Re: [RFC Patch Part2 V1 14/14] iommu/vt-d: update IOMMU state when memory hotplug happens

[Jiang Liu <jiang.liu@linux.intel.com>]

On 2014/1/8 14:14, Kai Huang wrote:
> On Wed, Jan 8, 2014 at 2:01 PM, Jiang Liu <jiang.liu@linux.intel.com> wrote:
>>
>>
>> On 2014/1/8 13:07, Kai Huang wrote:
>>> On Tue, Jan 7, 2014 at 5:00 PM, Jiang Liu <jiang.liu@linux.intel.com> wrote:
>>>> If static identity domain is created, IOMMU driver needs to update
>>>> si_domain page table when memory hotplug event happens. Otherwise
>>>> PCI device DMA operations can't access the hot-added memory regions.
>>>>
>>>> Signed-off-by: Jiang Liu <jiang.liu@linux.intel.com>
>>>> ---
>>>>  drivers/iommu/intel-iommu.c |   52 ++++++++++++++++++++++++++++++++++++++++++-
>>>>  1 file changed, 51 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
>>>> index 83e3ed4..35a987d 100644
>>>> --- a/drivers/iommu/intel-iommu.c
>>>> +++ b/drivers/iommu/intel-iommu.c
>>>> @@ -33,6 +33,7 @@
>>>>  #include <linux/dmar.h>
>>>>  #include <linux/dma-mapping.h>
>>>>  #include <linux/mempool.h>
>>>> +#include <linux/memory.h>
>>>>  #include <linux/timer.h>
>>>>  #include <linux/iova.h>
>>>>  #include <linux/iommu.h>
>>>> @@ -3689,6 +3690,54 @@ static struct notifier_block device_nb = {
>>>>         .notifier_call = device_notifier,
>>>>  };
>>>>
>>>> +static int intel_iommu_memory_notifier(struct notifier_block *nb,
>>>> +                                      unsigned long val, void *v)
>>>> +{
>>>> +       struct memory_notify *mhp = v;
>>>> +       unsigned long long start, end;
>>>> +       struct iova *iova;
>>>> +
>>>> +       switch (val) {
>>>> +       case MEM_GOING_ONLINE:
>>>> +               start = mhp->start_pfn << PAGE_SHIFT;
>>>> +               end = ((mhp->start_pfn + mhp->nr_pages) << PAGE_SHIFT) - 1;
>>>> +               if (iommu_domain_identity_map(si_domain, start, end)) {
>>>> +                       pr_warn("dmar: failed to build identity map for [%llx-%llx]\n",
>>>> +                               start, end);
>>>> +                       return NOTIFY_BAD;
>>>> +               }
>>>
>>> Better to use iommu_prepare_identity_map? For si_domain, if
>>> hw_pass_through is used, there's no page table.
>> Hi Kai,
>>         Good catch!
>> Seems function iommu_prepare_identity_map() is designed to handle
>> RMRRs. So how about avoiding of registering memory hotplug notifier
>> if hw_pass_through is true?
> 
> I think that's also fine :)
> 
> Btw, I have a related question to memory hotplug but not related to
> intel IOMMU specifically. For the devices use DMA remapping, suppose
> the device is already using the memory that we are trying to remove,
> is this case, looks we need to change the existing iova <-> pa
> mappings for the pa that is in the memory range about to be removed,
> and reset the mapping to different pa (iova remains the same). Does
> existing code have this covered? Is there a generic IOMMU layer memory
> hotplug notifier to handle memory removal?
That's a big issue about how to reclaim memory in use. Current rule
is that memory used by DMA won't be removed until released.

> 
> -Kai
>>
>> Thanks!
>> Gerry
>>
>>>
>>>> +               break;
>>>> +       case MEM_OFFLINE:
>>>> +       case MEM_CANCEL_ONLINE:
>>>> +               /* TODO: enhance RB-tree and IOVA code to support of splitting iova */
>>>> +               iova = find_iova(&si_domain->iovad, mhp->start_pfn);
>>>> +               if (iova) {
>>>> +                       unsigned long start_pfn, last_pfn;
>>>> +                       struct dmar_drhd_unit *drhd;
>>>> +                       struct intel_iommu *iommu;
>>>> +
>>>> +                       start_pfn = mm_to_dma_pfn(iova->pfn_lo);
>>>> +                       last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
>>>> +                       dma_pte_clear_range(si_domain, start_pfn, last_pfn);
>>>> +                       dma_pte_free_pagetable(si_domain, start_pfn, last_pfn);
>>>> +                       rcu_read_lock();
>>>> +                       for_each_active_iommu(iommu, drhd)
>>>> +                               iommu_flush_iotlb_psi(iommu, si_domain->id,
>>>> +                                       start_pfn, last_pfn - start_pfn + 1, 0);
>>>> +                       rcu_read_unlock();
>>>> +                       __free_iova(&si_domain->iovad, iova);
>>>> +               }
>>>
>>> The same as above. Looks we need to consider hw_pass_through for the si_domain.
>>>
>>> -Kai
>>>
>>>> +               break;
>>>> +       }
>>>> +
>>>> +       return NOTIFY_OK;
>>>> +}
>>>> +
>>>> +static struct notifier_block intel_iommu_memory_nb = {
>>>> +       .notifier_call = intel_iommu_memory_notifier,
>>>> +       .priority = 0
>>>> +};
>>>> +
>>>>  int __init intel_iommu_init(void)
>>>>  {
>>>>         int ret = -ENODEV;
>>>> @@ -3761,8 +3810,9 @@ int __init intel_iommu_init(void)
>>>>         init_iommu_pm_ops();
>>>>
>>>>         bus_set_iommu(&pci_bus_type, &intel_iommu_ops);
>>>> -
>>>>         bus_register_notifier(&pci_bus_type, &device_nb);
>>>> +       if (si_domain)
>>>> +               register_memory_notifier(&intel_iommu_memory_nb);
>>>>
>>>>         intel_iommu_enabled = 1;
>>>>
>>>> --
>>>> 1.7.10.4
>>>>
>>>> _______________________________________________
>>>> iommu mailing list
>>>> iommu@lists.linux-foundation.org
>>>> https://lists.linuxfoundation.org/mailman/listinfo/iommu