Re: [RFC PATCH] test-bdrv-drain: keep graph manipulations out of coroutines

[Emanuele Giuseppe Esposito]

Am 09/12/2022 um 13:18 schrieb Emanuele Giuseppe Esposito:
> 
> 
> Am 05/12/2022 um 14:01 schrieb Kevin Wolf:
>> Am 02.12.2022 um 18:22 hat Paolo Bonzini geschrieben:
>>> On 12/2/22 14:42, Emanuele Giuseppe Esposito wrote:
>>>>
>>>>
>>>> Am 02/12/2022 um 14:27 schrieb Paolo Bonzini:
>>>>> Changes to the BlockDriverState graph will have to take the
>>>>> corresponding lock for writing, and therefore cannot be done
>>>>> inside a coroutine.  Move them outside the test body.
>>>>>
>>>>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>>>>> ---
>>>>>   tests/unit/test-bdrv-drain.c | 63 ++++++++++++++++++++++++++----------
>>>>>   1 file changed, 46 insertions(+), 17 deletions(-)
>>>>>
>>>>> diff --git a/tests/unit/test-bdrv-drain.c b/tests/unit/test-bdrv-drain.c
>>>>> index 6ae44116fe79..d85083dd4f9e 100644
>>>>> --- a/tests/unit/test-bdrv-drain.c
>>>>> +++ b/tests/unit/test-bdrv-drain.c
>>>>> @@ -199,25 +199,40 @@ static void do_drain_end_unlocked(enum drain_type 
>>>>> drain_type, BlockDriverState *
>>>>>       }
>>>>>   }
>>>>> +static BlockBackend *blk;
>>>>> +static BlockDriverState *bs, *backing;
>>>>> +
>>>>> +static void test_drv_cb_init(void)
>>>>> +{
>>>>> +    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
>>>>> +    bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
>>>>> +                              &error_abort);
>>>>> +    blk_insert_bs(blk, bs, &error_abort);
>>>>> +
>>>>> +    backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, 
>>>>> &error_abort);
>>>>> +    bdrv_set_backing_hd(bs, backing, &error_abort);
>>>>> +}
>>>>> +
>>>>> +static void test_drv_cb_fini(void)
>>>>
>>>> fini stands for "finito"? :)
>>>
>>> No, for finish :)
>>> http://ftp.math.utah.edu/u/ma/hohn/linux/misc/elf/node3.html
>>>
>>>> Anyways, an alternative solution for this is also here (probably coming
>>>> from you too):
>>>> https://lists.nongnu.org/archive/html/qemu-devel/2022-03/msg03517.html
>>>
>>> Much better.  At least patches 7-8 from that series have to be salvaged,
>>> possibly 10 as well.
>>
>> I wonder if we need a more general solution for this because this test
>> is not the only place that calls this kind of functions in a coroutine.
>> The one I'm aware of in particular is all the .bdrv_co_create
>> implementations, but I'm almost sure there are more.
>>
>> Can we use a yield_to_drain()-like mechanism for these functions? Maybe
>> even something like the opposite of co_wrapper, a no_co_wrapper that
>> generates a foo_co() variant that drops out of coroutine context before
>> calling foo()?
>>
> 
> I implemented something like yield_to_drain as you suggested, but when
> thinking about it aren't we making a fix that will cost us even more
> work in the future? If we use a yield_to_drain-like function, we are
> doing something similar to g_c_w, and losing track of whether the caller
> is a coroutine or not. And the function could then be used potentially
> everywhere. Then we will realize "oh we need to get rid of this and
> split the functions differentiating the coroutine context" and
> eventually go through ALL the callers again to figure what is doing
> what, and implement the same fix of this patch or my series once again.
> 
> Instead, even though this is just a test, we have a clear separation and
> one less case to worry about in the future.
> 
At least the above is valid if the change you are proposing is the
following (tested already, works)


diff --git a/block.c b/block.c
index 6191ac1f44..8d28c1daa4 100644
--- a/block.c
+++ b/block.c
@@ -42,6 +42,7 @@
 #include "qapi/qobject-output-visitor.h"
 #include "qapi/qapi-visit-block-core.h"
 #include "sysemu/block-backend.h"
+#include "sysemu/replay.h"
 #include "qemu/notify.h"
 #include "qemu/option.h"
 #include "qemu/coroutine.h"
@@ -2831,6 +2832,94 @@ uint64_t
bdrv_qapi_perm_to_blk_perm(BlockPermission qapi_perm)
     return permissions[qapi_perm];
 }

+typedef struct {
+    Coroutine *co;
+    BlockDriverState *new_bs;
+    BdrvChild *child;
+    AioContext *ctx;
+    bool done;
+} BdrvCoGraphModData;
+
+static void bdrv_co_graph_mod_bh_cb(void *opaque)
+{
+    BdrvCoGraphModData *data = opaque;
+    Coroutine *co = data->co;
+    BlockDriverState *old_bs = data->child->bs;
+    BlockDriverState *new_bs = data->new_bs;
+
+    aio_context_acquire(data->ctx);
+
+     if (old_bs) {
+        bdrv_dec_in_flight(old_bs);
+    }
+
+    if (new_bs) {
+        bdrv_dec_in_flight(new_bs);
+    }
+    bdrv_replace_child_noperm(data->child, data->new_bs);
+    aio_context_release(data->ctx);
+
+    data->done = true;
+    aio_co_wake(co);
+}
+
+/*
+ * If bdrv_replace_child_noperm is called in a coroutine, defer the
work to the
+ * Main Loop by scheduling a BH.
+ */
+static void bdrv_co_yield_to_graph_mod(BdrvChild *child,
BlockDriverState *new_bs)
+{
+    BdrvCoGraphModData data;
+    Coroutine *self = qemu_coroutine_self();
+    AioContext *ctx = bdrv_get_aio_context(new_bs);
+    BlockDriverState *old_bs = child->bs;
+
+    GLOBAL_STATE_CODE();
+    assert(qemu_in_coroutine());
+
+    if (old_bs) {
+        bdrv_inc_in_flight(old_bs);
+        ctx = bdrv_get_aio_context(old_bs);
+    }
+
+    if (new_bs) {
+        bdrv_inc_in_flight(new_bs);
+    }
+
+    data = (BdrvCoGraphModData) {
+        .co = self,
+        .new_bs = new_bs,
+        .child = child,
+        .ctx = ctx,
+        .done = false,
+    };
+
+    /*
+     * Temporarily drop the lock across yield or we would get deadlocks.
+     * bdrv_co_drain_bh_cb() reaquires the lock as needed.
+     *
+     * When we yield below, the lock for the current context will be
+     * released, so if this is actually the lock that protects bs,
don't drop
+     * it a second time.
+     */
+    if (ctx != qemu_get_aio_context()) {
+        aio_context_release(ctx);
+    }
+    replay_bh_schedule_oneshot_event(ctx, bdrv_co_graph_mod_bh_cb, &data);
+
+    qemu_coroutine_yield();
+    /*
+     * If we are resumed from some other event (such as an aio
completion or a
+     * timer callback), it is a bug in the caller that should be fixed.
+     */
+    assert(data.done);
+
+    /* Reaquire the AioContext of bs if we dropped it */
+    if (ctx != qemu_get_aio_context()) {
+        aio_context_acquire(ctx);
+    }
+}
+
 /*
  * Replaces the node that a BdrvChild points to without updating
permissions.
  *
@@ -2875,6 +2964,12 @@ static void bdrv_replace_child_noperm(BdrvChild
*child,
         assert(bdrv_get_aio_context(old_bs) ==
bdrv_get_aio_context(new_bs));
     }

+    /* No graph modification is allowed in a coroutine! */
+    if (qemu_in_coroutine()) {
+        bdrv_co_yield_to_graph_mod(child, new_bs);
+        return;
+    }
+
     if (old_bs) {
         if (child->klass->detach) {
             child->klass->detach(child);