2010-02-18 22:45:07

by H. Peter Anvin

[permalink] [raw]
Subject: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

From: H. Peter Anvin <[email protected]>

x86-32 has had a static test for copy_on_user() overflow for a while.
This test currently fails in mm/migrate.c resulting in an
allyesconfig/allmodconfig build failure on x86-32:

In function ‘copy_from_user’,
inlined from ‘do_pages_stat’ at
/home/hpa/kernel/git/mm/migrate.c:1012:
/home/hpa/kernel/git/arch/x86/include/asm/uaccess_32.h:212: error:
call to ‘copy_from_user_overflow’ declared

Make the logic more explicit and therefore easier for gcc to
understand.

Reported-by: Len Brown <[email protected]>
Signed-off-by: H. Peter Anvin <[email protected]>
Cc: Arjan van de Ven <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: KOSAKI Motohiro <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Thomas Gleixner <[email protected]>
---

NOTE: Sending this patch directly rather than queueing it up since it is a
current build failure.

mm/migrate.c | 6 +++---
1 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/mm/migrate.c b/mm/migrate.c
index 9a0db5b..1a12f15 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1002,12 +1002,12 @@ static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
#define DO_PAGES_STAT_CHUNK_NR 16
const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
int chunk_status[DO_PAGES_STAT_CHUNK_NR];
- unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
+ unsigned long i, chunk_nr;
int err;

for (i = 0; i < nr_pages; i += chunk_nr) {
- if (chunk_nr > nr_pages - i)
- chunk_nr = nr_pages - i;
+ chunk_nr = min(nr_pages - i,
+ (unsigned long)DO_PAGES_STAT_CHUNK_NR);

err = copy_from_user(chunk_pages, &pages[i],
chunk_nr * sizeof(*chunk_pages));
--
1.6.5.2


2010-02-18 23:04:47

by Linus Torvalds

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable



On Thu, 18 Feb 2010, H. Peter Anvin wrote:
>
> Make the logic more explicit and therefore easier for gcc to
> understand.

Hmm. When making simplifications like this, I would really suggest you
also move the declaration of the variable itself into the block where it
is now used, rather than leaving it be function-wide.

Yes, it's used in the final condition of the for-loop, but that whole loop
is just screwy. The 'err' handling is insane. Sometimes 'err' is a return
value form copy_to/from_user, and sometimes it's a errno. The two are
_not_ the same thing, they don't even have the same type!

And 'i' is totally useless too.

So that whole loop should be rewritten.

I don't even have page migration enabled, so I haven't even compile-tested
this, but wouldn't something like this work? It's smaller, gets rid of two
pointless variables, and looks simpler to me. Hmm?

Linus

---
mm/migrate.c | 36 ++++++++++++++----------------------
1 files changed, 14 insertions(+), 22 deletions(-)

diff --git a/mm/migrate.c b/mm/migrate.c
index 9a0db5b..933d5b1 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -999,36 +999,28 @@ static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
const void __user * __user *pages,
int __user *status)
{
-#define DO_PAGES_STAT_CHUNK_NR 16
+#define DO_PAGES_STAT_CHUNK_NR 16ul
const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
int chunk_status[DO_PAGES_STAT_CHUNK_NR];
- unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
- int err;

- for (i = 0; i < nr_pages; i += chunk_nr) {
- if (chunk_nr > nr_pages - i)
- chunk_nr = nr_pages - i;
+ while (nr_pages) {
+ unsigned long chunk_nr;

- err = copy_from_user(chunk_pages, &pages[i],
- chunk_nr * sizeof(*chunk_pages));
- if (err) {
- err = -EFAULT;
- goto out;
- }
+ chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
+
+ if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*pages)))
+ break;

do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);

- err = copy_to_user(&status[i], chunk_status,
- chunk_nr * sizeof(*chunk_status));
- if (err) {
- err = -EFAULT;
- goto out;
- }
- }
- err = 0;
+ if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
+ break;

-out:
- return err;
+ pages += chunk_nr;
+ status += chunk_nr;
+ nr_pages -= chunk_nr;
+ }
+ return nr_pages ? -EFAULT : 0;
}

/*

2010-02-19 00:17:37

by H. Peter Anvin

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

On 02/18/2010 03:02 PM, Linus Torvalds wrote:
>
> Hmm. When making simplifications like this, I would really suggest you
> also move the declaration of the variable itself into the block where it
> is now used, rather than leaving it be function-wide.
>
> Yes, it's used in the final condition of the for-loop, but that whole loop
> is just screwy. The 'err' handling is insane. Sometimes 'err' is a return
> value form copy_to/from_user, and sometimes it's a errno. The two are
> _not_ the same thing, they don't even have the same type!
>
> And 'i' is totally useless too.
>
> So that whole loop should be rewritten.
>

OK, I was trying to make the minimal set of changes given the late -rc
status.

> I don't even have page migration enabled, so I haven't even compile-tested
> this, but wouldn't something like this work? It's smaller, gets rid of two
> pointless variables, and looks simpler to me. Hmm?

The code definitely looks cleaner, and it's a much more standard
"chunked data loop" form. Weirdly enough, though, gcc 4.4.2 can't
figure out the copy_from_user() that way... despite having the same
min() structure as my code.

However, if I change it to:

chunk_nr = nr_pages;
if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
chunk_nr = DO_PAGES_STAT_CHUNK_NR;

... then it works!

Overall, it looks like gcc is rather fragile with regards to its ability
to constant-propagate. It's probably no coincidence that chunked loops
is the place where we really have problems with this kind of stuff.

Updated patch, which compile-tests for me, attached.

-hpa


Attachments:
0001-mm-Make-copy_from_user-in-migrate.c-statically-predi.patch (2.75 kB)

2010-02-19 01:25:20

by KOSAKI Motohiro

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

> On 02/18/2010 03:02 PM, Linus Torvalds wrote:
> >
> > Hmm. When making simplifications like this, I would really suggest you
> > also move the declaration of the variable itself into the block where it
> > is now used, rather than leaving it be function-wide.
> >
> > Yes, it's used in the final condition of the for-loop, but that whole loop
> > is just screwy. The 'err' handling is insane. Sometimes 'err' is a return
> > value form copy_to/from_user, and sometimes it's a errno. The two are
> > _not_ the same thing, they don't even have the same type!
> >
> > And 'i' is totally useless too.
> >
> > So that whole loop should be rewritten.
> >
>
> OK, I was trying to make the minimal set of changes given the late -rc
> status.
>
> > I don't even have page migration enabled, so I haven't even compile-tested
> > this, but wouldn't something like this work? It's smaller, gets rid of two
> > pointless variables, and looks simpler to me. Hmm?
>
> The code definitely looks cleaner, and it's a much more standard
> "chunked data loop" form. Weirdly enough, though, gcc 4.4.2 can't
> figure out the copy_from_user() that way... despite having the same
> min() structure as my code.
>
> However, if I change it to:
>
> chunk_nr = nr_pages;
> if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
> chunk_nr = DO_PAGES_STAT_CHUNK_NR;
>
> ... then it works!
>
> Overall, it looks like gcc is rather fragile with regards to its ability
> to constant-propagate. It's probably no coincidence that chunked loops
> is the place where we really have problems with this kind of stuff.
>
> Updated patch, which compile-tests for me, attached.

hehe, I'm ESPer. I think you hope I do runtime-test, plz wait 12 hour :-)


2010-02-19 01:27:08

by H. Peter Anvin

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

On 02/18/2010 05:25 PM, KOSAKI Motohiro wrote:
>>
>> Updated patch, which compile-tests for me, attached.
>
> hehe, I'm ESPer. I think you hope I do runtime-test, plz wait 12 hour :-)
>

That would be most appreciated ;)

-hpa

2010-02-19 15:01:33

by Christoph Lameter

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

On Thu, 18 Feb 2010, H. Peter Anvin wrote:

> Updated patch, which compile-tests for me, attached.

Looks ok.

Acked-by: Christoph Lameter <[email protected]>


Attachments:
0001-mm-Make-copy_from_user-in-migrate.c-statically-predi.patch (4.84 kB)

2010-02-21 11:46:43

by KOSAKI Motohiro

[permalink] [raw]
Subject: Re: [PATCH] mm: Make copy_from_user() in migrate.c statically predictable

> > On 02/18/2010 03:02 PM, Linus Torvalds wrote:
> > >
> > > Hmm. When making simplifications like this, I would really suggest you
> > > also move the declaration of the variable itself into the block where it
> > > is now used, rather than leaving it be function-wide.
> > >
> > > Yes, it's used in the final condition of the for-loop, but that whole loop
> > > is just screwy. The 'err' handling is insane. Sometimes 'err' is a return
> > > value form copy_to/from_user, and sometimes it's a errno. The two are
> > > _not_ the same thing, they don't even have the same type!
> > >
> > > And 'i' is totally useless too.
> > >
> > > So that whole loop should be rewritten.
> > >
> >
> > OK, I was trying to make the minimal set of changes given the late -rc
> > status.
> >
> > > I don't even have page migration enabled, so I haven't even compile-tested
> > > this, but wouldn't something like this work? It's smaller, gets rid of two
> > > pointless variables, and looks simpler to me. Hmm?
> >
> > The code definitely looks cleaner, and it's a much more standard
> > "chunked data loop" form. Weirdly enough, though, gcc 4.4.2 can't
> > figure out the copy_from_user() that way... despite having the same
> > min() structure as my code.
> >
> > However, if I change it to:
> >
> > chunk_nr = nr_pages;
> > if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
> > chunk_nr = DO_PAGES_STAT_CHUNK_NR;
> >
> > ... then it works!
> >
> > Overall, it looks like gcc is rather fragile with regards to its ability
> > to constant-propagate. It's probably no coincidence that chunked loops
> > is the place where we really have problems with this kind of stuff.
> >
> > Updated patch, which compile-tests for me, attached.
>
> hehe, I'm ESPer. I think you hope I do runtime-test, plz wait 12 hour :-)

sorry for the delay. I had little machine trouble.
Reviewed-and-Tested-by: KOSAKI Motohiro <[email protected]>