[v8 PATCH] arm64: mm: show direct mapping use in /proc/meminfo

[Yang Shi]

On 7/22/26 4:31 AM, Will Deacon wrote:
> On Tue, Jun 09, 2026 at 05:06:31PM -0700, Yang Shi wrote:
>>
>> On 6/9/26 2:42 PM, Yang Shi wrote:
>>> Since commit a166563e7ec3 ("arm64: mm: support large block mapping when
>>> rodata=full"), the direct mapping may be split on some machines instead
>>> keeping static since boot. It makes more sense to show the direct mapping
>>> use in /proc/meminfo than before.
>>> This patch will make /proc/meminfo show the direct mapping use like the
>>> below (4K base page size):
>>> DirectMap4K:       94792 kB
>>> DirectMap64K:     134208 kB
>>> DirectMap2M:     1173504 kB
>>> DirectMap32M:    5636096 kB
>>> DirectMap1G:    529530880 kB
>>>
>>> Although just the machines which support BBML2_NOABORT can split the
>>> direct mapping, show it on all machines regardless of BBML2_NOABORT so
>>> that the users have consistent view in order to avoid confusion.
>>>
>>> Although ptdump also can tell the direct map use, but it needs to dump
>>> the whole kernel page table. It is costly and overkilling. It is also
>>> in debugfs which may not be enabled by all distros. So showing direct
>>> map use in /proc/meminfo seems more convenient and has less overhead.
>>>
>>> Signed-off-by: Yang Shi <yang at os.amperecomputing.com>
>>> ---
>>>    arch/arm64/mm/mmu.c | 200 +++++++++++++++++++++++++++++++++++++++-----
>>>    1 file changed, 179 insertions(+), 21 deletions(-)
>>>
>>> v8: * Fixed the double accounting per Sashiko
>>>       * Responded the review comments from Sashiko
>>> v7: * Rebased to v7.1-rc4
>>>       * Changed "dm" to "lm" to follow ARM convention per Will
>>>       * Used __is_lm_alias() instead of reinventing a new helper per Will
>>> v6: * Rebased to v7.0-rc3
>>>       * Rebased on top of Anshuman's v5 "arm64/mm: Enable batched TLB flush
>>>         in unmap_hotplug_range()"
>>>       * Used const for direct map type array per Will
>>>       * Defined PUD size for 16K/64K even though it is not used per Will
>>>       * Removed the misleading comment in init_pmd() per Will
>>> v5: * Rebased to v6.19-rc4
>>>       * Fixed the build error for !CONFIG_PROC_FS
>>> v4: * Used PAGE_END instead of _PAGE_END(VA_BITS_MIN) per Ryan
>>>       * Used shorter name for the helpers and variables per Ryan
>>>       * Fixed accounting for memory hotunplug
>>> v3: * Fixed the over-accounting problems per Ryan
>>>       * Introduced helpers for add/sub direct map use and #ifdef them with
>>>         CONFIG_PROC_FS per Ryan
>>>       * v3 is a fix patch on top of v2
>>> v2: * Counted in size instead of the number of entries per Ryan
>>>       * Removed shift array per Ryan
>>>       * Use lower case "k" per Ryan
>>>       * Fixed a couple of build warnings reported by kernel test robot
>>>       * Fixed a couple of poential miscounts
>> Aha, Sashiko is so fast. 2 comments this time.
>>
>> #1
>>> Will these updates suffer from data races?
>>> The lm_meminfo array tracks direct mapping statistics and is updated using
>>> non-atomic += and -= operations. These updates are invoked from multiple
>>> independent code paths that do not share a common lock.
>>> For example, runtime page permission changes call
>>> split_kernel_leaf_mapping_locked() which executes under
>>> pgtable_split_lock,
>>> while memory hotplug operations like arch_remove_memory() execute under
>>> mem_hotplug_lock. Because these paths can run concurrently on different
>>> CPUs,
>>> the non-atomic arithmetic could result in data races and lost updates.
>> Yes, it may race with memory hotplug. I missed memory hotplug for v7
>> Sashiko. Two options to solve it:
>>
>> 1. Use atomic variables. Make lm_meminfo[NR_LM_TYPE] atomic_long_t, then
>> manipulate it with atomic ops.
>> 2. Protect it with a spin lock.
>>
>> The contention for the cache line or the spin lock should be rare since
>> memory hotplug should happen rarely. Any preference?
> I'd vote for keeping it simple and using a lock. Might be worth looking
> at the ongoing work from Lorenzo:
>
> https://lore.kernel.org/all/20260717-series-vmap-race-fix-v5-0-606a0ac6d3e5@kernel.org/

Thanks, Will. Took a look at Lorenzo's work. IIUC he used init_mm 
mmap_lock to protect vmalloc area and linear mapping collapse on x86 in 
order to serialize against ptdump. I don't think we should use init_mm 
mmap_lock. We want to serialize linear mapping counter update between 
split and memory hotplug, but neither of them takes init_mm mmap_lock. 
We can let them take init_mm mmap_lock, but it sounds too overkilling.

So a dedicated spin lock may be better IMHO, which just protects linear 
mapping counter.

Yang

>
> Will