[libc] Add a config option to disable slab reclaiming

Summary:
Without slab reclaiming this interface is much simpler and it can speed
up cases with a lot of churn. Basically, wastes memory for performance.

Author: jhuber6

libc/config/config.json

@@ -119,6 +119,12 @@
       "doc": "Force the size of time_t to 64 bits, even on platforms where compatibility considerations would otherwise make it 32-bit."
     }
   },
+  "malloc": {
+    "LIBC_CONF_GPU_MALLOC_DISABLE_SLAB_RECLAIMING": {
+      "value": false,
+      "doc": "The malloc implementation will return unused slabs to system memory."
+    }
+  },
   "general": {
     "LIBC_ADD_NULL_CHECKS": {
       "value": true,

libc/config/gpu/amdgpu/config.json

@@ -36,5 +36,10 @@
     "LIBC_CONF_MATH_OPTIMIZATIONS": {
       "value": "(LIBC_MATH_SKIP_ACCURATE_PASS | LIBC_MATH_SMALL_TABLES | LIBC_MATH_NO_ERRNO | LIBC_MATH_NO_EXCEPT)"
     }
+  },
+  "malloc": {
+    "LIBC_CONF_GPU_MALLOC_DISABLE_SLAB_RECLAIMING": {
+      "value": false
+    }
   }
 }

libc/config/gpu/nvptx/config.json

@@ -36,5 +36,10 @@
     "LIBC_CONF_MATH_OPTIMIZATIONS": {
       "value": "(LIBC_MATH_SKIP_ACCURATE_PASS | LIBC_MATH_SMALL_TABLES | LIBC_MATH_NO_ERRNO | LIBC_MATH_NO_EXCEPT)"
     }
+  },
+  "malloc": {
+    "LIBC_CONF_GPU_MALLOC_DISABLE_SLAB_RECLAIMING": {
+      "value": false
+    }
   }
 }

libc/docs/configure.rst

@@ -32,6 +32,8 @@ to learn about the defaults for your platform and target.
     - ``LIBC_CONF_ERRNO_MODE``: The implementation used for errno, acceptable values are LIBC_ERRNO_MODE_DEFAULT, LIBC_ERRNO_MODE_UNDEFINED, LIBC_ERRNO_MODE_THREAD_LOCAL, LIBC_ERRNO_MODE_SHARED, LIBC_ERRNO_MODE_EXTERNAL, LIBC_ERRNO_MODE_SYSTEM, and LIBC_ERRNO_MODE_SYSTEM_INLINE.
 * **"general" options**
     - ``LIBC_ADD_NULL_CHECKS``: Add nullptr checks in the library's implementations to some functions for which passing nullptr is undefined behavior.
+* **"malloc" options**
+    - ``LIBC_CONF_GPU_MALLOC_DISABLE_SLAB_RECLAIMING``: The malloc implementation will return unused slabs to system memory.
 * **"math" options**
     - ``LIBC_CONF_FREXP_INF_NAN_EXPONENT``: The value written back to the second parameter when calling frexp/frexpf/frexpl` with `+/-Inf`/`NaN` is unspecified.  Configure an explicit exp value for Inf/NaN inputs.
     - ``LIBC_CONF_MATH_OPTIMIZATIONS``: Configures optimizations for math functions. Values accepted are LIBC_MATH_SKIP_ACCURATE_PASS, LIBC_MATH_SMALL_TABLES, LIBC_MATH_NO_ERRNO, LIBC_MATH_NO_EXCEPT, and LIBC_MATH_FAST.

libc/src/__support/GPU/CMakeLists.txt

@@ -3,6 +3,10 @@ if(NOT LIBC_TARGET_OS_IS_GPU)
   return()
 endif()
 
+if(LIBC_CONF_MALLOC_DISABLE_SLAB_RECLAIMING)
+  list(APPEND malloc_config_copts "-DLIBC_CONF_MALLOC_DISABLE_SLAB_RECLAIMING")
+endif()
+
 add_header_library(
   utils
   HDRS
@@ -23,4 +27,5 @@ add_object_library(
     libc.src.__support.CPP.bit
     libc.src.__support.CPP.new
     .utils
+  ${malloc_config_copts}
 )

libc/src/__support/GPU/allocator.cpp

@@ -39,6 +39,13 @@ constexpr static uint32_t MIN_ALIGNMENT = MIN_SIZE - 1;
 // The number of times to attempt claiming an in-progress slab allocation.
 constexpr static uint32_t MAX_TRIES = 1024;
 
+// Configuration for whether or not we will return unused slabs to memory.
+#ifdef LIBC_CONF_MALLOC_DISABLE_SLAB_RECLAIMING
+constexpr static bool RECLAIM = false;
+#else
+constexpr static bool RECLAIM = true;
+#endif
+
 static_assert(!(ARRAY_SIZE & (ARRAY_SIZE - 1)), "Must be a power of two");
 
 namespace impl {
@@ -368,7 +375,7 @@ struct GuardPtr {
       // and obtain exclusive rights to deconstruct it. If the CAS failed either
       // another thread resurrected the counter and we quit, or a parallel read
       // helped us invalidating it. For the latter, claim that flag and return.
-      if (counter.fetch_sub(n, cpp::MemoryOrder::RELAXED) == n) {
+      if (counter.fetch_sub(n, cpp::MemoryOrder::RELAXED) == n && RECLAIM) {
         uint32_t expected = 0;
         if (counter.compare_exchange_strong(expected, INVALID,
                                             cpp::MemoryOrder::RELAXED,
@@ -386,8 +393,9 @@ struct GuardPtr {
     // thread.
     uint64_t read() {
       auto val = counter.load(cpp::MemoryOrder::RELAXED);
-      if (val == 0 && counter.compare_exchange_strong(
-                          val, INVALID | HELPED, cpp::MemoryOrder::RELAXED))
+      if (val == 0 && RECLAIM &&
+          counter.compare_exchange_strong(val, INVALID | HELPED,
+                                          cpp::MemoryOrder::RELAXED))
         return 0;
       return (val & INVALID) ? 0 : val;
     }
@@ -421,7 +429,7 @@ struct GuardPtr {
       return nullptr;
 
     cpp::atomic_thread_fence(cpp::MemoryOrder::ACQUIRE);
-    return ptr.load(cpp::MemoryOrder::RELAXED);
+    return RECLAIM ? ptr.load(cpp::MemoryOrder::RELAXED) : expected;
   }
 
   // Finalize the associated memory and signal that it is ready to use by