Re: [PATCH 00/12] iommu: Remove IOMMU_DEV_FEAT_SVA/_IOPF

[Baolu Lu <baolu.lu@linux.intel.com>]

On 2/15/25 19:35, Zhangfei Gao wrote:
> On Sat, 15 Feb 2025 at 18:09, Baolu Lu<baolu.lu@linux.intel.com> wrote:
>> On 2/15/25 16:11, Zhangfei Gao wrote:
>>> It does not relate to multi devices, one device also happens when user
>>> page fault triggers.
>>>
>>> iopf_queue_remove_device is called.
>>> rcu_assign_pointer(param->fault_param, NULL);
>>>
>>> call trace
>>> [  304.961312] Call trace:
>>> [  304.961314]  show_stack+0x20/0x38 (C)
>>> [  304.961319]  dump_stack_lvl+0xc0/0xd0
>>> [  304.961324]  dump_stack+0x18/0x28
>>> [  304.961327]  iopf_queue_remove_device+0xb0/0x1f0
>>> [  304.961331]  arm_smmu_remove_master_domain+0x204/0x250
>>> [  304.961336]  arm_smmu_attach_commit+0x64/0x100
>>> [  304.961338]  arm_smmu_attach_dev_nested+0x104/0x1a8
>>> [  304.961340]  __iommu_attach_device+0x2c/0x110
>>> [  304.961343]  __iommu_device_set_domain.isra.0+0x78/0xe0
>>> [  304.961345]  __iommu_group_set_domain_internal+0x78/0x160
>>> [  304.961347]  iommu_replace_group_handle+0x9c/0x150
>>> [  304.961350]  iommufd_fault_domain_replace_dev+0x88/0x120
>>> [  304.961353]  iommufd_device_do_replace+0x190/0x3c0
>>> [  304.961355]  iommufd_device_change_pt+0x270/0x688
>>> [  304.961357]  iommufd_device_replace+0x20/0x38
>>> [  304.961359]  vfio_iommufd_physical_attach_ioas+0x30/0x78
>>> [  304.961363]  vfio_df_ioctl_attach_pt+0xa8/0x188
>>> [  304.961366]  vfio_device_fops_unl_ioctl+0x310/0x990
>>>
>>>
>>> When page fault triggers:
>>>
>>> [ 1016.383578] ------------[ cut here ]-----------
>>> [ 1016.388184] WARNING: CPU: 35 PID: 717 at
>>> drivers/iommu/io-pgfault.c:231 iommu_report_device_fault+0x2c8/0x470
>> It's likely that iopf_queue_add_device() was not called for this device.
> iopf_queue_add_device is called, but quickly iopf_queue_remove_device
> is called during guest bootup.
> Then fault_param is set to NULL.
> 
> arm_smmu_attach_commit
> arm_smmu_remove_master_domain
> // newly added in the first patch
>         if (master_domain) {
>                    if (master_domain->using_iopf)

It seems the above check is incorrect. We only need to disable iopf when
an iopf-capable domain is about to be removed. Will the following
additional change make any difference?

diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c 
b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
index 28e67a9e3861..9b9ef738d070 100644
--- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
+++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
@@ -2822,7 +2822,7 @@ static void arm_smmu_remove_master_domain(struct 
arm_smmu_master *master,
         spin_unlock_irqrestore(&smmu_domain->devices_lock, flags);

         if (master_domain) {
-               if (master_domain->using_iopf)
+               if (domain->iopf_handler)
                         arm_smmu_disable_iopf(master);
                 kfree(master_domain);
         }

>                            arm_smmu_disable_iopf(master); ->
> iopf_queue_remove_device
>                    kfree(master_domain);
>            }
> 
> As a comparison, without this patchset, only iopf_queue_add_device is
> called, not call iopf_queue_remove_device

Thanks,
baolu