2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/io_apic.h>
39 #include <asm/hw_irq.h>
40 #include <asm/msidef.h>
41 #include <asm/proto.h>
42 #include <asm/iommu.h>
46 #include "amd_iommu_proto.h"
47 #include "amd_iommu_types.h"
48 #include "irq_remapping.h"
50 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
52 #define LOOP_TIMEOUT 100000
55 * This bitmap is used to advertise the page sizes our hardware support
56 * to the IOMMU core, which will then use this information to split
57 * physically contiguous memory regions it is mapping into page sizes
60 * 512GB Pages are not supported due to a hardware bug
62 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
64 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
66 /* A list of preallocated protection domains */
67 static LIST_HEAD(iommu_pd_list);
68 static DEFINE_SPINLOCK(iommu_pd_list_lock);
70 /* List of all available dev_data structures */
71 static LIST_HEAD(dev_data_list);
72 static DEFINE_SPINLOCK(dev_data_list_lock);
74 LIST_HEAD(ioapic_map);
78 * Domain for untranslated devices - only allocated
79 * if iommu=pt passed on kernel cmd line.
81 static struct protection_domain *pt_domain;
83 static const struct iommu_ops amd_iommu_ops;
85 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
86 int amd_iommu_max_glx_val = -1;
88 static struct dma_map_ops amd_iommu_dma_ops;
91 * general struct to manage commands send to an IOMMU
97 struct kmem_cache *amd_iommu_irq_cache;
99 static void update_domain(struct protection_domain *domain);
100 static int __init alloc_passthrough_domain(void);
102 /****************************************************************************
106 ****************************************************************************/
108 static struct iommu_dev_data *alloc_dev_data(u16 devid)
110 struct iommu_dev_data *dev_data;
113 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
117 dev_data->devid = devid;
118 atomic_set(&dev_data->bind, 0);
120 spin_lock_irqsave(&dev_data_list_lock, flags);
121 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
122 spin_unlock_irqrestore(&dev_data_list_lock, flags);
127 static void free_dev_data(struct iommu_dev_data *dev_data)
131 spin_lock_irqsave(&dev_data_list_lock, flags);
132 list_del(&dev_data->dev_data_list);
133 spin_unlock_irqrestore(&dev_data_list_lock, flags);
138 static struct iommu_dev_data *search_dev_data(u16 devid)
140 struct iommu_dev_data *dev_data;
143 spin_lock_irqsave(&dev_data_list_lock, flags);
144 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
145 if (dev_data->devid == devid)
152 spin_unlock_irqrestore(&dev_data_list_lock, flags);
157 static struct iommu_dev_data *find_dev_data(u16 devid)
159 struct iommu_dev_data *dev_data;
161 dev_data = search_dev_data(devid);
163 if (dev_data == NULL)
164 dev_data = alloc_dev_data(devid);
169 static inline u16 get_device_id(struct device *dev)
171 struct pci_dev *pdev = to_pci_dev(dev);
173 return PCI_DEVID(pdev->bus->number, pdev->devfn);
176 static struct iommu_dev_data *get_dev_data(struct device *dev)
178 return dev->archdata.iommu;
181 static bool pci_iommuv2_capable(struct pci_dev *pdev)
183 static const int caps[] = {
186 PCI_EXT_CAP_ID_PASID,
190 for (i = 0; i < 3; ++i) {
191 pos = pci_find_ext_capability(pdev, caps[i]);
199 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
201 struct iommu_dev_data *dev_data;
203 dev_data = get_dev_data(&pdev->dev);
205 return dev_data->errata & (1 << erratum) ? true : false;
209 * In this function the list of preallocated protection domains is traversed to
210 * find the domain for a specific device
212 static struct dma_ops_domain *find_protection_domain(u16 devid)
214 struct dma_ops_domain *entry, *ret = NULL;
216 u16 alias = amd_iommu_alias_table[devid];
218 if (list_empty(&iommu_pd_list))
221 spin_lock_irqsave(&iommu_pd_list_lock, flags);
223 list_for_each_entry(entry, &iommu_pd_list, list) {
224 if (entry->target_dev == devid ||
225 entry->target_dev == alias) {
231 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
237 * This function checks if the driver got a valid device from the caller to
238 * avoid dereferencing invalid pointers.
240 static bool check_device(struct device *dev)
244 if (!dev || !dev->dma_mask)
248 if (!dev_is_pci(dev))
251 devid = get_device_id(dev);
253 /* Out of our scope? */
254 if (devid > amd_iommu_last_bdf)
257 if (amd_iommu_rlookup_table[devid] == NULL)
263 static void init_iommu_group(struct device *dev)
265 struct iommu_group *group;
267 group = iommu_group_get_for_dev(dev);
269 iommu_group_put(group);
272 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
274 *(u16 *)data = alias;
278 static u16 get_alias(struct device *dev)
280 struct pci_dev *pdev = to_pci_dev(dev);
281 u16 devid, ivrs_alias, pci_alias;
283 devid = get_device_id(dev);
284 ivrs_alias = amd_iommu_alias_table[devid];
285 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
287 if (ivrs_alias == pci_alias)
293 * The IVRS is fairly reliable in telling us about aliases, but it
294 * can't know about every screwy device. If we don't have an IVRS
295 * reported alias, use the PCI reported alias. In that case we may
296 * still need to initialize the rlookup and dev_table entries if the
297 * alias is to a non-existent device.
299 if (ivrs_alias == devid) {
300 if (!amd_iommu_rlookup_table[pci_alias]) {
301 amd_iommu_rlookup_table[pci_alias] =
302 amd_iommu_rlookup_table[devid];
303 memcpy(amd_iommu_dev_table[pci_alias].data,
304 amd_iommu_dev_table[devid].data,
305 sizeof(amd_iommu_dev_table[pci_alias].data));
311 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
312 "for device %s[%04x:%04x], kernel reported alias "
313 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
314 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
315 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
316 PCI_FUNC(pci_alias));
319 * If we don't have a PCI DMA alias and the IVRS alias is on the same
320 * bus, then the IVRS table may know about a quirk that we don't.
322 if (pci_alias == devid &&
323 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
324 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
325 pdev->dma_alias_devfn = ivrs_alias & 0xff;
326 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
327 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
334 static int iommu_init_device(struct device *dev)
336 struct pci_dev *pdev = to_pci_dev(dev);
337 struct iommu_dev_data *dev_data;
340 if (dev->archdata.iommu)
343 dev_data = find_dev_data(get_device_id(dev));
347 alias = get_alias(dev);
349 if (alias != dev_data->devid) {
350 struct iommu_dev_data *alias_data;
352 alias_data = find_dev_data(alias);
353 if (alias_data == NULL) {
354 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
356 free_dev_data(dev_data);
359 dev_data->alias_data = alias_data;
362 if (pci_iommuv2_capable(pdev)) {
363 struct amd_iommu *iommu;
365 iommu = amd_iommu_rlookup_table[dev_data->devid];
366 dev_data->iommu_v2 = iommu->is_iommu_v2;
369 dev->archdata.iommu = dev_data;
371 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
377 static void iommu_ignore_device(struct device *dev)
381 devid = get_device_id(dev);
382 alias = amd_iommu_alias_table[devid];
384 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
385 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
387 amd_iommu_rlookup_table[devid] = NULL;
388 amd_iommu_rlookup_table[alias] = NULL;
391 static void iommu_uninit_device(struct device *dev)
393 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
398 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
401 iommu_group_remove_device(dev);
403 /* Unlink from alias, it may change if another device is re-plugged */
404 dev_data->alias_data = NULL;
407 * We keep dev_data around for unplugged devices and reuse it when the
408 * device is re-plugged - not doing so would introduce a ton of races.
412 void __init amd_iommu_uninit_devices(void)
414 struct iommu_dev_data *dev_data, *n;
415 struct pci_dev *pdev = NULL;
417 for_each_pci_dev(pdev) {
419 if (!check_device(&pdev->dev))
422 iommu_uninit_device(&pdev->dev);
425 /* Free all of our dev_data structures */
426 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
427 free_dev_data(dev_data);
430 int __init amd_iommu_init_devices(void)
432 struct pci_dev *pdev = NULL;
435 for_each_pci_dev(pdev) {
437 if (!check_device(&pdev->dev))
440 ret = iommu_init_device(&pdev->dev);
441 if (ret == -ENOTSUPP)
442 iommu_ignore_device(&pdev->dev);
448 * Initialize IOMMU groups only after iommu_init_device() has
449 * had a chance to populate any IVRS defined aliases.
451 for_each_pci_dev(pdev) {
452 if (check_device(&pdev->dev))
453 init_iommu_group(&pdev->dev);
460 amd_iommu_uninit_devices();
464 #ifdef CONFIG_AMD_IOMMU_STATS
467 * Initialization code for statistics collection
470 DECLARE_STATS_COUNTER(compl_wait);
471 DECLARE_STATS_COUNTER(cnt_map_single);
472 DECLARE_STATS_COUNTER(cnt_unmap_single);
473 DECLARE_STATS_COUNTER(cnt_map_sg);
474 DECLARE_STATS_COUNTER(cnt_unmap_sg);
475 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
476 DECLARE_STATS_COUNTER(cnt_free_coherent);
477 DECLARE_STATS_COUNTER(cross_page);
478 DECLARE_STATS_COUNTER(domain_flush_single);
479 DECLARE_STATS_COUNTER(domain_flush_all);
480 DECLARE_STATS_COUNTER(alloced_io_mem);
481 DECLARE_STATS_COUNTER(total_map_requests);
482 DECLARE_STATS_COUNTER(complete_ppr);
483 DECLARE_STATS_COUNTER(invalidate_iotlb);
484 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
485 DECLARE_STATS_COUNTER(pri_requests);
487 static struct dentry *stats_dir;
488 static struct dentry *de_fflush;
490 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
492 if (stats_dir == NULL)
495 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
499 static void amd_iommu_stats_init(void)
501 stats_dir = debugfs_create_dir("amd-iommu", NULL);
502 if (stats_dir == NULL)
505 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
506 &amd_iommu_unmap_flush);
508 amd_iommu_stats_add(&compl_wait);
509 amd_iommu_stats_add(&cnt_map_single);
510 amd_iommu_stats_add(&cnt_unmap_single);
511 amd_iommu_stats_add(&cnt_map_sg);
512 amd_iommu_stats_add(&cnt_unmap_sg);
513 amd_iommu_stats_add(&cnt_alloc_coherent);
514 amd_iommu_stats_add(&cnt_free_coherent);
515 amd_iommu_stats_add(&cross_page);
516 amd_iommu_stats_add(&domain_flush_single);
517 amd_iommu_stats_add(&domain_flush_all);
518 amd_iommu_stats_add(&alloced_io_mem);
519 amd_iommu_stats_add(&total_map_requests);
520 amd_iommu_stats_add(&complete_ppr);
521 amd_iommu_stats_add(&invalidate_iotlb);
522 amd_iommu_stats_add(&invalidate_iotlb_all);
523 amd_iommu_stats_add(&pri_requests);
528 /****************************************************************************
530 * Interrupt handling functions
532 ****************************************************************************/
534 static void dump_dte_entry(u16 devid)
538 for (i = 0; i < 4; ++i)
539 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
540 amd_iommu_dev_table[devid].data[i]);
543 static void dump_command(unsigned long phys_addr)
545 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
548 for (i = 0; i < 4; ++i)
549 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
552 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
554 int type, devid, domid, flags;
555 volatile u32 *event = __evt;
560 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
561 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
562 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
563 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
564 address = (u64)(((u64)event[3]) << 32) | event[2];
567 /* Did we hit the erratum? */
568 if (++count == LOOP_TIMEOUT) {
569 pr_err("AMD-Vi: No event written to event log\n");
576 printk(KERN_ERR "AMD-Vi: Event logged [");
579 case EVENT_TYPE_ILL_DEV:
580 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
581 "address=0x%016llx flags=0x%04x]\n",
582 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
584 dump_dte_entry(devid);
586 case EVENT_TYPE_IO_FAULT:
587 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
588 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
589 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
590 domid, address, flags);
592 case EVENT_TYPE_DEV_TAB_ERR:
593 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
594 "address=0x%016llx flags=0x%04x]\n",
595 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
598 case EVENT_TYPE_PAGE_TAB_ERR:
599 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
600 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
601 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
602 domid, address, flags);
604 case EVENT_TYPE_ILL_CMD:
605 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
606 dump_command(address);
608 case EVENT_TYPE_CMD_HARD_ERR:
609 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
610 "flags=0x%04x]\n", address, flags);
612 case EVENT_TYPE_IOTLB_INV_TO:
613 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
614 "address=0x%016llx]\n",
615 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
618 case EVENT_TYPE_INV_DEV_REQ:
619 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
620 "address=0x%016llx flags=0x%04x]\n",
621 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
625 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
628 memset(__evt, 0, 4 * sizeof(u32));
631 static void iommu_poll_events(struct amd_iommu *iommu)
635 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
636 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
638 while (head != tail) {
639 iommu_print_event(iommu, iommu->evt_buf + head);
640 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
643 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
646 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
648 struct amd_iommu_fault fault;
650 INC_STATS_COUNTER(pri_requests);
652 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
653 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
657 fault.address = raw[1];
658 fault.pasid = PPR_PASID(raw[0]);
659 fault.device_id = PPR_DEVID(raw[0]);
660 fault.tag = PPR_TAG(raw[0]);
661 fault.flags = PPR_FLAGS(raw[0]);
663 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
666 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
670 if (iommu->ppr_log == NULL)
673 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
674 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
676 while (head != tail) {
681 raw = (u64 *)(iommu->ppr_log + head);
684 * Hardware bug: Interrupt may arrive before the entry is
685 * written to memory. If this happens we need to wait for the
688 for (i = 0; i < LOOP_TIMEOUT; ++i) {
689 if (PPR_REQ_TYPE(raw[0]) != 0)
694 /* Avoid memcpy function-call overhead */
699 * To detect the hardware bug we need to clear the entry
702 raw[0] = raw[1] = 0UL;
704 /* Update head pointer of hardware ring-buffer */
705 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
706 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
708 /* Handle PPR entry */
709 iommu_handle_ppr_entry(iommu, entry);
711 /* Refresh ring-buffer information */
712 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
713 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
717 irqreturn_t amd_iommu_int_thread(int irq, void *data)
719 struct amd_iommu *iommu = (struct amd_iommu *) data;
720 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
722 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
723 /* Enable EVT and PPR interrupts again */
724 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
725 iommu->mmio_base + MMIO_STATUS_OFFSET);
727 if (status & MMIO_STATUS_EVT_INT_MASK) {
728 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
729 iommu_poll_events(iommu);
732 if (status & MMIO_STATUS_PPR_INT_MASK) {
733 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
734 iommu_poll_ppr_log(iommu);
738 * Hardware bug: ERBT1312
739 * When re-enabling interrupt (by writing 1
740 * to clear the bit), the hardware might also try to set
741 * the interrupt bit in the event status register.
742 * In this scenario, the bit will be set, and disable
743 * subsequent interrupts.
745 * Workaround: The IOMMU driver should read back the
746 * status register and check if the interrupt bits are cleared.
747 * If not, driver will need to go through the interrupt handler
748 * again and re-clear the bits
750 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
755 irqreturn_t amd_iommu_int_handler(int irq, void *data)
757 return IRQ_WAKE_THREAD;
760 /****************************************************************************
762 * IOMMU command queuing functions
764 ****************************************************************************/
766 static int wait_on_sem(volatile u64 *sem)
770 while (*sem == 0 && i < LOOP_TIMEOUT) {
775 if (i == LOOP_TIMEOUT) {
776 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
783 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
784 struct iommu_cmd *cmd,
789 target = iommu->cmd_buf + tail;
790 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
792 /* Copy command to buffer */
793 memcpy(target, cmd, sizeof(*cmd));
795 /* Tell the IOMMU about it */
796 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
799 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
801 WARN_ON(address & 0x7ULL);
803 memset(cmd, 0, sizeof(*cmd));
804 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
805 cmd->data[1] = upper_32_bits(__pa(address));
807 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
810 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
812 memset(cmd, 0, sizeof(*cmd));
813 cmd->data[0] = devid;
814 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
817 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
818 size_t size, u16 domid, int pde)
823 pages = iommu_num_pages(address, size, PAGE_SIZE);
828 * If we have to flush more than one page, flush all
829 * TLB entries for this domain
831 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
835 address &= PAGE_MASK;
837 memset(cmd, 0, sizeof(*cmd));
838 cmd->data[1] |= domid;
839 cmd->data[2] = lower_32_bits(address);
840 cmd->data[3] = upper_32_bits(address);
841 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
842 if (s) /* size bit - we flush more than one 4kb page */
843 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
844 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
845 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
848 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
849 u64 address, size_t size)
854 pages = iommu_num_pages(address, size, PAGE_SIZE);
859 * If we have to flush more than one page, flush all
860 * TLB entries for this domain
862 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
866 address &= PAGE_MASK;
868 memset(cmd, 0, sizeof(*cmd));
869 cmd->data[0] = devid;
870 cmd->data[0] |= (qdep & 0xff) << 24;
871 cmd->data[1] = devid;
872 cmd->data[2] = lower_32_bits(address);
873 cmd->data[3] = upper_32_bits(address);
874 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
876 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
879 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
880 u64 address, bool size)
882 memset(cmd, 0, sizeof(*cmd));
884 address &= ~(0xfffULL);
886 cmd->data[0] = pasid;
887 cmd->data[1] = domid;
888 cmd->data[2] = lower_32_bits(address);
889 cmd->data[3] = upper_32_bits(address);
890 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
891 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
893 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
894 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
897 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
898 int qdep, u64 address, bool size)
900 memset(cmd, 0, sizeof(*cmd));
902 address &= ~(0xfffULL);
904 cmd->data[0] = devid;
905 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
906 cmd->data[0] |= (qdep & 0xff) << 24;
907 cmd->data[1] = devid;
908 cmd->data[1] |= (pasid & 0xff) << 16;
909 cmd->data[2] = lower_32_bits(address);
910 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
911 cmd->data[3] = upper_32_bits(address);
913 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
914 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
917 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
918 int status, int tag, bool gn)
920 memset(cmd, 0, sizeof(*cmd));
922 cmd->data[0] = devid;
924 cmd->data[1] = pasid;
925 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
927 cmd->data[3] = tag & 0x1ff;
928 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
930 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
933 static void build_inv_all(struct iommu_cmd *cmd)
935 memset(cmd, 0, sizeof(*cmd));
936 CMD_SET_TYPE(cmd, CMD_INV_ALL);
939 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
941 memset(cmd, 0, sizeof(*cmd));
942 cmd->data[0] = devid;
943 CMD_SET_TYPE(cmd, CMD_INV_IRT);
947 * Writes the command to the IOMMUs command buffer and informs the
948 * hardware about the new command.
950 static int iommu_queue_command_sync(struct amd_iommu *iommu,
951 struct iommu_cmd *cmd,
954 u32 left, tail, head, next_tail;
957 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
960 spin_lock_irqsave(&iommu->lock, flags);
962 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
963 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
964 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
965 left = (head - next_tail) % iommu->cmd_buf_size;
968 struct iommu_cmd sync_cmd;
969 volatile u64 sem = 0;
972 build_completion_wait(&sync_cmd, (u64)&sem);
973 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
975 spin_unlock_irqrestore(&iommu->lock, flags);
977 if ((ret = wait_on_sem(&sem)) != 0)
983 copy_cmd_to_buffer(iommu, cmd, tail);
985 /* We need to sync now to make sure all commands are processed */
986 iommu->need_sync = sync;
988 spin_unlock_irqrestore(&iommu->lock, flags);
993 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
995 return iommu_queue_command_sync(iommu, cmd, true);
999 * This function queues a completion wait command into the command
1000 * buffer of an IOMMU
1002 static int iommu_completion_wait(struct amd_iommu *iommu)
1004 struct iommu_cmd cmd;
1005 volatile u64 sem = 0;
1008 if (!iommu->need_sync)
1011 build_completion_wait(&cmd, (u64)&sem);
1013 ret = iommu_queue_command_sync(iommu, &cmd, false);
1017 return wait_on_sem(&sem);
1020 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1022 struct iommu_cmd cmd;
1024 build_inv_dte(&cmd, devid);
1026 return iommu_queue_command(iommu, &cmd);
1029 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1033 for (devid = 0; devid <= 0xffff; ++devid)
1034 iommu_flush_dte(iommu, devid);
1036 iommu_completion_wait(iommu);
1040 * This function uses heavy locking and may disable irqs for some time. But
1041 * this is no issue because it is only called during resume.
1043 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1047 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1048 struct iommu_cmd cmd;
1049 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1051 iommu_queue_command(iommu, &cmd);
1054 iommu_completion_wait(iommu);
1057 static void iommu_flush_all(struct amd_iommu *iommu)
1059 struct iommu_cmd cmd;
1061 build_inv_all(&cmd);
1063 iommu_queue_command(iommu, &cmd);
1064 iommu_completion_wait(iommu);
1067 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1069 struct iommu_cmd cmd;
1071 build_inv_irt(&cmd, devid);
1073 iommu_queue_command(iommu, &cmd);
1076 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1080 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1081 iommu_flush_irt(iommu, devid);
1083 iommu_completion_wait(iommu);
1086 void iommu_flush_all_caches(struct amd_iommu *iommu)
1088 if (iommu_feature(iommu, FEATURE_IA)) {
1089 iommu_flush_all(iommu);
1091 iommu_flush_dte_all(iommu);
1092 iommu_flush_irt_all(iommu);
1093 iommu_flush_tlb_all(iommu);
1098 * Command send function for flushing on-device TLB
1100 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1101 u64 address, size_t size)
1103 struct amd_iommu *iommu;
1104 struct iommu_cmd cmd;
1107 qdep = dev_data->ats.qdep;
1108 iommu = amd_iommu_rlookup_table[dev_data->devid];
1110 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1112 return iommu_queue_command(iommu, &cmd);
1116 * Command send function for invalidating a device table entry
1118 static int device_flush_dte(struct iommu_dev_data *dev_data)
1120 struct amd_iommu *iommu;
1123 iommu = amd_iommu_rlookup_table[dev_data->devid];
1125 ret = iommu_flush_dte(iommu, dev_data->devid);
1129 if (dev_data->ats.enabled)
1130 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1136 * TLB invalidation function which is called from the mapping functions.
1137 * It invalidates a single PTE if the range to flush is within a single
1138 * page. Otherwise it flushes the whole TLB of the IOMMU.
1140 static void __domain_flush_pages(struct protection_domain *domain,
1141 u64 address, size_t size, int pde)
1143 struct iommu_dev_data *dev_data;
1144 struct iommu_cmd cmd;
1147 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1149 for (i = 0; i < amd_iommus_present; ++i) {
1150 if (!domain->dev_iommu[i])
1154 * Devices of this domain are behind this IOMMU
1155 * We need a TLB flush
1157 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1160 list_for_each_entry(dev_data, &domain->dev_list, list) {
1162 if (!dev_data->ats.enabled)
1165 ret |= device_flush_iotlb(dev_data, address, size);
1171 static void domain_flush_pages(struct protection_domain *domain,
1172 u64 address, size_t size)
1174 __domain_flush_pages(domain, address, size, 0);
1177 /* Flush the whole IO/TLB for a given protection domain */
1178 static void domain_flush_tlb(struct protection_domain *domain)
1180 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1183 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1184 static void domain_flush_tlb_pde(struct protection_domain *domain)
1186 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1189 static void domain_flush_complete(struct protection_domain *domain)
1193 for (i = 0; i < amd_iommus_present; ++i) {
1194 if (!domain->dev_iommu[i])
1198 * Devices of this domain are behind this IOMMU
1199 * We need to wait for completion of all commands.
1201 iommu_completion_wait(amd_iommus[i]);
1207 * This function flushes the DTEs for all devices in domain
1209 static void domain_flush_devices(struct protection_domain *domain)
1211 struct iommu_dev_data *dev_data;
1213 list_for_each_entry(dev_data, &domain->dev_list, list)
1214 device_flush_dte(dev_data);
1217 /****************************************************************************
1219 * The functions below are used the create the page table mappings for
1220 * unity mapped regions.
1222 ****************************************************************************/
1225 * This function is used to add another level to an IO page table. Adding
1226 * another level increases the size of the address space by 9 bits to a size up
1229 static bool increase_address_space(struct protection_domain *domain,
1234 if (domain->mode == PAGE_MODE_6_LEVEL)
1235 /* address space already 64 bit large */
1238 pte = (void *)get_zeroed_page(gfp);
1242 *pte = PM_LEVEL_PDE(domain->mode,
1243 virt_to_phys(domain->pt_root));
1244 domain->pt_root = pte;
1246 domain->updated = true;
1251 static u64 *alloc_pte(struct protection_domain *domain,
1252 unsigned long address,
1253 unsigned long page_size,
1260 BUG_ON(!is_power_of_2(page_size));
1262 while (address > PM_LEVEL_SIZE(domain->mode))
1263 increase_address_space(domain, gfp);
1265 level = domain->mode - 1;
1266 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1267 address = PAGE_SIZE_ALIGN(address, page_size);
1268 end_lvl = PAGE_SIZE_LEVEL(page_size);
1270 while (level > end_lvl) {
1271 if (!IOMMU_PTE_PRESENT(*pte)) {
1272 page = (u64 *)get_zeroed_page(gfp);
1275 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1278 /* No level skipping support yet */
1279 if (PM_PTE_LEVEL(*pte) != level)
1284 pte = IOMMU_PTE_PAGE(*pte);
1286 if (pte_page && level == end_lvl)
1289 pte = &pte[PM_LEVEL_INDEX(level, address)];
1296 * This function checks if there is a PTE for a given dma address. If
1297 * there is one, it returns the pointer to it.
1299 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1304 if (address > PM_LEVEL_SIZE(domain->mode))
1307 level = domain->mode - 1;
1308 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1313 if (!IOMMU_PTE_PRESENT(*pte))
1317 if (PM_PTE_LEVEL(*pte) == 0x07) {
1318 unsigned long pte_mask, __pte;
1321 * If we have a series of large PTEs, make
1322 * sure to return a pointer to the first one.
1324 pte_mask = PTE_PAGE_SIZE(*pte);
1325 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1326 __pte = ((unsigned long)pte) & pte_mask;
1328 return (u64 *)__pte;
1331 /* No level skipping support yet */
1332 if (PM_PTE_LEVEL(*pte) != level)
1337 /* Walk to the next level */
1338 pte = IOMMU_PTE_PAGE(*pte);
1339 pte = &pte[PM_LEVEL_INDEX(level, address)];
1346 * Generic mapping functions. It maps a physical address into a DMA
1347 * address space. It allocates the page table pages if necessary.
1348 * In the future it can be extended to a generic mapping function
1349 * supporting all features of AMD IOMMU page tables like level skipping
1350 * and full 64 bit address spaces.
1352 static int iommu_map_page(struct protection_domain *dom,
1353 unsigned long bus_addr,
1354 unsigned long phys_addr,
1356 unsigned long page_size)
1361 if (!(prot & IOMMU_PROT_MASK))
1364 bus_addr = PAGE_ALIGN(bus_addr);
1365 phys_addr = PAGE_ALIGN(phys_addr);
1366 count = PAGE_SIZE_PTE_COUNT(page_size);
1367 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1369 for (i = 0; i < count; ++i)
1370 if (IOMMU_PTE_PRESENT(pte[i]))
1373 if (page_size > PAGE_SIZE) {
1374 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1375 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1377 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1379 if (prot & IOMMU_PROT_IR)
1380 __pte |= IOMMU_PTE_IR;
1381 if (prot & IOMMU_PROT_IW)
1382 __pte |= IOMMU_PTE_IW;
1384 for (i = 0; i < count; ++i)
1392 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1393 unsigned long bus_addr,
1394 unsigned long page_size)
1396 unsigned long long unmap_size, unmapped;
1399 BUG_ON(!is_power_of_2(page_size));
1403 while (unmapped < page_size) {
1405 pte = fetch_pte(dom, bus_addr);
1409 * No PTE for this address
1410 * move forward in 4kb steps
1412 unmap_size = PAGE_SIZE;
1413 } else if (PM_PTE_LEVEL(*pte) == 0) {
1414 /* 4kb PTE found for this address */
1415 unmap_size = PAGE_SIZE;
1420 /* Large PTE found which maps this address */
1421 unmap_size = PTE_PAGE_SIZE(*pte);
1423 /* Only unmap from the first pte in the page */
1424 if ((unmap_size - 1) & bus_addr)
1426 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1427 for (i = 0; i < count; i++)
1431 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1432 unmapped += unmap_size;
1435 BUG_ON(unmapped && !is_power_of_2(unmapped));
1441 * This function checks if a specific unity mapping entry is needed for
1442 * this specific IOMMU.
1444 static int iommu_for_unity_map(struct amd_iommu *iommu,
1445 struct unity_map_entry *entry)
1449 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1450 bdf = amd_iommu_alias_table[i];
1451 if (amd_iommu_rlookup_table[bdf] == iommu)
1459 * This function actually applies the mapping to the page table of the
1462 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1463 struct unity_map_entry *e)
1468 for (addr = e->address_start; addr < e->address_end;
1469 addr += PAGE_SIZE) {
1470 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1475 * if unity mapping is in aperture range mark the page
1476 * as allocated in the aperture
1478 if (addr < dma_dom->aperture_size)
1479 __set_bit(addr >> PAGE_SHIFT,
1480 dma_dom->aperture[0]->bitmap);
1487 * Init the unity mappings for a specific IOMMU in the system
1489 * Basically iterates over all unity mapping entries and applies them to
1490 * the default domain DMA of that IOMMU if necessary.
1492 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1494 struct unity_map_entry *entry;
1497 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1498 if (!iommu_for_unity_map(iommu, entry))
1500 ret = dma_ops_unity_map(iommu->default_dom, entry);
1509 * Inits the unity mappings required for a specific device
1511 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1514 struct unity_map_entry *e;
1517 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1518 if (!(devid >= e->devid_start && devid <= e->devid_end))
1520 ret = dma_ops_unity_map(dma_dom, e);
1528 /****************************************************************************
1530 * The next functions belong to the address allocator for the dma_ops
1531 * interface functions. They work like the allocators in the other IOMMU
1532 * drivers. Its basically a bitmap which marks the allocated pages in
1533 * the aperture. Maybe it could be enhanced in the future to a more
1534 * efficient allocator.
1536 ****************************************************************************/
1539 * The address allocator core functions.
1541 * called with domain->lock held
1545 * Used to reserve address ranges in the aperture (e.g. for exclusion
1548 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1549 unsigned long start_page,
1552 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1554 if (start_page + pages > last_page)
1555 pages = last_page - start_page;
1557 for (i = start_page; i < start_page + pages; ++i) {
1558 int index = i / APERTURE_RANGE_PAGES;
1559 int page = i % APERTURE_RANGE_PAGES;
1560 __set_bit(page, dom->aperture[index]->bitmap);
1565 * This function is used to add a new aperture range to an existing
1566 * aperture in case of dma_ops domain allocation or address allocation
1569 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1570 bool populate, gfp_t gfp)
1572 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1573 struct amd_iommu *iommu;
1574 unsigned long i, old_size;
1576 #ifdef CONFIG_IOMMU_STRESS
1580 if (index >= APERTURE_MAX_RANGES)
1583 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1584 if (!dma_dom->aperture[index])
1587 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1588 if (!dma_dom->aperture[index]->bitmap)
1591 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1594 unsigned long address = dma_dom->aperture_size;
1595 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1596 u64 *pte, *pte_page;
1598 for (i = 0; i < num_ptes; ++i) {
1599 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1604 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1606 address += APERTURE_RANGE_SIZE / 64;
1610 old_size = dma_dom->aperture_size;
1611 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1613 /* Reserve address range used for MSI messages */
1614 if (old_size < MSI_ADDR_BASE_LO &&
1615 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1616 unsigned long spage;
1619 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1620 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1622 dma_ops_reserve_addresses(dma_dom, spage, pages);
1625 /* Initialize the exclusion range if necessary */
1626 for_each_iommu(iommu) {
1627 if (iommu->exclusion_start &&
1628 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1629 && iommu->exclusion_start < dma_dom->aperture_size) {
1630 unsigned long startpage;
1631 int pages = iommu_num_pages(iommu->exclusion_start,
1632 iommu->exclusion_length,
1634 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1635 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1640 * Check for areas already mapped as present in the new aperture
1641 * range and mark those pages as reserved in the allocator. Such
1642 * mappings may already exist as a result of requested unity
1643 * mappings for devices.
1645 for (i = dma_dom->aperture[index]->offset;
1646 i < dma_dom->aperture_size;
1648 u64 *pte = fetch_pte(&dma_dom->domain, i);
1649 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1652 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1655 update_domain(&dma_dom->domain);
1660 update_domain(&dma_dom->domain);
1662 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1664 kfree(dma_dom->aperture[index]);
1665 dma_dom->aperture[index] = NULL;
1670 static unsigned long dma_ops_area_alloc(struct device *dev,
1671 struct dma_ops_domain *dom,
1673 unsigned long align_mask,
1675 unsigned long start)
1677 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1678 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1679 int i = start >> APERTURE_RANGE_SHIFT;
1680 unsigned long boundary_size;
1681 unsigned long address = -1;
1682 unsigned long limit;
1684 next_bit >>= PAGE_SHIFT;
1686 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1687 PAGE_SIZE) >> PAGE_SHIFT;
1689 for (;i < max_index; ++i) {
1690 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1692 if (dom->aperture[i]->offset >= dma_mask)
1695 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1696 dma_mask >> PAGE_SHIFT);
1698 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1699 limit, next_bit, pages, 0,
1700 boundary_size, align_mask);
1701 if (address != -1) {
1702 address = dom->aperture[i]->offset +
1703 (address << PAGE_SHIFT);
1704 dom->next_address = address + (pages << PAGE_SHIFT);
1714 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1715 struct dma_ops_domain *dom,
1717 unsigned long align_mask,
1720 unsigned long address;
1722 #ifdef CONFIG_IOMMU_STRESS
1723 dom->next_address = 0;
1724 dom->need_flush = true;
1727 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1728 dma_mask, dom->next_address);
1730 if (address == -1) {
1731 dom->next_address = 0;
1732 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1734 dom->need_flush = true;
1737 if (unlikely(address == -1))
1738 address = DMA_ERROR_CODE;
1740 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1746 * The address free function.
1748 * called with domain->lock held
1750 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1751 unsigned long address,
1754 unsigned i = address >> APERTURE_RANGE_SHIFT;
1755 struct aperture_range *range = dom->aperture[i];
1757 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1759 #ifdef CONFIG_IOMMU_STRESS
1764 if (address >= dom->next_address)
1765 dom->need_flush = true;
1767 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1769 bitmap_clear(range->bitmap, address, pages);
1773 /****************************************************************************
1775 * The next functions belong to the domain allocation. A domain is
1776 * allocated for every IOMMU as the default domain. If device isolation
1777 * is enabled, every device get its own domain. The most important thing
1778 * about domains is the page table mapping the DMA address space they
1781 ****************************************************************************/
1784 * This function adds a protection domain to the global protection domain list
1786 static void add_domain_to_list(struct protection_domain *domain)
1788 unsigned long flags;
1790 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1791 list_add(&domain->list, &amd_iommu_pd_list);
1792 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1796 * This function removes a protection domain to the global
1797 * protection domain list
1799 static void del_domain_from_list(struct protection_domain *domain)
1801 unsigned long flags;
1803 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1804 list_del(&domain->list);
1805 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1808 static u16 domain_id_alloc(void)
1810 unsigned long flags;
1813 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1814 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1816 if (id > 0 && id < MAX_DOMAIN_ID)
1817 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1820 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1825 static void domain_id_free(int id)
1827 unsigned long flags;
1829 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1830 if (id > 0 && id < MAX_DOMAIN_ID)
1831 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1832 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1835 #define DEFINE_FREE_PT_FN(LVL, FN) \
1836 static void free_pt_##LVL (unsigned long __pt) \
1844 for (i = 0; i < 512; ++i) { \
1845 if (!IOMMU_PTE_PRESENT(pt[i])) \
1848 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1851 free_page((unsigned long)pt); \
1854 DEFINE_FREE_PT_FN(l2, free_page)
1855 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1856 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1857 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1858 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1860 static void free_pagetable(struct protection_domain *domain)
1862 unsigned long root = (unsigned long)domain->pt_root;
1864 switch (domain->mode) {
1865 case PAGE_MODE_NONE:
1867 case PAGE_MODE_1_LEVEL:
1870 case PAGE_MODE_2_LEVEL:
1873 case PAGE_MODE_3_LEVEL:
1876 case PAGE_MODE_4_LEVEL:
1879 case PAGE_MODE_5_LEVEL:
1882 case PAGE_MODE_6_LEVEL:
1890 static void free_gcr3_tbl_level1(u64 *tbl)
1895 for (i = 0; i < 512; ++i) {
1896 if (!(tbl[i] & GCR3_VALID))
1899 ptr = __va(tbl[i] & PAGE_MASK);
1901 free_page((unsigned long)ptr);
1905 static void free_gcr3_tbl_level2(u64 *tbl)
1910 for (i = 0; i < 512; ++i) {
1911 if (!(tbl[i] & GCR3_VALID))
1914 ptr = __va(tbl[i] & PAGE_MASK);
1916 free_gcr3_tbl_level1(ptr);
1920 static void free_gcr3_table(struct protection_domain *domain)
1922 if (domain->glx == 2)
1923 free_gcr3_tbl_level2(domain->gcr3_tbl);
1924 else if (domain->glx == 1)
1925 free_gcr3_tbl_level1(domain->gcr3_tbl);
1926 else if (domain->glx != 0)
1929 free_page((unsigned long)domain->gcr3_tbl);
1933 * Free a domain, only used if something went wrong in the
1934 * allocation path and we need to free an already allocated page table
1936 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1943 del_domain_from_list(&dom->domain);
1945 free_pagetable(&dom->domain);
1947 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1948 if (!dom->aperture[i])
1950 free_page((unsigned long)dom->aperture[i]->bitmap);
1951 kfree(dom->aperture[i]);
1958 * Allocates a new protection domain usable for the dma_ops functions.
1959 * It also initializes the page table and the address allocator data
1960 * structures required for the dma_ops interface
1962 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1964 struct dma_ops_domain *dma_dom;
1966 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1970 spin_lock_init(&dma_dom->domain.lock);
1972 dma_dom->domain.id = domain_id_alloc();
1973 if (dma_dom->domain.id == 0)
1975 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1976 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1977 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1978 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1979 dma_dom->domain.priv = dma_dom;
1980 if (!dma_dom->domain.pt_root)
1983 dma_dom->need_flush = false;
1984 dma_dom->target_dev = 0xffff;
1986 add_domain_to_list(&dma_dom->domain);
1988 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1992 * mark the first page as allocated so we never return 0 as
1993 * a valid dma-address. So we can use 0 as error value
1995 dma_dom->aperture[0]->bitmap[0] = 1;
1996 dma_dom->next_address = 0;
2002 dma_ops_domain_free(dma_dom);
2008 * little helper function to check whether a given protection domain is a
2011 static bool dma_ops_domain(struct protection_domain *domain)
2013 return domain->flags & PD_DMA_OPS_MASK;
2016 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2021 if (domain->mode != PAGE_MODE_NONE)
2022 pte_root = virt_to_phys(domain->pt_root);
2024 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2025 << DEV_ENTRY_MODE_SHIFT;
2026 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2028 flags = amd_iommu_dev_table[devid].data[1];
2031 flags |= DTE_FLAG_IOTLB;
2033 if (domain->flags & PD_IOMMUV2_MASK) {
2034 u64 gcr3 = __pa(domain->gcr3_tbl);
2035 u64 glx = domain->glx;
2038 pte_root |= DTE_FLAG_GV;
2039 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2041 /* First mask out possible old values for GCR3 table */
2042 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2045 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2048 /* Encode GCR3 table into DTE */
2049 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2052 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2055 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2059 flags &= ~(0xffffUL);
2060 flags |= domain->id;
2062 amd_iommu_dev_table[devid].data[1] = flags;
2063 amd_iommu_dev_table[devid].data[0] = pte_root;
2066 static void clear_dte_entry(u16 devid)
2068 /* remove entry from the device table seen by the hardware */
2069 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2070 amd_iommu_dev_table[devid].data[1] = 0;
2072 amd_iommu_apply_erratum_63(devid);
2075 static void do_attach(struct iommu_dev_data *dev_data,
2076 struct protection_domain *domain)
2078 struct amd_iommu *iommu;
2081 iommu = amd_iommu_rlookup_table[dev_data->devid];
2082 ats = dev_data->ats.enabled;
2084 /* Update data structures */
2085 dev_data->domain = domain;
2086 list_add(&dev_data->list, &domain->dev_list);
2087 set_dte_entry(dev_data->devid, domain, ats);
2089 /* Do reference counting */
2090 domain->dev_iommu[iommu->index] += 1;
2091 domain->dev_cnt += 1;
2093 /* Flush the DTE entry */
2094 device_flush_dte(dev_data);
2097 static void do_detach(struct iommu_dev_data *dev_data)
2099 struct amd_iommu *iommu;
2101 iommu = amd_iommu_rlookup_table[dev_data->devid];
2103 /* decrease reference counters */
2104 dev_data->domain->dev_iommu[iommu->index] -= 1;
2105 dev_data->domain->dev_cnt -= 1;
2107 /* Update data structures */
2108 dev_data->domain = NULL;
2109 list_del(&dev_data->list);
2110 clear_dte_entry(dev_data->devid);
2112 /* Flush the DTE entry */
2113 device_flush_dte(dev_data);
2117 * If a device is not yet associated with a domain, this function does
2118 * assigns it visible for the hardware
2120 static int __attach_device(struct iommu_dev_data *dev_data,
2121 struct protection_domain *domain)
2126 spin_lock(&domain->lock);
2128 if (dev_data->alias_data != NULL) {
2129 struct iommu_dev_data *alias_data = dev_data->alias_data;
2131 /* Some sanity checks */
2133 if (alias_data->domain != NULL &&
2134 alias_data->domain != domain)
2137 if (dev_data->domain != NULL &&
2138 dev_data->domain != domain)
2141 /* Do real assignment */
2142 if (alias_data->domain == NULL)
2143 do_attach(alias_data, domain);
2145 atomic_inc(&alias_data->bind);
2148 if (dev_data->domain == NULL)
2149 do_attach(dev_data, domain);
2151 atomic_inc(&dev_data->bind);
2158 spin_unlock(&domain->lock);
2164 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2166 pci_disable_ats(pdev);
2167 pci_disable_pri(pdev);
2168 pci_disable_pasid(pdev);
2171 /* FIXME: Change generic reset-function to do the same */
2172 static int pri_reset_while_enabled(struct pci_dev *pdev)
2177 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2181 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2182 control |= PCI_PRI_CTRL_RESET;
2183 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2188 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2193 /* FIXME: Hardcode number of outstanding requests for now */
2195 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2197 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2199 /* Only allow access to user-accessible pages */
2200 ret = pci_enable_pasid(pdev, 0);
2204 /* First reset the PRI state of the device */
2205 ret = pci_reset_pri(pdev);
2210 ret = pci_enable_pri(pdev, reqs);
2215 ret = pri_reset_while_enabled(pdev);
2220 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2227 pci_disable_pri(pdev);
2228 pci_disable_pasid(pdev);
2233 /* FIXME: Move this to PCI code */
2234 #define PCI_PRI_TLP_OFF (1 << 15)
2236 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2241 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2245 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2247 return (status & PCI_PRI_TLP_OFF) ? true : false;
2251 * If a device is not yet associated with a domain, this function
2252 * assigns it visible for the hardware
2254 static int attach_device(struct device *dev,
2255 struct protection_domain *domain)
2257 struct pci_dev *pdev = to_pci_dev(dev);
2258 struct iommu_dev_data *dev_data;
2259 unsigned long flags;
2262 dev_data = get_dev_data(dev);
2264 if (domain->flags & PD_IOMMUV2_MASK) {
2265 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2268 if (pdev_iommuv2_enable(pdev) != 0)
2271 dev_data->ats.enabled = true;
2272 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2273 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2274 } else if (amd_iommu_iotlb_sup &&
2275 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2276 dev_data->ats.enabled = true;
2277 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2280 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2281 ret = __attach_device(dev_data, domain);
2282 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2285 * We might boot into a crash-kernel here. The crashed kernel
2286 * left the caches in the IOMMU dirty. So we have to flush
2287 * here to evict all dirty stuff.
2289 domain_flush_tlb_pde(domain);
2295 * Removes a device from a protection domain (unlocked)
2297 static void __detach_device(struct iommu_dev_data *dev_data)
2299 struct protection_domain *domain;
2300 unsigned long flags;
2302 BUG_ON(!dev_data->domain);
2304 domain = dev_data->domain;
2306 spin_lock_irqsave(&domain->lock, flags);
2308 if (dev_data->alias_data != NULL) {
2309 struct iommu_dev_data *alias_data = dev_data->alias_data;
2311 if (atomic_dec_and_test(&alias_data->bind))
2312 do_detach(alias_data);
2315 if (atomic_dec_and_test(&dev_data->bind))
2316 do_detach(dev_data);
2318 spin_unlock_irqrestore(&domain->lock, flags);
2321 * If we run in passthrough mode the device must be assigned to the
2322 * passthrough domain if it is detached from any other domain.
2323 * Make sure we can deassign from the pt_domain itself.
2325 if (dev_data->passthrough &&
2326 (dev_data->domain == NULL && domain != pt_domain))
2327 __attach_device(dev_data, pt_domain);
2331 * Removes a device from a protection domain (with devtable_lock held)
2333 static void detach_device(struct device *dev)
2335 struct protection_domain *domain;
2336 struct iommu_dev_data *dev_data;
2337 unsigned long flags;
2339 dev_data = get_dev_data(dev);
2340 domain = dev_data->domain;
2342 /* lock device table */
2343 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2344 __detach_device(dev_data);
2345 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2347 if (domain->flags & PD_IOMMUV2_MASK)
2348 pdev_iommuv2_disable(to_pci_dev(dev));
2349 else if (dev_data->ats.enabled)
2350 pci_disable_ats(to_pci_dev(dev));
2352 dev_data->ats.enabled = false;
2356 * Find out the protection domain structure for a given PCI device. This
2357 * will give us the pointer to the page table root for example.
2359 static struct protection_domain *domain_for_device(struct device *dev)
2361 struct iommu_dev_data *dev_data;
2362 struct protection_domain *dom = NULL;
2363 unsigned long flags;
2365 dev_data = get_dev_data(dev);
2367 if (dev_data->domain)
2368 return dev_data->domain;
2370 if (dev_data->alias_data != NULL) {
2371 struct iommu_dev_data *alias_data = dev_data->alias_data;
2373 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2374 if (alias_data->domain != NULL) {
2375 __attach_device(dev_data, alias_data->domain);
2376 dom = alias_data->domain;
2378 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2384 static int device_change_notifier(struct notifier_block *nb,
2385 unsigned long action, void *data)
2387 struct dma_ops_domain *dma_domain;
2388 struct protection_domain *domain;
2389 struct iommu_dev_data *dev_data;
2390 struct device *dev = data;
2391 struct amd_iommu *iommu;
2392 unsigned long flags;
2395 if (!check_device(dev))
2398 devid = get_device_id(dev);
2399 iommu = amd_iommu_rlookup_table[devid];
2400 dev_data = get_dev_data(dev);
2403 case BUS_NOTIFY_UNBOUND_DRIVER:
2405 domain = domain_for_device(dev);
2409 if (dev_data->passthrough)
2413 case BUS_NOTIFY_ADD_DEVICE:
2415 iommu_init_device(dev);
2416 init_iommu_group(dev);
2419 * dev_data is still NULL and
2420 * got initialized in iommu_init_device
2422 dev_data = get_dev_data(dev);
2424 if (iommu_pass_through || dev_data->iommu_v2) {
2425 dev_data->passthrough = true;
2426 attach_device(dev, pt_domain);
2430 domain = domain_for_device(dev);
2432 /* allocate a protection domain if a device is added */
2433 dma_domain = find_protection_domain(devid);
2435 dma_domain = dma_ops_domain_alloc();
2438 dma_domain->target_dev = devid;
2440 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2441 list_add_tail(&dma_domain->list, &iommu_pd_list);
2442 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2445 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2448 case BUS_NOTIFY_DEL_DEVICE:
2450 iommu_uninit_device(dev);
2456 iommu_completion_wait(iommu);
2462 static struct notifier_block device_nb = {
2463 .notifier_call = device_change_notifier,
2466 void amd_iommu_init_notifier(void)
2468 bus_register_notifier(&pci_bus_type, &device_nb);
2471 /*****************************************************************************
2473 * The next functions belong to the dma_ops mapping/unmapping code.
2475 *****************************************************************************/
2478 * In the dma_ops path we only have the struct device. This function
2479 * finds the corresponding IOMMU, the protection domain and the
2480 * requestor id for a given device.
2481 * If the device is not yet associated with a domain this is also done
2484 static struct protection_domain *get_domain(struct device *dev)
2486 struct protection_domain *domain;
2487 struct dma_ops_domain *dma_dom;
2488 u16 devid = get_device_id(dev);
2490 if (!check_device(dev))
2491 return ERR_PTR(-EINVAL);
2493 domain = domain_for_device(dev);
2494 if (domain != NULL && !dma_ops_domain(domain))
2495 return ERR_PTR(-EBUSY);
2500 /* Device not bound yet - bind it */
2501 dma_dom = find_protection_domain(devid);
2503 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2504 attach_device(dev, &dma_dom->domain);
2505 DUMP_printk("Using protection domain %d for device %s\n",
2506 dma_dom->domain.id, dev_name(dev));
2508 return &dma_dom->domain;
2511 static void update_device_table(struct protection_domain *domain)
2513 struct iommu_dev_data *dev_data;
2515 list_for_each_entry(dev_data, &domain->dev_list, list)
2516 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2519 static void update_domain(struct protection_domain *domain)
2521 if (!domain->updated)
2524 update_device_table(domain);
2526 domain_flush_devices(domain);
2527 domain_flush_tlb_pde(domain);
2529 domain->updated = false;
2533 * This function fetches the PTE for a given address in the aperture
2535 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2536 unsigned long address)
2538 struct aperture_range *aperture;
2539 u64 *pte, *pte_page;
2541 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2545 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2547 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2549 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2551 pte += PM_LEVEL_INDEX(0, address);
2553 update_domain(&dom->domain);
2559 * This is the generic map function. It maps one 4kb page at paddr to
2560 * the given address in the DMA address space for the domain.
2562 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2563 unsigned long address,
2569 WARN_ON(address > dom->aperture_size);
2573 pte = dma_ops_get_pte(dom, address);
2575 return DMA_ERROR_CODE;
2577 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2579 if (direction == DMA_TO_DEVICE)
2580 __pte |= IOMMU_PTE_IR;
2581 else if (direction == DMA_FROM_DEVICE)
2582 __pte |= IOMMU_PTE_IW;
2583 else if (direction == DMA_BIDIRECTIONAL)
2584 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2590 return (dma_addr_t)address;
2594 * The generic unmapping function for on page in the DMA address space.
2596 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2597 unsigned long address)
2599 struct aperture_range *aperture;
2602 if (address >= dom->aperture_size)
2605 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2609 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2613 pte += PM_LEVEL_INDEX(0, address);
2621 * This function contains common code for mapping of a physically
2622 * contiguous memory region into DMA address space. It is used by all
2623 * mapping functions provided with this IOMMU driver.
2624 * Must be called with the domain lock held.
2626 static dma_addr_t __map_single(struct device *dev,
2627 struct dma_ops_domain *dma_dom,
2634 dma_addr_t offset = paddr & ~PAGE_MASK;
2635 dma_addr_t address, start, ret;
2637 unsigned long align_mask = 0;
2640 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2643 INC_STATS_COUNTER(total_map_requests);
2646 INC_STATS_COUNTER(cross_page);
2649 align_mask = (1UL << get_order(size)) - 1;
2652 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2654 if (unlikely(address == DMA_ERROR_CODE)) {
2656 * setting next_address here will let the address
2657 * allocator only scan the new allocated range in the
2658 * first run. This is a small optimization.
2660 dma_dom->next_address = dma_dom->aperture_size;
2662 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2666 * aperture was successfully enlarged by 128 MB, try
2673 for (i = 0; i < pages; ++i) {
2674 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2675 if (ret == DMA_ERROR_CODE)
2683 ADD_STATS_COUNTER(alloced_io_mem, size);
2685 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2686 domain_flush_tlb(&dma_dom->domain);
2687 dma_dom->need_flush = false;
2688 } else if (unlikely(amd_iommu_np_cache))
2689 domain_flush_pages(&dma_dom->domain, address, size);
2696 for (--i; i >= 0; --i) {
2698 dma_ops_domain_unmap(dma_dom, start);
2701 dma_ops_free_addresses(dma_dom, address, pages);
2703 return DMA_ERROR_CODE;
2707 * Does the reverse of the __map_single function. Must be called with
2708 * the domain lock held too
2710 static void __unmap_single(struct dma_ops_domain *dma_dom,
2711 dma_addr_t dma_addr,
2715 dma_addr_t flush_addr;
2716 dma_addr_t i, start;
2719 if ((dma_addr == DMA_ERROR_CODE) ||
2720 (dma_addr + size > dma_dom->aperture_size))
2723 flush_addr = dma_addr;
2724 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2725 dma_addr &= PAGE_MASK;
2728 for (i = 0; i < pages; ++i) {
2729 dma_ops_domain_unmap(dma_dom, start);
2733 SUB_STATS_COUNTER(alloced_io_mem, size);
2735 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2737 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2738 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2739 dma_dom->need_flush = false;
2744 * The exported map_single function for dma_ops.
2746 static dma_addr_t map_page(struct device *dev, struct page *page,
2747 unsigned long offset, size_t size,
2748 enum dma_data_direction dir,
2749 struct dma_attrs *attrs)
2751 unsigned long flags;
2752 struct protection_domain *domain;
2755 phys_addr_t paddr = page_to_phys(page) + offset;
2757 INC_STATS_COUNTER(cnt_map_single);
2759 domain = get_domain(dev);
2760 if (PTR_ERR(domain) == -EINVAL)
2761 return (dma_addr_t)paddr;
2762 else if (IS_ERR(domain))
2763 return DMA_ERROR_CODE;
2765 dma_mask = *dev->dma_mask;
2767 spin_lock_irqsave(&domain->lock, flags);
2769 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2771 if (addr == DMA_ERROR_CODE)
2774 domain_flush_complete(domain);
2777 spin_unlock_irqrestore(&domain->lock, flags);
2783 * The exported unmap_single function for dma_ops.
2785 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2786 enum dma_data_direction dir, struct dma_attrs *attrs)
2788 unsigned long flags;
2789 struct protection_domain *domain;
2791 INC_STATS_COUNTER(cnt_unmap_single);
2793 domain = get_domain(dev);
2797 spin_lock_irqsave(&domain->lock, flags);
2799 __unmap_single(domain->priv, dma_addr, size, dir);
2801 domain_flush_complete(domain);
2803 spin_unlock_irqrestore(&domain->lock, flags);
2807 * The exported map_sg function for dma_ops (handles scatter-gather
2810 static int map_sg(struct device *dev, struct scatterlist *sglist,
2811 int nelems, enum dma_data_direction dir,
2812 struct dma_attrs *attrs)
2814 unsigned long flags;
2815 struct protection_domain *domain;
2817 struct scatterlist *s;
2819 int mapped_elems = 0;
2822 INC_STATS_COUNTER(cnt_map_sg);
2824 domain = get_domain(dev);
2828 dma_mask = *dev->dma_mask;
2830 spin_lock_irqsave(&domain->lock, flags);
2832 for_each_sg(sglist, s, nelems, i) {
2835 s->dma_address = __map_single(dev, domain->priv,
2836 paddr, s->length, dir, false,
2839 if (s->dma_address) {
2840 s->dma_length = s->length;
2846 domain_flush_complete(domain);
2849 spin_unlock_irqrestore(&domain->lock, flags);
2851 return mapped_elems;
2853 for_each_sg(sglist, s, mapped_elems, i) {
2855 __unmap_single(domain->priv, s->dma_address,
2856 s->dma_length, dir);
2857 s->dma_address = s->dma_length = 0;
2866 * The exported map_sg function for dma_ops (handles scatter-gather
2869 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2870 int nelems, enum dma_data_direction dir,
2871 struct dma_attrs *attrs)
2873 unsigned long flags;
2874 struct protection_domain *domain;
2875 struct scatterlist *s;
2878 INC_STATS_COUNTER(cnt_unmap_sg);
2880 domain = get_domain(dev);
2884 spin_lock_irqsave(&domain->lock, flags);
2886 for_each_sg(sglist, s, nelems, i) {
2887 __unmap_single(domain->priv, s->dma_address,
2888 s->dma_length, dir);
2889 s->dma_address = s->dma_length = 0;
2892 domain_flush_complete(domain);
2894 spin_unlock_irqrestore(&domain->lock, flags);
2898 * The exported alloc_coherent function for dma_ops.
2900 static void *alloc_coherent(struct device *dev, size_t size,
2901 dma_addr_t *dma_addr, gfp_t flag,
2902 struct dma_attrs *attrs)
2904 unsigned long flags;
2906 struct protection_domain *domain;
2908 u64 dma_mask = dev->coherent_dma_mask;
2910 INC_STATS_COUNTER(cnt_alloc_coherent);
2912 domain = get_domain(dev);
2913 if (PTR_ERR(domain) == -EINVAL) {
2914 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2915 *dma_addr = __pa(virt_addr);
2917 } else if (IS_ERR(domain))
2920 dma_mask = dev->coherent_dma_mask;
2921 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2924 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2928 paddr = virt_to_phys(virt_addr);
2931 dma_mask = *dev->dma_mask;
2933 spin_lock_irqsave(&domain->lock, flags);
2935 *dma_addr = __map_single(dev, domain->priv, paddr,
2936 size, DMA_BIDIRECTIONAL, true, dma_mask);
2938 if (*dma_addr == DMA_ERROR_CODE) {
2939 spin_unlock_irqrestore(&domain->lock, flags);
2943 domain_flush_complete(domain);
2945 spin_unlock_irqrestore(&domain->lock, flags);
2951 free_pages((unsigned long)virt_addr, get_order(size));
2957 * The exported free_coherent function for dma_ops.
2959 static void free_coherent(struct device *dev, size_t size,
2960 void *virt_addr, dma_addr_t dma_addr,
2961 struct dma_attrs *attrs)
2963 unsigned long flags;
2964 struct protection_domain *domain;
2966 INC_STATS_COUNTER(cnt_free_coherent);
2968 domain = get_domain(dev);
2972 spin_lock_irqsave(&domain->lock, flags);
2974 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2976 domain_flush_complete(domain);
2978 spin_unlock_irqrestore(&domain->lock, flags);
2981 free_pages((unsigned long)virt_addr, get_order(size));
2985 * This function is called by the DMA layer to find out if we can handle a
2986 * particular device. It is part of the dma_ops.
2988 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2990 return check_device(dev);
2994 * The function for pre-allocating protection domains.
2996 * If the driver core informs the DMA layer if a driver grabs a device
2997 * we don't need to preallocate the protection domains anymore.
2998 * For now we have to.
3000 static void __init prealloc_protection_domains(void)
3002 struct iommu_dev_data *dev_data;
3003 struct dma_ops_domain *dma_dom;
3004 struct pci_dev *dev = NULL;
3007 for_each_pci_dev(dev) {
3009 /* Do we handle this device? */
3010 if (!check_device(&dev->dev))
3013 dev_data = get_dev_data(&dev->dev);
3014 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
3015 /* Make sure passthrough domain is allocated */
3016 alloc_passthrough_domain();
3017 dev_data->passthrough = true;
3018 attach_device(&dev->dev, pt_domain);
3019 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3020 dev_name(&dev->dev));
3023 /* Is there already any domain for it? */
3024 if (domain_for_device(&dev->dev))
3027 devid = get_device_id(&dev->dev);
3029 dma_dom = dma_ops_domain_alloc();
3032 init_unity_mappings_for_device(dma_dom, devid);
3033 dma_dom->target_dev = devid;
3035 attach_device(&dev->dev, &dma_dom->domain);
3037 list_add_tail(&dma_dom->list, &iommu_pd_list);
3041 static struct dma_map_ops amd_iommu_dma_ops = {
3042 .alloc = alloc_coherent,
3043 .free = free_coherent,
3044 .map_page = map_page,
3045 .unmap_page = unmap_page,
3047 .unmap_sg = unmap_sg,
3048 .dma_supported = amd_iommu_dma_supported,
3051 static unsigned device_dma_ops_init(void)
3053 struct iommu_dev_data *dev_data;
3054 struct pci_dev *pdev = NULL;
3055 unsigned unhandled = 0;
3057 for_each_pci_dev(pdev) {
3058 if (!check_device(&pdev->dev)) {
3060 iommu_ignore_device(&pdev->dev);
3066 dev_data = get_dev_data(&pdev->dev);
3068 if (!dev_data->passthrough)
3069 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
3071 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3078 * The function which clues the AMD IOMMU driver into dma_ops.
3081 void __init amd_iommu_init_api(void)
3083 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3086 int __init amd_iommu_init_dma_ops(void)
3088 struct amd_iommu *iommu;
3092 * first allocate a default protection domain for every IOMMU we
3093 * found in the system. Devices not assigned to any other
3094 * protection domain will be assigned to the default one.
3096 for_each_iommu(iommu) {
3097 iommu->default_dom = dma_ops_domain_alloc();
3098 if (iommu->default_dom == NULL)
3100 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3101 ret = iommu_init_unity_mappings(iommu);
3107 * Pre-allocate the protection domains for each device.
3109 prealloc_protection_domains();
3114 /* Make the driver finally visible to the drivers */
3115 unhandled = device_dma_ops_init();
3116 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3117 /* There are unhandled devices - initialize swiotlb for them */
3121 amd_iommu_stats_init();
3123 if (amd_iommu_unmap_flush)
3124 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3126 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3132 for_each_iommu(iommu) {
3133 dma_ops_domain_free(iommu->default_dom);
3139 /*****************************************************************************
3141 * The following functions belong to the exported interface of AMD IOMMU
3143 * This interface allows access to lower level functions of the IOMMU
3144 * like protection domain handling and assignement of devices to domains
3145 * which is not possible with the dma_ops interface.
3147 *****************************************************************************/
3149 static void cleanup_domain(struct protection_domain *domain)
3151 struct iommu_dev_data *entry;
3152 unsigned long flags;
3154 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3156 while (!list_empty(&domain->dev_list)) {
3157 entry = list_first_entry(&domain->dev_list,
3158 struct iommu_dev_data, list);
3159 __detach_device(entry);
3160 atomic_set(&entry->bind, 0);
3163 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3166 static void protection_domain_free(struct protection_domain *domain)
3171 del_domain_from_list(domain);
3174 domain_id_free(domain->id);
3179 static struct protection_domain *protection_domain_alloc(void)
3181 struct protection_domain *domain;
3183 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3187 spin_lock_init(&domain->lock);
3188 mutex_init(&domain->api_lock);
3189 domain->id = domain_id_alloc();
3192 INIT_LIST_HEAD(&domain->dev_list);
3194 add_domain_to_list(domain);
3204 static int __init alloc_passthrough_domain(void)
3206 if (pt_domain != NULL)
3209 /* allocate passthrough domain */
3210 pt_domain = protection_domain_alloc();
3214 pt_domain->mode = PAGE_MODE_NONE;
3218 static int amd_iommu_domain_init(struct iommu_domain *dom)
3220 struct protection_domain *domain;
3222 domain = protection_domain_alloc();
3226 domain->mode = PAGE_MODE_3_LEVEL;
3227 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3228 if (!domain->pt_root)
3231 domain->iommu_domain = dom;
3235 dom->geometry.aperture_start = 0;
3236 dom->geometry.aperture_end = ~0ULL;
3237 dom->geometry.force_aperture = true;
3242 protection_domain_free(domain);
3247 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3249 struct protection_domain *domain = dom->priv;
3254 if (domain->dev_cnt > 0)
3255 cleanup_domain(domain);
3257 BUG_ON(domain->dev_cnt != 0);
3259 if (domain->mode != PAGE_MODE_NONE)
3260 free_pagetable(domain);
3262 if (domain->flags & PD_IOMMUV2_MASK)
3263 free_gcr3_table(domain);
3265 protection_domain_free(domain);
3270 static void amd_iommu_detach_device(struct iommu_domain *dom,
3273 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3274 struct amd_iommu *iommu;
3277 if (!check_device(dev))
3280 devid = get_device_id(dev);
3282 if (dev_data->domain != NULL)
3285 iommu = amd_iommu_rlookup_table[devid];
3289 iommu_completion_wait(iommu);
3292 static int amd_iommu_attach_device(struct iommu_domain *dom,
3295 struct protection_domain *domain = dom->priv;
3296 struct iommu_dev_data *dev_data;
3297 struct amd_iommu *iommu;
3300 if (!check_device(dev))
3303 dev_data = dev->archdata.iommu;
3305 iommu = amd_iommu_rlookup_table[dev_data->devid];
3309 if (dev_data->domain)
3312 ret = attach_device(dev, domain);
3314 iommu_completion_wait(iommu);
3319 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3320 phys_addr_t paddr, size_t page_size, int iommu_prot)
3322 struct protection_domain *domain = dom->priv;
3326 if (domain->mode == PAGE_MODE_NONE)
3329 if (iommu_prot & IOMMU_READ)
3330 prot |= IOMMU_PROT_IR;
3331 if (iommu_prot & IOMMU_WRITE)
3332 prot |= IOMMU_PROT_IW;
3334 mutex_lock(&domain->api_lock);
3335 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3336 mutex_unlock(&domain->api_lock);
3341 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3344 struct protection_domain *domain = dom->priv;
3347 if (domain->mode == PAGE_MODE_NONE)
3350 mutex_lock(&domain->api_lock);
3351 unmap_size = iommu_unmap_page(domain, iova, page_size);
3352 mutex_unlock(&domain->api_lock);
3354 domain_flush_tlb_pde(domain);
3359 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3362 struct protection_domain *domain = dom->priv;
3363 unsigned long offset_mask;
3367 if (domain->mode == PAGE_MODE_NONE)
3370 pte = fetch_pte(domain, iova);
3372 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3375 if (PM_PTE_LEVEL(*pte) == 0)
3376 offset_mask = PAGE_SIZE - 1;
3378 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3380 __pte = *pte & PM_ADDR_MASK;
3381 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3386 static bool amd_iommu_capable(enum iommu_cap cap)
3389 case IOMMU_CAP_CACHE_COHERENCY:
3391 case IOMMU_CAP_INTR_REMAP:
3392 return (irq_remapping_enabled == 1);
3398 static const struct iommu_ops amd_iommu_ops = {
3399 .capable = amd_iommu_capable,
3400 .domain_init = amd_iommu_domain_init,
3401 .domain_destroy = amd_iommu_domain_destroy,
3402 .attach_dev = amd_iommu_attach_device,
3403 .detach_dev = amd_iommu_detach_device,
3404 .map = amd_iommu_map,
3405 .unmap = amd_iommu_unmap,
3406 .iova_to_phys = amd_iommu_iova_to_phys,
3407 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3410 /*****************************************************************************
3412 * The next functions do a basic initialization of IOMMU for pass through
3415 * In passthrough mode the IOMMU is initialized and enabled but not used for
3416 * DMA-API translation.
3418 *****************************************************************************/
3420 int __init amd_iommu_init_passthrough(void)
3422 struct iommu_dev_data *dev_data;
3423 struct pci_dev *dev = NULL;
3426 ret = alloc_passthrough_domain();
3430 for_each_pci_dev(dev) {
3431 if (!check_device(&dev->dev))
3434 dev_data = get_dev_data(&dev->dev);
3435 dev_data->passthrough = true;
3437 attach_device(&dev->dev, pt_domain);
3440 amd_iommu_stats_init();
3442 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3447 /* IOMMUv2 specific functions */
3448 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3450 return atomic_notifier_chain_register(&ppr_notifier, nb);
3452 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3454 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3456 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3458 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3460 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3462 struct protection_domain *domain = dom->priv;
3463 unsigned long flags;
3465 spin_lock_irqsave(&domain->lock, flags);
3467 /* Update data structure */
3468 domain->mode = PAGE_MODE_NONE;
3469 domain->updated = true;
3471 /* Make changes visible to IOMMUs */
3472 update_domain(domain);
3474 /* Page-table is not visible to IOMMU anymore, so free it */
3475 free_pagetable(domain);
3477 spin_unlock_irqrestore(&domain->lock, flags);
3479 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3481 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3483 struct protection_domain *domain = dom->priv;
3484 unsigned long flags;
3487 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3490 /* Number of GCR3 table levels required */
3491 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3494 if (levels > amd_iommu_max_glx_val)
3497 spin_lock_irqsave(&domain->lock, flags);
3500 * Save us all sanity checks whether devices already in the
3501 * domain support IOMMUv2. Just force that the domain has no
3502 * devices attached when it is switched into IOMMUv2 mode.
3505 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3509 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3510 if (domain->gcr3_tbl == NULL)
3513 domain->glx = levels;
3514 domain->flags |= PD_IOMMUV2_MASK;
3515 domain->updated = true;
3517 update_domain(domain);
3522 spin_unlock_irqrestore(&domain->lock, flags);
3526 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3528 static int __flush_pasid(struct protection_domain *domain, int pasid,
3529 u64 address, bool size)
3531 struct iommu_dev_data *dev_data;
3532 struct iommu_cmd cmd;
3535 if (!(domain->flags & PD_IOMMUV2_MASK))
3538 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3541 * IOMMU TLB needs to be flushed before Device TLB to
3542 * prevent device TLB refill from IOMMU TLB
3544 for (i = 0; i < amd_iommus_present; ++i) {
3545 if (domain->dev_iommu[i] == 0)
3548 ret = iommu_queue_command(amd_iommus[i], &cmd);
3553 /* Wait until IOMMU TLB flushes are complete */
3554 domain_flush_complete(domain);
3556 /* Now flush device TLBs */
3557 list_for_each_entry(dev_data, &domain->dev_list, list) {
3558 struct amd_iommu *iommu;
3561 BUG_ON(!dev_data->ats.enabled);
3563 qdep = dev_data->ats.qdep;
3564 iommu = amd_iommu_rlookup_table[dev_data->devid];
3566 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3567 qdep, address, size);
3569 ret = iommu_queue_command(iommu, &cmd);
3574 /* Wait until all device TLBs are flushed */
3575 domain_flush_complete(domain);
3584 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3587 INC_STATS_COUNTER(invalidate_iotlb);
3589 return __flush_pasid(domain, pasid, address, false);
3592 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3595 struct protection_domain *domain = dom->priv;
3596 unsigned long flags;
3599 spin_lock_irqsave(&domain->lock, flags);
3600 ret = __amd_iommu_flush_page(domain, pasid, address);
3601 spin_unlock_irqrestore(&domain->lock, flags);
3605 EXPORT_SYMBOL(amd_iommu_flush_page);
3607 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3609 INC_STATS_COUNTER(invalidate_iotlb_all);
3611 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3615 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3617 struct protection_domain *domain = dom->priv;
3618 unsigned long flags;
3621 spin_lock_irqsave(&domain->lock, flags);
3622 ret = __amd_iommu_flush_tlb(domain, pasid);
3623 spin_unlock_irqrestore(&domain->lock, flags);
3627 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3629 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3636 index = (pasid >> (9 * level)) & 0x1ff;
3642 if (!(*pte & GCR3_VALID)) {
3646 root = (void *)get_zeroed_page(GFP_ATOMIC);
3650 *pte = __pa(root) | GCR3_VALID;
3653 root = __va(*pte & PAGE_MASK);
3661 static int __set_gcr3(struct protection_domain *domain, int pasid,
3666 if (domain->mode != PAGE_MODE_NONE)
3669 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3673 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3675 return __amd_iommu_flush_tlb(domain, pasid);
3678 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3682 if (domain->mode != PAGE_MODE_NONE)
3685 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3691 return __amd_iommu_flush_tlb(domain, pasid);
3694 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3697 struct protection_domain *domain = dom->priv;
3698 unsigned long flags;
3701 spin_lock_irqsave(&domain->lock, flags);
3702 ret = __set_gcr3(domain, pasid, cr3);
3703 spin_unlock_irqrestore(&domain->lock, flags);
3707 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3709 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3711 struct protection_domain *domain = dom->priv;
3712 unsigned long flags;
3715 spin_lock_irqsave(&domain->lock, flags);
3716 ret = __clear_gcr3(domain, pasid);
3717 spin_unlock_irqrestore(&domain->lock, flags);
3721 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3723 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3724 int status, int tag)
3726 struct iommu_dev_data *dev_data;
3727 struct amd_iommu *iommu;
3728 struct iommu_cmd cmd;
3730 INC_STATS_COUNTER(complete_ppr);
3732 dev_data = get_dev_data(&pdev->dev);
3733 iommu = amd_iommu_rlookup_table[dev_data->devid];
3735 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3736 tag, dev_data->pri_tlp);
3738 return iommu_queue_command(iommu, &cmd);
3740 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3742 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3744 struct protection_domain *domain;
3746 domain = get_domain(&pdev->dev);
3750 /* Only return IOMMUv2 domains */
3751 if (!(domain->flags & PD_IOMMUV2_MASK))
3754 return domain->iommu_domain;
3756 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3758 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3760 struct iommu_dev_data *dev_data;
3762 if (!amd_iommu_v2_supported())
3765 dev_data = get_dev_data(&pdev->dev);
3766 dev_data->errata |= (1 << erratum);
3768 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3770 int amd_iommu_device_info(struct pci_dev *pdev,
3771 struct amd_iommu_device_info *info)
3776 if (pdev == NULL || info == NULL)
3779 if (!amd_iommu_v2_supported())
3782 memset(info, 0, sizeof(*info));
3784 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3786 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3788 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3790 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3792 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3796 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3797 max_pasids = min(max_pasids, (1 << 20));
3799 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3800 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3802 features = pci_pasid_features(pdev);
3803 if (features & PCI_PASID_CAP_EXEC)
3804 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3805 if (features & PCI_PASID_CAP_PRIV)
3806 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3811 EXPORT_SYMBOL(amd_iommu_device_info);
3813 #ifdef CONFIG_IRQ_REMAP
3815 /*****************************************************************************
3817 * Interrupt Remapping Implementation
3819 *****************************************************************************/
3836 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3837 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3838 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3839 #define DTE_IRQ_REMAP_ENABLE 1ULL
3841 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3845 dte = amd_iommu_dev_table[devid].data[2];
3846 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3847 dte |= virt_to_phys(table->table);
3848 dte |= DTE_IRQ_REMAP_INTCTL;
3849 dte |= DTE_IRQ_TABLE_LEN;
3850 dte |= DTE_IRQ_REMAP_ENABLE;
3852 amd_iommu_dev_table[devid].data[2] = dte;
3855 #define IRTE_ALLOCATED (~1U)
3857 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3859 struct irq_remap_table *table = NULL;
3860 struct amd_iommu *iommu;
3861 unsigned long flags;
3864 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3866 iommu = amd_iommu_rlookup_table[devid];
3870 table = irq_lookup_table[devid];
3874 alias = amd_iommu_alias_table[devid];
3875 table = irq_lookup_table[alias];
3877 irq_lookup_table[devid] = table;
3878 set_dte_irq_entry(devid, table);
3879 iommu_flush_dte(iommu, devid);
3883 /* Nothing there yet, allocate new irq remapping table */
3884 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3888 /* Initialize table spin-lock */
3889 spin_lock_init(&table->lock);
3892 /* Keep the first 32 indexes free for IOAPIC interrupts */
3893 table->min_index = 32;
3895 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3896 if (!table->table) {
3902 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3907 for (i = 0; i < 32; ++i)
3908 table->table[i] = IRTE_ALLOCATED;
3911 irq_lookup_table[devid] = table;
3912 set_dte_irq_entry(devid, table);
3913 iommu_flush_dte(iommu, devid);
3914 if (devid != alias) {
3915 irq_lookup_table[alias] = table;
3916 set_dte_irq_entry(alias, table);
3917 iommu_flush_dte(iommu, alias);
3921 iommu_completion_wait(iommu);
3924 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3929 static int alloc_irq_index(struct irq_cfg *cfg, u16 devid, int count)
3931 struct irq_remap_table *table;
3932 unsigned long flags;
3935 table = get_irq_table(devid, false);
3939 spin_lock_irqsave(&table->lock, flags);
3941 /* Scan table for free entries */
3942 for (c = 0, index = table->min_index;
3943 index < MAX_IRQS_PER_TABLE;
3945 if (table->table[index] == 0)
3951 struct irq_2_irte *irte_info;
3954 table->table[index - c + 1] = IRTE_ALLOCATED;
3959 irte_info = &cfg->irq_2_irte;
3960 irte_info->devid = devid;
3961 irte_info->index = index;
3970 spin_unlock_irqrestore(&table->lock, flags);
3975 static int get_irte(u16 devid, int index, union irte *irte)
3977 struct irq_remap_table *table;
3978 unsigned long flags;
3980 table = get_irq_table(devid, false);
3984 spin_lock_irqsave(&table->lock, flags);
3985 irte->val = table->table[index];
3986 spin_unlock_irqrestore(&table->lock, flags);
3991 static int modify_irte(u16 devid, int index, union irte irte)
3993 struct irq_remap_table *table;
3994 struct amd_iommu *iommu;
3995 unsigned long flags;
3997 iommu = amd_iommu_rlookup_table[devid];
4001 table = get_irq_table(devid, false);
4005 spin_lock_irqsave(&table->lock, flags);
4006 table->table[index] = irte.val;
4007 spin_unlock_irqrestore(&table->lock, flags);
4009 iommu_flush_irt(iommu, devid);
4010 iommu_completion_wait(iommu);
4015 static void free_irte(u16 devid, int index)
4017 struct irq_remap_table *table;
4018 struct amd_iommu *iommu;
4019 unsigned long flags;
4021 iommu = amd_iommu_rlookup_table[devid];
4025 table = get_irq_table(devid, false);
4029 spin_lock_irqsave(&table->lock, flags);
4030 table->table[index] = 0;
4031 spin_unlock_irqrestore(&table->lock, flags);
4033 iommu_flush_irt(iommu, devid);
4034 iommu_completion_wait(iommu);
4037 static int setup_ioapic_entry(int irq, struct IO_APIC_route_entry *entry,
4038 unsigned int destination, int vector,
4039 struct io_apic_irq_attr *attr)
4041 struct irq_remap_table *table;
4042 struct irq_2_irte *irte_info;
4043 struct irq_cfg *cfg;
4050 cfg = irq_get_chip_data(irq);
4054 irte_info = &cfg->irq_2_irte;
4055 ioapic_id = mpc_ioapic_id(attr->ioapic);
4056 devid = get_ioapic_devid(ioapic_id);
4061 table = get_irq_table(devid, true);
4065 index = attr->ioapic_pin;
4067 /* Setup IRQ remapping info */
4069 irte_info->devid = devid;
4070 irte_info->index = index;
4072 /* Setup IRTE for IOMMU */
4074 irte.fields.vector = vector;
4075 irte.fields.int_type = apic->irq_delivery_mode;
4076 irte.fields.destination = destination;
4077 irte.fields.dm = apic->irq_dest_mode;
4078 irte.fields.valid = 1;
4080 ret = modify_irte(devid, index, irte);
4084 /* Setup IOAPIC entry */
4085 memset(entry, 0, sizeof(*entry));
4087 entry->vector = index;
4089 entry->trigger = attr->trigger;
4090 entry->polarity = attr->polarity;
4093 * Mask level triggered irqs.
4101 static int set_affinity(struct irq_data *data, const struct cpumask *mask,
4104 struct irq_2_irte *irte_info;
4105 unsigned int dest, irq;
4106 struct irq_cfg *cfg;
4110 if (!config_enabled(CONFIG_SMP))
4113 cfg = data->chip_data;
4115 irte_info = &cfg->irq_2_irte;
4117 if (!cpumask_intersects(mask, cpu_online_mask))
4120 if (get_irte(irte_info->devid, irte_info->index, &irte))
4123 if (assign_irq_vector(irq, cfg, mask))
4126 err = apic->cpu_mask_to_apicid_and(cfg->domain, mask, &dest);
4128 if (assign_irq_vector(irq, cfg, data->affinity))
4129 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq);
4133 irte.fields.vector = cfg->vector;
4134 irte.fields.destination = dest;
4136 modify_irte(irte_info->devid, irte_info->index, irte);
4138 if (cfg->move_in_progress)
4139 send_cleanup_vector(cfg);
4141 cpumask_copy(data->affinity, mask);
4146 static int free_irq(int irq)
4148 struct irq_2_irte *irte_info;
4149 struct irq_cfg *cfg;
4151 cfg = irq_get_chip_data(irq);
4155 irte_info = &cfg->irq_2_irte;
4157 free_irte(irte_info->devid, irte_info->index);
4162 static void compose_msi_msg(struct pci_dev *pdev,
4163 unsigned int irq, unsigned int dest,
4164 struct msi_msg *msg, u8 hpet_id)
4166 struct irq_2_irte *irte_info;
4167 struct irq_cfg *cfg;
4170 cfg = irq_get_chip_data(irq);
4174 irte_info = &cfg->irq_2_irte;
4177 irte.fields.vector = cfg->vector;
4178 irte.fields.int_type = apic->irq_delivery_mode;
4179 irte.fields.destination = dest;
4180 irte.fields.dm = apic->irq_dest_mode;
4181 irte.fields.valid = 1;
4183 modify_irte(irte_info->devid, irte_info->index, irte);
4185 msg->address_hi = MSI_ADDR_BASE_HI;
4186 msg->address_lo = MSI_ADDR_BASE_LO;
4187 msg->data = irte_info->index;
4190 static int msi_alloc_irq(struct pci_dev *pdev, int irq, int nvec)
4192 struct irq_cfg *cfg;
4199 cfg = irq_get_chip_data(irq);
4203 devid = get_device_id(&pdev->dev);
4204 index = alloc_irq_index(cfg, devid, nvec);
4206 return index < 0 ? MAX_IRQS_PER_TABLE : index;
4209 static int msi_setup_irq(struct pci_dev *pdev, unsigned int irq,
4210 int index, int offset)
4212 struct irq_2_irte *irte_info;
4213 struct irq_cfg *cfg;
4219 cfg = irq_get_chip_data(irq);
4223 if (index >= MAX_IRQS_PER_TABLE)
4226 devid = get_device_id(&pdev->dev);
4227 irte_info = &cfg->irq_2_irte;
4230 irte_info->devid = devid;
4231 irte_info->index = index + offset;
4236 static int alloc_hpet_msi(unsigned int irq, unsigned int id)
4238 struct irq_2_irte *irte_info;
4239 struct irq_cfg *cfg;
4242 cfg = irq_get_chip_data(irq);
4246 irte_info = &cfg->irq_2_irte;
4247 devid = get_hpet_devid(id);
4251 index = alloc_irq_index(cfg, devid, 1);
4256 irte_info->devid = devid;
4257 irte_info->index = index;
4262 struct irq_remap_ops amd_iommu_irq_ops = {
4263 .supported = amd_iommu_supported,
4264 .prepare = amd_iommu_prepare,
4265 .enable = amd_iommu_enable,
4266 .disable = amd_iommu_disable,
4267 .reenable = amd_iommu_reenable,
4268 .enable_faulting = amd_iommu_enable_faulting,
4269 .setup_ioapic_entry = setup_ioapic_entry,
4270 .set_affinity = set_affinity,
4271 .free_irq = free_irq,
4272 .compose_msi_msg = compose_msi_msg,
4273 .msi_alloc_irq = msi_alloc_irq,
4274 .msi_setup_irq = msi_setup_irq,
4275 .alloc_hpet_msi = alloc_hpet_msi,