On Thu, 10 Feb 2005, Andrea Arcangeli wrote: > On Thu, Feb 10, 2005 at 08:19:47PM +0000, Hugh Dickins wrote: > > > > get_user_pages broke COW in advance of the I/O. The reason for > > subsequent COW if the page gets unmapped is precisely because > > can_share_swap_page used page_count to judge exclusivity, and > > get_user_pages has bumped that up, so the page appears (in danger > > of being) non-exclusive when actually it is exclusive. By changing > > can_share_swap_page to use page_mapcount, that frustration vanishes. > > What if a second thread does a fork() while the rawio is in progress? > The page will be again no shareable, and if you unmap it the rawio will > be currupt if the thread that executed the fork while the I/O is in > progress writes to a part of the page again after it has been unmapped > (obviously the part of the page that isn't under read-rawio, rawio works > with the hardblocksize, not on the whole page).
I think there's no difference here between the two techniques. With the new can_share_swap_page, yes, "page_swapcount" goes up with the fork, the page will be judged non-exclusive, a user write to another part of the page will replace it by a copy page, and it's undefined how much of the I/O will be captured in the copy. But even with the old can_share_swap_page, and try_to_unmap_one refusing to unmap the page, copy_page_range's COW checking would still remove write permission from the pte of an anonymous page, so a user write to another part of the page would find raised page_count and replace it by a copy page: same behaviour. And it's fine for the behaviour to be somewhat undefined in this peculiar case: the important thing is just that the page must not be freed and reused while I/O occurs, hence get_user_page raising the page_count - which I'm _not_ proposing to change! Hugh - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/