2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/io_apic.h>
39 #include <asm/hw_irq.h>
40 #include <asm/msidef.h>
41 #include <asm/proto.h>
42 #include <asm/iommu.h>
46 #include "amd_iommu_proto.h"
47 #include "amd_iommu_types.h"
48 #include "irq_remapping.h"
50 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
52 #define LOOP_TIMEOUT 100000
55 * This bitmap is used to advertise the page sizes our hardware support
56 * to the IOMMU core, which will then use this information to split
57 * physically contiguous memory regions it is mapping into page sizes
60 * 512GB Pages are not supported due to a hardware bug
62 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
64 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
66 /* A list of preallocated protection domains */
67 static LIST_HEAD(iommu_pd_list);
68 static DEFINE_SPINLOCK(iommu_pd_list_lock);
70 /* List of all available dev_data structures */
71 static LIST_HEAD(dev_data_list);
72 static DEFINE_SPINLOCK(dev_data_list_lock);
74 LIST_HEAD(ioapic_map);
78 * Domain for untranslated devices - only allocated
79 * if iommu=pt passed on kernel cmd line.
81 static struct protection_domain *pt_domain;
83 static struct iommu_ops amd_iommu_ops;
85 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
86 int amd_iommu_max_glx_val = -1;
88 static struct dma_map_ops amd_iommu_dma_ops;
91 * general struct to manage commands send to an IOMMU
97 struct kmem_cache *amd_iommu_irq_cache;
99 static void update_domain(struct protection_domain *domain);
100 static int __init alloc_passthrough_domain(void);
102 /****************************************************************************
106 ****************************************************************************/
108 static struct iommu_dev_data *alloc_dev_data(u16 devid)
110 struct iommu_dev_data *dev_data;
113 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
117 dev_data->devid = devid;
118 atomic_set(&dev_data->bind, 0);
120 spin_lock_irqsave(&dev_data_list_lock, flags);
121 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
122 spin_unlock_irqrestore(&dev_data_list_lock, flags);
127 static void free_dev_data(struct iommu_dev_data *dev_data)
131 spin_lock_irqsave(&dev_data_list_lock, flags);
132 list_del(&dev_data->dev_data_list);
133 spin_unlock_irqrestore(&dev_data_list_lock, flags);
136 iommu_group_put(dev_data->group);
141 static struct iommu_dev_data *search_dev_data(u16 devid)
143 struct iommu_dev_data *dev_data;
146 spin_lock_irqsave(&dev_data_list_lock, flags);
147 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
148 if (dev_data->devid == devid)
155 spin_unlock_irqrestore(&dev_data_list_lock, flags);
160 static struct iommu_dev_data *find_dev_data(u16 devid)
162 struct iommu_dev_data *dev_data;
164 dev_data = search_dev_data(devid);
166 if (dev_data == NULL)
167 dev_data = alloc_dev_data(devid);
172 static inline u16 get_device_id(struct device *dev)
174 struct pci_dev *pdev = to_pci_dev(dev);
176 return calc_devid(pdev->bus->number, pdev->devfn);
179 static struct iommu_dev_data *get_dev_data(struct device *dev)
181 return dev->archdata.iommu;
184 static bool pci_iommuv2_capable(struct pci_dev *pdev)
186 static const int caps[] = {
189 PCI_EXT_CAP_ID_PASID,
193 for (i = 0; i < 3; ++i) {
194 pos = pci_find_ext_capability(pdev, caps[i]);
202 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
204 struct iommu_dev_data *dev_data;
206 dev_data = get_dev_data(&pdev->dev);
208 return dev_data->errata & (1 << erratum) ? true : false;
212 * In this function the list of preallocated protection domains is traversed to
213 * find the domain for a specific device
215 static struct dma_ops_domain *find_protection_domain(u16 devid)
217 struct dma_ops_domain *entry, *ret = NULL;
219 u16 alias = amd_iommu_alias_table[devid];
221 if (list_empty(&iommu_pd_list))
224 spin_lock_irqsave(&iommu_pd_list_lock, flags);
226 list_for_each_entry(entry, &iommu_pd_list, list) {
227 if (entry->target_dev == devid ||
228 entry->target_dev == alias) {
234 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
240 * This function checks if the driver got a valid device from the caller to
241 * avoid dereferencing invalid pointers.
243 static bool check_device(struct device *dev)
247 if (!dev || !dev->dma_mask)
250 /* No device or no PCI device */
251 if (dev->bus != &pci_bus_type)
254 devid = get_device_id(dev);
256 /* Out of our scope? */
257 if (devid > amd_iommu_last_bdf)
260 if (amd_iommu_rlookup_table[devid] == NULL)
266 static void swap_pci_ref(struct pci_dev **from, struct pci_dev *to)
272 static struct pci_bus *find_hosted_bus(struct pci_bus *bus)
275 if (!pci_is_root_bus(bus))
278 return ERR_PTR(-ENODEV);
284 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
286 static struct pci_dev *get_isolation_root(struct pci_dev *pdev)
288 struct pci_dev *dma_pdev = pdev;
290 /* Account for quirked devices */
291 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
294 * If it's a multifunction device that does not support our
295 * required ACS flags, add to the same group as function 0.
297 if (dma_pdev->multifunction &&
298 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
299 swap_pci_ref(&dma_pdev,
300 pci_get_slot(dma_pdev->bus,
301 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
305 * Devices on the root bus go through the iommu. If that's not us,
306 * find the next upstream device and test ACS up to the root bus.
307 * Finding the next device may require skipping virtual buses.
309 while (!pci_is_root_bus(dma_pdev->bus)) {
310 struct pci_bus *bus = find_hosted_bus(dma_pdev->bus);
314 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
317 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
323 static int use_pdev_iommu_group(struct pci_dev *pdev, struct device *dev)
325 struct iommu_group *group = iommu_group_get(&pdev->dev);
329 group = iommu_group_alloc();
331 return PTR_ERR(group);
333 WARN_ON(&pdev->dev != dev);
336 ret = iommu_group_add_device(group, dev);
337 iommu_group_put(group);
341 static int use_dev_data_iommu_group(struct iommu_dev_data *dev_data,
344 if (!dev_data->group) {
345 struct iommu_group *group = iommu_group_alloc();
347 return PTR_ERR(group);
349 dev_data->group = group;
352 return iommu_group_add_device(dev_data->group, dev);
355 static int init_iommu_group(struct device *dev)
357 struct iommu_dev_data *dev_data;
358 struct iommu_group *group;
359 struct pci_dev *dma_pdev;
362 group = iommu_group_get(dev);
364 iommu_group_put(group);
368 dev_data = find_dev_data(get_device_id(dev));
372 if (dev_data->alias_data) {
376 if (dev_data->alias_data->group)
380 * If the alias device exists, it's effectively just a first
381 * level quirk for finding the DMA source.
383 alias = amd_iommu_alias_table[dev_data->devid];
384 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
386 dma_pdev = get_isolation_root(dma_pdev);
391 * If the alias is virtual, try to find a parent device
392 * and test whether the IOMMU group is actualy rooted above
393 * the alias. Be careful to also test the parent device if
394 * we think the alias is the root of the group.
396 bus = pci_find_bus(0, alias >> 8);
400 bus = find_hosted_bus(bus);
401 if (IS_ERR(bus) || !bus->self)
404 dma_pdev = get_isolation_root(pci_dev_get(bus->self));
405 if (dma_pdev != bus->self || (dma_pdev->multifunction &&
406 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)))
409 pci_dev_put(dma_pdev);
413 dma_pdev = get_isolation_root(pci_dev_get(to_pci_dev(dev)));
415 ret = use_pdev_iommu_group(dma_pdev, dev);
416 pci_dev_put(dma_pdev);
419 return use_dev_data_iommu_group(dev_data->alias_data, dev);
422 static int iommu_init_device(struct device *dev)
424 struct pci_dev *pdev = to_pci_dev(dev);
425 struct iommu_dev_data *dev_data;
429 if (dev->archdata.iommu)
432 dev_data = find_dev_data(get_device_id(dev));
436 alias = amd_iommu_alias_table[dev_data->devid];
437 if (alias != dev_data->devid) {
438 struct iommu_dev_data *alias_data;
440 alias_data = find_dev_data(alias);
441 if (alias_data == NULL) {
442 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
444 free_dev_data(dev_data);
447 dev_data->alias_data = alias_data;
450 ret = init_iommu_group(dev);
454 if (pci_iommuv2_capable(pdev)) {
455 struct amd_iommu *iommu;
457 iommu = amd_iommu_rlookup_table[dev_data->devid];
458 dev_data->iommu_v2 = iommu->is_iommu_v2;
461 dev->archdata.iommu = dev_data;
466 static void iommu_ignore_device(struct device *dev)
470 devid = get_device_id(dev);
471 alias = amd_iommu_alias_table[devid];
473 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
474 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
476 amd_iommu_rlookup_table[devid] = NULL;
477 amd_iommu_rlookup_table[alias] = NULL;
480 static void iommu_uninit_device(struct device *dev)
482 iommu_group_remove_device(dev);
485 * Nothing to do here - we keep dev_data around for unplugged devices
486 * and reuse it when the device is re-plugged - not doing so would
487 * introduce a ton of races.
491 void __init amd_iommu_uninit_devices(void)
493 struct iommu_dev_data *dev_data, *n;
494 struct pci_dev *pdev = NULL;
496 for_each_pci_dev(pdev) {
498 if (!check_device(&pdev->dev))
501 iommu_uninit_device(&pdev->dev);
504 /* Free all of our dev_data structures */
505 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
506 free_dev_data(dev_data);
509 int __init amd_iommu_init_devices(void)
511 struct pci_dev *pdev = NULL;
514 for_each_pci_dev(pdev) {
516 if (!check_device(&pdev->dev))
519 ret = iommu_init_device(&pdev->dev);
520 if (ret == -ENOTSUPP)
521 iommu_ignore_device(&pdev->dev);
530 amd_iommu_uninit_devices();
534 #ifdef CONFIG_AMD_IOMMU_STATS
537 * Initialization code for statistics collection
540 DECLARE_STATS_COUNTER(compl_wait);
541 DECLARE_STATS_COUNTER(cnt_map_single);
542 DECLARE_STATS_COUNTER(cnt_unmap_single);
543 DECLARE_STATS_COUNTER(cnt_map_sg);
544 DECLARE_STATS_COUNTER(cnt_unmap_sg);
545 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
546 DECLARE_STATS_COUNTER(cnt_free_coherent);
547 DECLARE_STATS_COUNTER(cross_page);
548 DECLARE_STATS_COUNTER(domain_flush_single);
549 DECLARE_STATS_COUNTER(domain_flush_all);
550 DECLARE_STATS_COUNTER(alloced_io_mem);
551 DECLARE_STATS_COUNTER(total_map_requests);
552 DECLARE_STATS_COUNTER(complete_ppr);
553 DECLARE_STATS_COUNTER(invalidate_iotlb);
554 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
555 DECLARE_STATS_COUNTER(pri_requests);
557 static struct dentry *stats_dir;
558 static struct dentry *de_fflush;
560 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
562 if (stats_dir == NULL)
565 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
569 static void amd_iommu_stats_init(void)
571 stats_dir = debugfs_create_dir("amd-iommu", NULL);
572 if (stats_dir == NULL)
575 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
576 &amd_iommu_unmap_flush);
578 amd_iommu_stats_add(&compl_wait);
579 amd_iommu_stats_add(&cnt_map_single);
580 amd_iommu_stats_add(&cnt_unmap_single);
581 amd_iommu_stats_add(&cnt_map_sg);
582 amd_iommu_stats_add(&cnt_unmap_sg);
583 amd_iommu_stats_add(&cnt_alloc_coherent);
584 amd_iommu_stats_add(&cnt_free_coherent);
585 amd_iommu_stats_add(&cross_page);
586 amd_iommu_stats_add(&domain_flush_single);
587 amd_iommu_stats_add(&domain_flush_all);
588 amd_iommu_stats_add(&alloced_io_mem);
589 amd_iommu_stats_add(&total_map_requests);
590 amd_iommu_stats_add(&complete_ppr);
591 amd_iommu_stats_add(&invalidate_iotlb);
592 amd_iommu_stats_add(&invalidate_iotlb_all);
593 amd_iommu_stats_add(&pri_requests);
598 /****************************************************************************
600 * Interrupt handling functions
602 ****************************************************************************/
604 static void dump_dte_entry(u16 devid)
608 for (i = 0; i < 4; ++i)
609 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
610 amd_iommu_dev_table[devid].data[i]);
613 static void dump_command(unsigned long phys_addr)
615 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
618 for (i = 0; i < 4; ++i)
619 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
622 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
624 int type, devid, domid, flags;
625 volatile u32 *event = __evt;
630 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
631 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
632 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
633 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
634 address = (u64)(((u64)event[3]) << 32) | event[2];
637 /* Did we hit the erratum? */
638 if (++count == LOOP_TIMEOUT) {
639 pr_err("AMD-Vi: No event written to event log\n");
646 printk(KERN_ERR "AMD-Vi: Event logged [");
649 case EVENT_TYPE_ILL_DEV:
650 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
651 "address=0x%016llx flags=0x%04x]\n",
652 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
654 dump_dte_entry(devid);
656 case EVENT_TYPE_IO_FAULT:
657 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
658 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
659 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
660 domid, address, flags);
662 case EVENT_TYPE_DEV_TAB_ERR:
663 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
664 "address=0x%016llx flags=0x%04x]\n",
665 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
668 case EVENT_TYPE_PAGE_TAB_ERR:
669 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
670 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
671 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
672 domid, address, flags);
674 case EVENT_TYPE_ILL_CMD:
675 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
676 dump_command(address);
678 case EVENT_TYPE_CMD_HARD_ERR:
679 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
680 "flags=0x%04x]\n", address, flags);
682 case EVENT_TYPE_IOTLB_INV_TO:
683 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
684 "address=0x%016llx]\n",
685 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
688 case EVENT_TYPE_INV_DEV_REQ:
689 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
690 "address=0x%016llx flags=0x%04x]\n",
691 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
695 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
698 memset(__evt, 0, 4 * sizeof(u32));
701 static void iommu_poll_events(struct amd_iommu *iommu)
703 u32 head, tail, status;
706 spin_lock_irqsave(&iommu->lock, flags);
708 /* enable event interrupts again */
711 * Workaround for Erratum ERBT1312
712 * Clearing the EVT_INT bit may race in the hardware, so read
713 * it again and make sure it was really cleared
715 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
716 writel(MMIO_STATUS_EVT_INT_MASK,
717 iommu->mmio_base + MMIO_STATUS_OFFSET);
718 } while (status & MMIO_STATUS_EVT_INT_MASK);
720 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
721 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
723 while (head != tail) {
724 iommu_print_event(iommu, iommu->evt_buf + head);
725 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
728 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
730 spin_unlock_irqrestore(&iommu->lock, flags);
733 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
735 struct amd_iommu_fault fault;
737 INC_STATS_COUNTER(pri_requests);
739 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
740 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
744 fault.address = raw[1];
745 fault.pasid = PPR_PASID(raw[0]);
746 fault.device_id = PPR_DEVID(raw[0]);
747 fault.tag = PPR_TAG(raw[0]);
748 fault.flags = PPR_FLAGS(raw[0]);
750 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
753 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
756 u32 head, tail, status;
758 if (iommu->ppr_log == NULL)
761 spin_lock_irqsave(&iommu->lock, flags);
763 /* enable ppr interrupts again */
766 * Workaround for Erratum ERBT1312
767 * Clearing the PPR_INT bit may race in the hardware, so read
768 * it again and make sure it was really cleared
770 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
771 writel(MMIO_STATUS_PPR_INT_MASK,
772 iommu->mmio_base + MMIO_STATUS_OFFSET);
773 } while (status & MMIO_STATUS_PPR_INT_MASK);
775 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
776 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
778 while (head != tail) {
783 raw = (u64 *)(iommu->ppr_log + head);
786 * Hardware bug: Interrupt may arrive before the entry is
787 * written to memory. If this happens we need to wait for the
790 for (i = 0; i < LOOP_TIMEOUT; ++i) {
791 if (PPR_REQ_TYPE(raw[0]) != 0)
796 /* Avoid memcpy function-call overhead */
801 * To detect the hardware bug we need to clear the entry
804 raw[0] = raw[1] = 0UL;
806 /* Update head pointer of hardware ring-buffer */
807 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
808 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
811 * Release iommu->lock because ppr-handling might need to
814 spin_unlock_irqrestore(&iommu->lock, flags);
816 /* Handle PPR entry */
817 iommu_handle_ppr_entry(iommu, entry);
819 spin_lock_irqsave(&iommu->lock, flags);
821 /* Refresh ring-buffer information */
822 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
823 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
826 spin_unlock_irqrestore(&iommu->lock, flags);
829 irqreturn_t amd_iommu_int_thread(int irq, void *data)
831 struct amd_iommu *iommu;
833 for_each_iommu(iommu) {
834 iommu_poll_events(iommu);
835 iommu_poll_ppr_log(iommu);
841 irqreturn_t amd_iommu_int_handler(int irq, void *data)
843 return IRQ_WAKE_THREAD;
846 /****************************************************************************
848 * IOMMU command queuing functions
850 ****************************************************************************/
852 static int wait_on_sem(volatile u64 *sem)
856 while (*sem == 0 && i < LOOP_TIMEOUT) {
861 if (i == LOOP_TIMEOUT) {
862 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
869 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
870 struct iommu_cmd *cmd,
875 target = iommu->cmd_buf + tail;
876 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
878 /* Copy command to buffer */
879 memcpy(target, cmd, sizeof(*cmd));
881 /* Tell the IOMMU about it */
882 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
885 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
887 WARN_ON(address & 0x7ULL);
889 memset(cmd, 0, sizeof(*cmd));
890 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
891 cmd->data[1] = upper_32_bits(__pa(address));
893 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
896 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
898 memset(cmd, 0, sizeof(*cmd));
899 cmd->data[0] = devid;
900 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
903 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
904 size_t size, u16 domid, int pde)
909 pages = iommu_num_pages(address, size, PAGE_SIZE);
914 * If we have to flush more than one page, flush all
915 * TLB entries for this domain
917 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
921 address &= PAGE_MASK;
923 memset(cmd, 0, sizeof(*cmd));
924 cmd->data[1] |= domid;
925 cmd->data[2] = lower_32_bits(address);
926 cmd->data[3] = upper_32_bits(address);
927 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
928 if (s) /* size bit - we flush more than one 4kb page */
929 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
930 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
931 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
934 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
935 u64 address, size_t size)
940 pages = iommu_num_pages(address, size, PAGE_SIZE);
945 * If we have to flush more than one page, flush all
946 * TLB entries for this domain
948 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
952 address &= PAGE_MASK;
954 memset(cmd, 0, sizeof(*cmd));
955 cmd->data[0] = devid;
956 cmd->data[0] |= (qdep & 0xff) << 24;
957 cmd->data[1] = devid;
958 cmd->data[2] = lower_32_bits(address);
959 cmd->data[3] = upper_32_bits(address);
960 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
962 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
965 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
966 u64 address, bool size)
968 memset(cmd, 0, sizeof(*cmd));
970 address &= ~(0xfffULL);
972 cmd->data[0] = pasid & PASID_MASK;
973 cmd->data[1] = domid;
974 cmd->data[2] = lower_32_bits(address);
975 cmd->data[3] = upper_32_bits(address);
976 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
977 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
979 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
980 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
983 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
984 int qdep, u64 address, bool size)
986 memset(cmd, 0, sizeof(*cmd));
988 address &= ~(0xfffULL);
990 cmd->data[0] = devid;
991 cmd->data[0] |= (pasid & 0xff) << 16;
992 cmd->data[0] |= (qdep & 0xff) << 24;
993 cmd->data[1] = devid;
994 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
995 cmd->data[2] = lower_32_bits(address);
996 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
997 cmd->data[3] = upper_32_bits(address);
999 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1000 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1003 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
1004 int status, int tag, bool gn)
1006 memset(cmd, 0, sizeof(*cmd));
1008 cmd->data[0] = devid;
1010 cmd->data[1] = pasid & PASID_MASK;
1011 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1013 cmd->data[3] = tag & 0x1ff;
1014 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1016 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1019 static void build_inv_all(struct iommu_cmd *cmd)
1021 memset(cmd, 0, sizeof(*cmd));
1022 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1025 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1027 memset(cmd, 0, sizeof(*cmd));
1028 cmd->data[0] = devid;
1029 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1033 * Writes the command to the IOMMUs command buffer and informs the
1034 * hardware about the new command.
1036 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1037 struct iommu_cmd *cmd,
1040 u32 left, tail, head, next_tail;
1041 unsigned long flags;
1043 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
1046 spin_lock_irqsave(&iommu->lock, flags);
1048 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1049 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1050 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
1051 left = (head - next_tail) % iommu->cmd_buf_size;
1054 struct iommu_cmd sync_cmd;
1055 volatile u64 sem = 0;
1058 build_completion_wait(&sync_cmd, (u64)&sem);
1059 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1061 spin_unlock_irqrestore(&iommu->lock, flags);
1063 if ((ret = wait_on_sem(&sem)) != 0)
1069 copy_cmd_to_buffer(iommu, cmd, tail);
1071 /* We need to sync now to make sure all commands are processed */
1072 iommu->need_sync = sync;
1074 spin_unlock_irqrestore(&iommu->lock, flags);
1079 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1081 return iommu_queue_command_sync(iommu, cmd, true);
1085 * This function queues a completion wait command into the command
1086 * buffer of an IOMMU
1088 static int iommu_completion_wait(struct amd_iommu *iommu)
1090 struct iommu_cmd cmd;
1091 volatile u64 sem = 0;
1094 if (!iommu->need_sync)
1097 build_completion_wait(&cmd, (u64)&sem);
1099 ret = iommu_queue_command_sync(iommu, &cmd, false);
1103 return wait_on_sem(&sem);
1106 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1108 struct iommu_cmd cmd;
1110 build_inv_dte(&cmd, devid);
1112 return iommu_queue_command(iommu, &cmd);
1115 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1119 for (devid = 0; devid <= 0xffff; ++devid)
1120 iommu_flush_dte(iommu, devid);
1122 iommu_completion_wait(iommu);
1126 * This function uses heavy locking and may disable irqs for some time. But
1127 * this is no issue because it is only called during resume.
1129 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1133 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1134 struct iommu_cmd cmd;
1135 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1137 iommu_queue_command(iommu, &cmd);
1140 iommu_completion_wait(iommu);
1143 static void iommu_flush_all(struct amd_iommu *iommu)
1145 struct iommu_cmd cmd;
1147 build_inv_all(&cmd);
1149 iommu_queue_command(iommu, &cmd);
1150 iommu_completion_wait(iommu);
1153 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1155 struct iommu_cmd cmd;
1157 build_inv_irt(&cmd, devid);
1159 iommu_queue_command(iommu, &cmd);
1162 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1166 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1167 iommu_flush_irt(iommu, devid);
1169 iommu_completion_wait(iommu);
1172 void iommu_flush_all_caches(struct amd_iommu *iommu)
1174 if (iommu_feature(iommu, FEATURE_IA)) {
1175 iommu_flush_all(iommu);
1177 iommu_flush_dte_all(iommu);
1178 iommu_flush_irt_all(iommu);
1179 iommu_flush_tlb_all(iommu);
1184 * Command send function for flushing on-device TLB
1186 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1187 u64 address, size_t size)
1189 struct amd_iommu *iommu;
1190 struct iommu_cmd cmd;
1193 qdep = dev_data->ats.qdep;
1194 iommu = amd_iommu_rlookup_table[dev_data->devid];
1196 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1198 return iommu_queue_command(iommu, &cmd);
1202 * Command send function for invalidating a device table entry
1204 static int device_flush_dte(struct iommu_dev_data *dev_data)
1206 struct amd_iommu *iommu;
1209 iommu = amd_iommu_rlookup_table[dev_data->devid];
1211 ret = iommu_flush_dte(iommu, dev_data->devid);
1215 if (dev_data->ats.enabled)
1216 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1222 * TLB invalidation function which is called from the mapping functions.
1223 * It invalidates a single PTE if the range to flush is within a single
1224 * page. Otherwise it flushes the whole TLB of the IOMMU.
1226 static void __domain_flush_pages(struct protection_domain *domain,
1227 u64 address, size_t size, int pde)
1229 struct iommu_dev_data *dev_data;
1230 struct iommu_cmd cmd;
1233 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1235 for (i = 0; i < amd_iommus_present; ++i) {
1236 if (!domain->dev_iommu[i])
1240 * Devices of this domain are behind this IOMMU
1241 * We need a TLB flush
1243 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1246 list_for_each_entry(dev_data, &domain->dev_list, list) {
1248 if (!dev_data->ats.enabled)
1251 ret |= device_flush_iotlb(dev_data, address, size);
1257 static void domain_flush_pages(struct protection_domain *domain,
1258 u64 address, size_t size)
1260 __domain_flush_pages(domain, address, size, 0);
1263 /* Flush the whole IO/TLB for a given protection domain */
1264 static void domain_flush_tlb(struct protection_domain *domain)
1266 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1269 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1270 static void domain_flush_tlb_pde(struct protection_domain *domain)
1272 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1275 static void domain_flush_complete(struct protection_domain *domain)
1279 for (i = 0; i < amd_iommus_present; ++i) {
1280 if (!domain->dev_iommu[i])
1284 * Devices of this domain are behind this IOMMU
1285 * We need to wait for completion of all commands.
1287 iommu_completion_wait(amd_iommus[i]);
1293 * This function flushes the DTEs for all devices in domain
1295 static void domain_flush_devices(struct protection_domain *domain)
1297 struct iommu_dev_data *dev_data;
1299 list_for_each_entry(dev_data, &domain->dev_list, list)
1300 device_flush_dte(dev_data);
1303 /****************************************************************************
1305 * The functions below are used the create the page table mappings for
1306 * unity mapped regions.
1308 ****************************************************************************/
1311 * This function is used to add another level to an IO page table. Adding
1312 * another level increases the size of the address space by 9 bits to a size up
1315 static bool increase_address_space(struct protection_domain *domain,
1320 if (domain->mode == PAGE_MODE_6_LEVEL)
1321 /* address space already 64 bit large */
1324 pte = (void *)get_zeroed_page(gfp);
1328 *pte = PM_LEVEL_PDE(domain->mode,
1329 virt_to_phys(domain->pt_root));
1330 domain->pt_root = pte;
1332 domain->updated = true;
1337 static u64 *alloc_pte(struct protection_domain *domain,
1338 unsigned long address,
1339 unsigned long page_size,
1346 BUG_ON(!is_power_of_2(page_size));
1348 while (address > PM_LEVEL_SIZE(domain->mode))
1349 increase_address_space(domain, gfp);
1351 level = domain->mode - 1;
1352 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1353 address = PAGE_SIZE_ALIGN(address, page_size);
1354 end_lvl = PAGE_SIZE_LEVEL(page_size);
1356 while (level > end_lvl) {
1357 if (!IOMMU_PTE_PRESENT(*pte)) {
1358 page = (u64 *)get_zeroed_page(gfp);
1361 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1364 /* No level skipping support yet */
1365 if (PM_PTE_LEVEL(*pte) != level)
1370 pte = IOMMU_PTE_PAGE(*pte);
1372 if (pte_page && level == end_lvl)
1375 pte = &pte[PM_LEVEL_INDEX(level, address)];
1382 * This function checks if there is a PTE for a given dma address. If
1383 * there is one, it returns the pointer to it.
1385 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1390 if (address > PM_LEVEL_SIZE(domain->mode))
1393 level = domain->mode - 1;
1394 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1399 if (!IOMMU_PTE_PRESENT(*pte))
1403 if (PM_PTE_LEVEL(*pte) == 0x07) {
1404 unsigned long pte_mask, __pte;
1407 * If we have a series of large PTEs, make
1408 * sure to return a pointer to the first one.
1410 pte_mask = PTE_PAGE_SIZE(*pte);
1411 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1412 __pte = ((unsigned long)pte) & pte_mask;
1414 return (u64 *)__pte;
1417 /* No level skipping support yet */
1418 if (PM_PTE_LEVEL(*pte) != level)
1423 /* Walk to the next level */
1424 pte = IOMMU_PTE_PAGE(*pte);
1425 pte = &pte[PM_LEVEL_INDEX(level, address)];
1432 * Generic mapping functions. It maps a physical address into a DMA
1433 * address space. It allocates the page table pages if necessary.
1434 * In the future it can be extended to a generic mapping function
1435 * supporting all features of AMD IOMMU page tables like level skipping
1436 * and full 64 bit address spaces.
1438 static int iommu_map_page(struct protection_domain *dom,
1439 unsigned long bus_addr,
1440 unsigned long phys_addr,
1442 unsigned long page_size)
1447 if (!(prot & IOMMU_PROT_MASK))
1450 bus_addr = PAGE_ALIGN(bus_addr);
1451 phys_addr = PAGE_ALIGN(phys_addr);
1452 count = PAGE_SIZE_PTE_COUNT(page_size);
1453 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1455 for (i = 0; i < count; ++i)
1456 if (IOMMU_PTE_PRESENT(pte[i]))
1459 if (page_size > PAGE_SIZE) {
1460 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1461 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1463 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1465 if (prot & IOMMU_PROT_IR)
1466 __pte |= IOMMU_PTE_IR;
1467 if (prot & IOMMU_PROT_IW)
1468 __pte |= IOMMU_PTE_IW;
1470 for (i = 0; i < count; ++i)
1478 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1479 unsigned long bus_addr,
1480 unsigned long page_size)
1482 unsigned long long unmap_size, unmapped;
1485 BUG_ON(!is_power_of_2(page_size));
1489 while (unmapped < page_size) {
1491 pte = fetch_pte(dom, bus_addr);
1495 * No PTE for this address
1496 * move forward in 4kb steps
1498 unmap_size = PAGE_SIZE;
1499 } else if (PM_PTE_LEVEL(*pte) == 0) {
1500 /* 4kb PTE found for this address */
1501 unmap_size = PAGE_SIZE;
1506 /* Large PTE found which maps this address */
1507 unmap_size = PTE_PAGE_SIZE(*pte);
1508 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1509 for (i = 0; i < count; i++)
1513 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1514 unmapped += unmap_size;
1517 BUG_ON(!is_power_of_2(unmapped));
1523 * This function checks if a specific unity mapping entry is needed for
1524 * this specific IOMMU.
1526 static int iommu_for_unity_map(struct amd_iommu *iommu,
1527 struct unity_map_entry *entry)
1531 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1532 bdf = amd_iommu_alias_table[i];
1533 if (amd_iommu_rlookup_table[bdf] == iommu)
1541 * This function actually applies the mapping to the page table of the
1544 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1545 struct unity_map_entry *e)
1550 for (addr = e->address_start; addr < e->address_end;
1551 addr += PAGE_SIZE) {
1552 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1557 * if unity mapping is in aperture range mark the page
1558 * as allocated in the aperture
1560 if (addr < dma_dom->aperture_size)
1561 __set_bit(addr >> PAGE_SHIFT,
1562 dma_dom->aperture[0]->bitmap);
1569 * Init the unity mappings for a specific IOMMU in the system
1571 * Basically iterates over all unity mapping entries and applies them to
1572 * the default domain DMA of that IOMMU if necessary.
1574 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1576 struct unity_map_entry *entry;
1579 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1580 if (!iommu_for_unity_map(iommu, entry))
1582 ret = dma_ops_unity_map(iommu->default_dom, entry);
1591 * Inits the unity mappings required for a specific device
1593 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1596 struct unity_map_entry *e;
1599 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1600 if (!(devid >= e->devid_start && devid <= e->devid_end))
1602 ret = dma_ops_unity_map(dma_dom, e);
1610 /****************************************************************************
1612 * The next functions belong to the address allocator for the dma_ops
1613 * interface functions. They work like the allocators in the other IOMMU
1614 * drivers. Its basically a bitmap which marks the allocated pages in
1615 * the aperture. Maybe it could be enhanced in the future to a more
1616 * efficient allocator.
1618 ****************************************************************************/
1621 * The address allocator core functions.
1623 * called with domain->lock held
1627 * Used to reserve address ranges in the aperture (e.g. for exclusion
1630 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1631 unsigned long start_page,
1634 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1636 if (start_page + pages > last_page)
1637 pages = last_page - start_page;
1639 for (i = start_page; i < start_page + pages; ++i) {
1640 int index = i / APERTURE_RANGE_PAGES;
1641 int page = i % APERTURE_RANGE_PAGES;
1642 __set_bit(page, dom->aperture[index]->bitmap);
1647 * This function is used to add a new aperture range to an existing
1648 * aperture in case of dma_ops domain allocation or address allocation
1651 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1652 bool populate, gfp_t gfp)
1654 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1655 struct amd_iommu *iommu;
1656 unsigned long i, old_size;
1658 #ifdef CONFIG_IOMMU_STRESS
1662 if (index >= APERTURE_MAX_RANGES)
1665 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1666 if (!dma_dom->aperture[index])
1669 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1670 if (!dma_dom->aperture[index]->bitmap)
1673 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1676 unsigned long address = dma_dom->aperture_size;
1677 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1678 u64 *pte, *pte_page;
1680 for (i = 0; i < num_ptes; ++i) {
1681 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1686 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1688 address += APERTURE_RANGE_SIZE / 64;
1692 old_size = dma_dom->aperture_size;
1693 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1695 /* Reserve address range used for MSI messages */
1696 if (old_size < MSI_ADDR_BASE_LO &&
1697 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1698 unsigned long spage;
1701 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1702 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1704 dma_ops_reserve_addresses(dma_dom, spage, pages);
1707 /* Initialize the exclusion range if necessary */
1708 for_each_iommu(iommu) {
1709 if (iommu->exclusion_start &&
1710 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1711 && iommu->exclusion_start < dma_dom->aperture_size) {
1712 unsigned long startpage;
1713 int pages = iommu_num_pages(iommu->exclusion_start,
1714 iommu->exclusion_length,
1716 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1717 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1722 * Check for areas already mapped as present in the new aperture
1723 * range and mark those pages as reserved in the allocator. Such
1724 * mappings may already exist as a result of requested unity
1725 * mappings for devices.
1727 for (i = dma_dom->aperture[index]->offset;
1728 i < dma_dom->aperture_size;
1730 u64 *pte = fetch_pte(&dma_dom->domain, i);
1731 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1734 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1737 update_domain(&dma_dom->domain);
1742 update_domain(&dma_dom->domain);
1744 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1746 kfree(dma_dom->aperture[index]);
1747 dma_dom->aperture[index] = NULL;
1752 static unsigned long dma_ops_area_alloc(struct device *dev,
1753 struct dma_ops_domain *dom,
1755 unsigned long align_mask,
1757 unsigned long start)
1759 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1760 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1761 int i = start >> APERTURE_RANGE_SHIFT;
1762 unsigned long boundary_size;
1763 unsigned long address = -1;
1764 unsigned long limit;
1766 next_bit >>= PAGE_SHIFT;
1768 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1769 PAGE_SIZE) >> PAGE_SHIFT;
1771 for (;i < max_index; ++i) {
1772 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1774 if (dom->aperture[i]->offset >= dma_mask)
1777 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1778 dma_mask >> PAGE_SHIFT);
1780 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1781 limit, next_bit, pages, 0,
1782 boundary_size, align_mask);
1783 if (address != -1) {
1784 address = dom->aperture[i]->offset +
1785 (address << PAGE_SHIFT);
1786 dom->next_address = address + (pages << PAGE_SHIFT);
1796 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1797 struct dma_ops_domain *dom,
1799 unsigned long align_mask,
1802 unsigned long address;
1804 #ifdef CONFIG_IOMMU_STRESS
1805 dom->next_address = 0;
1806 dom->need_flush = true;
1809 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1810 dma_mask, dom->next_address);
1812 if (address == -1) {
1813 dom->next_address = 0;
1814 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1816 dom->need_flush = true;
1819 if (unlikely(address == -1))
1820 address = DMA_ERROR_CODE;
1822 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1828 * The address free function.
1830 * called with domain->lock held
1832 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1833 unsigned long address,
1836 unsigned i = address >> APERTURE_RANGE_SHIFT;
1837 struct aperture_range *range = dom->aperture[i];
1839 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1841 #ifdef CONFIG_IOMMU_STRESS
1846 if (address >= dom->next_address)
1847 dom->need_flush = true;
1849 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1851 bitmap_clear(range->bitmap, address, pages);
1855 /****************************************************************************
1857 * The next functions belong to the domain allocation. A domain is
1858 * allocated for every IOMMU as the default domain. If device isolation
1859 * is enabled, every device get its own domain. The most important thing
1860 * about domains is the page table mapping the DMA address space they
1863 ****************************************************************************/
1866 * This function adds a protection domain to the global protection domain list
1868 static void add_domain_to_list(struct protection_domain *domain)
1870 unsigned long flags;
1872 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1873 list_add(&domain->list, &amd_iommu_pd_list);
1874 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1878 * This function removes a protection domain to the global
1879 * protection domain list
1881 static void del_domain_from_list(struct protection_domain *domain)
1883 unsigned long flags;
1885 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1886 list_del(&domain->list);
1887 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1890 static u16 domain_id_alloc(void)
1892 unsigned long flags;
1895 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1896 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1898 if (id > 0 && id < MAX_DOMAIN_ID)
1899 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1902 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1907 static void domain_id_free(int id)
1909 unsigned long flags;
1911 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1912 if (id > 0 && id < MAX_DOMAIN_ID)
1913 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1914 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1917 static void free_pagetable(struct protection_domain *domain)
1922 p1 = domain->pt_root;
1927 for (i = 0; i < 512; ++i) {
1928 if (!IOMMU_PTE_PRESENT(p1[i]))
1931 p2 = IOMMU_PTE_PAGE(p1[i]);
1932 for (j = 0; j < 512; ++j) {
1933 if (!IOMMU_PTE_PRESENT(p2[j]))
1935 p3 = IOMMU_PTE_PAGE(p2[j]);
1936 free_page((unsigned long)p3);
1939 free_page((unsigned long)p2);
1942 free_page((unsigned long)p1);
1944 domain->pt_root = NULL;
1947 static void free_gcr3_tbl_level1(u64 *tbl)
1952 for (i = 0; i < 512; ++i) {
1953 if (!(tbl[i] & GCR3_VALID))
1956 ptr = __va(tbl[i] & PAGE_MASK);
1958 free_page((unsigned long)ptr);
1962 static void free_gcr3_tbl_level2(u64 *tbl)
1967 for (i = 0; i < 512; ++i) {
1968 if (!(tbl[i] & GCR3_VALID))
1971 ptr = __va(tbl[i] & PAGE_MASK);
1973 free_gcr3_tbl_level1(ptr);
1977 static void free_gcr3_table(struct protection_domain *domain)
1979 if (domain->glx == 2)
1980 free_gcr3_tbl_level2(domain->gcr3_tbl);
1981 else if (domain->glx == 1)
1982 free_gcr3_tbl_level1(domain->gcr3_tbl);
1983 else if (domain->glx != 0)
1986 free_page((unsigned long)domain->gcr3_tbl);
1990 * Free a domain, only used if something went wrong in the
1991 * allocation path and we need to free an already allocated page table
1993 static void dma_ops_domain_free(struct dma_ops_domain *dom)
2000 del_domain_from_list(&dom->domain);
2002 free_pagetable(&dom->domain);
2004 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
2005 if (!dom->aperture[i])
2007 free_page((unsigned long)dom->aperture[i]->bitmap);
2008 kfree(dom->aperture[i]);
2015 * Allocates a new protection domain usable for the dma_ops functions.
2016 * It also initializes the page table and the address allocator data
2017 * structures required for the dma_ops interface
2019 static struct dma_ops_domain *dma_ops_domain_alloc(void)
2021 struct dma_ops_domain *dma_dom;
2023 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2027 spin_lock_init(&dma_dom->domain.lock);
2029 dma_dom->domain.id = domain_id_alloc();
2030 if (dma_dom->domain.id == 0)
2032 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
2033 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2034 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2035 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2036 dma_dom->domain.priv = dma_dom;
2037 if (!dma_dom->domain.pt_root)
2040 dma_dom->need_flush = false;
2041 dma_dom->target_dev = 0xffff;
2043 add_domain_to_list(&dma_dom->domain);
2045 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2049 * mark the first page as allocated so we never return 0 as
2050 * a valid dma-address. So we can use 0 as error value
2052 dma_dom->aperture[0]->bitmap[0] = 1;
2053 dma_dom->next_address = 0;
2059 dma_ops_domain_free(dma_dom);
2065 * little helper function to check whether a given protection domain is a
2068 static bool dma_ops_domain(struct protection_domain *domain)
2070 return domain->flags & PD_DMA_OPS_MASK;
2073 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2078 if (domain->mode != PAGE_MODE_NONE)
2079 pte_root = virt_to_phys(domain->pt_root);
2081 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2082 << DEV_ENTRY_MODE_SHIFT;
2083 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2085 flags = amd_iommu_dev_table[devid].data[1];
2088 flags |= DTE_FLAG_IOTLB;
2090 if (domain->flags & PD_IOMMUV2_MASK) {
2091 u64 gcr3 = __pa(domain->gcr3_tbl);
2092 u64 glx = domain->glx;
2095 pte_root |= DTE_FLAG_GV;
2096 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2098 /* First mask out possible old values for GCR3 table */
2099 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2102 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2105 /* Encode GCR3 table into DTE */
2106 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2109 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2112 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2116 flags &= ~(0xffffUL);
2117 flags |= domain->id;
2119 amd_iommu_dev_table[devid].data[1] = flags;
2120 amd_iommu_dev_table[devid].data[0] = pte_root;
2123 static void clear_dte_entry(u16 devid)
2125 /* remove entry from the device table seen by the hardware */
2126 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2127 amd_iommu_dev_table[devid].data[1] = 0;
2129 amd_iommu_apply_erratum_63(devid);
2132 static void do_attach(struct iommu_dev_data *dev_data,
2133 struct protection_domain *domain)
2135 struct amd_iommu *iommu;
2138 iommu = amd_iommu_rlookup_table[dev_data->devid];
2139 ats = dev_data->ats.enabled;
2141 /* Update data structures */
2142 dev_data->domain = domain;
2143 list_add(&dev_data->list, &domain->dev_list);
2144 set_dte_entry(dev_data->devid, domain, ats);
2146 /* Do reference counting */
2147 domain->dev_iommu[iommu->index] += 1;
2148 domain->dev_cnt += 1;
2150 /* Flush the DTE entry */
2151 device_flush_dte(dev_data);
2154 static void do_detach(struct iommu_dev_data *dev_data)
2156 struct amd_iommu *iommu;
2158 iommu = amd_iommu_rlookup_table[dev_data->devid];
2160 /* decrease reference counters */
2161 dev_data->domain->dev_iommu[iommu->index] -= 1;
2162 dev_data->domain->dev_cnt -= 1;
2164 /* Update data structures */
2165 dev_data->domain = NULL;
2166 list_del(&dev_data->list);
2167 clear_dte_entry(dev_data->devid);
2169 /* Flush the DTE entry */
2170 device_flush_dte(dev_data);
2174 * If a device is not yet associated with a domain, this function does
2175 * assigns it visible for the hardware
2177 static int __attach_device(struct iommu_dev_data *dev_data,
2178 struct protection_domain *domain)
2183 spin_lock(&domain->lock);
2185 if (dev_data->alias_data != NULL) {
2186 struct iommu_dev_data *alias_data = dev_data->alias_data;
2188 /* Some sanity checks */
2190 if (alias_data->domain != NULL &&
2191 alias_data->domain != domain)
2194 if (dev_data->domain != NULL &&
2195 dev_data->domain != domain)
2198 /* Do real assignment */
2199 if (alias_data->domain == NULL)
2200 do_attach(alias_data, domain);
2202 atomic_inc(&alias_data->bind);
2205 if (dev_data->domain == NULL)
2206 do_attach(dev_data, domain);
2208 atomic_inc(&dev_data->bind);
2215 spin_unlock(&domain->lock);
2221 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2223 pci_disable_ats(pdev);
2224 pci_disable_pri(pdev);
2225 pci_disable_pasid(pdev);
2228 /* FIXME: Change generic reset-function to do the same */
2229 static int pri_reset_while_enabled(struct pci_dev *pdev)
2234 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2238 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2239 control |= PCI_PRI_CTRL_RESET;
2240 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2245 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2250 /* FIXME: Hardcode number of outstanding requests for now */
2252 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2254 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2256 /* Only allow access to user-accessible pages */
2257 ret = pci_enable_pasid(pdev, 0);
2261 /* First reset the PRI state of the device */
2262 ret = pci_reset_pri(pdev);
2267 ret = pci_enable_pri(pdev, reqs);
2272 ret = pri_reset_while_enabled(pdev);
2277 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2284 pci_disable_pri(pdev);
2285 pci_disable_pasid(pdev);
2290 /* FIXME: Move this to PCI code */
2291 #define PCI_PRI_TLP_OFF (1 << 15)
2293 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2298 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2302 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2304 return (status & PCI_PRI_TLP_OFF) ? true : false;
2308 * If a device is not yet associated with a domain, this function
2309 * assigns it visible for the hardware
2311 static int attach_device(struct device *dev,
2312 struct protection_domain *domain)
2314 struct pci_dev *pdev = to_pci_dev(dev);
2315 struct iommu_dev_data *dev_data;
2316 unsigned long flags;
2319 dev_data = get_dev_data(dev);
2321 if (domain->flags & PD_IOMMUV2_MASK) {
2322 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2325 if (pdev_iommuv2_enable(pdev) != 0)
2328 dev_data->ats.enabled = true;
2329 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2330 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2331 } else if (amd_iommu_iotlb_sup &&
2332 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2333 dev_data->ats.enabled = true;
2334 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2337 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2338 ret = __attach_device(dev_data, domain);
2339 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2342 * We might boot into a crash-kernel here. The crashed kernel
2343 * left the caches in the IOMMU dirty. So we have to flush
2344 * here to evict all dirty stuff.
2346 domain_flush_tlb_pde(domain);
2352 * Removes a device from a protection domain (unlocked)
2354 static void __detach_device(struct iommu_dev_data *dev_data)
2356 struct protection_domain *domain;
2357 unsigned long flags;
2359 BUG_ON(!dev_data->domain);
2361 domain = dev_data->domain;
2363 spin_lock_irqsave(&domain->lock, flags);
2365 if (dev_data->alias_data != NULL) {
2366 struct iommu_dev_data *alias_data = dev_data->alias_data;
2368 if (atomic_dec_and_test(&alias_data->bind))
2369 do_detach(alias_data);
2372 if (atomic_dec_and_test(&dev_data->bind))
2373 do_detach(dev_data);
2375 spin_unlock_irqrestore(&domain->lock, flags);
2378 * If we run in passthrough mode the device must be assigned to the
2379 * passthrough domain if it is detached from any other domain.
2380 * Make sure we can deassign from the pt_domain itself.
2382 if (dev_data->passthrough &&
2383 (dev_data->domain == NULL && domain != pt_domain))
2384 __attach_device(dev_data, pt_domain);
2388 * Removes a device from a protection domain (with devtable_lock held)
2390 static void detach_device(struct device *dev)
2392 struct protection_domain *domain;
2393 struct iommu_dev_data *dev_data;
2394 unsigned long flags;
2396 dev_data = get_dev_data(dev);
2397 domain = dev_data->domain;
2399 /* lock device table */
2400 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2401 __detach_device(dev_data);
2402 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2404 if (domain->flags & PD_IOMMUV2_MASK)
2405 pdev_iommuv2_disable(to_pci_dev(dev));
2406 else if (dev_data->ats.enabled)
2407 pci_disable_ats(to_pci_dev(dev));
2409 dev_data->ats.enabled = false;
2413 * Find out the protection domain structure for a given PCI device. This
2414 * will give us the pointer to the page table root for example.
2416 static struct protection_domain *domain_for_device(struct device *dev)
2418 struct iommu_dev_data *dev_data;
2419 struct protection_domain *dom = NULL;
2420 unsigned long flags;
2422 dev_data = get_dev_data(dev);
2424 if (dev_data->domain)
2425 return dev_data->domain;
2427 if (dev_data->alias_data != NULL) {
2428 struct iommu_dev_data *alias_data = dev_data->alias_data;
2430 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2431 if (alias_data->domain != NULL) {
2432 __attach_device(dev_data, alias_data->domain);
2433 dom = alias_data->domain;
2435 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2441 static int device_change_notifier(struct notifier_block *nb,
2442 unsigned long action, void *data)
2444 struct dma_ops_domain *dma_domain;
2445 struct protection_domain *domain;
2446 struct iommu_dev_data *dev_data;
2447 struct device *dev = data;
2448 struct amd_iommu *iommu;
2449 unsigned long flags;
2452 if (!check_device(dev))
2455 devid = get_device_id(dev);
2456 iommu = amd_iommu_rlookup_table[devid];
2457 dev_data = get_dev_data(dev);
2460 case BUS_NOTIFY_UNBOUND_DRIVER:
2462 domain = domain_for_device(dev);
2466 if (dev_data->passthrough)
2470 case BUS_NOTIFY_ADD_DEVICE:
2472 iommu_init_device(dev);
2475 * dev_data is still NULL and
2476 * got initialized in iommu_init_device
2478 dev_data = get_dev_data(dev);
2480 if (iommu_pass_through || dev_data->iommu_v2) {
2481 dev_data->passthrough = true;
2482 attach_device(dev, pt_domain);
2486 domain = domain_for_device(dev);
2488 /* allocate a protection domain if a device is added */
2489 dma_domain = find_protection_domain(devid);
2491 dma_domain = dma_ops_domain_alloc();
2494 dma_domain->target_dev = devid;
2496 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2497 list_add_tail(&dma_domain->list, &iommu_pd_list);
2498 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2501 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2504 case BUS_NOTIFY_DEL_DEVICE:
2506 iommu_uninit_device(dev);
2512 iommu_completion_wait(iommu);
2518 static struct notifier_block device_nb = {
2519 .notifier_call = device_change_notifier,
2522 void amd_iommu_init_notifier(void)
2524 bus_register_notifier(&pci_bus_type, &device_nb);
2527 /*****************************************************************************
2529 * The next functions belong to the dma_ops mapping/unmapping code.
2531 *****************************************************************************/
2534 * In the dma_ops path we only have the struct device. This function
2535 * finds the corresponding IOMMU, the protection domain and the
2536 * requestor id for a given device.
2537 * If the device is not yet associated with a domain this is also done
2540 static struct protection_domain *get_domain(struct device *dev)
2542 struct protection_domain *domain;
2543 struct dma_ops_domain *dma_dom;
2544 u16 devid = get_device_id(dev);
2546 if (!check_device(dev))
2547 return ERR_PTR(-EINVAL);
2549 domain = domain_for_device(dev);
2550 if (domain != NULL && !dma_ops_domain(domain))
2551 return ERR_PTR(-EBUSY);
2556 /* Device not bound yet - bind it */
2557 dma_dom = find_protection_domain(devid);
2559 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2560 attach_device(dev, &dma_dom->domain);
2561 DUMP_printk("Using protection domain %d for device %s\n",
2562 dma_dom->domain.id, dev_name(dev));
2564 return &dma_dom->domain;
2567 static void update_device_table(struct protection_domain *domain)
2569 struct iommu_dev_data *dev_data;
2571 list_for_each_entry(dev_data, &domain->dev_list, list)
2572 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2575 static void update_domain(struct protection_domain *domain)
2577 if (!domain->updated)
2580 update_device_table(domain);
2582 domain_flush_devices(domain);
2583 domain_flush_tlb_pde(domain);
2585 domain->updated = false;
2589 * This function fetches the PTE for a given address in the aperture
2591 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2592 unsigned long address)
2594 struct aperture_range *aperture;
2595 u64 *pte, *pte_page;
2597 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2601 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2603 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2605 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2607 pte += PM_LEVEL_INDEX(0, address);
2609 update_domain(&dom->domain);
2615 * This is the generic map function. It maps one 4kb page at paddr to
2616 * the given address in the DMA address space for the domain.
2618 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2619 unsigned long address,
2625 WARN_ON(address > dom->aperture_size);
2629 pte = dma_ops_get_pte(dom, address);
2631 return DMA_ERROR_CODE;
2633 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2635 if (direction == DMA_TO_DEVICE)
2636 __pte |= IOMMU_PTE_IR;
2637 else if (direction == DMA_FROM_DEVICE)
2638 __pte |= IOMMU_PTE_IW;
2639 else if (direction == DMA_BIDIRECTIONAL)
2640 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2646 return (dma_addr_t)address;
2650 * The generic unmapping function for on page in the DMA address space.
2652 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2653 unsigned long address)
2655 struct aperture_range *aperture;
2658 if (address >= dom->aperture_size)
2661 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2665 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2669 pte += PM_LEVEL_INDEX(0, address);
2677 * This function contains common code for mapping of a physically
2678 * contiguous memory region into DMA address space. It is used by all
2679 * mapping functions provided with this IOMMU driver.
2680 * Must be called with the domain lock held.
2682 static dma_addr_t __map_single(struct device *dev,
2683 struct dma_ops_domain *dma_dom,
2690 dma_addr_t offset = paddr & ~PAGE_MASK;
2691 dma_addr_t address, start, ret;
2693 unsigned long align_mask = 0;
2696 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2699 INC_STATS_COUNTER(total_map_requests);
2702 INC_STATS_COUNTER(cross_page);
2705 align_mask = (1UL << get_order(size)) - 1;
2708 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2710 if (unlikely(address == DMA_ERROR_CODE)) {
2712 * setting next_address here will let the address
2713 * allocator only scan the new allocated range in the
2714 * first run. This is a small optimization.
2716 dma_dom->next_address = dma_dom->aperture_size;
2718 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2722 * aperture was successfully enlarged by 128 MB, try
2729 for (i = 0; i < pages; ++i) {
2730 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2731 if (ret == DMA_ERROR_CODE)
2739 ADD_STATS_COUNTER(alloced_io_mem, size);
2741 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2742 domain_flush_tlb(&dma_dom->domain);
2743 dma_dom->need_flush = false;
2744 } else if (unlikely(amd_iommu_np_cache))
2745 domain_flush_pages(&dma_dom->domain, address, size);
2752 for (--i; i >= 0; --i) {
2754 dma_ops_domain_unmap(dma_dom, start);
2757 dma_ops_free_addresses(dma_dom, address, pages);
2759 return DMA_ERROR_CODE;
2763 * Does the reverse of the __map_single function. Must be called with
2764 * the domain lock held too
2766 static void __unmap_single(struct dma_ops_domain *dma_dom,
2767 dma_addr_t dma_addr,
2771 dma_addr_t flush_addr;
2772 dma_addr_t i, start;
2775 if ((dma_addr == DMA_ERROR_CODE) ||
2776 (dma_addr + size > dma_dom->aperture_size))
2779 flush_addr = dma_addr;
2780 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2781 dma_addr &= PAGE_MASK;
2784 for (i = 0; i < pages; ++i) {
2785 dma_ops_domain_unmap(dma_dom, start);
2789 SUB_STATS_COUNTER(alloced_io_mem, size);
2791 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2793 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2794 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2795 dma_dom->need_flush = false;
2800 * The exported map_single function for dma_ops.
2802 static dma_addr_t map_page(struct device *dev, struct page *page,
2803 unsigned long offset, size_t size,
2804 enum dma_data_direction dir,
2805 struct dma_attrs *attrs)
2807 unsigned long flags;
2808 struct protection_domain *domain;
2811 phys_addr_t paddr = page_to_phys(page) + offset;
2813 INC_STATS_COUNTER(cnt_map_single);
2815 domain = get_domain(dev);
2816 if (PTR_ERR(domain) == -EINVAL)
2817 return (dma_addr_t)paddr;
2818 else if (IS_ERR(domain))
2819 return DMA_ERROR_CODE;
2821 dma_mask = *dev->dma_mask;
2823 spin_lock_irqsave(&domain->lock, flags);
2825 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2827 if (addr == DMA_ERROR_CODE)
2830 domain_flush_complete(domain);
2833 spin_unlock_irqrestore(&domain->lock, flags);
2839 * The exported unmap_single function for dma_ops.
2841 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2842 enum dma_data_direction dir, struct dma_attrs *attrs)
2844 unsigned long flags;
2845 struct protection_domain *domain;
2847 INC_STATS_COUNTER(cnt_unmap_single);
2849 domain = get_domain(dev);
2853 spin_lock_irqsave(&domain->lock, flags);
2855 __unmap_single(domain->priv, dma_addr, size, dir);
2857 domain_flush_complete(domain);
2859 spin_unlock_irqrestore(&domain->lock, flags);
2863 * The exported map_sg function for dma_ops (handles scatter-gather
2866 static int map_sg(struct device *dev, struct scatterlist *sglist,
2867 int nelems, enum dma_data_direction dir,
2868 struct dma_attrs *attrs)
2870 unsigned long flags;
2871 struct protection_domain *domain;
2873 struct scatterlist *s;
2875 int mapped_elems = 0;
2878 INC_STATS_COUNTER(cnt_map_sg);
2880 domain = get_domain(dev);
2884 dma_mask = *dev->dma_mask;
2886 spin_lock_irqsave(&domain->lock, flags);
2888 for_each_sg(sglist, s, nelems, i) {
2891 s->dma_address = __map_single(dev, domain->priv,
2892 paddr, s->length, dir, false,
2895 if (s->dma_address) {
2896 s->dma_length = s->length;
2902 domain_flush_complete(domain);
2905 spin_unlock_irqrestore(&domain->lock, flags);
2907 return mapped_elems;
2909 for_each_sg(sglist, s, mapped_elems, i) {
2911 __unmap_single(domain->priv, s->dma_address,
2912 s->dma_length, dir);
2913 s->dma_address = s->dma_length = 0;
2922 * The exported map_sg function for dma_ops (handles scatter-gather
2925 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2926 int nelems, enum dma_data_direction dir,
2927 struct dma_attrs *attrs)
2929 unsigned long flags;
2930 struct protection_domain *domain;
2931 struct scatterlist *s;
2934 INC_STATS_COUNTER(cnt_unmap_sg);
2936 domain = get_domain(dev);
2940 spin_lock_irqsave(&domain->lock, flags);
2942 for_each_sg(sglist, s, nelems, i) {
2943 __unmap_single(domain->priv, s->dma_address,
2944 s->dma_length, dir);
2945 s->dma_address = s->dma_length = 0;
2948 domain_flush_complete(domain);
2950 spin_unlock_irqrestore(&domain->lock, flags);
2954 * The exported alloc_coherent function for dma_ops.
2956 static void *alloc_coherent(struct device *dev, size_t size,
2957 dma_addr_t *dma_addr, gfp_t flag,
2958 struct dma_attrs *attrs)
2960 unsigned long flags;
2962 struct protection_domain *domain;
2964 u64 dma_mask = dev->coherent_dma_mask;
2966 INC_STATS_COUNTER(cnt_alloc_coherent);
2968 domain = get_domain(dev);
2969 if (PTR_ERR(domain) == -EINVAL) {
2970 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2971 *dma_addr = __pa(virt_addr);
2973 } else if (IS_ERR(domain))
2976 dma_mask = dev->coherent_dma_mask;
2977 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2980 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2984 paddr = virt_to_phys(virt_addr);
2987 dma_mask = *dev->dma_mask;
2989 spin_lock_irqsave(&domain->lock, flags);
2991 *dma_addr = __map_single(dev, domain->priv, paddr,
2992 size, DMA_BIDIRECTIONAL, true, dma_mask);
2994 if (*dma_addr == DMA_ERROR_CODE) {
2995 spin_unlock_irqrestore(&domain->lock, flags);
2999 domain_flush_complete(domain);
3001 spin_unlock_irqrestore(&domain->lock, flags);
3007 free_pages((unsigned long)virt_addr, get_order(size));
3013 * The exported free_coherent function for dma_ops.
3015 static void free_coherent(struct device *dev, size_t size,
3016 void *virt_addr, dma_addr_t dma_addr,
3017 struct dma_attrs *attrs)
3019 unsigned long flags;
3020 struct protection_domain *domain;
3022 INC_STATS_COUNTER(cnt_free_coherent);
3024 domain = get_domain(dev);
3028 spin_lock_irqsave(&domain->lock, flags);
3030 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
3032 domain_flush_complete(domain);
3034 spin_unlock_irqrestore(&domain->lock, flags);
3037 free_pages((unsigned long)virt_addr, get_order(size));
3041 * This function is called by the DMA layer to find out if we can handle a
3042 * particular device. It is part of the dma_ops.
3044 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3046 return check_device(dev);
3050 * The function for pre-allocating protection domains.
3052 * If the driver core informs the DMA layer if a driver grabs a device
3053 * we don't need to preallocate the protection domains anymore.
3054 * For now we have to.
3056 static void __init prealloc_protection_domains(void)
3058 struct iommu_dev_data *dev_data;
3059 struct dma_ops_domain *dma_dom;
3060 struct pci_dev *dev = NULL;
3063 for_each_pci_dev(dev) {
3065 /* Do we handle this device? */
3066 if (!check_device(&dev->dev))
3069 dev_data = get_dev_data(&dev->dev);
3070 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3071 /* Make sure passthrough domain is allocated */
3072 alloc_passthrough_domain();
3073 dev_data->passthrough = true;
3074 attach_device(&dev->dev, pt_domain);
3075 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3076 dev_name(&dev->dev));
3079 /* Is there already any domain for it? */
3080 if (domain_for_device(&dev->dev))
3083 devid = get_device_id(&dev->dev);
3085 dma_dom = dma_ops_domain_alloc();
3088 init_unity_mappings_for_device(dma_dom, devid);
3089 dma_dom->target_dev = devid;
3091 attach_device(&dev->dev, &dma_dom->domain);
3093 list_add_tail(&dma_dom->list, &iommu_pd_list);
3097 static struct dma_map_ops amd_iommu_dma_ops = {
3098 .alloc = alloc_coherent,
3099 .free = free_coherent,
3100 .map_page = map_page,
3101 .unmap_page = unmap_page,
3103 .unmap_sg = unmap_sg,
3104 .dma_supported = amd_iommu_dma_supported,
3107 static unsigned device_dma_ops_init(void)
3109 struct iommu_dev_data *dev_data;
3110 struct pci_dev *pdev = NULL;
3111 unsigned unhandled = 0;
3113 for_each_pci_dev(pdev) {
3114 if (!check_device(&pdev->dev)) {
3116 iommu_ignore_device(&pdev->dev);
3122 dev_data = get_dev_data(&pdev->dev);
3124 if (!dev_data->passthrough)
3125 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3127 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3134 * The function which clues the AMD IOMMU driver into dma_ops.
3137 void __init amd_iommu_init_api(void)
3139 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3142 int __init amd_iommu_init_dma_ops(void)
3144 struct amd_iommu *iommu;
3148 * first allocate a default protection domain for every IOMMU we
3149 * found in the system. Devices not assigned to any other
3150 * protection domain will be assigned to the default one.
3152 for_each_iommu(iommu) {
3153 iommu->default_dom = dma_ops_domain_alloc();
3154 if (iommu->default_dom == NULL)
3156 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3157 ret = iommu_init_unity_mappings(iommu);
3163 * Pre-allocate the protection domains for each device.
3165 prealloc_protection_domains();
3170 /* Make the driver finally visible to the drivers */
3171 unhandled = device_dma_ops_init();
3172 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3173 /* There are unhandled devices - initialize swiotlb for them */
3177 amd_iommu_stats_init();
3179 if (amd_iommu_unmap_flush)
3180 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3182 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3188 for_each_iommu(iommu) {
3189 dma_ops_domain_free(iommu->default_dom);
3195 /*****************************************************************************
3197 * The following functions belong to the exported interface of AMD IOMMU
3199 * This interface allows access to lower level functions of the IOMMU
3200 * like protection domain handling and assignement of devices to domains
3201 * which is not possible with the dma_ops interface.
3203 *****************************************************************************/
3205 static void cleanup_domain(struct protection_domain *domain)
3207 struct iommu_dev_data *dev_data, *next;
3208 unsigned long flags;
3210 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3212 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3213 __detach_device(dev_data);
3214 atomic_set(&dev_data->bind, 0);
3217 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3220 static void protection_domain_free(struct protection_domain *domain)
3225 del_domain_from_list(domain);
3228 domain_id_free(domain->id);
3233 static struct protection_domain *protection_domain_alloc(void)
3235 struct protection_domain *domain;
3237 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3241 spin_lock_init(&domain->lock);
3242 mutex_init(&domain->api_lock);
3243 domain->id = domain_id_alloc();
3246 INIT_LIST_HEAD(&domain->dev_list);
3248 add_domain_to_list(domain);
3258 static int __init alloc_passthrough_domain(void)
3260 if (pt_domain != NULL)
3263 /* allocate passthrough domain */
3264 pt_domain = protection_domain_alloc();
3268 pt_domain->mode = PAGE_MODE_NONE;
3272 static int amd_iommu_domain_init(struct iommu_domain *dom)
3274 struct protection_domain *domain;
3276 domain = protection_domain_alloc();
3280 domain->mode = PAGE_MODE_3_LEVEL;
3281 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3282 if (!domain->pt_root)
3285 domain->iommu_domain = dom;
3289 dom->geometry.aperture_start = 0;
3290 dom->geometry.aperture_end = ~0ULL;
3291 dom->geometry.force_aperture = true;
3296 protection_domain_free(domain);
3301 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3303 struct protection_domain *domain = dom->priv;
3308 if (domain->dev_cnt > 0)
3309 cleanup_domain(domain);
3311 BUG_ON(domain->dev_cnt != 0);
3313 if (domain->mode != PAGE_MODE_NONE)
3314 free_pagetable(domain);
3316 if (domain->flags & PD_IOMMUV2_MASK)
3317 free_gcr3_table(domain);
3319 protection_domain_free(domain);
3324 static void amd_iommu_detach_device(struct iommu_domain *dom,
3327 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3328 struct amd_iommu *iommu;
3331 if (!check_device(dev))
3334 devid = get_device_id(dev);
3336 if (dev_data->domain != NULL)
3339 iommu = amd_iommu_rlookup_table[devid];
3343 iommu_completion_wait(iommu);
3346 static int amd_iommu_attach_device(struct iommu_domain *dom,
3349 struct protection_domain *domain = dom->priv;
3350 struct iommu_dev_data *dev_data;
3351 struct amd_iommu *iommu;
3354 if (!check_device(dev))
3357 dev_data = dev->archdata.iommu;
3359 iommu = amd_iommu_rlookup_table[dev_data->devid];
3363 if (dev_data->domain)
3366 ret = attach_device(dev, domain);
3368 iommu_completion_wait(iommu);
3373 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3374 phys_addr_t paddr, size_t page_size, int iommu_prot)
3376 struct protection_domain *domain = dom->priv;
3380 if (domain->mode == PAGE_MODE_NONE)
3383 if (iommu_prot & IOMMU_READ)
3384 prot |= IOMMU_PROT_IR;
3385 if (iommu_prot & IOMMU_WRITE)
3386 prot |= IOMMU_PROT_IW;
3388 mutex_lock(&domain->api_lock);
3389 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3390 mutex_unlock(&domain->api_lock);
3395 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3398 struct protection_domain *domain = dom->priv;
3401 if (domain->mode == PAGE_MODE_NONE)
3404 mutex_lock(&domain->api_lock);
3405 unmap_size = iommu_unmap_page(domain, iova, page_size);
3406 mutex_unlock(&domain->api_lock);
3408 domain_flush_tlb_pde(domain);
3413 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3416 struct protection_domain *domain = dom->priv;
3417 unsigned long offset_mask;
3421 if (domain->mode == PAGE_MODE_NONE)
3424 pte = fetch_pte(domain, iova);
3426 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3429 if (PM_PTE_LEVEL(*pte) == 0)
3430 offset_mask = PAGE_SIZE - 1;
3432 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3434 __pte = *pte & PM_ADDR_MASK;
3435 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3440 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3444 case IOMMU_CAP_CACHE_COHERENCY:
3446 case IOMMU_CAP_INTR_REMAP:
3447 return irq_remapping_enabled;
3453 static struct iommu_ops amd_iommu_ops = {
3454 .domain_init = amd_iommu_domain_init,
3455 .domain_destroy = amd_iommu_domain_destroy,
3456 .attach_dev = amd_iommu_attach_device,
3457 .detach_dev = amd_iommu_detach_device,
3458 .map = amd_iommu_map,
3459 .unmap = amd_iommu_unmap,
3460 .iova_to_phys = amd_iommu_iova_to_phys,
3461 .domain_has_cap = amd_iommu_domain_has_cap,
3462 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3465 /*****************************************************************************
3467 * The next functions do a basic initialization of IOMMU for pass through
3470 * In passthrough mode the IOMMU is initialized and enabled but not used for
3471 * DMA-API translation.
3473 *****************************************************************************/
3475 int __init amd_iommu_init_passthrough(void)
3477 struct iommu_dev_data *dev_data;
3478 struct pci_dev *dev = NULL;
3479 struct amd_iommu *iommu;
3483 ret = alloc_passthrough_domain();
3487 for_each_pci_dev(dev) {
3488 if (!check_device(&dev->dev))
3491 dev_data = get_dev_data(&dev->dev);
3492 dev_data->passthrough = true;
3494 devid = get_device_id(&dev->dev);
3496 iommu = amd_iommu_rlookup_table[devid];
3500 attach_device(&dev->dev, pt_domain);
3503 amd_iommu_stats_init();
3505 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3510 /* IOMMUv2 specific functions */
3511 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3513 return atomic_notifier_chain_register(&ppr_notifier, nb);
3515 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3517 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3519 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3521 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3523 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3525 struct protection_domain *domain = dom->priv;
3526 unsigned long flags;
3528 spin_lock_irqsave(&domain->lock, flags);
3530 /* Update data structure */
3531 domain->mode = PAGE_MODE_NONE;
3532 domain->updated = true;
3534 /* Make changes visible to IOMMUs */
3535 update_domain(domain);
3537 /* Page-table is not visible to IOMMU anymore, so free it */
3538 free_pagetable(domain);
3540 spin_unlock_irqrestore(&domain->lock, flags);
3542 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3544 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3546 struct protection_domain *domain = dom->priv;
3547 unsigned long flags;
3550 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3553 /* Number of GCR3 table levels required */
3554 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3557 if (levels > amd_iommu_max_glx_val)
3560 spin_lock_irqsave(&domain->lock, flags);
3563 * Save us all sanity checks whether devices already in the
3564 * domain support IOMMUv2. Just force that the domain has no
3565 * devices attached when it is switched into IOMMUv2 mode.
3568 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3572 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3573 if (domain->gcr3_tbl == NULL)
3576 domain->glx = levels;
3577 domain->flags |= PD_IOMMUV2_MASK;
3578 domain->updated = true;
3580 update_domain(domain);
3585 spin_unlock_irqrestore(&domain->lock, flags);
3589 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3591 static int __flush_pasid(struct protection_domain *domain, int pasid,
3592 u64 address, bool size)
3594 struct iommu_dev_data *dev_data;
3595 struct iommu_cmd cmd;
3598 if (!(domain->flags & PD_IOMMUV2_MASK))
3601 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3604 * IOMMU TLB needs to be flushed before Device TLB to
3605 * prevent device TLB refill from IOMMU TLB
3607 for (i = 0; i < amd_iommus_present; ++i) {
3608 if (domain->dev_iommu[i] == 0)
3611 ret = iommu_queue_command(amd_iommus[i], &cmd);
3616 /* Wait until IOMMU TLB flushes are complete */
3617 domain_flush_complete(domain);
3619 /* Now flush device TLBs */
3620 list_for_each_entry(dev_data, &domain->dev_list, list) {
3621 struct amd_iommu *iommu;
3624 BUG_ON(!dev_data->ats.enabled);
3626 qdep = dev_data->ats.qdep;
3627 iommu = amd_iommu_rlookup_table[dev_data->devid];
3629 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3630 qdep, address, size);
3632 ret = iommu_queue_command(iommu, &cmd);
3637 /* Wait until all device TLBs are flushed */
3638 domain_flush_complete(domain);
3647 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3650 INC_STATS_COUNTER(invalidate_iotlb);
3652 return __flush_pasid(domain, pasid, address, false);
3655 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3658 struct protection_domain *domain = dom->priv;
3659 unsigned long flags;
3662 spin_lock_irqsave(&domain->lock, flags);
3663 ret = __amd_iommu_flush_page(domain, pasid, address);
3664 spin_unlock_irqrestore(&domain->lock, flags);
3668 EXPORT_SYMBOL(amd_iommu_flush_page);
3670 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3672 INC_STATS_COUNTER(invalidate_iotlb_all);
3674 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3678 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3680 struct protection_domain *domain = dom->priv;
3681 unsigned long flags;
3684 spin_lock_irqsave(&domain->lock, flags);
3685 ret = __amd_iommu_flush_tlb(domain, pasid);
3686 spin_unlock_irqrestore(&domain->lock, flags);
3690 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3692 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3699 index = (pasid >> (9 * level)) & 0x1ff;
3705 if (!(*pte & GCR3_VALID)) {
3709 root = (void *)get_zeroed_page(GFP_ATOMIC);
3713 *pte = __pa(root) | GCR3_VALID;
3716 root = __va(*pte & PAGE_MASK);
3724 static int __set_gcr3(struct protection_domain *domain, int pasid,
3729 if (domain->mode != PAGE_MODE_NONE)
3732 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3736 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3738 return __amd_iommu_flush_tlb(domain, pasid);
3741 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3745 if (domain->mode != PAGE_MODE_NONE)
3748 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3754 return __amd_iommu_flush_tlb(domain, pasid);
3757 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3760 struct protection_domain *domain = dom->priv;
3761 unsigned long flags;
3764 spin_lock_irqsave(&domain->lock, flags);
3765 ret = __set_gcr3(domain, pasid, cr3);
3766 spin_unlock_irqrestore(&domain->lock, flags);
3770 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3772 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3774 struct protection_domain *domain = dom->priv;
3775 unsigned long flags;
3778 spin_lock_irqsave(&domain->lock, flags);
3779 ret = __clear_gcr3(domain, pasid);
3780 spin_unlock_irqrestore(&domain->lock, flags);
3784 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3786 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3787 int status, int tag)
3789 struct iommu_dev_data *dev_data;
3790 struct amd_iommu *iommu;
3791 struct iommu_cmd cmd;
3793 INC_STATS_COUNTER(complete_ppr);
3795 dev_data = get_dev_data(&pdev->dev);
3796 iommu = amd_iommu_rlookup_table[dev_data->devid];
3798 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3799 tag, dev_data->pri_tlp);
3801 return iommu_queue_command(iommu, &cmd);
3803 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3805 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3807 struct protection_domain *domain;
3809 domain = get_domain(&pdev->dev);
3813 /* Only return IOMMUv2 domains */
3814 if (!(domain->flags & PD_IOMMUV2_MASK))
3817 return domain->iommu_domain;
3819 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3821 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3823 struct iommu_dev_data *dev_data;
3825 if (!amd_iommu_v2_supported())
3828 dev_data = get_dev_data(&pdev->dev);
3829 dev_data->errata |= (1 << erratum);
3831 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3833 int amd_iommu_device_info(struct pci_dev *pdev,
3834 struct amd_iommu_device_info *info)
3839 if (pdev == NULL || info == NULL)
3842 if (!amd_iommu_v2_supported())
3845 memset(info, 0, sizeof(*info));
3847 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3849 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3851 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3853 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3855 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3859 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3860 max_pasids = min(max_pasids, (1 << 20));
3862 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3863 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3865 features = pci_pasid_features(pdev);
3866 if (features & PCI_PASID_CAP_EXEC)
3867 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3868 if (features & PCI_PASID_CAP_PRIV)
3869 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3874 EXPORT_SYMBOL(amd_iommu_device_info);
3876 #ifdef CONFIG_IRQ_REMAP
3878 /*****************************************************************************
3880 * Interrupt Remapping Implementation
3882 *****************************************************************************/
3899 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3900 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3901 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3902 #define DTE_IRQ_REMAP_ENABLE 1ULL
3904 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3908 dte = amd_iommu_dev_table[devid].data[2];
3909 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3910 dte |= virt_to_phys(table->table);
3911 dte |= DTE_IRQ_REMAP_INTCTL;
3912 dte |= DTE_IRQ_TABLE_LEN;
3913 dte |= DTE_IRQ_REMAP_ENABLE;
3915 amd_iommu_dev_table[devid].data[2] = dte;
3918 #define IRTE_ALLOCATED (~1U)
3920 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3922 struct irq_remap_table *table = NULL;
3923 struct amd_iommu *iommu;
3924 unsigned long flags;
3927 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3929 iommu = amd_iommu_rlookup_table[devid];
3933 table = irq_lookup_table[devid];
3937 alias = amd_iommu_alias_table[devid];
3938 table = irq_lookup_table[alias];
3940 irq_lookup_table[devid] = table;
3941 set_dte_irq_entry(devid, table);
3942 iommu_flush_dte(iommu, devid);
3946 /* Nothing there yet, allocate new irq remapping table */
3947 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3951 /* Initialize table spin-lock */
3952 spin_lock_init(&table->lock);
3955 /* Keep the first 32 indexes free for IOAPIC interrupts */
3956 table->min_index = 32;
3958 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3959 if (!table->table) {
3965 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3970 for (i = 0; i < 32; ++i)
3971 table->table[i] = IRTE_ALLOCATED;
3974 irq_lookup_table[devid] = table;
3975 set_dte_irq_entry(devid, table);
3976 iommu_flush_dte(iommu, devid);
3977 if (devid != alias) {
3978 irq_lookup_table[alias] = table;
3979 set_dte_irq_entry(devid, table);
3980 iommu_flush_dte(iommu, alias);
3984 iommu_completion_wait(iommu);
3987 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3992 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
3994 struct irq_remap_table *table;
3995 unsigned long flags;
3998 table = get_irq_table(devid, false);
4002 spin_lock_irqsave(&table->lock, flags);
4004 /* Scan table for free entries */
4005 for (c = 0, index = table->min_index;
4006 index < MAX_IRQS_PER_TABLE;
4008 if (table->table[index] == 0)
4014 struct irq_2_irte *irte_info;
4017 table->table[index - c + 1] = IRTE_ALLOCATED;
4022 irte_info = &cfg->irq_2_irte;
4023 irte_info->devid = devid;
4024 irte_info->index = index;
4033 spin_unlock_irqrestore(&table->lock, flags);
4038 static int get_irte(u16 devid, int index, union irte *irte)
4040 struct irq_remap_table *table;
4041 unsigned long flags;
4043 table = get_irq_table(devid, false);
4047 spin_lock_irqsave(&table->lock, flags);
4048 irte->val = table->table[index];
4049 spin_unlock_irqrestore(&table->lock, flags);
4054 static int modify_irte(u16 devid, int index, union irte irte)
4056 struct irq_remap_table *table;
4057 struct amd_iommu *iommu;
4058 unsigned long flags;
4060 iommu = amd_iommu_rlookup_table[devid];
4064 table = get_irq_table(devid, false);
4068 spin_lock_irqsave(&table->lock, flags);
4069 table->table[index] = irte.val;
4070 spin_unlock_irqrestore(&table->lock, flags);
4072 iommu_flush_irt(iommu, devid);
4073 iommu_completion_wait(iommu);
4078 static void free_irte(u16 devid, int index)
4080 struct irq_remap_table *table;
4081 struct amd_iommu *iommu;
4082 unsigned long flags;
4084 iommu = amd_iommu_rlookup_table[devid];
4088 table = get_irq_table(devid, false);
4092 spin_lock_irqsave(&table->lock, flags);
4093 table->table[index] = 0;
4094 spin_unlock_irqrestore(&table->lock, flags);
4096 iommu_flush_irt(iommu, devid);
4097 iommu_completion_wait(iommu);
4100 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4101 unsigned int destination, int vector,
4102 struct io_apic_irq_attr *attr)
4104 struct irq_remap_table *table;
4105 struct irq_2_irte *irte_info;
4106 struct irq_cfg *cfg;
4113 cfg = irq_get_chip_data(irq);
4117 irte_info = &cfg->irq_2_irte;
4118 ioapic_id = mpc_ioapic_id(attr->ioapic);
4119 devid = get_ioapic_devid(ioapic_id);
4124 table = get_irq_table(devid, true);
4128 index = attr->ioapic_pin;
4130 /* Setup IRQ remapping info */
4132 irte_info->devid = devid;
4133 irte_info->index = index;
4135 /* Setup IRTE for IOMMU */
4137 irte.fields.vector = vector;
4138 irte.fields.int_type = apic->irq_delivery_mode;
4139 irte.fields.destination = destination;
4140 irte.fields.dm = apic->irq_dest_mode;
4141 irte.fields.valid = 1;
4143 ret = modify_irte(devid, index, irte);
4147 /* Setup IOAPIC entry */
4148 memset(entry, 0, sizeof(*entry));
4150 entry->vector = index;
4152 entry->trigger = attr->trigger;
4153 entry->polarity = attr->polarity;
4156 * Mask level triggered irqs.
4164 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4167 struct irq_2_irte *irte_info;
4168 unsigned int dest, irq;
4169 struct irq_cfg *cfg;
4173 if (!config_enabled(CONFIG_SMP))
4176 cfg = data->chip_data;
4178 irte_info = &cfg->irq_2_irte;
4180 if (!cpumask_intersects(mask, cpu_online_mask))
4183 if (get_irte(irte_info->devid, irte_info->index, &irte))
4186 if (assign_irq_vector(irq, cfg, mask))
4189 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4191 if (assign_irq_vector(irq, cfg, data->affinity))
4192 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4196 irte.fields.vector = cfg->vector;
4197 irte.fields.destination = dest;
4199 modify_irte(irte_info->devid, irte_info->index, irte);
4201 if (cfg->move_in_progress)
4202 send_cleanup_vector(cfg);
4204 cpumask_copy(data->affinity, mask);
4209 static int free_irq(int irq)
4211 struct irq_2_irte *irte_info;
4212 struct irq_cfg *cfg;
4214 cfg = irq_get_chip_data(irq);
4218 irte_info = &cfg->irq_2_irte;
4220 free_irte(irte_info->devid, irte_info->index);
4225 static void compose_msi_msg(struct pci_dev *pdev,
4226 unsigned int irq, unsigned int dest,
4227 struct msi_msg *msg, u8 hpet_id)
4229 struct irq_2_irte *irte_info;
4230 struct irq_cfg *cfg;
4233 cfg = irq_get_chip_data(irq);
4237 irte_info = &cfg->irq_2_irte;
4240 irte.fields.vector = cfg->vector;
4241 irte.fields.int_type = apic->irq_delivery_mode;
4242 irte.fields.destination = dest;
4243 irte.fields.dm = apic->irq_dest_mode;
4244 irte.fields.valid = 1;
4246 modify_irte(irte_info->devid, irte_info->index, irte);
4248 msg->address_hi = MSI_ADDR_BASE_HI;
4249 msg->address_lo = MSI_ADDR_BASE_LO;
4250 msg->data = irte_info->index;
4253 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4255 struct irq_cfg *cfg;
4262 cfg = irq_get_chip_data(irq);
4266 devid = get_device_id(&pdev->dev);
4267 index = alloc_irq_index(cfg, devid, nvec);
4269 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4272 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4273 int index, int offset)
4275 struct irq_2_irte *irte_info;
4276 struct irq_cfg *cfg;
4282 cfg = irq_get_chip_data(irq);
4286 if (index >= MAX_IRQS_PER_TABLE)
4289 devid = get_device_id(&pdev->dev);
4290 irte_info = &cfg->irq_2_irte;
4293 irte_info->devid = devid;
4294 irte_info->index = index + offset;
4299 static int setup_hpet_msi(unsigned int irq, unsigned int id)
4301 struct irq_2_irte *irte_info;
4302 struct irq_cfg *cfg;
4305 cfg = irq_get_chip_data(irq);
4309 irte_info = &cfg->irq_2_irte;
4310 devid = get_hpet_devid(id);
4314 index = alloc_irq_index(cfg, devid, 1);
4319 irte_info->devid = devid;
4320 irte_info->index = index;
4325 struct irq_remap_ops amd_iommu_irq_ops = {
4326 .supported = amd_iommu_supported,
4327 .prepare = amd_iommu_prepare,
4328 .enable = amd_iommu_enable,
4329 .disable = amd_iommu_disable,
4330 .reenable = amd_iommu_reenable,
4331 .enable_faulting = amd_iommu_enable_faulting,
4332 .setup_ioapic_entry = setup_ioapic_entry,
4333 .set_affinity = set_affinity,
4334 .free_irq = free_irq,
4335 .compose_msi_msg = compose_msi_msg,
4336 .msi_alloc_irq = msi_alloc_irq,
4337 .msi_setup_irq = msi_setup_irq,
4338 .setup_hpet_msi = setup_hpet_msi,