>> >> @@ -547,7 +549,13 @@ int reuse_swap_page(struct page *page)
>> >> SetPageDirty(page);
>> >> }
>> >> }
>> >> - return count == 1;
>> >> +
>> >> + /*
>> >> + * If we can re-use the swap page _and_ the end
>> >> + * result has only one user (the mapping), then
>> >> + * we reuse the whole page
>> >> + */
>> >> + return count + page_count(page) == 2;
>> >> }
>> >
>> > I guess this patch does work to close the read-side race, but I slightly don't
>> > like using page_count for things like this. page_count can be temporarily
>> > raised for reasons other than access through their user mapping. Swapcache,
>> > page reclaim, LRU pagevecs, concurrent do_wp_page, etc.
>>
>> Yes, that's trade-off.
>> your early decow also can misjudge and make unnecessary copy.
>
> Yes indeed it can. Although it would only ever do so in case of pages
> that have had get_user_pages run against them previously, and not from
> random interactions from any other parts of the kernel.
Agreed.
> I would be interested, using an anon vma field as you say for keeping
> a gup count... it could potentially be used to avoid the extra copy.
> But hmm, I don't have much time to go down that path so long as the
> basic concept of my proposal is in question.
ok, I try to make it. thanks.
> + if (PageSwapCache(page) &&
> + page_count(page) != page_mapcount(page) + 2) {
> + ret = SWAP_FAIL;
> + goto out_unmap;
> + }
>
> Now if another thread does a get_user_pages_fast after it passes this
> check, it can take a gup reference to the page which is now about to
> be unmapped. Then after it is unmapped, if a wp fault is caused on the
> page, then it will not be reused and thus you lose data as explained
> in your big comment.
Grrr, I lose. I've misunderstood get_user_pages_fast() also grab pte_lock.
I must think it again.
I guess you dislike get_user_page_fast() grab pte_lock too, right?
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
