On Sat, 25 Sep 2021, Hugh Dickins wrote:
> On Fri, 24 Sep 2021, Hugh Dickins wrote:
> > On Thu, 9 Sep 2021, Liu Yuntao wrote:
> >
> > > In the case of SHMEM_HUGE_WITHIN_SIZE, the page index is not rounded
> > > up correctly. When the page index points to the first page in a huge
> > > page, round_up() cannot bring it to the end of the huge page, but
> > > to the end of the previous one.
> > >
> > > an example:
> > > HPAGE_PMD_NR on my machine is 512(2 MB huge page size).
> > > After allcoating a 3000 KB buffer, I access it at location 2050 KB.
> >
> > Your example is certainly helpful, but weird! It's not impossible,
> > but wouldn't it be easier to understand if you said "2048 KB" there?
I wanted to emphasize that access to any bit in the first page will
trigger this problem, so I didn't use "2048 KB".
> >
> > > In shmem_is_huge(), the corresponding index happens to be 512.
> > > After rounded up by HPAGE_PMD_NR, it will still be 512 which is
> > > smaller than i_size, and shmem_is_huge() will return true.
> > > As a result, my buffer takes an additional huge page, and that
> > > shouldn't happen when shmem_enabled is set to within_size.
> >
> > A colleague very recently opened my eyes to within_size on shmem_enabled:
> > I've always been dubious of both, but they can work quite well together.
> >
> > >
> > > Fixes: f3f0e1d2150b2b ("khugepaged: add support of collapse for tmpfs/shmem pages")
> > > Signed-off-by: Liu Yuntao <liuyuntao10@xxxxxxxxxx>
> >
> > Thanks, with a nice simplification from Kirill.
> >
> > Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>
>
> Andrew has just sent this on to Linus - thanks - and that's fine:
> no need to get in the way of that.
>
> But since replying, I have remembered more history, and there is
> something that we need to be aware of.
>
> Whereas to you this is a straightforward off-by-one (or off-by-page)
> fix, it also results in a significant change in behaviour - I'd say
> usually for the better, but some might be surprised. This patch has
> Kirill's Ack and my Ack, and I hope and believe that we can get away
> with the change in behaviour: but let's be aware of it.
>
> The change that concerns me is when, for example, copying a large
> file into a huge=within_size tmpfs (or more generally, just writing
> to the file by appending at EOF in the usual way).
>
> With the old WITHIN_SIZE code, the first 2MB was allocated in small
> pages, then subsequent 2MB extents were allocated with huge pages;
> including the final extent, even if it only needed a single byte.
>
> I always thought that was very clunky behaviour, the small pages
> coming at the wrong end of the file; and that's why I was dubious
> about it as a sensible filesystem mount option. But I was under
> the impression that it was the intended behaviour.
>
> With your new WITHIN_SIZE code, all those appending allocations
> are outside i_size, and the whole file is allocated in small pages.
> (Then maybe later on khugepaged can assemble huge pages for it.)
>
> Your patch makes within_size more sensible than it was for pre-sized
> files (and I think it's fair to say that the majority of files in
> shmem's internal mount, subject to thp/shmem_enabled, are likely to
> be fixed-size files); and better-defined than it used to be on
> growing files, but they won't get the huge pages they used to.
Although my patch changes shmem's behaviour, it makes shmem consistent
with the documentation. I think with the new code, it will be easier
for our users to understand.
>
> Hugh
