Re: [PATCH v3 1/3] arm64, vmcoreinfo : Append 'PTRS_PER_PGD' to vmcoreinfo

From: Bhupesh Sharma
Date: Sat May 04 2019 - 08:54:37 EST


On 04/03/2019 11:24 PM, Bhupesh Sharma wrote:
Hi James,

On 04/02/2019 10:56 PM, James Morse wrote:
Hi Bhupesh,

On 28/03/2019 11:42, Bhupesh Sharma wrote:
On 03/26/2019 10:06 PM, James Morse wrote:
On 20/03/2019 05:09, Bhupesh Sharma wrote:
With ARMv8.2-LVA architecture extension availability, arm64 hardware
which supports this extension can support a virtual address-space upto
52-bits.

Since at the moment we enable the support of this extension in kernel
via CONFIG flags, e.g.
- User-space 52-bit LVA via CONFIG_ARM64_USER_VA_BITS_52

so, there is no clear mechanism in the user-space right now to
determine these CONFIG flag values and hence determine the maximum
virtual address space supported by the underlying kernel.

User-space tools like 'makedumpfile' therefore are broken currently
as they have no proper method to calculate the 'PTRS_PER_PGD' value
which is required to perform a page table walk to determine the
physical address of a corresponding virtual address found in
kcore/vmcoreinfo.

If one appends 'PTRS_PER_PGD' number to vmcoreinfo for arm64,
it can be used in user-space to determine the maximum virtual address
supported by underlying kernel.

I don't think this really solves the problem, it feels fragile.

I can see how vmcoreinfo tells you VA_BITS==48, PAGE_SIZE==64K and PTRS_PER_PGD=1024.
You can use this to work out that the top level page table size isn't consistent with a
48bit VA, so 52bit VA must be in use...

But wasn't your problem walking the kernel page tables? In particular the offset that we
apply because the tables were based on a 48bit VA shifted up in swapper_pg_dir.

Where does the TTBR1_EL1 offset come from with this property? I assume makedumpfile
hard-codes it when it sees 52bit is in use ... somewhere.
We haven't solved the problem!

But isn't the TTBR1_EL1 offset already appended by the kernel via e842dfb5a2d3 ("arm64:
mm: Offset TTBR1 to allow 52-bit PTRS_PER_PGD")
in case of kernel configuration where 52-bit userspace VAs are possible.

Accordingly we have the following assembler helper in 'arch/arm64/include/asm/assembler.h':

.macro offset_ttbr1, ttbr
#ifdef CONFIG_ARM64_52BIT_VA
orr \ttbr, \ttbr, #TTBR1_BADDR_4852_OFFSET
#endif
.endm

where:
#ifdef CONFIG_ARM64_52BIT_VA
/* Must be at least 64-byte aligned to prevent corruption of the TTBR */
#define TTBR1_BADDR_4852_OFFSET (((UL(1) << (52 - PGDIR_SHIFT)) - \
(UL(1) << (48 - PGDIR_SHIFT))) * 8)
#endif

Sure, and all this would work today, because there is only one weird combination. But once
we support another combination of 52bit-va, you'd either need another value, or to start
using PTRS_PER_PGD as a flag for v5.1_FUNNY_BEHAVIOUR_ONE.

I completed my user-space experimentation with 52-bit kernel VA changes from Steve today and have shared a detailed review on his patchset (See <http://lists.infradead.org/pipermail/kexec/2019-April/022750.html>).

But first let me share some opinion on how we are adding the 52-bit address space changes for arm64 in the kernel.

I think we have ended up adding just a bit _too many_ CONFIG and MACRO values for the increased address space changes. For e.g. after the 52-bit kernel VA changes we have at-least 4 macros which explain the VA address range with CONFIG_ARM64_USER_KERNEL_VA_BITS_52=y:

VA_BITS = 52,
VA_BITS_ACTUAL = vabits_actual = 48,
VA_BITS_MIN = min (48, VA_BITS) = 48.
PTRS_PER_PGD = 64 (48-bit) or 1024 (52-bit)

Of these, VA_BITS, VA_BITS_ACTUAL and PTRS_PER_PGD are definitely of interest in the userspace as they define:

1.
/*
* VMEMMAP_SIZE - allows the whole linear region to be covered by
* a struct page array
*/
#define VMEMMAP_SIZE (UL(1) << (VA_BITS - PAGE_SHIFT - 1 + STRUCT_PAGE_MAX_SHIFT))

2. #define __is_lm_address(addr) (!((addr) & BIT(VA_BITS_ACTUAL - 1)))

We have discussed the usage of PTRS_PER_PGD in userspace already at quite some length, so I will focus on the other two below (VA_BITS and VA_BITS_ACUAL).

Both are critical for determining VMEMMAP_SIZE and whether a virtual address lies in the linear map range respectively.

I don't see any standard mechanism other than the following to achieve a working user-space with these changes:
- a sysfs node (may be a '/sys/devices/system/cpu/addressing-capabilities' node?) or HWCAP capability export for user-space utilities which perform a live analysis and use the above variables.
- exporting these variables in vmcoreinfo (for analysis of crash dump).

VA_BITS is already exported in vmcoreinfo, whereas I have proposed exporting PTRS_PER_PGD to vmcoreinfo via this patch.

For 52-bit kernel VA changes, VA_BITS_ACTUAL will also be needed in vmcoreinfo (See <http://lists.infradead.org/pipermail/kexec/2019-April/022750.html> for details).

Note that the above computation holds true both for PTRS_PER_PGD = 64 (48-bit kernel with
48-bit User VA) and 1024 (48-bit with 52-bit User VA) cases. And these are the
configurations for which we are trying to fix the user-space regressions reported (on
arm64) recently.

... and revisit it when there is another combination?

See above.

Today __cpu_setup() sets T0SZ and T1SZ differently for 52bit VA, but in the future it
could set them the same, or different the other-way-round.

Will makedumpfile using this value keep working once T1SZ is 52bit VA too? In this case
there would be no ttbr offset.

If you need another vmcoreinfo flag once that happens, we've done something wrong here.

I am currently experimenting with Steve's patches for 52-bit kernel VA
(<https://lwn.net/Articles/780093/>) and will comment more on the same when I am able to
get the user-space utilities like makedumpfile and kexec-tools to work with the same on
both ARMv8 Fast Simulator model and older CPUs which don't support ARMv8.2 extensions.

However, I think we should not hold up fixes for regressions already reported, because the
52-bit kernel VA changes probably still need some more rework.

Chucking things into vmcoreinfo isn't free: we need to keep them there forever, otherwise
yesterdays version of the tools breaks. Can we take the time to get this right for the
cases we know about?

Sure, but exporting variable(s) in vmcoreinfo in directly related to the information variable(s) we add in the kernel side without which the user-space would break.

I have added the requirements for 52-bit kernel VA above (i.e we need an additional VA_BITS_ACTUAL variable export'ed rather than any tinkering with already proposed PTRS_PER_PGD).

May be this is a good time to also talk about minimizing the kernel interfaces we are proposing to hold and indicate normal (48-bit) and extended (52-bit) address spaces on arm64.

Ideally, we would want to simplify it further to be on similar lines as x86:
CONFIG_X86_5LEVEL=y
vmcoreinfo_append_str("NUMBER(pgtable_l5_enabled)=%d\n",
pgtable_l5_enabled());

which seems much cleaner..

I am open to any suggestions on the same.

Yes the kernel code is going to move around, this is why the information we expose via
vmcoreinfo needs to be thought through: something we would always need, regardless of how
the kernel implements it.


(Not to mention what happens if the TTBR1_EL1 uses 52bit va, but TTBR0_EL1 doesn't)

I am wondering if there are any real users of the above combination.

Heh! Is there any hardware that supports this?

Pointer-auth changes all this again, as we may prefer to use the bits for pointer-auth in
one TTB or the other. PTRS_PER_PGD may show the 52bit value in this case, but neither TTBR
is mapping 52bits of VA.


So far, I have generally come across discussions where the following variations of the
address spaces have been proposed/requested:
- 48bit kernel VA + 48-bit User VA,
- 48-bit kernel VA + 52-bit User VA,

+ 52bit kernel, because there is excessive quantities of memory, and the kernel maps it
all, but 48-bit user, because it never maps all the memory, and we prefer the bits for
pointer-auth.

- 52-bit kernel VA + 52-bit User VA.

And... all four may happen with the same built image. I don't see how you can tell these
cases apart with the one (build-time-constant!) PTRS_PER_PGD value.

I'm sure some of these cases are hypothetical, but by considering it all now, we can avoid
three more kernel:vmcoreinfo updates, and three more fix-user-space-to-use-the-new-value.

Agree.

I think you probably do need PTRS_PER_PGD, as this is the one value the mm is using to
generate page tables. I'm pretty sure you also need T0SZ and T1SZ to know if that's
actually in use, or the kernel is bodging round it with an offset.

Sure, I am open to suggestions (as I realize that we need an additional VA_BITS_ACTUAL variable export'ed for 52-bit kernel VA changes).

Also how do we standardize reading T0SZ and T1SZ in user-space. Do you propose I make an enhancement in the cpu-feature-registers interface (see [1]) or the HWCAPS interface (see [2]) for the same?

[1]. https://www.kernel.org/doc/Documentation/arm64/cpu-feature-registers.txt
[2]. https://www.kernel.org/doc/Documentation/arm64/elf_hwcaps.txt

Thanks,
Bhupesh

Ping.

Hi James, Steve,

Any comments on the above points? At the moment we have to carry these fixes in the distribution kernels and I would like to have these fixed in upstream kernel itself.

Thanks,
Bhupesh