[PATCH 4/4] libpager: Do not throw away precious pages

[Sergey Bugaev]

The kernel invokes memory_object_data_return () with dirty = 0 in two
cases: if notification on eviction was requested, or when returning
precious pages. Properly distinguish between the two cases, and only
throw the clean page away in the former case.
---
 libpager/data-return.c | 54 ++++++++++++++----------------------------
 1 file changed, 18 insertions(+), 36 deletions(-)

diff --git a/libpager/data-return.c b/libpager/data-return.c
index 2cc9f1f8..304e88bc 100644
--- a/libpager/data-return.c
+++ b/libpager/data-return.c
@@ -93,23 +93,6 @@ _pager_do_write_request (struct pager *p,
 
   pm_entries = &p->pagemap[offset / __vm_page_size];
 
-  if (! dirty)
-    {
-      munmap ((void *) data, length);
-      if (!kcopy) {
-        /* Prepare notified array.  */
-        for (i = 0; i < npages; i++)
-          notified[i] = (p->notify_on_evict
-                         && ! (pm_entries[i] & PM_PAGEINWAIT));
-
-        goto notify;
-      }
-      else {
-        _pager_allow_termination (p);
-        goto release_out;
-      }
-    }
-
   /* Make sure there are no other in-progress writes for any of these
      pages before we begin.  This imposes a little more serialization
      than we really have to require (because *all* future writes on
@@ -118,26 +101,24 @@ _pager_do_write_request (struct pager *p,
   /* XXX: Is this still needed?  */
  retry:
   for (i = 0; i < npages; i++)
-    if (pm_entries[i] & PM_PAGINGOUT)
-      {
-       pm_entries[i] |= PM_WRITEWAIT;
-       pthread_cond_wait (&p->wakeup, &p->interlock);
-       goto retry;
+    {
+      if ((initializing && (pm_entries[i] & PM_INIT)) ||
+          (!dirty && !(pm_entries[i] & PM_PRECIOUS)))
+        {
+          omit_data[i] = 1;
+          continue;
+        }
+      if (pm_entries[i] & PM_PAGINGOUT)
+        {
+          pm_entries[i] |= PM_WRITEWAIT;
+          pthread_cond_wait (&p->wakeup, &p->interlock);
+          goto retry;
       }
+    }
 
   /* Mark these pages as being paged out.  */
-  if (initializing)
-    {
-      for (i = 0; i < npages; i++)
-       {
-         if (pm_entries[i] & PM_INIT)
-           omit_data[i] = 1;
-         else
-           pm_entries[i] |= PM_PAGINGOUT | PM_INIT;
-       }
-    }
-  else
-    for (i = 0; i < npages; i++)
+  for (i = 0; i < npages; i++)
+    if (!omit_data[i])
       pm_entries[i] |= PM_PAGINGOUT | PM_INIT;
 
   /* If this write occurs while a lock is pending, record
@@ -179,7 +160,9 @@ _pager_do_write_request (struct pager *p,
     {
       if (omit_data[i])
        {
-         notified[i] = 0;
+         notified[i] = (p->notify_on_evict
+                         && !kcopy
+                         && ! (pm_entries[i] & PM_PAGEINWAIT));
          continue;
        }
 
@@ -225,7 +208,6 @@ _pager_do_write_request (struct pager *p,
   if (wakeup)
     pthread_cond_broadcast (&p->wakeup);
 
- notify:
   _pager_allow_termination (p);
   pthread_mutex_unlock (&p->interlock);
 
-- 
2.31.1