accel/tcg/tcg-accel-ops-rr: ensure fairness with icount

The round-robin scheduler will iterate over the CPU list with an
assigned budget until the next timer expiry and may exit early because
of a TB exit.  This is fine under normal operation but with icount
enabled and SMP it is possible for a CPU to be starved of run time and
the system live-locks.

For example, booting a riscv64 platform with '-icount
shift=0,align=off,sleep=on -smp 2' we observe a livelock once the kernel
has timers enabled and starts performing TLB shootdowns.  In this case
we have CPU 0 in M-mode with interrupts disabled sending an IPI to CPU
1.  As we enter the TCG loop, we assign the icount budget to next timer
interrupt to CPU 0 and begin executing where the guest is sat in a busy
loop exhausting all of the budget before we try to execute CPU 1 which
is the target of the IPI but CPU 1 is left with no budget with which to
execute and the process repeats.

We try here to add some fairness by splitting the budget across all of
the CPUs on the thread fairly before entering each one.  The CPU count
is cached on CPU list generation ID to avoid iterating the list on each
loop iteration.  With this change it is possible to boot an SMP rv64
guest with icount enabled and no hangs.

Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Tested-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Jamie Iles <quic_jiles@quicinc.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20230427020925.51003-3-quic_jiles@quicinc.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

diff --git a/accel/tcg/tcg-accel-ops-icount.c b/accel/tcg/tcg-accel-ops-icount.c
index 84cc7421be88fa42132264da7329d4ee9882f7e1..3d2cfbbc97786368825a71958c7df090f256f146 100644 (file)
--- a/accel/tcg/tcg-accel-ops-icount.c
+++ b/accel/tcg/tcg-accel-ops-icount.c
@@ -89,7 +89,20 @@ void icount_handle_deadline(void)
     }
 }
 
-void icount_prepare_for_run(CPUState *cpu)
+/* Distribute the budget evenly across all CPUs */
+int64_t icount_percpu_budget(int cpu_count)
+{
+    int64_t limit = icount_get_limit();
+    int64_t timeslice = limit / cpu_count;
+
+    if (timeslice == 0) {
+        timeslice = limit;
+    }
+
+    return timeslice;
+}
+
+void icount_prepare_for_run(CPUState *cpu, int64_t cpu_budget)
 {
     int insns_left;
 
@@ -101,13 +114,13 @@ void icount_prepare_for_run(CPUState *cpu)
     g_assert(cpu_neg(cpu)->icount_decr.u16.low == 0);
     g_assert(cpu->icount_extra == 0);
 
-    cpu->icount_budget = icount_get_limit();
+    replay_mutex_lock();
+
+    cpu->icount_budget = MIN(icount_get_limit(), cpu_budget);
     insns_left = MIN(0xffff, cpu->icount_budget);
     cpu_neg(cpu)->icount_decr.u16.low = insns_left;
     cpu->icount_extra = cpu->icount_budget - insns_left;
 
-    replay_mutex_lock();
-
     if (cpu->icount_budget == 0) {
         /*
          * We're called without the iothread lock, so must take it while

diff --git a/accel/tcg/tcg-accel-ops-icount.h b/accel/tcg/tcg-accel-ops-icount.h
index 1b6fd9c607513717421870f9519f0f305e623f4d..16a301b6dc0ba515ecbc4073b28968f0a0604d14 100644 (file)
--- a/accel/tcg/tcg-accel-ops-icount.h
+++ b/accel/tcg/tcg-accel-ops-icount.h
@@ -11,7 +11,8 @@
 #define TCG_ACCEL_OPS_ICOUNT_H
 
 void icount_handle_deadline(void);
-void icount_prepare_for_run(CPUState *cpu);
+void icount_prepare_for_run(CPUState *cpu, int64_t cpu_budget);
+int64_t icount_percpu_budget(int cpu_count);
 void icount_process_data(CPUState *cpu);
 
 void icount_handle_interrupt(CPUState *cpu, int mask);

diff --git a/accel/tcg/tcg-accel-ops-rr.c b/accel/tcg/tcg-accel-ops-rr.c
index 290833a37fb28ee651a8d514cd867e3364c68b01..5788efa5ff4d1b8de2e365b6a23abbe1e46db217 100644 (file)
--- a/accel/tcg/tcg-accel-ops-rr.c
+++ b/accel/tcg/tcg-accel-ops-rr.c
@@ -24,6 +24,7 @@
  */
 
 #include "qemu/osdep.h"
+#include "qemu/lockable.h"
 #include "sysemu/tcg.h"
 #include "sysemu/replay.h"
 #include "sysemu/cpu-timers.h"
@@ -139,6 +140,33 @@ static void rr_force_rcu(Notifier *notify, void *data)
     rr_kick_next_cpu();
 }
 
+/*
+ * Calculate the number of CPUs that we will process in a single iteration of
+ * the main CPU thread loop so that we can fairly distribute the instruction
+ * count across CPUs.
+ *
+ * The CPU count is cached based on the CPU list generation ID to avoid
+ * iterating the list every time.
+ */
+static int rr_cpu_count(void)
+{
+    static unsigned int last_gen_id = ~0;
+    static int cpu_count;
+    CPUState *cpu;
+
+    QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
+
+    if (cpu_list_generation_id_get() != last_gen_id) {
+        cpu_count = 0;
+        CPU_FOREACH(cpu) {
+            ++cpu_count;
+        }
+        last_gen_id = cpu_list_generation_id_get();
+    }
+
+    return cpu_count;
+}
+
 /*
  * In the single-threaded case each vCPU is simulated in turn. If
  * there is more than a single vCPU we create a simple timer to kick
@@ -185,11 +213,16 @@ static void *rr_cpu_thread_fn(void *arg)
     cpu->exit_request = 1;
 
     while (1) {
+        /* Only used for icount_enabled() */
+        int64_t cpu_budget = 0;
+
         qemu_mutex_unlock_iothread();
         replay_mutex_lock();
         qemu_mutex_lock_iothread();
 
         if (icount_enabled()) {
+            int cpu_count = rr_cpu_count();
+
             /* Account partial waits to QEMU_CLOCK_VIRTUAL.  */
             icount_account_warp_timer();
             /*
@@ -197,6 +230,8 @@ static void *rr_cpu_thread_fn(void *arg)
              * waking up the I/O thread and waiting for completion.
              */
             icount_handle_deadline();
+
+            cpu_budget = icount_percpu_budget(cpu_count);
         }
 
         replay_mutex_unlock();
@@ -218,7 +253,7 @@ static void *rr_cpu_thread_fn(void *arg)
 
                 qemu_mutex_unlock_iothread();
                 if (icount_enabled()) {
-                    icount_prepare_for_run(cpu);
+                    icount_prepare_for_run(cpu, cpu_budget);
                 }
                 r = tcg_cpus_exec(cpu);
                 if (icount_enabled()) {

diff --git a/replay/replay.c b/replay/replay.c
index c39156c52221bd3fd0a4e96abb4d5f7ecc64d80e..0f7d766efe819b021403660f77d6eac347c71712 100644 (file)
--- a/replay/replay.c
+++ b/replay/replay.c
@@ -74,7 +74,7 @@ uint64_t replay_get_current_icount(void)
 int replay_get_instructions(void)
 {
     int res = 0;
-    replay_mutex_lock();
+    g_assert(replay_mutex_locked());
     if (replay_next_event_is(EVENT_INSTRUCTION)) {
         res = replay_state.instruction_count;
         if (replay_break_icount != -1LL) {
@@ -85,7 +85,6 @@ int replay_get_instructions(void)
             }
         }
     }
-    replay_mutex_unlock();
     return res;
 }