LinuxLists.cc - [GIT PULL] x86/mm changes for v4.4

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Tue, Nov 3, 2015 at 3:16 AM, Ingo Molnar <[email protected]> wrote:
>
> The new CONFIG_DEBUG_WX=y warning is marked default-y if CONFIG_DEBUG_RODATA=y is
> already eanbled, as a special exception, as these bugs are hard to notice and this
> check already found several live bugs.

So this seems to be not very useful.

...
Freeing unused kernel memory: 1068K (ffffffff8bcc0000 -
ffffffff8bdcb000)
Write protecting the kernel read-only data: 12288k
Freeing unused kernel memory: 1944K (ffff88000b61a000 - ffff88000b800000)
Freeing unused kernel memory: 1372K (ffff88000baa9000 - ffff88000bc00000)
------------[ cut here ]------------
WARNING: CPU: 7 PID: 1 at arch/x86/mm/dump_pagetables.c:225
note_page+0x5dc/0x780()
x86/mm: Found insecure W+X mapping at address
ffff88000005f000/0xffff88000005f000
...
x86/mm: Checked W+X mappings: FAILED, 165660 W+X pages found.
...

because it doesn't seem to give you any idea where to look for the
issue. Those 165660 pages come out to 647MB. Odd number.

Is this *supposed* to come out clean?

Linus

2015-11-04 23:39:43

by Dave Jones

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Wed, Nov 04, 2015 at 11:26:12AM -0800, Linus Torvalds wrote:
> On Tue, Nov 3, 2015 at 3:16 AM, Ingo Molnar <[email protected]> wrote:
> >
> > The new CONFIG_DEBUG_WX=y warning is marked default-y if CONFIG_DEBUG_RODATA=y is
> > already eanbled, as a special exception, as these bugs are hard to notice and this
> > check already found several live bugs.
>
> So this seems to be not very useful.
>
> ...
> Freeing unused kernel memory: 1068K (ffffffff8bcc0000 -
> ffffffff8bdcb000)
> Write protecting the kernel read-only data: 12288k
> Freeing unused kernel memory: 1944K (ffff88000b61a000 - ffff88000b800000)
> Freeing unused kernel memory: 1372K (ffff88000baa9000 - ffff88000bc00000)
> ------------[ cut here ]------------
> WARNING: CPU: 7 PID: 1 at arch/x86/mm/dump_pagetables.c:225
> note_page+0x5dc/0x780()
> x86/mm: Found insecure W+X mapping at address
> ffff88000005f000/0xffff88000005f000
> ...
> x86/mm: Checked W+X mappings: FAILED, 165660 W+X pages found.
> ...
>
> because it doesn't seem to give you any idea where to look for the
> issue. Those 165660 pages come out to 647MB. Odd number.
>
> Is this *supposed* to come out clean?

FWIW I'm seeing this too.

[ 3.293503] x86/mm: Found insecure W+X mapping at address ffff880000800000/0xffff880000800000

Looking further up the dmesg I see that range is..

[ 0.468368] ---[ Low Kernel Mapping ]---
[ 0.468381] 0xffff880000000000-0xffff880000800000 8M RW GLB NX pte
[ 0.468391] 0xffff880000800000-0xffff880000900000 1M RW GLB x pte

Linus, does that match your trace too ? The 2nd low kernel mapping?

Dave

2015-11-05 01:32:00

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Wed, Nov 4, 2015 at 3:39 PM, Dave Jones <[email protected]> wrote:
>
> FWIW I'm seeing this too.
>
> [ 0.468368] ---[ Low Kernel Mapping ]---
> [ 0.468381] 0xffff880000000000-0xffff880000800000 8M RW GLB NX pte
> [ 0.468391] 0xffff880000800000-0xffff880000900000 1M RW GLB x pte
>
> Linus, does that match your trace too ? The 2nd low kernel mapping?

I don't have that later debug output at all. Presumably some config difference.

Linus

2015-11-05 02:17:32

by Dave Jones

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
> On Wed, Nov 4, 2015 at 3:39 PM, Dave Jones <[email protected]> wrote:
> >
> > FWIW I'm seeing this too.
> >
> > [ 0.468368] ---[ Low Kernel Mapping ]---
> > [ 0.468381] 0xffff880000000000-0xffff880000800000 8M RW GLB NX pte
> > [ 0.468391] 0xffff880000800000-0xffff880000900000 1M RW GLB x pte
> >
> > Linus, does that match your trace too ? The 2nd low kernel mapping?
>
> I don't have that later debug output at all. Presumably some config difference.

CONFIG_X86_PTDUMP_CORE iirc.

Dave

2015-11-05 21:27:22

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Wed, Nov 4, 2015 at 6:17 PM, Dave Jones <[email protected]> wrote:
> On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
> >
> > I don't have that later debug output at all. Presumably some config difference.
>
> CONFIG_X86_PTDUMP_CORE iirc.

No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.

Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
not default to 'y' unless it is made more useful if it actually
triggers. Ingo?

Linus

2015-11-05 21:33:13

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, Nov 5, 2015 at 1:27 PM, Linus Torvalds
<[email protected]> wrote:
>
> I suspect CONFIG_EFI_PGT_DUMP instead.
>
> Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
> not default to 'y' unless it is made more useful if it actually
> triggers. Ingo?

Actually, I guess I should have cc'd Steven Smalley too, since he's the author.

Steven, the problem is that without other debug options, the
CONFIG_DEBUG_WX option doesn't seem useful. It shows a problem, but it
doesn't actually show any information about where that issue comes
from or any other details to even *guess* at what is going on.

And if this turns out to be due to EFI wanting those permissions, what
should we do? People have talked about running the EFI callbacks in
their own private page table setup, which sounds like the right idea,
but until that actually *happens*....

Linus

2015-11-05 22:04:57

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, Nov 5, 2015 at 1:27 PM, Linus Torvalds
<[email protected]> wrote:
>
> No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.

Yes, that seems to show the tables, and agrees with the problem address.

So for me I have:

WARNING: CPU: 1 PID: 1 at arch/x86/mm/dump_pagetables.c:225
note_page+0x5dc/0x780()
x86/mm: Found insecure W+X mapping at address
ffff88000005f000/0xffff88000005f000

and that 005f000 address also shows up in the firmware tables:

ACPI: 6 ACPI AML tables successfully acquired and loaded
---[ User Space ]---
0x0000000000000000-0x000000000005f000 380K RW
GLB NX pte
0x000000000005f000-0x0000000000060000 4K RW
GLB x pte
0x0000000000060000-0x0000000000200000 1664K RW
GLB NX pte
...
---[ Low Kernel Mapping ]---
0xffff880000000000-0xffff88000005f000 380K RW
GLB NX pte
0xffff88000005f000-0xffff880000060000 4K RW
GLB x pte
0xffff880000060000-0xffff880000200000 1664K RW
GLB NX pte
...

and there's quite a few other pages there that are RW but not marked
NX. I suspect they come from the EFI runtime services because the
pattern seems to match what I see in that area, but there's at least a
PSE mapping at START_KERNEL_map too, etc.

Linus

2015-11-05 22:27:56

by Borislav Petkov

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, Nov 05, 2015 at 02:04:55PM -0800, Linus Torvalds wrote:
> and there's quite a few other pages there that are RW but not marked
> NX. I suspect they come from the EFI runtime services because the

Yeah, at least the EFI mappings would need a bit more fiddling until
they're NX:

https://lkml.kernel.org/r/[email protected]

Perhaps we don't even have to make them NX - just set the PGD entry
containing them before the EFI call and unset it after its done.
I.e., something like that.

> pattern seems to match what I see in that area, but there's at least a
> PSE mapping at START_KERNEL_map too, etc.

Maybe we should default n CONFIG_DEBUG_WX until stuff has been fixed...?

--
Regards/Gruss,
Boris.

ECO tip #101: Trim your mails when you reply.

2015-11-06 06:55:56

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

* Linus Torvalds <[email protected]> wrote:

> On Wed, Nov 4, 2015 at 6:17 PM, Dave Jones <[email protected]> wrote:
> > On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
> > >
> > > I don't have that later debug output at all. Presumably some config difference.
> >
> > CONFIG_X86_PTDUMP_CORE iirc.
>
> No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.
>
> Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
> not default to 'y' unless it is made more useful if it actually
> triggers. Ingo?

Yeah, agreed absolutely.

So this is a bit sad because RWX pages are a real problem in practice, especially
since the EFI addresses are well predictable, but generating a warning without
being able to fix it quickly is counterproductive as well, as it only annoys
people and makes them turn off the option. (Which we could do as well to begin
with, without the annoyance factor...)

So the plan would be:

1) Make it default-n.

2) We should try to further improve the messages to make it easier to determine
what's wrong. We _do_ try to output symbolic information in the warning, to
make it easier to find buggy mappings, but these are not standard kernel
mappings. So I think we need an e820 mappings based semi-symbolic printout of
bad addresses - maybe even correlate it with the MMIO resource tree.

3) We should fix the EFI permission problem without relying on the firmware: it
appears we could just mark everything R-X optimistically, and if a write fault
happens (it's pretty rare in fact, only triggers when we write to an EFI
variable and so), we can mark the faulting page RW- on the fly, because it
appears that writable EFI sections, while not enumerated very well in 'old'
firmware, are still supposed to be page granular. (Even 'new' firmware I
wouldn't automatically trust to get the enumeration right...)

If that 'supposed to be' turns out to be 'not true' (not unheard of in
firmware land), then plan B would be to mark pages that generate write faults
RWX as well, to not break functionality. (This 'mark it RWX' is not something
that exploits would have easy access to, and we could also generate a warning
[after the EFI call has finished] if it ever triggers.)

Admittedly this approach might not be without its own complications, but it
looks reasonably simple (I don't think we need per EFI call page tables,
etc.), and does not assume much about the firmware being able to enumerate its
permissions properly. Were we to merge EFI support today I'd have insisted on
trying such an approach from day 1 on.

Thanks,

Ingo

2015-11-06 07:05:56

by Andy Lutomirski

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, Nov 5, 2015 at 10:55 PM, Ingo Molnar <[email protected]> wrote:
>
> * Linus Torvalds <[email protected]> wrote:
>
>> On Wed, Nov 4, 2015 at 6:17 PM, Dave Jones <[email protected]> wrote:
>> > On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
>> > >
>> > > I don't have that later debug output at all. Presumably some config difference.
>> >
>> > CONFIG_X86_PTDUMP_CORE iirc.
>>
>> No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.
>>
>> Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
>> not default to 'y' unless it is made more useful if it actually
>> triggers. Ingo?
>
> Yeah, agreed absolutely.
>
> So this is a bit sad because RWX pages are a real problem in practice, especially
> since the EFI addresses are well predictable, but generating a warning without
> being able to fix it quickly is counterproductive as well, as it only annoys
> people and makes them turn off the option. (Which we could do as well to begin
> with, without the annoyance factor...)
>
> So the plan would be:
>
> 1) Make it default-n.
>
> 2) We should try to further improve the messages to make it easier to determine
> what's wrong. We _do_ try to output symbolic information in the warning, to
> make it easier to find buggy mappings, but these are not standard kernel
> mappings. So I think we need an e820 mappings based semi-symbolic printout of
> bad addresses - maybe even correlate it with the MMIO resource tree.
>
> 3) We should fix the EFI permission problem without relying on the firmware: it
> appears we could just mark everything R-X optimistically, and if a write fault
> happens (it's pretty rare in fact, only triggers when we write to an EFI
> variable and so), we can mark the faulting page RW- on the fly, because it
> appears that writable EFI sections, while not enumerated very well in 'old'
> firmware, are still supposed to be page granular. (Even 'new' firmware I
> wouldn't automatically trust to get the enumeration right...)

I think it was Borislav who pointed out that this idea, which might
have been mine, is a bit silly. Why not just skip mapping the EFI
stuff in the init_pgd entirely and only map it in the EFI pgd?

We'll have RWX stuff in the EFI pgd, but so what? If we're exposing
anything that runs with the EFI pgd loaded to untrusted input, I think
we've already lost.

Admittedly, we might need to use a certain amount of care to avoid
interesting conflicts with the vmap mechanism. We might need to vmap
all of the EFI stuff, and possibly even all the top-level entries that
contain EFI stuff (i.e. exactly one of them unless EFI ends up *huge*)
as a blank not-present region to avoid overlaps, but that's not a big
deal.

>
> If that 'supposed to be' turns out to be 'not true' (not unheard of in
> firmware land), then plan B would be to mark pages that generate write faults
> RWX as well, to not break functionality. (This 'mark it RWX' is not something
> that exploits would have easy access to, and we could also generate a warning
> [after the EFI call has finished] if it ever triggers.)
>
> Admittedly this approach might not be without its own complications, but it
> looks reasonably simple (I don't think we need per EFI call page tables,
> etc.), and does not assume much about the firmware being able to enumerate its
> permissions properly. Were we to merge EFI support today I'd have insisted on
> trying such an approach from day 1 on.

I think we have separate EFI page tables already for other reasons. I
could be wrong -- I've never really understood the EFI mapping layout
very well.

--Andy

2015-11-06 07:44:30

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

(resent with Matt's email address fixed.)

* Ingo Molnar <[email protected]> wrote:

>
> * Linus Torvalds <[email protected]> wrote:
>
> > On Wed, Nov 4, 2015 at 6:17 PM, Dave Jones <[email protected]> wrote:
> > > On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
> > > >
> > > > I don't have that later debug output at all. Presumably some config difference.
> > >
> > > CONFIG_X86_PTDUMP_CORE iirc.
> >
> > No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.
> >
> > Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
> > not default to 'y' unless it is made more useful if it actually
> > triggers. Ingo?
>
> Yeah, agreed absolutely.
>
> So this is a bit sad because RWX pages are a real problem in practice, especially
> since the EFI addresses are well predictable, but generating a warning without
> being able to fix it quickly is counterproductive as well, as it only annoys
> people and makes them turn off the option. (Which we could do as well to begin
> with, without the annoyance factor...)
>
> So the plan would be:
>
> 1) Make it default-n.
>
> 2) We should try to further improve the messages to make it easier to determine
> what's wrong. We _do_ try to output symbolic information in the warning, to
> make it easier to find buggy mappings, but these are not standard kernel
> mappings. So I think we need an e820 mappings based semi-symbolic printout of
> bad addresses - maybe even correlate it with the MMIO resource tree.
>
> 3) We should fix the EFI permission problem without relying on the firmware: it
> appears we could just mark everything R-X optimistically, and if a write fault
> happens (it's pretty rare in fact, only triggers when we write to an EFI
> variable and so), we can mark the faulting page RW- on the fly, because it
> appears that writable EFI sections, while not enumerated very well in 'old'
> firmware, are still supposed to be page granular. (Even 'new' firmware I
> wouldn't automatically trust to get the enumeration right...)
>
> If that 'supposed to be' turns out to be 'not true' (not unheard of in
> firmware land), then plan B would be to mark pages that generate write faults
> RWX as well, to not break functionality. (This 'mark it RWX' is not something
> that exploits would have easy access to, and we could also generate a warning
> [after the EFI call has finished] if it ever triggers.)
>
> Admittedly this approach might not be without its own complications, but it
> looks reasonably simple (I don't think we need per EFI call page tables,
> etc.), and does not assume much about the firmware being able to enumerate its
> permissions properly. Were we to merge EFI support today I'd have insisted on
> trying such an approach from day 1 on.
>
> Thanks,
>
> Ingo

2015-11-06 11:39:48

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, 05 Nov, at 01:33:10PM, Linus Torvalds wrote:
>
> And if this turns out to be due to EFI wanting those permissions, what
> should we do? People have talked about running the EFI callbacks in
> their own private page table setup, which sounds like the right idea,
> but until that actually *happens*....

We have separate page tables today, for a few reasons, but mainly it's
so that we can have an identity mapping of memory present in the
region usually used by user processes - broken firmware still uses
those identity mappings even after the kernel tells it they're
invalid.

Note that when I say "separate" I'm talking about trampoline_pgd[]
which is also used by the x86 suspend/resume code.

However, turns out that the issue with the current scheme is the fact
that trampoline_pgd[] actually shares a couple of PGD entries with
swapper_pg_dir as can be seen in setup_real_mode(),

trampoline_pgd = (u64 *)__va(real_mode_header->trampoline_pgd);
trampoline_pgd[0] = init_level4_pgt[pgd_index(__PAGE_OFFSET)].pgd;
trampoline_pgd[511] = init_level4_pgt[511].pgd;

So when we map the EFI regions in efi_map_regions() we're inserting
them into swapper_pg_dir also, which is why you're seeing the
warnings.

If I remember correctly the rationale for using trampoline_pgd[] was
that it already did what we wanted (provided the identity mapping) and
would save us the overhead of maintaining more page tables for no good
reason. Obviously this entire thread is a good reason.

I suggest we stop using trampoline_pgd[] (since it has a good reason
for sharing the kernel mapping PGD entries) and create our own so that
we can isolate EFI completely.

For the immediate problem of the warnings spewing forth on all UEFI
machines, at the very least the config options needs to be disabled by
default, if not the patch reverted.

2015-11-06 12:39:17

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Fri, 06 Nov, at 07:55:50AM, Ingo Molnar wrote:
>
> 3) We should fix the EFI permission problem without relying on the firmware: it
> appears we could just mark everything R-X optimistically, and if a write fault
> happens (it's pretty rare in fact, only triggers when we write to an EFI
> variable and so), we can mark the faulting page RW- on the fly, because it
> appears that writable EFI sections, while not enumerated very well in 'old'
> firmware, are still supposed to be page granular. (Even 'new' firmware I
> wouldn't automatically trust to get the enumeration right...)

Sorry, this isn't true. I misled you with one of my earlier posts on
this topic. Let me try and clear things up...

Writing to EFI regions has to do with every invocation of the EFI
runtime services - it's not limited to when you read/write/delete EFI
variables. In fact, EFI variables really have nothing to do with this
discussion, they're a completely opaque concept to the OS, we have no
idea how the firmware implements them. Everything is done via the EFI
boot/runtime services.

The firmware itself will attempt to write to EFI regions when we
invoke the EFI services because that's where the PE/COFF ".data" and
".bss" sections live along with the heap. There's even some relocation
fixups that occur as SetVirtualAddressMap() time so it'll write to
".text" too.

Now, the above PE/COFF sections are usually (always?) contained within
EFI regions of type EfiRuntimeServicesCode. We know this is true
because the firmware folks have told us so, and because stopping that
is the motivation behind the new EFI_PROPERTIES_TABLE feature in UEFI
V2.5.

The data sections within the region are also *not* guaranteed to be
page granular because work was required in Tianocore for emitting
sections with 4k alignment as part of the EFI_PROPERTIES_TABLE
support.

Ultimately, what this means is that if you were to attempt to
dynamically fixup those regions that required write permission, you'd
have to modify the mappings for the majority of the EFI regions
anyway. And if you're blindly allowing write permission as a fixup,
there's not much security to be had.

> If that 'supposed to be' turns out to be 'not true' (not unheard of in
> firmware land), then plan B would be to mark pages that generate write faults
> RWX as well, to not break functionality. (This 'mark it RWX' is not something
> that exploits would have easy access to, and we could also generate a warning
> [after the EFI call has finished] if it ever triggers.)
>
> Admittedly this approach might not be without its own complications, but it
> looks reasonably simple (I don't think we need per EFI call page tables,
> etc.), and does not assume much about the firmware being able to enumerate its
> permissions properly. Were we to merge EFI support today I'd have insisted on
> trying such an approach from day 1 on.

We already have separate EFI page tables, though with the caveat that
we share some of swapper_pg_dir's PGD entries. The best solution would
be to stop sharing entries and isolate the EFI mappings from every
other page table structure, so that they're only used during the EFI
service calls.

2015-11-06 13:09:53

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Thu, 05 Nov, at 11:05:35PM, Andy Lutomirski wrote:
>
> Admittedly, we might need to use a certain amount of care to avoid
> interesting conflicts with the vmap mechanism. We might need to vmap
> all of the EFI stuff, and possibly even all the top-level entries that
> contain EFI stuff (i.e. exactly one of them unless EFI ends up *huge*)
> as a blank not-present region to avoid overlaps, but that's not a big
> deal.

There shouldn't be any room for conflicting with vmap() because the VA
region where we map EFI regions is still carved out especially for us.

Right Boris?

2015-11-06 13:24:57

by Borislav Petkov

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On Fri, Nov 06, 2015 at 01:09:48PM +0000, Matt Fleming wrote:
> On Thu, 05 Nov, at 11:05:35PM, Andy Lutomirski wrote:
> >
> > Admittedly, we might need to use a certain amount of care to avoid
> > interesting conflicts with the vmap mechanism. We might need to vmap
> > all of the EFI stuff, and possibly even all the top-level entries that
> > contain EFI stuff (i.e. exactly one of them unless EFI ends up *huge*)
> > as a blank not-present region to avoid overlaps, but that's not a big
> > deal.
>
> There shouldn't be any room for conflicting with vmap() because the VA
> region where we map EFI regions is still carved out especially for us.
>
> Right Boris?

Yap:

ffffea0000000000 - ffffeaffffffffff (=40 bits) virtual memory map (1TB)

vs

ffffffef00000000 - ffffffff00000000 EFI region in trampoline_pgd

the new pagetable will make that issue moot too.

--
Regards/Gruss,
Boris.

ECO tip #101: Trim your mails when you reply.

2015-11-07 07:03:46

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

* Andy Lutomirski <[email protected]> wrote:

> On Thu, Nov 5, 2015 at 10:55 PM, Ingo Molnar <[email protected]> wrote:
> >
> > * Linus Torvalds <[email protected]> wrote:
> >
> >> On Wed, Nov 4, 2015 at 6:17 PM, Dave Jones <[email protected]> wrote:
> >> > On Wed, Nov 04, 2015 at 05:31:59PM -0800, Linus Torvalds wrote:
> >> > >
> >> > > I don't have that later debug output at all. Presumably some config difference.
> >> >
> >> > CONFIG_X86_PTDUMP_CORE iirc.
> >>
> >> No, I have that. I suspect CONFIG_EFI_PGT_DUMP instead.
> >>
> >> Anyway, as it stands now, I think the CONFIG_DEBUG_WX option should
> >> not default to 'y' unless it is made more useful if it actually
> >> triggers. Ingo?
> >
> > Yeah, agreed absolutely.
> >
> > So this is a bit sad because RWX pages are a real problem in practice, especially
> > since the EFI addresses are well predictable, but generating a warning without
> > being able to fix it quickly is counterproductive as well, as it only annoys
> > people and makes them turn off the option. (Which we could do as well to begin
> > with, without the annoyance factor...)
> >
> > So the plan would be:
> >
> > 1) Make it default-n.
> >
> > 2) We should try to further improve the messages to make it easier to determine
> > what's wrong. We _do_ try to output symbolic information in the warning, to
> > make it easier to find buggy mappings, but these are not standard kernel
> > mappings. So I think we need an e820 mappings based semi-symbolic printout of
> > bad addresses - maybe even correlate it with the MMIO resource tree.
> >
> > 3) We should fix the EFI permission problem without relying on the firmware: it
> > appears we could just mark everything R-X optimistically, and if a write fault
> > happens (it's pretty rare in fact, only triggers when we write to an EFI
> > variable and so), we can mark the faulting page RW- on the fly, because it
> > appears that writable EFI sections, while not enumerated very well in 'old'
> > firmware, are still supposed to be page granular. (Even 'new' firmware I
> > wouldn't automatically trust to get the enumeration right...)
>
> I think it was Borislav who pointed out that this idea, which might
> have been mine, is a bit silly. Why not just skip mapping the EFI
> stuff in the init_pgd entirely and only map it in the EFI pgd?
>
> We'll have RWX stuff in the EFI pgd, but so what? If we're exposing
> anything that runs with the EFI pgd loaded to untrusted input, I think
> we've already lost.

That's certainly true, I was simply confused about the life time of these
mappings: I assumed they have to stay around. If they are meant to be and are
partly temporary today already, we should go the whole mile and make that really
so, because _today_ the mappings are permanent, so this is a real problem ...

Thanks,

Ingo

2015-11-07 07:06:03

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

* Matt Fleming <[email protected]> wrote:

> On Thu, 05 Nov, at 01:33:10PM, Linus Torvalds wrote:
> >
> > And if this turns out to be due to EFI wanting those permissions, what should
> > we do? People have talked about running the EFI callbacks in their own private
> > page table setup, which sounds like the right idea, but until that actually
> > *happens*....
>
> We have separate page tables today, for a few reasons, but mainly it's
> so that we can have an identity mapping of memory present in the
> region usually used by user processes - broken firmware still uses
> those identity mappings even after the kernel tells it they're
> invalid.
>
> Note that when I say "separate" I'm talking about trampoline_pgd[]
> which is also used by the x86 suspend/resume code.
>
> However, turns out that the issue with the current scheme is the fact
> that trampoline_pgd[] actually shares a couple of PGD entries with
> swapper_pg_dir as can be seen in setup_real_mode(),
>
>
> trampoline_pgd = (u64 *)__va(real_mode_header->trampoline_pgd);
> trampoline_pgd[0] = init_level4_pgt[pgd_index(__PAGE_OFFSET)].pgd;
> trampoline_pgd[511] = init_level4_pgt[511].pgd;
>
>
> So when we map the EFI regions in efi_map_regions() we're inserting
> them into swapper_pg_dir also, which is why you're seeing the
> warnings.
>
> If I remember correctly the rationale for using trampoline_pgd[] was
> that it already did what we wanted (provided the identity mapping) and
> would save us the overhead of maintaining more page tables for no good
> reason. Obviously this entire thread is a good reason.
>
> I suggest we stop using trampoline_pgd[] (since it has a good reason
> for sharing the kernel mapping PGD entries) and create our own so that
> we can isolate EFI completely.

Ok. Could you please make this fix a priority for upcoming EFI changes?

> For the immediate problem of the warnings spewing forth on all UEFI
> machines, at the very least the config options needs to be disabled by
> default, if not the patch reverted.

We'll certainly flip around the default, but reverting would be shooting
the messenger: the EFI code is endangering everyone else today, and for
no good reason as it appears... so the warning very much served its
purpose in pointing out a valid problem.

Thanks,

Ingo

2015-11-07 07:09:28

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

* Matt Fleming <[email protected]> wrote:

> On Fri, 06 Nov, at 07:55:50AM, Ingo Molnar wrote:
> >
> > 3) We should fix the EFI permission problem without relying on the firmware: it
> > appears we could just mark everything R-X optimistically, and if a write fault
> > happens (it's pretty rare in fact, only triggers when we write to an EFI
> > variable and so), we can mark the faulting page RW- on the fly, because it
> > appears that writable EFI sections, while not enumerated very well in 'old'
> > firmware, are still supposed to be page granular. (Even 'new' firmware I
> > wouldn't automatically trust to get the enumeration right...)
>
> Sorry, this isn't true. I misled you with one of my earlier posts on
> this topic. Let me try and clear things up...
>
> Writing to EFI regions has to do with every invocation of the EFI
> runtime services - it's not limited to when you read/write/delete EFI
> variables. In fact, EFI variables really have nothing to do with this
> discussion, they're a completely opaque concept to the OS, we have no
> idea how the firmware implements them. Everything is done via the EFI
> boot/runtime services.
>
> The firmware itself will attempt to write to EFI regions when we
> invoke the EFI services because that's where the PE/COFF ".data" and
> ".bss" sections live along with the heap. There's even some relocation
> fixups that occur as SetVirtualAddressMap() time so it'll write to
> ".text" too.
>
> Now, the above PE/COFF sections are usually (always?) contained within
> EFI regions of type EfiRuntimeServicesCode. We know this is true
> because the firmware folks have told us so, and because stopping that
> is the motivation behind the new EFI_PROPERTIES_TABLE feature in UEFI
> V2.5.
>
> The data sections within the region are also *not* guaranteed to be
> page granular because work was required in Tianocore for emitting
> sections with 4k alignment as part of the EFI_PROPERTIES_TABLE
> support.
>
> Ultimately, what this means is that if you were to attempt to
> dynamically fixup those regions that required write permission, you'd
> have to modify the mappings for the majority of the EFI regions
> anyway. And if you're blindly allowing write permission as a fixup,
> there's not much security to be had.

I think you misunderstood my suggestion: the 'fixup' would be changing it from R-X
to RW-, i.e. it would add 'write' permission but remove 'execute' permission.

Note that there would be no 'RWX' permission at any given moment - which is the
dangerous combination.

> > If that 'supposed to be' turns out to be 'not true' (not unheard of in
> > firmware land), then plan B would be to mark pages that generate write faults
> > RWX as well, to not break functionality. (This 'mark it RWX' is not something
> > that exploits would have easy access to, and we could also generate a warning
> > [after the EFI call has finished] if it ever triggers.)
> >
> > Admittedly this approach might not be without its own complications, but it
> > looks reasonably simple (I don't think we need per EFI call page tables,
> > etc.), and does not assume much about the firmware being able to enumerate its
> > permissions properly. Were we to merge EFI support today I'd have insisted on
> > trying such an approach from day 1 on.
>
> We already have separate EFI page tables, though with the caveat that
> we share some of swapper_pg_dir's PGD entries. The best solution would
> be to stop sharing entries and isolate the EFI mappings from every
> other page table structure, so that they're only used during the EFI
> service calls.

Absolutely. Can you try to fix this for v4.3?

Thanks,

Ingo

2015-11-07 07:39:37

by Ard Biesheuvel

[permalink] [raw]

Subject: Re: [GIT PULL] x86/mm changes for v4.4

On 7 November 2015 at 08:09, Ingo Molnar <[email protected]> wrote:
>
> * Matt Fleming <[email protected]> wrote:
>
>> On Fri, 06 Nov, at 07:55:50AM, Ingo Molnar wrote:
>> >
>> > 3) We should fix the EFI permission problem without relying on the firmware: it
>> > appears we could just mark everything R-X optimistically, and if a write fault
>> > happens (it's pretty rare in fact, only triggers when we write to an EFI
>> > variable and so), we can mark the faulting page RW- on the fly, because it
>> > appears that writable EFI sections, while not enumerated very well in 'old'
>> > firmware, are still supposed to be page granular. (Even 'new' firmware I
>> > wouldn't automatically trust to get the enumeration right...)
>>
>> Sorry, this isn't true. I misled you with one of my earlier posts on
>> this topic. Let me try and clear things up...
>>
>> Writing to EFI regions has to do with every invocation of the EFI
>> runtime services - it's not limited to when you read/write/delete EFI
>> variables. In fact, EFI variables really have nothing to do with this
>> discussion, they're a completely opaque concept to the OS, we have no
>> idea how the firmware implements them. Everything is done via the EFI
>> boot/runtime services.
>>
>> The firmware itself will attempt to write to EFI regions when we
>> invoke the EFI services because that's where the PE/COFF ".data" and
>> ".bss" sections live along with the heap. There's even some relocation
>> fixups that occur as SetVirtualAddressMap() time so it'll write to
>> ".text" too.
>>
>> Now, the above PE/COFF sections are usually (always?) contained within
>> EFI regions of type EfiRuntimeServicesCode. We know this is true
>> because the firmware folks have told us so, and because stopping that
>> is the motivation behind the new EFI_PROPERTIES_TABLE feature in UEFI
>> V2.5.
>>
>> The data sections within the region are also *not* guaranteed to be
>> page granular because work was required in Tianocore for emitting
>> sections with 4k alignment as part of the EFI_PROPERTIES_TABLE
>> support.
>>
>> Ultimately, what this means is that if you were to attempt to
>> dynamically fixup those regions that required write permission, you'd
>> have to modify the mappings for the majority of the EFI regions
>> anyway. And if you're blindly allowing write permission as a fixup,
>> there's not much security to be had.
>
> I think you misunderstood my suggestion: the 'fixup' would be changing it from R-X
> to RW-, i.e. it would add 'write' permission but remove 'execute' permission.
>
> Note that there would be no 'RWX' permission at any given moment - which is the
> dangerous combination.
>

The problem with that is that /any/ page in the UEFI runtime region
may intersect with both .text and .data of any of the PE/COFF images
that make up the runtime firmware (since the PE/COFF sections are not
necessarily page aligned). Such pages require RWX permissions. The
UEFI memory map does not provide the information to identify those
pages a priori (the entire region containing several PE/COFF images
could be covered by a single entry) so it is hard to guess which pages
should be allowed these RWX permissions.

>> > If that 'supposed to be' turns out to be 'not true' (not unheard of in
>> > firmware land), then plan B would be to mark pages that generate write faults
>> > RWX as well, to not break functionality. (This 'mark it RWX' is not something
>> > that exploits would have easy access to, and we could also generate a warning
>> > [after the EFI call has finished] if it ever triggers.)
>> >
>> > Admittedly this approach might not be without its own complications, but it
>> > looks reasonably simple (I don't think we need per EFI call page tables,
>> > etc.), and does not assume much about the firmware being able to enumerate its
>> > permissions properly. Were we to merge EFI support today I'd have insisted on
>> > trying such an approach from day 1 on.
>>
>> We already have separate EFI page tables, though with the caveat that
>> we share some of swapper_pg_dir's PGD entries. The best solution would
>> be to stop sharing entries and isolate the EFI mappings from every
>> other page table structure, so that they're only used during the EFI
>> service calls.
>
> Absolutely. Can you try to fix this for v4.3?
>
> Thanks,
>
> Ingo

2015-11-07 10:03:39