Slab: head now uses bitmask for used/free nodes info instead of lists

From now, there are no auxiliary pointers stored in the free slab nodes.
This led to strange debugging problems if use-after-free happened in
slab-allocated structures, especially if the structure's first member is
a next pointer.

This also reduces the memory needed by 1 pointer per allocated object.
OTOH, we now rely on pages being aligned to their size's multiple, which
is quite common anyway.

diff --git a/configure.ac b/configure.ac
index eabb3d562e584afa68aea5788a6660bd9193fe83..7d92a5d41a39087ac4c107290f77f25822389644 100644 (file)
--- a/configure.ac
+++ b/configure.ac
@@ -351,6 +351,7 @@ case $sysdesc in
 esac
 
 AC_CHECK_HEADERS_ONCE([alloca.h syslog.h])
+AC_CHECK_HEADER([sys/mman.h], [AC_DEFINE([HAVE_MMAP], [1], [Define to 1 if mmap() is available.])])
 AC_CHECK_MEMBERS([struct sockaddr.sa_len], [], [], [#include <sys/socket.h>])
 
 AC_C_BIGENDIAN(

diff --git a/lib/bitops.h b/lib/bitops.h
index a4c48a10f6c4ec8b5d128249615c6af715797e6b..0beda2a3525c3fe99bcac0b0c55e3311e5dbea3b 100644 (file)
--- a/lib/bitops.h
+++ b/lib/bitops.h
@@ -28,6 +28,7 @@ u32 u32_log2(u32 v);
 static inline u32 u32_hash(u32 v) { return v * 2902958171u; }
 
 static inline u8 u32_popcount(u32 v) { return __builtin_popcount(v); }
+static inline u8 u64_popcount(u64 v) { return __builtin_popcountll(v); }
 
 static inline int u32_clz(u32 v) { return __builtin_clz(v); }
 static inline int u32_ctz(u32 v) { return __builtin_ctz(v); }

diff --git a/lib/resource.h b/lib/resource.h
index b56bcff557faf03c80508c86dbb544f4ff3058cf..48e6298520ae7fa5c0bfcfac0d407c54ec553e0e 100644 (file)
--- a/lib/resource.h
+++ b/lib/resource.h
@@ -93,6 +93,10 @@ void sl_free(slab *, void *);
 
 void buffer_realloc(void **buf, unsigned *size, unsigned need, unsigned item_size);
 
+/* Allocator of whole pages; for use in slabs and other high-level allocators. */
+u64 get_page_size(void);
+void *alloc_page(void);
+void free_page(void *);
 
 #ifdef HAVE_LIBDMALLOC
 /*

diff --git a/lib/slab.c b/lib/slab.c
index f31355e02b567c5cba636bf5e36ca4e2c2c55d67..b0a01ae793edf84500b267800cb17077aeff35e6 100644 (file)
--- a/lib/slab.c
+++ b/lib/slab.c
@@ -4,6 +4,7 @@
  *     Heavily inspired by the original SLAB paper by Jeff Bonwick.
  *
  *     (c) 1998--2000 Martin Mares <mj@ucw.cz>
+ *     (c) 2020       Maria Matejka <mq@jmq.cz>
  *
  *     Can be freely distributed and used under the terms of the GNU GPL.
  */
@@ -147,12 +148,12 @@ slab_memsize(resource *r)
  *  Real efficient version.
  */
 
-#define SLAB_SIZE 4096
 #define MAX_EMPTY_HEADS 1
 
 struct slab {
   resource r;
-  uint obj_size, head_size, objs_per_slab, num_empty_heads, data_size;
+  uint obj_size, head_size, head_bitfield_len;
+  uint objs_per_slab, num_empty_heads, data_size;
   list empty_heads, partial_heads, full_heads;
 };
 
@@ -167,16 +168,8 @@ static struct resclass sl_class = {
 
 struct sl_head {
   node n;
-  struct sl_obj *first_free;
-  int num_full;
-};
-
-struct sl_obj {
-  struct sl_head *slab;
-  union {
-    struct sl_obj *next;
-    byte data[0];
-  } u;
+  u32 num_full;
+  u32 used_bits[0];
 };
 
 struct sl_alignment {                  /* Magic structure for testing of alignment */
@@ -184,6 +177,8 @@ struct sl_alignment {                       /* Magic structure for testing of alignment */
   int x[0];
 };
 
+#define SL_GET_HEAD(x) ((struct sl_head *) (((uintptr_t) (x)) & ~(get_page_size()-1)))
+
 /**
  * sl_new - create a new Slab
  * @p: resource pool
@@ -200,45 +195,32 @@ sl_new(pool *p, uint size)
   if (align < sizeof(int))
     align = sizeof(int);
   s->data_size = size;
-  size += OFFSETOF(struct sl_obj, u.data);
-  if (size < sizeof(struct sl_obj))
-    size = sizeof(struct sl_obj);
   size = (size + align - 1) / align * align;
   s->obj_size = size;
-  s->head_size = (sizeof(struct sl_head) + align - 1) / align * align;
-  s->objs_per_slab = (SLAB_SIZE - s->head_size) / size;
+
+  s->head_size = sizeof(struct sl_head);
+  u64 page_size = get_page_size();
+
+  do {
+    s->objs_per_slab = (page_size - s->head_size) / size;
+    s->head_bitfield_len = (s->objs_per_slab + 31) / 32;
+    s->head_size = (
+       sizeof(struct sl_head)
+      + sizeof(u32) * s->head_bitfield_len
+      + align - 1)
+    / align * align;
+  } while (s->objs_per_slab * size + s->head_size > page_size);
+
   if (!s->objs_per_slab)
     bug("Slab: object too large");
   s->num_empty_heads = 0;
+
   init_list(&s->empty_heads);
   init_list(&s->partial_heads);
   init_list(&s->full_heads);
   return s;
 }
 
-static struct sl_head *
-sl_new_head(slab *s)
-{
-  struct sl_head *h = xmalloc(SLAB_SIZE);
-  struct sl_obj *o = (struct sl_obj *)((byte *)h+s->head_size);
-  struct sl_obj *no;
-  uint n = s->objs_per_slab;
-
-  *h = (struct sl_head) {
-    .first_free = o,
-    .num_full = 0,
-  };
-
-  while (n--)
-    {
-      o->slab = h;
-      no = (struct sl_obj *)((char *) o+s->obj_size);
-      o->u.next = n ? no : NULL;
-      o = no;
-    }
-  return h;
-}
-
 /**
  * sl_alloc - allocate an object from Slab
  * @s: slab
@@ -250,24 +232,29 @@ void *
 sl_alloc(slab *s)
 {
   struct sl_head *h;
-  struct sl_obj *o;
 
 redo:
   h = HEAD(s->partial_heads);
   if (!h->n.next)
     goto no_partial;
 okay:
-  o = h->first_free;
-  if (!o)
-    goto full_partial;
-  h->first_free = o->u.next;
-  h->num_full++;
+  for (uint i=0; i<s->head_bitfield_len; i++)
+    if (~h->used_bits[i])
+    {
+      uint pos = u32_ctz(~h->used_bits[i]);
+      if (i * 32 + pos >= s->objs_per_slab)
+       break;
+
+      h->used_bits[i] |= 1 << pos;
+      h->num_full++;
+
+      void *out = ((void *) h) + s->head_size + (i * 32 + pos) * s->obj_size;
 #ifdef POISON
-  memset(o->u.data, 0xcd, s->data_size);
+      memset(out, 0xcd, s->data_size);
 #endif
-  return o->u.data;
+      return out;
+    }
 
-full_partial:
   rem_node(&h->n);
   add_tail(&s->full_heads, &h->n);
   goto redo;
@@ -281,7 +268,9 @@ no_partial:
       s->num_empty_heads--;
       goto okay;
     }
-  h = sl_new_head(s);
+  h = alloc_page();
+  ASSERT_DIE(SL_GET_HEAD(h) == h);
+  memset(h, 0, s->head_size);
   add_head(&s->partial_heads, &h->n);
   goto okay;
 }
@@ -313,30 +302,35 @@ sl_allocz(slab *s)
 void
 sl_free(slab *s, void *oo)
 {
-  struct sl_obj *o = SKIP_BACK(struct sl_obj, u.data, oo);
-  struct sl_head *h = o->slab;
+  struct sl_head *h = SL_GET_HEAD(oo);
 
 #ifdef POISON
   memset(oo, 0xdb, s->data_size);
 #endif
-  o->u.next = h->first_free;
-  h->first_free = o;
-  if (!--h->num_full)
+
+  uint offset = oo - ((void *) h) - s->head_size;
+  ASSERT_DIE(offset % s->obj_size == 0);
+  uint pos = offset / s->obj_size;
+  ASSERT_DIE(pos < s->objs_per_slab);
+
+  h->used_bits[pos / 32] &= ~(1 << (pos % 32));
+
+  if (h->num_full-- == s->objs_per_slab)
+    {
+      rem_node(&h->n);
+      add_head(&s->partial_heads, &h->n);
+    }
+  else if (!h->num_full)
     {
       rem_node(&h->n);
       if (s->num_empty_heads >= MAX_EMPTY_HEADS)
-       xfree(h);
+       free_page(h);
       else
        {
          add_head(&s->empty_heads, &h->n);
          s->num_empty_heads++;
        }
     }
-  else if (!o->u.next)
-    {
-      rem_node(&h->n);
-      add_head(&s->partial_heads, &h->n);
-    }
 }
 
 static void
@@ -346,11 +340,11 @@ slab_free(resource *r)
   struct sl_head *h, *g;
 
   WALK_LIST_DELSAFE(h, g, s->empty_heads)
-    xfree(h);
+    free_page(h);
   WALK_LIST_DELSAFE(h, g, s->partial_heads)
-    xfree(h);
+    free_page(h);
   WALK_LIST_DELSAFE(h, g, s->full_heads)
-    xfree(h);
+    free_page(h);
 }
 
 static void
@@ -383,7 +377,7 @@ slab_memsize(resource *r)
   WALK_LIST(h, s->full_heads)
     heads++;
 
-  return ALLOC_OVERHEAD + sizeof(struct slab) + heads * (ALLOC_OVERHEAD + SLAB_SIZE);
+  return ALLOC_OVERHEAD + sizeof(struct slab) + heads * (ALLOC_OVERHEAD + get_page_size());
 }
 
 static resource *
@@ -393,10 +387,10 @@ slab_lookup(resource *r, unsigned long a)
   struct sl_head *h;
 
   WALK_LIST(h, s->partial_heads)
-    if ((unsigned long) h < a && (unsigned long) h + SLAB_SIZE < a)
+    if ((unsigned long) h < a && (unsigned long) h + get_page_size() < a)
       return r;
   WALK_LIST(h, s->full_heads)
-    if ((unsigned long) h < a && (unsigned long) h + SLAB_SIZE < a)
+    if ((unsigned long) h < a && (unsigned long) h + get_page_size() < a)
       return r;
   return NULL;
 }

diff --git a/sysdep/unix/Makefile b/sysdep/unix/Makefile
index f592399c1ba85c8cbda5b4c899301bfddd025942..d0d36b5f7cdbe404c6a64a35220a3345ac504ab3 100644 (file)
--- a/sysdep/unix/Makefile
+++ b/sysdep/unix/Makefile
@@ -1,4 +1,4 @@
-src := io.c krt.c log.c main.c random.c
+src := alloc.c io.c krt.c log.c main.c random.c
 obj := $(src-o-files)
 $(all-daemon)
 $(cf-local)

diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c
new file mode 100644 (file)
index 0000000..c525f71
--- /dev/null
+++ b/sysdep/unix/alloc.c
@@ -0,0 +1,80 @@
+/*
+ *     BIRD Internet Routing Daemon -- Raw allocation
+ *
+ *     (c) 2020  Maria Matejka <mq@ucw.cz>
+ *
+ *     Can be freely distributed and used under the terms of the GNU GPL.
+ */
+
+#include "nest/bird.h"
+#include "lib/resource.h"
+
+#include <stdlib.h>
+#include <unistd.h>
+
+#ifdef HAVE_MMAP
+#include <sys/mman.h>
+#endif
+
+#ifdef HAVE_MMAP
+static u64 page_size = 0;
+static _Bool use_fake = 0;
+#else
+static const u64 page_size = 4096; /* Fake page size */
+#endif
+
+u64 get_page_size(void)
+{
+  if (page_size)
+    return page_size;
+
+#ifdef HAVE_MMAP
+  if (page_size = sysconf(_SC_PAGESIZE))
+  {
+    if ((u64_popcount(page_size) > 1) || (page_size > 16384))
+    {
+      /* Too big or strange page, use the aligned allocator instead */
+      page_size = 4096;
+      use_fake = 1;
+    }
+    return page_size;
+  }
+
+  bug("Page size must be non-zero");
+#endif
+}
+
+void *
+alloc_page(void)
+{
+#ifdef HAVE_MMAP
+  if (!use_fake)
+  {
+    void *ret = mmap(NULL, get_page_size(), PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    if (ret == MAP_FAILED)
+      bug("mmap(%lu) failed: %m", page_size);
+    return ret;
+  }
+  else
+#endif
+  {
+    void *ret = aligned_alloc(page_size, page_size);
+    if (!ret)
+      bug("aligned_alloc(%lu) failed", page_size);
+    return ret;
+  }
+}
+
+void
+free_page(void *ptr)
+{
+#ifdef HAVE_MMAP
+  if (!use_fake)
+  {
+    if (munmap(ptr, get_page_size()) < 0)
+      bug("munmap(%p) failed: %m", ptr);
+  }
+  else
+#endif
+    free(ptr);
+}