Re: [PATCH 2/2] mm: Make swap_readpage() for SWP_FS_OPS use ->direct_IO() not ->readpage()

From: Matthew Wilcox
Date: Thu Aug 12 2021 - 09:52:38 EST


On Thu, Aug 12, 2021 at 02:37:59PM +0100, David Howells wrote:
> David Howells <dhowells@xxxxxxxxxx> wrote:
>
> > Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > > After submitting the IO here ...
> > >
> > > > + if (ret != -EIOCBQUEUED)
> > > > + swapfile_read_complete(&ki->iocb, ret, 0);
> > >
> > > We only touch the 'ki' here ... if the caller didn't call read_complete
> > >
> > > > + swapfile_put_kiocb(ki);
> > >
> > > Except for here, which is only touched in order to put the refcount.
> > >
> > > So why can't swapfile_read_complete() do the work of freeing the ki?
> >
> > When I was doing something similar for cachefiles, I couldn't get it to work
> > like that. I'll have another look at that.
>
> Ah, yes. generic_file_direct_write() accesses in the kiocb *after* calling
> ->direct_IO(), so the kiocb *must not* go away until after
> generic_file_direct_write() has returned.

This is a read, not a write ... but we don't care about ki_pos being
updated, so that store can be conditioned on IOCB_SWAP being clear.
Or instead of storing directly to ki_pos, we take a pointer to ki_pos
and then redirect that pointer somewhere harmless.