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 function 0.
292 if (dma_pdev->multifunction &&
293 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
294 swap_pci_ref(&dma_pdev,
295 pci_get_slot(dma_pdev->bus,
296 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
300 * Devices on the root bus go through the iommu. If that's not us,
301 * find the next upstream device and test ACS up to the root bus.
302 * Finding the next device may require skipping virtual buses.
304 while (!pci_is_root_bus(dma_pdev->bus)) {
305 struct pci_bus *bus = find_hosted_bus(dma_pdev->bus);
309 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
312 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
318 static int use_pdev_iommu_group(struct pci_dev *pdev, struct device *dev)
320 struct iommu_group *group = iommu_group_get(&pdev->dev);
324 group = iommu_group_alloc();
326 return PTR_ERR(group);
328 WARN_ON(&pdev->dev != dev);
331 ret = iommu_group_add_device(group, dev);
332 iommu_group_put(group);
336 static int use_dev_data_iommu_group(struct iommu_dev_data *dev_data,
339 if (!dev_data->group) {
340 struct iommu_group *group = iommu_group_alloc();
342 return PTR_ERR(group);
344 dev_data->group = group;
347 return iommu_group_add_device(dev_data->group, dev);
350 static int init_iommu_group(struct device *dev)
352 struct iommu_dev_data *dev_data;
353 struct iommu_group *group;
354 struct pci_dev *dma_pdev;
357 group = iommu_group_get(dev);
359 iommu_group_put(group);
363 dev_data = find_dev_data(get_device_id(dev));
367 if (dev_data->alias_data) {
371 if (dev_data->alias_data->group)
375 * If the alias device exists, it's effectively just a first
376 * level quirk for finding the DMA source.
378 alias = amd_iommu_alias_table[dev_data->devid];
379 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
381 dma_pdev = get_isolation_root(dma_pdev);
386 * If the alias is virtual, try to find a parent device
387 * and test whether the IOMMU group is actualy rooted above
388 * the alias. Be careful to also test the parent device if
389 * we think the alias is the root of the group.
391 bus = pci_find_bus(0, alias >> 8);
395 bus = find_hosted_bus(bus);
396 if (IS_ERR(bus) || !bus->self)
399 dma_pdev = get_isolation_root(pci_dev_get(bus->self));
400 if (dma_pdev != bus->self || (dma_pdev->multifunction &&
401 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS)))
404 pci_dev_put(dma_pdev);
408 dma_pdev = get_isolation_root(pci_dev_get(to_pci_dev(dev)));
410 ret = use_pdev_iommu_group(dma_pdev, dev);
411 pci_dev_put(dma_pdev);
414 return use_dev_data_iommu_group(dev_data->alias_data, dev);
417 static int iommu_init_device(struct device *dev)
419 struct pci_dev *pdev = to_pci_dev(dev);
420 struct iommu_dev_data *dev_data;
424 if (dev->archdata.iommu)
427 dev_data = find_dev_data(get_device_id(dev));
431 alias = amd_iommu_alias_table[dev_data->devid];
432 if (alias != dev_data->devid) {
433 struct iommu_dev_data *alias_data;
435 alias_data = find_dev_data(alias);
436 if (alias_data == NULL) {
437 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
439 free_dev_data(dev_data);
442 dev_data->alias_data = alias_data;
445 ret = init_iommu_group(dev);
449 if (pci_iommuv2_capable(pdev)) {
450 struct amd_iommu *iommu;
452 iommu = amd_iommu_rlookup_table[dev_data->devid];
453 dev_data->iommu_v2 = iommu->is_iommu_v2;
456 dev->archdata.iommu = dev_data;
461 static void iommu_ignore_device(struct device *dev)
465 devid = get_device_id(dev);
466 alias = amd_iommu_alias_table[devid];
468 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
469 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
471 amd_iommu_rlookup_table[devid] = NULL;
472 amd_iommu_rlookup_table[alias] = NULL;
475 static void iommu_uninit_device(struct device *dev)
477 iommu_group_remove_device(dev);
480 * Nothing to do here - we keep dev_data around for unplugged devices
481 * and reuse it when the device is re-plugged - not doing so would
482 * introduce a ton of races.
486 void __init amd_iommu_uninit_devices(void)
488 struct iommu_dev_data *dev_data, *n;
489 struct pci_dev *pdev = NULL;
491 for_each_pci_dev(pdev) {
493 if (!check_device(&pdev->dev))
496 iommu_uninit_device(&pdev->dev);
499 /* Free all of our dev_data structures */
500 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
501 free_dev_data(dev_data);
504 int __init amd_iommu_init_devices(void)
506 struct pci_dev *pdev = NULL;
509 for_each_pci_dev(pdev) {
511 if (!check_device(&pdev->dev))
514 ret = iommu_init_device(&pdev->dev);
515 if (ret == -ENOTSUPP)
516 iommu_ignore_device(&pdev->dev);
525 amd_iommu_uninit_devices();
529 #ifdef CONFIG_AMD_IOMMU_STATS
532 * Initialization code for statistics collection
535 DECLARE_STATS_COUNTER(compl_wait);
536 DECLARE_STATS_COUNTER(cnt_map_single);
537 DECLARE_STATS_COUNTER(cnt_unmap_single);
538 DECLARE_STATS_COUNTER(cnt_map_sg);
539 DECLARE_STATS_COUNTER(cnt_unmap_sg);
540 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
541 DECLARE_STATS_COUNTER(cnt_free_coherent);
542 DECLARE_STATS_COUNTER(cross_page);
543 DECLARE_STATS_COUNTER(domain_flush_single);
544 DECLARE_STATS_COUNTER(domain_flush_all);
545 DECLARE_STATS_COUNTER(alloced_io_mem);
546 DECLARE_STATS_COUNTER(total_map_requests);
547 DECLARE_STATS_COUNTER(complete_ppr);
548 DECLARE_STATS_COUNTER(invalidate_iotlb);
549 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
550 DECLARE_STATS_COUNTER(pri_requests);
552 static struct dentry *stats_dir;
553 static struct dentry *de_fflush;
555 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
557 if (stats_dir == NULL)
560 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
564 static void amd_iommu_stats_init(void)
566 stats_dir = debugfs_create_dir("amd-iommu", NULL);
567 if (stats_dir == NULL)
570 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
571 &amd_iommu_unmap_flush);
573 amd_iommu_stats_add(&compl_wait);
574 amd_iommu_stats_add(&cnt_map_single);
575 amd_iommu_stats_add(&cnt_unmap_single);
576 amd_iommu_stats_add(&cnt_map_sg);
577 amd_iommu_stats_add(&cnt_unmap_sg);
578 amd_iommu_stats_add(&cnt_alloc_coherent);
579 amd_iommu_stats_add(&cnt_free_coherent);
580 amd_iommu_stats_add(&cross_page);
581 amd_iommu_stats_add(&domain_flush_single);
582 amd_iommu_stats_add(&domain_flush_all);
583 amd_iommu_stats_add(&alloced_io_mem);
584 amd_iommu_stats_add(&total_map_requests);
585 amd_iommu_stats_add(&complete_ppr);
586 amd_iommu_stats_add(&invalidate_iotlb);
587 amd_iommu_stats_add(&invalidate_iotlb_all);
588 amd_iommu_stats_add(&pri_requests);
593 /****************************************************************************
595 * Interrupt handling functions
597 ****************************************************************************/
599 static void dump_dte_entry(u16 devid)
603 for (i = 0; i < 4; ++i)
604 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
605 amd_iommu_dev_table[devid].data[i]);
608 static void dump_command(unsigned long phys_addr)
610 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
613 for (i = 0; i < 4; ++i)
614 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
617 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
619 int type, devid, domid, flags;
620 volatile u32 *event = __evt;
625 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
626 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
627 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
628 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
629 address = (u64)(((u64)event[3]) << 32) | event[2];
632 /* Did we hit the erratum? */
633 if (++count == LOOP_TIMEOUT) {
634 pr_err("AMD-Vi: No event written to event log\n");
641 printk(KERN_ERR "AMD-Vi: Event logged [");
644 case EVENT_TYPE_ILL_DEV:
645 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
646 "address=0x%016llx flags=0x%04x]\n",
647 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
649 dump_dte_entry(devid);
651 case EVENT_TYPE_IO_FAULT:
652 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
653 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
654 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
655 domid, address, flags);
657 case EVENT_TYPE_DEV_TAB_ERR:
658 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
659 "address=0x%016llx flags=0x%04x]\n",
660 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
663 case EVENT_TYPE_PAGE_TAB_ERR:
664 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
665 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
666 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
667 domid, address, flags);
669 case EVENT_TYPE_ILL_CMD:
670 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
671 dump_command(address);
673 case EVENT_TYPE_CMD_HARD_ERR:
674 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
675 "flags=0x%04x]\n", address, flags);
677 case EVENT_TYPE_IOTLB_INV_TO:
678 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
679 "address=0x%016llx]\n",
680 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
683 case EVENT_TYPE_INV_DEV_REQ:
684 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
685 "address=0x%016llx flags=0x%04x]\n",
686 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
690 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
693 memset(__evt, 0, 4 * sizeof(u32));
696 static void iommu_poll_events(struct amd_iommu *iommu)
700 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
701 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
703 while (head != tail) {
704 iommu_print_event(iommu, iommu->evt_buf + head);
705 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
708 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
711 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
713 struct amd_iommu_fault fault;
715 INC_STATS_COUNTER(pri_requests);
717 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
718 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
722 fault.address = raw[1];
723 fault.pasid = PPR_PASID(raw[0]);
724 fault.device_id = PPR_DEVID(raw[0]);
725 fault.tag = PPR_TAG(raw[0]);
726 fault.flags = PPR_FLAGS(raw[0]);
728 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
731 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
735 if (iommu->ppr_log == NULL)
738 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
739 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
741 while (head != tail) {
746 raw = (u64 *)(iommu->ppr_log + head);
749 * Hardware bug: Interrupt may arrive before the entry is
750 * written to memory. If this happens we need to wait for the
753 for (i = 0; i < LOOP_TIMEOUT; ++i) {
754 if (PPR_REQ_TYPE(raw[0]) != 0)
759 /* Avoid memcpy function-call overhead */
764 * To detect the hardware bug we need to clear the entry
767 raw[0] = raw[1] = 0UL;
769 /* Update head pointer of hardware ring-buffer */
770 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
771 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
773 /* Handle PPR entry */
774 iommu_handle_ppr_entry(iommu, entry);
776 /* Refresh ring-buffer information */
777 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
778 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
782 irqreturn_t amd_iommu_int_thread(int irq, void *data)
784 struct amd_iommu *iommu = (struct amd_iommu *) data;
785 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
787 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
788 /* Enable EVT and PPR interrupts again */
789 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
790 iommu->mmio_base + MMIO_STATUS_OFFSET);
792 if (status & MMIO_STATUS_EVT_INT_MASK) {
793 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
794 iommu_poll_events(iommu);
797 if (status & MMIO_STATUS_PPR_INT_MASK) {
798 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
799 iommu_poll_ppr_log(iommu);
803 * Hardware bug: ERBT1312
804 * When re-enabling interrupt (by writing 1
805 * to clear the bit), the hardware might also try to set
806 * the interrupt bit in the event status register.
807 * In this scenario, the bit will be set, and disable
808 * subsequent interrupts.
810 * Workaround: The IOMMU driver should read back the
811 * status register and check if the interrupt bits are cleared.
812 * If not, driver will need to go through the interrupt handler
813 * again and re-clear the bits
815 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
820 irqreturn_t amd_iommu_int_handler(int irq, void *data)
822 return IRQ_WAKE_THREAD;
825 /****************************************************************************
827 * IOMMU command queuing functions
829 ****************************************************************************/
831 static int wait_on_sem(volatile u64 *sem)
835 while (*sem == 0 && i < LOOP_TIMEOUT) {
840 if (i == LOOP_TIMEOUT) {
841 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
848 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
849 struct iommu_cmd *cmd,
854 target = iommu->cmd_buf + tail;
855 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
857 /* Copy command to buffer */
858 memcpy(target, cmd, sizeof(*cmd));
860 /* Tell the IOMMU about it */
861 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
864 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
866 WARN_ON(address & 0x7ULL);
868 memset(cmd, 0, sizeof(*cmd));
869 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
870 cmd->data[1] = upper_32_bits(__pa(address));
872 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
875 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
877 memset(cmd, 0, sizeof(*cmd));
878 cmd->data[0] = devid;
879 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
882 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
883 size_t size, u16 domid, int pde)
888 pages = iommu_num_pages(address, size, PAGE_SIZE);
893 * If we have to flush more than one page, flush all
894 * TLB entries for this domain
896 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
900 address &= PAGE_MASK;
902 memset(cmd, 0, sizeof(*cmd));
903 cmd->data[1] |= domid;
904 cmd->data[2] = lower_32_bits(address);
905 cmd->data[3] = upper_32_bits(address);
906 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
907 if (s) /* size bit - we flush more than one 4kb page */
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
909 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
910 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
913 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
914 u64 address, size_t size)
919 pages = iommu_num_pages(address, size, PAGE_SIZE);
924 * If we have to flush more than one page, flush all
925 * TLB entries for this domain
927 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
931 address &= PAGE_MASK;
933 memset(cmd, 0, sizeof(*cmd));
934 cmd->data[0] = devid;
935 cmd->data[0] |= (qdep & 0xff) << 24;
936 cmd->data[1] = devid;
937 cmd->data[2] = lower_32_bits(address);
938 cmd->data[3] = upper_32_bits(address);
939 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
941 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
944 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
945 u64 address, bool size)
947 memset(cmd, 0, sizeof(*cmd));
949 address &= ~(0xfffULL);
951 cmd->data[0] = pasid & PASID_MASK;
952 cmd->data[1] = domid;
953 cmd->data[2] = lower_32_bits(address);
954 cmd->data[3] = upper_32_bits(address);
955 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
958 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
959 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
962 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
963 int qdep, u64 address, bool size)
965 memset(cmd, 0, sizeof(*cmd));
967 address &= ~(0xfffULL);
969 cmd->data[0] = devid;
970 cmd->data[0] |= (pasid & 0xff) << 16;
971 cmd->data[0] |= (qdep & 0xff) << 24;
972 cmd->data[1] = devid;
973 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
974 cmd->data[2] = lower_32_bits(address);
975 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
976 cmd->data[3] = upper_32_bits(address);
978 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
979 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
982 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
983 int status, int tag, bool gn)
985 memset(cmd, 0, sizeof(*cmd));
987 cmd->data[0] = devid;
989 cmd->data[1] = pasid & PASID_MASK;
990 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
992 cmd->data[3] = tag & 0x1ff;
993 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
995 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
998 static void build_inv_all(struct iommu_cmd *cmd)
1000 memset(cmd, 0, sizeof(*cmd));
1001 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1004 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1006 memset(cmd, 0, sizeof(*cmd));
1007 cmd->data[0] = devid;
1008 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1012 * Writes the command to the IOMMUs command buffer and informs the
1013 * hardware about the new command.
1015 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1016 struct iommu_cmd *cmd,
1019 u32 left, tail, head, next_tail;
1020 unsigned long flags;
1022 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
1025 spin_lock_irqsave(&iommu->lock, flags);
1027 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1028 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1029 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
1030 left = (head - next_tail) % iommu->cmd_buf_size;
1033 struct iommu_cmd sync_cmd;
1034 volatile u64 sem = 0;
1037 build_completion_wait(&sync_cmd, (u64)&sem);
1038 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1040 spin_unlock_irqrestore(&iommu->lock, flags);
1042 if ((ret = wait_on_sem(&sem)) != 0)
1048 copy_cmd_to_buffer(iommu, cmd, tail);
1050 /* We need to sync now to make sure all commands are processed */
1051 iommu->need_sync = sync;
1053 spin_unlock_irqrestore(&iommu->lock, flags);
1058 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1060 return iommu_queue_command_sync(iommu, cmd, true);
1064 * This function queues a completion wait command into the command
1065 * buffer of an IOMMU
1067 static int iommu_completion_wait(struct amd_iommu *iommu)
1069 struct iommu_cmd cmd;
1070 volatile u64 sem = 0;
1073 if (!iommu->need_sync)
1076 build_completion_wait(&cmd, (u64)&sem);
1078 ret = iommu_queue_command_sync(iommu, &cmd, false);
1082 return wait_on_sem(&sem);
1085 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1087 struct iommu_cmd cmd;
1089 build_inv_dte(&cmd, devid);
1091 return iommu_queue_command(iommu, &cmd);
1094 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1098 for (devid = 0; devid <= 0xffff; ++devid)
1099 iommu_flush_dte(iommu, devid);
1101 iommu_completion_wait(iommu);
1105 * This function uses heavy locking and may disable irqs for some time. But
1106 * this is no issue because it is only called during resume.
1108 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1112 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1113 struct iommu_cmd cmd;
1114 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1116 iommu_queue_command(iommu, &cmd);
1119 iommu_completion_wait(iommu);
1122 static void iommu_flush_all(struct amd_iommu *iommu)
1124 struct iommu_cmd cmd;
1126 build_inv_all(&cmd);
1128 iommu_queue_command(iommu, &cmd);
1129 iommu_completion_wait(iommu);
1132 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1134 struct iommu_cmd cmd;
1136 build_inv_irt(&cmd, devid);
1138 iommu_queue_command(iommu, &cmd);
1141 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1145 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1146 iommu_flush_irt(iommu, devid);
1148 iommu_completion_wait(iommu);
1151 void iommu_flush_all_caches(struct amd_iommu *iommu)
1153 if (iommu_feature(iommu, FEATURE_IA)) {
1154 iommu_flush_all(iommu);
1156 iommu_flush_dte_all(iommu);
1157 iommu_flush_irt_all(iommu);
1158 iommu_flush_tlb_all(iommu);
1163 * Command send function for flushing on-device TLB
1165 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1166 u64 address, size_t size)
1168 struct amd_iommu *iommu;
1169 struct iommu_cmd cmd;
1172 qdep = dev_data->ats.qdep;
1173 iommu = amd_iommu_rlookup_table[dev_data->devid];
1175 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1177 return iommu_queue_command(iommu, &cmd);
1181 * Command send function for invalidating a device table entry
1183 static int device_flush_dte(struct iommu_dev_data *dev_data)
1185 struct amd_iommu *iommu;
1188 iommu = amd_iommu_rlookup_table[dev_data->devid];
1190 ret = iommu_flush_dte(iommu, dev_data->devid);
1194 if (dev_data->ats.enabled)
1195 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1201 * TLB invalidation function which is called from the mapping functions.
1202 * It invalidates a single PTE if the range to flush is within a single
1203 * page. Otherwise it flushes the whole TLB of the IOMMU.
1205 static void __domain_flush_pages(struct protection_domain *domain,
1206 u64 address, size_t size, int pde)
1208 struct iommu_dev_data *dev_data;
1209 struct iommu_cmd cmd;
1212 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1214 for (i = 0; i < amd_iommus_present; ++i) {
1215 if (!domain->dev_iommu[i])
1219 * Devices of this domain are behind this IOMMU
1220 * We need a TLB flush
1222 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1225 list_for_each_entry(dev_data, &domain->dev_list, list) {
1227 if (!dev_data->ats.enabled)
1230 ret |= device_flush_iotlb(dev_data, address, size);
1236 static void domain_flush_pages(struct protection_domain *domain,
1237 u64 address, size_t size)
1239 __domain_flush_pages(domain, address, size, 0);
1242 /* Flush the whole IO/TLB for a given protection domain */
1243 static void domain_flush_tlb(struct protection_domain *domain)
1245 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1248 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1249 static void domain_flush_tlb_pde(struct protection_domain *domain)
1251 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1254 static void domain_flush_complete(struct protection_domain *domain)
1258 for (i = 0; i < amd_iommus_present; ++i) {
1259 if (!domain->dev_iommu[i])
1263 * Devices of this domain are behind this IOMMU
1264 * We need to wait for completion of all commands.
1266 iommu_completion_wait(amd_iommus[i]);
1272 * This function flushes the DTEs for all devices in domain
1274 static void domain_flush_devices(struct protection_domain *domain)
1276 struct iommu_dev_data *dev_data;
1278 list_for_each_entry(dev_data, &domain->dev_list, list)
1279 device_flush_dte(dev_data);
1282 /****************************************************************************
1284 * The functions below are used the create the page table mappings for
1285 * unity mapped regions.
1287 ****************************************************************************/
1290 * This function is used to add another level to an IO page table. Adding
1291 * another level increases the size of the address space by 9 bits to a size up
1294 static bool increase_address_space(struct protection_domain *domain,
1299 if (domain->mode == PAGE_MODE_6_LEVEL)
1300 /* address space already 64 bit large */
1303 pte = (void *)get_zeroed_page(gfp);
1307 *pte = PM_LEVEL_PDE(domain->mode,
1308 virt_to_phys(domain->pt_root));
1309 domain->pt_root = pte;
1311 domain->updated = true;
1316 static u64 *alloc_pte(struct protection_domain *domain,
1317 unsigned long address,
1318 unsigned long page_size,
1325 BUG_ON(!is_power_of_2(page_size));
1327 while (address > PM_LEVEL_SIZE(domain->mode))
1328 increase_address_space(domain, gfp);
1330 level = domain->mode - 1;
1331 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1332 address = PAGE_SIZE_ALIGN(address, page_size);
1333 end_lvl = PAGE_SIZE_LEVEL(page_size);
1335 while (level > end_lvl) {
1336 if (!IOMMU_PTE_PRESENT(*pte)) {
1337 page = (u64 *)get_zeroed_page(gfp);
1340 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1343 /* No level skipping support yet */
1344 if (PM_PTE_LEVEL(*pte) != level)
1349 pte = IOMMU_PTE_PAGE(*pte);
1351 if (pte_page && level == end_lvl)
1354 pte = &pte[PM_LEVEL_INDEX(level, address)];
1361 * This function checks if there is a PTE for a given dma address. If
1362 * there is one, it returns the pointer to it.
1364 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1369 if (address > PM_LEVEL_SIZE(domain->mode))
1372 level = domain->mode - 1;
1373 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1378 if (!IOMMU_PTE_PRESENT(*pte))
1382 if (PM_PTE_LEVEL(*pte) == 0x07) {
1383 unsigned long pte_mask, __pte;
1386 * If we have a series of large PTEs, make
1387 * sure to return a pointer to the first one.
1389 pte_mask = PTE_PAGE_SIZE(*pte);
1390 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1391 __pte = ((unsigned long)pte) & pte_mask;
1393 return (u64 *)__pte;
1396 /* No level skipping support yet */
1397 if (PM_PTE_LEVEL(*pte) != level)
1402 /* Walk to the next level */
1403 pte = IOMMU_PTE_PAGE(*pte);
1404 pte = &pte[PM_LEVEL_INDEX(level, address)];
1411 * Generic mapping functions. It maps a physical address into a DMA
1412 * address space. It allocates the page table pages if necessary.
1413 * In the future it can be extended to a generic mapping function
1414 * supporting all features of AMD IOMMU page tables like level skipping
1415 * and full 64 bit address spaces.
1417 static int iommu_map_page(struct protection_domain *dom,
1418 unsigned long bus_addr,
1419 unsigned long phys_addr,
1421 unsigned long page_size)
1426 if (!(prot & IOMMU_PROT_MASK))
1429 bus_addr = PAGE_ALIGN(bus_addr);
1430 phys_addr = PAGE_ALIGN(phys_addr);
1431 count = PAGE_SIZE_PTE_COUNT(page_size);
1432 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1434 for (i = 0; i < count; ++i)
1435 if (IOMMU_PTE_PRESENT(pte[i]))
1438 if (page_size > PAGE_SIZE) {
1439 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1440 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1442 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1444 if (prot & IOMMU_PROT_IR)
1445 __pte |= IOMMU_PTE_IR;
1446 if (prot & IOMMU_PROT_IW)
1447 __pte |= IOMMU_PTE_IW;
1449 for (i = 0; i < count; ++i)
1457 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1458 unsigned long bus_addr,
1459 unsigned long page_size)
1461 unsigned long long unmap_size, unmapped;
1464 BUG_ON(!is_power_of_2(page_size));
1468 while (unmapped < page_size) {
1470 pte = fetch_pte(dom, bus_addr);
1474 * No PTE for this address
1475 * move forward in 4kb steps
1477 unmap_size = PAGE_SIZE;
1478 } else if (PM_PTE_LEVEL(*pte) == 0) {
1479 /* 4kb PTE found for this address */
1480 unmap_size = PAGE_SIZE;
1485 /* Large PTE found which maps this address */
1486 unmap_size = PTE_PAGE_SIZE(*pte);
1488 /* Only unmap from the first pte in the page */
1489 if ((unmap_size - 1) & bus_addr)
1491 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1492 for (i = 0; i < count; i++)
1496 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1497 unmapped += unmap_size;
1500 BUG_ON(unmapped && !is_power_of_2(unmapped));
1506 * This function checks if a specific unity mapping entry is needed for
1507 * this specific IOMMU.
1509 static int iommu_for_unity_map(struct amd_iommu *iommu,
1510 struct unity_map_entry *entry)
1514 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1515 bdf = amd_iommu_alias_table[i];
1516 if (amd_iommu_rlookup_table[bdf] == iommu)
1524 * This function actually applies the mapping to the page table of the
1527 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1528 struct unity_map_entry *e)
1533 for (addr = e->address_start; addr < e->address_end;
1534 addr += PAGE_SIZE) {
1535 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1540 * if unity mapping is in aperture range mark the page
1541 * as allocated in the aperture
1543 if (addr < dma_dom->aperture_size)
1544 __set_bit(addr >> PAGE_SHIFT,
1545 dma_dom->aperture[0]->bitmap);
1552 * Init the unity mappings for a specific IOMMU in the system
1554 * Basically iterates over all unity mapping entries and applies them to
1555 * the default domain DMA of that IOMMU if necessary.
1557 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1559 struct unity_map_entry *entry;
1562 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1563 if (!iommu_for_unity_map(iommu, entry))
1565 ret = dma_ops_unity_map(iommu->default_dom, entry);
1574 * Inits the unity mappings required for a specific device
1576 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1579 struct unity_map_entry *e;
1582 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1583 if (!(devid >= e->devid_start && devid <= e->devid_end))
1585 ret = dma_ops_unity_map(dma_dom, e);
1593 /****************************************************************************
1595 * The next functions belong to the address allocator for the dma_ops
1596 * interface functions. They work like the allocators in the other IOMMU
1597 * drivers. Its basically a bitmap which marks the allocated pages in
1598 * the aperture. Maybe it could be enhanced in the future to a more
1599 * efficient allocator.
1601 ****************************************************************************/
1604 * The address allocator core functions.
1606 * called with domain->lock held
1610 * Used to reserve address ranges in the aperture (e.g. for exclusion
1613 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1614 unsigned long start_page,
1617 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1619 if (start_page + pages > last_page)
1620 pages = last_page - start_page;
1622 for (i = start_page; i < start_page + pages; ++i) {
1623 int index = i / APERTURE_RANGE_PAGES;
1624 int page = i % APERTURE_RANGE_PAGES;
1625 __set_bit(page, dom->aperture[index]->bitmap);
1630 * This function is used to add a new aperture range to an existing
1631 * aperture in case of dma_ops domain allocation or address allocation
1634 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1635 bool populate, gfp_t gfp)
1637 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1638 struct amd_iommu *iommu;
1639 unsigned long i, old_size;
1641 #ifdef CONFIG_IOMMU_STRESS
1645 if (index >= APERTURE_MAX_RANGES)
1648 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1649 if (!dma_dom->aperture[index])
1652 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1653 if (!dma_dom->aperture[index]->bitmap)
1656 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1659 unsigned long address = dma_dom->aperture_size;
1660 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1661 u64 *pte, *pte_page;
1663 for (i = 0; i < num_ptes; ++i) {
1664 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1669 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1671 address += APERTURE_RANGE_SIZE / 64;
1675 old_size = dma_dom->aperture_size;
1676 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1678 /* Reserve address range used for MSI messages */
1679 if (old_size < MSI_ADDR_BASE_LO &&
1680 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1681 unsigned long spage;
1684 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1685 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1687 dma_ops_reserve_addresses(dma_dom, spage, pages);
1690 /* Initialize the exclusion range if necessary */
1691 for_each_iommu(iommu) {
1692 if (iommu->exclusion_start &&
1693 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1694 && iommu->exclusion_start < dma_dom->aperture_size) {
1695 unsigned long startpage;
1696 int pages = iommu_num_pages(iommu->exclusion_start,
1697 iommu->exclusion_length,
1699 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1700 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1705 * Check for areas already mapped as present in the new aperture
1706 * range and mark those pages as reserved in the allocator. Such
1707 * mappings may already exist as a result of requested unity
1708 * mappings for devices.
1710 for (i = dma_dom->aperture[index]->offset;
1711 i < dma_dom->aperture_size;
1713 u64 *pte = fetch_pte(&dma_dom->domain, i);
1714 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1717 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1720 update_domain(&dma_dom->domain);
1725 update_domain(&dma_dom->domain);
1727 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1729 kfree(dma_dom->aperture[index]);
1730 dma_dom->aperture[index] = NULL;
1735 static unsigned long dma_ops_area_alloc(struct device *dev,
1736 struct dma_ops_domain *dom,
1738 unsigned long align_mask,
1740 unsigned long start)
1742 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1743 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1744 int i = start >> APERTURE_RANGE_SHIFT;
1745 unsigned long boundary_size;
1746 unsigned long address = -1;
1747 unsigned long limit;
1749 next_bit >>= PAGE_SHIFT;
1751 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1752 PAGE_SIZE) >> PAGE_SHIFT;
1754 for (;i < max_index; ++i) {
1755 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1757 if (dom->aperture[i]->offset >= dma_mask)
1760 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1761 dma_mask >> PAGE_SHIFT);
1763 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1764 limit, next_bit, pages, 0,
1765 boundary_size, align_mask);
1766 if (address != -1) {
1767 address = dom->aperture[i]->offset +
1768 (address << PAGE_SHIFT);
1769 dom->next_address = address + (pages << PAGE_SHIFT);
1779 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1780 struct dma_ops_domain *dom,
1782 unsigned long align_mask,
1785 unsigned long address;
1787 #ifdef CONFIG_IOMMU_STRESS
1788 dom->next_address = 0;
1789 dom->need_flush = true;
1792 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1793 dma_mask, dom->next_address);
1795 if (address == -1) {
1796 dom->next_address = 0;
1797 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1799 dom->need_flush = true;
1802 if (unlikely(address == -1))
1803 address = DMA_ERROR_CODE;
1805 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1811 * The address free function.
1813 * called with domain->lock held
1815 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1816 unsigned long address,
1819 unsigned i = address >> APERTURE_RANGE_SHIFT;
1820 struct aperture_range *range = dom->aperture[i];
1822 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1824 #ifdef CONFIG_IOMMU_STRESS
1829 if (address >= dom->next_address)
1830 dom->need_flush = true;
1832 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1834 bitmap_clear(range->bitmap, address, pages);
1838 /****************************************************************************
1840 * The next functions belong to the domain allocation. A domain is
1841 * allocated for every IOMMU as the default domain. If device isolation
1842 * is enabled, every device get its own domain. The most important thing
1843 * about domains is the page table mapping the DMA address space they
1846 ****************************************************************************/
1849 * This function adds a protection domain to the global protection domain list
1851 static void add_domain_to_list(struct protection_domain *domain)
1853 unsigned long flags;
1855 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1856 list_add(&domain->list, &amd_iommu_pd_list);
1857 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1861 * This function removes a protection domain to the global
1862 * protection domain list
1864 static void del_domain_from_list(struct protection_domain *domain)
1866 unsigned long flags;
1868 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1869 list_del(&domain->list);
1870 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1873 static u16 domain_id_alloc(void)
1875 unsigned long flags;
1878 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1879 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1881 if (id > 0 && id < MAX_DOMAIN_ID)
1882 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1885 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1890 static void domain_id_free(int id)
1892 unsigned long flags;
1894 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1895 if (id > 0 && id < MAX_DOMAIN_ID)
1896 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1897 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1900 #define DEFINE_FREE_PT_FN(LVL, FN) \
1901 static void free_pt_##LVL (unsigned long __pt) \
1909 for (i = 0; i < 512; ++i) { \
1910 if (!IOMMU_PTE_PRESENT(pt[i])) \
1913 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1916 free_page((unsigned long)pt); \
1919 DEFINE_FREE_PT_FN(l2, free_page)
1920 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1921 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1922 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1923 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1925 static void free_pagetable(struct protection_domain *domain)
1927 unsigned long root = (unsigned long)domain->pt_root;
1929 switch (domain->mode) {
1930 case PAGE_MODE_NONE:
1932 case PAGE_MODE_1_LEVEL:
1935 case PAGE_MODE_2_LEVEL:
1938 case PAGE_MODE_3_LEVEL:
1941 case PAGE_MODE_4_LEVEL:
1944 case PAGE_MODE_5_LEVEL:
1947 case PAGE_MODE_6_LEVEL:
1955 static void free_gcr3_tbl_level1(u64 *tbl)
1960 for (i = 0; i < 512; ++i) {
1961 if (!(tbl[i] & GCR3_VALID))
1964 ptr = __va(tbl[i] & PAGE_MASK);
1966 free_page((unsigned long)ptr);
1970 static void free_gcr3_tbl_level2(u64 *tbl)
1975 for (i = 0; i < 512; ++i) {
1976 if (!(tbl[i] & GCR3_VALID))
1979 ptr = __va(tbl[i] & PAGE_MASK);
1981 free_gcr3_tbl_level1(ptr);
1985 static void free_gcr3_table(struct protection_domain *domain)
1987 if (domain->glx == 2)
1988 free_gcr3_tbl_level2(domain->gcr3_tbl);
1989 else if (domain->glx == 1)
1990 free_gcr3_tbl_level1(domain->gcr3_tbl);
1991 else if (domain->glx != 0)
1994 free_page((unsigned long)domain->gcr3_tbl);
1998 * Free a domain, only used if something went wrong in the
1999 * allocation path and we need to free an already allocated page table
2001 static void dma_ops_domain_free(struct dma_ops_domain *dom)
2008 del_domain_from_list(&dom->domain);
2010 free_pagetable(&dom->domain);
2012 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
2013 if (!dom->aperture[i])
2015 free_page((unsigned long)dom->aperture[i]->bitmap);
2016 kfree(dom->aperture[i]);
2023 * Allocates a new protection domain usable for the dma_ops functions.
2024 * It also initializes the page table and the address allocator data
2025 * structures required for the dma_ops interface
2027 static struct dma_ops_domain *dma_ops_domain_alloc(void)
2029 struct dma_ops_domain *dma_dom;
2031 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2035 spin_lock_init(&dma_dom->domain.lock);
2037 dma_dom->domain.id = domain_id_alloc();
2038 if (dma_dom->domain.id == 0)
2040 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
2041 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2042 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2043 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2044 dma_dom->domain.priv = dma_dom;
2045 if (!dma_dom->domain.pt_root)
2048 dma_dom->need_flush = false;
2049 dma_dom->target_dev = 0xffff;
2051 add_domain_to_list(&dma_dom->domain);
2053 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2057 * mark the first page as allocated so we never return 0 as
2058 * a valid dma-address. So we can use 0 as error value
2060 dma_dom->aperture[0]->bitmap[0] = 1;
2061 dma_dom->next_address = 0;
2067 dma_ops_domain_free(dma_dom);
2073 * little helper function to check whether a given protection domain is a
2076 static bool dma_ops_domain(struct protection_domain *domain)
2078 return domain->flags & PD_DMA_OPS_MASK;
2081 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2086 if (domain->mode != PAGE_MODE_NONE)
2087 pte_root = virt_to_phys(domain->pt_root);
2089 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2090 << DEV_ENTRY_MODE_SHIFT;
2091 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2093 flags = amd_iommu_dev_table[devid].data[1];
2096 flags |= DTE_FLAG_IOTLB;
2098 if (domain->flags & PD_IOMMUV2_MASK) {
2099 u64 gcr3 = __pa(domain->gcr3_tbl);
2100 u64 glx = domain->glx;
2103 pte_root |= DTE_FLAG_GV;
2104 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2106 /* First mask out possible old values for GCR3 table */
2107 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2110 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2113 /* Encode GCR3 table into DTE */
2114 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2117 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2120 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2124 flags &= ~(0xffffUL);
2125 flags |= domain->id;
2127 amd_iommu_dev_table[devid].data[1] = flags;
2128 amd_iommu_dev_table[devid].data[0] = pte_root;
2131 static void clear_dte_entry(u16 devid)
2133 /* remove entry from the device table seen by the hardware */
2134 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2135 amd_iommu_dev_table[devid].data[1] = 0;
2137 amd_iommu_apply_erratum_63(devid);
2140 static void do_attach(struct iommu_dev_data *dev_data,
2141 struct protection_domain *domain)
2143 struct amd_iommu *iommu;
2146 iommu = amd_iommu_rlookup_table[dev_data->devid];
2147 ats = dev_data->ats.enabled;
2149 /* Update data structures */
2150 dev_data->domain = domain;
2151 list_add(&dev_data->list, &domain->dev_list);
2152 set_dte_entry(dev_data->devid, domain, ats);
2154 /* Do reference counting */
2155 domain->dev_iommu[iommu->index] += 1;
2156 domain->dev_cnt += 1;
2158 /* Flush the DTE entry */
2159 device_flush_dte(dev_data);
2162 static void do_detach(struct iommu_dev_data *dev_data)
2164 struct amd_iommu *iommu;
2166 iommu = amd_iommu_rlookup_table[dev_data->devid];
2168 /* decrease reference counters */
2169 dev_data->domain->dev_iommu[iommu->index] -= 1;
2170 dev_data->domain->dev_cnt -= 1;
2172 /* Update data structures */
2173 dev_data->domain = NULL;
2174 list_del(&dev_data->list);
2175 clear_dte_entry(dev_data->devid);
2177 /* Flush the DTE entry */
2178 device_flush_dte(dev_data);
2182 * If a device is not yet associated with a domain, this function does
2183 * assigns it visible for the hardware
2185 static int __attach_device(struct iommu_dev_data *dev_data,
2186 struct protection_domain *domain)
2191 spin_lock(&domain->lock);
2193 if (dev_data->alias_data != NULL) {
2194 struct iommu_dev_data *alias_data = dev_data->alias_data;
2196 /* Some sanity checks */
2198 if (alias_data->domain != NULL &&
2199 alias_data->domain != domain)
2202 if (dev_data->domain != NULL &&
2203 dev_data->domain != domain)
2206 /* Do real assignment */
2207 if (alias_data->domain == NULL)
2208 do_attach(alias_data, domain);
2210 atomic_inc(&alias_data->bind);
2213 if (dev_data->domain == NULL)
2214 do_attach(dev_data, domain);
2216 atomic_inc(&dev_data->bind);
2223 spin_unlock(&domain->lock);
2229 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2231 pci_disable_ats(pdev);
2232 pci_disable_pri(pdev);
2233 pci_disable_pasid(pdev);
2236 /* FIXME: Change generic reset-function to do the same */
2237 static int pri_reset_while_enabled(struct pci_dev *pdev)
2242 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2246 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2247 control |= PCI_PRI_CTRL_RESET;
2248 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2253 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2258 /* FIXME: Hardcode number of outstanding requests for now */
2260 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2262 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2264 /* Only allow access to user-accessible pages */
2265 ret = pci_enable_pasid(pdev, 0);
2269 /* First reset the PRI state of the device */
2270 ret = pci_reset_pri(pdev);
2275 ret = pci_enable_pri(pdev, reqs);
2280 ret = pri_reset_while_enabled(pdev);
2285 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2292 pci_disable_pri(pdev);
2293 pci_disable_pasid(pdev);
2298 /* FIXME: Move this to PCI code */
2299 #define PCI_PRI_TLP_OFF (1 << 15)
2301 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2306 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2310 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2312 return (status & PCI_PRI_TLP_OFF) ? true : false;
2316 * If a device is not yet associated with a domain, this function
2317 * assigns it visible for the hardware
2319 static int attach_device(struct device *dev,
2320 struct protection_domain *domain)
2322 struct pci_dev *pdev = to_pci_dev(dev);
2323 struct iommu_dev_data *dev_data;
2324 unsigned long flags;
2327 dev_data = get_dev_data(dev);
2329 if (domain->flags & PD_IOMMUV2_MASK) {
2330 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2333 if (pdev_iommuv2_enable(pdev) != 0)
2336 dev_data->ats.enabled = true;
2337 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2338 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2339 } else if (amd_iommu_iotlb_sup &&
2340 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2341 dev_data->ats.enabled = true;
2342 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2345 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2346 ret = __attach_device(dev_data, domain);
2347 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2350 * We might boot into a crash-kernel here. The crashed kernel
2351 * left the caches in the IOMMU dirty. So we have to flush
2352 * here to evict all dirty stuff.
2354 domain_flush_tlb_pde(domain);
2360 * Removes a device from a protection domain (unlocked)
2362 static void __detach_device(struct iommu_dev_data *dev_data)
2364 struct protection_domain *domain;
2365 unsigned long flags;
2367 BUG_ON(!dev_data->domain);
2369 domain = dev_data->domain;
2371 spin_lock_irqsave(&domain->lock, flags);
2373 if (dev_data->alias_data != NULL) {
2374 struct iommu_dev_data *alias_data = dev_data->alias_data;
2376 if (atomic_dec_and_test(&alias_data->bind))
2377 do_detach(alias_data);
2380 if (atomic_dec_and_test(&dev_data->bind))
2381 do_detach(dev_data);
2383 spin_unlock_irqrestore(&domain->lock, flags);
2386 * If we run in passthrough mode the device must be assigned to the
2387 * passthrough domain if it is detached from any other domain.
2388 * Make sure we can deassign from the pt_domain itself.
2390 if (dev_data->passthrough &&
2391 (dev_data->domain == NULL && domain != pt_domain))
2392 __attach_device(dev_data, pt_domain);
2396 * Removes a device from a protection domain (with devtable_lock held)
2398 static void detach_device(struct device *dev)
2400 struct protection_domain *domain;
2401 struct iommu_dev_data *dev_data;
2402 unsigned long flags;
2404 dev_data = get_dev_data(dev);
2405 domain = dev_data->domain;
2407 /* lock device table */
2408 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2409 __detach_device(dev_data);
2410 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2412 if (domain->flags & PD_IOMMUV2_MASK)
2413 pdev_iommuv2_disable(to_pci_dev(dev));
2414 else if (dev_data->ats.enabled)
2415 pci_disable_ats(to_pci_dev(dev));
2417 dev_data->ats.enabled = false;
2421 * Find out the protection domain structure for a given PCI device. This
2422 * will give us the pointer to the page table root for example.
2424 static struct protection_domain *domain_for_device(struct device *dev)
2426 struct iommu_dev_data *dev_data;
2427 struct protection_domain *dom = NULL;
2428 unsigned long flags;
2430 dev_data = get_dev_data(dev);
2432 if (dev_data->domain)
2433 return dev_data->domain;
2435 if (dev_data->alias_data != NULL) {
2436 struct iommu_dev_data *alias_data = dev_data->alias_data;
2438 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2439 if (alias_data->domain != NULL) {
2440 __attach_device(dev_data, alias_data->domain);
2441 dom = alias_data->domain;
2443 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2449 static int device_change_notifier(struct notifier_block *nb,
2450 unsigned long action, void *data)
2452 struct dma_ops_domain *dma_domain;
2453 struct protection_domain *domain;
2454 struct iommu_dev_data *dev_data;
2455 struct device *dev = data;
2456 struct amd_iommu *iommu;
2457 unsigned long flags;
2460 if (!check_device(dev))
2463 devid = get_device_id(dev);
2464 iommu = amd_iommu_rlookup_table[devid];
2465 dev_data = get_dev_data(dev);
2468 case BUS_NOTIFY_UNBOUND_DRIVER:
2470 domain = domain_for_device(dev);
2474 if (dev_data->passthrough)
2478 case BUS_NOTIFY_ADD_DEVICE:
2480 iommu_init_device(dev);
2483 * dev_data is still NULL and
2484 * got initialized in iommu_init_device
2486 dev_data = get_dev_data(dev);
2488 if (iommu_pass_through || dev_data->iommu_v2) {
2489 dev_data->passthrough = true;
2490 attach_device(dev, pt_domain);
2494 domain = domain_for_device(dev);
2496 /* allocate a protection domain if a device is added */
2497 dma_domain = find_protection_domain(devid);
2499 dma_domain = dma_ops_domain_alloc();
2502 dma_domain->target_dev = devid;
2504 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2505 list_add_tail(&dma_domain->list, &iommu_pd_list);
2506 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2509 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2512 case BUS_NOTIFY_DEL_DEVICE:
2514 iommu_uninit_device(dev);
2520 iommu_completion_wait(iommu);
2526 static struct notifier_block device_nb = {
2527 .notifier_call = device_change_notifier,
2530 void amd_iommu_init_notifier(void)
2532 bus_register_notifier(&pci_bus_type, &device_nb);
2535 /*****************************************************************************
2537 * The next functions belong to the dma_ops mapping/unmapping code.
2539 *****************************************************************************/
2542 * In the dma_ops path we only have the struct device. This function
2543 * finds the corresponding IOMMU, the protection domain and the
2544 * requestor id for a given device.
2545 * If the device is not yet associated with a domain this is also done
2548 static struct protection_domain *get_domain(struct device *dev)
2550 struct protection_domain *domain;
2551 struct dma_ops_domain *dma_dom;
2552 u16 devid = get_device_id(dev);
2554 if (!check_device(dev))
2555 return ERR_PTR(-EINVAL);
2557 domain = domain_for_device(dev);
2558 if (domain != NULL && !dma_ops_domain(domain))
2559 return ERR_PTR(-EBUSY);
2564 /* Device not bound yet - bind it */
2565 dma_dom = find_protection_domain(devid);
2567 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2568 attach_device(dev, &dma_dom->domain);
2569 DUMP_printk("Using protection domain %d for device %s\n",
2570 dma_dom->domain.id, dev_name(dev));
2572 return &dma_dom->domain;
2575 static void update_device_table(struct protection_domain *domain)
2577 struct iommu_dev_data *dev_data;
2579 list_for_each_entry(dev_data, &domain->dev_list, list)
2580 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2583 static void update_domain(struct protection_domain *domain)
2585 if (!domain->updated)
2588 update_device_table(domain);
2590 domain_flush_devices(domain);
2591 domain_flush_tlb_pde(domain);
2593 domain->updated = false;
2597 * This function fetches the PTE for a given address in the aperture
2599 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2600 unsigned long address)
2602 struct aperture_range *aperture;
2603 u64 *pte, *pte_page;
2605 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2609 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2611 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2613 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2615 pte += PM_LEVEL_INDEX(0, address);
2617 update_domain(&dom->domain);
2623 * This is the generic map function. It maps one 4kb page at paddr to
2624 * the given address in the DMA address space for the domain.
2626 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2627 unsigned long address,
2633 WARN_ON(address > dom->aperture_size);
2637 pte = dma_ops_get_pte(dom, address);
2639 return DMA_ERROR_CODE;
2641 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2643 if (direction == DMA_TO_DEVICE)
2644 __pte |= IOMMU_PTE_IR;
2645 else if (direction == DMA_FROM_DEVICE)
2646 __pte |= IOMMU_PTE_IW;
2647 else if (direction == DMA_BIDIRECTIONAL)
2648 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2654 return (dma_addr_t)address;
2658 * The generic unmapping function for on page in the DMA address space.
2660 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2661 unsigned long address)
2663 struct aperture_range *aperture;
2666 if (address >= dom->aperture_size)
2669 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2673 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2677 pte += PM_LEVEL_INDEX(0, address);
2685 * This function contains common code for mapping of a physically
2686 * contiguous memory region into DMA address space. It is used by all
2687 * mapping functions provided with this IOMMU driver.
2688 * Must be called with the domain lock held.
2690 static dma_addr_t __map_single(struct device *dev,
2691 struct dma_ops_domain *dma_dom,
2698 dma_addr_t offset = paddr & ~PAGE_MASK;
2699 dma_addr_t address, start, ret;
2701 unsigned long align_mask = 0;
2704 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2707 INC_STATS_COUNTER(total_map_requests);
2710 INC_STATS_COUNTER(cross_page);
2713 align_mask = (1UL << get_order(size)) - 1;
2716 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2718 if (unlikely(address == DMA_ERROR_CODE)) {
2720 * setting next_address here will let the address
2721 * allocator only scan the new allocated range in the
2722 * first run. This is a small optimization.
2724 dma_dom->next_address = dma_dom->aperture_size;
2726 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2730 * aperture was successfully enlarged by 128 MB, try
2737 for (i = 0; i < pages; ++i) {
2738 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2739 if (ret == DMA_ERROR_CODE)
2747 ADD_STATS_COUNTER(alloced_io_mem, size);
2749 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2750 domain_flush_tlb(&dma_dom->domain);
2751 dma_dom->need_flush = false;
2752 } else if (unlikely(amd_iommu_np_cache))
2753 domain_flush_pages(&dma_dom->domain, address, size);
2760 for (--i; i >= 0; --i) {
2762 dma_ops_domain_unmap(dma_dom, start);
2765 dma_ops_free_addresses(dma_dom, address, pages);
2767 return DMA_ERROR_CODE;
2771 * Does the reverse of the __map_single function. Must be called with
2772 * the domain lock held too
2774 static void __unmap_single(struct dma_ops_domain *dma_dom,
2775 dma_addr_t dma_addr,
2779 dma_addr_t flush_addr;
2780 dma_addr_t i, start;
2783 if ((dma_addr == DMA_ERROR_CODE) ||
2784 (dma_addr + size > dma_dom->aperture_size))
2787 flush_addr = dma_addr;
2788 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2789 dma_addr &= PAGE_MASK;
2792 for (i = 0; i < pages; ++i) {
2793 dma_ops_domain_unmap(dma_dom, start);
2797 SUB_STATS_COUNTER(alloced_io_mem, size);
2799 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2801 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2802 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2803 dma_dom->need_flush = false;
2808 * The exported map_single function for dma_ops.
2810 static dma_addr_t map_page(struct device *dev, struct page *page,
2811 unsigned long offset, size_t size,
2812 enum dma_data_direction dir,
2813 struct dma_attrs *attrs)
2815 unsigned long flags;
2816 struct protection_domain *domain;
2819 phys_addr_t paddr = page_to_phys(page) + offset;
2821 INC_STATS_COUNTER(cnt_map_single);
2823 domain = get_domain(dev);
2824 if (PTR_ERR(domain) == -EINVAL)
2825 return (dma_addr_t)paddr;
2826 else if (IS_ERR(domain))
2827 return DMA_ERROR_CODE;
2829 dma_mask = *dev->dma_mask;
2831 spin_lock_irqsave(&domain->lock, flags);
2833 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2835 if (addr == DMA_ERROR_CODE)
2838 domain_flush_complete(domain);
2841 spin_unlock_irqrestore(&domain->lock, flags);
2847 * The exported unmap_single function for dma_ops.
2849 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2850 enum dma_data_direction dir, struct dma_attrs *attrs)
2852 unsigned long flags;
2853 struct protection_domain *domain;
2855 INC_STATS_COUNTER(cnt_unmap_single);
2857 domain = get_domain(dev);
2861 spin_lock_irqsave(&domain->lock, flags);
2863 __unmap_single(domain->priv, dma_addr, size, dir);
2865 domain_flush_complete(domain);
2867 spin_unlock_irqrestore(&domain->lock, flags);
2871 * The exported map_sg function for dma_ops (handles scatter-gather
2874 static int map_sg(struct device *dev, struct scatterlist *sglist,
2875 int nelems, enum dma_data_direction dir,
2876 struct dma_attrs *attrs)
2878 unsigned long flags;
2879 struct protection_domain *domain;
2881 struct scatterlist *s;
2883 int mapped_elems = 0;
2886 INC_STATS_COUNTER(cnt_map_sg);
2888 domain = get_domain(dev);
2892 dma_mask = *dev->dma_mask;
2894 spin_lock_irqsave(&domain->lock, flags);
2896 for_each_sg(sglist, s, nelems, i) {
2899 s->dma_address = __map_single(dev, domain->priv,
2900 paddr, s->length, dir, false,
2903 if (s->dma_address) {
2904 s->dma_length = s->length;
2910 domain_flush_complete(domain);
2913 spin_unlock_irqrestore(&domain->lock, flags);
2915 return mapped_elems;
2917 for_each_sg(sglist, s, mapped_elems, i) {
2919 __unmap_single(domain->priv, s->dma_address,
2920 s->dma_length, dir);
2921 s->dma_address = s->dma_length = 0;
2930 * The exported map_sg function for dma_ops (handles scatter-gather
2933 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2934 int nelems, enum dma_data_direction dir,
2935 struct dma_attrs *attrs)
2937 unsigned long flags;
2938 struct protection_domain *domain;
2939 struct scatterlist *s;
2942 INC_STATS_COUNTER(cnt_unmap_sg);
2944 domain = get_domain(dev);
2948 spin_lock_irqsave(&domain->lock, flags);
2950 for_each_sg(sglist, s, nelems, i) {
2951 __unmap_single(domain->priv, s->dma_address,
2952 s->dma_length, dir);
2953 s->dma_address = s->dma_length = 0;
2956 domain_flush_complete(domain);
2958 spin_unlock_irqrestore(&domain->lock, flags);
2962 * The exported alloc_coherent function for dma_ops.
2964 static void *alloc_coherent(struct device *dev, size_t size,
2965 dma_addr_t *dma_addr, gfp_t flag,
2966 struct dma_attrs *attrs)
2968 unsigned long flags;
2970 struct protection_domain *domain;
2972 u64 dma_mask = dev->coherent_dma_mask;
2974 INC_STATS_COUNTER(cnt_alloc_coherent);
2976 domain = get_domain(dev);
2977 if (PTR_ERR(domain) == -EINVAL) {
2978 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2979 *dma_addr = __pa(virt_addr);
2981 } else if (IS_ERR(domain))
2984 dma_mask = dev->coherent_dma_mask;
2985 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2988 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2992 paddr = virt_to_phys(virt_addr);
2995 dma_mask = *dev->dma_mask;
2997 spin_lock_irqsave(&domain->lock, flags);
2999 *dma_addr = __map_single(dev, domain->priv, paddr,
3000 size, DMA_BIDIRECTIONAL, true, dma_mask);
3002 if (*dma_addr == DMA_ERROR_CODE) {
3003 spin_unlock_irqrestore(&domain->lock, flags);
3007 domain_flush_complete(domain);
3009 spin_unlock_irqrestore(&domain->lock, flags);
3015 free_pages((unsigned long)virt_addr, get_order(size));
3021 * The exported free_coherent function for dma_ops.
3023 static void free_coherent(struct device *dev, size_t size,
3024 void *virt_addr, dma_addr_t dma_addr,
3025 struct dma_attrs *attrs)
3027 unsigned long flags;
3028 struct protection_domain *domain;
3030 INC_STATS_COUNTER(cnt_free_coherent);
3032 domain = get_domain(dev);
3036 spin_lock_irqsave(&domain->lock, flags);
3038 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
3040 domain_flush_complete(domain);
3042 spin_unlock_irqrestore(&domain->lock, flags);
3045 free_pages((unsigned long)virt_addr, get_order(size));
3049 * This function is called by the DMA layer to find out if we can handle a
3050 * particular device. It is part of the dma_ops.
3052 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3054 return check_device(dev);
3058 * The function for pre-allocating protection domains.
3060 * If the driver core informs the DMA layer if a driver grabs a device
3061 * we don't need to preallocate the protection domains anymore.
3062 * For now we have to.
3064 static void __init prealloc_protection_domains(void)
3066 struct iommu_dev_data *dev_data;
3067 struct dma_ops_domain *dma_dom;
3068 struct pci_dev *dev = NULL;
3071 for_each_pci_dev(dev) {
3073 /* Do we handle this device? */
3074 if (!check_device(&dev->dev))
3077 dev_data = get_dev_data(&dev->dev);
3078 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3079 /* Make sure passthrough domain is allocated */
3080 alloc_passthrough_domain();
3081 dev_data->passthrough = true;
3082 attach_device(&dev->dev, pt_domain);
3083 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3084 dev_name(&dev->dev));
3087 /* Is there already any domain for it? */
3088 if (domain_for_device(&dev->dev))
3091 devid = get_device_id(&dev->dev);
3093 dma_dom = dma_ops_domain_alloc();
3096 init_unity_mappings_for_device(dma_dom, devid);
3097 dma_dom->target_dev = devid;
3099 attach_device(&dev->dev, &dma_dom->domain);
3101 list_add_tail(&dma_dom->list, &iommu_pd_list);
3105 static struct dma_map_ops amd_iommu_dma_ops = {
3106 .alloc = alloc_coherent,
3107 .free = free_coherent,
3108 .map_page = map_page,
3109 .unmap_page = unmap_page,
3111 .unmap_sg = unmap_sg,
3112 .dma_supported = amd_iommu_dma_supported,
3115 static unsigned device_dma_ops_init(void)
3117 struct iommu_dev_data *dev_data;
3118 struct pci_dev *pdev = NULL;
3119 unsigned unhandled = 0;
3121 for_each_pci_dev(pdev) {
3122 if (!check_device(&pdev->dev)) {
3124 iommu_ignore_device(&pdev->dev);
3130 dev_data = get_dev_data(&pdev->dev);
3132 if (!dev_data->passthrough)
3133 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3135 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3142 * The function which clues the AMD IOMMU driver into dma_ops.
3145 void __init amd_iommu_init_api(void)
3147 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3150 int __init amd_iommu_init_dma_ops(void)
3152 struct amd_iommu *iommu;
3156 * first allocate a default protection domain for every IOMMU we
3157 * found in the system. Devices not assigned to any other
3158 * protection domain will be assigned to the default one.
3160 for_each_iommu(iommu) {
3161 iommu->default_dom = dma_ops_domain_alloc();
3162 if (iommu->default_dom == NULL)
3164 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3165 ret = iommu_init_unity_mappings(iommu);
3171 * Pre-allocate the protection domains for each device.
3173 prealloc_protection_domains();
3178 /* Make the driver finally visible to the drivers */
3179 unhandled = device_dma_ops_init();
3180 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3181 /* There are unhandled devices - initialize swiotlb for them */
3185 amd_iommu_stats_init();
3187 if (amd_iommu_unmap_flush)
3188 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3190 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3196 for_each_iommu(iommu) {
3197 dma_ops_domain_free(iommu->default_dom);
3203 /*****************************************************************************
3205 * The following functions belong to the exported interface of AMD IOMMU
3207 * This interface allows access to lower level functions of the IOMMU
3208 * like protection domain handling and assignement of devices to domains
3209 * which is not possible with the dma_ops interface.
3211 *****************************************************************************/
3213 static void cleanup_domain(struct protection_domain *domain)
3215 struct iommu_dev_data *dev_data, *next;
3216 unsigned long flags;
3218 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3220 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3221 __detach_device(dev_data);
3222 atomic_set(&dev_data->bind, 0);
3225 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3228 static void protection_domain_free(struct protection_domain *domain)
3233 del_domain_from_list(domain);
3236 domain_id_free(domain->id);
3241 static struct protection_domain *protection_domain_alloc(void)
3243 struct protection_domain *domain;
3245 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3249 spin_lock_init(&domain->lock);
3250 mutex_init(&domain->api_lock);
3251 domain->id = domain_id_alloc();
3254 INIT_LIST_HEAD(&domain->dev_list);
3256 add_domain_to_list(domain);
3266 static int __init alloc_passthrough_domain(void)
3268 if (pt_domain != NULL)
3271 /* allocate passthrough domain */
3272 pt_domain = protection_domain_alloc();
3276 pt_domain->mode = PAGE_MODE_NONE;
3280 static int amd_iommu_domain_init(struct iommu_domain *dom)
3282 struct protection_domain *domain;
3284 domain = protection_domain_alloc();
3288 domain->mode = PAGE_MODE_3_LEVEL;
3289 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3290 if (!domain->pt_root)
3293 domain->iommu_domain = dom;
3297 dom->geometry.aperture_start = 0;
3298 dom->geometry.aperture_end = ~0ULL;
3299 dom->geometry.force_aperture = true;
3304 protection_domain_free(domain);
3309 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3311 struct protection_domain *domain = dom->priv;
3316 if (domain->dev_cnt > 0)
3317 cleanup_domain(domain);
3319 BUG_ON(domain->dev_cnt != 0);
3321 if (domain->mode != PAGE_MODE_NONE)
3322 free_pagetable(domain);
3324 if (domain->flags & PD_IOMMUV2_MASK)
3325 free_gcr3_table(domain);
3327 protection_domain_free(domain);
3332 static void amd_iommu_detach_device(struct iommu_domain *dom,
3335 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3336 struct amd_iommu *iommu;
3339 if (!check_device(dev))
3342 devid = get_device_id(dev);
3344 if (dev_data->domain != NULL)
3347 iommu = amd_iommu_rlookup_table[devid];
3351 iommu_completion_wait(iommu);
3354 static int amd_iommu_attach_device(struct iommu_domain *dom,
3357 struct protection_domain *domain = dom->priv;
3358 struct iommu_dev_data *dev_data;
3359 struct amd_iommu *iommu;
3362 if (!check_device(dev))
3365 dev_data = dev->archdata.iommu;
3367 iommu = amd_iommu_rlookup_table[dev_data->devid];
3371 if (dev_data->domain)
3374 ret = attach_device(dev, domain);
3376 iommu_completion_wait(iommu);
3381 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3382 phys_addr_t paddr, size_t page_size, int iommu_prot)
3384 struct protection_domain *domain = dom->priv;
3388 if (domain->mode == PAGE_MODE_NONE)
3391 if (iommu_prot & IOMMU_READ)
3392 prot |= IOMMU_PROT_IR;
3393 if (iommu_prot & IOMMU_WRITE)
3394 prot |= IOMMU_PROT_IW;
3396 mutex_lock(&domain->api_lock);
3397 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3398 mutex_unlock(&domain->api_lock);
3403 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3406 struct protection_domain *domain = dom->priv;
3409 if (domain->mode == PAGE_MODE_NONE)
3412 mutex_lock(&domain->api_lock);
3413 unmap_size = iommu_unmap_page(domain, iova, page_size);
3414 mutex_unlock(&domain->api_lock);
3416 domain_flush_tlb_pde(domain);
3421 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3424 struct protection_domain *domain = dom->priv;
3425 unsigned long offset_mask;
3429 if (domain->mode == PAGE_MODE_NONE)
3432 pte = fetch_pte(domain, iova);
3434 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3437 if (PM_PTE_LEVEL(*pte) == 0)
3438 offset_mask = PAGE_SIZE - 1;
3440 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3442 __pte = *pte & PM_ADDR_MASK;
3443 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3448 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3452 case IOMMU_CAP_CACHE_COHERENCY:
3454 case IOMMU_CAP_INTR_REMAP:
3455 return irq_remapping_enabled;
3461 static struct iommu_ops amd_iommu_ops = {
3462 .domain_init = amd_iommu_domain_init,
3463 .domain_destroy = amd_iommu_domain_destroy,
3464 .attach_dev = amd_iommu_attach_device,
3465 .detach_dev = amd_iommu_detach_device,
3466 .map = amd_iommu_map,
3467 .unmap = amd_iommu_unmap,
3468 .iova_to_phys = amd_iommu_iova_to_phys,
3469 .domain_has_cap = amd_iommu_domain_has_cap,
3470 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3473 /*****************************************************************************
3475 * The next functions do a basic initialization of IOMMU for pass through
3478 * In passthrough mode the IOMMU is initialized and enabled but not used for
3479 * DMA-API translation.
3481 *****************************************************************************/
3483 int __init amd_iommu_init_passthrough(void)
3485 struct iommu_dev_data *dev_data;
3486 struct pci_dev *dev = NULL;
3487 struct amd_iommu *iommu;
3491 ret = alloc_passthrough_domain();
3495 for_each_pci_dev(dev) {
3496 if (!check_device(&dev->dev))
3499 dev_data = get_dev_data(&dev->dev);
3500 dev_data->passthrough = true;
3502 devid = get_device_id(&dev->dev);
3504 iommu = amd_iommu_rlookup_table[devid];
3508 attach_device(&dev->dev, pt_domain);
3511 amd_iommu_stats_init();
3513 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3518 /* IOMMUv2 specific functions */
3519 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3521 return atomic_notifier_chain_register(&ppr_notifier, nb);
3523 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3525 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3527 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3529 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3531 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3533 struct protection_domain *domain = dom->priv;
3534 unsigned long flags;
3536 spin_lock_irqsave(&domain->lock, flags);
3538 /* Update data structure */
3539 domain->mode = PAGE_MODE_NONE;
3540 domain->updated = true;
3542 /* Make changes visible to IOMMUs */
3543 update_domain(domain);
3545 /* Page-table is not visible to IOMMU anymore, so free it */
3546 free_pagetable(domain);
3548 spin_unlock_irqrestore(&domain->lock, flags);
3550 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3552 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3554 struct protection_domain *domain = dom->priv;
3555 unsigned long flags;
3558 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3561 /* Number of GCR3 table levels required */
3562 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3565 if (levels > amd_iommu_max_glx_val)
3568 spin_lock_irqsave(&domain->lock, flags);
3571 * Save us all sanity checks whether devices already in the
3572 * domain support IOMMUv2. Just force that the domain has no
3573 * devices attached when it is switched into IOMMUv2 mode.
3576 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3580 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3581 if (domain->gcr3_tbl == NULL)
3584 domain->glx = levels;
3585 domain->flags |= PD_IOMMUV2_MASK;
3586 domain->updated = true;
3588 update_domain(domain);
3593 spin_unlock_irqrestore(&domain->lock, flags);
3597 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3599 static int __flush_pasid(struct protection_domain *domain, int pasid,
3600 u64 address, bool size)
3602 struct iommu_dev_data *dev_data;
3603 struct iommu_cmd cmd;
3606 if (!(domain->flags & PD_IOMMUV2_MASK))
3609 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3612 * IOMMU TLB needs to be flushed before Device TLB to
3613 * prevent device TLB refill from IOMMU TLB
3615 for (i = 0; i < amd_iommus_present; ++i) {
3616 if (domain->dev_iommu[i] == 0)
3619 ret = iommu_queue_command(amd_iommus[i], &cmd);
3624 /* Wait until IOMMU TLB flushes are complete */
3625 domain_flush_complete(domain);
3627 /* Now flush device TLBs */
3628 list_for_each_entry(dev_data, &domain->dev_list, list) {
3629 struct amd_iommu *iommu;
3632 BUG_ON(!dev_data->ats.enabled);
3634 qdep = dev_data->ats.qdep;
3635 iommu = amd_iommu_rlookup_table[dev_data->devid];
3637 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3638 qdep, address, size);
3640 ret = iommu_queue_command(iommu, &cmd);
3645 /* Wait until all device TLBs are flushed */
3646 domain_flush_complete(domain);
3655 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3658 INC_STATS_COUNTER(invalidate_iotlb);
3660 return __flush_pasid(domain, pasid, address, false);
3663 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3666 struct protection_domain *domain = dom->priv;
3667 unsigned long flags;
3670 spin_lock_irqsave(&domain->lock, flags);
3671 ret = __amd_iommu_flush_page(domain, pasid, address);
3672 spin_unlock_irqrestore(&domain->lock, flags);
3676 EXPORT_SYMBOL(amd_iommu_flush_page);
3678 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3680 INC_STATS_COUNTER(invalidate_iotlb_all);
3682 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3686 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3688 struct protection_domain *domain = dom->priv;
3689 unsigned long flags;
3692 spin_lock_irqsave(&domain->lock, flags);
3693 ret = __amd_iommu_flush_tlb(domain, pasid);
3694 spin_unlock_irqrestore(&domain->lock, flags);
3698 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3700 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3707 index = (pasid >> (9 * level)) & 0x1ff;
3713 if (!(*pte & GCR3_VALID)) {
3717 root = (void *)get_zeroed_page(GFP_ATOMIC);
3721 *pte = __pa(root) | GCR3_VALID;
3724 root = __va(*pte & PAGE_MASK);
3732 static int __set_gcr3(struct protection_domain *domain, int pasid,
3737 if (domain->mode != PAGE_MODE_NONE)
3740 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3744 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3746 return __amd_iommu_flush_tlb(domain, pasid);
3749 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3753 if (domain->mode != PAGE_MODE_NONE)
3756 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3762 return __amd_iommu_flush_tlb(domain, pasid);
3765 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3768 struct protection_domain *domain = dom->priv;
3769 unsigned long flags;
3772 spin_lock_irqsave(&domain->lock, flags);
3773 ret = __set_gcr3(domain, pasid, cr3);
3774 spin_unlock_irqrestore(&domain->lock, flags);
3778 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3780 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3782 struct protection_domain *domain = dom->priv;
3783 unsigned long flags;
3786 spin_lock_irqsave(&domain->lock, flags);
3787 ret = __clear_gcr3(domain, pasid);
3788 spin_unlock_irqrestore(&domain->lock, flags);
3792 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3794 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3795 int status, int tag)
3797 struct iommu_dev_data *dev_data;
3798 struct amd_iommu *iommu;
3799 struct iommu_cmd cmd;
3801 INC_STATS_COUNTER(complete_ppr);
3803 dev_data = get_dev_data(&pdev->dev);
3804 iommu = amd_iommu_rlookup_table[dev_data->devid];
3806 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3807 tag, dev_data->pri_tlp);
3809 return iommu_queue_command(iommu, &cmd);
3811 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3813 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3815 struct protection_domain *domain;
3817 domain = get_domain(&pdev->dev);
3821 /* Only return IOMMUv2 domains */
3822 if (!(domain->flags & PD_IOMMUV2_MASK))
3825 return domain->iommu_domain;
3827 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3829 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3831 struct iommu_dev_data *dev_data;
3833 if (!amd_iommu_v2_supported())
3836 dev_data = get_dev_data(&pdev->dev);
3837 dev_data->errata |= (1 << erratum);
3839 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3841 int amd_iommu_device_info(struct pci_dev *pdev,
3842 struct amd_iommu_device_info *info)
3847 if (pdev == NULL || info == NULL)
3850 if (!amd_iommu_v2_supported())
3853 memset(info, 0, sizeof(*info));
3855 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3857 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3859 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3861 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3863 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3867 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3868 max_pasids = min(max_pasids, (1 << 20));
3870 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3871 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3873 features = pci_pasid_features(pdev);
3874 if (features & PCI_PASID_CAP_EXEC)
3875 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3876 if (features & PCI_PASID_CAP_PRIV)
3877 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3882 EXPORT_SYMBOL(amd_iommu_device_info);
3884 #ifdef CONFIG_IRQ_REMAP
3886 /*****************************************************************************
3888 * Interrupt Remapping Implementation
3890 *****************************************************************************/
3907 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3908 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3909 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3910 #define DTE_IRQ_REMAP_ENABLE 1ULL
3912 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3916 dte = amd_iommu_dev_table[devid].data[2];
3917 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3918 dte |= virt_to_phys(table->table);
3919 dte |= DTE_IRQ_REMAP_INTCTL;
3920 dte |= DTE_IRQ_TABLE_LEN;
3921 dte |= DTE_IRQ_REMAP_ENABLE;
3923 amd_iommu_dev_table[devid].data[2] = dte;
3926 #define IRTE_ALLOCATED (~1U)
3928 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3930 struct irq_remap_table *table = NULL;
3931 struct amd_iommu *iommu;
3932 unsigned long flags;
3935 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3937 iommu = amd_iommu_rlookup_table[devid];
3941 table = irq_lookup_table[devid];
3945 alias = amd_iommu_alias_table[devid];
3946 table = irq_lookup_table[alias];
3948 irq_lookup_table[devid] = table;
3949 set_dte_irq_entry(devid, table);
3950 iommu_flush_dte(iommu, devid);
3954 /* Nothing there yet, allocate new irq remapping table */
3955 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3959 /* Initialize table spin-lock */
3960 spin_lock_init(&table->lock);
3963 /* Keep the first 32 indexes free for IOAPIC interrupts */
3964 table->min_index = 32;
3966 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3967 if (!table->table) {
3973 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3978 for (i = 0; i < 32; ++i)
3979 table->table[i] = IRTE_ALLOCATED;
3982 irq_lookup_table[devid] = table;
3983 set_dte_irq_entry(devid, table);
3984 iommu_flush_dte(iommu, devid);
3985 if (devid != alias) {
3986 irq_lookup_table[alias] = table;
3987 set_dte_irq_entry(devid, table);
3988 iommu_flush_dte(iommu, alias);
3992 iommu_completion_wait(iommu);
3995 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
4000 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
4002 struct irq_remap_table *table;
4003 unsigned long flags;
4006 table = get_irq_table(devid, false);
4010 spin_lock_irqsave(&table->lock, flags);
4012 /* Scan table for free entries */
4013 for (c = 0, index = table->min_index;
4014 index < MAX_IRQS_PER_TABLE;
4016 if (table->table[index] == 0)
4022 struct irq_2_irte *irte_info;
4025 table->table[index - c + 1] = IRTE_ALLOCATED;
4030 irte_info = &cfg->irq_2_irte;
4031 irte_info->devid = devid;
4032 irte_info->index = index;
4041 spin_unlock_irqrestore(&table->lock, flags);
4046 static int get_irte(u16 devid, int index, union irte *irte)
4048 struct irq_remap_table *table;
4049 unsigned long flags;
4051 table = get_irq_table(devid, false);
4055 spin_lock_irqsave(&table->lock, flags);
4056 irte->val = table->table[index];
4057 spin_unlock_irqrestore(&table->lock, flags);
4062 static int modify_irte(u16 devid, int index, union irte irte)
4064 struct irq_remap_table *table;
4065 struct amd_iommu *iommu;
4066 unsigned long flags;
4068 iommu = amd_iommu_rlookup_table[devid];
4072 table = get_irq_table(devid, false);
4076 spin_lock_irqsave(&table->lock, flags);
4077 table->table[index] = irte.val;
4078 spin_unlock_irqrestore(&table->lock, flags);
4080 iommu_flush_irt(iommu, devid);
4081 iommu_completion_wait(iommu);
4086 static void free_irte(u16 devid, int index)
4088 struct irq_remap_table *table;
4089 struct amd_iommu *iommu;
4090 unsigned long flags;
4092 iommu = amd_iommu_rlookup_table[devid];
4096 table = get_irq_table(devid, false);
4100 spin_lock_irqsave(&table->lock, flags);
4101 table->table[index] = 0;
4102 spin_unlock_irqrestore(&table->lock, flags);
4104 iommu_flush_irt(iommu, devid);
4105 iommu_completion_wait(iommu);
4108 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4109 unsigned int destination, int vector,
4110 struct io_apic_irq_attr *attr)
4112 struct irq_remap_table *table;
4113 struct irq_2_irte *irte_info;
4114 struct irq_cfg *cfg;
4121 cfg = irq_get_chip_data(irq);
4125 irte_info = &cfg->irq_2_irte;
4126 ioapic_id = mpc_ioapic_id(attr->ioapic);
4127 devid = get_ioapic_devid(ioapic_id);
4132 table = get_irq_table(devid, true);
4136 index = attr->ioapic_pin;
4138 /* Setup IRQ remapping info */
4140 irte_info->devid = devid;
4141 irte_info->index = index;
4143 /* Setup IRTE for IOMMU */
4145 irte.fields.vector = vector;
4146 irte.fields.int_type = apic->irq_delivery_mode;
4147 irte.fields.destination = destination;
4148 irte.fields.dm = apic->irq_dest_mode;
4149 irte.fields.valid = 1;
4151 ret = modify_irte(devid, index, irte);
4155 /* Setup IOAPIC entry */
4156 memset(entry, 0, sizeof(*entry));
4158 entry->vector = index;
4160 entry->trigger = attr->trigger;
4161 entry->polarity = attr->polarity;
4164 * Mask level triggered irqs.
4172 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4175 struct irq_2_irte *irte_info;
4176 unsigned int dest, irq;
4177 struct irq_cfg *cfg;
4181 if (!config_enabled(CONFIG_SMP))
4184 cfg = data->chip_data;
4186 irte_info = &cfg->irq_2_irte;
4188 if (!cpumask_intersects(mask, cpu_online_mask))
4191 if (get_irte(irte_info->devid, irte_info->index, &irte))
4194 if (assign_irq_vector(irq, cfg, mask))
4197 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4199 if (assign_irq_vector(irq, cfg, data->affinity))
4200 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4204 irte.fields.vector = cfg->vector;
4205 irte.fields.destination = dest;
4207 modify_irte(irte_info->devid, irte_info->index, irte);
4209 if (cfg->move_in_progress)
4210 send_cleanup_vector(cfg);
4212 cpumask_copy(data->affinity, mask);
4217 static int free_irq(int irq)
4219 struct irq_2_irte *irte_info;
4220 struct irq_cfg *cfg;
4222 cfg = irq_get_chip_data(irq);
4226 irte_info = &cfg->irq_2_irte;
4228 free_irte(irte_info->devid, irte_info->index);
4233 static void compose_msi_msg(struct pci_dev *pdev,
4234 unsigned int irq, unsigned int dest,
4235 struct msi_msg *msg, u8 hpet_id)
4237 struct irq_2_irte *irte_info;
4238 struct irq_cfg *cfg;
4241 cfg = irq_get_chip_data(irq);
4245 irte_info = &cfg->irq_2_irte;
4248 irte.fields.vector = cfg->vector;
4249 irte.fields.int_type = apic->irq_delivery_mode;
4250 irte.fields.destination = dest;
4251 irte.fields.dm = apic->irq_dest_mode;
4252 irte.fields.valid = 1;
4254 modify_irte(irte_info->devid, irte_info->index, irte);
4256 msg->address_hi = MSI_ADDR_BASE_HI;
4257 msg->address_lo = MSI_ADDR_BASE_LO;
4258 msg->data = irte_info->index;
4261 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4263 struct irq_cfg *cfg;
4270 cfg = irq_get_chip_data(irq);
4274 devid = get_device_id(&pdev->dev);
4275 index = alloc_irq_index(cfg, devid, nvec);
4277 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4280 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4281 int index, int offset)
4283 struct irq_2_irte *irte_info;
4284 struct irq_cfg *cfg;
4290 cfg = irq_get_chip_data(irq);
4294 if (index >= MAX_IRQS_PER_TABLE)
4297 devid = get_device_id(&pdev->dev);
4298 irte_info = &cfg->irq_2_irte;
4301 irte_info->devid = devid;
4302 irte_info->index = index + offset;
4307 static int setup_hpet_msi(unsigned int irq, unsigned int id)
4309 struct irq_2_irte *irte_info;
4310 struct irq_cfg *cfg;
4313 cfg = irq_get_chip_data(irq);
4317 irte_info = &cfg->irq_2_irte;
4318 devid = get_hpet_devid(id);
4322 index = alloc_irq_index(cfg, devid, 1);
4327 irte_info->devid = devid;
4328 irte_info->index = index;
4333 struct irq_remap_ops amd_iommu_irq_ops = {
4334 .supported = amd_iommu_supported,
4335 .prepare = amd_iommu_prepare,
4336 .enable = amd_iommu_enable,
4337 .disable = amd_iommu_disable,
4338 .reenable = amd_iommu_reenable,
4339 .enable_faulting = amd_iommu_enable_faulting,
4340 .setup_ioapic_entry = setup_ioapic_entry,
4341 .set_affinity = set_affinity,
4342 .free_irq = free_irq,
4343 .compose_msi_msg = compose_msi_msg,
4344 .msi_alloc_irq = msi_alloc_irq,
4345 .msi_setup_irq = msi_setup_irq,
4346 .setup_hpet_msi = setup_hpet_msi,