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/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 <linux/dma-contiguous.h>
37 #include <linux/irqdomain.h>
38 #include <linux/percpu.h>
39 #include <asm/irq_remapping.h>
40 #include <asm/io_apic.h>
42 #include <asm/hw_irq.h>
43 #include <asm/msidef.h>
44 #include <asm/proto.h>
45 #include <asm/iommu.h>
49 #include "amd_iommu_proto.h"
50 #include "amd_iommu_types.h"
51 #include "irq_remapping.h"
53 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
55 #define LOOP_TIMEOUT 100000
58 * This bitmap is used to advertise the page sizes our hardware support
59 * to the IOMMU core, which will then use this information to split
60 * physically contiguous memory regions it is mapping into page sizes
63 * 512GB Pages are not supported due to a hardware bug
65 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
67 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
69 /* List of all available dev_data structures */
70 static LIST_HEAD(dev_data_list);
71 static DEFINE_SPINLOCK(dev_data_list_lock);
73 LIST_HEAD(ioapic_map);
77 * Domain for untranslated devices - only allocated
78 * if iommu=pt passed on kernel cmd line.
80 static const struct iommu_ops amd_iommu_ops;
82 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
83 int amd_iommu_max_glx_val = -1;
85 static struct dma_map_ops amd_iommu_dma_ops;
88 * This struct contains device specific data for the IOMMU
90 struct iommu_dev_data {
91 struct list_head list; /* For domain->dev_list */
92 struct list_head dev_data_list; /* For global dev_data_list */
93 struct protection_domain *domain; /* Domain the device is bound to */
94 u16 devid; /* PCI Device ID */
95 u16 alias; /* Alias Device ID */
96 bool iommu_v2; /* Device can make use of IOMMUv2 */
97 bool passthrough; /* Device is identity mapped */
101 } ats; /* ATS state */
102 bool pri_tlp; /* PASID TLB required for
104 u32 errata; /* Bitmap for errata to apply */
108 * general struct to manage commands send to an IOMMU
114 struct kmem_cache *amd_iommu_irq_cache;
116 static void update_domain(struct protection_domain *domain);
117 static int protection_domain_init(struct protection_domain *domain);
118 static void detach_device(struct device *dev);
121 * For dynamic growth the aperture size is split into ranges of 128MB of
122 * DMA address space each. This struct represents one such range.
124 struct aperture_range {
126 spinlock_t bitmap_lock;
128 /* address allocation bitmap */
129 unsigned long *bitmap;
130 unsigned long offset;
131 unsigned long next_bit;
134 * Array of PTE pages for the aperture. In this array we save all the
135 * leaf pages of the domain page table used for the aperture. This way
136 * we don't need to walk the page table to find a specific PTE. We can
137 * just calculate its address in constant time.
143 * Data container for a dma_ops specific protection domain
145 struct dma_ops_domain {
146 /* generic protection domain information */
147 struct protection_domain domain;
149 /* size of the aperture for the mappings */
150 unsigned long aperture_size;
152 /* aperture index we start searching for free addresses */
153 u32 __percpu *next_index;
155 /* address space relevant data */
156 struct aperture_range *aperture[APERTURE_MAX_RANGES];
159 /****************************************************************************
163 ****************************************************************************/
165 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
167 return container_of(dom, struct protection_domain, domain);
170 static inline u16 get_device_id(struct device *dev)
172 struct pci_dev *pdev = to_pci_dev(dev);
174 return PCI_DEVID(pdev->bus->number, pdev->devfn);
177 static struct iommu_dev_data *alloc_dev_data(u16 devid)
179 struct iommu_dev_data *dev_data;
182 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
186 dev_data->devid = devid;
188 spin_lock_irqsave(&dev_data_list_lock, flags);
189 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
190 spin_unlock_irqrestore(&dev_data_list_lock, flags);
195 static struct iommu_dev_data *search_dev_data(u16 devid)
197 struct iommu_dev_data *dev_data;
200 spin_lock_irqsave(&dev_data_list_lock, flags);
201 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
202 if (dev_data->devid == devid)
209 spin_unlock_irqrestore(&dev_data_list_lock, flags);
214 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
216 *(u16 *)data = alias;
220 static u16 get_alias(struct device *dev)
222 struct pci_dev *pdev = to_pci_dev(dev);
223 u16 devid, ivrs_alias, pci_alias;
225 devid = get_device_id(dev);
226 ivrs_alias = amd_iommu_alias_table[devid];
227 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
229 if (ivrs_alias == pci_alias)
235 * The IVRS is fairly reliable in telling us about aliases, but it
236 * can't know about every screwy device. If we don't have an IVRS
237 * reported alias, use the PCI reported alias. In that case we may
238 * still need to initialize the rlookup and dev_table entries if the
239 * alias is to a non-existent device.
241 if (ivrs_alias == devid) {
242 if (!amd_iommu_rlookup_table[pci_alias]) {
243 amd_iommu_rlookup_table[pci_alias] =
244 amd_iommu_rlookup_table[devid];
245 memcpy(amd_iommu_dev_table[pci_alias].data,
246 amd_iommu_dev_table[devid].data,
247 sizeof(amd_iommu_dev_table[pci_alias].data));
253 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
254 "for device %s[%04x:%04x], kernel reported alias "
255 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
256 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
257 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
258 PCI_FUNC(pci_alias));
261 * If we don't have a PCI DMA alias and the IVRS alias is on the same
262 * bus, then the IVRS table may know about a quirk that we don't.
264 if (pci_alias == devid &&
265 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
266 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
267 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
268 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
275 static struct iommu_dev_data *find_dev_data(u16 devid)
277 struct iommu_dev_data *dev_data;
279 dev_data = search_dev_data(devid);
281 if (dev_data == NULL)
282 dev_data = alloc_dev_data(devid);
287 static struct iommu_dev_data *get_dev_data(struct device *dev)
289 return dev->archdata.iommu;
292 static bool pci_iommuv2_capable(struct pci_dev *pdev)
294 static const int caps[] = {
297 PCI_EXT_CAP_ID_PASID,
301 for (i = 0; i < 3; ++i) {
302 pos = pci_find_ext_capability(pdev, caps[i]);
310 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
312 struct iommu_dev_data *dev_data;
314 dev_data = get_dev_data(&pdev->dev);
316 return dev_data->errata & (1 << erratum) ? true : false;
320 * This function actually applies the mapping to the page table of the
323 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
324 struct unity_map_entry *e)
328 for (addr = e->address_start; addr < e->address_end;
330 if (addr < dma_dom->aperture_size)
331 __set_bit(addr >> PAGE_SHIFT,
332 dma_dom->aperture[0]->bitmap);
337 * Inits the unity mappings required for a specific device
339 static void init_unity_mappings_for_device(struct device *dev,
340 struct dma_ops_domain *dma_dom)
342 struct unity_map_entry *e;
345 devid = get_device_id(dev);
347 list_for_each_entry(e, &amd_iommu_unity_map, list) {
348 if (!(devid >= e->devid_start && devid <= e->devid_end))
350 alloc_unity_mapping(dma_dom, e);
355 * This function checks if the driver got a valid device from the caller to
356 * avoid dereferencing invalid pointers.
358 static bool check_device(struct device *dev)
362 if (!dev || !dev->dma_mask)
366 if (!dev_is_pci(dev))
369 devid = get_device_id(dev);
371 /* Out of our scope? */
372 if (devid > amd_iommu_last_bdf)
375 if (amd_iommu_rlookup_table[devid] == NULL)
381 static void init_iommu_group(struct device *dev)
383 struct dma_ops_domain *dma_domain;
384 struct iommu_domain *domain;
385 struct iommu_group *group;
387 group = iommu_group_get_for_dev(dev);
391 domain = iommu_group_default_domain(group);
395 dma_domain = to_pdomain(domain)->priv;
397 init_unity_mappings_for_device(dev, dma_domain);
399 iommu_group_put(group);
402 static int iommu_init_device(struct device *dev)
404 struct pci_dev *pdev = to_pci_dev(dev);
405 struct iommu_dev_data *dev_data;
407 if (dev->archdata.iommu)
410 dev_data = find_dev_data(get_device_id(dev));
414 dev_data->alias = get_alias(dev);
416 if (pci_iommuv2_capable(pdev)) {
417 struct amd_iommu *iommu;
419 iommu = amd_iommu_rlookup_table[dev_data->devid];
420 dev_data->iommu_v2 = iommu->is_iommu_v2;
423 dev->archdata.iommu = dev_data;
425 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
431 static void iommu_ignore_device(struct device *dev)
435 devid = get_device_id(dev);
436 alias = get_alias(dev);
438 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
439 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
441 amd_iommu_rlookup_table[devid] = NULL;
442 amd_iommu_rlookup_table[alias] = NULL;
445 static void iommu_uninit_device(struct device *dev)
447 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
452 if (dev_data->domain)
455 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
458 iommu_group_remove_device(dev);
461 dev->archdata.dma_ops = NULL;
464 * We keep dev_data around for unplugged devices and reuse it when the
465 * device is re-plugged - not doing so would introduce a ton of races.
469 #ifdef CONFIG_AMD_IOMMU_STATS
472 * Initialization code for statistics collection
475 DECLARE_STATS_COUNTER(compl_wait);
476 DECLARE_STATS_COUNTER(cnt_map_single);
477 DECLARE_STATS_COUNTER(cnt_unmap_single);
478 DECLARE_STATS_COUNTER(cnt_map_sg);
479 DECLARE_STATS_COUNTER(cnt_unmap_sg);
480 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
481 DECLARE_STATS_COUNTER(cnt_free_coherent);
482 DECLARE_STATS_COUNTER(cross_page);
483 DECLARE_STATS_COUNTER(domain_flush_single);
484 DECLARE_STATS_COUNTER(domain_flush_all);
485 DECLARE_STATS_COUNTER(alloced_io_mem);
486 DECLARE_STATS_COUNTER(total_map_requests);
487 DECLARE_STATS_COUNTER(complete_ppr);
488 DECLARE_STATS_COUNTER(invalidate_iotlb);
489 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
490 DECLARE_STATS_COUNTER(pri_requests);
492 static struct dentry *stats_dir;
493 static struct dentry *de_fflush;
495 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
497 if (stats_dir == NULL)
500 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
504 static void amd_iommu_stats_init(void)
506 stats_dir = debugfs_create_dir("amd-iommu", NULL);
507 if (stats_dir == NULL)
510 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
511 &amd_iommu_unmap_flush);
513 amd_iommu_stats_add(&compl_wait);
514 amd_iommu_stats_add(&cnt_map_single);
515 amd_iommu_stats_add(&cnt_unmap_single);
516 amd_iommu_stats_add(&cnt_map_sg);
517 amd_iommu_stats_add(&cnt_unmap_sg);
518 amd_iommu_stats_add(&cnt_alloc_coherent);
519 amd_iommu_stats_add(&cnt_free_coherent);
520 amd_iommu_stats_add(&cross_page);
521 amd_iommu_stats_add(&domain_flush_single);
522 amd_iommu_stats_add(&domain_flush_all);
523 amd_iommu_stats_add(&alloced_io_mem);
524 amd_iommu_stats_add(&total_map_requests);
525 amd_iommu_stats_add(&complete_ppr);
526 amd_iommu_stats_add(&invalidate_iotlb);
527 amd_iommu_stats_add(&invalidate_iotlb_all);
528 amd_iommu_stats_add(&pri_requests);
533 /****************************************************************************
535 * Interrupt handling functions
537 ****************************************************************************/
539 static void dump_dte_entry(u16 devid)
543 for (i = 0; i < 4; ++i)
544 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
545 amd_iommu_dev_table[devid].data[i]);
548 static void dump_command(unsigned long phys_addr)
550 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
553 for (i = 0; i < 4; ++i)
554 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
557 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
559 int type, devid, domid, flags;
560 volatile u32 *event = __evt;
565 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
566 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
567 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
568 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
569 address = (u64)(((u64)event[3]) << 32) | event[2];
572 /* Did we hit the erratum? */
573 if (++count == LOOP_TIMEOUT) {
574 pr_err("AMD-Vi: No event written to event log\n");
581 printk(KERN_ERR "AMD-Vi: Event logged [");
584 case EVENT_TYPE_ILL_DEV:
585 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
586 "address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
589 dump_dte_entry(devid);
591 case EVENT_TYPE_IO_FAULT:
592 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
593 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
594 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
595 domid, address, flags);
597 case EVENT_TYPE_DEV_TAB_ERR:
598 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
599 "address=0x%016llx flags=0x%04x]\n",
600 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
603 case EVENT_TYPE_PAGE_TAB_ERR:
604 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
605 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
606 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
607 domid, address, flags);
609 case EVENT_TYPE_ILL_CMD:
610 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
611 dump_command(address);
613 case EVENT_TYPE_CMD_HARD_ERR:
614 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
615 "flags=0x%04x]\n", address, flags);
617 case EVENT_TYPE_IOTLB_INV_TO:
618 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
619 "address=0x%016llx]\n",
620 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
623 case EVENT_TYPE_INV_DEV_REQ:
624 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
625 "address=0x%016llx flags=0x%04x]\n",
626 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
630 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
633 memset(__evt, 0, 4 * sizeof(u32));
636 static void iommu_poll_events(struct amd_iommu *iommu)
640 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
641 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
643 while (head != tail) {
644 iommu_print_event(iommu, iommu->evt_buf + head);
645 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
648 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
651 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
653 struct amd_iommu_fault fault;
655 INC_STATS_COUNTER(pri_requests);
657 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
658 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
662 fault.address = raw[1];
663 fault.pasid = PPR_PASID(raw[0]);
664 fault.device_id = PPR_DEVID(raw[0]);
665 fault.tag = PPR_TAG(raw[0]);
666 fault.flags = PPR_FLAGS(raw[0]);
668 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
671 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
675 if (iommu->ppr_log == NULL)
678 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
679 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
681 while (head != tail) {
686 raw = (u64 *)(iommu->ppr_log + head);
689 * Hardware bug: Interrupt may arrive before the entry is
690 * written to memory. If this happens we need to wait for the
693 for (i = 0; i < LOOP_TIMEOUT; ++i) {
694 if (PPR_REQ_TYPE(raw[0]) != 0)
699 /* Avoid memcpy function-call overhead */
704 * To detect the hardware bug we need to clear the entry
707 raw[0] = raw[1] = 0UL;
709 /* Update head pointer of hardware ring-buffer */
710 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
711 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
713 /* Handle PPR entry */
714 iommu_handle_ppr_entry(iommu, entry);
716 /* Refresh ring-buffer information */
717 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
718 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
722 irqreturn_t amd_iommu_int_thread(int irq, void *data)
724 struct amd_iommu *iommu = (struct amd_iommu *) data;
725 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
727 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
728 /* Enable EVT and PPR interrupts again */
729 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
730 iommu->mmio_base + MMIO_STATUS_OFFSET);
732 if (status & MMIO_STATUS_EVT_INT_MASK) {
733 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
734 iommu_poll_events(iommu);
737 if (status & MMIO_STATUS_PPR_INT_MASK) {
738 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
739 iommu_poll_ppr_log(iommu);
743 * Hardware bug: ERBT1312
744 * When re-enabling interrupt (by writing 1
745 * to clear the bit), the hardware might also try to set
746 * the interrupt bit in the event status register.
747 * In this scenario, the bit will be set, and disable
748 * subsequent interrupts.
750 * Workaround: The IOMMU driver should read back the
751 * status register and check if the interrupt bits are cleared.
752 * If not, driver will need to go through the interrupt handler
753 * again and re-clear the bits
755 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
760 irqreturn_t amd_iommu_int_handler(int irq, void *data)
762 return IRQ_WAKE_THREAD;
765 /****************************************************************************
767 * IOMMU command queuing functions
769 ****************************************************************************/
771 static int wait_on_sem(volatile u64 *sem)
775 while (*sem == 0 && i < LOOP_TIMEOUT) {
780 if (i == LOOP_TIMEOUT) {
781 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
788 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
789 struct iommu_cmd *cmd,
794 target = iommu->cmd_buf + tail;
795 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
797 /* Copy command to buffer */
798 memcpy(target, cmd, sizeof(*cmd));
800 /* Tell the IOMMU about it */
801 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
804 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
806 WARN_ON(address & 0x7ULL);
808 memset(cmd, 0, sizeof(*cmd));
809 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
810 cmd->data[1] = upper_32_bits(__pa(address));
812 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
815 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
817 memset(cmd, 0, sizeof(*cmd));
818 cmd->data[0] = devid;
819 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
822 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
823 size_t size, u16 domid, int pde)
828 pages = iommu_num_pages(address, size, PAGE_SIZE);
833 * If we have to flush more than one page, flush all
834 * TLB entries for this domain
836 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
840 address &= PAGE_MASK;
842 memset(cmd, 0, sizeof(*cmd));
843 cmd->data[1] |= domid;
844 cmd->data[2] = lower_32_bits(address);
845 cmd->data[3] = upper_32_bits(address);
846 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
847 if (s) /* size bit - we flush more than one 4kb page */
848 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
849 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
850 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
853 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
854 u64 address, size_t size)
859 pages = iommu_num_pages(address, size, PAGE_SIZE);
864 * If we have to flush more than one page, flush all
865 * TLB entries for this domain
867 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
871 address &= PAGE_MASK;
873 memset(cmd, 0, sizeof(*cmd));
874 cmd->data[0] = devid;
875 cmd->data[0] |= (qdep & 0xff) << 24;
876 cmd->data[1] = devid;
877 cmd->data[2] = lower_32_bits(address);
878 cmd->data[3] = upper_32_bits(address);
879 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
881 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
884 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
885 u64 address, bool size)
887 memset(cmd, 0, sizeof(*cmd));
889 address &= ~(0xfffULL);
891 cmd->data[0] = pasid;
892 cmd->data[1] = domid;
893 cmd->data[2] = lower_32_bits(address);
894 cmd->data[3] = upper_32_bits(address);
895 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
896 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
898 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
899 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
902 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
903 int qdep, u64 address, bool size)
905 memset(cmd, 0, sizeof(*cmd));
907 address &= ~(0xfffULL);
909 cmd->data[0] = devid;
910 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
911 cmd->data[0] |= (qdep & 0xff) << 24;
912 cmd->data[1] = devid;
913 cmd->data[1] |= (pasid & 0xff) << 16;
914 cmd->data[2] = lower_32_bits(address);
915 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
916 cmd->data[3] = upper_32_bits(address);
918 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
919 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
922 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
923 int status, int tag, bool gn)
925 memset(cmd, 0, sizeof(*cmd));
927 cmd->data[0] = devid;
929 cmd->data[1] = pasid;
930 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
932 cmd->data[3] = tag & 0x1ff;
933 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
935 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
938 static void build_inv_all(struct iommu_cmd *cmd)
940 memset(cmd, 0, sizeof(*cmd));
941 CMD_SET_TYPE(cmd, CMD_INV_ALL);
944 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
946 memset(cmd, 0, sizeof(*cmd));
947 cmd->data[0] = devid;
948 CMD_SET_TYPE(cmd, CMD_INV_IRT);
952 * Writes the command to the IOMMUs command buffer and informs the
953 * hardware about the new command.
955 static int iommu_queue_command_sync(struct amd_iommu *iommu,
956 struct iommu_cmd *cmd,
959 u32 left, tail, head, next_tail;
963 spin_lock_irqsave(&iommu->lock, flags);
965 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
966 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
967 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
968 left = (head - next_tail) % CMD_BUFFER_SIZE;
971 struct iommu_cmd sync_cmd;
972 volatile u64 sem = 0;
975 build_completion_wait(&sync_cmd, (u64)&sem);
976 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
978 spin_unlock_irqrestore(&iommu->lock, flags);
980 if ((ret = wait_on_sem(&sem)) != 0)
986 copy_cmd_to_buffer(iommu, cmd, tail);
988 /* We need to sync now to make sure all commands are processed */
989 iommu->need_sync = sync;
991 spin_unlock_irqrestore(&iommu->lock, flags);
996 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
998 return iommu_queue_command_sync(iommu, cmd, true);
1002 * This function queues a completion wait command into the command
1003 * buffer of an IOMMU
1005 static int iommu_completion_wait(struct amd_iommu *iommu)
1007 struct iommu_cmd cmd;
1008 volatile u64 sem = 0;
1011 if (!iommu->need_sync)
1014 build_completion_wait(&cmd, (u64)&sem);
1016 ret = iommu_queue_command_sync(iommu, &cmd, false);
1020 return wait_on_sem(&sem);
1023 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1025 struct iommu_cmd cmd;
1027 build_inv_dte(&cmd, devid);
1029 return iommu_queue_command(iommu, &cmd);
1032 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1036 for (devid = 0; devid <= 0xffff; ++devid)
1037 iommu_flush_dte(iommu, devid);
1039 iommu_completion_wait(iommu);
1043 * This function uses heavy locking and may disable irqs for some time. But
1044 * this is no issue because it is only called during resume.
1046 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1050 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1051 struct iommu_cmd cmd;
1052 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1054 iommu_queue_command(iommu, &cmd);
1057 iommu_completion_wait(iommu);
1060 static void iommu_flush_all(struct amd_iommu *iommu)
1062 struct iommu_cmd cmd;
1064 build_inv_all(&cmd);
1066 iommu_queue_command(iommu, &cmd);
1067 iommu_completion_wait(iommu);
1070 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1072 struct iommu_cmd cmd;
1074 build_inv_irt(&cmd, devid);
1076 iommu_queue_command(iommu, &cmd);
1079 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1083 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1084 iommu_flush_irt(iommu, devid);
1086 iommu_completion_wait(iommu);
1089 void iommu_flush_all_caches(struct amd_iommu *iommu)
1091 if (iommu_feature(iommu, FEATURE_IA)) {
1092 iommu_flush_all(iommu);
1094 iommu_flush_dte_all(iommu);
1095 iommu_flush_irt_all(iommu);
1096 iommu_flush_tlb_all(iommu);
1101 * Command send function for flushing on-device TLB
1103 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1104 u64 address, size_t size)
1106 struct amd_iommu *iommu;
1107 struct iommu_cmd cmd;
1110 qdep = dev_data->ats.qdep;
1111 iommu = amd_iommu_rlookup_table[dev_data->devid];
1113 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1115 return iommu_queue_command(iommu, &cmd);
1119 * Command send function for invalidating a device table entry
1121 static int device_flush_dte(struct iommu_dev_data *dev_data)
1123 struct amd_iommu *iommu;
1127 iommu = amd_iommu_rlookup_table[dev_data->devid];
1128 alias = dev_data->alias;
1130 ret = iommu_flush_dte(iommu, dev_data->devid);
1131 if (!ret && alias != dev_data->devid)
1132 ret = iommu_flush_dte(iommu, alias);
1136 if (dev_data->ats.enabled)
1137 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1143 * TLB invalidation function which is called from the mapping functions.
1144 * It invalidates a single PTE if the range to flush is within a single
1145 * page. Otherwise it flushes the whole TLB of the IOMMU.
1147 static void __domain_flush_pages(struct protection_domain *domain,
1148 u64 address, size_t size, int pde)
1150 struct iommu_dev_data *dev_data;
1151 struct iommu_cmd cmd;
1154 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1156 for (i = 0; i < amd_iommus_present; ++i) {
1157 if (!domain->dev_iommu[i])
1161 * Devices of this domain are behind this IOMMU
1162 * We need a TLB flush
1164 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1167 list_for_each_entry(dev_data, &domain->dev_list, list) {
1169 if (!dev_data->ats.enabled)
1172 ret |= device_flush_iotlb(dev_data, address, size);
1178 static void domain_flush_pages(struct protection_domain *domain,
1179 u64 address, size_t size)
1181 __domain_flush_pages(domain, address, size, 0);
1184 /* Flush the whole IO/TLB for a given protection domain */
1185 static void domain_flush_tlb(struct protection_domain *domain)
1187 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1190 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1191 static void domain_flush_tlb_pde(struct protection_domain *domain)
1193 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1196 static void domain_flush_complete(struct protection_domain *domain)
1200 for (i = 0; i < amd_iommus_present; ++i) {
1201 if (!domain->dev_iommu[i])
1205 * Devices of this domain are behind this IOMMU
1206 * We need to wait for completion of all commands.
1208 iommu_completion_wait(amd_iommus[i]);
1214 * This function flushes the DTEs for all devices in domain
1216 static void domain_flush_devices(struct protection_domain *domain)
1218 struct iommu_dev_data *dev_data;
1220 list_for_each_entry(dev_data, &domain->dev_list, list)
1221 device_flush_dte(dev_data);
1224 /****************************************************************************
1226 * The functions below are used the create the page table mappings for
1227 * unity mapped regions.
1229 ****************************************************************************/
1232 * This function is used to add another level to an IO page table. Adding
1233 * another level increases the size of the address space by 9 bits to a size up
1236 static bool increase_address_space(struct protection_domain *domain,
1241 if (domain->mode == PAGE_MODE_6_LEVEL)
1242 /* address space already 64 bit large */
1245 pte = (void *)get_zeroed_page(gfp);
1249 *pte = PM_LEVEL_PDE(domain->mode,
1250 virt_to_phys(domain->pt_root));
1251 domain->pt_root = pte;
1253 domain->updated = true;
1258 static u64 *alloc_pte(struct protection_domain *domain,
1259 unsigned long address,
1260 unsigned long page_size,
1267 BUG_ON(!is_power_of_2(page_size));
1269 while (address > PM_LEVEL_SIZE(domain->mode))
1270 increase_address_space(domain, gfp);
1272 level = domain->mode - 1;
1273 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1274 address = PAGE_SIZE_ALIGN(address, page_size);
1275 end_lvl = PAGE_SIZE_LEVEL(page_size);
1277 while (level > end_lvl) {
1282 if (!IOMMU_PTE_PRESENT(__pte)) {
1283 page = (u64 *)get_zeroed_page(gfp);
1287 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1289 if (cmpxchg64(pte, __pte, __npte)) {
1290 free_page((unsigned long)page);
1295 /* No level skipping support yet */
1296 if (PM_PTE_LEVEL(*pte) != level)
1301 pte = IOMMU_PTE_PAGE(*pte);
1303 if (pte_page && level == end_lvl)
1306 pte = &pte[PM_LEVEL_INDEX(level, address)];
1313 * This function checks if there is a PTE for a given dma address. If
1314 * there is one, it returns the pointer to it.
1316 static u64 *fetch_pte(struct protection_domain *domain,
1317 unsigned long address,
1318 unsigned long *page_size)
1323 if (address > PM_LEVEL_SIZE(domain->mode))
1326 level = domain->mode - 1;
1327 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1328 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1333 if (!IOMMU_PTE_PRESENT(*pte))
1337 if (PM_PTE_LEVEL(*pte) == 7 ||
1338 PM_PTE_LEVEL(*pte) == 0)
1341 /* No level skipping support yet */
1342 if (PM_PTE_LEVEL(*pte) != level)
1347 /* Walk to the next level */
1348 pte = IOMMU_PTE_PAGE(*pte);
1349 pte = &pte[PM_LEVEL_INDEX(level, address)];
1350 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1353 if (PM_PTE_LEVEL(*pte) == 0x07) {
1354 unsigned long pte_mask;
1357 * If we have a series of large PTEs, make
1358 * sure to return a pointer to the first one.
1360 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1361 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1362 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1369 * Generic mapping functions. It maps a physical address into a DMA
1370 * address space. It allocates the page table pages if necessary.
1371 * In the future it can be extended to a generic mapping function
1372 * supporting all features of AMD IOMMU page tables like level skipping
1373 * and full 64 bit address spaces.
1375 static int iommu_map_page(struct protection_domain *dom,
1376 unsigned long bus_addr,
1377 unsigned long phys_addr,
1379 unsigned long page_size)
1384 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1385 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1387 if (!(prot & IOMMU_PROT_MASK))
1390 count = PAGE_SIZE_PTE_COUNT(page_size);
1391 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1396 for (i = 0; i < count; ++i)
1397 if (IOMMU_PTE_PRESENT(pte[i]))
1401 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1402 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1404 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1406 if (prot & IOMMU_PROT_IR)
1407 __pte |= IOMMU_PTE_IR;
1408 if (prot & IOMMU_PROT_IW)
1409 __pte |= IOMMU_PTE_IW;
1411 for (i = 0; i < count; ++i)
1419 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1420 unsigned long bus_addr,
1421 unsigned long page_size)
1423 unsigned long long unmapped;
1424 unsigned long unmap_size;
1427 BUG_ON(!is_power_of_2(page_size));
1431 while (unmapped < page_size) {
1433 pte = fetch_pte(dom, bus_addr, &unmap_size);
1438 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1439 for (i = 0; i < count; i++)
1443 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1444 unmapped += unmap_size;
1447 BUG_ON(unmapped && !is_power_of_2(unmapped));
1452 /****************************************************************************
1454 * The next functions belong to the address allocator for the dma_ops
1455 * interface functions. They work like the allocators in the other IOMMU
1456 * drivers. Its basically a bitmap which marks the allocated pages in
1457 * the aperture. Maybe it could be enhanced in the future to a more
1458 * efficient allocator.
1460 ****************************************************************************/
1463 * The address allocator core functions.
1465 * called with domain->lock held
1469 * Used to reserve address ranges in the aperture (e.g. for exclusion
1472 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1473 unsigned long start_page,
1476 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1478 if (start_page + pages > last_page)
1479 pages = last_page - start_page;
1481 for (i = start_page; i < start_page + pages; ++i) {
1482 int index = i / APERTURE_RANGE_PAGES;
1483 int page = i % APERTURE_RANGE_PAGES;
1484 __set_bit(page, dom->aperture[index]->bitmap);
1489 * This function is used to add a new aperture range to an existing
1490 * aperture in case of dma_ops domain allocation or address allocation
1493 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1494 bool populate, gfp_t gfp)
1496 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1497 unsigned long i, old_size, pte_pgsize;
1498 struct aperture_range *range;
1499 struct amd_iommu *iommu;
1500 unsigned long flags;
1502 #ifdef CONFIG_IOMMU_STRESS
1506 if (index >= APERTURE_MAX_RANGES)
1509 range = kzalloc(sizeof(struct aperture_range), gfp);
1513 range->bitmap = (void *)get_zeroed_page(gfp);
1517 range->offset = dma_dom->aperture_size;
1519 spin_lock_init(&range->bitmap_lock);
1522 unsigned long address = dma_dom->aperture_size;
1523 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1524 u64 *pte, *pte_page;
1526 for (i = 0; i < num_ptes; ++i) {
1527 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1532 range->pte_pages[i] = pte_page;
1534 address += APERTURE_RANGE_SIZE / 64;
1538 spin_lock_irqsave(&dma_dom->domain.lock, flags);
1540 /* First take the bitmap_lock and then publish the range */
1541 spin_lock(&range->bitmap_lock);
1543 old_size = dma_dom->aperture_size;
1544 dma_dom->aperture[index] = range;
1545 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1547 /* Reserve address range used for MSI messages */
1548 if (old_size < MSI_ADDR_BASE_LO &&
1549 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1550 unsigned long spage;
1553 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1554 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1556 dma_ops_reserve_addresses(dma_dom, spage, pages);
1559 /* Initialize the exclusion range if necessary */
1560 for_each_iommu(iommu) {
1561 if (iommu->exclusion_start &&
1562 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1563 && iommu->exclusion_start < dma_dom->aperture_size) {
1564 unsigned long startpage;
1565 int pages = iommu_num_pages(iommu->exclusion_start,
1566 iommu->exclusion_length,
1568 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1569 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1574 * Check for areas already mapped as present in the new aperture
1575 * range and mark those pages as reserved in the allocator. Such
1576 * mappings may already exist as a result of requested unity
1577 * mappings for devices.
1579 for (i = dma_dom->aperture[index]->offset;
1580 i < dma_dom->aperture_size;
1582 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1583 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1586 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1590 update_domain(&dma_dom->domain);
1592 spin_unlock(&range->bitmap_lock);
1594 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
1599 update_domain(&dma_dom->domain);
1601 free_page((unsigned long)range->bitmap);
1608 static dma_addr_t dma_ops_aperture_alloc(struct dma_ops_domain *dom,
1609 struct aperture_range *range,
1610 unsigned long pages,
1611 unsigned long dma_mask,
1612 unsigned long boundary_size,
1613 unsigned long align_mask,
1616 unsigned long offset, limit, flags;
1620 offset = range->offset >> PAGE_SHIFT;
1621 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1622 dma_mask >> PAGE_SHIFT);
1625 if (!spin_trylock_irqsave(&range->bitmap_lock, flags))
1628 spin_lock_irqsave(&range->bitmap_lock, flags);
1631 address = iommu_area_alloc(range->bitmap, limit, range->next_bit,
1632 pages, offset, boundary_size, align_mask);
1633 if (address == -1) {
1634 /* Nothing found, retry one time */
1635 address = iommu_area_alloc(range->bitmap, limit,
1636 0, pages, offset, boundary_size,
1642 range->next_bit = address + pages;
1644 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1647 domain_flush_tlb(&dom->domain);
1648 domain_flush_complete(&dom->domain);
1654 static unsigned long dma_ops_area_alloc(struct device *dev,
1655 struct dma_ops_domain *dom,
1657 unsigned long align_mask,
1660 unsigned long boundary_size, mask;
1661 unsigned long address = -1;
1667 mask = dma_get_seg_boundary(dev);
1670 start = this_cpu_read(*dom->next_index);
1672 /* Sanity check - is it really necessary? */
1673 if (unlikely(start > APERTURE_MAX_RANGES)) {
1675 this_cpu_write(*dom->next_index, 0);
1678 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1679 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1681 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1682 struct aperture_range *range;
1685 index = (start + i) % APERTURE_MAX_RANGES;
1687 range = dom->aperture[index];
1689 if (!range || range->offset >= dma_mask)
1692 address = dma_ops_aperture_alloc(dom, range, pages,
1693 dma_mask, boundary_size,
1695 if (address != -1) {
1696 address = range->offset + (address << PAGE_SHIFT);
1697 this_cpu_write(*dom->next_index, index);
1702 if (address == -1 && first) {
1712 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1713 struct dma_ops_domain *dom,
1715 unsigned long align_mask,
1718 unsigned long address = -1;
1720 while (address == -1) {
1721 address = dma_ops_area_alloc(dev, dom, pages,
1722 align_mask, dma_mask);
1724 if (address == -1 && alloc_new_range(dom, false, GFP_ATOMIC))
1728 if (unlikely(address == -1))
1729 address = DMA_ERROR_CODE;
1731 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1737 * The address free function.
1739 * called with domain->lock held
1741 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1742 unsigned long address,
1745 unsigned i = address >> APERTURE_RANGE_SHIFT;
1746 struct aperture_range *range = dom->aperture[i];
1747 unsigned long flags;
1749 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1751 #ifdef CONFIG_IOMMU_STRESS
1756 if (amd_iommu_unmap_flush) {
1757 domain_flush_tlb(&dom->domain);
1758 domain_flush_complete(&dom->domain);
1761 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1763 spin_lock_irqsave(&range->bitmap_lock, flags);
1764 if (address + pages > range->next_bit)
1765 range->next_bit = address + pages;
1766 bitmap_clear(range->bitmap, address, pages);
1767 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1771 /****************************************************************************
1773 * The next functions belong to the domain allocation. A domain is
1774 * allocated for every IOMMU as the default domain. If device isolation
1775 * is enabled, every device get its own domain. The most important thing
1776 * about domains is the page table mapping the DMA address space they
1779 ****************************************************************************/
1782 * This function adds a protection domain to the global protection domain list
1784 static void add_domain_to_list(struct protection_domain *domain)
1786 unsigned long flags;
1788 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1789 list_add(&domain->list, &amd_iommu_pd_list);
1790 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1794 * This function removes a protection domain to the global
1795 * protection domain list
1797 static void del_domain_from_list(struct protection_domain *domain)
1799 unsigned long flags;
1801 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1802 list_del(&domain->list);
1803 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1806 static u16 domain_id_alloc(void)
1808 unsigned long flags;
1811 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1812 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1814 if (id > 0 && id < MAX_DOMAIN_ID)
1815 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1818 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1823 static void domain_id_free(int id)
1825 unsigned long flags;
1827 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1828 if (id > 0 && id < MAX_DOMAIN_ID)
1829 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1830 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1833 #define DEFINE_FREE_PT_FN(LVL, FN) \
1834 static void free_pt_##LVL (unsigned long __pt) \
1842 for (i = 0; i < 512; ++i) { \
1843 /* PTE present? */ \
1844 if (!IOMMU_PTE_PRESENT(pt[i])) \
1848 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1849 PM_PTE_LEVEL(pt[i]) == 7) \
1852 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1855 free_page((unsigned long)pt); \
1858 DEFINE_FREE_PT_FN(l2, free_page)
1859 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1860 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1861 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1862 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1864 static void free_pagetable(struct protection_domain *domain)
1866 unsigned long root = (unsigned long)domain->pt_root;
1868 switch (domain->mode) {
1869 case PAGE_MODE_NONE:
1871 case PAGE_MODE_1_LEVEL:
1874 case PAGE_MODE_2_LEVEL:
1877 case PAGE_MODE_3_LEVEL:
1880 case PAGE_MODE_4_LEVEL:
1883 case PAGE_MODE_5_LEVEL:
1886 case PAGE_MODE_6_LEVEL:
1894 static void free_gcr3_tbl_level1(u64 *tbl)
1899 for (i = 0; i < 512; ++i) {
1900 if (!(tbl[i] & GCR3_VALID))
1903 ptr = __va(tbl[i] & PAGE_MASK);
1905 free_page((unsigned long)ptr);
1909 static void free_gcr3_tbl_level2(u64 *tbl)
1914 for (i = 0; i < 512; ++i) {
1915 if (!(tbl[i] & GCR3_VALID))
1918 ptr = __va(tbl[i] & PAGE_MASK);
1920 free_gcr3_tbl_level1(ptr);
1924 static void free_gcr3_table(struct protection_domain *domain)
1926 if (domain->glx == 2)
1927 free_gcr3_tbl_level2(domain->gcr3_tbl);
1928 else if (domain->glx == 1)
1929 free_gcr3_tbl_level1(domain->gcr3_tbl);
1931 BUG_ON(domain->glx != 0);
1933 free_page((unsigned long)domain->gcr3_tbl);
1937 * Free a domain, only used if something went wrong in the
1938 * allocation path and we need to free an already allocated page table
1940 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1947 free_percpu(dom->next_index);
1949 del_domain_from_list(&dom->domain);
1951 free_pagetable(&dom->domain);
1953 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1954 if (!dom->aperture[i])
1956 free_page((unsigned long)dom->aperture[i]->bitmap);
1957 kfree(dom->aperture[i]);
1963 static int dma_ops_domain_alloc_apertures(struct dma_ops_domain *dma_dom,
1966 int ret, i, apertures;
1968 apertures = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1971 for (i = apertures; i < max_apertures; ++i) {
1972 ret = alloc_new_range(dma_dom, false, GFP_KERNEL);
1981 * Allocates a new protection domain usable for the dma_ops functions.
1982 * It also initializes the page table and the address allocator data
1983 * structures required for the dma_ops interface
1985 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1987 struct dma_ops_domain *dma_dom;
1990 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1994 if (protection_domain_init(&dma_dom->domain))
1997 dma_dom->next_index = alloc_percpu(u32);
1998 if (!dma_dom->next_index)
2001 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2002 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2003 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2004 dma_dom->domain.priv = dma_dom;
2005 if (!dma_dom->domain.pt_root)
2008 add_domain_to_list(&dma_dom->domain);
2010 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2014 * mark the first page as allocated so we never return 0 as
2015 * a valid dma-address. So we can use 0 as error value
2017 dma_dom->aperture[0]->bitmap[0] = 1;
2019 for_each_possible_cpu(cpu)
2020 *per_cpu_ptr(dma_dom->next_index, cpu) = 0;
2025 dma_ops_domain_free(dma_dom);
2031 * little helper function to check whether a given protection domain is a
2034 static bool dma_ops_domain(struct protection_domain *domain)
2036 return domain->flags & PD_DMA_OPS_MASK;
2039 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2044 if (domain->mode != PAGE_MODE_NONE)
2045 pte_root = virt_to_phys(domain->pt_root);
2047 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2048 << DEV_ENTRY_MODE_SHIFT;
2049 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2051 flags = amd_iommu_dev_table[devid].data[1];
2054 flags |= DTE_FLAG_IOTLB;
2056 if (domain->flags & PD_IOMMUV2_MASK) {
2057 u64 gcr3 = __pa(domain->gcr3_tbl);
2058 u64 glx = domain->glx;
2061 pte_root |= DTE_FLAG_GV;
2062 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2064 /* First mask out possible old values for GCR3 table */
2065 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2068 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2071 /* Encode GCR3 table into DTE */
2072 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2075 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2078 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2082 flags &= ~(0xffffUL);
2083 flags |= domain->id;
2085 amd_iommu_dev_table[devid].data[1] = flags;
2086 amd_iommu_dev_table[devid].data[0] = pte_root;
2089 static void clear_dte_entry(u16 devid)
2091 /* remove entry from the device table seen by the hardware */
2092 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2093 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2095 amd_iommu_apply_erratum_63(devid);
2098 static void do_attach(struct iommu_dev_data *dev_data,
2099 struct protection_domain *domain)
2101 struct amd_iommu *iommu;
2105 iommu = amd_iommu_rlookup_table[dev_data->devid];
2106 alias = dev_data->alias;
2107 ats = dev_data->ats.enabled;
2109 /* Update data structures */
2110 dev_data->domain = domain;
2111 list_add(&dev_data->list, &domain->dev_list);
2113 /* Do reference counting */
2114 domain->dev_iommu[iommu->index] += 1;
2115 domain->dev_cnt += 1;
2117 /* Update device table */
2118 set_dte_entry(dev_data->devid, domain, ats);
2119 if (alias != dev_data->devid)
2120 set_dte_entry(alias, domain, ats);
2122 device_flush_dte(dev_data);
2125 static void do_detach(struct iommu_dev_data *dev_data)
2127 struct amd_iommu *iommu;
2131 * First check if the device is still attached. It might already
2132 * be detached from its domain because the generic
2133 * iommu_detach_group code detached it and we try again here in
2134 * our alias handling.
2136 if (!dev_data->domain)
2139 iommu = amd_iommu_rlookup_table[dev_data->devid];
2140 alias = dev_data->alias;
2142 /* decrease reference counters */
2143 dev_data->domain->dev_iommu[iommu->index] -= 1;
2144 dev_data->domain->dev_cnt -= 1;
2146 /* Update data structures */
2147 dev_data->domain = NULL;
2148 list_del(&dev_data->list);
2149 clear_dte_entry(dev_data->devid);
2150 if (alias != dev_data->devid)
2151 clear_dte_entry(alias);
2153 /* Flush the DTE entry */
2154 device_flush_dte(dev_data);
2158 * If a device is not yet associated with a domain, this function does
2159 * assigns it visible for the hardware
2161 static int __attach_device(struct iommu_dev_data *dev_data,
2162 struct protection_domain *domain)
2167 * Must be called with IRQs disabled. Warn here to detect early
2170 WARN_ON(!irqs_disabled());
2173 spin_lock(&domain->lock);
2176 if (dev_data->domain != NULL)
2179 /* Attach alias group root */
2180 do_attach(dev_data, domain);
2187 spin_unlock(&domain->lock);
2193 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2195 pci_disable_ats(pdev);
2196 pci_disable_pri(pdev);
2197 pci_disable_pasid(pdev);
2200 /* FIXME: Change generic reset-function to do the same */
2201 static int pri_reset_while_enabled(struct pci_dev *pdev)
2206 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2210 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2211 control |= PCI_PRI_CTRL_RESET;
2212 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2217 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2222 /* FIXME: Hardcode number of outstanding requests for now */
2224 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2226 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2228 /* Only allow access to user-accessible pages */
2229 ret = pci_enable_pasid(pdev, 0);
2233 /* First reset the PRI state of the device */
2234 ret = pci_reset_pri(pdev);
2239 ret = pci_enable_pri(pdev, reqs);
2244 ret = pri_reset_while_enabled(pdev);
2249 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2256 pci_disable_pri(pdev);
2257 pci_disable_pasid(pdev);
2262 /* FIXME: Move this to PCI code */
2263 #define PCI_PRI_TLP_OFF (1 << 15)
2265 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2270 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2274 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2276 return (status & PCI_PRI_TLP_OFF) ? true : false;
2280 * If a device is not yet associated with a domain, this function
2281 * assigns it visible for the hardware
2283 static int attach_device(struct device *dev,
2284 struct protection_domain *domain)
2286 struct pci_dev *pdev = to_pci_dev(dev);
2287 struct iommu_dev_data *dev_data;
2288 unsigned long flags;
2291 dev_data = get_dev_data(dev);
2293 if (domain->flags & PD_IOMMUV2_MASK) {
2294 if (!dev_data->passthrough)
2297 if (dev_data->iommu_v2) {
2298 if (pdev_iommuv2_enable(pdev) != 0)
2301 dev_data->ats.enabled = true;
2302 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2303 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2305 } else if (amd_iommu_iotlb_sup &&
2306 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2307 dev_data->ats.enabled = true;
2308 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2311 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2312 ret = __attach_device(dev_data, domain);
2313 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2316 * We might boot into a crash-kernel here. The crashed kernel
2317 * left the caches in the IOMMU dirty. So we have to flush
2318 * here to evict all dirty stuff.
2320 domain_flush_tlb_pde(domain);
2326 * Removes a device from a protection domain (unlocked)
2328 static void __detach_device(struct iommu_dev_data *dev_data)
2330 struct protection_domain *domain;
2333 * Must be called with IRQs disabled. Warn here to detect early
2336 WARN_ON(!irqs_disabled());
2338 if (WARN_ON(!dev_data->domain))
2341 domain = dev_data->domain;
2343 spin_lock(&domain->lock);
2345 do_detach(dev_data);
2347 spin_unlock(&domain->lock);
2351 * Removes a device from a protection domain (with devtable_lock held)
2353 static void detach_device(struct device *dev)
2355 struct protection_domain *domain;
2356 struct iommu_dev_data *dev_data;
2357 unsigned long flags;
2359 dev_data = get_dev_data(dev);
2360 domain = dev_data->domain;
2362 /* lock device table */
2363 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2364 __detach_device(dev_data);
2365 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2367 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2368 pdev_iommuv2_disable(to_pci_dev(dev));
2369 else if (dev_data->ats.enabled)
2370 pci_disable_ats(to_pci_dev(dev));
2372 dev_data->ats.enabled = false;
2375 static int amd_iommu_add_device(struct device *dev)
2377 struct iommu_dev_data *dev_data;
2378 struct iommu_domain *domain;
2379 struct amd_iommu *iommu;
2383 if (!check_device(dev) || get_dev_data(dev))
2386 devid = get_device_id(dev);
2387 iommu = amd_iommu_rlookup_table[devid];
2389 ret = iommu_init_device(dev);
2391 if (ret != -ENOTSUPP)
2392 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2395 iommu_ignore_device(dev);
2396 dev->archdata.dma_ops = &nommu_dma_ops;
2399 init_iommu_group(dev);
2401 dev_data = get_dev_data(dev);
2405 if (iommu_pass_through || dev_data->iommu_v2)
2406 iommu_request_dm_for_dev(dev);
2408 /* Domains are initialized for this device - have a look what we ended up with */
2409 domain = iommu_get_domain_for_dev(dev);
2410 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2411 dev_data->passthrough = true;
2413 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2416 iommu_completion_wait(iommu);
2421 static void amd_iommu_remove_device(struct device *dev)
2423 struct amd_iommu *iommu;
2426 if (!check_device(dev))
2429 devid = get_device_id(dev);
2430 iommu = amd_iommu_rlookup_table[devid];
2432 iommu_uninit_device(dev);
2433 iommu_completion_wait(iommu);
2436 /*****************************************************************************
2438 * The next functions belong to the dma_ops mapping/unmapping code.
2440 *****************************************************************************/
2443 * In the dma_ops path we only have the struct device. This function
2444 * finds the corresponding IOMMU, the protection domain and the
2445 * requestor id for a given device.
2446 * If the device is not yet associated with a domain this is also done
2449 static struct protection_domain *get_domain(struct device *dev)
2451 struct protection_domain *domain;
2452 struct iommu_domain *io_domain;
2454 if (!check_device(dev))
2455 return ERR_PTR(-EINVAL);
2457 io_domain = iommu_get_domain_for_dev(dev);
2461 domain = to_pdomain(io_domain);
2462 if (!dma_ops_domain(domain))
2463 return ERR_PTR(-EBUSY);
2468 static void update_device_table(struct protection_domain *domain)
2470 struct iommu_dev_data *dev_data;
2472 list_for_each_entry(dev_data, &domain->dev_list, list)
2473 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2476 static void update_domain(struct protection_domain *domain)
2478 if (!domain->updated)
2481 update_device_table(domain);
2483 domain_flush_devices(domain);
2484 domain_flush_tlb_pde(domain);
2486 domain->updated = false;
2490 * This function fetches the PTE for a given address in the aperture
2492 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2493 unsigned long address)
2495 struct aperture_range *aperture;
2496 u64 *pte, *pte_page;
2498 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2502 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2504 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2506 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2508 pte += PM_LEVEL_INDEX(0, address);
2510 update_domain(&dom->domain);
2516 * This is the generic map function. It maps one 4kb page at paddr to
2517 * the given address in the DMA address space for the domain.
2519 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2520 unsigned long address,
2526 WARN_ON(address > dom->aperture_size);
2530 pte = dma_ops_get_pte(dom, address);
2532 return DMA_ERROR_CODE;
2534 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2536 if (direction == DMA_TO_DEVICE)
2537 __pte |= IOMMU_PTE_IR;
2538 else if (direction == DMA_FROM_DEVICE)
2539 __pte |= IOMMU_PTE_IW;
2540 else if (direction == DMA_BIDIRECTIONAL)
2541 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2547 return (dma_addr_t)address;
2551 * The generic unmapping function for on page in the DMA address space.
2553 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2554 unsigned long address)
2556 struct aperture_range *aperture;
2559 if (address >= dom->aperture_size)
2562 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2566 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2570 pte += PM_LEVEL_INDEX(0, address);
2572 WARN_ON_ONCE(!*pte);
2578 * This function contains common code for mapping of a physically
2579 * contiguous memory region into DMA address space. It is used by all
2580 * mapping functions provided with this IOMMU driver.
2581 * Must be called with the domain lock held.
2583 static dma_addr_t __map_single(struct device *dev,
2584 struct dma_ops_domain *dma_dom,
2591 dma_addr_t offset = paddr & ~PAGE_MASK;
2592 dma_addr_t address, start, ret;
2594 unsigned long align_mask = 0;
2597 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2600 INC_STATS_COUNTER(total_map_requests);
2603 INC_STATS_COUNTER(cross_page);
2606 align_mask = (1UL << get_order(size)) - 1;
2608 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2611 if (address == DMA_ERROR_CODE)
2615 for (i = 0; i < pages; ++i) {
2616 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2617 if (ret == DMA_ERROR_CODE)
2625 ADD_STATS_COUNTER(alloced_io_mem, size);
2627 if (unlikely(amd_iommu_np_cache)) {
2628 domain_flush_pages(&dma_dom->domain, address, size);
2629 domain_flush_complete(&dma_dom->domain);
2637 for (--i; i >= 0; --i) {
2639 dma_ops_domain_unmap(dma_dom, start);
2642 dma_ops_free_addresses(dma_dom, address, pages);
2644 return DMA_ERROR_CODE;
2648 * Does the reverse of the __map_single function. Must be called with
2649 * the domain lock held too
2651 static void __unmap_single(struct dma_ops_domain *dma_dom,
2652 dma_addr_t dma_addr,
2656 dma_addr_t flush_addr;
2657 dma_addr_t i, start;
2660 if ((dma_addr == DMA_ERROR_CODE) ||
2661 (dma_addr + size > dma_dom->aperture_size))
2664 flush_addr = dma_addr;
2665 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2666 dma_addr &= PAGE_MASK;
2669 for (i = 0; i < pages; ++i) {
2670 dma_ops_domain_unmap(dma_dom, start);
2674 SUB_STATS_COUNTER(alloced_io_mem, size);
2676 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2680 * The exported map_single function for dma_ops.
2682 static dma_addr_t map_page(struct device *dev, struct page *page,
2683 unsigned long offset, size_t size,
2684 enum dma_data_direction dir,
2685 struct dma_attrs *attrs)
2687 phys_addr_t paddr = page_to_phys(page) + offset;
2688 struct protection_domain *domain;
2691 INC_STATS_COUNTER(cnt_map_single);
2693 domain = get_domain(dev);
2694 if (PTR_ERR(domain) == -EINVAL)
2695 return (dma_addr_t)paddr;
2696 else if (IS_ERR(domain))
2697 return DMA_ERROR_CODE;
2699 dma_mask = *dev->dma_mask;
2701 return __map_single(dev, domain->priv, paddr, size, dir, false,
2706 * The exported unmap_single function for dma_ops.
2708 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2709 enum dma_data_direction dir, struct dma_attrs *attrs)
2711 struct protection_domain *domain;
2713 INC_STATS_COUNTER(cnt_unmap_single);
2715 domain = get_domain(dev);
2719 __unmap_single(domain->priv, dma_addr, size, dir);
2723 * The exported map_sg function for dma_ops (handles scatter-gather
2726 static int map_sg(struct device *dev, struct scatterlist *sglist,
2727 int nelems, enum dma_data_direction dir,
2728 struct dma_attrs *attrs)
2730 struct protection_domain *domain;
2732 struct scatterlist *s;
2734 int mapped_elems = 0;
2737 INC_STATS_COUNTER(cnt_map_sg);
2739 domain = get_domain(dev);
2743 dma_mask = *dev->dma_mask;
2745 for_each_sg(sglist, s, nelems, i) {
2748 s->dma_address = __map_single(dev, domain->priv,
2749 paddr, s->length, dir, false,
2752 if (s->dma_address) {
2753 s->dma_length = s->length;
2759 return mapped_elems;
2762 for_each_sg(sglist, s, mapped_elems, i) {
2764 __unmap_single(domain->priv, s->dma_address,
2765 s->dma_length, dir);
2766 s->dma_address = s->dma_length = 0;
2773 * The exported map_sg function for dma_ops (handles scatter-gather
2776 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2777 int nelems, enum dma_data_direction dir,
2778 struct dma_attrs *attrs)
2780 struct protection_domain *domain;
2781 struct scatterlist *s;
2784 INC_STATS_COUNTER(cnt_unmap_sg);
2786 domain = get_domain(dev);
2790 for_each_sg(sglist, s, nelems, i) {
2791 __unmap_single(domain->priv, s->dma_address,
2792 s->dma_length, dir);
2793 s->dma_address = s->dma_length = 0;
2798 * The exported alloc_coherent function for dma_ops.
2800 static void *alloc_coherent(struct device *dev, size_t size,
2801 dma_addr_t *dma_addr, gfp_t flag,
2802 struct dma_attrs *attrs)
2804 u64 dma_mask = dev->coherent_dma_mask;
2805 struct protection_domain *domain;
2808 INC_STATS_COUNTER(cnt_alloc_coherent);
2810 domain = get_domain(dev);
2811 if (PTR_ERR(domain) == -EINVAL) {
2812 page = alloc_pages(flag, get_order(size));
2813 *dma_addr = page_to_phys(page);
2814 return page_address(page);
2815 } else if (IS_ERR(domain))
2818 size = PAGE_ALIGN(size);
2819 dma_mask = dev->coherent_dma_mask;
2820 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2823 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2825 if (!gfpflags_allow_blocking(flag))
2828 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2835 dma_mask = *dev->dma_mask;
2837 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2838 size, DMA_BIDIRECTIONAL, true, dma_mask);
2840 if (*dma_addr == DMA_ERROR_CODE)
2843 return page_address(page);
2847 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2848 __free_pages(page, get_order(size));
2854 * The exported free_coherent function for dma_ops.
2856 static void free_coherent(struct device *dev, size_t size,
2857 void *virt_addr, dma_addr_t dma_addr,
2858 struct dma_attrs *attrs)
2860 struct protection_domain *domain;
2863 INC_STATS_COUNTER(cnt_free_coherent);
2865 page = virt_to_page(virt_addr);
2866 size = PAGE_ALIGN(size);
2868 domain = get_domain(dev);
2872 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2875 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2876 __free_pages(page, get_order(size));
2880 * This function is called by the DMA layer to find out if we can handle a
2881 * particular device. It is part of the dma_ops.
2883 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2885 return check_device(dev);
2888 static int set_dma_mask(struct device *dev, u64 mask)
2890 struct protection_domain *domain;
2891 int max_apertures = 1;
2893 domain = get_domain(dev);
2895 return PTR_ERR(domain);
2897 if (mask == DMA_BIT_MASK(64))
2899 else if (mask > DMA_BIT_MASK(32))
2903 * To prevent lock contention it doesn't make sense to allocate more
2904 * apertures than online cpus
2906 if (max_apertures > num_online_cpus())
2907 max_apertures = num_online_cpus();
2909 if (dma_ops_domain_alloc_apertures(domain->priv, max_apertures))
2910 dev_err(dev, "Can't allocate %d iommu apertures\n",
2916 static struct dma_map_ops amd_iommu_dma_ops = {
2917 .alloc = alloc_coherent,
2918 .free = free_coherent,
2919 .map_page = map_page,
2920 .unmap_page = unmap_page,
2922 .unmap_sg = unmap_sg,
2923 .dma_supported = amd_iommu_dma_supported,
2924 .set_dma_mask = set_dma_mask,
2927 int __init amd_iommu_init_api(void)
2929 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2932 int __init amd_iommu_init_dma_ops(void)
2934 swiotlb = iommu_pass_through ? 1 : 0;
2938 * In case we don't initialize SWIOTLB (actually the common case
2939 * when AMD IOMMU is enabled), make sure there are global
2940 * dma_ops set as a fall-back for devices not handled by this
2941 * driver (for example non-PCI devices).
2944 dma_ops = &nommu_dma_ops;
2946 amd_iommu_stats_init();
2948 if (amd_iommu_unmap_flush)
2949 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2951 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2956 /*****************************************************************************
2958 * The following functions belong to the exported interface of AMD IOMMU
2960 * This interface allows access to lower level functions of the IOMMU
2961 * like protection domain handling and assignement of devices to domains
2962 * which is not possible with the dma_ops interface.
2964 *****************************************************************************/
2966 static void cleanup_domain(struct protection_domain *domain)
2968 struct iommu_dev_data *entry;
2969 unsigned long flags;
2971 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2973 while (!list_empty(&domain->dev_list)) {
2974 entry = list_first_entry(&domain->dev_list,
2975 struct iommu_dev_data, list);
2976 __detach_device(entry);
2979 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2982 static void protection_domain_free(struct protection_domain *domain)
2987 del_domain_from_list(domain);
2990 domain_id_free(domain->id);
2995 static int protection_domain_init(struct protection_domain *domain)
2997 spin_lock_init(&domain->lock);
2998 mutex_init(&domain->api_lock);
2999 domain->id = domain_id_alloc();
3002 INIT_LIST_HEAD(&domain->dev_list);
3007 static struct protection_domain *protection_domain_alloc(void)
3009 struct protection_domain *domain;
3011 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3015 if (protection_domain_init(domain))
3018 add_domain_to_list(domain);
3028 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
3030 struct protection_domain *pdomain;
3031 struct dma_ops_domain *dma_domain;
3034 case IOMMU_DOMAIN_UNMANAGED:
3035 pdomain = protection_domain_alloc();
3039 pdomain->mode = PAGE_MODE_3_LEVEL;
3040 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3041 if (!pdomain->pt_root) {
3042 protection_domain_free(pdomain);
3046 pdomain->domain.geometry.aperture_start = 0;
3047 pdomain->domain.geometry.aperture_end = ~0ULL;
3048 pdomain->domain.geometry.force_aperture = true;
3051 case IOMMU_DOMAIN_DMA:
3052 dma_domain = dma_ops_domain_alloc();
3054 pr_err("AMD-Vi: Failed to allocate\n");
3057 pdomain = &dma_domain->domain;
3059 case IOMMU_DOMAIN_IDENTITY:
3060 pdomain = protection_domain_alloc();
3064 pdomain->mode = PAGE_MODE_NONE;
3070 return &pdomain->domain;
3073 static void amd_iommu_domain_free(struct iommu_domain *dom)
3075 struct protection_domain *domain;
3080 domain = to_pdomain(dom);
3082 if (domain->dev_cnt > 0)
3083 cleanup_domain(domain);
3085 BUG_ON(domain->dev_cnt != 0);
3087 if (domain->mode != PAGE_MODE_NONE)
3088 free_pagetable(domain);
3090 if (domain->flags & PD_IOMMUV2_MASK)
3091 free_gcr3_table(domain);
3093 protection_domain_free(domain);
3096 static void amd_iommu_detach_device(struct iommu_domain *dom,
3099 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3100 struct amd_iommu *iommu;
3103 if (!check_device(dev))
3106 devid = get_device_id(dev);
3108 if (dev_data->domain != NULL)
3111 iommu = amd_iommu_rlookup_table[devid];
3115 iommu_completion_wait(iommu);
3118 static int amd_iommu_attach_device(struct iommu_domain *dom,
3121 struct protection_domain *domain = to_pdomain(dom);
3122 struct iommu_dev_data *dev_data;
3123 struct amd_iommu *iommu;
3126 if (!check_device(dev))
3129 dev_data = dev->archdata.iommu;
3131 iommu = amd_iommu_rlookup_table[dev_data->devid];
3135 if (dev_data->domain)
3138 ret = attach_device(dev, domain);
3140 iommu_completion_wait(iommu);
3145 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3146 phys_addr_t paddr, size_t page_size, int iommu_prot)
3148 struct protection_domain *domain = to_pdomain(dom);
3152 if (domain->mode == PAGE_MODE_NONE)
3155 if (iommu_prot & IOMMU_READ)
3156 prot |= IOMMU_PROT_IR;
3157 if (iommu_prot & IOMMU_WRITE)
3158 prot |= IOMMU_PROT_IW;
3160 mutex_lock(&domain->api_lock);
3161 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3162 mutex_unlock(&domain->api_lock);
3167 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3170 struct protection_domain *domain = to_pdomain(dom);
3173 if (domain->mode == PAGE_MODE_NONE)
3176 mutex_lock(&domain->api_lock);
3177 unmap_size = iommu_unmap_page(domain, iova, page_size);
3178 mutex_unlock(&domain->api_lock);
3180 domain_flush_tlb_pde(domain);
3185 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3188 struct protection_domain *domain = to_pdomain(dom);
3189 unsigned long offset_mask, pte_pgsize;
3192 if (domain->mode == PAGE_MODE_NONE)
3195 pte = fetch_pte(domain, iova, &pte_pgsize);
3197 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3200 offset_mask = pte_pgsize - 1;
3201 __pte = *pte & PM_ADDR_MASK;
3203 return (__pte & ~offset_mask) | (iova & offset_mask);
3206 static bool amd_iommu_capable(enum iommu_cap cap)
3209 case IOMMU_CAP_CACHE_COHERENCY:
3211 case IOMMU_CAP_INTR_REMAP:
3212 return (irq_remapping_enabled == 1);
3213 case IOMMU_CAP_NOEXEC:
3220 static void amd_iommu_get_dm_regions(struct device *dev,
3221 struct list_head *head)
3223 struct unity_map_entry *entry;
3226 devid = get_device_id(dev);
3228 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3229 struct iommu_dm_region *region;
3231 if (devid < entry->devid_start || devid > entry->devid_end)
3234 region = kzalloc(sizeof(*region), GFP_KERNEL);
3236 pr_err("Out of memory allocating dm-regions for %s\n",
3241 region->start = entry->address_start;
3242 region->length = entry->address_end - entry->address_start;
3243 if (entry->prot & IOMMU_PROT_IR)
3244 region->prot |= IOMMU_READ;
3245 if (entry->prot & IOMMU_PROT_IW)
3246 region->prot |= IOMMU_WRITE;
3248 list_add_tail(®ion->list, head);
3252 static void amd_iommu_put_dm_regions(struct device *dev,
3253 struct list_head *head)
3255 struct iommu_dm_region *entry, *next;
3257 list_for_each_entry_safe(entry, next, head, list)
3261 static const struct iommu_ops amd_iommu_ops = {
3262 .capable = amd_iommu_capable,
3263 .domain_alloc = amd_iommu_domain_alloc,
3264 .domain_free = amd_iommu_domain_free,
3265 .attach_dev = amd_iommu_attach_device,
3266 .detach_dev = amd_iommu_detach_device,
3267 .map = amd_iommu_map,
3268 .unmap = amd_iommu_unmap,
3269 .map_sg = default_iommu_map_sg,
3270 .iova_to_phys = amd_iommu_iova_to_phys,
3271 .add_device = amd_iommu_add_device,
3272 .remove_device = amd_iommu_remove_device,
3273 .device_group = pci_device_group,
3274 .get_dm_regions = amd_iommu_get_dm_regions,
3275 .put_dm_regions = amd_iommu_put_dm_regions,
3276 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3279 /*****************************************************************************
3281 * The next functions do a basic initialization of IOMMU for pass through
3284 * In passthrough mode the IOMMU is initialized and enabled but not used for
3285 * DMA-API translation.
3287 *****************************************************************************/
3289 /* IOMMUv2 specific functions */
3290 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3292 return atomic_notifier_chain_register(&ppr_notifier, nb);
3294 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3296 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3298 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3300 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3302 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3304 struct protection_domain *domain = to_pdomain(dom);
3305 unsigned long flags;
3307 spin_lock_irqsave(&domain->lock, flags);
3309 /* Update data structure */
3310 domain->mode = PAGE_MODE_NONE;
3311 domain->updated = true;
3313 /* Make changes visible to IOMMUs */
3314 update_domain(domain);
3316 /* Page-table is not visible to IOMMU anymore, so free it */
3317 free_pagetable(domain);
3319 spin_unlock_irqrestore(&domain->lock, flags);
3321 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3323 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3325 struct protection_domain *domain = to_pdomain(dom);
3326 unsigned long flags;
3329 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3332 /* Number of GCR3 table levels required */
3333 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3336 if (levels > amd_iommu_max_glx_val)
3339 spin_lock_irqsave(&domain->lock, flags);
3342 * Save us all sanity checks whether devices already in the
3343 * domain support IOMMUv2. Just force that the domain has no
3344 * devices attached when it is switched into IOMMUv2 mode.
3347 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3351 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3352 if (domain->gcr3_tbl == NULL)
3355 domain->glx = levels;
3356 domain->flags |= PD_IOMMUV2_MASK;
3357 domain->updated = true;
3359 update_domain(domain);
3364 spin_unlock_irqrestore(&domain->lock, flags);
3368 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3370 static int __flush_pasid(struct protection_domain *domain, int pasid,
3371 u64 address, bool size)
3373 struct iommu_dev_data *dev_data;
3374 struct iommu_cmd cmd;
3377 if (!(domain->flags & PD_IOMMUV2_MASK))
3380 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3383 * IOMMU TLB needs to be flushed before Device TLB to
3384 * prevent device TLB refill from IOMMU TLB
3386 for (i = 0; i < amd_iommus_present; ++i) {
3387 if (domain->dev_iommu[i] == 0)
3390 ret = iommu_queue_command(amd_iommus[i], &cmd);
3395 /* Wait until IOMMU TLB flushes are complete */
3396 domain_flush_complete(domain);
3398 /* Now flush device TLBs */
3399 list_for_each_entry(dev_data, &domain->dev_list, list) {
3400 struct amd_iommu *iommu;
3404 There might be non-IOMMUv2 capable devices in an IOMMUv2
3407 if (!dev_data->ats.enabled)
3410 qdep = dev_data->ats.qdep;
3411 iommu = amd_iommu_rlookup_table[dev_data->devid];
3413 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3414 qdep, address, size);
3416 ret = iommu_queue_command(iommu, &cmd);
3421 /* Wait until all device TLBs are flushed */
3422 domain_flush_complete(domain);
3431 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3434 INC_STATS_COUNTER(invalidate_iotlb);
3436 return __flush_pasid(domain, pasid, address, false);
3439 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3442 struct protection_domain *domain = to_pdomain(dom);
3443 unsigned long flags;
3446 spin_lock_irqsave(&domain->lock, flags);
3447 ret = __amd_iommu_flush_page(domain, pasid, address);
3448 spin_unlock_irqrestore(&domain->lock, flags);
3452 EXPORT_SYMBOL(amd_iommu_flush_page);
3454 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3456 INC_STATS_COUNTER(invalidate_iotlb_all);
3458 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3462 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3464 struct protection_domain *domain = to_pdomain(dom);
3465 unsigned long flags;
3468 spin_lock_irqsave(&domain->lock, flags);
3469 ret = __amd_iommu_flush_tlb(domain, pasid);
3470 spin_unlock_irqrestore(&domain->lock, flags);
3474 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3476 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3483 index = (pasid >> (9 * level)) & 0x1ff;
3489 if (!(*pte & GCR3_VALID)) {
3493 root = (void *)get_zeroed_page(GFP_ATOMIC);
3497 *pte = __pa(root) | GCR3_VALID;
3500 root = __va(*pte & PAGE_MASK);
3508 static int __set_gcr3(struct protection_domain *domain, int pasid,
3513 if (domain->mode != PAGE_MODE_NONE)
3516 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3520 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3522 return __amd_iommu_flush_tlb(domain, pasid);
3525 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3529 if (domain->mode != PAGE_MODE_NONE)
3532 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3538 return __amd_iommu_flush_tlb(domain, pasid);
3541 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3544 struct protection_domain *domain = to_pdomain(dom);
3545 unsigned long flags;
3548 spin_lock_irqsave(&domain->lock, flags);
3549 ret = __set_gcr3(domain, pasid, cr3);
3550 spin_unlock_irqrestore(&domain->lock, flags);
3554 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3556 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3558 struct protection_domain *domain = to_pdomain(dom);
3559 unsigned long flags;
3562 spin_lock_irqsave(&domain->lock, flags);
3563 ret = __clear_gcr3(domain, pasid);
3564 spin_unlock_irqrestore(&domain->lock, flags);
3568 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3570 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3571 int status, int tag)
3573 struct iommu_dev_data *dev_data;
3574 struct amd_iommu *iommu;
3575 struct iommu_cmd cmd;
3577 INC_STATS_COUNTER(complete_ppr);
3579 dev_data = get_dev_data(&pdev->dev);
3580 iommu = amd_iommu_rlookup_table[dev_data->devid];
3582 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3583 tag, dev_data->pri_tlp);
3585 return iommu_queue_command(iommu, &cmd);
3587 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3589 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3591 struct protection_domain *pdomain;
3593 pdomain = get_domain(&pdev->dev);
3594 if (IS_ERR(pdomain))
3597 /* Only return IOMMUv2 domains */
3598 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3601 return &pdomain->domain;
3603 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3605 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3607 struct iommu_dev_data *dev_data;
3609 if (!amd_iommu_v2_supported())
3612 dev_data = get_dev_data(&pdev->dev);
3613 dev_data->errata |= (1 << erratum);
3615 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3617 int amd_iommu_device_info(struct pci_dev *pdev,
3618 struct amd_iommu_device_info *info)
3623 if (pdev == NULL || info == NULL)
3626 if (!amd_iommu_v2_supported())
3629 memset(info, 0, sizeof(*info));
3631 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3633 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3635 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3637 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3639 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3643 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3644 max_pasids = min(max_pasids, (1 << 20));
3646 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3647 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3649 features = pci_pasid_features(pdev);
3650 if (features & PCI_PASID_CAP_EXEC)
3651 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3652 if (features & PCI_PASID_CAP_PRIV)
3653 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3658 EXPORT_SYMBOL(amd_iommu_device_info);
3660 #ifdef CONFIG_IRQ_REMAP
3662 /*****************************************************************************
3664 * Interrupt Remapping Implementation
3666 *****************************************************************************/
3684 u16 devid; /* Device ID for IRTE table */
3685 u16 index; /* Index into IRTE table*/
3688 struct amd_ir_data {
3689 struct irq_2_irte irq_2_irte;
3690 union irte irte_entry;
3692 struct msi_msg msi_entry;
3696 static struct irq_chip amd_ir_chip;
3698 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3699 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3700 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3701 #define DTE_IRQ_REMAP_ENABLE 1ULL
3703 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3707 dte = amd_iommu_dev_table[devid].data[2];
3708 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3709 dte |= virt_to_phys(table->table);
3710 dte |= DTE_IRQ_REMAP_INTCTL;
3711 dte |= DTE_IRQ_TABLE_LEN;
3712 dte |= DTE_IRQ_REMAP_ENABLE;
3714 amd_iommu_dev_table[devid].data[2] = dte;
3717 #define IRTE_ALLOCATED (~1U)
3719 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3721 struct irq_remap_table *table = NULL;
3722 struct amd_iommu *iommu;
3723 unsigned long flags;
3726 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3728 iommu = amd_iommu_rlookup_table[devid];
3732 table = irq_lookup_table[devid];
3736 alias = amd_iommu_alias_table[devid];
3737 table = irq_lookup_table[alias];
3739 irq_lookup_table[devid] = table;
3740 set_dte_irq_entry(devid, table);
3741 iommu_flush_dte(iommu, devid);
3745 /* Nothing there yet, allocate new irq remapping table */
3746 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3750 /* Initialize table spin-lock */
3751 spin_lock_init(&table->lock);
3754 /* Keep the first 32 indexes free for IOAPIC interrupts */
3755 table->min_index = 32;
3757 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3758 if (!table->table) {
3764 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3769 for (i = 0; i < 32; ++i)
3770 table->table[i] = IRTE_ALLOCATED;
3773 irq_lookup_table[devid] = table;
3774 set_dte_irq_entry(devid, table);
3775 iommu_flush_dte(iommu, devid);
3776 if (devid != alias) {
3777 irq_lookup_table[alias] = table;
3778 set_dte_irq_entry(alias, table);
3779 iommu_flush_dte(iommu, alias);
3783 iommu_completion_wait(iommu);
3786 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3791 static int alloc_irq_index(u16 devid, int count)
3793 struct irq_remap_table *table;
3794 unsigned long flags;
3797 table = get_irq_table(devid, false);
3801 spin_lock_irqsave(&table->lock, flags);
3803 /* Scan table for free entries */
3804 for (c = 0, index = table->min_index;
3805 index < MAX_IRQS_PER_TABLE;
3807 if (table->table[index] == 0)
3814 table->table[index - c + 1] = IRTE_ALLOCATED;
3824 spin_unlock_irqrestore(&table->lock, flags);
3829 static int modify_irte(u16 devid, int index, union irte irte)
3831 struct irq_remap_table *table;
3832 struct amd_iommu *iommu;
3833 unsigned long flags;
3835 iommu = amd_iommu_rlookup_table[devid];
3839 table = get_irq_table(devid, false);
3843 spin_lock_irqsave(&table->lock, flags);
3844 table->table[index] = irte.val;
3845 spin_unlock_irqrestore(&table->lock, flags);
3847 iommu_flush_irt(iommu, devid);
3848 iommu_completion_wait(iommu);
3853 static void free_irte(u16 devid, int index)
3855 struct irq_remap_table *table;
3856 struct amd_iommu *iommu;
3857 unsigned long flags;
3859 iommu = amd_iommu_rlookup_table[devid];
3863 table = get_irq_table(devid, false);
3867 spin_lock_irqsave(&table->lock, flags);
3868 table->table[index] = 0;
3869 spin_unlock_irqrestore(&table->lock, flags);
3871 iommu_flush_irt(iommu, devid);
3872 iommu_completion_wait(iommu);
3875 static int get_devid(struct irq_alloc_info *info)
3879 switch (info->type) {
3880 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3881 devid = get_ioapic_devid(info->ioapic_id);
3883 case X86_IRQ_ALLOC_TYPE_HPET:
3884 devid = get_hpet_devid(info->hpet_id);
3886 case X86_IRQ_ALLOC_TYPE_MSI:
3887 case X86_IRQ_ALLOC_TYPE_MSIX:
3888 devid = get_device_id(&info->msi_dev->dev);
3898 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3900 struct amd_iommu *iommu;
3906 devid = get_devid(info);
3908 iommu = amd_iommu_rlookup_table[devid];
3910 return iommu->ir_domain;
3916 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3918 struct amd_iommu *iommu;
3924 switch (info->type) {
3925 case X86_IRQ_ALLOC_TYPE_MSI:
3926 case X86_IRQ_ALLOC_TYPE_MSIX:
3927 devid = get_device_id(&info->msi_dev->dev);
3928 iommu = amd_iommu_rlookup_table[devid];
3930 return iommu->msi_domain;
3939 struct irq_remap_ops amd_iommu_irq_ops = {
3940 .prepare = amd_iommu_prepare,
3941 .enable = amd_iommu_enable,
3942 .disable = amd_iommu_disable,
3943 .reenable = amd_iommu_reenable,
3944 .enable_faulting = amd_iommu_enable_faulting,
3945 .get_ir_irq_domain = get_ir_irq_domain,
3946 .get_irq_domain = get_irq_domain,
3949 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3950 struct irq_cfg *irq_cfg,
3951 struct irq_alloc_info *info,
3952 int devid, int index, int sub_handle)
3954 struct irq_2_irte *irte_info = &data->irq_2_irte;
3955 struct msi_msg *msg = &data->msi_entry;
3956 union irte *irte = &data->irte_entry;
3957 struct IO_APIC_route_entry *entry;
3959 data->irq_2_irte.devid = devid;
3960 data->irq_2_irte.index = index + sub_handle;
3962 /* Setup IRTE for IOMMU */
3964 irte->fields.vector = irq_cfg->vector;
3965 irte->fields.int_type = apic->irq_delivery_mode;
3966 irte->fields.destination = irq_cfg->dest_apicid;
3967 irte->fields.dm = apic->irq_dest_mode;
3968 irte->fields.valid = 1;
3970 switch (info->type) {
3971 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3972 /* Setup IOAPIC entry */
3973 entry = info->ioapic_entry;
3974 info->ioapic_entry = NULL;
3975 memset(entry, 0, sizeof(*entry));
3976 entry->vector = index;
3978 entry->trigger = info->ioapic_trigger;
3979 entry->polarity = info->ioapic_polarity;
3980 /* Mask level triggered irqs. */
3981 if (info->ioapic_trigger)
3985 case X86_IRQ_ALLOC_TYPE_HPET:
3986 case X86_IRQ_ALLOC_TYPE_MSI:
3987 case X86_IRQ_ALLOC_TYPE_MSIX:
3988 msg->address_hi = MSI_ADDR_BASE_HI;
3989 msg->address_lo = MSI_ADDR_BASE_LO;
3990 msg->data = irte_info->index;
3999 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4000 unsigned int nr_irqs, void *arg)
4002 struct irq_alloc_info *info = arg;
4003 struct irq_data *irq_data;
4004 struct amd_ir_data *data;
4005 struct irq_cfg *cfg;
4011 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4012 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4016 * With IRQ remapping enabled, don't need contiguous CPU vectors
4017 * to support multiple MSI interrupts.
4019 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4020 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4022 devid = get_devid(info);
4026 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4030 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4031 if (get_irq_table(devid, true))
4032 index = info->ioapic_pin;
4036 index = alloc_irq_index(devid, nr_irqs);
4039 pr_warn("Failed to allocate IRTE\n");
4040 goto out_free_parent;
4043 for (i = 0; i < nr_irqs; i++) {
4044 irq_data = irq_domain_get_irq_data(domain, virq + i);
4045 cfg = irqd_cfg(irq_data);
4046 if (!irq_data || !cfg) {
4052 data = kzalloc(sizeof(*data), GFP_KERNEL);
4056 irq_data->hwirq = (devid << 16) + i;
4057 irq_data->chip_data = data;
4058 irq_data->chip = &amd_ir_chip;
4059 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4060 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4066 for (i--; i >= 0; i--) {
4067 irq_data = irq_domain_get_irq_data(domain, virq + i);
4069 kfree(irq_data->chip_data);
4071 for (i = 0; i < nr_irqs; i++)
4072 free_irte(devid, index + i);
4074 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4078 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4079 unsigned int nr_irqs)
4081 struct irq_2_irte *irte_info;
4082 struct irq_data *irq_data;
4083 struct amd_ir_data *data;
4086 for (i = 0; i < nr_irqs; i++) {
4087 irq_data = irq_domain_get_irq_data(domain, virq + i);
4088 if (irq_data && irq_data->chip_data) {
4089 data = irq_data->chip_data;
4090 irte_info = &data->irq_2_irte;
4091 free_irte(irte_info->devid, irte_info->index);
4095 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4098 static void irq_remapping_activate(struct irq_domain *domain,
4099 struct irq_data *irq_data)
4101 struct amd_ir_data *data = irq_data->chip_data;
4102 struct irq_2_irte *irte_info = &data->irq_2_irte;
4104 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4107 static void irq_remapping_deactivate(struct irq_domain *domain,
4108 struct irq_data *irq_data)
4110 struct amd_ir_data *data = irq_data->chip_data;
4111 struct irq_2_irte *irte_info = &data->irq_2_irte;
4115 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4118 static struct irq_domain_ops amd_ir_domain_ops = {
4119 .alloc = irq_remapping_alloc,
4120 .free = irq_remapping_free,
4121 .activate = irq_remapping_activate,
4122 .deactivate = irq_remapping_deactivate,
4125 static int amd_ir_set_affinity(struct irq_data *data,
4126 const struct cpumask *mask, bool force)
4128 struct amd_ir_data *ir_data = data->chip_data;
4129 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4130 struct irq_cfg *cfg = irqd_cfg(data);
4131 struct irq_data *parent = data->parent_data;
4134 ret = parent->chip->irq_set_affinity(parent, mask, force);
4135 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4139 * Atomically updates the IRTE with the new destination, vector
4140 * and flushes the interrupt entry cache.
4142 ir_data->irte_entry.fields.vector = cfg->vector;
4143 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4144 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4147 * After this point, all the interrupts will start arriving
4148 * at the new destination. So, time to cleanup the previous
4149 * vector allocation.
4151 send_cleanup_vector(cfg);
4153 return IRQ_SET_MASK_OK_DONE;
4156 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4158 struct amd_ir_data *ir_data = irq_data->chip_data;
4160 *msg = ir_data->msi_entry;
4163 static struct irq_chip amd_ir_chip = {
4164 .irq_ack = ir_ack_apic_edge,
4165 .irq_set_affinity = amd_ir_set_affinity,
4166 .irq_compose_msi_msg = ir_compose_msi_msg,
4169 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4171 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4172 if (!iommu->ir_domain)
4175 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4176 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);