From: Hanna Czenczek <hreitz@redhat.com>

scsi_device_for_each_req_async() currently does not provide any way to
be awaited.  One of its callers is scsi_device_purge_requests(), which
therefore currently does not guarantee that all requests are fully
settled when it returns.

We want all requests to be settled, because scsi_device_purge_requests()
is called through the unrealize path, including the one invoked by
virtio_scsi_hotunplug() through qdev_simple_device_unplug_cb(), which
most likely assumes that all SCSI requests are done then.

In fact, scsi_device_purge_requests() already contains a blk_drain(),
but this will not fully await scsi_device_for_each_req_async(), only the
I/O requests it potentially cancels (not the non-I/O requests).
However, we can have scsi_device_for_each_req_async() increment the BB
in-flight counter, and have scsi_device_for_each_req_async_bh()
decrement it when it is done.  This way, the blk_drain() will fully
await all SCSI requests to be purged.

This also removes the need for scsi_device_for_each_req_async_bh() to
double-check the current context and potentially re-schedule itself,
should it now differ from the BB's context: Changing a BB's AioContext
with a root node is done through bdrv_try_change_aio_context(), which
creates a drained section.  With this patch, we keep the BB in-flight
counter elevated throughout, so we know the BB's context cannot change.

Signed-off-by: Hanna Czenczek <hreitz@redhat.com>
Message-ID: <20240202144755.671354-3-hreitz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 hw/scsi/scsi-bus.c | 30 +++++++++++++++++++++---------
 1 file changed, 21 insertions(+), 9 deletions(-)

diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c
index 0a2eb11c56..230313022c 100644
--- a/hw/scsi/scsi-bus.c
+++ b/hw/scsi/scsi-bus.c
@@ -120,17 +120,13 @@ static void scsi_device_for_each_req_async_bh(void *opaque)
     SCSIRequest *next;
 
     /*
-     * If the AioContext changed before this BH was called then reschedule into
-     * the new AioContext before accessing ->requests. This can happen when
-     * scsi_device_for_each_req_async() is called and then the AioContext is
-     * changed before BHs are run.
+     * The BB cannot have changed contexts between this BH being scheduled and
+     * now: BBs' AioContexts, when they have a node attached, can only be
+     * changed via bdrv_try_change_aio_context(), in a drained section.  While
+     * we have the in-flight counter incremented, that drain must block.
      */
     ctx = blk_get_aio_context(s->conf.blk);
-    if (ctx != qemu_get_current_aio_context()) {
-        aio_bh_schedule_oneshot(ctx, scsi_device_for_each_req_async_bh,
-                                g_steal_pointer(&data));
-        return;
-    }
+    assert(ctx == qemu_get_current_aio_context());
 
     QTAILQ_FOREACH_SAFE(req, &s->requests, next, next) {
         data->fn(req, data->fn_opaque);
@@ -138,11 +134,16 @@ static void scsi_device_for_each_req_async_bh(void *opaque)
 
     /* Drop the reference taken by scsi_device_for_each_req_async() */
     object_unref(OBJECT(s));
+
+    /* Paired with blk_inc_in_flight() in scsi_device_for_each_req_async() */
+    blk_dec_in_flight(s->conf.blk);
 }
 
 /*
  * Schedule @fn() to be invoked for each enqueued request in device @s. @fn()
  * runs in the AioContext that is executing the request.
+ * Keeps the BlockBackend's in-flight counter incremented until everything is
+ * done, so draining it will settle all scheduled @fn() calls.
  */
 static void scsi_device_for_each_req_async(SCSIDevice *s,
                                            void (*fn)(SCSIRequest *, void *),
@@ -163,6 +164,8 @@ static void scsi_device_for_each_req_async(SCSIDevice *s,
      */
     object_ref(OBJECT(s));
 
+    /* Paired with blk_dec_in_flight() in scsi_device_for_each_req_async_bh() */
+    blk_inc_in_flight(s->conf.blk);
     aio_bh_schedule_oneshot(blk_get_aio_context(s->conf.blk),
                             scsi_device_for_each_req_async_bh,
                             data);
@@ -1728,11 +1731,20 @@ static void scsi_device_purge_one_req(SCSIRequest *req, void *opaque)
     scsi_req_cancel_async(req, NULL);
 }
 
+/**
+ * Cancel all requests, and block until they are deleted.
+ */
 void scsi_device_purge_requests(SCSIDevice *sdev, SCSISense sense)
 {
     scsi_device_for_each_req_async(sdev, scsi_device_purge_one_req, NULL);
 
+    /*
+     * Await all the scsi_device_purge_one_req() calls scheduled by
+     * scsi_device_for_each_req_async(), and all I/O requests that were
+     * cancelled this way, but may still take a bit of time to settle.
+     */
     blk_drain(sdev->conf.blk);
+
     scsi_device_set_ua(sdev, sense);
 }
 
-- 
2.43.0