[PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[H.J. Lu]

On Linux/x86-64, for elf/tst-align3, we now get

munmap(0x7f88f9401000, 1126424)         = 0

instead of

munmap(0x7f1615200018, 544768)          = -1 EINVAL (Invalid argument)
---
 elf/dl-map-segments.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
index 70a4c40695..54e606aa87 100644
--- a/elf/dl-map-segments.h
+++ b/elf/dl-map-segments.h
@@ -55,6 +55,7 @@ _dl_map_segment (const struct loadcmd *c, ElfW(Addr) mappref,
       if (delta)
 	__munmap ((void *) map_start, delta);
       ElfW(Addr) map_end = map_start_aligned + maplength;
+      map_end = ALIGN_UP (map_end, GLRO(dl_pagesize));
       delta = map_start + maplen - map_end;
       if (delta)
 	__munmap ((void *) map_end, delta);
-- 
2.33.1

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[Rongwei Wang]

Hi, HJ

Thanks for your help to fix this bug. And I just found out that the 
patch "elf: Properly align PT_LOAD segments" had been merged into 
glibc/master.

And I saw Florian's email:

 > We should check munmap failure though and rollback everything if
 > necessary.  It's possible we can undo the initial PROT_NONE mapping even
 > if future munmap calls fail because unmapping the first mapping does not
 > need to split a mapping.

It seems rollback initial PROT_NONE mapping when munmap fails is good 
idea. Nice to have. Of course, IMO, this your patch is also can fix this 
munmap failure well. what do you think?

When this fix is stable in glibc upstream, I will backport it into our 
internal glibc repo. So, if you have any improve this fix suggestions, I 
will respond immediately. Sorry for my speed is much slower than your.

Thanks.

On 12/13/21 11:20 PM, H.J. Lu via Libc-alpha wrote:
> On Linux/x86-64, for elf/tst-align3, we now get
> 
> munmap(0x7f88f9401000, 1126424)         = 0
> 
> instead of
> 
> munmap(0x7f1615200018, 544768)          = -1 EINVAL (Invalid argument)
> ---
>   elf/dl-map-segments.h | 1 +
>   1 file changed, 1 insertion(+)
> 
> diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
> index 70a4c40695..54e606aa87 100644
> --- a/elf/dl-map-segments.h
> +++ b/elf/dl-map-segments.h
> @@ -55,6 +55,7 @@ _dl_map_segment (const struct loadcmd *c, ElfW(Addr) mappref,
>         if (delta)
>   	__munmap ((void *) map_start, delta);
>         ElfW(Addr) map_end = map_start_aligned + maplength;
> +      map_end = ALIGN_UP (map_end, GLRO(dl_pagesize));
>         delta = map_start + maplen - map_end;
>         if (delta)
>   	__munmap ((void *) map_end, delta);
> 

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[H.J. Lu]

On Mon, Dec 13, 2021 at 6:49 PM Rongwei Wang
<rongwei.wang@linux.alibaba.com> wrote:
>
> Hi, HJ
>
> Thanks for your help to fix this bug. And I just found out that the
> patch "elf: Properly align PT_LOAD segments" had been merged into
> glibc/master.
>
> And I saw Florian's email:
>
>  > We should check munmap failure though and rollback everything if
>  > necessary.  It's possible we can undo the initial PROT_NONE mapping even
>  > if future munmap calls fail because unmapping the first mapping does not
>  > need to split a mapping.
>
> It seems rollback initial PROT_NONE mapping when munmap fails is good
> idea. Nice to have. Of course, IMO, this your patch is also can fix this
> munmap failure well. what do you think?
>
> When this fix is stable in glibc upstream, I will backport it into our
> internal glibc repo. So, if you have any improve this fix suggestions, I
> will respond immediately. Sorry for my speed is much slower than your.
>
>

Let's wait for a few days.

-- 
H.J.

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[Florian Weimer]

* H. J. Lu via Libc-alpha:

> On Mon, Dec 13, 2021 at 6:49 PM Rongwei Wang
> <rongwei.wang@linux.alibaba.com> wrote:
>>
>> Hi, HJ
>>
>> Thanks for your help to fix this bug. And I just found out that the
>> patch "elf: Properly align PT_LOAD segments" had been merged into
>> glibc/master.
>>
>> And I saw Florian's email:
>>
>>  > We should check munmap failure though and rollback everything if
>>  > necessary.  It's possible we can undo the initial PROT_NONE mapping even
>>  > if future munmap calls fail because unmapping the first mapping does not
>>  > need to split a mapping.
>>
>> It seems rollback initial PROT_NONE mapping when munmap fails is good
>> idea. Nice to have. Of course, IMO, this your patch is also can fix this
>> munmap failure well. what do you think?
>>
>> When this fix is stable in glibc upstream, I will backport it into our
>> internal glibc repo. So, if you have any improve this fix suggestions, I
>> will respond immediately. Sorry for my speed is much slower than your.
>>
>>
>
> Let's wait for a few days.

Can we please fix the regression in some way?  It impacts development,
particularly at this late stage during the 2.34 cycle.

Either revert the change, or at least add the unchecked __munmap, and
fix that later.

Thanks,
Florian

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[H.J. Lu]

On Mon, Dec 13, 2021 at 11:40 PM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> > On Mon, Dec 13, 2021 at 6:49 PM Rongwei Wang
> > <rongwei.wang@linux.alibaba.com> wrote:
> >>
> >> Hi, HJ
> >>
> >> Thanks for your help to fix this bug. And I just found out that the
> >> patch "elf: Properly align PT_LOAD segments" had been merged into
> >> glibc/master.
> >>
> >> And I saw Florian's email:
> >>
> >>  > We should check munmap failure though and rollback everything if
> >>  > necessary.  It's possible we can undo the initial PROT_NONE mapping even
> >>  > if future munmap calls fail because unmapping the first mapping does not
> >>  > need to split a mapping.
> >>
> >> It seems rollback initial PROT_NONE mapping when munmap fails is good
> >> idea. Nice to have. Of course, IMO, this your patch is also can fix this
> >> munmap failure well. what do you think?
> >>
> >> When this fix is stable in glibc upstream, I will backport it into our
> >> internal glibc repo. So, if you have any improve this fix suggestions, I
> >> will respond immediately. Sorry for my speed is much slower than your.
> >>
> >>
> >
> > Let's wait for a few days.
>
> Can we please fix the regression in some way?  It impacts development,
> particularly at this late stage during the 2.34 cycle.
>
> Either revert the change, or at least add the unchecked __munmap, and
> fix that later.

What is the unchecked __munmap?


-- 
H.J.

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[Florian Weimer]

* H. J. Lu:

> On Mon, Dec 13, 2021 at 11:40 PM Florian Weimer <fweimer@redhat.com> wrote:
>>
>> * H. J. Lu via Libc-alpha:
>>
>> > On Mon, Dec 13, 2021 at 6:49 PM Rongwei Wang
>> > <rongwei.wang@linux.alibaba.com> wrote:
>> >>
>> >> Hi, HJ
>> >>
>> >> Thanks for your help to fix this bug. And I just found out that the
>> >> patch "elf: Properly align PT_LOAD segments" had been merged into
>> >> glibc/master.
>> >>
>> >> And I saw Florian's email:
>> >>
>> >>  > We should check munmap failure though and rollback everything if
>> >>  > necessary.  It's possible we can undo the initial PROT_NONE mapping even
>> >>  > if future munmap calls fail because unmapping the first mapping does not
>> >>  > need to split a mapping.
>> >>
>> >> It seems rollback initial PROT_NONE mapping when munmap fails is good
>> >> idea. Nice to have. Of course, IMO, this your patch is also can fix this
>> >> munmap failure well. what do you think?
>> >>
>> >> When this fix is stable in glibc upstream, I will backport it into our
>> >> internal glibc repo. So, if you have any improve this fix suggestions, I
>> >> will respond immediately. Sorry for my speed is much slower than your.
>> >>
>> >>
>> >
>> > Let's wait for a few days.
>>
>> Can we please fix the regression in some way?  It impacts development,
>> particularly at this late stage during the 2.34 cycle.
>>
>> Either revert the change, or at least add the unchecked __munmap, and
>> fix that later.
>
> What is the unchecked __munmap?

The __munmap at the end:

diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
index 70a4c40695..54e606aa87 100644
--- a/elf/dl-map-segments.h
+++ b/elf/dl-map-segments.h
@@ -55,6 +55,7 @@ _dl_map_segment (const struct loadcmd *c, ElfW(Addr) mappref,
       if (delta)
 	__munmap ((void *) map_start, delta);
       ElfW(Addr) map_end = map_start_aligned + maplength;
+      map_end = ALIGN_UP (map_end, GLRO(dl_pagesize));
       delta = map_start + maplen - map_end;
       if (delta)
 	__munmap ((void *) map_end, delta);

Thanks,
Florian

Re: [PATCH] elf: Align argument of __munmap to page size [BZ #28676]

[H.J. Lu]

On Tue, Dec 14, 2021 at 6:24 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu:
>
> > On Mon, Dec 13, 2021 at 11:40 PM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * H. J. Lu via Libc-alpha:
> >>
> >> > On Mon, Dec 13, 2021 at 6:49 PM Rongwei Wang
> >> > <rongwei.wang@linux.alibaba.com> wrote:
> >> >>
> >> >> Hi, HJ
> >> >>
> >> >> Thanks for your help to fix this bug. And I just found out that the
> >> >> patch "elf: Properly align PT_LOAD segments" had been merged into
> >> >> glibc/master.
> >> >>
> >> >> And I saw Florian's email:
> >> >>
> >> >>  > We should check munmap failure though and rollback everything if
> >> >>  > necessary.  It's possible we can undo the initial PROT_NONE mapping even
> >> >>  > if future munmap calls fail because unmapping the first mapping does not
> >> >>  > need to split a mapping.
> >> >>
> >> >> It seems rollback initial PROT_NONE mapping when munmap fails is good
> >> >> idea. Nice to have. Of course, IMO, this your patch is also can fix this
> >> >> munmap failure well. what do you think?
> >> >>
> >> >> When this fix is stable in glibc upstream, I will backport it into our
> >> >> internal glibc repo. So, if you have any improve this fix suggestions, I
> >> >> will respond immediately. Sorry for my speed is much slower than your.
> >> >>
> >> >>
> >> >
> >> > Let's wait for a few days.
> >>
> >> Can we please fix the regression in some way?  It impacts development,
> >> particularly at this late stage during the 2.34 cycle.
> >>
> >> Either revert the change, or at least add the unchecked __munmap, and
> >> fix that later.
> >
> > What is the unchecked __munmap?
>
> The __munmap at the end:
>
> diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
> index 70a4c40695..54e606aa87 100644
> --- a/elf/dl-map-segments.h
> +++ b/elf/dl-map-segments.h
> @@ -55,6 +55,7 @@ _dl_map_segment (const struct loadcmd *c, ElfW(Addr) mappref,
>        if (delta)
>         __munmap ((void *) map_start, delta);
>        ElfW(Addr) map_end = map_start_aligned + maplength;
> +      map_end = ALIGN_UP (map_end, GLRO(dl_pagesize));
>        delta = map_start + maplen - map_end;
>        if (delta)
>         __munmap ((void *) map_end, delta);
>

This is

https://sourceware.org/pipermail/libc-alpha/2021-December/134097.html

I am checking it in now.

Thanks.

-- 
H.J.