[Qemu-devel] [PATCH 09/18] qed: Periodically flush and clear need check bit

[Kevin Wolf <kwolf@redhat.com>]

From: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>

One strategy to limit the startup delay of consistency check when
opening image files is to ensure that the file is marked dirty for as
little time as possible.

QED currently marks the image dirty when the first allocating write
request is issued and clears the dirty bit again when the image is
cleanly closed.  In practice that means the image is marked dirty for
most of a guest's lifetime and prone to being in a dirty state upon
crash or power failure.

It is safe to clear the dirty bit after all allocating write requests
have completed and a flush has been performed.  This patch adds a timer
after the last allocating write request completes.  When the timer fires
it will flush and then clear the dirty bit.  The timer is set to 5
seconds and is cancelled upon arrival of a new allocating write request.

Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 block/qed.c  |  104 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 block/qed.h  |    7 ++++
 trace-events |    3 ++
 3 files changed, 112 insertions(+), 2 deletions(-)

diff --git a/block/qed.c b/block/qed.c
index c8c5930..d8d6ea2 100644
--- a/block/qed.c
+++ b/block/qed.c
@@ -12,6 +12,7 @@
  *
  */
 
+#include "qemu-timer.h"
 #include "trace.h"
 #include "qed.h"
 #include "qerror.h"
@@ -291,6 +292,88 @@ static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
 
 static void qed_aio_next_io(void *opaque, int ret);
 
+static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
+{
+    assert(!s->allocating_write_reqs_plugged);
+
+    s->allocating_write_reqs_plugged = true;
+}
+
+static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
+{
+    QEDAIOCB *acb;
+
+    assert(s->allocating_write_reqs_plugged);
+
+    s->allocating_write_reqs_plugged = false;
+
+    acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
+    if (acb) {
+        qed_aio_next_io(acb, 0);
+    }
+}
+
+static void qed_finish_clear_need_check(void *opaque, int ret)
+{
+    /* Do nothing */
+}
+
+static void qed_flush_after_clear_need_check(void *opaque, int ret)
+{
+    BDRVQEDState *s = opaque;
+
+    bdrv_aio_flush(s->bs, qed_finish_clear_need_check, s);
+
+    /* No need to wait until flush completes */
+    qed_unplug_allocating_write_reqs(s);
+}
+
+static void qed_clear_need_check(void *opaque, int ret)
+{
+    BDRVQEDState *s = opaque;
+
+    if (ret) {
+        qed_unplug_allocating_write_reqs(s);
+        return;
+    }
+
+    s->header.features &= ~QED_F_NEED_CHECK;
+    qed_write_header(s, qed_flush_after_clear_need_check, s);
+}
+
+static void qed_need_check_timer_cb(void *opaque)
+{
+    BDRVQEDState *s = opaque;
+
+    /* The timer should only fire when allocating writes have drained */
+    assert(!QSIMPLEQ_FIRST(&s->allocating_write_reqs));
+
+    trace_qed_need_check_timer_cb(s);
+
+    qed_plug_allocating_write_reqs(s);
+
+    /* Ensure writes are on disk before clearing flag */
+    bdrv_aio_flush(s->bs, qed_clear_need_check, s);
+}
+
+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) +
+                   get_ticks_per_sec() * QED_NEED_CHECK_TIMEOUT);
+}
+
+/* It's okay to call this multiple times or when no timer is started */
+static void qed_cancel_need_check_timer(BDRVQEDState *s)
+{
+    trace_qed_cancel_need_check_timer(s);
+    qemu_del_timer(s->need_check_timer);
+}
+
 static int bdrv_qed_open(BlockDriverState *bs, int flags)
 {
     BDRVQEDState *s = bs->opaque;
@@ -406,7 +489,10 @@ static int bdrv_qed_open(BlockDriverState *bs, int flags)
             BdrvCheckResult result = {0};
 
             ret = qed_check(s, &result, true);
-            if (!ret && !result.corruptions && !result.check_errors) {
+            if (ret) {
+                goto out;
+            }
+            if (!result.corruptions && !result.check_errors) {
                 /* Ensure fixes reach storage before clearing check bit */
                 bdrv_flush(s->bs);
 
@@ -416,6 +502,9 @@ static int bdrv_qed_open(BlockDriverState *bs, int flags)
         }
     }
 
+    s->need_check_timer = qemu_new_timer_ns(vm_clock,
+                                            qed_need_check_timer_cb, s);
+
 out:
     if (ret) {
         qed_free_l2_cache(&s->l2_cache);
@@ -428,6 +517,9 @@ static void bdrv_qed_close(BlockDriverState *bs)
 {
     BDRVQEDState *s = bs->opaque;
 
+    qed_cancel_need_check_timer(s);
+    qemu_free_timer(s->need_check_timer);
+
     /* Ensure writes reach stable storage */
     bdrv_flush(bs->file);
 
@@ -809,6 +901,8 @@ static void qed_aio_complete(QEDAIOCB *acb, int ret)
         acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
         if (acb) {
             qed_aio_next_io(acb, 0);
+        } else if (s->header.features & QED_F_NEED_CHECK) {
+            qed_start_need_check_timer(s);
         }
     }
 }
@@ -1014,11 +1108,17 @@ static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
 {
     BDRVQEDState *s = acb_to_s(acb);
 
+    /* Cancel timer when the first allocating request comes in */
+    if (QSIMPLEQ_EMPTY(&s->allocating_write_reqs)) {
+        qed_cancel_need_check_timer(s);
+    }
+
     /* Freeze this request if another allocating write is in progress */
     if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
         QSIMPLEQ_INSERT_TAIL(&s->allocating_write_reqs, acb, next);
     }
-    if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs)) {
+    if (acb != QSIMPLEQ_FIRST(&s->allocating_write_reqs) ||
+        s->allocating_write_reqs_plugged) {
         return; /* wait for existing request to finish */
     }
 
diff --git a/block/qed.h b/block/qed.h
index 1d1421f..388fdb3 100644
--- a/block/qed.h
+++ b/block/qed.h
@@ -78,6 +78,9 @@ enum {
     QED_MIN_TABLE_SIZE = 1,        /* in clusters */
     QED_MAX_TABLE_SIZE = 16,
     QED_DEFAULT_TABLE_SIZE = 4,
+
+    /* Delay to flush and clean image after last allocating write completes */
+    QED_NEED_CHECK_TIMEOUT = 5,    /* in seconds */
 };
 
 typedef struct {
@@ -157,6 +160,10 @@ typedef struct {
 
     /* Allocating write request queue */
     QSIMPLEQ_HEAD(, QEDAIOCB) allocating_write_reqs;
+    bool allocating_write_reqs_plugged;
+
+    /* Periodic flush and clear need check flag */
+    QEMUTimer *need_check_timer;
 } BDRVQEDState;
 
 enum {
diff --git a/trace-events b/trace-events
index a00b63c..385cb00 100644
--- a/trace-events
+++ b/trace-events
@@ -220,6 +220,9 @@ disable qed_write_table(void *s, uint64_t offset, void *table, unsigned int inde
 disable qed_write_table_cb(void *s, void *table, int flush, int ret) "s %p table %p flush %d ret %d"
 
 # block/qed.c
+disable qed_need_check_timer_cb(void *s) "s %p"
+disable qed_start_need_check_timer(void *s) "s %p"
+disable qed_cancel_need_check_timer(void *s) "s %p"
 disable qed_aio_complete(void *s, void *acb, int ret) "s %p acb %p ret %d"
 disable qed_aio_setup(void *s, void *acb, int64_t sector_num, int nb_sectors, void *opaque, int is_write) "s %p acb %p sector_num %"PRId64" nb_sectors %d opaque %p is_write %d"
 disable qed_aio_next_io(void *s, void *acb, int ret, uint64_t cur_pos) "s %p acb %p ret %d cur_pos %"PRIu64""
-- 
1.7.2.3