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 <asm/irq_remapping.h>
38 #include <asm/io_apic.h>
40 #include <asm/hw_irq.h>
41 #include <asm/msidef.h>
42 #include <asm/proto.h>
43 #include <asm/iommu.h>
47 #include "amd_iommu_proto.h"
48 #include "amd_iommu_types.h"
49 #include "irq_remapping.h"
51 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
53 #define LOOP_TIMEOUT 100000
56 * This bitmap is used to advertise the page sizes our hardware support
57 * to the IOMMU core, which will then use this information to split
58 * physically contiguous memory regions it is mapping into page sizes
61 * 512GB Pages are not supported due to a hardware bug
63 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
65 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
67 /* A list of preallocated protection domains */
68 static LIST_HEAD(iommu_pd_list);
69 static DEFINE_SPINLOCK(iommu_pd_list_lock);
71 /* List of all available dev_data structures */
72 static LIST_HEAD(dev_data_list);
73 static DEFINE_SPINLOCK(dev_data_list_lock);
75 LIST_HEAD(ioapic_map);
79 * Domain for untranslated devices - only allocated
80 * if iommu=pt passed on kernel cmd line.
82 static struct protection_domain *pt_domain;
84 static const struct iommu_ops amd_iommu_ops;
86 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
87 int amd_iommu_max_glx_val = -1;
89 static struct dma_map_ops amd_iommu_dma_ops;
92 * This struct contains device specific data for the IOMMU
94 struct iommu_dev_data {
95 struct list_head list; /* For domain->dev_list */
96 struct list_head dev_data_list; /* For global dev_data_list */
97 struct list_head alias_list; /* Link alias-groups together */
98 struct iommu_dev_data *alias_data;/* The alias dev_data */
99 struct protection_domain *domain; /* Domain the device is bound to */
100 u16 devid; /* PCI Device ID */
101 bool iommu_v2; /* Device can make use of IOMMUv2 */
102 bool passthrough; /* Default for device is pt_domain */
106 } ats; /* ATS state */
107 bool pri_tlp; /* PASID TLB required for
109 u32 errata; /* Bitmap for errata to apply */
113 * general struct to manage commands send to an IOMMU
119 struct kmem_cache *amd_iommu_irq_cache;
121 static void update_domain(struct protection_domain *domain);
122 static int __init alloc_passthrough_domain(void);
124 /****************************************************************************
128 ****************************************************************************/
130 static struct iommu_dev_data *alloc_dev_data(u16 devid)
132 struct iommu_dev_data *dev_data;
135 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
139 INIT_LIST_HEAD(&dev_data->alias_list);
141 dev_data->devid = devid;
143 spin_lock_irqsave(&dev_data_list_lock, flags);
144 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
145 spin_unlock_irqrestore(&dev_data_list_lock, flags);
150 static void free_dev_data(struct iommu_dev_data *dev_data)
154 spin_lock_irqsave(&dev_data_list_lock, flags);
155 list_del(&dev_data->dev_data_list);
156 spin_unlock_irqrestore(&dev_data_list_lock, flags);
161 static struct iommu_dev_data *search_dev_data(u16 devid)
163 struct iommu_dev_data *dev_data;
166 spin_lock_irqsave(&dev_data_list_lock, flags);
167 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
168 if (dev_data->devid == devid)
175 spin_unlock_irqrestore(&dev_data_list_lock, flags);
180 static struct iommu_dev_data *find_dev_data(u16 devid)
182 struct iommu_dev_data *dev_data;
184 dev_data = search_dev_data(devid);
186 if (dev_data == NULL)
187 dev_data = alloc_dev_data(devid);
192 static inline u16 get_device_id(struct device *dev)
194 struct pci_dev *pdev = to_pci_dev(dev);
196 return PCI_DEVID(pdev->bus->number, pdev->devfn);
199 static struct iommu_dev_data *get_dev_data(struct device *dev)
201 return dev->archdata.iommu;
204 static bool pci_iommuv2_capable(struct pci_dev *pdev)
206 static const int caps[] = {
209 PCI_EXT_CAP_ID_PASID,
213 for (i = 0; i < 3; ++i) {
214 pos = pci_find_ext_capability(pdev, caps[i]);
222 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
224 struct iommu_dev_data *dev_data;
226 dev_data = get_dev_data(&pdev->dev);
228 return dev_data->errata & (1 << erratum) ? true : false;
232 * In this function the list of preallocated protection domains is traversed to
233 * find the domain for a specific device
235 static struct dma_ops_domain *find_protection_domain(u16 devid)
237 struct dma_ops_domain *entry, *ret = NULL;
239 u16 alias = amd_iommu_alias_table[devid];
241 if (list_empty(&iommu_pd_list))
244 spin_lock_irqsave(&iommu_pd_list_lock, flags);
246 list_for_each_entry(entry, &iommu_pd_list, list) {
247 if (entry->target_dev == devid ||
248 entry->target_dev == alias) {
254 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
260 * This function checks if the driver got a valid device from the caller to
261 * avoid dereferencing invalid pointers.
263 static bool check_device(struct device *dev)
267 if (!dev || !dev->dma_mask)
271 if (!dev_is_pci(dev))
274 devid = get_device_id(dev);
276 /* Out of our scope? */
277 if (devid > amd_iommu_last_bdf)
280 if (amd_iommu_rlookup_table[devid] == NULL)
286 static void init_iommu_group(struct device *dev)
288 struct iommu_group *group;
290 group = iommu_group_get_for_dev(dev);
292 iommu_group_put(group);
295 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
297 *(u16 *)data = alias;
301 static u16 get_alias(struct device *dev)
303 struct pci_dev *pdev = to_pci_dev(dev);
304 u16 devid, ivrs_alias, pci_alias;
306 devid = get_device_id(dev);
307 ivrs_alias = amd_iommu_alias_table[devid];
308 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
310 if (ivrs_alias == pci_alias)
316 * The IVRS is fairly reliable in telling us about aliases, but it
317 * can't know about every screwy device. If we don't have an IVRS
318 * reported alias, use the PCI reported alias. In that case we may
319 * still need to initialize the rlookup and dev_table entries if the
320 * alias is to a non-existent device.
322 if (ivrs_alias == devid) {
323 if (!amd_iommu_rlookup_table[pci_alias]) {
324 amd_iommu_rlookup_table[pci_alias] =
325 amd_iommu_rlookup_table[devid];
326 memcpy(amd_iommu_dev_table[pci_alias].data,
327 amd_iommu_dev_table[devid].data,
328 sizeof(amd_iommu_dev_table[pci_alias].data));
334 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
335 "for device %s[%04x:%04x], kernel reported alias "
336 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
337 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
338 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
339 PCI_FUNC(pci_alias));
342 * If we don't have a PCI DMA alias and the IVRS alias is on the same
343 * bus, then the IVRS table may know about a quirk that we don't.
345 if (pci_alias == devid &&
346 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
347 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
348 pdev->dma_alias_devfn = ivrs_alias & 0xff;
349 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
350 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
357 static int iommu_init_device(struct device *dev)
359 struct pci_dev *pdev = to_pci_dev(dev);
360 struct iommu_dev_data *dev_data;
363 if (dev->archdata.iommu)
366 dev_data = find_dev_data(get_device_id(dev));
370 alias = get_alias(dev);
372 if (alias != dev_data->devid) {
373 struct iommu_dev_data *alias_data;
375 alias_data = find_dev_data(alias);
376 if (alias_data == NULL) {
377 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
379 free_dev_data(dev_data);
382 dev_data->alias_data = alias_data;
384 /* Add device to the alias_list */
385 list_add(&dev_data->alias_list, &alias_data->alias_list);
388 if (pci_iommuv2_capable(pdev)) {
389 struct amd_iommu *iommu;
391 iommu = amd_iommu_rlookup_table[dev_data->devid];
392 dev_data->iommu_v2 = iommu->is_iommu_v2;
395 dev->archdata.iommu = dev_data;
397 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
403 static void iommu_ignore_device(struct device *dev)
407 devid = get_device_id(dev);
408 alias = amd_iommu_alias_table[devid];
410 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
411 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
413 amd_iommu_rlookup_table[devid] = NULL;
414 amd_iommu_rlookup_table[alias] = NULL;
417 static void iommu_uninit_device(struct device *dev)
419 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
424 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
427 iommu_group_remove_device(dev);
429 /* Unlink from alias, it may change if another device is re-plugged */
430 dev_data->alias_data = NULL;
433 * We keep dev_data around for unplugged devices and reuse it when the
434 * device is re-plugged - not doing so would introduce a ton of races.
438 void __init amd_iommu_uninit_devices(void)
440 struct iommu_dev_data *dev_data, *n;
441 struct pci_dev *pdev = NULL;
443 for_each_pci_dev(pdev) {
445 if (!check_device(&pdev->dev))
448 iommu_uninit_device(&pdev->dev);
451 /* Free all of our dev_data structures */
452 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
453 free_dev_data(dev_data);
456 int __init amd_iommu_init_devices(void)
458 struct pci_dev *pdev = NULL;
461 for_each_pci_dev(pdev) {
463 if (!check_device(&pdev->dev))
466 ret = iommu_init_device(&pdev->dev);
467 if (ret == -ENOTSUPP)
468 iommu_ignore_device(&pdev->dev);
474 * Initialize IOMMU groups only after iommu_init_device() has
475 * had a chance to populate any IVRS defined aliases.
477 for_each_pci_dev(pdev) {
478 if (check_device(&pdev->dev))
479 init_iommu_group(&pdev->dev);
486 amd_iommu_uninit_devices();
490 #ifdef CONFIG_AMD_IOMMU_STATS
493 * Initialization code for statistics collection
496 DECLARE_STATS_COUNTER(compl_wait);
497 DECLARE_STATS_COUNTER(cnt_map_single);
498 DECLARE_STATS_COUNTER(cnt_unmap_single);
499 DECLARE_STATS_COUNTER(cnt_map_sg);
500 DECLARE_STATS_COUNTER(cnt_unmap_sg);
501 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
502 DECLARE_STATS_COUNTER(cnt_free_coherent);
503 DECLARE_STATS_COUNTER(cross_page);
504 DECLARE_STATS_COUNTER(domain_flush_single);
505 DECLARE_STATS_COUNTER(domain_flush_all);
506 DECLARE_STATS_COUNTER(alloced_io_mem);
507 DECLARE_STATS_COUNTER(total_map_requests);
508 DECLARE_STATS_COUNTER(complete_ppr);
509 DECLARE_STATS_COUNTER(invalidate_iotlb);
510 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
511 DECLARE_STATS_COUNTER(pri_requests);
513 static struct dentry *stats_dir;
514 static struct dentry *de_fflush;
516 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
518 if (stats_dir == NULL)
521 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
525 static void amd_iommu_stats_init(void)
527 stats_dir = debugfs_create_dir("amd-iommu", NULL);
528 if (stats_dir == NULL)
531 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
532 &amd_iommu_unmap_flush);
534 amd_iommu_stats_add(&compl_wait);
535 amd_iommu_stats_add(&cnt_map_single);
536 amd_iommu_stats_add(&cnt_unmap_single);
537 amd_iommu_stats_add(&cnt_map_sg);
538 amd_iommu_stats_add(&cnt_unmap_sg);
539 amd_iommu_stats_add(&cnt_alloc_coherent);
540 amd_iommu_stats_add(&cnt_free_coherent);
541 amd_iommu_stats_add(&cross_page);
542 amd_iommu_stats_add(&domain_flush_single);
543 amd_iommu_stats_add(&domain_flush_all);
544 amd_iommu_stats_add(&alloced_io_mem);
545 amd_iommu_stats_add(&total_map_requests);
546 amd_iommu_stats_add(&complete_ppr);
547 amd_iommu_stats_add(&invalidate_iotlb);
548 amd_iommu_stats_add(&invalidate_iotlb_all);
549 amd_iommu_stats_add(&pri_requests);
554 /****************************************************************************
556 * Interrupt handling functions
558 ****************************************************************************/
560 static void dump_dte_entry(u16 devid)
564 for (i = 0; i < 4; ++i)
565 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
566 amd_iommu_dev_table[devid].data[i]);
569 static void dump_command(unsigned long phys_addr)
571 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
574 for (i = 0; i < 4; ++i)
575 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
578 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
580 int type, devid, domid, flags;
581 volatile u32 *event = __evt;
586 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
587 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
588 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
589 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
590 address = (u64)(((u64)event[3]) << 32) | event[2];
593 /* Did we hit the erratum? */
594 if (++count == LOOP_TIMEOUT) {
595 pr_err("AMD-Vi: No event written to event log\n");
602 printk(KERN_ERR "AMD-Vi: Event logged [");
605 case EVENT_TYPE_ILL_DEV:
606 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
607 "address=0x%016llx flags=0x%04x]\n",
608 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
610 dump_dte_entry(devid);
612 case EVENT_TYPE_IO_FAULT:
613 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
614 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
615 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 domid, address, flags);
618 case EVENT_TYPE_DEV_TAB_ERR:
619 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
620 "address=0x%016llx flags=0x%04x]\n",
621 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
624 case EVENT_TYPE_PAGE_TAB_ERR:
625 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
626 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
627 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
628 domid, address, flags);
630 case EVENT_TYPE_ILL_CMD:
631 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
632 dump_command(address);
634 case EVENT_TYPE_CMD_HARD_ERR:
635 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
636 "flags=0x%04x]\n", address, flags);
638 case EVENT_TYPE_IOTLB_INV_TO:
639 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
640 "address=0x%016llx]\n",
641 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
644 case EVENT_TYPE_INV_DEV_REQ:
645 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
646 "address=0x%016llx flags=0x%04x]\n",
647 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
651 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
654 memset(__evt, 0, 4 * sizeof(u32));
657 static void iommu_poll_events(struct amd_iommu *iommu)
661 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
662 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
664 while (head != tail) {
665 iommu_print_event(iommu, iommu->evt_buf + head);
666 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
669 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
672 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
674 struct amd_iommu_fault fault;
676 INC_STATS_COUNTER(pri_requests);
678 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
679 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
683 fault.address = raw[1];
684 fault.pasid = PPR_PASID(raw[0]);
685 fault.device_id = PPR_DEVID(raw[0]);
686 fault.tag = PPR_TAG(raw[0]);
687 fault.flags = PPR_FLAGS(raw[0]);
689 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
692 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
696 if (iommu->ppr_log == NULL)
699 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
700 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
702 while (head != tail) {
707 raw = (u64 *)(iommu->ppr_log + head);
710 * Hardware bug: Interrupt may arrive before the entry is
711 * written to memory. If this happens we need to wait for the
714 for (i = 0; i < LOOP_TIMEOUT; ++i) {
715 if (PPR_REQ_TYPE(raw[0]) != 0)
720 /* Avoid memcpy function-call overhead */
725 * To detect the hardware bug we need to clear the entry
728 raw[0] = raw[1] = 0UL;
730 /* Update head pointer of hardware ring-buffer */
731 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
732 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
734 /* Handle PPR entry */
735 iommu_handle_ppr_entry(iommu, entry);
737 /* Refresh ring-buffer information */
738 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
739 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
743 irqreturn_t amd_iommu_int_thread(int irq, void *data)
745 struct amd_iommu *iommu = (struct amd_iommu *) data;
746 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
748 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
749 /* Enable EVT and PPR interrupts again */
750 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
751 iommu->mmio_base + MMIO_STATUS_OFFSET);
753 if (status & MMIO_STATUS_EVT_INT_MASK) {
754 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
755 iommu_poll_events(iommu);
758 if (status & MMIO_STATUS_PPR_INT_MASK) {
759 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
760 iommu_poll_ppr_log(iommu);
764 * Hardware bug: ERBT1312
765 * When re-enabling interrupt (by writing 1
766 * to clear the bit), the hardware might also try to set
767 * the interrupt bit in the event status register.
768 * In this scenario, the bit will be set, and disable
769 * subsequent interrupts.
771 * Workaround: The IOMMU driver should read back the
772 * status register and check if the interrupt bits are cleared.
773 * If not, driver will need to go through the interrupt handler
774 * again and re-clear the bits
776 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
781 irqreturn_t amd_iommu_int_handler(int irq, void *data)
783 return IRQ_WAKE_THREAD;
786 /****************************************************************************
788 * IOMMU command queuing functions
790 ****************************************************************************/
792 static int wait_on_sem(volatile u64 *sem)
796 while (*sem == 0 && i < LOOP_TIMEOUT) {
801 if (i == LOOP_TIMEOUT) {
802 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
809 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
810 struct iommu_cmd *cmd,
815 target = iommu->cmd_buf + tail;
816 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
818 /* Copy command to buffer */
819 memcpy(target, cmd, sizeof(*cmd));
821 /* Tell the IOMMU about it */
822 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
825 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
827 WARN_ON(address & 0x7ULL);
829 memset(cmd, 0, sizeof(*cmd));
830 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
831 cmd->data[1] = upper_32_bits(__pa(address));
833 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
836 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
838 memset(cmd, 0, sizeof(*cmd));
839 cmd->data[0] = devid;
840 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
843 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
844 size_t size, u16 domid, int pde)
849 pages = iommu_num_pages(address, size, PAGE_SIZE);
854 * If we have to flush more than one page, flush all
855 * TLB entries for this domain
857 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
861 address &= PAGE_MASK;
863 memset(cmd, 0, sizeof(*cmd));
864 cmd->data[1] |= domid;
865 cmd->data[2] = lower_32_bits(address);
866 cmd->data[3] = upper_32_bits(address);
867 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
868 if (s) /* size bit - we flush more than one 4kb page */
869 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
870 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
871 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
874 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
875 u64 address, size_t size)
880 pages = iommu_num_pages(address, size, PAGE_SIZE);
885 * If we have to flush more than one page, flush all
886 * TLB entries for this domain
888 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
892 address &= PAGE_MASK;
894 memset(cmd, 0, sizeof(*cmd));
895 cmd->data[0] = devid;
896 cmd->data[0] |= (qdep & 0xff) << 24;
897 cmd->data[1] = devid;
898 cmd->data[2] = lower_32_bits(address);
899 cmd->data[3] = upper_32_bits(address);
900 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
902 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
905 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
906 u64 address, bool size)
908 memset(cmd, 0, sizeof(*cmd));
910 address &= ~(0xfffULL);
912 cmd->data[0] = pasid;
913 cmd->data[1] = domid;
914 cmd->data[2] = lower_32_bits(address);
915 cmd->data[3] = upper_32_bits(address);
916 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
917 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
919 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
920 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
923 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
924 int qdep, u64 address, bool size)
926 memset(cmd, 0, sizeof(*cmd));
928 address &= ~(0xfffULL);
930 cmd->data[0] = devid;
931 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
932 cmd->data[0] |= (qdep & 0xff) << 24;
933 cmd->data[1] = devid;
934 cmd->data[1] |= (pasid & 0xff) << 16;
935 cmd->data[2] = lower_32_bits(address);
936 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
937 cmd->data[3] = upper_32_bits(address);
939 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
940 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
943 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
944 int status, int tag, bool gn)
946 memset(cmd, 0, sizeof(*cmd));
948 cmd->data[0] = devid;
950 cmd->data[1] = pasid;
951 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
953 cmd->data[3] = tag & 0x1ff;
954 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
956 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
959 static void build_inv_all(struct iommu_cmd *cmd)
961 memset(cmd, 0, sizeof(*cmd));
962 CMD_SET_TYPE(cmd, CMD_INV_ALL);
965 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
967 memset(cmd, 0, sizeof(*cmd));
968 cmd->data[0] = devid;
969 CMD_SET_TYPE(cmd, CMD_INV_IRT);
973 * Writes the command to the IOMMUs command buffer and informs the
974 * hardware about the new command.
976 static int iommu_queue_command_sync(struct amd_iommu *iommu,
977 struct iommu_cmd *cmd,
980 u32 left, tail, head, next_tail;
983 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
986 spin_lock_irqsave(&iommu->lock, flags);
988 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
989 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
990 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
991 left = (head - next_tail) % iommu->cmd_buf_size;
994 struct iommu_cmd sync_cmd;
995 volatile u64 sem = 0;
998 build_completion_wait(&sync_cmd, (u64)&sem);
999 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1001 spin_unlock_irqrestore(&iommu->lock, flags);
1003 if ((ret = wait_on_sem(&sem)) != 0)
1009 copy_cmd_to_buffer(iommu, cmd, tail);
1011 /* We need to sync now to make sure all commands are processed */
1012 iommu->need_sync = sync;
1014 spin_unlock_irqrestore(&iommu->lock, flags);
1019 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1021 return iommu_queue_command_sync(iommu, cmd, true);
1025 * This function queues a completion wait command into the command
1026 * buffer of an IOMMU
1028 static int iommu_completion_wait(struct amd_iommu *iommu)
1030 struct iommu_cmd cmd;
1031 volatile u64 sem = 0;
1034 if (!iommu->need_sync)
1037 build_completion_wait(&cmd, (u64)&sem);
1039 ret = iommu_queue_command_sync(iommu, &cmd, false);
1043 return wait_on_sem(&sem);
1046 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1048 struct iommu_cmd cmd;
1050 build_inv_dte(&cmd, devid);
1052 return iommu_queue_command(iommu, &cmd);
1055 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1059 for (devid = 0; devid <= 0xffff; ++devid)
1060 iommu_flush_dte(iommu, devid);
1062 iommu_completion_wait(iommu);
1066 * This function uses heavy locking and may disable irqs for some time. But
1067 * this is no issue because it is only called during resume.
1069 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1073 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1074 struct iommu_cmd cmd;
1075 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1077 iommu_queue_command(iommu, &cmd);
1080 iommu_completion_wait(iommu);
1083 static void iommu_flush_all(struct amd_iommu *iommu)
1085 struct iommu_cmd cmd;
1087 build_inv_all(&cmd);
1089 iommu_queue_command(iommu, &cmd);
1090 iommu_completion_wait(iommu);
1093 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1095 struct iommu_cmd cmd;
1097 build_inv_irt(&cmd, devid);
1099 iommu_queue_command(iommu, &cmd);
1102 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1106 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1107 iommu_flush_irt(iommu, devid);
1109 iommu_completion_wait(iommu);
1112 void iommu_flush_all_caches(struct amd_iommu *iommu)
1114 if (iommu_feature(iommu, FEATURE_IA)) {
1115 iommu_flush_all(iommu);
1117 iommu_flush_dte_all(iommu);
1118 iommu_flush_irt_all(iommu);
1119 iommu_flush_tlb_all(iommu);
1124 * Command send function for flushing on-device TLB
1126 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1127 u64 address, size_t size)
1129 struct amd_iommu *iommu;
1130 struct iommu_cmd cmd;
1133 qdep = dev_data->ats.qdep;
1134 iommu = amd_iommu_rlookup_table[dev_data->devid];
1136 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1138 return iommu_queue_command(iommu, &cmd);
1142 * Command send function for invalidating a device table entry
1144 static int device_flush_dte(struct iommu_dev_data *dev_data)
1146 struct amd_iommu *iommu;
1149 iommu = amd_iommu_rlookup_table[dev_data->devid];
1151 ret = iommu_flush_dte(iommu, dev_data->devid);
1155 if (dev_data->ats.enabled)
1156 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1162 * TLB invalidation function which is called from the mapping functions.
1163 * It invalidates a single PTE if the range to flush is within a single
1164 * page. Otherwise it flushes the whole TLB of the IOMMU.
1166 static void __domain_flush_pages(struct protection_domain *domain,
1167 u64 address, size_t size, int pde)
1169 struct iommu_dev_data *dev_data;
1170 struct iommu_cmd cmd;
1173 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1175 for (i = 0; i < amd_iommus_present; ++i) {
1176 if (!domain->dev_iommu[i])
1180 * Devices of this domain are behind this IOMMU
1181 * We need a TLB flush
1183 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1186 list_for_each_entry(dev_data, &domain->dev_list, list) {
1188 if (!dev_data->ats.enabled)
1191 ret |= device_flush_iotlb(dev_data, address, size);
1197 static void domain_flush_pages(struct protection_domain *domain,
1198 u64 address, size_t size)
1200 __domain_flush_pages(domain, address, size, 0);
1203 /* Flush the whole IO/TLB for a given protection domain */
1204 static void domain_flush_tlb(struct protection_domain *domain)
1206 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1209 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1210 static void domain_flush_tlb_pde(struct protection_domain *domain)
1212 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1215 static void domain_flush_complete(struct protection_domain *domain)
1219 for (i = 0; i < amd_iommus_present; ++i) {
1220 if (!domain->dev_iommu[i])
1224 * Devices of this domain are behind this IOMMU
1225 * We need to wait for completion of all commands.
1227 iommu_completion_wait(amd_iommus[i]);
1233 * This function flushes the DTEs for all devices in domain
1235 static void domain_flush_devices(struct protection_domain *domain)
1237 struct iommu_dev_data *dev_data;
1239 list_for_each_entry(dev_data, &domain->dev_list, list)
1240 device_flush_dte(dev_data);
1243 /****************************************************************************
1245 * The functions below are used the create the page table mappings for
1246 * unity mapped regions.
1248 ****************************************************************************/
1251 * This function is used to add another level to an IO page table. Adding
1252 * another level increases the size of the address space by 9 bits to a size up
1255 static bool increase_address_space(struct protection_domain *domain,
1260 if (domain->mode == PAGE_MODE_6_LEVEL)
1261 /* address space already 64 bit large */
1264 pte = (void *)get_zeroed_page(gfp);
1268 *pte = PM_LEVEL_PDE(domain->mode,
1269 virt_to_phys(domain->pt_root));
1270 domain->pt_root = pte;
1272 domain->updated = true;
1277 static u64 *alloc_pte(struct protection_domain *domain,
1278 unsigned long address,
1279 unsigned long page_size,
1286 BUG_ON(!is_power_of_2(page_size));
1288 while (address > PM_LEVEL_SIZE(domain->mode))
1289 increase_address_space(domain, gfp);
1291 level = domain->mode - 1;
1292 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1293 address = PAGE_SIZE_ALIGN(address, page_size);
1294 end_lvl = PAGE_SIZE_LEVEL(page_size);
1296 while (level > end_lvl) {
1297 if (!IOMMU_PTE_PRESENT(*pte)) {
1298 page = (u64 *)get_zeroed_page(gfp);
1301 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1304 /* No level skipping support yet */
1305 if (PM_PTE_LEVEL(*pte) != level)
1310 pte = IOMMU_PTE_PAGE(*pte);
1312 if (pte_page && level == end_lvl)
1315 pte = &pte[PM_LEVEL_INDEX(level, address)];
1322 * This function checks if there is a PTE for a given dma address. If
1323 * there is one, it returns the pointer to it.
1325 static u64 *fetch_pte(struct protection_domain *domain,
1326 unsigned long address,
1327 unsigned long *page_size)
1332 if (address > PM_LEVEL_SIZE(domain->mode))
1335 level = domain->mode - 1;
1336 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1337 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1342 if (!IOMMU_PTE_PRESENT(*pte))
1346 if (PM_PTE_LEVEL(*pte) == 7 ||
1347 PM_PTE_LEVEL(*pte) == 0)
1350 /* No level skipping support yet */
1351 if (PM_PTE_LEVEL(*pte) != level)
1356 /* Walk to the next level */
1357 pte = IOMMU_PTE_PAGE(*pte);
1358 pte = &pte[PM_LEVEL_INDEX(level, address)];
1359 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1362 if (PM_PTE_LEVEL(*pte) == 0x07) {
1363 unsigned long pte_mask;
1366 * If we have a series of large PTEs, make
1367 * sure to return a pointer to the first one.
1369 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1370 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1371 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1378 * Generic mapping functions. It maps a physical address into a DMA
1379 * address space. It allocates the page table pages if necessary.
1380 * In the future it can be extended to a generic mapping function
1381 * supporting all features of AMD IOMMU page tables like level skipping
1382 * and full 64 bit address spaces.
1384 static int iommu_map_page(struct protection_domain *dom,
1385 unsigned long bus_addr,
1386 unsigned long phys_addr,
1388 unsigned long page_size)
1393 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1394 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1396 if (!(prot & IOMMU_PROT_MASK))
1399 count = PAGE_SIZE_PTE_COUNT(page_size);
1400 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1405 for (i = 0; i < count; ++i)
1406 if (IOMMU_PTE_PRESENT(pte[i]))
1410 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1411 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1413 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1415 if (prot & IOMMU_PROT_IR)
1416 __pte |= IOMMU_PTE_IR;
1417 if (prot & IOMMU_PROT_IW)
1418 __pte |= IOMMU_PTE_IW;
1420 for (i = 0; i < count; ++i)
1428 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1429 unsigned long bus_addr,
1430 unsigned long page_size)
1432 unsigned long long unmapped;
1433 unsigned long unmap_size;
1436 BUG_ON(!is_power_of_2(page_size));
1440 while (unmapped < page_size) {
1442 pte = fetch_pte(dom, bus_addr, &unmap_size);
1447 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1448 for (i = 0; i < count; i++)
1452 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1453 unmapped += unmap_size;
1456 BUG_ON(unmapped && !is_power_of_2(unmapped));
1462 * This function checks if a specific unity mapping entry is needed for
1463 * this specific IOMMU.
1465 static int iommu_for_unity_map(struct amd_iommu *iommu,
1466 struct unity_map_entry *entry)
1470 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1471 bdf = amd_iommu_alias_table[i];
1472 if (amd_iommu_rlookup_table[bdf] == iommu)
1480 * This function actually applies the mapping to the page table of the
1483 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1484 struct unity_map_entry *e)
1489 for (addr = e->address_start; addr < e->address_end;
1490 addr += PAGE_SIZE) {
1491 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1496 * if unity mapping is in aperture range mark the page
1497 * as allocated in the aperture
1499 if (addr < dma_dom->aperture_size)
1500 __set_bit(addr >> PAGE_SHIFT,
1501 dma_dom->aperture[0]->bitmap);
1508 * Init the unity mappings for a specific IOMMU in the system
1510 * Basically iterates over all unity mapping entries and applies them to
1511 * the default domain DMA of that IOMMU if necessary.
1513 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1515 struct unity_map_entry *entry;
1518 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1519 if (!iommu_for_unity_map(iommu, entry))
1521 ret = dma_ops_unity_map(iommu->default_dom, entry);
1530 * Inits the unity mappings required for a specific device
1532 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1535 struct unity_map_entry *e;
1538 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1539 if (!(devid >= e->devid_start && devid <= e->devid_end))
1541 ret = dma_ops_unity_map(dma_dom, e);
1549 /****************************************************************************
1551 * The next functions belong to the address allocator for the dma_ops
1552 * interface functions. They work like the allocators in the other IOMMU
1553 * drivers. Its basically a bitmap which marks the allocated pages in
1554 * the aperture. Maybe it could be enhanced in the future to a more
1555 * efficient allocator.
1557 ****************************************************************************/
1560 * The address allocator core functions.
1562 * called with domain->lock held
1566 * Used to reserve address ranges in the aperture (e.g. for exclusion
1569 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1570 unsigned long start_page,
1573 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1575 if (start_page + pages > last_page)
1576 pages = last_page - start_page;
1578 for (i = start_page; i < start_page + pages; ++i) {
1579 int index = i / APERTURE_RANGE_PAGES;
1580 int page = i % APERTURE_RANGE_PAGES;
1581 __set_bit(page, dom->aperture[index]->bitmap);
1586 * This function is used to add a new aperture range to an existing
1587 * aperture in case of dma_ops domain allocation or address allocation
1590 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1591 bool populate, gfp_t gfp)
1593 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1594 struct amd_iommu *iommu;
1595 unsigned long i, old_size, pte_pgsize;
1597 #ifdef CONFIG_IOMMU_STRESS
1601 if (index >= APERTURE_MAX_RANGES)
1604 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1605 if (!dma_dom->aperture[index])
1608 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1609 if (!dma_dom->aperture[index]->bitmap)
1612 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1615 unsigned long address = dma_dom->aperture_size;
1616 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1617 u64 *pte, *pte_page;
1619 for (i = 0; i < num_ptes; ++i) {
1620 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1625 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1627 address += APERTURE_RANGE_SIZE / 64;
1631 old_size = dma_dom->aperture_size;
1632 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1634 /* Reserve address range used for MSI messages */
1635 if (old_size < MSI_ADDR_BASE_LO &&
1636 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1637 unsigned long spage;
1640 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1641 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1643 dma_ops_reserve_addresses(dma_dom, spage, pages);
1646 /* Initialize the exclusion range if necessary */
1647 for_each_iommu(iommu) {
1648 if (iommu->exclusion_start &&
1649 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1650 && iommu->exclusion_start < dma_dom->aperture_size) {
1651 unsigned long startpage;
1652 int pages = iommu_num_pages(iommu->exclusion_start,
1653 iommu->exclusion_length,
1655 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1656 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1661 * Check for areas already mapped as present in the new aperture
1662 * range and mark those pages as reserved in the allocator. Such
1663 * mappings may already exist as a result of requested unity
1664 * mappings for devices.
1666 for (i = dma_dom->aperture[index]->offset;
1667 i < dma_dom->aperture_size;
1669 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1670 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1673 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1677 update_domain(&dma_dom->domain);
1682 update_domain(&dma_dom->domain);
1684 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1686 kfree(dma_dom->aperture[index]);
1687 dma_dom->aperture[index] = NULL;
1692 static unsigned long dma_ops_area_alloc(struct device *dev,
1693 struct dma_ops_domain *dom,
1695 unsigned long align_mask,
1697 unsigned long start)
1699 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1700 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1701 int i = start >> APERTURE_RANGE_SHIFT;
1702 unsigned long boundary_size;
1703 unsigned long address = -1;
1704 unsigned long limit;
1706 next_bit >>= PAGE_SHIFT;
1708 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1709 PAGE_SIZE) >> PAGE_SHIFT;
1711 for (;i < max_index; ++i) {
1712 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1714 if (dom->aperture[i]->offset >= dma_mask)
1717 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1718 dma_mask >> PAGE_SHIFT);
1720 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1721 limit, next_bit, pages, 0,
1722 boundary_size, align_mask);
1723 if (address != -1) {
1724 address = dom->aperture[i]->offset +
1725 (address << PAGE_SHIFT);
1726 dom->next_address = address + (pages << PAGE_SHIFT);
1736 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1737 struct dma_ops_domain *dom,
1739 unsigned long align_mask,
1742 unsigned long address;
1744 #ifdef CONFIG_IOMMU_STRESS
1745 dom->next_address = 0;
1746 dom->need_flush = true;
1749 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1750 dma_mask, dom->next_address);
1752 if (address == -1) {
1753 dom->next_address = 0;
1754 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1756 dom->need_flush = true;
1759 if (unlikely(address == -1))
1760 address = DMA_ERROR_CODE;
1762 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1768 * The address free function.
1770 * called with domain->lock held
1772 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1773 unsigned long address,
1776 unsigned i = address >> APERTURE_RANGE_SHIFT;
1777 struct aperture_range *range = dom->aperture[i];
1779 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1781 #ifdef CONFIG_IOMMU_STRESS
1786 if (address >= dom->next_address)
1787 dom->need_flush = true;
1789 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1791 bitmap_clear(range->bitmap, address, pages);
1795 /****************************************************************************
1797 * The next functions belong to the domain allocation. A domain is
1798 * allocated for every IOMMU as the default domain. If device isolation
1799 * is enabled, every device get its own domain. The most important thing
1800 * about domains is the page table mapping the DMA address space they
1803 ****************************************************************************/
1806 * This function adds a protection domain to the global protection domain list
1808 static void add_domain_to_list(struct protection_domain *domain)
1810 unsigned long flags;
1812 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1813 list_add(&domain->list, &amd_iommu_pd_list);
1814 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1818 * This function removes a protection domain to the global
1819 * protection domain list
1821 static void del_domain_from_list(struct protection_domain *domain)
1823 unsigned long flags;
1825 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1826 list_del(&domain->list);
1827 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1830 static u16 domain_id_alloc(void)
1832 unsigned long flags;
1835 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1836 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1838 if (id > 0 && id < MAX_DOMAIN_ID)
1839 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1842 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1847 static void domain_id_free(int id)
1849 unsigned long flags;
1851 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1852 if (id > 0 && id < MAX_DOMAIN_ID)
1853 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1854 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1857 #define DEFINE_FREE_PT_FN(LVL, FN) \
1858 static void free_pt_##LVL (unsigned long __pt) \
1866 for (i = 0; i < 512; ++i) { \
1867 if (!IOMMU_PTE_PRESENT(pt[i])) \
1870 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1873 free_page((unsigned long)pt); \
1876 DEFINE_FREE_PT_FN(l2, free_page)
1877 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1878 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1879 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1880 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1882 static void free_pagetable(struct protection_domain *domain)
1884 unsigned long root = (unsigned long)domain->pt_root;
1886 switch (domain->mode) {
1887 case PAGE_MODE_NONE:
1889 case PAGE_MODE_1_LEVEL:
1892 case PAGE_MODE_2_LEVEL:
1895 case PAGE_MODE_3_LEVEL:
1898 case PAGE_MODE_4_LEVEL:
1901 case PAGE_MODE_5_LEVEL:
1904 case PAGE_MODE_6_LEVEL:
1912 static void free_gcr3_tbl_level1(u64 *tbl)
1917 for (i = 0; i < 512; ++i) {
1918 if (!(tbl[i] & GCR3_VALID))
1921 ptr = __va(tbl[i] & PAGE_MASK);
1923 free_page((unsigned long)ptr);
1927 static void free_gcr3_tbl_level2(u64 *tbl)
1932 for (i = 0; i < 512; ++i) {
1933 if (!(tbl[i] & GCR3_VALID))
1936 ptr = __va(tbl[i] & PAGE_MASK);
1938 free_gcr3_tbl_level1(ptr);
1942 static void free_gcr3_table(struct protection_domain *domain)
1944 if (domain->glx == 2)
1945 free_gcr3_tbl_level2(domain->gcr3_tbl);
1946 else if (domain->glx == 1)
1947 free_gcr3_tbl_level1(domain->gcr3_tbl);
1948 else if (domain->glx != 0)
1951 free_page((unsigned long)domain->gcr3_tbl);
1955 * Free a domain, only used if something went wrong in the
1956 * allocation path and we need to free an already allocated page table
1958 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1965 del_domain_from_list(&dom->domain);
1967 free_pagetable(&dom->domain);
1969 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1970 if (!dom->aperture[i])
1972 free_page((unsigned long)dom->aperture[i]->bitmap);
1973 kfree(dom->aperture[i]);
1980 * Allocates a new protection domain usable for the dma_ops functions.
1981 * It also initializes the page table and the address allocator data
1982 * structures required for the dma_ops interface
1984 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1986 struct dma_ops_domain *dma_dom;
1988 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1992 spin_lock_init(&dma_dom->domain.lock);
1994 dma_dom->domain.id = domain_id_alloc();
1995 if (dma_dom->domain.id == 0)
1997 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1998 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1999 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2000 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2001 dma_dom->domain.priv = dma_dom;
2002 if (!dma_dom->domain.pt_root)
2005 dma_dom->need_flush = false;
2006 dma_dom->target_dev = 0xffff;
2008 add_domain_to_list(&dma_dom->domain);
2010 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
2014 * mark the first page as allocated so we never return 0 as
2015 * a valid dma-address. So we can use 0 as error value
2017 dma_dom->aperture[0]->bitmap[0] = 1;
2018 dma_dom->next_address = 0;
2024 dma_ops_domain_free(dma_dom);
2030 * little helper function to check whether a given protection domain is a
2033 static bool dma_ops_domain(struct protection_domain *domain)
2035 return domain->flags & PD_DMA_OPS_MASK;
2038 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2043 if (domain->mode != PAGE_MODE_NONE)
2044 pte_root = virt_to_phys(domain->pt_root);
2046 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2047 << DEV_ENTRY_MODE_SHIFT;
2048 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2050 flags = amd_iommu_dev_table[devid].data[1];
2053 flags |= DTE_FLAG_IOTLB;
2055 if (domain->flags & PD_IOMMUV2_MASK) {
2056 u64 gcr3 = __pa(domain->gcr3_tbl);
2057 u64 glx = domain->glx;
2060 pte_root |= DTE_FLAG_GV;
2061 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2063 /* First mask out possible old values for GCR3 table */
2064 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2067 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2070 /* Encode GCR3 table into DTE */
2071 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2074 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2077 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2081 flags &= ~(0xffffUL);
2082 flags |= domain->id;
2084 amd_iommu_dev_table[devid].data[1] = flags;
2085 amd_iommu_dev_table[devid].data[0] = pte_root;
2088 static void clear_dte_entry(u16 devid)
2090 /* remove entry from the device table seen by the hardware */
2091 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2092 amd_iommu_dev_table[devid].data[1] = 0;
2094 amd_iommu_apply_erratum_63(devid);
2097 static void do_attach(struct iommu_dev_data *dev_data,
2098 struct protection_domain *domain)
2100 struct amd_iommu *iommu;
2103 iommu = amd_iommu_rlookup_table[dev_data->devid];
2104 ats = dev_data->ats.enabled;
2106 /* Update data structures */
2107 dev_data->domain = domain;
2108 list_add(&dev_data->list, &domain->dev_list);
2109 set_dte_entry(dev_data->devid, domain, ats);
2111 /* Do reference counting */
2112 domain->dev_iommu[iommu->index] += 1;
2113 domain->dev_cnt += 1;
2115 /* Flush the DTE entry */
2116 device_flush_dte(dev_data);
2119 static void do_detach(struct iommu_dev_data *dev_data)
2121 struct amd_iommu *iommu;
2123 iommu = amd_iommu_rlookup_table[dev_data->devid];
2125 /* decrease reference counters */
2126 dev_data->domain->dev_iommu[iommu->index] -= 1;
2127 dev_data->domain->dev_cnt -= 1;
2129 /* Update data structures */
2130 dev_data->domain = NULL;
2131 list_del(&dev_data->list);
2132 clear_dte_entry(dev_data->devid);
2134 /* Flush the DTE entry */
2135 device_flush_dte(dev_data);
2139 * If a device is not yet associated with a domain, this function does
2140 * assigns it visible for the hardware
2142 static int __attach_device(struct iommu_dev_data *dev_data,
2143 struct protection_domain *domain)
2145 struct iommu_dev_data *head, *entry;
2149 spin_lock(&domain->lock);
2153 if (head->alias_data != NULL)
2154 head = head->alias_data;
2156 /* Now we have the root of the alias group, if any */
2159 if (head->domain != NULL)
2162 /* Attach alias group root */
2163 do_attach(head, domain);
2165 /* Attach other devices in the alias group */
2166 list_for_each_entry(entry, &head->alias_list, alias_list)
2167 do_attach(entry, domain);
2174 spin_unlock(&domain->lock);
2180 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2182 pci_disable_ats(pdev);
2183 pci_disable_pri(pdev);
2184 pci_disable_pasid(pdev);
2187 /* FIXME: Change generic reset-function to do the same */
2188 static int pri_reset_while_enabled(struct pci_dev *pdev)
2193 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2197 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2198 control |= PCI_PRI_CTRL_RESET;
2199 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2204 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2209 /* FIXME: Hardcode number of outstanding requests for now */
2211 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2213 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2215 /* Only allow access to user-accessible pages */
2216 ret = pci_enable_pasid(pdev, 0);
2220 /* First reset the PRI state of the device */
2221 ret = pci_reset_pri(pdev);
2226 ret = pci_enable_pri(pdev, reqs);
2231 ret = pri_reset_while_enabled(pdev);
2236 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2243 pci_disable_pri(pdev);
2244 pci_disable_pasid(pdev);
2249 /* FIXME: Move this to PCI code */
2250 #define PCI_PRI_TLP_OFF (1 << 15)
2252 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2257 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2261 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2263 return (status & PCI_PRI_TLP_OFF) ? true : false;
2267 * If a device is not yet associated with a domain, this function
2268 * assigns it visible for the hardware
2270 static int attach_device(struct device *dev,
2271 struct protection_domain *domain)
2273 struct pci_dev *pdev = to_pci_dev(dev);
2274 struct iommu_dev_data *dev_data;
2275 unsigned long flags;
2278 dev_data = get_dev_data(dev);
2280 if (domain->flags & PD_IOMMUV2_MASK) {
2281 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2284 if (pdev_iommuv2_enable(pdev) != 0)
2287 dev_data->ats.enabled = true;
2288 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2289 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2290 } else if (amd_iommu_iotlb_sup &&
2291 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2292 dev_data->ats.enabled = true;
2293 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2296 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2297 ret = __attach_device(dev_data, domain);
2298 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2301 * We might boot into a crash-kernel here. The crashed kernel
2302 * left the caches in the IOMMU dirty. So we have to flush
2303 * here to evict all dirty stuff.
2305 domain_flush_tlb_pde(domain);
2311 * Removes a device from a protection domain (unlocked)
2313 static void __detach_device(struct iommu_dev_data *dev_data)
2315 struct iommu_dev_data *head, *entry;
2316 struct protection_domain *domain;
2317 unsigned long flags;
2319 BUG_ON(!dev_data->domain);
2321 domain = dev_data->domain;
2323 spin_lock_irqsave(&domain->lock, flags);
2326 if (head->alias_data != NULL)
2327 head = head->alias_data;
2329 list_for_each_entry(entry, &head->alias_list, alias_list)
2334 spin_unlock_irqrestore(&domain->lock, flags);
2337 * If we run in passthrough mode the device must be assigned to the
2338 * passthrough domain if it is detached from any other domain.
2339 * Make sure we can deassign from the pt_domain itself.
2341 if (dev_data->passthrough &&
2342 (dev_data->domain == NULL && domain != pt_domain))
2343 __attach_device(dev_data, pt_domain);
2347 * Removes a device from a protection domain (with devtable_lock held)
2349 static void detach_device(struct device *dev)
2351 struct protection_domain *domain;
2352 struct iommu_dev_data *dev_data;
2353 unsigned long flags;
2355 dev_data = get_dev_data(dev);
2356 domain = dev_data->domain;
2358 /* lock device table */
2359 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2360 __detach_device(dev_data);
2361 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2363 if (domain->flags & PD_IOMMUV2_MASK)
2364 pdev_iommuv2_disable(to_pci_dev(dev));
2365 else if (dev_data->ats.enabled)
2366 pci_disable_ats(to_pci_dev(dev));
2368 dev_data->ats.enabled = false;
2372 * Find out the protection domain structure for a given PCI device. This
2373 * will give us the pointer to the page table root for example.
2375 static struct protection_domain *domain_for_device(struct device *dev)
2377 struct iommu_dev_data *dev_data;
2378 struct protection_domain *dom = NULL;
2379 unsigned long flags;
2381 dev_data = get_dev_data(dev);
2383 if (dev_data->domain)
2384 return dev_data->domain;
2386 if (dev_data->alias_data != NULL) {
2387 struct iommu_dev_data *alias_data = dev_data->alias_data;
2389 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2390 if (alias_data->domain != NULL) {
2391 __attach_device(dev_data, alias_data->domain);
2392 dom = alias_data->domain;
2394 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2400 static int device_change_notifier(struct notifier_block *nb,
2401 unsigned long action, void *data)
2403 struct dma_ops_domain *dma_domain;
2404 struct protection_domain *domain;
2405 struct iommu_dev_data *dev_data;
2406 struct device *dev = data;
2407 struct amd_iommu *iommu;
2408 unsigned long flags;
2411 if (!check_device(dev))
2414 devid = get_device_id(dev);
2415 iommu = amd_iommu_rlookup_table[devid];
2416 dev_data = get_dev_data(dev);
2419 case BUS_NOTIFY_ADD_DEVICE:
2421 iommu_init_device(dev);
2422 init_iommu_group(dev);
2425 * dev_data is still NULL and
2426 * got initialized in iommu_init_device
2428 dev_data = get_dev_data(dev);
2430 if (iommu_pass_through || dev_data->iommu_v2) {
2431 dev_data->passthrough = true;
2432 attach_device(dev, pt_domain);
2436 domain = domain_for_device(dev);
2438 /* allocate a protection domain if a device is added */
2439 dma_domain = find_protection_domain(devid);
2441 dma_domain = dma_ops_domain_alloc();
2444 dma_domain->target_dev = devid;
2446 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2447 list_add_tail(&dma_domain->list, &iommu_pd_list);
2448 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2451 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2454 case BUS_NOTIFY_REMOVED_DEVICE:
2456 iommu_uninit_device(dev);
2462 iommu_completion_wait(iommu);
2468 static struct notifier_block device_nb = {
2469 .notifier_call = device_change_notifier,
2472 void amd_iommu_init_notifier(void)
2474 bus_register_notifier(&pci_bus_type, &device_nb);
2477 /*****************************************************************************
2479 * The next functions belong to the dma_ops mapping/unmapping code.
2481 *****************************************************************************/
2484 * In the dma_ops path we only have the struct device. This function
2485 * finds the corresponding IOMMU, the protection domain and the
2486 * requestor id for a given device.
2487 * If the device is not yet associated with a domain this is also done
2490 static struct protection_domain *get_domain(struct device *dev)
2492 struct protection_domain *domain;
2493 struct dma_ops_domain *dma_dom;
2494 u16 devid = get_device_id(dev);
2496 if (!check_device(dev))
2497 return ERR_PTR(-EINVAL);
2499 domain = domain_for_device(dev);
2500 if (domain != NULL && !dma_ops_domain(domain))
2501 return ERR_PTR(-EBUSY);
2506 /* Device not bound yet - bind it */
2507 dma_dom = find_protection_domain(devid);
2509 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2510 attach_device(dev, &dma_dom->domain);
2511 DUMP_printk("Using protection domain %d for device %s\n",
2512 dma_dom->domain.id, dev_name(dev));
2514 return &dma_dom->domain;
2517 static void update_device_table(struct protection_domain *domain)
2519 struct iommu_dev_data *dev_data;
2521 list_for_each_entry(dev_data, &domain->dev_list, list)
2522 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2525 static void update_domain(struct protection_domain *domain)
2527 if (!domain->updated)
2530 update_device_table(domain);
2532 domain_flush_devices(domain);
2533 domain_flush_tlb_pde(domain);
2535 domain->updated = false;
2539 * This function fetches the PTE for a given address in the aperture
2541 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2542 unsigned long address)
2544 struct aperture_range *aperture;
2545 u64 *pte, *pte_page;
2547 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2551 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2553 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2555 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2557 pte += PM_LEVEL_INDEX(0, address);
2559 update_domain(&dom->domain);
2565 * This is the generic map function. It maps one 4kb page at paddr to
2566 * the given address in the DMA address space for the domain.
2568 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2569 unsigned long address,
2575 WARN_ON(address > dom->aperture_size);
2579 pte = dma_ops_get_pte(dom, address);
2581 return DMA_ERROR_CODE;
2583 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2585 if (direction == DMA_TO_DEVICE)
2586 __pte |= IOMMU_PTE_IR;
2587 else if (direction == DMA_FROM_DEVICE)
2588 __pte |= IOMMU_PTE_IW;
2589 else if (direction == DMA_BIDIRECTIONAL)
2590 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2596 return (dma_addr_t)address;
2600 * The generic unmapping function for on page in the DMA address space.
2602 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2603 unsigned long address)
2605 struct aperture_range *aperture;
2608 if (address >= dom->aperture_size)
2611 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2615 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2619 pte += PM_LEVEL_INDEX(0, address);
2627 * This function contains common code for mapping of a physically
2628 * contiguous memory region into DMA address space. It is used by all
2629 * mapping functions provided with this IOMMU driver.
2630 * Must be called with the domain lock held.
2632 static dma_addr_t __map_single(struct device *dev,
2633 struct dma_ops_domain *dma_dom,
2640 dma_addr_t offset = paddr & ~PAGE_MASK;
2641 dma_addr_t address, start, ret;
2643 unsigned long align_mask = 0;
2646 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2649 INC_STATS_COUNTER(total_map_requests);
2652 INC_STATS_COUNTER(cross_page);
2655 align_mask = (1UL << get_order(size)) - 1;
2658 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2660 if (unlikely(address == DMA_ERROR_CODE)) {
2662 * setting next_address here will let the address
2663 * allocator only scan the new allocated range in the
2664 * first run. This is a small optimization.
2666 dma_dom->next_address = dma_dom->aperture_size;
2668 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2672 * aperture was successfully enlarged by 128 MB, try
2679 for (i = 0; i < pages; ++i) {
2680 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2681 if (ret == DMA_ERROR_CODE)
2689 ADD_STATS_COUNTER(alloced_io_mem, size);
2691 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2692 domain_flush_tlb(&dma_dom->domain);
2693 dma_dom->need_flush = false;
2694 } else if (unlikely(amd_iommu_np_cache))
2695 domain_flush_pages(&dma_dom->domain, address, size);
2702 for (--i; i >= 0; --i) {
2704 dma_ops_domain_unmap(dma_dom, start);
2707 dma_ops_free_addresses(dma_dom, address, pages);
2709 return DMA_ERROR_CODE;
2713 * Does the reverse of the __map_single function. Must be called with
2714 * the domain lock held too
2716 static void __unmap_single(struct dma_ops_domain *dma_dom,
2717 dma_addr_t dma_addr,
2721 dma_addr_t flush_addr;
2722 dma_addr_t i, start;
2725 if ((dma_addr == DMA_ERROR_CODE) ||
2726 (dma_addr + size > dma_dom->aperture_size))
2729 flush_addr = dma_addr;
2730 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2731 dma_addr &= PAGE_MASK;
2734 for (i = 0; i < pages; ++i) {
2735 dma_ops_domain_unmap(dma_dom, start);
2739 SUB_STATS_COUNTER(alloced_io_mem, size);
2741 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2743 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2744 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2745 dma_dom->need_flush = false;
2750 * The exported map_single function for dma_ops.
2752 static dma_addr_t map_page(struct device *dev, struct page *page,
2753 unsigned long offset, size_t size,
2754 enum dma_data_direction dir,
2755 struct dma_attrs *attrs)
2757 unsigned long flags;
2758 struct protection_domain *domain;
2761 phys_addr_t paddr = page_to_phys(page) + offset;
2763 INC_STATS_COUNTER(cnt_map_single);
2765 domain = get_domain(dev);
2766 if (PTR_ERR(domain) == -EINVAL)
2767 return (dma_addr_t)paddr;
2768 else if (IS_ERR(domain))
2769 return DMA_ERROR_CODE;
2771 dma_mask = *dev->dma_mask;
2773 spin_lock_irqsave(&domain->lock, flags);
2775 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2777 if (addr == DMA_ERROR_CODE)
2780 domain_flush_complete(domain);
2783 spin_unlock_irqrestore(&domain->lock, flags);
2789 * The exported unmap_single function for dma_ops.
2791 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2792 enum dma_data_direction dir, struct dma_attrs *attrs)
2794 unsigned long flags;
2795 struct protection_domain *domain;
2797 INC_STATS_COUNTER(cnt_unmap_single);
2799 domain = get_domain(dev);
2803 spin_lock_irqsave(&domain->lock, flags);
2805 __unmap_single(domain->priv, dma_addr, size, dir);
2807 domain_flush_complete(domain);
2809 spin_unlock_irqrestore(&domain->lock, flags);
2813 * The exported map_sg function for dma_ops (handles scatter-gather
2816 static int map_sg(struct device *dev, struct scatterlist *sglist,
2817 int nelems, enum dma_data_direction dir,
2818 struct dma_attrs *attrs)
2820 unsigned long flags;
2821 struct protection_domain *domain;
2823 struct scatterlist *s;
2825 int mapped_elems = 0;
2828 INC_STATS_COUNTER(cnt_map_sg);
2830 domain = get_domain(dev);
2834 dma_mask = *dev->dma_mask;
2836 spin_lock_irqsave(&domain->lock, flags);
2838 for_each_sg(sglist, s, nelems, i) {
2841 s->dma_address = __map_single(dev, domain->priv,
2842 paddr, s->length, dir, false,
2845 if (s->dma_address) {
2846 s->dma_length = s->length;
2852 domain_flush_complete(domain);
2855 spin_unlock_irqrestore(&domain->lock, flags);
2857 return mapped_elems;
2859 for_each_sg(sglist, s, mapped_elems, i) {
2861 __unmap_single(domain->priv, s->dma_address,
2862 s->dma_length, dir);
2863 s->dma_address = s->dma_length = 0;
2872 * The exported map_sg function for dma_ops (handles scatter-gather
2875 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2876 int nelems, enum dma_data_direction dir,
2877 struct dma_attrs *attrs)
2879 unsigned long flags;
2880 struct protection_domain *domain;
2881 struct scatterlist *s;
2884 INC_STATS_COUNTER(cnt_unmap_sg);
2886 domain = get_domain(dev);
2890 spin_lock_irqsave(&domain->lock, flags);
2892 for_each_sg(sglist, s, nelems, i) {
2893 __unmap_single(domain->priv, s->dma_address,
2894 s->dma_length, dir);
2895 s->dma_address = s->dma_length = 0;
2898 domain_flush_complete(domain);
2900 spin_unlock_irqrestore(&domain->lock, flags);
2904 * The exported alloc_coherent function for dma_ops.
2906 static void *alloc_coherent(struct device *dev, size_t size,
2907 dma_addr_t *dma_addr, gfp_t flag,
2908 struct dma_attrs *attrs)
2910 u64 dma_mask = dev->coherent_dma_mask;
2911 struct protection_domain *domain;
2912 unsigned long flags;
2915 INC_STATS_COUNTER(cnt_alloc_coherent);
2917 domain = get_domain(dev);
2918 if (PTR_ERR(domain) == -EINVAL) {
2919 page = alloc_pages(flag, get_order(size));
2920 *dma_addr = page_to_phys(page);
2921 return page_address(page);
2922 } else if (IS_ERR(domain))
2925 size = PAGE_ALIGN(size);
2926 dma_mask = dev->coherent_dma_mask;
2927 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2929 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2931 if (!(flag & __GFP_WAIT))
2934 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2941 dma_mask = *dev->dma_mask;
2943 spin_lock_irqsave(&domain->lock, flags);
2945 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2946 size, DMA_BIDIRECTIONAL, true, dma_mask);
2948 if (*dma_addr == DMA_ERROR_CODE) {
2949 spin_unlock_irqrestore(&domain->lock, flags);
2953 domain_flush_complete(domain);
2955 spin_unlock_irqrestore(&domain->lock, flags);
2957 return page_address(page);
2961 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2962 __free_pages(page, get_order(size));
2968 * The exported free_coherent function for dma_ops.
2970 static void free_coherent(struct device *dev, size_t size,
2971 void *virt_addr, dma_addr_t dma_addr,
2972 struct dma_attrs *attrs)
2974 struct protection_domain *domain;
2975 unsigned long flags;
2978 INC_STATS_COUNTER(cnt_free_coherent);
2980 page = virt_to_page(virt_addr);
2981 size = PAGE_ALIGN(size);
2983 domain = get_domain(dev);
2987 spin_lock_irqsave(&domain->lock, flags);
2989 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2991 domain_flush_complete(domain);
2993 spin_unlock_irqrestore(&domain->lock, flags);
2996 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2997 __free_pages(page, get_order(size));
3001 * This function is called by the DMA layer to find out if we can handle a
3002 * particular device. It is part of the dma_ops.
3004 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
3006 return check_device(dev);
3010 * The function for pre-allocating protection domains.
3012 * If the driver core informs the DMA layer if a driver grabs a device
3013 * we don't need to preallocate the protection domains anymore.
3014 * For now we have to.
3016 static void __init prealloc_protection_domains(void)
3018 struct iommu_dev_data *dev_data;
3019 struct dma_ops_domain *dma_dom;
3020 struct pci_dev *dev = NULL;
3023 for_each_pci_dev(dev) {
3025 /* Do we handle this device? */
3026 if (!check_device(&dev->dev))
3029 dev_data = get_dev_data(&dev->dev);
3030 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3031 /* Make sure passthrough domain is allocated */
3032 alloc_passthrough_domain();
3033 dev_data->passthrough = true;
3034 attach_device(&dev->dev, pt_domain);
3035 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3036 dev_name(&dev->dev));
3039 /* Is there already any domain for it? */
3040 if (domain_for_device(&dev->dev))
3043 devid = get_device_id(&dev->dev);
3045 dma_dom = dma_ops_domain_alloc();
3048 init_unity_mappings_for_device(dma_dom, devid);
3049 dma_dom->target_dev = devid;
3051 attach_device(&dev->dev, &dma_dom->domain);
3053 list_add_tail(&dma_dom->list, &iommu_pd_list);
3057 static struct dma_map_ops amd_iommu_dma_ops = {
3058 .alloc = alloc_coherent,
3059 .free = free_coherent,
3060 .map_page = map_page,
3061 .unmap_page = unmap_page,
3063 .unmap_sg = unmap_sg,
3064 .dma_supported = amd_iommu_dma_supported,
3067 static unsigned device_dma_ops_init(void)
3069 struct iommu_dev_data *dev_data;
3070 struct pci_dev *pdev = NULL;
3071 unsigned unhandled = 0;
3073 for_each_pci_dev(pdev) {
3074 if (!check_device(&pdev->dev)) {
3076 iommu_ignore_device(&pdev->dev);
3082 dev_data = get_dev_data(&pdev->dev);
3084 if (!dev_data->passthrough)
3085 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3087 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3094 * The function which clues the AMD IOMMU driver into dma_ops.
3097 void __init amd_iommu_init_api(void)
3099 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3102 int __init amd_iommu_init_dma_ops(void)
3104 struct amd_iommu *iommu;
3108 * first allocate a default protection domain for every IOMMU we
3109 * found in the system. Devices not assigned to any other
3110 * protection domain will be assigned to the default one.
3112 for_each_iommu(iommu) {
3113 iommu->default_dom = dma_ops_domain_alloc();
3114 if (iommu->default_dom == NULL)
3116 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3117 ret = iommu_init_unity_mappings(iommu);
3123 * Pre-allocate the protection domains for each device.
3125 prealloc_protection_domains();
3130 /* Make the driver finally visible to the drivers */
3131 unhandled = device_dma_ops_init();
3132 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3133 /* There are unhandled devices - initialize swiotlb for them */
3137 amd_iommu_stats_init();
3139 if (amd_iommu_unmap_flush)
3140 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3142 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3148 for_each_iommu(iommu) {
3149 dma_ops_domain_free(iommu->default_dom);
3155 /*****************************************************************************
3157 * The following functions belong to the exported interface of AMD IOMMU
3159 * This interface allows access to lower level functions of the IOMMU
3160 * like protection domain handling and assignement of devices to domains
3161 * which is not possible with the dma_ops interface.
3163 *****************************************************************************/
3165 static void cleanup_domain(struct protection_domain *domain)
3167 struct iommu_dev_data *entry;
3168 unsigned long flags;
3170 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3172 while (!list_empty(&domain->dev_list)) {
3173 entry = list_first_entry(&domain->dev_list,
3174 struct iommu_dev_data, list);
3175 __detach_device(entry);
3178 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3181 static void protection_domain_free(struct protection_domain *domain)
3186 del_domain_from_list(domain);
3189 domain_id_free(domain->id);
3194 static struct protection_domain *protection_domain_alloc(void)
3196 struct protection_domain *domain;
3198 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3202 spin_lock_init(&domain->lock);
3203 mutex_init(&domain->api_lock);
3204 domain->id = domain_id_alloc();
3207 INIT_LIST_HEAD(&domain->dev_list);
3209 add_domain_to_list(domain);
3219 static int __init alloc_passthrough_domain(void)
3221 if (pt_domain != NULL)
3224 /* allocate passthrough domain */
3225 pt_domain = protection_domain_alloc();
3229 pt_domain->mode = PAGE_MODE_NONE;
3233 static int amd_iommu_domain_init(struct iommu_domain *dom)
3235 struct protection_domain *domain;
3237 domain = protection_domain_alloc();
3241 domain->mode = PAGE_MODE_3_LEVEL;
3242 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3243 if (!domain->pt_root)
3246 domain->iommu_domain = dom;
3250 dom->geometry.aperture_start = 0;
3251 dom->geometry.aperture_end = ~0ULL;
3252 dom->geometry.force_aperture = true;
3257 protection_domain_free(domain);
3262 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3264 struct protection_domain *domain = dom->priv;
3269 if (domain->dev_cnt > 0)
3270 cleanup_domain(domain);
3272 BUG_ON(domain->dev_cnt != 0);
3274 if (domain->mode != PAGE_MODE_NONE)
3275 free_pagetable(domain);
3277 if (domain->flags & PD_IOMMUV2_MASK)
3278 free_gcr3_table(domain);
3280 protection_domain_free(domain);
3285 static void amd_iommu_detach_device(struct iommu_domain *dom,
3288 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3289 struct amd_iommu *iommu;
3292 if (!check_device(dev))
3295 devid = get_device_id(dev);
3297 if (dev_data->domain != NULL)
3300 iommu = amd_iommu_rlookup_table[devid];
3304 iommu_completion_wait(iommu);
3307 static int amd_iommu_attach_device(struct iommu_domain *dom,
3310 struct protection_domain *domain = dom->priv;
3311 struct iommu_dev_data *dev_data;
3312 struct amd_iommu *iommu;
3315 if (!check_device(dev))
3318 dev_data = dev->archdata.iommu;
3320 iommu = amd_iommu_rlookup_table[dev_data->devid];
3324 if (dev_data->domain)
3327 ret = attach_device(dev, domain);
3329 iommu_completion_wait(iommu);
3334 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3335 phys_addr_t paddr, size_t page_size, int iommu_prot)
3337 struct protection_domain *domain = dom->priv;
3341 if (domain->mode == PAGE_MODE_NONE)
3344 if (iommu_prot & IOMMU_READ)
3345 prot |= IOMMU_PROT_IR;
3346 if (iommu_prot & IOMMU_WRITE)
3347 prot |= IOMMU_PROT_IW;
3349 mutex_lock(&domain->api_lock);
3350 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3351 mutex_unlock(&domain->api_lock);
3356 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3359 struct protection_domain *domain = dom->priv;
3362 if (domain->mode == PAGE_MODE_NONE)
3365 mutex_lock(&domain->api_lock);
3366 unmap_size = iommu_unmap_page(domain, iova, page_size);
3367 mutex_unlock(&domain->api_lock);
3369 domain_flush_tlb_pde(domain);
3374 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3377 struct protection_domain *domain = dom->priv;
3378 unsigned long offset_mask, pte_pgsize;
3381 if (domain->mode == PAGE_MODE_NONE)
3384 pte = fetch_pte(domain, iova, &pte_pgsize);
3386 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3389 offset_mask = pte_pgsize - 1;
3390 __pte = *pte & PM_ADDR_MASK;
3392 return (__pte & ~offset_mask) | (iova & offset_mask);
3395 static bool amd_iommu_capable(enum iommu_cap cap)
3398 case IOMMU_CAP_CACHE_COHERENCY:
3400 case IOMMU_CAP_INTR_REMAP:
3401 return (irq_remapping_enabled == 1);
3402 case IOMMU_CAP_NOEXEC:
3409 static const struct iommu_ops amd_iommu_ops = {
3410 .capable = amd_iommu_capable,
3411 .domain_init = amd_iommu_domain_init,
3412 .domain_destroy = amd_iommu_domain_destroy,
3413 .attach_dev = amd_iommu_attach_device,
3414 .detach_dev = amd_iommu_detach_device,
3415 .map = amd_iommu_map,
3416 .unmap = amd_iommu_unmap,
3417 .map_sg = default_iommu_map_sg,
3418 .iova_to_phys = amd_iommu_iova_to_phys,
3419 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3422 /*****************************************************************************
3424 * The next functions do a basic initialization of IOMMU for pass through
3427 * In passthrough mode the IOMMU is initialized and enabled but not used for
3428 * DMA-API translation.
3430 *****************************************************************************/
3432 int __init amd_iommu_init_passthrough(void)
3434 struct iommu_dev_data *dev_data;
3435 struct pci_dev *dev = NULL;
3438 ret = alloc_passthrough_domain();
3442 for_each_pci_dev(dev) {
3443 if (!check_device(&dev->dev))
3446 dev_data = get_dev_data(&dev->dev);
3447 dev_data->passthrough = true;
3449 attach_device(&dev->dev, pt_domain);
3452 amd_iommu_stats_init();
3454 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3459 /* IOMMUv2 specific functions */
3460 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3462 return atomic_notifier_chain_register(&ppr_notifier, nb);
3464 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3466 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3468 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3470 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3472 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3474 struct protection_domain *domain = dom->priv;
3475 unsigned long flags;
3477 spin_lock_irqsave(&domain->lock, flags);
3479 /* Update data structure */
3480 domain->mode = PAGE_MODE_NONE;
3481 domain->updated = true;
3483 /* Make changes visible to IOMMUs */
3484 update_domain(domain);
3486 /* Page-table is not visible to IOMMU anymore, so free it */
3487 free_pagetable(domain);
3489 spin_unlock_irqrestore(&domain->lock, flags);
3491 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3493 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3495 struct protection_domain *domain = dom->priv;
3496 unsigned long flags;
3499 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3502 /* Number of GCR3 table levels required */
3503 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3506 if (levels > amd_iommu_max_glx_val)
3509 spin_lock_irqsave(&domain->lock, flags);
3512 * Save us all sanity checks whether devices already in the
3513 * domain support IOMMUv2. Just force that the domain has no
3514 * devices attached when it is switched into IOMMUv2 mode.
3517 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3521 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3522 if (domain->gcr3_tbl == NULL)
3525 domain->glx = levels;
3526 domain->flags |= PD_IOMMUV2_MASK;
3527 domain->updated = true;
3529 update_domain(domain);
3534 spin_unlock_irqrestore(&domain->lock, flags);
3538 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3540 static int __flush_pasid(struct protection_domain *domain, int pasid,
3541 u64 address, bool size)
3543 struct iommu_dev_data *dev_data;
3544 struct iommu_cmd cmd;
3547 if (!(domain->flags & PD_IOMMUV2_MASK))
3550 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3553 * IOMMU TLB needs to be flushed before Device TLB to
3554 * prevent device TLB refill from IOMMU TLB
3556 for (i = 0; i < amd_iommus_present; ++i) {
3557 if (domain->dev_iommu[i] == 0)
3560 ret = iommu_queue_command(amd_iommus[i], &cmd);
3565 /* Wait until IOMMU TLB flushes are complete */
3566 domain_flush_complete(domain);
3568 /* Now flush device TLBs */
3569 list_for_each_entry(dev_data, &domain->dev_list, list) {
3570 struct amd_iommu *iommu;
3573 BUG_ON(!dev_data->ats.enabled);
3575 qdep = dev_data->ats.qdep;
3576 iommu = amd_iommu_rlookup_table[dev_data->devid];
3578 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3579 qdep, address, size);
3581 ret = iommu_queue_command(iommu, &cmd);
3586 /* Wait until all device TLBs are flushed */
3587 domain_flush_complete(domain);
3596 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3599 INC_STATS_COUNTER(invalidate_iotlb);
3601 return __flush_pasid(domain, pasid, address, false);
3604 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3607 struct protection_domain *domain = dom->priv;
3608 unsigned long flags;
3611 spin_lock_irqsave(&domain->lock, flags);
3612 ret = __amd_iommu_flush_page(domain, pasid, address);
3613 spin_unlock_irqrestore(&domain->lock, flags);
3617 EXPORT_SYMBOL(amd_iommu_flush_page);
3619 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3621 INC_STATS_COUNTER(invalidate_iotlb_all);
3623 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3627 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3629 struct protection_domain *domain = dom->priv;
3630 unsigned long flags;
3633 spin_lock_irqsave(&domain->lock, flags);
3634 ret = __amd_iommu_flush_tlb(domain, pasid);
3635 spin_unlock_irqrestore(&domain->lock, flags);
3639 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3641 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3648 index = (pasid >> (9 * level)) & 0x1ff;
3654 if (!(*pte & GCR3_VALID)) {
3658 root = (void *)get_zeroed_page(GFP_ATOMIC);
3662 *pte = __pa(root) | GCR3_VALID;
3665 root = __va(*pte & PAGE_MASK);
3673 static int __set_gcr3(struct protection_domain *domain, int pasid,
3678 if (domain->mode != PAGE_MODE_NONE)
3681 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3685 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3687 return __amd_iommu_flush_tlb(domain, pasid);
3690 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3694 if (domain->mode != PAGE_MODE_NONE)
3697 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3703 return __amd_iommu_flush_tlb(domain, pasid);
3706 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3709 struct protection_domain *domain = dom->priv;
3710 unsigned long flags;
3713 spin_lock_irqsave(&domain->lock, flags);
3714 ret = __set_gcr3(domain, pasid, cr3);
3715 spin_unlock_irqrestore(&domain->lock, flags);
3719 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3721 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3723 struct protection_domain *domain = dom->priv;
3724 unsigned long flags;
3727 spin_lock_irqsave(&domain->lock, flags);
3728 ret = __clear_gcr3(domain, pasid);
3729 spin_unlock_irqrestore(&domain->lock, flags);
3733 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3735 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3736 int status, int tag)
3738 struct iommu_dev_data *dev_data;
3739 struct amd_iommu *iommu;
3740 struct iommu_cmd cmd;
3742 INC_STATS_COUNTER(complete_ppr);
3744 dev_data = get_dev_data(&pdev->dev);
3745 iommu = amd_iommu_rlookup_table[dev_data->devid];
3747 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3748 tag, dev_data->pri_tlp);
3750 return iommu_queue_command(iommu, &cmd);
3752 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3754 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3756 struct protection_domain *domain;
3758 domain = get_domain(&pdev->dev);
3762 /* Only return IOMMUv2 domains */
3763 if (!(domain->flags & PD_IOMMUV2_MASK))
3766 return domain->iommu_domain;
3768 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3770 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3772 struct iommu_dev_data *dev_data;
3774 if (!amd_iommu_v2_supported())
3777 dev_data = get_dev_data(&pdev->dev);
3778 dev_data->errata |= (1 << erratum);
3780 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3782 int amd_iommu_device_info(struct pci_dev *pdev,
3783 struct amd_iommu_device_info *info)
3788 if (pdev == NULL || info == NULL)
3791 if (!amd_iommu_v2_supported())
3794 memset(info, 0, sizeof(*info));
3796 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3798 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3800 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3802 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3804 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3808 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3809 max_pasids = min(max_pasids, (1 << 20));
3811 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3812 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3814 features = pci_pasid_features(pdev);
3815 if (features & PCI_PASID_CAP_EXEC)
3816 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3817 if (features & PCI_PASID_CAP_PRIV)
3818 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3823 EXPORT_SYMBOL(amd_iommu_device_info);
3825 #ifdef CONFIG_IRQ_REMAP
3827 /*****************************************************************************
3829 * Interrupt Remapping Implementation
3831 *****************************************************************************/
3848 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3849 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3850 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3851 #define DTE_IRQ_REMAP_ENABLE 1ULL
3853 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3857 dte = amd_iommu_dev_table[devid].data[2];
3858 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3859 dte |= virt_to_phys(table->table);
3860 dte |= DTE_IRQ_REMAP_INTCTL;
3861 dte |= DTE_IRQ_TABLE_LEN;
3862 dte |= DTE_IRQ_REMAP_ENABLE;
3864 amd_iommu_dev_table[devid].data[2] = dte;
3867 #define IRTE_ALLOCATED (~1U)
3869 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3871 struct irq_remap_table *table = NULL;
3872 struct amd_iommu *iommu;
3873 unsigned long flags;
3876 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3878 iommu = amd_iommu_rlookup_table[devid];
3882 table = irq_lookup_table[devid];
3886 alias = amd_iommu_alias_table[devid];
3887 table = irq_lookup_table[alias];
3889 irq_lookup_table[devid] = table;
3890 set_dte_irq_entry(devid, table);
3891 iommu_flush_dte(iommu, devid);
3895 /* Nothing there yet, allocate new irq remapping table */
3896 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3900 /* Initialize table spin-lock */
3901 spin_lock_init(&table->lock);
3904 /* Keep the first 32 indexes free for IOAPIC interrupts */
3905 table->min_index = 32;
3907 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3908 if (!table->table) {
3914 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3919 for (i = 0; i < 32; ++i)
3920 table->table[i] = IRTE_ALLOCATED;
3923 irq_lookup_table[devid] = table;
3924 set_dte_irq_entry(devid, table);
3925 iommu_flush_dte(iommu, devid);
3926 if (devid != alias) {
3927 irq_lookup_table[alias] = table;
3928 set_dte_irq_entry(alias, table);
3929 iommu_flush_dte(iommu, alias);
3933 iommu_completion_wait(iommu);
3936 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3941 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
3943 struct irq_remap_table *table;
3944 unsigned long flags;
3947 table = get_irq_table(devid, false);
3951 spin_lock_irqsave(&table->lock, flags);
3953 /* Scan table for free entries */
3954 for (c = 0, index = table->min_index;
3955 index < MAX_IRQS_PER_TABLE;
3957 if (table->table[index] == 0)
3963 struct irq_2_irte *irte_info;
3966 table->table[index - c + 1] = IRTE_ALLOCATED;
3971 irte_info = &cfg->irq_2_irte;
3972 irte_info->devid = devid;
3973 irte_info->index = index;
3982 spin_unlock_irqrestore(&table->lock, flags);
3987 static int get_irte(u16 devid, int index, union irte *irte)
3989 struct irq_remap_table *table;
3990 unsigned long flags;
3992 table = get_irq_table(devid, false);
3996 spin_lock_irqsave(&table->lock, flags);
3997 irte->val = table->table[index];
3998 spin_unlock_irqrestore(&table->lock, flags);
4003 static int modify_irte(u16 devid, int index, union irte irte)
4005 struct irq_remap_table *table;
4006 struct amd_iommu *iommu;
4007 unsigned long flags;
4009 iommu = amd_iommu_rlookup_table[devid];
4013 table = get_irq_table(devid, false);
4017 spin_lock_irqsave(&table->lock, flags);
4018 table->table[index] = irte.val;
4019 spin_unlock_irqrestore(&table->lock, flags);
4021 iommu_flush_irt(iommu, devid);
4022 iommu_completion_wait(iommu);
4027 static void free_irte(u16 devid, int index)
4029 struct irq_remap_table *table;
4030 struct amd_iommu *iommu;
4031 unsigned long flags;
4033 iommu = amd_iommu_rlookup_table[devid];
4037 table = get_irq_table(devid, false);
4041 spin_lock_irqsave(&table->lock, flags);
4042 table->table[index] = 0;
4043 spin_unlock_irqrestore(&table->lock, flags);
4045 iommu_flush_irt(iommu, devid);
4046 iommu_completion_wait(iommu);
4049 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4050 unsigned int destination, int vector,
4051 struct io_apic_irq_attr *attr)
4053 struct irq_remap_table *table;
4054 struct irq_2_irte *irte_info;
4055 struct irq_cfg *cfg;
4066 irte_info = &cfg->irq_2_irte;
4067 ioapic_id = mpc_ioapic_id(attr->ioapic);
4068 devid = get_ioapic_devid(ioapic_id);
4073 table = get_irq_table(devid, true);
4077 index = attr->ioapic_pin;
4079 /* Setup IRQ remapping info */
4081 irte_info->devid = devid;
4082 irte_info->index = index;
4084 /* Setup IRTE for IOMMU */
4086 irte.fields.vector = vector;
4087 irte.fields.int_type = apic->irq_delivery_mode;
4088 irte.fields.destination = destination;
4089 irte.fields.dm = apic->irq_dest_mode;
4090 irte.fields.valid = 1;
4092 ret = modify_irte(devid, index, irte);
4096 /* Setup IOAPIC entry */
4097 memset(entry, 0, sizeof(*entry));
4099 entry->vector = index;
4101 entry->trigger = attr->trigger;
4102 entry->polarity = attr->polarity;
4105 * Mask level triggered irqs.
4113 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4116 struct irq_2_irte *irte_info;
4117 unsigned int dest, irq;
4118 struct irq_cfg *cfg;
4122 if (!config_enabled(CONFIG_SMP))
4125 cfg = irqd_cfg(data);
4127 irte_info = &cfg->irq_2_irte;
4129 if (!cpumask_intersects(mask, cpu_online_mask))
4132 if (get_irte(irte_info->devid, irte_info->index, &irte))
4135 if (assign_irq_vector(irq, cfg, mask))
4138 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4140 if (assign_irq_vector(irq, cfg, data->affinity))
4141 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4145 irte.fields.vector = cfg->vector;
4146 irte.fields.destination = dest;
4148 modify_irte(irte_info->devid, irte_info->index, irte);
4150 if (cfg->move_in_progress)
4151 send_cleanup_vector(cfg);
4153 cpumask_copy(data->affinity, mask);
4158 static int free_irq(int irq)
4160 struct irq_2_irte *irte_info;
4161 struct irq_cfg *cfg;
4167 irte_info = &cfg->irq_2_irte;
4169 free_irte(irte_info->devid, irte_info->index);
4174 static void compose_msi_msg(struct pci_dev *pdev,
4175 unsigned int irq, unsigned int dest,
4176 struct msi_msg *msg, u8 hpet_id)
4178 struct irq_2_irte *irte_info;
4179 struct irq_cfg *cfg;
4186 irte_info = &cfg->irq_2_irte;
4189 irte.fields.vector = cfg->vector;
4190 irte.fields.int_type = apic->irq_delivery_mode;
4191 irte.fields.destination = dest;
4192 irte.fields.dm = apic->irq_dest_mode;
4193 irte.fields.valid = 1;
4195 modify_irte(irte_info->devid, irte_info->index, irte);
4197 msg->address_hi = MSI_ADDR_BASE_HI;
4198 msg->address_lo = MSI_ADDR_BASE_LO;
4199 msg->data = irte_info->index;
4202 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4204 struct irq_cfg *cfg;
4215 devid = get_device_id(&pdev->dev);
4216 index = alloc_irq_index(cfg, devid, nvec);
4218 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4221 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4222 int index, int offset)
4224 struct irq_2_irte *irte_info;
4225 struct irq_cfg *cfg;
4235 if (index >= MAX_IRQS_PER_TABLE)
4238 devid = get_device_id(&pdev->dev);
4239 irte_info = &cfg->irq_2_irte;
4242 irte_info->devid = devid;
4243 irte_info->index = index + offset;
4248 static int alloc_hpet_msi(unsigned int irq, unsigned int id)
4250 struct irq_2_irte *irte_info;
4251 struct irq_cfg *cfg;
4258 irte_info = &cfg->irq_2_irte;
4259 devid = get_hpet_devid(id);
4263 index = alloc_irq_index(cfg, devid, 1);
4268 irte_info->devid = devid;
4269 irte_info->index = index;
4274 struct irq_remap_ops amd_iommu_irq_ops = {
4275 .prepare = amd_iommu_prepare,
4276 .enable = amd_iommu_enable,
4277 .disable = amd_iommu_disable,
4278 .reenable = amd_iommu_reenable,
4279 .enable_faulting = amd_iommu_enable_faulting,
4280 .setup_ioapic_entry = setup_ioapic_entry,
4281 .set_affinity = set_affinity,
4282 .free_irq = free_irq,
4283 .compose_msi_msg = compose_msi_msg,
4284 .msi_alloc_irq = msi_alloc_irq,
4285 .msi_setup_irq = msi_setup_irq,
4286 .alloc_hpet_msi = alloc_hpet_msi,