Thread (96 messages) 96 messages, 6 authors, 2021-06-08

Re: [PATCH v10 08/33] mm: Add folio_try_get_rcu

From: Christoph Hellwig <hch@infradead.org>
Date: 2021-05-27 08:17:10
Also in: linux-fsdevel, lkml

On Tue, May 11, 2021 at 10:47:10PM +0100, Matthew Wilcox (Oracle) wrote:
-static inline int page_ref_add_unless(struct page *page, int nr, int u)
+static inline bool page_ref_add_unless(struct page *page, int nr, int u)
 {
-	int ret = atomic_add_unless(&page->_refcount, nr, u);
+	bool ret = atomic_add_unless(&page->_refcount, nr, u);
 
 	if (page_ref_tracepoint_active(page_ref_mod_unless))
 		__page_ref_mod_unless(page, nr, ret);
 	return ret;
 }
Unrelated but neat cleanup.
 
-static inline int folio_ref_add_unless(struct folio *folio, int nr, int u)
+static inline bool folio_ref_add_unless(struct folio *folio, int nr, int u)
 {
 	return page_ref_add_unless(&folio->page, nr, u);
 }
This should probably go into the patch adding folio_ref_add_unless.
+static inline bool folio_ref_try_add_rcu(struct folio *folio, int count)
Should this have a __ prefix and/or a don't use direct comment?
+{
+#ifdef CONFIG_TINY_RCU
+	/*
+	 * The caller guarantees the folio will not be freed from interrupt
+	 * context, so (on !SMP) we only need preemption to be disabled
+	 * and TINY_RCU does that for us.
+	 */
+# ifdef CONFIG_PREEMPT_COUNT
+	VM_BUG_ON(!in_atomic() && !irqs_disabled());
+# endif
	VM_BUG_ON(IS_ENABLED(CONFIG_PREEMPT_COUNT) &&
		  !in_atomic() && !irqs_disabled());

?
+	VM_BUG_ON_FOLIO(folio_ref_count(folio) == 0, folio);
+	folio_ref_add(folio, count);
+#else
+	if (unlikely(!folio_ref_add_unless(folio, count, 0))) {
+		/* Either the folio has been freed, or will be freed. */
+		return false;
+	}
+#endif
+	return true;
but is this tiny rcu optimization really worth it?  I guess we're just
preserving it from the existing code and don't rock the boat..
quoted hunk ↗ jump to hunk
@@ -1746,6 +1746,26 @@ pgoff_t page_cache_prev_miss(struct address_space *mapping,
 }
 EXPORT_SYMBOL(page_cache_prev_miss);
 
+/*
+ * Lockless page cache protocol:
+ * On the lookup side:
+ * 1. Load the folio from i_pages
+ * 2. Increment the refcount if it's not zero
+ * 3. If the folio is not found by xas_reload(), put the refcount and retry
+ *
+ * On the removal side:
+ * A. Freeze the page (by zeroing the refcount if nobody else has a reference)
+ * B. Remove the page from i_pages
+ * C. Return the page to the page allocator
+ *
+ * This means that any page may have its reference count temporarily
+ * increased by a speculative page cache (or fast GUP) lookup as it can
+ * be allocated by another user before the RCU grace period expires.
+ * Because the refcount temporarily acquired here may end up being the
+ * last refcount on the page, any page allocation must be freeable by
+ * put_folio().
+ */
+
 /*
  * mapping_get_entry - Get a page cache entry.
  * @mapping: the address_space to search
Is this really a good place for the comment?  I'd expect it either near
a relevant function or at the top of a file.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help