iommu/vt-d: change intel-iommu to use IOVA frame numbers
authorOmer Peleg <omer@cs.technion.ac.il>
Wed, 20 Apr 2016 08:33:57 +0000 (11:33 +0300)
committerDavid Woodhouse <David.Woodhouse@intel.com>
Wed, 20 Apr 2016 19:07:22 +0000 (15:07 -0400)
Make intel-iommu map/unmap/invalidate work with IOVA pfns instead of
pointers to "struct iova". This avoids using the iova struct from the IOVA
red-black tree and the resulting explicit find_iova() on unmap.

This patch will allow us to cache IOVAs in the next patch, in order to
avoid rbtree operations for the majority of map/unmap operations.

Note: In eliminating the find_iova() operation, we have also eliminated
the sanity check previously done in the unmap flow. Arguably, this was
overhead that is better avoided in production code, but it could be
brought back as a debug option for driver development.

Signed-off-by: Omer Peleg <omer@cs.technion.ac.il>
[mad@cs.technion.ac.il: rebased, fixed to not break iova api, and reworded
 the commit message]
Signed-off-by: Adam Morrison <mad@cs.technion.ac.il>
Reviewed-by: Shaohua Li <shli@fb.com>
Reviewed-by: Ben Serebrin <serebrin@google.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
drivers/iommu/intel-iommu.c

index d100583e15e37599f21745e15bcfbf3eac522a38..a8babc43e6d40880110944f0a29423f71052eb6b 100644 (file)
@@ -459,7 +459,7 @@ static LIST_HEAD(dmar_rmrr_units);
 static void flush_unmaps_timeout(unsigned long data);
 
 struct deferred_flush_entry {
-       struct iova *iova;
+       unsigned long iova_pfn;
        unsigned long nrpages;
        struct dmar_domain *domain;
        struct page *freelist;
@@ -3353,7 +3353,7 @@ error:
 }
 
 /* This takes a number of _MM_ pages, not VTD pages */
-static struct iova *intel_alloc_iova(struct device *dev,
+static unsigned long intel_alloc_iova(struct device *dev,
                                     struct dmar_domain *domain,
                                     unsigned long nrpages, uint64_t dma_mask)
 {
@@ -3373,16 +3373,16 @@ static struct iova *intel_alloc_iova(struct device *dev,
                iova = alloc_iova(&domain->iovad, nrpages,
                                  IOVA_PFN(DMA_BIT_MASK(32)), 1);
                if (iova)
-                       return iova;
+                       return iova->pfn_lo;
        }
        iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
        if (unlikely(!iova)) {
                pr_err("Allocating %ld-page iova for %s failed",
                       nrpages, dev_name(dev));
-               return NULL;
+               return 0;
        }
 
-       return iova;
+       return iova->pfn_lo;
 }
 
 static struct dmar_domain *__get_valid_domain_for_dev(struct device *dev)
@@ -3480,7 +3480,7 @@ static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
 {
        struct dmar_domain *domain;
        phys_addr_t start_paddr;
-       struct iova *iova;
+       unsigned long iova_pfn;
        int prot = 0;
        int ret;
        struct intel_iommu *iommu;
@@ -3498,8 +3498,8 @@ static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
        iommu = domain_get_iommu(domain);
        size = aligned_nrpages(paddr, size);
 
-       iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
-       if (!iova)
+       iova_pfn = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
+       if (!iova_pfn)
                goto error;
 
        /*
@@ -3517,7 +3517,7 @@ static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
         * might have two guest_addr mapping to the same host paddr, but this
         * is not a big problem
         */
-       ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
+       ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova_pfn),
                                 mm_to_dma_pfn(paddr_pfn), size, prot);
        if (ret)
                goto error;
@@ -3525,18 +3525,18 @@ static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
        /* it's a non-present to present mapping. Only flush if caching mode */
        if (cap_caching_mode(iommu->cap))
                iommu_flush_iotlb_psi(iommu, domain,
-                                     mm_to_dma_pfn(iova->pfn_lo),
+                                     mm_to_dma_pfn(iova_pfn),
                                      size, 0, 1);
        else
                iommu_flush_write_buffer(iommu);
 
-       start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
+       start_paddr = (phys_addr_t)iova_pfn << PAGE_SHIFT;
        start_paddr += paddr & ~PAGE_MASK;
        return start_paddr;
 
 error:
-       if (iova)
-               __free_iova(&domain->iovad, iova);
+       if (iova_pfn)
+               free_iova(&domain->iovad, iova_pfn);
        pr_err("Device %s request: %zx@%llx dir %d --- failed\n",
                dev_name(dev), size, (unsigned long long)paddr, dir);
        return 0;
@@ -3576,7 +3576,7 @@ static void flush_unmaps(struct deferred_flush_data *flush_data)
                        unsigned long mask;
                        struct deferred_flush_entry *entry =
                                                &flush_table->entries[j];
-                       struct iova *iova = entry->iova;
+                       unsigned long iova_pfn = entry->iova_pfn;
                        unsigned long nrpages = entry->nrpages;
                        struct dmar_domain *domain = entry->domain;
                        struct page *freelist = entry->freelist;
@@ -3584,14 +3584,14 @@ static void flush_unmaps(struct deferred_flush_data *flush_data)
                        /* On real hardware multiple invalidations are expensive */
                        if (cap_caching_mode(iommu->cap))
                                iommu_flush_iotlb_psi(iommu, domain,
-                                       mm_to_dma_pfn(iova->pfn_lo),
+                                       mm_to_dma_pfn(iova_pfn),
                                        nrpages, !freelist, 0);
                        else {
                                mask = ilog2(nrpages);
                                iommu_flush_dev_iotlb(domain,
-                                               (uint64_t)iova->pfn_lo << PAGE_SHIFT, mask);
+                                               (uint64_t)iova_pfn << PAGE_SHIFT, mask);
                        }
-                       __free_iova(&domain->iovad, iova);
+                       free_iova(&domain->iovad, iova_pfn);
                        if (freelist)
                                dma_free_pagelist(freelist);
                }
@@ -3611,7 +3611,7 @@ static void flush_unmaps_timeout(unsigned long cpuid)
        spin_unlock_irqrestore(&flush_data->lock, flags);
 }
 
-static void add_unmap(struct dmar_domain *dom, struct iova *iova,
+static void add_unmap(struct dmar_domain *dom, unsigned long iova_pfn,
                      unsigned long nrpages, struct page *freelist)
 {
        unsigned long flags;
@@ -3645,7 +3645,7 @@ static void add_unmap(struct dmar_domain *dom, struct iova *iova,
 
        entry = &flush_data->tables[iommu_id].entries[entry_id];
        entry->domain = dom;
-       entry->iova = iova;
+       entry->iova_pfn = iova_pfn;
        entry->nrpages = nrpages;
        entry->freelist = freelist;
 
@@ -3664,7 +3664,7 @@ static void intel_unmap(struct device *dev, dma_addr_t dev_addr, size_t size)
        struct dmar_domain *domain;
        unsigned long start_pfn, last_pfn;
        unsigned long nrpages;
-       struct iova *iova;
+       unsigned long iova_pfn;
        struct intel_iommu *iommu;
        struct page *freelist;
 
@@ -3676,13 +3676,10 @@ static void intel_unmap(struct device *dev, dma_addr_t dev_addr, size_t size)
 
        iommu = domain_get_iommu(domain);
 
-       iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
-       if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
-                     (unsigned long long)dev_addr))
-               return;
+       iova_pfn = IOVA_PFN(dev_addr);
 
        nrpages = aligned_nrpages(dev_addr, size);
-       start_pfn = mm_to_dma_pfn(iova->pfn_lo);
+       start_pfn = mm_to_dma_pfn(iova_pfn);
        last_pfn = start_pfn + nrpages - 1;
 
        pr_debug("Device %s unmapping: pfn %lx-%lx\n",
@@ -3694,10 +3691,10 @@ static void intel_unmap(struct device *dev, dma_addr_t dev_addr, size_t size)
                iommu_flush_iotlb_psi(iommu, domain, start_pfn,
                                      nrpages, !freelist, 0);
                /* free iova */
-               __free_iova(&domain->iovad, iova);
+               free_iova(&domain->iovad, iova_pfn);
                dma_free_pagelist(freelist);
        } else {
-               add_unmap(domain, iova, nrpages, freelist);
+               add_unmap(domain, iova_pfn, nrpages, freelist);
                /*
                 * queue up the release of the unmap to save the 1/6th of the
                 * cpu used up by the iotlb flush operation...
@@ -3810,7 +3807,7 @@ static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nele
        struct dmar_domain *domain;
        size_t size = 0;
        int prot = 0;
-       struct iova *iova = NULL;
+       unsigned long iova_pfn;
        int ret;
        struct scatterlist *sg;
        unsigned long start_vpfn;
@@ -3829,9 +3826,9 @@ static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nele
        for_each_sg(sglist, sg, nelems, i)
                size += aligned_nrpages(sg->offset, sg->length);
 
-       iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
+       iova_pfn = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
                                *dev->dma_mask);
-       if (!iova) {
+       if (!iova_pfn) {
                sglist->dma_length = 0;
                return 0;
        }
@@ -3846,13 +3843,13 @@ static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nele
        if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
                prot |= DMA_PTE_WRITE;
 
-       start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
+       start_vpfn = mm_to_dma_pfn(iova_pfn);
 
        ret = domain_sg_mapping(domain, start_vpfn, sglist, size, prot);
        if (unlikely(ret)) {
                dma_pte_free_pagetable(domain, start_vpfn,
                                       start_vpfn + size - 1);
-               __free_iova(&domain->iovad, iova);
+               free_iova(&domain->iovad, iova_pfn);
                return 0;
        }