[PATCH v6 04/12] VT-d: allow use of superpage mappings

[Jan Beulich <jbeulich@xxxxxxxx>]

... depending on feature availability (and absence of quirks).

Also make the page table dumping function aware of superpages.

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
Reviewed-by: Kevin Tian <kevin.tian@xxxxxxxxx>
Reviewed-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
---
v6: Re-base over addition of "iommu=no-superpages" command line option.
v5: In intel_iommu_{,un}map_page() assert page order is supported.
v4: Change type of queue_free_pt()'s 1st parameter. Re-base.
v3: Rename queue_free_pt()'s last parameter. Replace "level > 1" checks
    where possible. Tighten assertion.

--- a/xen/drivers/passthrough/vtd/iommu.c
+++ b/xen/drivers/passthrough/vtd/iommu.c
@@ -779,18 +779,37 @@ static int __must_check cf_check iommu_f
     return ret;
 }
 
+static void queue_free_pt(struct domain_iommu *hd, mfn_t mfn, unsigned int 
level)
+{
+    if ( level > 1 )
+    {
+        struct dma_pte *pt = map_domain_page(mfn);
+        unsigned int i;
+
+        for ( i = 0; i < PTE_NUM; ++i )
+            if ( dma_pte_present(pt[i]) && !dma_pte_superpage(pt[i]) )
+                queue_free_pt(hd, maddr_to_mfn(dma_pte_addr(pt[i])),
+                              level - 1);
+
+        unmap_domain_page(pt);
+    }
+
+    iommu_queue_free_pgtable(hd, mfn_to_page(mfn));
+}
+
 /* clear one page's page table */
 static int dma_pte_clear_one(struct domain *domain, daddr_t addr,
                              unsigned int order,
                              unsigned int *flush_flags)
 {
     struct domain_iommu *hd = dom_iommu(domain);
-    struct dma_pte *page = NULL, *pte = NULL;
+    struct dma_pte *page = NULL, *pte = NULL, old;
     u64 pg_maddr;
+    unsigned int level = (order / LEVEL_STRIDE) + 1;
 
     spin_lock(&hd->arch.mapping_lock);
-    /* get last level pte */
-    pg_maddr = addr_to_dma_page_maddr(domain, addr, 1, flush_flags, false);
+    /* get target level pte */
+    pg_maddr = addr_to_dma_page_maddr(domain, addr, level, flush_flags, false);
     if ( pg_maddr < PAGE_SIZE )
     {
         spin_unlock(&hd->arch.mapping_lock);
@@ -798,7 +817,7 @@ static int dma_pte_clear_one(struct doma
     }
 
     page = (struct dma_pte *)map_vtd_domain_page(pg_maddr);
-    pte = page + address_level_offset(addr, 1);
+    pte = &page[address_level_offset(addr, level)];
 
     if ( !dma_pte_present(*pte) )
     {
@@ -807,14 +826,20 @@ static int dma_pte_clear_one(struct doma
         return 0;
     }
 
+    old = *pte;
     dma_clear_pte(*pte);
-    *flush_flags |= IOMMU_FLUSHF_modified;
 
     spin_unlock(&hd->arch.mapping_lock);
     iommu_sync_cache(pte, sizeof(struct dma_pte));
 
     unmap_vtd_domain_page(page);
 
+    *flush_flags |= IOMMU_FLUSHF_modified;
+
+    if ( order && !dma_pte_superpage(old) )
+        queue_free_pt(hd, maddr_to_mfn(dma_pte_addr(old)),
+                      order / LEVEL_STRIDE);
+
     return 0;
 }
 
@@ -2092,8 +2117,12 @@ static int __must_check cf_check intel_i
     struct domain_iommu *hd = dom_iommu(d);
     struct dma_pte *page, *pte, old, new = {};
     u64 pg_maddr;
+    unsigned int level = (IOMMUF_order(flags) / LEVEL_STRIDE) + 1;
     int rc = 0;
 
+    ASSERT((hd->platform_ops->page_sizes >> IOMMUF_order(flags)) &
+           PAGE_SIZE_4K);
+
     /* Do nothing if VT-d shares EPT page table */
     if ( iommu_use_hap_pt(d) )
         return 0;
@@ -2116,7 +2145,7 @@ static int __must_check cf_check intel_i
         return 0;
     }
 
-    pg_maddr = addr_to_dma_page_maddr(d, dfn_to_daddr(dfn), 1, flush_flags,
+    pg_maddr = addr_to_dma_page_maddr(d, dfn_to_daddr(dfn), level, flush_flags,
                                       true);
     if ( pg_maddr < PAGE_SIZE )
     {
@@ -2125,13 +2154,15 @@ static int __must_check cf_check intel_i
     }
 
     page = (struct dma_pte *)map_vtd_domain_page(pg_maddr);
-    pte = &page[dfn_x(dfn) & LEVEL_MASK];
+    pte = &page[address_level_offset(dfn_to_daddr(dfn), level)];
     old = *pte;
 
     dma_set_pte_addr(new, mfn_to_maddr(mfn));
     dma_set_pte_prot(new,
                      ((flags & IOMMUF_readable) ? DMA_PTE_READ  : 0) |
                      ((flags & IOMMUF_writable) ? DMA_PTE_WRITE : 0));
+    if ( IOMMUF_order(flags) )
+        dma_set_pte_superpage(new);
 
     /* Set the SNP on leaf page table if Snoop Control available */
     if ( iommu_snoop )
@@ -2152,14 +2183,26 @@ static int __must_check cf_check intel_i
 
     *flush_flags |= IOMMU_FLUSHF_added;
     if ( dma_pte_present(old) )
+    {
         *flush_flags |= IOMMU_FLUSHF_modified;
 
+        if ( IOMMUF_order(flags) && !dma_pte_superpage(old) )
+            queue_free_pt(hd, maddr_to_mfn(dma_pte_addr(old)),
+                          IOMMUF_order(flags) / LEVEL_STRIDE);
+    }
+
     return rc;
 }
 
 static int __must_check cf_check intel_iommu_unmap_page(
     struct domain *d, dfn_t dfn, unsigned int order, unsigned int *flush_flags)
 {
+    /*
+     * While really we could unmap at any granularity, for now we assume unmaps
+     * are issued by common code only at the same granularity as maps.
+     */
+    ASSERT((dom_iommu(d)->platform_ops->page_sizes >> order) & PAGE_SIZE_4K);
+
     /* Do nothing if VT-d shares EPT page table */
     if ( iommu_use_hap_pt(d) )
         return 0;
@@ -2515,6 +2558,7 @@ static int __init cf_check vtd_setup(voi
 {
     struct acpi_drhd_unit *drhd;
     struct vtd_iommu *iommu;
+    unsigned int large_sizes = iommu_superpages ? PAGE_SIZE_2M | PAGE_SIZE_1G 
: 0;
     int ret;
     bool reg_inval_supported = true;
 
@@ -2557,6 +2601,11 @@ static int __init cf_check vtd_setup(voi
                cap_sps_2mb(iommu->cap) ? ", 2MB" : "",
                cap_sps_1gb(iommu->cap) ? ", 1GB" : "");
 
+        if ( !cap_sps_2mb(iommu->cap) )
+            large_sizes &= ~PAGE_SIZE_2M;
+        if ( !cap_sps_1gb(iommu->cap) )
+            large_sizes &= ~PAGE_SIZE_1G;
+
 #ifndef iommu_snoop
         if ( iommu_snoop && !ecap_snp_ctl(iommu->ecap) )
             iommu_snoop = false;
@@ -2628,6 +2677,9 @@ static int __init cf_check vtd_setup(voi
     if ( ret )
         goto error;
 
+    ASSERT(iommu_ops.page_sizes == PAGE_SIZE_4K);
+    iommu_ops.page_sizes |= large_sizes;
+
     register_keyhandler('V', vtd_dump_iommu_info, "dump iommu info", 1);
 
     return 0;
@@ -2960,7 +3012,7 @@ static void vtd_dump_page_table_level(pa
             continue;
 
         address = gpa + offset_level_address(i, level);
-        if ( next_level >= 1 ) 
+        if ( next_level && !dma_pte_superpage(*pte) )
             vtd_dump_page_table_level(dma_pte_addr(*pte), next_level,
                                       address, indent + 1);
         else