cpu-exec: simplify jump cache management

Unless I'm missing something egregious, the jmp cache is only every
populated with a valid entry by the same thread that reads the cache.
Therefore, the contents of any valid entry are always consistent and
there is no need for any acquire/release magic.

Indeed ->tb has to be accessed with atomics, because concurrent
invalidations would otherwise cause data races.  But ->pc is only ever
accessed by one thread, and accesses to ->tb and ->pc within tb_lookup
can never race with another tb_lookup.  While the TranslationBlock
(especially the flags) could be modified by a concurrent invalidation,
store-release and load-acquire operations on the cache entry would
not add any additional ordering beyond what you get from performing
the accesses within a single thread.

Because of this, there is really nothing to win in splitting the CF_PCREL
and !CF_PCREL paths.  It is easier to just always use the ->pc field in
the jump cache.

I noticed this while working on splitting commit 8ed558ec0cb
("accel/tcg: Introduce TARGET_TB_PCREL", 2022-10-04) into multiple
pieces, for the sake of finding a more fine-grained bisection
result for https://gitlab.com/qemu-project/qemu/-/issues/2092.
It does not (and does not intend to) fix that issue; therefore
it may make sense to not commit it until the root cause
of issue #2092 is found.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20240122153409.351959-1-pbonzini@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

diff --git a/accel/tcg/cpu-exec.c b/accel/tcg/cpu-exec.c
index 67eda9865ee2a89f7d5a4f3a4142efb1a869d4fd..40c268bfa1f3060eda62061b3263ab3c4af3abda 100644 (file)
--- a/accel/tcg/cpu-exec.c
+++ b/accel/tcg/cpu-exec.c
@@ -253,43 +253,29 @@ static inline TranslationBlock *tb_lookup(CPUState *cpu, vaddr pc,
     hash = tb_jmp_cache_hash_func(pc);
     jc = cpu->tb_jmp_cache;
 
-    if (cflags & CF_PCREL) {
-        /* Use acquire to ensure current load of pc from jc. */
-        tb = qatomic_load_acquire(&jc->array[hash].tb);
-
-        if (likely(tb &&
-                   jc->array[hash].pc == pc &&
-                   tb->cs_base == cs_base &&
-                   tb->flags == flags &&
-                   tb_cflags(tb) == cflags)) {
-            return tb;
-        }
-        tb = tb_htable_lookup(cpu, pc, cs_base, flags, cflags);
-        if (tb == NULL) {
-            return NULL;
-        }
-        jc->array[hash].pc = pc;
-        /* Ensure pc is written first. */
-        qatomic_store_release(&jc->array[hash].tb, tb);
-    } else {
-        /* Use rcu_read to ensure current load of pc from *tb. */
-        tb = qatomic_rcu_read(&jc->array[hash].tb);
-
-        if (likely(tb &&
-                   tb->pc == pc &&
-                   tb->cs_base == cs_base &&
-                   tb->flags == flags &&
-                   tb_cflags(tb) == cflags)) {
-            return tb;
-        }
-        tb = tb_htable_lookup(cpu, pc, cs_base, flags, cflags);
-        if (tb == NULL) {
-            return NULL;
-        }
-        /* Use the pc value already stored in tb->pc. */
-        qatomic_set(&jc->array[hash].tb, tb);
+    tb = qatomic_read(&jc->array[hash].tb);
+    if (likely(tb &&
+               jc->array[hash].pc == pc &&
+               tb->cs_base == cs_base &&
+               tb->flags == flags &&
+               tb_cflags(tb) == cflags)) {
+        goto hit;
     }
 
+    tb = tb_htable_lookup(cpu, pc, cs_base, flags, cflags);
+    if (tb == NULL) {
+        return NULL;
+    }
+
+    jc->array[hash].pc = pc;
+    qatomic_set(&jc->array[hash].tb, tb);
+
+hit:
+    /*
+     * As long as tb is not NULL, the contents are consistent.  Therefore,
+     * the virtual PC has to match for non-CF_PCREL translations.
+     */
+    assert((tb_cflags(tb) & CF_PCREL) || tb->pc == pc);
     return tb;
 }
 
@@ -1012,14 +998,8 @@ cpu_exec_loop(CPUState *cpu, SyncClocks *sc)
                  */
                 h = tb_jmp_cache_hash_func(pc);
                 jc = cpu->tb_jmp_cache;
-                if (cflags & CF_PCREL) {
-                    jc->array[h].pc = pc;
-                    /* Ensure pc is written first. */
-                    qatomic_store_release(&jc->array[h].tb, tb);
-                } else {
-                    /* Use the pc value already stored in tb->pc. */
-                    qatomic_set(&jc->array[h].tb, tb);
-                }
+                jc->array[h].pc = pc;
+                qatomic_set(&jc->array[h].tb, tb);
             }
 
 #ifndef CONFIG_USER_ONLY

diff --git a/accel/tcg/tb-jmp-cache.h b/accel/tcg/tb-jmp-cache.h
index bb424c8a05b75c47e99ffe9bbe5c849725b6a549..4ab8553afcc02ff2cc754d1ff545bda9b29830f9 100644 (file)
--- a/accel/tcg/tb-jmp-cache.h
+++ b/accel/tcg/tb-jmp-cache.h
@@ -13,9 +13,11 @@
 #define TB_JMP_CACHE_SIZE (1 << TB_JMP_CACHE_BITS)
 
 /*
- * Accessed in parallel; all accesses to 'tb' must be atomic.
- * For CF_PCREL, accesses to 'pc' must be protected by a
- * load_acquire/store_release to 'tb'.
+ * Invalidated in parallel; all accesses to 'tb' must be atomic.
+ * A valid entry is read/written by a single CPU, therefore there is
+ * no need for qatomic_rcu_read() and pc is always consistent with a
+ * non-NULL value of 'tb'.  Strictly speaking pc is only needed for
+ * CF_PCREL, but it's used always for simplicity.
  */
 struct CPUJumpCache {
     struct rcu_head rcu;