2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <asm/msidef.h>
35 #include <asm/proto.h>
36 #include <asm/iommu.h>
40 #include "amd_iommu_proto.h"
41 #include "amd_iommu_types.h"
43 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
45 #define LOOP_TIMEOUT 100000
48 * This bitmap is used to advertise the page sizes our hardware support
49 * to the IOMMU core, which will then use this information to split
50 * physically contiguous memory regions it is mapping into page sizes
53 * Traditionally the IOMMU core just handed us the mappings directly,
54 * after making sure the size is an order of a 4KiB page and that the
55 * mapping has natural alignment.
57 * To retain this behavior, we currently advertise that we support
58 * all page sizes that are an order of 4KiB.
60 * If at some point we'd like to utilize the IOMMU core's new behavior,
61 * we could change this to advertise the real page sizes we support.
63 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
65 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
67 /* A list of preallocated protection domains */
68 static LIST_HEAD(iommu_pd_list);
69 static DEFINE_SPINLOCK(iommu_pd_list_lock);
71 /* List of all available dev_data structures */
72 static LIST_HEAD(dev_data_list);
73 static DEFINE_SPINLOCK(dev_data_list_lock);
75 LIST_HEAD(ioapic_map);
79 * Domain for untranslated devices - only allocated
80 * if iommu=pt passed on kernel cmd line.
82 static struct protection_domain *pt_domain;
84 static struct iommu_ops amd_iommu_ops;
86 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
87 int amd_iommu_max_glx_val = -1;
89 static struct dma_map_ops amd_iommu_dma_ops;
92 * general struct to manage commands send to an IOMMU
98 struct kmem_cache *amd_iommu_irq_cache;
100 static void update_domain(struct protection_domain *domain);
101 static int __init alloc_passthrough_domain(void);
103 /****************************************************************************
107 ****************************************************************************/
109 static struct iommu_dev_data *alloc_dev_data(u16 devid)
111 struct iommu_dev_data *dev_data;
114 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
118 dev_data->devid = devid;
119 atomic_set(&dev_data->bind, 0);
121 spin_lock_irqsave(&dev_data_list_lock, flags);
122 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
123 spin_unlock_irqrestore(&dev_data_list_lock, flags);
128 static void free_dev_data(struct iommu_dev_data *dev_data)
132 spin_lock_irqsave(&dev_data_list_lock, flags);
133 list_del(&dev_data->dev_data_list);
134 spin_unlock_irqrestore(&dev_data_list_lock, flags);
139 static struct iommu_dev_data *search_dev_data(u16 devid)
141 struct iommu_dev_data *dev_data;
144 spin_lock_irqsave(&dev_data_list_lock, flags);
145 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
146 if (dev_data->devid == devid)
153 spin_unlock_irqrestore(&dev_data_list_lock, flags);
158 static struct iommu_dev_data *find_dev_data(u16 devid)
160 struct iommu_dev_data *dev_data;
162 dev_data = search_dev_data(devid);
164 if (dev_data == NULL)
165 dev_data = alloc_dev_data(devid);
170 static inline u16 get_device_id(struct device *dev)
172 struct pci_dev *pdev = to_pci_dev(dev);
174 return calc_devid(pdev->bus->number, pdev->devfn);
177 static struct iommu_dev_data *get_dev_data(struct device *dev)
179 return dev->archdata.iommu;
182 static bool pci_iommuv2_capable(struct pci_dev *pdev)
184 static const int caps[] = {
187 PCI_EXT_CAP_ID_PASID,
191 for (i = 0; i < 3; ++i) {
192 pos = pci_find_ext_capability(pdev, caps[i]);
200 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
202 struct iommu_dev_data *dev_data;
204 dev_data = get_dev_data(&pdev->dev);
206 return dev_data->errata & (1 << erratum) ? true : false;
210 * In this function the list of preallocated protection domains is traversed to
211 * find the domain for a specific device
213 static struct dma_ops_domain *find_protection_domain(u16 devid)
215 struct dma_ops_domain *entry, *ret = NULL;
217 u16 alias = amd_iommu_alias_table[devid];
219 if (list_empty(&iommu_pd_list))
222 spin_lock_irqsave(&iommu_pd_list_lock, flags);
224 list_for_each_entry(entry, &iommu_pd_list, list) {
225 if (entry->target_dev == devid ||
226 entry->target_dev == alias) {
232 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
238 * This function checks if the driver got a valid device from the caller to
239 * avoid dereferencing invalid pointers.
241 static bool check_device(struct device *dev)
245 if (!dev || !dev->dma_mask)
248 /* No device or no PCI device */
249 if (dev->bus != &pci_bus_type)
252 devid = get_device_id(dev);
254 /* Out of our scope? */
255 if (devid > amd_iommu_last_bdf)
258 if (amd_iommu_rlookup_table[devid] == NULL)
264 static void swap_pci_ref(struct pci_dev **from, struct pci_dev *to)
270 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
272 static int iommu_init_device(struct device *dev)
274 struct pci_dev *dma_pdev, *pdev = to_pci_dev(dev);
275 struct iommu_dev_data *dev_data;
276 struct iommu_group *group;
280 if (dev->archdata.iommu)
283 dev_data = find_dev_data(get_device_id(dev));
287 alias = amd_iommu_alias_table[dev_data->devid];
288 if (alias != dev_data->devid) {
289 struct iommu_dev_data *alias_data;
291 alias_data = find_dev_data(alias);
292 if (alias_data == NULL) {
293 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
295 free_dev_data(dev_data);
298 dev_data->alias_data = alias_data;
300 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
302 dma_pdev = pci_dev_get(pdev);
304 /* Account for quirked devices */
305 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
308 * If it's a multifunction device that does not support our
309 * required ACS flags, add to the same group as function 0.
311 if (dma_pdev->multifunction &&
312 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
313 swap_pci_ref(&dma_pdev,
314 pci_get_slot(dma_pdev->bus,
315 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
319 * Devices on the root bus go through the iommu. If that's not us,
320 * find the next upstream device and test ACS up to the root bus.
321 * Finding the next device may require skipping virtual buses.
323 while (!pci_is_root_bus(dma_pdev->bus)) {
324 struct pci_bus *bus = dma_pdev->bus;
327 if (!pci_is_root_bus(bus))
333 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
336 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
340 group = iommu_group_get(&dma_pdev->dev);
341 pci_dev_put(dma_pdev);
343 group = iommu_group_alloc();
345 return PTR_ERR(group);
348 ret = iommu_group_add_device(group, dev);
350 iommu_group_put(group);
355 if (pci_iommuv2_capable(pdev)) {
356 struct amd_iommu *iommu;
358 iommu = amd_iommu_rlookup_table[dev_data->devid];
359 dev_data->iommu_v2 = iommu->is_iommu_v2;
362 dev->archdata.iommu = dev_data;
367 static void iommu_ignore_device(struct device *dev)
371 devid = get_device_id(dev);
372 alias = amd_iommu_alias_table[devid];
374 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
375 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
377 amd_iommu_rlookup_table[devid] = NULL;
378 amd_iommu_rlookup_table[alias] = NULL;
381 static void iommu_uninit_device(struct device *dev)
383 iommu_group_remove_device(dev);
386 * Nothing to do here - we keep dev_data around for unplugged devices
387 * and reuse it when the device is re-plugged - not doing so would
388 * introduce a ton of races.
392 void __init amd_iommu_uninit_devices(void)
394 struct iommu_dev_data *dev_data, *n;
395 struct pci_dev *pdev = NULL;
397 for_each_pci_dev(pdev) {
399 if (!check_device(&pdev->dev))
402 iommu_uninit_device(&pdev->dev);
405 /* Free all of our dev_data structures */
406 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
407 free_dev_data(dev_data);
410 int __init amd_iommu_init_devices(void)
412 struct pci_dev *pdev = NULL;
415 for_each_pci_dev(pdev) {
417 if (!check_device(&pdev->dev))
420 ret = iommu_init_device(&pdev->dev);
421 if (ret == -ENOTSUPP)
422 iommu_ignore_device(&pdev->dev);
431 amd_iommu_uninit_devices();
435 #ifdef CONFIG_AMD_IOMMU_STATS
438 * Initialization code for statistics collection
441 DECLARE_STATS_COUNTER(compl_wait);
442 DECLARE_STATS_COUNTER(cnt_map_single);
443 DECLARE_STATS_COUNTER(cnt_unmap_single);
444 DECLARE_STATS_COUNTER(cnt_map_sg);
445 DECLARE_STATS_COUNTER(cnt_unmap_sg);
446 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
447 DECLARE_STATS_COUNTER(cnt_free_coherent);
448 DECLARE_STATS_COUNTER(cross_page);
449 DECLARE_STATS_COUNTER(domain_flush_single);
450 DECLARE_STATS_COUNTER(domain_flush_all);
451 DECLARE_STATS_COUNTER(alloced_io_mem);
452 DECLARE_STATS_COUNTER(total_map_requests);
453 DECLARE_STATS_COUNTER(complete_ppr);
454 DECLARE_STATS_COUNTER(invalidate_iotlb);
455 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
456 DECLARE_STATS_COUNTER(pri_requests);
458 static struct dentry *stats_dir;
459 static struct dentry *de_fflush;
461 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
463 if (stats_dir == NULL)
466 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
470 static void amd_iommu_stats_init(void)
472 stats_dir = debugfs_create_dir("amd-iommu", NULL);
473 if (stats_dir == NULL)
476 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
477 &amd_iommu_unmap_flush);
479 amd_iommu_stats_add(&compl_wait);
480 amd_iommu_stats_add(&cnt_map_single);
481 amd_iommu_stats_add(&cnt_unmap_single);
482 amd_iommu_stats_add(&cnt_map_sg);
483 amd_iommu_stats_add(&cnt_unmap_sg);
484 amd_iommu_stats_add(&cnt_alloc_coherent);
485 amd_iommu_stats_add(&cnt_free_coherent);
486 amd_iommu_stats_add(&cross_page);
487 amd_iommu_stats_add(&domain_flush_single);
488 amd_iommu_stats_add(&domain_flush_all);
489 amd_iommu_stats_add(&alloced_io_mem);
490 amd_iommu_stats_add(&total_map_requests);
491 amd_iommu_stats_add(&complete_ppr);
492 amd_iommu_stats_add(&invalidate_iotlb);
493 amd_iommu_stats_add(&invalidate_iotlb_all);
494 amd_iommu_stats_add(&pri_requests);
499 /****************************************************************************
501 * Interrupt handling functions
503 ****************************************************************************/
505 static void dump_dte_entry(u16 devid)
509 for (i = 0; i < 4; ++i)
510 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
511 amd_iommu_dev_table[devid].data[i]);
514 static void dump_command(unsigned long phys_addr)
516 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
519 for (i = 0; i < 4; ++i)
520 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
523 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
525 int type, devid, domid, flags;
526 volatile u32 *event = __evt;
531 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
532 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
533 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
534 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
535 address = (u64)(((u64)event[3]) << 32) | event[2];
538 /* Did we hit the erratum? */
539 if (++count == LOOP_TIMEOUT) {
540 pr_err("AMD-Vi: No event written to event log\n");
547 printk(KERN_ERR "AMD-Vi: Event logged [");
550 case EVENT_TYPE_ILL_DEV:
551 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
552 "address=0x%016llx flags=0x%04x]\n",
553 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
555 dump_dte_entry(devid);
557 case EVENT_TYPE_IO_FAULT:
558 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
559 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
560 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
561 domid, address, flags);
563 case EVENT_TYPE_DEV_TAB_ERR:
564 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
565 "address=0x%016llx flags=0x%04x]\n",
566 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
569 case EVENT_TYPE_PAGE_TAB_ERR:
570 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
571 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
572 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
573 domid, address, flags);
575 case EVENT_TYPE_ILL_CMD:
576 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
577 dump_command(address);
579 case EVENT_TYPE_CMD_HARD_ERR:
580 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
581 "flags=0x%04x]\n", address, flags);
583 case EVENT_TYPE_IOTLB_INV_TO:
584 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
585 "address=0x%016llx]\n",
586 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
589 case EVENT_TYPE_INV_DEV_REQ:
590 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
591 "address=0x%016llx flags=0x%04x]\n",
592 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
596 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
599 memset(__evt, 0, 4 * sizeof(u32));
602 static void iommu_poll_events(struct amd_iommu *iommu)
607 spin_lock_irqsave(&iommu->lock, flags);
609 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
610 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
612 while (head != tail) {
613 iommu_print_event(iommu, iommu->evt_buf + head);
614 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
617 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
619 spin_unlock_irqrestore(&iommu->lock, flags);
622 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
624 struct amd_iommu_fault fault;
626 INC_STATS_COUNTER(pri_requests);
628 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
629 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
633 fault.address = raw[1];
634 fault.pasid = PPR_PASID(raw[0]);
635 fault.device_id = PPR_DEVID(raw[0]);
636 fault.tag = PPR_TAG(raw[0]);
637 fault.flags = PPR_FLAGS(raw[0]);
639 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
642 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
647 if (iommu->ppr_log == NULL)
650 /* enable ppr interrupts again */
651 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
653 spin_lock_irqsave(&iommu->lock, flags);
655 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
656 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
658 while (head != tail) {
663 raw = (u64 *)(iommu->ppr_log + head);
666 * Hardware bug: Interrupt may arrive before the entry is
667 * written to memory. If this happens we need to wait for the
670 for (i = 0; i < LOOP_TIMEOUT; ++i) {
671 if (PPR_REQ_TYPE(raw[0]) != 0)
676 /* Avoid memcpy function-call overhead */
681 * To detect the hardware bug we need to clear the entry
684 raw[0] = raw[1] = 0UL;
686 /* Update head pointer of hardware ring-buffer */
687 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
688 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
691 * Release iommu->lock because ppr-handling might need to
694 spin_unlock_irqrestore(&iommu->lock, flags);
696 /* Handle PPR entry */
697 iommu_handle_ppr_entry(iommu, entry);
699 spin_lock_irqsave(&iommu->lock, flags);
701 /* Refresh ring-buffer information */
702 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
703 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
706 spin_unlock_irqrestore(&iommu->lock, flags);
709 irqreturn_t amd_iommu_int_thread(int irq, void *data)
711 struct amd_iommu *iommu;
713 for_each_iommu(iommu) {
714 iommu_poll_events(iommu);
715 iommu_poll_ppr_log(iommu);
721 irqreturn_t amd_iommu_int_handler(int irq, void *data)
723 return IRQ_WAKE_THREAD;
726 /****************************************************************************
728 * IOMMU command queuing functions
730 ****************************************************************************/
732 static int wait_on_sem(volatile u64 *sem)
736 while (*sem == 0 && i < LOOP_TIMEOUT) {
741 if (i == LOOP_TIMEOUT) {
742 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
749 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
750 struct iommu_cmd *cmd,
755 target = iommu->cmd_buf + tail;
756 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
758 /* Copy command to buffer */
759 memcpy(target, cmd, sizeof(*cmd));
761 /* Tell the IOMMU about it */
762 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
765 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
767 WARN_ON(address & 0x7ULL);
769 memset(cmd, 0, sizeof(*cmd));
770 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
771 cmd->data[1] = upper_32_bits(__pa(address));
773 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
776 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
778 memset(cmd, 0, sizeof(*cmd));
779 cmd->data[0] = devid;
780 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
783 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
784 size_t size, u16 domid, int pde)
789 pages = iommu_num_pages(address, size, PAGE_SIZE);
794 * If we have to flush more than one page, flush all
795 * TLB entries for this domain
797 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
801 address &= PAGE_MASK;
803 memset(cmd, 0, sizeof(*cmd));
804 cmd->data[1] |= domid;
805 cmd->data[2] = lower_32_bits(address);
806 cmd->data[3] = upper_32_bits(address);
807 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
808 if (s) /* size bit - we flush more than one 4kb page */
809 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
810 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
811 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
814 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
815 u64 address, size_t size)
820 pages = iommu_num_pages(address, size, PAGE_SIZE);
825 * If we have to flush more than one page, flush all
826 * TLB entries for this domain
828 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
832 address &= PAGE_MASK;
834 memset(cmd, 0, sizeof(*cmd));
835 cmd->data[0] = devid;
836 cmd->data[0] |= (qdep & 0xff) << 24;
837 cmd->data[1] = devid;
838 cmd->data[2] = lower_32_bits(address);
839 cmd->data[3] = upper_32_bits(address);
840 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
842 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
845 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
846 u64 address, bool size)
848 memset(cmd, 0, sizeof(*cmd));
850 address &= ~(0xfffULL);
852 cmd->data[0] = pasid & PASID_MASK;
853 cmd->data[1] = domid;
854 cmd->data[2] = lower_32_bits(address);
855 cmd->data[3] = upper_32_bits(address);
856 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
857 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
859 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
860 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
863 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
864 int qdep, u64 address, bool size)
866 memset(cmd, 0, sizeof(*cmd));
868 address &= ~(0xfffULL);
870 cmd->data[0] = devid;
871 cmd->data[0] |= (pasid & 0xff) << 16;
872 cmd->data[0] |= (qdep & 0xff) << 24;
873 cmd->data[1] = devid;
874 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
875 cmd->data[2] = lower_32_bits(address);
876 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
877 cmd->data[3] = upper_32_bits(address);
879 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
880 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
883 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
884 int status, int tag, bool gn)
886 memset(cmd, 0, sizeof(*cmd));
888 cmd->data[0] = devid;
890 cmd->data[1] = pasid & PASID_MASK;
891 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
893 cmd->data[3] = tag & 0x1ff;
894 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
896 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
899 static void build_inv_all(struct iommu_cmd *cmd)
901 memset(cmd, 0, sizeof(*cmd));
902 CMD_SET_TYPE(cmd, CMD_INV_ALL);
906 * Writes the command to the IOMMUs command buffer and informs the
907 * hardware about the new command.
909 static int iommu_queue_command_sync(struct amd_iommu *iommu,
910 struct iommu_cmd *cmd,
913 u32 left, tail, head, next_tail;
916 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
919 spin_lock_irqsave(&iommu->lock, flags);
921 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
922 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
923 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
924 left = (head - next_tail) % iommu->cmd_buf_size;
927 struct iommu_cmd sync_cmd;
928 volatile u64 sem = 0;
931 build_completion_wait(&sync_cmd, (u64)&sem);
932 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
934 spin_unlock_irqrestore(&iommu->lock, flags);
936 if ((ret = wait_on_sem(&sem)) != 0)
942 copy_cmd_to_buffer(iommu, cmd, tail);
944 /* We need to sync now to make sure all commands are processed */
945 iommu->need_sync = sync;
947 spin_unlock_irqrestore(&iommu->lock, flags);
952 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
954 return iommu_queue_command_sync(iommu, cmd, true);
958 * This function queues a completion wait command into the command
961 static int iommu_completion_wait(struct amd_iommu *iommu)
963 struct iommu_cmd cmd;
964 volatile u64 sem = 0;
967 if (!iommu->need_sync)
970 build_completion_wait(&cmd, (u64)&sem);
972 ret = iommu_queue_command_sync(iommu, &cmd, false);
976 return wait_on_sem(&sem);
979 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
981 struct iommu_cmd cmd;
983 build_inv_dte(&cmd, devid);
985 return iommu_queue_command(iommu, &cmd);
988 static void iommu_flush_dte_all(struct amd_iommu *iommu)
992 for (devid = 0; devid <= 0xffff; ++devid)
993 iommu_flush_dte(iommu, devid);
995 iommu_completion_wait(iommu);
999 * This function uses heavy locking and may disable irqs for some time. But
1000 * this is no issue because it is only called during resume.
1002 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1006 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1007 struct iommu_cmd cmd;
1008 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1010 iommu_queue_command(iommu, &cmd);
1013 iommu_completion_wait(iommu);
1016 static void iommu_flush_all(struct amd_iommu *iommu)
1018 struct iommu_cmd cmd;
1020 build_inv_all(&cmd);
1022 iommu_queue_command(iommu, &cmd);
1023 iommu_completion_wait(iommu);
1026 void iommu_flush_all_caches(struct amd_iommu *iommu)
1028 if (iommu_feature(iommu, FEATURE_IA)) {
1029 iommu_flush_all(iommu);
1031 iommu_flush_dte_all(iommu);
1032 iommu_flush_tlb_all(iommu);
1037 * Command send function for flushing on-device TLB
1039 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1040 u64 address, size_t size)
1042 struct amd_iommu *iommu;
1043 struct iommu_cmd cmd;
1046 qdep = dev_data->ats.qdep;
1047 iommu = amd_iommu_rlookup_table[dev_data->devid];
1049 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1051 return iommu_queue_command(iommu, &cmd);
1055 * Command send function for invalidating a device table entry
1057 static int device_flush_dte(struct iommu_dev_data *dev_data)
1059 struct amd_iommu *iommu;
1062 iommu = amd_iommu_rlookup_table[dev_data->devid];
1064 ret = iommu_flush_dte(iommu, dev_data->devid);
1068 if (dev_data->ats.enabled)
1069 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1075 * TLB invalidation function which is called from the mapping functions.
1076 * It invalidates a single PTE if the range to flush is within a single
1077 * page. Otherwise it flushes the whole TLB of the IOMMU.
1079 static void __domain_flush_pages(struct protection_domain *domain,
1080 u64 address, size_t size, int pde)
1082 struct iommu_dev_data *dev_data;
1083 struct iommu_cmd cmd;
1086 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1088 for (i = 0; i < amd_iommus_present; ++i) {
1089 if (!domain->dev_iommu[i])
1093 * Devices of this domain are behind this IOMMU
1094 * We need a TLB flush
1096 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1099 list_for_each_entry(dev_data, &domain->dev_list, list) {
1101 if (!dev_data->ats.enabled)
1104 ret |= device_flush_iotlb(dev_data, address, size);
1110 static void domain_flush_pages(struct protection_domain *domain,
1111 u64 address, size_t size)
1113 __domain_flush_pages(domain, address, size, 0);
1116 /* Flush the whole IO/TLB for a given protection domain */
1117 static void domain_flush_tlb(struct protection_domain *domain)
1119 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1122 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1123 static void domain_flush_tlb_pde(struct protection_domain *domain)
1125 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1128 static void domain_flush_complete(struct protection_domain *domain)
1132 for (i = 0; i < amd_iommus_present; ++i) {
1133 if (!domain->dev_iommu[i])
1137 * Devices of this domain are behind this IOMMU
1138 * We need to wait for completion of all commands.
1140 iommu_completion_wait(amd_iommus[i]);
1146 * This function flushes the DTEs for all devices in domain
1148 static void domain_flush_devices(struct protection_domain *domain)
1150 struct iommu_dev_data *dev_data;
1152 list_for_each_entry(dev_data, &domain->dev_list, list)
1153 device_flush_dte(dev_data);
1156 /****************************************************************************
1158 * The functions below are used the create the page table mappings for
1159 * unity mapped regions.
1161 ****************************************************************************/
1164 * This function is used to add another level to an IO page table. Adding
1165 * another level increases the size of the address space by 9 bits to a size up
1168 static bool increase_address_space(struct protection_domain *domain,
1173 if (domain->mode == PAGE_MODE_6_LEVEL)
1174 /* address space already 64 bit large */
1177 pte = (void *)get_zeroed_page(gfp);
1181 *pte = PM_LEVEL_PDE(domain->mode,
1182 virt_to_phys(domain->pt_root));
1183 domain->pt_root = pte;
1185 domain->updated = true;
1190 static u64 *alloc_pte(struct protection_domain *domain,
1191 unsigned long address,
1192 unsigned long page_size,
1199 BUG_ON(!is_power_of_2(page_size));
1201 while (address > PM_LEVEL_SIZE(domain->mode))
1202 increase_address_space(domain, gfp);
1204 level = domain->mode - 1;
1205 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1206 address = PAGE_SIZE_ALIGN(address, page_size);
1207 end_lvl = PAGE_SIZE_LEVEL(page_size);
1209 while (level > end_lvl) {
1210 if (!IOMMU_PTE_PRESENT(*pte)) {
1211 page = (u64 *)get_zeroed_page(gfp);
1214 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1217 /* No level skipping support yet */
1218 if (PM_PTE_LEVEL(*pte) != level)
1223 pte = IOMMU_PTE_PAGE(*pte);
1225 if (pte_page && level == end_lvl)
1228 pte = &pte[PM_LEVEL_INDEX(level, address)];
1235 * This function checks if there is a PTE for a given dma address. If
1236 * there is one, it returns the pointer to it.
1238 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1243 if (address > PM_LEVEL_SIZE(domain->mode))
1246 level = domain->mode - 1;
1247 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1252 if (!IOMMU_PTE_PRESENT(*pte))
1256 if (PM_PTE_LEVEL(*pte) == 0x07) {
1257 unsigned long pte_mask, __pte;
1260 * If we have a series of large PTEs, make
1261 * sure to return a pointer to the first one.
1263 pte_mask = PTE_PAGE_SIZE(*pte);
1264 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1265 __pte = ((unsigned long)pte) & pte_mask;
1267 return (u64 *)__pte;
1270 /* No level skipping support yet */
1271 if (PM_PTE_LEVEL(*pte) != level)
1276 /* Walk to the next level */
1277 pte = IOMMU_PTE_PAGE(*pte);
1278 pte = &pte[PM_LEVEL_INDEX(level, address)];
1285 * Generic mapping functions. It maps a physical address into a DMA
1286 * address space. It allocates the page table pages if necessary.
1287 * In the future it can be extended to a generic mapping function
1288 * supporting all features of AMD IOMMU page tables like level skipping
1289 * and full 64 bit address spaces.
1291 static int iommu_map_page(struct protection_domain *dom,
1292 unsigned long bus_addr,
1293 unsigned long phys_addr,
1295 unsigned long page_size)
1300 if (!(prot & IOMMU_PROT_MASK))
1303 bus_addr = PAGE_ALIGN(bus_addr);
1304 phys_addr = PAGE_ALIGN(phys_addr);
1305 count = PAGE_SIZE_PTE_COUNT(page_size);
1306 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1308 for (i = 0; i < count; ++i)
1309 if (IOMMU_PTE_PRESENT(pte[i]))
1312 if (page_size > PAGE_SIZE) {
1313 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1314 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1316 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1318 if (prot & IOMMU_PROT_IR)
1319 __pte |= IOMMU_PTE_IR;
1320 if (prot & IOMMU_PROT_IW)
1321 __pte |= IOMMU_PTE_IW;
1323 for (i = 0; i < count; ++i)
1331 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1332 unsigned long bus_addr,
1333 unsigned long page_size)
1335 unsigned long long unmap_size, unmapped;
1338 BUG_ON(!is_power_of_2(page_size));
1342 while (unmapped < page_size) {
1344 pte = fetch_pte(dom, bus_addr);
1348 * No PTE for this address
1349 * move forward in 4kb steps
1351 unmap_size = PAGE_SIZE;
1352 } else if (PM_PTE_LEVEL(*pte) == 0) {
1353 /* 4kb PTE found for this address */
1354 unmap_size = PAGE_SIZE;
1359 /* Large PTE found which maps this address */
1360 unmap_size = PTE_PAGE_SIZE(*pte);
1361 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1362 for (i = 0; i < count; i++)
1366 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1367 unmapped += unmap_size;
1370 BUG_ON(!is_power_of_2(unmapped));
1376 * This function checks if a specific unity mapping entry is needed for
1377 * this specific IOMMU.
1379 static int iommu_for_unity_map(struct amd_iommu *iommu,
1380 struct unity_map_entry *entry)
1384 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1385 bdf = amd_iommu_alias_table[i];
1386 if (amd_iommu_rlookup_table[bdf] == iommu)
1394 * This function actually applies the mapping to the page table of the
1397 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1398 struct unity_map_entry *e)
1403 for (addr = e->address_start; addr < e->address_end;
1404 addr += PAGE_SIZE) {
1405 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1410 * if unity mapping is in aperture range mark the page
1411 * as allocated in the aperture
1413 if (addr < dma_dom->aperture_size)
1414 __set_bit(addr >> PAGE_SHIFT,
1415 dma_dom->aperture[0]->bitmap);
1422 * Init the unity mappings for a specific IOMMU in the system
1424 * Basically iterates over all unity mapping entries and applies them to
1425 * the default domain DMA of that IOMMU if necessary.
1427 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1429 struct unity_map_entry *entry;
1432 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1433 if (!iommu_for_unity_map(iommu, entry))
1435 ret = dma_ops_unity_map(iommu->default_dom, entry);
1444 * Inits the unity mappings required for a specific device
1446 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1449 struct unity_map_entry *e;
1452 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1453 if (!(devid >= e->devid_start && devid <= e->devid_end))
1455 ret = dma_ops_unity_map(dma_dom, e);
1463 /****************************************************************************
1465 * The next functions belong to the address allocator for the dma_ops
1466 * interface functions. They work like the allocators in the other IOMMU
1467 * drivers. Its basically a bitmap which marks the allocated pages in
1468 * the aperture. Maybe it could be enhanced in the future to a more
1469 * efficient allocator.
1471 ****************************************************************************/
1474 * The address allocator core functions.
1476 * called with domain->lock held
1480 * Used to reserve address ranges in the aperture (e.g. for exclusion
1483 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1484 unsigned long start_page,
1487 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1489 if (start_page + pages > last_page)
1490 pages = last_page - start_page;
1492 for (i = start_page; i < start_page + pages; ++i) {
1493 int index = i / APERTURE_RANGE_PAGES;
1494 int page = i % APERTURE_RANGE_PAGES;
1495 __set_bit(page, dom->aperture[index]->bitmap);
1500 * This function is used to add a new aperture range to an existing
1501 * aperture in case of dma_ops domain allocation or address allocation
1504 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1505 bool populate, gfp_t gfp)
1507 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1508 struct amd_iommu *iommu;
1509 unsigned long i, old_size;
1511 #ifdef CONFIG_IOMMU_STRESS
1515 if (index >= APERTURE_MAX_RANGES)
1518 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1519 if (!dma_dom->aperture[index])
1522 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1523 if (!dma_dom->aperture[index]->bitmap)
1526 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1529 unsigned long address = dma_dom->aperture_size;
1530 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1531 u64 *pte, *pte_page;
1533 for (i = 0; i < num_ptes; ++i) {
1534 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1539 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1541 address += APERTURE_RANGE_SIZE / 64;
1545 old_size = dma_dom->aperture_size;
1546 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1548 /* Reserve address range used for MSI messages */
1549 if (old_size < MSI_ADDR_BASE_LO &&
1550 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1551 unsigned long spage;
1554 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1555 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1557 dma_ops_reserve_addresses(dma_dom, spage, pages);
1560 /* Initialize the exclusion range if necessary */
1561 for_each_iommu(iommu) {
1562 if (iommu->exclusion_start &&
1563 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1564 && iommu->exclusion_start < dma_dom->aperture_size) {
1565 unsigned long startpage;
1566 int pages = iommu_num_pages(iommu->exclusion_start,
1567 iommu->exclusion_length,
1569 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1570 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1575 * Check for areas already mapped as present in the new aperture
1576 * range and mark those pages as reserved in the allocator. Such
1577 * mappings may already exist as a result of requested unity
1578 * mappings for devices.
1580 for (i = dma_dom->aperture[index]->offset;
1581 i < dma_dom->aperture_size;
1583 u64 *pte = fetch_pte(&dma_dom->domain, i);
1584 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1587 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1590 update_domain(&dma_dom->domain);
1595 update_domain(&dma_dom->domain);
1597 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1599 kfree(dma_dom->aperture[index]);
1600 dma_dom->aperture[index] = NULL;
1605 static unsigned long dma_ops_area_alloc(struct device *dev,
1606 struct dma_ops_domain *dom,
1608 unsigned long align_mask,
1610 unsigned long start)
1612 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1613 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1614 int i = start >> APERTURE_RANGE_SHIFT;
1615 unsigned long boundary_size;
1616 unsigned long address = -1;
1617 unsigned long limit;
1619 next_bit >>= PAGE_SHIFT;
1621 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1622 PAGE_SIZE) >> PAGE_SHIFT;
1624 for (;i < max_index; ++i) {
1625 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1627 if (dom->aperture[i]->offset >= dma_mask)
1630 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1631 dma_mask >> PAGE_SHIFT);
1633 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1634 limit, next_bit, pages, 0,
1635 boundary_size, align_mask);
1636 if (address != -1) {
1637 address = dom->aperture[i]->offset +
1638 (address << PAGE_SHIFT);
1639 dom->next_address = address + (pages << PAGE_SHIFT);
1649 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1650 struct dma_ops_domain *dom,
1652 unsigned long align_mask,
1655 unsigned long address;
1657 #ifdef CONFIG_IOMMU_STRESS
1658 dom->next_address = 0;
1659 dom->need_flush = true;
1662 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1663 dma_mask, dom->next_address);
1665 if (address == -1) {
1666 dom->next_address = 0;
1667 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1669 dom->need_flush = true;
1672 if (unlikely(address == -1))
1673 address = DMA_ERROR_CODE;
1675 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1681 * The address free function.
1683 * called with domain->lock held
1685 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1686 unsigned long address,
1689 unsigned i = address >> APERTURE_RANGE_SHIFT;
1690 struct aperture_range *range = dom->aperture[i];
1692 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1694 #ifdef CONFIG_IOMMU_STRESS
1699 if (address >= dom->next_address)
1700 dom->need_flush = true;
1702 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1704 bitmap_clear(range->bitmap, address, pages);
1708 /****************************************************************************
1710 * The next functions belong to the domain allocation. A domain is
1711 * allocated for every IOMMU as the default domain. If device isolation
1712 * is enabled, every device get its own domain. The most important thing
1713 * about domains is the page table mapping the DMA address space they
1716 ****************************************************************************/
1719 * This function adds a protection domain to the global protection domain list
1721 static void add_domain_to_list(struct protection_domain *domain)
1723 unsigned long flags;
1725 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1726 list_add(&domain->list, &amd_iommu_pd_list);
1727 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1731 * This function removes a protection domain to the global
1732 * protection domain list
1734 static void del_domain_from_list(struct protection_domain *domain)
1736 unsigned long flags;
1738 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1739 list_del(&domain->list);
1740 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1743 static u16 domain_id_alloc(void)
1745 unsigned long flags;
1748 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1749 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1751 if (id > 0 && id < MAX_DOMAIN_ID)
1752 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1755 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1760 static void domain_id_free(int id)
1762 unsigned long flags;
1764 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1765 if (id > 0 && id < MAX_DOMAIN_ID)
1766 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1767 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1770 static void free_pagetable(struct protection_domain *domain)
1775 p1 = domain->pt_root;
1780 for (i = 0; i < 512; ++i) {
1781 if (!IOMMU_PTE_PRESENT(p1[i]))
1784 p2 = IOMMU_PTE_PAGE(p1[i]);
1785 for (j = 0; j < 512; ++j) {
1786 if (!IOMMU_PTE_PRESENT(p2[j]))
1788 p3 = IOMMU_PTE_PAGE(p2[j]);
1789 free_page((unsigned long)p3);
1792 free_page((unsigned long)p2);
1795 free_page((unsigned long)p1);
1797 domain->pt_root = NULL;
1800 static void free_gcr3_tbl_level1(u64 *tbl)
1805 for (i = 0; i < 512; ++i) {
1806 if (!(tbl[i] & GCR3_VALID))
1809 ptr = __va(tbl[i] & PAGE_MASK);
1811 free_page((unsigned long)ptr);
1815 static void free_gcr3_tbl_level2(u64 *tbl)
1820 for (i = 0; i < 512; ++i) {
1821 if (!(tbl[i] & GCR3_VALID))
1824 ptr = __va(tbl[i] & PAGE_MASK);
1826 free_gcr3_tbl_level1(ptr);
1830 static void free_gcr3_table(struct protection_domain *domain)
1832 if (domain->glx == 2)
1833 free_gcr3_tbl_level2(domain->gcr3_tbl);
1834 else if (domain->glx == 1)
1835 free_gcr3_tbl_level1(domain->gcr3_tbl);
1836 else if (domain->glx != 0)
1839 free_page((unsigned long)domain->gcr3_tbl);
1843 * Free a domain, only used if something went wrong in the
1844 * allocation path and we need to free an already allocated page table
1846 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1853 del_domain_from_list(&dom->domain);
1855 free_pagetable(&dom->domain);
1857 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1858 if (!dom->aperture[i])
1860 free_page((unsigned long)dom->aperture[i]->bitmap);
1861 kfree(dom->aperture[i]);
1868 * Allocates a new protection domain usable for the dma_ops functions.
1869 * It also initializes the page table and the address allocator data
1870 * structures required for the dma_ops interface
1872 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1874 struct dma_ops_domain *dma_dom;
1876 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1880 spin_lock_init(&dma_dom->domain.lock);
1882 dma_dom->domain.id = domain_id_alloc();
1883 if (dma_dom->domain.id == 0)
1885 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1886 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1887 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1888 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1889 dma_dom->domain.priv = dma_dom;
1890 if (!dma_dom->domain.pt_root)
1893 dma_dom->need_flush = false;
1894 dma_dom->target_dev = 0xffff;
1896 add_domain_to_list(&dma_dom->domain);
1898 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1902 * mark the first page as allocated so we never return 0 as
1903 * a valid dma-address. So we can use 0 as error value
1905 dma_dom->aperture[0]->bitmap[0] = 1;
1906 dma_dom->next_address = 0;
1912 dma_ops_domain_free(dma_dom);
1918 * little helper function to check whether a given protection domain is a
1921 static bool dma_ops_domain(struct protection_domain *domain)
1923 return domain->flags & PD_DMA_OPS_MASK;
1926 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1931 if (domain->mode != PAGE_MODE_NONE)
1932 pte_root = virt_to_phys(domain->pt_root);
1934 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1935 << DEV_ENTRY_MODE_SHIFT;
1936 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1938 flags = amd_iommu_dev_table[devid].data[1];
1941 flags |= DTE_FLAG_IOTLB;
1943 if (domain->flags & PD_IOMMUV2_MASK) {
1944 u64 gcr3 = __pa(domain->gcr3_tbl);
1945 u64 glx = domain->glx;
1948 pte_root |= DTE_FLAG_GV;
1949 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1951 /* First mask out possible old values for GCR3 table */
1952 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1955 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1958 /* Encode GCR3 table into DTE */
1959 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1962 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1965 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1969 flags &= ~(0xffffUL);
1970 flags |= domain->id;
1972 amd_iommu_dev_table[devid].data[1] = flags;
1973 amd_iommu_dev_table[devid].data[0] = pte_root;
1976 static void clear_dte_entry(u16 devid)
1978 /* remove entry from the device table seen by the hardware */
1979 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1980 amd_iommu_dev_table[devid].data[1] = 0;
1982 amd_iommu_apply_erratum_63(devid);
1985 static void do_attach(struct iommu_dev_data *dev_data,
1986 struct protection_domain *domain)
1988 struct amd_iommu *iommu;
1991 iommu = amd_iommu_rlookup_table[dev_data->devid];
1992 ats = dev_data->ats.enabled;
1994 /* Update data structures */
1995 dev_data->domain = domain;
1996 list_add(&dev_data->list, &domain->dev_list);
1997 set_dte_entry(dev_data->devid, domain, ats);
1999 /* Do reference counting */
2000 domain->dev_iommu[iommu->index] += 1;
2001 domain->dev_cnt += 1;
2003 /* Flush the DTE entry */
2004 device_flush_dte(dev_data);
2007 static void do_detach(struct iommu_dev_data *dev_data)
2009 struct amd_iommu *iommu;
2011 iommu = amd_iommu_rlookup_table[dev_data->devid];
2013 /* decrease reference counters */
2014 dev_data->domain->dev_iommu[iommu->index] -= 1;
2015 dev_data->domain->dev_cnt -= 1;
2017 /* Update data structures */
2018 dev_data->domain = NULL;
2019 list_del(&dev_data->list);
2020 clear_dte_entry(dev_data->devid);
2022 /* Flush the DTE entry */
2023 device_flush_dte(dev_data);
2027 * If a device is not yet associated with a domain, this function does
2028 * assigns it visible for the hardware
2030 static int __attach_device(struct iommu_dev_data *dev_data,
2031 struct protection_domain *domain)
2036 spin_lock(&domain->lock);
2038 if (dev_data->alias_data != NULL) {
2039 struct iommu_dev_data *alias_data = dev_data->alias_data;
2041 /* Some sanity checks */
2043 if (alias_data->domain != NULL &&
2044 alias_data->domain != domain)
2047 if (dev_data->domain != NULL &&
2048 dev_data->domain != domain)
2051 /* Do real assignment */
2052 if (alias_data->domain == NULL)
2053 do_attach(alias_data, domain);
2055 atomic_inc(&alias_data->bind);
2058 if (dev_data->domain == NULL)
2059 do_attach(dev_data, domain);
2061 atomic_inc(&dev_data->bind);
2068 spin_unlock(&domain->lock);
2074 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2076 pci_disable_ats(pdev);
2077 pci_disable_pri(pdev);
2078 pci_disable_pasid(pdev);
2081 /* FIXME: Change generic reset-function to do the same */
2082 static int pri_reset_while_enabled(struct pci_dev *pdev)
2087 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2091 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2092 control |= PCI_PRI_CTRL_RESET;
2093 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2098 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2103 /* FIXME: Hardcode number of outstanding requests for now */
2105 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2107 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2109 /* Only allow access to user-accessible pages */
2110 ret = pci_enable_pasid(pdev, 0);
2114 /* First reset the PRI state of the device */
2115 ret = pci_reset_pri(pdev);
2120 ret = pci_enable_pri(pdev, reqs);
2125 ret = pri_reset_while_enabled(pdev);
2130 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2137 pci_disable_pri(pdev);
2138 pci_disable_pasid(pdev);
2143 /* FIXME: Move this to PCI code */
2144 #define PCI_PRI_TLP_OFF (1 << 15)
2146 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2151 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2155 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2157 return (status & PCI_PRI_TLP_OFF) ? true : false;
2161 * If a device is not yet associated with a domain, this function does
2162 * assigns it visible for the hardware
2164 static int attach_device(struct device *dev,
2165 struct protection_domain *domain)
2167 struct pci_dev *pdev = to_pci_dev(dev);
2168 struct iommu_dev_data *dev_data;
2169 unsigned long flags;
2172 dev_data = get_dev_data(dev);
2174 if (domain->flags & PD_IOMMUV2_MASK) {
2175 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2178 if (pdev_iommuv2_enable(pdev) != 0)
2181 dev_data->ats.enabled = true;
2182 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2183 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2184 } else if (amd_iommu_iotlb_sup &&
2185 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2186 dev_data->ats.enabled = true;
2187 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2190 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2191 ret = __attach_device(dev_data, domain);
2192 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2195 * We might boot into a crash-kernel here. The crashed kernel
2196 * left the caches in the IOMMU dirty. So we have to flush
2197 * here to evict all dirty stuff.
2199 domain_flush_tlb_pde(domain);
2205 * Removes a device from a protection domain (unlocked)
2207 static void __detach_device(struct iommu_dev_data *dev_data)
2209 struct protection_domain *domain;
2210 unsigned long flags;
2212 BUG_ON(!dev_data->domain);
2214 domain = dev_data->domain;
2216 spin_lock_irqsave(&domain->lock, flags);
2218 if (dev_data->alias_data != NULL) {
2219 struct iommu_dev_data *alias_data = dev_data->alias_data;
2221 if (atomic_dec_and_test(&alias_data->bind))
2222 do_detach(alias_data);
2225 if (atomic_dec_and_test(&dev_data->bind))
2226 do_detach(dev_data);
2228 spin_unlock_irqrestore(&domain->lock, flags);
2231 * If we run in passthrough mode the device must be assigned to the
2232 * passthrough domain if it is detached from any other domain.
2233 * Make sure we can deassign from the pt_domain itself.
2235 if (dev_data->passthrough &&
2236 (dev_data->domain == NULL && domain != pt_domain))
2237 __attach_device(dev_data, pt_domain);
2241 * Removes a device from a protection domain (with devtable_lock held)
2243 static void detach_device(struct device *dev)
2245 struct protection_domain *domain;
2246 struct iommu_dev_data *dev_data;
2247 unsigned long flags;
2249 dev_data = get_dev_data(dev);
2250 domain = dev_data->domain;
2252 /* lock device table */
2253 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2254 __detach_device(dev_data);
2255 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2257 if (domain->flags & PD_IOMMUV2_MASK)
2258 pdev_iommuv2_disable(to_pci_dev(dev));
2259 else if (dev_data->ats.enabled)
2260 pci_disable_ats(to_pci_dev(dev));
2262 dev_data->ats.enabled = false;
2266 * Find out the protection domain structure for a given PCI device. This
2267 * will give us the pointer to the page table root for example.
2269 static struct protection_domain *domain_for_device(struct device *dev)
2271 struct iommu_dev_data *dev_data;
2272 struct protection_domain *dom = NULL;
2273 unsigned long flags;
2275 dev_data = get_dev_data(dev);
2277 if (dev_data->domain)
2278 return dev_data->domain;
2280 if (dev_data->alias_data != NULL) {
2281 struct iommu_dev_data *alias_data = dev_data->alias_data;
2283 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2284 if (alias_data->domain != NULL) {
2285 __attach_device(dev_data, alias_data->domain);
2286 dom = alias_data->domain;
2288 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2294 static int device_change_notifier(struct notifier_block *nb,
2295 unsigned long action, void *data)
2297 struct dma_ops_domain *dma_domain;
2298 struct protection_domain *domain;
2299 struct iommu_dev_data *dev_data;
2300 struct device *dev = data;
2301 struct amd_iommu *iommu;
2302 unsigned long flags;
2305 if (!check_device(dev))
2308 devid = get_device_id(dev);
2309 iommu = amd_iommu_rlookup_table[devid];
2310 dev_data = get_dev_data(dev);
2313 case BUS_NOTIFY_UNBOUND_DRIVER:
2315 domain = domain_for_device(dev);
2319 if (dev_data->passthrough)
2323 case BUS_NOTIFY_ADD_DEVICE:
2325 iommu_init_device(dev);
2328 * dev_data is still NULL and
2329 * got initialized in iommu_init_device
2331 dev_data = get_dev_data(dev);
2333 if (iommu_pass_through || dev_data->iommu_v2) {
2334 dev_data->passthrough = true;
2335 attach_device(dev, pt_domain);
2339 domain = domain_for_device(dev);
2341 /* allocate a protection domain if a device is added */
2342 dma_domain = find_protection_domain(devid);
2345 dma_domain = dma_ops_domain_alloc();
2348 dma_domain->target_dev = devid;
2350 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2351 list_add_tail(&dma_domain->list, &iommu_pd_list);
2352 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2354 dev_data = get_dev_data(dev);
2356 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2359 case BUS_NOTIFY_DEL_DEVICE:
2361 iommu_uninit_device(dev);
2367 iommu_completion_wait(iommu);
2373 static struct notifier_block device_nb = {
2374 .notifier_call = device_change_notifier,
2377 void amd_iommu_init_notifier(void)
2379 bus_register_notifier(&pci_bus_type, &device_nb);
2382 /*****************************************************************************
2384 * The next functions belong to the dma_ops mapping/unmapping code.
2386 *****************************************************************************/
2389 * In the dma_ops path we only have the struct device. This function
2390 * finds the corresponding IOMMU, the protection domain and the
2391 * requestor id for a given device.
2392 * If the device is not yet associated with a domain this is also done
2395 static struct protection_domain *get_domain(struct device *dev)
2397 struct protection_domain *domain;
2398 struct dma_ops_domain *dma_dom;
2399 u16 devid = get_device_id(dev);
2401 if (!check_device(dev))
2402 return ERR_PTR(-EINVAL);
2404 domain = domain_for_device(dev);
2405 if (domain != NULL && !dma_ops_domain(domain))
2406 return ERR_PTR(-EBUSY);
2411 /* Device not bount yet - bind it */
2412 dma_dom = find_protection_domain(devid);
2414 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2415 attach_device(dev, &dma_dom->domain);
2416 DUMP_printk("Using protection domain %d for device %s\n",
2417 dma_dom->domain.id, dev_name(dev));
2419 return &dma_dom->domain;
2422 static void update_device_table(struct protection_domain *domain)
2424 struct iommu_dev_data *dev_data;
2426 list_for_each_entry(dev_data, &domain->dev_list, list)
2427 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2430 static void update_domain(struct protection_domain *domain)
2432 if (!domain->updated)
2435 update_device_table(domain);
2437 domain_flush_devices(domain);
2438 domain_flush_tlb_pde(domain);
2440 domain->updated = false;
2444 * This function fetches the PTE for a given address in the aperture
2446 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2447 unsigned long address)
2449 struct aperture_range *aperture;
2450 u64 *pte, *pte_page;
2452 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2456 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2458 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2460 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2462 pte += PM_LEVEL_INDEX(0, address);
2464 update_domain(&dom->domain);
2470 * This is the generic map function. It maps one 4kb page at paddr to
2471 * the given address in the DMA address space for the domain.
2473 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2474 unsigned long address,
2480 WARN_ON(address > dom->aperture_size);
2484 pte = dma_ops_get_pte(dom, address);
2486 return DMA_ERROR_CODE;
2488 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2490 if (direction == DMA_TO_DEVICE)
2491 __pte |= IOMMU_PTE_IR;
2492 else if (direction == DMA_FROM_DEVICE)
2493 __pte |= IOMMU_PTE_IW;
2494 else if (direction == DMA_BIDIRECTIONAL)
2495 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2501 return (dma_addr_t)address;
2505 * The generic unmapping function for on page in the DMA address space.
2507 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2508 unsigned long address)
2510 struct aperture_range *aperture;
2513 if (address >= dom->aperture_size)
2516 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2520 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2524 pte += PM_LEVEL_INDEX(0, address);
2532 * This function contains common code for mapping of a physically
2533 * contiguous memory region into DMA address space. It is used by all
2534 * mapping functions provided with this IOMMU driver.
2535 * Must be called with the domain lock held.
2537 static dma_addr_t __map_single(struct device *dev,
2538 struct dma_ops_domain *dma_dom,
2545 dma_addr_t offset = paddr & ~PAGE_MASK;
2546 dma_addr_t address, start, ret;
2548 unsigned long align_mask = 0;
2551 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2554 INC_STATS_COUNTER(total_map_requests);
2557 INC_STATS_COUNTER(cross_page);
2560 align_mask = (1UL << get_order(size)) - 1;
2563 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2565 if (unlikely(address == DMA_ERROR_CODE)) {
2567 * setting next_address here will let the address
2568 * allocator only scan the new allocated range in the
2569 * first run. This is a small optimization.
2571 dma_dom->next_address = dma_dom->aperture_size;
2573 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2577 * aperture was successfully enlarged by 128 MB, try
2584 for (i = 0; i < pages; ++i) {
2585 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2586 if (ret == DMA_ERROR_CODE)
2594 ADD_STATS_COUNTER(alloced_io_mem, size);
2596 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2597 domain_flush_tlb(&dma_dom->domain);
2598 dma_dom->need_flush = false;
2599 } else if (unlikely(amd_iommu_np_cache))
2600 domain_flush_pages(&dma_dom->domain, address, size);
2607 for (--i; i >= 0; --i) {
2609 dma_ops_domain_unmap(dma_dom, start);
2612 dma_ops_free_addresses(dma_dom, address, pages);
2614 return DMA_ERROR_CODE;
2618 * Does the reverse of the __map_single function. Must be called with
2619 * the domain lock held too
2621 static void __unmap_single(struct dma_ops_domain *dma_dom,
2622 dma_addr_t dma_addr,
2626 dma_addr_t flush_addr;
2627 dma_addr_t i, start;
2630 if ((dma_addr == DMA_ERROR_CODE) ||
2631 (dma_addr + size > dma_dom->aperture_size))
2634 flush_addr = dma_addr;
2635 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2636 dma_addr &= PAGE_MASK;
2639 for (i = 0; i < pages; ++i) {
2640 dma_ops_domain_unmap(dma_dom, start);
2644 SUB_STATS_COUNTER(alloced_io_mem, size);
2646 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2648 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2649 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2650 dma_dom->need_flush = false;
2655 * The exported map_single function for dma_ops.
2657 static dma_addr_t map_page(struct device *dev, struct page *page,
2658 unsigned long offset, size_t size,
2659 enum dma_data_direction dir,
2660 struct dma_attrs *attrs)
2662 unsigned long flags;
2663 struct protection_domain *domain;
2666 phys_addr_t paddr = page_to_phys(page) + offset;
2668 INC_STATS_COUNTER(cnt_map_single);
2670 domain = get_domain(dev);
2671 if (PTR_ERR(domain) == -EINVAL)
2672 return (dma_addr_t)paddr;
2673 else if (IS_ERR(domain))
2674 return DMA_ERROR_CODE;
2676 dma_mask = *dev->dma_mask;
2678 spin_lock_irqsave(&domain->lock, flags);
2680 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2682 if (addr == DMA_ERROR_CODE)
2685 domain_flush_complete(domain);
2688 spin_unlock_irqrestore(&domain->lock, flags);
2694 * The exported unmap_single function for dma_ops.
2696 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2697 enum dma_data_direction dir, struct dma_attrs *attrs)
2699 unsigned long flags;
2700 struct protection_domain *domain;
2702 INC_STATS_COUNTER(cnt_unmap_single);
2704 domain = get_domain(dev);
2708 spin_lock_irqsave(&domain->lock, flags);
2710 __unmap_single(domain->priv, dma_addr, size, dir);
2712 domain_flush_complete(domain);
2714 spin_unlock_irqrestore(&domain->lock, flags);
2718 * This is a special map_sg function which is used if we should map a
2719 * device which is not handled by an AMD IOMMU in the system.
2721 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2722 int nelems, int dir)
2724 struct scatterlist *s;
2727 for_each_sg(sglist, s, nelems, i) {
2728 s->dma_address = (dma_addr_t)sg_phys(s);
2729 s->dma_length = s->length;
2736 * The exported map_sg function for dma_ops (handles scatter-gather
2739 static int map_sg(struct device *dev, struct scatterlist *sglist,
2740 int nelems, enum dma_data_direction dir,
2741 struct dma_attrs *attrs)
2743 unsigned long flags;
2744 struct protection_domain *domain;
2746 struct scatterlist *s;
2748 int mapped_elems = 0;
2751 INC_STATS_COUNTER(cnt_map_sg);
2753 domain = get_domain(dev);
2754 if (PTR_ERR(domain) == -EINVAL)
2755 return map_sg_no_iommu(dev, sglist, nelems, dir);
2756 else if (IS_ERR(domain))
2759 dma_mask = *dev->dma_mask;
2761 spin_lock_irqsave(&domain->lock, flags);
2763 for_each_sg(sglist, s, nelems, i) {
2766 s->dma_address = __map_single(dev, domain->priv,
2767 paddr, s->length, dir, false,
2770 if (s->dma_address) {
2771 s->dma_length = s->length;
2777 domain_flush_complete(domain);
2780 spin_unlock_irqrestore(&domain->lock, flags);
2782 return mapped_elems;
2784 for_each_sg(sglist, s, mapped_elems, i) {
2786 __unmap_single(domain->priv, s->dma_address,
2787 s->dma_length, dir);
2788 s->dma_address = s->dma_length = 0;
2797 * The exported map_sg function for dma_ops (handles scatter-gather
2800 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2801 int nelems, enum dma_data_direction dir,
2802 struct dma_attrs *attrs)
2804 unsigned long flags;
2805 struct protection_domain *domain;
2806 struct scatterlist *s;
2809 INC_STATS_COUNTER(cnt_unmap_sg);
2811 domain = get_domain(dev);
2815 spin_lock_irqsave(&domain->lock, flags);
2817 for_each_sg(sglist, s, nelems, i) {
2818 __unmap_single(domain->priv, s->dma_address,
2819 s->dma_length, dir);
2820 s->dma_address = s->dma_length = 0;
2823 domain_flush_complete(domain);
2825 spin_unlock_irqrestore(&domain->lock, flags);
2829 * The exported alloc_coherent function for dma_ops.
2831 static void *alloc_coherent(struct device *dev, size_t size,
2832 dma_addr_t *dma_addr, gfp_t flag,
2833 struct dma_attrs *attrs)
2835 unsigned long flags;
2837 struct protection_domain *domain;
2839 u64 dma_mask = dev->coherent_dma_mask;
2841 INC_STATS_COUNTER(cnt_alloc_coherent);
2843 domain = get_domain(dev);
2844 if (PTR_ERR(domain) == -EINVAL) {
2845 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2846 *dma_addr = __pa(virt_addr);
2848 } else if (IS_ERR(domain))
2851 dma_mask = dev->coherent_dma_mask;
2852 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2855 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2859 paddr = virt_to_phys(virt_addr);
2862 dma_mask = *dev->dma_mask;
2864 spin_lock_irqsave(&domain->lock, flags);
2866 *dma_addr = __map_single(dev, domain->priv, paddr,
2867 size, DMA_BIDIRECTIONAL, true, dma_mask);
2869 if (*dma_addr == DMA_ERROR_CODE) {
2870 spin_unlock_irqrestore(&domain->lock, flags);
2874 domain_flush_complete(domain);
2876 spin_unlock_irqrestore(&domain->lock, flags);
2882 free_pages((unsigned long)virt_addr, get_order(size));
2888 * The exported free_coherent function for dma_ops.
2890 static void free_coherent(struct device *dev, size_t size,
2891 void *virt_addr, dma_addr_t dma_addr,
2892 struct dma_attrs *attrs)
2894 unsigned long flags;
2895 struct protection_domain *domain;
2897 INC_STATS_COUNTER(cnt_free_coherent);
2899 domain = get_domain(dev);
2903 spin_lock_irqsave(&domain->lock, flags);
2905 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2907 domain_flush_complete(domain);
2909 spin_unlock_irqrestore(&domain->lock, flags);
2912 free_pages((unsigned long)virt_addr, get_order(size));
2916 * This function is called by the DMA layer to find out if we can handle a
2917 * particular device. It is part of the dma_ops.
2919 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2921 return check_device(dev);
2925 * The function for pre-allocating protection domains.
2927 * If the driver core informs the DMA layer if a driver grabs a device
2928 * we don't need to preallocate the protection domains anymore.
2929 * For now we have to.
2931 static void __init prealloc_protection_domains(void)
2933 struct iommu_dev_data *dev_data;
2934 struct dma_ops_domain *dma_dom;
2935 struct pci_dev *dev = NULL;
2938 for_each_pci_dev(dev) {
2940 /* Do we handle this device? */
2941 if (!check_device(&dev->dev))
2944 dev_data = get_dev_data(&dev->dev);
2945 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2946 /* Make sure passthrough domain is allocated */
2947 alloc_passthrough_domain();
2948 dev_data->passthrough = true;
2949 attach_device(&dev->dev, pt_domain);
2950 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2951 dev_name(&dev->dev));
2954 /* Is there already any domain for it? */
2955 if (domain_for_device(&dev->dev))
2958 devid = get_device_id(&dev->dev);
2960 dma_dom = dma_ops_domain_alloc();
2963 init_unity_mappings_for_device(dma_dom, devid);
2964 dma_dom->target_dev = devid;
2966 attach_device(&dev->dev, &dma_dom->domain);
2968 list_add_tail(&dma_dom->list, &iommu_pd_list);
2972 static struct dma_map_ops amd_iommu_dma_ops = {
2973 .alloc = alloc_coherent,
2974 .free = free_coherent,
2975 .map_page = map_page,
2976 .unmap_page = unmap_page,
2978 .unmap_sg = unmap_sg,
2979 .dma_supported = amd_iommu_dma_supported,
2982 static unsigned device_dma_ops_init(void)
2984 struct iommu_dev_data *dev_data;
2985 struct pci_dev *pdev = NULL;
2986 unsigned unhandled = 0;
2988 for_each_pci_dev(pdev) {
2989 if (!check_device(&pdev->dev)) {
2991 iommu_ignore_device(&pdev->dev);
2997 dev_data = get_dev_data(&pdev->dev);
2999 if (!dev_data->passthrough)
3000 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3002 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3009 * The function which clues the AMD IOMMU driver into dma_ops.
3012 void __init amd_iommu_init_api(void)
3014 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3017 int __init amd_iommu_init_dma_ops(void)
3019 struct amd_iommu *iommu;
3023 * first allocate a default protection domain for every IOMMU we
3024 * found in the system. Devices not assigned to any other
3025 * protection domain will be assigned to the default one.
3027 for_each_iommu(iommu) {
3028 iommu->default_dom = dma_ops_domain_alloc();
3029 if (iommu->default_dom == NULL)
3031 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3032 ret = iommu_init_unity_mappings(iommu);
3038 * Pre-allocate the protection domains for each device.
3040 prealloc_protection_domains();
3045 /* Make the driver finally visible to the drivers */
3046 unhandled = device_dma_ops_init();
3047 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3048 /* There are unhandled devices - initialize swiotlb for them */
3052 amd_iommu_stats_init();
3054 if (amd_iommu_unmap_flush)
3055 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3057 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3063 for_each_iommu(iommu) {
3064 if (iommu->default_dom)
3065 dma_ops_domain_free(iommu->default_dom);
3071 /*****************************************************************************
3073 * The following functions belong to the exported interface of AMD IOMMU
3075 * This interface allows access to lower level functions of the IOMMU
3076 * like protection domain handling and assignement of devices to domains
3077 * which is not possible with the dma_ops interface.
3079 *****************************************************************************/
3081 static void cleanup_domain(struct protection_domain *domain)
3083 struct iommu_dev_data *dev_data, *next;
3084 unsigned long flags;
3086 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3088 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3089 __detach_device(dev_data);
3090 atomic_set(&dev_data->bind, 0);
3093 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3096 static void protection_domain_free(struct protection_domain *domain)
3101 del_domain_from_list(domain);
3104 domain_id_free(domain->id);
3109 static struct protection_domain *protection_domain_alloc(void)
3111 struct protection_domain *domain;
3113 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3117 spin_lock_init(&domain->lock);
3118 mutex_init(&domain->api_lock);
3119 domain->id = domain_id_alloc();
3122 INIT_LIST_HEAD(&domain->dev_list);
3124 add_domain_to_list(domain);
3134 static int __init alloc_passthrough_domain(void)
3136 if (pt_domain != NULL)
3139 /* allocate passthrough domain */
3140 pt_domain = protection_domain_alloc();
3144 pt_domain->mode = PAGE_MODE_NONE;
3148 static int amd_iommu_domain_init(struct iommu_domain *dom)
3150 struct protection_domain *domain;
3152 domain = protection_domain_alloc();
3156 domain->mode = PAGE_MODE_3_LEVEL;
3157 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3158 if (!domain->pt_root)
3161 domain->iommu_domain = dom;
3165 dom->geometry.aperture_start = 0;
3166 dom->geometry.aperture_end = ~0ULL;
3167 dom->geometry.force_aperture = true;
3172 protection_domain_free(domain);
3177 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3179 struct protection_domain *domain = dom->priv;
3184 if (domain->dev_cnt > 0)
3185 cleanup_domain(domain);
3187 BUG_ON(domain->dev_cnt != 0);
3189 if (domain->mode != PAGE_MODE_NONE)
3190 free_pagetable(domain);
3192 if (domain->flags & PD_IOMMUV2_MASK)
3193 free_gcr3_table(domain);
3195 protection_domain_free(domain);
3200 static void amd_iommu_detach_device(struct iommu_domain *dom,
3203 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3204 struct amd_iommu *iommu;
3207 if (!check_device(dev))
3210 devid = get_device_id(dev);
3212 if (dev_data->domain != NULL)
3215 iommu = amd_iommu_rlookup_table[devid];
3219 iommu_completion_wait(iommu);
3222 static int amd_iommu_attach_device(struct iommu_domain *dom,
3225 struct protection_domain *domain = dom->priv;
3226 struct iommu_dev_data *dev_data;
3227 struct amd_iommu *iommu;
3230 if (!check_device(dev))
3233 dev_data = dev->archdata.iommu;
3235 iommu = amd_iommu_rlookup_table[dev_data->devid];
3239 if (dev_data->domain)
3242 ret = attach_device(dev, domain);
3244 iommu_completion_wait(iommu);
3249 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3250 phys_addr_t paddr, size_t page_size, int iommu_prot)
3252 struct protection_domain *domain = dom->priv;
3256 if (domain->mode == PAGE_MODE_NONE)
3259 if (iommu_prot & IOMMU_READ)
3260 prot |= IOMMU_PROT_IR;
3261 if (iommu_prot & IOMMU_WRITE)
3262 prot |= IOMMU_PROT_IW;
3264 mutex_lock(&domain->api_lock);
3265 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3266 mutex_unlock(&domain->api_lock);
3271 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3274 struct protection_domain *domain = dom->priv;
3277 if (domain->mode == PAGE_MODE_NONE)
3280 mutex_lock(&domain->api_lock);
3281 unmap_size = iommu_unmap_page(domain, iova, page_size);
3282 mutex_unlock(&domain->api_lock);
3284 domain_flush_tlb_pde(domain);
3289 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3292 struct protection_domain *domain = dom->priv;
3293 unsigned long offset_mask;
3297 if (domain->mode == PAGE_MODE_NONE)
3300 pte = fetch_pte(domain, iova);
3302 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3305 if (PM_PTE_LEVEL(*pte) == 0)
3306 offset_mask = PAGE_SIZE - 1;
3308 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3310 __pte = *pte & PM_ADDR_MASK;
3311 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3316 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3320 case IOMMU_CAP_CACHE_COHERENCY:
3327 static struct iommu_ops amd_iommu_ops = {
3328 .domain_init = amd_iommu_domain_init,
3329 .domain_destroy = amd_iommu_domain_destroy,
3330 .attach_dev = amd_iommu_attach_device,
3331 .detach_dev = amd_iommu_detach_device,
3332 .map = amd_iommu_map,
3333 .unmap = amd_iommu_unmap,
3334 .iova_to_phys = amd_iommu_iova_to_phys,
3335 .domain_has_cap = amd_iommu_domain_has_cap,
3336 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3339 /*****************************************************************************
3341 * The next functions do a basic initialization of IOMMU for pass through
3344 * In passthrough mode the IOMMU is initialized and enabled but not used for
3345 * DMA-API translation.
3347 *****************************************************************************/
3349 int __init amd_iommu_init_passthrough(void)
3351 struct iommu_dev_data *dev_data;
3352 struct pci_dev *dev = NULL;
3353 struct amd_iommu *iommu;
3357 ret = alloc_passthrough_domain();
3361 for_each_pci_dev(dev) {
3362 if (!check_device(&dev->dev))
3365 dev_data = get_dev_data(&dev->dev);
3366 dev_data->passthrough = true;
3368 devid = get_device_id(&dev->dev);
3370 iommu = amd_iommu_rlookup_table[devid];
3374 attach_device(&dev->dev, pt_domain);
3377 amd_iommu_stats_init();
3379 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3384 /* IOMMUv2 specific functions */
3385 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3387 return atomic_notifier_chain_register(&ppr_notifier, nb);
3389 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3391 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3393 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3395 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3397 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3399 struct protection_domain *domain = dom->priv;
3400 unsigned long flags;
3402 spin_lock_irqsave(&domain->lock, flags);
3404 /* Update data structure */
3405 domain->mode = PAGE_MODE_NONE;
3406 domain->updated = true;
3408 /* Make changes visible to IOMMUs */
3409 update_domain(domain);
3411 /* Page-table is not visible to IOMMU anymore, so free it */
3412 free_pagetable(domain);
3414 spin_unlock_irqrestore(&domain->lock, flags);
3416 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3418 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3420 struct protection_domain *domain = dom->priv;
3421 unsigned long flags;
3424 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3427 /* Number of GCR3 table levels required */
3428 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3431 if (levels > amd_iommu_max_glx_val)
3434 spin_lock_irqsave(&domain->lock, flags);
3437 * Save us all sanity checks whether devices already in the
3438 * domain support IOMMUv2. Just force that the domain has no
3439 * devices attached when it is switched into IOMMUv2 mode.
3442 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3446 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3447 if (domain->gcr3_tbl == NULL)
3450 domain->glx = levels;
3451 domain->flags |= PD_IOMMUV2_MASK;
3452 domain->updated = true;
3454 update_domain(domain);
3459 spin_unlock_irqrestore(&domain->lock, flags);
3463 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3465 static int __flush_pasid(struct protection_domain *domain, int pasid,
3466 u64 address, bool size)
3468 struct iommu_dev_data *dev_data;
3469 struct iommu_cmd cmd;
3472 if (!(domain->flags & PD_IOMMUV2_MASK))
3475 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3478 * IOMMU TLB needs to be flushed before Device TLB to
3479 * prevent device TLB refill from IOMMU TLB
3481 for (i = 0; i < amd_iommus_present; ++i) {
3482 if (domain->dev_iommu[i] == 0)
3485 ret = iommu_queue_command(amd_iommus[i], &cmd);
3490 /* Wait until IOMMU TLB flushes are complete */
3491 domain_flush_complete(domain);
3493 /* Now flush device TLBs */
3494 list_for_each_entry(dev_data, &domain->dev_list, list) {
3495 struct amd_iommu *iommu;
3498 BUG_ON(!dev_data->ats.enabled);
3500 qdep = dev_data->ats.qdep;
3501 iommu = amd_iommu_rlookup_table[dev_data->devid];
3503 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3504 qdep, address, size);
3506 ret = iommu_queue_command(iommu, &cmd);
3511 /* Wait until all device TLBs are flushed */
3512 domain_flush_complete(domain);
3521 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3524 INC_STATS_COUNTER(invalidate_iotlb);
3526 return __flush_pasid(domain, pasid, address, false);
3529 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3532 struct protection_domain *domain = dom->priv;
3533 unsigned long flags;
3536 spin_lock_irqsave(&domain->lock, flags);
3537 ret = __amd_iommu_flush_page(domain, pasid, address);
3538 spin_unlock_irqrestore(&domain->lock, flags);
3542 EXPORT_SYMBOL(amd_iommu_flush_page);
3544 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3546 INC_STATS_COUNTER(invalidate_iotlb_all);
3548 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3552 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3554 struct protection_domain *domain = dom->priv;
3555 unsigned long flags;
3558 spin_lock_irqsave(&domain->lock, flags);
3559 ret = __amd_iommu_flush_tlb(domain, pasid);
3560 spin_unlock_irqrestore(&domain->lock, flags);
3564 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3566 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3573 index = (pasid >> (9 * level)) & 0x1ff;
3579 if (!(*pte & GCR3_VALID)) {
3583 root = (void *)get_zeroed_page(GFP_ATOMIC);
3587 *pte = __pa(root) | GCR3_VALID;
3590 root = __va(*pte & PAGE_MASK);
3598 static int __set_gcr3(struct protection_domain *domain, int pasid,
3603 if (domain->mode != PAGE_MODE_NONE)
3606 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3610 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3612 return __amd_iommu_flush_tlb(domain, pasid);
3615 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3619 if (domain->mode != PAGE_MODE_NONE)
3622 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3628 return __amd_iommu_flush_tlb(domain, pasid);
3631 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3634 struct protection_domain *domain = dom->priv;
3635 unsigned long flags;
3638 spin_lock_irqsave(&domain->lock, flags);
3639 ret = __set_gcr3(domain, pasid, cr3);
3640 spin_unlock_irqrestore(&domain->lock, flags);
3644 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3646 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3648 struct protection_domain *domain = dom->priv;
3649 unsigned long flags;
3652 spin_lock_irqsave(&domain->lock, flags);
3653 ret = __clear_gcr3(domain, pasid);
3654 spin_unlock_irqrestore(&domain->lock, flags);
3658 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3660 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3661 int status, int tag)
3663 struct iommu_dev_data *dev_data;
3664 struct amd_iommu *iommu;
3665 struct iommu_cmd cmd;
3667 INC_STATS_COUNTER(complete_ppr);
3669 dev_data = get_dev_data(&pdev->dev);
3670 iommu = amd_iommu_rlookup_table[dev_data->devid];
3672 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3673 tag, dev_data->pri_tlp);
3675 return iommu_queue_command(iommu, &cmd);
3677 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3679 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3681 struct protection_domain *domain;
3683 domain = get_domain(&pdev->dev);
3687 /* Only return IOMMUv2 domains */
3688 if (!(domain->flags & PD_IOMMUV2_MASK))
3691 return domain->iommu_domain;
3693 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3695 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3697 struct iommu_dev_data *dev_data;
3699 if (!amd_iommu_v2_supported())
3702 dev_data = get_dev_data(&pdev->dev);
3703 dev_data->errata |= (1 << erratum);
3705 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3707 int amd_iommu_device_info(struct pci_dev *pdev,
3708 struct amd_iommu_device_info *info)
3713 if (pdev == NULL || info == NULL)
3716 if (!amd_iommu_v2_supported())
3719 memset(info, 0, sizeof(*info));
3721 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3723 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3725 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3727 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3729 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3733 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3734 max_pasids = min(max_pasids, (1 << 20));
3736 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3737 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3739 features = pci_pasid_features(pdev);
3740 if (features & PCI_PASID_CAP_EXEC)
3741 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3742 if (features & PCI_PASID_CAP_PRIV)
3743 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3748 EXPORT_SYMBOL(amd_iommu_device_info);