Am 09.02.2015 um 15:48 hat Stefan Hajnoczi geschrieben:
> On Mon, Feb 09, 2015 at 02:50:39PM +0800, Bin Wu wrote:
> > From: Bin Wu <wu.wu...@huawei.com>
> >
> > 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 <wu.wu...@huawei.com>
> > ---
> > 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.
This one-liner is the core of the real fix:
--- a/qemu-coroutine.c
+++ b/qemu-coroutine.c
@@ -148,5 +148,5 @@ void coroutine_fn qemu_coroutine_yield(void)
}
self->caller = NULL;
- coroutine_swap(self, to);
+ qemu_coroutine_switch(self, to, COROUTINE_YIELD);
}
There is no reason at all for qemu_coroutine_yield() to go through
coroutine_swap(). The cleanups that can be made on top is to inline
coroutine_swap() in qemu_coroutine_enter(), and to change its use of
COROUTINE_YIELD into a new COROUTINE_ENTER (which is never looked at
anyway, but we're not yielding here).
All the other solutions take a unnecessarily complicated design as
granted and further complicate it to distinguish two unrelated code
paths.
Kevin
pgp3zaxTZrE9A.pgp
Description: PGP signature
