Re: [RFC PATCH v4 01/13] iommu: Introduce dirty log tracking framework

[Keqian Zhu <zhukeqian1@huawei.com>]

Hi Baolu,

On 2021/5/8 11:46, Lu Baolu wrote:
> Hi Keqian,
> 
> On 5/7/21 6:21 PM, Keqian Zhu wrote:
>> Some types of IOMMU are capable of tracking DMA dirty log, such as
>> ARM SMMU with HTTU or Intel IOMMU with SLADE. This introduces the
>> dirty log tracking framework in the IOMMU base layer.
>>
>> Four new essential interfaces are added, and we maintaince the status
>> of dirty log tracking in iommu_domain.
>> 1. iommu_support_dirty_log: Check whether domain supports dirty log tracking
>> 2. iommu_switch_dirty_log: Perform actions to start|stop dirty log tracking
>> 3. iommu_sync_dirty_log: Sync dirty log from IOMMU into a dirty bitmap
>> 4. iommu_clear_dirty_log: Clear dirty log of IOMMU by a mask bitmap
>>
>> Note: Don't concurrently call these interfaces with other ops that
>> access underlying page table.
>>
>> Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
>> Signed-off-by: Kunkun Jiang <jiangkunkun@huawei.com>
>> ---
>>   drivers/iommu/iommu.c        | 201 +++++++++++++++++++++++++++++++++++
>>   include/linux/iommu.h        |  63 +++++++++++
>>   include/trace/events/iommu.h |  63 +++++++++++
>>   3 files changed, 327 insertions(+)
>>
>> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
>> index 808ab70d5df5..0d15620d1e90 100644
>> --- a/drivers/iommu/iommu.c
>> +++ b/drivers/iommu/iommu.c
>> @@ -1940,6 +1940,7 @@ static struct iommu_domain *__iommu_domain_alloc(struct bus_type *bus,
>>       domain->type = type;
>>       /* Assume all sizes by default; the driver may override this later */
>>       domain->pgsize_bitmap  = bus->iommu_ops->pgsize_bitmap;
>> +    mutex_init(&domain->switch_log_lock);
>>         return domain;
>>   }
>> @@ -2703,6 +2704,206 @@ int iommu_set_pgtable_quirks(struct iommu_domain *domain,
>>   }
>>   EXPORT_SYMBOL_GPL(iommu_set_pgtable_quirks);
>>   +bool iommu_support_dirty_log(struct iommu_domain *domain)
>> +{
>> +    const struct iommu_ops *ops = domain->ops;
>> +
>> +    return ops->support_dirty_log && ops->support_dirty_log(domain);
>> +}
>> +EXPORT_SYMBOL_GPL(iommu_support_dirty_log);
> 
> I suppose this interface is to ask the vendor IOMMU driver to check
> whether each device/iommu in the domain supports dirty bit tracking.
> But what will happen if new devices with different tracking capability
> are added afterward?
Yep, this is considered in the vfio part. We will query again after attaching or
detaching devices from the domain.  When the domain becomes capable, we enable
dirty log for it. When it becomes not capable, we disable dirty log for it.

> 
> To make things simple, is it possible to support this tracking only when
> all underlying IOMMUs support dirty bit tracking?
IIUC, all underlying IOMMUs you refer is of system wide. I think this idea may has
two issues. 1) The target domain may just contains part of system IOMMUs. 2) The
dirty tracking capability can be related to the capability of devices. For example,
we can track dirty log based on IOPF, which needs the capability of devices. That's
to say, we can make this framework more common.

> 
> Or, the more crazy idea is that we don't need to check this capability
> at all. If dirty bit tracking is not supported by hardware, just mark
> all pages dirty?
Yeah, I think this idea is nice :).

Still one concern is that we may have other dirty tracking methods in the future,
if we can't track dirty through iommu, we can still try other methods.

If there is no interface to check this capability, we have no chance to try
other methods. What do you think?

> 
>> +
>> +int iommu_switch_dirty_log(struct iommu_domain *domain, bool enable,
>> +               unsigned long iova, size_t size, int prot)
>> +{
>> +    const struct iommu_ops *ops = domain->ops;
>> +    unsigned long orig_iova = iova;
>> +    unsigned int min_pagesz;
>> +    size_t orig_size = size;
>> +    bool flush = false;
>> +    int ret = 0;
>> +
>> +    if (unlikely(!ops->switch_dirty_log))
>> +        return -ENODEV;
>> +
>> +    min_pagesz = 1 << __ffs(domain->pgsize_bitmap);
>> +    if (!IS_ALIGNED(iova | size, min_pagesz)) {
>> +        pr_err("unaligned: iova 0x%lx size 0x%zx min_pagesz 0x%x\n",
>> +               iova, size, min_pagesz);
>> +        return -EINVAL;
>> +    }
>> +
>> +    mutex_lock(&domain->switch_log_lock);
>> +    if (enable && domain->dirty_log_tracking) {
>> +        ret = -EBUSY;
>> +        goto out;
>> +    } else if (!enable && !domain->dirty_log_tracking) {
>> +        ret = -EINVAL;
>> +        goto out;
>> +    }
>> +
>> +    pr_debug("switch_dirty_log %s for: iova 0x%lx size 0x%zx\n",
>> +         enable ? "enable" : "disable", iova, size);
>> +
>> +    while (size) {
>> +        size_t pgsize = iommu_pgsize(domain, iova, size);
>> +
>> +        flush = true;
>> +        ret = ops->switch_dirty_log(domain, enable, iova, pgsize, prot);
> 
> Per minimal page callback is much expensive. How about using (pagesize,
> count), so that all pages with the same page size could be handled in a
> single indirect call? I remember I commented this during last review,
> but I don't mind doing it again.
Thanks for reminding me again :). I'll do that in next version.

Thanks,
Keqian