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 <asm/irq_remapping.h>
39 #include <asm/io_apic.h>
41 #include <asm/hw_irq.h>
42 #include <asm/msidef.h>
43 #include <asm/proto.h>
44 #include <asm/iommu.h>
48 #include "amd_iommu_proto.h"
49 #include "amd_iommu_types.h"
50 #include "irq_remapping.h"
52 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
54 #define LOOP_TIMEOUT 100000
57 * This bitmap is used to advertise the page sizes our hardware support
58 * to the IOMMU core, which will then use this information to split
59 * physically contiguous memory regions it is mapping into page sizes
62 * 512GB Pages are not supported due to a hardware bug
64 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
66 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
68 /* A list of preallocated protection domains */
69 static LIST_HEAD(iommu_pd_list);
70 static DEFINE_SPINLOCK(iommu_pd_list_lock);
72 /* List of all available dev_data structures */
73 static LIST_HEAD(dev_data_list);
74 static DEFINE_SPINLOCK(dev_data_list_lock);
76 LIST_HEAD(ioapic_map);
80 * Domain for untranslated devices - only allocated
81 * if iommu=pt passed on kernel cmd line.
83 static struct protection_domain *pt_domain;
85 static const struct iommu_ops amd_iommu_ops;
87 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
88 int amd_iommu_max_glx_val = -1;
90 static struct dma_map_ops amd_iommu_dma_ops;
93 * This struct contains device specific data for the IOMMU
95 struct iommu_dev_data {
96 struct list_head list; /* For domain->dev_list */
97 struct list_head dev_data_list; /* For global dev_data_list */
98 struct list_head alias_list; /* Link alias-groups together */
99 struct iommu_dev_data *alias_data;/* The alias dev_data */
100 struct protection_domain *domain; /* Domain the device is bound to */
101 u16 devid; /* PCI Device ID */
102 bool iommu_v2; /* Device can make use of IOMMUv2 */
103 bool passthrough; /* Default for device is pt_domain */
107 } ats; /* ATS state */
108 bool pri_tlp; /* PASID TLB required for
110 u32 errata; /* Bitmap for errata to apply */
114 * general struct to manage commands send to an IOMMU
120 struct kmem_cache *amd_iommu_irq_cache;
122 static void update_domain(struct protection_domain *domain);
123 static int __init alloc_passthrough_domain(void);
125 /****************************************************************************
129 ****************************************************************************/
131 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
133 return container_of(dom, struct protection_domain, domain);
136 static struct iommu_dev_data *alloc_dev_data(u16 devid)
138 struct iommu_dev_data *dev_data;
141 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
145 INIT_LIST_HEAD(&dev_data->alias_list);
147 dev_data->devid = devid;
149 spin_lock_irqsave(&dev_data_list_lock, flags);
150 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
151 spin_unlock_irqrestore(&dev_data_list_lock, flags);
156 static void free_dev_data(struct iommu_dev_data *dev_data)
160 spin_lock_irqsave(&dev_data_list_lock, flags);
161 list_del(&dev_data->dev_data_list);
162 spin_unlock_irqrestore(&dev_data_list_lock, flags);
167 static struct iommu_dev_data *search_dev_data(u16 devid)
169 struct iommu_dev_data *dev_data;
172 spin_lock_irqsave(&dev_data_list_lock, flags);
173 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
174 if (dev_data->devid == devid)
181 spin_unlock_irqrestore(&dev_data_list_lock, flags);
186 static struct iommu_dev_data *find_dev_data(u16 devid)
188 struct iommu_dev_data *dev_data;
190 dev_data = search_dev_data(devid);
192 if (dev_data == NULL)
193 dev_data = alloc_dev_data(devid);
198 static inline u16 get_device_id(struct device *dev)
200 struct pci_dev *pdev = to_pci_dev(dev);
202 return PCI_DEVID(pdev->bus->number, pdev->devfn);
205 static struct iommu_dev_data *get_dev_data(struct device *dev)
207 return dev->archdata.iommu;
210 static bool pci_iommuv2_capable(struct pci_dev *pdev)
212 static const int caps[] = {
215 PCI_EXT_CAP_ID_PASID,
219 for (i = 0; i < 3; ++i) {
220 pos = pci_find_ext_capability(pdev, caps[i]);
228 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
230 struct iommu_dev_data *dev_data;
232 dev_data = get_dev_data(&pdev->dev);
234 return dev_data->errata & (1 << erratum) ? true : false;
238 * In this function the list of preallocated protection domains is traversed to
239 * find the domain for a specific device
241 static struct dma_ops_domain *find_protection_domain(u16 devid)
243 struct dma_ops_domain *entry, *ret = NULL;
245 u16 alias = amd_iommu_alias_table[devid];
247 if (list_empty(&iommu_pd_list))
250 spin_lock_irqsave(&iommu_pd_list_lock, flags);
252 list_for_each_entry(entry, &iommu_pd_list, list) {
253 if (entry->target_dev == devid ||
254 entry->target_dev == alias) {
260 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
266 * This function checks if the driver got a valid device from the caller to
267 * avoid dereferencing invalid pointers.
269 static bool check_device(struct device *dev)
273 if (!dev || !dev->dma_mask)
277 if (!dev_is_pci(dev))
280 devid = get_device_id(dev);
282 /* Out of our scope? */
283 if (devid > amd_iommu_last_bdf)
286 if (amd_iommu_rlookup_table[devid] == NULL)
292 static void init_iommu_group(struct device *dev)
294 struct iommu_group *group;
296 group = iommu_group_get_for_dev(dev);
298 iommu_group_put(group);
301 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
303 *(u16 *)data = alias;
307 static u16 get_alias(struct device *dev)
309 struct pci_dev *pdev = to_pci_dev(dev);
310 u16 devid, ivrs_alias, pci_alias;
312 devid = get_device_id(dev);
313 ivrs_alias = amd_iommu_alias_table[devid];
314 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
316 if (ivrs_alias == pci_alias)
322 * The IVRS is fairly reliable in telling us about aliases, but it
323 * can't know about every screwy device. If we don't have an IVRS
324 * reported alias, use the PCI reported alias. In that case we may
325 * still need to initialize the rlookup and dev_table entries if the
326 * alias is to a non-existent device.
328 if (ivrs_alias == devid) {
329 if (!amd_iommu_rlookup_table[pci_alias]) {
330 amd_iommu_rlookup_table[pci_alias] =
331 amd_iommu_rlookup_table[devid];
332 memcpy(amd_iommu_dev_table[pci_alias].data,
333 amd_iommu_dev_table[devid].data,
334 sizeof(amd_iommu_dev_table[pci_alias].data));
340 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
341 "for device %s[%04x:%04x], kernel reported alias "
342 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
343 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
344 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
345 PCI_FUNC(pci_alias));
348 * If we don't have a PCI DMA alias and the IVRS alias is on the same
349 * bus, then the IVRS table may know about a quirk that we don't.
351 if (pci_alias == devid &&
352 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
353 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
354 pdev->dma_alias_devfn = ivrs_alias & 0xff;
355 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
356 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
363 static int iommu_init_device(struct device *dev)
365 struct pci_dev *pdev = to_pci_dev(dev);
366 struct iommu_dev_data *dev_data;
369 if (dev->archdata.iommu)
372 dev_data = find_dev_data(get_device_id(dev));
376 alias = get_alias(dev);
378 if (alias != dev_data->devid) {
379 struct iommu_dev_data *alias_data;
381 alias_data = find_dev_data(alias);
382 if (alias_data == NULL) {
383 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
385 free_dev_data(dev_data);
388 dev_data->alias_data = alias_data;
390 /* Add device to the alias_list */
391 list_add(&dev_data->alias_list, &alias_data->alias_list);
394 if (pci_iommuv2_capable(pdev)) {
395 struct amd_iommu *iommu;
397 iommu = amd_iommu_rlookup_table[dev_data->devid];
398 dev_data->iommu_v2 = iommu->is_iommu_v2;
401 dev->archdata.iommu = dev_data;
403 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
409 static void iommu_ignore_device(struct device *dev)
413 devid = get_device_id(dev);
414 alias = amd_iommu_alias_table[devid];
416 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
417 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
419 amd_iommu_rlookup_table[devid] = NULL;
420 amd_iommu_rlookup_table[alias] = NULL;
423 static void iommu_uninit_device(struct device *dev)
425 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
430 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
433 iommu_group_remove_device(dev);
435 /* Unlink from alias, it may change if another device is re-plugged */
436 dev_data->alias_data = NULL;
439 * We keep dev_data around for unplugged devices and reuse it when the
440 * device is re-plugged - not doing so would introduce a ton of races.
444 void __init amd_iommu_uninit_devices(void)
446 struct iommu_dev_data *dev_data, *n;
447 struct pci_dev *pdev = NULL;
449 for_each_pci_dev(pdev) {
451 if (!check_device(&pdev->dev))
454 iommu_uninit_device(&pdev->dev);
457 /* Free all of our dev_data structures */
458 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
459 free_dev_data(dev_data);
462 int __init amd_iommu_init_devices(void)
464 struct pci_dev *pdev = NULL;
467 for_each_pci_dev(pdev) {
469 if (!check_device(&pdev->dev))
472 ret = iommu_init_device(&pdev->dev);
473 if (ret == -ENOTSUPP)
474 iommu_ignore_device(&pdev->dev);
480 * Initialize IOMMU groups only after iommu_init_device() has
481 * had a chance to populate any IVRS defined aliases.
483 for_each_pci_dev(pdev) {
484 if (check_device(&pdev->dev))
485 init_iommu_group(&pdev->dev);
492 amd_iommu_uninit_devices();
496 #ifdef CONFIG_AMD_IOMMU_STATS
499 * Initialization code for statistics collection
502 DECLARE_STATS_COUNTER(compl_wait);
503 DECLARE_STATS_COUNTER(cnt_map_single);
504 DECLARE_STATS_COUNTER(cnt_unmap_single);
505 DECLARE_STATS_COUNTER(cnt_map_sg);
506 DECLARE_STATS_COUNTER(cnt_unmap_sg);
507 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
508 DECLARE_STATS_COUNTER(cnt_free_coherent);
509 DECLARE_STATS_COUNTER(cross_page);
510 DECLARE_STATS_COUNTER(domain_flush_single);
511 DECLARE_STATS_COUNTER(domain_flush_all);
512 DECLARE_STATS_COUNTER(alloced_io_mem);
513 DECLARE_STATS_COUNTER(total_map_requests);
514 DECLARE_STATS_COUNTER(complete_ppr);
515 DECLARE_STATS_COUNTER(invalidate_iotlb);
516 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
517 DECLARE_STATS_COUNTER(pri_requests);
519 static struct dentry *stats_dir;
520 static struct dentry *de_fflush;
522 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
524 if (stats_dir == NULL)
527 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
531 static void amd_iommu_stats_init(void)
533 stats_dir = debugfs_create_dir("amd-iommu", NULL);
534 if (stats_dir == NULL)
537 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
538 &amd_iommu_unmap_flush);
540 amd_iommu_stats_add(&compl_wait);
541 amd_iommu_stats_add(&cnt_map_single);
542 amd_iommu_stats_add(&cnt_unmap_single);
543 amd_iommu_stats_add(&cnt_map_sg);
544 amd_iommu_stats_add(&cnt_unmap_sg);
545 amd_iommu_stats_add(&cnt_alloc_coherent);
546 amd_iommu_stats_add(&cnt_free_coherent);
547 amd_iommu_stats_add(&cross_page);
548 amd_iommu_stats_add(&domain_flush_single);
549 amd_iommu_stats_add(&domain_flush_all);
550 amd_iommu_stats_add(&alloced_io_mem);
551 amd_iommu_stats_add(&total_map_requests);
552 amd_iommu_stats_add(&complete_ppr);
553 amd_iommu_stats_add(&invalidate_iotlb);
554 amd_iommu_stats_add(&invalidate_iotlb_all);
555 amd_iommu_stats_add(&pri_requests);
560 /****************************************************************************
562 * Interrupt handling functions
564 ****************************************************************************/
566 static void dump_dte_entry(u16 devid)
570 for (i = 0; i < 4; ++i)
571 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
572 amd_iommu_dev_table[devid].data[i]);
575 static void dump_command(unsigned long phys_addr)
577 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
580 for (i = 0; i < 4; ++i)
581 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
584 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
586 int type, devid, domid, flags;
587 volatile u32 *event = __evt;
592 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
593 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
594 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
595 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
596 address = (u64)(((u64)event[3]) << 32) | event[2];
599 /* Did we hit the erratum? */
600 if (++count == LOOP_TIMEOUT) {
601 pr_err("AMD-Vi: No event written to event log\n");
608 printk(KERN_ERR "AMD-Vi: Event logged [");
611 case EVENT_TYPE_ILL_DEV:
612 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
613 "address=0x%016llx flags=0x%04x]\n",
614 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 dump_dte_entry(devid);
618 case EVENT_TYPE_IO_FAULT:
619 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
620 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
621 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
622 domid, address, flags);
624 case EVENT_TYPE_DEV_TAB_ERR:
625 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
626 "address=0x%016llx flags=0x%04x]\n",
627 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
630 case EVENT_TYPE_PAGE_TAB_ERR:
631 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
632 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
633 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
634 domid, address, flags);
636 case EVENT_TYPE_ILL_CMD:
637 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
638 dump_command(address);
640 case EVENT_TYPE_CMD_HARD_ERR:
641 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
642 "flags=0x%04x]\n", address, flags);
644 case EVENT_TYPE_IOTLB_INV_TO:
645 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
646 "address=0x%016llx]\n",
647 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
650 case EVENT_TYPE_INV_DEV_REQ:
651 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
652 "address=0x%016llx flags=0x%04x]\n",
653 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
657 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
660 memset(__evt, 0, 4 * sizeof(u32));
663 static void iommu_poll_events(struct amd_iommu *iommu)
667 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
668 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
670 while (head != tail) {
671 iommu_print_event(iommu, iommu->evt_buf + head);
672 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
675 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
678 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
680 struct amd_iommu_fault fault;
682 INC_STATS_COUNTER(pri_requests);
684 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
685 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
689 fault.address = raw[1];
690 fault.pasid = PPR_PASID(raw[0]);
691 fault.device_id = PPR_DEVID(raw[0]);
692 fault.tag = PPR_TAG(raw[0]);
693 fault.flags = PPR_FLAGS(raw[0]);
695 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
698 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
702 if (iommu->ppr_log == NULL)
705 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
706 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
708 while (head != tail) {
713 raw = (u64 *)(iommu->ppr_log + head);
716 * Hardware bug: Interrupt may arrive before the entry is
717 * written to memory. If this happens we need to wait for the
720 for (i = 0; i < LOOP_TIMEOUT; ++i) {
721 if (PPR_REQ_TYPE(raw[0]) != 0)
726 /* Avoid memcpy function-call overhead */
731 * To detect the hardware bug we need to clear the entry
734 raw[0] = raw[1] = 0UL;
736 /* Update head pointer of hardware ring-buffer */
737 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
738 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
740 /* Handle PPR entry */
741 iommu_handle_ppr_entry(iommu, entry);
743 /* Refresh ring-buffer information */
744 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
745 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
749 irqreturn_t amd_iommu_int_thread(int irq, void *data)
751 struct amd_iommu *iommu = (struct amd_iommu *) data;
752 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
754 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
755 /* Enable EVT and PPR interrupts again */
756 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
757 iommu->mmio_base + MMIO_STATUS_OFFSET);
759 if (status & MMIO_STATUS_EVT_INT_MASK) {
760 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
761 iommu_poll_events(iommu);
764 if (status & MMIO_STATUS_PPR_INT_MASK) {
765 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
766 iommu_poll_ppr_log(iommu);
770 * Hardware bug: ERBT1312
771 * When re-enabling interrupt (by writing 1
772 * to clear the bit), the hardware might also try to set
773 * the interrupt bit in the event status register.
774 * In this scenario, the bit will be set, and disable
775 * subsequent interrupts.
777 * Workaround: The IOMMU driver should read back the
778 * status register and check if the interrupt bits are cleared.
779 * If not, driver will need to go through the interrupt handler
780 * again and re-clear the bits
782 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
787 irqreturn_t amd_iommu_int_handler(int irq, void *data)
789 return IRQ_WAKE_THREAD;
792 /****************************************************************************
794 * IOMMU command queuing functions
796 ****************************************************************************/
798 static int wait_on_sem(volatile u64 *sem)
802 while (*sem == 0 && i < LOOP_TIMEOUT) {
807 if (i == LOOP_TIMEOUT) {
808 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
815 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
816 struct iommu_cmd *cmd,
821 target = iommu->cmd_buf + tail;
822 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
824 /* Copy command to buffer */
825 memcpy(target, cmd, sizeof(*cmd));
827 /* Tell the IOMMU about it */
828 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
831 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
833 WARN_ON(address & 0x7ULL);
835 memset(cmd, 0, sizeof(*cmd));
836 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
837 cmd->data[1] = upper_32_bits(__pa(address));
839 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
842 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
844 memset(cmd, 0, sizeof(*cmd));
845 cmd->data[0] = devid;
846 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
849 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
850 size_t size, u16 domid, int pde)
855 pages = iommu_num_pages(address, size, PAGE_SIZE);
860 * If we have to flush more than one page, flush all
861 * TLB entries for this domain
863 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
867 address &= PAGE_MASK;
869 memset(cmd, 0, sizeof(*cmd));
870 cmd->data[1] |= domid;
871 cmd->data[2] = lower_32_bits(address);
872 cmd->data[3] = upper_32_bits(address);
873 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
874 if (s) /* size bit - we flush more than one 4kb page */
875 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
876 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
877 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
880 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
881 u64 address, size_t size)
886 pages = iommu_num_pages(address, size, PAGE_SIZE);
891 * If we have to flush more than one page, flush all
892 * TLB entries for this domain
894 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
898 address &= PAGE_MASK;
900 memset(cmd, 0, sizeof(*cmd));
901 cmd->data[0] = devid;
902 cmd->data[0] |= (qdep & 0xff) << 24;
903 cmd->data[1] = devid;
904 cmd->data[2] = lower_32_bits(address);
905 cmd->data[3] = upper_32_bits(address);
906 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
911 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
912 u64 address, bool size)
914 memset(cmd, 0, sizeof(*cmd));
916 address &= ~(0xfffULL);
918 cmd->data[0] = pasid;
919 cmd->data[1] = domid;
920 cmd->data[2] = lower_32_bits(address);
921 cmd->data[3] = upper_32_bits(address);
922 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
923 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
925 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
926 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
929 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
930 int qdep, u64 address, bool size)
932 memset(cmd, 0, sizeof(*cmd));
934 address &= ~(0xfffULL);
936 cmd->data[0] = devid;
937 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
938 cmd->data[0] |= (qdep & 0xff) << 24;
939 cmd->data[1] = devid;
940 cmd->data[1] |= (pasid & 0xff) << 16;
941 cmd->data[2] = lower_32_bits(address);
942 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
943 cmd->data[3] = upper_32_bits(address);
945 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
946 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
949 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
950 int status, int tag, bool gn)
952 memset(cmd, 0, sizeof(*cmd));
954 cmd->data[0] = devid;
956 cmd->data[1] = pasid;
957 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
959 cmd->data[3] = tag & 0x1ff;
960 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
962 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
965 static void build_inv_all(struct iommu_cmd *cmd)
967 memset(cmd, 0, sizeof(*cmd));
968 CMD_SET_TYPE(cmd, CMD_INV_ALL);
971 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
973 memset(cmd, 0, sizeof(*cmd));
974 cmd->data[0] = devid;
975 CMD_SET_TYPE(cmd, CMD_INV_IRT);
979 * Writes the command to the IOMMUs command buffer and informs the
980 * hardware about the new command.
982 static int iommu_queue_command_sync(struct amd_iommu *iommu,
983 struct iommu_cmd *cmd,
986 u32 left, tail, head, next_tail;
989 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
992 spin_lock_irqsave(&iommu->lock, flags);
994 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
995 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
996 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
997 left = (head - next_tail) % iommu->cmd_buf_size;
1000 struct iommu_cmd sync_cmd;
1001 volatile u64 sem = 0;
1004 build_completion_wait(&sync_cmd, (u64)&sem);
1005 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1007 spin_unlock_irqrestore(&iommu->lock, flags);
1009 if ((ret = wait_on_sem(&sem)) != 0)
1015 copy_cmd_to_buffer(iommu, cmd, tail);
1017 /* We need to sync now to make sure all commands are processed */
1018 iommu->need_sync = sync;
1020 spin_unlock_irqrestore(&iommu->lock, flags);
1025 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1027 return iommu_queue_command_sync(iommu, cmd, true);
1031 * This function queues a completion wait command into the command
1032 * buffer of an IOMMU
1034 static int iommu_completion_wait(struct amd_iommu *iommu)
1036 struct iommu_cmd cmd;
1037 volatile u64 sem = 0;
1040 if (!iommu->need_sync)
1043 build_completion_wait(&cmd, (u64)&sem);
1045 ret = iommu_queue_command_sync(iommu, &cmd, false);
1049 return wait_on_sem(&sem);
1052 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1054 struct iommu_cmd cmd;
1056 build_inv_dte(&cmd, devid);
1058 return iommu_queue_command(iommu, &cmd);
1061 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1065 for (devid = 0; devid <= 0xffff; ++devid)
1066 iommu_flush_dte(iommu, devid);
1068 iommu_completion_wait(iommu);
1072 * This function uses heavy locking and may disable irqs for some time. But
1073 * this is no issue because it is only called during resume.
1075 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1079 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1080 struct iommu_cmd cmd;
1081 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1083 iommu_queue_command(iommu, &cmd);
1086 iommu_completion_wait(iommu);
1089 static void iommu_flush_all(struct amd_iommu *iommu)
1091 struct iommu_cmd cmd;
1093 build_inv_all(&cmd);
1095 iommu_queue_command(iommu, &cmd);
1096 iommu_completion_wait(iommu);
1099 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1101 struct iommu_cmd cmd;
1103 build_inv_irt(&cmd, devid);
1105 iommu_queue_command(iommu, &cmd);
1108 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1112 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1113 iommu_flush_irt(iommu, devid);
1115 iommu_completion_wait(iommu);
1118 void iommu_flush_all_caches(struct amd_iommu *iommu)
1120 if (iommu_feature(iommu, FEATURE_IA)) {
1121 iommu_flush_all(iommu);
1123 iommu_flush_dte_all(iommu);
1124 iommu_flush_irt_all(iommu);
1125 iommu_flush_tlb_all(iommu);
1130 * Command send function for flushing on-device TLB
1132 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1133 u64 address, size_t size)
1135 struct amd_iommu *iommu;
1136 struct iommu_cmd cmd;
1139 qdep = dev_data->ats.qdep;
1140 iommu = amd_iommu_rlookup_table[dev_data->devid];
1142 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1144 return iommu_queue_command(iommu, &cmd);
1148 * Command send function for invalidating a device table entry
1150 static int device_flush_dte(struct iommu_dev_data *dev_data)
1152 struct amd_iommu *iommu;
1155 iommu = amd_iommu_rlookup_table[dev_data->devid];
1157 ret = iommu_flush_dte(iommu, dev_data->devid);
1161 if (dev_data->ats.enabled)
1162 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1168 * TLB invalidation function which is called from the mapping functions.
1169 * It invalidates a single PTE if the range to flush is within a single
1170 * page. Otherwise it flushes the whole TLB of the IOMMU.
1172 static void __domain_flush_pages(struct protection_domain *domain,
1173 u64 address, size_t size, int pde)
1175 struct iommu_dev_data *dev_data;
1176 struct iommu_cmd cmd;
1179 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1181 for (i = 0; i < amd_iommus_present; ++i) {
1182 if (!domain->dev_iommu[i])
1186 * Devices of this domain are behind this IOMMU
1187 * We need a TLB flush
1189 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1192 list_for_each_entry(dev_data, &domain->dev_list, list) {
1194 if (!dev_data->ats.enabled)
1197 ret |= device_flush_iotlb(dev_data, address, size);
1203 static void domain_flush_pages(struct protection_domain *domain,
1204 u64 address, size_t size)
1206 __domain_flush_pages(domain, address, size, 0);
1209 /* Flush the whole IO/TLB for a given protection domain */
1210 static void domain_flush_tlb(struct protection_domain *domain)
1212 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1215 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1216 static void domain_flush_tlb_pde(struct protection_domain *domain)
1218 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1221 static void domain_flush_complete(struct protection_domain *domain)
1225 for (i = 0; i < amd_iommus_present; ++i) {
1226 if (!domain->dev_iommu[i])
1230 * Devices of this domain are behind this IOMMU
1231 * We need to wait for completion of all commands.
1233 iommu_completion_wait(amd_iommus[i]);
1239 * This function flushes the DTEs for all devices in domain
1241 static void domain_flush_devices(struct protection_domain *domain)
1243 struct iommu_dev_data *dev_data;
1245 list_for_each_entry(dev_data, &domain->dev_list, list)
1246 device_flush_dte(dev_data);
1249 /****************************************************************************
1251 * The functions below are used the create the page table mappings for
1252 * unity mapped regions.
1254 ****************************************************************************/
1257 * This function is used to add another level to an IO page table. Adding
1258 * another level increases the size of the address space by 9 bits to a size up
1261 static bool increase_address_space(struct protection_domain *domain,
1266 if (domain->mode == PAGE_MODE_6_LEVEL)
1267 /* address space already 64 bit large */
1270 pte = (void *)get_zeroed_page(gfp);
1274 *pte = PM_LEVEL_PDE(domain->mode,
1275 virt_to_phys(domain->pt_root));
1276 domain->pt_root = pte;
1278 domain->updated = true;
1283 static u64 *alloc_pte(struct protection_domain *domain,
1284 unsigned long address,
1285 unsigned long page_size,
1292 BUG_ON(!is_power_of_2(page_size));
1294 while (address > PM_LEVEL_SIZE(domain->mode))
1295 increase_address_space(domain, gfp);
1297 level = domain->mode - 1;
1298 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1299 address = PAGE_SIZE_ALIGN(address, page_size);
1300 end_lvl = PAGE_SIZE_LEVEL(page_size);
1302 while (level > end_lvl) {
1303 if (!IOMMU_PTE_PRESENT(*pte)) {
1304 page = (u64 *)get_zeroed_page(gfp);
1307 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1310 /* No level skipping support yet */
1311 if (PM_PTE_LEVEL(*pte) != level)
1316 pte = IOMMU_PTE_PAGE(*pte);
1318 if (pte_page && level == end_lvl)
1321 pte = &pte[PM_LEVEL_INDEX(level, address)];
1328 * This function checks if there is a PTE for a given dma address. If
1329 * there is one, it returns the pointer to it.
1331 static u64 *fetch_pte(struct protection_domain *domain,
1332 unsigned long address,
1333 unsigned long *page_size)
1338 if (address > PM_LEVEL_SIZE(domain->mode))
1341 level = domain->mode - 1;
1342 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1343 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1348 if (!IOMMU_PTE_PRESENT(*pte))
1352 if (PM_PTE_LEVEL(*pte) == 7 ||
1353 PM_PTE_LEVEL(*pte) == 0)
1356 /* No level skipping support yet */
1357 if (PM_PTE_LEVEL(*pte) != level)
1362 /* Walk to the next level */
1363 pte = IOMMU_PTE_PAGE(*pte);
1364 pte = &pte[PM_LEVEL_INDEX(level, address)];
1365 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1368 if (PM_PTE_LEVEL(*pte) == 0x07) {
1369 unsigned long pte_mask;
1372 * If we have a series of large PTEs, make
1373 * sure to return a pointer to the first one.
1375 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1376 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1377 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1384 * Generic mapping functions. It maps a physical address into a DMA
1385 * address space. It allocates the page table pages if necessary.
1386 * In the future it can be extended to a generic mapping function
1387 * supporting all features of AMD IOMMU page tables like level skipping
1388 * and full 64 bit address spaces.
1390 static int iommu_map_page(struct protection_domain *dom,
1391 unsigned long bus_addr,
1392 unsigned long phys_addr,
1394 unsigned long page_size)
1399 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1400 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1402 if (!(prot & IOMMU_PROT_MASK))
1405 count = PAGE_SIZE_PTE_COUNT(page_size);
1406 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1411 for (i = 0; i < count; ++i)
1412 if (IOMMU_PTE_PRESENT(pte[i]))
1416 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1417 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1419 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1421 if (prot & IOMMU_PROT_IR)
1422 __pte |= IOMMU_PTE_IR;
1423 if (prot & IOMMU_PROT_IW)
1424 __pte |= IOMMU_PTE_IW;
1426 for (i = 0; i < count; ++i)
1434 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1435 unsigned long bus_addr,
1436 unsigned long page_size)
1438 unsigned long long unmapped;
1439 unsigned long unmap_size;
1442 BUG_ON(!is_power_of_2(page_size));
1446 while (unmapped < page_size) {
1448 pte = fetch_pte(dom, bus_addr, &unmap_size);
1453 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1454 for (i = 0; i < count; i++)
1458 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1459 unmapped += unmap_size;
1462 BUG_ON(unmapped && !is_power_of_2(unmapped));
1468 * This function checks if a specific unity mapping entry is needed for
1469 * this specific IOMMU.
1471 static int iommu_for_unity_map(struct amd_iommu *iommu,
1472 struct unity_map_entry *entry)
1476 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1477 bdf = amd_iommu_alias_table[i];
1478 if (amd_iommu_rlookup_table[bdf] == iommu)
1486 * This function actually applies the mapping to the page table of the
1489 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1490 struct unity_map_entry *e)
1495 for (addr = e->address_start; addr < e->address_end;
1496 addr += PAGE_SIZE) {
1497 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1502 * if unity mapping is in aperture range mark the page
1503 * as allocated in the aperture
1505 if (addr < dma_dom->aperture_size)
1506 __set_bit(addr >> PAGE_SHIFT,
1507 dma_dom->aperture[0]->bitmap);
1514 * Init the unity mappings for a specific IOMMU in the system
1516 * Basically iterates over all unity mapping entries and applies them to
1517 * the default domain DMA of that IOMMU if necessary.
1519 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1521 struct unity_map_entry *entry;
1524 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1525 if (!iommu_for_unity_map(iommu, entry))
1527 ret = dma_ops_unity_map(iommu->default_dom, entry);
1536 * Inits the unity mappings required for a specific device
1538 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1541 struct unity_map_entry *e;
1544 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1545 if (!(devid >= e->devid_start && devid <= e->devid_end))
1547 ret = dma_ops_unity_map(dma_dom, e);
1555 /****************************************************************************
1557 * The next functions belong to the address allocator for the dma_ops
1558 * interface functions. They work like the allocators in the other IOMMU
1559 * drivers. Its basically a bitmap which marks the allocated pages in
1560 * the aperture. Maybe it could be enhanced in the future to a more
1561 * efficient allocator.
1563 ****************************************************************************/
1566 * The address allocator core functions.
1568 * called with domain->lock held
1572 * Used to reserve address ranges in the aperture (e.g. for exclusion
1575 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1576 unsigned long start_page,
1579 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1581 if (start_page + pages > last_page)
1582 pages = last_page - start_page;
1584 for (i = start_page; i < start_page + pages; ++i) {
1585 int index = i / APERTURE_RANGE_PAGES;
1586 int page = i % APERTURE_RANGE_PAGES;
1587 __set_bit(page, dom->aperture[index]->bitmap);
1592 * This function is used to add a new aperture range to an existing
1593 * aperture in case of dma_ops domain allocation or address allocation
1596 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1597 bool populate, gfp_t gfp)
1599 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1600 struct amd_iommu *iommu;
1601 unsigned long i, old_size, pte_pgsize;
1603 #ifdef CONFIG_IOMMU_STRESS
1607 if (index >= APERTURE_MAX_RANGES)
1610 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1611 if (!dma_dom->aperture[index])
1614 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1615 if (!dma_dom->aperture[index]->bitmap)
1618 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1621 unsigned long address = dma_dom->aperture_size;
1622 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1623 u64 *pte, *pte_page;
1625 for (i = 0; i < num_ptes; ++i) {
1626 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1631 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1633 address += APERTURE_RANGE_SIZE / 64;
1637 old_size = dma_dom->aperture_size;
1638 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1640 /* Reserve address range used for MSI messages */
1641 if (old_size < MSI_ADDR_BASE_LO &&
1642 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1643 unsigned long spage;
1646 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1647 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1649 dma_ops_reserve_addresses(dma_dom, spage, pages);
1652 /* Initialize the exclusion range if necessary */
1653 for_each_iommu(iommu) {
1654 if (iommu->exclusion_start &&
1655 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1656 && iommu->exclusion_start < dma_dom->aperture_size) {
1657 unsigned long startpage;
1658 int pages = iommu_num_pages(iommu->exclusion_start,
1659 iommu->exclusion_length,
1661 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1662 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1667 * Check for areas already mapped as present in the new aperture
1668 * range and mark those pages as reserved in the allocator. Such
1669 * mappings may already exist as a result of requested unity
1670 * mappings for devices.
1672 for (i = dma_dom->aperture[index]->offset;
1673 i < dma_dom->aperture_size;
1675 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1676 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1679 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1683 update_domain(&dma_dom->domain);
1688 update_domain(&dma_dom->domain);
1690 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1692 kfree(dma_dom->aperture[index]);
1693 dma_dom->aperture[index] = NULL;
1698 static unsigned long dma_ops_area_alloc(struct device *dev,
1699 struct dma_ops_domain *dom,
1701 unsigned long align_mask,
1703 unsigned long start)
1705 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1706 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1707 int i = start >> APERTURE_RANGE_SHIFT;
1708 unsigned long boundary_size;
1709 unsigned long address = -1;
1710 unsigned long limit;
1712 next_bit >>= PAGE_SHIFT;
1714 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1715 PAGE_SIZE) >> PAGE_SHIFT;
1717 for (;i < max_index; ++i) {
1718 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1720 if (dom->aperture[i]->offset >= dma_mask)
1723 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1724 dma_mask >> PAGE_SHIFT);
1726 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1727 limit, next_bit, pages, 0,
1728 boundary_size, align_mask);
1729 if (address != -1) {
1730 address = dom->aperture[i]->offset +
1731 (address << PAGE_SHIFT);
1732 dom->next_address = address + (pages << PAGE_SHIFT);
1742 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1743 struct dma_ops_domain *dom,
1745 unsigned long align_mask,
1748 unsigned long address;
1750 #ifdef CONFIG_IOMMU_STRESS
1751 dom->next_address = 0;
1752 dom->need_flush = true;
1755 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1756 dma_mask, dom->next_address);
1758 if (address == -1) {
1759 dom->next_address = 0;
1760 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1762 dom->need_flush = true;
1765 if (unlikely(address == -1))
1766 address = DMA_ERROR_CODE;
1768 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1774 * The address free function.
1776 * called with domain->lock held
1778 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1779 unsigned long address,
1782 unsigned i = address >> APERTURE_RANGE_SHIFT;
1783 struct aperture_range *range = dom->aperture[i];
1785 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1787 #ifdef CONFIG_IOMMU_STRESS
1792 if (address >= dom->next_address)
1793 dom->need_flush = true;
1795 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1797 bitmap_clear(range->bitmap, address, pages);
1801 /****************************************************************************
1803 * The next functions belong to the domain allocation. A domain is
1804 * allocated for every IOMMU as the default domain. If device isolation
1805 * is enabled, every device get its own domain. The most important thing
1806 * about domains is the page table mapping the DMA address space they
1809 ****************************************************************************/
1812 * This function adds a protection domain to the global protection domain list
1814 static void add_domain_to_list(struct protection_domain *domain)
1816 unsigned long flags;
1818 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1819 list_add(&domain->list, &amd_iommu_pd_list);
1820 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1824 * This function removes a protection domain to the global
1825 * protection domain list
1827 static void del_domain_from_list(struct protection_domain *domain)
1829 unsigned long flags;
1831 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1832 list_del(&domain->list);
1833 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1836 static u16 domain_id_alloc(void)
1838 unsigned long flags;
1841 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1842 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1844 if (id > 0 && id < MAX_DOMAIN_ID)
1845 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1848 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1853 static void domain_id_free(int id)
1855 unsigned long flags;
1857 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1858 if (id > 0 && id < MAX_DOMAIN_ID)
1859 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1860 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1863 #define DEFINE_FREE_PT_FN(LVL, FN) \
1864 static void free_pt_##LVL (unsigned long __pt) \
1872 for (i = 0; i < 512; ++i) { \
1873 if (!IOMMU_PTE_PRESENT(pt[i])) \
1876 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1879 free_page((unsigned long)pt); \
1882 DEFINE_FREE_PT_FN(l2, free_page)
1883 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1884 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1885 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1886 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1888 static void free_pagetable(struct protection_domain *domain)
1890 unsigned long root = (unsigned long)domain->pt_root;
1892 switch (domain->mode) {
1893 case PAGE_MODE_NONE:
1895 case PAGE_MODE_1_LEVEL:
1898 case PAGE_MODE_2_LEVEL:
1901 case PAGE_MODE_3_LEVEL:
1904 case PAGE_MODE_4_LEVEL:
1907 case PAGE_MODE_5_LEVEL:
1910 case PAGE_MODE_6_LEVEL:
1918 static void free_gcr3_tbl_level1(u64 *tbl)
1923 for (i = 0; i < 512; ++i) {
1924 if (!(tbl[i] & GCR3_VALID))
1927 ptr = __va(tbl[i] & PAGE_MASK);
1929 free_page((unsigned long)ptr);
1933 static void free_gcr3_tbl_level2(u64 *tbl)
1938 for (i = 0; i < 512; ++i) {
1939 if (!(tbl[i] & GCR3_VALID))
1942 ptr = __va(tbl[i] & PAGE_MASK);
1944 free_gcr3_tbl_level1(ptr);
1948 static void free_gcr3_table(struct protection_domain *domain)
1950 if (domain->glx == 2)
1951 free_gcr3_tbl_level2(domain->gcr3_tbl);
1952 else if (domain->glx == 1)
1953 free_gcr3_tbl_level1(domain->gcr3_tbl);
1954 else if (domain->glx != 0)
1957 free_page((unsigned long)domain->gcr3_tbl);
1961 * Free a domain, only used if something went wrong in the
1962 * allocation path and we need to free an already allocated page table
1964 static void dma_ops_domain_free(struct dma_ops_domain *dom)
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]);
1986 * Allocates a new protection domain usable for the dma_ops functions.
1987 * It also initializes the page table and the address allocator data
1988 * structures required for the dma_ops interface
1990 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1992 struct dma_ops_domain *dma_dom;
1994 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1998 spin_lock_init(&dma_dom->domain.lock);
2000 dma_dom->domain.id = domain_id_alloc();
2001 if (dma_dom->domain.id == 0)
2003 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
2004 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
2005 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2006 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2007 dma_dom->domain.priv = dma_dom;
2008 if (!dma_dom->domain.pt_root)
2011 dma_dom->need_flush = false;
2012 dma_dom->target_dev = 0xffff;
2014 add_domain_to_list(&dma_dom->domain);
2016 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2020 * mark the first page as allocated so we never return 0 as
2021 * a valid dma-address. So we can use 0 as error value
2023 dma_dom->aperture[0]->bitmap[0] = 1;
2024 dma_dom->next_address = 0;
2030 dma_ops_domain_free(dma_dom);
2036 * little helper function to check whether a given protection domain is a
2039 static bool dma_ops_domain(struct protection_domain *domain)
2041 return domain->flags & PD_DMA_OPS_MASK;
2044 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2049 if (domain->mode != PAGE_MODE_NONE)
2050 pte_root = virt_to_phys(domain->pt_root);
2052 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2053 << DEV_ENTRY_MODE_SHIFT;
2054 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2056 flags = amd_iommu_dev_table[devid].data[1];
2059 flags |= DTE_FLAG_IOTLB;
2061 if (domain->flags & PD_IOMMUV2_MASK) {
2062 u64 gcr3 = __pa(domain->gcr3_tbl);
2063 u64 glx = domain->glx;
2066 pte_root |= DTE_FLAG_GV;
2067 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2069 /* First mask out possible old values for GCR3 table */
2070 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2073 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2076 /* Encode GCR3 table into DTE */
2077 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2080 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2083 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2087 flags &= ~(0xffffUL);
2088 flags |= domain->id;
2090 amd_iommu_dev_table[devid].data[1] = flags;
2091 amd_iommu_dev_table[devid].data[0] = pte_root;
2094 static void clear_dte_entry(u16 devid)
2096 /* remove entry from the device table seen by the hardware */
2097 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2098 amd_iommu_dev_table[devid].data[1] = 0;
2100 amd_iommu_apply_erratum_63(devid);
2103 static void do_attach(struct iommu_dev_data *dev_data,
2104 struct protection_domain *domain)
2106 struct amd_iommu *iommu;
2109 iommu = amd_iommu_rlookup_table[dev_data->devid];
2110 ats = dev_data->ats.enabled;
2112 /* Update data structures */
2113 dev_data->domain = domain;
2114 list_add(&dev_data->list, &domain->dev_list);
2115 set_dte_entry(dev_data->devid, domain, ats);
2117 /* Do reference counting */
2118 domain->dev_iommu[iommu->index] += 1;
2119 domain->dev_cnt += 1;
2121 /* Flush the DTE entry */
2122 device_flush_dte(dev_data);
2125 static void do_detach(struct iommu_dev_data *dev_data)
2127 struct amd_iommu *iommu;
2129 iommu = amd_iommu_rlookup_table[dev_data->devid];
2131 /* decrease reference counters */
2132 dev_data->domain->dev_iommu[iommu->index] -= 1;
2133 dev_data->domain->dev_cnt -= 1;
2135 /* Update data structures */
2136 dev_data->domain = NULL;
2137 list_del(&dev_data->list);
2138 clear_dte_entry(dev_data->devid);
2140 /* Flush the DTE entry */
2141 device_flush_dte(dev_data);
2145 * If a device is not yet associated with a domain, this function does
2146 * assigns it visible for the hardware
2148 static int __attach_device(struct iommu_dev_data *dev_data,
2149 struct protection_domain *domain)
2151 struct iommu_dev_data *head, *entry;
2155 spin_lock(&domain->lock);
2159 if (head->alias_data != NULL)
2160 head = head->alias_data;
2162 /* Now we have the root of the alias group, if any */
2165 if (head->domain != NULL)
2168 /* Attach alias group root */
2169 do_attach(head, domain);
2171 /* Attach other devices in the alias group */
2172 list_for_each_entry(entry, &head->alias_list, alias_list)
2173 do_attach(entry, domain);
2180 spin_unlock(&domain->lock);
2186 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2188 pci_disable_ats(pdev);
2189 pci_disable_pri(pdev);
2190 pci_disable_pasid(pdev);
2193 /* FIXME: Change generic reset-function to do the same */
2194 static int pri_reset_while_enabled(struct pci_dev *pdev)
2199 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2203 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2204 control |= PCI_PRI_CTRL_RESET;
2205 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2210 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2215 /* FIXME: Hardcode number of outstanding requests for now */
2217 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2219 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2221 /* Only allow access to user-accessible pages */
2222 ret = pci_enable_pasid(pdev, 0);
2226 /* First reset the PRI state of the device */
2227 ret = pci_reset_pri(pdev);
2232 ret = pci_enable_pri(pdev, reqs);
2237 ret = pri_reset_while_enabled(pdev);
2242 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2249 pci_disable_pri(pdev);
2250 pci_disable_pasid(pdev);
2255 /* FIXME: Move this to PCI code */
2256 #define PCI_PRI_TLP_OFF (1 << 15)
2258 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2263 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2267 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2269 return (status & PCI_PRI_TLP_OFF) ? true : false;
2273 * If a device is not yet associated with a domain, this function
2274 * assigns it visible for the hardware
2276 static int attach_device(struct device *dev,
2277 struct protection_domain *domain)
2279 struct pci_dev *pdev = to_pci_dev(dev);
2280 struct iommu_dev_data *dev_data;
2281 unsigned long flags;
2284 dev_data = get_dev_data(dev);
2286 if (domain->flags & PD_IOMMUV2_MASK) {
2287 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2290 if (pdev_iommuv2_enable(pdev) != 0)
2293 dev_data->ats.enabled = true;
2294 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2295 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2296 } else if (amd_iommu_iotlb_sup &&
2297 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2298 dev_data->ats.enabled = true;
2299 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2302 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2303 ret = __attach_device(dev_data, domain);
2304 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2307 * We might boot into a crash-kernel here. The crashed kernel
2308 * left the caches in the IOMMU dirty. So we have to flush
2309 * here to evict all dirty stuff.
2311 domain_flush_tlb_pde(domain);
2317 * Removes a device from a protection domain (unlocked)
2319 static void __detach_device(struct iommu_dev_data *dev_data)
2321 struct iommu_dev_data *head, *entry;
2322 struct protection_domain *domain;
2323 unsigned long flags;
2325 BUG_ON(!dev_data->domain);
2327 domain = dev_data->domain;
2329 spin_lock_irqsave(&domain->lock, flags);
2332 if (head->alias_data != NULL)
2333 head = head->alias_data;
2335 list_for_each_entry(entry, &head->alias_list, alias_list)
2340 spin_unlock_irqrestore(&domain->lock, flags);
2343 * If we run in passthrough mode the device must be assigned to the
2344 * passthrough domain if it is detached from any other domain.
2345 * Make sure we can deassign from the pt_domain itself.
2347 if (dev_data->passthrough &&
2348 (dev_data->domain == NULL && domain != pt_domain))
2349 __attach_device(dev_data, pt_domain);
2353 * Removes a device from a protection domain (with devtable_lock held)
2355 static void detach_device(struct device *dev)
2357 struct protection_domain *domain;
2358 struct iommu_dev_data *dev_data;
2359 unsigned long flags;
2361 dev_data = get_dev_data(dev);
2362 domain = dev_data->domain;
2364 /* lock device table */
2365 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2366 __detach_device(dev_data);
2367 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2369 if (domain->flags & PD_IOMMUV2_MASK)
2370 pdev_iommuv2_disable(to_pci_dev(dev));
2371 else if (dev_data->ats.enabled)
2372 pci_disable_ats(to_pci_dev(dev));
2374 dev_data->ats.enabled = false;
2378 * Find out the protection domain structure for a given PCI device. This
2379 * will give us the pointer to the page table root for example.
2381 static struct protection_domain *domain_for_device(struct device *dev)
2383 struct iommu_dev_data *dev_data;
2384 struct protection_domain *dom = NULL;
2385 unsigned long flags;
2387 dev_data = get_dev_data(dev);
2389 if (dev_data->domain)
2390 return dev_data->domain;
2392 if (dev_data->alias_data != NULL) {
2393 struct iommu_dev_data *alias_data = dev_data->alias_data;
2395 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2396 if (alias_data->domain != NULL) {
2397 __attach_device(dev_data, alias_data->domain);
2398 dom = alias_data->domain;
2400 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2406 static int device_change_notifier(struct notifier_block *nb,
2407 unsigned long action, void *data)
2409 struct dma_ops_domain *dma_domain;
2410 struct protection_domain *domain;
2411 struct iommu_dev_data *dev_data;
2412 struct device *dev = data;
2413 struct amd_iommu *iommu;
2414 unsigned long flags;
2417 if (!check_device(dev))
2420 devid = get_device_id(dev);
2421 iommu = amd_iommu_rlookup_table[devid];
2422 dev_data = get_dev_data(dev);
2425 case BUS_NOTIFY_ADD_DEVICE:
2427 iommu_init_device(dev);
2428 init_iommu_group(dev);
2431 * dev_data is still NULL and
2432 * got initialized in iommu_init_device
2434 dev_data = get_dev_data(dev);
2436 if (iommu_pass_through || dev_data->iommu_v2) {
2437 dev_data->passthrough = true;
2438 attach_device(dev, pt_domain);
2442 domain = domain_for_device(dev);
2444 /* allocate a protection domain if a device is added */
2445 dma_domain = find_protection_domain(devid);
2447 dma_domain = dma_ops_domain_alloc();
2450 dma_domain->target_dev = devid;
2452 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2453 list_add_tail(&dma_domain->list, &iommu_pd_list);
2454 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2457 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2460 case BUS_NOTIFY_REMOVED_DEVICE:
2462 iommu_uninit_device(dev);
2468 iommu_completion_wait(iommu);
2474 static struct notifier_block device_nb = {
2475 .notifier_call = device_change_notifier,
2478 void amd_iommu_init_notifier(void)
2480 bus_register_notifier(&pci_bus_type, &device_nb);
2483 /*****************************************************************************
2485 * The next functions belong to the dma_ops mapping/unmapping code.
2487 *****************************************************************************/
2490 * In the dma_ops path we only have the struct device. This function
2491 * finds the corresponding IOMMU, the protection domain and the
2492 * requestor id for a given device.
2493 * If the device is not yet associated with a domain this is also done
2496 static struct protection_domain *get_domain(struct device *dev)
2498 struct protection_domain *domain;
2499 struct dma_ops_domain *dma_dom;
2500 u16 devid = get_device_id(dev);
2502 if (!check_device(dev))
2503 return ERR_PTR(-EINVAL);
2505 domain = domain_for_device(dev);
2506 if (domain != NULL && !dma_ops_domain(domain))
2507 return ERR_PTR(-EBUSY);
2512 /* Device not bound yet - bind it */
2513 dma_dom = find_protection_domain(devid);
2515 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2516 attach_device(dev, &dma_dom->domain);
2517 DUMP_printk("Using protection domain %d for device %s\n",
2518 dma_dom->domain.id, dev_name(dev));
2520 return &dma_dom->domain;
2523 static void update_device_table(struct protection_domain *domain)
2525 struct iommu_dev_data *dev_data;
2527 list_for_each_entry(dev_data, &domain->dev_list, list)
2528 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2531 static void update_domain(struct protection_domain *domain)
2533 if (!domain->updated)
2536 update_device_table(domain);
2538 domain_flush_devices(domain);
2539 domain_flush_tlb_pde(domain);
2541 domain->updated = false;
2545 * This function fetches the PTE for a given address in the aperture
2547 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2548 unsigned long address)
2550 struct aperture_range *aperture;
2551 u64 *pte, *pte_page;
2553 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2557 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2559 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2561 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2563 pte += PM_LEVEL_INDEX(0, address);
2565 update_domain(&dom->domain);
2571 * This is the generic map function. It maps one 4kb page at paddr to
2572 * the given address in the DMA address space for the domain.
2574 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2575 unsigned long address,
2581 WARN_ON(address > dom->aperture_size);
2585 pte = dma_ops_get_pte(dom, address);
2587 return DMA_ERROR_CODE;
2589 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2591 if (direction == DMA_TO_DEVICE)
2592 __pte |= IOMMU_PTE_IR;
2593 else if (direction == DMA_FROM_DEVICE)
2594 __pte |= IOMMU_PTE_IW;
2595 else if (direction == DMA_BIDIRECTIONAL)
2596 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2602 return (dma_addr_t)address;
2606 * The generic unmapping function for on page in the DMA address space.
2608 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2609 unsigned long address)
2611 struct aperture_range *aperture;
2614 if (address >= dom->aperture_size)
2617 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2621 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2625 pte += PM_LEVEL_INDEX(0, address);
2633 * This function contains common code for mapping of a physically
2634 * contiguous memory region into DMA address space. It is used by all
2635 * mapping functions provided with this IOMMU driver.
2636 * Must be called with the domain lock held.
2638 static dma_addr_t __map_single(struct device *dev,
2639 struct dma_ops_domain *dma_dom,
2646 dma_addr_t offset = paddr & ~PAGE_MASK;
2647 dma_addr_t address, start, ret;
2649 unsigned long align_mask = 0;
2652 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2655 INC_STATS_COUNTER(total_map_requests);
2658 INC_STATS_COUNTER(cross_page);
2661 align_mask = (1UL << get_order(size)) - 1;
2664 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2666 if (unlikely(address == DMA_ERROR_CODE)) {
2668 * setting next_address here will let the address
2669 * allocator only scan the new allocated range in the
2670 * first run. This is a small optimization.
2672 dma_dom->next_address = dma_dom->aperture_size;
2674 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2678 * aperture was successfully enlarged by 128 MB, try
2685 for (i = 0; i < pages; ++i) {
2686 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2687 if (ret == DMA_ERROR_CODE)
2695 ADD_STATS_COUNTER(alloced_io_mem, size);
2697 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2698 domain_flush_tlb(&dma_dom->domain);
2699 dma_dom->need_flush = false;
2700 } else if (unlikely(amd_iommu_np_cache))
2701 domain_flush_pages(&dma_dom->domain, address, size);
2708 for (--i; i >= 0; --i) {
2710 dma_ops_domain_unmap(dma_dom, start);
2713 dma_ops_free_addresses(dma_dom, address, pages);
2715 return DMA_ERROR_CODE;
2719 * Does the reverse of the __map_single function. Must be called with
2720 * the domain lock held too
2722 static void __unmap_single(struct dma_ops_domain *dma_dom,
2723 dma_addr_t dma_addr,
2727 dma_addr_t flush_addr;
2728 dma_addr_t i, start;
2731 if ((dma_addr == DMA_ERROR_CODE) ||
2732 (dma_addr + size > dma_dom->aperture_size))
2735 flush_addr = dma_addr;
2736 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2737 dma_addr &= PAGE_MASK;
2740 for (i = 0; i < pages; ++i) {
2741 dma_ops_domain_unmap(dma_dom, start);
2745 SUB_STATS_COUNTER(alloced_io_mem, size);
2747 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2749 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2750 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2751 dma_dom->need_flush = false;
2756 * The exported map_single function for dma_ops.
2758 static dma_addr_t map_page(struct device *dev, struct page *page,
2759 unsigned long offset, size_t size,
2760 enum dma_data_direction dir,
2761 struct dma_attrs *attrs)
2763 unsigned long flags;
2764 struct protection_domain *domain;
2767 phys_addr_t paddr = page_to_phys(page) + offset;
2769 INC_STATS_COUNTER(cnt_map_single);
2771 domain = get_domain(dev);
2772 if (PTR_ERR(domain) == -EINVAL)
2773 return (dma_addr_t)paddr;
2774 else if (IS_ERR(domain))
2775 return DMA_ERROR_CODE;
2777 dma_mask = *dev->dma_mask;
2779 spin_lock_irqsave(&domain->lock, flags);
2781 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2783 if (addr == DMA_ERROR_CODE)
2786 domain_flush_complete(domain);
2789 spin_unlock_irqrestore(&domain->lock, flags);
2795 * The exported unmap_single function for dma_ops.
2797 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2798 enum dma_data_direction dir, struct dma_attrs *attrs)
2800 unsigned long flags;
2801 struct protection_domain *domain;
2803 INC_STATS_COUNTER(cnt_unmap_single);
2805 domain = get_domain(dev);
2809 spin_lock_irqsave(&domain->lock, flags);
2811 __unmap_single(domain->priv, dma_addr, size, dir);
2813 domain_flush_complete(domain);
2815 spin_unlock_irqrestore(&domain->lock, flags);
2819 * The exported map_sg function for dma_ops (handles scatter-gather
2822 static int map_sg(struct device *dev, struct scatterlist *sglist,
2823 int nelems, enum dma_data_direction dir,
2824 struct dma_attrs *attrs)
2826 unsigned long flags;
2827 struct protection_domain *domain;
2829 struct scatterlist *s;
2831 int mapped_elems = 0;
2834 INC_STATS_COUNTER(cnt_map_sg);
2836 domain = get_domain(dev);
2840 dma_mask = *dev->dma_mask;
2842 spin_lock_irqsave(&domain->lock, flags);
2844 for_each_sg(sglist, s, nelems, i) {
2847 s->dma_address = __map_single(dev, domain->priv,
2848 paddr, s->length, dir, false,
2851 if (s->dma_address) {
2852 s->dma_length = s->length;
2858 domain_flush_complete(domain);
2861 spin_unlock_irqrestore(&domain->lock, flags);
2863 return mapped_elems;
2865 for_each_sg(sglist, s, mapped_elems, i) {
2867 __unmap_single(domain->priv, s->dma_address,
2868 s->dma_length, dir);
2869 s->dma_address = s->dma_length = 0;
2878 * The exported map_sg function for dma_ops (handles scatter-gather
2881 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2882 int nelems, enum dma_data_direction dir,
2883 struct dma_attrs *attrs)
2885 unsigned long flags;
2886 struct protection_domain *domain;
2887 struct scatterlist *s;
2890 INC_STATS_COUNTER(cnt_unmap_sg);
2892 domain = get_domain(dev);
2896 spin_lock_irqsave(&domain->lock, flags);
2898 for_each_sg(sglist, s, nelems, i) {
2899 __unmap_single(domain->priv, s->dma_address,
2900 s->dma_length, dir);
2901 s->dma_address = s->dma_length = 0;
2904 domain_flush_complete(domain);
2906 spin_unlock_irqrestore(&domain->lock, flags);
2910 * The exported alloc_coherent function for dma_ops.
2912 static void *alloc_coherent(struct device *dev, size_t size,
2913 dma_addr_t *dma_addr, gfp_t flag,
2914 struct dma_attrs *attrs)
2916 u64 dma_mask = dev->coherent_dma_mask;
2917 struct protection_domain *domain;
2918 unsigned long flags;
2921 INC_STATS_COUNTER(cnt_alloc_coherent);
2923 domain = get_domain(dev);
2924 if (PTR_ERR(domain) == -EINVAL) {
2925 page = alloc_pages(flag, get_order(size));
2926 *dma_addr = page_to_phys(page);
2927 return page_address(page);
2928 } else if (IS_ERR(domain))
2931 size = PAGE_ALIGN(size);
2932 dma_mask = dev->coherent_dma_mask;
2933 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2936 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2938 if (!(flag & __GFP_WAIT))
2941 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2948 dma_mask = *dev->dma_mask;
2950 spin_lock_irqsave(&domain->lock, flags);
2952 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2953 size, DMA_BIDIRECTIONAL, true, dma_mask);
2955 if (*dma_addr == DMA_ERROR_CODE) {
2956 spin_unlock_irqrestore(&domain->lock, flags);
2960 domain_flush_complete(domain);
2962 spin_unlock_irqrestore(&domain->lock, flags);
2964 return page_address(page);
2968 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2969 __free_pages(page, get_order(size));
2975 * The exported free_coherent function for dma_ops.
2977 static void free_coherent(struct device *dev, size_t size,
2978 void *virt_addr, dma_addr_t dma_addr,
2979 struct dma_attrs *attrs)
2981 struct protection_domain *domain;
2982 unsigned long flags;
2985 INC_STATS_COUNTER(cnt_free_coherent);
2987 page = virt_to_page(virt_addr);
2988 size = PAGE_ALIGN(size);
2990 domain = get_domain(dev);
2994 spin_lock_irqsave(&domain->lock, flags);
2996 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2998 domain_flush_complete(domain);
3000 spin_unlock_irqrestore(&domain->lock, flags);
3003 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3004 __free_pages(page, get_order(size));
3008 * This function is called by the DMA layer to find out if we can handle a
3009 * particular device. It is part of the dma_ops.
3011 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3013 return check_device(dev);
3017 * The function for pre-allocating protection domains.
3019 * If the driver core informs the DMA layer if a driver grabs a device
3020 * we don't need to preallocate the protection domains anymore.
3021 * For now we have to.
3023 static void __init prealloc_protection_domains(void)
3025 struct iommu_dev_data *dev_data;
3026 struct dma_ops_domain *dma_dom;
3027 struct pci_dev *dev = NULL;
3030 for_each_pci_dev(dev) {
3032 /* Do we handle this device? */
3033 if (!check_device(&dev->dev))
3036 dev_data = get_dev_data(&dev->dev);
3037 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3038 /* Make sure passthrough domain is allocated */
3039 alloc_passthrough_domain();
3040 dev_data->passthrough = true;
3041 attach_device(&dev->dev, pt_domain);
3042 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3043 dev_name(&dev->dev));
3046 /* Is there already any domain for it? */
3047 if (domain_for_device(&dev->dev))
3050 devid = get_device_id(&dev->dev);
3052 dma_dom = dma_ops_domain_alloc();
3055 init_unity_mappings_for_device(dma_dom, devid);
3056 dma_dom->target_dev = devid;
3058 attach_device(&dev->dev, &dma_dom->domain);
3060 list_add_tail(&dma_dom->list, &iommu_pd_list);
3064 static struct dma_map_ops amd_iommu_dma_ops = {
3065 .alloc = alloc_coherent,
3066 .free = free_coherent,
3067 .map_page = map_page,
3068 .unmap_page = unmap_page,
3070 .unmap_sg = unmap_sg,
3071 .dma_supported = amd_iommu_dma_supported,
3074 static unsigned device_dma_ops_init(void)
3076 struct iommu_dev_data *dev_data;
3077 struct pci_dev *pdev = NULL;
3078 unsigned unhandled = 0;
3080 for_each_pci_dev(pdev) {
3081 if (!check_device(&pdev->dev)) {
3083 iommu_ignore_device(&pdev->dev);
3089 dev_data = get_dev_data(&pdev->dev);
3091 if (!dev_data->passthrough)
3092 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3094 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3101 * The function which clues the AMD IOMMU driver into dma_ops.
3104 void __init amd_iommu_init_api(void)
3106 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3109 int __init amd_iommu_init_dma_ops(void)
3111 struct amd_iommu *iommu;
3115 * first allocate a default protection domain for every IOMMU we
3116 * found in the system. Devices not assigned to any other
3117 * protection domain will be assigned to the default one.
3119 for_each_iommu(iommu) {
3120 iommu->default_dom = dma_ops_domain_alloc();
3121 if (iommu->default_dom == NULL)
3123 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3124 ret = iommu_init_unity_mappings(iommu);
3130 * Pre-allocate the protection domains for each device.
3132 prealloc_protection_domains();
3137 /* Make the driver finally visible to the drivers */
3138 unhandled = device_dma_ops_init();
3139 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3140 /* There are unhandled devices - initialize swiotlb for them */
3144 amd_iommu_stats_init();
3146 if (amd_iommu_unmap_flush)
3147 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3149 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3155 for_each_iommu(iommu) {
3156 dma_ops_domain_free(iommu->default_dom);
3162 /*****************************************************************************
3164 * The following functions belong to the exported interface of AMD IOMMU
3166 * This interface allows access to lower level functions of the IOMMU
3167 * like protection domain handling and assignement of devices to domains
3168 * which is not possible with the dma_ops interface.
3170 *****************************************************************************/
3172 static void cleanup_domain(struct protection_domain *domain)
3174 struct iommu_dev_data *entry;
3175 unsigned long flags;
3177 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3179 while (!list_empty(&domain->dev_list)) {
3180 entry = list_first_entry(&domain->dev_list,
3181 struct iommu_dev_data, list);
3182 __detach_device(entry);
3185 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3188 static void protection_domain_free(struct protection_domain *domain)
3193 del_domain_from_list(domain);
3196 domain_id_free(domain->id);
3201 static struct protection_domain *protection_domain_alloc(void)
3203 struct protection_domain *domain;
3205 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3209 spin_lock_init(&domain->lock);
3210 mutex_init(&domain->api_lock);
3211 domain->id = domain_id_alloc();
3214 INIT_LIST_HEAD(&domain->dev_list);
3216 add_domain_to_list(domain);
3226 static int __init alloc_passthrough_domain(void)
3228 if (pt_domain != NULL)
3231 /* allocate passthrough domain */
3232 pt_domain = protection_domain_alloc();
3236 pt_domain->mode = PAGE_MODE_NONE;
3241 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
3243 struct protection_domain *pdomain;
3245 /* We only support unmanaged domains for now */
3246 if (type != IOMMU_DOMAIN_UNMANAGED)
3249 pdomain = protection_domain_alloc();
3253 pdomain->mode = PAGE_MODE_3_LEVEL;
3254 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3255 if (!pdomain->pt_root)
3258 pdomain->domain.geometry.aperture_start = 0;
3259 pdomain->domain.geometry.aperture_end = ~0ULL;
3260 pdomain->domain.geometry.force_aperture = true;
3262 return &pdomain->domain;
3265 protection_domain_free(pdomain);
3270 static void amd_iommu_domain_free(struct iommu_domain *dom)
3272 struct protection_domain *domain;
3277 domain = to_pdomain(dom);
3279 if (domain->dev_cnt > 0)
3280 cleanup_domain(domain);
3282 BUG_ON(domain->dev_cnt != 0);
3284 if (domain->mode != PAGE_MODE_NONE)
3285 free_pagetable(domain);
3287 if (domain->flags & PD_IOMMUV2_MASK)
3288 free_gcr3_table(domain);
3290 protection_domain_free(domain);
3293 static void amd_iommu_detach_device(struct iommu_domain *dom,
3296 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3297 struct amd_iommu *iommu;
3300 if (!check_device(dev))
3303 devid = get_device_id(dev);
3305 if (dev_data->domain != NULL)
3308 iommu = amd_iommu_rlookup_table[devid];
3312 iommu_completion_wait(iommu);
3315 static int amd_iommu_attach_device(struct iommu_domain *dom,
3318 struct protection_domain *domain = to_pdomain(dom);
3319 struct iommu_dev_data *dev_data;
3320 struct amd_iommu *iommu;
3323 if (!check_device(dev))
3326 dev_data = dev->archdata.iommu;
3328 iommu = amd_iommu_rlookup_table[dev_data->devid];
3332 if (dev_data->domain)
3335 ret = attach_device(dev, domain);
3337 iommu_completion_wait(iommu);
3342 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3343 phys_addr_t paddr, size_t page_size, int iommu_prot)
3345 struct protection_domain *domain = to_pdomain(dom);
3349 if (domain->mode == PAGE_MODE_NONE)
3352 if (iommu_prot & IOMMU_READ)
3353 prot |= IOMMU_PROT_IR;
3354 if (iommu_prot & IOMMU_WRITE)
3355 prot |= IOMMU_PROT_IW;
3357 mutex_lock(&domain->api_lock);
3358 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3359 mutex_unlock(&domain->api_lock);
3364 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3367 struct protection_domain *domain = to_pdomain(dom);
3370 if (domain->mode == PAGE_MODE_NONE)
3373 mutex_lock(&domain->api_lock);
3374 unmap_size = iommu_unmap_page(domain, iova, page_size);
3375 mutex_unlock(&domain->api_lock);
3377 domain_flush_tlb_pde(domain);
3382 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3385 struct protection_domain *domain = to_pdomain(dom);
3386 unsigned long offset_mask, pte_pgsize;
3389 if (domain->mode == PAGE_MODE_NONE)
3392 pte = fetch_pte(domain, iova, &pte_pgsize);
3394 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3397 offset_mask = pte_pgsize - 1;
3398 __pte = *pte & PM_ADDR_MASK;
3400 return (__pte & ~offset_mask) | (iova & offset_mask);
3403 static bool amd_iommu_capable(enum iommu_cap cap)
3406 case IOMMU_CAP_CACHE_COHERENCY:
3408 case IOMMU_CAP_INTR_REMAP:
3409 return (irq_remapping_enabled == 1);
3410 case IOMMU_CAP_NOEXEC:
3417 static const struct iommu_ops amd_iommu_ops = {
3418 .capable = amd_iommu_capable,
3419 .domain_alloc = amd_iommu_domain_alloc,
3420 .domain_free = amd_iommu_domain_free,
3421 .attach_dev = amd_iommu_attach_device,
3422 .detach_dev = amd_iommu_detach_device,
3423 .map = amd_iommu_map,
3424 .unmap = amd_iommu_unmap,
3425 .map_sg = default_iommu_map_sg,
3426 .iova_to_phys = amd_iommu_iova_to_phys,
3427 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3430 /*****************************************************************************
3432 * The next functions do a basic initialization of IOMMU for pass through
3435 * In passthrough mode the IOMMU is initialized and enabled but not used for
3436 * DMA-API translation.
3438 *****************************************************************************/
3440 int __init amd_iommu_init_passthrough(void)
3442 struct iommu_dev_data *dev_data;
3443 struct pci_dev *dev = NULL;
3446 ret = alloc_passthrough_domain();
3450 for_each_pci_dev(dev) {
3451 if (!check_device(&dev->dev))
3454 dev_data = get_dev_data(&dev->dev);
3455 dev_data->passthrough = true;
3457 attach_device(&dev->dev, pt_domain);
3460 amd_iommu_stats_init();
3462 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3467 /* IOMMUv2 specific functions */
3468 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3470 return atomic_notifier_chain_register(&ppr_notifier, nb);
3472 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3474 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3476 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3478 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3480 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3482 struct protection_domain *domain = to_pdomain(dom);
3483 unsigned long flags;
3485 spin_lock_irqsave(&domain->lock, flags);
3487 /* Update data structure */
3488 domain->mode = PAGE_MODE_NONE;
3489 domain->updated = true;
3491 /* Make changes visible to IOMMUs */
3492 update_domain(domain);
3494 /* Page-table is not visible to IOMMU anymore, so free it */
3495 free_pagetable(domain);
3497 spin_unlock_irqrestore(&domain->lock, flags);
3499 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3501 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3503 struct protection_domain *domain = to_pdomain(dom);
3504 unsigned long flags;
3507 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3510 /* Number of GCR3 table levels required */
3511 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3514 if (levels > amd_iommu_max_glx_val)
3517 spin_lock_irqsave(&domain->lock, flags);
3520 * Save us all sanity checks whether devices already in the
3521 * domain support IOMMUv2. Just force that the domain has no
3522 * devices attached when it is switched into IOMMUv2 mode.
3525 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3529 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3530 if (domain->gcr3_tbl == NULL)
3533 domain->glx = levels;
3534 domain->flags |= PD_IOMMUV2_MASK;
3535 domain->updated = true;
3537 update_domain(domain);
3542 spin_unlock_irqrestore(&domain->lock, flags);
3546 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3548 static int __flush_pasid(struct protection_domain *domain, int pasid,
3549 u64 address, bool size)
3551 struct iommu_dev_data *dev_data;
3552 struct iommu_cmd cmd;
3555 if (!(domain->flags & PD_IOMMUV2_MASK))
3558 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3561 * IOMMU TLB needs to be flushed before Device TLB to
3562 * prevent device TLB refill from IOMMU TLB
3564 for (i = 0; i < amd_iommus_present; ++i) {
3565 if (domain->dev_iommu[i] == 0)
3568 ret = iommu_queue_command(amd_iommus[i], &cmd);
3573 /* Wait until IOMMU TLB flushes are complete */
3574 domain_flush_complete(domain);
3576 /* Now flush device TLBs */
3577 list_for_each_entry(dev_data, &domain->dev_list, list) {
3578 struct amd_iommu *iommu;
3581 BUG_ON(!dev_data->ats.enabled);
3583 qdep = dev_data->ats.qdep;
3584 iommu = amd_iommu_rlookup_table[dev_data->devid];
3586 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3587 qdep, address, size);
3589 ret = iommu_queue_command(iommu, &cmd);
3594 /* Wait until all device TLBs are flushed */
3595 domain_flush_complete(domain);
3604 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3607 INC_STATS_COUNTER(invalidate_iotlb);
3609 return __flush_pasid(domain, pasid, address, false);
3612 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3615 struct protection_domain *domain = to_pdomain(dom);
3616 unsigned long flags;
3619 spin_lock_irqsave(&domain->lock, flags);
3620 ret = __amd_iommu_flush_page(domain, pasid, address);
3621 spin_unlock_irqrestore(&domain->lock, flags);
3625 EXPORT_SYMBOL(amd_iommu_flush_page);
3627 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3629 INC_STATS_COUNTER(invalidate_iotlb_all);
3631 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3635 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3637 struct protection_domain *domain = to_pdomain(dom);
3638 unsigned long flags;
3641 spin_lock_irqsave(&domain->lock, flags);
3642 ret = __amd_iommu_flush_tlb(domain, pasid);
3643 spin_unlock_irqrestore(&domain->lock, flags);
3647 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3649 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3656 index = (pasid >> (9 * level)) & 0x1ff;
3662 if (!(*pte & GCR3_VALID)) {
3666 root = (void *)get_zeroed_page(GFP_ATOMIC);
3670 *pte = __pa(root) | GCR3_VALID;
3673 root = __va(*pte & PAGE_MASK);
3681 static int __set_gcr3(struct protection_domain *domain, int pasid,
3686 if (domain->mode != PAGE_MODE_NONE)
3689 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3693 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3695 return __amd_iommu_flush_tlb(domain, pasid);
3698 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3702 if (domain->mode != PAGE_MODE_NONE)
3705 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3711 return __amd_iommu_flush_tlb(domain, pasid);
3714 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3717 struct protection_domain *domain = to_pdomain(dom);
3718 unsigned long flags;
3721 spin_lock_irqsave(&domain->lock, flags);
3722 ret = __set_gcr3(domain, pasid, cr3);
3723 spin_unlock_irqrestore(&domain->lock, flags);
3727 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3729 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3731 struct protection_domain *domain = to_pdomain(dom);
3732 unsigned long flags;
3735 spin_lock_irqsave(&domain->lock, flags);
3736 ret = __clear_gcr3(domain, pasid);
3737 spin_unlock_irqrestore(&domain->lock, flags);
3741 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3743 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3744 int status, int tag)
3746 struct iommu_dev_data *dev_data;
3747 struct amd_iommu *iommu;
3748 struct iommu_cmd cmd;
3750 INC_STATS_COUNTER(complete_ppr);
3752 dev_data = get_dev_data(&pdev->dev);
3753 iommu = amd_iommu_rlookup_table[dev_data->devid];
3755 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3756 tag, dev_data->pri_tlp);
3758 return iommu_queue_command(iommu, &cmd);
3760 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3762 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3764 struct protection_domain *pdomain;
3766 pdomain = get_domain(&pdev->dev);
3767 if (IS_ERR(pdomain))
3770 /* Only return IOMMUv2 domains */
3771 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3774 return &pdomain->domain;
3776 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3778 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3780 struct iommu_dev_data *dev_data;
3782 if (!amd_iommu_v2_supported())
3785 dev_data = get_dev_data(&pdev->dev);
3786 dev_data->errata |= (1 << erratum);
3788 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3790 int amd_iommu_device_info(struct pci_dev *pdev,
3791 struct amd_iommu_device_info *info)
3796 if (pdev == NULL || info == NULL)
3799 if (!amd_iommu_v2_supported())
3802 memset(info, 0, sizeof(*info));
3804 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3806 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3808 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3810 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3812 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3816 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3817 max_pasids = min(max_pasids, (1 << 20));
3819 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3820 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3822 features = pci_pasid_features(pdev);
3823 if (features & PCI_PASID_CAP_EXEC)
3824 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3825 if (features & PCI_PASID_CAP_PRIV)
3826 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3831 EXPORT_SYMBOL(amd_iommu_device_info);
3833 #ifdef CONFIG_IRQ_REMAP
3835 /*****************************************************************************
3837 * Interrupt Remapping Implementation
3839 *****************************************************************************/
3857 u16 devid; /* Device ID for IRTE table */
3858 u16 index; /* Index into IRTE table*/
3861 struct amd_ir_data {
3862 struct irq_2_irte irq_2_irte;
3863 union irte irte_entry;
3865 struct msi_msg msi_entry;
3869 static struct irq_chip amd_ir_chip;
3871 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3872 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3873 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3874 #define DTE_IRQ_REMAP_ENABLE 1ULL
3876 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3880 dte = amd_iommu_dev_table[devid].data[2];
3881 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3882 dte |= virt_to_phys(table->table);
3883 dte |= DTE_IRQ_REMAP_INTCTL;
3884 dte |= DTE_IRQ_TABLE_LEN;
3885 dte |= DTE_IRQ_REMAP_ENABLE;
3887 amd_iommu_dev_table[devid].data[2] = dte;
3890 #define IRTE_ALLOCATED (~1U)
3892 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3894 struct irq_remap_table *table = NULL;
3895 struct amd_iommu *iommu;
3896 unsigned long flags;
3899 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3901 iommu = amd_iommu_rlookup_table[devid];
3905 table = irq_lookup_table[devid];
3909 alias = amd_iommu_alias_table[devid];
3910 table = irq_lookup_table[alias];
3912 irq_lookup_table[devid] = table;
3913 set_dte_irq_entry(devid, table);
3914 iommu_flush_dte(iommu, devid);
3918 /* Nothing there yet, allocate new irq remapping table */
3919 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3923 /* Initialize table spin-lock */
3924 spin_lock_init(&table->lock);
3927 /* Keep the first 32 indexes free for IOAPIC interrupts */
3928 table->min_index = 32;
3930 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3931 if (!table->table) {
3937 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3942 for (i = 0; i < 32; ++i)
3943 table->table[i] = IRTE_ALLOCATED;
3946 irq_lookup_table[devid] = table;
3947 set_dte_irq_entry(devid, table);
3948 iommu_flush_dte(iommu, devid);
3949 if (devid != alias) {
3950 irq_lookup_table[alias] = table;
3951 set_dte_irq_entry(alias, table);
3952 iommu_flush_dte(iommu, alias);
3956 iommu_completion_wait(iommu);
3959 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3964 static int alloc_irq_index(u16 devid, int count)
3966 struct irq_remap_table *table;
3967 unsigned long flags;
3970 table = get_irq_table(devid, false);
3974 spin_lock_irqsave(&table->lock, flags);
3976 /* Scan table for free entries */
3977 for (c = 0, index = table->min_index;
3978 index < MAX_IRQS_PER_TABLE;
3980 if (table->table[index] == 0)
3987 table->table[index - c + 1] = IRTE_ALLOCATED;
3997 spin_unlock_irqrestore(&table->lock, flags);
4002 static int modify_irte(u16 devid, int index, union irte irte)
4004 struct irq_remap_table *table;
4005 struct amd_iommu *iommu;
4006 unsigned long flags;
4008 iommu = amd_iommu_rlookup_table[devid];
4012 table = get_irq_table(devid, false);
4016 spin_lock_irqsave(&table->lock, flags);
4017 table->table[index] = irte.val;
4018 spin_unlock_irqrestore(&table->lock, flags);
4020 iommu_flush_irt(iommu, devid);
4021 iommu_completion_wait(iommu);
4026 static void free_irte(u16 devid, int index)
4028 struct irq_remap_table *table;
4029 struct amd_iommu *iommu;
4030 unsigned long flags;
4032 iommu = amd_iommu_rlookup_table[devid];
4036 table = get_irq_table(devid, false);
4040 spin_lock_irqsave(&table->lock, flags);
4041 table->table[index] = 0;
4042 spin_unlock_irqrestore(&table->lock, flags);
4044 iommu_flush_irt(iommu, devid);
4045 iommu_completion_wait(iommu);
4048 static int get_devid(struct irq_alloc_info *info)
4052 switch (info->type) {
4053 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4054 devid = get_ioapic_devid(info->ioapic_id);
4056 case X86_IRQ_ALLOC_TYPE_HPET:
4057 devid = get_hpet_devid(info->hpet_id);
4059 case X86_IRQ_ALLOC_TYPE_MSI:
4060 case X86_IRQ_ALLOC_TYPE_MSIX:
4061 devid = get_device_id(&info->msi_dev->dev);
4071 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4073 struct amd_iommu *iommu;
4079 devid = get_devid(info);
4081 iommu = amd_iommu_rlookup_table[devid];
4083 return iommu->ir_domain;
4089 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4091 struct amd_iommu *iommu;
4097 switch (info->type) {
4098 case X86_IRQ_ALLOC_TYPE_MSI:
4099 case X86_IRQ_ALLOC_TYPE_MSIX:
4100 devid = get_device_id(&info->msi_dev->dev);
4102 iommu = amd_iommu_rlookup_table[devid];
4104 return iommu->msi_domain;
4114 struct irq_remap_ops amd_iommu_irq_ops = {
4115 .prepare = amd_iommu_prepare,
4116 .enable = amd_iommu_enable,
4117 .disable = amd_iommu_disable,
4118 .reenable = amd_iommu_reenable,
4119 .enable_faulting = amd_iommu_enable_faulting,
4120 .get_ir_irq_domain = get_ir_irq_domain,
4121 .get_irq_domain = get_irq_domain,
4124 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4125 struct irq_cfg *irq_cfg,
4126 struct irq_alloc_info *info,
4127 int devid, int index, int sub_handle)
4129 struct irq_2_irte *irte_info = &data->irq_2_irte;
4130 struct msi_msg *msg = &data->msi_entry;
4131 union irte *irte = &data->irte_entry;
4132 struct IO_APIC_route_entry *entry;
4134 data->irq_2_irte.devid = devid;
4135 data->irq_2_irte.index = index + sub_handle;
4137 /* Setup IRTE for IOMMU */
4139 irte->fields.vector = irq_cfg->vector;
4140 irte->fields.int_type = apic->irq_delivery_mode;
4141 irte->fields.destination = irq_cfg->dest_apicid;
4142 irte->fields.dm = apic->irq_dest_mode;
4143 irte->fields.valid = 1;
4145 switch (info->type) {
4146 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4147 /* Setup IOAPIC entry */
4148 entry = info->ioapic_entry;
4149 info->ioapic_entry = NULL;
4150 memset(entry, 0, sizeof(*entry));
4151 entry->vector = index;
4153 entry->trigger = info->ioapic_trigger;
4154 entry->polarity = info->ioapic_polarity;
4155 /* Mask level triggered irqs. */
4156 if (info->ioapic_trigger)
4160 case X86_IRQ_ALLOC_TYPE_HPET:
4161 case X86_IRQ_ALLOC_TYPE_MSI:
4162 case X86_IRQ_ALLOC_TYPE_MSIX:
4163 msg->address_hi = MSI_ADDR_BASE_HI;
4164 msg->address_lo = MSI_ADDR_BASE_LO;
4165 msg->data = irte_info->index;
4174 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4175 unsigned int nr_irqs, void *arg)
4177 struct irq_alloc_info *info = arg;
4178 struct irq_data *irq_data;
4179 struct amd_ir_data *data;
4180 struct irq_cfg *cfg;
4186 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4187 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4191 * With IRQ remapping enabled, don't need contiguous CPU vectors
4192 * to support multiple MSI interrupts.
4194 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4195 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4197 devid = get_devid(info);
4201 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4206 data = kzalloc(sizeof(*data), GFP_KERNEL);
4208 goto out_free_parent;
4210 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4211 if (get_irq_table(devid, true))
4212 index = info->ioapic_pin;
4216 index = alloc_irq_index(devid, nr_irqs);
4219 pr_warn("Failed to allocate IRTE\n");
4221 goto out_free_parent;
4224 for (i = 0; i < nr_irqs; i++) {
4225 irq_data = irq_domain_get_irq_data(domain, virq + i);
4226 cfg = irqd_cfg(irq_data);
4227 if (!irq_data || !cfg) {
4233 data = kzalloc(sizeof(*data), GFP_KERNEL);
4237 irq_data->hwirq = (devid << 16) + i;
4238 irq_data->chip_data = data;
4239 irq_data->chip = &amd_ir_chip;
4240 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4241 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4246 for (i--; i >= 0; i--) {
4247 irq_data = irq_domain_get_irq_data(domain, virq + i);
4249 kfree(irq_data->chip_data);
4251 for (i = 0; i < nr_irqs; i++)
4252 free_irte(devid, index + i);
4254 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4258 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4259 unsigned int nr_irqs)
4261 struct irq_2_irte *irte_info;
4262 struct irq_data *irq_data;
4263 struct amd_ir_data *data;
4266 for (i = 0; i < nr_irqs; i++) {
4267 irq_data = irq_domain_get_irq_data(domain, virq + i);
4268 if (irq_data && irq_data->chip_data) {
4269 data = irq_data->chip_data;
4270 irte_info = &data->irq_2_irte;
4271 free_irte(irte_info->devid, irte_info->index);
4275 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4278 static void irq_remapping_activate(struct irq_domain *domain,
4279 struct irq_data *irq_data)
4281 struct amd_ir_data *data = irq_data->chip_data;
4282 struct irq_2_irte *irte_info = &data->irq_2_irte;
4284 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4287 static void irq_remapping_deactivate(struct irq_domain *domain,
4288 struct irq_data *irq_data)
4290 struct amd_ir_data *data = irq_data->chip_data;
4291 struct irq_2_irte *irte_info = &data->irq_2_irte;
4295 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4298 static struct irq_domain_ops amd_ir_domain_ops = {
4299 .alloc = irq_remapping_alloc,
4300 .free = irq_remapping_free,
4301 .activate = irq_remapping_activate,
4302 .deactivate = irq_remapping_deactivate,
4305 static int amd_ir_set_affinity(struct irq_data *data,
4306 const struct cpumask *mask, bool force)
4308 struct amd_ir_data *ir_data = data->chip_data;
4309 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4310 struct irq_cfg *cfg = irqd_cfg(data);
4311 struct irq_data *parent = data->parent_data;
4314 ret = parent->chip->irq_set_affinity(parent, mask, force);
4315 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4319 * Atomically updates the IRTE with the new destination, vector
4320 * and flushes the interrupt entry cache.
4322 ir_data->irte_entry.fields.vector = cfg->vector;
4323 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4324 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4327 * After this point, all the interrupts will start arriving
4328 * at the new destination. So, time to cleanup the previous
4329 * vector allocation.
4331 send_cleanup_vector(cfg);
4333 return IRQ_SET_MASK_OK_DONE;
4336 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4338 struct amd_ir_data *ir_data = irq_data->chip_data;
4340 *msg = ir_data->msi_entry;
4343 static struct irq_chip amd_ir_chip = {
4344 .irq_ack = ir_ack_apic_edge,
4345 .irq_set_affinity = amd_ir_set_affinity,
4346 .irq_compose_msi_msg = ir_compose_msi_msg,
4349 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4351 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4352 if (!iommu->ir_domain)
4355 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4356 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);