target/arm: Complete TBI clearing for user-only for SVE

There are a number of paths by which the TBI is still intact
for user-only in the SVE helpers.

Because we currently always set TBI for user-only, we do not
need to pass down the actual TBI setting from above, and we
can remove the top byte in the inner-most primitives, so that
none are forgotten.  Moreover, this keeps the "dirty" pointer
around at the higher levels, where we need it for any MTE checking.

Since the normal case, especially for user-only, goes through
RAM, this clearing merely adds two insns per page lookup, which
will be completely in the noise.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20200626033144.790098-39-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

3 files changed, 25 insertions, 2 deletions

diff --git a/target/arm/cpu.c b/target/arm/cpu.c
index d987633..afe81e9 100644
--- a/target/arm/cpu.c
+++ b/target/arm/cpu.c
@@ -203,6 +203,9 @@ static void arm_cpu_reset(DeviceState *dev)
          * Enable TBI0 and TBI1.  While the real kernel only enables TBI0,
          * turning on both here will produce smaller code and otherwise
          * make no difference to the user-level emulation.
+         *
+         * In sve_probe_page, we assume that this is set.
+         * Do not modify this without other changes.
          */
         env->cp15.tcr_el[1].raw_tcr = (3ULL << 37);
 #else
diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c
index ad974c2..382fa82 100644
--- a/target/arm/sve_helper.c
+++ b/target/arm/sve_helper.c
@@ -3966,14 +3966,16 @@ static void sve_##NAME##_host(void *vd, intptr_t reg_off, void *host)  \
 static void sve_##NAME##_tlb(CPUARMState *env, void *vd, intptr_t reg_off,  \
                              target_ulong addr, uintptr_t ra)               \
 {                                                                           \
-    *(TYPEE *)(vd + H(reg_off)) = (TYPEM)TLB(env, addr, ra);                \
+    *(TYPEE *)(vd + H(reg_off)) =                                           \
+        (TYPEM)TLB(env, useronly_clean_ptr(addr), ra);                      \
 }
 
 #define DO_ST_TLB(NAME, H, TYPEE, TYPEM, TLB) \
 static void sve_##NAME##_tlb(CPUARMState *env, void *vd, intptr_t reg_off,  \
                              target_ulong addr, uintptr_t ra)               \
 {                                                                           \
-    TLB(env, addr, (TYPEM)*(TYPEE *)(vd + H(reg_off)), ra);                 \
+    TLB(env, useronly_clean_ptr(addr),                                      \
+        (TYPEM)*(TYPEE *)(vd + H(reg_off)), ra);                            \
 }
 
 #define DO_LD_PRIM_1(NAME, H, TE, TM)                   \
@@ -4091,6 +4093,19 @@ static bool sve_probe_page(SVEHostPage *info, bool nofault,
     int flags;
 
     addr += mem_off;
+
+    /*
+     * User-only currently always issues with TBI.  See the comment
+     * above useronly_clean_ptr.  Usually we clean this top byte away
+     * during translation, but we can't do that for e.g. vector + imm
+     * addressing modes.
+     *
+     * We currently always enable TBI for user-only, and do not provide
+     * a way to turn it off.  So clean the pointer unconditionally here,
+     * rather than look it up here, or pass it down from above.
+     */
+    addr = useronly_clean_ptr(addr);
+
     flags = probe_access_flags(env, addr, access_type, mmu_idx, nofault,
                                &info->host, retaddr);
     info->flags = flags;
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index e46c4a4..c20af6e 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -14634,6 +14634,11 @@ static void aarch64_tr_init_disas_context(DisasContextBase *dcbase,
     dc->features = env->features;
     dc->dcz_blocksize = arm_cpu->dcz_blocksize;
 
+#ifdef CONFIG_USER_ONLY
+    /* In sve_probe_page, we assume TBI is enabled. */
+    tcg_debug_assert(dc->tbid & 1);
+#endif
+
     /* Single step state. The code-generation logic here is:
      *  SS_ACTIVE == 0:
      *   generate code with no special handling for single-stepping (except