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 <linux/irq.h>
35 #include <linux/msi.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/io_apic.h>
39 #include <asm/hw_irq.h>
40 #include <asm/msidef.h>
41 #include <asm/proto.h>
42 #include <asm/iommu.h>
46 #include "amd_iommu_proto.h"
47 #include "amd_iommu_types.h"
48 #include "irq_remapping.h"
50 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
52 #define LOOP_TIMEOUT 100000
55 * This bitmap is used to advertise the page sizes our hardware support
56 * to the IOMMU core, which will then use this information to split
57 * physically contiguous memory regions it is mapping into page sizes
60 * Traditionally the IOMMU core just handed us the mappings directly,
61 * after making sure the size is an order of a 4KiB page and that the
62 * mapping has natural alignment.
64 * To retain this behavior, we currently advertise that we support
65 * all page sizes that are an order of 4KiB.
67 * If at some point we'd like to utilize the IOMMU core's new behavior,
68 * we could change this to advertise the real page sizes we support.
70 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
72 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
74 /* A list of preallocated protection domains */
75 static LIST_HEAD(iommu_pd_list);
76 static DEFINE_SPINLOCK(iommu_pd_list_lock);
78 /* List of all available dev_data structures */
79 static LIST_HEAD(dev_data_list);
80 static DEFINE_SPINLOCK(dev_data_list_lock);
82 LIST_HEAD(ioapic_map);
86 * Domain for untranslated devices - only allocated
87 * if iommu=pt passed on kernel cmd line.
89 static struct protection_domain *pt_domain;
91 static struct iommu_ops amd_iommu_ops;
93 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
94 int amd_iommu_max_glx_val = -1;
96 static struct dma_map_ops amd_iommu_dma_ops;
99 * general struct to manage commands send to an IOMMU
105 struct kmem_cache *amd_iommu_irq_cache;
107 static void update_domain(struct protection_domain *domain);
108 static int __init alloc_passthrough_domain(void);
110 /****************************************************************************
114 ****************************************************************************/
116 static struct iommu_dev_data *alloc_dev_data(u16 devid)
118 struct iommu_dev_data *dev_data;
121 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
125 dev_data->devid = devid;
126 atomic_set(&dev_data->bind, 0);
128 spin_lock_irqsave(&dev_data_list_lock, flags);
129 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
130 spin_unlock_irqrestore(&dev_data_list_lock, flags);
135 static void free_dev_data(struct iommu_dev_data *dev_data)
139 spin_lock_irqsave(&dev_data_list_lock, flags);
140 list_del(&dev_data->dev_data_list);
141 spin_unlock_irqrestore(&dev_data_list_lock, flags);
146 static struct iommu_dev_data *search_dev_data(u16 devid)
148 struct iommu_dev_data *dev_data;
151 spin_lock_irqsave(&dev_data_list_lock, flags);
152 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
153 if (dev_data->devid == devid)
160 spin_unlock_irqrestore(&dev_data_list_lock, flags);
165 static struct iommu_dev_data *find_dev_data(u16 devid)
167 struct iommu_dev_data *dev_data;
169 dev_data = search_dev_data(devid);
171 if (dev_data == NULL)
172 dev_data = alloc_dev_data(devid);
177 static inline u16 get_device_id(struct device *dev)
179 struct pci_dev *pdev = to_pci_dev(dev);
181 return calc_devid(pdev->bus->number, pdev->devfn);
184 static struct iommu_dev_data *get_dev_data(struct device *dev)
186 return dev->archdata.iommu;
189 static bool pci_iommuv2_capable(struct pci_dev *pdev)
191 static const int caps[] = {
194 PCI_EXT_CAP_ID_PASID,
198 for (i = 0; i < 3; ++i) {
199 pos = pci_find_ext_capability(pdev, caps[i]);
207 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
209 struct iommu_dev_data *dev_data;
211 dev_data = get_dev_data(&pdev->dev);
213 return dev_data->errata & (1 << erratum) ? true : false;
217 * In this function the list of preallocated protection domains is traversed to
218 * find the domain for a specific device
220 static struct dma_ops_domain *find_protection_domain(u16 devid)
222 struct dma_ops_domain *entry, *ret = NULL;
224 u16 alias = amd_iommu_alias_table[devid];
226 if (list_empty(&iommu_pd_list))
229 spin_lock_irqsave(&iommu_pd_list_lock, flags);
231 list_for_each_entry(entry, &iommu_pd_list, list) {
232 if (entry->target_dev == devid ||
233 entry->target_dev == alias) {
239 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
245 * This function checks if the driver got a valid device from the caller to
246 * avoid dereferencing invalid pointers.
248 static bool check_device(struct device *dev)
252 if (!dev || !dev->dma_mask)
255 /* No device or no PCI device */
256 if (dev->bus != &pci_bus_type)
259 devid = get_device_id(dev);
261 /* Out of our scope? */
262 if (devid > amd_iommu_last_bdf)
265 if (amd_iommu_rlookup_table[devid] == NULL)
271 static void swap_pci_ref(struct pci_dev **from, struct pci_dev *to)
277 static struct pci_bus *find_hosted_bus(struct pci_bus *bus)
280 if (!pci_is_root_bus(bus))
283 return ERR_PTR(-ENODEV);
289 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
291 static struct pci_dev *get_isolation_root(struct pci_dev *pdev)
293 struct pci_dev *dma_pdev = pdev;
295 /* Account for quirked devices */
296 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
299 * If it's a multifunction device that does not support our
300 * required ACS flags, add to the same group as function 0.
302 if (dma_pdev->multifunction &&
303 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
304 swap_pci_ref(&dma_pdev,
305 pci_get_slot(dma_pdev->bus,
306 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
310 * Devices on the root bus go through the iommu. If that's not us,
311 * find the next upstream device and test ACS up to the root bus.
312 * Finding the next device may require skipping virtual buses.
314 while (!pci_is_root_bus(dma_pdev->bus)) {
315 struct pci_bus *bus = find_hosted_bus(dma_pdev->bus);
319 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
322 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
328 static int use_pdev_iommu_group(struct pci_dev *pdev, struct device *dev)
330 struct iommu_group *group = iommu_group_get(&pdev->dev);
334 group = iommu_group_alloc();
336 return PTR_ERR(group);
338 WARN_ON(&pdev->dev != dev);
341 ret = iommu_group_add_device(group, dev);
342 iommu_group_put(group);
346 static int init_iommu_group(struct device *dev)
348 struct iommu_dev_data *dev_data;
349 struct iommu_group *group;
350 struct pci_dev *dma_pdev = NULL;
353 group = iommu_group_get(dev);
355 iommu_group_put(group);
359 dev_data = find_dev_data(get_device_id(dev));
363 if (dev_data->alias_data) {
366 alias = amd_iommu_alias_table[dev_data->devid];
367 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
371 dma_pdev = pci_dev_get(to_pci_dev(dev));
373 dma_pdev = get_isolation_root(dma_pdev);
374 ret = use_pdev_iommu_group(dma_pdev, dev);
375 pci_dev_put(dma_pdev);
379 static int iommu_init_device(struct device *dev)
381 struct pci_dev *pdev = to_pci_dev(dev);
382 struct iommu_dev_data *dev_data;
386 if (dev->archdata.iommu)
389 dev_data = find_dev_data(get_device_id(dev));
393 alias = amd_iommu_alias_table[dev_data->devid];
394 if (alias != dev_data->devid) {
395 struct iommu_dev_data *alias_data;
397 alias_data = find_dev_data(alias);
398 if (alias_data == NULL) {
399 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
401 free_dev_data(dev_data);
404 dev_data->alias_data = alias_data;
407 ret = init_iommu_group(dev);
411 if (pci_iommuv2_capable(pdev)) {
412 struct amd_iommu *iommu;
414 iommu = amd_iommu_rlookup_table[dev_data->devid];
415 dev_data->iommu_v2 = iommu->is_iommu_v2;
418 dev->archdata.iommu = dev_data;
423 static void iommu_ignore_device(struct device *dev)
427 devid = get_device_id(dev);
428 alias = amd_iommu_alias_table[devid];
430 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
431 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
433 amd_iommu_rlookup_table[devid] = NULL;
434 amd_iommu_rlookup_table[alias] = NULL;
437 static void iommu_uninit_device(struct device *dev)
439 iommu_group_remove_device(dev);
442 * Nothing to do here - we keep dev_data around for unplugged devices
443 * and reuse it when the device is re-plugged - not doing so would
444 * introduce a ton of races.
448 void __init amd_iommu_uninit_devices(void)
450 struct iommu_dev_data *dev_data, *n;
451 struct pci_dev *pdev = NULL;
453 for_each_pci_dev(pdev) {
455 if (!check_device(&pdev->dev))
458 iommu_uninit_device(&pdev->dev);
461 /* Free all of our dev_data structures */
462 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
463 free_dev_data(dev_data);
466 int __init amd_iommu_init_devices(void)
468 struct pci_dev *pdev = NULL;
471 for_each_pci_dev(pdev) {
473 if (!check_device(&pdev->dev))
476 ret = iommu_init_device(&pdev->dev);
477 if (ret == -ENOTSUPP)
478 iommu_ignore_device(&pdev->dev);
487 amd_iommu_uninit_devices();
491 #ifdef CONFIG_AMD_IOMMU_STATS
494 * Initialization code for statistics collection
497 DECLARE_STATS_COUNTER(compl_wait);
498 DECLARE_STATS_COUNTER(cnt_map_single);
499 DECLARE_STATS_COUNTER(cnt_unmap_single);
500 DECLARE_STATS_COUNTER(cnt_map_sg);
501 DECLARE_STATS_COUNTER(cnt_unmap_sg);
502 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
503 DECLARE_STATS_COUNTER(cnt_free_coherent);
504 DECLARE_STATS_COUNTER(cross_page);
505 DECLARE_STATS_COUNTER(domain_flush_single);
506 DECLARE_STATS_COUNTER(domain_flush_all);
507 DECLARE_STATS_COUNTER(alloced_io_mem);
508 DECLARE_STATS_COUNTER(total_map_requests);
509 DECLARE_STATS_COUNTER(complete_ppr);
510 DECLARE_STATS_COUNTER(invalidate_iotlb);
511 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
512 DECLARE_STATS_COUNTER(pri_requests);
514 static struct dentry *stats_dir;
515 static struct dentry *de_fflush;
517 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
519 if (stats_dir == NULL)
522 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
526 static void amd_iommu_stats_init(void)
528 stats_dir = debugfs_create_dir("amd-iommu", NULL);
529 if (stats_dir == NULL)
532 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
533 &amd_iommu_unmap_flush);
535 amd_iommu_stats_add(&compl_wait);
536 amd_iommu_stats_add(&cnt_map_single);
537 amd_iommu_stats_add(&cnt_unmap_single);
538 amd_iommu_stats_add(&cnt_map_sg);
539 amd_iommu_stats_add(&cnt_unmap_sg);
540 amd_iommu_stats_add(&cnt_alloc_coherent);
541 amd_iommu_stats_add(&cnt_free_coherent);
542 amd_iommu_stats_add(&cross_page);
543 amd_iommu_stats_add(&domain_flush_single);
544 amd_iommu_stats_add(&domain_flush_all);
545 amd_iommu_stats_add(&alloced_io_mem);
546 amd_iommu_stats_add(&total_map_requests);
547 amd_iommu_stats_add(&complete_ppr);
548 amd_iommu_stats_add(&invalidate_iotlb);
549 amd_iommu_stats_add(&invalidate_iotlb_all);
550 amd_iommu_stats_add(&pri_requests);
555 /****************************************************************************
557 * Interrupt handling functions
559 ****************************************************************************/
561 static void dump_dte_entry(u16 devid)
565 for (i = 0; i < 4; ++i)
566 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
567 amd_iommu_dev_table[devid].data[i]);
570 static void dump_command(unsigned long phys_addr)
572 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
575 for (i = 0; i < 4; ++i)
576 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
579 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
581 int type, devid, domid, flags;
582 volatile u32 *event = __evt;
587 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
588 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
589 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
590 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
591 address = (u64)(((u64)event[3]) << 32) | event[2];
594 /* Did we hit the erratum? */
595 if (++count == LOOP_TIMEOUT) {
596 pr_err("AMD-Vi: No event written to event log\n");
603 printk(KERN_ERR "AMD-Vi: Event logged [");
606 case EVENT_TYPE_ILL_DEV:
607 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
608 "address=0x%016llx flags=0x%04x]\n",
609 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
611 dump_dte_entry(devid);
613 case EVENT_TYPE_IO_FAULT:
614 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
615 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
616 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
617 domid, address, flags);
619 case EVENT_TYPE_DEV_TAB_ERR:
620 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
621 "address=0x%016llx flags=0x%04x]\n",
622 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
625 case EVENT_TYPE_PAGE_TAB_ERR:
626 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
627 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
628 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
629 domid, address, flags);
631 case EVENT_TYPE_ILL_CMD:
632 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
633 dump_command(address);
635 case EVENT_TYPE_CMD_HARD_ERR:
636 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
637 "flags=0x%04x]\n", address, flags);
639 case EVENT_TYPE_IOTLB_INV_TO:
640 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
641 "address=0x%016llx]\n",
642 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
645 case EVENT_TYPE_INV_DEV_REQ:
646 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
647 "address=0x%016llx flags=0x%04x]\n",
648 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
652 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
655 memset(__evt, 0, 4 * sizeof(u32));
658 static void iommu_poll_events(struct amd_iommu *iommu)
663 spin_lock_irqsave(&iommu->lock, flags);
665 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
666 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
668 while (head != tail) {
669 iommu_print_event(iommu, iommu->evt_buf + head);
670 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
673 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
675 spin_unlock_irqrestore(&iommu->lock, flags);
678 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
680 struct amd_iommu_fault fault;
682 INC_STATS_COUNTER(pri_requests);
684 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
685 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
689 fault.address = raw[1];
690 fault.pasid = PPR_PASID(raw[0]);
691 fault.device_id = PPR_DEVID(raw[0]);
692 fault.tag = PPR_TAG(raw[0]);
693 fault.flags = PPR_FLAGS(raw[0]);
695 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
698 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
703 if (iommu->ppr_log == NULL)
706 /* enable ppr interrupts again */
707 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
709 spin_lock_irqsave(&iommu->lock, flags);
711 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
712 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
714 while (head != tail) {
719 raw = (u64 *)(iommu->ppr_log + head);
722 * Hardware bug: Interrupt may arrive before the entry is
723 * written to memory. If this happens we need to wait for the
726 for (i = 0; i < LOOP_TIMEOUT; ++i) {
727 if (PPR_REQ_TYPE(raw[0]) != 0)
732 /* Avoid memcpy function-call overhead */
737 * To detect the hardware bug we need to clear the entry
740 raw[0] = raw[1] = 0UL;
742 /* Update head pointer of hardware ring-buffer */
743 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
744 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
747 * Release iommu->lock because ppr-handling might need to
750 spin_unlock_irqrestore(&iommu->lock, flags);
752 /* Handle PPR entry */
753 iommu_handle_ppr_entry(iommu, entry);
755 spin_lock_irqsave(&iommu->lock, flags);
757 /* Refresh ring-buffer information */
758 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
759 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
762 spin_unlock_irqrestore(&iommu->lock, flags);
765 irqreturn_t amd_iommu_int_thread(int irq, void *data)
767 struct amd_iommu *iommu;
769 for_each_iommu(iommu) {
770 iommu_poll_events(iommu);
771 iommu_poll_ppr_log(iommu);
777 irqreturn_t amd_iommu_int_handler(int irq, void *data)
779 return IRQ_WAKE_THREAD;
782 /****************************************************************************
784 * IOMMU command queuing functions
786 ****************************************************************************/
788 static int wait_on_sem(volatile u64 *sem)
792 while (*sem == 0 && i < LOOP_TIMEOUT) {
797 if (i == LOOP_TIMEOUT) {
798 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
805 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
806 struct iommu_cmd *cmd,
811 target = iommu->cmd_buf + tail;
812 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
814 /* Copy command to buffer */
815 memcpy(target, cmd, sizeof(*cmd));
817 /* Tell the IOMMU about it */
818 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
821 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
823 WARN_ON(address & 0x7ULL);
825 memset(cmd, 0, sizeof(*cmd));
826 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
827 cmd->data[1] = upper_32_bits(__pa(address));
829 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
832 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
834 memset(cmd, 0, sizeof(*cmd));
835 cmd->data[0] = devid;
836 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
839 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
840 size_t size, u16 domid, int pde)
845 pages = iommu_num_pages(address, size, PAGE_SIZE);
850 * If we have to flush more than one page, flush all
851 * TLB entries for this domain
853 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
857 address &= PAGE_MASK;
859 memset(cmd, 0, sizeof(*cmd));
860 cmd->data[1] |= domid;
861 cmd->data[2] = lower_32_bits(address);
862 cmd->data[3] = upper_32_bits(address);
863 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
864 if (s) /* size bit - we flush more than one 4kb page */
865 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
866 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
867 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
870 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
871 u64 address, size_t size)
876 pages = iommu_num_pages(address, size, PAGE_SIZE);
881 * If we have to flush more than one page, flush all
882 * TLB entries for this domain
884 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
888 address &= PAGE_MASK;
890 memset(cmd, 0, sizeof(*cmd));
891 cmd->data[0] = devid;
892 cmd->data[0] |= (qdep & 0xff) << 24;
893 cmd->data[1] = devid;
894 cmd->data[2] = lower_32_bits(address);
895 cmd->data[3] = upper_32_bits(address);
896 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
898 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
901 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
902 u64 address, bool size)
904 memset(cmd, 0, sizeof(*cmd));
906 address &= ~(0xfffULL);
908 cmd->data[0] = pasid & PASID_MASK;
909 cmd->data[1] = domid;
910 cmd->data[2] = lower_32_bits(address);
911 cmd->data[3] = upper_32_bits(address);
912 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
913 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
915 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
916 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
919 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
920 int qdep, u64 address, bool size)
922 memset(cmd, 0, sizeof(*cmd));
924 address &= ~(0xfffULL);
926 cmd->data[0] = devid;
927 cmd->data[0] |= (pasid & 0xff) << 16;
928 cmd->data[0] |= (qdep & 0xff) << 24;
929 cmd->data[1] = devid;
930 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
931 cmd->data[2] = lower_32_bits(address);
932 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
933 cmd->data[3] = upper_32_bits(address);
935 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
936 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
939 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
940 int status, int tag, bool gn)
942 memset(cmd, 0, sizeof(*cmd));
944 cmd->data[0] = devid;
946 cmd->data[1] = pasid & PASID_MASK;
947 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
949 cmd->data[3] = tag & 0x1ff;
950 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
952 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
955 static void build_inv_all(struct iommu_cmd *cmd)
957 memset(cmd, 0, sizeof(*cmd));
958 CMD_SET_TYPE(cmd, CMD_INV_ALL);
961 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
963 memset(cmd, 0, sizeof(*cmd));
964 cmd->data[0] = devid;
965 CMD_SET_TYPE(cmd, CMD_INV_IRT);
969 * Writes the command to the IOMMUs command buffer and informs the
970 * hardware about the new command.
972 static int iommu_queue_command_sync(struct amd_iommu *iommu,
973 struct iommu_cmd *cmd,
976 u32 left, tail, head, next_tail;
979 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
982 spin_lock_irqsave(&iommu->lock, flags);
984 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
985 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
986 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
987 left = (head - next_tail) % iommu->cmd_buf_size;
990 struct iommu_cmd sync_cmd;
991 volatile u64 sem = 0;
994 build_completion_wait(&sync_cmd, (u64)&sem);
995 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
997 spin_unlock_irqrestore(&iommu->lock, flags);
999 if ((ret = wait_on_sem(&sem)) != 0)
1005 copy_cmd_to_buffer(iommu, cmd, tail);
1007 /* We need to sync now to make sure all commands are processed */
1008 iommu->need_sync = sync;
1010 spin_unlock_irqrestore(&iommu->lock, flags);
1015 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1017 return iommu_queue_command_sync(iommu, cmd, true);
1021 * This function queues a completion wait command into the command
1022 * buffer of an IOMMU
1024 static int iommu_completion_wait(struct amd_iommu *iommu)
1026 struct iommu_cmd cmd;
1027 volatile u64 sem = 0;
1030 if (!iommu->need_sync)
1033 build_completion_wait(&cmd, (u64)&sem);
1035 ret = iommu_queue_command_sync(iommu, &cmd, false);
1039 return wait_on_sem(&sem);
1042 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1044 struct iommu_cmd cmd;
1046 build_inv_dte(&cmd, devid);
1048 return iommu_queue_command(iommu, &cmd);
1051 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1055 for (devid = 0; devid <= 0xffff; ++devid)
1056 iommu_flush_dte(iommu, devid);
1058 iommu_completion_wait(iommu);
1062 * This function uses heavy locking and may disable irqs for some time. But
1063 * this is no issue because it is only called during resume.
1065 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1069 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1070 struct iommu_cmd cmd;
1071 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1073 iommu_queue_command(iommu, &cmd);
1076 iommu_completion_wait(iommu);
1079 static void iommu_flush_all(struct amd_iommu *iommu)
1081 struct iommu_cmd cmd;
1083 build_inv_all(&cmd);
1085 iommu_queue_command(iommu, &cmd);
1086 iommu_completion_wait(iommu);
1089 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1091 struct iommu_cmd cmd;
1093 build_inv_irt(&cmd, devid);
1095 iommu_queue_command(iommu, &cmd);
1098 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1102 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1103 iommu_flush_irt(iommu, devid);
1105 iommu_completion_wait(iommu);
1108 void iommu_flush_all_caches(struct amd_iommu *iommu)
1110 if (iommu_feature(iommu, FEATURE_IA)) {
1111 iommu_flush_all(iommu);
1113 iommu_flush_dte_all(iommu);
1114 iommu_flush_irt_all(iommu);
1115 iommu_flush_tlb_all(iommu);
1120 * Command send function for flushing on-device TLB
1122 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1123 u64 address, size_t size)
1125 struct amd_iommu *iommu;
1126 struct iommu_cmd cmd;
1129 qdep = dev_data->ats.qdep;
1130 iommu = amd_iommu_rlookup_table[dev_data->devid];
1132 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1134 return iommu_queue_command(iommu, &cmd);
1138 * Command send function for invalidating a device table entry
1140 static int device_flush_dte(struct iommu_dev_data *dev_data)
1142 struct amd_iommu *iommu;
1145 iommu = amd_iommu_rlookup_table[dev_data->devid];
1147 ret = iommu_flush_dte(iommu, dev_data->devid);
1151 if (dev_data->ats.enabled)
1152 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1158 * TLB invalidation function which is called from the mapping functions.
1159 * It invalidates a single PTE if the range to flush is within a single
1160 * page. Otherwise it flushes the whole TLB of the IOMMU.
1162 static void __domain_flush_pages(struct protection_domain *domain,
1163 u64 address, size_t size, int pde)
1165 struct iommu_dev_data *dev_data;
1166 struct iommu_cmd cmd;
1169 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1171 for (i = 0; i < amd_iommus_present; ++i) {
1172 if (!domain->dev_iommu[i])
1176 * Devices of this domain are behind this IOMMU
1177 * We need a TLB flush
1179 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1182 list_for_each_entry(dev_data, &domain->dev_list, list) {
1184 if (!dev_data->ats.enabled)
1187 ret |= device_flush_iotlb(dev_data, address, size);
1193 static void domain_flush_pages(struct protection_domain *domain,
1194 u64 address, size_t size)
1196 __domain_flush_pages(domain, address, size, 0);
1199 /* Flush the whole IO/TLB for a given protection domain */
1200 static void domain_flush_tlb(struct protection_domain *domain)
1202 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1205 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1206 static void domain_flush_tlb_pde(struct protection_domain *domain)
1208 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1211 static void domain_flush_complete(struct protection_domain *domain)
1215 for (i = 0; i < amd_iommus_present; ++i) {
1216 if (!domain->dev_iommu[i])
1220 * Devices of this domain are behind this IOMMU
1221 * We need to wait for completion of all commands.
1223 iommu_completion_wait(amd_iommus[i]);
1229 * This function flushes the DTEs for all devices in domain
1231 static void domain_flush_devices(struct protection_domain *domain)
1233 struct iommu_dev_data *dev_data;
1235 list_for_each_entry(dev_data, &domain->dev_list, list)
1236 device_flush_dte(dev_data);
1239 /****************************************************************************
1241 * The functions below are used the create the page table mappings for
1242 * unity mapped regions.
1244 ****************************************************************************/
1247 * This function is used to add another level to an IO page table. Adding
1248 * another level increases the size of the address space by 9 bits to a size up
1251 static bool increase_address_space(struct protection_domain *domain,
1256 if (domain->mode == PAGE_MODE_6_LEVEL)
1257 /* address space already 64 bit large */
1260 pte = (void *)get_zeroed_page(gfp);
1264 *pte = PM_LEVEL_PDE(domain->mode,
1265 virt_to_phys(domain->pt_root));
1266 domain->pt_root = pte;
1268 domain->updated = true;
1273 static u64 *alloc_pte(struct protection_domain *domain,
1274 unsigned long address,
1275 unsigned long page_size,
1282 BUG_ON(!is_power_of_2(page_size));
1284 while (address > PM_LEVEL_SIZE(domain->mode))
1285 increase_address_space(domain, gfp);
1287 level = domain->mode - 1;
1288 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1289 address = PAGE_SIZE_ALIGN(address, page_size);
1290 end_lvl = PAGE_SIZE_LEVEL(page_size);
1292 while (level > end_lvl) {
1293 if (!IOMMU_PTE_PRESENT(*pte)) {
1294 page = (u64 *)get_zeroed_page(gfp);
1297 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1300 /* No level skipping support yet */
1301 if (PM_PTE_LEVEL(*pte) != level)
1306 pte = IOMMU_PTE_PAGE(*pte);
1308 if (pte_page && level == end_lvl)
1311 pte = &pte[PM_LEVEL_INDEX(level, address)];
1318 * This function checks if there is a PTE for a given dma address. If
1319 * there is one, it returns the pointer to it.
1321 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1326 if (address > PM_LEVEL_SIZE(domain->mode))
1329 level = domain->mode - 1;
1330 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1335 if (!IOMMU_PTE_PRESENT(*pte))
1339 if (PM_PTE_LEVEL(*pte) == 0x07) {
1340 unsigned long pte_mask, __pte;
1343 * If we have a series of large PTEs, make
1344 * sure to return a pointer to the first one.
1346 pte_mask = PTE_PAGE_SIZE(*pte);
1347 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1348 __pte = ((unsigned long)pte) & pte_mask;
1350 return (u64 *)__pte;
1353 /* No level skipping support yet */
1354 if (PM_PTE_LEVEL(*pte) != level)
1359 /* Walk to the next level */
1360 pte = IOMMU_PTE_PAGE(*pte);
1361 pte = &pte[PM_LEVEL_INDEX(level, address)];
1368 * Generic mapping functions. It maps a physical address into a DMA
1369 * address space. It allocates the page table pages if necessary.
1370 * In the future it can be extended to a generic mapping function
1371 * supporting all features of AMD IOMMU page tables like level skipping
1372 * and full 64 bit address spaces.
1374 static int iommu_map_page(struct protection_domain *dom,
1375 unsigned long bus_addr,
1376 unsigned long phys_addr,
1378 unsigned long page_size)
1383 if (!(prot & IOMMU_PROT_MASK))
1386 bus_addr = PAGE_ALIGN(bus_addr);
1387 phys_addr = PAGE_ALIGN(phys_addr);
1388 count = PAGE_SIZE_PTE_COUNT(page_size);
1389 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1391 for (i = 0; i < count; ++i)
1392 if (IOMMU_PTE_PRESENT(pte[i]))
1395 if (page_size > PAGE_SIZE) {
1396 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1397 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1399 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1401 if (prot & IOMMU_PROT_IR)
1402 __pte |= IOMMU_PTE_IR;
1403 if (prot & IOMMU_PROT_IW)
1404 __pte |= IOMMU_PTE_IW;
1406 for (i = 0; i < count; ++i)
1414 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1415 unsigned long bus_addr,
1416 unsigned long page_size)
1418 unsigned long long unmap_size, unmapped;
1421 BUG_ON(!is_power_of_2(page_size));
1425 while (unmapped < page_size) {
1427 pte = fetch_pte(dom, bus_addr);
1431 * No PTE for this address
1432 * move forward in 4kb steps
1434 unmap_size = PAGE_SIZE;
1435 } else if (PM_PTE_LEVEL(*pte) == 0) {
1436 /* 4kb PTE found for this address */
1437 unmap_size = PAGE_SIZE;
1442 /* Large PTE found which maps this address */
1443 unmap_size = PTE_PAGE_SIZE(*pte);
1444 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1445 for (i = 0; i < count; i++)
1449 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1450 unmapped += unmap_size;
1453 BUG_ON(!is_power_of_2(unmapped));
1459 * This function checks if a specific unity mapping entry is needed for
1460 * this specific IOMMU.
1462 static int iommu_for_unity_map(struct amd_iommu *iommu,
1463 struct unity_map_entry *entry)
1467 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1468 bdf = amd_iommu_alias_table[i];
1469 if (amd_iommu_rlookup_table[bdf] == iommu)
1477 * This function actually applies the mapping to the page table of the
1480 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1481 struct unity_map_entry *e)
1486 for (addr = e->address_start; addr < e->address_end;
1487 addr += PAGE_SIZE) {
1488 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1493 * if unity mapping is in aperture range mark the page
1494 * as allocated in the aperture
1496 if (addr < dma_dom->aperture_size)
1497 __set_bit(addr >> PAGE_SHIFT,
1498 dma_dom->aperture[0]->bitmap);
1505 * Init the unity mappings for a specific IOMMU in the system
1507 * Basically iterates over all unity mapping entries and applies them to
1508 * the default domain DMA of that IOMMU if necessary.
1510 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1512 struct unity_map_entry *entry;
1515 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1516 if (!iommu_for_unity_map(iommu, entry))
1518 ret = dma_ops_unity_map(iommu->default_dom, entry);
1527 * Inits the unity mappings required for a specific device
1529 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1532 struct unity_map_entry *e;
1535 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1536 if (!(devid >= e->devid_start && devid <= e->devid_end))
1538 ret = dma_ops_unity_map(dma_dom, e);
1546 /****************************************************************************
1548 * The next functions belong to the address allocator for the dma_ops
1549 * interface functions. They work like the allocators in the other IOMMU
1550 * drivers. Its basically a bitmap which marks the allocated pages in
1551 * the aperture. Maybe it could be enhanced in the future to a more
1552 * efficient allocator.
1554 ****************************************************************************/
1557 * The address allocator core functions.
1559 * called with domain->lock held
1563 * Used to reserve address ranges in the aperture (e.g. for exclusion
1566 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1567 unsigned long start_page,
1570 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1572 if (start_page + pages > last_page)
1573 pages = last_page - start_page;
1575 for (i = start_page; i < start_page + pages; ++i) {
1576 int index = i / APERTURE_RANGE_PAGES;
1577 int page = i % APERTURE_RANGE_PAGES;
1578 __set_bit(page, dom->aperture[index]->bitmap);
1583 * This function is used to add a new aperture range to an existing
1584 * aperture in case of dma_ops domain allocation or address allocation
1587 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1588 bool populate, gfp_t gfp)
1590 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1591 struct amd_iommu *iommu;
1592 unsigned long i, old_size;
1594 #ifdef CONFIG_IOMMU_STRESS
1598 if (index >= APERTURE_MAX_RANGES)
1601 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1602 if (!dma_dom->aperture[index])
1605 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1606 if (!dma_dom->aperture[index]->bitmap)
1609 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1612 unsigned long address = dma_dom->aperture_size;
1613 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1614 u64 *pte, *pte_page;
1616 for (i = 0; i < num_ptes; ++i) {
1617 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1622 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1624 address += APERTURE_RANGE_SIZE / 64;
1628 old_size = dma_dom->aperture_size;
1629 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1631 /* Reserve address range used for MSI messages */
1632 if (old_size < MSI_ADDR_BASE_LO &&
1633 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1634 unsigned long spage;
1637 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1638 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1640 dma_ops_reserve_addresses(dma_dom, spage, pages);
1643 /* Initialize the exclusion range if necessary */
1644 for_each_iommu(iommu) {
1645 if (iommu->exclusion_start &&
1646 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1647 && iommu->exclusion_start < dma_dom->aperture_size) {
1648 unsigned long startpage;
1649 int pages = iommu_num_pages(iommu->exclusion_start,
1650 iommu->exclusion_length,
1652 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1653 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1658 * Check for areas already mapped as present in the new aperture
1659 * range and mark those pages as reserved in the allocator. Such
1660 * mappings may already exist as a result of requested unity
1661 * mappings for devices.
1663 for (i = dma_dom->aperture[index]->offset;
1664 i < dma_dom->aperture_size;
1666 u64 *pte = fetch_pte(&dma_dom->domain, i);
1667 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1670 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1673 update_domain(&dma_dom->domain);
1678 update_domain(&dma_dom->domain);
1680 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1682 kfree(dma_dom->aperture[index]);
1683 dma_dom->aperture[index] = NULL;
1688 static unsigned long dma_ops_area_alloc(struct device *dev,
1689 struct dma_ops_domain *dom,
1691 unsigned long align_mask,
1693 unsigned long start)
1695 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1696 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1697 int i = start >> APERTURE_RANGE_SHIFT;
1698 unsigned long boundary_size;
1699 unsigned long address = -1;
1700 unsigned long limit;
1702 next_bit >>= PAGE_SHIFT;
1704 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1705 PAGE_SIZE) >> PAGE_SHIFT;
1707 for (;i < max_index; ++i) {
1708 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1710 if (dom->aperture[i]->offset >= dma_mask)
1713 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1714 dma_mask >> PAGE_SHIFT);
1716 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1717 limit, next_bit, pages, 0,
1718 boundary_size, align_mask);
1719 if (address != -1) {
1720 address = dom->aperture[i]->offset +
1721 (address << PAGE_SHIFT);
1722 dom->next_address = address + (pages << PAGE_SHIFT);
1732 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1733 struct dma_ops_domain *dom,
1735 unsigned long align_mask,
1738 unsigned long address;
1740 #ifdef CONFIG_IOMMU_STRESS
1741 dom->next_address = 0;
1742 dom->need_flush = true;
1745 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1746 dma_mask, dom->next_address);
1748 if (address == -1) {
1749 dom->next_address = 0;
1750 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1752 dom->need_flush = true;
1755 if (unlikely(address == -1))
1756 address = DMA_ERROR_CODE;
1758 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1764 * The address free function.
1766 * called with domain->lock held
1768 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1769 unsigned long address,
1772 unsigned i = address >> APERTURE_RANGE_SHIFT;
1773 struct aperture_range *range = dom->aperture[i];
1775 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1777 #ifdef CONFIG_IOMMU_STRESS
1782 if (address >= dom->next_address)
1783 dom->need_flush = true;
1785 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1787 bitmap_clear(range->bitmap, address, pages);
1791 /****************************************************************************
1793 * The next functions belong to the domain allocation. A domain is
1794 * allocated for every IOMMU as the default domain. If device isolation
1795 * is enabled, every device get its own domain. The most important thing
1796 * about domains is the page table mapping the DMA address space they
1799 ****************************************************************************/
1802 * This function adds a protection domain to the global protection domain list
1804 static void add_domain_to_list(struct protection_domain *domain)
1806 unsigned long flags;
1808 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1809 list_add(&domain->list, &amd_iommu_pd_list);
1810 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1814 * This function removes a protection domain to the global
1815 * protection domain list
1817 static void del_domain_from_list(struct protection_domain *domain)
1819 unsigned long flags;
1821 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1822 list_del(&domain->list);
1823 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1826 static u16 domain_id_alloc(void)
1828 unsigned long flags;
1831 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1832 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1834 if (id > 0 && id < MAX_DOMAIN_ID)
1835 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1838 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1843 static void domain_id_free(int id)
1845 unsigned long flags;
1847 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1848 if (id > 0 && id < MAX_DOMAIN_ID)
1849 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1850 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1853 static void free_pagetable(struct protection_domain *domain)
1858 p1 = domain->pt_root;
1863 for (i = 0; i < 512; ++i) {
1864 if (!IOMMU_PTE_PRESENT(p1[i]))
1867 p2 = IOMMU_PTE_PAGE(p1[i]);
1868 for (j = 0; j < 512; ++j) {
1869 if (!IOMMU_PTE_PRESENT(p2[j]))
1871 p3 = IOMMU_PTE_PAGE(p2[j]);
1872 free_page((unsigned long)p3);
1875 free_page((unsigned long)p2);
1878 free_page((unsigned long)p1);
1880 domain->pt_root = NULL;
1883 static void free_gcr3_tbl_level1(u64 *tbl)
1888 for (i = 0; i < 512; ++i) {
1889 if (!(tbl[i] & GCR3_VALID))
1892 ptr = __va(tbl[i] & PAGE_MASK);
1894 free_page((unsigned long)ptr);
1898 static void free_gcr3_tbl_level2(u64 *tbl)
1903 for (i = 0; i < 512; ++i) {
1904 if (!(tbl[i] & GCR3_VALID))
1907 ptr = __va(tbl[i] & PAGE_MASK);
1909 free_gcr3_tbl_level1(ptr);
1913 static void free_gcr3_table(struct protection_domain *domain)
1915 if (domain->glx == 2)
1916 free_gcr3_tbl_level2(domain->gcr3_tbl);
1917 else if (domain->glx == 1)
1918 free_gcr3_tbl_level1(domain->gcr3_tbl);
1919 else if (domain->glx != 0)
1922 free_page((unsigned long)domain->gcr3_tbl);
1926 * Free a domain, only used if something went wrong in the
1927 * allocation path and we need to free an already allocated page table
1929 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1936 del_domain_from_list(&dom->domain);
1938 free_pagetable(&dom->domain);
1940 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1941 if (!dom->aperture[i])
1943 free_page((unsigned long)dom->aperture[i]->bitmap);
1944 kfree(dom->aperture[i]);
1951 * Allocates a new protection domain usable for the dma_ops functions.
1952 * It also initializes the page table and the address allocator data
1953 * structures required for the dma_ops interface
1955 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1957 struct dma_ops_domain *dma_dom;
1959 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1963 spin_lock_init(&dma_dom->domain.lock);
1965 dma_dom->domain.id = domain_id_alloc();
1966 if (dma_dom->domain.id == 0)
1968 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1969 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1970 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1971 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1972 dma_dom->domain.priv = dma_dom;
1973 if (!dma_dom->domain.pt_root)
1976 dma_dom->need_flush = false;
1977 dma_dom->target_dev = 0xffff;
1979 add_domain_to_list(&dma_dom->domain);
1981 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1985 * mark the first page as allocated so we never return 0 as
1986 * a valid dma-address. So we can use 0 as error value
1988 dma_dom->aperture[0]->bitmap[0] = 1;
1989 dma_dom->next_address = 0;
1995 dma_ops_domain_free(dma_dom);
2001 * little helper function to check whether a given protection domain is a
2004 static bool dma_ops_domain(struct protection_domain *domain)
2006 return domain->flags & PD_DMA_OPS_MASK;
2009 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2014 if (domain->mode != PAGE_MODE_NONE)
2015 pte_root = virt_to_phys(domain->pt_root);
2017 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2018 << DEV_ENTRY_MODE_SHIFT;
2019 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2021 flags = amd_iommu_dev_table[devid].data[1];
2024 flags |= DTE_FLAG_IOTLB;
2026 if (domain->flags & PD_IOMMUV2_MASK) {
2027 u64 gcr3 = __pa(domain->gcr3_tbl);
2028 u64 glx = domain->glx;
2031 pte_root |= DTE_FLAG_GV;
2032 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2034 /* First mask out possible old values for GCR3 table */
2035 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2038 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2041 /* Encode GCR3 table into DTE */
2042 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2045 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2048 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2052 flags &= ~(0xffffUL);
2053 flags |= domain->id;
2055 amd_iommu_dev_table[devid].data[1] = flags;
2056 amd_iommu_dev_table[devid].data[0] = pte_root;
2059 static void clear_dte_entry(u16 devid)
2061 /* remove entry from the device table seen by the hardware */
2062 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2063 amd_iommu_dev_table[devid].data[1] = 0;
2065 amd_iommu_apply_erratum_63(devid);
2068 static void do_attach(struct iommu_dev_data *dev_data,
2069 struct protection_domain *domain)
2071 struct amd_iommu *iommu;
2074 iommu = amd_iommu_rlookup_table[dev_data->devid];
2075 ats = dev_data->ats.enabled;
2077 /* Update data structures */
2078 dev_data->domain = domain;
2079 list_add(&dev_data->list, &domain->dev_list);
2080 set_dte_entry(dev_data->devid, domain, ats);
2082 /* Do reference counting */
2083 domain->dev_iommu[iommu->index] += 1;
2084 domain->dev_cnt += 1;
2086 /* Flush the DTE entry */
2087 device_flush_dte(dev_data);
2090 static void do_detach(struct iommu_dev_data *dev_data)
2092 struct amd_iommu *iommu;
2094 iommu = amd_iommu_rlookup_table[dev_data->devid];
2096 /* decrease reference counters */
2097 dev_data->domain->dev_iommu[iommu->index] -= 1;
2098 dev_data->domain->dev_cnt -= 1;
2100 /* Update data structures */
2101 dev_data->domain = NULL;
2102 list_del(&dev_data->list);
2103 clear_dte_entry(dev_data->devid);
2105 /* Flush the DTE entry */
2106 device_flush_dte(dev_data);
2110 * If a device is not yet associated with a domain, this function does
2111 * assigns it visible for the hardware
2113 static int __attach_device(struct iommu_dev_data *dev_data,
2114 struct protection_domain *domain)
2119 spin_lock(&domain->lock);
2121 if (dev_data->alias_data != NULL) {
2122 struct iommu_dev_data *alias_data = dev_data->alias_data;
2124 /* Some sanity checks */
2126 if (alias_data->domain != NULL &&
2127 alias_data->domain != domain)
2130 if (dev_data->domain != NULL &&
2131 dev_data->domain != domain)
2134 /* Do real assignment */
2135 if (alias_data->domain == NULL)
2136 do_attach(alias_data, domain);
2138 atomic_inc(&alias_data->bind);
2141 if (dev_data->domain == NULL)
2142 do_attach(dev_data, domain);
2144 atomic_inc(&dev_data->bind);
2151 spin_unlock(&domain->lock);
2157 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2159 pci_disable_ats(pdev);
2160 pci_disable_pri(pdev);
2161 pci_disable_pasid(pdev);
2164 /* FIXME: Change generic reset-function to do the same */
2165 static int pri_reset_while_enabled(struct pci_dev *pdev)
2170 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2174 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2175 control |= PCI_PRI_CTRL_RESET;
2176 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2181 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2186 /* FIXME: Hardcode number of outstanding requests for now */
2188 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2190 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2192 /* Only allow access to user-accessible pages */
2193 ret = pci_enable_pasid(pdev, 0);
2197 /* First reset the PRI state of the device */
2198 ret = pci_reset_pri(pdev);
2203 ret = pci_enable_pri(pdev, reqs);
2208 ret = pri_reset_while_enabled(pdev);
2213 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2220 pci_disable_pri(pdev);
2221 pci_disable_pasid(pdev);
2226 /* FIXME: Move this to PCI code */
2227 #define PCI_PRI_TLP_OFF (1 << 15)
2229 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2234 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2238 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2240 return (status & PCI_PRI_TLP_OFF) ? true : false;
2244 * If a device is not yet associated with a domain, this function
2245 * assigns it visible for the hardware
2247 static int attach_device(struct device *dev,
2248 struct protection_domain *domain)
2250 struct pci_dev *pdev = to_pci_dev(dev);
2251 struct iommu_dev_data *dev_data;
2252 unsigned long flags;
2255 dev_data = get_dev_data(dev);
2257 if (domain->flags & PD_IOMMUV2_MASK) {
2258 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2261 if (pdev_iommuv2_enable(pdev) != 0)
2264 dev_data->ats.enabled = true;
2265 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2266 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2267 } else if (amd_iommu_iotlb_sup &&
2268 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2269 dev_data->ats.enabled = true;
2270 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2273 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2274 ret = __attach_device(dev_data, domain);
2275 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2278 * We might boot into a crash-kernel here. The crashed kernel
2279 * left the caches in the IOMMU dirty. So we have to flush
2280 * here to evict all dirty stuff.
2282 domain_flush_tlb_pde(domain);
2288 * Removes a device from a protection domain (unlocked)
2290 static void __detach_device(struct iommu_dev_data *dev_data)
2292 struct protection_domain *domain;
2293 unsigned long flags;
2295 BUG_ON(!dev_data->domain);
2297 domain = dev_data->domain;
2299 spin_lock_irqsave(&domain->lock, flags);
2301 if (dev_data->alias_data != NULL) {
2302 struct iommu_dev_data *alias_data = dev_data->alias_data;
2304 if (atomic_dec_and_test(&alias_data->bind))
2305 do_detach(alias_data);
2308 if (atomic_dec_and_test(&dev_data->bind))
2309 do_detach(dev_data);
2311 spin_unlock_irqrestore(&domain->lock, flags);
2314 * If we run in passthrough mode the device must be assigned to the
2315 * passthrough domain if it is detached from any other domain.
2316 * Make sure we can deassign from the pt_domain itself.
2318 if (dev_data->passthrough &&
2319 (dev_data->domain == NULL && domain != pt_domain))
2320 __attach_device(dev_data, pt_domain);
2324 * Removes a device from a protection domain (with devtable_lock held)
2326 static void detach_device(struct device *dev)
2328 struct protection_domain *domain;
2329 struct iommu_dev_data *dev_data;
2330 unsigned long flags;
2332 dev_data = get_dev_data(dev);
2333 domain = dev_data->domain;
2335 /* lock device table */
2336 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2337 __detach_device(dev_data);
2338 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2340 if (domain->flags & PD_IOMMUV2_MASK)
2341 pdev_iommuv2_disable(to_pci_dev(dev));
2342 else if (dev_data->ats.enabled)
2343 pci_disable_ats(to_pci_dev(dev));
2345 dev_data->ats.enabled = false;
2349 * Find out the protection domain structure for a given PCI device. This
2350 * will give us the pointer to the page table root for example.
2352 static struct protection_domain *domain_for_device(struct device *dev)
2354 struct iommu_dev_data *dev_data;
2355 struct protection_domain *dom = NULL;
2356 unsigned long flags;
2358 dev_data = get_dev_data(dev);
2360 if (dev_data->domain)
2361 return dev_data->domain;
2363 if (dev_data->alias_data != NULL) {
2364 struct iommu_dev_data *alias_data = dev_data->alias_data;
2366 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2367 if (alias_data->domain != NULL) {
2368 __attach_device(dev_data, alias_data->domain);
2369 dom = alias_data->domain;
2371 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2377 static int device_change_notifier(struct notifier_block *nb,
2378 unsigned long action, void *data)
2380 struct dma_ops_domain *dma_domain;
2381 struct protection_domain *domain;
2382 struct iommu_dev_data *dev_data;
2383 struct device *dev = data;
2384 struct amd_iommu *iommu;
2385 unsigned long flags;
2388 if (!check_device(dev))
2391 devid = get_device_id(dev);
2392 iommu = amd_iommu_rlookup_table[devid];
2393 dev_data = get_dev_data(dev);
2396 case BUS_NOTIFY_UNBOUND_DRIVER:
2398 domain = domain_for_device(dev);
2402 if (dev_data->passthrough)
2406 case BUS_NOTIFY_ADD_DEVICE:
2408 iommu_init_device(dev);
2411 * dev_data is still NULL and
2412 * got initialized in iommu_init_device
2414 dev_data = get_dev_data(dev);
2416 if (iommu_pass_through || dev_data->iommu_v2) {
2417 dev_data->passthrough = true;
2418 attach_device(dev, pt_domain);
2422 domain = domain_for_device(dev);
2424 /* allocate a protection domain if a device is added */
2425 dma_domain = find_protection_domain(devid);
2428 dma_domain = dma_ops_domain_alloc();
2431 dma_domain->target_dev = devid;
2433 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2434 list_add_tail(&dma_domain->list, &iommu_pd_list);
2435 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2437 dev_data = get_dev_data(dev);
2439 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2442 case BUS_NOTIFY_DEL_DEVICE:
2444 iommu_uninit_device(dev);
2450 iommu_completion_wait(iommu);
2456 static struct notifier_block device_nb = {
2457 .notifier_call = device_change_notifier,
2460 void amd_iommu_init_notifier(void)
2462 bus_register_notifier(&pci_bus_type, &device_nb);
2465 /*****************************************************************************
2467 * The next functions belong to the dma_ops mapping/unmapping code.
2469 *****************************************************************************/
2472 * In the dma_ops path we only have the struct device. This function
2473 * finds the corresponding IOMMU, the protection domain and the
2474 * requestor id for a given device.
2475 * If the device is not yet associated with a domain this is also done
2478 static struct protection_domain *get_domain(struct device *dev)
2480 struct protection_domain *domain;
2481 struct dma_ops_domain *dma_dom;
2482 u16 devid = get_device_id(dev);
2484 if (!check_device(dev))
2485 return ERR_PTR(-EINVAL);
2487 domain = domain_for_device(dev);
2488 if (domain != NULL && !dma_ops_domain(domain))
2489 return ERR_PTR(-EBUSY);
2494 /* Device not bound yet - bind it */
2495 dma_dom = find_protection_domain(devid);
2497 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2498 attach_device(dev, &dma_dom->domain);
2499 DUMP_printk("Using protection domain %d for device %s\n",
2500 dma_dom->domain.id, dev_name(dev));
2502 return &dma_dom->domain;
2505 static void update_device_table(struct protection_domain *domain)
2507 struct iommu_dev_data *dev_data;
2509 list_for_each_entry(dev_data, &domain->dev_list, list)
2510 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2513 static void update_domain(struct protection_domain *domain)
2515 if (!domain->updated)
2518 update_device_table(domain);
2520 domain_flush_devices(domain);
2521 domain_flush_tlb_pde(domain);
2523 domain->updated = false;
2527 * This function fetches the PTE for a given address in the aperture
2529 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2530 unsigned long address)
2532 struct aperture_range *aperture;
2533 u64 *pte, *pte_page;
2535 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2539 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2541 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2543 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2545 pte += PM_LEVEL_INDEX(0, address);
2547 update_domain(&dom->domain);
2553 * This is the generic map function. It maps one 4kb page at paddr to
2554 * the given address in the DMA address space for the domain.
2556 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2557 unsigned long address,
2563 WARN_ON(address > dom->aperture_size);
2567 pte = dma_ops_get_pte(dom, address);
2569 return DMA_ERROR_CODE;
2571 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2573 if (direction == DMA_TO_DEVICE)
2574 __pte |= IOMMU_PTE_IR;
2575 else if (direction == DMA_FROM_DEVICE)
2576 __pte |= IOMMU_PTE_IW;
2577 else if (direction == DMA_BIDIRECTIONAL)
2578 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2584 return (dma_addr_t)address;
2588 * The generic unmapping function for on page in the DMA address space.
2590 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2591 unsigned long address)
2593 struct aperture_range *aperture;
2596 if (address >= dom->aperture_size)
2599 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2603 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2607 pte += PM_LEVEL_INDEX(0, address);
2615 * This function contains common code for mapping of a physically
2616 * contiguous memory region into DMA address space. It is used by all
2617 * mapping functions provided with this IOMMU driver.
2618 * Must be called with the domain lock held.
2620 static dma_addr_t __map_single(struct device *dev,
2621 struct dma_ops_domain *dma_dom,
2628 dma_addr_t offset = paddr & ~PAGE_MASK;
2629 dma_addr_t address, start, ret;
2631 unsigned long align_mask = 0;
2634 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2637 INC_STATS_COUNTER(total_map_requests);
2640 INC_STATS_COUNTER(cross_page);
2643 align_mask = (1UL << get_order(size)) - 1;
2646 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2648 if (unlikely(address == DMA_ERROR_CODE)) {
2650 * setting next_address here will let the address
2651 * allocator only scan the new allocated range in the
2652 * first run. This is a small optimization.
2654 dma_dom->next_address = dma_dom->aperture_size;
2656 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2660 * aperture was successfully enlarged by 128 MB, try
2667 for (i = 0; i < pages; ++i) {
2668 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2669 if (ret == DMA_ERROR_CODE)
2677 ADD_STATS_COUNTER(alloced_io_mem, size);
2679 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2680 domain_flush_tlb(&dma_dom->domain);
2681 dma_dom->need_flush = false;
2682 } else if (unlikely(amd_iommu_np_cache))
2683 domain_flush_pages(&dma_dom->domain, address, size);
2690 for (--i; i >= 0; --i) {
2692 dma_ops_domain_unmap(dma_dom, start);
2695 dma_ops_free_addresses(dma_dom, address, pages);
2697 return DMA_ERROR_CODE;
2701 * Does the reverse of the __map_single function. Must be called with
2702 * the domain lock held too
2704 static void __unmap_single(struct dma_ops_domain *dma_dom,
2705 dma_addr_t dma_addr,
2709 dma_addr_t flush_addr;
2710 dma_addr_t i, start;
2713 if ((dma_addr == DMA_ERROR_CODE) ||
2714 (dma_addr + size > dma_dom->aperture_size))
2717 flush_addr = dma_addr;
2718 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2719 dma_addr &= PAGE_MASK;
2722 for (i = 0; i < pages; ++i) {
2723 dma_ops_domain_unmap(dma_dom, start);
2727 SUB_STATS_COUNTER(alloced_io_mem, size);
2729 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2731 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2732 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2733 dma_dom->need_flush = false;
2738 * The exported map_single function for dma_ops.
2740 static dma_addr_t map_page(struct device *dev, struct page *page,
2741 unsigned long offset, size_t size,
2742 enum dma_data_direction dir,
2743 struct dma_attrs *attrs)
2745 unsigned long flags;
2746 struct protection_domain *domain;
2749 phys_addr_t paddr = page_to_phys(page) + offset;
2751 INC_STATS_COUNTER(cnt_map_single);
2753 domain = get_domain(dev);
2754 if (PTR_ERR(domain) == -EINVAL)
2755 return (dma_addr_t)paddr;
2756 else if (IS_ERR(domain))
2757 return DMA_ERROR_CODE;
2759 dma_mask = *dev->dma_mask;
2761 spin_lock_irqsave(&domain->lock, flags);
2763 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2765 if (addr == DMA_ERROR_CODE)
2768 domain_flush_complete(domain);
2771 spin_unlock_irqrestore(&domain->lock, flags);
2777 * The exported unmap_single function for dma_ops.
2779 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2780 enum dma_data_direction dir, struct dma_attrs *attrs)
2782 unsigned long flags;
2783 struct protection_domain *domain;
2785 INC_STATS_COUNTER(cnt_unmap_single);
2787 domain = get_domain(dev);
2791 spin_lock_irqsave(&domain->lock, flags);
2793 __unmap_single(domain->priv, dma_addr, size, dir);
2795 domain_flush_complete(domain);
2797 spin_unlock_irqrestore(&domain->lock, flags);
2801 * This is a special map_sg function which is used if we should map a
2802 * device which is not handled by an AMD IOMMU in the system.
2804 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2805 int nelems, int dir)
2807 struct scatterlist *s;
2810 for_each_sg(sglist, s, nelems, i) {
2811 s->dma_address = (dma_addr_t)sg_phys(s);
2812 s->dma_length = s->length;
2819 * The exported map_sg function for dma_ops (handles scatter-gather
2822 static int map_sg(struct device *dev, struct scatterlist *sglist,
2823 int nelems, enum dma_data_direction dir,
2824 struct dma_attrs *attrs)
2826 unsigned long flags;
2827 struct protection_domain *domain;
2829 struct scatterlist *s;
2831 int mapped_elems = 0;
2834 INC_STATS_COUNTER(cnt_map_sg);
2836 domain = get_domain(dev);
2837 if (PTR_ERR(domain) == -EINVAL)
2838 return map_sg_no_iommu(dev, sglist, nelems, dir);
2839 else if (IS_ERR(domain))
2842 dma_mask = *dev->dma_mask;
2844 spin_lock_irqsave(&domain->lock, flags);
2846 for_each_sg(sglist, s, nelems, i) {
2849 s->dma_address = __map_single(dev, domain->priv,
2850 paddr, s->length, dir, false,
2853 if (s->dma_address) {
2854 s->dma_length = s->length;
2860 domain_flush_complete(domain);
2863 spin_unlock_irqrestore(&domain->lock, flags);
2865 return mapped_elems;
2867 for_each_sg(sglist, s, mapped_elems, i) {
2869 __unmap_single(domain->priv, s->dma_address,
2870 s->dma_length, dir);
2871 s->dma_address = s->dma_length = 0;
2880 * The exported map_sg function for dma_ops (handles scatter-gather
2883 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2884 int nelems, enum dma_data_direction dir,
2885 struct dma_attrs *attrs)
2887 unsigned long flags;
2888 struct protection_domain *domain;
2889 struct scatterlist *s;
2892 INC_STATS_COUNTER(cnt_unmap_sg);
2894 domain = get_domain(dev);
2898 spin_lock_irqsave(&domain->lock, flags);
2900 for_each_sg(sglist, s, nelems, i) {
2901 __unmap_single(domain->priv, s->dma_address,
2902 s->dma_length, dir);
2903 s->dma_address = s->dma_length = 0;
2906 domain_flush_complete(domain);
2908 spin_unlock_irqrestore(&domain->lock, flags);
2912 * The exported alloc_coherent function for dma_ops.
2914 static void *alloc_coherent(struct device *dev, size_t size,
2915 dma_addr_t *dma_addr, gfp_t flag,
2916 struct dma_attrs *attrs)
2918 unsigned long flags;
2920 struct protection_domain *domain;
2922 u64 dma_mask = dev->coherent_dma_mask;
2924 INC_STATS_COUNTER(cnt_alloc_coherent);
2926 domain = get_domain(dev);
2927 if (PTR_ERR(domain) == -EINVAL) {
2928 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2929 *dma_addr = __pa(virt_addr);
2931 } else if (IS_ERR(domain))
2934 dma_mask = dev->coherent_dma_mask;
2935 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2938 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2942 paddr = virt_to_phys(virt_addr);
2945 dma_mask = *dev->dma_mask;
2947 spin_lock_irqsave(&domain->lock, flags);
2949 *dma_addr = __map_single(dev, domain->priv, paddr,
2950 size, DMA_BIDIRECTIONAL, true, dma_mask);
2952 if (*dma_addr == DMA_ERROR_CODE) {
2953 spin_unlock_irqrestore(&domain->lock, flags);
2957 domain_flush_complete(domain);
2959 spin_unlock_irqrestore(&domain->lock, flags);
2965 free_pages((unsigned long)virt_addr, get_order(size));
2971 * The exported free_coherent function for dma_ops.
2973 static void free_coherent(struct device *dev, size_t size,
2974 void *virt_addr, dma_addr_t dma_addr,
2975 struct dma_attrs *attrs)
2977 unsigned long flags;
2978 struct protection_domain *domain;
2980 INC_STATS_COUNTER(cnt_free_coherent);
2982 domain = get_domain(dev);
2986 spin_lock_irqsave(&domain->lock, flags);
2988 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2990 domain_flush_complete(domain);
2992 spin_unlock_irqrestore(&domain->lock, flags);
2995 free_pages((unsigned long)virt_addr, get_order(size));
2999 * This function is called by the DMA layer to find out if we can handle a
3000 * particular device. It is part of the dma_ops.
3002 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3004 return check_device(dev);
3008 * The function for pre-allocating protection domains.
3010 * If the driver core informs the DMA layer if a driver grabs a device
3011 * we don't need to preallocate the protection domains anymore.
3012 * For now we have to.
3014 static void __init prealloc_protection_domains(void)
3016 struct iommu_dev_data *dev_data;
3017 struct dma_ops_domain *dma_dom;
3018 struct pci_dev *dev = NULL;
3021 for_each_pci_dev(dev) {
3023 /* Do we handle this device? */
3024 if (!check_device(&dev->dev))
3027 dev_data = get_dev_data(&dev->dev);
3028 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3029 /* Make sure passthrough domain is allocated */
3030 alloc_passthrough_domain();
3031 dev_data->passthrough = true;
3032 attach_device(&dev->dev, pt_domain);
3033 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3034 dev_name(&dev->dev));
3037 /* Is there already any domain for it? */
3038 if (domain_for_device(&dev->dev))
3041 devid = get_device_id(&dev->dev);
3043 dma_dom = dma_ops_domain_alloc();
3046 init_unity_mappings_for_device(dma_dom, devid);
3047 dma_dom->target_dev = devid;
3049 attach_device(&dev->dev, &dma_dom->domain);
3051 list_add_tail(&dma_dom->list, &iommu_pd_list);
3055 static struct dma_map_ops amd_iommu_dma_ops = {
3056 .alloc = alloc_coherent,
3057 .free = free_coherent,
3058 .map_page = map_page,
3059 .unmap_page = unmap_page,
3061 .unmap_sg = unmap_sg,
3062 .dma_supported = amd_iommu_dma_supported,
3065 static unsigned device_dma_ops_init(void)
3067 struct iommu_dev_data *dev_data;
3068 struct pci_dev *pdev = NULL;
3069 unsigned unhandled = 0;
3071 for_each_pci_dev(pdev) {
3072 if (!check_device(&pdev->dev)) {
3074 iommu_ignore_device(&pdev->dev);
3080 dev_data = get_dev_data(&pdev->dev);
3082 if (!dev_data->passthrough)
3083 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3085 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3092 * The function which clues the AMD IOMMU driver into dma_ops.
3095 void __init amd_iommu_init_api(void)
3097 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3100 int __init amd_iommu_init_dma_ops(void)
3102 struct amd_iommu *iommu;
3106 * first allocate a default protection domain for every IOMMU we
3107 * found in the system. Devices not assigned to any other
3108 * protection domain will be assigned to the default one.
3110 for_each_iommu(iommu) {
3111 iommu->default_dom = dma_ops_domain_alloc();
3112 if (iommu->default_dom == NULL)
3114 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3115 ret = iommu_init_unity_mappings(iommu);
3121 * Pre-allocate the protection domains for each device.
3123 prealloc_protection_domains();
3128 /* Make the driver finally visible to the drivers */
3129 unhandled = device_dma_ops_init();
3130 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3131 /* There are unhandled devices - initialize swiotlb for them */
3135 amd_iommu_stats_init();
3137 if (amd_iommu_unmap_flush)
3138 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3140 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3146 for_each_iommu(iommu) {
3147 if (iommu->default_dom)
3148 dma_ops_domain_free(iommu->default_dom);
3154 /*****************************************************************************
3156 * The following functions belong to the exported interface of AMD IOMMU
3158 * This interface allows access to lower level functions of the IOMMU
3159 * like protection domain handling and assignement of devices to domains
3160 * which is not possible with the dma_ops interface.
3162 *****************************************************************************/
3164 static void cleanup_domain(struct protection_domain *domain)
3166 struct iommu_dev_data *dev_data, *next;
3167 unsigned long flags;
3169 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3171 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3172 __detach_device(dev_data);
3173 atomic_set(&dev_data->bind, 0);
3176 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3179 static void protection_domain_free(struct protection_domain *domain)
3184 del_domain_from_list(domain);
3187 domain_id_free(domain->id);
3192 static struct protection_domain *protection_domain_alloc(void)
3194 struct protection_domain *domain;
3196 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3200 spin_lock_init(&domain->lock);
3201 mutex_init(&domain->api_lock);
3202 domain->id = domain_id_alloc();
3205 INIT_LIST_HEAD(&domain->dev_list);
3207 add_domain_to_list(domain);
3217 static int __init alloc_passthrough_domain(void)
3219 if (pt_domain != NULL)
3222 /* allocate passthrough domain */
3223 pt_domain = protection_domain_alloc();
3227 pt_domain->mode = PAGE_MODE_NONE;
3231 static int amd_iommu_domain_init(struct iommu_domain *dom)
3233 struct protection_domain *domain;
3235 domain = protection_domain_alloc();
3239 domain->mode = PAGE_MODE_3_LEVEL;
3240 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3241 if (!domain->pt_root)
3244 domain->iommu_domain = dom;
3248 dom->geometry.aperture_start = 0;
3249 dom->geometry.aperture_end = ~0ULL;
3250 dom->geometry.force_aperture = true;
3255 protection_domain_free(domain);
3260 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3262 struct protection_domain *domain = dom->priv;
3267 if (domain->dev_cnt > 0)
3268 cleanup_domain(domain);
3270 BUG_ON(domain->dev_cnt != 0);
3272 if (domain->mode != PAGE_MODE_NONE)
3273 free_pagetable(domain);
3275 if (domain->flags & PD_IOMMUV2_MASK)
3276 free_gcr3_table(domain);
3278 protection_domain_free(domain);
3283 static void amd_iommu_detach_device(struct iommu_domain *dom,
3286 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3287 struct amd_iommu *iommu;
3290 if (!check_device(dev))
3293 devid = get_device_id(dev);
3295 if (dev_data->domain != NULL)
3298 iommu = amd_iommu_rlookup_table[devid];
3302 iommu_completion_wait(iommu);
3305 static int amd_iommu_attach_device(struct iommu_domain *dom,
3308 struct protection_domain *domain = dom->priv;
3309 struct iommu_dev_data *dev_data;
3310 struct amd_iommu *iommu;
3313 if (!check_device(dev))
3316 dev_data = dev->archdata.iommu;
3318 iommu = amd_iommu_rlookup_table[dev_data->devid];
3322 if (dev_data->domain)
3325 ret = attach_device(dev, domain);
3327 iommu_completion_wait(iommu);
3332 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3333 phys_addr_t paddr, size_t page_size, int iommu_prot)
3335 struct protection_domain *domain = dom->priv;
3339 if (domain->mode == PAGE_MODE_NONE)
3342 if (iommu_prot & IOMMU_READ)
3343 prot |= IOMMU_PROT_IR;
3344 if (iommu_prot & IOMMU_WRITE)
3345 prot |= IOMMU_PROT_IW;
3347 mutex_lock(&domain->api_lock);
3348 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3349 mutex_unlock(&domain->api_lock);
3354 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3357 struct protection_domain *domain = dom->priv;
3360 if (domain->mode == PAGE_MODE_NONE)
3363 mutex_lock(&domain->api_lock);
3364 unmap_size = iommu_unmap_page(domain, iova, page_size);
3365 mutex_unlock(&domain->api_lock);
3367 domain_flush_tlb_pde(domain);
3372 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3375 struct protection_domain *domain = dom->priv;
3376 unsigned long offset_mask;
3380 if (domain->mode == PAGE_MODE_NONE)
3383 pte = fetch_pte(domain, iova);
3385 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3388 if (PM_PTE_LEVEL(*pte) == 0)
3389 offset_mask = PAGE_SIZE - 1;
3391 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3393 __pte = *pte & PM_ADDR_MASK;
3394 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3399 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3403 case IOMMU_CAP_CACHE_COHERENCY:
3405 case IOMMU_CAP_INTR_REMAP:
3406 return irq_remapping_enabled;
3412 static struct iommu_ops amd_iommu_ops = {
3413 .domain_init = amd_iommu_domain_init,
3414 .domain_destroy = amd_iommu_domain_destroy,
3415 .attach_dev = amd_iommu_attach_device,
3416 .detach_dev = amd_iommu_detach_device,
3417 .map = amd_iommu_map,
3418 .unmap = amd_iommu_unmap,
3419 .iova_to_phys = amd_iommu_iova_to_phys,
3420 .domain_has_cap = amd_iommu_domain_has_cap,
3421 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3424 /*****************************************************************************
3426 * The next functions do a basic initialization of IOMMU for pass through
3429 * In passthrough mode the IOMMU is initialized and enabled but not used for
3430 * DMA-API translation.
3432 *****************************************************************************/
3434 int __init amd_iommu_init_passthrough(void)
3436 struct iommu_dev_data *dev_data;
3437 struct pci_dev *dev = NULL;
3438 struct amd_iommu *iommu;
3442 ret = alloc_passthrough_domain();
3446 for_each_pci_dev(dev) {
3447 if (!check_device(&dev->dev))
3450 dev_data = get_dev_data(&dev->dev);
3451 dev_data->passthrough = true;
3453 devid = get_device_id(&dev->dev);
3455 iommu = amd_iommu_rlookup_table[devid];
3459 attach_device(&dev->dev, pt_domain);
3462 amd_iommu_stats_init();
3464 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3469 /* IOMMUv2 specific functions */
3470 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3472 return atomic_notifier_chain_register(&ppr_notifier, nb);
3474 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3476 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3478 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3480 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3482 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3484 struct protection_domain *domain = dom->priv;
3485 unsigned long flags;
3487 spin_lock_irqsave(&domain->lock, flags);
3489 /* Update data structure */
3490 domain->mode = PAGE_MODE_NONE;
3491 domain->updated = true;
3493 /* Make changes visible to IOMMUs */
3494 update_domain(domain);
3496 /* Page-table is not visible to IOMMU anymore, so free it */
3497 free_pagetable(domain);
3499 spin_unlock_irqrestore(&domain->lock, flags);
3501 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3503 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3505 struct protection_domain *domain = dom->priv;
3506 unsigned long flags;
3509 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3512 /* Number of GCR3 table levels required */
3513 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3516 if (levels > amd_iommu_max_glx_val)
3519 spin_lock_irqsave(&domain->lock, flags);
3522 * Save us all sanity checks whether devices already in the
3523 * domain support IOMMUv2. Just force that the domain has no
3524 * devices attached when it is switched into IOMMUv2 mode.
3527 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3531 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3532 if (domain->gcr3_tbl == NULL)
3535 domain->glx = levels;
3536 domain->flags |= PD_IOMMUV2_MASK;
3537 domain->updated = true;
3539 update_domain(domain);
3544 spin_unlock_irqrestore(&domain->lock, flags);
3548 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3550 static int __flush_pasid(struct protection_domain *domain, int pasid,
3551 u64 address, bool size)
3553 struct iommu_dev_data *dev_data;
3554 struct iommu_cmd cmd;
3557 if (!(domain->flags & PD_IOMMUV2_MASK))
3560 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3563 * IOMMU TLB needs to be flushed before Device TLB to
3564 * prevent device TLB refill from IOMMU TLB
3566 for (i = 0; i < amd_iommus_present; ++i) {
3567 if (domain->dev_iommu[i] == 0)
3570 ret = iommu_queue_command(amd_iommus[i], &cmd);
3575 /* Wait until IOMMU TLB flushes are complete */
3576 domain_flush_complete(domain);
3578 /* Now flush device TLBs */
3579 list_for_each_entry(dev_data, &domain->dev_list, list) {
3580 struct amd_iommu *iommu;
3583 BUG_ON(!dev_data->ats.enabled);
3585 qdep = dev_data->ats.qdep;
3586 iommu = amd_iommu_rlookup_table[dev_data->devid];
3588 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3589 qdep, address, size);
3591 ret = iommu_queue_command(iommu, &cmd);
3596 /* Wait until all device TLBs are flushed */
3597 domain_flush_complete(domain);
3606 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3609 INC_STATS_COUNTER(invalidate_iotlb);
3611 return __flush_pasid(domain, pasid, address, false);
3614 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3617 struct protection_domain *domain = dom->priv;
3618 unsigned long flags;
3621 spin_lock_irqsave(&domain->lock, flags);
3622 ret = __amd_iommu_flush_page(domain, pasid, address);
3623 spin_unlock_irqrestore(&domain->lock, flags);
3627 EXPORT_SYMBOL(amd_iommu_flush_page);
3629 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3631 INC_STATS_COUNTER(invalidate_iotlb_all);
3633 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3637 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3639 struct protection_domain *domain = dom->priv;
3640 unsigned long flags;
3643 spin_lock_irqsave(&domain->lock, flags);
3644 ret = __amd_iommu_flush_tlb(domain, pasid);
3645 spin_unlock_irqrestore(&domain->lock, flags);
3649 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3651 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3658 index = (pasid >> (9 * level)) & 0x1ff;
3664 if (!(*pte & GCR3_VALID)) {
3668 root = (void *)get_zeroed_page(GFP_ATOMIC);
3672 *pte = __pa(root) | GCR3_VALID;
3675 root = __va(*pte & PAGE_MASK);
3683 static int __set_gcr3(struct protection_domain *domain, int pasid,
3688 if (domain->mode != PAGE_MODE_NONE)
3691 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3695 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3697 return __amd_iommu_flush_tlb(domain, pasid);
3700 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3704 if (domain->mode != PAGE_MODE_NONE)
3707 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3713 return __amd_iommu_flush_tlb(domain, pasid);
3716 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3719 struct protection_domain *domain = dom->priv;
3720 unsigned long flags;
3723 spin_lock_irqsave(&domain->lock, flags);
3724 ret = __set_gcr3(domain, pasid, cr3);
3725 spin_unlock_irqrestore(&domain->lock, flags);
3729 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3731 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3733 struct protection_domain *domain = dom->priv;
3734 unsigned long flags;
3737 spin_lock_irqsave(&domain->lock, flags);
3738 ret = __clear_gcr3(domain, pasid);
3739 spin_unlock_irqrestore(&domain->lock, flags);
3743 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3745 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3746 int status, int tag)
3748 struct iommu_dev_data *dev_data;
3749 struct amd_iommu *iommu;
3750 struct iommu_cmd cmd;
3752 INC_STATS_COUNTER(complete_ppr);
3754 dev_data = get_dev_data(&pdev->dev);
3755 iommu = amd_iommu_rlookup_table[dev_data->devid];
3757 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3758 tag, dev_data->pri_tlp);
3760 return iommu_queue_command(iommu, &cmd);
3762 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3764 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3766 struct protection_domain *domain;
3768 domain = get_domain(&pdev->dev);
3772 /* Only return IOMMUv2 domains */
3773 if (!(domain->flags & PD_IOMMUV2_MASK))
3776 return domain->iommu_domain;
3778 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3780 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3782 struct iommu_dev_data *dev_data;
3784 if (!amd_iommu_v2_supported())
3787 dev_data = get_dev_data(&pdev->dev);
3788 dev_data->errata |= (1 << erratum);
3790 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3792 int amd_iommu_device_info(struct pci_dev *pdev,
3793 struct amd_iommu_device_info *info)
3798 if (pdev == NULL || info == NULL)
3801 if (!amd_iommu_v2_supported())
3804 memset(info, 0, sizeof(*info));
3806 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3808 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3810 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3812 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3814 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3818 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3819 max_pasids = min(max_pasids, (1 << 20));
3821 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3822 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3824 features = pci_pasid_features(pdev);
3825 if (features & PCI_PASID_CAP_EXEC)
3826 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3827 if (features & PCI_PASID_CAP_PRIV)
3828 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3833 EXPORT_SYMBOL(amd_iommu_device_info);
3835 #ifdef CONFIG_IRQ_REMAP
3837 /*****************************************************************************
3839 * Interrupt Remapping Implementation
3841 *****************************************************************************/
3858 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3859 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3860 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3861 #define DTE_IRQ_REMAP_ENABLE 1ULL
3863 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3867 dte = amd_iommu_dev_table[devid].data[2];
3868 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3869 dte |= virt_to_phys(table->table);
3870 dte |= DTE_IRQ_REMAP_INTCTL;
3871 dte |= DTE_IRQ_TABLE_LEN;
3872 dte |= DTE_IRQ_REMAP_ENABLE;
3874 amd_iommu_dev_table[devid].data[2] = dte;
3877 #define IRTE_ALLOCATED (~1U)
3879 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3881 struct irq_remap_table *table = NULL;
3882 struct amd_iommu *iommu;
3883 unsigned long flags;
3886 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3888 iommu = amd_iommu_rlookup_table[devid];
3892 table = irq_lookup_table[devid];
3896 alias = amd_iommu_alias_table[devid];
3897 table = irq_lookup_table[alias];
3899 irq_lookup_table[devid] = table;
3900 set_dte_irq_entry(devid, table);
3901 iommu_flush_dte(iommu, devid);
3905 /* Nothing there yet, allocate new irq remapping table */
3906 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3911 /* Keep the first 32 indexes free for IOAPIC interrupts */
3912 table->min_index = 32;
3914 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3915 if (!table->table) {
3921 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3926 for (i = 0; i < 32; ++i)
3927 table->table[i] = IRTE_ALLOCATED;
3930 irq_lookup_table[devid] = table;
3931 set_dte_irq_entry(devid, table);
3932 iommu_flush_dte(iommu, devid);
3933 if (devid != alias) {
3934 irq_lookup_table[alias] = table;
3935 set_dte_irq_entry(devid, table);
3936 iommu_flush_dte(iommu, alias);
3940 iommu_completion_wait(iommu);
3943 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3948 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
3950 struct irq_remap_table *table;
3951 unsigned long flags;
3954 table = get_irq_table(devid, false);
3958 spin_lock_irqsave(&table->lock, flags);
3960 /* Scan table for free entries */
3961 for (c = 0, index = table->min_index;
3962 index < MAX_IRQS_PER_TABLE;
3964 if (table->table[index] == 0)
3970 struct irq_2_iommu *irte_info;
3973 table->table[index - c + 1] = IRTE_ALLOCATED;
3977 irte_info = &cfg->irq_2_iommu;
3978 irte_info->sub_handle = devid;
3979 irte_info->irte_index = index;
3980 irte_info->iommu = (void *)cfg;
3989 spin_unlock_irqrestore(&table->lock, flags);
3994 static int get_irte(u16 devid, int index, union irte *irte)
3996 struct irq_remap_table *table;
3997 unsigned long flags;
3999 table = get_irq_table(devid, false);
4003 spin_lock_irqsave(&table->lock, flags);
4004 irte->val = table->table[index];
4005 spin_unlock_irqrestore(&table->lock, flags);
4010 static int modify_irte(u16 devid, int index, union irte irte)
4012 struct irq_remap_table *table;
4013 struct amd_iommu *iommu;
4014 unsigned long flags;
4016 iommu = amd_iommu_rlookup_table[devid];
4020 table = get_irq_table(devid, false);
4024 spin_lock_irqsave(&table->lock, flags);
4025 table->table[index] = irte.val;
4026 spin_unlock_irqrestore(&table->lock, flags);
4028 iommu_flush_irt(iommu, devid);
4029 iommu_completion_wait(iommu);
4034 static void free_irte(u16 devid, int index)
4036 struct irq_remap_table *table;
4037 struct amd_iommu *iommu;
4038 unsigned long flags;
4040 iommu = amd_iommu_rlookup_table[devid];
4044 table = get_irq_table(devid, false);
4048 spin_lock_irqsave(&table->lock, flags);
4049 table->table[index] = 0;
4050 spin_unlock_irqrestore(&table->lock, flags);
4052 iommu_flush_irt(iommu, devid);
4053 iommu_completion_wait(iommu);
4056 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4057 unsigned int destination, int vector,
4058 struct io_apic_irq_attr *attr)
4060 struct irq_remap_table *table;
4061 struct irq_2_iommu *irte_info;
4062 struct irq_cfg *cfg;
4069 cfg = irq_get_chip_data(irq);
4073 irte_info = &cfg->irq_2_iommu;
4074 ioapic_id = mpc_ioapic_id(attr->ioapic);
4075 devid = get_ioapic_devid(ioapic_id);
4080 table = get_irq_table(devid, true);
4084 index = attr->ioapic_pin;
4086 /* Setup IRQ remapping info */
4087 irte_info->sub_handle = devid;
4088 irte_info->irte_index = index;
4089 irte_info->iommu = (void *)cfg;
4091 /* Setup IRTE for IOMMU */
4093 irte.fields.vector = vector;
4094 irte.fields.int_type = apic->irq_delivery_mode;
4095 irte.fields.destination = destination;
4096 irte.fields.dm = apic->irq_dest_mode;
4097 irte.fields.valid = 1;
4099 ret = modify_irte(devid, index, irte);
4103 /* Setup IOAPIC entry */
4104 memset(entry, 0, sizeof(*entry));
4106 entry->vector = index;
4108 entry->trigger = attr->trigger;
4109 entry->polarity = attr->polarity;
4112 * Mask level triggered irqs.
4120 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4123 struct irq_2_iommu *irte_info;
4124 unsigned int dest, irq;
4125 struct irq_cfg *cfg;
4129 if (!config_enabled(CONFIG_SMP))
4132 cfg = data->chip_data;
4134 irte_info = &cfg->irq_2_iommu;
4136 if (!cpumask_intersects(mask, cpu_online_mask))
4139 if (get_irte(irte_info->sub_handle, irte_info->irte_index, &irte))
4142 if (assign_irq_vector(irq, cfg, mask))
4145 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4147 if (assign_irq_vector(irq, cfg, data->affinity))
4148 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4152 irte.fields.vector = cfg->vector;
4153 irte.fields.destination = dest;
4155 modify_irte(irte_info->sub_handle, irte_info->irte_index, irte);
4157 if (cfg->move_in_progress)
4158 send_cleanup_vector(cfg);
4160 cpumask_copy(data->affinity, mask);
4165 static int free_irq(int irq)
4167 struct irq_2_iommu *irte_info;
4168 struct irq_cfg *cfg;
4170 cfg = irq_get_chip_data(irq);
4174 irte_info = &cfg->irq_2_iommu;
4176 free_irte(irte_info->sub_handle, irte_info->irte_index);
4181 static void compose_msi_msg(struct pci_dev *pdev,
4182 unsigned int irq, unsigned int dest,
4183 struct msi_msg *msg, u8 hpet_id)
4185 struct irq_2_iommu *irte_info;
4186 struct irq_cfg *cfg;
4189 cfg = irq_get_chip_data(irq);
4193 irte_info = &cfg->irq_2_iommu;
4196 irte.fields.vector = cfg->vector;
4197 irte.fields.int_type = apic->irq_delivery_mode;
4198 irte.fields.destination = dest;
4199 irte.fields.dm = apic->irq_dest_mode;
4200 irte.fields.valid = 1;
4202 modify_irte(irte_info->sub_handle, irte_info->irte_index, irte);
4204 msg->address_hi = MSI_ADDR_BASE_HI;
4205 msg->address_lo = MSI_ADDR_BASE_LO;
4206 msg->data = irte_info->irte_index;
4209 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4211 struct irq_cfg *cfg;
4218 cfg = irq_get_chip_data(irq);
4222 devid = get_device_id(&pdev->dev);
4223 index = alloc_irq_index(cfg, devid, nvec);
4225 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4228 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4229 int index, int offset)
4231 struct irq_2_iommu *irte_info;
4232 struct irq_cfg *cfg;
4238 cfg = irq_get_chip_data(irq);
4242 if (index >= MAX_IRQS_PER_TABLE)
4245 devid = get_device_id(&pdev->dev);
4246 irte_info = &cfg->irq_2_iommu;
4248 irte_info->sub_handle = devid;
4249 irte_info->irte_index = index + offset;
4250 irte_info->iommu = (void *)cfg;
4255 static int setup_hpet_msi(unsigned int irq, unsigned int id)
4257 struct irq_2_iommu *irte_info;
4258 struct irq_cfg *cfg;
4261 cfg = irq_get_chip_data(irq);
4265 irte_info = &cfg->irq_2_iommu;
4266 devid = get_hpet_devid(id);
4270 index = alloc_irq_index(cfg, devid, 1);
4274 irte_info->sub_handle = devid;
4275 irte_info->irte_index = index;
4276 irte_info->iommu = (void *)cfg;
4281 struct irq_remap_ops amd_iommu_irq_ops = {
4282 .supported = amd_iommu_supported,
4283 .prepare = amd_iommu_prepare,
4284 .enable = amd_iommu_enable,
4285 .disable = amd_iommu_disable,
4286 .reenable = amd_iommu_reenable,
4287 .enable_faulting = amd_iommu_enable_faulting,
4288 .setup_ioapic_entry = setup_ioapic_entry,
4289 .set_affinity = set_affinity,
4290 .free_irq = free_irq,
4291 .compose_msi_msg = compose_msi_msg,
4292 .msi_alloc_irq = msi_alloc_irq,
4293 .msi_setup_irq = msi_setup_irq,
4294 .setup_hpet_msi = setup_hpet_msi,