Re: [PATCH v14 08/13] dma-iommu: Implement NESTED_MSI cookie

From: Auger Eric
Date: Sat Apr 10 2021 - 09:34:48 EST


Hi Zenghui,

On 4/7/21 9:39 AM, Zenghui Yu wrote:
> Hi Eric,
>
> On 2021/2/24 4:56, Eric Auger wrote:
>> Up to now, when the type was UNMANAGED, we used to
>> allocate IOVA pages within a reserved IOVA MSI range.
>>
>> If both the host and the guest are exposed with SMMUs, each
>> would allocate an IOVA. The guest allocates an IOVA (gIOVA)
>> to map onto the guest MSI doorbell (gDB). The Host allocates
>> another IOVA (hIOVA) to map onto the physical doorbell (hDB).
>>
>> So we end up with 2 unrelated mappings, at S1 and S2:
>>           S1             S2
>> gIOVA    ->     gDB
>>                 hIOVA    ->    hDB
>>
>> The PCI device would be programmed with hIOVA.
>> No stage 1 mapping would existing, causing the MSIs to fault.
>>
>> iommu_dma_bind_guest_msi() allows to pass gIOVA/gDB
>> to the host so that gIOVA can be used by the host instead of
>> re-allocating a new hIOVA.
>>
>>           S1           S2
>> gIOVA    ->    gDB    ->    hDB
>>
>> this time, the PCI device can be programmed with the gIOVA MSI
>> doorbell which is correctly mapped through both stages.
>>
>> Nested mode is not compatible with HW MSI regions as in that
>> case gDB and hDB should have a 1-1 mapping. This check will
>> be done when attaching each device to the IOMMU domain.
>>
>> Signed-off-by: Eric Auger <eric.auger@xxxxxxxxxx>
>
> [...]
>
>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>> index f659395e7959..d25eb7cecaa7 100644
>> --- a/drivers/iommu/dma-iommu.c
>> +++ b/drivers/iommu/dma-iommu.c
>> @@ -19,6 +19,7 @@
>>   #include <linux/irq.h>
>>   #include <linux/mm.h>
>>   #include <linux/mutex.h>
>> +#include <linux/mutex.h>
>
> Duplicated include.
sure
>
>>   #include <linux/pci.h>
>>   #include <linux/swiotlb.h>
>>   #include <linux/scatterlist.h>
>> @@ -29,12 +30,15 @@
>>   struct iommu_dma_msi_page {
>>       struct list_head    list;
>>       dma_addr_t        iova;
>> +    dma_addr_t        gpa;
>>       phys_addr_t        phys;
>> +    size_t            s1_granule;
>>   };
>>     enum iommu_dma_cookie_type {
>>       IOMMU_DMA_IOVA_COOKIE,
>>       IOMMU_DMA_MSI_COOKIE,
>> +    IOMMU_DMA_NESTED_MSI_COOKIE,
>>   };
>>     struct iommu_dma_cookie {
>> @@ -46,6 +50,7 @@ struct iommu_dma_cookie {
>>           dma_addr_t        msi_iova;
>
> msi_iova is unused in the nested mode, but we still set it to the start
> address of the RESV_SW_MSI region (in iommu_get_msi_cookie()), which
> looks a bit strange to me.
I agree with you
>
>>       };
>>       struct list_head        msi_page_list;
>> +    spinlock_t            msi_lock;
>
> Should msi_lock be grabbed everywhere msi_page_list is populated?
> Especially in iommu_dma_get_msi_page(), which can be invoked from the
> irqchip driver.
Yes I agree
>
>>         /* Domain for flush queue callback; NULL if flush queue not in
>> use */
>>       struct iommu_domain        *fq_domain;
>> @@ -87,6 +92,7 @@ static struct iommu_dma_cookie *cookie_alloc(enum
>> iommu_dma_cookie_type type)
>>         cookie = kzalloc(sizeof(*cookie), GFP_KERNEL);
>>       if (cookie) {
>> +        spin_lock_init(&cookie->msi_lock);
>>           INIT_LIST_HEAD(&cookie->msi_page_list);
>>           cookie->type = type;
>>       }
>> @@ -120,14 +126,17 @@ EXPORT_SYMBOL(iommu_get_dma_cookie);
>>    *
>>    * Users who manage their own IOVA allocation and do not want DMA
>> API support,
>>    * but would still like to take advantage of automatic MSI
>> remapping, can use
>> - * this to initialise their own domain appropriately. Users should
>> reserve a
>> + * this to initialise their own domain appropriately. Users may
>> reserve a
>>    * contiguous IOVA region, starting at @base, large enough to
>> accommodate the
>>    * number of PAGE_SIZE mappings necessary to cover every MSI
>> doorbell address
>> - * used by the devices attached to @domain.
>> + * used by the devices attached to @domain. The other way round is to
>> provide
>> + * usable iova pages through the iommu_dma_bind_doorbell API (nested
>> stages
>
> s/iommu_dma_bind_doorbell/iommu_dma_bind_guest_msi/ ?
correct
>
>> + * use case)
>>    */
>>   int iommu_get_msi_cookie(struct iommu_domain *domain, dma_addr_t base)
>>   {
>>       struct iommu_dma_cookie *cookie;
>> +    int nesting, ret;
>>         if (domain->type != IOMMU_DOMAIN_UNMANAGED)
>>           return -EINVAL;
>> @@ -135,7 +144,12 @@ int iommu_get_msi_cookie(struct iommu_domain
>> *domain, dma_addr_t base)
>>       if (domain->iova_cookie)
>>           return -EEXIST;
>>   -    cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
>> +    ret =  iommu_domain_get_attr(domain, DOMAIN_ATTR_NESTING, &nesting);
>
> Redundant space.
yep
>
>> +    if (!ret && nesting)
>> +        cookie = cookie_alloc(IOMMU_DMA_NESTED_MSI_COOKIE);
>> +    else
>> +        cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
>> +
>>       if (!cookie)
>>           return -ENOMEM;
>>   @@ -156,6 +170,7 @@ void iommu_put_dma_cookie(struct iommu_domain
>> *domain)
>>   {
>>       struct iommu_dma_cookie *cookie = domain->iova_cookie;
>>       struct iommu_dma_msi_page *msi, *tmp;
>> +    bool s2_unmap = false;
>>         if (!cookie)
>>           return;
>> @@ -163,7 +178,15 @@ void iommu_put_dma_cookie(struct iommu_domain
>> *domain)
>>       if (cookie->type == IOMMU_DMA_IOVA_COOKIE && cookie->iovad.granule)
>>           put_iova_domain(&cookie->iovad);
>>   +    if (cookie->type == IOMMU_DMA_NESTED_MSI_COOKIE)
>> +        s2_unmap = true;
>> +
>>       list_for_each_entry_safe(msi, tmp, &cookie->msi_page_list, list) {
>> +        if (s2_unmap && msi->phys) {
>
> I don't think @s2_unmap is necessary. Checking 'cookie->type==NESTED'
> directly shouldn't be that expensive.
agreed
>
>> +            size_t size = cookie_msi_granule(cookie);
>> +
>> +            WARN_ON(iommu_unmap(domain, msi->gpa, size) != size);
>> +        }
>>           list_del(&msi->list);
>>           kfree(msi);
>>       }
>> @@ -172,6 +195,92 @@ void iommu_put_dma_cookie(struct iommu_domain
>> *domain)
>>   }
>>   EXPORT_SYMBOL(iommu_put_dma_cookie);
>>   +/**
>> + * iommu_dma_bind_guest_msi - Allows to pass the stage 1
>> + * binding of a virtual MSI doorbell used by @dev.
>> + *
>> + * @domain: domain handle
>> + * @iova: guest iova
>
> Can we change it to 'giova' (to match the unbind side)?
sure
>
>> + * @gpa: gpa of the virtual doorbell
>> + * @size: size of the granule used for the stage1 mapping
>> + *
>> + * In nested stage use case, the user can provide IOVA/IPA bindings
>> + * corresponding to a guest MSI stage 1 mapping. When the host needs
>> + * to map its own MSI doorbells, it can use @gpa as stage 2 input
>> + * and map it onto the physical MSI doorbell.
>> + */
>> +int iommu_dma_bind_guest_msi(struct iommu_domain *domain,
>> +                 dma_addr_t iova, phys_addr_t gpa, size_t size)
>> +{
>> +    struct iommu_dma_cookie *cookie = domain->iova_cookie;
>> +    struct iommu_dma_msi_page *msi;
>> +    int ret = 0;
>> +
>> +    if (!cookie)
>> +        return -EINVAL;
>> +
>> +    if (cookie->type != IOMMU_DMA_NESTED_MSI_COOKIE)
>> +        return -EINVAL;
>> +
>> +    iova = iova & ~(dma_addr_t)(size - 1);
>> +    gpa = gpa & ~(phys_addr_t)(size - 1);
>> +
>> +    spin_lock(&cookie->msi_lock);
>> +
>> +    list_for_each_entry(msi, &cookie->msi_page_list, list) {
>> +        if (msi->iova == iova)
>> +            goto unlock; /* this page is already registered */
>> +    }
>> +
>> +    msi = kzalloc(sizeof(*msi), GFP_ATOMIC);
>> +    if (!msi) {
>> +        ret = -ENOMEM;
>> +        goto unlock;
>> +    }
>> +
>> +    msi->iova = iova;
>> +    msi->gpa = gpa;
>> +    msi->s1_granule = size;
>> +    list_add(&msi->list, &cookie->msi_page_list);
>> +unlock:
>> +    spin_unlock(&cookie->msi_lock);
>> +    return ret;
>> +}
>> +EXPORT_SYMBOL(iommu_dma_bind_guest_msi);
>> +
>> +void iommu_dma_unbind_guest_msi(struct iommu_domain *domain,
>> dma_addr_t giova)
>> +{
>> +    struct iommu_dma_cookie *cookie = domain->iova_cookie;
>> +    struct iommu_dma_msi_page *msi;
>> +
>> +    if (!cookie)
>> +        return;
>> +
>> +    if (cookie->type != IOMMU_DMA_NESTED_MSI_COOKIE)
>> +        return;
>> +
>> +    spin_lock(&cookie->msi_lock);
>> +
>> +    list_for_each_entry(msi, &cookie->msi_page_list, list) {
>> +        dma_addr_t aligned_giova =
>> +            giova & ~(dma_addr_t)(msi->s1_granule - 1);
>> +
>> +        if (msi->iova == aligned_giova) {
>> +            if (msi->phys) {
>> +                /* unmap the stage 2 */
>> +                size_t size = cookie_msi_granule(cookie);
>> +
>> +                WARN_ON(iommu_unmap(domain, msi->gpa, size) != size);
>> +            }
>> +            list_del(&msi->list);
>> +            kfree(msi);
>> +            break;
>> +        }
>> +    }
>> +    spin_unlock(&cookie->msi_lock);
>> +}
>> +EXPORT_SYMBOL(iommu_dma_unbind_guest_msi);
>> +
>>   /**
>>    * iommu_dma_get_resv_regions - Reserved region driver helper
>>    * @dev: Device from iommu_get_resv_regions()
>> @@ -1343,6 +1452,33 @@ static struct iommu_dma_msi_page
>> *iommu_dma_get_msi_page(struct device *dev,
>>           if (msi_page->phys == msi_addr)
>>               return msi_page;
>>   +    /*
>> +     * In nested stage mode, we do not allocate an MSI page in
>> +     * a range provided by the user. Instead, IOVA/IPA bindings are
>> +     * individually provided. We reuse thise IOVAs to build the
>
> s/thise/these/
yep
>
>> +     * GIOVA -> GPA -> MSI HPA nested stage mapping.
>> +     */
>> +    if (cookie->type == IOMMU_DMA_NESTED_MSI_COOKIE) {
>> +        list_for_each_entry(msi_page, &cookie->msi_page_list, list)
>> +            if (!msi_page->phys) {
>> +                int ret;
>> +
>> +                /* do the stage 2 mapping */
>> +                ret = iommu_map(domain,
>> +                        msi_page->gpa, msi_addr, size,
>
> Shouldn't we make sure that the size of S2 mapping is not less than
> s1_granule? Although what we need is actually a 32-bit TRANSLATER
> register, we don't know where it is mapped in S1.
The GITS_TRANSLATER is at offset 0x40 of the page. That's what I had in
mind. But indeed the doorbell may belong to another MSI controller with
different layout. So yes that's wort checking.

I will add this check in iommu_dma_bind_guest_msi()

>
>> +                        IOMMU_MMIO | IOMMU_WRITE);
>
> Is it intentional to drop the IOMMU_NOEXEC flag (from @prot)?
no restoring it.
>
>> +                if (ret) {
>> +                    pr_warn("MSI S2 mapping failed (%d)\n",
>> +                        ret);
>> +                    return NULL;
>> +                }
>> +                msi_page->phys = msi_addr;
>> +                return msi_page;
>> +            }
>> +        pr_warn("%s no MSI binding found\n", __func__);
>> +        return NULL;
>> +    }
>> +
>>       msi_page = kzalloc(sizeof(*msi_page), GFP_KERNEL);
>>       if (!msi_page)
>>           return NULL;
>
>
> Thanks,
> Zenghui
>

Thank you for your careful review!

Best Regards

Eric