On Wed, Mar 03, 2021 at 01:51:42PM -0800, Andrew Morton wrote: > On Wed, 3 Mar 2021 20:57:36 +0000 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > On Wed, Mar 03, 2021 at 12:12:53PM -0800, Andrew Morton wrote: > > > On Wed, 3 Mar 2021 13:26:40 +0000 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > > > > > But here's the thing ... invalidate_mapping_pages() doesn't > > > > ClearPageUptodate. The only places where we ClearPageUptodate is on an > > > > I/O error. > > > > > > yup. > > > > > > > So ... as far as I can tell, the only way to hit this is: > > > > > > > > - Get an I/O error during the wait > > > > - Have another thread cause the page to be removed from the page cache > > > > (eg do direct I/O to the file) before this thread is run. > > > > > > > > and the consequence to this change is that we have another attempt to > > > > read the page instead of returning an error immediately. I'm OK with > > > > that unintentional change, although I think the previous behaviour was > > > > also perfectly acceptable (after all, there was an I/O error while trying > > > > to read this page). > > > > > > > > Delving into the linux-fullhistory tree, this code was introduced by ... > > > > > > > > commit 56f0d5fe6851037214a041a5cb4fc66199256544 > > > > Author: Andrew Morton <akpm@xxxxxxxx> > > > > Date: Fri Jan 7 22:03:01 2005 -0800 > > > > > > > > [PATCH] readpage-vs-invalidate fix > > > > > > > > A while ago we merged a patch which tried to solve a problem wherein a > > > > concurrent read() and invalidate_inode_pages() would cause the read() to > > > > return -EIO because invalidate cleared PageUptodate() at the wrong time. > > > > > > > > We no longer clear PageUptodate, so I think this is stale code? Perhaps > > > > you could check with the original author ... > > > > > > Which code do you think might be stale? We need the !PageUptodate > > > check to catch IO errors and we need the !page->mapping check to catch > > > invalidates. Am a bit confused. > > > > I think the check of !page->mapping here: > > > > if (PageUptodate(page)) > > return 0; > > if (!page->mapping) /* page truncated */ > > return AOP_TRUNCATED_PAGE; > > > > is no longer needed. If we didn't see an error, the page will be Uptodate, > > regardless of whether it's been removed from the page cache. If we did > > see an error, it's OK to return -EIO, even if the page has been removed > > from the page cache in the interim. > > OK. > > Checking page->mapping of an unlocked page seems meaningless anyway - > what's to prevent it from being truncated just after we checked? Originally it prevented the opposite race from being misinterpreted; we didn't care if it was truncated after we saw it was Uptodate; we cared if it was !Uptodate because it had been truncated. Now, it's just pointless.
