)engine;
+	if (memtx->checkpoint != NULL)
+		checkpoint_cancel(memtx->checkpoint);
 	mempool_destroy(&memtx->iterator_pool);
 	if (mempool_is_initialized(&memtx->rtree_iterator_pool))
 		mempool_destroy(&memtx->rtree_iterator_pool);
@@ -605,6 +610,20 @@ checkpoint_delete(struct checkpoint *ckpt)
 	free(ckpt);
 }
 
+static void
+checkpoint_cancel(struct checkpoint *ckpt)
+{
+	/*
+	 * Cancel the checkpoint thread if it's running and wait
+	 * for it to terminate so as to eliminate the possibility
+	 * of use-after-free.
+	 */
+	if (ckpt->waiting_for_snap_thread) {
+		tt_pthread_cancel(ckpt->cord.id);
+		tt_pthread_join(ckpt->cord.id, NULL);
+	}
+	checkpoint_delete(ckpt);
+}
 
 static int
 checkpoint_add_space(struct space *sp, void *data)