1 files changed, 294 insertions, 170 deletions

diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
index 63cacf5d6cf2..83a55df17dfe 100644
--- a/drivers/iommu/amd_iommu.c
+++ b/drivers/iommu/amd_iommu.c
@@ -89,25 +89,6 @@ LIST_HEAD(ioapic_map);
 LIST_HEAD(hpet_map);
 LIST_HEAD(acpihid_map);
 
-#define FLUSH_QUEUE_SIZE 256
-
-struct flush_queue_entry {
-	unsigned long iova_pfn;
-	unsigned long pages;
-	struct dma_ops_domain *dma_dom;
-};
-
-struct flush_queue {
-	spinlock_t lock;
-	unsigned next;
-	struct flush_queue_entry *entries;
-};
-
-static DEFINE_PER_CPU(struct flush_queue, flush_queue);
-
-static atomic_t queue_timer_on;
-static struct timer_list queue_timer;
-
 /*
  * Domain for untranslated devices - only allocated
  * if iommu=pt passed on kernel cmd line.
@@ -138,6 +119,8 @@ struct iommu_dev_data {
 					     PPR completions */
 	u32 errata;			  /* Bitmap for errata to apply */
 	bool use_vapic;			  /* Enable device to use vapic mode */
+
+	struct ratelimit_state rs;	  /* Ratelimit IOPF messages */
 };
 
 /*
@@ -153,6 +136,20 @@ static void update_domain(struct protection_domain *domain);
 static int protection_domain_init(struct protection_domain *domain);
 static void detach_device(struct device *dev);
 
+#define FLUSH_QUEUE_SIZE 256
+
+struct flush_queue_entry {
+	unsigned long iova_pfn;
+	unsigned long pages;
+	u64 counter; /* Flush counter when this entry was added to the queue */
+};
+
+struct flush_queue {
+	struct flush_queue_entry *entries;
+	unsigned head, tail;
+	spinlock_t lock;
+};
+
 /*
  * Data container for a dma_ops specific protection domain
  */
@@ -162,6 +159,36 @@ struct dma_ops_domain {
 
 	/* IOVA RB-Tree */
 	struct iova_domain iovad;
+
+	struct flush_queue __percpu *flush_queue;
+
+	/*
+	 * We need two counter here to be race-free wrt. IOTLB flushing and
+	 * adding entries to the flush queue.
+	 *
+	 * The flush_start_cnt is incremented _before_ the IOTLB flush starts.
+	 * New entries added to the flush ring-buffer get their 'counter' value
+	 * from here. This way we can make sure that entries added to the queue
+	 * (or other per-cpu queues of the same domain) while the TLB is about
+	 * to be flushed are not considered to be flushed already.
+	 */
+	atomic64_t flush_start_cnt;
+
+	/*
+	 * The flush_finish_cnt is incremented when an IOTLB flush is complete.
+	 * This value is always smaller than flush_start_cnt. The queue_add
+	 * function frees all IOVAs that have a counter value smaller than
+	 * flush_finish_cnt. This makes sure that we only free IOVAs that are
+	 * flushed out of the IOTLB of the domain.
+	 */
+	atomic64_t flush_finish_cnt;
+
+	/*
+	 * Timer to make sure we don't keep IOVAs around unflushed
+	 * for too long
+	 */
+	struct timer_list flush_timer;
+	atomic_t flush_timer_on;
 };
 
 static struct iova_domain reserved_iova_ranges;
@@ -253,6 +280,8 @@ static struct iommu_dev_data *alloc_dev_data(u16 devid)
 	list_add_tail(&dev_data->dev_data_list, &dev_data_list);
 	spin_unlock_irqrestore(&dev_data_list_lock, flags);
 
+	ratelimit_default_init(&dev_data->rs);
+
 	return dev_data;
 }
 
@@ -551,6 +580,29 @@ static void dump_command(unsigned long phys_addr)
 		pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
 }
 
+static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
+					u64 address, int flags)
+{
+	struct iommu_dev_data *dev_data = NULL;
+	struct pci_dev *pdev;
+
+	pdev = pci_get_bus_and_slot(PCI_BUS_NUM(devid), devid & 0xff);
+	if (pdev)
+		dev_data = get_dev_data(&pdev->dev);
+
+	if (dev_data && __ratelimit(&dev_data->rs)) {
+		dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
+			domain_id, address, flags);
+	} else if (printk_ratelimit()) {
+		pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
+			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
+			domain_id, address, flags);
+	}
+
+	if (pdev)
+		pci_dev_put(pdev);
+}
+
 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
 {
 	int type, devid, domid, flags;
@@ -575,7 +627,12 @@ retry:
 		goto retry;
 	}
 
-	printk(KERN_ERR "AMD-Vi: Event logged [");
+	if (type == EVENT_TYPE_IO_FAULT) {
+		amd_iommu_report_page_fault(devid, domid, address, flags);
+		return;
+	} else {
+		printk(KERN_ERR "AMD-Vi: Event logged [");
+	}
 
 	switch (type) {
 	case EVENT_TYPE_ILL_DEV:
@@ -585,12 +642,6 @@ retry:
 		       address, flags);
 		dump_dte_entry(devid);
 		break;
-	case EVENT_TYPE_IO_FAULT:
-		printk("IO_PAGE_FAULT device=%02x:%02x.%x "
-		       "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
-		       PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
-		       domid, address, flags);
-		break;
 	case EVENT_TYPE_DEV_TAB_ERR:
 		printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
 		       "address=0x%016llx flags=0x%04x]\n",
@@ -848,19 +899,20 @@ static int wait_on_sem(volatile u64 *sem)
 }
 
 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
-			       struct iommu_cmd *cmd,
-			       u32 tail)
+			       struct iommu_cmd *cmd)
 {
 	u8 *target;
 
-	target = iommu->cmd_buf + tail;
-	tail   = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
+	target = iommu->cmd_buf + iommu->cmd_buf_tail;
+
+	iommu->cmd_buf_tail += sizeof(*cmd);
+	iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
 
 	/* Copy command to buffer */
 	memcpy(target, cmd, sizeof(*cmd));
 
 	/* Tell the IOMMU about it */
-	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
+	writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
 }
 
 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
@@ -1018,33 +1070,34 @@ static int __iommu_queue_command_sync(struct amd_iommu *iommu,
 				      struct iommu_cmd *cmd,
 				      bool sync)
 {
-	u32 left, tail, head, next_tail;
+	unsigned int count = 0;
+	u32 left, next_tail;
 
+	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
 again:
-
-	head      = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
-	tail      = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
-	next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
-	left      = (head - next_tail) % CMD_BUFFER_SIZE;
+	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
 
 	if (left <= 0x20) {
-		struct iommu_cmd sync_cmd;
-		int ret;
-
-		iommu->cmd_sem = 0;
+		/* Skip udelay() the first time around */
+		if (count++) {
+			if (count == LOOP_TIMEOUT) {
+				pr_err("AMD-Vi: Command buffer timeout\n");
+				return -EIO;
+			}
 
-		build_completion_wait(&sync_cmd, (u64)&iommu->cmd_sem);
-		copy_cmd_to_buffer(iommu, &sync_cmd, tail);
+			udelay(1);
+		}
 
-		if ((ret = wait_on_sem(&iommu->cmd_sem)) != 0)
-			return ret;
+		/* Update head and recheck remaining space */
+		iommu->cmd_buf_head = readl(iommu->mmio_base +
+					    MMIO_CMD_HEAD_OFFSET);
 
 		goto again;
 	}
 
-	copy_cmd_to_buffer(iommu, cmd, tail);
+	copy_cmd_to_buffer(iommu, cmd);
 
-	/* We need to sync now to make sure all commands are processed */
+	/* Do we need to make sure all commands are processed? */
 	iommu->need_sync = sync;
 
 	return 0;
@@ -1733,6 +1786,180 @@ static void free_gcr3_table(struct protection_domain *domain)
 	free_page((unsigned long)domain->gcr3_tbl);
 }
 
+static void dma_ops_domain_free_flush_queue(struct dma_ops_domain *dom)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		struct flush_queue *queue;
+
+		queue = per_cpu_ptr(dom->flush_queue, cpu);
+		kfree(queue->entries);
+	}
+
+	free_percpu(dom->flush_queue);
+
+	dom->flush_queue = NULL;
+}
+
+static int dma_ops_domain_alloc_flush_queue(struct dma_ops_domain *dom)
+{
+	int cpu;
+
+	atomic64_set(&dom->flush_start_cnt,  0);
+	atomic64_set(&dom->flush_finish_cnt, 0);
+
+	dom->flush_queue = alloc_percpu(struct flush_queue);
+	if (!dom->flush_queue)
+		return -ENOMEM;
+
+	/* First make sure everything is cleared */
+	for_each_possible_cpu(cpu) {
+		struct flush_queue *queue;
+
+		queue = per_cpu_ptr(dom->flush_queue, cpu);
+		queue->head    = 0;
+		queue->tail    = 0;
+		queue->entries = NULL;
+	}
+
+	/* Now start doing the allocation */
+	for_each_possible_cpu(cpu) {
+		struct flush_queue *queue;
+
+		queue = per_cpu_ptr(dom->flush_queue, cpu);
+		queue->entries = kzalloc(FLUSH_QUEUE_SIZE * sizeof(*queue->entries),
+					 GFP_KERNEL);
+		if (!queue->entries) {
+			dma_ops_domain_free_flush_queue(dom);
+			return -ENOMEM;
+		}
+
+		spin_lock_init(&queue->lock);
+	}
+
+	return 0;
+}
+
+static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
+{
+	atomic64_inc(&dom->flush_start_cnt);
+	domain_flush_tlb(&dom->domain);
+	domain_flush_complete(&dom->domain);
+	atomic64_inc(&dom->flush_finish_cnt);
+}
+
+static inline bool queue_ring_full(struct flush_queue *queue)
+{
+	assert_spin_locked(&queue->lock);
+
+	return (((queue->tail + 1) % FLUSH_QUEUE_SIZE) == queue->head);
+}
+
+#define queue_ring_for_each(i, q) \
+	for (i = (q)->head; i != (q)->tail; i = (i + 1) % FLUSH_QUEUE_SIZE)
+
+static inline unsigned queue_ring_add(struct flush_queue *queue)
+{
+	unsigned idx = queue->tail;
+
+	assert_spin_locked(&queue->lock);
+	queue->tail = (idx + 1) % FLUSH_QUEUE_SIZE;
+
+	return idx;
+}
+
+static inline void queue_ring_remove_head(struct flush_queue *queue)
+{
+	assert_spin_locked(&queue->lock);
+	queue->head = (queue->head + 1) % FLUSH_QUEUE_SIZE;
+}
+
+static void queue_ring_free_flushed(struct dma_ops_domain *dom,
+				    struct flush_queue *queue)
+{
+	u64 counter = atomic64_read(&dom->flush_finish_cnt);
+	int idx;
+
+	queue_ring_for_each(idx, queue) {
+		/*
+		 * This assumes that counter values in the ring-buffer are
+		 * monotonously rising.
+		 */
+		if (queue->entries[idx].counter >= counter)
+			break;
+
+		free_iova_fast(&dom->iovad,
+			       queue->entries[idx].iova_pfn,
+			       queue->entries[idx].pages);
+
+		queue_ring_remove_head(queue);
+	}
+}
+
+static void queue_add(struct dma_ops_domain *dom,
+		      unsigned long address, unsigned long pages)
+{
+	struct flush_queue *queue;
+	unsigned long flags;
+	int idx;
+
+	pages     = __roundup_pow_of_two(pages);
+	address >>= PAGE_SHIFT;
+
+	queue = get_cpu_ptr(dom->flush_queue);
+	spin_lock_irqsave(&queue->lock, flags);
+
+	/*
+	 * First remove the enries from the ring-buffer that are already
+	 * flushed to make the below queue_ring_full() check less likely
+	 */
+	queue_ring_free_flushed(dom, queue);
+
+	/*
+	 * When ring-queue is full, flush the entries from the IOTLB so
+	 * that we can free all entries with queue_ring_free_flushed()
+	 * below.
+	 */
+	if (queue_ring_full(queue)) {
+		dma_ops_domain_flush_tlb(dom);
+		queue_ring_free_flushed(dom, queue);
+	}
+
+	idx = queue_ring_add(queue);
+
+	queue->entries[idx].iova_pfn = address;
+	queue->entries[idx].pages    = pages;
+	queue->entries[idx].counter  = atomic64_read(&dom->flush_start_cnt);
+
+	spin_unlock_irqrestore(&queue->lock, flags);
+
+	if (atomic_cmpxchg(&dom->flush_timer_on, 0, 1) == 0)
+		mod_timer(&dom->flush_timer, jiffies + msecs_to_jiffies(10));
+
+	put_cpu_ptr(dom->flush_queue);
+}
+
+static void queue_flush_timeout(unsigned long data)
+{
+	struct dma_ops_domain *dom = (struct dma_ops_domain *)data;
+	int cpu;
+
+	atomic_set(&dom->flush_timer_on, 0);
+
+	dma_ops_domain_flush_tlb(dom);
+
+	for_each_possible_cpu(cpu) {
+		struct flush_queue *queue;
+		unsigned long flags;
+
+		queue = per_cpu_ptr(dom->flush_queue, cpu);
+		spin_lock_irqsave(&queue->lock, flags);
+		queue_ring_free_flushed(dom, queue);
+		spin_unlock_irqrestore(&queue->lock, flags);
+	}
+}
+
 /*
  * Free a domain, only used if something went wrong in the
  * allocation path and we need to free an already allocated page table
@@ -1744,6 +1971,11 @@ static void dma_ops_domain_free(struct dma_ops_domain *dom)
 
 	del_domain_from_list(&dom->domain);
 
+	if (timer_pending(&dom->flush_timer))
+		del_timer(&dom->flush_timer);
+
+	dma_ops_domain_free_flush_queue(dom);
+
 	put_iova_domain(&dom->iovad);
 
 	free_pagetable(&dom->domain);
@@ -1782,6 +2014,14 @@ static struct dma_ops_domain *dma_ops_domain_alloc(void)
 	/* Initialize reserved ranges */
 	copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
 
+	if (dma_ops_domain_alloc_flush_queue(dma_dom))
+		goto free_dma_dom;
+
+	setup_timer(&dma_dom->flush_timer, queue_flush_timeout,
+		    (unsigned long)dma_dom);
+
+	atomic_set(&dma_dom->flush_timer_on, 0);
+
 	add_domain_to_list(&dma_dom->domain);
 
 	return dma_dom;
@@ -1844,7 +2084,8 @@ static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
 		flags    |= tmp;
 	}
 
-	flags &= ~(0xffffUL);
+
+	flags &= ~(DTE_FLAG_SA | 0xffffULL);
 	flags |= domain->id;
 
 	amd_iommu_dev_table[devid].data[1]  = flags;
@@ -2225,92 +2466,6 @@ static struct iommu_group *amd_iommu_device_group(struct device *dev)
  *
  *****************************************************************************/
 
-static void __queue_flush(struct flush_queue *queue)
-{
-	struct protection_domain *domain;
-	unsigned long flags;
-	int idx;
-
-	/* First flush TLB of all known domains */
-	spin_lock_irqsave(&amd_iommu_pd_lock, flags);
-	list_for_each_entry(domain, &amd_iommu_pd_list, list)
-		domain_flush_tlb(domain);
-	spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
-
-	/* Wait until flushes have completed */
-	domain_flush_complete(NULL);
-
-	for (idx = 0; idx < queue->next; ++idx) {
-		struct flush_queue_entry *entry;
-
-		entry = queue->entries + idx;
-
-		free_iova_fast(&entry->dma_dom->iovad,
-				entry->iova_pfn,
-				entry->pages);
-
-		/* Not really necessary, just to make sure we catch any bugs */
-		entry->dma_dom = NULL;
-	}
-
-	queue->next = 0;
-}
-
-static void queue_flush_all(void)
-{
-	int cpu;
-
-	for_each_possible_cpu(cpu) {
-		struct flush_queue *queue;
-		unsigned long flags;
-
-		queue = per_cpu_ptr(&flush_queue, cpu);
-		spin_lock_irqsave(&queue->lock, flags);
-		if (queue->next > 0)
-			__queue_flush(queue);
-		spin_unlock_irqrestore(&queue->lock, flags);
-	}
-}
-
-static void queue_flush_timeout(unsigned long unsused)
-{
-	atomic_set(&queue_timer_on, 0);
-	queue_flush_all();
-}
-
-static void queue_add(struct dma_ops_domain *dma_dom,
-		      unsigned long address, unsigned long pages)
-{
-	struct flush_queue_entry *entry;
-	struct flush_queue *queue;
-	unsigned long flags;
-	int idx;
-
-	pages     = __roundup_pow_of_two(pages);
-	address >>= PAGE_SHIFT;
-
-	queue = get_cpu_ptr(&flush_queue);
-	spin_lock_irqsave(&queue->lock, flags);
-
-	if (queue->next == FLUSH_QUEUE_SIZE)
-		__queue_flush(queue);
-
-	idx   = queue->next++;
-	entry = queue->entries + idx;
-
-	entry->iova_pfn = address;
-	entry->pages    = pages;
-	entry->dma_dom  = dma_dom;
-
-	spin_unlock_irqrestore(&queue->lock, flags);
-
-	if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
-		mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
-
-	put_cpu_ptr(&flush_queue);
-}
-
-
 /*
  * In the dma_ops path we only have the struct device. This function
  * finds the corresponding IOMMU, the protection domain and the
@@ -2797,7 +2952,7 @@ static int init_reserved_iova_ranges(void)
 
 int __init amd_iommu_init_api(void)
 {
-	int ret, cpu, err = 0;
+	int ret, err = 0;
 
 	ret = iova_cache_get();
 	if (ret)
@@ -2807,18 +2962,6 @@ int __init amd_iommu_init_api(void)
 	if (ret)
 		return ret;
 
-	for_each_possible_cpu(cpu) {
-		struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
-
-		queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
-					 sizeof(*queue->entries),
-					 GFP_KERNEL);
-		if (!queue->entries)
-			goto out_put_iova;
-
-		spin_lock_init(&queue->lock);
-	}
-
 	err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
 	if (err)
 		return err;
@@ -2830,23 +2973,12 @@ int __init amd_iommu_init_api(void)
 	err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
 	if (err)
 		return err;
-	return 0;
-
-out_put_iova:
-	for_each_possible_cpu(cpu) {
-		struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
-
-		kfree(queue->entries);
-	}
 
-	return -ENOMEM;
+	return 0;
 }
 
 int __init amd_iommu_init_dma_ops(void)
 {
-	setup_timer(&queue_timer, queue_flush_timeout, 0);
-	atomic_set(&queue_timer_on, 0);
-
 	swiotlb        = iommu_pass_through ? 1 : 0;
 	iommu_detected = 1;
 
@@ -3002,12 +3134,6 @@ static void amd_iommu_domain_free(struct iommu_domain *dom)
 
 	switch (dom->type) {
 	case IOMMU_DOMAIN_DMA:
-		/*
-		 * First make sure the domain is no longer referenced from the
-		 * flush queue
-		 */
-		queue_flush_all();
-
 		/* Now release the domain */
 		dma_dom = to_dma_ops_domain(domain);
 		dma_ops_domain_free(dma_dom);
@@ -3879,11 +4005,9 @@ static void irte_ga_prepare(void *entry,
 			    u8 vector, u32 dest_apicid, int devid)
 {
 	struct irte_ga *irte = (struct irte_ga *) entry;
-	struct iommu_dev_data *dev_data = search_dev_data(devid);
 
 	irte->lo.val                      = 0;
 	irte->hi.val                      = 0;
-	irte->lo.fields_remap.guest_mode  = dev_data ? dev_data->use_vapic : 0;
 	irte->lo.fields_remap.int_type    = delivery_mode;
 	irte->lo.fields_remap.dm          = dest_mode;
 	irte->hi.fields.vector            = vector;
@@ -3939,10 +4063,10 @@ static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
 	struct irte_ga *irte = (struct irte_ga *) entry;
 	struct iommu_dev_data *dev_data = search_dev_data(devid);
 
-	if (!dev_data || !dev_data->use_vapic) {
+	if (!dev_data || !dev_data->use_vapic ||
+	    !irte->lo.fields_remap.guest_mode) {
 		irte->hi.fields.vector = vector;
 		irte->lo.fields_remap.destination = dest_apicid;
-		irte->lo.fields_remap.guest_mode = 0;
 		modify_irte_ga(devid, index, irte, NULL);
 	}
 }
@@ -4273,7 +4397,7 @@ static void irq_remapping_deactivate(struct irq_domain *domain,
 					    irte_info->index);
 }
 
-static struct irq_domain_ops amd_ir_domain_ops = {
+static const struct irq_domain_ops amd_ir_domain_ops = {
 	.alloc = irq_remapping_alloc,
 	.free = irq_remapping_free,
 	.activate = irq_remapping_activate,