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"
51 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
53 #define LOOP_TIMEOUT 100000
56 * This bitmap is used to advertise the page sizes our hardware support
57 * to the IOMMU core, which will then use this information to split
58 * physically contiguous memory regions it is mapping into page sizes
61 * 512GB Pages are not supported due to a hardware bug
63 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
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);
137 iommu_group_put(dev_data->group);
142 static struct iommu_dev_data *search_dev_data(u16 devid)
144 struct iommu_dev_data *dev_data;
147 spin_lock_irqsave(&dev_data_list_lock, flags);
148 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
149 if (dev_data->devid == devid)
156 spin_unlock_irqrestore(&dev_data_list_lock, flags);
161 static struct iommu_dev_data *find_dev_data(u16 devid)
163 struct iommu_dev_data *dev_data;
165 dev_data = search_dev_data(devid);
167 if (dev_data == NULL)
168 dev_data = alloc_dev_data(devid);
173 static inline u16 get_device_id(struct device *dev)
175 struct pci_dev *pdev = to_pci_dev(dev);
177 return calc_devid(pdev->bus->number, pdev->devfn);
180 static struct iommu_dev_data *get_dev_data(struct device *dev)
182 return dev->archdata.iommu;
185 static bool pci_iommuv2_capable(struct pci_dev *pdev)
187 static const int caps[] = {
190 PCI_EXT_CAP_ID_PASID,
194 for (i = 0; i < 3; ++i) {
195 pos = pci_find_ext_capability(pdev, caps[i]);
203 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
205 struct iommu_dev_data *dev_data;
207 dev_data = get_dev_data(&pdev->dev);
209 return dev_data->errata & (1 << erratum) ? true : false;
213 * In this function the list of preallocated protection domains is traversed to
214 * find the domain for a specific device
216 static struct dma_ops_domain *find_protection_domain(u16 devid)
218 struct dma_ops_domain *entry, *ret = NULL;
220 u16 alias = amd_iommu_alias_table[devid];
222 if (list_empty(&iommu_pd_list))
225 spin_lock_irqsave(&iommu_pd_list_lock, flags);
227 list_for_each_entry(entry, &iommu_pd_list, list) {
228 if (entry->target_dev == devid ||
229 entry->target_dev == alias) {
235 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
241 * This function checks if the driver got a valid device from the caller to
242 * avoid dereferencing invalid pointers.
244 static bool check_device(struct device *dev)
248 if (!dev || !dev->dma_mask)
251 /* No device or no PCI device */
252 if (dev->bus != &pci_bus_type)
255 devid = get_device_id(dev);
257 /* Out of our scope? */
258 if (devid > amd_iommu_last_bdf)
261 if (amd_iommu_rlookup_table[devid] == NULL)
267 static struct pci_bus *find_hosted_bus(struct pci_bus *bus)
270 if (!pci_is_root_bus(bus))
273 return ERR_PTR(-ENODEV);
279 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
281 static struct pci_dev *get_isolation_root(struct pci_dev *pdev)
283 struct pci_dev *dma_pdev = pdev;
285 /* Account for quirked devices */
286 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
289 * If it's a multifunction device that does not support our
290 * required ACS flags, add to the same group as function 0.
292 if (dma_pdev->multifunction &&
293 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
294 swap_pci_ref(&dma_pdev,
295 pci_get_slot(dma_pdev->bus,
296 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
300 * Devices on the root bus go through the iommu. If that's not us,
301 * find the next upstream device and test ACS up to the root bus.
302 * Finding the next device may require skipping virtual buses.
304 while (!pci_is_root_bus(dma_pdev->bus)) {
305 struct pci_bus *bus = find_hosted_bus(dma_pdev->bus);
309 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
312 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
318 static int use_pdev_iommu_group(struct pci_dev *pdev, struct device *dev)
320 struct iommu_group *group = iommu_group_get(&pdev->dev);
324 group = iommu_group_alloc();
326 return PTR_ERR(group);
328 WARN_ON(&pdev->dev != dev);
331 ret = iommu_group_add_device(group, dev);
332 iommu_group_put(group);
336 static int use_dev_data_iommu_group(struct iommu_dev_data *dev_data,
339 if (!dev_data->group) {
340 struct iommu_group *group = iommu_group_alloc();
342 return PTR_ERR(group);
344 dev_data->group = group;
347 return iommu_group_add_device(dev_data->group, dev);
350 static int init_iommu_group(struct device *dev)
352 struct iommu_dev_data *dev_data;
353 struct iommu_group *group;
354 struct pci_dev *dma_pdev;
357 group = iommu_group_get(dev);
359 iommu_group_put(group);
363 dev_data = find_dev_data(get_device_id(dev));
367 if (dev_data->alias_data) {
371 if (dev_data->alias_data->group)
375 * If the alias device exists, it's effectively just a first
376 * level quirk for finding the DMA source.
378 alias = amd_iommu_alias_table[dev_data->devid];
379 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
381 dma_pdev = get_isolation_root(dma_pdev);
386 * If the alias is virtual, try to find a parent device
387 * and test whether the IOMMU group is actualy rooted above
388 * the alias. Be careful to also test the parent device if
389 * we think the alias is the root of the group.
391 bus = pci_find_bus(0, alias >> 8);
395 bus = find_hosted_bus(bus);
396 if (IS_ERR(bus) || !bus->self)
399 dma_pdev = get_isolation_root(pci_dev_get(bus->self));
400 if (dma_pdev != bus->self || (dma_pdev->multifunction &&
401 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)))
404 pci_dev_put(dma_pdev);
408 dma_pdev = get_isolation_root(pci_dev_get(to_pci_dev(dev)));
410 ret = use_pdev_iommu_group(dma_pdev, dev);
411 pci_dev_put(dma_pdev);
414 return use_dev_data_iommu_group(dev_data->alias_data, dev);
417 static int iommu_init_device(struct device *dev)
419 struct pci_dev *pdev = to_pci_dev(dev);
420 struct iommu_dev_data *dev_data;
424 if (dev->archdata.iommu)
427 dev_data = find_dev_data(get_device_id(dev));
431 alias = amd_iommu_alias_table[dev_data->devid];
432 if (alias != dev_data->devid) {
433 struct iommu_dev_data *alias_data;
435 alias_data = find_dev_data(alias);
436 if (alias_data == NULL) {
437 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
439 free_dev_data(dev_data);
442 dev_data->alias_data = alias_data;
445 ret = init_iommu_group(dev);
449 if (pci_iommuv2_capable(pdev)) {
450 struct amd_iommu *iommu;
452 iommu = amd_iommu_rlookup_table[dev_data->devid];
453 dev_data->iommu_v2 = iommu->is_iommu_v2;
456 dev->archdata.iommu = dev_data;
461 static void iommu_ignore_device(struct device *dev)
465 devid = get_device_id(dev);
466 alias = amd_iommu_alias_table[devid];
468 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
469 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
471 amd_iommu_rlookup_table[devid] = NULL;
472 amd_iommu_rlookup_table[alias] = NULL;
475 static void iommu_uninit_device(struct device *dev)
477 iommu_group_remove_device(dev);
480 * Nothing to do here - we keep dev_data around for unplugged devices
481 * and reuse it when the device is re-plugged - not doing so would
482 * introduce a ton of races.
486 void __init amd_iommu_uninit_devices(void)
488 struct iommu_dev_data *dev_data, *n;
489 struct pci_dev *pdev = NULL;
491 for_each_pci_dev(pdev) {
493 if (!check_device(&pdev->dev))
496 iommu_uninit_device(&pdev->dev);
499 /* Free all of our dev_data structures */
500 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
501 free_dev_data(dev_data);
504 int __init amd_iommu_init_devices(void)
506 struct pci_dev *pdev = NULL;
509 for_each_pci_dev(pdev) {
511 if (!check_device(&pdev->dev))
514 ret = iommu_init_device(&pdev->dev);
515 if (ret == -ENOTSUPP)
516 iommu_ignore_device(&pdev->dev);
525 amd_iommu_uninit_devices();
529 #ifdef CONFIG_AMD_IOMMU_STATS
532 * Initialization code for statistics collection
535 DECLARE_STATS_COUNTER(compl_wait);
536 DECLARE_STATS_COUNTER(cnt_map_single);
537 DECLARE_STATS_COUNTER(cnt_unmap_single);
538 DECLARE_STATS_COUNTER(cnt_map_sg);
539 DECLARE_STATS_COUNTER(cnt_unmap_sg);
540 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
541 DECLARE_STATS_COUNTER(cnt_free_coherent);
542 DECLARE_STATS_COUNTER(cross_page);
543 DECLARE_STATS_COUNTER(domain_flush_single);
544 DECLARE_STATS_COUNTER(domain_flush_all);
545 DECLARE_STATS_COUNTER(alloced_io_mem);
546 DECLARE_STATS_COUNTER(total_map_requests);
547 DECLARE_STATS_COUNTER(complete_ppr);
548 DECLARE_STATS_COUNTER(invalidate_iotlb);
549 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
550 DECLARE_STATS_COUNTER(pri_requests);
552 static struct dentry *stats_dir;
553 static struct dentry *de_fflush;
555 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
557 if (stats_dir == NULL)
560 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
564 static void amd_iommu_stats_init(void)
566 stats_dir = debugfs_create_dir("amd-iommu", NULL);
567 if (stats_dir == NULL)
570 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
571 &amd_iommu_unmap_flush);
573 amd_iommu_stats_add(&compl_wait);
574 amd_iommu_stats_add(&cnt_map_single);
575 amd_iommu_stats_add(&cnt_unmap_single);
576 amd_iommu_stats_add(&cnt_map_sg);
577 amd_iommu_stats_add(&cnt_unmap_sg);
578 amd_iommu_stats_add(&cnt_alloc_coherent);
579 amd_iommu_stats_add(&cnt_free_coherent);
580 amd_iommu_stats_add(&cross_page);
581 amd_iommu_stats_add(&domain_flush_single);
582 amd_iommu_stats_add(&domain_flush_all);
583 amd_iommu_stats_add(&alloced_io_mem);
584 amd_iommu_stats_add(&total_map_requests);
585 amd_iommu_stats_add(&complete_ppr);
586 amd_iommu_stats_add(&invalidate_iotlb);
587 amd_iommu_stats_add(&invalidate_iotlb_all);
588 amd_iommu_stats_add(&pri_requests);
593 /****************************************************************************
595 * Interrupt handling functions
597 ****************************************************************************/
599 static void dump_dte_entry(u16 devid)
603 for (i = 0; i < 4; ++i)
604 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
605 amd_iommu_dev_table[devid].data[i]);
608 static void dump_command(unsigned long phys_addr)
610 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
613 for (i = 0; i < 4; ++i)
614 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
617 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
619 int type, devid, domid, flags;
620 volatile u32 *event = __evt;
625 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
626 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
627 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
628 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
629 address = (u64)(((u64)event[3]) << 32) | event[2];
632 /* Did we hit the erratum? */
633 if (++count == LOOP_TIMEOUT) {
634 pr_err("AMD-Vi: No event written to event log\n");
641 printk(KERN_ERR "AMD-Vi: Event logged [");
644 case EVENT_TYPE_ILL_DEV:
645 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
646 "address=0x%016llx flags=0x%04x]\n",
647 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
649 dump_dte_entry(devid);
651 case EVENT_TYPE_IO_FAULT:
652 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
653 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
654 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
655 domid, address, flags);
657 case EVENT_TYPE_DEV_TAB_ERR:
658 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
659 "address=0x%016llx flags=0x%04x]\n",
660 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
663 case EVENT_TYPE_PAGE_TAB_ERR:
664 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
665 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
666 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
667 domid, address, flags);
669 case EVENT_TYPE_ILL_CMD:
670 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
671 dump_command(address);
673 case EVENT_TYPE_CMD_HARD_ERR:
674 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
675 "flags=0x%04x]\n", address, flags);
677 case EVENT_TYPE_IOTLB_INV_TO:
678 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
679 "address=0x%016llx]\n",
680 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
683 case EVENT_TYPE_INV_DEV_REQ:
684 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
685 "address=0x%016llx flags=0x%04x]\n",
686 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
690 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
693 memset(__evt, 0, 4 * sizeof(u32));
696 static void iommu_poll_events(struct amd_iommu *iommu)
701 spin_lock_irqsave(&iommu->lock, flags);
703 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
704 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
706 while (head != tail) {
707 iommu_print_event(iommu, iommu->evt_buf + head);
708 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
711 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
713 spin_unlock_irqrestore(&iommu->lock, flags);
716 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
718 struct amd_iommu_fault fault;
720 INC_STATS_COUNTER(pri_requests);
722 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
723 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
727 fault.address = raw[1];
728 fault.pasid = PPR_PASID(raw[0]);
729 fault.device_id = PPR_DEVID(raw[0]);
730 fault.tag = PPR_TAG(raw[0]);
731 fault.flags = PPR_FLAGS(raw[0]);
733 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
736 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
741 if (iommu->ppr_log == NULL)
744 /* enable ppr interrupts again */
745 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
747 spin_lock_irqsave(&iommu->lock, flags);
749 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
750 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
752 while (head != tail) {
757 raw = (u64 *)(iommu->ppr_log + head);
760 * Hardware bug: Interrupt may arrive before the entry is
761 * written to memory. If this happens we need to wait for the
764 for (i = 0; i < LOOP_TIMEOUT; ++i) {
765 if (PPR_REQ_TYPE(raw[0]) != 0)
770 /* Avoid memcpy function-call overhead */
775 * To detect the hardware bug we need to clear the entry
778 raw[0] = raw[1] = 0UL;
780 /* Update head pointer of hardware ring-buffer */
781 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
782 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
785 * Release iommu->lock because ppr-handling might need to
788 spin_unlock_irqrestore(&iommu->lock, flags);
790 /* Handle PPR entry */
791 iommu_handle_ppr_entry(iommu, entry);
793 spin_lock_irqsave(&iommu->lock, flags);
795 /* Refresh ring-buffer information */
796 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
797 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
800 spin_unlock_irqrestore(&iommu->lock, flags);
803 irqreturn_t amd_iommu_int_thread(int irq, void *data)
805 struct amd_iommu *iommu;
807 for_each_iommu(iommu) {
808 iommu_poll_events(iommu);
809 iommu_poll_ppr_log(iommu);
815 irqreturn_t amd_iommu_int_handler(int irq, void *data)
817 return IRQ_WAKE_THREAD;
820 /****************************************************************************
822 * IOMMU command queuing functions
824 ****************************************************************************/
826 static int wait_on_sem(volatile u64 *sem)
830 while (*sem == 0 && i < LOOP_TIMEOUT) {
835 if (i == LOOP_TIMEOUT) {
836 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
843 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
844 struct iommu_cmd *cmd,
849 target = iommu->cmd_buf + tail;
850 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
852 /* Copy command to buffer */
853 memcpy(target, cmd, sizeof(*cmd));
855 /* Tell the IOMMU about it */
856 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
859 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
861 WARN_ON(address & 0x7ULL);
863 memset(cmd, 0, sizeof(*cmd));
864 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
865 cmd->data[1] = upper_32_bits(__pa(address));
867 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
870 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
872 memset(cmd, 0, sizeof(*cmd));
873 cmd->data[0] = devid;
874 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
877 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
878 size_t size, u16 domid, int pde)
883 pages = iommu_num_pages(address, size, PAGE_SIZE);
888 * If we have to flush more than one page, flush all
889 * TLB entries for this domain
891 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
895 address &= PAGE_MASK;
897 memset(cmd, 0, sizeof(*cmd));
898 cmd->data[1] |= domid;
899 cmd->data[2] = lower_32_bits(address);
900 cmd->data[3] = upper_32_bits(address);
901 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
902 if (s) /* size bit - we flush more than one 4kb page */
903 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
904 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
905 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
908 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
909 u64 address, size_t size)
914 pages = iommu_num_pages(address, size, PAGE_SIZE);
919 * If we have to flush more than one page, flush all
920 * TLB entries for this domain
922 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
926 address &= PAGE_MASK;
928 memset(cmd, 0, sizeof(*cmd));
929 cmd->data[0] = devid;
930 cmd->data[0] |= (qdep & 0xff) << 24;
931 cmd->data[1] = devid;
932 cmd->data[2] = lower_32_bits(address);
933 cmd->data[3] = upper_32_bits(address);
934 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
936 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
939 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
940 u64 address, bool size)
942 memset(cmd, 0, sizeof(*cmd));
944 address &= ~(0xfffULL);
946 cmd->data[0] = pasid & PASID_MASK;
947 cmd->data[1] = domid;
948 cmd->data[2] = lower_32_bits(address);
949 cmd->data[3] = upper_32_bits(address);
950 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
951 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
953 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
954 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
957 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
958 int qdep, u64 address, bool size)
960 memset(cmd, 0, sizeof(*cmd));
962 address &= ~(0xfffULL);
964 cmd->data[0] = devid;
965 cmd->data[0] |= (pasid & 0xff) << 16;
966 cmd->data[0] |= (qdep & 0xff) << 24;
967 cmd->data[1] = devid;
968 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
969 cmd->data[2] = lower_32_bits(address);
970 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
971 cmd->data[3] = upper_32_bits(address);
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
974 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
977 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
978 int status, int tag, bool gn)
980 memset(cmd, 0, sizeof(*cmd));
982 cmd->data[0] = devid;
984 cmd->data[1] = pasid & PASID_MASK;
985 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
987 cmd->data[3] = tag & 0x1ff;
988 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
990 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
993 static void build_inv_all(struct iommu_cmd *cmd)
995 memset(cmd, 0, sizeof(*cmd));
996 CMD_SET_TYPE(cmd, CMD_INV_ALL);
999 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1001 memset(cmd, 0, sizeof(*cmd));
1002 cmd->data[0] = devid;
1003 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1007 * Writes the command to the IOMMUs command buffer and informs the
1008 * hardware about the new command.
1010 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1011 struct iommu_cmd *cmd,
1014 u32 left, tail, head, next_tail;
1015 unsigned long flags;
1017 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
1020 spin_lock_irqsave(&iommu->lock, flags);
1022 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1023 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1024 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
1025 left = (head - next_tail) % iommu->cmd_buf_size;
1028 struct iommu_cmd sync_cmd;
1029 volatile u64 sem = 0;
1032 build_completion_wait(&sync_cmd, (u64)&sem);
1033 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1035 spin_unlock_irqrestore(&iommu->lock, flags);
1037 if ((ret = wait_on_sem(&sem)) != 0)
1043 copy_cmd_to_buffer(iommu, cmd, tail);
1045 /* We need to sync now to make sure all commands are processed */
1046 iommu->need_sync = sync;
1048 spin_unlock_irqrestore(&iommu->lock, flags);
1053 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1055 return iommu_queue_command_sync(iommu, cmd, true);
1059 * This function queues a completion wait command into the command
1060 * buffer of an IOMMU
1062 static int iommu_completion_wait(struct amd_iommu *iommu)
1064 struct iommu_cmd cmd;
1065 volatile u64 sem = 0;
1068 if (!iommu->need_sync)
1071 build_completion_wait(&cmd, (u64)&sem);
1073 ret = iommu_queue_command_sync(iommu, &cmd, false);
1077 return wait_on_sem(&sem);
1080 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1082 struct iommu_cmd cmd;
1084 build_inv_dte(&cmd, devid);
1086 return iommu_queue_command(iommu, &cmd);
1089 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1093 for (devid = 0; devid <= 0xffff; ++devid)
1094 iommu_flush_dte(iommu, devid);
1096 iommu_completion_wait(iommu);
1100 * This function uses heavy locking and may disable irqs for some time. But
1101 * this is no issue because it is only called during resume.
1103 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1107 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1108 struct iommu_cmd cmd;
1109 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1111 iommu_queue_command(iommu, &cmd);
1114 iommu_completion_wait(iommu);
1117 static void iommu_flush_all(struct amd_iommu *iommu)
1119 struct iommu_cmd cmd;
1121 build_inv_all(&cmd);
1123 iommu_queue_command(iommu, &cmd);
1124 iommu_completion_wait(iommu);
1127 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1129 struct iommu_cmd cmd;
1131 build_inv_irt(&cmd, devid);
1133 iommu_queue_command(iommu, &cmd);
1136 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1140 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1141 iommu_flush_irt(iommu, devid);
1143 iommu_completion_wait(iommu);
1146 void iommu_flush_all_caches(struct amd_iommu *iommu)
1148 if (iommu_feature(iommu, FEATURE_IA)) {
1149 iommu_flush_all(iommu);
1151 iommu_flush_dte_all(iommu);
1152 iommu_flush_irt_all(iommu);
1153 iommu_flush_tlb_all(iommu);
1158 * Command send function for flushing on-device TLB
1160 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1161 u64 address, size_t size)
1163 struct amd_iommu *iommu;
1164 struct iommu_cmd cmd;
1167 qdep = dev_data->ats.qdep;
1168 iommu = amd_iommu_rlookup_table[dev_data->devid];
1170 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1172 return iommu_queue_command(iommu, &cmd);
1176 * Command send function for invalidating a device table entry
1178 static int device_flush_dte(struct iommu_dev_data *dev_data)
1180 struct amd_iommu *iommu;
1183 iommu = amd_iommu_rlookup_table[dev_data->devid];
1185 ret = iommu_flush_dte(iommu, dev_data->devid);
1189 if (dev_data->ats.enabled)
1190 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1196 * TLB invalidation function which is called from the mapping functions.
1197 * It invalidates a single PTE if the range to flush is within a single
1198 * page. Otherwise it flushes the whole TLB of the IOMMU.
1200 static void __domain_flush_pages(struct protection_domain *domain,
1201 u64 address, size_t size, int pde)
1203 struct iommu_dev_data *dev_data;
1204 struct iommu_cmd cmd;
1207 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1209 for (i = 0; i < amd_iommus_present; ++i) {
1210 if (!domain->dev_iommu[i])
1214 * Devices of this domain are behind this IOMMU
1215 * We need a TLB flush
1217 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1220 list_for_each_entry(dev_data, &domain->dev_list, list) {
1222 if (!dev_data->ats.enabled)
1225 ret |= device_flush_iotlb(dev_data, address, size);
1231 static void domain_flush_pages(struct protection_domain *domain,
1232 u64 address, size_t size)
1234 __domain_flush_pages(domain, address, size, 0);
1237 /* Flush the whole IO/TLB for a given protection domain */
1238 static void domain_flush_tlb(struct protection_domain *domain)
1240 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1243 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1244 static void domain_flush_tlb_pde(struct protection_domain *domain)
1246 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1249 static void domain_flush_complete(struct protection_domain *domain)
1253 for (i = 0; i < amd_iommus_present; ++i) {
1254 if (!domain->dev_iommu[i])
1258 * Devices of this domain are behind this IOMMU
1259 * We need to wait for completion of all commands.
1261 iommu_completion_wait(amd_iommus[i]);
1267 * This function flushes the DTEs for all devices in domain
1269 static void domain_flush_devices(struct protection_domain *domain)
1271 struct iommu_dev_data *dev_data;
1273 list_for_each_entry(dev_data, &domain->dev_list, list)
1274 device_flush_dte(dev_data);
1277 /****************************************************************************
1279 * The functions below are used the create the page table mappings for
1280 * unity mapped regions.
1282 ****************************************************************************/
1285 * This function is used to add another level to an IO page table. Adding
1286 * another level increases the size of the address space by 9 bits to a size up
1289 static bool increase_address_space(struct protection_domain *domain,
1294 if (domain->mode == PAGE_MODE_6_LEVEL)
1295 /* address space already 64 bit large */
1298 pte = (void *)get_zeroed_page(gfp);
1302 *pte = PM_LEVEL_PDE(domain->mode,
1303 virt_to_phys(domain->pt_root));
1304 domain->pt_root = pte;
1306 domain->updated = true;
1311 static u64 *alloc_pte(struct protection_domain *domain,
1312 unsigned long address,
1313 unsigned long page_size,
1320 BUG_ON(!is_power_of_2(page_size));
1322 while (address > PM_LEVEL_SIZE(domain->mode))
1323 increase_address_space(domain, gfp);
1325 level = domain->mode - 1;
1326 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1327 address = PAGE_SIZE_ALIGN(address, page_size);
1328 end_lvl = PAGE_SIZE_LEVEL(page_size);
1330 while (level > end_lvl) {
1331 if (!IOMMU_PTE_PRESENT(*pte)) {
1332 page = (u64 *)get_zeroed_page(gfp);
1335 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1338 /* No level skipping support yet */
1339 if (PM_PTE_LEVEL(*pte) != level)
1344 pte = IOMMU_PTE_PAGE(*pte);
1346 if (pte_page && level == end_lvl)
1349 pte = &pte[PM_LEVEL_INDEX(level, address)];
1356 * This function checks if there is a PTE for a given dma address. If
1357 * there is one, it returns the pointer to it.
1359 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1364 if (address > PM_LEVEL_SIZE(domain->mode))
1367 level = domain->mode - 1;
1368 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1373 if (!IOMMU_PTE_PRESENT(*pte))
1377 if (PM_PTE_LEVEL(*pte) == 0x07) {
1378 unsigned long pte_mask, __pte;
1381 * If we have a series of large PTEs, make
1382 * sure to return a pointer to the first one.
1384 pte_mask = PTE_PAGE_SIZE(*pte);
1385 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1386 __pte = ((unsigned long)pte) & pte_mask;
1388 return (u64 *)__pte;
1391 /* No level skipping support yet */
1392 if (PM_PTE_LEVEL(*pte) != level)
1397 /* Walk to the next level */
1398 pte = IOMMU_PTE_PAGE(*pte);
1399 pte = &pte[PM_LEVEL_INDEX(level, address)];
1406 * Generic mapping functions. It maps a physical address into a DMA
1407 * address space. It allocates the page table pages if necessary.
1408 * In the future it can be extended to a generic mapping function
1409 * supporting all features of AMD IOMMU page tables like level skipping
1410 * and full 64 bit address spaces.
1412 static int iommu_map_page(struct protection_domain *dom,
1413 unsigned long bus_addr,
1414 unsigned long phys_addr,
1416 unsigned long page_size)
1421 if (!(prot & IOMMU_PROT_MASK))
1424 bus_addr = PAGE_ALIGN(bus_addr);
1425 phys_addr = PAGE_ALIGN(phys_addr);
1426 count = PAGE_SIZE_PTE_COUNT(page_size);
1427 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1429 for (i = 0; i < count; ++i)
1430 if (IOMMU_PTE_PRESENT(pte[i]))
1433 if (page_size > PAGE_SIZE) {
1434 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1435 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1437 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1439 if (prot & IOMMU_PROT_IR)
1440 __pte |= IOMMU_PTE_IR;
1441 if (prot & IOMMU_PROT_IW)
1442 __pte |= IOMMU_PTE_IW;
1444 for (i = 0; i < count; ++i)
1452 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1453 unsigned long bus_addr,
1454 unsigned long page_size)
1456 unsigned long long unmap_size, unmapped;
1459 BUG_ON(!is_power_of_2(page_size));
1463 while (unmapped < page_size) {
1465 pte = fetch_pte(dom, bus_addr);
1469 * No PTE for this address
1470 * move forward in 4kb steps
1472 unmap_size = PAGE_SIZE;
1473 } else if (PM_PTE_LEVEL(*pte) == 0) {
1474 /* 4kb PTE found for this address */
1475 unmap_size = PAGE_SIZE;
1480 /* Large PTE found which maps this address */
1481 unmap_size = PTE_PAGE_SIZE(*pte);
1482 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1483 for (i = 0; i < count; i++)
1487 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1488 unmapped += unmap_size;
1491 BUG_ON(!is_power_of_2(unmapped));
1497 * This function checks if a specific unity mapping entry is needed for
1498 * this specific IOMMU.
1500 static int iommu_for_unity_map(struct amd_iommu *iommu,
1501 struct unity_map_entry *entry)
1505 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1506 bdf = amd_iommu_alias_table[i];
1507 if (amd_iommu_rlookup_table[bdf] == iommu)
1515 * This function actually applies the mapping to the page table of the
1518 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1519 struct unity_map_entry *e)
1524 for (addr = e->address_start; addr < e->address_end;
1525 addr += PAGE_SIZE) {
1526 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1531 * if unity mapping is in aperture range mark the page
1532 * as allocated in the aperture
1534 if (addr < dma_dom->aperture_size)
1535 __set_bit(addr >> PAGE_SHIFT,
1536 dma_dom->aperture[0]->bitmap);
1543 * Init the unity mappings for a specific IOMMU in the system
1545 * Basically iterates over all unity mapping entries and applies them to
1546 * the default domain DMA of that IOMMU if necessary.
1548 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1550 struct unity_map_entry *entry;
1553 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1554 if (!iommu_for_unity_map(iommu, entry))
1556 ret = dma_ops_unity_map(iommu->default_dom, entry);
1565 * Inits the unity mappings required for a specific device
1567 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1570 struct unity_map_entry *e;
1573 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1574 if (!(devid >= e->devid_start && devid <= e->devid_end))
1576 ret = dma_ops_unity_map(dma_dom, e);
1584 /****************************************************************************
1586 * The next functions belong to the address allocator for the dma_ops
1587 * interface functions. They work like the allocators in the other IOMMU
1588 * drivers. Its basically a bitmap which marks the allocated pages in
1589 * the aperture. Maybe it could be enhanced in the future to a more
1590 * efficient allocator.
1592 ****************************************************************************/
1595 * The address allocator core functions.
1597 * called with domain->lock held
1601 * Used to reserve address ranges in the aperture (e.g. for exclusion
1604 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1605 unsigned long start_page,
1608 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1610 if (start_page + pages > last_page)
1611 pages = last_page - start_page;
1613 for (i = start_page; i < start_page + pages; ++i) {
1614 int index = i / APERTURE_RANGE_PAGES;
1615 int page = i % APERTURE_RANGE_PAGES;
1616 __set_bit(page, dom->aperture[index]->bitmap);
1621 * This function is used to add a new aperture range to an existing
1622 * aperture in case of dma_ops domain allocation or address allocation
1625 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1626 bool populate, gfp_t gfp)
1628 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1629 struct amd_iommu *iommu;
1630 unsigned long i, old_size;
1632 #ifdef CONFIG_IOMMU_STRESS
1636 if (index >= APERTURE_MAX_RANGES)
1639 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1640 if (!dma_dom->aperture[index])
1643 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1644 if (!dma_dom->aperture[index]->bitmap)
1647 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1650 unsigned long address = dma_dom->aperture_size;
1651 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1652 u64 *pte, *pte_page;
1654 for (i = 0; i < num_ptes; ++i) {
1655 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1660 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1662 address += APERTURE_RANGE_SIZE / 64;
1666 old_size = dma_dom->aperture_size;
1667 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1669 /* Reserve address range used for MSI messages */
1670 if (old_size < MSI_ADDR_BASE_LO &&
1671 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1672 unsigned long spage;
1675 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1676 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1678 dma_ops_reserve_addresses(dma_dom, spage, pages);
1681 /* Initialize the exclusion range if necessary */
1682 for_each_iommu(iommu) {
1683 if (iommu->exclusion_start &&
1684 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1685 && iommu->exclusion_start < dma_dom->aperture_size) {
1686 unsigned long startpage;
1687 int pages = iommu_num_pages(iommu->exclusion_start,
1688 iommu->exclusion_length,
1690 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1691 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1696 * Check for areas already mapped as present in the new aperture
1697 * range and mark those pages as reserved in the allocator. Such
1698 * mappings may already exist as a result of requested unity
1699 * mappings for devices.
1701 for (i = dma_dom->aperture[index]->offset;
1702 i < dma_dom->aperture_size;
1704 u64 *pte = fetch_pte(&dma_dom->domain, i);
1705 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1708 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1711 update_domain(&dma_dom->domain);
1716 update_domain(&dma_dom->domain);
1718 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1720 kfree(dma_dom->aperture[index]);
1721 dma_dom->aperture[index] = NULL;
1726 static unsigned long dma_ops_area_alloc(struct device *dev,
1727 struct dma_ops_domain *dom,
1729 unsigned long align_mask,
1731 unsigned long start)
1733 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1734 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1735 int i = start >> APERTURE_RANGE_SHIFT;
1736 unsigned long boundary_size;
1737 unsigned long address = -1;
1738 unsigned long limit;
1740 next_bit >>= PAGE_SHIFT;
1742 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1743 PAGE_SIZE) >> PAGE_SHIFT;
1745 for (;i < max_index; ++i) {
1746 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1748 if (dom->aperture[i]->offset >= dma_mask)
1751 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1752 dma_mask >> PAGE_SHIFT);
1754 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1755 limit, next_bit, pages, 0,
1756 boundary_size, align_mask);
1757 if (address != -1) {
1758 address = dom->aperture[i]->offset +
1759 (address << PAGE_SHIFT);
1760 dom->next_address = address + (pages << PAGE_SHIFT);
1770 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1771 struct dma_ops_domain *dom,
1773 unsigned long align_mask,
1776 unsigned long address;
1778 #ifdef CONFIG_IOMMU_STRESS
1779 dom->next_address = 0;
1780 dom->need_flush = true;
1783 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1784 dma_mask, dom->next_address);
1786 if (address == -1) {
1787 dom->next_address = 0;
1788 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1790 dom->need_flush = true;
1793 if (unlikely(address == -1))
1794 address = DMA_ERROR_CODE;
1796 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1802 * The address free function.
1804 * called with domain->lock held
1806 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1807 unsigned long address,
1810 unsigned i = address >> APERTURE_RANGE_SHIFT;
1811 struct aperture_range *range = dom->aperture[i];
1813 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1815 #ifdef CONFIG_IOMMU_STRESS
1820 if (address >= dom->next_address)
1821 dom->need_flush = true;
1823 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1825 bitmap_clear(range->bitmap, address, pages);
1829 /****************************************************************************
1831 * The next functions belong to the domain allocation. A domain is
1832 * allocated for every IOMMU as the default domain. If device isolation
1833 * is enabled, every device get its own domain. The most important thing
1834 * about domains is the page table mapping the DMA address space they
1837 ****************************************************************************/
1840 * This function adds a protection domain to the global protection domain list
1842 static void add_domain_to_list(struct protection_domain *domain)
1844 unsigned long flags;
1846 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1847 list_add(&domain->list, &amd_iommu_pd_list);
1848 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1852 * This function removes a protection domain to the global
1853 * protection domain list
1855 static void del_domain_from_list(struct protection_domain *domain)
1857 unsigned long flags;
1859 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1860 list_del(&domain->list);
1861 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1864 static u16 domain_id_alloc(void)
1866 unsigned long flags;
1869 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1870 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1872 if (id > 0 && id < MAX_DOMAIN_ID)
1873 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1876 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1881 static void domain_id_free(int id)
1883 unsigned long flags;
1885 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1886 if (id > 0 && id < MAX_DOMAIN_ID)
1887 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1888 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1891 static void free_pagetable(struct protection_domain *domain)
1896 p1 = domain->pt_root;
1901 for (i = 0; i < 512; ++i) {
1902 if (!IOMMU_PTE_PRESENT(p1[i]))
1905 p2 = IOMMU_PTE_PAGE(p1[i]);
1906 for (j = 0; j < 512; ++j) {
1907 if (!IOMMU_PTE_PRESENT(p2[j]))
1909 p3 = IOMMU_PTE_PAGE(p2[j]);
1910 free_page((unsigned long)p3);
1913 free_page((unsigned long)p2);
1916 free_page((unsigned long)p1);
1918 domain->pt_root = NULL;
1921 static void free_gcr3_tbl_level1(u64 *tbl)
1926 for (i = 0; i < 512; ++i) {
1927 if (!(tbl[i] & GCR3_VALID))
1930 ptr = __va(tbl[i] & PAGE_MASK);
1932 free_page((unsigned long)ptr);
1936 static void free_gcr3_tbl_level2(u64 *tbl)
1941 for (i = 0; i < 512; ++i) {
1942 if (!(tbl[i] & GCR3_VALID))
1945 ptr = __va(tbl[i] & PAGE_MASK);
1947 free_gcr3_tbl_level1(ptr);
1951 static void free_gcr3_table(struct protection_domain *domain)
1953 if (domain->glx == 2)
1954 free_gcr3_tbl_level2(domain->gcr3_tbl);
1955 else if (domain->glx == 1)
1956 free_gcr3_tbl_level1(domain->gcr3_tbl);
1957 else if (domain->glx != 0)
1960 free_page((unsigned long)domain->gcr3_tbl);
1964 * Free a domain, only used if something went wrong in the
1965 * allocation path and we need to free an already allocated page table
1967 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1974 del_domain_from_list(&dom->domain);
1976 free_pagetable(&dom->domain);
1978 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1979 if (!dom->aperture[i])
1981 free_page((unsigned long)dom->aperture[i]->bitmap);
1982 kfree(dom->aperture[i]);
1989 * Allocates a new protection domain usable for the dma_ops functions.
1990 * It also initializes the page table and the address allocator data
1991 * structures required for the dma_ops interface
1993 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1995 struct dma_ops_domain *dma_dom;
1997 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2001 spin_lock_init(&dma_dom->domain.lock);
2003 dma_dom->domain.id = domain_id_alloc();
2004 if (dma_dom->domain.id == 0)
2006 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
2007 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2008 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2009 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2010 dma_dom->domain.priv = dma_dom;
2011 if (!dma_dom->domain.pt_root)
2014 dma_dom->need_flush = false;
2015 dma_dom->target_dev = 0xffff;
2017 add_domain_to_list(&dma_dom->domain);
2019 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2023 * mark the first page as allocated so we never return 0 as
2024 * a valid dma-address. So we can use 0 as error value
2026 dma_dom->aperture[0]->bitmap[0] = 1;
2027 dma_dom->next_address = 0;
2033 dma_ops_domain_free(dma_dom);
2039 * little helper function to check whether a given protection domain is a
2042 static bool dma_ops_domain(struct protection_domain *domain)
2044 return domain->flags & PD_DMA_OPS_MASK;
2047 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2052 if (domain->mode != PAGE_MODE_NONE)
2053 pte_root = virt_to_phys(domain->pt_root);
2055 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2056 << DEV_ENTRY_MODE_SHIFT;
2057 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2059 flags = amd_iommu_dev_table[devid].data[1];
2062 flags |= DTE_FLAG_IOTLB;
2064 if (domain->flags & PD_IOMMUV2_MASK) {
2065 u64 gcr3 = __pa(domain->gcr3_tbl);
2066 u64 glx = domain->glx;
2069 pte_root |= DTE_FLAG_GV;
2070 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2072 /* First mask out possible old values for GCR3 table */
2073 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2076 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2079 /* Encode GCR3 table into DTE */
2080 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2083 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2086 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2090 flags &= ~(0xffffUL);
2091 flags |= domain->id;
2093 amd_iommu_dev_table[devid].data[1] = flags;
2094 amd_iommu_dev_table[devid].data[0] = pte_root;
2097 static void clear_dte_entry(u16 devid)
2099 /* remove entry from the device table seen by the hardware */
2100 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2101 amd_iommu_dev_table[devid].data[1] = 0;
2103 amd_iommu_apply_erratum_63(devid);
2106 static void do_attach(struct iommu_dev_data *dev_data,
2107 struct protection_domain *domain)
2109 struct amd_iommu *iommu;
2112 iommu = amd_iommu_rlookup_table[dev_data->devid];
2113 ats = dev_data->ats.enabled;
2115 /* Update data structures */
2116 dev_data->domain = domain;
2117 list_add(&dev_data->list, &domain->dev_list);
2118 set_dte_entry(dev_data->devid, domain, ats);
2120 /* Do reference counting */
2121 domain->dev_iommu[iommu->index] += 1;
2122 domain->dev_cnt += 1;
2124 /* Flush the DTE entry */
2125 device_flush_dte(dev_data);
2128 static void do_detach(struct iommu_dev_data *dev_data)
2130 struct amd_iommu *iommu;
2132 iommu = amd_iommu_rlookup_table[dev_data->devid];
2134 /* decrease reference counters */
2135 dev_data->domain->dev_iommu[iommu->index] -= 1;
2136 dev_data->domain->dev_cnt -= 1;
2138 /* Update data structures */
2139 dev_data->domain = NULL;
2140 list_del(&dev_data->list);
2141 clear_dte_entry(dev_data->devid);
2143 /* Flush the DTE entry */
2144 device_flush_dte(dev_data);
2148 * If a device is not yet associated with a domain, this function does
2149 * assigns it visible for the hardware
2151 static int __attach_device(struct iommu_dev_data *dev_data,
2152 struct protection_domain *domain)
2157 spin_lock(&domain->lock);
2159 if (dev_data->alias_data != NULL) {
2160 struct iommu_dev_data *alias_data = dev_data->alias_data;
2162 /* Some sanity checks */
2164 if (alias_data->domain != NULL &&
2165 alias_data->domain != domain)
2168 if (dev_data->domain != NULL &&
2169 dev_data->domain != domain)
2172 /* Do real assignment */
2173 if (alias_data->domain == NULL)
2174 do_attach(alias_data, domain);
2176 atomic_inc(&alias_data->bind);
2179 if (dev_data->domain == NULL)
2180 do_attach(dev_data, domain);
2182 atomic_inc(&dev_data->bind);
2189 spin_unlock(&domain->lock);
2195 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2197 pci_disable_ats(pdev);
2198 pci_disable_pri(pdev);
2199 pci_disable_pasid(pdev);
2202 /* FIXME: Change generic reset-function to do the same */
2203 static int pri_reset_while_enabled(struct pci_dev *pdev)
2208 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2212 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2213 control |= PCI_PRI_CTRL_RESET;
2214 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2219 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2224 /* FIXME: Hardcode number of outstanding requests for now */
2226 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2228 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2230 /* Only allow access to user-accessible pages */
2231 ret = pci_enable_pasid(pdev, 0);
2235 /* First reset the PRI state of the device */
2236 ret = pci_reset_pri(pdev);
2241 ret = pci_enable_pri(pdev, reqs);
2246 ret = pri_reset_while_enabled(pdev);
2251 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2258 pci_disable_pri(pdev);
2259 pci_disable_pasid(pdev);
2264 /* FIXME: Move this to PCI code */
2265 #define PCI_PRI_TLP_OFF (1 << 15)
2267 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2272 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2276 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2278 return (status & PCI_PRI_TLP_OFF) ? true : false;
2282 * If a device is not yet associated with a domain, this function
2283 * assigns it visible for the hardware
2285 static int attach_device(struct device *dev,
2286 struct protection_domain *domain)
2288 struct pci_dev *pdev = to_pci_dev(dev);
2289 struct iommu_dev_data *dev_data;
2290 unsigned long flags;
2293 dev_data = get_dev_data(dev);
2295 if (domain->flags & PD_IOMMUV2_MASK) {
2296 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2299 if (pdev_iommuv2_enable(pdev) != 0)
2302 dev_data->ats.enabled = true;
2303 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2304 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2305 } else if (amd_iommu_iotlb_sup &&
2306 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2307 dev_data->ats.enabled = true;
2308 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2311 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2312 ret = __attach_device(dev_data, domain);
2313 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2316 * We might boot into a crash-kernel here. The crashed kernel
2317 * left the caches in the IOMMU dirty. So we have to flush
2318 * here to evict all dirty stuff.
2320 domain_flush_tlb_pde(domain);
2326 * Removes a device from a protection domain (unlocked)
2328 static void __detach_device(struct iommu_dev_data *dev_data)
2330 struct protection_domain *domain;
2331 unsigned long flags;
2333 BUG_ON(!dev_data->domain);
2335 domain = dev_data->domain;
2337 spin_lock_irqsave(&domain->lock, flags);
2339 if (dev_data->alias_data != NULL) {
2340 struct iommu_dev_data *alias_data = dev_data->alias_data;
2342 if (atomic_dec_and_test(&alias_data->bind))
2343 do_detach(alias_data);
2346 if (atomic_dec_and_test(&dev_data->bind))
2347 do_detach(dev_data);
2349 spin_unlock_irqrestore(&domain->lock, flags);
2352 * If we run in passthrough mode the device must be assigned to the
2353 * passthrough domain if it is detached from any other domain.
2354 * Make sure we can deassign from the pt_domain itself.
2356 if (dev_data->passthrough &&
2357 (dev_data->domain == NULL && domain != pt_domain))
2358 __attach_device(dev_data, pt_domain);
2362 * Removes a device from a protection domain (with devtable_lock held)
2364 static void detach_device(struct device *dev)
2366 struct protection_domain *domain;
2367 struct iommu_dev_data *dev_data;
2368 unsigned long flags;
2370 dev_data = get_dev_data(dev);
2371 domain = dev_data->domain;
2373 /* lock device table */
2374 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2375 __detach_device(dev_data);
2376 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2378 if (domain->flags & PD_IOMMUV2_MASK)
2379 pdev_iommuv2_disable(to_pci_dev(dev));
2380 else if (dev_data->ats.enabled)
2381 pci_disable_ats(to_pci_dev(dev));
2383 dev_data->ats.enabled = false;
2387 * Find out the protection domain structure for a given PCI device. This
2388 * will give us the pointer to the page table root for example.
2390 static struct protection_domain *domain_for_device(struct device *dev)
2392 struct iommu_dev_data *dev_data;
2393 struct protection_domain *dom = NULL;
2394 unsigned long flags;
2396 dev_data = get_dev_data(dev);
2398 if (dev_data->domain)
2399 return dev_data->domain;
2401 if (dev_data->alias_data != NULL) {
2402 struct iommu_dev_data *alias_data = dev_data->alias_data;
2404 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2405 if (alias_data->domain != NULL) {
2406 __attach_device(dev_data, alias_data->domain);
2407 dom = alias_data->domain;
2409 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2415 static int device_change_notifier(struct notifier_block *nb,
2416 unsigned long action, void *data)
2418 struct dma_ops_domain *dma_domain;
2419 struct protection_domain *domain;
2420 struct iommu_dev_data *dev_data;
2421 struct device *dev = data;
2422 struct amd_iommu *iommu;
2423 unsigned long flags;
2426 if (!check_device(dev))
2429 devid = get_device_id(dev);
2430 iommu = amd_iommu_rlookup_table[devid];
2431 dev_data = get_dev_data(dev);
2434 case BUS_NOTIFY_UNBOUND_DRIVER:
2436 domain = domain_for_device(dev);
2440 if (dev_data->passthrough)
2444 case BUS_NOTIFY_ADD_DEVICE:
2446 iommu_init_device(dev);
2449 * dev_data is still NULL and
2450 * got initialized in iommu_init_device
2452 dev_data = get_dev_data(dev);
2454 if (iommu_pass_through || dev_data->iommu_v2) {
2455 dev_data->passthrough = true;
2456 attach_device(dev, pt_domain);
2460 domain = domain_for_device(dev);
2462 /* allocate a protection domain if a device is added */
2463 dma_domain = find_protection_domain(devid);
2465 dma_domain = dma_ops_domain_alloc();
2468 dma_domain->target_dev = devid;
2470 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2471 list_add_tail(&dma_domain->list, &iommu_pd_list);
2472 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2475 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2478 case BUS_NOTIFY_DEL_DEVICE:
2480 iommu_uninit_device(dev);
2486 iommu_completion_wait(iommu);
2492 static struct notifier_block device_nb = {
2493 .notifier_call = device_change_notifier,
2496 void amd_iommu_init_notifier(void)
2498 bus_register_notifier(&pci_bus_type, &device_nb);
2501 /*****************************************************************************
2503 * The next functions belong to the dma_ops mapping/unmapping code.
2505 *****************************************************************************/
2508 * In the dma_ops path we only have the struct device. This function
2509 * finds the corresponding IOMMU, the protection domain and the
2510 * requestor id for a given device.
2511 * If the device is not yet associated with a domain this is also done
2514 static struct protection_domain *get_domain(struct device *dev)
2516 struct protection_domain *domain;
2517 struct dma_ops_domain *dma_dom;
2518 u16 devid = get_device_id(dev);
2520 if (!check_device(dev))
2521 return ERR_PTR(-EINVAL);
2523 domain = domain_for_device(dev);
2524 if (domain != NULL && !dma_ops_domain(domain))
2525 return ERR_PTR(-EBUSY);
2530 /* Device not bound yet - bind it */
2531 dma_dom = find_protection_domain(devid);
2533 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2534 attach_device(dev, &dma_dom->domain);
2535 DUMP_printk("Using protection domain %d for device %s\n",
2536 dma_dom->domain.id, dev_name(dev));
2538 return &dma_dom->domain;
2541 static void update_device_table(struct protection_domain *domain)
2543 struct iommu_dev_data *dev_data;
2545 list_for_each_entry(dev_data, &domain->dev_list, list)
2546 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2549 static void update_domain(struct protection_domain *domain)
2551 if (!domain->updated)
2554 update_device_table(domain);
2556 domain_flush_devices(domain);
2557 domain_flush_tlb_pde(domain);
2559 domain->updated = false;
2563 * This function fetches the PTE for a given address in the aperture
2565 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2566 unsigned long address)
2568 struct aperture_range *aperture;
2569 u64 *pte, *pte_page;
2571 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2575 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2577 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2579 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2581 pte += PM_LEVEL_INDEX(0, address);
2583 update_domain(&dom->domain);
2589 * This is the generic map function. It maps one 4kb page at paddr to
2590 * the given address in the DMA address space for the domain.
2592 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2593 unsigned long address,
2599 WARN_ON(address > dom->aperture_size);
2603 pte = dma_ops_get_pte(dom, address);
2605 return DMA_ERROR_CODE;
2607 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2609 if (direction == DMA_TO_DEVICE)
2610 __pte |= IOMMU_PTE_IR;
2611 else if (direction == DMA_FROM_DEVICE)
2612 __pte |= IOMMU_PTE_IW;
2613 else if (direction == DMA_BIDIRECTIONAL)
2614 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2620 return (dma_addr_t)address;
2624 * The generic unmapping function for on page in the DMA address space.
2626 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2627 unsigned long address)
2629 struct aperture_range *aperture;
2632 if (address >= dom->aperture_size)
2635 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2639 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2643 pte += PM_LEVEL_INDEX(0, address);
2651 * This function contains common code for mapping of a physically
2652 * contiguous memory region into DMA address space. It is used by all
2653 * mapping functions provided with this IOMMU driver.
2654 * Must be called with the domain lock held.
2656 static dma_addr_t __map_single(struct device *dev,
2657 struct dma_ops_domain *dma_dom,
2664 dma_addr_t offset = paddr & ~PAGE_MASK;
2665 dma_addr_t address, start, ret;
2667 unsigned long align_mask = 0;
2670 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2673 INC_STATS_COUNTER(total_map_requests);
2676 INC_STATS_COUNTER(cross_page);
2679 align_mask = (1UL << get_order(size)) - 1;
2682 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2684 if (unlikely(address == DMA_ERROR_CODE)) {
2686 * setting next_address here will let the address
2687 * allocator only scan the new allocated range in the
2688 * first run. This is a small optimization.
2690 dma_dom->next_address = dma_dom->aperture_size;
2692 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2696 * aperture was successfully enlarged by 128 MB, try
2703 for (i = 0; i < pages; ++i) {
2704 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2705 if (ret == DMA_ERROR_CODE)
2713 ADD_STATS_COUNTER(alloced_io_mem, size);
2715 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2716 domain_flush_tlb(&dma_dom->domain);
2717 dma_dom->need_flush = false;
2718 } else if (unlikely(amd_iommu_np_cache))
2719 domain_flush_pages(&dma_dom->domain, address, size);
2726 for (--i; i >= 0; --i) {
2728 dma_ops_domain_unmap(dma_dom, start);
2731 dma_ops_free_addresses(dma_dom, address, pages);
2733 return DMA_ERROR_CODE;
2737 * Does the reverse of the __map_single function. Must be called with
2738 * the domain lock held too
2740 static void __unmap_single(struct dma_ops_domain *dma_dom,
2741 dma_addr_t dma_addr,
2745 dma_addr_t flush_addr;
2746 dma_addr_t i, start;
2749 if ((dma_addr == DMA_ERROR_CODE) ||
2750 (dma_addr + size > dma_dom->aperture_size))
2753 flush_addr = dma_addr;
2754 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2755 dma_addr &= PAGE_MASK;
2758 for (i = 0; i < pages; ++i) {
2759 dma_ops_domain_unmap(dma_dom, start);
2763 SUB_STATS_COUNTER(alloced_io_mem, size);
2765 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2767 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2768 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2769 dma_dom->need_flush = false;
2774 * The exported map_single function for dma_ops.
2776 static dma_addr_t map_page(struct device *dev, struct page *page,
2777 unsigned long offset, size_t size,
2778 enum dma_data_direction dir,
2779 struct dma_attrs *attrs)
2781 unsigned long flags;
2782 struct protection_domain *domain;
2785 phys_addr_t paddr = page_to_phys(page) + offset;
2787 INC_STATS_COUNTER(cnt_map_single);
2789 domain = get_domain(dev);
2790 if (PTR_ERR(domain) == -EINVAL)
2791 return (dma_addr_t)paddr;
2792 else if (IS_ERR(domain))
2793 return DMA_ERROR_CODE;
2795 dma_mask = *dev->dma_mask;
2797 spin_lock_irqsave(&domain->lock, flags);
2799 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2801 if (addr == DMA_ERROR_CODE)
2804 domain_flush_complete(domain);
2807 spin_unlock_irqrestore(&domain->lock, flags);
2813 * The exported unmap_single function for dma_ops.
2815 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2816 enum dma_data_direction dir, struct dma_attrs *attrs)
2818 unsigned long flags;
2819 struct protection_domain *domain;
2821 INC_STATS_COUNTER(cnt_unmap_single);
2823 domain = get_domain(dev);
2827 spin_lock_irqsave(&domain->lock, flags);
2829 __unmap_single(domain->priv, dma_addr, size, dir);
2831 domain_flush_complete(domain);
2833 spin_unlock_irqrestore(&domain->lock, flags);
2837 * This is a special map_sg function which is used if we should map a
2838 * device which is not handled by an AMD IOMMU in the system.
2840 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2841 int nelems, int dir)
2843 struct scatterlist *s;
2846 for_each_sg(sglist, s, nelems, i) {
2847 s->dma_address = (dma_addr_t)sg_phys(s);
2848 s->dma_length = s->length;
2855 * The exported map_sg function for dma_ops (handles scatter-gather
2858 static int map_sg(struct device *dev, struct scatterlist *sglist,
2859 int nelems, enum dma_data_direction dir,
2860 struct dma_attrs *attrs)
2862 unsigned long flags;
2863 struct protection_domain *domain;
2865 struct scatterlist *s;
2867 int mapped_elems = 0;
2870 INC_STATS_COUNTER(cnt_map_sg);
2872 domain = get_domain(dev);
2873 if (PTR_ERR(domain) == -EINVAL)
2874 return map_sg_no_iommu(dev, sglist, nelems, dir);
2875 else if (IS_ERR(domain))
2878 dma_mask = *dev->dma_mask;
2880 spin_lock_irqsave(&domain->lock, flags);
2882 for_each_sg(sglist, s, nelems, i) {
2885 s->dma_address = __map_single(dev, domain->priv,
2886 paddr, s->length, dir, false,
2889 if (s->dma_address) {
2890 s->dma_length = s->length;
2896 domain_flush_complete(domain);
2899 spin_unlock_irqrestore(&domain->lock, flags);
2901 return mapped_elems;
2903 for_each_sg(sglist, s, mapped_elems, i) {
2905 __unmap_single(domain->priv, s->dma_address,
2906 s->dma_length, dir);
2907 s->dma_address = s->dma_length = 0;
2916 * The exported map_sg function for dma_ops (handles scatter-gather
2919 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2920 int nelems, enum dma_data_direction dir,
2921 struct dma_attrs *attrs)
2923 unsigned long flags;
2924 struct protection_domain *domain;
2925 struct scatterlist *s;
2928 INC_STATS_COUNTER(cnt_unmap_sg);
2930 domain = get_domain(dev);
2934 spin_lock_irqsave(&domain->lock, flags);
2936 for_each_sg(sglist, s, nelems, i) {
2937 __unmap_single(domain->priv, s->dma_address,
2938 s->dma_length, dir);
2939 s->dma_address = s->dma_length = 0;
2942 domain_flush_complete(domain);
2944 spin_unlock_irqrestore(&domain->lock, flags);
2948 * The exported alloc_coherent function for dma_ops.
2950 static void *alloc_coherent(struct device *dev, size_t size,
2951 dma_addr_t *dma_addr, gfp_t flag,
2952 struct dma_attrs *attrs)
2954 unsigned long flags;
2956 struct protection_domain *domain;
2958 u64 dma_mask = dev->coherent_dma_mask;
2960 INC_STATS_COUNTER(cnt_alloc_coherent);
2962 domain = get_domain(dev);
2963 if (PTR_ERR(domain) == -EINVAL) {
2964 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2965 *dma_addr = __pa(virt_addr);
2967 } else if (IS_ERR(domain))
2970 dma_mask = dev->coherent_dma_mask;
2971 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2974 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2978 paddr = virt_to_phys(virt_addr);
2981 dma_mask = *dev->dma_mask;
2983 spin_lock_irqsave(&domain->lock, flags);
2985 *dma_addr = __map_single(dev, domain->priv, paddr,
2986 size, DMA_BIDIRECTIONAL, true, dma_mask);
2988 if (*dma_addr == DMA_ERROR_CODE) {
2989 spin_unlock_irqrestore(&domain->lock, flags);
2993 domain_flush_complete(domain);
2995 spin_unlock_irqrestore(&domain->lock, flags);
3001 free_pages((unsigned long)virt_addr, get_order(size));
3007 * The exported free_coherent function for dma_ops.
3009 static void free_coherent(struct device *dev, size_t size,
3010 void *virt_addr, dma_addr_t dma_addr,
3011 struct dma_attrs *attrs)
3013 unsigned long flags;
3014 struct protection_domain *domain;
3016 INC_STATS_COUNTER(cnt_free_coherent);
3018 domain = get_domain(dev);
3022 spin_lock_irqsave(&domain->lock, flags);
3024 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
3026 domain_flush_complete(domain);
3028 spin_unlock_irqrestore(&domain->lock, flags);
3031 free_pages((unsigned long)virt_addr, get_order(size));
3035 * This function is called by the DMA layer to find out if we can handle a
3036 * particular device. It is part of the dma_ops.
3038 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3040 return check_device(dev);
3044 * The function for pre-allocating protection domains.
3046 * If the driver core informs the DMA layer if a driver grabs a device
3047 * we don't need to preallocate the protection domains anymore.
3048 * For now we have to.
3050 static void __init prealloc_protection_domains(void)
3052 struct iommu_dev_data *dev_data;
3053 struct dma_ops_domain *dma_dom;
3054 struct pci_dev *dev = NULL;
3057 for_each_pci_dev(dev) {
3059 /* Do we handle this device? */
3060 if (!check_device(&dev->dev))
3063 dev_data = get_dev_data(&dev->dev);
3064 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3065 /* Make sure passthrough domain is allocated */
3066 alloc_passthrough_domain();
3067 dev_data->passthrough = true;
3068 attach_device(&dev->dev, pt_domain);
3069 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3070 dev_name(&dev->dev));
3073 /* Is there already any domain for it? */
3074 if (domain_for_device(&dev->dev))
3077 devid = get_device_id(&dev->dev);
3079 dma_dom = dma_ops_domain_alloc();
3082 init_unity_mappings_for_device(dma_dom, devid);
3083 dma_dom->target_dev = devid;
3085 attach_device(&dev->dev, &dma_dom->domain);
3087 list_add_tail(&dma_dom->list, &iommu_pd_list);
3091 static struct dma_map_ops amd_iommu_dma_ops = {
3092 .alloc = alloc_coherent,
3093 .free = free_coherent,
3094 .map_page = map_page,
3095 .unmap_page = unmap_page,
3097 .unmap_sg = unmap_sg,
3098 .dma_supported = amd_iommu_dma_supported,
3101 static unsigned device_dma_ops_init(void)
3103 struct iommu_dev_data *dev_data;
3104 struct pci_dev *pdev = NULL;
3105 unsigned unhandled = 0;
3107 for_each_pci_dev(pdev) {
3108 if (!check_device(&pdev->dev)) {
3110 iommu_ignore_device(&pdev->dev);
3116 dev_data = get_dev_data(&pdev->dev);
3118 if (!dev_data->passthrough)
3119 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3121 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3128 * The function which clues the AMD IOMMU driver into dma_ops.
3131 void __init amd_iommu_init_api(void)
3133 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3136 int __init amd_iommu_init_dma_ops(void)
3138 struct amd_iommu *iommu;
3142 * first allocate a default protection domain for every IOMMU we
3143 * found in the system. Devices not assigned to any other
3144 * protection domain will be assigned to the default one.
3146 for_each_iommu(iommu) {
3147 iommu->default_dom = dma_ops_domain_alloc();
3148 if (iommu->default_dom == NULL)
3150 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3151 ret = iommu_init_unity_mappings(iommu);
3157 * Pre-allocate the protection domains for each device.
3159 prealloc_protection_domains();
3164 /* Make the driver finally visible to the drivers */
3165 unhandled = device_dma_ops_init();
3166 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3167 /* There are unhandled devices - initialize swiotlb for them */
3171 amd_iommu_stats_init();
3173 if (amd_iommu_unmap_flush)
3174 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3176 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3182 for_each_iommu(iommu) {
3183 dma_ops_domain_free(iommu->default_dom);
3189 /*****************************************************************************
3191 * The following functions belong to the exported interface of AMD IOMMU
3193 * This interface allows access to lower level functions of the IOMMU
3194 * like protection domain handling and assignement of devices to domains
3195 * which is not possible with the dma_ops interface.
3197 *****************************************************************************/
3199 static void cleanup_domain(struct protection_domain *domain)
3201 struct iommu_dev_data *dev_data, *next;
3202 unsigned long flags;
3204 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3206 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3207 __detach_device(dev_data);
3208 atomic_set(&dev_data->bind, 0);
3211 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3214 static void protection_domain_free(struct protection_domain *domain)
3219 del_domain_from_list(domain);
3222 domain_id_free(domain->id);
3227 static struct protection_domain *protection_domain_alloc(void)
3229 struct protection_domain *domain;
3231 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3235 spin_lock_init(&domain->lock);
3236 mutex_init(&domain->api_lock);
3237 domain->id = domain_id_alloc();
3240 INIT_LIST_HEAD(&domain->dev_list);
3242 add_domain_to_list(domain);
3252 static int __init alloc_passthrough_domain(void)
3254 if (pt_domain != NULL)
3257 /* allocate passthrough domain */
3258 pt_domain = protection_domain_alloc();
3262 pt_domain->mode = PAGE_MODE_NONE;
3266 static int amd_iommu_domain_init(struct iommu_domain *dom)
3268 struct protection_domain *domain;
3270 domain = protection_domain_alloc();
3274 domain->mode = PAGE_MODE_3_LEVEL;
3275 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3276 if (!domain->pt_root)
3279 domain->iommu_domain = dom;
3283 dom->geometry.aperture_start = 0;
3284 dom->geometry.aperture_end = ~0ULL;
3285 dom->geometry.force_aperture = true;
3290 protection_domain_free(domain);
3295 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3297 struct protection_domain *domain = dom->priv;
3302 if (domain->dev_cnt > 0)
3303 cleanup_domain(domain);
3305 BUG_ON(domain->dev_cnt != 0);
3307 if (domain->mode != PAGE_MODE_NONE)
3308 free_pagetable(domain);
3310 if (domain->flags & PD_IOMMUV2_MASK)
3311 free_gcr3_table(domain);
3313 protection_domain_free(domain);
3318 static void amd_iommu_detach_device(struct iommu_domain *dom,
3321 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3322 struct amd_iommu *iommu;
3325 if (!check_device(dev))
3328 devid = get_device_id(dev);
3330 if (dev_data->domain != NULL)
3333 iommu = amd_iommu_rlookup_table[devid];
3337 iommu_completion_wait(iommu);
3340 static int amd_iommu_attach_device(struct iommu_domain *dom,
3343 struct protection_domain *domain = dom->priv;
3344 struct iommu_dev_data *dev_data;
3345 struct amd_iommu *iommu;
3348 if (!check_device(dev))
3351 dev_data = dev->archdata.iommu;
3353 iommu = amd_iommu_rlookup_table[dev_data->devid];
3357 if (dev_data->domain)
3360 ret = attach_device(dev, domain);
3362 iommu_completion_wait(iommu);
3367 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3368 phys_addr_t paddr, size_t page_size, int iommu_prot)
3370 struct protection_domain *domain = dom->priv;
3374 if (domain->mode == PAGE_MODE_NONE)
3377 if (iommu_prot & IOMMU_READ)
3378 prot |= IOMMU_PROT_IR;
3379 if (iommu_prot & IOMMU_WRITE)
3380 prot |= IOMMU_PROT_IW;
3382 mutex_lock(&domain->api_lock);
3383 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3384 mutex_unlock(&domain->api_lock);
3389 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3392 struct protection_domain *domain = dom->priv;
3395 if (domain->mode == PAGE_MODE_NONE)
3398 mutex_lock(&domain->api_lock);
3399 unmap_size = iommu_unmap_page(domain, iova, page_size);
3400 mutex_unlock(&domain->api_lock);
3402 domain_flush_tlb_pde(domain);
3407 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3410 struct protection_domain *domain = dom->priv;
3411 unsigned long offset_mask;
3415 if (domain->mode == PAGE_MODE_NONE)
3418 pte = fetch_pte(domain, iova);
3420 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3423 if (PM_PTE_LEVEL(*pte) == 0)
3424 offset_mask = PAGE_SIZE - 1;
3426 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3428 __pte = *pte & PM_ADDR_MASK;
3429 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3434 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3438 case IOMMU_CAP_CACHE_COHERENCY:
3440 case IOMMU_CAP_INTR_REMAP:
3441 return irq_remapping_enabled;
3447 static struct iommu_ops amd_iommu_ops = {
3448 .domain_init = amd_iommu_domain_init,
3449 .domain_destroy = amd_iommu_domain_destroy,
3450 .attach_dev = amd_iommu_attach_device,
3451 .detach_dev = amd_iommu_detach_device,
3452 .map = amd_iommu_map,
3453 .unmap = amd_iommu_unmap,
3454 .iova_to_phys = amd_iommu_iova_to_phys,
3455 .domain_has_cap = amd_iommu_domain_has_cap,
3456 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3459 /*****************************************************************************
3461 * The next functions do a basic initialization of IOMMU for pass through
3464 * In passthrough mode the IOMMU is initialized and enabled but not used for
3465 * DMA-API translation.
3467 *****************************************************************************/
3469 int __init amd_iommu_init_passthrough(void)
3471 struct iommu_dev_data *dev_data;
3472 struct pci_dev *dev = NULL;
3473 struct amd_iommu *iommu;
3477 ret = alloc_passthrough_domain();
3481 for_each_pci_dev(dev) {
3482 if (!check_device(&dev->dev))
3485 dev_data = get_dev_data(&dev->dev);
3486 dev_data->passthrough = true;
3488 devid = get_device_id(&dev->dev);
3490 iommu = amd_iommu_rlookup_table[devid];
3494 attach_device(&dev->dev, pt_domain);
3497 amd_iommu_stats_init();
3499 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3504 /* IOMMUv2 specific functions */
3505 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3507 return atomic_notifier_chain_register(&ppr_notifier, nb);
3509 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3511 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3513 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3515 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3517 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3519 struct protection_domain *domain = dom->priv;
3520 unsigned long flags;
3522 spin_lock_irqsave(&domain->lock, flags);
3524 /* Update data structure */
3525 domain->mode = PAGE_MODE_NONE;
3526 domain->updated = true;
3528 /* Make changes visible to IOMMUs */
3529 update_domain(domain);
3531 /* Page-table is not visible to IOMMU anymore, so free it */
3532 free_pagetable(domain);
3534 spin_unlock_irqrestore(&domain->lock, flags);
3536 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3538 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3540 struct protection_domain *domain = dom->priv;
3541 unsigned long flags;
3544 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3547 /* Number of GCR3 table levels required */
3548 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3551 if (levels > amd_iommu_max_glx_val)
3554 spin_lock_irqsave(&domain->lock, flags);
3557 * Save us all sanity checks whether devices already in the
3558 * domain support IOMMUv2. Just force that the domain has no
3559 * devices attached when it is switched into IOMMUv2 mode.
3562 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3566 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3567 if (domain->gcr3_tbl == NULL)
3570 domain->glx = levels;
3571 domain->flags |= PD_IOMMUV2_MASK;
3572 domain->updated = true;
3574 update_domain(domain);
3579 spin_unlock_irqrestore(&domain->lock, flags);
3583 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3585 static int __flush_pasid(struct protection_domain *domain, int pasid,
3586 u64 address, bool size)
3588 struct iommu_dev_data *dev_data;
3589 struct iommu_cmd cmd;
3592 if (!(domain->flags & PD_IOMMUV2_MASK))
3595 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3598 * IOMMU TLB needs to be flushed before Device TLB to
3599 * prevent device TLB refill from IOMMU TLB
3601 for (i = 0; i < amd_iommus_present; ++i) {
3602 if (domain->dev_iommu[i] == 0)
3605 ret = iommu_queue_command(amd_iommus[i], &cmd);
3610 /* Wait until IOMMU TLB flushes are complete */
3611 domain_flush_complete(domain);
3613 /* Now flush device TLBs */
3614 list_for_each_entry(dev_data, &domain->dev_list, list) {
3615 struct amd_iommu *iommu;
3618 BUG_ON(!dev_data->ats.enabled);
3620 qdep = dev_data->ats.qdep;
3621 iommu = amd_iommu_rlookup_table[dev_data->devid];
3623 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3624 qdep, address, size);
3626 ret = iommu_queue_command(iommu, &cmd);
3631 /* Wait until all device TLBs are flushed */
3632 domain_flush_complete(domain);
3641 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3644 INC_STATS_COUNTER(invalidate_iotlb);
3646 return __flush_pasid(domain, pasid, address, false);
3649 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3652 struct protection_domain *domain = dom->priv;
3653 unsigned long flags;
3656 spin_lock_irqsave(&domain->lock, flags);
3657 ret = __amd_iommu_flush_page(domain, pasid, address);
3658 spin_unlock_irqrestore(&domain->lock, flags);
3662 EXPORT_SYMBOL(amd_iommu_flush_page);
3664 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3666 INC_STATS_COUNTER(invalidate_iotlb_all);
3668 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3672 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3674 struct protection_domain *domain = dom->priv;
3675 unsigned long flags;
3678 spin_lock_irqsave(&domain->lock, flags);
3679 ret = __amd_iommu_flush_tlb(domain, pasid);
3680 spin_unlock_irqrestore(&domain->lock, flags);
3684 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3686 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3693 index = (pasid >> (9 * level)) & 0x1ff;
3699 if (!(*pte & GCR3_VALID)) {
3703 root = (void *)get_zeroed_page(GFP_ATOMIC);
3707 *pte = __pa(root) | GCR3_VALID;
3710 root = __va(*pte & PAGE_MASK);
3718 static int __set_gcr3(struct protection_domain *domain, int pasid,
3723 if (domain->mode != PAGE_MODE_NONE)
3726 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3730 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3732 return __amd_iommu_flush_tlb(domain, pasid);
3735 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3739 if (domain->mode != PAGE_MODE_NONE)
3742 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3748 return __amd_iommu_flush_tlb(domain, pasid);
3751 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3754 struct protection_domain *domain = dom->priv;
3755 unsigned long flags;
3758 spin_lock_irqsave(&domain->lock, flags);
3759 ret = __set_gcr3(domain, pasid, cr3);
3760 spin_unlock_irqrestore(&domain->lock, flags);
3764 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3766 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3768 struct protection_domain *domain = dom->priv;
3769 unsigned long flags;
3772 spin_lock_irqsave(&domain->lock, flags);
3773 ret = __clear_gcr3(domain, pasid);
3774 spin_unlock_irqrestore(&domain->lock, flags);
3778 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3780 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3781 int status, int tag)
3783 struct iommu_dev_data *dev_data;
3784 struct amd_iommu *iommu;
3785 struct iommu_cmd cmd;
3787 INC_STATS_COUNTER(complete_ppr);
3789 dev_data = get_dev_data(&pdev->dev);
3790 iommu = amd_iommu_rlookup_table[dev_data->devid];
3792 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3793 tag, dev_data->pri_tlp);
3795 return iommu_queue_command(iommu, &cmd);
3797 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3799 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3801 struct protection_domain *domain;
3803 domain = get_domain(&pdev->dev);
3807 /* Only return IOMMUv2 domains */
3808 if (!(domain->flags & PD_IOMMUV2_MASK))
3811 return domain->iommu_domain;
3813 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3815 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3817 struct iommu_dev_data *dev_data;
3819 if (!amd_iommu_v2_supported())
3822 dev_data = get_dev_data(&pdev->dev);
3823 dev_data->errata |= (1 << erratum);
3825 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3827 int amd_iommu_device_info(struct pci_dev *pdev,
3828 struct amd_iommu_device_info *info)
3833 if (pdev == NULL || info == NULL)
3836 if (!amd_iommu_v2_supported())
3839 memset(info, 0, sizeof(*info));
3841 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3843 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3845 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3847 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3849 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3853 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3854 max_pasids = min(max_pasids, (1 << 20));
3856 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3857 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3859 features = pci_pasid_features(pdev);
3860 if (features & PCI_PASID_CAP_EXEC)
3861 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3862 if (features & PCI_PASID_CAP_PRIV)
3863 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3868 EXPORT_SYMBOL(amd_iommu_device_info);
3870 #ifdef CONFIG_IRQ_REMAP
3872 /*****************************************************************************
3874 * Interrupt Remapping Implementation
3876 *****************************************************************************/
3893 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3894 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3895 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3896 #define DTE_IRQ_REMAP_ENABLE 1ULL
3898 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3902 dte = amd_iommu_dev_table[devid].data[2];
3903 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3904 dte |= virt_to_phys(table->table);
3905 dte |= DTE_IRQ_REMAP_INTCTL;
3906 dte |= DTE_IRQ_TABLE_LEN;
3907 dte |= DTE_IRQ_REMAP_ENABLE;
3909 amd_iommu_dev_table[devid].data[2] = dte;
3912 #define IRTE_ALLOCATED (~1U)
3914 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3916 struct irq_remap_table *table = NULL;
3917 struct amd_iommu *iommu;
3918 unsigned long flags;
3921 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3923 iommu = amd_iommu_rlookup_table[devid];
3927 table = irq_lookup_table[devid];
3931 alias = amd_iommu_alias_table[devid];
3932 table = irq_lookup_table[alias];
3934 irq_lookup_table[devid] = table;
3935 set_dte_irq_entry(devid, table);
3936 iommu_flush_dte(iommu, devid);
3940 /* Nothing there yet, allocate new irq remapping table */
3941 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3946 /* Keep the first 32 indexes free for IOAPIC interrupts */
3947 table->min_index = 32;
3949 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3950 if (!table->table) {
3956 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3961 for (i = 0; i < 32; ++i)
3962 table->table[i] = IRTE_ALLOCATED;
3965 irq_lookup_table[devid] = table;
3966 set_dte_irq_entry(devid, table);
3967 iommu_flush_dte(iommu, devid);
3968 if (devid != alias) {
3969 irq_lookup_table[alias] = table;
3970 set_dte_irq_entry(devid, table);
3971 iommu_flush_dte(iommu, alias);
3975 iommu_completion_wait(iommu);
3978 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3983 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
3985 struct irq_remap_table *table;
3986 unsigned long flags;
3989 table = get_irq_table(devid, false);
3993 spin_lock_irqsave(&table->lock, flags);
3995 /* Scan table for free entries */
3996 for (c = 0, index = table->min_index;
3997 index < MAX_IRQS_PER_TABLE;
3999 if (table->table[index] == 0)
4005 struct irq_2_iommu *irte_info;
4008 table->table[index - c + 1] = IRTE_ALLOCATED;
4013 irte_info = &cfg->irq_2_iommu;
4014 irte_info->sub_handle = devid;
4015 irte_info->irte_index = index;
4024 spin_unlock_irqrestore(&table->lock, flags);
4029 static int get_irte(u16 devid, int index, union irte *irte)
4031 struct irq_remap_table *table;
4032 unsigned long flags;
4034 table = get_irq_table(devid, false);
4038 spin_lock_irqsave(&table->lock, flags);
4039 irte->val = table->table[index];
4040 spin_unlock_irqrestore(&table->lock, flags);
4045 static int modify_irte(u16 devid, int index, union irte irte)
4047 struct irq_remap_table *table;
4048 struct amd_iommu *iommu;
4049 unsigned long flags;
4051 iommu = amd_iommu_rlookup_table[devid];
4055 table = get_irq_table(devid, false);
4059 spin_lock_irqsave(&table->lock, flags);
4060 table->table[index] = irte.val;
4061 spin_unlock_irqrestore(&table->lock, flags);
4063 iommu_flush_irt(iommu, devid);
4064 iommu_completion_wait(iommu);
4069 static void free_irte(u16 devid, int index)
4071 struct irq_remap_table *table;
4072 struct amd_iommu *iommu;
4073 unsigned long flags;
4075 iommu = amd_iommu_rlookup_table[devid];
4079 table = get_irq_table(devid, false);
4083 spin_lock_irqsave(&table->lock, flags);
4084 table->table[index] = 0;
4085 spin_unlock_irqrestore(&table->lock, flags);
4087 iommu_flush_irt(iommu, devid);
4088 iommu_completion_wait(iommu);
4091 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4092 unsigned int destination, int vector,
4093 struct io_apic_irq_attr *attr)
4095 struct irq_remap_table *table;
4096 struct irq_2_iommu *irte_info;
4097 struct irq_cfg *cfg;
4104 cfg = irq_get_chip_data(irq);
4108 irte_info = &cfg->irq_2_iommu;
4109 ioapic_id = mpc_ioapic_id(attr->ioapic);
4110 devid = get_ioapic_devid(ioapic_id);
4115 table = get_irq_table(devid, true);
4119 index = attr->ioapic_pin;
4121 /* Setup IRQ remapping info */
4123 irte_info->sub_handle = devid;
4124 irte_info->irte_index = index;
4126 /* Setup IRTE for IOMMU */
4128 irte.fields.vector = vector;
4129 irte.fields.int_type = apic->irq_delivery_mode;
4130 irte.fields.destination = destination;
4131 irte.fields.dm = apic->irq_dest_mode;
4132 irte.fields.valid = 1;
4134 ret = modify_irte(devid, index, irte);
4138 /* Setup IOAPIC entry */
4139 memset(entry, 0, sizeof(*entry));
4141 entry->vector = index;
4143 entry->trigger = attr->trigger;
4144 entry->polarity = attr->polarity;
4147 * Mask level triggered irqs.
4155 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4158 struct irq_2_iommu *irte_info;
4159 unsigned int dest, irq;
4160 struct irq_cfg *cfg;
4164 if (!config_enabled(CONFIG_SMP))
4167 cfg = data->chip_data;
4169 irte_info = &cfg->irq_2_iommu;
4171 if (!cpumask_intersects(mask, cpu_online_mask))
4174 if (get_irte(irte_info->sub_handle, irte_info->irte_index, &irte))
4177 if (assign_irq_vector(irq, cfg, mask))
4180 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4182 if (assign_irq_vector(irq, cfg, data->affinity))
4183 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4187 irte.fields.vector = cfg->vector;
4188 irte.fields.destination = dest;
4190 modify_irte(irte_info->sub_handle, irte_info->irte_index, irte);
4192 if (cfg->move_in_progress)
4193 send_cleanup_vector(cfg);
4195 cpumask_copy(data->affinity, mask);
4200 static int free_irq(int irq)
4202 struct irq_2_iommu *irte_info;
4203 struct irq_cfg *cfg;
4205 cfg = irq_get_chip_data(irq);
4209 irte_info = &cfg->irq_2_iommu;
4211 free_irte(irte_info->sub_handle, irte_info->irte_index);
4216 static void compose_msi_msg(struct pci_dev *pdev,
4217 unsigned int irq, unsigned int dest,
4218 struct msi_msg *msg, u8 hpet_id)
4220 struct irq_2_iommu *irte_info;
4221 struct irq_cfg *cfg;
4224 cfg = irq_get_chip_data(irq);
4228 irte_info = &cfg->irq_2_iommu;
4231 irte.fields.vector = cfg->vector;
4232 irte.fields.int_type = apic->irq_delivery_mode;
4233 irte.fields.destination = dest;
4234 irte.fields.dm = apic->irq_dest_mode;
4235 irte.fields.valid = 1;
4237 modify_irte(irte_info->sub_handle, irte_info->irte_index, irte);
4239 msg->address_hi = MSI_ADDR_BASE_HI;
4240 msg->address_lo = MSI_ADDR_BASE_LO;
4241 msg->data = irte_info->irte_index;
4244 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4246 struct irq_cfg *cfg;
4253 cfg = irq_get_chip_data(irq);
4257 devid = get_device_id(&pdev->dev);
4258 index = alloc_irq_index(cfg, devid, nvec);
4260 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4263 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4264 int index, int offset)
4266 struct irq_2_iommu *irte_info;
4267 struct irq_cfg *cfg;
4273 cfg = irq_get_chip_data(irq);
4277 if (index >= MAX_IRQS_PER_TABLE)
4280 devid = get_device_id(&pdev->dev);
4281 irte_info = &cfg->irq_2_iommu;
4284 irte_info->sub_handle = devid;
4285 irte_info->irte_index = index + offset;
4290 static int setup_hpet_msi(unsigned int irq, unsigned int id)
4292 struct irq_2_iommu *irte_info;
4293 struct irq_cfg *cfg;
4296 cfg = irq_get_chip_data(irq);
4300 irte_info = &cfg->irq_2_iommu;
4301 devid = get_hpet_devid(id);
4305 index = alloc_irq_index(cfg, devid, 1);
4310 irte_info->sub_handle = devid;
4311 irte_info->irte_index = index;
4316 struct irq_remap_ops amd_iommu_irq_ops = {
4317 .supported = amd_iommu_supported,
4318 .prepare = amd_iommu_prepare,
4319 .enable = amd_iommu_enable,
4320 .disable = amd_iommu_disable,
4321 .reenable = amd_iommu_reenable,
4322 .enable_faulting = amd_iommu_enable_faulting,
4323 .setup_ioapic_entry = setup_ioapic_entry,
4324 .set_affinity = set_affinity,
4325 .free_irq = free_irq,
4326 .compose_msi_msg = compose_msi_msg,
4327 .msi_alloc_irq = msi_alloc_irq,
4328 .msi_setup_irq = msi_setup_irq,
4329 .setup_hpet_msi = setup_hpet_msi,