optimize shrinking reallocs

`arena_realloc_tail` exceeding PAGE_SIZE would create a big page,
preceeding calls to `arena_realloc_tail` would then cause memory bugs.

.clang-format

@@ -1,10 +1,2 @@
 
 BasedOnStyle: WebKit
-
-AlignAfterOpenBracket: Align
-AlignConsecutiveAssignments: true
-AlignConsecutiveDeclarations: true
-AlignEscapedNewlines: Left
-AlignOperands: true
-AlignTrailingComments: false
-PointerAlignment: Right

.gitignore

@@ -1,5 +1,5 @@
 obj/
+test/
 *.o
 *.a
 *.so
-build/

CMakeLists.txt

@@ -1,15 +0,0 @@
-cmake_minimum_required(VERSION 3.16)
-
-project("ArenaAllocator"
-	VERSION 1.0
-	LANGUAGES C
-	DESCRIPTION "Arena Allocator in C"
-)
-
-add_subdirectory(src)
-
-if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME)
-	message(STATUS "Building tests")
-	add_subdirectory(test)
-endif()
-

Makefile

@@ -0,0 +1,72 @@
+
+all:
+
+BUILD ?= debug
+COMPILER ?= gcc
+
+# ==== set compiler flags ====
+# credits Maxim Egorushkin:
+# https://stackoverflow.com/questions/48791883/best-practice-for-building-a-make-file/48793058#48793058
+
+CC := $(COMPILER)
+
+# -fsanitize={address,undefined}
+CFLAGS.gcc.debug := -Og -ggdb -pg -fsanitize=address -fno-omit-frame-pointer
+CFLAGS.gcc.release := -O3 -g -march=native -DNDEBUG
+CFLAGS.gcc := -fanalyzer
+
+CFLAGS.clang.debug :=-O0 -g3 -DBACKTRACE -rdynamic
+CFLAGS.clang.release :=-O3 -g -march=native -DNDEBUG
+CFLAGS.clang :=
+
+CFLAGS := ${CFLAGS.${COMPILER}} ${CFLAGS.${COMPILER}.${BUILD}} \
+		  -Iinclude \
+		  -std=c2x \
+		  -Wall -Wextra -Wpedantic -Werror
+
+BUILD_DIR := build
+BIN_DIR := lib
+TEST_DIR := test
+
+OBJS := $(BUILD_DIR)/arena.o
+FPIC_OBJS := $(BUILD_DIR)/fpic/arena.o
+
+all : static
+
+static : $(BIN_DIR)/libarena.a
+
+dynamic : $(BIN_DIR)/libarena.so
+
+tests : test/test_arena
+
+$(TEST_DIR)/test_arena : src/test_arena.c $(BIN_DIR)/libarena.a | $(TEST_DIR)
+	$(CC) -o $@ $(CFLAGS) $^
+
+$(BIN_DIR)/libarena.a : $(OBJS) | $(BIN_DIR)
+	ar cr $@ $^
+
+$(BIN_DIR)/libarena.so : $(FPIC_OBJS) | $(BIN_DIR)
+	$(CC) -o $@ -shared $(CFLAGS) $^
+
+$(BUILD_DIR)/%.o : src/%.c | $(BUILD_DIR)
+	$(CC) -o $@ -c $(CFLAGS) $<
+
+$(BUILD_DIR)/fpic/%.o : src/%.c  | $(BUILD_DIR)/fpic
+	$(CC) -o $@ -c -fPIC $(CFLAGS) $<
+
+$(BUILD_DIR):
+	mkdir -p $@
+
+$(BUILD_DIR)/fpic:
+	mkdir -p $@
+
+$(BIN_DIR) :
+	mkdir -p $@
+
+$(TEST_DIR) :
+	mkdir -p $@
+
+clean:
+	rm -rf $(BUILD_DIR) $(BIN_DIR) $(TEST_DIR)
+
+.PHONY: clean $(BUILD_DIR) test bin all

README.md

@@ -1,65 +1,13 @@
 
-# Arena Allocator
-
-Arena allocators are region based allocators that tie many allocations to a
-single region of memory. Benefits are massivly simplified allocation and
-deallocation for complex structures, increased performance due to improved
-cache locality, and reduced memory fragmentation as long as individual items
-don't need to be deallocated. For programs that need to micro-manage individual
-allocations this is not an ideal solution.
-
-Deallocating a region of memory with arenas is extremely fast, because the
-arena length is just set to 0. Allocating memory is also extremely fast.  This
-implementation also grows the arena on demand using mmap and mprotect.
-Extending the library with a new type of allocation strategy should be easy.
-
-# Reference
-
 
-## `arena_t arena_new()`
+`arena_t* arena_new()`
 
-Allocate a new arena.
+Initializes and returns new arena
 
-If `arena_new()` fails, the struct member `data` is set to NULL.  This can be
-checked with `arena_new_failed()`
+`void arena_reset(arena_t *a)`
 
+Free memory allocated in arena
 
-## `bool arena_new_failed(arena_t *a)`
-
-Returns true if creating a new arena failed
-
-
-## `int arena_delete(arena_t *a)`
-
-Delete underlying buffer of arena.  Should only be used with arenas from
-`arena_new()`.
-
-Returns 0 on success, -1 on failure
-
-
-## `arena_t arena_attach(void *ptr, size_t size)`
-
-Attach an arena to an existing memory region.  The arena will not expand if
-capacity is exceeded.
-
-
-## `void *arena_detatch(arena_t arena)`
-
-Detach an arena from an existing memory region. Returns the underlying data.
-
-
-## `void arena_reset(arena_t *a)`
-
-Resets an arena.
-
-
-## `void *arena_alloc(arena_t *a, size_t len)`
-
-Allocate memory with an arena. Returns NULL and sets errno on failure.
-
-
-## `void *arena_calloc(arena_t *a, size_t nmemb, size_t size)`
-
-Allocate and zero memory with an arena.  Returns NULL and sets errno on
-failure.
+`void* arena_alloc(arena_t *a, size_t len)`
 
+Allocate new memory using arena

include/arena.h

@@ -1,81 +1,24 @@
 
-#pragma once
 #ifndef ARENA_H
 #define ARENA_H
 
-#include <stdbool.h>
-#include <stddef.h>
+#include <stddef.h> // ptrdiff_t
+#include <stdint.h> // uintptr_t
 
-// flags = 0 should be considered the default
-enum arena_flags {
-    ARENA_GROW      = 1 << 0,
-    ARENA_DONTALIGN = 1 << 1,
-};
+#define _WORD_SIZE (sizeof(intptr_t))
 
+// linked list terminology:
+// https://web.archive.org/web/20230604145332/https://i.stack.imgur.com/2FbXf.png
 typedef struct arena {
-    char*  data;
-    size_t size;
-    size_t cap;
-    char   flags;
-} arena_t;
+    struct page *head, *last;
+} __attribute__((aligned(_WORD_SIZE))) arena_t;
 
-/**
- * Allocate a new arena.
- * The underlying memory is allocated with mmap.
- * Errors can be checked with `arena_new_failed()`
- */
-arena_t arena_new();
-
-/**
- * Delete memory mapped for arena.
- * Should only be used with arenas from arena_new().
- * Returns 0 on success, -1 on failure
- */
-int arena_delete(arena_t *a);
-
-/**
- * Attach an arena to an existing memory region.
- * The arena will not expand the region if space is exceeded.
- */
-static inline arena_t arena_attach(void *ptr, size_t size)
-{
-    return (arena_t) { .data = ptr, .size = 0, .cap = size, .flags = 0 };
-}
-
-/**
- * Detach an arena from an existing memory region.
- */
-static inline void *arena_detatch(arena_t arena)
-{
-    return arena.data;
-}
-
-/**
- * Returns true if creating new arena failed
- */
-static inline bool arena_new_failed(arena_t *a)
-{
-    return a->data == NULL;
-}
-
-/**
- * Reset an arena.
- */
-static inline void arena_reset(arena_t *a)
-{
-    a->size = 0;
-}
-
-/**
- * Allocate memory from an arena.
- * Returns NULL and sets errno on failure.
- */
+arena_t arena_new(void);
+void arena_reset(arena_t* a);
 void* arena_alloc(arena_t* a, size_t len);
-
-/**
- * Allocate and zero memory from an arena.
- * Returns NULL and sets errno on failure.
- */
 void* arena_calloc(arena_t* a, size_t nmemb, size_t size);
+void* arena_realloc_tail(arena_t* a, size_t len);
+void arena_delete(arena_t* a);
+void arena_free(arena_t* a, void* p);
 
 #endif

include/knob.h

@@ -1,6 +0,0 @@
-
-#pragma once
-
-#define KNOB_MMAP_SIZE (1UL << 36UL)
-#define KNOB_ALIGNMENT (sizeof(char*))
-

src/CMakeLists.txt

@@ -1,4 +0,0 @@
-
-add_library(arena arena.c)
-
-target_include_directories(arena PUBLIC ../include)

src/arena.c

@@ -1,125 +1,158 @@
 
 #include "arena.h"
-#include "knob.h"
 
-#include <errno.h> // errno
-#include <stdbool.h>
-#include <stdio.h> // fprintf
-#include <string.h> // strerror
-#include <sys/mman.h>
+#include <errno.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
 #include <unistd.h>
 
+// (sizeof(intptr_t) isn't guaranteed to be the machine word size but on most
+// compilers it is)
+#define SYS_PAGE_SIZE ((size_t)sysconf(_SC_PAGE_SIZE))
 
-#ifndef NDEBUG
-#define arena_err(msg) \
-    fprintf(stderr, "%s (%s:%d): %s\n", msg, __func__, __LINE__, strerror(errno))
-#else
-#define arena_err(msg)
-#endif
+#define BIG_PAGE (SYS_PAGE_SIZE + 1)
 
-static bool arena_grow(struct arena *a, size_t min_size)
+typedef unsigned char byte_t;
+
+struct page {
+    void* data;
+    size_t offset, prev_offset;
+    struct page* next;
+};
+
+// linked list terminology:
+// https://web.archive.org/web/20230604145332/https://i.stack.imgur.com/2FbXf.png
+
+arena_t arena_new(void)
 {
-    if (!(a->flags & ARENA_GROW)) {
-        return false;
-    }
+    struct page* p = malloc(sizeof *p);
 
-    size_t new_cap = a->cap * 2;
-    while (new_cap < min_size) {
-        new_cap *= 2;
-    }
+    if (p == NULL)
+        exit(errno);
 
-    int ok = mprotect(
-        a->data + a->cap,
-        new_cap - a->cap,
-        PROT_READ | PROT_WRITE
-    );
+    p->next = NULL;
+    p->offset = 0;
+    p->prev_offset = 0;
+    p->data = malloc(SYS_PAGE_SIZE);
 
-    if (ok == -1) {
-        arena_err("mprotect");
-        return false;
-    }
+    if (p->data == NULL)
+        exit(errno);
 
-    a->cap = new_cap;
-
-    return true;
-}
-
-
-arena_t arena_new()
-{
-    size_t size = sysconf(_SC_PAGE_SIZE);
-    if (size == -1) {
-        arena_err("sysconf");
-        goto sysconf_failed;
-    }
-
-    void *p = mmap(NULL, KNOB_MMAP_SIZE, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
-    if (p == MAP_FAILED) {
-        arena_err("mmap");
-        goto mmap_failed;
-    }
-
-    int ok = mprotect(p, size, PROT_READ | PROT_WRITE);
-    if (ok == -1) {
-        arena_err("mprotect");
-        goto mprotect_failed;
-    }
-
-    struct arena a = {
-        .data = p,
-        .size = 0,
-        .cap = size,
-        .flags = ARENA_GROW,
+    arena_t a = {
+        .head = p,
+        .last = p,
     };
 
     return a;
-
-mprotect_failed:
-    ok = munmap(p, KNOB_MMAP_SIZE);
-    if (ok == -1) {
-        arena_err("munmap");
 }
-mmap_failed:
-sysconf_failed:
-    return (arena_t) { 0 };
+
+void _arena_new_page(arena_t* a, size_t size)
+{
+    /* potentially reuse page from previously
+       reset arena */
+    if (a->head->next != NULL) {
+        a->head = a->head->next;
+        a->head->offset = 0;
+        a->head->prev_offset = 0;
+        return;
+    }
+
+    void* tmp = calloc(1, sizeof *(a->head->next));
+
+    if (tmp == NULL)
+        exit(errno);
+
+    a->head->next = tmp;
+
+    a->head = a->head->next;
+    a->head->data = malloc(size);
+
+    if (a->head->data == NULL)
+        exit(errno);
+}
+
+void arena_reset(arena_t* a)
+{
+    a->head = a->last;
+    a->head->offset = 0;
+    a->head->prev_offset = 0;
+}
+
+void* _arena_big_alloc(arena_t* a, size_t size)
+{
+    _arena_new_page(a, size);
+    a->head->offset = BIG_PAGE;
+    a->head->prev_offset = BIG_PAGE;
+
+    return a->head->data;
 }
 
 void* arena_alloc(arena_t* a, size_t size)
 {
-    // align
-    if (!(a->flags & ARENA_DONTALIGN)) {
-        size = (size + KNOB_ALIGNMENT - 1) & ~(KNOB_ALIGNMENT - 1);
+    if (size > SYS_PAGE_SIZE)
+        return _arena_big_alloc(a, size);
+
+    // align size to machine word size
+    size = (size + _WORD_SIZE - 1) & ~(_WORD_SIZE - 1);
+
+    if (a->head->offset > SYS_PAGE_SIZE - size) {
+        _arena_new_page(a, SYS_PAGE_SIZE);
+        return arena_alloc(a, size);
     }
 
-    void *p = a->data + a->size;
-    if (a->size + size > a->cap) {
-        if (!arena_grow(a, a->size))
-            return NULL;
-    }
-    a->size += size;
-    return p;
+    a->head->prev_offset = a->head->offset;
+    a->head->offset += size;
+
+    return (byte_t*)(a->head->data) + a->head->prev_offset;
 }
 
 void* arena_calloc(arena_t* a, size_t nmemb, size_t size)
 {
     void* p = arena_alloc(a, nmemb * size);
-    if (p == NULL)
-        return p;
     memset(p, 0, nmemb * size);
     return p;
 }
 
-int arena_delete(struct arena *a)
+void* arena_realloc_tail(arena_t* a, size_t len)
 {
-    if (!(a->flags & ARENA_GROW)) {
-        return -1;
+    if (len <= (a->head->offset - a->head->prev_offset)) {
+        a->head->offset = a->head->prev_offset + len;
+        return (byte_t*)(a->head->data) + a->head->prev_offset;
     }
-    int ok = munmap(a->data, KNOB_MMAP_SIZE);
-    if (ok == -1) {
-        arena_err("munmap");
-        return -1;
+
+    if (a->head->offset == BIG_PAGE) {
+        void* tmp = realloc(a->head->data, len);
+
+        if (tmp == NULL)
+            exit(errno);
+
+        a->head->data = tmp;
+
+        return tmp;
     }
-    a->cap  = -1;
-    a->size = -1;
-    return 0;
+
+    a->head->offset = a->head->prev_offset;
+
+    return arena_alloc(a, len);
+}
+
+void arena_delete(arena_t* a)
+{
+    struct page* p = a->last;
+
+    while (p != NULL) {
+        struct page* next = p->next;
+        free(p->data);
+        free(p);
+        p = next;
+    }
+}
+
+// included for completeness
+void arena_free(arena_t* a, void* p)
+{
+    (void)p;
+    (void)a;
+    return;
 }

src/test_arena.c

@@ -0,0 +1,42 @@
+
+
+// _start test_arena_alloc.c
+#include "arena.h"
+
+#include <err.h>
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+int main()
+{
+    arena_t default_arena = arena_new();
+
+    printf("\nAttempting to allocate and write to 1 byte 1024 times");
+    for (size_t i = 0; i < 1024; i++) {
+        char* c = arena_alloc(&default_arena, sizeof *c);
+
+        if (c == NULL)
+            err(EXIT_FAILURE, "failed to allocate memory");
+
+        *c = i & 0xFF;
+    }
+    printf("\n    OK!\n");
+
+    printf("\nAttempting to allocate and write to `_SC_PAGESIZE+1` bytes");
+    size_t psz = sysconf(_SC_PAGESIZE) + 1;
+    char* c = arena_alloc(&default_arena, psz);
+
+    if (c == NULL)
+        err(EXIT_FAILURE, "failed to allocate memory");
+
+    for (size_t i = 0; i < psz; i++)
+        *c = i & 0xFF;
+
+    printf("\n    OK!\n");
+
+    arena_delete(&default_arena);
+
+    return 0;
+}

test/CMakeLists.txt

@@ -1,6 +0,0 @@
-
-include_directories(${CMAKE_SOURCE_DIR}/include)
-
-add_executable(test_arena test_arena.c)
-
-target_link_libraries(test_arena arena)

test/README

@@ -1,6 +0,0 @@
-
-run compiled test in conjunction with other tools, e.g.
-
-/usr/bin/time strace -e trace=\!write ./test/test_arena
-
- 

test/test_arena.c

@@ -1,153 +0,0 @@
-
-#include "arena.h"
-#include "knob.h"
-
-#include <signal.h>
-#include <stdio.h>  // fprintf
-#include <stdlib.h> // exit, EXIT_FAILURE
-#include <string.h> // memset
-#include <sys/wait.h> // waitpid
-#include <unistd.h> // sysconf
-
-int main()
-{
-    size_t page_size = sysconf(_SC_PAGE_SIZE);
-    if (page_size == -1) {
-        perror("sysconf");
-        exit(EXIT_FAILURE);
-    }
-
-    /*
-     * test arena from arena_new();
-     * ===============================
-     */
-    {
-        fprintf(stderr, "creating new arena with arena_new()\n    ");
-        arena_t a = arena_new();
-        if (arena_new_failed(&a)) {
-            fprintf(stderr, "arena_new failed\n");
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "OK\n");
-
-        fprintf(stderr, "attempting many small allocations\n    ");
-        for (int i = 0; i < page_size * 8; i++) {
-            char* s = arena_alloc(&a, 4 * sizeof *s);
-            if (!s) {
-                fprintf(stderr, "arena_alloc failed\n");
-                exit(EXIT_FAILURE);
-            }
-            memset(s, 'a', 4);
-        }
-        fprintf(stderr, "OK\n");
-
-        fprintf(stderr, "testing allocations of cap * 3 + 123\n    ");
-        for (int i = 0; i < 2; i++) {
-            size_t n = a.cap * 3 + 123;
-            volatile char* s = arena_alloc(&a, n);
-            if (!s) {
-                fprintf(stderr, "arena_alloc failed\n");
-                exit(EXIT_FAILURE);
-            }
-        }
-        fprintf(stderr, "OK\n");
-
-        fprintf(stderr, "calling arena_delete() on arena from arena_new()\n    ");
-        int ok = arena_delete(&a);
-        if (ok == -1) {
-            fprintf(stderr, "arena_delete failed\n");
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "OK\n");
-    }
-
-    /*
-     * test arena made by arena_attach();
-     * ===============================
-     */
-    {
-        fprintf(stderr, "creating new arena with arena_attach() and malloc()\n    ");
-        char* p = malloc(page_size);
-        if (!p) {
-            perror("malloc");
-            exit(EXIT_FAILURE);
-        }
-        arena_t a = arena_attach(p, page_size);
-        fprintf(stderr, "OK\n");
-
-        bool failed = false;
-
-        fprintf(stderr, "try to grow more than the buffer size (should fail)\n    ");
-        int i;
-        for (i=0; i < page_size + 1; i++) {
-            char* s = arena_alloc(&a, 8);
-            if (!s) {
-                failed = true;
-                break;
-            }
-        }
-        if (!failed) {
-            fprintf(stderr, "allocation was supposed to fail, but didn't\n");
-            exit(EXIT_FAILURE);
-        } else if (i != page_size / KNOB_ALIGNMENT ) {
-            fprintf(stderr, "allocation failed after %d allocations, but should fail after %d!\n", i, page_size/sizeof(void*));
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "OK\n");
-
-        fprintf(stderr, "deleting an arena not made with arena_new (should fail)\n    ");
-        int ok = arena_delete(&a);
-        if (ok != -1) {
-            fprintf(stderr, "arena_delete was supposed to fail, but didn't\n");
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "OK\n");
-
-        free(arena_detatch(a));
-    }
-
-    /*
-     * test the memory protection
-     * ===============================
-     */
-    {
-        fprintf(stderr, "attempting to access memory beyond arena cap (should sigsegv)\n    ");
-        pid_t p;
-        switch (p = fork()) {
-        case -1: /* error */
-            perror("fork()");
-            exit(EXIT_FAILURE);
-
-        case 0: /* child */
-            arena_t a = arena_new();
-            if (arena_new_failed(&a)) {
-                fprintf(stderr, "arena_new failed\n");
-                exit(EXIT_FAILURE);
-            }
-
-            char* s = arena_alloc(&a, 12345);
-            if (!s) {
-                fprintf(stderr, "arena_alloc failed\n");
-                exit(EXIT_FAILURE);
-            }
-
-            // should sigsegv
-            a.data[a.cap] = 1;
-            exit(EXIT_SUCCESS);
-        }
-
-        int status;
-        int ok = waitpid(p, &status, 0);
-        if (!ok) {
-            perror("waitpid");
-        }
-
-        if (!( WIFSIGNALED(status) && WTERMSIG(status) == SIGSEGV)) {
-            fprintf(stderr, "allocating beyond arena cap succeeded when it shouldn't!\n");
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "OK\n");
-    }
-
-    return EXIT_SUCCESS;
-}