Implement cache index randomization for large allocations.

Extract szad size quantization into {extent,run}_quantize(), and .
quantize szad run sizes to the union of valid small region run sizes and
large run sizes.

Refactor iteration in arena_run_first_fit() to use
run_quantize{,_first,_next(), and add support for padded large runs.

For large allocations that have no specified alignment constraints,
compute a pseudo-random offset from the beginning of the first backing
page that is a multiple of the cache line size.  Under typical
configurations with 4-KiB pages and 64-byte cache lines this results in
a uniform distribution among 64 page boundary offsets.

Add the --disable-cache-oblivious option, primarily intended for
performance testing.

This resolves #13.

Author: jasone

ChangeLog

@@ -101,6 +101,9 @@ found in the git revision history:
     run fragmentation, smaller runs reduce external fragmentation for small size
     classes, and packed (less uniformly aligned) metadata layout improves CPU
     cache set distribution.
+  - Randomly distribute large allocation base pointer alignment relative to page
+    boundaries in order to more uniformly utilize CPU cache sets.  This can be
+    disabled via the --disable-cache-oblivious configure option.
   - Micro-optimize the fast paths for the public API functions.
   - Refactor thread-specific data to reside in a single structure.  This assures
     that only a single TLS read is necessary per call into the public API.

INSTALL

@@ -185,6 +185,15 @@ any of the following arguments (not a definitive list) to 'configure':
     thread-local variables via the __thread keyword.  If TLS is available,
     jemalloc uses it for several purposes.
 
+--disable-cache-oblivious
+    Disable cache-oblivious large allocation alignment for large allocation
+    requests with no alignment constraints.  If this feature is disabled, all
+    large allocations are page-aligned as an implementation artifact, which can
+    severely harm CPU cache utilization.  However, the cache-oblivious layout
+    comes at the cost of one extra page per large allocation, which in the
+    most extreme case increases physical memory usage for the 16 KiB size class
+    to 20 KiB.
+
 --with-xslroot=<path>
     Specify where to find DocBook XSL stylesheets when building the
     documentation.

configure.ac

@@ -952,6 +952,23 @@ if test "x$enable_xmalloc" = "x1" ; then
 fi
 AC_SUBST([enable_xmalloc])
 
+dnl Support cache-oblivious allocation alignment by default.
+AC_ARG_ENABLE([cache-oblivious],
+  [AS_HELP_STRING([--disable-cache-oblivious],
+                  [Disable support for cache-oblivious allocation alignment])],
+[if test "x$enable_cache_oblivious" = "xno" ; then
+  enable_cache_oblivious="0"
+else
+  enable_cache_oblivious="1"
+fi
+],
+[enable_cache_oblivious="1"]
+)
+if test "x$enable_cache_oblivious" = "x1" ; then
+  AC_DEFINE([JEMALLOC_CACHE_OBLIVIOUS], [ ])
+fi
+AC_SUBST([enable_cache_oblivious])
+
 dnl ============================================================================
 dnl Check for  __builtin_ffsl(), then ffsl(3), and fail if neither are found.
 dnl One of those two functions should (theoretically) exist on all platforms
@@ -1663,4 +1680,5 @@ AC_MSG_RESULT([xmalloc            : ${enable_xmalloc}])
 AC_MSG_RESULT([munmap             : ${enable_munmap}])
 AC_MSG_RESULT([lazy_lock          : ${enable_lazy_lock}])
 AC_MSG_RESULT([tls                : ${enable_tls}])
+AC_MSG_RESULT([cache-oblivious    : ${enable_cache_oblivious}])
 AC_MSG_RESULT([===============================================================================])

include/jemalloc/internal/arena.h

@@ -290,6 +290,12 @@ struct arena_s {
 
 	uint64_t		prof_accumbytes;
 
+	/*
+	 * PRNG state for cache index randomization of large allocation base
+	 * pointers.
+	 */
+	uint64_t		offset_state;
+
 	dss_prec_t		dss_prec;
 
 	/*
@@ -394,7 +400,15 @@ struct arena_s {
 /******************************************************************************/
 #ifdef JEMALLOC_H_EXTERNS
 
-extern ssize_t	opt_lg_dirty_mult;
+static const size_t	large_pad =
+#ifdef JEMALLOC_CACHE_OBLIVIOUS
+    PAGE
+#else
+    0
+#endif
+    ;
+
+extern ssize_t		opt_lg_dirty_mult;
 
 extern arena_bin_info_t	arena_bin_info[NBINS];
 
@@ -475,7 +489,7 @@ void	arena_stats_merge(arena_t *arena, const char **dss,
     arena_stats_t *astats, malloc_bin_stats_t *bstats,
     malloc_large_stats_t *lstats, malloc_huge_stats_t *hstats);
 arena_t	*arena_new(unsigned ind);
-void	arena_boot(void);
+bool	arena_boot(void);
 void	arena_prefork(arena_t *arena);
 void	arena_postfork_parent(arena_t *arena);
 void	arena_postfork_child(arena_t *arena);
@@ -721,7 +735,7 @@ arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size,
 {
 	size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind);
 
-	assert((size & PAGE_MASK) == 0);
+	assert(size == PAGE_CEILING(size));
 	assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0);
 	assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags);
 	arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags);
@@ -734,7 +748,7 @@ arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind,
 	size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind);
 	size_t mapbits = arena_mapbitsp_read(mapbitsp);
 
-	assert((size & PAGE_MASK) == 0);
+	assert(size == PAGE_CEILING(size));
 	assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0);
 	arena_mapbitsp_write(mapbitsp, size | (mapbits & PAGE_MASK));
 }
@@ -747,7 +761,7 @@ arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size,
 	size_t mapbits = arena_mapbitsp_read(mapbitsp);
 	size_t unzeroed;
 
-	assert((size & PAGE_MASK) == 0);
+	assert(size == PAGE_CEILING(size));
 	assert((flags & CHUNK_MAP_DIRTY) == flags);
 	unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */
 	arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags
@@ -762,7 +776,8 @@ arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind,
 	size_t mapbits = arena_mapbitsp_read(mapbitsp);
 
 	assert(binind <= BININD_INVALID);
-	assert(arena_mapbits_large_size_get(chunk, pageind) == LARGE_MINCLASS);
+	assert(arena_mapbits_large_size_get(chunk, pageind) == LARGE_MINCLASS +
+	    large_pad);
 	arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_BININD_MASK) |
 	    (binind << CHUNK_MAP_BININD_SHIFT));
 }
@@ -1107,13 +1122,16 @@ arena_salloc(const void *ptr, bool demote)
 			 * end up looking at binind to determine that ptr is a
 			 * small allocation.
 			 */
-			assert(((uintptr_t)ptr & PAGE_MASK) == 0);
-			ret = arena_mapbits_large_size_get(chunk, pageind);
+			assert(config_cache_oblivious || ((uintptr_t)ptr &
+			    PAGE_MASK) == 0);
+			ret = arena_mapbits_large_size_get(chunk, pageind) -
+			    large_pad;
 			assert(ret != 0);
-			assert(pageind + (ret>>LG_PAGE) <= chunk_npages);
+			assert(pageind + ((ret+large_pad)>>LG_PAGE) <=
+			    chunk_npages);
 			assert(arena_mapbits_dirty_get(chunk, pageind) ==
 			    arena_mapbits_dirty_get(chunk,
-			    pageind+(ret>>LG_PAGE)-1));
+			    pageind+((ret+large_pad)>>LG_PAGE)-1));
 		} else {
 			/*
 			 * Small allocation (possibly promoted to a large
@@ -1157,11 +1175,13 @@ arena_dalloc(tsd_t *tsd, void *ptr, tcache_t *tcache)
 			size_t size = arena_mapbits_large_size_get(chunk,
 			    pageind);
 
-			assert(((uintptr_t)ptr & PAGE_MASK) == 0);
+			assert(config_cache_oblivious || ((uintptr_t)ptr &
+			    PAGE_MASK) == 0);
 
-			if (likely(tcache != NULL) && size <= tcache_maxclass)
-				tcache_dalloc_large(tsd, tcache, ptr, size);
-			else {
+			if (likely(tcache != NULL) && size <= tcache_maxclass) {
+				tcache_dalloc_large(tsd, tcache, ptr, size -
+				    large_pad);
+			} else {
 				arena_dalloc_large(extent_node_arena_get(
 				    &chunk->node), chunk, ptr);
 			}
@@ -1188,7 +1208,7 @@ arena_sdalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache)
 				 */
 				assert(((uintptr_t)ptr & PAGE_MASK) == 0);
 				size = arena_mapbits_large_size_get(chunk,
-				    pageind);
+				    pageind) - large_pad;
 			}
 		}
 		assert(s2u(size) == s2u(arena_salloc(ptr, false)));
@@ -1205,7 +1225,8 @@ arena_sdalloc(tsd_t *tsd, void *ptr, size_t size, tcache_t *tcache)
 				    &chunk->node), chunk, ptr, pageind);
 			}
 		} else {
-			assert(((uintptr_t)ptr & PAGE_MASK) == 0);
+			assert(config_cache_oblivious || ((uintptr_t)ptr &
+			    PAGE_MASK) == 0);
 
 			if (likely(tcache != NULL) && size <= tcache_maxclass)
 				tcache_dalloc_large(tsd, tcache, ptr, size);

include/jemalloc/internal/jemalloc_internal.h.in

@@ -126,6 +126,13 @@ static const bool config_ivsalloc =
     false
 #endif
     ;
+static const bool config_cache_oblivious =
+#ifdef JEMALLOC_CACHE_OBLIVIOUS
+    true
+#else
+    false
+#endif
+    ;
 
 #ifdef JEMALLOC_C11ATOMICS
 #include <stdatomic.h>

include/jemalloc/internal/jemalloc_internal_defs.h.in

@@ -192,6 +192,12 @@
  */
 #undef JEMALLOC_IVSALLOC
 
+/*
+ * If defined, explicitly attempt to more uniformly distribute large allocation
+ * pointer alignments across all cache indices.
+ */
+#undef JEMALLOC_CACHE_OBLIVIOUS
+
 /*
  * Darwin (OS X) uses zones to work around Mach-O symbol override shortcomings.
  */

include/jemalloc/internal/prng.h

@@ -26,22 +26,22 @@
  *   const uint32_t a, c : See above discussion.
  */
 #define	prng32(r, lg_range, state, a, c) do {				\
-	assert(lg_range > 0);						\
-	assert(lg_range <= 32);						\
+	assert((lg_range) > 0);						\
+	assert((lg_range) <= 32);					\
 									\
 	r = (state * (a)) + (c);					\
 	state = r;							\
-	r >>= (32 - lg_range);						\
+	r >>= (32 - (lg_range));					\
 } while (false)
 
 /* Same as prng32(), but 64 bits of pseudo-randomness, using uint64_t. */
 #define	prng64(r, lg_range, state, a, c) do {				\
-	assert(lg_range > 0);						\
-	assert(lg_range <= 64);						\
+	assert((lg_range) > 0);						\
+	assert((lg_range) <= 64);					\
 									\
 	r = (state * (a)) + (c);					\
 	state = r;							\
-	r >>= (64 - lg_range);						\
+	r >>= (64 - (lg_range));					\
 } while (false)
 
 #endif /* JEMALLOC_H_TYPES */