Single-threaded version of sark-branch memory page management

5 files changed, 137 insertions, 79 deletions

diff --git a/lib/resource.c b/lib/resource.c
index 5636872c..a179afe3 100644
--- a/lib/resource.c
+++ b/lib/resource.c
@@ -270,6 +270,8 @@ 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);
diff --git a/lib/resource.h b/lib/resource.h
index 0e4c44d8..992fa5b2 100644
--- a/lib/resource.h
+++ b/lib/resource.h
@@ -110,10 +110,11 @@ 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);
+extern long page_size;
 void *alloc_page(void);
 void free_page(void *);
-extern uint pages_kept;
+
+void resource_sys_init(void);
 
 #ifdef HAVE_LIBDMALLOC
 /*
diff --git a/lib/slab.c b/lib/slab.c
index 9a4c3ee2..9be9844d 100644
--- a/lib/slab.c
+++ b/lib/slab.c
@@ -180,7 +180,7 @@ 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)))
+#define SL_GET_HEAD(x)	((struct sl_head *) (((uintptr_t) (x)) & ~(page_size-1)))
 
 /**
  * sl_new - create a new Slab
@@ -202,7 +202,6 @@ 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;
@@ -273,7 +272,7 @@ no_partial:
     }
   h = alloc_page();
 #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);
@@ -332,7 +331,7 @@ 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);
       }
@@ -399,7 +398,7 @@ slab_memsize(resource *r)
 
   return (struct resmem) {
     .effective = eff,
-    .overhead = ALLOC_OVERHEAD + sizeof(struct slab) + heads * get_page_size() - eff,
+    .overhead = ALLOC_OVERHEAD + sizeof(struct slab) + heads * page_size - eff,
   };
 }
 
@@ -410,10 +409,10 @@ 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;
 }
diff --git a/nest/cmds.c b/nest/cmds.c
index 1a16f9c7..8481bf96 100644
--- a/nest/cmds.c
+++ b/nest/cmds.c
@@ -108,6 +108,7 @@ print_size(char *dsc, struct resmem vals)
 
 extern pool *rt_table_pool;
 extern pool *rta_pool;
+extern uint *pages_kept;
 
 void
 cmd_show_memory(void)
@@ -119,8 +120,8 @@ cmd_show_memory(void)
   print_size("Protocols:", rmemsize(proto_pool));
   struct resmem total = rmemsize(&root_pool);
 #ifdef HAVE_MMAP
-  print_size("Standby memory:", (struct resmem) { .overhead = get_page_size() * pages_kept });
-  total.overhead += get_page_size() * pages_kept;
+  print_size("Standby memory:", (struct resmem) { .overhead = page_size * *pages_kept });
+  total.overhead += page_size * *pages_kept;
 #endif
   print_size("Total:", total);
   cli_msg(0, "");
diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c
index 0e944d57..f96c0fcf 100644
--- a/sysdep/unix/alloc.c
+++ b/sysdep/unix/alloc.c
@@ -11,6 +11,7 @@
 #include "lib/lists.h"
 #include "lib/event.h"
 
+#include <errno.h>
 #include <stdlib.h>
 #include <unistd.h>
 
@@ -18,113 +19,167 @@
 #include <sys/mman.h>
 #endif
 
-#ifdef HAVE_MMAP
-#define KEEP_PAGES  512
+long page_size = 0;
 
-static u64 page_size = 0;
-static _Bool use_fake = 0;
+#ifdef HAVE_MMAP
+#define KEEP_PAGES_MAIN_MAX	256
+#define KEEP_PAGES_MAIN_MIN	8
+#define CLEANUP_PAGES_BULK	256
 
-uint pages_kept = 0;
-static list pages_list;
+_Static_assert(KEEP_PAGES_MAIN_MIN * 4 < KEEP_PAGES_MAIN_MAX);
 
-static void cleanup_pages(void *data);
-static event page_cleanup_event = { .hook = cleanup_pages };
+static _Bool use_fake = 0;
 
+#if DEBUGGING
+struct free_page {
+  node unused[42];
+  node n;
+};
 #else
-static const u64 page_size = 4096; /* Fake page size */
+struct free_page {
+  node n;
+};
 #endif
 
-u64 get_page_size(void)
+struct free_pages {
+  list pages;
+  u16 min, max;		/* Minimal and maximal number of free pages kept */
+  uint cnt;		/* Number of empty pages */
+  event cleanup;
+};
+
+static void global_free_pages_cleanup_event(void *);
+
+static struct free_pages global_free_pages = {
+  .min = KEEP_PAGES_MAIN_MIN,
+  .max = KEEP_PAGES_MAIN_MAX,
+  .cleanup = { .hook = global_free_pages_cleanup_event },
+};
+
+uint *pages_kept = &global_free_pages.cnt;
+
+static void *
+alloc_sys_page(void)
 {
-  if (page_size)
-    return page_size;
+  void *ptr = mmap(NULL, page_size, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
-#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;
-  }
+  if (ptr == MAP_FAILED)
+    bug("mmap(%lu) failed: %m", page_size);
 
-  bug("Page size must be non-zero");
-#endif
+  return ptr;
 }
 
+extern int shutting_down; /* Shutdown requested. */
+
+#else // ! HAVE_MMAP
+#define use_fake  1
+#endif
+
 void *
 alloc_page(void)
 {
-#ifdef HAVE_MMAP
-  if (pages_kept)
-  {
-    node *page = TAIL(pages_list);
-    rem_node(page);
-    pages_kept--;
-    memset(page, 0, get_page_size());
-    return page;
-  }
-
-  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", (long unsigned int) page_size);
-    return ret;
-  }
-  else
-#endif
+  if (use_fake)
   {
     void *ptr = NULL;
     int err = posix_memalign(&ptr, page_size, page_size);
+
     if (err || !ptr)
       bug("posix_memalign(%lu) failed", (long unsigned int) page_size);
+
     return ptr;
   }
+
+#ifdef HAVE_MMAP
+  struct free_pages *fps = &global_free_pages;
+
+  if (fps->cnt)
+  {
+    struct free_page *fp = SKIP_BACK(struct free_page, n, HEAD(fps->pages));
+    rem_node(&fp->n);
+    if ((--fps->cnt < fps->min) && !shutting_down)
+      ev_schedule(&fps->cleanup);
+
+    bzero(fp, page_size);
+    return fp;
+  }
+
+  return alloc_sys_page();
+#endif
 }
 
 void
 free_page(void *ptr)
 {
-#ifdef HAVE_MMAP
-  if (!use_fake)
+  if (use_fake)
   {
-    if (!pages_kept)
-      init_list(&pages_list);
+    free(ptr);
+    return;
+  }
+
+#ifdef HAVE_MMAP
+  struct free_pages *fps = &global_free_pages;
+  struct free_page *fp = ptr;
 
-    memset(ptr, 0, sizeof(node));
-    add_tail(&pages_list, ptr);
+  fp->n = (node) {};
+  add_tail(&fps->pages, &fp->n);
 
-    if (++pages_kept > KEEP_PAGES)
-      ev_schedule(&page_cleanup_event);
-  }
-  else
+  if ((++fps->cnt > fps->max) && !shutting_down)
+    ev_schedule(&fps->cleanup);
 #endif
-    free(ptr);
 }
 
 #ifdef HAVE_MMAP
 static void
-cleanup_pages(void *data UNUSED)
+global_free_pages_cleanup_event(void *data UNUSED)
 {
-  for (uint seen = 0; (pages_kept > KEEP_PAGES) && (seen < KEEP_PAGES); seen++)
+  if (shutting_down)
+    return;
+
+  struct free_pages *fps = &global_free_pages;
+
+  while (fps->cnt / 2 < fps->min)
   {
-    void *ptr = HEAD(pages_list);
-    rem_node(ptr);
-    if (munmap(ptr, get_page_size()) == 0)
-      pages_kept--;
-#ifdef ENOMEM
+    struct free_page *fp = alloc_sys_page();
+    fp->n = (node) {};
+    add_tail(&fps->pages, &fp->n);
+    fps->cnt++;
+  }
+
+  for (uint seen = 0; (seen < CLEANUP_PAGES_BULK) && (fps->cnt > fps->max / 2); seen++)
+  {
+    struct free_page *fp = SKIP_BACK(struct free_page, n, TAIL(fps->pages));
+    rem_node(&fp->n);
+
+    if (munmap(fp, page_size) == 0)
+      fps->cnt--;
     else if (errno == ENOMEM)
-      add_tail(&pages_list, ptr);
-#endif
+      add_head(&fps->pages, &fp->n);
     else
-      bug("munmap(%p) failed: %m", ptr);
+      bug("munmap(%p) failed: %m", fp);
   }
-
-  if (pages_kept > KEEP_PAGES)
-    ev_schedule(&page_cleanup_event);
 }
 #endif
+
+void
+resource_sys_init(void)
+{
+#ifdef HAVE_MMAP
+  if (!(page_size = sysconf(_SC_PAGESIZE)))
+    die("System page size must be non-zero");
+
+  if (u64_popcount(page_size) == 1)
+  {
+    struct free_pages *fps = &global_free_pages;
+
+    init_list(&fps->pages);
+    global_free_pages_cleanup_event(NULL);
+    return;
+  }
+
+  /* Too big or strange page, use the aligned allocator instead */
+  log(L_WARN "Got strange memory page size (%lu), using the aligned allocator instead", page_size);
+  use_fake = 1;
+#endif
+
+  page_size = 4096;
+}