Bound allocated pages to resource pools with page caches to avoid unnecessary syscalls

lib/resource.c

@@ -31,9 +31,18 @@
 struct pool {
   resource r;
   list inside;
+  struct pool_pages *pages;
   const char *name;
 };
 
+struct pool_pages {
+  uint free;
+  uint used;
+  void *ptr[0];
+};
+
+#define POOL_PAGES_MAX	((page_size - sizeof(struct pool_pages)) / sizeof (void *))
+
 static void pool_dump(resource *);
 static void pool_free(resource *);
 static resource *pool_lookup(resource *, unsigned long);
@@ -50,6 +59,10 @@ static struct resclass pool_class = {
 
 pool root_pool;
 
+void *alloc_sys_page(void);
+void free_sys_page(void *);
+void resource_sys_init(void);
+
 static int indent;
 
 /**
@@ -82,6 +95,14 @@ pool_free(resource *P)
       xfree(r);
       r = rr;
     }
+
+  if (p->pages)
+    {
+      ASSERT_DIE(!p->pages->used);
+      for (uint i=0; i<p->pages->free; i++)
+	free_sys_page(p->pages->ptr[i]);
+      free_sys_page(p->pages);
+    }
 }
 
 static void
@@ -107,6 +128,9 @@ pool_memsize(resource *P)
   WALK_LIST(r, p->inside)
     sum += rmemsize(r);
 
+  if (p->pages)
+    sum += page_size * (p->pages->used + p->pages->free + 1);
+
   return sum;
 }
 
@@ -259,6 +283,7 @@ rlookup(unsigned long a)
 void
 resource_init(void)
 {
+  resource_sys_init();
   root_pool.r.class = &pool_class;
   root_pool.name = "Root";
   init_list(&root_pool.inside);
@@ -425,6 +450,39 @@ mb_free(void *m)
   rfree(b);
 }
 
+void *
+alloc_page(pool *p)
+{
+  if (!p->pages)
+  {
+    p->pages = alloc_sys_page();
+    p->pages->free = 0;
+    p->pages->used = 1;
+  }
+  else
+    p->pages->used++;
+
+  if (p->pages->free)
+  {
+    void *ptr = p->pages->ptr[--p->pages->free];
+    bzero(ptr, page_size);
+    return ptr;
+  }
+  else
+    return alloc_sys_page();
+}
+
+void
+free_page(pool *p, void *ptr)
+{
+  ASSERT_DIE(p->pages);
+  p->pages->used--;
+
+  if (p->pages->free >= POOL_PAGES_MAX)
+    return free_sys_page(ptr);
+  else
+    p->pages->ptr[p->pages->free++] = ptr;
+}
 
 
 #define STEP_UP(x) ((x) + (x)/2 + 4)

lib/resource.h

@@ -94,10 +94,12 @@ void sl_free(slab *, void *);
 
 void buffer_realloc(void **buf, unsigned *size, unsigned need, unsigned item_size);
 
+extern long page_size;
+
 /* Allocator of whole pages; for use in slabs and other high-level allocators. */
-u64 get_page_size(void);
+void *alloc_page(pool *);
-void *alloc_page(void);
+void free_page(pool *, void *);
-void free_page(void *);
+#define PAGE_HEAD(x)	((void *) (((intptr_t) (x)) & ~(page_size-1)))
 
 #ifdef HAVE_LIBDMALLOC
 /*

lib/slab.c

@@ -152,6 +152,7 @@ slab_memsize(resource *r)
 
 struct slab {
   resource r;
+  pool *p;
   uint obj_size, head_size, head_bitfield_len;
   uint objs_per_slab, num_empty_heads, data_size;
   list empty_heads, partial_heads, full_heads;
@@ -191,6 +192,7 @@ slab *
 sl_new(pool *p, uint size)
 {
   slab *s = ralloc(p, &sl_class);
+  s->p = p;
   uint align = sizeof(struct sl_alignment);
   if (align < sizeof(int))
     align = sizeof(int);
@@ -199,7 +201,6 @@ sl_new(pool *p, uint size)
   s->obj_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;
@@ -268,9 +269,9 @@ no_partial:
       s->num_empty_heads--;
       goto okay;
     }
-  h = alloc_page();
+  h = alloc_page(s->p);
 #ifdef POISON
-  memset(h, 0xba, get_page_size());
+  memset(h, 0xba, page_size);
 #endif
   ASSERT_DIE(SL_GET_HEAD(h) == h);
   memset(h, 0, s->head_size);
@@ -329,9 +330,9 @@ sl_free(slab *s, void *oo)
       if (s->num_empty_heads >= MAX_EMPTY_HEADS)
       {
 #ifdef POISON
-	memset(h, 0xde, get_page_size());
+	memset(h, 0xde, page_size);
 #endif
-	free_page(h);
+	free_page(s->p, h);
       }
       else
 	{
@@ -348,11 +349,11 @@ slab_free(resource *r)
   struct sl_head *h, *g;
 
   WALK_LIST_DELSAFE(h, g, s->empty_heads)
-    free_page(h);
+    free_page(s->p, h);
   WALK_LIST_DELSAFE(h, g, s->partial_heads)
-    free_page(h);
+    free_page(s->p, h);
   WALK_LIST_DELSAFE(h, g, s->full_heads)
-    free_page(h);
+    free_page(s->p, h);
 }
 
 static void
@@ -385,7 +386,8 @@ slab_memsize(resource *r)
   WALK_LIST(h, s->full_heads)
     heads++;
 
-  return ALLOC_OVERHEAD + sizeof(struct slab) + heads * (ALLOC_OVERHEAD + get_page_size());
+//  return ALLOC_OVERHEAD + sizeof(struct slab) + heads * (ALLOC_OVERHEAD + page_size);
+  return ALLOC_OVERHEAD + sizeof(struct slab); /* The page sizes are accounted for in the pool */
 }
 
 static resource *
@@ -395,10 +397,10 @@ 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 + get_page_size() < a)
+    if ((unsigned long) h < a && (unsigned long) h + page_size < a)
       return r;
   WALK_LIST(h, s->full_heads)
-    if ((unsigned long) h < a && (unsigned long) h + get_page_size() < a)
+    if ((unsigned long) h < a && (unsigned long) h + page_size < a)
       return r;
   return NULL;
 }

sysdep/unix/alloc.c

@@ -16,41 +16,36 @@
 #include <sys/mman.h>
 #endif
 
+long page_size = 0;
+
 #ifdef HAVE_MMAP
-static u64 page_size = 0;
 static _Bool use_fake = 0;
 #else
-static const u64 page_size = 4096; /* Fake page size */
+static _Bool use_fake = 1;
 #endif
 
-u64 get_page_size(void)
+void resource_sys_init(void)
 {
-  if (page_size)
-    return page_size;
-
 #ifdef HAVE_MMAP
-  if (page_size = sysconf(_SC_PAGESIZE))
+  if (!(page_size = sysconf(_SC_PAGESIZE)))
-  {
+    die("System page size must be non-zero");
-    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");
+  if ((u64_popcount(page_size) > 1) || (page_size > 16384))
+  {
 #endif
+    /* Too big or strange page, use the aligned allocator instead */
+    page_size = 4096;
+    use_fake = 1;
+  }
 }
 
 void *
-alloc_page(void)
+alloc_sys_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);
+    void *ret = mmap(NULL, 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;
@@ -66,12 +61,12 @@ alloc_page(void)
 }
 
 void
-free_page(void *ptr)
+free_sys_page(void *ptr)
 {
 #ifdef HAVE_MMAP
   if (!use_fake)
   {
-    if (munmap(ptr, get_page_size()) < 0)
+    if (munmap(ptr, page_size) < 0)
       bug("munmap(%p) failed: %m", ptr);
   }
   else