[RFC PATCH 5/5] iommu/arm-smmu-v3: Add Context Descriptor display to debugfs

[Qinxin Xia]

On 2026/3/14 05:04:10, Nicolin Chen <nicolinc at nvidia.com> wrote:
> On Fri, Mar 13, 2026 at 06:43:51PM +0800, Qinxin Xia wrote:
>>    * Directory Structure:
>>    * /sys/kernel/debug/iommu/arm_smmu_v3/
>>    * └── smmu0/
>> @@ -35,6 +42,8 @@
>>    *     └── stream_table
>>    *	   └── 0000:01:00.0:0/                    # PCI device with Stream ID 0
>>    *             ├── ste                           # Stream Table Entry
>> + *             └── context_descriptors/
>> + *                 └── all                       # All Context Descriptors
> 
> I wonder if we should do per-RID/PASID v.s. all. Otherwise, cd_dir
> seems unnecessary, as it could be a file instead?
> 

That's exactly what I'm thinking about.
Per-PASID could be better expanded, but now all maybe is enough.

>> +/**
>> + * smmu_debug_dump_cd() - Dump Context Descriptor details to seq_file
>> + * @seq: seq_file to write to
>> + * @dev: device associated with the CD
>> + * @ssid: Substream ID
>> + */
>> +static void smmu_debug_dump_cd(struct seq_file *seq, struct device *dev, u32 ssid)
>> +{
>> +	struct arm_smmu_master *master = dev_iommu_priv_get(dev);
>> +	struct arm_smmu_cd *cd;
>> +	u64 data;
>> +	int i;
>> +
>> +	if (!master) {
>> +		seq_puts(seq, "No master data\n");
>> +		return;
>> +	}
>> +
>> +	cd = arm_smmu_get_cd_ptr(master, ssid);
>> +	if (!cd) {
>> +		seq_printf(seq, "CD not available for SSID %u\n", ssid);
>> +		return;
>> +	}
> 
> master->cd_table can be free-ed. We could hit UAF in a race. This
> very likely needs arm_smmu_asid_lock.
> 
>> +	seq_printf(seq, "CD for Substream ID %u:\n", ssid);
>> +
>> +	/* CD 0 */
>> +	data = le64_to_cpu(cd->data[0]);
>> +	seq_printf(seq, "  Valid: %s\n", data & CTXDESC_CD_0_V ? "Yes" : "No");
>> +	seq_printf(seq, "  T0SZ: 0x%llx\n", data & CTXDESC_CD_0_TCR_T0SZ);
>> +	seq_printf(seq, "  EPD0: %s\n", data & CTXDESC_CD_0_TCR_EPD0 ? "Yes" : "No");
>> +	seq_printf(seq, "  EPD1: %s\n", data & CTXDESC_CD_0_TCR_EPD1 ? "Yes" : "No");
> 
> All these are unnecessary if V=0? Maybe the per-SSID output should
> depend on V=1 at all. Same might apply to STE.
> 
I made this judgment before dump:

		if (cd && (le64_to_cpu(cd->data[0]) & CTXDESC_CD_0_V)) {
			seq_printf(seq, "\n--- SSID %u ---\n", ssid);
			smmu_debug_dump_cd(seq, dev, ssid);
		}

So, the CDs dumped here are valid.
You remind me that "Valid" print can be omitted here :-)

>> +/* All CDs debugfs file operations */
>> +static int smmu_debugfs_all_cds_show(struct seq_file *seq, void *v)
>> +{
>> +	struct device *dev = seq->private;
>> +
>> +	smmu_debug_dump_all_cds(seq, dev);
>> +	return 0;
> 
> Just unwrap the smmu_debug_dump_all_cds().
> 
> Nicolin
> 
-- 
Thanks,
Qinxin