Re: [kvm-unit-tests PATCH v3 15/27] arm/arm64: mmu_disable: Clean and invalidate before disabling

[Nikos Nikoleris <nikos.nikoleris@arm.com>]

On 01/07/2022 11:24, Alexandru Elisei wrote:
> Hi,
> 
> On Fri, Jul 01, 2022 at 11:12:14AM +0200, Andrew Jones wrote:
>> On Thu, Jun 30, 2022 at 04:57:39PM +0100, Alexandru Elisei wrote:
>>> Hi,
>>>
>>> On Thu, Jun 30, 2022 at 04:16:09PM +0100, Nikos Nikoleris wrote:
>>>> Hi Alex,
>>>>
>>>> On 30/06/2022 12:24, Alexandru Elisei wrote:
>>>>> Hi,
>>>>>
>>>>> On Thu, Jun 30, 2022 at 12:08:41PM +0100, Nikos Nikoleris wrote:
>>>>>> Hi Alex,
>>>>>>
>>>>>> On 30/06/2022 11:20, Alexandru Elisei wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> On Thu, Jun 30, 2022 at 11:03:12AM +0100, Nikos Nikoleris wrote:
>>>>>>>> From: Andrew Jones <drjones@redhat.com>
>>>>>>>>
>>>>>>>> The commit message of commit 410b3bf09e76 ("arm/arm64: Perform dcache
>>>>>>>> clean + invalidate after turning MMU off") justifies cleaning and
>>>>>>>> invalidating the dcache after disabling the MMU by saying it's nice
>>>>>>>> not to rely on the current page tables and that it should still work
>>>>>>>> (per the spec), as long as there's an identity map in the current
>>>>>>>> tables. Doing the invalidation after also somewhat helped with
>>>>>>>> reenabling the MMU without seeing stale data, but the real problem
>>>>>>>> with reenabling was because the cache needs to be disabled with
>>>>>>>> the MMU, but it wasn't.
>>>>>>>>
>>>>>>>> Since we have to trust/validate that the current page tables have an
>>>>>>>> identity map anyway, then there's no harm in doing the clean
>>>>>>>> and invalidate first (it feels a little better to do so, anyway,
>>>>>>>> considering the cache maintenance instructions take virtual
>>>>>>>> addresses). Then, also disable the cache with the MMU to avoid
>>>>>>>> problems when reenabling. We invalidate the Icache and disable
>>>>>>>> that too for good measure. And, a final TLB invalidation ensures
>>>>>>>> we're crystal clean when we return from asm_mmu_disable().
>>>>>>>
>>>>>>> I'll point you to my previous reply [1] to this exact patch which explains
>>>>>>> why it's incorrect and is only papering over another problem.
>>>>>>>
>>>>>>> [1] https://lore.kernel.org/all/Yn5Z6Kyj62cUNgRN@monolith.localdoman/
>>>>>>>
>>>>>>
>>>>>> Apologies, I didn't mean to ignore your feedback on this. There was a
>>>>>> parallel discussion in [2] which I thought makes the problem more concrete.
>>>>>
>>>>> No problem, I figured as much :).
>>>>>
>>>>>>
>>>>>> This is Drew's patch as soon as he confirms he's also happy with the change
>>>>>> you suggested in the patch description I am happy to make it.
>>>>>>
>>>>>> Generally, a test will start off with the MMU enabled. At this point, we
>>>>>> access code, use and modify data (EfiLoaderData, EfiLoaderCode). Any of the
>>>>>> two regions could be mapped as any type of memory (I need to have another
>>>>>> look to confirm if it's Normal Memory). Then we want to take over control of
>>>>>> the page tables and for that reason we have to switch off the MMU. And any
>>>>>> access to code or data will be with Device-nGnRnE as you pointed out. If we
>>>>>> don't clean and invalidate, instructions and data might be in the cache and
>>>>>> we will be mixing memory attributes, won't we?
>>>>>
>>>>> I missed that comment, sorry. I've replied to that comment made in v2,
>>>>> here, in this ieration, in patch #19 ("arm/arm64: Add a setup sequence for
>>>>> systems that boot through EFI").
>>>>>
>>>>> This is the second time you've mentioned mixed memory attributes, so I'm
>>>>> going to reiterate the question I asked in patch #19: what do you mean by
>>>>> "mixing memory attributes" and what is wrong with it? Because it looks to
>>>>> me like you're saying that you cannot access data written with the MMU on
>>>>> when the MMU is off (and I assume the other way around, you cannot data
>>>>> written with the MMU off when the MMU is on).
>>>>>
>>>>
>>>> What I mean by mixing memory attributes is illustrated by the following
>>>> example.
>>>>
>>>> Take a memory location x, for which the page table entry maps to a physical
>>>> location as Normal, Inner-Shareable, Inner-writeback and Outer-writeback. If
>>>> we access it when the MMU is on and subquently when the MMU is off (treated
>>>> as Device-nGnRnE), then we have two accesses with mismatched memory
>>>> attributes to the same location. There is a whole section in the Arm ARM on
>>>> why this needs to be avoided (B2.8 Mismatched memory attributes) but the
>>>> result is "a loss of the uniprocessor semantics, ordering, or coherency". As
>>>> I understand, the solution to this is:
>>>>
>>>> "If the mismatched attributes for a Location mean that multiple cacheable
>>>> accesses to the Location might be made with different shareability
>>>> attributes, then uniprocessor semantics, ordering, and coherency are
>>>> guaranteed only if:
>>>> • Software running on a PE cleans and invalidates a Location from cache
>>>> before and after each read or write to that Location by that PE.
>>>> • A DMB barrier with scope that covers the full shareability of the accesses
>>>> is placed between any accesses to the same memory Location that use
>>>> different attributes."
>>>
>>> Ok, so this is about *mismatched* memory attributes. I searched the Arm ARM
>>> for the string "mixed" and nothing relevant came up.
>>>
>>> Device-whatever memory is outer shareable and kvm-unit-tests maps memory as
>>> inner shareable, so that matches the "different shareability attributes"
>>> part of the paragraph.
>>>
>>> But I would like to point out that there is only one type of cacheable
>>> access that is being performed, when the MMU is on. When the MMU is off,
>>> the access is not cacheable. So there are two types of accesses being
>>> performed:
>>>
>>> - cacheable + inner-shareable (MMU on)
>>> - non-cacheable + outer-shareable (MMU off)
>>>
>>> It looks to me like the paragraph doesn't apply to our case, because there
>>> are no "multiple cacheable accesses [..] made with different shareability
>>> attributes". Do you agree?
>>>

No, I think the rules on mismatched memory attributes still apply.

"Physical memory locations are accessed with mismatched attributes if 
all accesses to the location do not use a common definition of all of 
the following attributes of that location:
• Memory type: Device-nGnRnE, Device-nGnRE, Device-nGRE, Device-GRE or 
Normal."

When we're switching off the MMU, we're changing the type of memory for 
many of the efi regions and therefore the rules on mismatched memory 
attributes still apply.

>>>>
>>>> So unless UEFI maps all memory as Device-nGnRnE we have to do something. I
>>>> will try to find out more about UEFI's page tables.
>>>
>>> That's important to know, especially regarding the text section of the
>>> image. If UEFI doesnt' clean it to PoC, kvm-unit-tests must do it in order
>>> to execute correctly with the MMU off.
>>>

Drew confirmed that UEFI page tables map most memory as Normal, Cacheable.

>>
>> Hi Alex and Nikos,
>>
>> Indeed my experiments on bare-metal made this change necessary. I'm happy
>> to see this discussion, though, as this patch could be tweaked or at least
>> the commit message improved in order to better explain what's going on and
>> why the changes are necessary. IOW, I have no problem with this patch
> 
> If you fix the commit message to be architecturally correct then you will
> come to the conclusion that the patch is architecturally incorrect because
> while it fixes the problem you were seeing, it breaks asm_mmu_disable in
> all other cases.
> > The problem you were seeing according to my investigation was this:
> 
> __phys_offset and __phys_end are written with the MMU on and the most up to
> date value is in the cache. When the MMU is turned off, the value that
> asm_mmu_disable reads is the stale value from main memory and it will not
> clean + invalidate all the memory, which is what we want. This assumes that
> UEFI cleaned the image to PoC, otherwise, that will need to be cleaned too
> by kvm-unit-tests before turning off the MMU.

The clean and/or invalidate instructions are also affected by the Memory 
type attributes so any operation after we switch off the MMU will be 
affected. Having said that, we might be fine:

"For Device memory and Normal memory that is Inner Non-cacheable, Outer 
Non-cacheable, these instructions must affect the caches of all PEs in 
the Outer Shareable shareability domain of the PE on which the 
instruction is operating."

> 
> This was explained before, both on your original UEFI support series on
> github [1], and on this list.
> 
> As for why it breaks asm_mmu_disable for all other cases:
> 
> The purpose of the clean in asm_mmu_disable is for the CPU to sync the
> caches with main memory when the MMU is turned off (to propagate the most
> up-to-date value from the cache to main memory); the purpose of the
> invalidate is to make sure that the CPU reads from main memory instead of
> the cache once the MMU is turned back on - if the cache line is still
> valid, the CPU wll read the values written *before* the MMU was turned
> off, not the values written *after* the MMU was turned off.
> 
> If you do the dcache clean + invalidate *before* turning the MMU on, the
> CPU can speculate a read and allocate a new cache line before the MMU is
> turned off, which would make the invalidate useless. Speculation is
> prohibited with the MMU off, that's why the invalidate must be done with
> the MMU off.
> 
> Because of this reason I believe the patch is incorrect.
> 

Given the effect of the DC instructions for Device Memory maybe keeping 
the invalidation after switching the MMU off seems to be the right 
approach. We only need to clean anything we use during the clean and 
invalidate as you suggest.

>> being dropped and replaced by one of you with something that "makes
>> more sense" as long as the outcome (coherent execution on bare-metal)
>> still works.
> 
> Hmm... maybe an experiment will work. I propose the following:
> > 1. Revert this patch.
> 2. Apply this diff on top of the series:
> 
> diff --git a/lib/arm/setup.c b/lib/arm/setup.c
> index 30d04d0eb100..913f4088d96c 100644
> --- a/lib/arm/setup.c
> +++ b/lib/arm/setup.c
> @@ -374,6 +374,11 @@ static efi_status_t efi_mem_init(efi_bootinfo_t *efi_bootinfo)
>                  }
>          }
>          __phys_end &= PHYS_MASK;
> +
> +       asm volatile("dc cvau, %0\n" :: "r" (&__phys_offset) : "memory");
> +       asm volatile("dc cvau, %0\n" :: "r" (&__phys_end) : "memory");

These need to be dc cvac.

> +       dsb(sy);
> +
>          asm_mmu_disable();
> 
>          if (free_mem_pages == 0)
> 
> This is the solution, based on an architectural explanation of what we were
> observing, that I proposed on your github branch, a solution that you've
> tested with the result:
> 
> "I tested at least 10 times (lost count) with a build where "arm/arm64:
> mmu_disable: Clean and invalidate before disabling" was reverted from the
> target-efi branch and your hack was applied. It worked every time."
>

FWIW, I don't think running 10 times on one machine shows much about the 
architectural correctness of either solution.

Thanks,

Nikos

> [1] https://github.com/rhdrjones/kvm-unit-tests/commit/fc58684bc47b7d07d75098fdfddb6083e9b12104#commitcomment-44222926
> 
> Thanks,
> Alex