[libc] implement cached process/thread identity (#98989)

migrated from #95965 due to
corrupted git history

Author: SchrodingerZhu

libc/config/config.json

@@ -75,6 +75,16 @@
     "LIBC_CONF_FREELIST_MALLOC_BUFFER_SIZE": {
       "value": 1073741824,
       "doc": "Default size for the constinit freelist buffer used for the freelist malloc implementation (default 1o 1GB)."
+    },
+  },
+  "unistd": {
+    "LIBC_CONF_ENABLE_TID_CACHE": {
+      "value": true,
+      "doc": "Enable caching mechanism for gettid to avoid syscall (only effective in fullbuild mode, default to true). Please refer to Undefined Behavior documentation for implications."
+    },
+    "LIBC_CONF_ENABLE_PID_CACHE": {
+      "value": true,
+      "doc": "Enable caching mechanism for getpid to avoid syscall (default to true). Please refer to Undefined Behavior documentation for implications."
     }
   },
   "math": {

libc/config/linux/aarch64/entrypoints.txt

@@ -297,6 +297,7 @@ set(TARGET_LIBC_ENTRYPOINTS
     libc.src.unistd.geteuid
     libc.src.unistd.getpid
     libc.src.unistd.getppid
+    libc.src.unistd.gettid
     libc.src.unistd.getuid
     libc.src.unistd.isatty
     libc.src.unistd.link

libc/config/linux/riscv/entrypoints.txt

@@ -296,6 +296,7 @@ set(TARGET_LIBC_ENTRYPOINTS
     libc.src.unistd.geteuid
     libc.src.unistd.getpid
     libc.src.unistd.getppid
+    libc.src.unistd.gettid
     libc.src.unistd.getuid
     libc.src.unistd.isatty
     libc.src.unistd.link

libc/config/linux/x86_64/entrypoints.txt

@@ -315,6 +315,7 @@ set(TARGET_LIBC_ENTRYPOINTS
     libc.src.unistd.geteuid
     libc.src.unistd.getpid
     libc.src.unistd.getppid
+    libc.src.unistd.gettid
     libc.src.unistd.getuid
     libc.src.unistd.isatty
     libc.src.unistd.link

libc/docs/configure.rst

@@ -52,3 +52,6 @@ to learn about the defaults for your platform and target.
 * **"string" options**
     - ``LIBC_CONF_MEMSET_X86_USE_SOFTWARE_PREFETCHING``: Inserts prefetch for write instructions (PREFETCHW) for memset on x86 to recover performance when hardware prefetcher is disabled.
     - ``LIBC_CONF_STRING_UNSAFE_WIDE_READ``: Read more than a byte at a time to perform byte-string operations like strlen.
+* **"unistd" options**
+    - ``LIBC_CONF_ENABLE_PID_CACHE``: Enable caching mechanism for getpid to avoid syscall (default to true). Please refer to Undefined Behavior documentation for implications.
+    - ``LIBC_CONF_ENABLE_TID_CACHE``: Enable caching mechanism for gettid to avoid syscall (only effective in fullbuild mode, default to true). Please refer to Undefined Behavior documentation for implications.

libc/docs/dev/undefined_behavior.rst

@@ -93,3 +93,26 @@ direction in this case.
 Non-const Constant Return Values
 --------------------------------
 Some libc functions, like ``dlerror()``, return ``char *`` instead of ``const char *`` and then tell the caller they promise not to to modify this value. Any modification of this value is undefined behavior.
+
+Cached ``getpid/gettid``
+------------------------
+Since version ``2.25``, glibc removes its cache mechanism for ``getpid/gettid`` 
+(See the history section in https://man7.org/linux/man-pages/man2/getpid.2.html).
+LLVM's libc still implements the cache as it is useful for fast deadlock detection.
+The cache mechanism is also implemented in MUSL and bionic. The tid/pid cache can 
+be disabled by setting ``LIBC_CONF_ENABLE_TID_CACHE`` and ``LIBC_CONF_ENABLE_PID_CACHE``
+to ``false`` respectively.
+
+Unwrapped ``SYS_clone/SYS_fork/SYS_vfork``
+------------------------------------------
+It is highly discouraged to use unwrapped ``SYS_clone/SYS_fork/SYS_vfork``. 
+First, calling such syscalls without provided libc wrappers ignores 
+all the ``pthread_atfork`` entries as libc can no longer detect the ``fork``. 
+Second, libc relies on the ``fork/clone`` wrappers to correctly maintain cache for
+process id and thread id, and other important process-specific states such as the list 
+of robust mutexes. Third, even if the user is to call ``exec*`` functions immediately, 
+there can still be other unexpected issues. For instance, there can be signal handlers 
+inherited from parent process triggered inside the instruction window between ``fork`` 
+and ``exec*``. As libc failed to maintain its internal states correctly, even though the
+functions used inside the signal handlers are marked as ``async-signal-safe`` (such as
+``getpid``), they will still return wrong values or lead to other even worse situations.

libc/spec/posix.td

@@ -546,6 +546,11 @@ def POSIX : StandardSpec<"POSIX"> {
           RetValSpec<PidT>,
           [ArgSpec<VoidType>]
         >,
+        FunctionSpec<
+          "gettid",
+          RetValSpec<PidT>,
+          [ArgSpec<VoidType>]
+        >,
         FunctionSpec<
           "getuid",
           RetValSpec<UidT>,
@@ -601,16 +606,6 @@ def POSIX : StandardSpec<"POSIX"> {
           RetValSpec<IntType>,
           [ArgSpec<ConstCharPtr>]
         >,
-        FunctionSpec<
-          "getpid",
-          RetValSpec<IntType>,
-          [ArgSpec<VoidType>]
-        >,
-        FunctionSpec<
-          "getppid",
-          RetValSpec<IntType>,
-          [ArgSpec<VoidType>]
-        >,
         FunctionSpec<
           "link",
           RetValSpec<IntType>,

libc/src/__support/OSUtil/CMakeLists.txt

@@ -15,3 +15,20 @@ add_object_library(
   DEPENDS
     ${target_os_util}
 )
+
+if (LIBC_CONF_ENABLE_PID_CACHE)
+  set(libc_copt_enable_pid_cache 1)
+else()
+  set(libc_copt_enable_pid_cache 0)
+endif()
+
+if(TARGET libc.src.__support.OSUtil.${LIBC_TARGET_OS}.pid)
+  add_object_library(
+    pid
+    ALIAS
+    DEPENDS
+      .${LIBC_TARGET_OS}.pid
+    COMPILE_OPTIONS
+      -DLIBC_COPT_ENABLE_PID_CACHE=${libc_copt_enable_pid_cache}
+  )
+endif()

libc/src/__support/OSUtil/linux/CMakeLists.txt

@@ -23,3 +23,16 @@ add_object_library(
     libc.hdr.types.struct_f_owner_ex
     libc.hdr.types.off_t
 )
+
+add_object_library(
+  pid
+  SRCS
+    pid.cpp
+  HDRS
+    ../pid.h
+  DEPENDS
+    libc.src.__support.OSUtil.osutil
+    libc.src.__support.common
+    libc.hdr.types.pid_t
+    libc.include.sys_syscall
+)

libc/src/__support/OSUtil/linux/pid.cpp

@@ -0,0 +1,20 @@
+//===------------ pid_t utilities implementation ----------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "src/__support/OSUtil/pid.h"
+#include "src/__support/OSUtil/syscall.h"
+#include <sys/syscall.h>
+
+namespace LIBC_NAMESPACE_DECL {
+
+pid_t ProcessIdentity::cache = -1;
+pid_t ProcessIdentity::get_uncached() {
+  return syscall_impl<pid_t>(SYS_getpid);
+}
+
+} // namespace LIBC_NAMESPACE_DECL