i#3544 RV64: Refine fault translation and signal handling

Refine fault translation and signal handling, part of the linux.signalXXXX tests works on real hardware; however, they do not work on QEMU, so we can't enable them on CI.

Issue: #3544

core/arch/arch.h

@@ -1569,7 +1569,7 @@ translate_x86_to_x64(dcontext_t *dcontext, instrlist_t *ilist,
                      DR_PARAM_INOUT instr_t **instr);
 #endif
 
-#ifdef AARCHXX
+#if defined( AARCHXX) || defined(RISCV64)
 bool
 instr_is_ldstex_mangling(dcontext_t *dcontext, instr_t *inst);
 #endif

core/arch/riscv64/mangle.c

@@ -648,6 +648,47 @@ mangle_special_registers(dcontext_t *dcontext, instrlist_t *ilist, instr_t *inst
     return next_instr;
 }
 
+/***************************************************************************
+ * LR/SC sequence mangling.
+ */
+
+bool
+instr_is_ldstex_mangling(dcontext_t *dcontext, instr_t *instr)
+{
+    /* This should be kept in sync with mangle_exclusive_monitor_op(). */
+    if (!instr_is_our_mangling(instr))
+        return false;
+
+    opnd_t memop = opnd_create_null();
+    if (instr_get_opcode(instr) == OP_sd)
+        memop = instr_get_src(instr, 0);
+    else if (instr_get_opcode(instr) == OP_ld)
+        memop = instr_get_dst(instr, 0);
+    if (opnd_is_base_disp(memop)) {
+        ASSERT(opnd_get_index(memop) == DR_REG_NULL && opnd_get_scale(memop) == 0);
+        uint offs = opnd_get_disp(memop);
+        if (opnd_get_base(memop) == dr_reg_stolen && offs >= TLS_LRSC_ADDR_SLOT &&
+            offs <= TLS_LRSC_SIZE_SLOT)
+            return true;
+    }
+
+    ptr_int_t val;
+    if (instr_get_opcode(instr) == OP_fence || instr_get_opcode(instr) == OP_bne ||
+        instr_get_opcode(instr) == OP_bne ||
+        /* Check for sc.w/d+bne+jal pattern.  */
+        (instr_get_opcode(instr) == OP_jal && instr_get_prev(instr) != NULL &&
+         instr_get_opcode(instr_get_prev(instr)) == OP_bne &&
+         instr_get_prev(instr_get_prev(instr)) != NULL &&
+         instr_is_exclusive_store(instr_get_prev(instr_get_prev(instr)))) ||
+        instr_is_exclusive_load(instr) || instr_is_exclusive_store(instr) ||
+        (instr_is_mov_constant(instr, &val) &&
+         (val == 1 /* cas fail */ || val == -1 /* reservation invalidation */ ||
+          val == 4 /* lr.w/sc.w size */ || val == 8 /* lr.d/sc.d size */)))
+        return true;
+
+    return false;
+}
+
 static instr_t *
 mangle_exclusive_load(dcontext_t *dcontext, instrlist_t *ilist, instr_t *instr,
                       instr_t *next_instr)

core/arch/x86_code.c

@@ -306,7 +306,7 @@ new_thread_setup(priv_mcontext_t *mc)
     ASSERT(rc != -1); /* this better be a new thread */
     dcontext = get_thread_private_dcontext();
     ASSERT(dcontext != NULL);
-#    ifdef AARCHXX
+#    if defined(AARCHXX) || defined(RISCV64)
     set_app_lib_tls_base_from_clone_record(dcontext, crec);
 #    endif
 #    ifdef ARM

core/translate.c

@@ -132,6 +132,8 @@ instr_is_inline_syscall_jmp(dcontext_t *dcontext, instr_t *inst)
              /* A32 uses a regular jump */
              instr_get_opcode(inst) == OP_b) &&
             opnd_is_instr(instr_get_target(inst)));
+#    elif defined(RISCV64)
+    return (instr_get_opcode(inst) == OP_jal && opnd_is_instr(instr_get_target(inst)));
 #    else
     ASSERT_NOT_IMPLEMENTED(false);
     return false;
@@ -304,14 +306,14 @@ translate_walk_track_pre_instr(dcontext_t *tdcontext, instr_t *inst,
         walk->unsupported_mangle = false;
         walk->xsp_adjust = 0;
         for (reg_id_t r = 0; r < REG_SPILL_NUM; r++) {
-#ifndef AARCHXX
+#ifdef X86
             /* we should have seen a restore for every spill, unless at
              * fragment-ending jump to ibl, which shouldn't come here
              */
             ASSERT(walk->reg_spill_offs[r] == UINT_MAX);
             walk->reg_spill_offs[r] = UINT_MAX; /* be paranoid */
 #else
-            /* On AArchXX we do spill registers across app instrs and mangle
+            /* On AArchXX/RISCV64 we do spill registers across app instrs and mangle
              * regions, though right now only the following routines do this:
              * - mangle_stolen_reg()
              * - mangle_gpr_list_read()
@@ -399,9 +401,6 @@ translate_walk_track_post_instr(dcontext_t *tdcontext, instr_t *inst,
          * comment above for post-mangling traces), and so for local
          * spills like rip-rel and ind branches this is fine.
          */
-#if defined(RISCV64)
-        ASSERT_NOT_IMPLEMENTED(false);
-#endif
         if (instr_is_cti(inst) &&
 #ifdef X86
             /* Do not reset for a trace-cmp jecxz or jmp (32-bit) or
@@ -420,10 +419,7 @@ translate_walk_track_post_instr(dcontext_t *tdcontext, instr_t *inst,
               (!opnd_is_pc(instr_get_target(inst)) ||
                (opnd_get_pc(instr_get_target(inst)) >= walk->start_cache &&
                 opnd_get_pc(instr_get_target(inst)) < walk->end_cache))))
-#elif defined(RISCV64)
-            /* FIXME i#3544: Not implemented */
-            false
-#else
+#elif defined(AARCHXX)
             /* Do not reset for cbnz/bne in ldstex mangling, nor for the b after strex. */
             !(instr_get_opcode(inst) == OP_cbnz ||
               (instr_get_opcode(inst) == OP_b &&
@@ -432,6 +428,17 @@ translate_walk_track_post_instr(dcontext_t *tdcontext, instr_t *inst,
               (instr_get_opcode(inst) == OP_b &&
                (instr_get_prev(inst) != NULL &&
                 instr_is_exclusive_store(instr_get_prev(inst)))))
+#elif defined(RISCV64)
+            /* Do not reset for bne in LR/SC mangling, nor for the jal after SC.
+             * This should be kept in sync with mangle_exclusive_monitor_op().
+             */
+            !(instr_get_opcode(inst) == OP_bne ||
+              (instr_get_opcode(inst) == OP_jal && instr_get_prev(inst) != NULL &&
+               instr_get_opcode(instr_get_prev(inst)) == OP_bne &&
+               instr_get_prev(instr_get_prev(inst)) != NULL &&
+               instr_is_exclusive_store(instr_get_prev(instr_get_prev(inst)))))
+#else
+#    error Unsupported architecture
 #endif
         ) {
             /* FIXME i#1551: add ARM version of the series of trace cti checks above */
@@ -547,7 +554,7 @@ translate_walk_track_post_instr(dcontext_t *tdcontext, instr_t *inst,
             /* nothing to do */
         }
 #endif
-#ifdef AARCHXX
+#if defined(AARCHXX) || defined(RISCV64)
         else if (instr_is_ldstex_mangling(tdcontext, inst)) {
             /* nothing to do */
         }

core/unix/signal.c

@@ -2826,7 +2826,6 @@ sigcontext_to_mcontext(priv_mcontext_t *mc, sig_full_cxt_t *sc_full,
     if (TEST(DR_MC_INTEGER, flags)) {
         mc->ra = sc->SC_FIELD(sc_regs.ra);
         mc->gp = sc->SC_FIELD(sc_regs.gp);
-        mc->tp = sc->SC_FIELD(sc_regs.tp);
         mc->t0 = sc->SC_FIELD(sc_regs.t0);
         mc->t1 = sc->SC_FIELD(sc_regs.t1);
         mc->t2 = sc->SC_FIELD(sc_regs.t2);
@@ -2921,6 +2920,13 @@ sigcontext_to_mcontext(priv_mcontext_t *mc, sig_full_cxt_t *sc_full,
  * thread_get_mcontext() sets the dcontext isa_mode from cpsr, and thread_set_mcontext()
  * sets it.
  */
+/* XXX: on RISC-V, the tp register is a general-purpose register, but as its name
+ * suggests, it is conventionally used as the thread register.
+ * We save DR's tls by keeping tp's value unchanged, regardless in DR or in application.
+ * Therefore, when converting between sigcontext and mcontext, we should keep tp
+ * unchanged, otherwise for example when we sigreturn to app tp will change, causing an
+ * error.
+ */
 void
 mcontext_to_sigcontext(sig_full_cxt_t *sc_full, priv_mcontext_t *mc,
                        dr_mcontext_flags_t flags)
@@ -2956,7 +2962,6 @@ mcontext_to_sigcontext(sig_full_cxt_t *sc_full, priv_mcontext_t *mc,
     if (TEST(DR_MC_INTEGER, flags)) {
         sc->SC_FIELD(sc_regs.ra) = mc->ra;
         sc->SC_FIELD(sc_regs.gp) = mc->gp;
-        sc->SC_FIELD(sc_regs.tp) = mc->tp;
         sc->SC_FIELD(sc_regs.t0) = mc->t0;
         sc->SC_FIELD(sc_regs.t1) = mc->t1;
         sc->SC_FIELD(sc_regs.t2) = mc->t2;
@@ -3051,20 +3056,22 @@ mcontext_to_ucontext(kernel_ucontext_t *uc, priv_mcontext_t *mc)
     mcontext_to_sigcontext(&sc_full, mc, DR_MC_ALL);
 }
 
-#ifdef AARCHXX
+#if defined(AARCHXX) || defined(RISCV64)
 static void
 set_sigcxt_stolen_reg(sigcontext_t *sc, reg_t val)
 {
-    *(&sc->SC_R0 + (dr_reg_stolen - DR_REG_R0)) = val;
+    *(&sc->IF_AARCHXX_ELSE(SC_R0, SC_A0) +
+      (dr_reg_stolen - IF_AARCHXX_ELSE(DR_REG_R0, DR_REG_A0))) = val;
 }
 
 static reg_t
 get_sigcxt_stolen_reg(sigcontext_t *sc)
 {
-    return *(&sc->SC_R0 + (dr_reg_stolen - DR_REG_R0));
+    return *(&sc->IF_AARCHXX_ELSE(SC_R0, SC_A0) +
+             (dr_reg_stolen - IF_AARCHXX_ELSE(DR_REG_R0, DR_REG_A0)));
 }
 
-#    ifndef AARCH64
+#    ifdef ARM
 static dr_isa_mode_t
 get_pc_mode_from_cpsr(sigcontext_t *sc)
 {
@@ -3251,7 +3258,11 @@ thread_set_self_context(void *cxt)
 #elif defined(ARM)
     asm("ldr  " ASM_XSP ", %0" : : "m"(xsp_for_sigreturn));
     asm("b    dynamorio_sigreturn");
-#endif /* X86/AARCH64/ARM */
+#elif defined(RISCV)
+    ASSERT_NOT_TESTED();
+    asm("addi   " ASM_XSP ", %0, 0" : : "r"(xsp_for_sigreturn));
+    asm("j      dynamorio_sigreturn");
+#endif /* X86/AARCH64/ARM/RISCV64 */
     ASSERT_NOT_REACHED();
 }
 
@@ -3370,7 +3381,6 @@ sig_has_restorer(thread_sig_info_t *info, int sig)
 #    elif defined(RISCV64)
         static const byte SIGRET_NONRT[8] = { 0 }; /* unused */
         static const byte SIGRET_RT[8] = { 0 };    /* unused */
-        ;
 #    endif
         byte buf[MAX(sizeof(SIGRET_NONRT), sizeof(SIGRET_RT))] = { 0 };
         if (d_r_safe_read(info->sighand->action[sig]->restorer, sizeof(buf), buf) &&
@@ -4626,7 +4636,7 @@ adjust_syscall_for_restart(dcontext_t *dcontext, thread_sig_info_t *info, int si
     } else {
         ASSERT_NOT_REACHED(); /* Inlined syscalls no longer come here. */
     }
-#ifdef AARCHXX
+#if defined(AARCHXX) || defined(RISCV64)
     /* dr_reg_stolen is holding DR's TLS on receiving a signal,
      * so we need to put the app's reg value into the ucontext instead.
      * The translation process normally does this for us, but here we're doing
@@ -4734,7 +4744,8 @@ find_next_fragment_from_gencode(dcontext_t *dcontext, sigcontext_t *sc)
         if (f == NULL && sc->SC_XCX != 0)
             f = fragment_lookup(dcontext, (app_pc)sc->SC_XCX);
 #elif defined(RISCV64)
-/* FIXME i#3544: Not implemented */
+        /* FIXME i#3544: Not implemented */
+        ASSERT_NOT_IMPLEMENTED(false);
 #else
 #    error Unsupported arch.
 #endif
@@ -6292,7 +6303,9 @@ execute_handler_from_dispatch(dcontext_t *dcontext, int sig)
         dump_sigcontext(dcontext, sc);
         LOG(THREAD, LOG_ASYNCH, 3, "\n");
     }
-    IF_AARCHXX(ASSERT(get_sigcxt_stolen_reg(sc) != (reg_t)*get_dr_tls_base_addr()));
+#    if defined(AARCHXX) || defined(RISCV64)
+    ASSERT(get_sigcxt_stolen_reg(sc) != (reg_t)*get_dr_tls_base_addr());
+#    endif
 #endif
     /* FIXME: other state?  debug regs?
      * if no syscall allowed between main_ (when frame created) and
@@ -7370,26 +7383,26 @@ handle_sigreturn(dcontext_t *dcontext, void *ucxt_param, int style)
      * look like whatever would happen to the app...
      */
     ASSERT((app_pc)sc->SC_XIP != next_pc);
-#    if defined(AARCHXX)
+#    if defined(AARCHXX) || defined(RISCV64)
     ASSERT(get_sigcxt_stolen_reg(sc) != (reg_t)*get_dr_tls_base_addr());
     /* We're called from DR and are not yet in the cache, so we want to set the
      * mcontext slot, not the TLS slot, to set the stolen reg value.
      */
     set_stolen_reg_val(get_mcontext(dcontext), get_sigcxt_stolen_reg(sc));
     /* The linkstub expects DR's TLS to be in the actual register. */
     set_sigcxt_stolen_reg(sc, (reg_t)*get_dr_tls_base_addr());
-#        ifdef AARCH64
+#        if defined(AARCH64)
     /* On entry to the do_syscall gencode, we save X1 into TLS_REG1_SLOT.
      * Then the sigreturn would redirect the flow to the fcache_return gencode.
      * In fcache_return it recovers the values of x0 and x1 from TLS_SLOT 0 and 1.
      */
     get_mcontext(dcontext)->r1 = sc->SC_FIELD_AARCH64(1);
+#        elif defined(RISCV64)
+    get_mcontext(dcontext)->a1 = sc->SC_FIELD(sc_regs.a1);
 #        else
     /* We're going to our fcache_return gencode which uses DEFAULT_ISA_MODE */
     set_pc_mode_in_cpsr(sc, DEFAULT_ISA_MODE);
 #        endif
-#    elif defined(RISCV64)
-    ASSERT_NOT_IMPLEMENTED(false);
 #    endif
 #endif
 

core/unix/signal_linux_riscv64.c

@@ -57,7 +57,39 @@ save_fpstate(dcontext_t *dcontext, sigframe_rt_t *frame)
 void
 dump_sigcontext(dcontext_t *dcontext, sigcontext_t *sc)
 {
-    LOG(THREAD, LOG_ASYNCH, 1, "FIXME i#3544: NYI on RISCV64");
+
+    LOG(THREAD, LOG_ASYNCH, 1, "\tpc  = " PFX "\n", sc->sc_regs.pc);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tra  = " PFX "\n", sc->sc_regs.ra);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tsp  = " PFX "\n", sc->sc_regs.sp);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tgp  = " PFX "\n", sc->sc_regs.gp);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ttp  = " PFX "\n", sc->sc_regs.tp);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt0  = " PFX "\n", sc->sc_regs.t0);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt1  = " PFX "\n", sc->sc_regs.t1);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt2  = " PFX "\n", sc->sc_regs.t2);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts0  = " PFX "\n", sc->sc_regs.s0);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts1  = " PFX "\n", sc->sc_regs.s1);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta0  = " PFX "\n", sc->sc_regs.a0);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta1  = " PFX "\n", sc->sc_regs.a1);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta2  = " PFX "\n", sc->sc_regs.a2);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta3  = " PFX "\n", sc->sc_regs.a3);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta4  = " PFX "\n", sc->sc_regs.a4);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta5  = " PFX "\n", sc->sc_regs.a5);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta6  = " PFX "\n", sc->sc_regs.a6);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ta7  = " PFX "\n", sc->sc_regs.a7);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts2  = " PFX "\n", sc->sc_regs.s2);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts3  = " PFX "\n", sc->sc_regs.s3);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts4  = " PFX "\n", sc->sc_regs.s4);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts5  = " PFX "\n", sc->sc_regs.s5);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts6  = " PFX "\n", sc->sc_regs.s6);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts7  = " PFX "\n", sc->sc_regs.s7);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts8  = " PFX "\n", sc->sc_regs.s8);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts9  = " PFX "\n", sc->sc_regs.s9);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts10 = " PFX "\n", sc->sc_regs.s10);
+    LOG(THREAD, LOG_ASYNCH, 1, "\ts11 = " PFX "\n", sc->sc_regs.s11);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt3  = " PFX "\n", sc->sc_regs.t3);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt4  = " PFX "\n", sc->sc_regs.t4);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt5  = " PFX "\n", sc->sc_regs.t5);
+    LOG(THREAD, LOG_ASYNCH, 1, "\tt6  = " PFX "\n", sc->sc_regs.t6);
 }
 #endif /* DEBUG */
 

suite/tests/CMakeLists.txt

@@ -5870,6 +5870,7 @@ if (RISCV64)
     code_api|client.stack-overflow
     code_api|client.truncate
     code_api|common.broadfun
+    code_api|common.fib
     code_api|libutil.drconfig_test
     code_api|libutil.frontend_test
     code_api|linux.exit
@@ -5880,6 +5881,7 @@ if (RISCV64)
     code_api|linux.signalfd
     code_api|pthreads.pthreads
     code_api|pthreads.pthreads_exit
+    code_api|pthreads.pthreads_exit
     code_api|security-linux.trampoline
     no_code_api,no_intercept_all_signals|linux.sigaction
     PROPERTIES LABELS RUNS_ON_QEMU)