iommu/amd: Use only natural aligned flushes in a VM

When running on an AMD vIOMMU, it is better to avoid TLB flushes
of unmodified PTEs. vIOMMUs require the hypervisor to synchronize the
virtualized IOMMU's PTEs with the physical ones. This process induce
overheads.

AMD IOMMU allows us to flush any range that is aligned to the power of
2. So when running on top of a vIOMMU, break the range into sub-ranges
that are naturally aligned, and flush each one separately. This apporach
is better when running with a vIOMMU, but on physical IOMMUs, the
penalty of IOTLB misses due to unnecessary flushed entries is likely to
be low.

Repurpose (i.e., keeping the name, changing the logic)
domain_flush_pages() so it is used to choose whether to perform one
flush of the whole range or multiple ones to avoid flushing unnecessary
ranges. Use NpCache, as usual, to infer whether the IOMMU is physical or
virtual.

Cc: Joerg Roedel <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Jiajun Cao <[email protected]>
Cc: Lu Baolu <[email protected]>
Cc: [email protected]
Cc: [email protected]
Suggested-by: Robin Murphy <[email protected]>
Signed-off-by: Nadav Amit <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Joerg Roedel <[email protected]>

Author: anadav

drivers/iommu/amd/iommu.c

@@ -1267,15 +1267,52 @@ static void __domain_flush_pages(struct protection_domain *domain,
 }
 
 static void domain_flush_pages(struct protection_domain *domain,
-			       u64 address, size_t size)
+			       u64 address, size_t size, int pde)
 {
-	__domain_flush_pages(domain, address, size, 0);
+	if (likely(!amd_iommu_np_cache)) {
+		__domain_flush_pages(domain, address, size, pde);
+		return;
+	}
+
+	/*
+	 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
+	 * In such setups it is best to avoid flushes of ranges which are not
+	 * naturally aligned, since it would lead to flushes of unmodified
+	 * PTEs. Such flushes would require the hypervisor to do more work than
+	 * necessary. Therefore, perform repeated flushes of aligned ranges
+	 * until you cover the range. Each iteration flushes the smaller
+	 * between the natural alignment of the address that we flush and the
+	 * greatest naturally aligned region that fits in the range.
+	 */
+	while (size != 0) {
+		int addr_alignment = __ffs(address);
+		int size_alignment = __fls(size);
+		int min_alignment;
+		size_t flush_size;
+
+		/*
+		 * size is always non-zero, but address might be zero, causing
+		 * addr_alignment to be negative. As the casting of the
+		 * argument in __ffs(address) to long might trim the high bits
+		 * of the address on x86-32, cast to long when doing the check.
+		 */
+		if (likely((unsigned long)address != 0))
+			min_alignment = min(addr_alignment, size_alignment);
+		else
+			min_alignment = size_alignment;
+
+		flush_size = 1ul << min_alignment;
+
+		__domain_flush_pages(domain, address, flush_size, pde);
+		address += flush_size;
+		size -= flush_size;
+	}
 }
 
 /* Flush the whole IO/TLB for a given protection domain - including PDE */
 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
 {
-	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
+	domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
 }
 
 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
@@ -1302,7 +1339,7 @@ static void domain_flush_np_cache(struct protection_domain *domain,
 		unsigned long flags;
 
 		spin_lock_irqsave(&domain->lock, flags);
-		domain_flush_pages(domain, iova, size);
+		domain_flush_pages(domain, iova, size, 1);
 		amd_iommu_domain_flush_complete(domain);
 		spin_unlock_irqrestore(&domain->lock, flags);
 	}
@@ -2206,7 +2243,7 @@ static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
 	unsigned long flags;
 
 	spin_lock_irqsave(&dom->lock, flags);
-	__domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
+	domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
 	amd_iommu_domain_flush_complete(dom);
 	spin_unlock_irqrestore(&dom->lock, flags);
 }