By holding off updates to timer_state.qemu_icount we can run into
trouble when the non-vCPU thread needs to know the time. This helper
ensures we atomically update timers_state.qemu_icount based on what
has been currently executed.

Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
---
 cpus.c               | 16 ++++++++++++++--
 include/qemu/timer.h |  1 +
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/cpus.c b/cpus.c
index 88eabdc19f..71c3baba87 100644
--- a/cpus.c
+++ b/cpus.c
@@ -232,12 +232,24 @@ static int64_t cpu_get_icount_executed(CPUState *cpu)
     return cpu->icount_budget - (cpu->icount_decr.u16.low + cpu->icount_extra);
 }
 
+/*
+ * Update the global shared timer_state.qemu_icount to take into
+ * account executed instructions. This is done by the TCG vCPU
+ * thread so the main-loop can see time has moved forward.
+ */
+void cpu_update_icount(CPUState *cpu)
+{
+    int64_t executed = cpu_get_icount_executed(cpu);
+    cpu->icount_budget -= executed;
+    atomic_add(&timers_state.qemu_icount, executed);
+}
+
 int64_t cpu_get_icount_raw(void)
 {
     int64_t icount;
     CPUState *cpu = current_cpu;
 
-    icount = timers_state.qemu_icount;
+    icount = atomic_read(&timers_state.qemu_icount);
     if (cpu && cpu->running) {
         if (!cpu->can_do_io) {
             fprintf(stderr, "Bad icount read\n");
@@ -1220,7 +1232,7 @@ static void process_icount_data(CPUState *cpu)
 {
     if (use_icount) {
         /* Account for executed instructions */
-        timers_state.qemu_icount += cpu_get_icount_executed(cpu);
+        cpu_update_icount(cpu);
 
         /* Reset the counters */
         cpu->icount_decr.u16.low = 0;
diff --git a/include/qemu/timer.h b/include/qemu/timer.h
index e1742f2f3d..8a1eb74839 100644
--- a/include/qemu/timer.h
+++ b/include/qemu/timer.h
@@ -869,6 +869,7 @@ int64_t cpu_get_icount_raw(void);
 int64_t cpu_get_icount(void);
 int64_t cpu_get_clock(void);
 int64_t cpu_icount_to_ns(int64_t icount);
+void    cpu_update_icount(CPUState *cpu);
 
 /*******************************************/
 /* host CPU ticks (if available) */
-- 
2.11.0

On 05/04/2017 15:25, Alex Bennée wrote:
> +{
> +    int64_t executed = cpu_get_icount_executed(cpu);
> +    cpu->icount_budget -= executed;
> +    atomic_add(&timers_state.qemu_icount, executed);
> +}

Since there's only one writer, it's also okay to do

	atomic_set(&timers_state.qemu_icount,
		   atomic_read(&timers_state.qemu_icount) + executed);

(also not just faster, but also simpler to turn into TCG code if we do
the update in gen_io_start).

Paolo

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 05/04/2017 15:25, Alex Bennée wrote:
>> +{
>> +    int64_t executed = cpu_get_icount_executed(cpu);
>> +    cpu->icount_budget -= executed;
>> +    atomic_add(&timers_state.qemu_icount, executed);
>> +}
>
> Since there's only one writer, it's also okay to do
>
> 	atomic_set(&timers_state.qemu_icount,
> 		   atomic_read(&timers_state.qemu_icount) + executed);
>
> (also not just faster, but also simpler to turn into TCG code if we do
> the update in gen_io_start).

OK fair enough. Annoyingly I've just noticed this breaks for 32 bit
hosts because we might not have CONFIG_ATOMIC64. What's the best
approach? #ifdef and handwave?

--
Alex Bennée

On 05/04/2017 16:34, Alex Bennée wrote:
>> On 05/04/2017 15:25, Alex Bennée wrote:
>>> +{
>>> +    int64_t executed = cpu_get_icount_executed(cpu);
>>> +    cpu->icount_budget -= executed;
>>> +    atomic_add(&timers_state.qemu_icount, executed);
>>> +}
>> Since there's only one writer, it's also okay to do
>>
>> 	atomic_set(&timers_state.qemu_icount,
>> 		   atomic_read(&timers_state.qemu_icount) + executed);
>>
>> (also not just faster, but also simpler to turn into TCG code if we do
>> the update in gen_io_start).
>
> OK fair enough. Annoyingly I've just noticed this breaks for 32 bit
> hosts because we might not have CONFIG_ATOMIC64. What's the best
> approach? #ifdef and handwave?

For 2.9 yes... For 2.10 I'm going to add a "Stat64" type that can
compute max or add of 63-bit values, and blocks on the read side if a
doubleword change is in progress.

Paolo