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.
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.
Fixes: f3f0e1d2150b2b ("khugepaged: add support of collapse for tmpfs/shmem pages")
Signed-off-by: Liu Yuntao <[email protected]>
---
V1->V2:
add simplification of the condition after round_up()
---
mm/shmem.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/shmem.c b/mm/shmem.c
index 88742953532c..b5860f4a2738 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -490,9 +490,9 @@ bool shmem_is_huge(struct vm_area_struct *vma,
case SHMEM_HUGE_ALWAYS:
return true;
case SHMEM_HUGE_WITHIN_SIZE:
- index = round_up(index, HPAGE_PMD_NR);
+ index = round_up(index + 1, HPAGE_PMD_NR);
i_size = round_up(i_size_read(inode), PAGE_SIZE);
- if (i_size >= HPAGE_PMD_SIZE && (i_size >> PAGE_SHIFT) >= index)
+ if (i_size >> PAGE_SHIFT >= index)
return true;
fallthrough;
case SHMEM_HUGE_ADVISE:
--
2.23.0
On Thu, Sep 09, 2021 at 11:20:07AM +0800, 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.
> 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.
>
> Fixes: f3f0e1d2150b2b ("khugepaged: add support of collapse for tmpfs/shmem pages")
> Signed-off-by: Liu Yuntao <[email protected]>
Acked-by: Kirill A. Shutemov <[email protected]>
--
Kirill A. Shutemov
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?
> 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 <[email protected]>
Thanks, with a nice simplification from Kirill.
Acked-by: Hugh Dickins <[email protected]>
Ignore the comment I've added below - it's not worth worrying about.
> ---
> V1->V2:
> add simplification of the condition after round_up()
> ---
> mm/shmem.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 88742953532c..b5860f4a2738 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -490,9 +490,9 @@ bool shmem_is_huge(struct vm_area_struct *vma,
> case SHMEM_HUGE_ALWAYS:
> return true;
> case SHMEM_HUGE_WITHIN_SIZE:
> - index = round_up(index, HPAGE_PMD_NR);
> + index = round_up(index + 1, HPAGE_PMD_NR);
Even without your change, I notice now that there's a possibility of
index wrapping to 0 on 32-bit architecture here. But nothing goes
terribly wrong in that case: it is not worth worrying about here.
> i_size = round_up(i_size_read(inode), PAGE_SIZE);
> - if (i_size >= HPAGE_PMD_SIZE && (i_size >> PAGE_SHIFT) >= index)
> + if (i_size >> PAGE_SHIFT >= index)
> return true;
> fallthrough;
> case SHMEM_HUGE_ADVISE:
> --
> 2.23.0
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?
>
> > 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 <[email protected]>
>
> Thanks, with a nice simplification from Kirill.
>
> Acked-by: Hugh Dickins <[email protected]>
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.
Hugh
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 <[email protected]>
> >
> > Thanks, with a nice simplification from Kirill.
> >
> > Acked-by: Hugh Dickins <[email protected]>
>
> 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
On Sun, 26 Sep 2021, liuyuntao wrote:
> 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".
Okay, thanks, I see your point now. (And I have to admit that, in my
confusion, I had thought 2050 KB would be index 514 - of course not!)
> > > > In shmem_is_huge(), the corresponding index happens to be 512.
...
> > 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.
Yes, I do agree with you. But the change in behaviour when appending at
EOF is significant, and needed to be called out - I think none of quite
realized that effect at first.
Hugh