[Qemu-devel] [PATCH] qed: replace vm_clock with rt_clock for qemu-tool compatibility

[Stefan Hajnoczi]

The QED dirty bit timer marks the file clean after allocating writes
have drained.  This is cheaper than clearing/setting the dirty bit on
each allocating write because the timer introduces a grace period which
can be extended if more allocating writes arrive.

The vm_clock was used in an attempt to prevent modifying the image file
when live migration has stopped the VM.  Unfortunately vm_clock is
unavailable in the qemu-tool environment and will abort(3)!

Since QED currently does not support live migration, just replace
vm_clock with rt_clock and add comments explaining the migration
blocker.

Signed-off-by: Stefan Hajnoczi <address@hidden>
---
Zhi Yong: This patch is needed in addition to the qemu_init_main_loop() patches
you sent recently.  Without this patch QED may read the vm_clock, which calls
abort(3) in qemu-tool.c.  Together, our patches make QED work again in qemu-img
and qemu-io.

 block/qed.c |   16 +++++++++++-----
 1 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/block/qed.c b/block/qed.c
index a041d31..fdb90e3 100644
--- a/block/qed.c
+++ b/block/qed.c
@@ -353,10 +353,7 @@ static void qed_start_need_check_timer(BDRVQEDState *s)
 {
     trace_qed_start_need_check_timer(s);
 
-    /* Use vm_clock so we don't alter the image file while suspended for
-     * migration.
-     */
-    qemu_mod_timer(s->need_check_timer, qemu_get_clock_ns(vm_clock) +
+    qemu_mod_timer(s->need_check_timer, qemu_get_clock_ns(rt_clock) +
                    get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
 }
 
@@ -494,9 +491,18 @@ static int bdrv_qed_open(BlockDriverState *bs, int flags)
         }
     }
 
-    s->need_check_timer = qemu_new_timer_ns(vm_clock,
+    s->need_check_timer = qemu_new_timer_ns(rt_clock,
                                             qed_need_check_timer_cb, s);
 
+    /* There are two issues with live migration:
+     *
+     * 1. The destination will open the image file and see the dirty bit is
+     *    set, causing it to "repair" the image while the source still has it
+     *    open for writing.
+     *
+     * 2. The timer used for clearing the dirty bit uses rt_clock and can in
+     *    theory fire when the VM is not running during migration.
+     */
     error_set(&s->migration_blocker,
               QERR_BLOCK_FORMAT_FEATURE_NOT_SUPPORTED,
               "qed", bs->device_name, "live migration");
-- 
1.7.9