qemu/0001-Handle-CPU-interrupts-by-inline-checking-of-a-flag.patch

From 4cc09e9530fb08123594be6c72dfc381df5dcddc Mon Sep 17 00:00:00 2001
From: Peter Maydell <peter.maydell@linaro.org>
Date: Wed, 5 Oct 2011 10:04:02 +0100
Subject: [PATCH] Handle CPU interrupts by inline checking of a flag

Fix the nasty TCG race conditions and crashes by implementing cpu_exit
as setting a flag which is checked at the start of each TB. This is
slightly slower than the attempt to have cpu_exit alter the graph of
TBs, but it doesn't crash if a thread or signal handler calls cpu_exit
while the execution thread is itself modifying the TB graph.

This version of the patch includes command line option "-no-stopflag"
which reverts to the previous racy behaviour. This is intended for
convenience in testing and comparative benchmarking and won't be
in the final patch.

It's probably worth experimenting with whether the flag-testing
code has the branch in a sense which confuses branch-prediction
and thus whether flipping it might change performance.

Mostly this needs benchmarking to determine what the actual speed
hit is, which I never got round to. Feel free to do some :-)
---
 cpu-exec.c        |   11 ++++++++++-
 exec.c            |   14 ++++++++++++--
 gen-icount.h      |   16 ++++++++++++++++
 linux-user/main.c |    8 ++++++++
 qemu-options.hx   |    9 +++++++++
 vl.c              |    5 +++++
 6 files changed, 60 insertions(+), 3 deletions(-)

diff --git a/cpu-exec.c b/cpu-exec.c
index a9fa608..5f7982f 100644
--- a/cpu-exec.c
+++ b/cpu-exec.c
@@ -564,7 +564,16 @@ int cpu_exec(CPUState *env)
                     tc_ptr = tb->tc_ptr;
                 /* execute the generated code */
                     next_tb = tcg_qemu_tb_exec(env, tc_ptr);
-                    if ((next_tb & 3) == 2) {
+                    if ((next_tb & 3) == 3) {
+                        /* hit stopflag check */
+                        tb = (TranslationBlock *)(long)(next_tb & ~3);
+                        /* Restore PC.  */
+                        cpu_pc_from_tb(env, tb);
+                        next_tb = 0;
+                        env->exit_request = 0;
+                        env->exception_index = EXCP_INTERRUPT;
+                        cpu_loop_exit(env);
+                    } else if ((next_tb & 3) == 2) {
                         /* Instruction counter expired.  */
                         int insns_left;
                         tb = (TranslationBlock *)(long)(next_tb & ~3);
diff --git a/exec.c b/exec.c
index 6b92198..6c923f2 100644
--- a/exec.c
+++ b/exec.c
@@ -125,6 +125,8 @@ DEFINE_TLS(CPUState *,cpu_single_env);
    1 = Precise instruction counting.
    2 = Adaptive rate instruction counting.  */
 int use_icount = 0;
+/* 1 to do cpu_exit by inline flag check rather than tb link breaking */
+int use_stopflag = 1;
 
 typedef struct PageDesc {
     /* list of TBs intersecting this ram page */
@@ -1670,7 +1672,13 @@ static void tcg_handle_interrupt(CPUState *env, int mask)
             cpu_abort(env, "Raised interrupt while not in I/O function");
         }
     } else {
-        cpu_unlink_tb(env);
+        // XXX just call cpu_exit ?
+        if (use_stopflag) {
+            // XXX is this OK?
+            env->exit_request = 1;
+        } else {
+            cpu_unlink_tb(env);
+        }
     }
 }
 
@@ -1693,7 +1701,9 @@ void cpu_reset_interrupt(CPUState *env, int mask)
 void cpu_exit(CPUState *env)
 {
     env->exit_request = 1;
-    cpu_unlink_tb(env);
+    if (!use_stopflag) {
+        cpu_unlink_tb(env);
+    }
 }
 
 const CPULogItem cpu_log_items[] = {
diff --git a/gen-icount.h b/gen-icount.h
index 5fb3829..060f814 100644
--- a/gen-icount.h
+++ b/gen-icount.h
@@ -2,13 +2,25 @@
 
 /* Helpers for instruction counting code generation.  */
 
+extern int use_stopflag;
+
 static TCGArg *icount_arg;
 static int icount_label;
+static int stopflag_label;
 
 static inline void gen_icount_start(void)
 {
     TCGv_i32 count;
 
+    if (use_stopflag) {
+        TCGv_i32 flag;
+        stopflag_label = gen_new_label();
+        flag = tcg_temp_local_new_i32();
+        tcg_gen_ld_i32(flag, cpu_env, offsetof(CPUState, exit_request));
+        tcg_gen_brcondi_i32(TCG_COND_NE, flag, 0, stopflag_label);
+        tcg_temp_free_i32(flag);
+    }
+
     if (!use_icount)
         return;
 
@@ -26,6 +38,10 @@ static inline void gen_icount_start(void)
 
 static void gen_icount_end(TranslationBlock *tb, int num_insns)
 {
+    if (use_stopflag) {
+        gen_set_label(stopflag_label);
+        tcg_gen_exit_tb((long)tb + 3); // XXX
+    }
     if (use_icount) {
         *icount_arg = num_insns;
         gen_set_label(icount_label);
diff --git a/linux-user/main.c b/linux-user/main.c
index d1bbc57..1cd8eb7 100644
--- a/linux-user/main.c
+++ b/linux-user/main.c
@@ -52,6 +52,7 @@ unsigned long reserved_va;
 #endif
 
 static void usage(void);
+extern int use_stopflag;
 
 static const char *interp_prefix = CONFIG_QEMU_INTERP_PREFIX;
 const char *qemu_uname_release = CONFIG_UNAME_RELEASE;
@@ -3072,6 +3073,11 @@ static void handle_arg_reserved_va(const char *arg)
 }
 #endif
 
+static void handle_arg_nostopflag(const char *arg)
+{
+    use_stopflag = 0;
+}
+
 static void handle_arg_singlestep(const char *arg)
 {
     singlestep = 1;
@@ -3125,6 +3131,8 @@ struct qemu_argument arg_table[] = {
 #endif
     {"d",          "QEMU_LOG",         true,  handle_arg_log,
      "options",    "activate log"},
+    {"no-stopflag", "QEMU_NOSTOPFLAG", false, handle_arg_nostopflag,
+     "",           "run in singlestep mode"},
     {"p",          "QEMU_PAGESIZE",    true,  handle_arg_pagesize,
      "pagesize",   "set the host page size to 'pagesize'"},
     {"singlestep", "QEMU_SINGLESTEP",  false, handle_arg_singlestep,
diff --git a/qemu-options.hx b/qemu-options.hx
index 681eaf1..83b1f38 100644
--- a/qemu-options.hx
+++ b/qemu-options.hx
@@ -1112,6 +1112,15 @@ STEXI
 Disable HPET support.
 ETEXI
 
+DEF("no-stopflag", 0, QEMU_OPTION_no_stopflag,
+    "-no-stopflag   use old behaviour, not inline stopflag checks\n", QEMU_ARCH_ALL)
+STEXI
+@item -no-stopflag
+@findex -no-stopflag
+Implement cpu-exit by the old tb link breaking method rather than inline checks
+(this is slightly faster but racy!)
+ETEXI
+
 DEF("balloon", HAS_ARG, QEMU_OPTION_balloon,
     "-balloon none   disable balloon device\n"
     "-balloon virtio[,addr=str]\n"
diff --git a/vl.c b/vl.c
index a50842b..7fdd80f 100644
--- a/vl.c
+++ b/vl.c
@@ -174,6 +174,8 @@ int main(int argc, char **argv)
 
 #define MAX_VIRTIO_CONSOLES 1
 
+extern int use_stopflag;
+
 static const char *data_dir;
 const char *bios_name = NULL;
 enum vga_retrace_method vga_retrace_method = VGA_RETRACE_DUMB;
@@ -2819,6 +2821,9 @@ int main(int argc, char **argv, char **envp)
             case QEMU_OPTION_rtc_td_hack:
                 rtc_td_hack = 1;
                 break;
+            case QEMU_OPTION_no_stopflag:
+                use_stopflag = 0;
+                break;
             case QEMU_OPTION_acpitable:
                 do_acpitable_option(optarg);
                 break;
-- 
1.6.0.2