2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
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/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/pci-ats.h>
25 #include <linux/bitmap.h>
26 #include <linux/slab.h>
27 #include <linux/debugfs.h>
28 #include <linux/scatterlist.h>
29 #include <linux/dma-mapping.h>
30 #include <linux/iommu-helper.h>
31 #include <linux/iommu.h>
32 #include <linux/delay.h>
33 #include <linux/amd-iommu.h>
34 #include <linux/notifier.h>
35 #include <linux/export.h>
36 #include <linux/irq.h>
37 #include <linux/msi.h>
38 #include <linux/dma-contiguous.h>
39 #include <linux/irqdomain.h>
40 #include <linux/percpu.h>
41 #include <asm/irq_remapping.h>
42 #include <asm/io_apic.h>
44 #include <asm/hw_irq.h>
45 #include <asm/msidef.h>
46 #include <asm/proto.h>
47 #include <asm/iommu.h>
51 #include "amd_iommu_proto.h"
52 #include "amd_iommu_types.h"
53 #include "irq_remapping.h"
55 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
57 #define LOOP_TIMEOUT 100000
60 * This bitmap is used to advertise the page sizes our hardware support
61 * to the IOMMU core, which will then use this information to split
62 * physically contiguous memory regions it is mapping into page sizes
65 * 512GB Pages are not supported due to a hardware bug
67 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
69 static DEFINE_RWLOCK(amd_iommu_devtable_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);
77 LIST_HEAD(acpihid_map);
80 * Domain for untranslated devices - only allocated
81 * if iommu=pt passed on kernel cmd line.
83 static const 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 * This struct contains device specific data for the IOMMU
93 struct iommu_dev_data {
94 struct list_head list; /* For domain->dev_list */
95 struct list_head dev_data_list; /* For global dev_data_list */
96 struct protection_domain *domain; /* Domain the device is bound to */
97 u16 devid; /* PCI Device ID */
98 u16 alias; /* Alias Device ID */
99 bool iommu_v2; /* Device can make use of IOMMUv2 */
100 bool passthrough; /* Device is identity mapped */
104 } ats; /* ATS state */
105 bool pri_tlp; /* PASID TLB required for
107 u32 errata; /* Bitmap for errata to apply */
111 * general struct to manage commands send to an IOMMU
117 struct kmem_cache *amd_iommu_irq_cache;
119 static void update_domain(struct protection_domain *domain);
120 static int protection_domain_init(struct protection_domain *domain);
121 static void detach_device(struct device *dev);
124 * For dynamic growth the aperture size is split into ranges of 128MB of
125 * DMA address space each. This struct represents one such range.
127 struct aperture_range {
129 spinlock_t bitmap_lock;
131 /* address allocation bitmap */
132 unsigned long *bitmap;
133 unsigned long offset;
134 unsigned long next_bit;
137 * Array of PTE pages for the aperture. In this array we save all the
138 * leaf pages of the domain page table used for the aperture. This way
139 * we don't need to walk the page table to find a specific PTE. We can
140 * just calculate its address in constant time.
146 * Data container for a dma_ops specific protection domain
148 struct dma_ops_domain {
149 /* generic protection domain information */
150 struct protection_domain domain;
152 /* size of the aperture for the mappings */
153 unsigned long aperture_size;
155 /* aperture index we start searching for free addresses */
156 u32 __percpu *next_index;
158 /* address space relevant data */
159 struct aperture_range *aperture[APERTURE_MAX_RANGES];
162 /****************************************************************************
166 ****************************************************************************/
168 static inline int match_hid_uid(struct device *dev,
169 struct acpihid_map_entry *entry)
171 const char *hid, *uid;
173 hid = acpi_device_hid(ACPI_COMPANION(dev));
174 uid = acpi_device_uid(ACPI_COMPANION(dev));
180 return strcmp(hid, entry->hid);
183 return strcmp(hid, entry->hid);
185 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
188 static inline u16 get_pci_device_id(struct device *dev)
190 struct pci_dev *pdev = to_pci_dev(dev);
192 return PCI_DEVID(pdev->bus->number, pdev->devfn);
195 static inline int get_acpihid_device_id(struct device *dev,
196 struct acpihid_map_entry **entry)
198 struct acpihid_map_entry *p;
200 list_for_each_entry(p, &acpihid_map, list) {
201 if (!match_hid_uid(dev, p)) {
210 static inline int get_device_id(struct device *dev)
215 devid = get_pci_device_id(dev);
217 devid = get_acpihid_device_id(dev, NULL);
222 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
224 return container_of(dom, struct protection_domain, domain);
227 static struct iommu_dev_data *alloc_dev_data(u16 devid)
229 struct iommu_dev_data *dev_data;
232 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
236 dev_data->devid = devid;
238 spin_lock_irqsave(&dev_data_list_lock, flags);
239 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
240 spin_unlock_irqrestore(&dev_data_list_lock, flags);
245 static struct iommu_dev_data *search_dev_data(u16 devid)
247 struct iommu_dev_data *dev_data;
250 spin_lock_irqsave(&dev_data_list_lock, flags);
251 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
252 if (dev_data->devid == devid)
259 spin_unlock_irqrestore(&dev_data_list_lock, flags);
264 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
266 *(u16 *)data = alias;
270 static u16 get_alias(struct device *dev)
272 struct pci_dev *pdev = to_pci_dev(dev);
273 u16 devid, ivrs_alias, pci_alias;
275 /* The callers make sure that get_device_id() does not fail here */
276 devid = get_device_id(dev);
277 ivrs_alias = amd_iommu_alias_table[devid];
278 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
280 if (ivrs_alias == pci_alias)
286 * The IVRS is fairly reliable in telling us about aliases, but it
287 * can't know about every screwy device. If we don't have an IVRS
288 * reported alias, use the PCI reported alias. In that case we may
289 * still need to initialize the rlookup and dev_table entries if the
290 * alias is to a non-existent device.
292 if (ivrs_alias == devid) {
293 if (!amd_iommu_rlookup_table[pci_alias]) {
294 amd_iommu_rlookup_table[pci_alias] =
295 amd_iommu_rlookup_table[devid];
296 memcpy(amd_iommu_dev_table[pci_alias].data,
297 amd_iommu_dev_table[devid].data,
298 sizeof(amd_iommu_dev_table[pci_alias].data));
304 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
305 "for device %s[%04x:%04x], kernel reported alias "
306 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
307 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
308 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
309 PCI_FUNC(pci_alias));
312 * If we don't have a PCI DMA alias and the IVRS alias is on the same
313 * bus, then the IVRS table may know about a quirk that we don't.
315 if (pci_alias == devid &&
316 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
317 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
318 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
319 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
326 static struct iommu_dev_data *find_dev_data(u16 devid)
328 struct iommu_dev_data *dev_data;
330 dev_data = search_dev_data(devid);
332 if (dev_data == NULL)
333 dev_data = alloc_dev_data(devid);
338 static struct iommu_dev_data *get_dev_data(struct device *dev)
340 return dev->archdata.iommu;
344 * Find or create an IOMMU group for a acpihid device.
346 static struct iommu_group *acpihid_device_group(struct device *dev)
348 struct acpihid_map_entry *p, *entry = NULL;
351 devid = get_acpihid_device_id(dev, &entry);
353 return ERR_PTR(devid);
355 list_for_each_entry(p, &acpihid_map, list) {
356 if ((devid == p->devid) && p->group)
357 entry->group = p->group;
361 entry->group = generic_device_group(dev);
366 static bool pci_iommuv2_capable(struct pci_dev *pdev)
368 static const int caps[] = {
371 PCI_EXT_CAP_ID_PASID,
375 for (i = 0; i < 3; ++i) {
376 pos = pci_find_ext_capability(pdev, caps[i]);
384 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
386 struct iommu_dev_data *dev_data;
388 dev_data = get_dev_data(&pdev->dev);
390 return dev_data->errata & (1 << erratum) ? true : false;
394 * This function actually applies the mapping to the page table of the
397 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
398 struct unity_map_entry *e)
402 for (addr = e->address_start; addr < e->address_end;
404 if (addr < dma_dom->aperture_size)
405 __set_bit(addr >> PAGE_SHIFT,
406 dma_dom->aperture[0]->bitmap);
411 * Inits the unity mappings required for a specific device
413 static void init_unity_mappings_for_device(struct device *dev,
414 struct dma_ops_domain *dma_dom)
416 struct unity_map_entry *e;
419 devid = get_device_id(dev);
423 list_for_each_entry(e, &amd_iommu_unity_map, list) {
424 if (!(devid >= e->devid_start && devid <= e->devid_end))
426 alloc_unity_mapping(dma_dom, e);
431 * This function checks if the driver got a valid device from the caller to
432 * avoid dereferencing invalid pointers.
434 static bool check_device(struct device *dev)
438 if (!dev || !dev->dma_mask)
441 devid = get_device_id(dev);
445 /* Out of our scope? */
446 if (devid > amd_iommu_last_bdf)
449 if (amd_iommu_rlookup_table[devid] == NULL)
455 static void init_iommu_group(struct device *dev)
457 struct dma_ops_domain *dma_domain;
458 struct iommu_domain *domain;
459 struct iommu_group *group;
461 group = iommu_group_get_for_dev(dev);
465 domain = iommu_group_default_domain(group);
469 dma_domain = to_pdomain(domain)->priv;
471 init_unity_mappings_for_device(dev, dma_domain);
473 iommu_group_put(group);
476 static int iommu_init_device(struct device *dev)
478 struct iommu_dev_data *dev_data;
481 if (dev->archdata.iommu)
484 devid = get_device_id(dev);
488 dev_data = find_dev_data(devid);
492 dev_data->alias = get_alias(dev);
494 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
495 struct amd_iommu *iommu;
497 iommu = amd_iommu_rlookup_table[dev_data->devid];
498 dev_data->iommu_v2 = iommu->is_iommu_v2;
501 dev->archdata.iommu = dev_data;
503 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
509 static void iommu_ignore_device(struct device *dev)
514 devid = get_device_id(dev);
518 alias = get_alias(dev);
520 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
521 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
523 amd_iommu_rlookup_table[devid] = NULL;
524 amd_iommu_rlookup_table[alias] = NULL;
527 static void iommu_uninit_device(struct device *dev)
530 struct iommu_dev_data *dev_data;
532 devid = get_device_id(dev);
536 dev_data = search_dev_data(devid);
540 if (dev_data->domain)
543 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
546 iommu_group_remove_device(dev);
549 dev->archdata.dma_ops = NULL;
552 * We keep dev_data around for unplugged devices and reuse it when the
553 * device is re-plugged - not doing so would introduce a ton of races.
557 /****************************************************************************
559 * Interrupt handling functions
561 ****************************************************************************/
563 static void dump_dte_entry(u16 devid)
567 for (i = 0; i < 4; ++i)
568 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
569 amd_iommu_dev_table[devid].data[i]);
572 static void dump_command(unsigned long phys_addr)
574 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
577 for (i = 0; i < 4; ++i)
578 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
581 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
583 int type, devid, domid, flags;
584 volatile u32 *event = __evt;
589 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
590 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
591 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
592 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
593 address = (u64)(((u64)event[3]) << 32) | event[2];
596 /* Did we hit the erratum? */
597 if (++count == LOOP_TIMEOUT) {
598 pr_err("AMD-Vi: No event written to event log\n");
605 printk(KERN_ERR "AMD-Vi: Event logged [");
608 case EVENT_TYPE_ILL_DEV:
609 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
610 "address=0x%016llx flags=0x%04x]\n",
611 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
613 dump_dte_entry(devid);
615 case EVENT_TYPE_IO_FAULT:
616 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
617 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
618 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
619 domid, address, flags);
621 case EVENT_TYPE_DEV_TAB_ERR:
622 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
623 "address=0x%016llx flags=0x%04x]\n",
624 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
627 case EVENT_TYPE_PAGE_TAB_ERR:
628 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
629 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
630 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
631 domid, address, flags);
633 case EVENT_TYPE_ILL_CMD:
634 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
635 dump_command(address);
637 case EVENT_TYPE_CMD_HARD_ERR:
638 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
639 "flags=0x%04x]\n", address, flags);
641 case EVENT_TYPE_IOTLB_INV_TO:
642 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
643 "address=0x%016llx]\n",
644 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
647 case EVENT_TYPE_INV_DEV_REQ:
648 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
649 "address=0x%016llx flags=0x%04x]\n",
650 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
654 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
657 memset(__evt, 0, 4 * sizeof(u32));
660 static void iommu_poll_events(struct amd_iommu *iommu)
664 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
665 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
667 while (head != tail) {
668 iommu_print_event(iommu, iommu->evt_buf + head);
669 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
672 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
675 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
677 struct amd_iommu_fault fault;
679 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
680 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
684 fault.address = raw[1];
685 fault.pasid = PPR_PASID(raw[0]);
686 fault.device_id = PPR_DEVID(raw[0]);
687 fault.tag = PPR_TAG(raw[0]);
688 fault.flags = PPR_FLAGS(raw[0]);
690 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
693 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
697 if (iommu->ppr_log == NULL)
700 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
701 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
703 while (head != tail) {
708 raw = (u64 *)(iommu->ppr_log + head);
711 * Hardware bug: Interrupt may arrive before the entry is
712 * written to memory. If this happens we need to wait for the
715 for (i = 0; i < LOOP_TIMEOUT; ++i) {
716 if (PPR_REQ_TYPE(raw[0]) != 0)
721 /* Avoid memcpy function-call overhead */
726 * To detect the hardware bug we need to clear the entry
729 raw[0] = raw[1] = 0UL;
731 /* Update head pointer of hardware ring-buffer */
732 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
733 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
735 /* Handle PPR entry */
736 iommu_handle_ppr_entry(iommu, entry);
738 /* Refresh ring-buffer information */
739 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
740 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
744 irqreturn_t amd_iommu_int_thread(int irq, void *data)
746 struct amd_iommu *iommu = (struct amd_iommu *) data;
747 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
749 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
750 /* Enable EVT and PPR interrupts again */
751 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
752 iommu->mmio_base + MMIO_STATUS_OFFSET);
754 if (status & MMIO_STATUS_EVT_INT_MASK) {
755 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
756 iommu_poll_events(iommu);
759 if (status & MMIO_STATUS_PPR_INT_MASK) {
760 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
761 iommu_poll_ppr_log(iommu);
765 * Hardware bug: ERBT1312
766 * When re-enabling interrupt (by writing 1
767 * to clear the bit), the hardware might also try to set
768 * the interrupt bit in the event status register.
769 * In this scenario, the bit will be set, and disable
770 * subsequent interrupts.
772 * Workaround: The IOMMU driver should read back the
773 * status register and check if the interrupt bits are cleared.
774 * If not, driver will need to go through the interrupt handler
775 * again and re-clear the bits
777 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
782 irqreturn_t amd_iommu_int_handler(int irq, void *data)
784 return IRQ_WAKE_THREAD;
787 /****************************************************************************
789 * IOMMU command queuing functions
791 ****************************************************************************/
793 static int wait_on_sem(volatile u64 *sem)
797 while (*sem == 0 && i < LOOP_TIMEOUT) {
802 if (i == LOOP_TIMEOUT) {
803 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
810 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
811 struct iommu_cmd *cmd,
816 target = iommu->cmd_buf + tail;
817 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
819 /* Copy command to buffer */
820 memcpy(target, cmd, sizeof(*cmd));
822 /* Tell the IOMMU about it */
823 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
826 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
828 WARN_ON(address & 0x7ULL);
830 memset(cmd, 0, sizeof(*cmd));
831 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
832 cmd->data[1] = upper_32_bits(__pa(address));
834 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
837 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
839 memset(cmd, 0, sizeof(*cmd));
840 cmd->data[0] = devid;
841 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
844 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
845 size_t size, u16 domid, int pde)
850 pages = iommu_num_pages(address, size, PAGE_SIZE);
855 * If we have to flush more than one page, flush all
856 * TLB entries for this domain
858 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
862 address &= PAGE_MASK;
864 memset(cmd, 0, sizeof(*cmd));
865 cmd->data[1] |= domid;
866 cmd->data[2] = lower_32_bits(address);
867 cmd->data[3] = upper_32_bits(address);
868 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
869 if (s) /* size bit - we flush more than one 4kb page */
870 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
871 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
872 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
875 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
876 u64 address, size_t size)
881 pages = iommu_num_pages(address, size, PAGE_SIZE);
886 * If we have to flush more than one page, flush all
887 * TLB entries for this domain
889 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
893 address &= PAGE_MASK;
895 memset(cmd, 0, sizeof(*cmd));
896 cmd->data[0] = devid;
897 cmd->data[0] |= (qdep & 0xff) << 24;
898 cmd->data[1] = devid;
899 cmd->data[2] = lower_32_bits(address);
900 cmd->data[3] = upper_32_bits(address);
901 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
903 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
906 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
907 u64 address, bool size)
909 memset(cmd, 0, sizeof(*cmd));
911 address &= ~(0xfffULL);
913 cmd->data[0] = pasid;
914 cmd->data[1] = domid;
915 cmd->data[2] = lower_32_bits(address);
916 cmd->data[3] = upper_32_bits(address);
917 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
918 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
920 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
921 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
924 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
925 int qdep, u64 address, bool size)
927 memset(cmd, 0, sizeof(*cmd));
929 address &= ~(0xfffULL);
931 cmd->data[0] = devid;
932 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
933 cmd->data[0] |= (qdep & 0xff) << 24;
934 cmd->data[1] = devid;
935 cmd->data[1] |= (pasid & 0xff) << 16;
936 cmd->data[2] = lower_32_bits(address);
937 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
938 cmd->data[3] = upper_32_bits(address);
940 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
941 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
944 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
945 int status, int tag, bool gn)
947 memset(cmd, 0, sizeof(*cmd));
949 cmd->data[0] = devid;
951 cmd->data[1] = pasid;
952 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
954 cmd->data[3] = tag & 0x1ff;
955 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
957 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
960 static void build_inv_all(struct iommu_cmd *cmd)
962 memset(cmd, 0, sizeof(*cmd));
963 CMD_SET_TYPE(cmd, CMD_INV_ALL);
966 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
968 memset(cmd, 0, sizeof(*cmd));
969 cmd->data[0] = devid;
970 CMD_SET_TYPE(cmd, CMD_INV_IRT);
974 * Writes the command to the IOMMUs command buffer and informs the
975 * hardware about the new command.
977 static int iommu_queue_command_sync(struct amd_iommu *iommu,
978 struct iommu_cmd *cmd,
981 u32 left, tail, head, next_tail;
985 spin_lock_irqsave(&iommu->lock, flags);
987 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
988 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
989 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
990 left = (head - next_tail) % CMD_BUFFER_SIZE;
993 struct iommu_cmd sync_cmd;
994 volatile u64 sem = 0;
997 build_completion_wait(&sync_cmd, (u64)&sem);
998 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1000 spin_unlock_irqrestore(&iommu->lock, flags);
1002 if ((ret = wait_on_sem(&sem)) != 0)
1008 copy_cmd_to_buffer(iommu, cmd, tail);
1010 /* We need to sync now to make sure all commands are processed */
1011 iommu->need_sync = sync;
1013 spin_unlock_irqrestore(&iommu->lock, flags);
1018 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1020 return iommu_queue_command_sync(iommu, cmd, true);
1024 * This function queues a completion wait command into the command
1025 * buffer of an IOMMU
1027 static int iommu_completion_wait(struct amd_iommu *iommu)
1029 struct iommu_cmd cmd;
1030 volatile u64 sem = 0;
1033 if (!iommu->need_sync)
1036 build_completion_wait(&cmd, (u64)&sem);
1038 ret = iommu_queue_command_sync(iommu, &cmd, false);
1042 return wait_on_sem(&sem);
1045 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1047 struct iommu_cmd cmd;
1049 build_inv_dte(&cmd, devid);
1051 return iommu_queue_command(iommu, &cmd);
1054 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1058 for (devid = 0; devid <= 0xffff; ++devid)
1059 iommu_flush_dte(iommu, devid);
1061 iommu_completion_wait(iommu);
1065 * This function uses heavy locking and may disable irqs for some time. But
1066 * this is no issue because it is only called during resume.
1068 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1072 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1073 struct iommu_cmd cmd;
1074 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1076 iommu_queue_command(iommu, &cmd);
1079 iommu_completion_wait(iommu);
1082 static void iommu_flush_all(struct amd_iommu *iommu)
1084 struct iommu_cmd cmd;
1086 build_inv_all(&cmd);
1088 iommu_queue_command(iommu, &cmd);
1089 iommu_completion_wait(iommu);
1092 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1094 struct iommu_cmd cmd;
1096 build_inv_irt(&cmd, devid);
1098 iommu_queue_command(iommu, &cmd);
1101 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1105 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1106 iommu_flush_irt(iommu, devid);
1108 iommu_completion_wait(iommu);
1111 void iommu_flush_all_caches(struct amd_iommu *iommu)
1113 if (iommu_feature(iommu, FEATURE_IA)) {
1114 iommu_flush_all(iommu);
1116 iommu_flush_dte_all(iommu);
1117 iommu_flush_irt_all(iommu);
1118 iommu_flush_tlb_all(iommu);
1123 * Command send function for flushing on-device TLB
1125 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1126 u64 address, size_t size)
1128 struct amd_iommu *iommu;
1129 struct iommu_cmd cmd;
1132 qdep = dev_data->ats.qdep;
1133 iommu = amd_iommu_rlookup_table[dev_data->devid];
1135 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1137 return iommu_queue_command(iommu, &cmd);
1141 * Command send function for invalidating a device table entry
1143 static int device_flush_dte(struct iommu_dev_data *dev_data)
1145 struct amd_iommu *iommu;
1149 iommu = amd_iommu_rlookup_table[dev_data->devid];
1150 alias = dev_data->alias;
1152 ret = iommu_flush_dte(iommu, dev_data->devid);
1153 if (!ret && alias != dev_data->devid)
1154 ret = iommu_flush_dte(iommu, alias);
1158 if (dev_data->ats.enabled)
1159 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1165 * TLB invalidation function which is called from the mapping functions.
1166 * It invalidates a single PTE if the range to flush is within a single
1167 * page. Otherwise it flushes the whole TLB of the IOMMU.
1169 static void __domain_flush_pages(struct protection_domain *domain,
1170 u64 address, size_t size, int pde)
1172 struct iommu_dev_data *dev_data;
1173 struct iommu_cmd cmd;
1176 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1178 for (i = 0; i < amd_iommus_present; ++i) {
1179 if (!domain->dev_iommu[i])
1183 * Devices of this domain are behind this IOMMU
1184 * We need a TLB flush
1186 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1189 list_for_each_entry(dev_data, &domain->dev_list, list) {
1191 if (!dev_data->ats.enabled)
1194 ret |= device_flush_iotlb(dev_data, address, size);
1200 static void domain_flush_pages(struct protection_domain *domain,
1201 u64 address, size_t size)
1203 __domain_flush_pages(domain, address, size, 0);
1206 /* Flush the whole IO/TLB for a given protection domain */
1207 static void domain_flush_tlb(struct protection_domain *domain)
1209 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1212 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1213 static void domain_flush_tlb_pde(struct protection_domain *domain)
1215 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1218 static void domain_flush_complete(struct protection_domain *domain)
1222 for (i = 0; i < amd_iommus_present; ++i) {
1223 if (!domain->dev_iommu[i])
1227 * Devices of this domain are behind this IOMMU
1228 * We need to wait for completion of all commands.
1230 iommu_completion_wait(amd_iommus[i]);
1236 * This function flushes the DTEs for all devices in domain
1238 static void domain_flush_devices(struct protection_domain *domain)
1240 struct iommu_dev_data *dev_data;
1242 list_for_each_entry(dev_data, &domain->dev_list, list)
1243 device_flush_dte(dev_data);
1246 /****************************************************************************
1248 * The functions below are used the create the page table mappings for
1249 * unity mapped regions.
1251 ****************************************************************************/
1254 * This function is used to add another level to an IO page table. Adding
1255 * another level increases the size of the address space by 9 bits to a size up
1258 static bool increase_address_space(struct protection_domain *domain,
1263 if (domain->mode == PAGE_MODE_6_LEVEL)
1264 /* address space already 64 bit large */
1267 pte = (void *)get_zeroed_page(gfp);
1271 *pte = PM_LEVEL_PDE(domain->mode,
1272 virt_to_phys(domain->pt_root));
1273 domain->pt_root = pte;
1275 domain->updated = true;
1280 static u64 *alloc_pte(struct protection_domain *domain,
1281 unsigned long address,
1282 unsigned long page_size,
1289 BUG_ON(!is_power_of_2(page_size));
1291 while (address > PM_LEVEL_SIZE(domain->mode))
1292 increase_address_space(domain, gfp);
1294 level = domain->mode - 1;
1295 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1296 address = PAGE_SIZE_ALIGN(address, page_size);
1297 end_lvl = PAGE_SIZE_LEVEL(page_size);
1299 while (level > end_lvl) {
1304 if (!IOMMU_PTE_PRESENT(__pte)) {
1305 page = (u64 *)get_zeroed_page(gfp);
1309 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1311 if (cmpxchg64(pte, __pte, __npte)) {
1312 free_page((unsigned long)page);
1317 /* No level skipping support yet */
1318 if (PM_PTE_LEVEL(*pte) != level)
1323 pte = IOMMU_PTE_PAGE(*pte);
1325 if (pte_page && level == end_lvl)
1328 pte = &pte[PM_LEVEL_INDEX(level, address)];
1335 * This function checks if there is a PTE for a given dma address. If
1336 * there is one, it returns the pointer to it.
1338 static u64 *fetch_pte(struct protection_domain *domain,
1339 unsigned long address,
1340 unsigned long *page_size)
1345 if (address > PM_LEVEL_SIZE(domain->mode))
1348 level = domain->mode - 1;
1349 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1350 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1355 if (!IOMMU_PTE_PRESENT(*pte))
1359 if (PM_PTE_LEVEL(*pte) == 7 ||
1360 PM_PTE_LEVEL(*pte) == 0)
1363 /* No level skipping support yet */
1364 if (PM_PTE_LEVEL(*pte) != level)
1369 /* Walk to the next level */
1370 pte = IOMMU_PTE_PAGE(*pte);
1371 pte = &pte[PM_LEVEL_INDEX(level, address)];
1372 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1375 if (PM_PTE_LEVEL(*pte) == 0x07) {
1376 unsigned long pte_mask;
1379 * If we have a series of large PTEs, make
1380 * sure to return a pointer to the first one.
1382 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1383 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1384 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1391 * Generic mapping functions. It maps a physical address into a DMA
1392 * address space. It allocates the page table pages if necessary.
1393 * In the future it can be extended to a generic mapping function
1394 * supporting all features of AMD IOMMU page tables like level skipping
1395 * and full 64 bit address spaces.
1397 static int iommu_map_page(struct protection_domain *dom,
1398 unsigned long bus_addr,
1399 unsigned long phys_addr,
1401 unsigned long page_size)
1406 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1407 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1409 if (!(prot & IOMMU_PROT_MASK))
1412 count = PAGE_SIZE_PTE_COUNT(page_size);
1413 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1418 for (i = 0; i < count; ++i)
1419 if (IOMMU_PTE_PRESENT(pte[i]))
1423 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1424 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1426 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1428 if (prot & IOMMU_PROT_IR)
1429 __pte |= IOMMU_PTE_IR;
1430 if (prot & IOMMU_PROT_IW)
1431 __pte |= IOMMU_PTE_IW;
1433 for (i = 0; i < count; ++i)
1441 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1442 unsigned long bus_addr,
1443 unsigned long page_size)
1445 unsigned long long unmapped;
1446 unsigned long unmap_size;
1449 BUG_ON(!is_power_of_2(page_size));
1453 while (unmapped < page_size) {
1455 pte = fetch_pte(dom, bus_addr, &unmap_size);
1460 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1461 for (i = 0; i < count; i++)
1465 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1466 unmapped += unmap_size;
1469 BUG_ON(unmapped && !is_power_of_2(unmapped));
1474 /****************************************************************************
1476 * The next functions belong to the address allocator for the dma_ops
1477 * interface functions. They work like the allocators in the other IOMMU
1478 * drivers. Its basically a bitmap which marks the allocated pages in
1479 * the aperture. Maybe it could be enhanced in the future to a more
1480 * efficient allocator.
1482 ****************************************************************************/
1485 * The address allocator core functions.
1487 * called with domain->lock held
1491 * Used to reserve address ranges in the aperture (e.g. for exclusion
1494 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1495 unsigned long start_page,
1498 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1500 if (start_page + pages > last_page)
1501 pages = last_page - start_page;
1503 for (i = start_page; i < start_page + pages; ++i) {
1504 int index = i / APERTURE_RANGE_PAGES;
1505 int page = i % APERTURE_RANGE_PAGES;
1506 __set_bit(page, dom->aperture[index]->bitmap);
1511 * This function is used to add a new aperture range to an existing
1512 * aperture in case of dma_ops domain allocation or address allocation
1515 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1516 bool populate, gfp_t gfp)
1518 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1519 unsigned long i, old_size, pte_pgsize;
1520 struct aperture_range *range;
1521 struct amd_iommu *iommu;
1522 unsigned long flags;
1524 #ifdef CONFIG_IOMMU_STRESS
1528 if (index >= APERTURE_MAX_RANGES)
1531 range = kzalloc(sizeof(struct aperture_range), gfp);
1535 range->bitmap = (void *)get_zeroed_page(gfp);
1539 range->offset = dma_dom->aperture_size;
1541 spin_lock_init(&range->bitmap_lock);
1544 unsigned long address = dma_dom->aperture_size;
1545 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1546 u64 *pte, *pte_page;
1548 for (i = 0; i < num_ptes; ++i) {
1549 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1554 range->pte_pages[i] = pte_page;
1556 address += APERTURE_RANGE_SIZE / 64;
1560 spin_lock_irqsave(&dma_dom->domain.lock, flags);
1562 /* First take the bitmap_lock and then publish the range */
1563 spin_lock(&range->bitmap_lock);
1565 old_size = dma_dom->aperture_size;
1566 dma_dom->aperture[index] = range;
1567 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1569 /* Reserve address range used for MSI messages */
1570 if (old_size < MSI_ADDR_BASE_LO &&
1571 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1572 unsigned long spage;
1575 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1576 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1578 dma_ops_reserve_addresses(dma_dom, spage, pages);
1581 /* Initialize the exclusion range if necessary */
1582 for_each_iommu(iommu) {
1583 if (iommu->exclusion_start &&
1584 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1585 && iommu->exclusion_start < dma_dom->aperture_size) {
1586 unsigned long startpage;
1587 int pages = iommu_num_pages(iommu->exclusion_start,
1588 iommu->exclusion_length,
1590 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1591 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1596 * Check for areas already mapped as present in the new aperture
1597 * range and mark those pages as reserved in the allocator. Such
1598 * mappings may already exist as a result of requested unity
1599 * mappings for devices.
1601 for (i = dma_dom->aperture[index]->offset;
1602 i < dma_dom->aperture_size;
1604 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1605 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1608 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1612 update_domain(&dma_dom->domain);
1614 spin_unlock(&range->bitmap_lock);
1616 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
1621 update_domain(&dma_dom->domain);
1623 free_page((unsigned long)range->bitmap);
1630 static dma_addr_t dma_ops_aperture_alloc(struct dma_ops_domain *dom,
1631 struct aperture_range *range,
1632 unsigned long pages,
1633 unsigned long dma_mask,
1634 unsigned long boundary_size,
1635 unsigned long align_mask,
1638 unsigned long offset, limit, flags;
1642 offset = range->offset >> PAGE_SHIFT;
1643 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1644 dma_mask >> PAGE_SHIFT);
1647 if (!spin_trylock_irqsave(&range->bitmap_lock, flags))
1650 spin_lock_irqsave(&range->bitmap_lock, flags);
1653 address = iommu_area_alloc(range->bitmap, limit, range->next_bit,
1654 pages, offset, boundary_size, align_mask);
1655 if (address == -1) {
1656 /* Nothing found, retry one time */
1657 address = iommu_area_alloc(range->bitmap, limit,
1658 0, pages, offset, boundary_size,
1664 range->next_bit = address + pages;
1666 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1669 domain_flush_tlb(&dom->domain);
1670 domain_flush_complete(&dom->domain);
1676 static unsigned long dma_ops_area_alloc(struct device *dev,
1677 struct dma_ops_domain *dom,
1679 unsigned long align_mask,
1682 unsigned long boundary_size, mask;
1683 unsigned long address = -1;
1689 mask = dma_get_seg_boundary(dev);
1692 start = this_cpu_read(*dom->next_index);
1694 /* Sanity check - is it really necessary? */
1695 if (unlikely(start > APERTURE_MAX_RANGES)) {
1697 this_cpu_write(*dom->next_index, 0);
1700 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1701 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1703 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1704 struct aperture_range *range;
1707 index = (start + i) % APERTURE_MAX_RANGES;
1709 range = dom->aperture[index];
1711 if (!range || range->offset >= dma_mask)
1714 address = dma_ops_aperture_alloc(dom, range, pages,
1715 dma_mask, boundary_size,
1717 if (address != -1) {
1718 address = range->offset + (address << PAGE_SHIFT);
1719 this_cpu_write(*dom->next_index, index);
1724 if (address == -1 && first) {
1734 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1735 struct dma_ops_domain *dom,
1737 unsigned long align_mask,
1740 unsigned long address = -1;
1742 while (address == -1) {
1743 address = dma_ops_area_alloc(dev, dom, pages,
1744 align_mask, dma_mask);
1746 if (address == -1 && alloc_new_range(dom, false, GFP_ATOMIC))
1750 if (unlikely(address == -1))
1751 address = DMA_ERROR_CODE;
1753 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1759 * The address free function.
1761 * called with domain->lock held
1763 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1764 unsigned long address,
1767 unsigned i = address >> APERTURE_RANGE_SHIFT;
1768 struct aperture_range *range = dom->aperture[i];
1769 unsigned long flags;
1771 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1773 #ifdef CONFIG_IOMMU_STRESS
1778 if (amd_iommu_unmap_flush) {
1779 domain_flush_tlb(&dom->domain);
1780 domain_flush_complete(&dom->domain);
1783 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1785 spin_lock_irqsave(&range->bitmap_lock, flags);
1786 if (address + pages > range->next_bit)
1787 range->next_bit = address + pages;
1788 bitmap_clear(range->bitmap, address, pages);
1789 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1793 /****************************************************************************
1795 * The next functions belong to the domain allocation. A domain is
1796 * allocated for every IOMMU as the default domain. If device isolation
1797 * is enabled, every device get its own domain. The most important thing
1798 * about domains is the page table mapping the DMA address space they
1801 ****************************************************************************/
1804 * This function adds a protection domain to the global protection domain list
1806 static void add_domain_to_list(struct protection_domain *domain)
1808 unsigned long flags;
1810 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1811 list_add(&domain->list, &amd_iommu_pd_list);
1812 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1816 * This function removes a protection domain to the global
1817 * protection domain list
1819 static void del_domain_from_list(struct protection_domain *domain)
1821 unsigned long flags;
1823 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1824 list_del(&domain->list);
1825 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1828 static u16 domain_id_alloc(void)
1830 unsigned long flags;
1833 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1834 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1836 if (id > 0 && id < MAX_DOMAIN_ID)
1837 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1840 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1845 static void domain_id_free(int id)
1847 unsigned long flags;
1849 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1850 if (id > 0 && id < MAX_DOMAIN_ID)
1851 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1852 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1855 #define DEFINE_FREE_PT_FN(LVL, FN) \
1856 static void free_pt_##LVL (unsigned long __pt) \
1864 for (i = 0; i < 512; ++i) { \
1865 /* PTE present? */ \
1866 if (!IOMMU_PTE_PRESENT(pt[i])) \
1870 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1871 PM_PTE_LEVEL(pt[i]) == 7) \
1874 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1877 free_page((unsigned long)pt); \
1880 DEFINE_FREE_PT_FN(l2, free_page)
1881 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1882 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1883 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1884 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1886 static void free_pagetable(struct protection_domain *domain)
1888 unsigned long root = (unsigned long)domain->pt_root;
1890 switch (domain->mode) {
1891 case PAGE_MODE_NONE:
1893 case PAGE_MODE_1_LEVEL:
1896 case PAGE_MODE_2_LEVEL:
1899 case PAGE_MODE_3_LEVEL:
1902 case PAGE_MODE_4_LEVEL:
1905 case PAGE_MODE_5_LEVEL:
1908 case PAGE_MODE_6_LEVEL:
1916 static void free_gcr3_tbl_level1(u64 *tbl)
1921 for (i = 0; i < 512; ++i) {
1922 if (!(tbl[i] & GCR3_VALID))
1925 ptr = __va(tbl[i] & PAGE_MASK);
1927 free_page((unsigned long)ptr);
1931 static void free_gcr3_tbl_level2(u64 *tbl)
1936 for (i = 0; i < 512; ++i) {
1937 if (!(tbl[i] & GCR3_VALID))
1940 ptr = __va(tbl[i] & PAGE_MASK);
1942 free_gcr3_tbl_level1(ptr);
1946 static void free_gcr3_table(struct protection_domain *domain)
1948 if (domain->glx == 2)
1949 free_gcr3_tbl_level2(domain->gcr3_tbl);
1950 else if (domain->glx == 1)
1951 free_gcr3_tbl_level1(domain->gcr3_tbl);
1953 BUG_ON(domain->glx != 0);
1955 free_page((unsigned long)domain->gcr3_tbl);
1959 * Free a domain, only used if something went wrong in the
1960 * allocation path and we need to free an already allocated page table
1962 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1969 free_percpu(dom->next_index);
1971 del_domain_from_list(&dom->domain);
1973 free_pagetable(&dom->domain);
1975 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1976 if (!dom->aperture[i])
1978 free_page((unsigned long)dom->aperture[i]->bitmap);
1979 kfree(dom->aperture[i]);
1985 static int dma_ops_domain_alloc_apertures(struct dma_ops_domain *dma_dom,
1988 int ret, i, apertures;
1990 apertures = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1993 for (i = apertures; i < max_apertures; ++i) {
1994 ret = alloc_new_range(dma_dom, false, GFP_KERNEL);
2003 * Allocates a new protection domain usable for the dma_ops functions.
2004 * It also initializes the page table and the address allocator data
2005 * structures required for the dma_ops interface
2007 static struct dma_ops_domain *dma_ops_domain_alloc(void)
2009 struct dma_ops_domain *dma_dom;
2012 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
2016 if (protection_domain_init(&dma_dom->domain))
2019 dma_dom->next_index = alloc_percpu(u32);
2020 if (!dma_dom->next_index)
2023 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2024 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2025 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2026 dma_dom->domain.priv = dma_dom;
2027 if (!dma_dom->domain.pt_root)
2030 add_domain_to_list(&dma_dom->domain);
2032 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2036 * mark the first page as allocated so we never return 0 as
2037 * a valid dma-address. So we can use 0 as error value
2039 dma_dom->aperture[0]->bitmap[0] = 1;
2041 for_each_possible_cpu(cpu)
2042 *per_cpu_ptr(dma_dom->next_index, cpu) = 0;
2047 dma_ops_domain_free(dma_dom);
2053 * little helper function to check whether a given protection domain is a
2056 static bool dma_ops_domain(struct protection_domain *domain)
2058 return domain->flags & PD_DMA_OPS_MASK;
2061 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2066 if (domain->mode != PAGE_MODE_NONE)
2067 pte_root = virt_to_phys(domain->pt_root);
2069 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2070 << DEV_ENTRY_MODE_SHIFT;
2071 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2073 flags = amd_iommu_dev_table[devid].data[1];
2076 flags |= DTE_FLAG_IOTLB;
2078 if (domain->flags & PD_IOMMUV2_MASK) {
2079 u64 gcr3 = __pa(domain->gcr3_tbl);
2080 u64 glx = domain->glx;
2083 pte_root |= DTE_FLAG_GV;
2084 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2086 /* First mask out possible old values for GCR3 table */
2087 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2090 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2093 /* Encode GCR3 table into DTE */
2094 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2097 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2100 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2104 flags &= ~(0xffffUL);
2105 flags |= domain->id;
2107 amd_iommu_dev_table[devid].data[1] = flags;
2108 amd_iommu_dev_table[devid].data[0] = pte_root;
2111 static void clear_dte_entry(u16 devid)
2113 /* remove entry from the device table seen by the hardware */
2114 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2115 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2117 amd_iommu_apply_erratum_63(devid);
2120 static void do_attach(struct iommu_dev_data *dev_data,
2121 struct protection_domain *domain)
2123 struct amd_iommu *iommu;
2127 iommu = amd_iommu_rlookup_table[dev_data->devid];
2128 alias = dev_data->alias;
2129 ats = dev_data->ats.enabled;
2131 /* Update data structures */
2132 dev_data->domain = domain;
2133 list_add(&dev_data->list, &domain->dev_list);
2135 /* Do reference counting */
2136 domain->dev_iommu[iommu->index] += 1;
2137 domain->dev_cnt += 1;
2139 /* Update device table */
2140 set_dte_entry(dev_data->devid, domain, ats);
2141 if (alias != dev_data->devid)
2142 set_dte_entry(alias, domain, ats);
2144 device_flush_dte(dev_data);
2147 static void do_detach(struct iommu_dev_data *dev_data)
2149 struct amd_iommu *iommu;
2153 * First check if the device is still attached. It might already
2154 * be detached from its domain because the generic
2155 * iommu_detach_group code detached it and we try again here in
2156 * our alias handling.
2158 if (!dev_data->domain)
2161 iommu = amd_iommu_rlookup_table[dev_data->devid];
2162 alias = dev_data->alias;
2164 /* decrease reference counters */
2165 dev_data->domain->dev_iommu[iommu->index] -= 1;
2166 dev_data->domain->dev_cnt -= 1;
2168 /* Update data structures */
2169 dev_data->domain = NULL;
2170 list_del(&dev_data->list);
2171 clear_dte_entry(dev_data->devid);
2172 if (alias != dev_data->devid)
2173 clear_dte_entry(alias);
2175 /* Flush the DTE entry */
2176 device_flush_dte(dev_data);
2180 * If a device is not yet associated with a domain, this function does
2181 * assigns it visible for the hardware
2183 static int __attach_device(struct iommu_dev_data *dev_data,
2184 struct protection_domain *domain)
2189 * Must be called with IRQs disabled. Warn here to detect early
2192 WARN_ON(!irqs_disabled());
2195 spin_lock(&domain->lock);
2198 if (dev_data->domain != NULL)
2201 /* Attach alias group root */
2202 do_attach(dev_data, domain);
2209 spin_unlock(&domain->lock);
2215 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2217 pci_disable_ats(pdev);
2218 pci_disable_pri(pdev);
2219 pci_disable_pasid(pdev);
2222 /* FIXME: Change generic reset-function to do the same */
2223 static int pri_reset_while_enabled(struct pci_dev *pdev)
2228 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2232 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2233 control |= PCI_PRI_CTRL_RESET;
2234 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2239 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2244 /* FIXME: Hardcode number of outstanding requests for now */
2246 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2248 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2250 /* Only allow access to user-accessible pages */
2251 ret = pci_enable_pasid(pdev, 0);
2255 /* First reset the PRI state of the device */
2256 ret = pci_reset_pri(pdev);
2261 ret = pci_enable_pri(pdev, reqs);
2266 ret = pri_reset_while_enabled(pdev);
2271 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2278 pci_disable_pri(pdev);
2279 pci_disable_pasid(pdev);
2284 /* FIXME: Move this to PCI code */
2285 #define PCI_PRI_TLP_OFF (1 << 15)
2287 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2292 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2296 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2298 return (status & PCI_PRI_TLP_OFF) ? true : false;
2302 * If a device is not yet associated with a domain, this function
2303 * assigns it visible for the hardware
2305 static int attach_device(struct device *dev,
2306 struct protection_domain *domain)
2308 struct pci_dev *pdev;
2309 struct iommu_dev_data *dev_data;
2310 unsigned long flags;
2313 dev_data = get_dev_data(dev);
2315 if (!dev_is_pci(dev))
2316 goto skip_ats_check;
2318 pdev = to_pci_dev(dev);
2319 if (domain->flags & PD_IOMMUV2_MASK) {
2320 if (!dev_data->passthrough)
2323 if (dev_data->iommu_v2) {
2324 if (pdev_iommuv2_enable(pdev) != 0)
2327 dev_data->ats.enabled = true;
2328 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2329 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);
2338 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2339 ret = __attach_device(dev_data, domain);
2340 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2343 * We might boot into a crash-kernel here. The crashed kernel
2344 * left the caches in the IOMMU dirty. So we have to flush
2345 * here to evict all dirty stuff.
2347 domain_flush_tlb_pde(domain);
2353 * Removes a device from a protection domain (unlocked)
2355 static void __detach_device(struct iommu_dev_data *dev_data)
2357 struct protection_domain *domain;
2360 * Must be called with IRQs disabled. Warn here to detect early
2363 WARN_ON(!irqs_disabled());
2365 if (WARN_ON(!dev_data->domain))
2368 domain = dev_data->domain;
2370 spin_lock(&domain->lock);
2372 do_detach(dev_data);
2374 spin_unlock(&domain->lock);
2378 * Removes a device from a protection domain (with devtable_lock held)
2380 static void detach_device(struct device *dev)
2382 struct protection_domain *domain;
2383 struct iommu_dev_data *dev_data;
2384 unsigned long flags;
2386 dev_data = get_dev_data(dev);
2387 domain = dev_data->domain;
2389 /* lock device table */
2390 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2391 __detach_device(dev_data);
2392 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2394 if (!dev_is_pci(dev))
2397 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2398 pdev_iommuv2_disable(to_pci_dev(dev));
2399 else if (dev_data->ats.enabled)
2400 pci_disable_ats(to_pci_dev(dev));
2402 dev_data->ats.enabled = false;
2405 static int amd_iommu_add_device(struct device *dev)
2407 struct iommu_dev_data *dev_data;
2408 struct iommu_domain *domain;
2409 struct amd_iommu *iommu;
2412 if (!check_device(dev) || get_dev_data(dev))
2415 devid = get_device_id(dev);
2419 iommu = amd_iommu_rlookup_table[devid];
2421 ret = iommu_init_device(dev);
2423 if (ret != -ENOTSUPP)
2424 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2427 iommu_ignore_device(dev);
2428 dev->archdata.dma_ops = &nommu_dma_ops;
2431 init_iommu_group(dev);
2433 dev_data = get_dev_data(dev);
2437 if (iommu_pass_through || dev_data->iommu_v2)
2438 iommu_request_dm_for_dev(dev);
2440 /* Domains are initialized for this device - have a look what we ended up with */
2441 domain = iommu_get_domain_for_dev(dev);
2442 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2443 dev_data->passthrough = true;
2445 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2448 iommu_completion_wait(iommu);
2453 static void amd_iommu_remove_device(struct device *dev)
2455 struct amd_iommu *iommu;
2458 if (!check_device(dev))
2461 devid = get_device_id(dev);
2465 iommu = amd_iommu_rlookup_table[devid];
2467 iommu_uninit_device(dev);
2468 iommu_completion_wait(iommu);
2471 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2473 if (dev_is_pci(dev))
2474 return pci_device_group(dev);
2476 return acpihid_device_group(dev);
2479 /*****************************************************************************
2481 * The next functions belong to the dma_ops mapping/unmapping code.
2483 *****************************************************************************/
2486 * In the dma_ops path we only have the struct device. This function
2487 * finds the corresponding IOMMU, the protection domain and the
2488 * requestor id for a given device.
2489 * If the device is not yet associated with a domain this is also done
2492 static struct protection_domain *get_domain(struct device *dev)
2494 struct protection_domain *domain;
2495 struct iommu_domain *io_domain;
2497 if (!check_device(dev))
2498 return ERR_PTR(-EINVAL);
2500 io_domain = iommu_get_domain_for_dev(dev);
2504 domain = to_pdomain(io_domain);
2505 if (!dma_ops_domain(domain))
2506 return ERR_PTR(-EBUSY);
2511 static void update_device_table(struct protection_domain *domain)
2513 struct iommu_dev_data *dev_data;
2515 list_for_each_entry(dev_data, &domain->dev_list, list)
2516 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2519 static void update_domain(struct protection_domain *domain)
2521 if (!domain->updated)
2524 update_device_table(domain);
2526 domain_flush_devices(domain);
2527 domain_flush_tlb_pde(domain);
2529 domain->updated = false;
2533 * This function fetches the PTE for a given address in the aperture
2535 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2536 unsigned long address)
2538 struct aperture_range *aperture;
2539 u64 *pte, *pte_page;
2541 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2545 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2547 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2549 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2551 pte += PM_LEVEL_INDEX(0, address);
2553 update_domain(&dom->domain);
2559 * This is the generic map function. It maps one 4kb page at paddr to
2560 * the given address in the DMA address space for the domain.
2562 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2563 unsigned long address,
2569 WARN_ON(address > dom->aperture_size);
2573 pte = dma_ops_get_pte(dom, address);
2575 return DMA_ERROR_CODE;
2577 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2579 if (direction == DMA_TO_DEVICE)
2580 __pte |= IOMMU_PTE_IR;
2581 else if (direction == DMA_FROM_DEVICE)
2582 __pte |= IOMMU_PTE_IW;
2583 else if (direction == DMA_BIDIRECTIONAL)
2584 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2590 return (dma_addr_t)address;
2594 * The generic unmapping function for on page in the DMA address space.
2596 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2597 unsigned long address)
2599 struct aperture_range *aperture;
2602 if (address >= dom->aperture_size)
2605 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2609 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2613 pte += PM_LEVEL_INDEX(0, address);
2615 WARN_ON_ONCE(!*pte);
2621 * This function contains common code for mapping of a physically
2622 * contiguous memory region into DMA address space. It is used by all
2623 * mapping functions provided with this IOMMU driver.
2624 * Must be called with the domain lock held.
2626 static dma_addr_t __map_single(struct device *dev,
2627 struct dma_ops_domain *dma_dom,
2634 dma_addr_t offset = paddr & ~PAGE_MASK;
2635 dma_addr_t address, start, ret;
2637 unsigned long align_mask = 0;
2640 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2644 align_mask = (1UL << get_order(size)) - 1;
2646 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2649 if (address == DMA_ERROR_CODE)
2653 for (i = 0; i < pages; ++i) {
2654 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2655 if (ret == DMA_ERROR_CODE)
2663 if (unlikely(amd_iommu_np_cache)) {
2664 domain_flush_pages(&dma_dom->domain, address, size);
2665 domain_flush_complete(&dma_dom->domain);
2673 for (--i; i >= 0; --i) {
2675 dma_ops_domain_unmap(dma_dom, start);
2678 dma_ops_free_addresses(dma_dom, address, pages);
2680 return DMA_ERROR_CODE;
2684 * Does the reverse of the __map_single function. Must be called with
2685 * the domain lock held too
2687 static void __unmap_single(struct dma_ops_domain *dma_dom,
2688 dma_addr_t dma_addr,
2692 dma_addr_t flush_addr;
2693 dma_addr_t i, start;
2696 if ((dma_addr == DMA_ERROR_CODE) ||
2697 (dma_addr + size > dma_dom->aperture_size))
2700 flush_addr = dma_addr;
2701 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2702 dma_addr &= PAGE_MASK;
2705 for (i = 0; i < pages; ++i) {
2706 dma_ops_domain_unmap(dma_dom, start);
2710 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2714 * The exported map_single function for dma_ops.
2716 static dma_addr_t map_page(struct device *dev, struct page *page,
2717 unsigned long offset, size_t size,
2718 enum dma_data_direction dir,
2719 struct dma_attrs *attrs)
2721 phys_addr_t paddr = page_to_phys(page) + offset;
2722 struct protection_domain *domain;
2725 domain = get_domain(dev);
2726 if (PTR_ERR(domain) == -EINVAL)
2727 return (dma_addr_t)paddr;
2728 else if (IS_ERR(domain))
2729 return DMA_ERROR_CODE;
2731 dma_mask = *dev->dma_mask;
2733 return __map_single(dev, domain->priv, paddr, size, dir, false,
2738 * The exported unmap_single function for dma_ops.
2740 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2741 enum dma_data_direction dir, struct dma_attrs *attrs)
2743 struct protection_domain *domain;
2745 domain = get_domain(dev);
2749 __unmap_single(domain->priv, dma_addr, size, dir);
2753 * The exported map_sg function for dma_ops (handles scatter-gather
2756 static int map_sg(struct device *dev, struct scatterlist *sglist,
2757 int nelems, enum dma_data_direction dir,
2758 struct dma_attrs *attrs)
2760 struct protection_domain *domain;
2762 struct scatterlist *s;
2764 int mapped_elems = 0;
2767 domain = get_domain(dev);
2771 dma_mask = *dev->dma_mask;
2773 for_each_sg(sglist, s, nelems, i) {
2776 s->dma_address = __map_single(dev, domain->priv,
2777 paddr, s->length, dir, false,
2780 if (s->dma_address) {
2781 s->dma_length = s->length;
2787 return mapped_elems;
2790 for_each_sg(sglist, s, mapped_elems, i) {
2792 __unmap_single(domain->priv, s->dma_address,
2793 s->dma_length, dir);
2794 s->dma_address = s->dma_length = 0;
2801 * The exported map_sg function for dma_ops (handles scatter-gather
2804 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2805 int nelems, enum dma_data_direction dir,
2806 struct dma_attrs *attrs)
2808 struct protection_domain *domain;
2809 struct scatterlist *s;
2812 domain = get_domain(dev);
2816 for_each_sg(sglist, s, nelems, i) {
2817 __unmap_single(domain->priv, s->dma_address,
2818 s->dma_length, dir);
2819 s->dma_address = s->dma_length = 0;
2824 * The exported alloc_coherent function for dma_ops.
2826 static void *alloc_coherent(struct device *dev, size_t size,
2827 dma_addr_t *dma_addr, gfp_t flag,
2828 struct dma_attrs *attrs)
2830 u64 dma_mask = dev->coherent_dma_mask;
2831 struct protection_domain *domain;
2834 domain = get_domain(dev);
2835 if (PTR_ERR(domain) == -EINVAL) {
2836 page = alloc_pages(flag, get_order(size));
2837 *dma_addr = page_to_phys(page);
2838 return page_address(page);
2839 } else if (IS_ERR(domain))
2842 size = PAGE_ALIGN(size);
2843 dma_mask = dev->coherent_dma_mask;
2844 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2847 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2849 if (!gfpflags_allow_blocking(flag))
2852 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2859 dma_mask = *dev->dma_mask;
2861 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2862 size, DMA_BIDIRECTIONAL, true, dma_mask);
2864 if (*dma_addr == DMA_ERROR_CODE)
2867 return page_address(page);
2871 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2872 __free_pages(page, get_order(size));
2878 * The exported free_coherent function for dma_ops.
2880 static void free_coherent(struct device *dev, size_t size,
2881 void *virt_addr, dma_addr_t dma_addr,
2882 struct dma_attrs *attrs)
2884 struct protection_domain *domain;
2887 page = virt_to_page(virt_addr);
2888 size = PAGE_ALIGN(size);
2890 domain = get_domain(dev);
2894 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2897 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2898 __free_pages(page, get_order(size));
2902 * This function is called by the DMA layer to find out if we can handle a
2903 * particular device. It is part of the dma_ops.
2905 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2907 return check_device(dev);
2910 static int set_dma_mask(struct device *dev, u64 mask)
2912 struct protection_domain *domain;
2913 int max_apertures = 1;
2915 domain = get_domain(dev);
2917 return PTR_ERR(domain);
2919 if (mask == DMA_BIT_MASK(64))
2921 else if (mask > DMA_BIT_MASK(32))
2925 * To prevent lock contention it doesn't make sense to allocate more
2926 * apertures than online cpus
2928 if (max_apertures > num_online_cpus())
2929 max_apertures = num_online_cpus();
2931 if (dma_ops_domain_alloc_apertures(domain->priv, max_apertures))
2932 dev_err(dev, "Can't allocate %d iommu apertures\n",
2938 static struct dma_map_ops amd_iommu_dma_ops = {
2939 .alloc = alloc_coherent,
2940 .free = free_coherent,
2941 .map_page = map_page,
2942 .unmap_page = unmap_page,
2944 .unmap_sg = unmap_sg,
2945 .dma_supported = amd_iommu_dma_supported,
2946 .set_dma_mask = set_dma_mask,
2949 int __init amd_iommu_init_api(void)
2953 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2956 #ifdef CONFIG_ARM_AMBA
2957 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2964 int __init amd_iommu_init_dma_ops(void)
2966 swiotlb = iommu_pass_through ? 1 : 0;
2970 * In case we don't initialize SWIOTLB (actually the common case
2971 * when AMD IOMMU is enabled), make sure there are global
2972 * dma_ops set as a fall-back for devices not handled by this
2973 * driver (for example non-PCI devices).
2976 dma_ops = &nommu_dma_ops;
2978 if (amd_iommu_unmap_flush)
2979 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2981 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2986 /*****************************************************************************
2988 * The following functions belong to the exported interface of AMD IOMMU
2990 * This interface allows access to lower level functions of the IOMMU
2991 * like protection domain handling and assignement of devices to domains
2992 * which is not possible with the dma_ops interface.
2994 *****************************************************************************/
2996 static void cleanup_domain(struct protection_domain *domain)
2998 struct iommu_dev_data *entry;
2999 unsigned long flags;
3001 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3003 while (!list_empty(&domain->dev_list)) {
3004 entry = list_first_entry(&domain->dev_list,
3005 struct iommu_dev_data, list);
3006 __detach_device(entry);
3009 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3012 static void protection_domain_free(struct protection_domain *domain)
3017 del_domain_from_list(domain);
3020 domain_id_free(domain->id);
3025 static int protection_domain_init(struct protection_domain *domain)
3027 spin_lock_init(&domain->lock);
3028 mutex_init(&domain->api_lock);
3029 domain->id = domain_id_alloc();
3032 INIT_LIST_HEAD(&domain->dev_list);
3037 static struct protection_domain *protection_domain_alloc(void)
3039 struct protection_domain *domain;
3041 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3045 if (protection_domain_init(domain))
3048 add_domain_to_list(domain);
3058 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
3060 struct protection_domain *pdomain;
3061 struct dma_ops_domain *dma_domain;
3064 case IOMMU_DOMAIN_UNMANAGED:
3065 pdomain = protection_domain_alloc();
3069 pdomain->mode = PAGE_MODE_3_LEVEL;
3070 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3071 if (!pdomain->pt_root) {
3072 protection_domain_free(pdomain);
3076 pdomain->domain.geometry.aperture_start = 0;
3077 pdomain->domain.geometry.aperture_end = ~0ULL;
3078 pdomain->domain.geometry.force_aperture = true;
3081 case IOMMU_DOMAIN_DMA:
3082 dma_domain = dma_ops_domain_alloc();
3084 pr_err("AMD-Vi: Failed to allocate\n");
3087 pdomain = &dma_domain->domain;
3089 case IOMMU_DOMAIN_IDENTITY:
3090 pdomain = protection_domain_alloc();
3094 pdomain->mode = PAGE_MODE_NONE;
3100 return &pdomain->domain;
3103 static void amd_iommu_domain_free(struct iommu_domain *dom)
3105 struct protection_domain *domain;
3110 domain = to_pdomain(dom);
3112 if (domain->dev_cnt > 0)
3113 cleanup_domain(domain);
3115 BUG_ON(domain->dev_cnt != 0);
3117 if (domain->mode != PAGE_MODE_NONE)
3118 free_pagetable(domain);
3120 if (domain->flags & PD_IOMMUV2_MASK)
3121 free_gcr3_table(domain);
3123 protection_domain_free(domain);
3126 static void amd_iommu_detach_device(struct iommu_domain *dom,
3129 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3130 struct amd_iommu *iommu;
3133 if (!check_device(dev))
3136 devid = get_device_id(dev);
3140 if (dev_data->domain != NULL)
3143 iommu = amd_iommu_rlookup_table[devid];
3147 iommu_completion_wait(iommu);
3150 static int amd_iommu_attach_device(struct iommu_domain *dom,
3153 struct protection_domain *domain = to_pdomain(dom);
3154 struct iommu_dev_data *dev_data;
3155 struct amd_iommu *iommu;
3158 if (!check_device(dev))
3161 dev_data = dev->archdata.iommu;
3163 iommu = amd_iommu_rlookup_table[dev_data->devid];
3167 if (dev_data->domain)
3170 ret = attach_device(dev, domain);
3172 iommu_completion_wait(iommu);
3177 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3178 phys_addr_t paddr, size_t page_size, int iommu_prot)
3180 struct protection_domain *domain = to_pdomain(dom);
3184 if (domain->mode == PAGE_MODE_NONE)
3187 if (iommu_prot & IOMMU_READ)
3188 prot |= IOMMU_PROT_IR;
3189 if (iommu_prot & IOMMU_WRITE)
3190 prot |= IOMMU_PROT_IW;
3192 mutex_lock(&domain->api_lock);
3193 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3194 mutex_unlock(&domain->api_lock);
3199 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3202 struct protection_domain *domain = to_pdomain(dom);
3205 if (domain->mode == PAGE_MODE_NONE)
3208 mutex_lock(&domain->api_lock);
3209 unmap_size = iommu_unmap_page(domain, iova, page_size);
3210 mutex_unlock(&domain->api_lock);
3212 domain_flush_tlb_pde(domain);
3217 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3220 struct protection_domain *domain = to_pdomain(dom);
3221 unsigned long offset_mask, pte_pgsize;
3224 if (domain->mode == PAGE_MODE_NONE)
3227 pte = fetch_pte(domain, iova, &pte_pgsize);
3229 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3232 offset_mask = pte_pgsize - 1;
3233 __pte = *pte & PM_ADDR_MASK;
3235 return (__pte & ~offset_mask) | (iova & offset_mask);
3238 static bool amd_iommu_capable(enum iommu_cap cap)
3241 case IOMMU_CAP_CACHE_COHERENCY:
3243 case IOMMU_CAP_INTR_REMAP:
3244 return (irq_remapping_enabled == 1);
3245 case IOMMU_CAP_NOEXEC:
3252 static void amd_iommu_get_dm_regions(struct device *dev,
3253 struct list_head *head)
3255 struct unity_map_entry *entry;
3258 devid = get_device_id(dev);
3262 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3263 struct iommu_dm_region *region;
3265 if (devid < entry->devid_start || devid > entry->devid_end)
3268 region = kzalloc(sizeof(*region), GFP_KERNEL);
3270 pr_err("Out of memory allocating dm-regions for %s\n",
3275 region->start = entry->address_start;
3276 region->length = entry->address_end - entry->address_start;
3277 if (entry->prot & IOMMU_PROT_IR)
3278 region->prot |= IOMMU_READ;
3279 if (entry->prot & IOMMU_PROT_IW)
3280 region->prot |= IOMMU_WRITE;
3282 list_add_tail(®ion->list, head);
3286 static void amd_iommu_put_dm_regions(struct device *dev,
3287 struct list_head *head)
3289 struct iommu_dm_region *entry, *next;
3291 list_for_each_entry_safe(entry, next, head, list)
3295 static const struct iommu_ops amd_iommu_ops = {
3296 .capable = amd_iommu_capable,
3297 .domain_alloc = amd_iommu_domain_alloc,
3298 .domain_free = amd_iommu_domain_free,
3299 .attach_dev = amd_iommu_attach_device,
3300 .detach_dev = amd_iommu_detach_device,
3301 .map = amd_iommu_map,
3302 .unmap = amd_iommu_unmap,
3303 .map_sg = default_iommu_map_sg,
3304 .iova_to_phys = amd_iommu_iova_to_phys,
3305 .add_device = amd_iommu_add_device,
3306 .remove_device = amd_iommu_remove_device,
3307 .device_group = amd_iommu_device_group,
3308 .get_dm_regions = amd_iommu_get_dm_regions,
3309 .put_dm_regions = amd_iommu_put_dm_regions,
3310 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3313 /*****************************************************************************
3315 * The next functions do a basic initialization of IOMMU for pass through
3318 * In passthrough mode the IOMMU is initialized and enabled but not used for
3319 * DMA-API translation.
3321 *****************************************************************************/
3323 /* IOMMUv2 specific functions */
3324 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3326 return atomic_notifier_chain_register(&ppr_notifier, nb);
3328 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3330 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3332 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3334 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3336 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3338 struct protection_domain *domain = to_pdomain(dom);
3339 unsigned long flags;
3341 spin_lock_irqsave(&domain->lock, flags);
3343 /* Update data structure */
3344 domain->mode = PAGE_MODE_NONE;
3345 domain->updated = true;
3347 /* Make changes visible to IOMMUs */
3348 update_domain(domain);
3350 /* Page-table is not visible to IOMMU anymore, so free it */
3351 free_pagetable(domain);
3353 spin_unlock_irqrestore(&domain->lock, flags);
3355 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3357 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3359 struct protection_domain *domain = to_pdomain(dom);
3360 unsigned long flags;
3363 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3366 /* Number of GCR3 table levels required */
3367 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3370 if (levels > amd_iommu_max_glx_val)
3373 spin_lock_irqsave(&domain->lock, flags);
3376 * Save us all sanity checks whether devices already in the
3377 * domain support IOMMUv2. Just force that the domain has no
3378 * devices attached when it is switched into IOMMUv2 mode.
3381 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3385 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3386 if (domain->gcr3_tbl == NULL)
3389 domain->glx = levels;
3390 domain->flags |= PD_IOMMUV2_MASK;
3391 domain->updated = true;
3393 update_domain(domain);
3398 spin_unlock_irqrestore(&domain->lock, flags);
3402 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3404 static int __flush_pasid(struct protection_domain *domain, int pasid,
3405 u64 address, bool size)
3407 struct iommu_dev_data *dev_data;
3408 struct iommu_cmd cmd;
3411 if (!(domain->flags & PD_IOMMUV2_MASK))
3414 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3417 * IOMMU TLB needs to be flushed before Device TLB to
3418 * prevent device TLB refill from IOMMU TLB
3420 for (i = 0; i < amd_iommus_present; ++i) {
3421 if (domain->dev_iommu[i] == 0)
3424 ret = iommu_queue_command(amd_iommus[i], &cmd);
3429 /* Wait until IOMMU TLB flushes are complete */
3430 domain_flush_complete(domain);
3432 /* Now flush device TLBs */
3433 list_for_each_entry(dev_data, &domain->dev_list, list) {
3434 struct amd_iommu *iommu;
3438 There might be non-IOMMUv2 capable devices in an IOMMUv2
3441 if (!dev_data->ats.enabled)
3444 qdep = dev_data->ats.qdep;
3445 iommu = amd_iommu_rlookup_table[dev_data->devid];
3447 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3448 qdep, address, size);
3450 ret = iommu_queue_command(iommu, &cmd);
3455 /* Wait until all device TLBs are flushed */
3456 domain_flush_complete(domain);
3465 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3468 return __flush_pasid(domain, pasid, address, false);
3471 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3474 struct protection_domain *domain = to_pdomain(dom);
3475 unsigned long flags;
3478 spin_lock_irqsave(&domain->lock, flags);
3479 ret = __amd_iommu_flush_page(domain, pasid, address);
3480 spin_unlock_irqrestore(&domain->lock, flags);
3484 EXPORT_SYMBOL(amd_iommu_flush_page);
3486 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3488 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3492 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3494 struct protection_domain *domain = to_pdomain(dom);
3495 unsigned long flags;
3498 spin_lock_irqsave(&domain->lock, flags);
3499 ret = __amd_iommu_flush_tlb(domain, pasid);
3500 spin_unlock_irqrestore(&domain->lock, flags);
3504 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3506 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3513 index = (pasid >> (9 * level)) & 0x1ff;
3519 if (!(*pte & GCR3_VALID)) {
3523 root = (void *)get_zeroed_page(GFP_ATOMIC);
3527 *pte = __pa(root) | GCR3_VALID;
3530 root = __va(*pte & PAGE_MASK);
3538 static int __set_gcr3(struct protection_domain *domain, int pasid,
3543 if (domain->mode != PAGE_MODE_NONE)
3546 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3550 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3552 return __amd_iommu_flush_tlb(domain, pasid);
3555 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3559 if (domain->mode != PAGE_MODE_NONE)
3562 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3568 return __amd_iommu_flush_tlb(domain, pasid);
3571 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3574 struct protection_domain *domain = to_pdomain(dom);
3575 unsigned long flags;
3578 spin_lock_irqsave(&domain->lock, flags);
3579 ret = __set_gcr3(domain, pasid, cr3);
3580 spin_unlock_irqrestore(&domain->lock, flags);
3584 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3586 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3588 struct protection_domain *domain = to_pdomain(dom);
3589 unsigned long flags;
3592 spin_lock_irqsave(&domain->lock, flags);
3593 ret = __clear_gcr3(domain, pasid);
3594 spin_unlock_irqrestore(&domain->lock, flags);
3598 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3600 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3601 int status, int tag)
3603 struct iommu_dev_data *dev_data;
3604 struct amd_iommu *iommu;
3605 struct iommu_cmd cmd;
3607 dev_data = get_dev_data(&pdev->dev);
3608 iommu = amd_iommu_rlookup_table[dev_data->devid];
3610 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3611 tag, dev_data->pri_tlp);
3613 return iommu_queue_command(iommu, &cmd);
3615 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3617 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3619 struct protection_domain *pdomain;
3621 pdomain = get_domain(&pdev->dev);
3622 if (IS_ERR(pdomain))
3625 /* Only return IOMMUv2 domains */
3626 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3629 return &pdomain->domain;
3631 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3633 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3635 struct iommu_dev_data *dev_data;
3637 if (!amd_iommu_v2_supported())
3640 dev_data = get_dev_data(&pdev->dev);
3641 dev_data->errata |= (1 << erratum);
3643 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3645 int amd_iommu_device_info(struct pci_dev *pdev,
3646 struct amd_iommu_device_info *info)
3651 if (pdev == NULL || info == NULL)
3654 if (!amd_iommu_v2_supported())
3657 memset(info, 0, sizeof(*info));
3659 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3661 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3663 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3665 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3667 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3671 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3672 max_pasids = min(max_pasids, (1 << 20));
3674 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3675 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3677 features = pci_pasid_features(pdev);
3678 if (features & PCI_PASID_CAP_EXEC)
3679 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3680 if (features & PCI_PASID_CAP_PRIV)
3681 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3686 EXPORT_SYMBOL(amd_iommu_device_info);
3688 #ifdef CONFIG_IRQ_REMAP
3690 /*****************************************************************************
3692 * Interrupt Remapping Implementation
3694 *****************************************************************************/
3712 u16 devid; /* Device ID for IRTE table */
3713 u16 index; /* Index into IRTE table*/
3716 struct amd_ir_data {
3717 struct irq_2_irte irq_2_irte;
3718 union irte irte_entry;
3720 struct msi_msg msi_entry;
3724 static struct irq_chip amd_ir_chip;
3726 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3727 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3728 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3729 #define DTE_IRQ_REMAP_ENABLE 1ULL
3731 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3735 dte = amd_iommu_dev_table[devid].data[2];
3736 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3737 dte |= virt_to_phys(table->table);
3738 dte |= DTE_IRQ_REMAP_INTCTL;
3739 dte |= DTE_IRQ_TABLE_LEN;
3740 dte |= DTE_IRQ_REMAP_ENABLE;
3742 amd_iommu_dev_table[devid].data[2] = dte;
3745 #define IRTE_ALLOCATED (~1U)
3747 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3749 struct irq_remap_table *table = NULL;
3750 struct amd_iommu *iommu;
3751 unsigned long flags;
3754 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3756 iommu = amd_iommu_rlookup_table[devid];
3760 table = irq_lookup_table[devid];
3764 alias = amd_iommu_alias_table[devid];
3765 table = irq_lookup_table[alias];
3767 irq_lookup_table[devid] = table;
3768 set_dte_irq_entry(devid, table);
3769 iommu_flush_dte(iommu, devid);
3773 /* Nothing there yet, allocate new irq remapping table */
3774 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3778 /* Initialize table spin-lock */
3779 spin_lock_init(&table->lock);
3782 /* Keep the first 32 indexes free for IOAPIC interrupts */
3783 table->min_index = 32;
3785 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3786 if (!table->table) {
3792 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3797 for (i = 0; i < 32; ++i)
3798 table->table[i] = IRTE_ALLOCATED;
3801 irq_lookup_table[devid] = table;
3802 set_dte_irq_entry(devid, table);
3803 iommu_flush_dte(iommu, devid);
3804 if (devid != alias) {
3805 irq_lookup_table[alias] = table;
3806 set_dte_irq_entry(alias, table);
3807 iommu_flush_dte(iommu, alias);
3811 iommu_completion_wait(iommu);
3814 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3819 static int alloc_irq_index(u16 devid, int count)
3821 struct irq_remap_table *table;
3822 unsigned long flags;
3825 table = get_irq_table(devid, false);
3829 spin_lock_irqsave(&table->lock, flags);
3831 /* Scan table for free entries */
3832 for (c = 0, index = table->min_index;
3833 index < MAX_IRQS_PER_TABLE;
3835 if (table->table[index] == 0)
3842 table->table[index - c + 1] = IRTE_ALLOCATED;
3852 spin_unlock_irqrestore(&table->lock, flags);
3857 static int modify_irte(u16 devid, int index, union irte irte)
3859 struct irq_remap_table *table;
3860 struct amd_iommu *iommu;
3861 unsigned long flags;
3863 iommu = amd_iommu_rlookup_table[devid];
3867 table = get_irq_table(devid, false);
3871 spin_lock_irqsave(&table->lock, flags);
3872 table->table[index] = irte.val;
3873 spin_unlock_irqrestore(&table->lock, flags);
3875 iommu_flush_irt(iommu, devid);
3876 iommu_completion_wait(iommu);
3881 static void free_irte(u16 devid, int index)
3883 struct irq_remap_table *table;
3884 struct amd_iommu *iommu;
3885 unsigned long flags;
3887 iommu = amd_iommu_rlookup_table[devid];
3891 table = get_irq_table(devid, false);
3895 spin_lock_irqsave(&table->lock, flags);
3896 table->table[index] = 0;
3897 spin_unlock_irqrestore(&table->lock, flags);
3899 iommu_flush_irt(iommu, devid);
3900 iommu_completion_wait(iommu);
3903 static int get_devid(struct irq_alloc_info *info)
3907 switch (info->type) {
3908 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3909 devid = get_ioapic_devid(info->ioapic_id);
3911 case X86_IRQ_ALLOC_TYPE_HPET:
3912 devid = get_hpet_devid(info->hpet_id);
3914 case X86_IRQ_ALLOC_TYPE_MSI:
3915 case X86_IRQ_ALLOC_TYPE_MSIX:
3916 devid = get_device_id(&info->msi_dev->dev);
3926 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3928 struct amd_iommu *iommu;
3934 devid = get_devid(info);
3936 iommu = amd_iommu_rlookup_table[devid];
3938 return iommu->ir_domain;
3944 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3946 struct amd_iommu *iommu;
3952 switch (info->type) {
3953 case X86_IRQ_ALLOC_TYPE_MSI:
3954 case X86_IRQ_ALLOC_TYPE_MSIX:
3955 devid = get_device_id(&info->msi_dev->dev);
3959 iommu = amd_iommu_rlookup_table[devid];
3961 return iommu->msi_domain;
3970 struct irq_remap_ops amd_iommu_irq_ops = {
3971 .prepare = amd_iommu_prepare,
3972 .enable = amd_iommu_enable,
3973 .disable = amd_iommu_disable,
3974 .reenable = amd_iommu_reenable,
3975 .enable_faulting = amd_iommu_enable_faulting,
3976 .get_ir_irq_domain = get_ir_irq_domain,
3977 .get_irq_domain = get_irq_domain,
3980 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3981 struct irq_cfg *irq_cfg,
3982 struct irq_alloc_info *info,
3983 int devid, int index, int sub_handle)
3985 struct irq_2_irte *irte_info = &data->irq_2_irte;
3986 struct msi_msg *msg = &data->msi_entry;
3987 union irte *irte = &data->irte_entry;
3988 struct IO_APIC_route_entry *entry;
3990 data->irq_2_irte.devid = devid;
3991 data->irq_2_irte.index = index + sub_handle;
3993 /* Setup IRTE for IOMMU */
3995 irte->fields.vector = irq_cfg->vector;
3996 irte->fields.int_type = apic->irq_delivery_mode;
3997 irte->fields.destination = irq_cfg->dest_apicid;
3998 irte->fields.dm = apic->irq_dest_mode;
3999 irte->fields.valid = 1;
4001 switch (info->type) {
4002 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4003 /* Setup IOAPIC entry */
4004 entry = info->ioapic_entry;
4005 info->ioapic_entry = NULL;
4006 memset(entry, 0, sizeof(*entry));
4007 entry->vector = index;
4009 entry->trigger = info->ioapic_trigger;
4010 entry->polarity = info->ioapic_polarity;
4011 /* Mask level triggered irqs. */
4012 if (info->ioapic_trigger)
4016 case X86_IRQ_ALLOC_TYPE_HPET:
4017 case X86_IRQ_ALLOC_TYPE_MSI:
4018 case X86_IRQ_ALLOC_TYPE_MSIX:
4019 msg->address_hi = MSI_ADDR_BASE_HI;
4020 msg->address_lo = MSI_ADDR_BASE_LO;
4021 msg->data = irte_info->index;
4030 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4031 unsigned int nr_irqs, void *arg)
4033 struct irq_alloc_info *info = arg;
4034 struct irq_data *irq_data;
4035 struct amd_ir_data *data;
4036 struct irq_cfg *cfg;
4042 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4043 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4047 * With IRQ remapping enabled, don't need contiguous CPU vectors
4048 * to support multiple MSI interrupts.
4050 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4051 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4053 devid = get_devid(info);
4057 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4061 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4062 if (get_irq_table(devid, true))
4063 index = info->ioapic_pin;
4067 index = alloc_irq_index(devid, nr_irqs);
4070 pr_warn("Failed to allocate IRTE\n");
4071 goto out_free_parent;
4074 for (i = 0; i < nr_irqs; i++) {
4075 irq_data = irq_domain_get_irq_data(domain, virq + i);
4076 cfg = irqd_cfg(irq_data);
4077 if (!irq_data || !cfg) {
4083 data = kzalloc(sizeof(*data), GFP_KERNEL);
4087 irq_data->hwirq = (devid << 16) + i;
4088 irq_data->chip_data = data;
4089 irq_data->chip = &amd_ir_chip;
4090 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4091 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4097 for (i--; i >= 0; i--) {
4098 irq_data = irq_domain_get_irq_data(domain, virq + i);
4100 kfree(irq_data->chip_data);
4102 for (i = 0; i < nr_irqs; i++)
4103 free_irte(devid, index + i);
4105 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4109 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4110 unsigned int nr_irqs)
4112 struct irq_2_irte *irte_info;
4113 struct irq_data *irq_data;
4114 struct amd_ir_data *data;
4117 for (i = 0; i < nr_irqs; i++) {
4118 irq_data = irq_domain_get_irq_data(domain, virq + i);
4119 if (irq_data && irq_data->chip_data) {
4120 data = irq_data->chip_data;
4121 irte_info = &data->irq_2_irte;
4122 free_irte(irte_info->devid, irte_info->index);
4126 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4129 static void irq_remapping_activate(struct irq_domain *domain,
4130 struct irq_data *irq_data)
4132 struct amd_ir_data *data = irq_data->chip_data;
4133 struct irq_2_irte *irte_info = &data->irq_2_irte;
4135 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4138 static void irq_remapping_deactivate(struct irq_domain *domain,
4139 struct irq_data *irq_data)
4141 struct amd_ir_data *data = irq_data->chip_data;
4142 struct irq_2_irte *irte_info = &data->irq_2_irte;
4146 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4149 static struct irq_domain_ops amd_ir_domain_ops = {
4150 .alloc = irq_remapping_alloc,
4151 .free = irq_remapping_free,
4152 .activate = irq_remapping_activate,
4153 .deactivate = irq_remapping_deactivate,
4156 static int amd_ir_set_affinity(struct irq_data *data,
4157 const struct cpumask *mask, bool force)
4159 struct amd_ir_data *ir_data = data->chip_data;
4160 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4161 struct irq_cfg *cfg = irqd_cfg(data);
4162 struct irq_data *parent = data->parent_data;
4165 ret = parent->chip->irq_set_affinity(parent, mask, force);
4166 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4170 * Atomically updates the IRTE with the new destination, vector
4171 * and flushes the interrupt entry cache.
4173 ir_data->irte_entry.fields.vector = cfg->vector;
4174 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4175 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4178 * After this point, all the interrupts will start arriving
4179 * at the new destination. So, time to cleanup the previous
4180 * vector allocation.
4182 send_cleanup_vector(cfg);
4184 return IRQ_SET_MASK_OK_DONE;
4187 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4189 struct amd_ir_data *ir_data = irq_data->chip_data;
4191 *msg = ir_data->msi_entry;
4194 static struct irq_chip amd_ir_chip = {
4195 .irq_ack = ir_ack_apic_edge,
4196 .irq_set_affinity = amd_ir_set_affinity,
4197 .irq_compose_msi_msg = ir_compose_msi_msg,
4200 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4202 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4203 if (!iommu->ir_domain)
4206 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4207 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);