Support 4-byte atomic instruction fetch when instruction is natural
aligned.

Current implementation is not atomic because it loads instruction twice
for first and last 2 bytes. We load 4 bytes at once to keep the
atomicity. This instruction preload method only applys when instruction
is 4-byte aligned. If instruction is unaligned, it could be across pages
so that preload will trigger additional page fault.

We encounter this issue when doing pressure test of enabling & disabling
Linux kernel ftrace. Ftrace with kernel preemption requires concurrent
modification and execution of instruction, so non-atomic instruction
fetch will cause the race condition. We may fetch the wrong instruction
which is the mixing of 2 instructions.

Also, RISC-V Profile wants to provide this feature by HW. RVA20U64
Ziccif protects the atomicity of instruction fetch when it is
natural aligned.

Signed-off-by: Jim Shu <jim.shu@sifive.com>
Reviewed-by: Frank Chang <frank.chang@sifive.com>
---
 target/riscv/translate.c | 45 ++++++++++++++++++++++++++++++----------
 1 file changed, 34 insertions(+), 11 deletions(-)

diff --git a/target/riscv/translate.c b/target/riscv/translate.c
index 0569224e53..2be8ef63e6 100644
--- a/target/riscv/translate.c
+++ b/target/riscv/translate.c
@@ -1133,13 +1133,37 @@ const RISCVDecoder decoder_table[] = {
 
 const size_t decoder_table_size = ARRAY_SIZE(decoder_table);
 
-static void decode_opc(CPURISCVState *env, DisasContext *ctx, uint16_t opcode)
+static void decode_opc(CPURISCVState *env, DisasContext *ctx)
 {
     ctx->virt_inst_excp = false;
+
+    uint32_t opcode;
+    bool is_4byte_align = false;
+
+    if ((ctx->base.pc_next % 4) == 0) {
+        /*
+         * Load 4 bytes at once to make instruction fetch atomically.
+         *
+         * Note: When pc is 4-byte aligned, 4-byte instruction wouldn't be
+         * across pages. We could preload 4 bytes instruction no matter
+         * real one is 2 or 4 bytes. Instruction preload wouldn't trigger
+         * additional page fault.
+         */
+        opcode = translator_ldl(env, &ctx->base, ctx->base.pc_next);
+        is_4byte_align = true;
+    } else {
+        /*
+         * For unaligned pc, instruction preload may trigger additional
+         * page fault so we only load 2 bytes here.
+         */
+        opcode = (uint32_t) translator_lduw(env, &ctx->base, ctx->base.pc_next);
+    }
+    ctx->ol = ctx->xl;
+
     ctx->cur_insn_len = insn_len(opcode);
     /* Check for compressed insn */
     if (ctx->cur_insn_len == 2) {
-        ctx->opcode = opcode;
+        ctx->opcode = (uint16_t)opcode;
         /*
          * The Zca extension is added as way to refer to instructions in the C
          * extension that do not include the floating-point loads and stores
@@ -1149,15 +1173,16 @@ static void decode_opc(CPURISCVState *env, DisasContext *ctx, uint16_t opcode)
             return;
         }
     } else {
-        uint32_t opcode32 = opcode;
-        opcode32 = deposit32(opcode32, 16, 16,
-                             translator_lduw(env, &ctx->base,
-                                             ctx->base.pc_next + 2));
-        ctx->opcode = opcode32;
+        if (!is_4byte_align) {
+            /* Load last 2 bytes of instruction here */
+            opcode = deposit32(opcode, 16, 16,
+                               translator_lduw(env, &ctx->base,
+                                               ctx->base.pc_next + 2));
+        }
 
         for (guint i = 0; i < ctx->decoders->len; ++i) {
             riscv_cpu_decode_fn func = g_ptr_array_index(ctx->decoders, i);
-            if (func(ctx, opcode32)) {
+            if (func(ctx, opcode)) {
                 return;
             }
         }
@@ -1226,10 +1251,8 @@ static void riscv_tr_translate_insn(DisasContextBase *dcbase, CPUState *cpu)
 {
     DisasContext *ctx = container_of(dcbase, DisasContext, base);
     CPURISCVState *env = cpu_env(cpu);
-    uint16_t opcode16 = translator_lduw(env, &ctx->base, ctx->base.pc_next);
 
-    ctx->ol = ctx->xl;
-    decode_opc(env, ctx, opcode16);
+    decode_opc(env, ctx);
     ctx->base.pc_next += ctx->cur_insn_len;
 
     /* Only the first insn within a TB is allowed to cross a page boundary. */
-- 
2.17.1

On 6/7/24 03:14, Jim Shu wrote:
> Support 4-byte atomic instruction fetch when instruction is natural
> aligned.
> 
> Current implementation is not atomic because it loads instruction twice
> for first and last 2 bytes. We load 4 bytes at once to keep the
> atomicity. This instruction preload method only applys when instruction
> is 4-byte aligned. If instruction is unaligned, it could be across pages
> so that preload will trigger additional page fault.
> 
> We encounter this issue when doing pressure test of enabling & disabling
> Linux kernel ftrace. Ftrace with kernel preemption requires concurrent
> modification and execution of instruction, so non-atomic instruction
> fetch will cause the race condition. We may fetch the wrong instruction
> which is the mixing of 2 instructions.

Hmm.  This is insufficient, as you have no guarantee that translator_ldl itself is atomic. 
  Indeed, it is not: it uses memcpy as the innermost read.

Let me think about this...


r~