fix: process-wide spinlock in userspace heap

This commit is contained in:
2026-05-25 11:42:44 +02:00
parent 29d5daaa62
commit f13e691e06
+39 -10
View File
@@ -37,6 +37,22 @@ namespace heap_detail {
inline FreeNode g_overflow{0, nullptr};
inline bool g_initialized = false;
// Process-wide heap lock. Userspace threads share the heap, so the
// public malloc/mfree/realloc entry points must serialize access to
// g_buckets/g_overflow. Kept inline here (not in thread.h) because
// thread.h depends on heap.h, and the internal helpers below are not
// reentrant into the public API, so a plain spinlock suffices.
inline volatile uint32_t g_heap_lock = 0;
static inline void heap_lock_acquire() {
while (__atomic_exchange_n(&g_heap_lock, 1, __ATOMIC_ACQUIRE) != 0) {
syscall0(Montauk::SYS_YIELD);
}
}
static inline void heap_lock_release() {
__atomic_store_n(&g_heap_lock, 0, __ATOMIC_RELEASE);
}
static inline Header* get_header(void* block) {
return (Header*)((uint8_t*)block - sizeof(Header));
}
@@ -146,15 +162,17 @@ namespace heap_detail {
inline void* malloc(uint64_t size) {
using namespace heap_detail;
// Guard against overflow: size + Header must not wrap
if (size > UINT64_MAX - sizeof(Header) - 15)
return nullptr;
heap_lock_acquire();
if (!g_initialized) {
grow(16 * 0x1000); // seed with 64 KiB
g_initialized = true;
}
// Guard against overflow: size + Header must not wrap
if (size > UINT64_MAX - sizeof(Header) - 15)
return nullptr;
uint64_t needed = size + sizeof(Header);
needed = (needed + 15) & ~15ULL;
@@ -162,8 +180,10 @@ namespace heap_detail {
if (idx >= 0) {
// Small allocation — use segregated bucket (O(1))
if (g_buckets[idx] == nullptr && !refill_bucket(idx))
if (g_buckets[idx] == nullptr && !refill_bucket(idx)) {
heap_lock_release();
return nullptr;
}
FreeNode* node = g_buckets[idx];
g_buckets[idx] = node->next;
@@ -171,20 +191,22 @@ namespace heap_detail {
Header* header = (Header*)node;
header->magic = HEADER_MAGIC;
header->size = size;
heap_lock_release();
return (void*)((uint8_t*)header + sizeof(Header));
}
// Large allocation — search overflow list
void* block = take_from_overflow(needed);
if (block == nullptr) {
if (!grow(needed)) return nullptr;
if (!grow(needed)) { heap_lock_release(); return nullptr; }
block = take_from_overflow(needed);
if (block == nullptr) return nullptr;
if (block == nullptr) { heap_lock_release(); return nullptr; }
}
Header* header = (Header*)block;
header->magic = HEADER_MAGIC;
header->size = size;
heap_lock_release();
return (void*)((uint8_t*)header + sizeof(Header));
}
@@ -195,8 +217,10 @@ namespace heap_detail {
Header* header = get_header(ptr);
if (header->magic == FREED_MAGIC) return; // double-free
if (header->magic != HEADER_MAGIC) return; // corrupt
heap_lock_acquire();
if (header->magic == FREED_MAGIC) { heap_lock_release(); return; } // double-free
if (header->magic != HEADER_MAGIC) { heap_lock_release(); return; } // corrupt
header->magic = FREED_MAGIC;
uint64_t blockSize = header->size + sizeof(Header);
@@ -214,15 +238,18 @@ namespace heap_detail {
// Large block — sorted insert with coalescing
insert_overflow((void*)header, blockSize);
}
heap_lock_release();
}
inline void* realloc(void* ptr, uint64_t size) {
if (ptr == nullptr) return malloc(size);
// Read old size under the lock to avoid racing with another
// thread that might be freeing/recycling this header.
heap_detail::heap_lock_acquire();
auto* header = heap_detail::get_header(ptr);
uint64_t old = header->size;
// Compute actual block size (accounting for bucket rounding)
uint64_t oldBlock = (old + sizeof(heap_detail::Header) + 15) & ~15ULL;
int idx = heap_detail::bucket_index(oldBlock);
if (idx >= 0) oldBlock = heap_detail::BUCKET_SIZES[idx];
@@ -230,8 +257,10 @@ namespace heap_detail {
uint64_t newNeed = (size + sizeof(heap_detail::Header) + 15) & ~15ULL;
if (newNeed <= oldBlock) {
header->size = size;
heap_detail::heap_lock_release();
return ptr;
}
heap_detail::heap_lock_release();
void* newBlock = malloc(size);
if (newBlock == nullptr) return nullptr;