Re: [Qemu-devel] [PATCH v2] qemu-coroutine: segfault when restarting co_queue

[Stefan Hajnoczi]

On Mon, Feb 09, 2015 at 02:50:39PM +0800, Bin Wu wrote:
> From: Bin Wu <address@hidden>
> 
> We tested VMs migration with their disk images by drive_mirror. With
> migration, two VMs copyed large files between each other. During the
> test, a segfault occured. The stack was as follow:
> 
> (gdb) bt
> qemu-coroutine-lock.c:66
> to=0x7fa5a1798648) at qemu-coroutine.c:97
> request=0x7fa28c2ffa10, reply=0x7fa28c2ffa30, qiov=0x0, offset=0) at
> block/nbd-client.c:165
> sector_num=8552704, nb_sectors=2040, qiov=0x7fa5a1757468, offset=0) at
> block/nbd-client.c:262
> sector_num=8552704, nb_sectors=2048, qiov=0x7fa5a1757468) at
> block/nbd-client.c:296
> nb_sectors=2048, qiov=0x7fa5a1757468) at block/nbd.c:291
> req=0x7fa28c2ffbb0, offset=4378984448, bytes=1048576, qiov=0x7fa5a1757468,
> flags=0) at block.c:3321
> offset=4378984448, bytes=1048576, qiov=0x7fa5a1757468, flags=(unknown: 0)) at
> block.c:3447
> sector_num=8552704, nb_sectors=2048, qiov=0x7fa5a1757468, flags=(unknown: 0)) 
> at
> block.c:3471
> nb_sectors=2048, qiov=0x7fa5a1757468) at block.c:3480
> nb_sectors=2048, qiov=0x7fa5a1757468) at block/raw_bsd.c:62
> req=0x7fa28c2ffe30, offset=4378984448, bytes=1048576, qiov=0x7fa5a1757468,
> flags=0) at block.c:3321
> offset=4378984448, bytes=1048576, qiov=0x7fa5a1757468, flags=(unknown: 0)) at
> block.c:3447
> sector_num=8552704, nb_sectors=2048, qiov=0x7fa5a1757468, flags=(unknown: 0)) 
> at
> block.c:3471
> coroutine-ucontext.c:121

This backtrace is incomplete.  Where are the function names?  The
parameter lists appear incomplete too.

> After analyzing the stack and reviewing the code, we find the
> qemu_co_queue_run_restart should not be put in the coroutine_swap function 
> which
> can be invoked by qemu_coroutine_enter or qemu_coroutine_yield. Only
> qemu_coroutine_enter needs to restart the co_queue.
> 
> The error scenario is as follow: coroutine C1 enters C2, C2 yields
> back to C1, then C1 ternimates and the related coroutine memory
> becomes invalid. After a while, the C2 coroutine is entered again.
> At this point, C1 is used as a parameter passed to
> qemu_co_queue_run_restart. Therefore, qemu_co_queue_run_restart
> accesses an invalid memory and a segfault error ocurrs.
> 
> The qemu_co_queue_run_restart function re-enters coroutines waiting
> in the co_queue. However, this function should be only used int the
> qemu_coroutine_enter context. Only in this context, when the current
> coroutine gets execution control again(after the execution of
> qemu_coroutine_switch), we can restart the target coutine because the
> target coutine has yielded back to the current coroutine or it has
> terminated.
> 
> First we want to put qemu_co_queue_run_restart in qemu_coroutine_enter,
> but we find we can not access the target coroutine if it terminates.

This example captures the scenario you describe:

diff --git a/qemu-coroutine.c b/qemu-coroutine.c
index 525247b..883cbf5 100644
--- a/qemu-coroutine.c
+++ b/qemu-coroutine.c
@@ -103,7 +103,10 @@ static void coroutine_swap(Coroutine *from, Coroutine *to)
 {
     CoroutineAction ret;
 
+    fprintf(stderr, "> %s from %p to %p\n", __func__, from, to);
     ret = qemu_coroutine_switch(from, to, COROUTINE_YIELD);
+    fprintf(stderr, "< %s from %p to %p switch %s\n", __func__, from, to,
+            ret == COROUTINE_YIELD ? "yield" : "terminate");
 
     qemu_co_queue_run_restart(to);
 
@@ -111,6 +114,7 @@ static void coroutine_swap(Coroutine *from, Coroutine *to)
     case COROUTINE_YIELD:
         return;
     case COROUTINE_TERMINATE:
+        fprintf(stderr, "coroutine_delete %p\n", to);
         trace_qemu_coroutine_terminate(to);
         coroutine_delete(to);
         return;
diff --git a/tests/test-coroutine.c b/tests/test-coroutine.c
index 27d1b6f..d44c428 100644
--- a/tests/test-coroutine.c
+++ b/tests/test-coroutine.c
@@ -13,6 +13,7 @@
 
 #include <glib.h>
 #include "block/coroutine.h"
+#include "block/coroutine_int.h"
 
 /*
  * Check that qemu_in_coroutine() works
@@ -122,6 +123,35 @@ static void test_yield(void)
     g_assert_cmpint(i, ==, 5); /* coroutine must yield 5 times */
 }
 
+static void coroutine_fn c2_fn(void *opaque)
+{
+    fprintf(stderr, "c2 Part 1\n");
+    qemu_coroutine_yield();
+    fprintf(stderr, "c2 Part 2\n");
+}
+
+static void coroutine_fn c1_fn(void *opaque)
+{
+    Coroutine *c2 = opaque;
+
+    fprintf(stderr, "c1 Part 1\n");
+    qemu_coroutine_enter(c2, NULL);
+    fprintf(stderr, "c1 Part 2\n");
+}
+
+static void test_co_queue(void)
+{
+    Coroutine *c1;
+    Coroutine *c2;
+
+    c1 = qemu_coroutine_create(c1_fn);
+    c2 = qemu_coroutine_create(c2_fn);
+
+    qemu_coroutine_enter(c1, c2);
+    memset(c1, 0xff, sizeof(Coroutine));
+    qemu_coroutine_enter(c2, NULL);
+}
+
 /*
  * Check that creation, enter, and return work
  */
@@ -343,6 +373,7 @@ static void perf_cost(void)
 int main(int argc, char **argv)
 {
     g_test_init(&argc, &argv, NULL);
+    g_test_add_func("/basic/co_queue", test_co_queue);
     g_test_add_func("/basic/lifecycle", test_lifecycle);
     g_test_add_func("/basic/yield", test_yield);
     g_test_add_func("/basic/nesting", test_nesting);

Here is the output (with printfs in coroutine_swap):

-> coroutine_swap from MAIN to C1
c1 Part 1
-> coroutine_swap from C1 to C2
c2 Part 1
-> coroutine_swap from C2 to C1
<- coroutine_swap from C1 to C2 switch yield
c1 Part 2
<- coroutine_swap from MAIN to C1 switch terminate
coroutine_delete C1
-> coroutine_swap from MAIN to C2
<- coroutine_swap from C2 to C1 switch yield  !!!
c2 Part 2
<- coroutine_swap from MAIN to C2 switch terminate
coroutine_delete C2

I have marked the problematic line with "!!!".  The to=C1 variable is
used after C1 has been deleted.

The test crashes since it writes 0xff to C1 after it has terminated.

> Signed-off-by: Bin Wu <address@hidden>
> ---
>  qemu-coroutine.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
> 
> diff --git a/qemu-coroutine.c b/qemu-coroutine.c
> index 525247b..cc0bdfa 100644
> --- a/qemu-coroutine.c
> +++ b/qemu-coroutine.c
> @@ -99,29 +99,31 @@ static void coroutine_delete(Coroutine *co)
>      qemu_coroutine_delete(co);
>  }
>  
> -static void coroutine_swap(Coroutine *from, Coroutine *to)
> +static CoroutineAction coroutine_swap(Coroutine *from, Coroutine *to)
>  {
>      CoroutineAction ret;
>  
>      ret = qemu_coroutine_switch(from, to, COROUTINE_YIELD);
>  
> -    qemu_co_queue_run_restart(to);
> -
>      switch (ret) {
>      case COROUTINE_YIELD:
> -        return;
> +        break;
>      case COROUTINE_TERMINATE:
>          trace_qemu_coroutine_terminate(to);
> +        qemu_co_queue_run_restart(to);
>          coroutine_delete(to);
> -        return;
> +        break;
>      default:
>          abort();
>      }
> +
> +    return ret;
>  }
>  
>  void qemu_coroutine_enter(Coroutine *co, void *opaque)
>  {
>      Coroutine *self = qemu_coroutine_self();
> +    CoroutineAction ret;
>  
>      trace_qemu_coroutine_enter(self, co, opaque);
>  
> @@ -132,7 +134,9 @@ void qemu_coroutine_enter(Coroutine *co, void *opaque)
>  
>      co->caller = self;
>      co->entry_arg = opaque;
> -    coroutine_swap(self, co);
> +    ret = coroutine_swap(self, co);
> +    if (ret == COROUTINE_YIELD)
> +        qemu_co_queue_run_restart(co);
>  }

Your fix looks correct although QEMU coding style requires {}.

I tried to think of a simpler solution that keeps a single
qemu_co_queue_run_restart() call but was unable to find one.

Please send another revision with a test-coroutine.c test case so we can
protect against regressions.

Thanks,
Stefan

pgp9n8F45zl_5.pgp
Description: PGP signature