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

@@ -0,0 +1,2 @@
+
+BasedOnStyle: WebKit

.gitignore

@@ -0,0 +1,5 @@
+obj/
+test/
+*.o
+*.a
+*.so

Makefile

@@ -1,27 +1,72 @@
 
-CC := gcc
-CFLAGS := -ggdb -O0 -std=c99 -Wall -Wextra -Wpedantic
+all:
 
-OBJS := obj/arena.o obj/alloc_backend.o
+BUILD ?= debug
+COMPILER ?= gcc
 
-all : test/test_arena
+# ==== set compiler flags ====
+# credits Maxim Egorushkin:
+# https://stackoverflow.com/questions/48791883/best-practice-for-building-a-make-file/48793058#48793058
 
-test/test_arena : src/test_arena.c $(OBJS) | test
+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) $^
 
-obj/%.o : src/%.c | obj
+$(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) $<
 
-obj:
+$(BUILD_DIR)/fpic/%.o : src/%.c  | $(BUILD_DIR)/fpic
+	$(CC) -o $@ -c -fPIC $(CFLAGS) $<
+
+$(BUILD_DIR):
 	mkdir -p $@
 
-bin:
+$(BUILD_DIR)/fpic:
 	mkdir -p $@
 
-test:
+$(BIN_DIR) :
+	mkdir -p $@
+
+$(TEST_DIR) :
 	mkdir -p $@
 
 clean:
-	rm -rf obj bin test
+	rm -rf $(BUILD_DIR) $(BIN_DIR) $(TEST_DIR)
 
-.PHONY: clean obj test bin all
+.PHONY: clean $(BUILD_DIR) test bin all

README.md

@@ -0,0 +1,13 @@
+
+
+`arena_t* arena_new()`
+
+Initializes and returns new arena
+
+`void arena_reset(arena_t *a)`
+
+Free memory allocated in arena
+
+`void* arena_alloc(arena_t *a, size_t len)`
+
+Allocate new memory using arena

compile_flags.txt

@@ -0,0 +1,5 @@
+-Wall
+-Wextra
+-Wpedantic
+-std=c2x
+-Iinclude

include/arena.h

@@ -0,0 +1,24 @@
+
+#ifndef ARENA_H
+#define ARENA_H
+
+#include <stddef.h> // ptrdiff_t
+#include <stdint.h> // uintptr_t
+
+#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 {
+    struct page *head, *last;
+} __attribute__((aligned(_WORD_SIZE))) arena_t;
+
+arena_t arena_new(void);
+void arena_reset(arena_t* a);
+void* arena_alloc(arena_t* a, size_t len);
+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

src/alloc_backend.c

@@ -1,30 +0,0 @@
-
-#include "alloc_backend.h"
-
-#define _GNU_SOURCE
-#include <sys/mman.h>
-
-/*
- * Separate function to make switching
- * allocation backend easier
- *
- * must take size_t size as an argument
- * must return NULL on failure
- */
-void* call_alloc_backend(size_t size)
-{
-    void *p = mmap(
-            NULL,
-            size,
-            PROT_READ | PROT_WRITE, 
-            MAP_ANONYMOUS | MAP_PRIVATE,
-            -1, /* man mmap(2): "[...], some implementations require fd to be
-                   -1 if MAP_ANONYMOUS is specified [...]" */
-            0
-    );
-
-    if (p == MAP_FAILED)
-        return NULL;
-
-    return p;
-}

src/alloc_backend.h

@@ -1,9 +0,0 @@
-
-#ifndef ALLOC_BACKEND_H
-#define ALLOC_BACKEND_H
-
-#include <stddef.h>
-
-void* call_alloc_backend(size_t size);
-
-#endif

src/arena.c

@@ -1,60 +1,158 @@
 
-
 #include "arena.h"
-#include "alloc_backend.h"
 
 #include <errno.h>
-#include <unistd.h>
+#include <stdint.h>
+#include <stdlib.h>
 #include <string.h>
+#include <unistd.h>
 
-#define ARENA_SIZE ((size_t)(128*sysconf(_SC_PAGE_SIZE)))
+// (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))
 
-/*
- * Allocates and returns new arena
- */
-struct arena* arena_new()
+#define BIG_PAGE (SYS_PAGE_SIZE + 1)
+
+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)
 {
-    size_t size = ARENA_SIZE;
-
-    unsigned char *p = call_alloc_backend(size);
+    struct page* p = malloc(sizeof *p);
 
     if (p == NULL)
-        return NULL;
+        exit(errno);
 
-    struct arena a = {
-        .begin = p + sizeof(struct arena),
-        .next  = p + sizeof(struct arena),
-        .cap   = size
+    p->next = NULL;
+    p->offset = 0;
+    p->prev_offset = 0;
+    p->data = malloc(SYS_PAGE_SIZE);
+
+    if (p->data == NULL)
+        exit(errno);
+
+    arena_t a = {
+        .head = p,
+        .last = p,
     };
 
-    memcpy(p, &a, sizeof a);
-
-    return (struct arena*)p;
+    return a;
 }
 
-
-/*
- * Frees all memory in arena
- */
-void arena_reset(struct arena *a)
+void _arena_new_page(arena_t* a, size_t size)
 {
-    a->next = a->begin;
-}
-
-
-/*
- * Allocate new memory using arena
- */
-void* arena_alloc(struct arena *a, size_t len)
-{
-    void *p = a->next;
-    a->next += len;
-
-    if (a->next - a->begin >= a->cap) {
-        errno = ENOMEM;
-        return NULL;
+    /* 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)
+{
+    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);
+    }
+
+    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);
+    memset(p, 0, nmemb * size);
     return p;
 }
 
+void* arena_realloc_tail(arena_t* a, size_t len)
+{
+    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;
+    }
+
+    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->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/arena.h

@@ -1,19 +0,0 @@
-
-#ifndef ARENA_H
-#define ARENA_H
-
-
-#include <stddef.h> // ptrdiff_t
-
-struct arena {
-    unsigned char *begin,
-                  *next;
-    ptrdiff_t cap;
-};
-
-struct arena* arena_new();
-void arena_reset(struct arena *a);
-void* arena_alloc(struct arena *a, size_t len);
-
-
-#endif

src/test_arena.c

@@ -1,44 +1,42 @@
 
 
-
 // _start test_arena_alloc.c
 #include "arena.h"
-#include <err.h>
-#include <stdio.h>
-#include <errno.h>
 
-static struct arena *default_arena = NULL;
+#include <err.h>
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
 
 int main()
 {
-    default_arena = arena_new();
+    arena_t default_arena = arena_new();
 
-    if (default_arena == NULL) {
-        err(errno, "failed to allocate arena");
+    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("\n%p\n", default_arena->next);
+    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);
 
-    char *ten_A = arena_alloc(default_arena, 11 * sizeof *ten_A);
-    char *ten_B = arena_alloc(default_arena, 11 * sizeof *ten_B);
-    char *ten_C = arena_alloc(default_arena, 11 * sizeof *ten_C);
+    if (c == NULL)
+        err(EXIT_FAILURE, "failed to allocate memory");
 
-    for (size_t i=0; i<10; ++i) {
-        ten_A[i] = 'A';
-        ten_B[i] = 'B';
-        ten_C[i] = 'C';
-    }
+    for (size_t i = 0; i < psz; i++)
+        *c = i & 0xFF;
 
-    ten_A[10] = '\0';
-    ten_B[10] = '\0';
-    ten_C[10] = '\0';
+    printf("\n    OK!\n");
 
-    printf("\n%s %s %s", ten_A, ten_B, ten_C);
-	printf("\n%p\n", default_arena->next);
-
-	arena_reset(default_arena);
-
-	printf("\n%p\n", default_arena->next);
+    arena_delete(&default_arena);
 
     return 0;
 }