Re: [PATCH 0/2] iommu: Add io_ptdump debug interface for iommu

[Qinxin Xia <xiaqinxin@huawei.com>]

On 2025/9/10 22:15:47, Jason Gunthorpe <jgg@ziepe.ca> wrote:
> On Wed, Sep 10, 2025 at 11:20:08AM +0800, Qinxin Xia wrote:
>> Ok, I see, my colleague Wang Zhou also released a version of io_ptdump
>> a long time ago, which is implemented in smmu debugfs. Will recommends that
>> io_ptdump be implemented in a way similar to CPU page table dump. Using
>> debugfs to expose the data and using format-specific callbacks to implement
>> specific data dumps, I'll talk to him about this as well.
> 
> I feel we should have a iommu subsystem debugfs and per-iommu_domain
> directories to dump the page tables.
> 
> The smmu debugfs can report what iommu_domains each STE/CD is
> referencing.
> 
> This also needs RCU freeing of page table levels as a locking
> strategy.
> 
> Jason
> 

Thanks, I'll add RCU in the next version, but there's some confusion, Do 
you mean to create a directory for each domain? like：

/sys/kernel/debug/io_page_tables/domain_xxxx (xxxx=domain addr)

tree domain_xxxx like:
domain_xxxx
└── group x
│ └── device
└── io_ptdump

And the current design is such:

User Space Interface
└── DebugFS file: /sys/kernel/debug/io_page_tables
└── Operation: Reading this file triggers the entire debug information 
collection process

Kernel Space Components
├── Configuration option (CONFIG_IO_PTDUMP_DEBUGFS)
├── Initialization module (mm/io_ptdump.c)
│ └── device_initcall(io_ptdump_init)
│ └── io_ptdump_debugfs_register("io_page_tables")
├── DebugFS backend (mm/io_ptdump_debugfs.c)
│ └── DEFINE_SHOW_ATTRIBUTE(io_ptdump)
│ └── .show = io_ptdump_show
│ └── iommu_group_and_iova_dump(m)
└── IOMMU core extension (drivers/iommu/iommu.c)
└── iommu_group_and_iova_dump()
├── Traverse all IOMMU groups (via iommu_group_kset)
├── Filter groups with default domain using DMA_IOVA cookie
├── Build domain-group hierarchy
│ ├── domain_list: list head
│ ├── dump_domain: domain entry
│ │ ├── domain: pointer to iommu_domain
│ │ └── groups: list head of groups
│ └── dump_group: group entry
│ └── group: pointer to iommu_group
├── Output domain information
├── Output group and device information
└── Call iommu_iova_info_dump() to output IOVA mappings
└── Traverse IOVA red-black tree
└── Call domain->ops->dump_iova_prot() to get protection information
└── ARM SMMUv3 implementation (drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c)
└── arm_smmu_dump_iova_prot()
└── Call io_pgtable_ops->dump_iova_prot()
└── ARM LPAE implementation (drivers/iommu/io-pgtable-arm.c)
└── arm_lpae_dump_iova_prot()
├── Use __arm_lpae_iopte_walk() to traverse page tables
├── Obtain page table entry and level information
├── Format and output mapping range and information
└── Call dump_prot() to output protection flags
└── Use prot_bits array to parse permission bits

Do you mean that the interface in io-pgtable-arm.c is directly invoked 
during the process of obtaining page table information without passing 
through arm-smmu-v3.c?

I'll add STE and CD dumps to the next release. Any other suggestions?