[compiler-rt][hwasan] Support for Intel LAM v6 API

[glider]

Summary

Version 6 of Intel LAM kernel patches
(https://lore.kernel.org/all/[e-mail]/)
introduces arch_prctl(ARCH_GET_MAX_TAG_BITS), which (unlike
ARCH_GET_UNTAG_MASK) can be used to determine if the kernel really
supports pointer tagging.

Depends on D132545

[kstoimenov]

compiler-rt/lib/hwasan/hwasan_linux.cpp

137

Does it have to kTagBits? Could it be 'num_bits < kTagBits' because you would be able to fit tag if it is >= kTagBits?

[vitalybuka]

compiler-rt/lib/hwasan/hwasan_linux.cpp

137

Even more, why do we use this call at all if patches support ARCH_GET_UNTAG_MASK?
We expect particular bits for hwasan, so better check exact mask.
if ((kAddressTagMask & returned_mask) != kAddressTagMask) return false;

[glider]

compiler-rt/lib/hwasan/hwasan_linux.cpp

137

Does it have to kTagBits? Could it be 'num_bits < kTagBits' because you would be able to fit tag if it is >= kTagBits?

Fair enough, we need to check for >=

137

Even more, why do we use this call at all if patches support ARCH_GET_UNTAG_MASK?
We expect particular bits for hwasan, so better check exact mask.
if ((kAddressTagMask & returned_mask) != kAddressTagMask) return false;

Good catch, but we need to check both.
ARCH_GET_UNTAG_MASK will return a nonzero mask even if the hardware doesn't support LAM.

[vitalybuka]

compiler-rt/lib/hwasan/hwasan_linux.cpp

137

my mistake, ARCH_GET_UNTAG_MASK is for enabled tagging
so (num_bits < kTagBits) is appropriate here

[vitalybuka]

Assuming the comment is straightforward

[kstoimenov]

Feel free to push after addressing the comments.

[glider]

Use ARCH_GET_UNTAG_MASK to double check HWASan can place tags in this bit range.

[glider]

compiler-rt/lib/hwasan/hwasan_linux.cpp

137

Ah, you are right, we can only check the mask in EnableTaggingAbi().

[glider]

Compare the mask returned by ARCH_GET_UNTAG_MASK to what HWASan expects.

Also use unsigned long variables to get values from arch_prctl.

[glider]

PTAL

[kstoimenov]

LGTM

compiler-rt/lib/hwasan/hwasan_linux.cpp

static void MaybeDieIfNoTaggingAbi(const char *message) {
   Die();
 }
 
 #  define PR_SET_TAGGED_ADDR_CTRL 55
 #  define PR_GET_TAGGED_ADDR_CTRL 56
 #  define PR_TAGGED_ADDR_ENABLE (1UL << 0)
 #  define ARCH_GET_UNTAG_MASK 0x4001
 #  define ARCH_ENABLE_TAGGED_ADDR 0x4002
+#  define ARCH_GET_MAX_TAG_BITS 0x4003
 
 static bool CanUseTaggingAbi() {
 #  if defined(__x86_64__)
+  unsigned long num_bits = 0;
   // Check for x86 LAM support. This API is based on a currently unsubmitted
-  // patch to the Linux kernel (as of July 2022) and is thus subject to
-  // change. Patch is here:
-  // https://lore.kernel.org/linux-mm/[e-mail]/
-  return !internal_iserror(internal_arch_prctl(ARCH_GET_UNTAG_MASK, 0));
+  // patch to the Linux kernel (as of August 2022) and is thus subject to
+  // change. The patch is here:
+  // https://lore.kernel.org/all/[e-mail]/
+  //
+  // arch_prctl(ARCH_GET_MAX_TAG_BITS, &bits) returns the maximum number of tag
+  // bits the user can request, or zero if LAM is not supported by the hardware.
+  if (internal_iserror(internal_arch_prctl(ARCH_GET_MAX_TAG_BITS,
+                                           reinterpret_cast<uptr>(&num_bits))))
+    return false;
+  // The platform must provide enough bits for HWASan tags.
+  if (num_bits < kTagBits)
+    return false;
+  return true;
 #  else
   // Check for ARM TBI support.
   return !internal_iserror(internal_prctl(PR_GET_TAGGED_ADDR_CTRL, 0, 0, 0, 0));
 #  endif // __x86_64__
 }
 
 static bool EnableTaggingAbi() {
 #  if defined(__x86_64__)
   // Enable x86 LAM tagging for the process.
-  return !internal_iserror(internal_arch_prctl(ARCH_ENABLE_TAGGED_ADDR, kTagBits));
+  //
+  // arch_prctl(ARCH_ENABLE_TAGGED_ADDR, bits) enables tagging if the number of
+  // tag bits requested by the user does not exceed that provided by the system.
+  // arch_prctl(ARCH_GET_UNTAG_MASK, &mask) returns the mask of significant
+  // address bits. It is ~0ULL if either LAM is disabled for the process or LAM
+  // is not supported by the hardware.
+  if (internal_iserror(internal_arch_prctl(ARCH_ENABLE_TAGGED_ADDR, kTagBits)))
+    return false;
+  unsigned long mask = 0;
+  // Make sure the tag bits are where we expect them to be.
+  if (internal_iserror(internal_arch_prctl(ARCH_GET_UNTAG_MASK,
+                                           reinterpret_cast<uptr>(&mask))))
+    return false;
+  // @mask has ones for non-tag bits, whereas @kAddressTagMask has ones for tag
+  // bits. Therefore these masks must not overlap.
+  if (mask & kAddressTagMask)
+    return false;
+  return true;
 #  else
   // Enable ARM TBI tagging for the process. If for some reason tagging is not
   // supported, prctl(PR_SET_TAGGED_ADDR_CTRL, PR_TAGGED_ADDR_ENABLE) returns
   // -EINVAL.
   if (internal_iserror(internal_prctl(PR_SET_TAGGED_ADDR_CTRL,
                                       PR_TAGGED_ADDR_ENABLE, 0, 0, 0)))
     return false;
   // Ensure that TBI is enabled.