Export of internal Abseil changes

--
07240ca7822d007cdcc79f2c40bd58b2c2010348 by Abseil Team <absl-team@google.com>:

Correct the comment from "AlphaNum" to "Arg".

PiperOrigin-RevId: 416139192

--
adcba4a6b3763626e1db7b1e8c108b3114903557 by Martijn Vels <mvels@google.com>:

Fix NewExternalRep() to require data being non-empty, and remove nullptr return.

PiperOrigin-RevId: 416135865

--
c0d14cd918fb16f15d1d84de9284b5c5ecc1f8f2 by Abseil Team <absl-team@google.com>:

Fix doc comment for absl::ascii_isprint().

The comment was incorrectly saying that it includes all whitespace.
It doesn't; the only whitespace char it includes is ' '.

PiperOrigin-RevId: 416112524

--
d83327800159c07002b6865e21232a12463e02dd by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 416099978

--
baf11e9ca42ca9140cdbf8075f971db8d65b1195 by Ilya Tokar <tokarip@google.com>:

Prevent compiler from optimizing Group_Match* benchmarks away.
Currently we benchmark single store of precomputed value.

Not all affected benchmarks show performance changes:

BM_Group_Match                                          0.53ns ± 1%  0.53ns ± 0%   -0.42%  (p=0.038 n=10+10)
BM_Group_MatchEmpty                                     0.26ns ± 1%  0.26ns ± 1%     ~     (p=1.000 n=10+10)
BM_Group_MatchEmptyOrDeleted                            0.26ns ± 1%  0.26ns ± 1%     ~     (p=0.121 n=10+10)
BM_Group_CountLeadingEmptyOrDeleted                     0.26ns ± 1%  0.45ns ± 0%  +70.05%   (p=0.000 n=10+8)
BM_Group_MatchFirstEmptyOrDeleted                       0.26ns ± 0%  0.44ns ± 1%  +65.91%    (p=0.000 n=8+9)

But inspecting the generated code shows the difference,
e. g. BM_Group_MatchFirstEmptyOrDeleted

Before:
add  $0xffffffffffffffff,%rbx
jne  30

After:
pcmpeqd  %xmm0,%xmm0
pcmpgtb  -0x30(%rbp),%xmm0
pmovmskb %xmm0,%eax
add: 0x23$0xffffffffffffffff,%rbx
jne      40
PiperOrigin-RevId: 416083515

--
122fbff893dc4571b3e75e4b241eb4495b925610 by Abseil Team <absl-team@google.com>:

Put namespace guard in ABSL_DECLARE_FLAG to make declaring a flag in a namespace a compiler error instead of a linker error.

PiperOrigin-RevId: 416036072

--
020fd8a20f5fa319e948846e003391fcb9e03868 by Ilya Tokar <tokarip@google.com>:

Make Cord::InlineRep::set_data unconditionally zero out memory.

Currently there is a single case where we don't zero out memory
as an optimization. Unconditional zeroing doesn't show any changes
in benchmarks, except for the unrelated improvement:

BM_CordPartialCopyToCord/1M/1              12.6ns ± 4%   12.6ns ± 4%     ~     (p=0.857 n=16+19)
BM_CordPartialCopyToCord/1M/128            44.9ns ± 7%   45.0ns ± 3%     ~     (p=0.468 n=18+17)
BM_CordPartialCopyToCord/1M/1k             64.5ns ± 4%   61.4ns ± 4%   -4.82%  (p=0.000 n=19+17)
BM_CordPartialCopyToCord/1M/8k              139ns ± 3%    128ns ±15%   -7.76%  (p=0.009 n=17+20)
BM_CordPartialCopyToCord/1M/16k             193ns ± 6%    168ns ± 6%  -13.17%  (p=0.000 n=17+17)
BM_CordPartialCopyToCord/4M/16k             199ns ± 4%    177ns ± 4%  -11.36%  (p=0.000 n=17+18)
BM_CordPartialCopyToCord/4M/32k             275ns ± 3%    250ns ± 4%   -9.00%  (p=0.000 n=18+18)
BM_CordPartialCopyToCord/4M/64k             291ns ± 4%    266ns ± 5%   -8.53%  (p=0.000 n=18+16)
BM_CordPartialCopyToCord/4M/128k            322ns ± 5%    291ns ± 4%   -9.43%  (p=0.000 n=20+18)
BM_CordPartialCopyToCord/8M/32k             281ns ± 5%    251ns ± 4%  -10.38%  (p=0.000 n=20+16)
BM_CordPartialCopyToCord/8M/64k             293ns ± 6%    267ns ± 4%   -8.87%  (p=0.000 n=16+19)
BM_CordPartialCopyToCord/8M/128k            334ns ± 3%    305ns ± 2%   -8.56%  (p=0.000 n=17+16)

This is clearly an alignmnet effect since number of the executed instructions is the same:
M_CordPartialCopyToCord/1M/1                 155 ± 0%                155 ± 0%     ~     (all samples are equal)
BM_CordPartialCopyToCord/1M/128               446 ± 0%                446 ± 0%     ~           (p=0.332 n=36+39)
BM_CordPartialCopyToCord/1M/1k                473 ± 0%                473 ± 0%     ~           (p=0.969 n=40+40)
BM_CordPartialCopyToCord/1M/8k                808 ± 0%                808 ± 0%     ~           (p=0.127 n=40+39)
BM_CordPartialCopyToCord/1M/16k               957 ± 0%                957 ± 0%     ~           (p=0.532 n=40+40)
BM_CordPartialCopyToCord/4M/16k               952 ± 0%                952 ± 0%     ~           (p=0.686 n=39+39)
BM_CordPartialCopyToCord/4M/32k             1.12k ± 0%              1.12k ± 0%     ~           (p=0.690 n=40+40)
BM_CordPartialCopyToCord/4M/64k             1.23k ± 0%              1.23k ± 0%     ~           (p=0.182 n=40+39)
BM_CordPartialCopyToCord/4M/128k            1.44k ± 0%              1.44k ± 0%     ~           (p=0.711 n=40+40)
BM_CordPartialCopyToCord/8M/32k             1.12k ± 0%              1.12k ± 0%     ~           (p=0.697 n=40+40)
BM_CordPartialCopyToCord/8M/64k             1.23k ± 0%              1.23k ± 0%   +0.00%        (p=0.049 n=40+40)
BM_CordPartialCopyToCord/8M/128k            1.44k ± 0%              1.44k ± 0%     ~           (p=0.507 n=40+40)

This makes code simpler and doesn't regress performance.

PiperOrigin-RevId: 415560574

--
37305b2690b31682088749e4d62f40d7095bdc54 by Derek Mauro <dmauro@google.com>:

Internal change

PiperOrigin-RevId: 415558737

--
86aaed569b9e743c1eb813a5f48def978a793db3 by Martijn Vels <mvels@google.com>:

Internal change

PiperOrigin-RevId: 415515201

--
6cdb8786cdcb4fa0b8a4b72fc98940877d1fdeff by Abseil Team <absl-team@google.com>:

Update SubmitMutexProfileData to accept wait_cycles instead of wait_timestamp

PiperOrigin-RevId: 415360871

--
9f979d307aa16ad09f214e04876cbe84395c0901 by Abseil Team <absl-team@google.com>:

absl::flat_hash_set compiles with -Wconversion -Wsign-compare

PiperOrigin-RevId: 415357498

--
9eceb14174708f15e61259d449b214a8a4c7f9e7 by Abseil Team <absl-team@google.com>:

Fix AddressIsReadable for the corner case of (aligned) addr == NULL.

PiperOrigin-RevId: 415307792

--
1a39ffe55898375e2d7f88c17c99db5a1b95b313 by Martijn Vels <mvels@google.com>:

Internal change

PiperOrigin-RevId: 415162872

--
64378549b110d5f5762185a5906c520fba70f0e7 by Abseil Team <absl-team@google.com>:

Fix a typo in the comments

PiperOrigin-RevId: 415088461

--
41aae8322e913b82710153c22b97c611fdb6e1fb by Abseil Team <absl-team@google.com>:

Switch from `connect` to `rt_sigreturn` -- the latter is much less problematic
for system call sandboxes.

PiperOrigin-RevId: 415073965

--
870c5e3388b6a35611bff538626fe7a1c8c87171 by Abseil Team <absl-team@google.com>:

Add ABSL_HAVE_HWADDRESS_SANITIZER and ABSL_HAVE_LEAK_SANITIZER

PiperOrigin-RevId: 414871189

--
f213ed60a66b58da7ac40555adfb1d529ff0a4db by Derek Mauro <dmauro@google.com>:

Remove reference to __SANITIZE_MEMORY__, which does not exist

It appears to have been copied by pattern matching from the ASAN/TSAN
code blocks.

https://github.com/gcc-mirror/gcc/blob/f47662204de27f7685699eeef89aa173ccf32d85/gcc/cppbuiltin.c#L79-L126

PiperOrigin-RevId: 414806587

--
b152891e73ab515f397ceb53f66c8ee2f33863ea by Abseil Team <absl-team@google.com>:

Rollback previous commit: SYS_open is not defined in certain environments.

PiperOrigin-RevId: 414521820

--
5a1cbb282331023902e1374dd0d920c4effbe47f by Abseil Team <absl-team@google.com>:

Use syscall(SYS_open, ...) instead of open() to avoid possible symbol
interposition.

Also add some warning notes.

PiperOrigin-RevId: 414508186

--
1824d6593612710aafdc599a89b0adced7d787f6 by Abseil Team <absl-team@google.com>:

Correct aarch64 macro check

The macro is __aarch64__, not __arch64__.

PiperOrigin-RevId: 414446225

--
a1536a57b64dfd53945d33a01cfc08b18c99c97b by Abseil Team <absl-team@google.com>:

Fix backwards comment in the last commit.

PiperOrigin-RevId: 414281214

--
11ac021ba779513667a31cf2563ddafc57d6d913 by Abseil Team <absl-team@google.com>:

AddressIsReadable() didn't work correctly on ARM when the given pointer was
misaligned at the end of the page.

Fix that by aligning the pointer on an 8-byte boundary before checking it.

PiperOrigin-RevId: 414203863
GitOrigin-RevId: 07240ca7822d007cdcc79f2c40bd58b2c2010348
Change-Id: If5f129194d59f5c9e5d84efd8cd9e17a70e072ab

1 files changed, 46 insertions, 56 deletions

diff --git a/absl/debugging/internal/address_is_readable.cc b/absl/debugging/internal/address_is_readable.cc
index 26df3289..4be6256b 100644
--- a/absl/debugging/internal/address_is_readable.cc
+++ b/absl/debugging/internal/address_is_readable.cc
@@ -30,16 +30,12 @@ bool AddressIsReadable(const void* /* addr */) { return true; }
 ABSL_NAMESPACE_END
 }  // namespace absl
 
-#else
+#else  // __linux__ && !__ANDROID__
 
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <sys/syscall.h>
-#include <sys/types.h>
+#include <stdint.h>
+#include <syscall.h>
 #include <unistd.h>
 
-#include <cerrno>
-
 #include "absl/base/internal/errno_saver.h"
 #include "absl/base/internal/raw_logging.h"
 
@@ -47,60 +43,54 @@ namespace absl {
 ABSL_NAMESPACE_BEGIN
 namespace debugging_internal {
 
+// NOTE: be extra careful about adding any interposable function calls here
+// (such as open(), read(), etc.). These symbols may be interposed and will get
+// invoked in contexts they don't expect.
+//
+// NOTE: any new system calls here may also require sandbox reconfiguration.
+//
 bool AddressIsReadable(const void *addr) {
-  int fd = 0;
+  // Align address on 8-byte boundary. On aarch64, checking last
+  // byte before inaccessible page returned unexpected EFAULT.
+  const uintptr_t u_addr = reinterpret_cast<uintptr_t>(addr) & ~7;
+  addr = reinterpret_cast<const void *>(u_addr);
+
+  // rt_sigprocmask below will succeed for this input.
+  if (addr == nullptr) return false;
+
   absl::base_internal::ErrnoSaver errno_saver;
-  for (int j = 0; j < 2; j++) {
-    // Here we probe with some syscall which
-    // - accepts a one-byte region of user memory as input
-    // - tests for EFAULT before other validation
-    // - has no problematic side-effects
-    //
-    // connect(2) works for this.  It copies the address into kernel
-    // memory before any validation beyond requiring an open fd.
-    // But a one byte address is never valid (sa_family is two bytes),
-    // so the call cannot succeed and change any state.
-    //
-    // This strategy depends on Linux implementation details,
-    // so we rely on the test to alert us if it stops working.
-    //
-    // Some discarded past approaches:
-    // - msync() doesn't reject PROT_NONE regions
-    // - write() on /dev/null doesn't return EFAULT
-    // - write() on a pipe requires creating it and draining the writes
-    //
-    // Use syscall(SYS_connect, ...) instead of connect() to prevent ASAN
-    // and other checkers from complaining about accesses to arbitrary memory.
-    do {
-      ABSL_RAW_CHECK(syscall(SYS_connect, fd, addr, 1) == -1,
-                     "should never succeed");
-    } while (errno == EINTR);
-    if (errno == EFAULT) return false;
-    if (errno == EBADF) {
-      if (j != 0) {
-        // Unclear what happened.
-        ABSL_RAW_LOG(ERROR, "unexpected EBADF on fd %d", fd);
-        return false;
-      }
-      // fd 0 must have been closed. Try opening it again.
-      // Note: we shouldn't leak too many file descriptors here, since we expect
-      // to get fd==0 reopened.
-      fd = open("/dev/null", O_RDONLY);
-      if (fd == -1) {
-        ABSL_RAW_LOG(ERROR, "can't open /dev/null");
-        return false;
-      }
-    } else {
-      // probably EINVAL or ENOTSOCK; we got past EFAULT validation.
-      return true;
-    }
-  }
-  ABSL_RAW_CHECK(false, "unreachable");
-  return false;
+
+  // Here we probe with some syscall which
+  // - accepts an 8-byte region of user memory as input
+  // - tests for EFAULT before other validation
+  // - has no problematic side-effects
+  //
+  // rt_sigprocmask(2) works for this.  It copies sizeof(kernel_sigset_t)==8
+  // bytes from the address into the kernel memory before any validation.
+  //
+  // The call can never succeed, since the `how` parameter is not one of
+  // SIG_BLOCK, SIG_UNBLOCK, SIG_SETMASK.
+  //
+  // This strategy depends on Linux implementation details,
+  // so we rely on the test to alert us if it stops working.
+  //
+  // Some discarded past approaches:
+  // - msync() doesn't reject PROT_NONE regions
+  // - write() on /dev/null doesn't return EFAULT
+  // - write() on a pipe requires creating it and draining the writes
+  // - connect() works but is problematic for sandboxes and needs a valid
+  //   file descriptor
+  //
+  // This can never succeed (invalid first argument to sigprocmask).
+  ABSL_RAW_CHECK(syscall(SYS_rt_sigprocmask, ~0, addr, nullptr,
+                         /*sizeof(kernel_sigset_t)*/ 8) == -1,
+                 "unexpected success");
+  ABSL_RAW_CHECK(errno == EFAULT || errno == EINVAL, "unexpected errno");
+  return errno != EFAULT;
 }
 
 }  // namespace debugging_internal
 ABSL_NAMESPACE_END
 }  // namespace absl
 
-#endif
+#endif  // __linux__ && !__ANDROID__