Re: [PATCH v2 10/18] AMD/IOMMU: walk trees upon page fault

[Jan Beulich <jbeulich@suse.com>]

On 03.12.2021 10:03, Roger Pau Monné wrote:
> On Fri, Sep 24, 2021 at 11:51:15AM +0200, Jan Beulich wrote:
>> This is to aid diagnosing issues and largely matches VT-d's behavior.
>> Since I'm adding permissions output here as well, take the opportunity
>> and also add their displaying to amd_dump_page_table_level().
>>
>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
>>
>> --- a/xen/drivers/passthrough/amd/iommu.h
>> +++ b/xen/drivers/passthrough/amd/iommu.h
>> @@ -243,6 +243,8 @@ int __must_check amd_iommu_flush_iotlb_p
>>                                               unsigned long page_count,
>>                                               unsigned int flush_flags);
>>  int __must_check amd_iommu_flush_iotlb_all(struct domain *d);
>> +void amd_iommu_print_entries(const struct amd_iommu *iommu, unsigned int dev_id,
>> +                             dfn_t dfn);
>>  
>>  /* device table functions */
>>  int get_dma_requestor_id(uint16_t seg, uint16_t bdf);
>> --- a/xen/drivers/passthrough/amd/iommu_init.c
>> +++ b/xen/drivers/passthrough/amd/iommu_init.c
>> @@ -573,6 +573,9 @@ static void parse_event_log_entry(struct
>>                 (flags & 0x002) ? " NX" : "",
>>                 (flags & 0x001) ? " GN" : "");
>>  
>> +        if ( iommu_verbose )
>> +            amd_iommu_print_entries(iommu, device_id, daddr_to_dfn(addr));
>> +
>>          for ( bdf = 0; bdf < ivrs_bdf_entries; bdf++ )
>>              if ( get_dma_requestor_id(iommu->seg, bdf) == device_id )
>>                  pci_check_disable_device(iommu->seg, PCI_BUS(bdf),
>> --- a/xen/drivers/passthrough/amd/iommu_map.c
>> +++ b/xen/drivers/passthrough/amd/iommu_map.c
>> @@ -363,6 +363,50 @@ int amd_iommu_unmap_page(struct domain *
>>      return 0;
>>  }
>>  
>> +void amd_iommu_print_entries(const struct amd_iommu *iommu, unsigned int dev_id,
>> +                             dfn_t dfn)
>> +{
>> +    mfn_t pt_mfn;
>> +    unsigned int level;
>> +    const struct amd_iommu_dte *dt = iommu->dev_table.buffer;
>> +
>> +    if ( !dt[dev_id].tv )
>> +    {
>> +        printk("%pp: no root\n", &PCI_SBDF2(iommu->seg, dev_id));
>> +        return;
>> +    }
>> +
>> +    pt_mfn = _mfn(dt[dev_id].pt_root);
>> +    level = dt[dev_id].paging_mode;
>> +    printk("%pp root @ %"PRI_mfn" (%u levels) dfn=%"PRI_dfn"\n",
>> +           &PCI_SBDF2(iommu->seg, dev_id), mfn_x(pt_mfn), level, dfn_x(dfn));
>> +
>> +    while ( level )
>> +    {
>> +        const union amd_iommu_pte *pt = map_domain_page(pt_mfn);
>> +        unsigned int idx = pfn_to_pde_idx(dfn_x(dfn), level);
>> +        union amd_iommu_pte pte = pt[idx];
> 
> Don't you need to take a lock here (mapping_lock maybe?) in order to
> prevent changes to the IOMMU page tables while walking them?

Generally speaking - yes. But see the description saying "largely
matches VT-d's behavior". On VT-d both the IOMMU lock and the mapping
lock would need acquiring to be safe (the former could perhaps be
dropped early). Likewise here. If I wanted to do so here, I ought to
add a prereq patch adjusting the VT-d function. The main "excuse" not
to do so is/was probably the size of the series.

Jan