Avoid a race condition with munmap in another thread.
Use around blocks that exclusively use "host_fn".
Keep the blocks as small as possible, but without setting
and clearing for every operation on one page.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/tcg/sme_helper.c | 16 ++++++++++++++++
 target/arm/tcg/sve_helper.c | 26 ++++++++++++++++++++++++++
 2 files changed, 42 insertions(+)

diff --git a/target/arm/tcg/sme_helper.c b/target/arm/tcg/sme_helper.c
index e2e0575039..ab40ced38f 100644
--- a/target/arm/tcg/sme_helper.c
+++ b/target/arm/tcg/sme_helper.c
@@ -517,6 +517,8 @@ void sme_ld1(CPUARMState *env, void *za, uint64_t *vg,
         clr_fn(za, 0, reg_off);
     }
 
+    set_helper_retaddr(ra);
+
     while (reg_off <= reg_last) {
         uint64_t pg = vg[reg_off >> 6];
         do {
@@ -529,6 +531,8 @@ void sme_ld1(CPUARMState *env, void *za, uint64_t *vg,
         } while (reg_off <= reg_last && (reg_off & 63));
     }
 
+    clear_helper_retaddr();
+
     /*
      * Use the slow path to manage the cross-page misalignment.
      * But we know this is RAM and cannot trap.
@@ -543,6 +547,8 @@ void sme_ld1(CPUARMState *env, void *za, uint64_t *vg,
         reg_last = info.reg_off_last[1];
         host = info.page[1].host;
 
+        set_helper_retaddr(ra);
+
         do {
             uint64_t pg = vg[reg_off >> 6];
             do {
@@ -554,6 +560,8 @@ void sme_ld1(CPUARMState *env, void *za, uint64_t *vg,
                 reg_off += esize;
             } while (reg_off & 63);
         } while (reg_off <= reg_last);
+
+        clear_helper_retaddr();
     }
 }
 
@@ -701,6 +709,8 @@ void sme_st1(CPUARMState *env, void *za, uint64_t *vg,
     reg_last = info.reg_off_last[0];
     host = info.page[0].host;
 
+    set_helper_retaddr(ra);
+
     while (reg_off <= reg_last) {
         uint64_t pg = vg[reg_off >> 6];
         do {
@@ -711,6 +721,8 @@ void sme_st1(CPUARMState *env, void *za, uint64_t *vg,
         } while (reg_off <= reg_last && (reg_off & 63));
     }
 
+    clear_helper_retaddr();
+
     /*
      * Use the slow path to manage the cross-page misalignment.
      * But we know this is RAM and cannot trap.
@@ -725,6 +737,8 @@ void sme_st1(CPUARMState *env, void *za, uint64_t *vg,
         reg_last = info.reg_off_last[1];
         host = info.page[1].host;
 
+        set_helper_retaddr(ra);
+
         do {
             uint64_t pg = vg[reg_off >> 6];
             do {
@@ -734,6 +748,8 @@ void sme_st1(CPUARMState *env, void *za, uint64_t *vg,
                 reg_off += 1 << esz;
             } while (reg_off & 63);
         } while (reg_off <= reg_last);
+
+        clear_helper_retaddr();
     }
 }
 
diff --git a/target/arm/tcg/sve_helper.c b/target/arm/tcg/sve_helper.c
index dd49e67d7a..8d0af4bb1c 100644
--- a/target/arm/tcg/sve_helper.c
+++ b/target/arm/tcg/sve_helper.c
@@ -5738,6 +5738,8 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const target_ulong addr,
     reg_last = info.reg_off_last[0];
     host = info.page[0].host;
 
+    set_helper_retaddr(retaddr);
+
     while (reg_off <= reg_last) {
         uint64_t pg = vg[reg_off >> 6];
         do {
@@ -5752,6 +5754,8 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const target_ulong addr,
         } while (reg_off <= reg_last && (reg_off & 63));
     }
 
+    clear_helper_retaddr();
+
     /*
      * Use the slow path to manage the cross-page misalignment.
      * But we know this is RAM and cannot trap.
@@ -5771,6 +5775,8 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const target_ulong addr,
         reg_last = info.reg_off_last[1];
         host = info.page[1].host;
 
+        set_helper_retaddr(retaddr);
+
         do {
             uint64_t pg = vg[reg_off >> 6];
             do {
@@ -5784,6 +5790,8 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const target_ulong addr,
                 mem_off += N << msz;
             } while (reg_off & 63);
         } while (reg_off <= reg_last);
+
+        clear_helper_retaddr();
     }
 }
 
@@ -6093,6 +6101,8 @@ void sve_ldnfff1_r(CPUARMState *env, void *vg, const target_ulong addr,
     reg_last = info.reg_off_last[0];
     host = info.page[0].host;
 
+    set_helper_retaddr(retaddr);
+
     do {
         uint64_t pg = *(uint64_t *)(vg + (reg_off >> 3));
         do {
@@ -6113,6 +6123,8 @@ void sve_ldnfff1_r(CPUARMState *env, void *vg, const target_ulong addr,
         } while (reg_off <= reg_last && (reg_off & 63));
     } while (reg_off <= reg_last);
 
+    clear_helper_retaddr();
+
     /*
      * MemSingleNF is allowed to fail for any reason.  We have special
      * code above to handle the first element crossing a page boundary.
@@ -6348,6 +6360,8 @@ void sve_stN_r(CPUARMState *env, uint64_t *vg, target_ulong addr,
     reg_last = info.reg_off_last[0];
     host = info.page[0].host;
 
+    set_helper_retaddr(retaddr);
+
     while (reg_off <= reg_last) {
         uint64_t pg = vg[reg_off >> 6];
         do {
@@ -6362,6 +6376,8 @@ void sve_stN_r(CPUARMState *env, uint64_t *vg, target_ulong addr,
         } while (reg_off <= reg_last && (reg_off & 63));
     }
 
+    clear_helper_retaddr();
+
     /*
      * Use the slow path to manage the cross-page misalignment.
      * But we know this is RAM and cannot trap.
@@ -6381,6 +6397,8 @@ void sve_stN_r(CPUARMState *env, uint64_t *vg, target_ulong addr,
         reg_last = info.reg_off_last[1];
         host = info.page[1].host;
 
+        set_helper_retaddr(retaddr);
+
         do {
             uint64_t pg = vg[reg_off >> 6];
             do {
@@ -6394,6 +6412,8 @@ void sve_stN_r(CPUARMState *env, uint64_t *vg, target_ulong addr,
                 mem_off += N << msz;
             } while (reg_off & 63);
         } while (reg_off <= reg_last);
+
+        clear_helper_retaddr();
     }
 }
 
@@ -6560,7 +6580,9 @@ void sve_ld1_z(CPUARMState *env, void *vd, uint64_t *vg, void *vm,
                     if (unlikely(info.flags & TLB_MMIO)) {
                         tlb_fn(env, &scratch, reg_off, addr, retaddr);
                     } else {
+                        set_helper_retaddr(retaddr);
                         host_fn(&scratch, reg_off, info.host);
+                        clear_helper_retaddr();
                     }
                 } else {
                     /* Element crosses the page boundary. */
@@ -6782,7 +6804,9 @@ void sve_ldff1_z(CPUARMState *env, void *vd, uint64_t *vg, void *vm,
                     goto fault;
                 }
 
+                set_helper_retaddr(retaddr);
                 host_fn(vd, reg_off, info.host);
+                clear_helper_retaddr();
             }
             reg_off += esize;
         } while (reg_off & 63);
@@ -6986,7 +7010,9 @@ void sve_st1_z(CPUARMState *env, void *vd, uint64_t *vg, void *vm,
     do {
         void *h = host[i];
         if (likely(h != NULL)) {
+            set_helper_retaddr(retaddr);
             host_fn(vd, reg_off, h);
+            clear_helper_retaddr();
         } else if ((vg[reg_off >> 6] >> (reg_off & 63)) & 1) {
             target_ulong addr = base + (off_fn(vm, reg_off) << scale);
             tlb_fn(env, vd, reg_off, addr, retaddr);
-- 
2.43.0

On Wed, 10 Jul 2024 at 04:31, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Avoid a race condition with munmap in another thread.
> Use around blocks that exclusively use "host_fn".
> Keep the blocks as small as possible, but without setting
> and clearing for every operation on one page.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  target/arm/tcg/sme_helper.c | 16 ++++++++++++++++
>  target/arm/tcg/sve_helper.c | 26 ++++++++++++++++++++++++++
>  2 files changed, 42 insertions(+)


> @@ -6093,6 +6101,8 @@ void sve_ldnfff1_r(CPUARMState *env, void *vg, const target_ulong addr,
>      reg_last = info.reg_off_last[0];
>      host = info.page[0].host;
>
> +    set_helper_retaddr(retaddr);
> +
>      do {
>          uint64_t pg = *(uint64_t *)(vg + (reg_off >> 3));
>          do {
> @@ -6113,6 +6123,8 @@ void sve_ldnfff1_r(CPUARMState *env, void *vg, const target_ulong addr,
>          } while (reg_off <= reg_last && (reg_off & 63));
>      } while (reg_off <= reg_last);
>
> +    clear_helper_retaddr();
> +
>      /*
>       * MemSingleNF is allowed to fail for any reason.  We have special
>       * code above to handle the first element crossing a page boundary.

There's a "goto do_fault" inside the loop that we've bracketed here with
the set/clear_helper_retaddr() calls -- don't we need to call
clear_helper_retaddr() on that failure path too?

There's a TODO comment at the top of this file:

/*
 * Load contiguous data, first-fault and no-fault.
 *
 * For user-only, one could argue that we should hold the mmap_lock during
 * the operation so that there is no race between page_check_range and the
 * load operation.  However, unmapping pages out from under a running thread
 * is extraordinarily unlikely.  This theoretical race condition also affects
 * linux-user/ in its get_user/put_user macros.
 *
 * TODO: Construct some helpers, written in assembly, that interact with
 * host_signal_handler to produce memory ops which can properly report errors
 * without racing.
 */

Should we update it to note that we make at least some attempt to
handle the pages-unmapped-from-under-a-running-thread situation now?

thanks
-- PMM