[Qemu-devel] [RFC 5/8] timer: associate timer with AioContext

[Liu Ping Fan <qemulist@gmail.com>]

Currently, timers run on iothread inside QBL, this limits the usage
of timers in some case, e.g. virtio-blk-dataplane. In order to run
timers on private thread, we arm AioContext with three lists in
according to three QemuClock (and later, run timers in aio_poll).

Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>

------ issue to fix ---
Note: before this patch, there should be another one to fix the race
issue by qemu_mod_timer() and _run_timers().  I plan to adopt the BH
method for timers to fix it.
---
 async.c              |  9 +++++++
 include/block/aio.h  | 13 ++++++++++
 include/qemu/timer.h | 20 ++++++++++++++++
 qemu-timer.c         | 67 +++++++++++++++++++++++++++++-----------------------
 4 files changed, 80 insertions(+), 29 deletions(-)

diff --git a/async.c b/async.c
index 8209cea..36df208 100644
--- a/async.c
+++ b/async.c
@@ -202,12 +202,16 @@ static void
 aio_ctx_finalize(GSource     *source)
 {
     AioContext *ctx = (AioContext *) source;
+    int i;
 
     thread_pool_free(ctx->thread_pool);
     aio_set_event_notifier(ctx, &ctx->notifier, NULL, NULL);
     event_notifier_cleanup(&ctx->notifier);
     g_array_free(ctx->pollfds, TRUE);
     alarm_timer_destroy(ctx->alarm_timer);
+    for (i = 0; i < QEMU_CLOCK_MAXCNT; i++) {
+        timer_list_finalize(&ctx->timer_list[i]);
+    }
 }
 
 static GSourceFuncs aio_source_funcs = {
@@ -239,6 +243,8 @@ void aio_notify(AioContext *ctx)
 AioContext *aio_context_new(void)
 {
     AioContext *ctx;
+    int i;
+
     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
     ctx->pollfds = g_array_new(FALSE, FALSE, sizeof(GPollFD));
     ctx->thread_pool = NULL;
@@ -248,6 +254,9 @@ AioContext *aio_context_new(void)
                            (EventNotifierHandler *)
                            event_notifier_test_and_clear, NULL);
     ctx->alarm_timer = alarm_timer_create(ctx);
+    for (i = 0; i < QEMU_CLOCK_MAXCNT; i++) {
+        timer_list_init(&ctx->timer_list[i]);
+    }
 
     return ctx;
 }
diff --git a/include/block/aio.h b/include/block/aio.h
index 84537a2..3f1a7b4 100644
--- a/include/block/aio.h
+++ b/include/block/aio.h
@@ -43,6 +43,18 @@ typedef struct AioHandler AioHandler;
 typedef void QEMUBHFunc(void *opaque);
 typedef void IOHandler(void *opaque);
 
+/* Related timer with AioContext */
+typedef struct QEMUTimer QEMUTimer;
+#define QEMU_CLOCK_MAXCNT 3
+
+typedef struct TimerList {
+    QEMUTimer *active_timers;
+    QemuMutex active_timers_lock;
+} TimerList;
+
+void timer_list_init(TimerList *tlist);
+void timer_list_finalize(TimerList *tlist);
+
 typedef struct AioContext {
     GSource source;
 
@@ -74,6 +86,7 @@ typedef struct AioContext {
     /* Thread pool for performing work and receiving completion callbacks */
     struct ThreadPool *thread_pool;
     struct qemu_alarm_timer *alarm_timer;
+    TimerList timer_list[QEMU_CLOCK_MAXCNT];
 } AioContext;
 
 /* Returns 1 if there are still outstanding AIO requests; 0 otherwise */
diff --git a/include/qemu/timer.h b/include/qemu/timer.h
index 4a72c99..a5acfe0 100644
--- a/include/qemu/timer.h
+++ b/include/qemu/timer.h
@@ -33,9 +33,14 @@ extern QEMUClock *vm_clock;
 extern QEMUClock *host_clock;
 
 int64_t qemu_get_clock_ns(QEMUClock *clock);
+/* qemu_clock_has_timers, qemu_clock_expired, qemu_clock_deadline
+ * run In tcg icount mode. There is only one AioContext i.e. qemu_aio_context.
+ * So we only count the timers on qemu_aio_context.
+ */
 int64_t qemu_clock_has_timers(QEMUClock *clock);
 int64_t qemu_clock_expired(QEMUClock *clock);
 int64_t qemu_clock_deadline(QEMUClock *clock);
+
 void qemu_clock_enable(QEMUClock *clock, bool enabled);
 void qemu_clock_warp(QEMUClock *clock);
 
@@ -45,6 +50,9 @@ void qemu_unregister_clock_reset_notifier(QEMUClock *clock,
 
 QEMUTimer *qemu_new_timer(QEMUClock *clock, int scale,
                           QEMUTimerCB *cb, void *opaque);
+QEMUTimer *aioctx_new_timer(QEMUClock *clock, int scale,
+                          QEMUTimerCB *cb, void *opaque, AioContext *ctx);
+
 void qemu_free_timer(QEMUTimer *ts);
 void qemu_del_timer(QEMUTimer *ts);
 void qemu_mod_timer_ns(QEMUTimer *ts, int64_t expire_time);
@@ -77,6 +85,18 @@ static inline QEMUTimer *qemu_new_timer_ms(QEMUClock *clock, QEMUTimerCB *cb,
     return qemu_new_timer(clock, SCALE_MS, cb, opaque);
 }
 
+static inline QEMUTimer *aioctx_new_timer_ns(QEMUClock *clock, QEMUTimerCB *cb,
+                                           void *opaque, AioContext *ctx)
+{
+    return aioctx_new_timer(clock, SCALE_NS, cb, opaque, ctx);
+}
+
+static inline QEMUTimer *aioctx_new_timer_ms(QEMUClock *clock, QEMUTimerCB *cb,
+                                           void *opaque, AioContext *ctx)
+{
+    return aioctx_new_timer(clock, SCALE_MS, cb, opaque, ctx);
+}
+
 static inline int64_t qemu_get_clock_ms(QEMUClock *clock)
 {
     return qemu_get_clock_ns(clock) / SCALE_MS;
diff --git a/qemu-timer.c b/qemu-timer.c
index 0ee68dc..1a0cbae 100644
--- a/qemu-timer.c
+++ b/qemu-timer.c
@@ -45,14 +45,6 @@
 #define QEMU_CLOCK_REALTIME 0
 #define QEMU_CLOCK_VIRTUAL  1
 #define QEMU_CLOCK_HOST     2
-#define QEMU_CLOCK_MAXCNT 3
-
-typedef struct TimerList {
-    QEMUTimer *active_timers;
-    QemuMutex active_timers_lock;
-} TimerList;
-
-static TimerList timer_list[QEMU_CLOCK_MAXCNT];
 
 struct QEMUClock {
     NotifierList reset_notifiers;
@@ -64,7 +56,9 @@ struct QEMUClock {
 
 struct QEMUTimer {
     int64_t expire_time;	/* in nanoseconds */
+    /* quick link to AioContext timer list */
     TimerList *list;
+    AioContext *ctx;
     QEMUTimerCB *cb;
     void *opaque;
     QEMUTimer *next;
@@ -128,11 +122,12 @@ void alarm_timer_destroy(struct qemu_alarm_timer *t)
     g_free(t);
 }
 
-static TimerList *clock_to_timerlist(QEMUClock *clock)
+static TimerList *clock_to_timerlist(QEMUClock *clock, AioContext *ctx)
 {
     int type = clock->type;
 
-    return  &timer_list[type];
+    assert(ctx);
+    return  &ctx->timer_list[type];
 }
 
 static bool qemu_timer_expired_ns(QEMUTimer *timer_head, int64_t current_time)
@@ -140,7 +135,8 @@ static bool qemu_timer_expired_ns(QEMUTimer *timer_head, int64_t current_time)
     return timer_head && (timer_head->expire_time <= current_time);
 }
 
-static int64_t qemu_next_clock_deadline(QEMUClock *clock, int64_t delta)
+static int64_t qemu_next_clock_deadline(QEMUClock *clock, int64_t delta,
+    AioContext *ctx)
 {
     int64_t expire_time, next;
     bool has_timer = false;
@@ -150,7 +146,7 @@ static int64_t qemu_next_clock_deadline(QEMUClock *clock, int64_t delta)
         return delta;
     }
 
-    tlist = clock_to_timerlist(clock);
+    tlist = clock_to_timerlist(clock, ctx);
     qemu_mutex_lock(&tlist->active_timers_lock);
     if (tlist->active_timers) {
         has_timer = true;
@@ -165,18 +161,15 @@ static int64_t qemu_next_clock_deadline(QEMUClock *clock, int64_t delta)
     return MIN(next, delta);
 }
 
-/* Soon this will be fixed: till now, timer list is not associated with
- * AioContext, so @ctx has no effect on deadline currently.
- */
 static int64_t qemu_next_alarm_deadline(AioContext *ctx)
 {
     int64_t delta = INT64_MAX;
 
     if (!use_icount) {
-        delta = qemu_next_clock_deadline(vm_clock, delta);
+        delta = qemu_next_clock_deadline(vm_clock, delta, ctx);
     }
-    delta = qemu_next_clock_deadline(host_clock, delta);
-    return qemu_next_clock_deadline(rt_clock, delta);
+    delta = qemu_next_clock_deadline(host_clock, delta, ctx);
+    return qemu_next_clock_deadline(rt_clock, delta, ctx);
 }
 
 static void qemu_rearm_alarm_timer(struct qemu_alarm_timer *t)
@@ -309,24 +302,27 @@ QEMUClock *rt_clock;
 QEMUClock *vm_clock;
 QEMUClock *host_clock;
 
-static void timer_list_init(TimerList *tlist)
+void timer_list_init(TimerList *tlist)
 {
     qemu_mutex_init(&tlist->active_timers_lock);
     tlist->active_timers = NULL;
 }
 
+void timer_list_finalize(TimerList *tlist)
+{
+    qemu_mutex_destroy(&tlist->active_timers_lock);
+    assert(!tlist->active_timers);
+}
+
 static QEMUClock *qemu_new_clock(int type)
 {
     QEMUClock *clock;
-    TimerList *tlist;
 
     clock = g_malloc0(sizeof(QEMUClock));
     clock->type = type;
     clock->enabled = true;
     clock->last = INT64_MIN;
     notifier_list_init(&clock->reset_notifiers);
-    tlist = clock_to_timerlist(clock);
-    timer_list_init(tlist);
 
     return clock;
 }
@@ -340,10 +336,14 @@ void qemu_clock_enable(QEMUClock *clock, bool enabled)
     }
 }
 
+/* qemu_clock_has_timers, qemu_clock_expired, qemu_clock_deadline
+  * run In tcg icount mode. There is only one AioContext i.e. qemu_aio_context.
+ * So we only count the timers on qemu_aio_context.
+*/
 int64_t qemu_clock_has_timers(QEMUClock *clock)
 {
     bool has_timers;
-    TimerList *tlist = clock_to_timerlist(clock);
+    TimerList *tlist = clock_to_timerlist(clock, qemu_get_aio_context());
 
     qemu_mutex_lock(&tlist->active_timers_lock);
     has_timers = !!tlist->active_timers;
@@ -355,7 +355,7 @@ int64_t qemu_clock_expired(QEMUClock *clock)
 {
     bool has_timers;
     int64_t expire_time;
-    TimerList *tlist = clock_to_timerlist(clock);
+    TimerList *tlist = clock_to_timerlist(clock, qemu_get_aio_context());
 
     qemu_mutex_lock(&tlist->active_timers_lock);
     has_timers = tlist->active_timers;
@@ -371,7 +371,7 @@ int64_t qemu_clock_deadline(QEMUClock *clock)
     int64_t delta = INT32_MAX;
     bool has_timers;
     int64_t expire_time;
-    TimerList *tlist = clock_to_timerlist(clock);
+    TimerList *tlist = clock_to_timerlist(clock, qemu_get_aio_context());
 
     qemu_mutex_lock(&tlist->active_timers_lock);
     has_timers = tlist->active_timers;
@@ -387,19 +387,26 @@ int64_t qemu_clock_deadline(QEMUClock *clock)
     return delta;
 }
 
-QEMUTimer *qemu_new_timer(QEMUClock *clock, int scale,
-                          QEMUTimerCB *cb, void *opaque)
+QEMUTimer *aioctx_new_timer(QEMUClock *clock, int scale,
+                          QEMUTimerCB *cb, void *opaque, AioContext *ctx)
 {
     QEMUTimer *ts;
 
     ts = g_malloc0(sizeof(QEMUTimer));
-    ts->list = clock_to_timerlist(clock);
+    ts->list = clock_to_timerlist(clock, ctx);
     ts->cb = cb;
     ts->opaque = opaque;
     ts->scale = scale;
+    ts->ctx = ctx;
     return ts;
 }
 
+QEMUTimer *qemu_new_timer(QEMUClock *clock, int scale,
+                          QEMUTimerCB *cb, void *opaque)
+{
+    return aioctx_new_timer(clock, scale, cb, opaque, qemu_get_aio_context());
+}
+
 void qemu_free_timer(QEMUTimer *ts)
 {
     g_free(ts);
@@ -491,12 +498,14 @@ void qemu_run_timers(QEMUClock *clock)
     QEMUTimer *ts;
     int64_t current_time;
     TimerList *tlist;
+    AioContext *ctx;
 
     if (!clock->enabled)
         return;
 
     current_time = qemu_get_clock_ns(clock);
-    tlist = clock_to_timerlist(clock);
+    ctx = *tls_get_thread_aio_context();
+    tlist = clock_to_timerlist(clock, ctx);
 
     for(;;) {
         qemu_mutex_lock(&tlist->active_timers_lock);
-- 
1.8.1.4