#include "atomic.h"
#include "pthread_impl.h"
-uintptr_t __brk(uintptr_t);
+#if defined(__GNUC__) && defined(__PIC__)
+#define inline inline __attribute__((always_inline))
+#endif
+
void *__mmap(void *, size_t, int, int, int, off_t);
int __munmap(void *, size_t);
void *__mremap(void *, size_t, size_t, int, ...);
};
struct bin {
- int lock[2];
+ volatile int lock[2];
struct chunk *head;
struct chunk *tail;
};
static struct {
- uintptr_t brk;
- size_t *heap;
- uint64_t binmap;
+ volatile uint64_t binmap;
struct bin bins[64];
- int brk_lock[2];
- int free_lock[2];
+ volatile int free_lock[2];
} mal;
#define DONTCARE 16
#define RECLAIM 163840
-#define CHUNK_SIZE(c) ((c)->csize & SIZE_MASK)
-#define CHUNK_PSIZE(c) ((c)->psize & SIZE_MASK)
+#define CHUNK_SIZE(c) ((c)->csize & -2)
+#define CHUNK_PSIZE(c) ((c)->psize & -2)
#define PREV_CHUNK(c) ((struct chunk *)((char *)(c) - CHUNK_PSIZE(c)))
#define NEXT_CHUNK(c) ((struct chunk *)((char *)(c) + CHUNK_SIZE(c)))
#define MEM_TO_CHUNK(p) (struct chunk *)((char *)(p) - OVERHEAD)
#define BIN_TO_CHUNK(i) (MEM_TO_CHUNK(&mal.bins[i].head))
#define C_INUSE ((size_t)1)
-#define C_FLAGS ((size_t)3)
-#define C_SIZE SIZE_MASK
#define IS_MMAPPED(c) !((c)->csize & (C_INUSE))
/* Synchronization tools */
-static void lock(volatile int *lk)
+static inline void lock(volatile int *lk)
{
- if (!libc.threads_minus_1) return;
- while(a_swap(lk, 1)) __wait(lk, lk+1, 1, 1);
+ if (libc.threads_minus_1)
+ while(a_swap(lk, 1)) __wait(lk, lk+1, 1, 1);
}
-static void unlock(volatile int *lk)
+static inline void unlock(volatile int *lk)
{
- if (!libc.threads_minus_1) return;
- a_store(lk, 0);
- if (lk[1]) __wake(lk, 1, 1);
+ if (lk[0]) {
+ a_store(lk, 0);
+ if (lk[1]) __wake(lk, 1, 1);
+ }
}
-static void lock_bin(int i)
+static inline void lock_bin(int i)
{
- if (libc.threads_minus_1)
- lock(mal.bins[i].lock);
+ lock(mal.bins[i].lock);
if (!mal.bins[i].head)
mal.bins[i].head = mal.bins[i].tail = BIN_TO_CHUNK(i);
}
-static void unlock_bin(int i)
+static inline void unlock_bin(int i)
{
- if (!libc.threads_minus_1) return;
unlock(mal.bins[i].lock);
}
}
#endif
+void *__expand_heap(size_t *);
+
static struct chunk *expand_heap(size_t n)
{
+ static int heap_lock[2];
+ static void *end;
+ void *p;
struct chunk *w;
- uintptr_t new;
- lock(mal.brk_lock);
+ /* The argument n already accounts for the caller's chunk
+ * overhead needs, but if the heap can't be extended in-place,
+ * we need room for an extra zero-sized sentinel chunk. */
+ n += SIZE_ALIGN;
- if (n > SIZE_MAX - mal.brk - 2*PAGE_SIZE) goto fail;
- new = mal.brk + n + SIZE_ALIGN + PAGE_SIZE - 1 & -PAGE_SIZE;
- n = new - mal.brk;
+ lock(heap_lock);
- if (__brk(new) != new) goto fail;
-
- w = MEM_TO_CHUNK(new);
- w->psize = n | C_INUSE;
- w->csize = 0 | C_INUSE;
-
- w = MEM_TO_CHUNK(mal.brk);
- w->csize = n | C_INUSE;
- mal.brk = new;
-
- unlock(mal.brk_lock);
-
- return w;
-fail:
- unlock(mal.brk_lock);
- return 0;
-}
-
-static int init_malloc(size_t n)
-{
- static int init, waiters;
- int state;
- struct chunk *c;
-
- if (init == 2) return 0;
-
- while ((state=a_swap(&init, 1)) == 1)
- __wait(&init, &waiters, 1, 1);
- if (state) {
- a_store(&init, 2);
+ p = __expand_heap(&n);
+ if (!p) {
+ unlock(heap_lock);
return 0;
}
- mal.brk = __brk(0) + 2*SIZE_ALIGN-1 & -SIZE_ALIGN;
+ /* If not just expanding existing space, we need to make a
+ * new sentinel chunk below the allocated space. */
+ if (p != end) {
+ /* Valid/safe because of the prologue increment. */
+ n -= SIZE_ALIGN;
+ p = (char *)p + SIZE_ALIGN;
+ w = MEM_TO_CHUNK(p);
+ w->psize = 0 | C_INUSE;
+ }
- c = expand_heap(n);
+ /* Record new heap end and fill in footer. */
+ end = (char *)p + n;
+ w = MEM_TO_CHUNK(end);
+ w->psize = n | C_INUSE;
+ w->csize = 0 | C_INUSE;
- if (!c) {
- a_store(&init, 0);
- if (waiters) __wake(&init, 1, 1);
- return -1;
- }
+ /* Fill in header, which may be new or may be replacing a
+ * zero-size sentinel header at the old end-of-heap. */
+ w = MEM_TO_CHUNK(p);
+ w->csize = n | C_INUSE;
- mal.heap = (void *)c;
- c->psize = 0 | C_INUSE;
- free(CHUNK_TO_MEM(c));
+ unlock(heap_lock);
- a_store(&init, 2);
- if (waiters) __wake(&init, -1, 1);
- return 1;
+ return w;
}
static int adjust_size(size_t *n)
for (;;) {
uint64_t mask = mal.binmap & -(1ULL<<i);
if (!mask) {
- if (init_malloc(n) > 0) continue;
c = expand_heap(n);
if (!c) return 0;
if (alloc_rev(c)) {
j = first_set(mask);
lock_bin(j);
c = mal.bins[j].head;
- if (c != BIN_TO_CHUNK(j) && j == bin_index(c->csize)) {
+ if (c != BIN_TO_CHUNK(j)) {
if (!pretrim(c, n, i, j)) unbin(c, j);
unlock_bin(j);
break;
size_t oldlen = n0 + extra;
size_t newlen = n + extra;
/* Crash on realloc of freed chunk */
- if ((uintptr_t)base < mal.brk) *(volatile char *)0=0;
+ if (extra & 1) a_crash();
if (newlen < PAGE_SIZE && (new = malloc(n))) {
memcpy(new, p, n-OVERHEAD);
free(p);
next = NEXT_CHUNK(self);
+ /* Crash on corrupted footer (likely from buffer overflow) */
+ if (next->psize != self->csize) a_crash();
+
/* Merge adjacent chunks if we need more space. This is not
* a waste of time even if we fail to get enough space, because our
* subsequent call to free would otherwise have to do the merge. */
char *base = (char *)self - extra;
size_t len = CHUNK_SIZE(self) + extra;
/* Crash on double free */
- if ((uintptr_t)base < mal.brk) *(volatile char *)0=0;
+ if (extra & 1) a_crash();
__munmap(base, len);
return;
}
final_size = new_size = CHUNK_SIZE(self);
next = NEXT_CHUNK(self);
+ /* Crash on corrupted footer (likely from buffer overflow) */
+ if (next->psize != self->csize) a_crash();
+
for (;;) {
/* Replace middle of large chunks with fresh zero pages */
if (reclaim && (self->psize & next->csize & C_INUSE)) {
projects / musl / blobdiff