From 794a4eefa11f3166404d91edddd0f4f19458f652 Mon Sep 17 00:00:00 2001
From: Maria Matejka <mq@ucw.cz>
Date: Thu, 11 Nov 2021 16:25:59 +0100
Subject: [PATCH] Keeping un-unmmappable pages until they can be reused

On Linux, munmap() may fail with ENOMEM when virtual memory is too
fragmented. Working this around by just keeping such blocks for future
use.
---
 lib/locking.h       |  1 +
 lib/resource.c      | 23 +++++++++++++++++------
 sysdep/unix/alloc.c | 43 +++++++++++++++++++++++++++++++++++++++++--
 3 files changed, 59 insertions(+), 8 deletions(-)

diff --git a/lib/locking.h b/lib/locking.h
index 0a69f50f..1a8bdcd4 100644
--- a/lib/locking.h
+++ b/lib/locking.h
@@ -19,6 +19,7 @@ struct lock_order {
   struct domain_generic *attrs;
   struct domain_generic *cork;
   struct domain_generic *event;
+  struct domain_generic *resource;
 };
 
 extern _Thread_local struct lock_order locking_stack;
diff --git a/lib/resource.c b/lib/resource.c
index e80b315b..2d041ad5 100644
--- a/lib/resource.c
+++ b/lib/resource.c
@@ -60,7 +60,7 @@ static struct resclass pool_class = {
 pool root_pool;
 
 void *alloc_sys_page(void);
-void free_sys_page(void *);
+int free_sys_page(void *);
 
 static int indent;
 
@@ -98,8 +98,10 @@ pool_free(resource *P)
   if (p->pages)
     {
       ASSERT_DIE(!p->pages->used);
-      for (uint i=0; i<p->pages->free; i++)
+
+      for (uint i = 0; i < p->pages->free; i++)
 	free_sys_page(p->pages->ptr[i]);
+
       free_sys_page(p->pages);
     }
 }
@@ -476,10 +478,19 @@ 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;
+  ASSERT_DIE(p->pages->free <= POOL_PAGES_MAX);
+
+  if (p->pages->free == POOL_PAGES_MAX)
+  {
+    const unsigned long keep = POOL_PAGES_MAX / 4;
+
+    for (uint i = keep; i < p->pages->free; i++)
+      free_sys_page(p->pages->ptr[i]);
+
+    p->pages->free = keep;
+  }
+
+  p->pages->ptr[p->pages->free++] = ptr;
 }
 
 
diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c
index 4c9d5eb5..4ae1a9db 100644
--- a/sysdep/unix/alloc.c
+++ b/sysdep/unix/alloc.c
@@ -11,6 +11,8 @@
 
 #include <stdlib.h>
 #include <unistd.h>
+#include <stdatomic.h>
+#include <errno.h>
 
 #ifdef HAVE_MMAP
 #include <sys/mman.h>
@@ -19,6 +21,13 @@
 long page_size = 0;
 _Bool alloc_multipage = 0;
 
+static _Atomic int global_page_list_not_empty;
+static list global_page_list;
+static _Atomic int global_page_spinlock;
+
+#define	GLOBAL_PAGE_SPIN_LOCK	for (int v = 0; !atomic_compare_exchange_weak_explicit(&global_page_spinlock, &v, 1, memory_order_acq_rel, memory_order_acquire); v = 0)
+#define GLOBAL_PAGE_SPIN_UNLOCK	do { int v = 1; ASSERT_DIE(atomic_compare_exchange_strong_explicit(&global_page_spinlock, &v, 0, memory_order_acq_rel, memory_order_acquire)); } while (0)
+
 #ifdef HAVE_MMAP
 static _Bool use_fake = 0;
 #else
@@ -28,12 +37,14 @@ static _Bool use_fake = 1;
 void resource_sys_init(void)
 {
 #ifdef HAVE_MMAP
+  init_list(&global_page_list);
+
   if (!(page_size = sysconf(_SC_PAGESIZE)))
     die("System 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;
@@ -46,6 +57,22 @@ alloc_sys_page(void)
 #ifdef HAVE_MMAP
   if (!use_fake)
   {
+    if (atomic_load_explicit(&global_page_list_not_empty, memory_order_relaxed))
+    {
+      GLOBAL_PAGE_SPIN_LOCK;
+      if (!EMPTY_LIST(global_page_list))
+      {
+	node *ret = HEAD(global_page_list);
+	rem_node(ret);
+	if (EMPTY_LIST(global_page_list))
+	  atomic_store_explicit(&global_page_list_not_empty, 0, memory_order_relaxed);
+	GLOBAL_PAGE_SPIN_UNLOCK;
+	memset(ret, 0, sizeof(node));
+	return (void *) ret;
+      }
+      GLOBAL_PAGE_SPIN_UNLOCK;
+    }
+
     if (alloc_multipage)
     {
       void *big = mmap(NULL, page_size * 2, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
@@ -90,7 +117,19 @@ free_sys_page(void *ptr)
   if (!use_fake)
   {
     if (munmap(ptr, page_size) < 0)
-      bug("munmap(%p) failed: %m", ptr);
+#ifdef ENOMEM
+      if (errno == ENOMEM)
+      {
+	memset(ptr, 0, page_size);
+
+	GLOBAL_PAGE_SPIN_LOCK;
+	add_tail(&global_page_list, (node *) ptr);
+	atomic_store_explicit(&global_page_list_not_empty, 1, memory_order_relaxed);
+	GLOBAL_PAGE_SPIN_UNLOCK;
+      }
+      else
+#endif
+	bug("munmap(%p) failed: %m", ptr);
   }
   else
 #endif