Restructure source tree.

jemalloc/include/jemalloc/internal/arena.h

@@ -0,0 +1,487 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+/*
+ * Subpages are an artificially designated partitioning of pages.  Their only
+ * purpose is to support subpage-spaced size classes.
+ *
+ * There must be at least 4 subpages per page, due to the way size classes are
+ * handled.
+ */
+#define	LG_SUBPAGE		8
+#define	SUBPAGE			((size_t)(1U << LG_SUBPAGE))
+#define	SUBPAGE_MASK		(SUBPAGE - 1)
+
+/* Return the smallest subpage multiple that is >= s. */
+#define	SUBPAGE_CEILING(s)						\
+	(((s) + SUBPAGE_MASK) & ~SUBPAGE_MASK)
+
+#ifdef JEMALLOC_TINY
+   /* Smallest size class to support. */
+#  define LG_TINY_MIN		1
+#endif
+
+/*
+ * Maximum size class that is a multiple of the quantum, but not (necessarily)
+ * a power of 2.  Above this size, allocations are rounded up to the nearest
+ * power of 2.
+ */
+#define	LG_QSPACE_MAX_DEFAULT	7
+
+/*
+ * Maximum size class that is a multiple of the cacheline, but not (necessarily)
+ * a power of 2.  Above this size, allocations are rounded up to the nearest
+ * power of 2.
+ */
+#define	LG_CSPACE_MAX_DEFAULT	9
+
+/*
+ * Maximum medium size class.  This must not be more than 1/4 of a chunk
+ * (LG_MEDIUM_MAX_DEFAULT <= LG_CHUNK_DEFAULT - 2).
+ */
+#define	LG_MEDIUM_MAX_DEFAULT	15
+
+/* Return the smallest medium size class that is >= s. */
+#define	MEDIUM_CEILING(s)						\
+	(((s) + mspace_mask) & ~mspace_mask)
+
+/*
+ * Soft limit on the number of medium size classes.  Spacing between medium
+ * size classes never exceeds pagesize, which can force more than NBINS_MAX
+ * medium size classes.
+ */
+#define	NMBINS_MAX	16
+
+/*
+ * RUN_MAX_OVRHD indicates maximum desired run header overhead.  Runs are sized
+ * as small as possible such that this setting is still honored, without
+ * violating other constraints.  The goal is to make runs as small as possible
+ * without exceeding a per run external fragmentation threshold.
+ *
+ * We use binary fixed point math for overhead computations, where the binary
+ * point is implicitly RUN_BFP bits to the left.
+ *
+ * Note that it is possible to set RUN_MAX_OVRHD low enough that it cannot be
+ * honored for some/all object sizes, since there is one bit of header overhead
+ * per object (plus a constant).  This constraint is relaxed (ignored) for runs
+ * that are so small that the per-region overhead is greater than:
+ *
+ *   (RUN_MAX_OVRHD / (reg_size << (3+RUN_BFP))
+ */
+#define	RUN_BFP			12
+/*                                    \/   Implicit binary fixed point. */
+#define	RUN_MAX_OVRHD		0x0000003dU
+#define	RUN_MAX_OVRHD_RELAX	0x00001800U
+
+/* Put a cap on small object run size.  This overrides RUN_MAX_OVRHD. */
+#define	RUN_MAX_SMALL							\
+	(arena_maxclass <= (1U << (CHUNK_MAP_LG_PG_RANGE + PAGE_SHIFT))	\
+	    ? arena_maxclass : (1U << (CHUNK_MAP_LG_PG_RANGE +		\
+	    PAGE_SHIFT)))
+
+/*
+ * The minimum ratio of active:dirty pages per arena is computed as:
+ *
+ *   (nactive >> opt_lg_dirty_mult) >= ndirty
+ *
+ * So, supposing that opt_lg_dirty_mult is 5, there can be no less than 32
+ * times as many active pages as dirty pages.
+ */
+#define	LG_DIRTY_MULT_DEFAULT	5
+
+typedef struct arena_chunk_map_s arena_chunk_map_t;
+typedef struct arena_chunk_s arena_chunk_t;
+typedef struct arena_run_s arena_run_t;
+typedef struct arena_bin_s arena_bin_t;
+typedef struct arena_s arena_t;
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+/* Each element of the chunk map corresponds to one page within the chunk. */
+struct arena_chunk_map_s {
+	/*
+	 * Linkage for run trees.  There are two disjoint uses:
+	 *
+	 * 1) arena_t's runs_avail tree.
+	 * 2) arena_run_t conceptually uses this linkage for in-use non-full
+	 *    runs, rather than directly embedding linkage.
+	 */
+	rb_node(arena_chunk_map_t)	link;
+
+#ifdef JEMALLOC_PROF
+	/* Profile counters, used for large object runs. */
+	prof_thr_cnt_t			*prof_cnt;
+#endif
+
+	/*
+	 * Run address (or size) and various flags are stored together.  The bit
+	 * layout looks like (assuming 32-bit system):
+	 *
+	 *   ???????? ???????? ????cccc ccccdzla
+	 *
+	 * ? : Unallocated: Run address for first/last pages, unset for internal
+	 *                  pages.
+	 *     Small/medium: Don't care.
+	 *     Large: Run size for first page, unset for trailing pages.
+	 * - : Unused.
+	 * c : refcount (could overflow for PAGE_SIZE >= 128 KiB)
+	 * d : dirty?
+	 * z : zeroed?
+	 * l : large?
+	 * a : allocated?
+	 *
+	 * Following are example bit patterns for the three types of runs.
+	 *
+	 * p : run page offset
+	 * s : run size
+	 * x : don't care
+	 * - : 0
+	 * [dzla] : bit set
+	 *
+	 *   Unallocated:
+	 *     ssssssss ssssssss ssss---- --------
+	 *     xxxxxxxx xxxxxxxx xxxx---- ----d---
+	 *     ssssssss ssssssss ssss---- -----z--
+	 *
+	 *   Small/medium:
+	 *     pppppppp ppppcccc cccccccc cccc---a
+	 *     pppppppp ppppcccc cccccccc cccc---a
+	 *     pppppppp ppppcccc cccccccc cccc---a
+	 *
+	 *   Large:
+	 *     ssssssss ssssssss ssss---- ------la
+	 *     -------- -------- -------- ------la
+	 *     -------- -------- -------- ------la
+	 */
+	size_t				bits;
+#define	CHUNK_MAP_PG_MASK	((size_t)0xfff00000U)
+#define	CHUNK_MAP_PG_SHIFT	20
+#define	CHUNK_MAP_LG_PG_RANGE	12
+
+#define	CHUNK_MAP_RC_MASK	((size_t)0xffff0U)
+#define	CHUNK_MAP_RC_ONE	((size_t)0x00010U)
+
+#define	CHUNK_MAP_FLAGS_MASK	((size_t)0xfU)
+#define	CHUNK_MAP_DIRTY		((size_t)0x8U)
+#define	CHUNK_MAP_ZEROED	((size_t)0x4U)
+#define	CHUNK_MAP_LARGE		((size_t)0x2U)
+#define	CHUNK_MAP_ALLOCATED	((size_t)0x1U)
+#define	CHUNK_MAP_KEY		(CHUNK_MAP_DIRTY | CHUNK_MAP_ALLOCATED)
+};
+typedef rb_tree(arena_chunk_map_t) arena_avail_tree_t;
+typedef rb_tree(arena_chunk_map_t) arena_run_tree_t;
+
+/* Arena chunk header. */
+struct arena_chunk_s {
+	/* Arena that owns the chunk. */
+	arena_t		*arena;
+
+	/* Linkage for the arena's chunks_dirty tree. */
+	rb_node(arena_chunk_t) link_dirty;
+
+	/*
+	 * True if the chunk is currently in the chunks_dirty tree, due to
+	 * having at some point contained one or more dirty pages.  Removal
+	 * from chunks_dirty is lazy, so (dirtied && ndirty == 0) is possible.
+	 */
+	bool		dirtied;
+
+	/* Number of dirty pages. */
+	size_t		ndirty;
+
+	/* Map of pages within chunk that keeps track of free/large/small. */
+	arena_chunk_map_t map[1]; /* Dynamically sized. */
+};
+typedef rb_tree(arena_chunk_t) arena_chunk_tree_t;
+
+struct arena_run_s {
+#ifdef JEMALLOC_DEBUG
+	uint32_t	magic;
+#  define ARENA_RUN_MAGIC 0x384adf93
+#endif
+
+	/* Bin this run is associated with. */
+	arena_bin_t	*bin;
+
+	/* Index of first element that might have a free region. */
+	unsigned	regs_minelm;
+
+	/* Number of free regions in run. */
+	unsigned	nfree;
+
+	/* Bitmask of in-use regions (0: in use, 1: free). */
+	unsigned	regs_mask[1]; /* Dynamically sized. */
+};
+
+struct arena_bin_s {
+	/*
+	 * Current run being used to service allocations of this bin's size
+	 * class.
+	 */
+	arena_run_t	*runcur;
+
+	/*
+	 * Tree of non-full runs.  This tree is used when looking for an
+	 * existing run when runcur is no longer usable.  We choose the
+	 * non-full run that is lowest in memory; this policy tends to keep
+	 * objects packed well, and it can also help reduce the number of
+	 * almost-empty chunks.
+	 */
+	arena_run_tree_t runs;
+
+	/* Size of regions in a run for this bin's size class. */
+	size_t		reg_size;
+
+	/* Total size of a run for this bin's size class. */
+	size_t		run_size;
+
+	/* Total number of regions in a run for this bin's size class. */
+	uint32_t	nregs;
+
+	/* Number of elements in a run's regs_mask for this bin's size class. */
+	uint32_t	regs_mask_nelms;
+
+#ifdef JEMALLOC_PROF
+	/*
+	 * Offset of first (prof_cnt_t *) in a run header for this bin's size
+	 * class, or 0 if (opt_prof == false).
+	 */
+	uint32_t	cnt0_offset;
+#endif
+
+	/* Offset of first region in a run for this bin's size class. */
+	uint32_t	reg0_offset;
+
+#ifdef JEMALLOC_STATS
+	/* Bin statistics. */
+	malloc_bin_stats_t stats;
+#endif
+};
+
+struct arena_s {
+#ifdef JEMALLOC_DEBUG
+	uint32_t		magic;
+#  define ARENA_MAGIC 0x947d3d24
+#endif
+
+	/* This arena's index within the arenas array. */
+	unsigned		ind;
+
+	/* All operations on this arena require that lock be locked. */
+	malloc_mutex_t		lock;
+
+#ifdef JEMALLOC_STATS
+	arena_stats_t		stats;
+#  ifdef JEMALLOC_TCACHE
+	/*
+	 * List of tcaches for extant threads associated with this arena.
+	 * Stats from these are merged incrementally, and at exit.
+	 */
+	ql_head(tcache_t)	tcache_ql;
+#  endif
+#endif
+
+#ifdef JEMALLOC_PROF
+	uint64_t		prof_accumbytes;
+#endif
+
+	/* Tree of dirty-page-containing chunks this arena manages. */
+	arena_chunk_tree_t	chunks_dirty;
+
+	/*
+	 * In order to avoid rapid chunk allocation/deallocation when an arena
+	 * oscillates right on the cusp of needing a new chunk, cache the most
+	 * recently freed chunk.  The spare is left in the arena's chunk trees
+	 * until it is deleted.
+	 *
+	 * There is one spare chunk per arena, rather than one spare total, in
+	 * order to avoid interactions between multiple threads that could make
+	 * a single spare inadequate.
+	 */
+	arena_chunk_t		*spare;
+
+	/* Number of pages in active runs. */
+	size_t			nactive;
+
+	/*
+	 * Current count of pages within unused runs that are potentially
+	 * dirty, and for which madvise(... MADV_DONTNEED) has not been called.
+	 * By tracking this, we can institute a limit on how much dirty unused
+	 * memory is mapped for each arena.
+	 */
+	size_t			ndirty;
+
+	/*
+	 * Size/address-ordered tree of this arena's available runs.  This tree
+	 * is used for first-best-fit run allocation.
+	 */
+	arena_avail_tree_t	runs_avail;
+
+	/*
+	 * bins is used to store trees of free regions of the following sizes,
+	 * assuming a 16-byte quantum, 4 KiB page size, and default
+	 * JEMALLOC_OPTIONS.
+	 *
+	 *   bins[i] |   size |
+	 *   --------+--------+
+	 *        0  |      2 |
+	 *        1  |      4 |
+	 *        2  |      8 |
+	 *   --------+--------+
+	 *        3  |     16 |
+	 *        4  |     32 |
+	 *        5  |     48 |
+	 *           :        :
+	 *        8  |     96 |
+	 *        9  |    112 |
+	 *       10  |    128 |
+	 *   --------+--------+
+	 *       11  |    192 |
+	 *       12  |    256 |
+	 *       13  |    320 |
+	 *       14  |    384 |
+	 *       15  |    448 |
+	 *       16  |    512 |
+	 *   --------+--------+
+	 *       17  |    768 |
+	 *       18  |   1024 |
+	 *       19  |   1280 |
+	 *           :        :
+	 *       27  |   3328 |
+	 *       28  |   3584 |
+	 *       29  |   3840 |
+	 *   --------+--------+
+	 *       30  |  4 KiB |
+	 *       31  |  6 KiB |
+	 *       33  |  8 KiB |
+	 *           :        :
+	 *       43  | 28 KiB |
+	 *       44  | 30 KiB |
+	 *       45  | 32 KiB |
+	 *   --------+--------+
+	 */
+	arena_bin_t		bins[1]; /* Dynamically sized. */
+};
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern size_t	opt_lg_qspace_max;
+extern size_t	opt_lg_cspace_max;
+extern size_t	opt_lg_medium_max;
+extern ssize_t		opt_lg_dirty_mult;
+extern uint8_t const	*small_size2bin;
+
+/* Various bin-related settings. */
+#ifdef JEMALLOC_TINY		/* Number of (2^n)-spaced tiny bins. */
+#  define		ntbins	((unsigned)(LG_QUANTUM - LG_TINY_MIN))
+#else
+#  define		ntbins	0
+#endif
+extern unsigned		nqbins; /* Number of quantum-spaced bins. */
+extern unsigned		ncbins; /* Number of cacheline-spaced bins. */
+extern unsigned		nsbins; /* Number of subpage-spaced bins. */
+extern unsigned		nmbins; /* Number of medium bins. */
+extern unsigned		nbins;
+extern unsigned		mbin0; /* mbin offset (nbins - nmbins). */
+#ifdef JEMALLOC_TINY
+#  define		tspace_max	((size_t)(QUANTUM >> 1))
+#endif
+#define			qspace_min	QUANTUM
+extern size_t		qspace_max;
+extern size_t		cspace_min;
+extern size_t		cspace_max;
+extern size_t		sspace_min;
+extern size_t		sspace_max;
+#define			small_maxclass	sspace_max
+#define			medium_min	PAGE_SIZE
+extern size_t		medium_max;
+#define			bin_maxclass	medium_max
+
+/* Spacing between medium size classes. */
+extern size_t		lg_mspace;
+extern size_t		mspace_mask;
+
+#define			nlclasses	((chunksize - PAGE_SIZE) >> PAGE_SHIFT)
+
+#ifdef JEMALLOC_TCACHE
+void	arena_tcache_fill(arena_t *arena, tcache_bin_t *tbin, size_t binind
+#  ifdef JEMALLOC_PROF
+    , uint64_t prof_accumbytes
+#  endif
+    );
+#endif
+#ifdef JEMALLOC_PROF
+void	arena_prof_accum(arena_t *arena, uint64_t accumbytes);
+#endif
+void	*arena_malloc_small(arena_t *arena, size_t size, bool zero);
+void	*arena_malloc_medium(arena_t *arena, size_t size, bool zero);
+void	*arena_malloc(size_t size, bool zero);
+void	*arena_palloc(arena_t *arena, size_t alignment, size_t size,
+    size_t alloc_size);
+size_t	arena_salloc(const void *ptr);
+#ifdef JEMALLOC_PROF
+prof_thr_cnt_t	*arena_prof_cnt_get(const void *ptr);
+void	arena_prof_cnt_set(const void *ptr, prof_thr_cnt_t *cnt);
+#endif
+void	arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr,
+    arena_chunk_map_t *mapelm);
+void	arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr);
+#ifdef JEMALLOC_STATS
+void	arena_stats_merge(arena_t *arena, size_t *nactive, size_t *ndirty,
+    arena_stats_t *astats, malloc_bin_stats_t *bstats,
+    malloc_large_stats_t *lstats);
+#endif
+void	*arena_ralloc(void *ptr, size_t size, size_t oldsize);
+bool	arena_new(arena_t *arena, unsigned ind);
+bool	arena_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_))
+JEMALLOC_INLINE void
+arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr)
+{
+	size_t pageind;
+	arena_chunk_map_t *mapelm;
+
+	assert(arena != NULL);
+	assert(arena->magic == ARENA_MAGIC);
+	assert(chunk->arena == arena);
+	assert(ptr != NULL);
+	assert(CHUNK_ADDR2BASE(ptr) != ptr);
+
+	pageind = (((uintptr_t)ptr - (uintptr_t)chunk) >> PAGE_SHIFT);
+	mapelm = &chunk->map[pageind];
+	assert((mapelm->bits & CHUNK_MAP_ALLOCATED) != 0);
+	if ((mapelm->bits & CHUNK_MAP_LARGE) == 0) {
+		/* Small allocation. */
+#ifdef JEMALLOC_TCACHE
+		tcache_t *tcache;
+
+		if ((tcache = tcache_get()) != NULL)
+			tcache_dalloc(tcache, ptr);
+		else {
+#endif
+			malloc_mutex_lock(&arena->lock);
+			arena_dalloc_bin(arena, chunk, ptr, mapelm);
+			malloc_mutex_unlock(&arena->lock);
+#ifdef JEMALLOC_TCACHE
+		}
+#endif
+	} else
+		arena_dalloc_large(arena, chunk, ptr);
+}
+#endif
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/base.h

@@ -0,0 +1,24 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern malloc_mutex_t	base_mtx;
+
+void	*base_alloc(size_t size);
+extent_node_t *base_node_alloc(void);
+void	base_node_dealloc(extent_node_t *node);
+bool	base_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/chunk.h

@@ -0,0 +1,61 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+/*
+ * Size and alignment of memory chunks that are allocated by the OS's virtual
+ * memory system.
+ */
+#define	LG_CHUNK_DEFAULT	22
+
+/* Return the chunk address for allocation address a. */
+#define	CHUNK_ADDR2BASE(a)						\
+	((void *)((uintptr_t)(a) & ~chunksize_mask))
+
+/* Return the chunk offset of address a. */
+#define	CHUNK_ADDR2OFFSET(a)						\
+	((size_t)((uintptr_t)(a) & chunksize_mask))
+
+/* Return the smallest chunk multiple that is >= s. */
+#define	CHUNK_CEILING(s)						\
+	(((s) + chunksize_mask) & ~chunksize_mask)
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern size_t		opt_lg_chunk;
+#ifdef JEMALLOC_SWAP
+extern bool		opt_overcommit;
+#endif
+
+#if (defined(JEMALLOC_STATS) || defined(JEMALLOC_PROF))
+/* Protects stats_chunks; currently not used for any other purpose. */
+extern malloc_mutex_t	chunks_mtx;
+/* Chunk statistics. */
+extern chunk_stats_t	stats_chunks;
+#endif
+
+extern size_t		chunksize;
+extern size_t		chunksize_mask; /* (chunksize - 1). */
+extern size_t		chunk_npages;
+extern size_t		arena_chunk_header_npages;
+extern size_t		arena_maxclass; /* Max size class for arenas. */
+
+void	*chunk_alloc(size_t size, bool *zero);
+void	chunk_dealloc(void *chunk, size_t size);
+bool	chunk_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+
+#include "jemalloc/internal/chunk_swap.h"
+#include "jemalloc/internal/chunk_dss.h"
+#include "jemalloc/internal/chunk_mmap.h"

jemalloc/include/jemalloc/internal/chunk_dss.h

@@ -0,0 +1,29 @@
+#ifdef JEMALLOC_DSS
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+/*
+ * Protects sbrk() calls.  This avoids malloc races among threads, though it
+ * does not protect against races with threads that call sbrk() directly.
+ */
+extern malloc_mutex_t	dss_mtx;
+
+void	*chunk_alloc_dss(size_t size, bool *zero);
+bool	chunk_dealloc_dss(void *chunk, size_t size);
+bool	chunk_dss_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+#endif /* JEMALLOC_DSS */

jemalloc/include/jemalloc/internal/chunk_mmap.h

@@ -0,0 +1,20 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+void	*chunk_alloc_mmap(size_t size);
+void	chunk_dealloc_mmap(void *chunk, size_t size);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/chunk_swap.h

@@ -0,0 +1,33 @@
+#ifdef JEMALLOC_SWAP
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern malloc_mutex_t	swap_mtx;
+extern bool		swap_enabled;
+extern bool		swap_prezeroed;
+extern size_t		swap_nfds;
+extern int		*swap_fds;
+#ifdef JEMALLOC_STATS
+extern size_t		swap_avail;
+#endif
+
+void	*chunk_alloc_swap(size_t size, bool *zero);
+bool	chunk_dealloc_swap(void *chunk, size_t size);
+bool	chunk_swap_enable(const int *fds, unsigned nfds, bool prezeroed);
+bool	chunk_swap_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+#endif /* JEMALLOC_SWAP */

jemalloc/include/jemalloc/internal/ckh.h

@@ -0,0 +1,95 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef struct ckh_s ckh_t;
+typedef struct ckhc_s ckhc_t;
+
+/* Typedefs to allow easy function pointer passing. */
+typedef void ckh_hash_t (const void *, unsigned, size_t *, size_t *);
+typedef bool ckh_keycomp_t (const void *, const void *);
+
+/* Maintain counters used to get an idea of performance. */
+/* #define	CKH_COUNT */
+/* Print counter values in ckh_delete() (requires CKH_COUNT). */
+/* #define	CKH_VERBOSE */
+
+/*
+ * There are 2^LG_CKH_BUCKET_CELLS cells in each hash table bucket.  Try to fit
+ * one bucket per L1 cache line.
+ */
+#define LG_CKH_BUCKET_CELLS (LG_CACHELINE - LG_SIZEOF_PTR - 1)
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+/* Hash table cell. */
+struct ckhc_s {
+	const void	*key;
+	const void	*data;
+};
+
+struct ckh_s {
+#ifdef JEMALLOC_DEBUG
+#define	CKH_MAGIG	0x3af2489d
+	uint32_t	magic;
+#endif
+
+#ifdef CKH_COUNT
+	/* Counters used to get an idea of performance. */
+	uint64_t	ngrows;
+	uint64_t	nshrinks;
+	uint64_t	nshrinkfails;
+	uint64_t	ninserts;
+	uint64_t	nrelocs;
+#endif
+
+	/* Used for pseudo-random number generation. */
+#define	CKH_A		12345
+#define	CKH_C		12347
+	uint32_t	prn_state;
+
+	/* Total number of items. */
+	size_t		count;
+
+	/*
+	 * Minimum and current number of hash table buckets.  There are
+	 * 2^LG_CKH_BUCKET_CELLS cells per bucket.
+	 */
+	unsigned	lg_minbuckets;
+	unsigned	lg_curbuckets;
+
+	/* Hash and comparison functions. */
+	ckh_hash_t	*hash;
+	ckh_keycomp_t	*keycomp;
+
+	/* Hash table with 2^lg_curbuckets buckets. */
+	ckhc_t		*tab;
+};
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+bool	ckh_new(ckh_t *ckh, size_t minitems, ckh_hash_t *hash,
+    ckh_keycomp_t *keycomp);
+void	ckh_delete(ckh_t *ckh);
+size_t	ckh_count(ckh_t *ckh);
+bool	ckh_iter(ckh_t *ckh, size_t *tabind, void **key, void **data);
+bool	ckh_insert(ckh_t *ckh, const void *key, const void *data);
+bool	ckh_remove(ckh_t *ckh, const void *searchkey, void **key,
+    void **data);
+bool	ckh_search(ckh_t *ckh, const void *seachkey, void **key, void **data);
+void	ckh_string_hash(const void *key, unsigned minbits, size_t *hash1,
+    size_t *hash2);
+bool	ckh_string_keycomp(const void *k1, const void *k2);
+void	ckh_pointer_hash(const void *key, unsigned minbits, size_t *hash1,
+    size_t *hash2);
+bool	ckh_pointer_keycomp(const void *k1, const void *k2);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/ctl.h

@@ -0,0 +1,108 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef struct ctl_node_s ctl_node_t;
+typedef struct ctl_arena_stats_s ctl_arena_stats_t;
+typedef struct ctl_stats_s ctl_stats_t;
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+struct ctl_node_s {
+	bool			named;
+	union {
+		struct {
+			const char	*name;
+			/* If (nchildren == 0), this is a terminal node. */
+			unsigned	nchildren;
+			const	ctl_node_t *children;
+		} named;
+		struct {
+			const ctl_node_t *(*index)(const size_t *, size_t,
+			    size_t);
+		} indexed;
+	} u;
+	int	(*ctl)(const size_t *, size_t, void *, size_t *, void *,
+	    size_t);
+};
+
+struct ctl_arena_stats_s {
+	bool			initialized;
+	size_t			pactive;
+	size_t			pdirty;
+#ifdef JEMALLOC_STATS
+	arena_stats_t		astats;
+	malloc_bin_stats_t	*bstats;	/* nbins elements. */
+	malloc_large_stats_t	*lstats;	/* nlclasses elements. */
+#endif
+};
+
+struct ctl_stats_s {
+#ifdef JEMALLOC_STATS
+	size_t			allocated;
+	size_t			active;
+	size_t			mapped;
+	struct {
+		size_t		current;	/* stats_chunks.curchunks */
+		uint64_t	total;		/* stats_chunks.nchunks */
+		size_t		high;		/* stats_chunks.highchunks */
+	} chunks;
+	struct {
+		size_t		allocated;	/* huge_allocated */
+		uint64_t	nmalloc;	/* huge_nmalloc */
+		uint64_t	ndalloc;	/* huge_ndalloc */
+	} huge;
+#endif
+	ctl_arena_stats_t	*arenas;	/* (narenas + 1) elements. */
+#ifdef JEMALLOC_SWAP
+	size_t			swap_avail;
+#endif
+};
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+int	ctl_byname(const char *name, void *oldp, size_t *oldlenp, void *newp,
+    size_t newlen);
+int	ctl_nametomib(const char *name, size_t *mibp, size_t *miblenp);
+
+int	ctl_bymib(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
+    void *newp, size_t newlen);
+bool	ctl_boot(void);
+
+#define	xmallctl(name, oldp, oldlenp, newp, newlen) do {		\
+	if (mallctl(name, oldp, oldlenp, newp, newlen) != 0) {		\
+		malloc_write4("<jemalloc>: Invalid xmallctl(\"", name,	\
+		    "\", ...) call\n", "");				\
+		abort();						\
+	}								\
+} while (0)
+
+#define	xmallctlnametomib(name, mibp, miblenp) do {			\
+	if (mallctlnametomib(name, mibp, miblenp) != 0) {		\
+		malloc_write4(						\
+		    "<jemalloc>: Invalid xmallctlnametomib(\"", name,	\
+		    "\", ...) call\n", "");				\
+		abort();						\
+	}								\
+} while (0)
+
+#define	xmallctlbymib(mib, miblen, oldp, oldlenp, newp, newlen) do {	\
+	if (mallctlbymib(mib, miblen, oldp, oldlenp, newp, newlen)	\
+	    != 0) {							\
+		malloc_write4(						\
+		    "<jemalloc>: Invalid xmallctlbymib() call\n", "",	\
+		    "", "");						\
+		abort();						\
+	}								\
+} while (0)
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+

jemalloc/include/jemalloc/internal/extent.h

@@ -0,0 +1,49 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef struct extent_node_s extent_node_t;
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+/* Tree of extents. */
+struct extent_node_s {
+#if (defined(JEMALLOC_SWAP) || defined(JEMALLOC_DSS))
+	/* Linkage for the size/address-ordered tree. */
+	rb_node(extent_node_t)	link_szad;
+#endif
+
+	/* Linkage for the address-ordered tree. */
+	rb_node(extent_node_t)	link_ad;
+
+#ifdef JEMALLOC_PROF
+	/* Profile counters, used for huge objects. */
+	prof_thr_cnt_t		*prof_cnt;
+#endif
+
+	/* Pointer to the extent that this tree node is responsible for. */
+	void			*addr;
+
+	/* Total region size. */
+	size_t			size;
+};
+typedef rb_tree(extent_node_t) extent_tree_t;
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#if (defined(JEMALLOC_SWAP) || defined(JEMALLOC_DSS))
+rb_proto(, extent_tree_szad_, extent_tree_t, extent_node_t)
+#endif
+
+rb_proto(, extent_tree_ad_, extent_tree_t, extent_node_t)
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+

jemalloc/include/jemalloc/internal/hash.h

@@ -0,0 +1,70 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#ifndef JEMALLOC_ENABLE_INLINE
+uint64_t	hash(const void *key, size_t len, uint64_t seed);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(HASH_C_))
+/*
+ * The following hash function is based on MurmurHash64A(), placed into the
+ * public domain by Austin Appleby.  See http://murmurhash.googlepages.com/ for
+ * details.
+ */
+JEMALLOC_INLINE uint64_t
+hash(const void *key, size_t len, uint64_t seed)
+{
+	const uint64_t m = 0xc6a4a7935bd1e995;
+	const int r = 47;
+	uint64_t h = seed ^ (len * m);
+	const uint64_t *data = (const uint64_t *)key;
+	const uint64_t *end = data + (len/8);
+	const unsigned char *data2;
+
+	assert(((uintptr_t)key & 0x7) == 0);
+
+	while(data != end) {
+		uint64_t k = *data++;
+
+		k *= m;
+		k ^= k >> r;
+		k *= m;
+
+		h ^= k;
+		h *= m;
+	}
+
+	data2 = (const unsigned char *)data;
+	switch(len & 7) {
+		case 7: h ^= ((uint64_t)(data2[6])) << 48;
+		case 6: h ^= ((uint64_t)(data2[5])) << 40;
+		case 5: h ^= ((uint64_t)(data2[4])) << 32;
+		case 4: h ^= ((uint64_t)(data2[3])) << 24;
+		case 3: h ^= ((uint64_t)(data2[2])) << 16;
+		case 2: h ^= ((uint64_t)(data2[1])) << 8;
+		case 1: h ^= ((uint64_t)(data2[0]));
+			h *= m;
+	}
+
+	h ^= h >> r;
+	h *= m;
+	h ^= h >> r;
+
+	return h;
+}
+#endif
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/huge.h

@@ -0,0 +1,38 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#ifdef JEMALLOC_STATS
+/* Huge allocation statistics. */
+extern uint64_t		huge_nmalloc;
+extern uint64_t		huge_ndalloc;
+extern size_t		huge_allocated;
+#endif
+
+/* Protects chunk-related data structures. */
+extern malloc_mutex_t	huge_mtx;
+
+void	*huge_malloc(size_t size, bool zero);
+void	*huge_palloc(size_t alignment, size_t size);
+void	*huge_ralloc(void *ptr, size_t size, size_t oldsize);
+void	huge_dalloc(void *ptr);
+size_t	huge_salloc(const void *ptr);
+#ifdef JEMALLOC_PROF
+prof_thr_cnt_t	*huge_prof_cnt_get(const void *ptr);
+void	huge_prof_cnt_set(const void *ptr, prof_thr_cnt_t *cnt);
+#endif
+bool	huge_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/jemalloc_internal.h.in

@@ -0,0 +1,554 @@
+#include <sys/mman.h>
+#include <sys/param.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#include <sys/uio.h>
+
+#include <errno.h>
+#include <limits.h>
+#ifndef SIZE_T_MAX
+#  define SIZE_T_MAX	SIZE_MAX
+#endif
+#include <pthread.h>
+#include <sched.h>
+#include <stdarg.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <strings.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <pthread.h>
+
+#define	JEMALLOC_MANGLE
+#include "../jemalloc@install_suffix@.h"
+
+#ifdef JEMALLOC_LAZY_LOCK
+#include <dlfcn.h>
+#endif
+
+#include "jemalloc/internal/rb.h"
+#include "jemalloc/internal/qr.h"
+#include "jemalloc/internal/ql.h"
+
+extern void	(*JEMALLOC_P(malloc_message))(void *w4opaque, const char *p1,
+    const char *p2, const char *p3, const char *p4);
+
+/*
+ * Define a custom assert() in order to reduce the chances of deadlock during
+ * assertion failure.
+ */
+#ifdef JEMALLOC_DEBUG
+#  define assert(e) do {						\
+	if (!(e)) {							\
+		char line_buf[UMAX2S_BUFSIZE];				\
+		malloc_write4("<jemalloc>: ", __FILE__, ":",		\
+		    umax2s(__LINE__, 10, line_buf));			\
+		malloc_write4(": Failed assertion: ", "\"", #e,		\
+		    "\"\n");						\
+		abort();						\
+	}								\
+} while (0)
+#else
+#define assert(e)
+#endif
+
+/*
+ * jemalloc can conceptually be broken into components (arena, tcache, etc.),
+ * but there are circular dependencies that cannot be broken without
+ * substantial performance degradation.  In order to reduce the effect on
+ * visual code flow, read the header files in multiple passes, with one of the
+ * following cpp variables defined during each pass:
+ *
+ *   JEMALLOC_H_TYPES   : Preprocessor-defined constants and psuedo-opaque data
+ *                        types.
+ *   JEMALLOC_H_STRUCTS : Data structures.
+ *   JEMALLOC_H_EXTERNS : Extern data declarations and function prototypes.
+ *   JEMALLOC_H_INLINES : Inline functions.
+ */
+/******************************************************************************/
+#define JEMALLOC_H_TYPES
+
+#define	ZU(z)	((size_t)z)
+
+#ifndef __DECONST
+#  define	__DECONST(type, var)	((type)(uintptr_t)(const void *)(var))
+#endif
+
+#ifdef JEMALLOC_DEBUG
+   /* Disable inlining to make debugging easier. */
+#  define JEMALLOC_INLINE
+#  define inline
+#else
+#  define JEMALLOC_ENABLE_INLINE
+#  define JEMALLOC_INLINE static inline
+#endif
+
+/* Size of stack-allocated buffer passed to strerror_r(). */
+#define	STRERROR_BUF		64
+
+/* Minimum alignment of allocations is 2^LG_QUANTUM bytes. */
+#ifdef __i386__
+#  define LG_QUANTUM		4
+#endif
+#ifdef __ia64__
+#  define LG_QUANTUM		4
+#endif
+#ifdef __alpha__
+#  define LG_QUANTUM		4
+#endif
+#ifdef __sparc64__
+#  define LG_QUANTUM		4
+#endif
+#if (defined(__amd64__) || defined(__x86_64__))
+#  define LG_QUANTUM		4
+#endif
+#ifdef __arm__
+#  define LG_QUANTUM		3
+#endif
+#ifdef __mips__
+#  define LG_QUANTUM		3
+#endif
+#ifdef __powerpc__
+#  define LG_QUANTUM		4
+#endif
+#ifdef __s390x__
+#  define LG_QUANTUM		4
+#endif
+
+#define	QUANTUM			((size_t)(1U << LG_QUANTUM))
+#define	QUANTUM_MASK		(QUANTUM - 1)
+
+/* Return the smallest quantum multiple that is >= a. */
+#define	QUANTUM_CEILING(a)						\
+	(((a) + QUANTUM_MASK) & ~QUANTUM_MASK)
+
+#define	SIZEOF_PTR		(1U << LG_SIZEOF_PTR)
+
+/* We can't use TLS in non-PIC programs, since TLS relies on loader magic. */
+#if (!defined(PIC) && !defined(NO_TLS))
+#  define NO_TLS
+#endif
+
+/*
+ * Maximum size of L1 cache line.  This is used to avoid cache line aliasing.
+ * In addition, this controls the spacing of cacheline-spaced size classes.
+ */
+#define	LG_CACHELINE		6
+#define	CACHELINE		((size_t)(1U << LG_CACHELINE))
+#define	CACHELINE_MASK		(CACHELINE - 1)
+
+/* Return the smallest cacheline multiple that is >= s. */
+#define	CACHELINE_CEILING(s)						\
+	(((s) + CACHELINE_MASK) & ~CACHELINE_MASK)
+
+/*
+ * Page size.  STATIC_PAGE_SHIFT is determined by the configure script.  If
+ * DYNAMIC_PAGE_SHIFT is enabled, only use the STATIC_PAGE_* macros where
+ * compile-time values are required for the purposes of defining data
+ * structures.
+ */
+#define	STATIC_PAGE_SIZE ((size_t)(1U << STATIC_PAGE_SHIFT))
+#define	STATIC_PAGE_MASK ((size_t)(STATIC_PAGE_SIZE - 1))
+
+#ifdef DYNAMIC_PAGE_SHIFT
+#  define PAGE_SHIFT	lg_pagesize
+#  define PAGE_SIZE	pagesize
+#  define PAGE_MASK	pagesize_mask
+#else
+#  define PAGE_SHIFT	STATIC_PAGE_SHIFT
+#  define PAGE_SIZE	STATIC_PAGE_SIZE
+#  define PAGE_MASK	STATIC_PAGE_MASK
+#endif
+
+/* Return the smallest pagesize multiple that is >= s. */
+#define	PAGE_CEILING(s)							\
+	(((s) + PAGE_MASK) & ~PAGE_MASK)
+
+#include "jemalloc/internal/prn.h"
+#include "jemalloc/internal/mb.h"
+#include "jemalloc/internal/ckh.h"
+#include "jemalloc/internal/stats.h"
+#include "jemalloc/internal/ctl.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/extent.h"
+#include "jemalloc/internal/arena.h"
+#include "jemalloc/internal/base.h"
+#include "jemalloc/internal/chunk.h"
+#include "jemalloc/internal/huge.h"
+#include "jemalloc/internal/tcache.h"
+#include "jemalloc/internal/hash.h"
+#include "jemalloc/internal/prof.h"
+
+#undef JEMALLOC_H_TYPES
+/******************************************************************************/
+#define JEMALLOC_H_STRUCTS
+
+#include "jemalloc/internal/prn.h"
+#include "jemalloc/internal/mb.h"
+#include "jemalloc/internal/ckh.h"
+#include "jemalloc/internal/stats.h"
+#include "jemalloc/internal/ctl.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/extent.h"
+#include "jemalloc/internal/arena.h"
+#include "jemalloc/internal/base.h"
+#include "jemalloc/internal/chunk.h"
+#include "jemalloc/internal/huge.h"
+#include "jemalloc/internal/tcache.h"
+#include "jemalloc/internal/hash.h"
+#include "jemalloc/internal/prof.h"
+
+#undef JEMALLOC_H_STRUCTS
+/******************************************************************************/
+#define JEMALLOC_H_EXTERNS
+
+extern bool	opt_abort;
+#ifdef JEMALLOC_FILL
+extern bool	opt_junk;
+#endif
+#ifdef JEMALLOC_SYSV
+extern bool	opt_sysv;
+#endif
+#ifdef JEMALLOC_XMALLOC
+extern bool	opt_xmalloc;
+#endif
+#ifdef JEMALLOC_FILL
+extern bool	opt_zero;
+#endif
+
+#ifdef DYNAMIC_PAGE_SHIFT
+extern size_t		pagesize;
+extern size_t		pagesize_mask;
+extern size_t		lg_pagesize;
+#endif
+
+/* Number of CPUs. */
+extern unsigned		ncpus;
+
+extern malloc_mutex_t	arenas_lock; /* Protects arenas initialization. */
+#ifndef NO_TLS
+/*
+ * Map of pthread_self() --> arenas[???], used for selecting an arena to use
+ * for allocations.
+ */
+extern __thread arena_t	*arenas_map JEMALLOC_ATTR(tls_model("initial-exec"));
+#endif
+/*
+ * Arenas that are used to service external requests.  Not all elements of the
+ * arenas array are necessarily used; arenas are created lazily as needed.
+ */
+extern arena_t		**arenas;
+extern unsigned		narenas;
+
+arena_t	*arenas_extend(unsigned ind);
+#ifndef NO_TLS
+arena_t	*choose_arena_hard(void);
+#endif
+
+#include "jemalloc/internal/prn.h"
+#include "jemalloc/internal/mb.h"
+#include "jemalloc/internal/ckh.h"
+#include "jemalloc/internal/stats.h"
+#include "jemalloc/internal/ctl.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/extent.h"
+#include "jemalloc/internal/arena.h"
+#include "jemalloc/internal/base.h"
+#include "jemalloc/internal/chunk.h"
+#include "jemalloc/internal/huge.h"
+#include "jemalloc/internal/tcache.h"
+#include "jemalloc/internal/hash.h"
+#include "jemalloc/internal/prof.h"
+
+#undef JEMALLOC_H_EXTERNS
+/******************************************************************************/
+#define JEMALLOC_H_INLINES
+
+#include "jemalloc/internal/prn.h"
+#include "jemalloc/internal/mb.h"
+#include "jemalloc/internal/ckh.h"
+#include "jemalloc/internal/stats.h"
+#include "jemalloc/internal/ctl.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/extent.h"
+#include "jemalloc/internal/base.h"
+#include "jemalloc/internal/chunk.h"
+#include "jemalloc/internal/huge.h"
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	malloc_write4(const char *p1, const char *p2, const char *p3,
+    const char *p4);
+arena_t	*choose_arena(void);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_C_))
+/*
+ * Wrapper around malloc_message() that avoids the need for
+ * JEMALLOC_P(malloc_message)(...) throughout the code.
+ */
+JEMALLOC_INLINE void
+malloc_write4(const char *p1, const char *p2, const char *p3, const char *p4)
+{
+
+	JEMALLOC_P(malloc_message)(NULL, p1, p2, p3, p4);
+}
+
+/*
+ * Choose an arena based on a per-thread value (fast-path code, calls slow-path
+ * code if necessary).
+ */
+JEMALLOC_INLINE arena_t *
+choose_arena(void)
+{
+	arena_t *ret;
+
+	/*
+	 * We can only use TLS if this is a PIC library, since for the static
+	 * library version, libc's malloc is used by TLS allocation, which
+	 * introduces a bootstrapping issue.
+	 */
+#ifndef NO_TLS
+	ret = arenas_map;
+	if (ret == NULL) {
+		ret = choose_arena_hard();
+		assert(ret != NULL);
+	}
+#else
+	if (isthreaded && narenas > 1) {
+		unsigned long ind;
+
+		/*
+		 * Hash pthread_self() to one of the arenas.  There is a prime
+		 * number of arenas, so this has a reasonable chance of
+		 * working.  Even so, the hashing can be easily thwarted by
+		 * inconvenient pthread_self() values.  Without specific
+		 * knowledge of how pthread_self() calculates values, we can't
+		 * easily do much better than this.
+		 */
+		ind = (unsigned long) pthread_self() % narenas;
+
+		/*
+		 * Optimistially assume that arenas[ind] has been initialized.
+		 * At worst, we find out that some other thread has already
+		 * done so, after acquiring the lock in preparation.  Note that
+		 * this lazy locking also has the effect of lazily forcing
+		 * cache coherency; without the lock acquisition, there's no
+		 * guarantee that modification of arenas[ind] by another thread
+		 * would be seen on this CPU for an arbitrary amount of time.
+		 *
+		 * In general, this approach to modifying a synchronized value
+		 * isn't a good idea, but in this case we only ever modify the
+		 * value once, so things work out well.
+		 */
+		ret = arenas[ind];
+		if (ret == NULL) {
+			/*
+			 * Avoid races with another thread that may have already
+			 * initialized arenas[ind].
+			 */
+			malloc_mutex_lock(&arenas_lock);
+			if (arenas[ind] == NULL)
+				ret = arenas_extend((unsigned)ind);
+			else
+				ret = arenas[ind];
+			malloc_mutex_unlock(&arenas_lock);
+		}
+	} else
+		ret = arenas[0];
+#endif
+
+	assert(ret != NULL);
+	return (ret);
+}
+#endif
+
+#include "jemalloc/internal/tcache.h"
+#include "jemalloc/internal/arena.h"
+#include "jemalloc/internal/hash.h"
+#include "jemalloc/internal/prof.h"
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	*imalloc(size_t size);
+void	*icalloc(size_t size);
+void	*ipalloc(size_t alignment, size_t size);
+size_t	isalloc(const void *ptr);
+void	*iralloc(void *ptr, size_t size);
+void	idalloc(void *ptr);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_C_))
+JEMALLOC_INLINE void *
+imalloc(size_t size)
+{
+
+	assert(size != 0);
+
+	if (size <= arena_maxclass)
+		return (arena_malloc(size, false));
+	else
+		return (huge_malloc(size, false));
+}
+
+JEMALLOC_INLINE void *
+icalloc(size_t size)
+{
+
+	if (size <= arena_maxclass)
+		return (arena_malloc(size, true));
+	else
+		return (huge_malloc(size, true));
+}
+
+JEMALLOC_INLINE void *
+ipalloc(size_t alignment, size_t size)
+{
+	void *ret;
+	size_t ceil_size;
+
+	/*
+	 * Round size up to the nearest multiple of alignment.
+	 *
+	 * This done, we can take advantage of the fact that for each small
+	 * size class, every object is aligned at the smallest power of two
+	 * that is non-zero in the base two representation of the size.  For
+	 * example:
+	 *
+	 *   Size |   Base 2 | Minimum alignment
+	 *   -----+----------+------------------
+	 *     96 |  1100000 |  32
+	 *    144 | 10100000 |  32
+	 *    192 | 11000000 |  64
+	 *
+	 * Depending on runtime settings, it is possible that arena_malloc()
+	 * will further round up to a power of two, but that never causes
+	 * correctness issues.
+	 */
+	ceil_size = (size + (alignment - 1)) & (-alignment);
+	/*
+	 * (ceil_size < size) protects against the combination of maximal
+	 * alignment and size greater than maximal alignment.
+	 */
+	if (ceil_size < size) {
+		/* size_t overflow. */
+		return (NULL);
+	}
+
+	if (ceil_size <= PAGE_SIZE || (alignment <= PAGE_SIZE
+	    && ceil_size <= arena_maxclass))
+		ret = arena_malloc(ceil_size, false);
+	else {
+		size_t run_size;
+
+		/*
+		 * We can't achieve subpage alignment, so round up alignment
+		 * permanently; it makes later calculations simpler.
+		 */
+		alignment = PAGE_CEILING(alignment);
+		ceil_size = PAGE_CEILING(size);
+		/*
+		 * (ceil_size < size) protects against very large sizes within
+		 * PAGE_SIZE of SIZE_T_MAX.
+		 *
+		 * (ceil_size + alignment < ceil_size) protects against the
+		 * combination of maximal alignment and ceil_size large enough
+		 * to cause overflow.  This is similar to the first overflow
+		 * check above, but it needs to be repeated due to the new
+		 * ceil_size value, which may now be *equal* to maximal
+		 * alignment, whereas before we only detected overflow if the
+		 * original size was *greater* than maximal alignment.
+		 */
+		if (ceil_size < size || ceil_size + alignment < ceil_size) {
+			/* size_t overflow. */
+			return (NULL);
+		}
+
+		/*
+		 * Calculate the size of the over-size run that arena_palloc()
+		 * would need to allocate in order to guarantee the alignment.
+		 */
+		if (ceil_size >= alignment)
+			run_size = ceil_size + alignment - PAGE_SIZE;
+		else {
+			/*
+			 * It is possible that (alignment << 1) will cause
+			 * overflow, but it doesn't matter because we also
+			 * subtract PAGE_SIZE, which in the case of overflow
+			 * leaves us with a very large run_size.  That causes
+			 * the first conditional below to fail, which means
+			 * that the bogus run_size value never gets used for
+			 * anything important.
+			 */
+			run_size = (alignment << 1) - PAGE_SIZE;
+		}
+
+		if (run_size <= arena_maxclass) {
+			ret = arena_palloc(choose_arena(), alignment, ceil_size,
+			    run_size);
+		} else if (alignment <= chunksize)
+			ret = huge_malloc(ceil_size, false);
+		else
+			ret = huge_palloc(alignment, ceil_size);
+	}
+
+	assert(((uintptr_t)ret & (alignment - 1)) == 0);
+	return (ret);
+}
+
+JEMALLOC_INLINE size_t
+isalloc(const void *ptr)
+{
+	size_t ret;
+	arena_chunk_t *chunk;
+
+	assert(ptr != NULL);
+
+	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
+	if (chunk != ptr) {
+		/* Region. */
+		assert(chunk->arena->magic == ARENA_MAGIC);
+
+		ret = arena_salloc(ptr);
+	} else
+		ret = huge_salloc(ptr);
+
+	return (ret);
+}
+
+JEMALLOC_INLINE void *
+iralloc(void *ptr, size_t size)
+{
+	size_t oldsize;
+
+	assert(ptr != NULL);
+	assert(size != 0);
+
+	oldsize = isalloc(ptr);
+
+	if (size <= arena_maxclass)
+		return (arena_ralloc(ptr, size, oldsize));
+	else
+		return (huge_ralloc(ptr, size, oldsize));
+}
+
+JEMALLOC_INLINE void
+idalloc(void *ptr)
+{
+	arena_chunk_t *chunk;
+
+	assert(ptr != NULL);
+
+	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
+	if (chunk != ptr)
+		arena_dalloc(chunk->arena, chunk, ptr);
+	else
+		huge_dalloc(ptr);
+}
+#endif
+
+#undef JEMALLOC_H_INLINES
+/******************************************************************************/

jemalloc/include/jemalloc/internal/mb.h

@@ -0,0 +1,108 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	mb_write(void);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(MB_C_))
+#ifdef __i386__
+/*
+ * According to the Intel Architecture Software Developer's Manual, current
+ * processors execute instructions in order from the perspective of other
+ * processors in a multiprocessor system, but 1) Intel reserves the right to
+ * change that, and 2) the compiler's optimizer could re-order instructions if
+ * there weren't some form of barrier.  Therefore, even if running on an
+ * architecture that does not need memory barriers (everything through at least
+ * i686), an "optimizer barrier" is necessary.
+ */
+JEMALLOC_INLINE void
+mb_write(void)
+{
+
+#  if 0
+	/* This is a true memory barrier. */
+	asm volatile ("pusha;"
+	    "xor  %%eax,%%eax;"
+	    "cpuid;"
+	    "popa;"
+	    : /* Outputs. */
+	    : /* Inputs. */
+	    : "memory" /* Clobbers. */
+	    );
+#else
+	/*
+	 * This is hopefully enough to keep the compiler from reordering
+	 * instructions around this one.
+	 */
+	asm volatile ("nop;"
+	    : /* Outputs. */
+	    : /* Inputs. */
+	    : "memory" /* Clobbers. */
+	    );
+#endif
+}
+#elif (defined(__amd64_) || defined(__x86_64__))
+JEMALLOC_INLINE void
+mb_write(void)
+{
+
+	asm volatile ("sfence"
+	    : /* Outputs. */
+	    : /* Inputs. */
+	    : "memory" /* Clobbers. */
+	    );
+}
+#elif defined(__powerpc__)
+JEMALLOC_INLINE void
+mb_write(void)
+{
+
+	asm volatile ("eieio"
+	    : /* Outputs. */
+	    : /* Inputs. */
+	    : "memory" /* Clobbers. */
+	    );
+}
+#elif defined(__sparc64__)
+JEMALLOC_INLINE void
+mb_write(void)
+{
+
+	asm volatile ("membar #StoreStore"
+	    : /* Outputs. */
+	    : /* Inputs. */
+	    : "memory" /* Clobbers. */
+	    );
+}
+#else
+/*
+ * This is much slower than a simple memory barrier, but the semantics of mutex
+ * unlock make this work.
+ */
+JEMALLOC_INLINE void
+mb_write(void)
+{
+	malloc_mutex_t mtx;
+
+	malloc_mutex_init(&mtx);
+	malloc_mutex_lock(&mtx);
+	malloc_mutex_unlock(&mtx);
+}
+#endif
+#endif
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/mutex.h

@@ -0,0 +1,50 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef pthread_mutex_t malloc_mutex_t;
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#ifdef JEMALLOC_LAZY_LOCK
+extern bool isthreaded;
+#else
+#  define isthreaded true
+#endif
+
+bool	malloc_mutex_init(malloc_mutex_t *mutex);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	malloc_mutex_lock(malloc_mutex_t *mutex);
+void	malloc_mutex_unlock(malloc_mutex_t *mutex);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_MUTEX_C_))
+JEMALLOC_INLINE void
+malloc_mutex_lock(malloc_mutex_t *mutex)
+{
+
+	if (isthreaded)
+		pthread_mutex_lock(mutex);
+}
+
+JEMALLOC_INLINE void
+malloc_mutex_unlock(malloc_mutex_t *mutex)
+{
+
+	if (isthreaded)
+		pthread_mutex_unlock(mutex);
+}
+#endif
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/prn.h

@@ -0,0 +1,50 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+/*
+ * Simple linear congruential pseudo-random number generator:
+ *
+ *   prn(y) = (a*x + c) % m
+ *
+ * where the following constants ensure maximal period:
+ *
+ *   a == Odd number (relatively prime to 2^n), and (a-1) is a multiple of 4.
+ *   c == Odd number (relatively prime to 2^n).
+ *   m == 2^32
+ *
+ * See Knuth's TAOCP 3rd Ed., Vol. 2, pg. 17 for details on these constraints.
+ *
+ * This choice of m has the disadvantage that the quality of the bits is
+ * proportional to bit position.  For example. the lowest bit has a cycle of 2,
+ * the next has a cycle of 4, etc.  For this reason, we prefer to use the upper
+ * bits.
+ *
+ * Macro parameters:
+ *   uint32_t r          : Result.
+ *   unsigned lg_range   : (0..32], number of least significant bits to return.
+ *   uint32_t state      : Seed value.
+ *   const uint32_t a, c : See above discussion.
+ */
+#define prn(r, lg_range, state, a, c) do {				\
+	assert(lg_range > 0);						\
+	assert(lg_range <= 32);						\
+									\
+	r = (state * (a)) + (c);					\
+	state = r;							\
+	r >>= (32 - lg_range);						\
+} while (false)
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/prof.h

@@ -0,0 +1,154 @@
+#ifdef JEMALLOC_PROF
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef struct prof_bt_s prof_bt_t;
+typedef struct prof_cnt_s prof_cnt_t;
+typedef struct prof_thr_cnt_s prof_thr_cnt_t;
+typedef struct prof_ctx_s prof_ctx_t;
+typedef struct prof_s prof_t;
+
+#define	LG_PROF_INTERVAL_DEFAULT	30
+
+/*
+ * Hard limit on stack backtrace depth.  Note that the version of
+ * prof_backtrace() that is based on __builtin_return_address() necessarily has
+ * a hard-coded number of backtrace frame handlers, so increasing
+ * LG_PROF_BT_MAX requires changing prof_backtrace().
+ */
+#define	LG_PROF_BT_MAX		7
+#define	PROF_BT_MAX		(1U << LG_PROF_BT_MAX)
+
+/* Initial hash table size. */
+#define	PROF_CKH_MINITEMS	64
+
+/* Size of memory buffer to use when writing dump files. */
+#define	PROF_DUMP_BUF_SIZE	65536
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+struct prof_bt_s {
+	/* Backtrace, stored as len program counters. */
+	void			**vec;
+	unsigned		len;
+};
+
+#ifdef JEMALLOC_PROF_LIBGCC
+/* Data structure passed to libgcc _Unwind_Backtrace() callback functions. */
+typedef struct {
+	prof_bt_t *bt;
+	unsigned nignore;
+	unsigned max;
+} prof_unwind_data_t;
+#endif
+
+struct prof_cnt_s {
+	/*
+	 * Profiling counters.  An allocation/deallocation pair can operate on
+	 * different prof_thr_cnt_t objects that are linked into the same
+	 * prof_ctx_t sets_ql, so it is possible for the cur* counters to go
+	 * negative.  In principle it is possible for the *bytes counters to
+	 * overflow/underflow, but a general solution would require some form
+	 * of 128-bit counter solution; this implementation doesn't bother to
+	 * solve that problem.
+	 */
+	int64_t		curobjs;
+	int64_t		curbytes;
+	uint64_t	accumobjs;
+	uint64_t	accumbytes;
+};
+
+struct prof_thr_cnt_s {
+	/* Linkage into prof_ctx_t's sets_ql. */
+	ql_elm(prof_thr_cnt_t)	link;
+
+	/*
+	 * Associated context.  If a thread frees an object that it did not
+	 * allocate, it is possible that the context is not cached in the
+	 * thread's hash table, in which case it must be able to look up the
+	 * context, insert a new prof_thr_cnt_t into the thread's hash table,
+	 * and link it into the prof_ctx_t's sets_ql.
+	 */
+	prof_ctx_t		*ctx;
+
+	/*
+	 * Threads use memory barriers to update the counters.  Since there is
+	 * only ever one writer, the only challenge is for the reader to get a
+	 * consistent read of the counters.
+	 *
+	 * The writer uses this series of operations:
+	 *
+	 * 1) Increment epoch to an odd number.
+	 * 2) Update counters.
+	 * 3) Increment epoch to an even number.
+	 *
+	 * The reader must assure 1) that the epoch is even while it reads the
+	 * counters, and 2) that the epoch doesn't change between the time it
+	 * starts and finishes reading the counters.
+	 */
+	unsigned		epoch;
+
+	/* Profiling counters. */
+	prof_cnt_t		cnts;
+};
+
+struct prof_ctx_s {
+	/* Protects cnt_merged and sets_ql. */
+	malloc_mutex_t		lock;
+
+	/* Temporary storage for aggregation during dump. */
+	prof_cnt_t		cnt_dump;
+
+	/* When threads exit, they merge their stats into cnt_merged. */
+	prof_cnt_t		cnt_merged;
+
+	/*
+	 * List of profile counters, one for each thread that has allocated in
+	 * this context.
+	 */
+	ql_head(prof_thr_cnt_t)	cnts_ql;
+};
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern bool	opt_prof;
+extern size_t	opt_lg_prof_bt_max; /* Maximum backtrace depth. */
+extern size_t	opt_lg_prof_interval;
+extern bool	opt_prof_udump; /* High-water memory dumping. */
+extern bool	opt_prof_leak; /* Dump leak summary at exit. */
+
+/*
+ * Profile dump interval, measured in bytes allocated.  Each arena triggers a
+ * profile dump when it reaches this threshold.  The effect is that the
+ * interval between profile dumps averages prof_interval, though the actual
+ * interval between dumps will tend to be sporadic, and the interval will be a
+ * maximum of approximately (prof_interval * narenas).
+ */
+extern uint64_t	prof_interval;
+
+bool	prof_init(prof_t *prof, bool master);
+void	prof_destroy(prof_t *prof);
+
+prof_thr_cnt_t	*prof_alloc_prep(void);
+prof_thr_cnt_t	*prof_cnt_get(const void *ptr);
+void	prof_malloc(const void *ptr, prof_thr_cnt_t *cnt);
+void	prof_realloc(const void *ptr, prof_thr_cnt_t *cnt, const void *old_ptr,
+    size_t old_size, prof_thr_cnt_t *old_cnt);
+void	prof_free(const void *ptr);
+void	prof_idump(void);
+void	prof_mdump(void);
+void	prof_udump(void);
+void	prof_boot0(void);
+bool	prof_boot1(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+#endif /* JEMALLOC_PROF */

jemalloc/include/jemalloc/internal/ql.h

@@ -0,0 +1,83 @@
+/*
+ * List definitions.
+ */
+#define ql_head(a_type)							\
+struct {								\
+	a_type *qlh_first;						\
+}
+
+#define ql_head_initializer(a_head) {NULL}
+
+#define ql_elm(a_type)	qr(a_type)
+
+/* List functions. */
+#define ql_new(a_head) do {						\
+	(a_head)->qlh_first = NULL;					\
+} while (0)
+
+#define ql_elm_new(a_elm, a_field) qr_new((a_elm), a_field)
+
+#define ql_first(a_head) ((a_head)->qlh_first)
+
+#define ql_last(a_head, a_field)					\
+	((ql_first(a_head) != NULL)					\
+	    ? qr_prev(ql_first(a_head), a_field) : NULL)
+
+#define ql_next(a_head, a_elm, a_field)					\
+	((ql_last(a_head, a_field) != (a_elm))				\
+	    ? qr_next((a_elm), a_field)	: NULL)
+
+#define ql_prev(a_head, a_elm, a_field)					\
+	((ql_first(a_head) != (a_elm)) ? qr_prev((a_elm), a_field)	\
+				       : NULL)
+
+#define ql_before_insert(a_head, a_qlelm, a_elm, a_field) do {		\
+	qr_before_insert((a_qlelm), (a_elm), a_field);			\
+	if (ql_first(a_head) == (a_qlelm)) {				\
+		ql_first(a_head) = (a_elm);				\
+	}								\
+} while (0)
+
+#define ql_after_insert(a_qlelm, a_elm, a_field)			\
+	qr_after_insert((a_qlelm), (a_elm), a_field)
+
+#define ql_head_insert(a_head, a_elm, a_field) do {			\
+	if (ql_first(a_head) != NULL) {					\
+		qr_before_insert(ql_first(a_head), (a_elm), a_field);	\
+	}								\
+	ql_first(a_head) = (a_elm);					\
+} while (0)
+
+#define ql_tail_insert(a_head, a_elm, a_field) do {			\
+	if (ql_first(a_head) != NULL) {					\
+		qr_before_insert(ql_first(a_head), (a_elm), a_field);	\
+	}								\
+	ql_first(a_head) = qr_next((a_elm), a_field);			\
+} while (0)
+
+#define ql_remove(a_head, a_elm, a_field) do {				\
+	if (ql_first(a_head) == (a_elm)) {				\
+		ql_first(a_head) = qr_next(ql_first(a_head), a_field);	\
+	}								\
+	if (ql_first(a_head) != (a_elm)) {				\
+		qr_remove((a_elm), a_field);				\
+	} else {							\
+		ql_first(a_head) = NULL;				\
+	}								\
+} while (0)
+
+#define ql_head_remove(a_head, a_type, a_field) do {			\
+	a_type *t = ql_first(a_head);					\
+	ql_remove((a_head), t, a_field);				\
+} while (0)
+
+#define ql_tail_remove(a_head, a_type, a_field) do {			\
+	a_type *t = ql_last(a_head, a_field);				\
+	ql_remove((a_head), t, a_field);				\
+} while (0)
+
+#define ql_foreach(a_var, a_head, a_field)				\
+	qr_foreach((a_var), ql_first(a_head), a_field)
+
+#define ql_reverse_foreach(a_var, a_head, a_field)			\
+	qr_reverse_foreach((a_var), ql_first(a_head), a_field)

jemalloc/include/jemalloc/internal/qr.h

@@ -0,0 +1,67 @@
+/* Ring definitions. */
+#define qr(a_type)							\
+struct {								\
+	a_type	*qre_next;						\
+	a_type	*qre_prev;						\
+}
+
+/* Ring functions. */
+#define qr_new(a_qr, a_field) do {					\
+	(a_qr)->a_field.qre_next = (a_qr);				\
+	(a_qr)->a_field.qre_prev = (a_qr);				\
+} while (0)
+
+#define qr_next(a_qr, a_field) ((a_qr)->a_field.qre_next)
+
+#define qr_prev(a_qr, a_field) ((a_qr)->a_field.qre_prev)
+
+#define qr_before_insert(a_qrelm, a_qr, a_field) do {			\
+	(a_qr)->a_field.qre_prev = (a_qrelm)->a_field.qre_prev;		\
+	(a_qr)->a_field.qre_next = (a_qrelm);				\
+	(a_qr)->a_field.qre_prev->a_field.qre_next = (a_qr);		\
+	(a_qrelm)->a_field.qre_prev = (a_qr);				\
+} while (0)
+
+#define qr_after_insert(a_qrelm, a_qr, a_field)				\
+    do									\
+    {									\
+	(a_qr)->a_field.qre_next = (a_qrelm)->a_field.qre_next;		\
+	(a_qr)->a_field.qre_prev = (a_qrelm);				\
+	(a_qr)->a_field.qre_next->a_field.qre_prev = (a_qr);		\
+	(a_qrelm)->a_field.qre_next = (a_qr);				\
+    } while (0)
+
+#define qr_meld(a_qr_a, a_qr_b, a_field) do {				\
+	void *t;							\
+	(a_qr_a)->a_field.qre_prev->a_field.qre_next = (a_qr_b);	\
+	(a_qr_b)->a_field.qre_prev->a_field.qre_next = (a_qr_a);	\
+	t = (a_qr_a)->a_field.qre_prev;					\
+	(a_qr_a)->a_field.qre_prev = (a_qr_b)->a_field.qre_prev;	\
+	(a_qr_b)->a_field.qre_prev = t;					\
+} while (0)
+
+/* qr_meld() and qr_split() are functionally equivalent, so there's no need to
+ * have two copies of the code. */
+#define qr_split(a_qr_a, a_qr_b, a_field)				\
+	qr_meld((a_qr_a), (a_qr_b), a_field)
+
+#define qr_remove(a_qr, a_field) do {					\
+	(a_qr)->a_field.qre_prev->a_field.qre_next			\
+	    = (a_qr)->a_field.qre_next;					\
+	(a_qr)->a_field.qre_next->a_field.qre_prev			\
+	    = (a_qr)->a_field.qre_prev;					\
+	(a_qr)->a_field.qre_next = (a_qr);				\
+	(a_qr)->a_field.qre_prev = (a_qr);				\
+} while (0)
+
+#define qr_foreach(var, a_qr, a_field)					\
+	for ((var) = (a_qr);						\
+	    (var) != NULL;						\
+	    (var) = (((var)->a_field.qre_next != (a_qr))		\
+	    ? (var)->a_field.qre_next : NULL))
+
+#define qr_reverse_foreach(var, a_qr, a_field)				\
+	for ((var) = ((a_qr) != NULL) ? qr_prev(a_qr, a_field) : NULL;	\
+	    (var) != NULL;						\
+	    (var) = (((var) != (a_qr))					\
+	    ? (var)->a_field.qre_prev : NULL))

jemalloc/include/jemalloc/internal/rb.h

@@ -0,0 +1,945 @@
+/*-
+ * cpp macro implementation of left-leaning red-black trees.
+ *
+ * Usage:
+ *
+ *   (Optional, see assert(3).)
+ *   #define NDEBUG
+ *
+ *   (Required.)
+ *   #include <assert.h>
+ *   #include <rb.h>
+ *   ...
+ *
+ * All operations are done non-recursively.  Parent pointers are not used, and
+ * color bits are stored in the least significant bit of right-child pointers,
+ * thus making node linkage as compact as is possible for red-black trees.
+ *
+ * Some macros use a comparison function pointer, which is expected to have the
+ * following prototype:
+ *
+ *   int (a_cmp *)(a_type *a_node, a_type *a_other);
+ *                         ^^^^^^
+ *                      or a_key
+ *
+ * Interpretation of comparision function return values:
+ *
+ *   -1 : a_node <  a_other
+ *    0 : a_node == a_other
+ *    1 : a_node >  a_other
+ *
+ * In all cases, the a_node or a_key macro argument is the first argument to the
+ * comparison function, which makes it possible to write comparison functions
+ * that treat the first argument specially.
+ *
+ ******************************************************************************/
+
+#ifndef RB_H_
+#define	RB_H_
+
+#if 0
+__FBSDID("$FreeBSD: src/lib/libc/stdlib/rb.h,v 1.4 2008/05/14 18:33:13 jasone Exp $");
+#endif
+
+/* Node structure. */
+#define	rb_node(a_type)							\
+struct {								\
+    a_type *rbn_left;							\
+    a_type *rbn_right_red;						\
+}
+
+/* Root structure. */
+#define	rb_tree(a_type)							\
+struct {								\
+    a_type *rbt_root;							\
+    a_type rbt_nil;							\
+}
+
+/* Left accessors. */
+#define	rbp_left_get(a_type, a_field, a_node)				\
+    ((a_node)->a_field.rbn_left)
+#define	rbp_left_set(a_type, a_field, a_node, a_left) do {		\
+    (a_node)->a_field.rbn_left = a_left;				\
+} while (0)
+
+/* Right accessors. */
+#define	rbp_right_get(a_type, a_field, a_node)				\
+    ((a_type *) (((intptr_t) (a_node)->a_field.rbn_right_red)		\
+      & ((ssize_t)-2)))
+#define	rbp_right_set(a_type, a_field, a_node, a_right) do {		\
+    (a_node)->a_field.rbn_right_red = (a_type *) (((uintptr_t) a_right)	\
+      | (((uintptr_t) (a_node)->a_field.rbn_right_red) & ((size_t)1)));	\
+} while (0)
+
+/* Color accessors. */
+#define	rbp_red_get(a_type, a_field, a_node)				\
+    ((bool) (((uintptr_t) (a_node)->a_field.rbn_right_red)		\
+      & ((size_t)1)))
+#define	rbp_color_set(a_type, a_field, a_node, a_red) do {		\
+    (a_node)->a_field.rbn_right_red = (a_type *) ((((intptr_t)		\
+      (a_node)->a_field.rbn_right_red) & ((ssize_t)-2))			\
+      | ((ssize_t)a_red));						\
+} while (0)
+#define	rbp_red_set(a_type, a_field, a_node) do {			\
+    (a_node)->a_field.rbn_right_red = (a_type *) (((uintptr_t)		\
+      (a_node)->a_field.rbn_right_red) | ((size_t)1));			\
+} while (0)
+#define	rbp_black_set(a_type, a_field, a_node) do {			\
+    (a_node)->a_field.rbn_right_red = (a_type *) (((intptr_t)		\
+      (a_node)->a_field.rbn_right_red) & ((ssize_t)-2));		\
+} while (0)
+
+/* Node initializer. */
+#define	rbp_node_new(a_type, a_field, a_tree, a_node) do {		\
+    rbp_left_set(a_type, a_field, (a_node), &(a_tree)->rbt_nil);	\
+    rbp_right_set(a_type, a_field, (a_node), &(a_tree)->rbt_nil);	\
+    rbp_red_set(a_type, a_field, (a_node));				\
+} while (0)
+
+/* Tree initializer. */
+#define	rb_new(a_type, a_field, a_tree) do {				\
+    (a_tree)->rbt_root = &(a_tree)->rbt_nil;				\
+    rbp_node_new(a_type, a_field, a_tree, &(a_tree)->rbt_nil);		\
+    rbp_black_set(a_type, a_field, &(a_tree)->rbt_nil);			\
+} while (0)
+
+/* Tree operations. */
+#define	rbp_black_height(a_type, a_field, a_tree, r_height) do {	\
+    a_type *rbp_bh_t;							\
+    for (rbp_bh_t = (a_tree)->rbt_root, (r_height) = 0;			\
+      rbp_bh_t != &(a_tree)->rbt_nil;					\
+      rbp_bh_t = rbp_left_get(a_type, a_field, rbp_bh_t)) {		\
+	if (rbp_red_get(a_type, a_field, rbp_bh_t) == false) {		\
+	    (r_height)++;						\
+	}								\
+    }									\
+} while (0)
+
+#define	rbp_first(a_type, a_field, a_tree, a_root, r_node) do {		\
+    for ((r_node) = (a_root);						\
+      rbp_left_get(a_type, a_field, (r_node)) != &(a_tree)->rbt_nil;	\
+      (r_node) = rbp_left_get(a_type, a_field, (r_node))) {		\
+    }									\
+} while (0)
+
+#define	rbp_last(a_type, a_field, a_tree, a_root, r_node) do {		\
+    for ((r_node) = (a_root);						\
+      rbp_right_get(a_type, a_field, (r_node)) != &(a_tree)->rbt_nil;	\
+      (r_node) = rbp_right_get(a_type, a_field, (r_node))) {		\
+    }									\
+} while (0)
+
+#define	rbp_next(a_type, a_field, a_cmp, a_tree, a_node, r_node) do {	\
+    if (rbp_right_get(a_type, a_field, (a_node))			\
+      != &(a_tree)->rbt_nil) {						\
+	rbp_first(a_type, a_field, a_tree, rbp_right_get(a_type,	\
+	  a_field, (a_node)), (r_node));				\
+    } else {								\
+	a_type *rbp_n_t = (a_tree)->rbt_root;				\
+	assert(rbp_n_t != &(a_tree)->rbt_nil);				\
+	(r_node) = &(a_tree)->rbt_nil;					\
+	while (true) {							\
+	    int rbp_n_cmp = (a_cmp)((a_node), rbp_n_t);			\
+	    if (rbp_n_cmp < 0) {					\
+		(r_node) = rbp_n_t;					\
+		rbp_n_t = rbp_left_get(a_type, a_field, rbp_n_t);	\
+	    } else if (rbp_n_cmp > 0) {					\
+		rbp_n_t = rbp_right_get(a_type, a_field, rbp_n_t);	\
+	    } else {							\
+		break;							\
+	    }								\
+	    assert(rbp_n_t != &(a_tree)->rbt_nil);			\
+	}								\
+    }									\
+} while (0)
+
+#define	rbp_prev(a_type, a_field, a_cmp, a_tree, a_node, r_node) do {	\
+    if (rbp_left_get(a_type, a_field, (a_node)) != &(a_tree)->rbt_nil) {\
+	rbp_last(a_type, a_field, a_tree, rbp_left_get(a_type,		\
+	  a_field, (a_node)), (r_node));				\
+    } else {								\
+	a_type *rbp_p_t = (a_tree)->rbt_root;				\
+	assert(rbp_p_t != &(a_tree)->rbt_nil);				\
+	(r_node) = &(a_tree)->rbt_nil;					\
+	while (true) {							\
+	    int rbp_p_cmp = (a_cmp)((a_node), rbp_p_t);			\
+	    if (rbp_p_cmp < 0) {					\
+		rbp_p_t = rbp_left_get(a_type, a_field, rbp_p_t);	\
+	    } else if (rbp_p_cmp > 0) {					\
+		(r_node) = rbp_p_t;					\
+		rbp_p_t = rbp_right_get(a_type, a_field, rbp_p_t);	\
+	    } else {							\
+		break;							\
+	    }								\
+	    assert(rbp_p_t != &(a_tree)->rbt_nil);			\
+	}								\
+    }									\
+} while (0)
+
+#define	rb_first(a_type, a_field, a_tree, r_node) do {			\
+    rbp_first(a_type, a_field, a_tree, (a_tree)->rbt_root, (r_node));	\
+    if ((r_node) == &(a_tree)->rbt_nil) {				\
+	(r_node) = NULL;						\
+    }									\
+} while (0)
+
+#define	rb_last(a_type, a_field, a_tree, r_node) do {			\
+    rbp_last(a_type, a_field, a_tree, (a_tree)->rbt_root, r_node);	\
+    if ((r_node) == &(a_tree)->rbt_nil) {				\
+	(r_node) = NULL;						\
+    }									\
+} while (0)
+
+#define	rb_next(a_type, a_field, a_cmp, a_tree, a_node, r_node) do {	\
+    rbp_next(a_type, a_field, a_cmp, a_tree, (a_node), (r_node));	\
+    if ((r_node) == &(a_tree)->rbt_nil) {				\
+	(r_node) = NULL;						\
+    }									\
+} while (0)
+
+#define	rb_prev(a_type, a_field, a_cmp, a_tree, a_node, r_node) do {	\
+    rbp_prev(a_type, a_field, a_cmp, a_tree, (a_node), (r_node));	\
+    if ((r_node) == &(a_tree)->rbt_nil) {				\
+	(r_node) = NULL;						\
+    }									\
+} while (0)
+
+#define	rb_search(a_type, a_field, a_cmp, a_tree, a_key, r_node) do {	\
+    int rbp_se_cmp;							\
+    (r_node) = (a_tree)->rbt_root;					\
+    while ((r_node) != &(a_tree)->rbt_nil				\
+      && (rbp_se_cmp = (a_cmp)((a_key), (r_node))) != 0) {		\
+	if (rbp_se_cmp < 0) {						\
+	    (r_node) = rbp_left_get(a_type, a_field, (r_node));		\
+	} else {							\
+	    (r_node) = rbp_right_get(a_type, a_field, (r_node));	\
+	}								\
+    }									\
+    if ((r_node) == &(a_tree)->rbt_nil) {				\
+	(r_node) = NULL;						\
+    }									\
+} while (0)
+
+/*
+ * Find a match if it exists.  Otherwise, find the next greater node, if one
+ * exists.
+ */
+#define	rb_nsearch(a_type, a_field, a_cmp, a_tree, a_key, r_node) do {	\
+    a_type *rbp_ns_t = (a_tree)->rbt_root;				\
+    (r_node) = NULL;							\
+    while (rbp_ns_t != &(a_tree)->rbt_nil) {				\
+	int rbp_ns_cmp = (a_cmp)((a_key), rbp_ns_t);			\
+	if (rbp_ns_cmp < 0) {						\
+	    (r_node) = rbp_ns_t;					\
+	    rbp_ns_t = rbp_left_get(a_type, a_field, rbp_ns_t);		\
+	} else if (rbp_ns_cmp > 0) {					\
+	    rbp_ns_t = rbp_right_get(a_type, a_field, rbp_ns_t);	\
+	} else {							\
+	    (r_node) = rbp_ns_t;					\
+	    break;							\
+	}								\
+    }									\
+} while (0)
+
+/*
+ * Find a match if it exists.  Otherwise, find the previous lesser node, if one
+ * exists.
+ */
+#define	rb_psearch(a_type, a_field, a_cmp, a_tree, a_key, r_node) do {	\
+    a_type *rbp_ps_t = (a_tree)->rbt_root;				\
+    (r_node) = NULL;							\
+    while (rbp_ps_t != &(a_tree)->rbt_nil) {				\
+	int rbp_ps_cmp = (a_cmp)((a_key), rbp_ps_t);			\
+	if (rbp_ps_cmp < 0) {						\
+	    rbp_ps_t = rbp_left_get(a_type, a_field, rbp_ps_t);		\
+	} else if (rbp_ps_cmp > 0) {					\
+	    (r_node) = rbp_ps_t;					\
+	    rbp_ps_t = rbp_right_get(a_type, a_field, rbp_ps_t);	\
+	} else {							\
+	    (r_node) = rbp_ps_t;					\
+	    break;							\
+	}								\
+    }									\
+} while (0)
+
+#define	rbp_rotate_left(a_type, a_field, a_node, r_node) do {		\
+    (r_node) = rbp_right_get(a_type, a_field, (a_node));		\
+    rbp_right_set(a_type, a_field, (a_node),				\
+      rbp_left_get(a_type, a_field, (r_node)));				\
+    rbp_left_set(a_type, a_field, (r_node), (a_node));			\
+} while (0)
+
+#define	rbp_rotate_right(a_type, a_field, a_node, r_node) do {		\
+    (r_node) = rbp_left_get(a_type, a_field, (a_node));			\
+    rbp_left_set(a_type, a_field, (a_node),				\
+      rbp_right_get(a_type, a_field, (r_node)));			\
+    rbp_right_set(a_type, a_field, (r_node), (a_node));			\
+} while (0)
+
+#define	rbp_lean_left(a_type, a_field, a_node, r_node) do {		\
+    bool rbp_ll_red;							\
+    rbp_rotate_left(a_type, a_field, (a_node), (r_node));		\
+    rbp_ll_red = rbp_red_get(a_type, a_field, (a_node));		\
+    rbp_color_set(a_type, a_field, (r_node), rbp_ll_red);		\
+    rbp_red_set(a_type, a_field, (a_node));				\
+} while (0)
+
+#define	rbp_lean_right(a_type, a_field, a_node, r_node) do {		\
+    bool rbp_lr_red;							\
+    rbp_rotate_right(a_type, a_field, (a_node), (r_node));		\
+    rbp_lr_red = rbp_red_get(a_type, a_field, (a_node));		\
+    rbp_color_set(a_type, a_field, (r_node), rbp_lr_red);		\
+    rbp_red_set(a_type, a_field, (a_node));				\
+} while (0)
+
+#define	rbp_move_red_left(a_type, a_field, a_node, r_node) do {		\
+    a_type *rbp_mrl_t, *rbp_mrl_u;					\
+    rbp_mrl_t = rbp_left_get(a_type, a_field, (a_node));		\
+    rbp_red_set(a_type, a_field, rbp_mrl_t);				\
+    rbp_mrl_t = rbp_right_get(a_type, a_field, (a_node));		\
+    rbp_mrl_u = rbp_left_get(a_type, a_field, rbp_mrl_t);		\
+    if (rbp_red_get(a_type, a_field, rbp_mrl_u)) {			\
+	rbp_rotate_right(a_type, a_field, rbp_mrl_t, rbp_mrl_u);	\
+	rbp_right_set(a_type, a_field, (a_node), rbp_mrl_u);		\
+	rbp_rotate_left(a_type, a_field, (a_node), (r_node));		\
+	rbp_mrl_t = rbp_right_get(a_type, a_field, (a_node));		\
+	if (rbp_red_get(a_type, a_field, rbp_mrl_t)) {			\
+	    rbp_black_set(a_type, a_field, rbp_mrl_t);			\
+	    rbp_red_set(a_type, a_field, (a_node));			\
+	    rbp_rotate_left(a_type, a_field, (a_node), rbp_mrl_t);	\
+	    rbp_left_set(a_type, a_field, (r_node), rbp_mrl_t);		\
+	} else {							\
+	    rbp_black_set(a_type, a_field, (a_node));			\
+	}								\
+    } else {								\
+	rbp_red_set(a_type, a_field, (a_node));				\
+	rbp_rotate_left(a_type, a_field, (a_node), (r_node));		\
+    }									\
+} while (0)
+
+#define	rbp_move_red_right(a_type, a_field, a_node, r_node) do {	\
+    a_type *rbp_mrr_t;							\
+    rbp_mrr_t = rbp_left_get(a_type, a_field, (a_node));		\
+    if (rbp_red_get(a_type, a_field, rbp_mrr_t)) {			\
+	a_type *rbp_mrr_u, *rbp_mrr_v;					\
+	rbp_mrr_u = rbp_right_get(a_type, a_field, rbp_mrr_t);		\
+	rbp_mrr_v = rbp_left_get(a_type, a_field, rbp_mrr_u);		\
+	if (rbp_red_get(a_type, a_field, rbp_mrr_v)) {			\
+	    rbp_color_set(a_type, a_field, rbp_mrr_u,			\
+	      rbp_red_get(a_type, a_field, (a_node)));			\
+	    rbp_black_set(a_type, a_field, rbp_mrr_v);			\
+	    rbp_rotate_left(a_type, a_field, rbp_mrr_t, rbp_mrr_u);	\
+	    rbp_left_set(a_type, a_field, (a_node), rbp_mrr_u);		\
+	    rbp_rotate_right(a_type, a_field, (a_node), (r_node));	\
+	    rbp_rotate_left(a_type, a_field, (a_node), rbp_mrr_t);	\
+	    rbp_right_set(a_type, a_field, (r_node), rbp_mrr_t);	\
+	} else {							\
+	    rbp_color_set(a_type, a_field, rbp_mrr_t,			\
+	      rbp_red_get(a_type, a_field, (a_node)));			\
+	    rbp_red_set(a_type, a_field, rbp_mrr_u);			\
+	    rbp_rotate_right(a_type, a_field, (a_node), (r_node));	\
+	    rbp_rotate_left(a_type, a_field, (a_node), rbp_mrr_t);	\
+	    rbp_right_set(a_type, a_field, (r_node), rbp_mrr_t);	\
+	}								\
+	rbp_red_set(a_type, a_field, (a_node));				\
+    } else {								\
+	rbp_red_set(a_type, a_field, rbp_mrr_t);			\
+	rbp_mrr_t = rbp_left_get(a_type, a_field, rbp_mrr_t);		\
+	if (rbp_red_get(a_type, a_field, rbp_mrr_t)) {			\
+	    rbp_black_set(a_type, a_field, rbp_mrr_t);			\
+	    rbp_rotate_right(a_type, a_field, (a_node), (r_node));	\
+	    rbp_rotate_left(a_type, a_field, (a_node), rbp_mrr_t);	\
+	    rbp_right_set(a_type, a_field, (r_node), rbp_mrr_t);	\
+	} else {							\
+	    rbp_rotate_left(a_type, a_field, (a_node), (r_node));	\
+	}								\
+    }									\
+} while (0)
+
+#define	rb_insert(a_type, a_field, a_cmp, a_tree, a_node) do {		\
+    a_type rbp_i_s;							\
+    a_type *rbp_i_g, *rbp_i_p, *rbp_i_c, *rbp_i_t, *rbp_i_u;		\
+    int rbp_i_cmp = 0;							\
+    rbp_i_g = &(a_tree)->rbt_nil;					\
+    rbp_left_set(a_type, a_field, &rbp_i_s, (a_tree)->rbt_root);	\
+    rbp_right_set(a_type, a_field, &rbp_i_s, &(a_tree)->rbt_nil);	\
+    rbp_black_set(a_type, a_field, &rbp_i_s);				\
+    rbp_i_p = &rbp_i_s;							\
+    rbp_i_c = (a_tree)->rbt_root;					\
+    /* Iteratively search down the tree for the insertion point,      */\
+    /* splitting 4-nodes as they are encountered.  At the end of each */\
+    /* iteration, rbp_i_g->rbp_i_p->rbp_i_c is a 3-level path down    */\
+    /* the tree, assuming a sufficiently deep tree.                   */\
+    while (rbp_i_c != &(a_tree)->rbt_nil) {				\
+	rbp_i_t = rbp_left_get(a_type, a_field, rbp_i_c);		\
+	rbp_i_u = rbp_left_get(a_type, a_field, rbp_i_t);		\
+	if (rbp_red_get(a_type, a_field, rbp_i_t)			\
+	  && rbp_red_get(a_type, a_field, rbp_i_u)) {			\
+	    /* rbp_i_c is the top of a logical 4-node, so split it.   */\
+	    /* This iteration does not move down the tree, due to the */\
+	    /* disruptiveness of node splitting.                      */\
+	    /*                                                        */\
+	    /* Rotate right.                                          */\
+	    rbp_rotate_right(a_type, a_field, rbp_i_c, rbp_i_t);	\
+	    /* Pass red links up one level.                           */\
+	    rbp_i_u = rbp_left_get(a_type, a_field, rbp_i_t);		\
+	    rbp_black_set(a_type, a_field, rbp_i_u);			\
+	    if (rbp_left_get(a_type, a_field, rbp_i_p) == rbp_i_c) {	\
+		rbp_left_set(a_type, a_field, rbp_i_p, rbp_i_t);	\
+		rbp_i_c = rbp_i_t;					\
+	    } else {							\
+		/* rbp_i_c was the right child of rbp_i_p, so rotate  */\
+		/* left in order to maintain the left-leaning         */\
+		/* invariant.                                         */\
+		assert(rbp_right_get(a_type, a_field, rbp_i_p)		\
+		  == rbp_i_c);						\
+		rbp_right_set(a_type, a_field, rbp_i_p, rbp_i_t);	\
+		rbp_lean_left(a_type, a_field, rbp_i_p, rbp_i_u);	\
+		if (rbp_left_get(a_type, a_field, rbp_i_g) == rbp_i_p) {\
+		    rbp_left_set(a_type, a_field, rbp_i_g, rbp_i_u);	\
+		} else {						\
+		    assert(rbp_right_get(a_type, a_field, rbp_i_g)	\
+		      == rbp_i_p);					\
+		    rbp_right_set(a_type, a_field, rbp_i_g, rbp_i_u);	\
+		}							\
+		rbp_i_p = rbp_i_u;					\
+		rbp_i_cmp = (a_cmp)((a_node), rbp_i_p);			\
+		if (rbp_i_cmp < 0) {					\
+		    rbp_i_c = rbp_left_get(a_type, a_field, rbp_i_p);	\
+		} else {						\
+		    assert(rbp_i_cmp > 0);				\
+		    rbp_i_c = rbp_right_get(a_type, a_field, rbp_i_p);	\
+		}							\
+		continue;						\
+	    }								\
+	}								\
+	rbp_i_g = rbp_i_p;						\
+	rbp_i_p = rbp_i_c;						\
+	rbp_i_cmp = (a_cmp)((a_node), rbp_i_c);				\
+	if (rbp_i_cmp < 0) {						\
+	    rbp_i_c = rbp_left_get(a_type, a_field, rbp_i_c);		\
+	} else {							\
+	    assert(rbp_i_cmp > 0);					\
+	    rbp_i_c = rbp_right_get(a_type, a_field, rbp_i_c);		\
+	}								\
+    }									\
+    /* rbp_i_p now refers to the node under which to insert.          */\
+    rbp_node_new(a_type, a_field, a_tree, (a_node));			\
+    if (rbp_i_cmp > 0) {						\
+	rbp_right_set(a_type, a_field, rbp_i_p, (a_node));		\
+	rbp_lean_left(a_type, a_field, rbp_i_p, rbp_i_t);		\
+	if (rbp_left_get(a_type, a_field, rbp_i_g) == rbp_i_p) {	\
+	    rbp_left_set(a_type, a_field, rbp_i_g, rbp_i_t);		\
+	} else if (rbp_right_get(a_type, a_field, rbp_i_g) == rbp_i_p) {\
+	    rbp_right_set(a_type, a_field, rbp_i_g, rbp_i_t);		\
+	}								\
+    } else {								\
+	rbp_left_set(a_type, a_field, rbp_i_p, (a_node));		\
+    }									\
+    /* Update the root and make sure that it is black.                */\
+    (a_tree)->rbt_root = rbp_left_get(a_type, a_field, &rbp_i_s);	\
+    rbp_black_set(a_type, a_field, (a_tree)->rbt_root);			\
+} while (0)
+
+#define	rb_remove(a_type, a_field, a_cmp, a_tree, a_node) do {		\
+    a_type rbp_r_s;							\
+    a_type *rbp_r_p, *rbp_r_c, *rbp_r_xp, *rbp_r_t, *rbp_r_u;		\
+    int rbp_r_cmp;							\
+    rbp_left_set(a_type, a_field, &rbp_r_s, (a_tree)->rbt_root);	\
+    rbp_right_set(a_type, a_field, &rbp_r_s, &(a_tree)->rbt_nil);	\
+    rbp_black_set(a_type, a_field, &rbp_r_s);				\
+    rbp_r_p = &rbp_r_s;							\
+    rbp_r_c = (a_tree)->rbt_root;					\
+    rbp_r_xp = &(a_tree)->rbt_nil;					\
+    /* Iterate down the tree, but always transform 2-nodes to 3- or   */\
+    /* 4-nodes in order to maintain the invariant that the current    */\
+    /* node is not a 2-node.  This allows simple deletion once a leaf */\
+    /* is reached.  Handle the root specially though, since there may */\
+    /* be no way to convert it from a 2-node to a 3-node.             */\
+    rbp_r_cmp = (a_cmp)((a_node), rbp_r_c);				\
+    if (rbp_r_cmp < 0) {						\
+	rbp_r_t = rbp_left_get(a_type, a_field, rbp_r_c);		\
+	rbp_r_u = rbp_left_get(a_type, a_field, rbp_r_t);		\
+	if (rbp_red_get(a_type, a_field, rbp_r_t) == false		\
+	  && rbp_red_get(a_type, a_field, rbp_r_u) == false) {		\
+	    /* Apply standard transform to prepare for left move.     */\
+	    rbp_move_red_left(a_type, a_field, rbp_r_c, rbp_r_t);	\
+	    rbp_black_set(a_type, a_field, rbp_r_t);			\
+	    rbp_left_set(a_type, a_field, rbp_r_p, rbp_r_t);		\
+	    rbp_r_c = rbp_r_t;						\
+	} else {							\
+	    /* Move left.                                             */\
+	    rbp_r_p = rbp_r_c;						\
+	    rbp_r_c = rbp_left_get(a_type, a_field, rbp_r_c);		\
+	}								\
+    } else {								\
+	if (rbp_r_cmp == 0) {						\
+	    assert((a_node) == rbp_r_c);				\
+	    if (rbp_right_get(a_type, a_field, rbp_r_c)			\
+	      == &(a_tree)->rbt_nil) {					\
+		/* Delete root node (which is also a leaf node).      */\
+		if (rbp_left_get(a_type, a_field, rbp_r_c)		\
+		  != &(a_tree)->rbt_nil) {				\
+		    rbp_lean_right(a_type, a_field, rbp_r_c, rbp_r_t);	\
+		    rbp_right_set(a_type, a_field, rbp_r_t,		\
+		      &(a_tree)->rbt_nil);				\
+		} else {						\
+		    rbp_r_t = &(a_tree)->rbt_nil;			\
+		}							\
+		rbp_left_set(a_type, a_field, rbp_r_p, rbp_r_t);	\
+	    } else {							\
+		/* This is the node we want to delete, but we will    */\
+		/* instead swap it with its successor and delete the  */\
+		/* successor.  Record enough information to do the    */\
+		/* swap later.  rbp_r_xp is the a_node's parent.      */\
+		rbp_r_xp = rbp_r_p;					\
+		rbp_r_cmp = 1; /* Note that deletion is incomplete.   */\
+	    }								\
+	}								\
+	if (rbp_r_cmp == 1) {						\
+	    if (rbp_red_get(a_type, a_field, rbp_left_get(a_type,	\
+	      a_field, rbp_right_get(a_type, a_field, rbp_r_c)))	\
+	      == false) {						\
+		rbp_r_t = rbp_left_get(a_type, a_field, rbp_r_c);	\
+		if (rbp_red_get(a_type, a_field, rbp_r_t)) {		\
+		    /* Standard transform.                            */\
+		    rbp_move_red_right(a_type, a_field, rbp_r_c,	\
+		      rbp_r_t);						\
+		} else {						\
+		    /* Root-specific transform.                       */\
+		    rbp_red_set(a_type, a_field, rbp_r_c);		\
+		    rbp_r_u = rbp_left_get(a_type, a_field, rbp_r_t);	\
+		    if (rbp_red_get(a_type, a_field, rbp_r_u)) {	\
+			rbp_black_set(a_type, a_field, rbp_r_u);	\
+			rbp_rotate_right(a_type, a_field, rbp_r_c,	\
+			  rbp_r_t);					\
+			rbp_rotate_left(a_type, a_field, rbp_r_c,	\
+			  rbp_r_u);					\
+			rbp_right_set(a_type, a_field, rbp_r_t,		\
+			  rbp_r_u);					\
+		    } else {						\
+			rbp_red_set(a_type, a_field, rbp_r_t);		\
+			rbp_rotate_left(a_type, a_field, rbp_r_c,	\
+			  rbp_r_t);					\
+		    }							\
+		}							\
+		rbp_left_set(a_type, a_field, rbp_r_p, rbp_r_t);	\
+		rbp_r_c = rbp_r_t;					\
+	    } else {							\
+		/* Move right.                                        */\
+		rbp_r_p = rbp_r_c;					\
+		rbp_r_c = rbp_right_get(a_type, a_field, rbp_r_c);	\
+	    }								\
+	}								\
+    }									\
+    if (rbp_r_cmp != 0) {						\
+	while (true) {							\
+	    assert(rbp_r_p != &(a_tree)->rbt_nil);			\
+	    rbp_r_cmp = (a_cmp)((a_node), rbp_r_c);			\
+	    if (rbp_r_cmp < 0) {					\
+		rbp_r_t = rbp_left_get(a_type, a_field, rbp_r_c);	\
+		if (rbp_r_t == &(a_tree)->rbt_nil) {			\
+		    /* rbp_r_c now refers to the successor node to    */\
+		    /* relocate, and rbp_r_xp/a_node refer to the     */\
+		    /* context for the relocation.                    */\
+		    if (rbp_left_get(a_type, a_field, rbp_r_xp)		\
+		      == (a_node)) {					\
+			rbp_left_set(a_type, a_field, rbp_r_xp,		\
+			  rbp_r_c);					\
+		    } else {						\
+			assert(rbp_right_get(a_type, a_field,		\
+			  rbp_r_xp) == (a_node));			\
+			rbp_right_set(a_type, a_field, rbp_r_xp,	\
+			  rbp_r_c);					\
+		    }							\
+		    rbp_left_set(a_type, a_field, rbp_r_c,		\
+		      rbp_left_get(a_type, a_field, (a_node)));		\
+		    rbp_right_set(a_type, a_field, rbp_r_c,		\
+		      rbp_right_get(a_type, a_field, (a_node)));	\
+		    rbp_color_set(a_type, a_field, rbp_r_c,		\
+		      rbp_red_get(a_type, a_field, (a_node)));		\
+		    if (rbp_left_get(a_type, a_field, rbp_r_p)		\
+		      == rbp_r_c) {					\
+			rbp_left_set(a_type, a_field, rbp_r_p,		\
+			  &(a_tree)->rbt_nil);				\
+		    } else {						\
+			assert(rbp_right_get(a_type, a_field, rbp_r_p)	\
+			  == rbp_r_c);					\
+			rbp_right_set(a_type, a_field, rbp_r_p,		\
+			  &(a_tree)->rbt_nil);				\
+		    }							\
+		    break;						\
+		}							\
+		rbp_r_u = rbp_left_get(a_type, a_field, rbp_r_t);	\
+		if (rbp_red_get(a_type, a_field, rbp_r_t) == false	\
+		  && rbp_red_get(a_type, a_field, rbp_r_u) == false) {	\
+		    rbp_move_red_left(a_type, a_field, rbp_r_c,		\
+		      rbp_r_t);						\
+		    if (rbp_left_get(a_type, a_field, rbp_r_p)		\
+		      == rbp_r_c) {					\
+			rbp_left_set(a_type, a_field, rbp_r_p, rbp_r_t);\
+		    } else {						\
+			rbp_right_set(a_type, a_field, rbp_r_p,		\
+			  rbp_r_t);					\
+		    }							\
+		    rbp_r_c = rbp_r_t;					\
+		} else {						\
+		    rbp_r_p = rbp_r_c;					\
+		    rbp_r_c = rbp_left_get(a_type, a_field, rbp_r_c);	\
+		}							\
+	    } else {							\
+		/* Check whether to delete this node (it has to be    */\
+		/* the correct node and a leaf node).                 */\
+		if (rbp_r_cmp == 0) {					\
+		    assert((a_node) == rbp_r_c);			\
+		    if (rbp_right_get(a_type, a_field, rbp_r_c)		\
+		      == &(a_tree)->rbt_nil) {				\
+			/* Delete leaf node.                          */\
+			if (rbp_left_get(a_type, a_field, rbp_r_c)	\
+			  != &(a_tree)->rbt_nil) {			\
+			    rbp_lean_right(a_type, a_field, rbp_r_c,	\
+			      rbp_r_t);					\
+			    rbp_right_set(a_type, a_field, rbp_r_t,	\
+			      &(a_tree)->rbt_nil);			\
+			} else {					\
+			    rbp_r_t = &(a_tree)->rbt_nil;		\
+			}						\
+			if (rbp_left_get(a_type, a_field, rbp_r_p)	\
+			  == rbp_r_c) {					\
+			    rbp_left_set(a_type, a_field, rbp_r_p,	\
+			      rbp_r_t);					\
+			} else {					\
+			    rbp_right_set(a_type, a_field, rbp_r_p,	\
+			      rbp_r_t);					\
+			}						\
+			break;						\
+		    } else {						\
+			/* This is the node we want to delete, but we */\
+			/* will instead swap it with its successor    */\
+			/* and delete the successor.  Record enough   */\
+			/* information to do the swap later.          */\
+			/* rbp_r_xp is a_node's parent.               */\
+			rbp_r_xp = rbp_r_p;				\
+		    }							\
+		}							\
+		rbp_r_t = rbp_right_get(a_type, a_field, rbp_r_c);	\
+		rbp_r_u = rbp_left_get(a_type, a_field, rbp_r_t);	\
+		if (rbp_red_get(a_type, a_field, rbp_r_u) == false) {	\
+		    rbp_move_red_right(a_type, a_field, rbp_r_c,	\
+		      rbp_r_t);						\
+		    if (rbp_left_get(a_type, a_field, rbp_r_p)		\
+		      == rbp_r_c) {					\
+			rbp_left_set(a_type, a_field, rbp_r_p, rbp_r_t);\
+		    } else {						\
+			rbp_right_set(a_type, a_field, rbp_r_p,		\
+			  rbp_r_t);					\
+		    }							\
+		    rbp_r_c = rbp_r_t;					\
+		} else {						\
+		    rbp_r_p = rbp_r_c;					\
+		    rbp_r_c = rbp_right_get(a_type, a_field, rbp_r_c);	\
+		}							\
+	    }								\
+	}								\
+    }									\
+    /* Update root.                                                   */\
+    (a_tree)->rbt_root = rbp_left_get(a_type, a_field, &rbp_r_s);	\
+} while (0)
+
+/*
+ * The rb_proto() macro generates function prototypes that correspond to the
+ * functions generated by an equivalently parameterized call to rb_wrap().
+ */
+
+#define	rb_proto(a_attr, a_prefix, a_tree_type, a_type)			\
+a_attr void								\
+a_prefix##new(a_tree_type *tree);					\
+a_attr a_type *								\
+a_prefix##first(a_tree_type *tree);					\
+a_attr a_type *								\
+a_prefix##last(a_tree_type *tree);					\
+a_attr a_type *								\
+a_prefix##next(a_tree_type *tree, a_type *node);			\
+a_attr a_type *								\
+a_prefix##prev(a_tree_type *tree, a_type *node);			\
+a_attr a_type *								\
+a_prefix##search(a_tree_type *tree, a_type *key);			\
+a_attr a_type *								\
+a_prefix##nsearch(a_tree_type *tree, a_type *key);			\
+a_attr a_type *								\
+a_prefix##psearch(a_tree_type *tree, a_type *key);			\
+a_attr void								\
+a_prefix##insert(a_tree_type *tree, a_type *node);			\
+a_attr void								\
+a_prefix##remove(a_tree_type *tree, a_type *node);
+
+/*
+ * The rb_wrap() macro provides a convenient way to wrap functions around the
+ * cpp macros.  The main benefits of wrapping are that 1) repeated macro
+ * expansion can cause code bloat, especially for rb_{insert,remove)(), and
+ * 2) type, linkage, comparison functions, etc. need not be specified at every
+ * call point.
+ */
+
+#define	rb_wrap(a_attr, a_prefix, a_tree_type, a_type, a_field, a_cmp)	\
+a_attr void								\
+a_prefix##new(a_tree_type *tree) {					\
+    rb_new(a_type, a_field, tree);					\
+}									\
+a_attr a_type *								\
+a_prefix##first(a_tree_type *tree) {					\
+    a_type *ret;							\
+    rb_first(a_type, a_field, tree, ret);				\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##last(a_tree_type *tree) {					\
+    a_type *ret;							\
+    rb_last(a_type, a_field, tree, ret);				\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##next(a_tree_type *tree, a_type *node) {			\
+    a_type *ret;							\
+    rb_next(a_type, a_field, a_cmp, tree, node, ret);			\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##prev(a_tree_type *tree, a_type *node) {			\
+    a_type *ret;							\
+    rb_prev(a_type, a_field, a_cmp, tree, node, ret);			\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##search(a_tree_type *tree, a_type *key) {			\
+    a_type *ret;							\
+    rb_search(a_type, a_field, a_cmp, tree, key, ret);			\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##nsearch(a_tree_type *tree, a_type *key) {			\
+    a_type *ret;							\
+    rb_nsearch(a_type, a_field, a_cmp, tree, key, ret);			\
+    return (ret);							\
+}									\
+a_attr a_type *								\
+a_prefix##psearch(a_tree_type *tree, a_type *key) {			\
+    a_type *ret;							\
+    rb_psearch(a_type, a_field, a_cmp, tree, key, ret);			\
+    return (ret);							\
+}									\
+a_attr void								\
+a_prefix##insert(a_tree_type *tree, a_type *node) {			\
+    rb_insert(a_type, a_field, a_cmp, tree, node);			\
+}									\
+a_attr void								\
+a_prefix##remove(a_tree_type *tree, a_type *node) {			\
+    rb_remove(a_type, a_field, a_cmp, tree, node);			\
+}
+
+/*
+ * The iterators simulate recursion via an array of pointers that store the
+ * current path.  This is critical to performance, since a series of calls to
+ * rb_{next,prev}() would require time proportional to (n lg n), whereas this
+ * implementation only requires time proportional to (n).
+ *
+ * Since the iterators cache a path down the tree, any tree modification may
+ * cause the cached path to become invalid.  In order to continue iteration,
+ * use something like the following sequence:
+ *
+ *   {
+ *       a_type *node, *tnode;
+ *
+ *       rb_foreach_begin(a_type, a_field, a_tree, node) {
+ *           ...
+ *           rb_next(a_type, a_field, a_cmp, a_tree, node, tnode);
+ *           rb_remove(a_type, a_field, a_cmp, a_tree, node);
+ *           rb_foreach_next(a_type, a_field, a_cmp, a_tree, tnode);
+ *           ...
+ *       } rb_foreach_end(a_type, a_field, a_tree, node)
+ *   }
+ *
+ * Note that this idiom is not advised if every iteration modifies the tree,
+ * since in that case there is no algorithmic complexity improvement over a
+ * series of rb_{next,prev}() calls, thus making the setup overhead wasted
+ * effort.
+ */
+
+#define	rb_foreach_begin(a_type, a_field, a_tree, a_var) {		\
+    /* Compute the maximum possible tree depth (3X the black height). */\
+    unsigned rbp_f_height;						\
+    rbp_black_height(a_type, a_field, a_tree, rbp_f_height);		\
+    rbp_f_height *= 3;							\
+    {									\
+	/* Initialize the path to contain the left spine.             */\
+	a_type *rbp_f_path[rbp_f_height];				\
+	a_type *rbp_f_node;						\
+	bool rbp_f_synced = false;					\
+	unsigned rbp_f_depth = 0;					\
+	if ((a_tree)->rbt_root != &(a_tree)->rbt_nil) {			\
+	    rbp_f_path[rbp_f_depth] = (a_tree)->rbt_root;		\
+	    rbp_f_depth++;						\
+	    while ((rbp_f_node = rbp_left_get(a_type, a_field,		\
+	      rbp_f_path[rbp_f_depth-1])) != &(a_tree)->rbt_nil) {	\
+		rbp_f_path[rbp_f_depth] = rbp_f_node;			\
+		rbp_f_depth++;						\
+	    }								\
+	}								\
+	/* While the path is non-empty, iterate.                      */\
+	while (rbp_f_depth > 0) {					\
+	    (a_var) = rbp_f_path[rbp_f_depth-1];
+
+/* Only use if modifying the tree during iteration. */
+#define	rb_foreach_next(a_type, a_field, a_cmp, a_tree, a_node)		\
+	    /* Re-initialize the path to contain the path to a_node.  */\
+	    rbp_f_depth = 0;						\
+	    if (a_node != NULL) {					\
+		if ((a_tree)->rbt_root != &(a_tree)->rbt_nil) {		\
+		    rbp_f_path[rbp_f_depth] = (a_tree)->rbt_root;	\
+		    rbp_f_depth++;					\
+		    rbp_f_node = rbp_f_path[0];				\
+		    while (true) {					\
+			int rbp_f_cmp = (a_cmp)((a_node),		\
+			  rbp_f_path[rbp_f_depth-1]);			\
+			if (rbp_f_cmp < 0) {				\
+			    rbp_f_node = rbp_left_get(a_type, a_field,	\
+			      rbp_f_path[rbp_f_depth-1]);		\
+			} else if (rbp_f_cmp > 0) {			\
+			    rbp_f_node = rbp_right_get(a_type, a_field,	\
+			      rbp_f_path[rbp_f_depth-1]);		\
+			} else {					\
+			    break;					\
+			}						\
+			assert(rbp_f_node != &(a_tree)->rbt_nil);	\
+			rbp_f_path[rbp_f_depth] = rbp_f_node;		\
+			rbp_f_depth++;					\
+		    }							\
+		}							\
+	    }								\
+	    rbp_f_synced = true;
+
+#define	rb_foreach_end(a_type, a_field, a_tree, a_var)			\
+	    if (rbp_f_synced) {						\
+		rbp_f_synced = false;					\
+		continue;						\
+	    }								\
+	    /* Find the successor.                                    */\
+	    if ((rbp_f_node = rbp_right_get(a_type, a_field,		\
+	      rbp_f_path[rbp_f_depth-1])) != &(a_tree)->rbt_nil) {	\
+	        /* The successor is the left-most node in the right   */\
+		/* subtree.                                           */\
+		rbp_f_path[rbp_f_depth] = rbp_f_node;			\
+		rbp_f_depth++;						\
+		while ((rbp_f_node = rbp_left_get(a_type, a_field,	\
+		  rbp_f_path[rbp_f_depth-1])) != &(a_tree)->rbt_nil) {	\
+		    rbp_f_path[rbp_f_depth] = rbp_f_node;		\
+		    rbp_f_depth++;					\
+		}							\
+	    } else {							\
+		/* The successor is above the current node.  Unwind   */\
+		/* until a left-leaning edge is removed from the      */\
+		/* path, or the path is empty.                        */\
+		for (rbp_f_depth--; rbp_f_depth > 0; rbp_f_depth--) {	\
+		    if (rbp_left_get(a_type, a_field,			\
+		      rbp_f_path[rbp_f_depth-1])			\
+		      == rbp_f_path[rbp_f_depth]) {			\
+			break;						\
+		    }							\
+		}							\
+	    }								\
+	}								\
+    }									\
+}
+
+#define	rb_foreach_reverse_begin(a_type, a_field, a_tree, a_var) {	\
+    /* Compute the maximum possible tree depth (3X the black height). */\
+    unsigned rbp_fr_height;						\
+    rbp_black_height(a_type, a_field, a_tree, rbp_fr_height);		\
+    rbp_fr_height *= 3;							\
+    {									\
+	/* Initialize the path to contain the right spine.            */\
+	a_type *rbp_fr_path[rbp_fr_height];				\
+	a_type *rbp_fr_node;						\
+	bool rbp_fr_synced = false;					\
+	unsigned rbp_fr_depth = 0;					\
+	if ((a_tree)->rbt_root != &(a_tree)->rbt_nil) {			\
+	    rbp_fr_path[rbp_fr_depth] = (a_tree)->rbt_root;		\
+	    rbp_fr_depth++;						\
+	    while ((rbp_fr_node = rbp_right_get(a_type, a_field,	\
+	      rbp_fr_path[rbp_fr_depth-1])) != &(a_tree)->rbt_nil) {	\
+		rbp_fr_path[rbp_fr_depth] = rbp_fr_node;		\
+		rbp_fr_depth++;						\
+	    }								\
+	}								\
+	/* While the path is non-empty, iterate.                      */\
+	while (rbp_fr_depth > 0) {					\
+	    (a_var) = rbp_fr_path[rbp_fr_depth-1];
+
+/* Only use if modifying the tree during iteration. */
+#define	rb_foreach_reverse_prev(a_type, a_field, a_cmp, a_tree, a_node)	\
+	    /* Re-initialize the path to contain the path to a_node.  */\
+	    rbp_fr_depth = 0;						\
+	    if (a_node != NULL) {					\
+		if ((a_tree)->rbt_root != &(a_tree)->rbt_nil) {		\
+		    rbp_fr_path[rbp_fr_depth] = (a_tree)->rbt_root;	\
+		    rbp_fr_depth++;					\
+		    rbp_fr_node = rbp_fr_path[0];			\
+		    while (true) {					\
+			int rbp_fr_cmp = (a_cmp)((a_node),		\
+			  rbp_fr_path[rbp_fr_depth-1]);			\
+			if (rbp_fr_cmp < 0) {				\
+			    rbp_fr_node = rbp_left_get(a_type, a_field,	\
+			      rbp_fr_path[rbp_fr_depth-1]);		\
+			} else if (rbp_fr_cmp > 0) {			\
+			    rbp_fr_node = rbp_right_get(a_type, a_field,\
+			      rbp_fr_path[rbp_fr_depth-1]);		\
+			} else {					\
+			    break;					\
+			}						\
+			assert(rbp_fr_node != &(a_tree)->rbt_nil);	\
+			rbp_fr_path[rbp_fr_depth] = rbp_fr_node;	\
+			rbp_fr_depth++;					\
+		    }							\
+		}							\
+	    }								\
+	    rbp_fr_synced = true;
+
+#define	rb_foreach_reverse_end(a_type, a_field, a_tree, a_var)		\
+	    if (rbp_fr_synced) {					\
+		rbp_fr_synced = false;					\
+		continue;						\
+	    }								\
+	    if (rbp_fr_depth == 0) {					\
+		/* rb_foreach_reverse_sync() was called with a NULL   */\
+		/* a_node.                                            */\
+		break;							\
+	    }								\
+	    /* Find the predecessor.                                  */\
+	    if ((rbp_fr_node = rbp_left_get(a_type, a_field,		\
+	      rbp_fr_path[rbp_fr_depth-1])) != &(a_tree)->rbt_nil) {	\
+	        /* The predecessor is the right-most node in the left */\
+		/* subtree.                                           */\
+		rbp_fr_path[rbp_fr_depth] = rbp_fr_node;		\
+		rbp_fr_depth++;						\
+		while ((rbp_fr_node = rbp_right_get(a_type, a_field,	\
+		  rbp_fr_path[rbp_fr_depth-1])) != &(a_tree)->rbt_nil) {\
+		    rbp_fr_path[rbp_fr_depth] = rbp_fr_node;		\
+		    rbp_fr_depth++;					\
+		}							\
+	    } else {							\
+		/* The predecessor is above the current node.  Unwind */\
+		/* until a right-leaning edge is removed from the     */\
+		/* path, or the path is empty.                        */\
+		for (rbp_fr_depth--; rbp_fr_depth > 0; rbp_fr_depth--) {\
+		    if (rbp_right_get(a_type, a_field,			\
+		      rbp_fr_path[rbp_fr_depth-1])			\
+		      == rbp_fr_path[rbp_fr_depth]) {			\
+			break;						\
+		    }							\
+		}							\
+	    }								\
+	}								\
+    }									\
+}
+
+#endif /* RB_H_ */

jemalloc/include/jemalloc/internal/stats.h

@@ -0,0 +1,155 @@
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+#define	UMAX2S_BUFSIZE	65
+
+#ifdef JEMALLOC_STATS
+typedef struct tcache_bin_stats_s tcache_bin_stats_t;
+typedef struct malloc_bin_stats_s malloc_bin_stats_t;
+typedef struct malloc_large_stats_s malloc_large_stats_t;
+typedef struct arena_stats_s arena_stats_t;
+#endif
+#if (defined(JEMALLOC_STATS) || defined(JEMALLOC_PROF))
+typedef struct chunk_stats_s chunk_stats_t;
+#endif
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+#ifdef JEMALLOC_STATS
+
+#ifdef JEMALLOC_TCACHE
+struct tcache_bin_stats_s {
+	/*
+	 * Number of allocation requests that corresponded to the size of this
+	 * bin.
+	 */
+	uint64_t	nrequests;
+};
+#endif
+
+struct malloc_bin_stats_s {
+	/*
+	 * Number of allocation requests that corresponded to the size of this
+	 * bin.
+	 */
+	uint64_t	nrequests;
+
+#ifdef JEMALLOC_TCACHE
+	/* Number of tcache fills from this bin. */
+	uint64_t	nfills;
+
+	/* Number of tcache flushes to this bin. */
+	uint64_t	nflushes;
+#endif
+
+	/* Total number of runs created for this bin's size class. */
+	uint64_t	nruns;
+
+	/*
+	 * Total number of runs reused by extracting them from the runs tree for
+	 * this bin's size class.
+	 */
+	uint64_t	reruns;
+
+	/* High-water mark for this bin. */
+	size_t		highruns;
+
+	/* Current number of runs in this bin. */
+	size_t		curruns;
+};
+
+struct malloc_large_stats_s {
+	/*
+	 * Number of allocation requests that corresponded to this size class.
+	 */
+	uint64_t	nrequests;
+
+	/* High-water mark for this size class. */
+	size_t		highruns;
+
+	/* Current number of runs of this size class. */
+	size_t		curruns;
+};
+
+struct arena_stats_s {
+	/* Number of bytes currently mapped. */
+	size_t		mapped;
+
+	/*
+	 * Total number of purge sweeps, total number of madvise calls made,
+	 * and total pages purged in order to keep dirty unused memory under
+	 * control.
+	 */
+	uint64_t	npurge;
+	uint64_t	nmadvise;
+	uint64_t	purged;
+
+	/* Per-size-category statistics. */
+	size_t		allocated_small;
+	uint64_t	nmalloc_small;
+	uint64_t	ndalloc_small;
+
+	size_t		allocated_medium;
+	uint64_t	nmalloc_medium;
+	uint64_t	ndalloc_medium;
+
+	size_t		allocated_large;
+	uint64_t	nmalloc_large;
+	uint64_t	ndalloc_large;
+
+	/*
+	 * One element for each possible size class, including sizes that
+	 * overlap with bin size classes.  This is necessary because ipalloc()
+	 * sometimes has to use such large objects in order to assure proper
+	 * alignment.
+	 */
+	malloc_large_stats_t	*lstats;
+};
+#endif /* JEMALLOC_STATS */
+
+#if (defined(JEMALLOC_STATS) || defined(JEMALLOC_PROF))
+struct chunk_stats_s {
+#  ifdef JEMALLOC_STATS
+	/* Number of chunks that were allocated. */
+	uint64_t	nchunks;
+#  endif
+
+	/* High-water mark for number of chunks allocated. */
+	size_t		highchunks;
+
+	/*
+	 * Current number of chunks allocated.  This value isn't maintained for
+	 * any other purpose, so keep track of it in order to be able to set
+	 * highchunks.
+	 */
+	size_t		curchunks;
+};
+#endif /* JEMALLOC_STATS */
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern bool	opt_stats_print;
+
+char	*umax2s(uintmax_t x, unsigned base, char *s);
+#ifdef JEMALLOC_STATS
+void malloc_cprintf(void (*write4)(void *, const char *, const char *,
+    const char *, const char *), void *w4opaque, const char *format, ...)
+    JEMALLOC_ATTR(format(printf, 3, 4));
+void	malloc_printf(const char *format, ...)
+    JEMALLOC_ATTR(format(printf, 1, 2));
+#endif
+void	stats_print(void (*write4)(void *, const char *, const char *,
+    const char *, const char *), void *w4opaque, const char *opts);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_STATS
+#ifdef JEMALLOC_H_INLINES
+
+#endif /* JEMALLOC_H_INLINES */
+#endif /* JEMALLOC_STATS */
+/******************************************************************************/

jemalloc/include/jemalloc/internal/tcache.h

@@ -0,0 +1,286 @@
+#ifdef JEMALLOC_TCACHE
+/******************************************************************************/
+#ifdef JEMALLOC_H_TYPES
+
+typedef struct tcache_bin_s tcache_bin_t;
+typedef struct tcache_s tcache_t;
+
+/*
+ * Default number of cache slots for each bin in the thread cache (0:
+ * disabled).
+ */
+#define LG_TCACHE_NSLOTS_DEFAULT	7
+ /*
+  * (1U << opt_lg_tcache_gc_sweep) is the approximate number of allocation
+  * events between full GC sweeps (-1: disabled).  Integer rounding may cause
+  * the actual number to be slightly higher, since GC is performed
+  * incrementally.
+  */
+#define LG_TCACHE_GC_SWEEP_DEFAULT	13
+
+#endif /* JEMALLOC_H_TYPES */
+/******************************************************************************/
+#ifdef JEMALLOC_H_STRUCTS
+
+struct tcache_bin_s {
+#  ifdef JEMALLOC_STATS
+	tcache_bin_stats_t tstats;
+#  endif
+	unsigned	low_water;	/* Min # cached since last GC. */
+	unsigned	high_water;	/* Max # cached since last GC. */
+	unsigned	ncached;	/* # of cached objects. */
+	void		*slots[1];	/* Dynamically sized. */
+};
+
+struct tcache_s {
+#  ifdef JEMALLOC_STATS
+	ql_elm(tcache_t) link;		/* Used for aggregating stats. */
+#  endif
+#  ifdef JEMALLOC_PROF
+	uint64_t	prof_accumbytes;/* Cleared after arena_prof_accum() */
+#  endif
+	arena_t		*arena;		/* This thread's arena. */
+	unsigned	ev_cnt;		/* Event count since incremental GC. */
+	unsigned	next_gc_bin;	/* Next bin to GC. */
+	tcache_bin_t	*tbins[1];	/* Dynamically sized. */
+};
+
+#endif /* JEMALLOC_H_STRUCTS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_EXTERNS
+
+extern size_t	opt_lg_tcache_nslots;
+extern ssize_t	opt_lg_tcache_gc_sweep;
+
+/* Map of thread-specific caches. */
+extern __thread tcache_t	*tcache_tls
+    JEMALLOC_ATTR(tls_model("initial-exec"));
+
+/*
+ * Number of cache slots for each bin in the thread cache, or 0 if tcache is
+ * disabled.
+ */
+extern size_t			tcache_nslots;
+
+/* Number of tcache allocation/deallocation events between incremental GCs. */
+extern unsigned			tcache_gc_incr;
+
+void	tcache_bin_flush(tcache_bin_t *tbin, size_t binind, unsigned rem
+#ifdef JEMALLOC_PROF
+    , tcache_t *tcache
+#endif
+    );
+tcache_t *tcache_create(arena_t *arena);
+void	tcache_bin_destroy(tcache_t *tcache, tcache_bin_t *tbin,
+    unsigned binind);
+void	*tcache_alloc_hard(tcache_t *tcache, tcache_bin_t *tbin, size_t binind);
+tcache_bin_t *tcache_bin_create(arena_t *arena);
+void	tcache_destroy(tcache_t *tcache);
+#ifdef JEMALLOC_STATS
+void	tcache_stats_merge(tcache_t *tcache, arena_t *arena);
+#endif
+void	tcache_boot(void);
+
+#endif /* JEMALLOC_H_EXTERNS */
+/******************************************************************************/
+#ifdef JEMALLOC_H_INLINES
+
+#ifndef JEMALLOC_ENABLE_INLINE
+void	tcache_event(tcache_t *tcache);
+tcache_t *tcache_get(void);
+void	*tcache_bin_alloc(tcache_bin_t *tbin);
+void	*tcache_alloc(tcache_t *tcache, size_t size, bool zero);
+void	tcache_dalloc(tcache_t *tcache, void *ptr);
+#endif
+
+#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_TCACHE_C_))
+JEMALLOC_INLINE tcache_t *
+tcache_get(void)
+{
+	tcache_t *tcache;
+
+	if (isthreaded == false || tcache_nslots == 0)
+		return (NULL);
+
+	tcache = tcache_tls;
+	if ((uintptr_t)tcache <= (uintptr_t)1) {
+		if (tcache == NULL) {
+			tcache = tcache_create(choose_arena());
+			if (tcache == NULL)
+				return (NULL);
+		} else
+			return (NULL);
+	}
+
+	return (tcache);
+}
+
+JEMALLOC_INLINE void
+tcache_event(tcache_t *tcache)
+{
+
+	if (tcache_gc_incr == 0)
+		return;
+
+	tcache->ev_cnt++;
+	assert(tcache->ev_cnt <= tcache_gc_incr);
+	if (tcache->ev_cnt >= tcache_gc_incr) {
+		size_t binind = tcache->next_gc_bin;
+		tcache_bin_t *tbin = tcache->tbins[binind];
+
+		if (tbin != NULL) {
+			if (tbin->high_water == 0) {
+				/*
+				 * This bin went completely unused for an
+				 * entire GC cycle, so throw away the tbin.
+				 */
+				assert(tbin->ncached == 0);
+				tcache_bin_destroy(tcache, tbin, binind);
+				tcache->tbins[binind] = NULL;
+			} else {
+				if (tbin->low_water > 0) {
+					/*
+					 * Flush (ceiling) half of the objects
+					 * below the low water mark.
+					 */
+					tcache_bin_flush(tbin, binind,
+					    tbin->ncached - (tbin->low_water >>
+					    1) - (tbin->low_water & 1)
+#ifdef JEMALLOC_PROF
+					    , tcache
+#endif
+					    );
+				}
+				tbin->low_water = tbin->ncached;
+				tbin->high_water = tbin->ncached;
+			}
+		}
+
+		tcache->next_gc_bin++;
+		if (tcache->next_gc_bin == nbins)
+			tcache->next_gc_bin = 0;
+		tcache->ev_cnt = 0;
+	}
+}
+
+JEMALLOC_INLINE void *
+tcache_bin_alloc(tcache_bin_t *tbin)
+{
+
+	if (tbin->ncached == 0)
+		return (NULL);
+	tbin->ncached--;
+	if (tbin->ncached < tbin->low_water)
+		tbin->low_water = tbin->ncached;
+	return (tbin->slots[tbin->ncached]);
+}
+
+JEMALLOC_INLINE void *
+tcache_alloc(tcache_t *tcache, size_t size, bool zero)
+{
+	void *ret;
+	tcache_bin_t *tbin;
+	size_t binind;
+
+	if (size <= small_maxclass)
+		binind = small_size2bin[size];
+	else {
+		binind = mbin0 + ((MEDIUM_CEILING(size) - medium_min) >>
+		    lg_mspace);
+	}
+	assert(binind < nbins);
+	tbin = tcache->tbins[binind];
+	if (tbin == NULL) {
+		tbin = tcache_bin_create(tcache->arena);
+		if (tbin == NULL)
+			return (NULL);
+		tcache->tbins[binind] = tbin;
+	}
+
+	ret = tcache_bin_alloc(tbin);
+	if (ret == NULL) {
+		ret = tcache_alloc_hard(tcache, tbin, binind);
+		if (ret == NULL)
+			return (NULL);
+	}
+
+	if (zero == false) {
+#ifdef JEMALLOC_FILL
+		if (opt_junk)
+			memset(ret, 0xa5, size);
+		else if (opt_zero)
+			memset(ret, 0, size);
+#endif
+	} else
+		memset(ret, 0, size);
+
+#ifdef JEMALLOC_STATS
+	tbin->tstats.nrequests++;
+#endif
+#ifdef JEMALLOC_PROF
+	tcache->prof_accumbytes += tcache->arena->bins[binind].reg_size;
+#endif
+	tcache_event(tcache);
+	return (ret);
+}
+
+JEMALLOC_INLINE void
+tcache_dalloc(tcache_t *tcache, void *ptr)
+{
+	arena_t *arena;
+	arena_chunk_t *chunk;
+	arena_run_t *run;
+	arena_bin_t *bin;
+	tcache_bin_t *tbin;
+	size_t pageind, binind;
+	arena_chunk_map_t *mapelm;
+
+	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
+	arena = chunk->arena;
+	pageind = (((uintptr_t)ptr - (uintptr_t)chunk) >> PAGE_SHIFT);
+	mapelm = &chunk->map[pageind];
+	run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind -
+	    ((mapelm->bits & CHUNK_MAP_PG_MASK) >> CHUNK_MAP_PG_SHIFT)) <<
+	    PAGE_SHIFT));
+	assert(run->magic == ARENA_RUN_MAGIC);
+	bin = run->bin;
+	binind = ((uintptr_t)bin - (uintptr_t)&arena->bins) /
+	    sizeof(arena_bin_t);
+	assert(binind < nbins);
+
+#ifdef JEMALLOC_FILL
+	if (opt_junk)
+		memset(ptr, 0x5a, arena->bins[binind].reg_size);
+#endif
+
+	tbin = tcache->tbins[binind];
+	if (tbin == NULL) {
+		tbin = tcache_bin_create(choose_arena());
+		if (tbin == NULL) {
+			malloc_mutex_lock(&arena->lock);
+			arena_dalloc_bin(arena, chunk, ptr, mapelm);
+			malloc_mutex_unlock(&arena->lock);
+			return;
+		}
+		tcache->tbins[binind] = tbin;
+	}
+
+	if (tbin->ncached == tcache_nslots)
+		tcache_bin_flush(tbin, binind, (tcache_nslots >> 1)
+#ifdef JEMALLOC_PROF
+		    , tcache
+#endif
+		    );
+	assert(tbin->ncached < tcache_nslots);
+	tbin->slots[tbin->ncached] = ptr;
+	tbin->ncached++;
+	if (tbin->ncached > tbin->high_water)
+		tbin->high_water = tbin->ncached;
+
+	tcache_event(tcache);
+}
+#endif
+
+#endif /* JEMALLOC_H_INLINES */
+/******************************************************************************/
+#endif /* JEMALLOC_TCACHE */

jemalloc/include/jemalloc/jemalloc.h.in

@@ -0,0 +1,37 @@
+#ifndef JEMALLOC_H_
+#define	JEMALLOC_H_
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "jemalloc_defs@install_suffix@.h"
+#ifndef JEMALLOC_P
+#  define JEMALLOC_P(s) s
+#endif
+
+extern const char	*JEMALLOC_P(malloc_options);
+extern void		(*JEMALLOC_P(malloc_message))(void *, const char *p1,
+    const char *p2, const char *p3, const char *p4);
+
+void	*JEMALLOC_P(malloc)(size_t size) JEMALLOC_ATTR(malloc);
+void	*JEMALLOC_P(calloc)(size_t num, size_t size) JEMALLOC_ATTR(malloc);
+int	JEMALLOC_P(posix_memalign)(void **memptr, size_t alignment, size_t size)
+    JEMALLOC_ATTR(nonnull(1));
+void	*JEMALLOC_P(realloc)(void *ptr, size_t size);
+void	JEMALLOC_P(free)(void *ptr);
+
+size_t	JEMALLOC_P(malloc_usable_size)(const void *ptr);
+void	JEMALLOC_P(malloc_stats_print)(void (*write4)(void *, const char *,
+    const char *, const char *, const char *), void *w4opaque,
+    const char *opts);
+int	JEMALLOC_P(mallctl)(const char *name, void *oldp, size_t *oldlenp,
+    void *newp, size_t newlen);
+int	JEMALLOC_P(mallctlnametomib)(const char *name, size_t *mibp,
+    size_t *miblenp);
+int	JEMALLOC_P(mallctlbymib)(const size_t *mib, size_t miblen, void *oldp,
+    size_t *oldlenp, void *newp, size_t newlen);
+
+#ifdef __cplusplus
+};
+#endif
+#endif /* JEMALLOC_H_ */

jemalloc/include/jemalloc/jemalloc_defs.h.in

@@ -0,0 +1,106 @@
+#ifndef JEMALLOC_DEFS_H_
+#define	JEMALLOC_DEFS_H_
+
+/*
+ * jemalloc version string.
+ */
+#undef JEMALLOC_VERSION
+
+/*
+ * If JEMALLOC_PREFIX is defined, it will cause all public APIs to be prefixed.
+ * This makes it possible, with some care, to use multiple allocators
+ * simultaneously.
+ *
+ * In many cases it is more convenient to manually prefix allocator function
+ * calls than to let macros do it automatically, particularly when using
+ * multiple allocators simultaneously.  Define JEMALLOC_MANGLE before
+ * #include'ing jemalloc.h in order to cause name mangling that corresponds to
+ * the API prefixing.
+ */
+#undef JEMALLOC_PREFIX
+#if (defined(JEMALLOC_PREFIX) && defined(JEMALLOC_MANGLE))
+#undef JEMALLOC_P
+#endif
+
+/*
+ * Hyper-threaded CPUs may need a special instruction inside spin loops in
+ * order to yield to another virtual CPU.
+ */
+#undef CPU_SPINWAIT
+
+/* Defined if __attribute__((...)) syntax is supported. */
+#undef JEMALLOC_HAVE_ATTR
+#ifdef JEMALLOC_HAVE_ATTR
+#  define JEMALLOC_ATTR(s) __attribute__((s))
+#else
+#  define JEMALLOC_ATTR(s)
+#endif
+
+/*
+ * JEMALLOC_DEBUG enables assertions and other sanity checks, and disables
+ * inline functions.
+ */
+#undef JEMALLOC_DEBUG
+
+/* JEMALLOC_STATS enables statistics calculation. */
+#undef JEMALLOC_STATS
+
+/* JEMALLOC_PROF enables allocation profiling. */
+#undef JEMALLOC_PROF
+
+/* Use libunwind for profile backtracing if defined. */
+#undef JEMALLOC_PROF_LIBUNWIND
+
+/* Use libgcc for profile backtracing if defined. */
+#undef JEMALLOC_PROF_LIBGCC
+
+/*
+ * JEMALLOC_TINY enables support for tiny objects, which are smaller than one
+ * quantum.
+ */
+#undef JEMALLOC_TINY
+
+/*
+ * JEMALLOC_TCACHE enables a thread-specific caching layer for small and medium
+ * objects.  This makes it possible to allocate/deallocate objects without any
+ * locking when the cache is in the steady state.
+ */
+#undef JEMALLOC_TCACHE
+
+/*
+ * JEMALLOC_DSS enables use of sbrk(2) to allocate chunks from the data storage
+ * segment (DSS).
+ */
+#undef JEMALLOC_DSS
+
+/* JEMALLOC_SWAP enables mmap()ed swap file support. */
+#undef JEMALLOC_SWAP
+
+/* Support memory filling (junk/zero). */
+#undef JEMALLOC_FILL
+
+/* Support optional abort() on OOM. */
+#undef JEMALLOC_XMALLOC
+
+/* Support SYSV semantics. */
+#undef JEMALLOC_SYSV
+
+/* Support lazy locking (avoid locking unless a second thread is launched). */
+#undef JEMALLOC_LAZY_LOCK
+
+/* Determine page size at run time if defined. */
+#undef DYNAMIC_PAGE_SHIFT
+
+/* One page is 2^STATIC_PAGE_SHIFT bytes. */
+#undef STATIC_PAGE_SHIFT
+
+/* TLS is used to map arenas and magazine caches to threads. */
+#undef NO_TLS
+
+/* sizeof(void *) == 2^LG_SIZEOF_PTR. */
+#undef LG_SIZEOF_PTR
+
+/* sizeof(int) == 2^LG_SIZEOF_INT. */
+#undef LG_SIZEOF_INT
+
+#endif /* JEMALLOC_DEFS_H_ */