[ powerpc maintainers: Thoughts on replacing the ptesync in
vmemmap_set_pmd() with a flush_cache_vmap() post-vmemmap-hotplug in the
generic code? ]

On 7/6/26 11:00, Muchun Song wrote:
>> On Jun 30, 2026, at 15:51, Vivian Wang <[email protected]> wrote:
>>
>> section_activate() does not flush TLB after populating new vmemmap
>> pages. On most architectures, this is okay. However it is a problem on
>> RISC-V since there the TLB caching non-present entries is permitted,
>> which causes spurious faults on some hardwares.
>>
>> This seems to be most easily reproduced with DEBUG_VM=y and
>> PAGE_POISONING=y, which causes these newly mapped struct pages to be
>> poisoned i.e. written to immediately after mapping.
>>
>> Add a hook vmemmap_populate_finalize() in __populate_section_memmap()
>> after population, to allow architectures to handle such situations as
>> needed. Then implement it on RISC-V to arrange for the existing
>> exception handler code to deal with these faults if they happen.
>>
>> Signed-off-by: Vivian Wang <[email protected]>
>> ---
>> Changes in v4:
>> - Rebase on v7.2-rc1, drop dependencies
>> - (No code changes otherwise)
>> - (A concurrency fix for mark_new_valid_map was sent independently)
>>  
>> https://lore.kernel.org/linux-riscv/20260629-riscv-mm-new-valid-map-ordering-v1-1-60d8c10c6...@iscas.ac.cn/
>> - Link to v3: 
>> https://patch.msgid.link/20260605-mark-after-vmemmap-populate-v3-1-a06001ac9...@iscas.ac.cn
>>
>> Changes in v3:
>> - Merged back into one patch (Mike)
>> - (No code changes otherwise.)
>> - Link to v2: 
>> https://patch.msgid.link/20260604-mark-after-vmemmap-populate-v2-0-ab6a7d03b...@iscas.ac.cn
>>
>> Changes in v2:
>> - Split patch in two, hook point and riscv hook 
>> - Explain hook necessity in patch 1 message (Mike)
>> - Make hook #define based (Mike)
>> - Call finalize hook only on populate success
>> - Link to v1: 
>> https://patch.msgid.link/20260525-mark-after-vmemmap-populate-v1-1-e698d859b...@iscas.ac.cn
>> ---
>> arch/riscv/include/asm/pgtable.h | 4 ++++
>> arch/riscv/mm/init.c             | 6 ++++++
>> mm/sparse-vmemmap.c              | 8 ++++++++
>> 3 files changed, 18 insertions(+)
>>
>> diff --git a/arch/riscv/include/asm/pgtable.h 
>> b/arch/riscv/include/asm/pgtable.h
>> index 5d5756bda82e..6b000c990ba7 100644
>> --- a/arch/riscv/include/asm/pgtable.h
>> +++ b/arch/riscv/include/asm/pgtable.h
>> @@ -1253,6 +1253,10 @@ static inline pte_t pte_swp_clear_exclusive(pte_t pte)
>> #define TASK_SIZE FIXADDR_START
>> #endif
>>
>> +/* Needed on SPARSEMEM_VMEMMAP */
>> +#define vmemmap_populate_finalize vmemmap_populate_finalize
>> +void __meminit vmemmap_populate_finalize(void);
>> +
>> #else /* CONFIG_MMU */
>>
>> #define PAGE_SHARED __pgprot(0)
>> diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
>> index 5b1b3c88b4d1..800cb5c007d1 100644
>> --- a/arch/riscv/mm/init.c
>> +++ b/arch/riscv/mm/init.c
>> @@ -1372,6 +1372,12 @@ int __meminit vmemmap_populate(unsigned long start, 
>> unsigned long end, int node,
>> */
>>      return vmemmap_populate_hugepages(start, end, node, altmap);
>> }
>> +
>> +void __meminit vmemmap_populate_finalize(void)
>> +{
>> +    /* Avoid faults on cached non-present TLB entries. */
>> +    mark_new_valid_map();
>> +}
>> #endif
>>
>> #if defined(CONFIG_MMU) && defined(CONFIG_64BIT)
>> diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
>> index 99e2be39671b..290cafcfd723 100644
>> --- a/mm/sparse-vmemmap.c
>> +++ b/mm/sparse-vmemmap.c
>> @@ -544,6 +544,12 @@ static int __meminit 
>> vmemmap_populate_compound_pages(unsigned long start_pfn,
>>
>> #endif
>>
>> +#ifndef vmemmap_populate_finalize
>> +static void __meminit vmemmap_populate_finalize(void)
>> +{
>> +}
>> +#endif
>> +
>> struct page * __meminit __populate_section_memmap(unsigned long pfn,
>> unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
>> struct dev_pagemap *pgmap)
>> @@ -564,6 +570,8 @@ struct page * __meminit 
>> __populate_section_memmap(unsigned long pfn,
>>      if (r < 0)
>>              return NULL;
>>
>> +    vmemmap_populate_finalize();
> Does the hook of flush_cache_vmap() work for you?  From the document of
> cachetlb.rst, it said: 
>
>       6) ``void flush_cache_vmap(unsigned long start, unsigned long end)``
>          ``void flush_cache_vunmap(unsigned long start, unsigned long end)``
>
>               Here in these two interfaces we are flushing a specific range
>               of (kernel) virtual addresses from the cache. After running,
>               there will be no entries in the cache for the kernel address
>               space for virtual addresses in the range 'start' to 'end-1'.
>
> It seems that flush_cache_vmap() is supposed to be called after a new kernel
> mapping has been installed, but before this address range is accessed.

Good point. I had initially avoided flush_cache_vmap() because, to be
honest, it looked scary. But the combination of "have vmemmap" and "have
flush_cache_vmap()" is much less scary.

I'd need to drop the address range check in RISC-V's flush_cache_vmap(),
but might as well.

> Therefore, invoking flush_cache_vmap() here aligns perfectly with its 
> interface
> definition. But how would this change impact other architectures? Currently,
> among architectures where CONFIG_SPARSEMEM=y, only RISC-V and PowerPC provide
> a concrete implementation for flush_cache_vmap().
>
> For PowerPC, the implementation is as follows:
>
>       static inline void flush_cache_vmap(unsigned long start, unsigned long 
> end)
>       {
>               asm volatile("ptesync" ::: "memory");
>       }
>
> Here, flush_cache_vmap() is also intended to supply the ptesync instruction
> that the kernel address mapping lacks. However, flush_cache_vmap() is not
> called in the __populate_section_memmap() path. So, how does PowerPC ensure
> ptesync is executed there? The answer lies in the architecture-specific
> vmemmap_set_pmd(), which injects the necessary ptesync.
>
> If we choose to reuse flush_cache_vmap() in the generic path, it yields the
> following benefits:
>
> - Avoids architectural churn: We do not need to introduce a new hook
>   like vmemmap_populate_finalize().
> - Cleans up arch code: PowerPC can deprecate its custom vmemmap_set_pmd()
>   and migrate to the generic implementation.
>
> Let me know if this rationale makes sense, or if there are any hidden edge
> cases I might have overlooked regarding other architectures.
>
> Thanks,
> Muchun

At least on the face of it, it makes sense. The main thing is I don't
know much about powerpc. So I'm also hoping powerpc maintainers could
chime in. 

I also don't know how this would affect new archs, but if they need
flush_cache_vmap() they *probably* also want this treatment for vmemmap
hotplug? I guess we can deal with that when it happens.

Thanks,
Vivian "dramforever" Wang


Reply via email to