From: "Emilio G. Cota" <cota@braap.org>

Add a comment about how atomic_read works here. The comment refers to
a "BQL-less CPU loop", which will materialize toward the end
of this series.

Note that the modifications to cpu_reset_interrupt are there to
avoid deadlock during the CPU lock transition; once that is complete,
cpu_interrupt_request will be simple again.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Robert Foley <robert.foley@linaro.org>
---
 hw/core/cpu.c         | 27 +++++++++++++++++++++------
 include/hw/core/cpu.h | 37 +++++++++++++++++++++++++++++++++++++
 2 files changed, 58 insertions(+), 6 deletions(-)

diff --git a/hw/core/cpu.c b/hw/core/cpu.c
index 9683d6408a..01091d0623 100644
--- a/hw/core/cpu.c
+++ b/hw/core/cpu.c
@@ -99,14 +99,29 @@ static void cpu_common_get_memory_mapping(CPUState *cpu,
  * BQL here if we need to.  cpu_interrupt assumes it is held.*/
 void cpu_reset_interrupt(CPUState *cpu, int mask)
 {
-    bool need_lock = !qemu_mutex_iothread_locked();
+    bool has_bql = qemu_mutex_iothread_locked();
+    bool has_cpu_lock = cpu_mutex_locked(cpu);
 
-    if (need_lock) {
-        qemu_mutex_lock_iothread();
+    if (has_bql) {
+        if (has_cpu_lock) {
+            atomic_set(&cpu->interrupt_request, cpu->interrupt_request & ~mask);
+        } else {
+            cpu_mutex_lock(cpu);
+            atomic_set(&cpu->interrupt_request, cpu->interrupt_request & ~mask);
+            cpu_mutex_unlock(cpu);
+        }
+        return;
+    }
+
+    if (has_cpu_lock) {
+        cpu_mutex_unlock(cpu);
     }
-    cpu->interrupt_request &= ~mask;
-    if (need_lock) {
-        qemu_mutex_unlock_iothread();
+    qemu_mutex_lock_iothread();
+    cpu_mutex_lock(cpu);
+    atomic_set(&cpu->interrupt_request, cpu->interrupt_request & ~mask);
+    qemu_mutex_unlock_iothread();
+    if (!has_cpu_lock) {
+        cpu_mutex_unlock(cpu);
     }
 }
 
diff --git a/include/hw/core/cpu.h b/include/hw/core/cpu.h
index 48b46c90ed..1c3701b142 100644
--- a/include/hw/core/cpu.h
+++ b/include/hw/core/cpu.h
@@ -522,6 +522,43 @@ static inline void cpu_halted_set(CPUState *cpu, uint32_t val)
     cpu_mutex_unlock(cpu);
 }
 
+/*
+ * When sending an interrupt, setters OR the appropriate bit and kick the
+ * destination vCPU. The latter can then read interrupt_request without
+ * acquiring the CPU lock, because once the kick-induced completes, they'll read
+ * an up-to-date interrupt_request.
+ * Setters always acquire the lock, which guarantees that (1) concurrent
+ * updates from different threads won't result in data races, and (2) the
+ * BQL-less CPU loop will always see an up-to-date interrupt_request, since the
+ * loop holds the CPU lock.
+ */
+static inline uint32_t cpu_interrupt_request(CPUState *cpu)
+{
+    return atomic_read(&cpu->interrupt_request);
+}
+
+static inline void cpu_interrupt_request_or(CPUState *cpu, uint32_t mask)
+{
+    if (cpu_mutex_locked(cpu)) {
+        atomic_set(&cpu->interrupt_request, cpu->interrupt_request | mask);
+        return;
+    }
+    cpu_mutex_lock(cpu);
+    atomic_set(&cpu->interrupt_request, cpu->interrupt_request | mask);
+    cpu_mutex_unlock(cpu);
+}
+
+static inline void cpu_interrupt_request_set(CPUState *cpu, uint32_t val)
+{
+    if (cpu_mutex_locked(cpu)) {
+        atomic_set(&cpu->interrupt_request, val);
+        return;
+    }
+    cpu_mutex_lock(cpu);
+    atomic_set(&cpu->interrupt_request, val);
+    cpu_mutex_unlock(cpu);
+}
+
 static inline void cpu_tb_jmp_cache_clear(CPUState *cpu)
 {
     unsigned int i;
-- 
2.17.1