fix red-zone support in memorypool

src/microhttpd/memorypool.c

@@ -96,6 +96,25 @@
 #define ROUND_TO_ALIGN(n) (((n) + (ALIGN_SIZE - 1)) \
                            / (ALIGN_SIZE) *(ALIGN_SIZE))
 
+/* ROUND_TO_ALIGN_PLUS_RED_ZONE() rounds up and, when user poisoning is
+   active, adds a red zone, so its result is always >= its argument --
+   unless it wrapped.  Every caller that passes a caller-supplied size
+   must therefore reject the result with
+
+     if (asize < size)
+
+   and not by testing for zero.  Testing for zero is correct only in the
+   build where the red zone is 0: ROUND_TO_ALIGN() lands exactly on 0 for
+   the top ALIGN_SIZE-1 values of size_t, so a zero result is the only
+   wrapped outcome there.  Adding a red zone moves that outcome to
+   _MHD_RED_ZONE_SIZE, which is non-zero and small enough to pass the
+   remaining-space test below it, so a zero test silently stops firing in
+   exactly the build that has the extra instrumentation.  Sizes that are
+   already bounded by pool->size (MHD_pool_reset(), the block_offset +
+   block_size expressions) cannot wrap and need no such test -- but they
+   do need the trailing red zone clamped to pool->size, see the two sites
+   that do so. */
+
 
 #ifndef MHD_ASAN_POISON_ACTIVE
 #define _MHD_NOSANITIZE_PTRS /**/
@@ -408,8 +427,8 @@ MHD_pool_allocate (struct MemoryPool *pool,
   mhd_assert (pool->size >= pool->end - pool->pos);
   mhd_assert (pool->pos == ROUND_TO_ALIGN (pool->pos));
   asize = ROUND_TO_ALIGN_PLUS_RED_ZONE (size);
-  if ( (0 == asize) && (0 != size) )
-    return NULL; /* size too close to SIZE_MAX */
+  if (asize < size)
+    return NULL; /* Value wrap, @a size is too close to SIZE_MAX */
   if (asize > pool->end - pool->pos)
     return NULL;
   if (from_end)
@@ -490,8 +509,8 @@ MHD_pool_try_alloc (struct MemoryPool *pool,
   mhd_assert (pool->size >= pool->end - pool->pos);
   mhd_assert (pool->pos == ROUND_TO_ALIGN (pool->pos));
   asize = ROUND_TO_ALIGN_PLUS_RED_ZONE (size);
-  if ( (0 == asize) && (0 != size) )
-  { /* size is too close to SIZE_MAX, very unlikely */
+  if (asize < size)
+  { /* Value wrap, @a size is too close to SIZE_MAX, very unlikely */
     *required_bytes = SIZE_MAX;
     return NULL;
   }
@@ -623,8 +642,7 @@ MHD_pool_reallocate (struct MemoryPool *pool,
   }
   /* Need to allocate new block */
   asize = ROUND_TO_ALIGN_PLUS_RED_ZONE (new_size);
-  if ( ( (0 == asize) &&
-         (0 != new_size) ) || /* Value wrap, too large new_size. */
+  if ( (asize < new_size) || /* Value wrap, too large new_size. */
        (asize > pool->end - pool->pos) ) /* Not enough space */
     return NULL;
 
@@ -695,8 +713,15 @@ MHD_pool_deallocate (struct MemoryPool *pool,
          exactly full the two are equal, and a block allocated "from the
          end" then also satisfies 'block_offset <= pool->pos', so it
          would be mistaken for a normal block and never returned. */
-      const size_t alg_end =
+      size_t alg_end =
         ROUND_TO_ALIGN_PLUS_RED_ZONE (block_offset + block_size);
+      /* Clamped exactly as MHD_pool_reset() clamps pool->pos, and for the
+         same reason: a block that reaches the end of the pool has no room
+         for a trailing red zone, so its recorded end is pool->size.  The
+         two have to agree or the "is this the last block" test below can
+         never match such a block. */
+      if (alg_end > pool->size)
+        alg_end = pool->size;
       mhd_assert (alg_end <= pool->pos);
       if (alg_end == pool->pos)
       {
@@ -834,6 +859,16 @@ MHD_pool_reset (struct MemoryPool *pool,
             to_zero);
   }
   pool->pos = ROUND_TO_ALIGN_PLUS_RED_ZONE (new_size);
+  /* The red zone is a gap kept *inside* the pool so that an overrun of the
+     block lands in poisoned bytes.  There is no room for it once the kept
+     block reaches the end of the pool, and none is needed there either --
+     past pool->size the allocation's own ASAN red zone takes over.  Without
+     this clamp pool->pos would be set past pool->end for any @a new_size
+     within a red zone of pool->size, breaking the pool->end >= pool->pos
+     invariant that every other entry point asserts.  In the build with no
+     user poisoning _MHD_RED_ZONE_SIZE is 0 and this can never fire. */
+  if (pool->pos > pool->size)
+    pool->pos = pool->size;
   pool->end = pool->size;
   _MHD_POISON_MEMORY (((uint8_t *) pool->memory) + new_size, \
                       pool->size - new_size);