Skip to contentSkip to conversation

target/s390x: Fix MVC not always invalidating translation blocks

PATCH In review Ilya Leoshkevich 3 participants 6 messages Jan 28, 2025

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 target/s390x: Fix MVC not always invalidating translation blocks

Node.js crashes in qemu-system-s390x with random SIGSEGVs / SIGILLs.

The v8 JIT used by Node.js can garbage collect and overwrite unused code. Overwriting is performed by WritableJitAllocation::CopyCode(), which ultimately calls memcpy(). For certain sizes, memcpy() uses the MVC instruction.

QEMU implements MVC and other similar instructions using helpers. While TCG store ops invalidate affected translation blocks automatically, helpers must do this manually by calling probe_access_flags(). The MVC helper does this using the access_prepare() -> access_prepare_nf() -> s390_probe_access() -> probe_access_flags() call chain.

At the last step of this chain, the store size is replaced with 0. This causes the probe_access_flags() -> notdirty_write() -> tb_invalidate_phys_range_fast() chain to miss some translation blocks.

When this happens, QEMU executes a mix of old and new code. This quickly leads to either a SIGSEGV or a SIGILL in case the old code ends in the middle of a new instruction.

Fix by passing the true size.

---
 target/s390x/tcg/mem_helper.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
target/s390x/tcg/mem_helper.c +1 -1
index 32717acb7d1..c6ab2901e5a 100644
@@ -149,7 +149,7 @@ static inline int s390_probe_access(CPUArchState *env, target_ulong addr,
                                     int mmu_idx, bool nonfault,
                                     void **phost, uintptr_t ra)
 {
-    int flags = probe_access_flags(env, addr, 0, access_type, mmu_idx,
+    int flags = probe_access_flags(env, addr, size, access_type, mmu_idx,
                                    nonfault, phost, ra);
 
     if (unlikely(flags & TLB_INVALID_MASK)) {
Signature · 1 line
2.48.1
#1 lore raw

PATCH 2/2 tests/tcg/s390x: Test modifying code using the MVC instruction

Add a small test to prevent regressions.

---
 tests/tcg/s390x/Makefile.softmmu-target |  1 +
 tests/tcg/s390x/mvc-smc.c               | 82 +++++++++++++++++++++++++
 2 files changed, 83 insertions(+)
 create mode 100644 tests/tcg/s390x/mvc-smc.c
tests/tcg/s390x/Makefile.softmmu-target +1 -0
index 7adde2fa087..8cd4667c63b 100644
@@ -42,6 +42,7 @@ $(ASM_TESTS): LDFLAGS += -Wl,-T$(LINK_SCRIPT) -Wl,--build-id=none
 $(ASM_TESTS): $(LINK_SCRIPT)
 TESTS += $(ASM_TESTS)
 
+MULTIARCH_TESTS += mvc-smc
 S390X_MULTIARCH_RUNTIME_OBJS = head64.o console.o $(MINILIB_OBJS)
 $(MULTIARCH_TESTS): $(S390X_MULTIARCH_RUNTIME_OBJS)
 $(MULTIARCH_TESTS): LDFLAGS += $(S390X_MULTIARCH_RUNTIME_OBJS)
tests/tcg/s390x/mvc-smc.c new file +82 -0

82 lines not on this page · raw message

Signature · 1 line
2.48.1
Reviewed-by on v1 1/2
#3 lore raw
In reply to Ilya Leoshkevich
Quoted 17 lines · Ilya Leoshkevich

Node.js crashes in qemu-system-s390x with random SIGSEGVs / SIGILLs.

The v8 JIT used by Node.js can garbage collect and overwrite unused code. Overwriting is performed by WritableJitAllocation::CopyCode(), which ultimately calls memcpy(). For certain sizes, memcpy() uses the MVC instruction.

QEMU implements MVC and other similar instructions using helpers. While TCG store ops invalidate affected translation blocks automatically, helpers must do this manually by calling probe_access_flags(). The MVC helper does this using the access_prepare() -> access_prepare_nf() -> s390_probe_access() -> probe_access_flags() call chain.

At the last step of this chain, the store size is replaced with 0. This causes the probe_access_flags() -> notdirty_write() -> tb_invalidate_phys_range_fast() chain to miss some translation blocks.

Quoted 5 lines

When this happens, QEMU executes a mix of old and new code. This quickly leads to either a SIGSEGV or a SIGILL in case the old code ends in the middle of a new instruction.

Fix by passing the true size.

r~

Quoted 23 lines
---
  target/s390x/tcg/mem_helper.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
target/s390x/tcg/mem_helper.c +1 -1
index 32717acb7d1..c6ab2901e5a 100644
@@ -149,7 +149,7 @@ static inline int s390_probe_access(CPUArchState *env, target_ulong addr,
                                      int mmu_idx, bool nonfault,
                                      void **phost, uintptr_t ra)
  {
-    int flags = probe_access_flags(env, addr, 0, access_type, mmu_idx,
+    int flags = probe_access_flags(env, addr, size, access_type, mmu_idx,
                                     nonfault, phost, ra);
 
      if (unlikely(flags & TLB_INVALID_MASK)) {
Reviewed-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] tests/tcg/s390x: Test modifying code using the MVC instruction

In reply to Ilya Leoshkevich
Quoted 8 lines · Ilya Leoshkevich

Add a small test to prevent regressions.

---
  tests/tcg/s390x/Makefile.softmmu-target |  1 +
  tests/tcg/s390x/mvc-smc.c               | 82 +++++++++++++++++++++++++
  2 files changed, 83 insertions(+)

create mode 100644 tests/tcg/s390x/mvc-smc.c

r~

Quoted 101 lines
tests/tcg/s390x/Makefile.softmmu-target +1 -0
index 7adde2fa087..8cd4667c63b 100644
@@ -42,6 +42,7 @@ $(ASM_TESTS): LDFLAGS += -Wl,-T$(LINK_SCRIPT) -Wl,--build-id=none
  $(ASM_TESTS): $(LINK_SCRIPT)
  TESTS += $(ASM_TESTS)
 
+MULTIARCH_TESTS += mvc-smc
  S390X_MULTIARCH_RUNTIME_OBJS = head64.o console.o $(MINILIB_OBJS)
  $(MULTIARCH_TESTS): $(S390X_MULTIARCH_RUNTIME_OBJS)
  $(MULTIARCH_TESTS): LDFLAGS += $(S390X_MULTIARCH_RUNTIME_OBJS)
tests/tcg/s390x/mvc-smc.c new file +82 -0
new file mode 100644
index 00000000000..d68f60caa85
@@ -0,0 +1,82 @@
+/*
+ * Test modifying code using the MVC instruction.
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ */
+
+#include <minilib.h>
+
+#define PAGE_SIZE 4096
+#define BR_14_SIZE 2
+#define RWX_OFFSET 2
+
+static unsigned char rw[PAGE_SIZE + BR_14_SIZE];
+static unsigned char rwx[RWX_OFFSET + sizeof(rw)]
+    __attribute__((aligned(PAGE_SIZE)));
+
+typedef unsigned long (*function_t)(unsigned long);
+
+static int emit_function(unsigned char *p, int n)
+{
+    int i = 0, val = 0;
+
+    while (i < n - 2) {
+        /* aghi %r2,1 */
+        p[i++] = 0xa7;
+        p[i++] = 0x2b;
+        p[i++] = 0x00;
+        p[i++] = 0x01;
+        val++;
+    }
+
+    /* br %r14 */
+    p[i++] = 0x07;
+    p[i++] = 0xfe;
+
+    return val;
+}
+
+static void memcpy_mvc(void *dest, void *src, unsigned long n)
+{
+    while (n >= 256) {
+        asm("mvc 0(256,%[dest]),0(%[src])"
+            :
+            : [dest] "a" (dest)
+            , [src] "a" (src)
+            : "memory");
+        dest += 256;
+        src += 256;
+        n -= 256;
+    }
+    asm("exrl %[n],0f\n"
+        "j 1f\n"
+        "0: mvc 0(1,%[dest]),0(%[src])\n"
+        "1:"
+        :
+        : [dest] "a" (dest)
+        , [src] "a" (src)
+        , [n] "a" (n)
+        : "memory");
+}
+
+int main(void)
+{
+    int expected, size;
+
+    /* Create a TB. */
+    size = sizeof(rwx) - RWX_OFFSET - 4;
+    expected = emit_function(rwx + RWX_OFFSET, size);
+    if (((function_t)(rwx + RWX_OFFSET))(0) != expected) {
+        return 1;
+    }
+
+    /* Overwrite the TB. */
+    size += 4;
+    expected = emit_function(rw, size);
+    memcpy_mvc(rwx + RWX_OFFSET, rw, size);
+    if (((function_t)(rwx + RWX_OFFSET))(0) != expected) {
+        return 2;
+    }
+
+    return 0;
+}
Reviewed-by on v1 1/2
#4 lore raw
In reply to Ilya Leoshkevich
Quoted 13 lines · Ilya Leoshkevich

Node.js crashes in qemu-system-s390x with random SIGSEGVs / SIGILLs.

The v8 JIT used by Node.js can garbage collect and overwrite unused code. Overwriting is performed by WritableJitAllocation::CopyCode(), which ultimately calls memcpy(). For certain sizes, memcpy() uses the MVC instruction.

QEMU implements MVC and other similar instructions using helpers. While TCG store ops invalidate affected translation blocks automatically, helpers must do this manually by calling probe_access_flags(). The MVC helper does this using the access_prepare() -> access_prepare_nf() -> s390_probe_access() -> probe_access_flags() call chain.

Quoted 3 lines

At the last step of this chain, the store size is replaced with 0. This causes the probe_access_flags() -> notdirty_write() -> tb_invalidate_phys_range_fast() chain to miss some translation blocks.

We added the size parameter in:

commit 1770b2f2d3d6fe8f1e2d61692692264cac44340d

Author: Daniel Henrique Barboza <[email protected]>
Date:   Thu Feb 23 20:44:24 2023 -0300

accel/tcg: Add 'size' param to probe_access_flags()

probe_access_flags() as it is today uses probe_access_full(), which in turn uses probe_access_internal() with size = 0. probe_access_internal() then uses the size to call the tlb_fill() callback for the given CPU. This size param ('fault_size' as probe_access_internal() calls it) is ignored by most existing .tlb_fill callback implementations, e.g. arm_cpu_tlb_fill(), ppc_cpu_tlb_fill(), x86_cpu_tlb_fill() and mips_cpu_tlb_fill() to name a few.

And added support for more than one byte for the notdirty case in

commit e2faabee78ff127848f59892747d4c07c56de033

Author: Jessica Clarke <[email protected]>
Date:   Fri Nov 10 21:43:03 2023 -0800

accel/tcg: Forward probe size on to notdirty_write

Without this, we just dirty a single byte, and so if the caller writes more than one byte to the host memory then we won't have invalidated any translation blocks that start after the first byte and overlap those writes.

So I guess it's rather hard to find a "Fixes:" tag, because likely it's been broken forever?

Quoted 6 lines

When this happens, QEMU executes a mix of old and new code. This quickly leads to either a SIGSEGV or a SIGILL in case the old code ends in the middle of a new instruction.

Fix by passing the true size.

Signature · 3 lines
Cheers,

David / dhildenb
#5 lore raw
Quoted 64 lines · David Hildenbrand
Quoted 19 lines · Ilya Leoshkevich

Node.js crashes in qemu-system-s390x with random SIGSEGVs / SIGILLs.

The v8 JIT used by Node.js can garbage collect and overwrite unused code. Overwriting is performed by WritableJitAllocation::CopyCode(),
which ultimately calls memcpy(). For certain sizes, memcpy() uses the
MVC instruction.

QEMU implements MVC and other similar instructions using helpers. While
TCG store ops invalidate affected translation blocks automatically, helpers must do this manually by calling probe_access_flags(). The MVC
helper does this using the access_prepare() -> access_prepare_nf() ->
s390_probe_access() -> probe_access_flags() call chain.

Quoted 5 lines

At the last step of this chain, the store size is replaced with 0. This
causes the probe_access_flags() -> notdirty_write() -> tb_invalidate_phys_range_fast() chain to miss some translation blocks.

We added the size parameter in:

commit 1770b2f2d3d6fe8f1e2d61692692264cac44340d
Author: Daniel Henrique Barboza <[email protected]> Date:   Thu Feb 23 20:44:24 2023 -0300

     accel/tcg: Add 'size' param to probe_access_flags()

     probe_access_flags() as it is today uses probe_access_full(), which in
     turn uses probe_access_internal() with size = 0. probe_access_internal()
     then uses the size to call the tlb_fill() callback for the given CPU.
     This size param ('fault_size' as probe_access_internal() calls it) is
     ignored by most existing .tlb_fill callback implementations, e.g.
     arm_cpu_tlb_fill(), ppc_cpu_tlb_fill(), x86_cpu_tlb_fill() and mips_cpu_tlb_fill() to name a few.

And added support for more than one byte for the notdirty case in

commit e2faabee78ff127848f59892747d4c07c56de033
Author: Jessica Clarke <[email protected]>
Date:   Fri Nov 10 21:43:03 2023 -0800

     accel/tcg: Forward probe size on to notdirty_write

     Without this, we just dirty a single byte, and so if the caller writes
     more than one byte to the host memory then we won't have invalidated any
     translation blocks that start after the first byte and overlap those
     writes.

Quoted 3 lines

So I guess it's rather hard to find a "Fixes:" tag, because likely it's been
broken forever?

Yes, I thought about this too and gave up.
However, I now wonder if we still need one for non-philosophical, but rather practical backporting reasons? Then

Fixes e2faabee78ff ("accel/tcg: Forward probe size on to

notdirty_write")

(v8.2.0+) should be the right one I think?

Quoted 8 lines
Quoted 5 lines

When this happens, QEMU executes a mix of old and new code. This quickly leads to either a SIGSEGV or a SIGILL in case the old code ends in the middle of a new instruction.

Fix by passing the true size.