scsi: fix device removal race vs IO restart callback on resume
authorMaxim Levitsky <mlevitsk@redhat.com>
Thu, 10 Dec 2020 12:59:29 +0000 (14:59 +0200)
committerPaolo Bonzini <pbonzini@redhat.com>
Tue, 15 Dec 2020 17:52:04 +0000 (12:52 -0500)
There is (mostly theoretical) race between removal of a scsi device and
scsi_dma_restart_bh.

It used to be easier to hit this race prior to my / Paulo's patch series
that added rcu to scsi bus device handling code, but IMHO this race
should still be possible to hit, at least in theory.

Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=1854811
Fix it anyway with a patch that was proposed by Paulo in the above bugzilla.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20201210125929.1136390-2-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
hw/scsi/scsi-bus.c

index b901e701f081ffc5637fac0cdedbb285d4b4eeaa..edb5c3492a83a2e86be2523366f0a85485a972f6 100644 (file)
@@ -170,6 +170,8 @@ static void scsi_dma_restart_bh(void *opaque)
         scsi_req_unref(req);
     }
     aio_context_release(blk_get_aio_context(s->conf.blk));
+    /* Drop the reference that was acquired in scsi_dma_restart_cb */
+    object_unref(OBJECT(s));
 }
 
 void scsi_req_retry(SCSIRequest *req)
@@ -188,6 +190,8 @@ static void scsi_dma_restart_cb(void *opaque, int running, RunState state)
     }
     if (!s->bh) {
         AioContext *ctx = blk_get_aio_context(s->conf.blk);
+        /* The reference is dropped in scsi_dma_restart_bh.*/
+        object_ref(OBJECT(s));
         s->bh = aio_bh_new(ctx, scsi_dma_restart_bh, s);
         qemu_bh_schedule(s->bh);
     }