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 PCI_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 lowest numbered
291 * function that also does not suport the required ACS flags.
293 if (dma_pdev->multifunction &&
294 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)) {
295 u8 i, slot = PCI_SLOT(dma_pdev->devfn);
297 for (i = 0; i < 8; i++) {
300 tmp = pci_get_slot(dma_pdev->bus, PCI_DEVFN(slot, i));
304 if (!pci_acs_enabled(tmp, REQ_ACS_FLAGS)) {
305 swap_pci_ref(&dma_pdev, tmp);
313 * Devices on the root bus go through the iommu. If that's not us,
314 * find the next upstream device and test ACS up to the root bus.
315 * Finding the next device may require skipping virtual buses.
317 while (!pci_is_root_bus(dma_pdev->bus)) {
318 struct pci_bus *bus = find_hosted_bus(dma_pdev->bus);
322 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
325 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
331 static int use_pdev_iommu_group(struct pci_dev *pdev, struct device *dev)
333 struct iommu_group *group = iommu_group_get(&pdev->dev);
337 group = iommu_group_alloc();
339 return PTR_ERR(group);
341 WARN_ON(&pdev->dev != dev);
344 ret = iommu_group_add_device(group, dev);
345 iommu_group_put(group);
349 static int use_dev_data_iommu_group(struct iommu_dev_data *dev_data,
352 if (!dev_data->group) {
353 struct iommu_group *group = iommu_group_alloc();
355 return PTR_ERR(group);
357 dev_data->group = group;
360 return iommu_group_add_device(dev_data->group, dev);
363 static int init_iommu_group(struct device *dev)
365 struct iommu_dev_data *dev_data;
366 struct iommu_group *group;
367 struct pci_dev *dma_pdev;
370 group = iommu_group_get(dev);
372 iommu_group_put(group);
376 dev_data = find_dev_data(get_device_id(dev));
380 if (dev_data->alias_data) {
384 if (dev_data->alias_data->group)
388 * If the alias device exists, it's effectively just a first
389 * level quirk for finding the DMA source.
391 alias = amd_iommu_alias_table[dev_data->devid];
392 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
394 dma_pdev = get_isolation_root(dma_pdev);
399 * If the alias is virtual, try to find a parent device
400 * and test whether the IOMMU group is actualy rooted above
401 * the alias. Be careful to also test the parent device if
402 * we think the alias is the root of the group.
404 bus = pci_find_bus(0, alias >> 8);
408 bus = find_hosted_bus(bus);
409 if (IS_ERR(bus) || !bus->self)
412 dma_pdev = get_isolation_root(pci_dev_get(bus->self));
413 if (dma_pdev != bus->self || (dma_pdev->multifunction &&
414 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)))
417 pci_dev_put(dma_pdev);
421 dma_pdev = get_isolation_root(pci_dev_get(to_pci_dev(dev)));
423 ret = use_pdev_iommu_group(dma_pdev, dev);
424 pci_dev_put(dma_pdev);
427 return use_dev_data_iommu_group(dev_data->alias_data, dev);
430 static int iommu_init_device(struct device *dev)
432 struct pci_dev *pdev = to_pci_dev(dev);
433 struct iommu_dev_data *dev_data;
437 if (dev->archdata.iommu)
440 dev_data = find_dev_data(get_device_id(dev));
444 alias = amd_iommu_alias_table[dev_data->devid];
445 if (alias != dev_data->devid) {
446 struct iommu_dev_data *alias_data;
448 alias_data = find_dev_data(alias);
449 if (alias_data == NULL) {
450 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
452 free_dev_data(dev_data);
455 dev_data->alias_data = alias_data;
458 ret = init_iommu_group(dev);
460 free_dev_data(dev_data);
464 if (pci_iommuv2_capable(pdev)) {
465 struct amd_iommu *iommu;
467 iommu = amd_iommu_rlookup_table[dev_data->devid];
468 dev_data->iommu_v2 = iommu->is_iommu_v2;
471 dev->archdata.iommu = dev_data;
476 static void iommu_ignore_device(struct device *dev)
480 devid = get_device_id(dev);
481 alias = amd_iommu_alias_table[devid];
483 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
484 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
486 amd_iommu_rlookup_table[devid] = NULL;
487 amd_iommu_rlookup_table[alias] = NULL;
490 static void iommu_uninit_device(struct device *dev)
492 iommu_group_remove_device(dev);
495 * Nothing to do here - we keep dev_data around for unplugged devices
496 * and reuse it when the device is re-plugged - not doing so would
497 * introduce a ton of races.
501 void __init amd_iommu_uninit_devices(void)
503 struct iommu_dev_data *dev_data, *n;
504 struct pci_dev *pdev = NULL;
506 for_each_pci_dev(pdev) {
508 if (!check_device(&pdev->dev))
511 iommu_uninit_device(&pdev->dev);
514 /* Free all of our dev_data structures */
515 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
516 free_dev_data(dev_data);
519 int __init amd_iommu_init_devices(void)
521 struct pci_dev *pdev = NULL;
524 for_each_pci_dev(pdev) {
526 if (!check_device(&pdev->dev))
529 ret = iommu_init_device(&pdev->dev);
530 if (ret == -ENOTSUPP)
531 iommu_ignore_device(&pdev->dev);
540 amd_iommu_uninit_devices();
544 #ifdef CONFIG_AMD_IOMMU_STATS
547 * Initialization code for statistics collection
550 DECLARE_STATS_COUNTER(compl_wait);
551 DECLARE_STATS_COUNTER(cnt_map_single);
552 DECLARE_STATS_COUNTER(cnt_unmap_single);
553 DECLARE_STATS_COUNTER(cnt_map_sg);
554 DECLARE_STATS_COUNTER(cnt_unmap_sg);
555 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
556 DECLARE_STATS_COUNTER(cnt_free_coherent);
557 DECLARE_STATS_COUNTER(cross_page);
558 DECLARE_STATS_COUNTER(domain_flush_single);
559 DECLARE_STATS_COUNTER(domain_flush_all);
560 DECLARE_STATS_COUNTER(alloced_io_mem);
561 DECLARE_STATS_COUNTER(total_map_requests);
562 DECLARE_STATS_COUNTER(complete_ppr);
563 DECLARE_STATS_COUNTER(invalidate_iotlb);
564 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
565 DECLARE_STATS_COUNTER(pri_requests);
567 static struct dentry *stats_dir;
568 static struct dentry *de_fflush;
570 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
572 if (stats_dir == NULL)
575 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
579 static void amd_iommu_stats_init(void)
581 stats_dir = debugfs_create_dir("amd-iommu", NULL);
582 if (stats_dir == NULL)
585 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
586 &amd_iommu_unmap_flush);
588 amd_iommu_stats_add(&compl_wait);
589 amd_iommu_stats_add(&cnt_map_single);
590 amd_iommu_stats_add(&cnt_unmap_single);
591 amd_iommu_stats_add(&cnt_map_sg);
592 amd_iommu_stats_add(&cnt_unmap_sg);
593 amd_iommu_stats_add(&cnt_alloc_coherent);
594 amd_iommu_stats_add(&cnt_free_coherent);
595 amd_iommu_stats_add(&cross_page);
596 amd_iommu_stats_add(&domain_flush_single);
597 amd_iommu_stats_add(&domain_flush_all);
598 amd_iommu_stats_add(&alloced_io_mem);
599 amd_iommu_stats_add(&total_map_requests);
600 amd_iommu_stats_add(&complete_ppr);
601 amd_iommu_stats_add(&invalidate_iotlb);
602 amd_iommu_stats_add(&invalidate_iotlb_all);
603 amd_iommu_stats_add(&pri_requests);
608 /****************************************************************************
610 * Interrupt handling functions
612 ****************************************************************************/
614 static void dump_dte_entry(u16 devid)
618 for (i = 0; i < 4; ++i)
619 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
620 amd_iommu_dev_table[devid].data[i]);
623 static void dump_command(unsigned long phys_addr)
625 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
628 for (i = 0; i < 4; ++i)
629 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
632 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
634 int type, devid, domid, flags;
635 volatile u32 *event = __evt;
640 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
641 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
642 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
643 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
644 address = (u64)(((u64)event[3]) << 32) | event[2];
647 /* Did we hit the erratum? */
648 if (++count == LOOP_TIMEOUT) {
649 pr_err("AMD-Vi: No event written to event log\n");
656 printk(KERN_ERR "AMD-Vi: Event logged [");
659 case EVENT_TYPE_ILL_DEV:
660 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
661 "address=0x%016llx flags=0x%04x]\n",
662 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
664 dump_dte_entry(devid);
666 case EVENT_TYPE_IO_FAULT:
667 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
668 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
669 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
670 domid, address, flags);
672 case EVENT_TYPE_DEV_TAB_ERR:
673 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
674 "address=0x%016llx flags=0x%04x]\n",
675 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
678 case EVENT_TYPE_PAGE_TAB_ERR:
679 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
680 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
681 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
682 domid, address, flags);
684 case EVENT_TYPE_ILL_CMD:
685 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
686 dump_command(address);
688 case EVENT_TYPE_CMD_HARD_ERR:
689 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
690 "flags=0x%04x]\n", address, flags);
692 case EVENT_TYPE_IOTLB_INV_TO:
693 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
694 "address=0x%016llx]\n",
695 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
698 case EVENT_TYPE_INV_DEV_REQ:
699 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
700 "address=0x%016llx flags=0x%04x]\n",
701 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
705 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
708 memset(__evt, 0, 4 * sizeof(u32));
711 static void iommu_poll_events(struct amd_iommu *iommu)
715 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
716 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
718 while (head != tail) {
719 iommu_print_event(iommu, iommu->evt_buf + head);
720 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
723 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
726 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
728 struct amd_iommu_fault fault;
730 INC_STATS_COUNTER(pri_requests);
732 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
733 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
737 fault.address = raw[1];
738 fault.pasid = PPR_PASID(raw[0]);
739 fault.device_id = PPR_DEVID(raw[0]);
740 fault.tag = PPR_TAG(raw[0]);
741 fault.flags = PPR_FLAGS(raw[0]);
743 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
746 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
750 if (iommu->ppr_log == NULL)
753 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
754 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
756 while (head != tail) {
761 raw = (u64 *)(iommu->ppr_log + head);
764 * Hardware bug: Interrupt may arrive before the entry is
765 * written to memory. If this happens we need to wait for the
768 for (i = 0; i < LOOP_TIMEOUT; ++i) {
769 if (PPR_REQ_TYPE(raw[0]) != 0)
774 /* Avoid memcpy function-call overhead */
779 * To detect the hardware bug we need to clear the entry
782 raw[0] = raw[1] = 0UL;
784 /* Update head pointer of hardware ring-buffer */
785 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
786 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
788 /* Handle PPR entry */
789 iommu_handle_ppr_entry(iommu, entry);
791 /* Refresh ring-buffer information */
792 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
793 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
797 irqreturn_t amd_iommu_int_thread(int irq, void *data)
799 struct amd_iommu *iommu = (struct amd_iommu *) data;
800 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
802 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
803 /* Enable EVT and PPR interrupts again */
804 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
805 iommu->mmio_base + MMIO_STATUS_OFFSET);
807 if (status & MMIO_STATUS_EVT_INT_MASK) {
808 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
809 iommu_poll_events(iommu);
812 if (status & MMIO_STATUS_PPR_INT_MASK) {
813 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
814 iommu_poll_ppr_log(iommu);
818 * Hardware bug: ERBT1312
819 * When re-enabling interrupt (by writing 1
820 * to clear the bit), the hardware might also try to set
821 * the interrupt bit in the event status register.
822 * In this scenario, the bit will be set, and disable
823 * subsequent interrupts.
825 * Workaround: The IOMMU driver should read back the
826 * status register and check if the interrupt bits are cleared.
827 * If not, driver will need to go through the interrupt handler
828 * again and re-clear the bits
830 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
835 irqreturn_t amd_iommu_int_handler(int irq, void *data)
837 return IRQ_WAKE_THREAD;
840 /****************************************************************************
842 * IOMMU command queuing functions
844 ****************************************************************************/
846 static int wait_on_sem(volatile u64 *sem)
850 while (*sem == 0 && i < LOOP_TIMEOUT) {
855 if (i == LOOP_TIMEOUT) {
856 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
863 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
864 struct iommu_cmd *cmd,
869 target = iommu->cmd_buf + tail;
870 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
872 /* Copy command to buffer */
873 memcpy(target, cmd, sizeof(*cmd));
875 /* Tell the IOMMU about it */
876 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
879 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
881 WARN_ON(address & 0x7ULL);
883 memset(cmd, 0, sizeof(*cmd));
884 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
885 cmd->data[1] = upper_32_bits(__pa(address));
887 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
890 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
892 memset(cmd, 0, sizeof(*cmd));
893 cmd->data[0] = devid;
894 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
897 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
898 size_t size, u16 domid, int pde)
903 pages = iommu_num_pages(address, size, PAGE_SIZE);
908 * If we have to flush more than one page, flush all
909 * TLB entries for this domain
911 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
915 address &= PAGE_MASK;
917 memset(cmd, 0, sizeof(*cmd));
918 cmd->data[1] |= domid;
919 cmd->data[2] = lower_32_bits(address);
920 cmd->data[3] = upper_32_bits(address);
921 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
922 if (s) /* size bit - we flush more than one 4kb page */
923 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
924 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
925 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
928 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
929 u64 address, size_t size)
934 pages = iommu_num_pages(address, size, PAGE_SIZE);
939 * If we have to flush more than one page, flush all
940 * TLB entries for this domain
942 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
946 address &= PAGE_MASK;
948 memset(cmd, 0, sizeof(*cmd));
949 cmd->data[0] = devid;
950 cmd->data[0] |= (qdep & 0xff) << 24;
951 cmd->data[1] = devid;
952 cmd->data[2] = lower_32_bits(address);
953 cmd->data[3] = upper_32_bits(address);
954 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
959 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
960 u64 address, bool size)
962 memset(cmd, 0, sizeof(*cmd));
964 address &= ~(0xfffULL);
966 cmd->data[0] = pasid & PASID_MASK;
967 cmd->data[1] = domid;
968 cmd->data[2] = lower_32_bits(address);
969 cmd->data[3] = upper_32_bits(address);
970 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
971 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
974 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
977 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
978 int qdep, u64 address, bool size)
980 memset(cmd, 0, sizeof(*cmd));
982 address &= ~(0xfffULL);
984 cmd->data[0] = devid;
985 cmd->data[0] |= (pasid & 0xff) << 16;
986 cmd->data[0] |= (qdep & 0xff) << 24;
987 cmd->data[1] = devid;
988 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
989 cmd->data[2] = lower_32_bits(address);
990 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
991 cmd->data[3] = upper_32_bits(address);
993 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
994 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
997 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
998 int status, int tag, bool gn)
1000 memset(cmd, 0, sizeof(*cmd));
1002 cmd->data[0] = devid;
1004 cmd->data[1] = pasid & PASID_MASK;
1005 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1007 cmd->data[3] = tag & 0x1ff;
1008 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1010 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1013 static void build_inv_all(struct iommu_cmd *cmd)
1015 memset(cmd, 0, sizeof(*cmd));
1016 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1019 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1021 memset(cmd, 0, sizeof(*cmd));
1022 cmd->data[0] = devid;
1023 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1027 * Writes the command to the IOMMUs command buffer and informs the
1028 * hardware about the new command.
1030 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1031 struct iommu_cmd *cmd,
1034 u32 left, tail, head, next_tail;
1035 unsigned long flags;
1037 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
1040 spin_lock_irqsave(&iommu->lock, flags);
1042 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1043 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1044 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
1045 left = (head - next_tail) % iommu->cmd_buf_size;
1048 struct iommu_cmd sync_cmd;
1049 volatile u64 sem = 0;
1052 build_completion_wait(&sync_cmd, (u64)&sem);
1053 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1055 spin_unlock_irqrestore(&iommu->lock, flags);
1057 if ((ret = wait_on_sem(&sem)) != 0)
1063 copy_cmd_to_buffer(iommu, cmd, tail);
1065 /* We need to sync now to make sure all commands are processed */
1066 iommu->need_sync = sync;
1068 spin_unlock_irqrestore(&iommu->lock, flags);
1073 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1075 return iommu_queue_command_sync(iommu, cmd, true);
1079 * This function queues a completion wait command into the command
1080 * buffer of an IOMMU
1082 static int iommu_completion_wait(struct amd_iommu *iommu)
1084 struct iommu_cmd cmd;
1085 volatile u64 sem = 0;
1088 if (!iommu->need_sync)
1091 build_completion_wait(&cmd, (u64)&sem);
1093 ret = iommu_queue_command_sync(iommu, &cmd, false);
1097 return wait_on_sem(&sem);
1100 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1102 struct iommu_cmd cmd;
1104 build_inv_dte(&cmd, devid);
1106 return iommu_queue_command(iommu, &cmd);
1109 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1113 for (devid = 0; devid <= 0xffff; ++devid)
1114 iommu_flush_dte(iommu, devid);
1116 iommu_completion_wait(iommu);
1120 * This function uses heavy locking and may disable irqs for some time. But
1121 * this is no issue because it is only called during resume.
1123 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1127 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1128 struct iommu_cmd cmd;
1129 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1131 iommu_queue_command(iommu, &cmd);
1134 iommu_completion_wait(iommu);
1137 static void iommu_flush_all(struct amd_iommu *iommu)
1139 struct iommu_cmd cmd;
1141 build_inv_all(&cmd);
1143 iommu_queue_command(iommu, &cmd);
1144 iommu_completion_wait(iommu);
1147 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1149 struct iommu_cmd cmd;
1151 build_inv_irt(&cmd, devid);
1153 iommu_queue_command(iommu, &cmd);
1156 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1160 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1161 iommu_flush_irt(iommu, devid);
1163 iommu_completion_wait(iommu);
1166 void iommu_flush_all_caches(struct amd_iommu *iommu)
1168 if (iommu_feature(iommu, FEATURE_IA)) {
1169 iommu_flush_all(iommu);
1171 iommu_flush_dte_all(iommu);
1172 iommu_flush_irt_all(iommu);
1173 iommu_flush_tlb_all(iommu);
1178 * Command send function for flushing on-device TLB
1180 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1181 u64 address, size_t size)
1183 struct amd_iommu *iommu;
1184 struct iommu_cmd cmd;
1187 qdep = dev_data->ats.qdep;
1188 iommu = amd_iommu_rlookup_table[dev_data->devid];
1190 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1192 return iommu_queue_command(iommu, &cmd);
1196 * Command send function for invalidating a device table entry
1198 static int device_flush_dte(struct iommu_dev_data *dev_data)
1200 struct amd_iommu *iommu;
1203 iommu = amd_iommu_rlookup_table[dev_data->devid];
1205 ret = iommu_flush_dte(iommu, dev_data->devid);
1209 if (dev_data->ats.enabled)
1210 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1216 * TLB invalidation function which is called from the mapping functions.
1217 * It invalidates a single PTE if the range to flush is within a single
1218 * page. Otherwise it flushes the whole TLB of the IOMMU.
1220 static void __domain_flush_pages(struct protection_domain *domain,
1221 u64 address, size_t size, int pde)
1223 struct iommu_dev_data *dev_data;
1224 struct iommu_cmd cmd;
1227 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1229 for (i = 0; i < amd_iommus_present; ++i) {
1230 if (!domain->dev_iommu[i])
1234 * Devices of this domain are behind this IOMMU
1235 * We need a TLB flush
1237 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1240 list_for_each_entry(dev_data, &domain->dev_list, list) {
1242 if (!dev_data->ats.enabled)
1245 ret |= device_flush_iotlb(dev_data, address, size);
1251 static void domain_flush_pages(struct protection_domain *domain,
1252 u64 address, size_t size)
1254 __domain_flush_pages(domain, address, size, 0);
1257 /* Flush the whole IO/TLB for a given protection domain */
1258 static void domain_flush_tlb(struct protection_domain *domain)
1260 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1263 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1264 static void domain_flush_tlb_pde(struct protection_domain *domain)
1266 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1269 static void domain_flush_complete(struct protection_domain *domain)
1273 for (i = 0; i < amd_iommus_present; ++i) {
1274 if (!domain->dev_iommu[i])
1278 * Devices of this domain are behind this IOMMU
1279 * We need to wait for completion of all commands.
1281 iommu_completion_wait(amd_iommus[i]);
1287 * This function flushes the DTEs for all devices in domain
1289 static void domain_flush_devices(struct protection_domain *domain)
1291 struct iommu_dev_data *dev_data;
1293 list_for_each_entry(dev_data, &domain->dev_list, list)
1294 device_flush_dte(dev_data);
1297 /****************************************************************************
1299 * The functions below are used the create the page table mappings for
1300 * unity mapped regions.
1302 ****************************************************************************/
1305 * This function is used to add another level to an IO page table. Adding
1306 * another level increases the size of the address space by 9 bits to a size up
1309 static bool increase_address_space(struct protection_domain *domain,
1314 if (domain->mode == PAGE_MODE_6_LEVEL)
1315 /* address space already 64 bit large */
1318 pte = (void *)get_zeroed_page(gfp);
1322 *pte = PM_LEVEL_PDE(domain->mode,
1323 virt_to_phys(domain->pt_root));
1324 domain->pt_root = pte;
1326 domain->updated = true;
1331 static u64 *alloc_pte(struct protection_domain *domain,
1332 unsigned long address,
1333 unsigned long page_size,
1340 BUG_ON(!is_power_of_2(page_size));
1342 while (address > PM_LEVEL_SIZE(domain->mode))
1343 increase_address_space(domain, gfp);
1345 level = domain->mode - 1;
1346 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1347 address = PAGE_SIZE_ALIGN(address, page_size);
1348 end_lvl = PAGE_SIZE_LEVEL(page_size);
1350 while (level > end_lvl) {
1351 if (!IOMMU_PTE_PRESENT(*pte)) {
1352 page = (u64 *)get_zeroed_page(gfp);
1355 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1358 /* No level skipping support yet */
1359 if (PM_PTE_LEVEL(*pte) != level)
1364 pte = IOMMU_PTE_PAGE(*pte);
1366 if (pte_page && level == end_lvl)
1369 pte = &pte[PM_LEVEL_INDEX(level, address)];
1376 * This function checks if there is a PTE for a given dma address. If
1377 * there is one, it returns the pointer to it.
1379 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1384 if (address > PM_LEVEL_SIZE(domain->mode))
1387 level = domain->mode - 1;
1388 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1393 if (!IOMMU_PTE_PRESENT(*pte))
1397 if (PM_PTE_LEVEL(*pte) == 0x07) {
1398 unsigned long pte_mask, __pte;
1401 * If we have a series of large PTEs, make
1402 * sure to return a pointer to the first one.
1404 pte_mask = PTE_PAGE_SIZE(*pte);
1405 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1406 __pte = ((unsigned long)pte) & pte_mask;
1408 return (u64 *)__pte;
1411 /* No level skipping support yet */
1412 if (PM_PTE_LEVEL(*pte) != level)
1417 /* Walk to the next level */
1418 pte = IOMMU_PTE_PAGE(*pte);
1419 pte = &pte[PM_LEVEL_INDEX(level, address)];
1426 * Generic mapping functions. It maps a physical address into a DMA
1427 * address space. It allocates the page table pages if necessary.
1428 * In the future it can be extended to a generic mapping function
1429 * supporting all features of AMD IOMMU page tables like level skipping
1430 * and full 64 bit address spaces.
1432 static int iommu_map_page(struct protection_domain *dom,
1433 unsigned long bus_addr,
1434 unsigned long phys_addr,
1436 unsigned long page_size)
1441 if (!(prot & IOMMU_PROT_MASK))
1444 bus_addr = PAGE_ALIGN(bus_addr);
1445 phys_addr = PAGE_ALIGN(phys_addr);
1446 count = PAGE_SIZE_PTE_COUNT(page_size);
1447 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1449 for (i = 0; i < count; ++i)
1450 if (IOMMU_PTE_PRESENT(pte[i]))
1453 if (page_size > PAGE_SIZE) {
1454 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1455 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1457 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1459 if (prot & IOMMU_PROT_IR)
1460 __pte |= IOMMU_PTE_IR;
1461 if (prot & IOMMU_PROT_IW)
1462 __pte |= IOMMU_PTE_IW;
1464 for (i = 0; i < count; ++i)
1472 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1473 unsigned long bus_addr,
1474 unsigned long page_size)
1476 unsigned long long unmap_size, unmapped;
1479 BUG_ON(!is_power_of_2(page_size));
1483 while (unmapped < page_size) {
1485 pte = fetch_pte(dom, bus_addr);
1489 * No PTE for this address
1490 * move forward in 4kb steps
1492 unmap_size = PAGE_SIZE;
1493 } else if (PM_PTE_LEVEL(*pte) == 0) {
1494 /* 4kb PTE found for this address */
1495 unmap_size = PAGE_SIZE;
1500 /* Large PTE found which maps this address */
1501 unmap_size = PTE_PAGE_SIZE(*pte);
1503 /* Only unmap from the first pte in the page */
1504 if ((unmap_size - 1) & bus_addr)
1506 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1507 for (i = 0; i < count; i++)
1511 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1512 unmapped += unmap_size;
1515 BUG_ON(unmapped && !is_power_of_2(unmapped));
1521 * This function checks if a specific unity mapping entry is needed for
1522 * this specific IOMMU.
1524 static int iommu_for_unity_map(struct amd_iommu *iommu,
1525 struct unity_map_entry *entry)
1529 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1530 bdf = amd_iommu_alias_table[i];
1531 if (amd_iommu_rlookup_table[bdf] == iommu)
1539 * This function actually applies the mapping to the page table of the
1542 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1543 struct unity_map_entry *e)
1548 for (addr = e->address_start; addr < e->address_end;
1549 addr += PAGE_SIZE) {
1550 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1555 * if unity mapping is in aperture range mark the page
1556 * as allocated in the aperture
1558 if (addr < dma_dom->aperture_size)
1559 __set_bit(addr >> PAGE_SHIFT,
1560 dma_dom->aperture[0]->bitmap);
1567 * Init the unity mappings for a specific IOMMU in the system
1569 * Basically iterates over all unity mapping entries and applies them to
1570 * the default domain DMA of that IOMMU if necessary.
1572 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1574 struct unity_map_entry *entry;
1577 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1578 if (!iommu_for_unity_map(iommu, entry))
1580 ret = dma_ops_unity_map(iommu->default_dom, entry);
1589 * Inits the unity mappings required for a specific device
1591 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1594 struct unity_map_entry *e;
1597 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1598 if (!(devid >= e->devid_start && devid <= e->devid_end))
1600 ret = dma_ops_unity_map(dma_dom, e);
1608 /****************************************************************************
1610 * The next functions belong to the address allocator for the dma_ops
1611 * interface functions. They work like the allocators in the other IOMMU
1612 * drivers. Its basically a bitmap which marks the allocated pages in
1613 * the aperture. Maybe it could be enhanced in the future to a more
1614 * efficient allocator.
1616 ****************************************************************************/
1619 * The address allocator core functions.
1621 * called with domain->lock held
1625 * Used to reserve address ranges in the aperture (e.g. for exclusion
1628 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1629 unsigned long start_page,
1632 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1634 if (start_page + pages > last_page)
1635 pages = last_page - start_page;
1637 for (i = start_page; i < start_page + pages; ++i) {
1638 int index = i / APERTURE_RANGE_PAGES;
1639 int page = i % APERTURE_RANGE_PAGES;
1640 __set_bit(page, dom->aperture[index]->bitmap);
1645 * This function is used to add a new aperture range to an existing
1646 * aperture in case of dma_ops domain allocation or address allocation
1649 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1650 bool populate, gfp_t gfp)
1652 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1653 struct amd_iommu *iommu;
1654 unsigned long i, old_size;
1656 #ifdef CONFIG_IOMMU_STRESS
1660 if (index >= APERTURE_MAX_RANGES)
1663 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1664 if (!dma_dom->aperture[index])
1667 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1668 if (!dma_dom->aperture[index]->bitmap)
1671 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1674 unsigned long address = dma_dom->aperture_size;
1675 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1676 u64 *pte, *pte_page;
1678 for (i = 0; i < num_ptes; ++i) {
1679 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1684 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1686 address += APERTURE_RANGE_SIZE / 64;
1690 old_size = dma_dom->aperture_size;
1691 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1693 /* Reserve address range used for MSI messages */
1694 if (old_size < MSI_ADDR_BASE_LO &&
1695 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1696 unsigned long spage;
1699 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1700 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1702 dma_ops_reserve_addresses(dma_dom, spage, pages);
1705 /* Initialize the exclusion range if necessary */
1706 for_each_iommu(iommu) {
1707 if (iommu->exclusion_start &&
1708 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1709 && iommu->exclusion_start < dma_dom->aperture_size) {
1710 unsigned long startpage;
1711 int pages = iommu_num_pages(iommu->exclusion_start,
1712 iommu->exclusion_length,
1714 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1715 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1720 * Check for areas already mapped as present in the new aperture
1721 * range and mark those pages as reserved in the allocator. Such
1722 * mappings may already exist as a result of requested unity
1723 * mappings for devices.
1725 for (i = dma_dom->aperture[index]->offset;
1726 i < dma_dom->aperture_size;
1728 u64 *pte = fetch_pte(&dma_dom->domain, i);
1729 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1732 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1735 update_domain(&dma_dom->domain);
1740 update_domain(&dma_dom->domain);
1742 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1744 kfree(dma_dom->aperture[index]);
1745 dma_dom->aperture[index] = NULL;
1750 static unsigned long dma_ops_area_alloc(struct device *dev,
1751 struct dma_ops_domain *dom,
1753 unsigned long align_mask,
1755 unsigned long start)
1757 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1758 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1759 int i = start >> APERTURE_RANGE_SHIFT;
1760 unsigned long boundary_size;
1761 unsigned long address = -1;
1762 unsigned long limit;
1764 next_bit >>= PAGE_SHIFT;
1766 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1767 PAGE_SIZE) >> PAGE_SHIFT;
1769 for (;i < max_index; ++i) {
1770 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1772 if (dom->aperture[i]->offset >= dma_mask)
1775 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1776 dma_mask >> PAGE_SHIFT);
1778 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1779 limit, next_bit, pages, 0,
1780 boundary_size, align_mask);
1781 if (address != -1) {
1782 address = dom->aperture[i]->offset +
1783 (address << PAGE_SHIFT);
1784 dom->next_address = address + (pages << PAGE_SHIFT);
1794 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1795 struct dma_ops_domain *dom,
1797 unsigned long align_mask,
1800 unsigned long address;
1802 #ifdef CONFIG_IOMMU_STRESS
1803 dom->next_address = 0;
1804 dom->need_flush = true;
1807 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1808 dma_mask, dom->next_address);
1810 if (address == -1) {
1811 dom->next_address = 0;
1812 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1814 dom->need_flush = true;
1817 if (unlikely(address == -1))
1818 address = DMA_ERROR_CODE;
1820 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1826 * The address free function.
1828 * called with domain->lock held
1830 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1831 unsigned long address,
1834 unsigned i = address >> APERTURE_RANGE_SHIFT;
1835 struct aperture_range *range = dom->aperture[i];
1837 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1839 #ifdef CONFIG_IOMMU_STRESS
1844 if (address >= dom->next_address)
1845 dom->need_flush = true;
1847 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1849 bitmap_clear(range->bitmap, address, pages);
1853 /****************************************************************************
1855 * The next functions belong to the domain allocation. A domain is
1856 * allocated for every IOMMU as the default domain. If device isolation
1857 * is enabled, every device get its own domain. The most important thing
1858 * about domains is the page table mapping the DMA address space they
1861 ****************************************************************************/
1864 * This function adds a protection domain to the global protection domain list
1866 static void add_domain_to_list(struct protection_domain *domain)
1868 unsigned long flags;
1870 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1871 list_add(&domain->list, &amd_iommu_pd_list);
1872 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1876 * This function removes a protection domain to the global
1877 * protection domain list
1879 static void del_domain_from_list(struct protection_domain *domain)
1881 unsigned long flags;
1883 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1884 list_del(&domain->list);
1885 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1888 static u16 domain_id_alloc(void)
1890 unsigned long flags;
1893 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1894 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1896 if (id > 0 && id < MAX_DOMAIN_ID)
1897 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1900 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1905 static void domain_id_free(int id)
1907 unsigned long flags;
1909 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1910 if (id > 0 && id < MAX_DOMAIN_ID)
1911 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1912 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1915 #define DEFINE_FREE_PT_FN(LVL, FN) \
1916 static void free_pt_##LVL (unsigned long __pt) \
1924 for (i = 0; i < 512; ++i) { \
1925 if (!IOMMU_PTE_PRESENT(pt[i])) \
1928 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1931 free_page((unsigned long)pt); \
1934 DEFINE_FREE_PT_FN(l2, free_page)
1935 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1936 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1937 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1938 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1940 static void free_pagetable(struct protection_domain *domain)
1942 unsigned long root = (unsigned long)domain->pt_root;
1944 switch (domain->mode) {
1945 case PAGE_MODE_NONE:
1947 case PAGE_MODE_1_LEVEL:
1950 case PAGE_MODE_2_LEVEL:
1953 case PAGE_MODE_3_LEVEL:
1956 case PAGE_MODE_4_LEVEL:
1959 case PAGE_MODE_5_LEVEL:
1962 case PAGE_MODE_6_LEVEL:
1970 static void free_gcr3_tbl_level1(u64 *tbl)
1975 for (i = 0; i < 512; ++i) {
1976 if (!(tbl[i] & GCR3_VALID))
1979 ptr = __va(tbl[i] & PAGE_MASK);
1981 free_page((unsigned long)ptr);
1985 static void free_gcr3_tbl_level2(u64 *tbl)
1990 for (i = 0; i < 512; ++i) {
1991 if (!(tbl[i] & GCR3_VALID))
1994 ptr = __va(tbl[i] & PAGE_MASK);
1996 free_gcr3_tbl_level1(ptr);
2000 static void free_gcr3_table(struct protection_domain *domain)
2002 if (domain->glx == 2)
2003 free_gcr3_tbl_level2(domain->gcr3_tbl);
2004 else if (domain->glx == 1)
2005 free_gcr3_tbl_level1(domain->gcr3_tbl);
2006 else if (domain->glx != 0)
2009 free_page((unsigned long)domain->gcr3_tbl);
2013 * Free a domain, only used if something went wrong in the
2014 * allocation path and we need to free an already allocated page table
2016 static void dma_ops_domain_free(struct dma_ops_domain *dom)
2023 del_domain_from_list(&dom->domain);
2025 free_pagetable(&dom->domain);
2027 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
2028 if (!dom->aperture[i])
2030 free_page((unsigned long)dom->aperture[i]->bitmap);
2031 kfree(dom->aperture[i]);
2038 * Allocates a new protection domain usable for the dma_ops functions.
2039 * It also initializes the page table and the address allocator data
2040 * structures required for the dma_ops interface
2042 static struct dma_ops_domain *dma_ops_domain_alloc(void)
2044 struct dma_ops_domain *dma_dom;
2046 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2050 spin_lock_init(&dma_dom->domain.lock);
2052 dma_dom->domain.id = domain_id_alloc();
2053 if (dma_dom->domain.id == 0)
2055 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
2056 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2057 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2058 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2059 dma_dom->domain.priv = dma_dom;
2060 if (!dma_dom->domain.pt_root)
2063 dma_dom->need_flush = false;
2064 dma_dom->target_dev = 0xffff;
2066 add_domain_to_list(&dma_dom->domain);
2068 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2072 * mark the first page as allocated so we never return 0 as
2073 * a valid dma-address. So we can use 0 as error value
2075 dma_dom->aperture[0]->bitmap[0] = 1;
2076 dma_dom->next_address = 0;
2082 dma_ops_domain_free(dma_dom);
2088 * little helper function to check whether a given protection domain is a
2091 static bool dma_ops_domain(struct protection_domain *domain)
2093 return domain->flags & PD_DMA_OPS_MASK;
2096 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2101 if (domain->mode != PAGE_MODE_NONE)
2102 pte_root = virt_to_phys(domain->pt_root);
2104 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2105 << DEV_ENTRY_MODE_SHIFT;
2106 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2108 flags = amd_iommu_dev_table[devid].data[1];
2111 flags |= DTE_FLAG_IOTLB;
2113 if (domain->flags & PD_IOMMUV2_MASK) {
2114 u64 gcr3 = __pa(domain->gcr3_tbl);
2115 u64 glx = domain->glx;
2118 pte_root |= DTE_FLAG_GV;
2119 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2121 /* First mask out possible old values for GCR3 table */
2122 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2125 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2128 /* Encode GCR3 table into DTE */
2129 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2132 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2135 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2139 flags &= ~(0xffffUL);
2140 flags |= domain->id;
2142 amd_iommu_dev_table[devid].data[1] = flags;
2143 amd_iommu_dev_table[devid].data[0] = pte_root;
2146 static void clear_dte_entry(u16 devid)
2148 /* remove entry from the device table seen by the hardware */
2149 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2150 amd_iommu_dev_table[devid].data[1] = 0;
2152 amd_iommu_apply_erratum_63(devid);
2155 static void do_attach(struct iommu_dev_data *dev_data,
2156 struct protection_domain *domain)
2158 struct amd_iommu *iommu;
2161 iommu = amd_iommu_rlookup_table[dev_data->devid];
2162 ats = dev_data->ats.enabled;
2164 /* Update data structures */
2165 dev_data->domain = domain;
2166 list_add(&dev_data->list, &domain->dev_list);
2167 set_dte_entry(dev_data->devid, domain, ats);
2169 /* Do reference counting */
2170 domain->dev_iommu[iommu->index] += 1;
2171 domain->dev_cnt += 1;
2173 /* Flush the DTE entry */
2174 device_flush_dte(dev_data);
2177 static void do_detach(struct iommu_dev_data *dev_data)
2179 struct amd_iommu *iommu;
2181 iommu = amd_iommu_rlookup_table[dev_data->devid];
2183 /* decrease reference counters */
2184 dev_data->domain->dev_iommu[iommu->index] -= 1;
2185 dev_data->domain->dev_cnt -= 1;
2187 /* Update data structures */
2188 dev_data->domain = NULL;
2189 list_del(&dev_data->list);
2190 clear_dte_entry(dev_data->devid);
2192 /* Flush the DTE entry */
2193 device_flush_dte(dev_data);
2197 * If a device is not yet associated with a domain, this function does
2198 * assigns it visible for the hardware
2200 static int __attach_device(struct iommu_dev_data *dev_data,
2201 struct protection_domain *domain)
2206 spin_lock(&domain->lock);
2208 if (dev_data->alias_data != NULL) {
2209 struct iommu_dev_data *alias_data = dev_data->alias_data;
2211 /* Some sanity checks */
2213 if (alias_data->domain != NULL &&
2214 alias_data->domain != domain)
2217 if (dev_data->domain != NULL &&
2218 dev_data->domain != domain)
2221 /* Do real assignment */
2222 if (alias_data->domain == NULL)
2223 do_attach(alias_data, domain);
2225 atomic_inc(&alias_data->bind);
2228 if (dev_data->domain == NULL)
2229 do_attach(dev_data, domain);
2231 atomic_inc(&dev_data->bind);
2238 spin_unlock(&domain->lock);
2244 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2246 pci_disable_ats(pdev);
2247 pci_disable_pri(pdev);
2248 pci_disable_pasid(pdev);
2251 /* FIXME: Change generic reset-function to do the same */
2252 static int pri_reset_while_enabled(struct pci_dev *pdev)
2257 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2261 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2262 control |= PCI_PRI_CTRL_RESET;
2263 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2268 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2273 /* FIXME: Hardcode number of outstanding requests for now */
2275 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2277 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2279 /* Only allow access to user-accessible pages */
2280 ret = pci_enable_pasid(pdev, 0);
2284 /* First reset the PRI state of the device */
2285 ret = pci_reset_pri(pdev);
2290 ret = pci_enable_pri(pdev, reqs);
2295 ret = pri_reset_while_enabled(pdev);
2300 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2307 pci_disable_pri(pdev);
2308 pci_disable_pasid(pdev);
2313 /* FIXME: Move this to PCI code */
2314 #define PCI_PRI_TLP_OFF (1 << 15)
2316 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2321 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2325 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2327 return (status & PCI_PRI_TLP_OFF) ? true : false;
2331 * If a device is not yet associated with a domain, this function
2332 * assigns it visible for the hardware
2334 static int attach_device(struct device *dev,
2335 struct protection_domain *domain)
2337 struct pci_dev *pdev = to_pci_dev(dev);
2338 struct iommu_dev_data *dev_data;
2339 unsigned long flags;
2342 dev_data = get_dev_data(dev);
2344 if (domain->flags & PD_IOMMUV2_MASK) {
2345 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2348 if (pdev_iommuv2_enable(pdev) != 0)
2351 dev_data->ats.enabled = true;
2352 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2353 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2354 } else if (amd_iommu_iotlb_sup &&
2355 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2356 dev_data->ats.enabled = true;
2357 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2360 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2361 ret = __attach_device(dev_data, domain);
2362 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2365 * We might boot into a crash-kernel here. The crashed kernel
2366 * left the caches in the IOMMU dirty. So we have to flush
2367 * here to evict all dirty stuff.
2369 domain_flush_tlb_pde(domain);
2375 * Removes a device from a protection domain (unlocked)
2377 static void __detach_device(struct iommu_dev_data *dev_data)
2379 struct protection_domain *domain;
2380 unsigned long flags;
2382 BUG_ON(!dev_data->domain);
2384 domain = dev_data->domain;
2386 spin_lock_irqsave(&domain->lock, flags);
2388 if (dev_data->alias_data != NULL) {
2389 struct iommu_dev_data *alias_data = dev_data->alias_data;
2391 if (atomic_dec_and_test(&alias_data->bind))
2392 do_detach(alias_data);
2395 if (atomic_dec_and_test(&dev_data->bind))
2396 do_detach(dev_data);
2398 spin_unlock_irqrestore(&domain->lock, flags);
2401 * If we run in passthrough mode the device must be assigned to the
2402 * passthrough domain if it is detached from any other domain.
2403 * Make sure we can deassign from the pt_domain itself.
2405 if (dev_data->passthrough &&
2406 (dev_data->domain == NULL && domain != pt_domain))
2407 __attach_device(dev_data, pt_domain);
2411 * Removes a device from a protection domain (with devtable_lock held)
2413 static void detach_device(struct device *dev)
2415 struct protection_domain *domain;
2416 struct iommu_dev_data *dev_data;
2417 unsigned long flags;
2419 dev_data = get_dev_data(dev);
2420 domain = dev_data->domain;
2422 /* lock device table */
2423 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2424 __detach_device(dev_data);
2425 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2427 if (domain->flags & PD_IOMMUV2_MASK)
2428 pdev_iommuv2_disable(to_pci_dev(dev));
2429 else if (dev_data->ats.enabled)
2430 pci_disable_ats(to_pci_dev(dev));
2432 dev_data->ats.enabled = false;
2436 * Find out the protection domain structure for a given PCI device. This
2437 * will give us the pointer to the page table root for example.
2439 static struct protection_domain *domain_for_device(struct device *dev)
2441 struct iommu_dev_data *dev_data;
2442 struct protection_domain *dom = NULL;
2443 unsigned long flags;
2445 dev_data = get_dev_data(dev);
2447 if (dev_data->domain)
2448 return dev_data->domain;
2450 if (dev_data->alias_data != NULL) {
2451 struct iommu_dev_data *alias_data = dev_data->alias_data;
2453 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2454 if (alias_data->domain != NULL) {
2455 __attach_device(dev_data, alias_data->domain);
2456 dom = alias_data->domain;
2458 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2464 static int device_change_notifier(struct notifier_block *nb,
2465 unsigned long action, void *data)
2467 struct dma_ops_domain *dma_domain;
2468 struct protection_domain *domain;
2469 struct iommu_dev_data *dev_data;
2470 struct device *dev = data;
2471 struct amd_iommu *iommu;
2472 unsigned long flags;
2475 if (!check_device(dev))
2478 devid = get_device_id(dev);
2479 iommu = amd_iommu_rlookup_table[devid];
2480 dev_data = get_dev_data(dev);
2483 case BUS_NOTIFY_UNBOUND_DRIVER:
2485 domain = domain_for_device(dev);
2489 if (dev_data->passthrough)
2493 case BUS_NOTIFY_ADD_DEVICE:
2495 iommu_init_device(dev);
2498 * dev_data is still NULL and
2499 * got initialized in iommu_init_device
2501 dev_data = get_dev_data(dev);
2503 if (iommu_pass_through || dev_data->iommu_v2) {
2504 dev_data->passthrough = true;
2505 attach_device(dev, pt_domain);
2509 domain = domain_for_device(dev);
2511 /* allocate a protection domain if a device is added */
2512 dma_domain = find_protection_domain(devid);
2514 dma_domain = dma_ops_domain_alloc();
2517 dma_domain->target_dev = devid;
2519 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2520 list_add_tail(&dma_domain->list, &iommu_pd_list);
2521 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2524 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2527 case BUS_NOTIFY_DEL_DEVICE:
2529 iommu_uninit_device(dev);
2535 iommu_completion_wait(iommu);
2541 static struct notifier_block device_nb = {
2542 .notifier_call = device_change_notifier,
2545 void amd_iommu_init_notifier(void)
2547 bus_register_notifier(&pci_bus_type, &device_nb);
2550 /*****************************************************************************
2552 * The next functions belong to the dma_ops mapping/unmapping code.
2554 *****************************************************************************/
2557 * In the dma_ops path we only have the struct device. This function
2558 * finds the corresponding IOMMU, the protection domain and the
2559 * requestor id for a given device.
2560 * If the device is not yet associated with a domain this is also done
2563 static struct protection_domain *get_domain(struct device *dev)
2565 struct protection_domain *domain;
2566 struct dma_ops_domain *dma_dom;
2567 u16 devid = get_device_id(dev);
2569 if (!check_device(dev))
2570 return ERR_PTR(-EINVAL);
2572 domain = domain_for_device(dev);
2573 if (domain != NULL && !dma_ops_domain(domain))
2574 return ERR_PTR(-EBUSY);
2579 /* Device not bound yet - bind it */
2580 dma_dom = find_protection_domain(devid);
2582 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2583 attach_device(dev, &dma_dom->domain);
2584 DUMP_printk("Using protection domain %d for device %s\n",
2585 dma_dom->domain.id, dev_name(dev));
2587 return &dma_dom->domain;
2590 static void update_device_table(struct protection_domain *domain)
2592 struct iommu_dev_data *dev_data;
2594 list_for_each_entry(dev_data, &domain->dev_list, list)
2595 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2598 static void update_domain(struct protection_domain *domain)
2600 if (!domain->updated)
2603 update_device_table(domain);
2605 domain_flush_devices(domain);
2606 domain_flush_tlb_pde(domain);
2608 domain->updated = false;
2612 * This function fetches the PTE for a given address in the aperture
2614 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2615 unsigned long address)
2617 struct aperture_range *aperture;
2618 u64 *pte, *pte_page;
2620 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2624 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2626 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2628 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2630 pte += PM_LEVEL_INDEX(0, address);
2632 update_domain(&dom->domain);
2638 * This is the generic map function. It maps one 4kb page at paddr to
2639 * the given address in the DMA address space for the domain.
2641 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2642 unsigned long address,
2648 WARN_ON(address > dom->aperture_size);
2652 pte = dma_ops_get_pte(dom, address);
2654 return DMA_ERROR_CODE;
2656 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2658 if (direction == DMA_TO_DEVICE)
2659 __pte |= IOMMU_PTE_IR;
2660 else if (direction == DMA_FROM_DEVICE)
2661 __pte |= IOMMU_PTE_IW;
2662 else if (direction == DMA_BIDIRECTIONAL)
2663 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2669 return (dma_addr_t)address;
2673 * The generic unmapping function for on page in the DMA address space.
2675 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2676 unsigned long address)
2678 struct aperture_range *aperture;
2681 if (address >= dom->aperture_size)
2684 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2688 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2692 pte += PM_LEVEL_INDEX(0, address);
2700 * This function contains common code for mapping of a physically
2701 * contiguous memory region into DMA address space. It is used by all
2702 * mapping functions provided with this IOMMU driver.
2703 * Must be called with the domain lock held.
2705 static dma_addr_t __map_single(struct device *dev,
2706 struct dma_ops_domain *dma_dom,
2713 dma_addr_t offset = paddr & ~PAGE_MASK;
2714 dma_addr_t address, start, ret;
2716 unsigned long align_mask = 0;
2719 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2722 INC_STATS_COUNTER(total_map_requests);
2725 INC_STATS_COUNTER(cross_page);
2728 align_mask = (1UL << get_order(size)) - 1;
2731 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2733 if (unlikely(address == DMA_ERROR_CODE)) {
2735 * setting next_address here will let the address
2736 * allocator only scan the new allocated range in the
2737 * first run. This is a small optimization.
2739 dma_dom->next_address = dma_dom->aperture_size;
2741 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2745 * aperture was successfully enlarged by 128 MB, try
2752 for (i = 0; i < pages; ++i) {
2753 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2754 if (ret == DMA_ERROR_CODE)
2762 ADD_STATS_COUNTER(alloced_io_mem, size);
2764 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2765 domain_flush_tlb(&dma_dom->domain);
2766 dma_dom->need_flush = false;
2767 } else if (unlikely(amd_iommu_np_cache))
2768 domain_flush_pages(&dma_dom->domain, address, size);
2775 for (--i; i >= 0; --i) {
2777 dma_ops_domain_unmap(dma_dom, start);
2780 dma_ops_free_addresses(dma_dom, address, pages);
2782 return DMA_ERROR_CODE;
2786 * Does the reverse of the __map_single function. Must be called with
2787 * the domain lock held too
2789 static void __unmap_single(struct dma_ops_domain *dma_dom,
2790 dma_addr_t dma_addr,
2794 dma_addr_t flush_addr;
2795 dma_addr_t i, start;
2798 if ((dma_addr == DMA_ERROR_CODE) ||
2799 (dma_addr + size > dma_dom->aperture_size))
2802 flush_addr = dma_addr;
2803 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2804 dma_addr &= PAGE_MASK;
2807 for (i = 0; i < pages; ++i) {
2808 dma_ops_domain_unmap(dma_dom, start);
2812 SUB_STATS_COUNTER(alloced_io_mem, size);
2814 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2816 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2817 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2818 dma_dom->need_flush = false;
2823 * The exported map_single function for dma_ops.
2825 static dma_addr_t map_page(struct device *dev, struct page *page,
2826 unsigned long offset, size_t size,
2827 enum dma_data_direction dir,
2828 struct dma_attrs *attrs)
2830 unsigned long flags;
2831 struct protection_domain *domain;
2834 phys_addr_t paddr = page_to_phys(page) + offset;
2836 INC_STATS_COUNTER(cnt_map_single);
2838 domain = get_domain(dev);
2839 if (PTR_ERR(domain) == -EINVAL)
2840 return (dma_addr_t)paddr;
2841 else if (IS_ERR(domain))
2842 return DMA_ERROR_CODE;
2844 dma_mask = *dev->dma_mask;
2846 spin_lock_irqsave(&domain->lock, flags);
2848 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2850 if (addr == DMA_ERROR_CODE)
2853 domain_flush_complete(domain);
2856 spin_unlock_irqrestore(&domain->lock, flags);
2862 * The exported unmap_single function for dma_ops.
2864 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2865 enum dma_data_direction dir, struct dma_attrs *attrs)
2867 unsigned long flags;
2868 struct protection_domain *domain;
2870 INC_STATS_COUNTER(cnt_unmap_single);
2872 domain = get_domain(dev);
2876 spin_lock_irqsave(&domain->lock, flags);
2878 __unmap_single(domain->priv, dma_addr, size, dir);
2880 domain_flush_complete(domain);
2882 spin_unlock_irqrestore(&domain->lock, flags);
2886 * The exported map_sg function for dma_ops (handles scatter-gather
2889 static int map_sg(struct device *dev, struct scatterlist *sglist,
2890 int nelems, enum dma_data_direction dir,
2891 struct dma_attrs *attrs)
2893 unsigned long flags;
2894 struct protection_domain *domain;
2896 struct scatterlist *s;
2898 int mapped_elems = 0;
2901 INC_STATS_COUNTER(cnt_map_sg);
2903 domain = get_domain(dev);
2907 dma_mask = *dev->dma_mask;
2909 spin_lock_irqsave(&domain->lock, flags);
2911 for_each_sg(sglist, s, nelems, i) {
2914 s->dma_address = __map_single(dev, domain->priv,
2915 paddr, s->length, dir, false,
2918 if (s->dma_address) {
2919 s->dma_length = s->length;
2925 domain_flush_complete(domain);
2928 spin_unlock_irqrestore(&domain->lock, flags);
2930 return mapped_elems;
2932 for_each_sg(sglist, s, mapped_elems, i) {
2934 __unmap_single(domain->priv, s->dma_address,
2935 s->dma_length, dir);
2936 s->dma_address = s->dma_length = 0;
2945 * The exported map_sg function for dma_ops (handles scatter-gather
2948 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2949 int nelems, enum dma_data_direction dir,
2950 struct dma_attrs *attrs)
2952 unsigned long flags;
2953 struct protection_domain *domain;
2954 struct scatterlist *s;
2957 INC_STATS_COUNTER(cnt_unmap_sg);
2959 domain = get_domain(dev);
2963 spin_lock_irqsave(&domain->lock, flags);
2965 for_each_sg(sglist, s, nelems, i) {
2966 __unmap_single(domain->priv, s->dma_address,
2967 s->dma_length, dir);
2968 s->dma_address = s->dma_length = 0;
2971 domain_flush_complete(domain);
2973 spin_unlock_irqrestore(&domain->lock, flags);
2977 * The exported alloc_coherent function for dma_ops.
2979 static void *alloc_coherent(struct device *dev, size_t size,
2980 dma_addr_t *dma_addr, gfp_t flag,
2981 struct dma_attrs *attrs)
2983 unsigned long flags;
2985 struct protection_domain *domain;
2987 u64 dma_mask = dev->coherent_dma_mask;
2989 INC_STATS_COUNTER(cnt_alloc_coherent);
2991 domain = get_domain(dev);
2992 if (PTR_ERR(domain) == -EINVAL) {
2993 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2994 *dma_addr = __pa(virt_addr);
2996 } else if (IS_ERR(domain))
2999 dma_mask = dev->coherent_dma_mask;
3000 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
3003 virt_addr = (void *)__get_free_pages(flag, get_order(size));
3007 paddr = virt_to_phys(virt_addr);
3010 dma_mask = *dev->dma_mask;
3012 spin_lock_irqsave(&domain->lock, flags);
3014 *dma_addr = __map_single(dev, domain->priv, paddr,
3015 size, DMA_BIDIRECTIONAL, true, dma_mask);
3017 if (*dma_addr == DMA_ERROR_CODE) {
3018 spin_unlock_irqrestore(&domain->lock, flags);
3022 domain_flush_complete(domain);
3024 spin_unlock_irqrestore(&domain->lock, flags);
3030 free_pages((unsigned long)virt_addr, get_order(size));
3036 * The exported free_coherent function for dma_ops.
3038 static void free_coherent(struct device *dev, size_t size,
3039 void *virt_addr, dma_addr_t dma_addr,
3040 struct dma_attrs *attrs)
3042 unsigned long flags;
3043 struct protection_domain *domain;
3045 INC_STATS_COUNTER(cnt_free_coherent);
3047 domain = get_domain(dev);
3051 spin_lock_irqsave(&domain->lock, flags);
3053 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
3055 domain_flush_complete(domain);
3057 spin_unlock_irqrestore(&domain->lock, flags);
3060 free_pages((unsigned long)virt_addr, get_order(size));
3064 * This function is called by the DMA layer to find out if we can handle a
3065 * particular device. It is part of the dma_ops.
3067 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3069 return check_device(dev);
3073 * The function for pre-allocating protection domains.
3075 * If the driver core informs the DMA layer if a driver grabs a device
3076 * we don't need to preallocate the protection domains anymore.
3077 * For now we have to.
3079 static void __init prealloc_protection_domains(void)
3081 struct iommu_dev_data *dev_data;
3082 struct dma_ops_domain *dma_dom;
3083 struct pci_dev *dev = NULL;
3086 for_each_pci_dev(dev) {
3088 /* Do we handle this device? */
3089 if (!check_device(&dev->dev))
3092 dev_data = get_dev_data(&dev->dev);
3093 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3094 /* Make sure passthrough domain is allocated */
3095 alloc_passthrough_domain();
3096 dev_data->passthrough = true;
3097 attach_device(&dev->dev, pt_domain);
3098 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3099 dev_name(&dev->dev));
3102 /* Is there already any domain for it? */
3103 if (domain_for_device(&dev->dev))
3106 devid = get_device_id(&dev->dev);
3108 dma_dom = dma_ops_domain_alloc();
3111 init_unity_mappings_for_device(dma_dom, devid);
3112 dma_dom->target_dev = devid;
3114 attach_device(&dev->dev, &dma_dom->domain);
3116 list_add_tail(&dma_dom->list, &iommu_pd_list);
3120 static struct dma_map_ops amd_iommu_dma_ops = {
3121 .alloc = alloc_coherent,
3122 .free = free_coherent,
3123 .map_page = map_page,
3124 .unmap_page = unmap_page,
3126 .unmap_sg = unmap_sg,
3127 .dma_supported = amd_iommu_dma_supported,
3130 static unsigned device_dma_ops_init(void)
3132 struct iommu_dev_data *dev_data;
3133 struct pci_dev *pdev = NULL;
3134 unsigned unhandled = 0;
3136 for_each_pci_dev(pdev) {
3137 if (!check_device(&pdev->dev)) {
3139 iommu_ignore_device(&pdev->dev);
3145 dev_data = get_dev_data(&pdev->dev);
3147 if (!dev_data->passthrough)
3148 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3150 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3157 * The function which clues the AMD IOMMU driver into dma_ops.
3160 void __init amd_iommu_init_api(void)
3162 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3165 int __init amd_iommu_init_dma_ops(void)
3167 struct amd_iommu *iommu;
3171 * first allocate a default protection domain for every IOMMU we
3172 * found in the system. Devices not assigned to any other
3173 * protection domain will be assigned to the default one.
3175 for_each_iommu(iommu) {
3176 iommu->default_dom = dma_ops_domain_alloc();
3177 if (iommu->default_dom == NULL)
3179 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3180 ret = iommu_init_unity_mappings(iommu);
3186 * Pre-allocate the protection domains for each device.
3188 prealloc_protection_domains();
3193 /* Make the driver finally visible to the drivers */
3194 unhandled = device_dma_ops_init();
3195 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3196 /* There are unhandled devices - initialize swiotlb for them */
3200 amd_iommu_stats_init();
3202 if (amd_iommu_unmap_flush)
3203 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3205 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3211 for_each_iommu(iommu) {
3212 dma_ops_domain_free(iommu->default_dom);
3218 /*****************************************************************************
3220 * The following functions belong to the exported interface of AMD IOMMU
3222 * This interface allows access to lower level functions of the IOMMU
3223 * like protection domain handling and assignement of devices to domains
3224 * which is not possible with the dma_ops interface.
3226 *****************************************************************************/
3228 static void cleanup_domain(struct protection_domain *domain)
3230 struct iommu_dev_data *dev_data, *next;
3231 unsigned long flags;
3233 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3235 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3236 __detach_device(dev_data);
3237 atomic_set(&dev_data->bind, 0);
3240 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3243 static void protection_domain_free(struct protection_domain *domain)
3248 del_domain_from_list(domain);
3251 domain_id_free(domain->id);
3256 static struct protection_domain *protection_domain_alloc(void)
3258 struct protection_domain *domain;
3260 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3264 spin_lock_init(&domain->lock);
3265 mutex_init(&domain->api_lock);
3266 domain->id = domain_id_alloc();
3269 INIT_LIST_HEAD(&domain->dev_list);
3271 add_domain_to_list(domain);
3281 static int __init alloc_passthrough_domain(void)
3283 if (pt_domain != NULL)
3286 /* allocate passthrough domain */
3287 pt_domain = protection_domain_alloc();
3291 pt_domain->mode = PAGE_MODE_NONE;
3295 static int amd_iommu_domain_init(struct iommu_domain *dom)
3297 struct protection_domain *domain;
3299 domain = protection_domain_alloc();
3303 domain->mode = PAGE_MODE_3_LEVEL;
3304 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3305 if (!domain->pt_root)
3308 domain->iommu_domain = dom;
3312 dom->geometry.aperture_start = 0;
3313 dom->geometry.aperture_end = ~0ULL;
3314 dom->geometry.force_aperture = true;
3319 protection_domain_free(domain);
3324 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3326 struct protection_domain *domain = dom->priv;
3331 if (domain->dev_cnt > 0)
3332 cleanup_domain(domain);
3334 BUG_ON(domain->dev_cnt != 0);
3336 if (domain->mode != PAGE_MODE_NONE)
3337 free_pagetable(domain);
3339 if (domain->flags & PD_IOMMUV2_MASK)
3340 free_gcr3_table(domain);
3342 protection_domain_free(domain);
3347 static void amd_iommu_detach_device(struct iommu_domain *dom,
3350 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3351 struct amd_iommu *iommu;
3354 if (!check_device(dev))
3357 devid = get_device_id(dev);
3359 if (dev_data->domain != NULL)
3362 iommu = amd_iommu_rlookup_table[devid];
3366 iommu_completion_wait(iommu);
3369 static int amd_iommu_attach_device(struct iommu_domain *dom,
3372 struct protection_domain *domain = dom->priv;
3373 struct iommu_dev_data *dev_data;
3374 struct amd_iommu *iommu;
3377 if (!check_device(dev))
3380 dev_data = dev->archdata.iommu;
3382 iommu = amd_iommu_rlookup_table[dev_data->devid];
3386 if (dev_data->domain)
3389 ret = attach_device(dev, domain);
3391 iommu_completion_wait(iommu);
3396 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3397 phys_addr_t paddr, size_t page_size, int iommu_prot)
3399 struct protection_domain *domain = dom->priv;
3403 if (domain->mode == PAGE_MODE_NONE)
3406 if (iommu_prot & IOMMU_READ)
3407 prot |= IOMMU_PROT_IR;
3408 if (iommu_prot & IOMMU_WRITE)
3409 prot |= IOMMU_PROT_IW;
3411 mutex_lock(&domain->api_lock);
3412 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3413 mutex_unlock(&domain->api_lock);
3418 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3421 struct protection_domain *domain = dom->priv;
3424 if (domain->mode == PAGE_MODE_NONE)
3427 mutex_lock(&domain->api_lock);
3428 unmap_size = iommu_unmap_page(domain, iova, page_size);
3429 mutex_unlock(&domain->api_lock);
3431 domain_flush_tlb_pde(domain);
3436 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3439 struct protection_domain *domain = dom->priv;
3440 unsigned long offset_mask;
3444 if (domain->mode == PAGE_MODE_NONE)
3447 pte = fetch_pte(domain, iova);
3449 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3452 if (PM_PTE_LEVEL(*pte) == 0)
3453 offset_mask = PAGE_SIZE - 1;
3455 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3457 __pte = *pte & PM_ADDR_MASK;
3458 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3463 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3467 case IOMMU_CAP_CACHE_COHERENCY:
3469 case IOMMU_CAP_INTR_REMAP:
3470 return irq_remapping_enabled;
3476 static struct iommu_ops amd_iommu_ops = {
3477 .domain_init = amd_iommu_domain_init,
3478 .domain_destroy = amd_iommu_domain_destroy,
3479 .attach_dev = amd_iommu_attach_device,
3480 .detach_dev = amd_iommu_detach_device,
3481 .map = amd_iommu_map,
3482 .unmap = amd_iommu_unmap,
3483 .iova_to_phys = amd_iommu_iova_to_phys,
3484 .domain_has_cap = amd_iommu_domain_has_cap,
3485 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3488 /*****************************************************************************
3490 * The next functions do a basic initialization of IOMMU for pass through
3493 * In passthrough mode the IOMMU is initialized and enabled but not used for
3494 * DMA-API translation.
3496 *****************************************************************************/
3498 int __init amd_iommu_init_passthrough(void)
3500 struct iommu_dev_data *dev_data;
3501 struct pci_dev *dev = NULL;
3502 struct amd_iommu *iommu;
3506 ret = alloc_passthrough_domain();
3510 for_each_pci_dev(dev) {
3511 if (!check_device(&dev->dev))
3514 dev_data = get_dev_data(&dev->dev);
3515 dev_data->passthrough = true;
3517 devid = get_device_id(&dev->dev);
3519 iommu = amd_iommu_rlookup_table[devid];
3523 attach_device(&dev->dev, pt_domain);
3526 amd_iommu_stats_init();
3528 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3533 /* IOMMUv2 specific functions */
3534 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3536 return atomic_notifier_chain_register(&ppr_notifier, nb);
3538 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3540 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3542 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3544 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3546 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3548 struct protection_domain *domain = dom->priv;
3549 unsigned long flags;
3551 spin_lock_irqsave(&domain->lock, flags);
3553 /* Update data structure */
3554 domain->mode = PAGE_MODE_NONE;
3555 domain->updated = true;
3557 /* Make changes visible to IOMMUs */
3558 update_domain(domain);
3560 /* Page-table is not visible to IOMMU anymore, so free it */
3561 free_pagetable(domain);
3563 spin_unlock_irqrestore(&domain->lock, flags);
3565 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3567 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3569 struct protection_domain *domain = dom->priv;
3570 unsigned long flags;
3573 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3576 /* Number of GCR3 table levels required */
3577 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3580 if (levels > amd_iommu_max_glx_val)
3583 spin_lock_irqsave(&domain->lock, flags);
3586 * Save us all sanity checks whether devices already in the
3587 * domain support IOMMUv2. Just force that the domain has no
3588 * devices attached when it is switched into IOMMUv2 mode.
3591 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3595 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3596 if (domain->gcr3_tbl == NULL)
3599 domain->glx = levels;
3600 domain->flags |= PD_IOMMUV2_MASK;
3601 domain->updated = true;
3603 update_domain(domain);
3608 spin_unlock_irqrestore(&domain->lock, flags);
3612 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3614 static int __flush_pasid(struct protection_domain *domain, int pasid,
3615 u64 address, bool size)
3617 struct iommu_dev_data *dev_data;
3618 struct iommu_cmd cmd;
3621 if (!(domain->flags & PD_IOMMUV2_MASK))
3624 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3627 * IOMMU TLB needs to be flushed before Device TLB to
3628 * prevent device TLB refill from IOMMU TLB
3630 for (i = 0; i < amd_iommus_present; ++i) {
3631 if (domain->dev_iommu[i] == 0)
3634 ret = iommu_queue_command(amd_iommus[i], &cmd);
3639 /* Wait until IOMMU TLB flushes are complete */
3640 domain_flush_complete(domain);
3642 /* Now flush device TLBs */
3643 list_for_each_entry(dev_data, &domain->dev_list, list) {
3644 struct amd_iommu *iommu;
3647 BUG_ON(!dev_data->ats.enabled);
3649 qdep = dev_data->ats.qdep;
3650 iommu = amd_iommu_rlookup_table[dev_data->devid];
3652 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3653 qdep, address, size);
3655 ret = iommu_queue_command(iommu, &cmd);
3660 /* Wait until all device TLBs are flushed */
3661 domain_flush_complete(domain);
3670 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3673 INC_STATS_COUNTER(invalidate_iotlb);
3675 return __flush_pasid(domain, pasid, address, false);
3678 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3681 struct protection_domain *domain = dom->priv;
3682 unsigned long flags;
3685 spin_lock_irqsave(&domain->lock, flags);
3686 ret = __amd_iommu_flush_page(domain, pasid, address);
3687 spin_unlock_irqrestore(&domain->lock, flags);
3691 EXPORT_SYMBOL(amd_iommu_flush_page);
3693 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3695 INC_STATS_COUNTER(invalidate_iotlb_all);
3697 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3701 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3703 struct protection_domain *domain = dom->priv;
3704 unsigned long flags;
3707 spin_lock_irqsave(&domain->lock, flags);
3708 ret = __amd_iommu_flush_tlb(domain, pasid);
3709 spin_unlock_irqrestore(&domain->lock, flags);
3713 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3715 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3722 index = (pasid >> (9 * level)) & 0x1ff;
3728 if (!(*pte & GCR3_VALID)) {
3732 root = (void *)get_zeroed_page(GFP_ATOMIC);
3736 *pte = __pa(root) | GCR3_VALID;
3739 root = __va(*pte & PAGE_MASK);
3747 static int __set_gcr3(struct protection_domain *domain, int pasid,
3752 if (domain->mode != PAGE_MODE_NONE)
3755 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3759 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3761 return __amd_iommu_flush_tlb(domain, pasid);
3764 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3768 if (domain->mode != PAGE_MODE_NONE)
3771 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3777 return __amd_iommu_flush_tlb(domain, pasid);
3780 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3783 struct protection_domain *domain = dom->priv;
3784 unsigned long flags;
3787 spin_lock_irqsave(&domain->lock, flags);
3788 ret = __set_gcr3(domain, pasid, cr3);
3789 spin_unlock_irqrestore(&domain->lock, flags);
3793 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3795 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3797 struct protection_domain *domain = dom->priv;
3798 unsigned long flags;
3801 spin_lock_irqsave(&domain->lock, flags);
3802 ret = __clear_gcr3(domain, pasid);
3803 spin_unlock_irqrestore(&domain->lock, flags);
3807 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3809 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3810 int status, int tag)
3812 struct iommu_dev_data *dev_data;
3813 struct amd_iommu *iommu;
3814 struct iommu_cmd cmd;
3816 INC_STATS_COUNTER(complete_ppr);
3818 dev_data = get_dev_data(&pdev->dev);
3819 iommu = amd_iommu_rlookup_table[dev_data->devid];
3821 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3822 tag, dev_data->pri_tlp);
3824 return iommu_queue_command(iommu, &cmd);
3826 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3828 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3830 struct protection_domain *domain;
3832 domain = get_domain(&pdev->dev);
3836 /* Only return IOMMUv2 domains */
3837 if (!(domain->flags & PD_IOMMUV2_MASK))
3840 return domain->iommu_domain;
3842 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3844 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3846 struct iommu_dev_data *dev_data;
3848 if (!amd_iommu_v2_supported())
3851 dev_data = get_dev_data(&pdev->dev);
3852 dev_data->errata |= (1 << erratum);
3854 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3856 int amd_iommu_device_info(struct pci_dev *pdev,
3857 struct amd_iommu_device_info *info)
3862 if (pdev == NULL || info == NULL)
3865 if (!amd_iommu_v2_supported())
3868 memset(info, 0, sizeof(*info));
3870 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3872 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3874 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3876 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3878 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3882 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3883 max_pasids = min(max_pasids, (1 << 20));
3885 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3886 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3888 features = pci_pasid_features(pdev);
3889 if (features & PCI_PASID_CAP_EXEC)
3890 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3891 if (features & PCI_PASID_CAP_PRIV)
3892 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3897 EXPORT_SYMBOL(amd_iommu_device_info);
3899 #ifdef CONFIG_IRQ_REMAP
3901 /*****************************************************************************
3903 * Interrupt Remapping Implementation
3905 *****************************************************************************/
3922 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3923 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3924 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3925 #define DTE_IRQ_REMAP_ENABLE 1ULL
3927 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3931 dte = amd_iommu_dev_table[devid].data[2];
3932 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3933 dte |= virt_to_phys(table->table);
3934 dte |= DTE_IRQ_REMAP_INTCTL;
3935 dte |= DTE_IRQ_TABLE_LEN;
3936 dte |= DTE_IRQ_REMAP_ENABLE;
3938 amd_iommu_dev_table[devid].data[2] = dte;
3941 #define IRTE_ALLOCATED (~1U)
3943 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3945 struct irq_remap_table *table = NULL;
3946 struct amd_iommu *iommu;
3947 unsigned long flags;
3950 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3952 iommu = amd_iommu_rlookup_table[devid];
3956 table = irq_lookup_table[devid];
3960 alias = amd_iommu_alias_table[devid];
3961 table = irq_lookup_table[alias];
3963 irq_lookup_table[devid] = table;
3964 set_dte_irq_entry(devid, table);
3965 iommu_flush_dte(iommu, devid);
3969 /* Nothing there yet, allocate new irq remapping table */
3970 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3974 /* Initialize table spin-lock */
3975 spin_lock_init(&table->lock);
3978 /* Keep the first 32 indexes free for IOAPIC interrupts */
3979 table->min_index = 32;
3981 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3982 if (!table->table) {
3988 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3993 for (i = 0; i < 32; ++i)
3994 table->table[i] = IRTE_ALLOCATED;
3997 irq_lookup_table[devid] = table;
3998 set_dte_irq_entry(devid, table);
3999 iommu_flush_dte(iommu, devid);
4000 if (devid != alias) {
4001 irq_lookup_table[alias] = table;
4002 set_dte_irq_entry(devid, table);
4003 iommu_flush_dte(iommu, alias);
4007 iommu_completion_wait(iommu);
4010 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
4015 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
4017 struct irq_remap_table *table;
4018 unsigned long flags;
4021 table = get_irq_table(devid, false);
4025 spin_lock_irqsave(&table->lock, flags);
4027 /* Scan table for free entries */
4028 for (c = 0, index = table->min_index;
4029 index < MAX_IRQS_PER_TABLE;
4031 if (table->table[index] == 0)
4037 struct irq_2_irte *irte_info;
4040 table->table[index - c + 1] = IRTE_ALLOCATED;
4045 irte_info = &cfg->irq_2_irte;
4046 irte_info->devid = devid;
4047 irte_info->index = index;
4056 spin_unlock_irqrestore(&table->lock, flags);
4061 static int get_irte(u16 devid, int index, union irte *irte)
4063 struct irq_remap_table *table;
4064 unsigned long flags;
4066 table = get_irq_table(devid, false);
4070 spin_lock_irqsave(&table->lock, flags);
4071 irte->val = table->table[index];
4072 spin_unlock_irqrestore(&table->lock, flags);
4077 static int modify_irte(u16 devid, int index, union irte irte)
4079 struct irq_remap_table *table;
4080 struct amd_iommu *iommu;
4081 unsigned long flags;
4083 iommu = amd_iommu_rlookup_table[devid];
4087 table = get_irq_table(devid, false);
4091 spin_lock_irqsave(&table->lock, flags);
4092 table->table[index] = irte.val;
4093 spin_unlock_irqrestore(&table->lock, flags);
4095 iommu_flush_irt(iommu, devid);
4096 iommu_completion_wait(iommu);
4101 static void free_irte(u16 devid, int index)
4103 struct irq_remap_table *table;
4104 struct amd_iommu *iommu;
4105 unsigned long flags;
4107 iommu = amd_iommu_rlookup_table[devid];
4111 table = get_irq_table(devid, false);
4115 spin_lock_irqsave(&table->lock, flags);
4116 table->table[index] = 0;
4117 spin_unlock_irqrestore(&table->lock, flags);
4119 iommu_flush_irt(iommu, devid);
4120 iommu_completion_wait(iommu);
4123 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4124 unsigned int destination, int vector,
4125 struct io_apic_irq_attr *attr)
4127 struct irq_remap_table *table;
4128 struct irq_2_irte *irte_info;
4129 struct irq_cfg *cfg;
4136 cfg = irq_get_chip_data(irq);
4140 irte_info = &cfg->irq_2_irte;
4141 ioapic_id = mpc_ioapic_id(attr->ioapic);
4142 devid = get_ioapic_devid(ioapic_id);
4147 table = get_irq_table(devid, true);
4151 index = attr->ioapic_pin;
4153 /* Setup IRQ remapping info */
4155 irte_info->devid = devid;
4156 irte_info->index = index;
4158 /* Setup IRTE for IOMMU */
4160 irte.fields.vector = vector;
4161 irte.fields.int_type = apic->irq_delivery_mode;
4162 irte.fields.destination = destination;
4163 irte.fields.dm = apic->irq_dest_mode;
4164 irte.fields.valid = 1;
4166 ret = modify_irte(devid, index, irte);
4170 /* Setup IOAPIC entry */
4171 memset(entry, 0, sizeof(*entry));
4173 entry->vector = index;
4175 entry->trigger = attr->trigger;
4176 entry->polarity = attr->polarity;
4179 * Mask level triggered irqs.
4187 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4190 struct irq_2_irte *irte_info;
4191 unsigned int dest, irq;
4192 struct irq_cfg *cfg;
4196 if (!config_enabled(CONFIG_SMP))
4199 cfg = data->chip_data;
4201 irte_info = &cfg->irq_2_irte;
4203 if (!cpumask_intersects(mask, cpu_online_mask))
4206 if (get_irte(irte_info->devid, irte_info->index, &irte))
4209 if (assign_irq_vector(irq, cfg, mask))
4212 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4214 if (assign_irq_vector(irq, cfg, data->affinity))
4215 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4219 irte.fields.vector = cfg->vector;
4220 irte.fields.destination = dest;
4222 modify_irte(irte_info->devid, irte_info->index, irte);
4224 if (cfg->move_in_progress)
4225 send_cleanup_vector(cfg);
4227 cpumask_copy(data->affinity, mask);
4232 static int free_irq(int irq)
4234 struct irq_2_irte *irte_info;
4235 struct irq_cfg *cfg;
4237 cfg = irq_get_chip_data(irq);
4241 irte_info = &cfg->irq_2_irte;
4243 free_irte(irte_info->devid, irte_info->index);
4248 static void compose_msi_msg(struct pci_dev *pdev,
4249 unsigned int irq, unsigned int dest,
4250 struct msi_msg *msg, u8 hpet_id)
4252 struct irq_2_irte *irte_info;
4253 struct irq_cfg *cfg;
4256 cfg = irq_get_chip_data(irq);
4260 irte_info = &cfg->irq_2_irte;
4263 irte.fields.vector = cfg->vector;
4264 irte.fields.int_type = apic->irq_delivery_mode;
4265 irte.fields.destination = dest;
4266 irte.fields.dm = apic->irq_dest_mode;
4267 irte.fields.valid = 1;
4269 modify_irte(irte_info->devid, irte_info->index, irte);
4271 msg->address_hi = MSI_ADDR_BASE_HI;
4272 msg->address_lo = MSI_ADDR_BASE_LO;
4273 msg->data = irte_info->index;
4276 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4278 struct irq_cfg *cfg;
4285 cfg = irq_get_chip_data(irq);
4289 devid = get_device_id(&pdev->dev);
4290 index = alloc_irq_index(cfg, devid, nvec);
4292 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4295 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4296 int index, int offset)
4298 struct irq_2_irte *irte_info;
4299 struct irq_cfg *cfg;
4305 cfg = irq_get_chip_data(irq);
4309 if (index >= MAX_IRQS_PER_TABLE)
4312 devid = get_device_id(&pdev->dev);
4313 irte_info = &cfg->irq_2_irte;
4316 irte_info->devid = devid;
4317 irte_info->index = index + offset;
4322 static int setup_hpet_msi(unsigned int irq, unsigned int id)
4324 struct irq_2_irte *irte_info;
4325 struct irq_cfg *cfg;
4328 cfg = irq_get_chip_data(irq);
4332 irte_info = &cfg->irq_2_irte;
4333 devid = get_hpet_devid(id);
4337 index = alloc_irq_index(cfg, devid, 1);
4342 irte_info->devid = devid;
4343 irte_info->index = index;
4348 struct irq_remap_ops amd_iommu_irq_ops = {
4349 .supported = amd_iommu_supported,
4350 .prepare = amd_iommu_prepare,
4351 .enable = amd_iommu_enable,
4352 .disable = amd_iommu_disable,
4353 .reenable = amd_iommu_reenable,
4354 .enable_faulting = amd_iommu_enable_faulting,
4355 .setup_ioapic_entry = setup_ioapic_entry,
4356 .set_affinity = set_affinity,
4357 .free_irq = free_irq,
4358 .compose_msi_msg = compose_msi_msg,
4359 .msi_alloc_irq = msi_alloc_irq,
4360 .msi_setup_irq = msi_setup_irq,
4361 .setup_hpet_msi = setup_hpet_msi,