Merge commit '67256d50' into HEAD

1 files changed, 120 insertions, 58 deletions

diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c
index a2384ca8..847def30 100644
--- a/sysdep/unix/alloc.c
+++ b/sysdep/unix/alloc.c
@@ -10,6 +10,7 @@
 #include "lib/resource.h"
 #include "lib/lists.h"
 #include "lib/event.h"
+#include "lib/rcu.h"
 
 #include <errno.h>
 #include <stdlib.h>
@@ -22,46 +23,43 @@
 long page_size = 0;
 
 #ifdef HAVE_MMAP
-#define KEEP_PAGES_MAIN_MAX	256
-#define KEEP_PAGES_MAIN_MIN	8
-#define CLEANUP_PAGES_BULK	256
+#define KEEP_PAGES_MAX	512
+#define KEEP_PAGES_MIN	32
+#define KEEP_PAGES_MAX_LOCAL	16
+#define ALLOC_PAGES_AT_ONCE	8
 
-STATIC_ASSERT(KEEP_PAGES_MAIN_MIN * 4 < KEEP_PAGES_MAIN_MAX);
+STATIC_ASSERT(KEEP_PAGES_MIN * 4 < KEEP_PAGES_MAX);
+STATIC_ASSERT(ALLOC_PAGES_AT_ONCE < KEEP_PAGES_MAX_LOCAL);
 
 static _Bool use_fake = 0;
+static _Bool initialized = 0;
 
 #if DEBUGGING
 struct free_page {
   node unused[42];
-  node n;
+  struct free_page * _Atomic next;
 };
 #else
 struct free_page {
-  node n;
+  struct free_page * _Atomic next;
 };
 #endif
 
-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_page * _Atomic page_stack = NULL;
+static _Thread_local struct free_page * local_page_stack = NULL;
 
-static struct free_pages global_free_pages = {
-  .min = KEEP_PAGES_MAIN_MIN,
-  .max = KEEP_PAGES_MAIN_MAX,
-  .cleanup = { .hook = global_free_pages_cleanup_event },
-};
+static void page_cleanup(void *);
+static event page_cleanup_event = { .hook = page_cleanup, };
+#define SCHEDULE_CLEANUP  do if (initialized && !shutting_down) ev_send(&global_event_list, &page_cleanup_event); while (0)
 
-uint *pages_kept = &global_free_pages.cnt;
+_Atomic int pages_kept = 0;
+_Atomic int pages_kept_locally = 0;
+static int pages_kept_here = 0;
 
 static void *
 alloc_sys_page(void)
 {
-  void *ptr = mmap(NULL, page_size, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+  void *ptr = mmap(NULL, page_size * ALLOC_PAGES_AT_ONCE, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 
   if (ptr == MAP_FAILED)
     bug("mmap(%lu) failed: %m", page_size);
@@ -90,20 +88,33 @@ alloc_page(void)
   }
 
 #ifdef HAVE_MMAP
-  struct free_pages *fps = &global_free_pages;
-
-  if (fps->cnt)
+  struct free_page *fp = local_page_stack;
+  if (fp)
   {
-    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_send(&global_work_list, &fps->cleanup);
-
-    bzero(fp, page_size);
+    local_page_stack = atomic_load_explicit(&fp->next, memory_order_acquire);
+    atomic_fetch_sub_explicit(&pages_kept_locally, 1, memory_order_relaxed);
+    pages_kept_here--;
     return fp;
   }
 
-  return alloc_sys_page();
+  rcu_read_lock();
+  fp = atomic_load_explicit(&page_stack, memory_order_acquire);
+  while (fp && !atomic_compare_exchange_strong_explicit(
+	&page_stack, &fp, atomic_load_explicit(&fp->next, memory_order_acquire),
+	memory_order_acq_rel, memory_order_acquire))
+    ;
+  rcu_read_unlock();
+
+  if (!fp)
+  {
+    void *ptr = alloc_sys_page();
+    for (int i=1; i<ALLOC_PAGES_AT_ONCE; i++)
+      free_page(ptr + page_size * i);
+    return ptr;
+  }
+
+  atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed);
+  return fp;
 #endif
 }
 
@@ -117,45 +128,95 @@ free_page(void *ptr)
   }
 
 #ifdef HAVE_MMAP
-  struct free_pages *fps = &global_free_pages;
   struct free_page *fp = ptr;
+  if (shutting_down || (pages_kept_here < KEEP_PAGES_MAX_LOCAL))
+  {
+    atomic_store_explicit(&fp->next, local_page_stack, memory_order_relaxed);
+    atomic_fetch_add_explicit(&pages_kept_locally, 1, memory_order_relaxed);
+    pages_kept_here++;
+    return;
+  }
+
+  rcu_read_lock();
+  struct free_page *next = atomic_load_explicit(&page_stack, memory_order_acquire);
 
-  fp->n = (node) {};
-  add_tail(&fps->pages, &fp->n);
+  do atomic_store_explicit(&fp->next, next, memory_order_release);
+  while (!atomic_compare_exchange_strong_explicit(
+	&page_stack, &next, fp,
+	memory_order_acq_rel, memory_order_acquire));
+  rcu_read_unlock();
 
-  if ((++fps->cnt > fps->max) && !shutting_down)
-    ev_send(&global_work_list, &fps->cleanup);
+  if (atomic_fetch_add_explicit(&pages_kept, 1, memory_order_relaxed) >= KEEP_PAGES_MAX)
+    SCHEDULE_CLEANUP;
 #endif
 }
 
+void
+flush_local_pages(void)
+{
+  if (use_fake || !local_page_stack || shutting_down)
+    return;
+
+  struct free_page *last = local_page_stack, *next;
+  int check_count = 1;
+  while (next = atomic_load_explicit(&last->next, memory_order_acquire))
+  {
+    check_count++;
+    last = next;
+  }
+
+  ASSERT_DIE(check_count == pages_kept_here);
+
+  rcu_read_lock();
+  next = atomic_load_explicit(&page_stack, memory_order_acquire);
+
+  do atomic_store_explicit(&last->next, next, memory_order_release);
+  while (!atomic_compare_exchange_strong_explicit(
+	&page_stack, &next, local_page_stack,
+	memory_order_acq_rel, memory_order_acquire));
+  rcu_read_unlock();
+
+  local_page_stack = NULL;
+  pages_kept_here = 0;
+
+  atomic_fetch_sub_explicit(&pages_kept_locally, check_count, memory_order_relaxed);
+  if (atomic_fetch_add_explicit(&pages_kept, check_count, memory_order_relaxed) >= KEEP_PAGES_MAX)
+    SCHEDULE_CLEANUP;
+}
+
 #ifdef HAVE_MMAP
 static void
-global_free_pages_cleanup_event(void *data UNUSED)
+page_cleanup(void *_ UNUSED)
 {
   if (shutting_down)
     return;
 
-  struct free_pages *fps = &global_free_pages;
+  struct free_page *stack = atomic_exchange_explicit(&page_stack, NULL, memory_order_acq_rel);
+  if (!stack)
+    return;
 
-  while (fps->cnt / 2 < fps->min)
-  {
-    struct free_page *fp = alloc_sys_page();
-    fp->n = (node) {};
-    add_tail(&fps->pages, &fp->n);
-    fps->cnt++;
+  synchronize_rcu();
+
+  do {
+    struct free_page *f = stack;
+    stack = atomic_load_explicit(&f->next, memory_order_acquire);
+
+    if (munmap(f, page_size) == 0)
+      continue;
+    else if (errno != ENOMEM)
+      bug("munmap(%p) failed: %m", f);
+    else
+      free_page(f);
   }
+  while ((atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed) >= KEEP_PAGES_MAX / 2) && stack);
 
-  for (uint seen = 0; (seen < CLEANUP_PAGES_BULK) && (fps->cnt > fps->max / 2); seen++)
+  while (stack)
   {
-    struct free_page *fp = SKIP_BACK(struct free_page, n, TAIL(fps->pages));
-    rem_node(&fp->n);
+    struct free_page *f = stack;
+    stack = atomic_load_explicit(&f->next, memory_order_acquire);
+    free_page(f);
 
-    if (munmap(fp, page_size) == 0)
-      fps->cnt--;
-    else if (errno == ENOMEM)
-      add_head(&fps->pages, &fp->n);
-    else
-      bug("munmap(%p) failed: %m", fp);
+    atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed);
   }
 }
 #endif
@@ -164,17 +225,17 @@ void
 resource_sys_init(void)
 {
 #ifdef HAVE_MMAP
-  ASSERT_DIE(global_free_pages.cnt == 0);
-
   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);
+    for (int i = 0; i < (KEEP_PAGES_MIN * 2); i++)
+      free_page(alloc_page());
+
+    page_cleanup(NULL);
+    initialized = 1;
     return;
   }
 
@@ -184,4 +245,5 @@ resource_sys_init(void)
 #endif
 
   page_size = 4096;
+  initialized = 1;
 }