Re: [patch 1/3] mm: fix PageUptodate memorder

[Nick Piggin <nickpiggin@yahoo.com.au>]

Andrew Morton wrote:
> On Tue,  6 Feb 2007 09:02:11 +0100 (CET) Nick Piggin <npiggin@suse.de> wrote:
> 
> 
>>+static inline void __SetPageUptodate(struct page *page)
>>+{
>>+#ifdef CONFIG_S390
>> 	if (!test_and_set_bit(PG_uptodate, &page->flags))
>> 		page_test_and_clear_dirty(page);
>>-}
>> #else
>>-#define SetPageUptodate(page)	set_bit(PG_uptodate, &(page)->flags)
>>+	/*
>>+	 * Memory barrier must be issued before setting the PG_uptodate bit,
>>+	 * so all previous writes that served to bring the page uptodate are
>>+	 * visible before PageUptodate becomes true.
>>+	 *
>>+	 * S390 is guaranteed to have a barrier in the test_and_set operation
>>+	 * (see Documentation/atomic_ops.txt).
>>+	 *
>>+	 * XXX: does this memory barrier need to be anything special to
>>+	 * handle things like DMA writes into the page?
>>+	 */
>>+	smp_wmb();
>>+	set_bit(PG_uptodate, &(page)->flags);
>> #endif
>>+}
>>+
>>+static inline void SetPageUptodate(struct page *page)
>>+{
>>+	WARN_ON(!PageLocked(page));
>>+	__SetPageUptodate(page);
>>+}
>>+
>>+static inline void SetNewPageUptodate(struct page *page)
>>+{
>>+	__SetPageUptodate(page);
>>+}
> 
> 
> I was panicing for a minute when I saw that __SetPageUptodate() in there.
> 
> Conventionally the __SetPageFoo namespace is for nonatomic updates to
> page->flags.  Can we call this something different?

Duh, of course, sorry.

> What a fugly patchset :(

Fugly problem. One could fix it by always locking the page, but I was
worried about Hugh flaming me if I tried that ;)

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com