PCI: iommu: iotlb flushing

This patch is for batching up the flushing of the IOTLB for the DMAR
implementation found in the Intel VT-d hardware.  It works by building a list
of to be flushed IOTLB entries and a bitmap list of which DMAR engine they are
from.

After either a high water mark (250 accessible via debugfs) or 10ms the list
of iova's will be reclaimed and the DMAR engines associated are IOTLB-flushed.

This approach recovers 15 to 20% of the performance lost when using the IOMMU
for my netperf udp stream benchmark with small packets.  It can be disabled
with a kernel boot parameter "intel_iommu=strict".

Its use does weaken the IOMMU protections a bit.
Signed-off-by: Mark Gross <mgross@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

Documentation/kernel-parameters.txt

@@ -847,6 +847,10 @@ and is between 256 and 4096 characters. It is defined in the file
 			than 32 bit addressing. The default is to look
 			for translation below 32 bit and if not available
 			then look in the higher range.
+		strict [Default Off]
+			With this option on every unmap_single operation will
+			result in a hardware IOTLB flush operation as opposed
+			to batching them for performance.
 
 	io_delay=	[X86-32,X86-64] I/O delay method
 		0x80

drivers/pci/intel-iommu.c

@@ -22,6 +22,7 @@
 
 #include <linux/init.h>
 #include <linux/bitmap.h>
+#include <linux/debugfs.h>
 #include <linux/slab.h>
 #include <linux/irq.h>
 #include <linux/interrupt.h>
@@ -31,6 +32,7 @@
 #include <linux/dmar.h>
 #include <linux/dma-mapping.h>
 #include <linux/mempool.h>
+#include <linux/timer.h>
 #include "iova.h"
 #include "intel-iommu.h"
 #include <asm/proto.h> /* force_iommu in this header in x86-64*/
@@ -51,11 +53,32 @@
 
 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
 
+
+static void flush_unmaps_timeout(unsigned long data);
+
+DEFINE_TIMER(unmap_timer,  flush_unmaps_timeout, 0, 0);
+
+static struct intel_iommu *g_iommus;
+/* bitmap for indexing intel_iommus */
+static unsigned long 	*g_iommus_to_flush;
+static int g_num_of_iommus;
+
+static DEFINE_SPINLOCK(async_umap_flush_lock);
+static LIST_HEAD(unmaps_to_do);
+
+static int timer_on;
+static long list_size;
+static int high_watermark;
+
+static struct dentry *intel_iommu_debug, *debug;
+
+
 static void domain_remove_dev_info(struct dmar_domain *domain);
 
 static int dmar_disabled;
 static int __initdata dmar_map_gfx = 1;
 static int dmar_forcedac;
+static int intel_iommu_strict;
 
 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
 static DEFINE_SPINLOCK(device_domain_lock);
@@ -74,9 +97,13 @@ static int __init intel_iommu_setup(char *str)
 			printk(KERN_INFO
 				"Intel-IOMMU: disable GFX device mapping\n");
 		} else if (!strncmp(str, "forcedac", 8)) {
-			printk (KERN_INFO
+			printk(KERN_INFO
 				"Intel-IOMMU: Forcing DAC for PCI devices\n");
 			dmar_forcedac = 1;
+		} else if (!strncmp(str, "strict", 6)) {
+			printk(KERN_INFO
+				"Intel-IOMMU: disable batched IOTLB flush\n");
+			intel_iommu_strict = 1;
 		}
 
 		str += strcspn(str, ",");
@@ -966,17 +993,13 @@ static int iommu_init_domains(struct intel_iommu *iommu)
 		set_bit(0, iommu->domain_ids);
 	return 0;
 }
-
-static struct intel_iommu *alloc_iommu(struct dmar_drhd_unit *drhd)
+static struct intel_iommu *alloc_iommu(struct intel_iommu *iommu,
+					struct dmar_drhd_unit *drhd)
 {
-	struct intel_iommu *iommu;
 	int ret;
 	int map_size;
 	u32 ver;
 
-	iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
-	if (!iommu)
-		return NULL;
 	iommu->reg = ioremap(drhd->reg_base_addr, PAGE_SIZE_4K);
 	if (!iommu->reg) {
 		printk(KERN_ERR "IOMMU: can't map the region\n");
@@ -1404,7 +1427,7 @@ static int dmar_pci_device_match(struct pci_dev *devices[], int cnt,
 	int index;
 
 	while (dev) {
-		for (index = 0; index < cnt; index ++)
+		for (index = 0; index < cnt; index++)
 			if (dev == devices[index])
 				return 1;
 
@@ -1669,7 +1692,7 @@ int __init init_dmars(void)
 	struct dmar_rmrr_unit *rmrr;
 	struct pci_dev *pdev;
 	struct intel_iommu *iommu;
-	int ret, unit = 0;
+	int nlongs, i, ret, unit = 0;
 
 	/*
 	 * for each drhd
@@ -1680,7 +1703,35 @@ int __init init_dmars(void)
 	for_each_drhd_unit(drhd) {
 		if (drhd->ignored)
 			continue;
-		iommu = alloc_iommu(drhd);
+		g_num_of_iommus++;
+		/*
+		 * lock not needed as this is only incremented in the single
+		 * threaded kernel __init code path all other access are read
+		 * only
+		 */
+	}
+
+	nlongs = BITS_TO_LONGS(g_num_of_iommus);
+	g_iommus_to_flush = kzalloc(nlongs * sizeof(unsigned long), GFP_KERNEL);
+	if (!g_iommus_to_flush) {
+		printk(KERN_ERR "Intel-IOMMU: "
+			"Allocating bitmap array failed\n");
+		return -ENOMEM;
+	}
+
+	g_iommus = kzalloc(g_num_of_iommus * sizeof(*iommu), GFP_KERNEL);
+	if (!g_iommus) {
+		kfree(g_iommus_to_flush);
+		ret = -ENOMEM;
+		goto error;
+	}
+
+	i = 0;
+	for_each_drhd_unit(drhd) {
+		if (drhd->ignored)
+			continue;
+		iommu = alloc_iommu(&g_iommus[i], drhd);
+		i++;
 		if (!iommu) {
 			ret = -ENOMEM;
 			goto error;
@@ -1713,7 +1764,6 @@ int __init init_dmars(void)
 	 * endfor
 	 */
 	for_each_rmrr_units(rmrr) {
-		int i;
 		for (i = 0; i < rmrr->devices_cnt; i++) {
 			pdev = rmrr->devices[i];
 			/* some BIOS lists non-exist devices in DMAR table */
@@ -1769,6 +1819,7 @@ int __init init_dmars(void)
 		iommu = drhd->iommu;
 		free_iommu(iommu);
 	}
+	kfree(g_iommus);
 	return ret;
 }
 
@@ -1917,6 +1968,53 @@ static dma_addr_t intel_map_single(struct device *hwdev, void *addr,
 	return 0;
 }
 
+static void flush_unmaps(void)
+{
+	struct iova *node, *n;
+	unsigned long flags;
+	int i;
+
+	spin_lock_irqsave(&async_umap_flush_lock, flags);
+	timer_on = 0;
+
+	/* just flush them all */
+	for (i = 0; i < g_num_of_iommus; i++) {
+		if (test_and_clear_bit(i, g_iommus_to_flush))
+			iommu_flush_iotlb_global(&g_iommus[i], 0);
+	}
+
+	list_for_each_entry_safe(node, n, &unmaps_to_do, list) {
+		/* free iova */
+		list_del(&node->list);
+		__free_iova(&((struct dmar_domain *)node->dmar)->iovad, node);
+
+	}
+	list_size = 0;
+	spin_unlock_irqrestore(&async_umap_flush_lock, flags);
+}
+
+static void flush_unmaps_timeout(unsigned long data)
+{
+	flush_unmaps();
+}
+
+static void add_unmap(struct dmar_domain *dom, struct iova *iova)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&async_umap_flush_lock, flags);
+	iova->dmar = dom;
+	list_add(&iova->list, &unmaps_to_do);
+	set_bit((dom->iommu - g_iommus), g_iommus_to_flush);
+
+	if (!timer_on) {
+		mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
+		timer_on = 1;
+	}
+	list_size++;
+	spin_unlock_irqrestore(&async_umap_flush_lock, flags);
+}
+
 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr,
 	size_t size, int dir)
 {
@@ -1944,13 +2042,21 @@ static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr,
 	dma_pte_clear_range(domain, start_addr, start_addr + size);
 	/* free page tables */
 	dma_pte_free_pagetable(domain, start_addr, start_addr + size);
-
-	if (iommu_flush_iotlb_psi(domain->iommu, domain->id, start_addr,
-			size >> PAGE_SHIFT_4K, 0))
-		iommu_flush_write_buffer(domain->iommu);
-
-	/* free iova */
-	__free_iova(&domain->iovad, iova);
+	if (intel_iommu_strict) {
+		if (iommu_flush_iotlb_psi(domain->iommu,
+			domain->id, start_addr, size >> PAGE_SHIFT_4K, 0))
+			iommu_flush_write_buffer(domain->iommu);
+		/* free iova */
+		__free_iova(&domain->iovad, iova);
+	} else {
+		add_unmap(domain, iova);
+		/*
+		 * queue up the release of the unmap to save the 1/6th of the
+		 * cpu used up by the iotlb flush operation...
+		 */
+		if (list_size > high_watermark)
+			flush_unmaps();
+	}
 }
 
 static void * intel_alloc_coherent(struct device *hwdev, size_t size,
@@ -2274,6 +2380,10 @@ int __init intel_iommu_init(void)
 	if (dmar_table_init())
 		return 	-ENODEV;
 
+	high_watermark = 250;
+	intel_iommu_debug = debugfs_create_dir("intel_iommu", NULL);
+	debug = debugfs_create_u32("high_watermark", S_IWUGO | S_IRUGO,
+					intel_iommu_debug, &high_watermark);
 	iommu_init_mempool();
 	dmar_init_reserved_ranges();
 
@@ -2289,6 +2399,7 @@ int __init intel_iommu_init(void)
 	printk(KERN_INFO
 	"PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
 
+	init_timer(&unmap_timer);
 	force_iommu = 1;
 	dma_ops = &intel_dma_ops;
 	return 0;

drivers/pci/iova.h

@@ -24,6 +24,8 @@ struct iova {
 	struct rb_node	node;
 	unsigned long	pfn_hi; /* IOMMU dish out addr hi */
 	unsigned long	pfn_lo; /* IOMMU dish out addr lo */
+	struct list_head list;
+	void *dmar;
 };
 
 /* holds all the iova translations for a domain */