Re: [RFC PATCH 3/4] KVM: arm64: Install the block entry before unmapping the page mappings

[Alexandru Elisei]

Hi Yanan,

Sorry for taking so long to reply, been busy with other things unfortunately. I
did notice that you sent a new version of this series, but I would like to
continue our discussion on this patch, since it's easier to get the full context.

On 3/4/21 7:07 AM, wangyanan (Y) wrote:
> Hi Alex,
>
> On 2021/3/4 1:27, Alexandru Elisei wrote:
>> Hi Yanan,
>>
>> On 3/3/21 11:04 AM, wangyanan (Y) wrote:
>>> Hi Alex,
>>>
>>> On 2021/3/3 1:13, Alexandru Elisei wrote:
>>>> Hello,
>>>>
>>>> On 2/8/21 11:22 AM, Yanan Wang wrote:
>>>>> When KVM needs to coalesce the normal page mappings into a block mapping,
>>>>> we currently invalidate the old table entry first followed by invalidation
>>>>> of TLB, then unmap the page mappings, and install the block entry at last.
>>>>>
>>>>> It will cost a long time to unmap the numerous page mappings, which means
>>>>> there will be a long period when the table entry can be found invalid.
>>>>> If other vCPUs access any guest page within the block range and find the
>>>>> table entry invalid, they will all exit from guest with a translation fault
>>>>> which is not necessary. And KVM will make efforts to handle these faults,
>>>>> especially when performing CMOs by block range.
>>>>>
>>>>> So let's quickly install the block entry at first to ensure uninterrupted
>>>>> memory access of the other vCPUs, and then unmap the page mappings after
>>>>> installation. This will reduce most of the time when the table entry is
>>>>> invalid, and avoid most of the unnecessary translation faults.
>>>> I'm not convinced I've fully understood what is going on yet, but it seems to me
>>>> that the idea is sound. Some questions and comments below.
>>> What I am trying to do in this patch is to adjust the order of rebuilding block
>>> mappings from page mappings.
>>> Take the rebuilding of 1G block mappings as an example.
>>> Before this patch, the order is like:
>>> 1) invalidate the table entry of the 1st level(PUD)
>>> 2) flush TLB by VMID
>>> 3) unmap the old PMD/PTE tables
>>> 4) install the new block entry to the 1st level(PUD)
>>>
>>> So entry in the 1st level can be found invalid by other vcpus in 1), 2), and 3),
>>> and it's a long time in 3) to unmap
>>> the numerous old PMD/PTE tables, which means the total time of the entry being
>>> invalid is long enough to
>>> affect the performance.
>>>
>>> After this patch, the order is like:
>>> 1) invalidate the table ebtry of the 1st level(PUD)
>>> 2) flush TLB by VMID
>>> 3) install the new block entry to the 1st level(PUD)
>>> 4) unmap the old PMD/PTE tables
>>>
>>> The change ensures that period of entry in the 1st level(PUD) being invalid is
>>> only in 1) and 2),
>>> so if other vcpus access memory within 1G, there will be less chance to find the
>>> entry invalid
>>> and as a result trigger an unnecessary translation fault.
>> Thank you for the explanation, that was my understand of it also, and I believe
>> your idea is correct. I was more concerned that I got some of the details wrong,
>> and you have kindly corrected me below.
>>
>>>>> Signed-off-by: Yanan Wang <wangyanan55@xxxxxxxxxx>
>>>>> ---
>>>>>    arch/arm64/kvm/hyp/pgtable.c | 26 ++++++++++++--------------
>>>>>    1 file changed, 12 insertions(+), 14 deletions(-)
>>>>>
>>>>> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
>>>>> index 78a560446f80..308c36b9cd21 100644
>>>>> --- a/arch/arm64/kvm/hyp/pgtable.c
>>>>> +++ b/arch/arm64/kvm/hyp/pgtable.c
>>>>> @@ -434,6 +434,7 @@ struct stage2_map_data {
>>>>>        kvm_pte_t            attr;
>>>>>          kvm_pte_t            *anchor;
>>>>> +    kvm_pte_t            *follow;
>>>>>          struct kvm_s2_mmu        *mmu;
>>>>>        struct kvm_mmu_memory_cache    *memcache;
>>>>> @@ -553,15 +554,14 @@ static int stage2_map_walk_table_pre(u64 addr, u64 end,
>>>>> u32 level,
>>>>>        if (!kvm_block_mapping_supported(addr, end, data->phys, level))
>>>>>            return 0;
>>>>>    -    kvm_set_invalid_pte(ptep);
>>>>> -
>>>>>        /*
>>>>> -     * Invalidate the whole stage-2, as we may have numerous leaf
>>>>> -     * entries below us which would otherwise need invalidating
>>>>> -     * individually.
>>>>> +     * If we need to coalesce existing table entries into a block here,
>>>>> +     * then install the block entry first and the sub-level page mappings
>>>>> +     * will be unmapped later.
>>>>>         */
>>>>> -    kvm_call_hyp(__kvm_tlb_flush_vmid, data->mmu);
>>>>>        data->anchor = ptep;
>>>>> +    data->follow = kvm_pte_follow(*ptep);
>>>>> +    stage2_coalesce_tables_into_block(addr, level, ptep, data);
>>>> Here's how stage2_coalesce_tables_into_block() is implemented from the previous
>>>> patch (it might be worth merging it with this patch, I found it impossible to
>>>> judge if the function is correct without seeing how it is used and what is
>>>> replacing):
>>> Ok, will do this if v2 is going to be post.
>>>> static void stage2_coalesce_tables_into_block(u64 addr, u32 level,
>>>>                             kvm_pte_t *ptep,
>>>>                             struct stage2_map_data *data)
>>>> {
>>>>       u64 granule = kvm_granule_size(level), phys = data->phys;
>>>>       kvm_pte_t new = kvm_init_valid_leaf_pte(phys, data->attr, level);
>>>>
>>>>       kvm_set_invalid_pte(ptep);
>>>>
>>>>       /*
>>>>        * Invalidate the whole stage-2, as we may have numerous leaf entries
>>>>        * below us which would otherwise need invalidating individually.
>>>>        */
>>>>       kvm_call_hyp(__kvm_tlb_flush_vmid, data->mmu);
>>>>       smp_store_release(ptep, new);
>>>>       data->phys += granule;
>>>> }
>>>>
>>>> This works because __kvm_pgtable_visit() saves the *ptep value before calling
>>>> the
>>>> pre callback, and it visits the next level table based on the initial pte value,
>>>> not the new value written by stage2_coalesce_tables_into_block().
>>> Right. So before replacing the initial pte value with the new value, we have
>>> to use
>>> *data->follow = kvm_pte_follow(*ptep)* in stage2_map_walk_table_pre() to save
>>> the initial pte value in advance. And data->follow will be used when  we start to
>>> unmap the old sub-level tables later.
>> Right, stage2_map_walk_table_post() will use data->follow to free the table page
>> which is no longer needed because we've replaced the entire next level table with
>> a block mapping.
>>
>>>> Assuming the first patch in the series is merged ("KVM: arm64: Move the clean of
>>>> dcache to the map handler"), this function is missing the CMOs from
>>>> stage2_map_walker_try_leaf().
>>> Yes, the CMOs are not performed in stage2_coalesce_tables_into_block() currently,
>>> because I thought they were not needed when we rebuild the block mappings from
>>> normal page mappings.
>> This assumes that the *only* situation when we replace a table entry with a block
>> mapping is when the next level table (or tables) is *fully* populated. Is there a
>> way to prove that this is true? I think it's important to prove it unequivocally,
>> because if there's a corner case where this doesn't happen and we remove the
>> dcache maintenance, we can end up with hard to reproduce and hard to diagnose
>> errors in a guest.
> So there is still one thing left about this patch to determine, and that is
> whether we can straightly
> discard CMOs in stage2_coalesce_tables_into_block() or we should distinguish
> different situations.
>
> Now we know that the situation you have described won't happen, then I think we
> will only end up
> in stage2_coalesce_tables_into_block() in the following situation:
> 1) KVM create a new block mapping in stage2_map_walker_try_leaf() for the first
> time, if guest accesses
>     memory backed by a THP/HUGETLB huge page. And CMOs will be performed here.
> 2) KVM split this block mapping in dirty logging, and build only one new page
> mapping.
> 3) KVM will build other new page mappings in dirty logging lazily, if guest
> access any other pages
>     within the block. *In this stage, pages in this block may be fully mapped,
> or may be not.*
> 4) After dirty logging is disabled, KVM decides to rebuild the block mapping.
>
> Do we still have to perform CMOs when rebuilding the block mapping in step 4, if
> pages in the block
> were not fully mapped in step 3 ? I'm not completely sure about this.

Did some digging and this is my understanding of what is happening. Please correct
me if I get something wrong.

When the kernel coalesces the userspace PTEs into a transparent hugepage, KVM will
unmap the old mappings and mark the PMD table as invalidated via the MMU
notifiers. To have a table at the PMD level while the corresponding entry is a
block mapping in the userspace translation tables, it means that the table was
created *after* the userspace block mapping was created.

user_mem_abort() will create a PAGE_SIZE mapping when the backing userspace
mapping is a block mapping in the following situations:

1. The start of the userspace block mapping is not aligned to the start of the
stage 2 block mapping (see fault_supports_stage2_huge_mapping()).

2. The stage 2 block mapping falls outside the memslot (see
fault_supports_stage2_huge_mapping()).

3. The memslot logs dirty pages.

For 1 and 2, the only scenario in which we can use a stage 2 block mapping for the
faulting IPA is if the memslot is modified, and that means the IPA range will have
been unmapped first, which destroys the PMD table entry (kvm_set_memslot() will
call kvm_arch_flush_shadow_memslot because change == KVM_MR_MOVE).

This leaves us with scenario 3. We can get in this scenario if the memslot is
logging and the userspace mapping has been coalesced into a transparent huge page
before dirty logging was set or if the userspace mapping is a hugetlb page. To
allow a block mapping at stage 2, we first need to remove the
KVM_MEM_LOG_DIRTY_PAGES flag from the memslot. Then we need to get a dabt in the
IPA range backed by the userspace block mapping. At this point there's nothing to
guarantee that the *entire* IPA range backed by the userspace block mapping is
mapped at stage 2.

In this case, we definitely need to do dcache maintenance because the guest might
be running with the MMU off and doing loads from from PoC (assuming not FWB), and
whatever userspace wrote in the guest memory (like the kernel image) might still
be in the dcache. We also need to do the icache inval after the dcache clean +
inval because instruction fetches can be cached even if the MMU is off.

Thanks,

Alex

>
> Thanks,
>
> Yanan
>>> At least, they are not needed if we rebuild the block mappings backed by
>>> hugetlbfs
>>> pages, because we must have built the new block mappings for the first time
>>> before
>>> and now need to rebuild them after they were split in dirty logging. Can we
>>> agree on this?
>>> Then let's see the following situation.
>>>> I can think of the following situation where they
>>>> are needed:
>>>>
>>>> 1. The 2nd level (PMD) table that will be turned into a block is mapped at
>>>> stage 2
>>>> because one of the pages in the 3rd level (PTE) table it points to is
>>>> accessed by
>>>> the guest.
>>>>
>>>> 2. The kernel decides to turn the userspace mapping into a transparent huge page
>>>> and calls the mmu notifier to remove the mapping from stage 2. The 2nd level
>>>> table
>>>> is still valid.
>>> I have a question here. Won't the PMD entry been invalidated too in this case?
>>> If remove of the stage2 mapping by mmu notifier is an unmap operation of a range,
>>> then it's correct and reasonable to both invalidate the PMD entry and free the
>>> PTE table.
>>> As I know, kvm_pgtable_stage2_unmap() does so when unmapping a range.
>>>
>>> And if I was right about this, we will not end up in
>>> stage2_coalesce_tables_into_block()
>>> like step 3 describes, but in stage2_map_walker_try_leaf() instead. Because the
>>> PMD entry
>>> is invalid, so KVM will create the new 2M block mapping.
>> Looking at the code for stage2_unmap_walker(), I believe you are correct. After
>> the entire PTE table has been unmapped, the function will mark the PMD entry as
>> invalid. In the situation I described, at step 3 we would end up in the leaf
>> mapper function because the PMD entry is invalid. My example was wrong.
>>
>>> If I'm wrong about this, then I think this is a valid situation.
>>>> 3. Guest accesses a page which is not the page it accessed at step 1, which
>>>> causes
>>>> a translation fault. KVM decides we can use a PMD block mapping to map the
>>>> address
>>>> and we end up in stage2_coalesce_tables_into_block(). We need CMOs in this case
>>>> because the guest accesses memory it didn't access before.
>>>>
>>>> What do you think, is that a valid situation?
>>>>>        return 0;
>>>>>    }
>>>>>    @@ -614,20 +614,18 @@ static int stage2_map_walk_table_post(u64 addr, u64
>>>>> end, u32 level,
>>>>>                          kvm_pte_t *ptep,
>>>>>                          struct stage2_map_data *data)
>>>>>    {
>>>>> -    int ret = 0;
>>>>> -
>>>>>        if (!data->anchor)
>>>>>            return 0;
>>>>>    -    free_page((unsigned long)kvm_pte_follow(*ptep));
>>>>> -    put_page(virt_to_page(ptep));
>>>>> -
>>>>> -    if (data->anchor == ptep) {
>>>>> +    if (data->anchor != ptep) {
>>>>> +        free_page((unsigned long)kvm_pte_follow(*ptep));
>>>>> +        put_page(virt_to_page(ptep));
>>>>> +    } else {
>>>>> +        free_page((unsigned long)data->follow);
>>>>>            data->anchor = NULL;
>>>>> -        ret = stage2_map_walk_leaf(addr, end, level, ptep, data);
>>>> stage2_map_walk_leaf() -> stage2_map_walk_table_post calls put_page() and
>>>> get_page() once in our case (valid old mapping). It looks to me like we're
>>>> missing
>>>> a put_page() call when the function is called for the anchor. Have you found the
>>>> call to be unnecessary?
>>> Before this patch:
>>> When we find data->anchor == ptep, put_page() has been called once in advance
>>> for the anchor
>>> in stage2_map_walk_table_post(). Then we call stage2_map_walk_leaf() ->
>>> stage2_map_walker_try_leaf()
>>> to install the block entry, and only get_page() will be called once in
>>> stage2_map_walker_try_leaf().
>>> There is a put_page() followed by a get_page() for the anchor, and there will
>>> not be a problem about
>>> page_counts.
>> This is how I'm reading the code before your patch:
>>
>> - stage2_map_walk_table_post() returns early if there is no anchor.
>>
>> - stage2_map_walk_table_pre() sets the anchor and marks the entry as invalid. The
>> entry was a table so the leaf visitor is not called in __kvm_pgtable_visit().
>>
>> - __kvm_pgtable_visit() visits the next level table.
>>
>> - stage2_map_walk_table_post() calls put_page(), calls stage2_map_walk_leaf() ->
>> stage2_map_walker_try_leaf(). The old entry was invalidated by the pre visitor, so
>> it only calls get_page() (and not put_page() + get_page().
>>
>> I agree with your conclusion, I didn't realize that because the pre visitor marks
>> the entry as invalid, stage2_map_walker_try_leaf() will not call put_page().
>>
>>> After this patch:
>>> Before we find data->anchor == ptep and after it, there is not a put_page() call
>>> for the anchor.
>>> This is because that we didn't call get_page() either in
>>> stage2_coalesce_tables_into_block() when
>>> install the block entry. So I think there will not be a problem too.
>> I agree, the refcount will be identical.
>>
>>> Is above the right answer for your point?
>> Yes, thank you clearing that up for me.
>>
>> Thanks,
>>
>> Alex
>>
>>>>>        }
>>>>>    -    return ret;
>>>>> +    return 0;
>>>> I think it's correct for this function to succeed unconditionally. The error was
>>>> coming from stage2_map_walk_leaf() -> stage2_map_walker_try_leaf(). The function
>>>> can return an error code if block mapping is not supported, which we know is
>>>> supported because we have an anchor, and if only the permissions are different
>>>> between the old and the new entry, but in our case we've changed both the valid
>>>> and type bits.
>>> Agreed. Besides, we will definitely not end up updating an old valid entry for
>>> the anchor
>>> in stage2_map_walker_try_leaf(), because *anchor has already been invalidated in
>>> stage2_map_walk_table_pre() before set the anchor, so it will look like a build
>>> of new mapping.
>>>
>>> Thanks,
>>>
>>> Yanan
>>>> Thanks,
>>>>
>>>> Alex
>>>>
>>>>>    }
>>>>>      /*
>>>> .
>> .