Re: warning from domain_get_iommu

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

Hi,

On 2020/2/8 18:19, Jerry Snitselaar wrote:
> On Sat Feb 08 20, Lu Baolu wrote:
>> Hi Jerry,
>>
>> On 2020/2/7 17:34, Jerry Snitselaar wrote:
>>> On Thu Feb 06 20, Jerry Snitselaar wrote:
>>>> On Tue Feb 04 20, Jerry Snitselaar wrote:
>>>>> I'm working on getting a system to reproduce this, and verify it 
>>>>> also occurs
>>>>> with 5.5, but I have a report of a case where the kdump kernel gives
>>>>> warnings like the following on a hp dl360 gen9:
>>>>>
>>>>> [    2.830589] ehci_hcd: USB 2.0 'Enhanced' Host Controller (EHCI) 
>>>>> Driver
>>>>> [    2.832615] ehci-pci: EHCI PCI platform driver
>>>>> [    2.834190] ehci-pci 0000:00:1a.0: EHCI Host Controller
>>>>> [    2.835974] ehci-pci 0000:00:1a.0: new USB bus registered, 
>>>>> assigned bus number 1
>>>>> [    2.838276] ehci-pci 0000:00:1a.0: debug port 2
>>>>> [    2.839700] WARNING: CPU: 0 PID: 1 at 
>>>>> drivers/iommu/intel-iommu.c:598 domain_get_iommu+0x55/0x60
>>>>> [    2.840671] Modules linked in:
>>>>> [    2.840671] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 
>>>>> 4.18.0-170.el8.kdump2.x86_64 #1
>>>>> [    2.840671] Hardware name: HP ProLiant DL360 Gen9/ProLiant DL360 
>>>>> Gen9, BIOS P89 07/21/2019
>>>>> [    2.840671] RIP: 0010:domain_get_iommu+0x55/0x60
>>>>> [    2.840671] Code: c2 01 eb 0b 48 83 c0 01 8b 34 87 85 f6 75 0b 
>>>>> 48 63 c8 48 39 c2 75 ed 31 c0 c3 48 c1 e1 03 48 8b 05 70 f3 91 01 
>>>>> 48 8b 04 08 c3 <0f> 0b 31 c0 c3 31 c9 eb eb 66 90 0f 1f 44 00 00 41 
>>>>> 55 40 0f b6 f6
>>>>> [    2.840671] RSP: 0018:ffffc900000dfab8 EFLAGS: 00010202
>>>>> [    2.840671] RAX: ffff88ec7f1c8000 RBX: 0000006c7c867000 RCX: 
>>>>> 0000000000000000
>>>>> [    2.840671] RDX: 00000000fffffff0 RSI: 0000000000000000 RDI: 
>>>>> ffff88ec7f1c8000
>>>>> [    2.840671] RBP: ffff88ec6f7000b0 R08: ffff88ec7f19d000 R09: 
>>>>> ffff88ec7cbfcd00
>>>>> [    2.840671] R10: 0000000000000095 R11: ffffc900000df928 R12: 
>>>>> 0000000000000000
>>>>> [    2.840671] R13: ffff88ec7f1c8000 R14: 0000000000001000 R15: 
>>>>> 00000000ffffffff
>>>>> [    2.840671] FS:  0000000000000000(0000) 
>>>>> GS:ffff88ec7f600000(0000) knlGS:0000000000000000
>>>>> [    2.840671] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>> [    2.840671] CR2: 00007ff3e1713000 CR3: 0000006c7de0a004 CR4: 
>>>>> 00000000001606b0
>>>>> [    2.840671] Call Trace:
>>>>> [    2.840671]  __intel_map_single+0x62/0x140
>>>>> [    2.840671]  intel_alloc_coherent+0xa6/0x130
>>>>> [    2.840671]  dma_pool_alloc+0xd8/0x1e0
>>>>> [    2.840671]  e_qh_alloc+0x55/0x130
>>>>> [    2.840671]  ehci_setup+0x284/0x7b0
>>>>> [    2.840671]  ehci_pci_setup+0xa3/0x530
>>>>> [    2.840671]  usb_add_hcd+0x2b6/0x800
>>>>> [    2.840671]  usb_hcd_pci_probe+0x375/0x460
>>>>> [    2.840671]  local_pci_probe+0x41/0x90
>>>>> [    2.840671]  pci_device_probe+0x105/0x1b0
>>>>> [    2.840671]  driver_probe_device+0x12d/0x460
>>>>> [    2.840671]  device_driver_attach+0x50/0x60
>>>>> [    2.840671]  __driver_attach+0x61/0x130
>>>>> [    2.840671]  ? device_driver_attach+0x60/0x60
>>>>> [    2.840671]  bus_for_each_dev+0x77/0xc0
>>>>> [    2.840671]  ? klist_add_tail+0x3b/0x70
>>>>> [    2.840671]  bus_add_driver+0x14d/0x1e0
>>>>> [    2.840671]  ? ehci_hcd_init+0xaa/0xaa
>>>>> [    2.840671]  ? do_early_param+0x91/0x91
>>>>> [    2.840671]  driver_register+0x6b/0xb0
>>>>> [    2.840671]  ? ehci_hcd_init+0xaa/0xaa
>>>>> [    2.840671]  do_one_initcall+0x46/0x1c3
>>>>> [    2.840671]  ? do_early_param+0x91/0x91
>>>>> [    2.840671]  kernel_init_freeable+0x1af/0x258
>>>>> [    2.840671]  ? rest_init+0xaa/0xaa
>>>>> [    2.840671]  kernel_init+0xa/0xf9
>>>>> [    2.840671]  ret_from_fork+0x35/0x40
>>>>> [    2.840671] ---[ end trace e87b0d9a1c8135c4 ]---
>>>>> [    3.010848] ehci-pci 0000:00:1a.0: Using iommu dma mapping
>>>>> [    3.012551] ehci-pci 0000:00:1a.0: 32bit DMA uses non-identity 
>>>>> mapping
>>>>> [    3.018537] ehci-pci 0000:00:1a.0: cache line size of 64 is not 
>>>>> supported
>>>>> [    3.021188] ehci-pci 0000:00:1a.0: irq 18, io mem 0x93002000
>>>>> [    3.029006] ehci-pci 0000:00:1a.0: USB 2.0 started, EHCI 1.00
>>>>> [    3.030918] usb usb1: New USB device found, idVendor=1d6b, 
>>>>> idProduct=0002, bcdDevice= 4.18
>>>>> [    3.033491] usb usb1: New USB device strings: Mfr=3, Product=2, 
>>>>> SerialNumber=1
>>>>> [    3.035900] usb usb1: Product: EHCI Host Controller
>>>>> [    3.037423] usb usb1: Manufacturer: Linux 
>>>>> 4.18.0-170.el8.kdump2.x86_64 ehci_hcd
>>>>> [    3.039691] usb usb1: SerialNumber: 0000:00:1a.0
>>>>>
>>>>> It looks like the device finishes initializing once it figures out it
>>>>> needs dma mapping instead of the default
>>>>> passthrough. intel_alloc_coherent calls iommu_need_mapping, before it
>>>>> calls __intel_map_single, so I'm not sure why it is tripping over the
>>>>> WARN_ON in domain_get_iommu.
>>>>>
>>>>> one thing I noticed while looking at this is that domain_get_iommu can
>>>>> return NULL. So should there be something like the following in
>>>>> __intel_map_single after the domain_get_iommu call?
>>>>>
>>>>> if (!iommu)
>>>>>  goto error;
>>>>>
>>>>> It is possible to deref the null pointer later otherwise.
>>>>>
>>>>> Regards,
>>>>> Jerry
>>>>
>>>> I reproduced the warning with a 5.5 kernel on an Intel NUC5i5MYBE.
>>>
>>> Hi Baolu,
>>>
>>> I think I understand what is happening here. With the kdump boot
>>> translation is pre-enabled, so in intel_iommu_add_device things are
>>> getting set to DEFER_DEVICE_DOMAIN_INFO. When intel_alloc_coherent
>>> calls iommu_need_mapping it returns true, but doesn't do the dma
>>> domain switch because of DEFER_DEVICE_DOMAIN_INFO. Then
>>> __intel_map_single gets called and it calls deferred_attach_domain,
>>> which sets the domain to the group domain, which in this case is the
>>> identity domain. Then it calls domain_get_iommu, which spits out the
>>> warning because the domain type was dma and returns null. My
>>> workaround was to add a call to iommu_need_mapping and find_domain
>>> after the deferred_attach_domain, but I don't know if that is the
>>> correct solution. There are a couple other spots like intel_map_sg
>>> that have the deferred_attach_domain after iommu_need_mapping that
>>> possibly will suffer from the same problem.
>>>
>>> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
>>> index b5c5ab58d395..063f45323cfc 100644
>>> --- a/drivers/iommu/intel-iommu.c
>>> +++ b/drivers/iommu/intel-iommu.c
>>> @@ -3515,6 +3515,10 @@ static dma_addr_t __intel_map_single(struct 
>>> device *dev, phys_addr_t paddr,
>>>         if (!domain)
>>>                 return DMA_MAPPING_ERROR;
>>>
>>> +       if (!iommu_need_mapping(dev))
>>> +               return paddr;
>>> +
>>> +       domain = find_domain(dev);
>>>         iommu = domain_get_iommu(domain);
>>>         size = aligned_nrpages(paddr, size);
>>>
>>>
>>> I finally got a git repo over to one of these systems, and was
>>> able to reproduce the issue with the head of linus's tree. With commit
>>> 9235cb13d7d1 ("iommu/vt-d: Allow devices with RMRRs to use identity 
>>> domain")
>>> there are more of the warnings, because devices are using identity that
>>> weren't before.
>>>
>>
>> Is it possible to move deferred domain attachment to identity_mapping()?
>>
>> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
>> index 9dc37672bf89..234ab346198e 100644
>> --- a/drivers/iommu/intel-iommu.c
>> +++ b/drivers/iommu/intel-iommu.c
>> @@ -2913,13 +2913,11 @@ static int __init si_domain_init(int hw)
>>
>> static int identity_mapping(struct device *dev)
>> {
>> -       struct device_domain_info *info;
>> +       struct dmar_domain *domain;
>>
>> -       info = dev->archdata.iommu;
>> -       if (info && info != DUMMY_DEVICE_DOMAIN_INFO && info != 
>> DEFER_DEVICE_DOMAIN_INFO)
>> -               return (info->domain == si_domain);
>> +       domain = deferred_attach_domain(dev);
>>
>> -       return 0;
>> +       return (!domain || domain_type_is_si(domain));
>> }
>>
>> static int domain_add_dev_info(struct dmar_domain *domain, struct 
>> device *dev)
>>
>> Best regards,
>> baolu
> 
> Hi Baolu,
> 
> I think that would work, and then change the deferred_attach_domain
> calls in __intel_map_single and intel_map_sg to find_domain?
> 

Yes.

> I did a quick test with it on the system where I've been looking at this.
> 

Thanks!

Best regards,
baolu