[PATCH RFC 7/8] mm: Keep pages available for allocation while scrubbing

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

Instead of scrubbing pages while holding heap lock we can mark
buddy's head as being scrubbed and drop the lock temporarily.
If someone (most likely alloc_heap_pages()) tries to access
this chunk it will signal the scrubber to abort scrub by setting
head's PAGE_SCRUB_ABORT bit. The scrubber checks this bit after
processing each page and stops its work as soon as it sees it.

Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
---
 xen/common/page_alloc.c  |   67 +++++++++++++++++++++++++++++++++++++++++----
 xen/include/asm-x86/mm.h |    4 +++
 2 files changed, 65 insertions(+), 6 deletions(-)

diff --git a/xen/common/page_alloc.c b/xen/common/page_alloc.c
index 3469185..a39afd4 100644
--- a/xen/common/page_alloc.c
+++ b/xen/common/page_alloc.c
@@ -687,6 +687,18 @@ static void check_low_mem_virq(void)
     }
 }
 
+static void check_and_stop_scrub(struct page_info *head)
+{
+    if ( head->u.free.scrub_state & PAGE_SCRUBBING )
+    {
+        head->u.free.scrub_state |= PAGE_SCRUB_ABORT;
+        smp_mb();
+        spink_lock_kick();
+        while ( ACCESS_ONCE(head->u.free.scrub_state) & PAGE_SCRUB_ABORT )
+            cpu_relax();
+    }
+}
+
 /* Allocate 2^@order contiguous pages. */
 static struct page_info *alloc_heap_pages(
     unsigned int zone_lo, unsigned int zone_hi,
@@ -773,10 +785,15 @@ static struct page_info *alloc_heap_pages(
             {
                 if ( (pg = page_list_remove_head(&heap(node, zone, j))) )
                 {
-                    if ( (order == 0) || use_unscrubbed ||
-                         !test_bit(_PGC_need_scrub, &pg[0].count_info) )
+                    if ( !test_bit(_PGC_need_scrub, &pg[0].count_info) )
                         goto found;
 
+                    if ( (order == 0) || use_unscrubbed )
+                    {
+                        check_and_stop_scrub(pg);
+                        goto found;
+                    }
+
                     page_list_add_tail(pg, &heap(node, zone, j));
                 }
             }
@@ -911,6 +928,8 @@ static int reserve_offlined_page(struct page_info *head)
 
     cur_head = head;
 
+    check_and_stop_scrub(head);
+
     page_list_del(head, &heap(node, zone, head_order));
 
     while ( cur_head < (head + (1 << head_order)) )
@@ -990,6 +1009,9 @@ static bool_t can_merge(struct page_info *head, unsigned int node,
          !!test_bit(_PGC_need_scrub, &head->count_info) )
         return 0;
 
+    if ( head->u.free.scrub_state & PAGE_SCRUBBING )
+        return 0;
+
     return 1;
 }
 
@@ -1033,6 +1055,14 @@ static void merge_chunks(struct page_info *pg, unsigned int node,
 }
 
 #define SCRUB_CHUNK_ORDER  8
+
+static bool_t scrub_continue(void *data)
+{
+    struct page_info *head = (struct page_info *)data;
+
+    return !(ACCESS_ONCE(head->u.free.scrub_state) & PAGE_SCRUB_ABORT);
+}
+
 bool_t scrub_free_pages(unsigned int node)
 {
     struct page_info *pg;
@@ -1064,14 +1094,25 @@ bool_t scrub_free_pages(unsigned int node)
                 if ( !test_bit(_PGC_need_scrub, &pg[0].count_info) )
                     break;
 
-                page_list_del(pg, &heap(node, zone, order));
+                ASSERT(!pg[0].u.free.scrub_state);
+                pg[0].u.free.scrub_state = PAGE_SCRUBBING;
+
+                spin_unlock(&heap_lock);
 
                 scrub_order = (order > SCRUB_CHUNK_ORDER) ? SCRUB_CHUNK_ORDER : order;
                 num_scrubbed = 0;
-                while ( num_scrubbed < (1 << order) )
+                while ( scrub_continue(&pg[0]) && num_scrubbed < (1 << order) )
                 {
                     for ( i = 0; i < (1 << scrub_order); i++ )
+                    {
                         scrub_one_page(&pg[num_scrubbed + i]);
+                        if ( !scrub_continue(&pg[0]) )
+                        {
+                            /* Someone wants this chunk. Drop everything. */
+                            pg[0].u.free.scrub_state = 0;
+                            goto out_nolock;
+                        }
+                    }
 
                     num_scrubbed += (1 << scrub_order);
                     if ( softirq_pending(cpu) )
@@ -1080,7 +1121,15 @@ bool_t scrub_free_pages(unsigned int node)
                         break;
                     }
                  }
- 
+
+		if ( !spin_lock_cond(&heap_lock, scrub_continue, &pg[0]) )
+                {
+                    pg[0].u.free.scrub_state = 0;
+                    goto out_nolock;
+                }
+
+                page_list_del(pg, &heap(node, zone, order));
+
                 start = 0;
                 end = num_scrubbed;
 
@@ -1118,7 +1167,9 @@ bool_t scrub_free_pages(unsigned int node)
                     merge_chunks(ppg, node, zone, chunk_order, 1);
                     end += (1 << chunk_order);
                 }
- 
+
+                pg[0].u.free.scrub_state = 0;
+
                 if ( preempt )
                     goto out;
             }
@@ -1127,6 +1178,8 @@ bool_t scrub_free_pages(unsigned int node)
 
  out:
     spin_unlock(&heap_lock);
+
+ out_nolock:
     clear_bit(node, &node_scrubbing);
     return (node_need_scrub[node] != 0);
 }
@@ -1165,6 +1218,8 @@ static void free_heap_pages(
         if ( page_state_is(&pg[i], offlined) )
             tainted = 1;
 
+        pg[i].u.free.scrub_state=0;
+
         /* If a page has no owner it will need no safety TLB flush. */
         pg[i].u.free.need_tlbflush = (page_get_owner(&pg[i]) != NULL);
         if ( pg[i].u.free.need_tlbflush )
diff --git a/xen/include/asm-x86/mm.h b/xen/include/asm-x86/mm.h
index b11124f..dd84835 100644
--- a/xen/include/asm-x86/mm.h
+++ b/xen/include/asm-x86/mm.h
@@ -87,6 +87,10 @@ struct page_info
 
         /* Page is on a free list: ((count_info & PGC_count_mask) == 0). */
         struct {
+#define PAGE_SCRUBBING      (1<<1)
+#define PAGE_SCRUB_ABORT    (1<<2)
+            unsigned char scrub_state;
+
             /* Do TLBs need flushing for safety before next page use? */
             bool_t need_tlbflush;
         } free;
-- 
1.7.1


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel