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 <asm/msidef.h>
35 #include <asm/proto.h>
36 #include <asm/iommu.h>
40 #include "amd_iommu_proto.h"
41 #include "amd_iommu_types.h"
43 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
45 #define LOOP_TIMEOUT 100000
48 * This bitmap is used to advertise the page sizes our hardware support
49 * to the IOMMU core, which will then use this information to split
50 * physically contiguous memory regions it is mapping into page sizes
53 * Traditionally the IOMMU core just handed us the mappings directly,
54 * after making sure the size is an order of a 4KiB page and that the
55 * mapping has natural alignment.
57 * To retain this behavior, we currently advertise that we support
58 * all page sizes that are an order of 4KiB.
60 * If at some point we'd like to utilize the IOMMU core's new behavior,
61 * we could change this to advertise the real page sizes we support.
63 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
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);
76 * Domain for untranslated devices - only allocated
77 * if iommu=pt passed on kernel cmd line.
79 static struct protection_domain *pt_domain;
81 static struct iommu_ops amd_iommu_ops;
83 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
84 int amd_iommu_max_glx_val = -1;
87 * general struct to manage commands send to an IOMMU
93 static void update_domain(struct protection_domain *domain);
94 static int __init alloc_passthrough_domain(void);
96 /****************************************************************************
100 ****************************************************************************/
102 static struct iommu_dev_data *alloc_dev_data(u16 devid)
104 struct iommu_dev_data *dev_data;
107 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
111 dev_data->devid = devid;
112 atomic_set(&dev_data->bind, 0);
114 spin_lock_irqsave(&dev_data_list_lock, flags);
115 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
116 spin_unlock_irqrestore(&dev_data_list_lock, flags);
121 static void free_dev_data(struct iommu_dev_data *dev_data)
125 spin_lock_irqsave(&dev_data_list_lock, flags);
126 list_del(&dev_data->dev_data_list);
127 spin_unlock_irqrestore(&dev_data_list_lock, flags);
132 static struct iommu_dev_data *search_dev_data(u16 devid)
134 struct iommu_dev_data *dev_data;
137 spin_lock_irqsave(&dev_data_list_lock, flags);
138 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
139 if (dev_data->devid == devid)
146 spin_unlock_irqrestore(&dev_data_list_lock, flags);
151 static struct iommu_dev_data *find_dev_data(u16 devid)
153 struct iommu_dev_data *dev_data;
155 dev_data = search_dev_data(devid);
157 if (dev_data == NULL)
158 dev_data = alloc_dev_data(devid);
163 static inline u16 get_device_id(struct device *dev)
165 struct pci_dev *pdev = to_pci_dev(dev);
167 return calc_devid(pdev->bus->number, pdev->devfn);
170 static struct iommu_dev_data *get_dev_data(struct device *dev)
172 return dev->archdata.iommu;
175 static bool pci_iommuv2_capable(struct pci_dev *pdev)
177 static const int caps[] = {
180 PCI_EXT_CAP_ID_PASID,
184 for (i = 0; i < 3; ++i) {
185 pos = pci_find_ext_capability(pdev, caps[i]);
193 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
195 struct iommu_dev_data *dev_data;
197 dev_data = get_dev_data(&pdev->dev);
199 return dev_data->errata & (1 << erratum) ? true : false;
203 * In this function the list of preallocated protection domains is traversed to
204 * find the domain for a specific device
206 static struct dma_ops_domain *find_protection_domain(u16 devid)
208 struct dma_ops_domain *entry, *ret = NULL;
210 u16 alias = amd_iommu_alias_table[devid];
212 if (list_empty(&iommu_pd_list))
215 spin_lock_irqsave(&iommu_pd_list_lock, flags);
217 list_for_each_entry(entry, &iommu_pd_list, list) {
218 if (entry->target_dev == devid ||
219 entry->target_dev == alias) {
225 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
231 * This function checks if the driver got a valid device from the caller to
232 * avoid dereferencing invalid pointers.
234 static bool check_device(struct device *dev)
238 if (!dev || !dev->dma_mask)
241 /* No device or no PCI device */
242 if (dev->bus != &pci_bus_type)
245 devid = get_device_id(dev);
247 /* Out of our scope? */
248 if (devid > amd_iommu_last_bdf)
251 if (amd_iommu_rlookup_table[devid] == NULL)
257 static void swap_pci_ref(struct pci_dev **from, struct pci_dev *to)
263 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
265 static int iommu_init_device(struct device *dev)
267 struct pci_dev *dma_pdev, *pdev = to_pci_dev(dev);
268 struct iommu_dev_data *dev_data;
269 struct iommu_group *group;
273 if (dev->archdata.iommu)
276 dev_data = find_dev_data(get_device_id(dev));
280 alias = amd_iommu_alias_table[dev_data->devid];
281 if (alias != dev_data->devid) {
282 struct iommu_dev_data *alias_data;
284 alias_data = find_dev_data(alias);
285 if (alias_data == NULL) {
286 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
288 free_dev_data(dev_data);
291 dev_data->alias_data = alias_data;
293 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
295 dma_pdev = pci_dev_get(pdev);
297 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
299 if (dma_pdev->multifunction &&
300 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
301 swap_pci_ref(&dma_pdev,
302 pci_get_slot(dma_pdev->bus,
303 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
306 while (!pci_is_root_bus(dma_pdev->bus)) {
307 if (pci_acs_path_enabled(dma_pdev->bus->self,
308 NULL, REQ_ACS_FLAGS))
311 swap_pci_ref(&dma_pdev, pci_dev_get(dma_pdev->bus->self));
314 group = iommu_group_get(&dma_pdev->dev);
315 pci_dev_put(dma_pdev);
317 group = iommu_group_alloc();
319 return PTR_ERR(group);
322 ret = iommu_group_add_device(group, dev);
324 iommu_group_put(group);
329 if (pci_iommuv2_capable(pdev)) {
330 struct amd_iommu *iommu;
332 iommu = amd_iommu_rlookup_table[dev_data->devid];
333 dev_data->iommu_v2 = iommu->is_iommu_v2;
336 dev->archdata.iommu = dev_data;
341 static void iommu_ignore_device(struct device *dev)
345 devid = get_device_id(dev);
346 alias = amd_iommu_alias_table[devid];
348 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
349 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
351 amd_iommu_rlookup_table[devid] = NULL;
352 amd_iommu_rlookup_table[alias] = NULL;
355 static void iommu_uninit_device(struct device *dev)
357 iommu_group_remove_device(dev);
360 * Nothing to do here - we keep dev_data around for unplugged devices
361 * and reuse it when the device is re-plugged - not doing so would
362 * introduce a ton of races.
366 void __init amd_iommu_uninit_devices(void)
368 struct iommu_dev_data *dev_data, *n;
369 struct pci_dev *pdev = NULL;
371 for_each_pci_dev(pdev) {
373 if (!check_device(&pdev->dev))
376 iommu_uninit_device(&pdev->dev);
379 /* Free all of our dev_data structures */
380 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
381 free_dev_data(dev_data);
384 int __init amd_iommu_init_devices(void)
386 struct pci_dev *pdev = NULL;
389 for_each_pci_dev(pdev) {
391 if (!check_device(&pdev->dev))
394 ret = iommu_init_device(&pdev->dev);
395 if (ret == -ENOTSUPP)
396 iommu_ignore_device(&pdev->dev);
405 amd_iommu_uninit_devices();
409 #ifdef CONFIG_AMD_IOMMU_STATS
412 * Initialization code for statistics collection
415 DECLARE_STATS_COUNTER(compl_wait);
416 DECLARE_STATS_COUNTER(cnt_map_single);
417 DECLARE_STATS_COUNTER(cnt_unmap_single);
418 DECLARE_STATS_COUNTER(cnt_map_sg);
419 DECLARE_STATS_COUNTER(cnt_unmap_sg);
420 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
421 DECLARE_STATS_COUNTER(cnt_free_coherent);
422 DECLARE_STATS_COUNTER(cross_page);
423 DECLARE_STATS_COUNTER(domain_flush_single);
424 DECLARE_STATS_COUNTER(domain_flush_all);
425 DECLARE_STATS_COUNTER(alloced_io_mem);
426 DECLARE_STATS_COUNTER(total_map_requests);
427 DECLARE_STATS_COUNTER(complete_ppr);
428 DECLARE_STATS_COUNTER(invalidate_iotlb);
429 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
430 DECLARE_STATS_COUNTER(pri_requests);
433 static struct dentry *stats_dir;
434 static struct dentry *de_fflush;
436 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
438 if (stats_dir == NULL)
441 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
445 static void amd_iommu_stats_init(void)
447 stats_dir = debugfs_create_dir("amd-iommu", NULL);
448 if (stats_dir == NULL)
451 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
452 (u32 *)&amd_iommu_unmap_flush);
454 amd_iommu_stats_add(&compl_wait);
455 amd_iommu_stats_add(&cnt_map_single);
456 amd_iommu_stats_add(&cnt_unmap_single);
457 amd_iommu_stats_add(&cnt_map_sg);
458 amd_iommu_stats_add(&cnt_unmap_sg);
459 amd_iommu_stats_add(&cnt_alloc_coherent);
460 amd_iommu_stats_add(&cnt_free_coherent);
461 amd_iommu_stats_add(&cross_page);
462 amd_iommu_stats_add(&domain_flush_single);
463 amd_iommu_stats_add(&domain_flush_all);
464 amd_iommu_stats_add(&alloced_io_mem);
465 amd_iommu_stats_add(&total_map_requests);
466 amd_iommu_stats_add(&complete_ppr);
467 amd_iommu_stats_add(&invalidate_iotlb);
468 amd_iommu_stats_add(&invalidate_iotlb_all);
469 amd_iommu_stats_add(&pri_requests);
474 /****************************************************************************
476 * Interrupt handling functions
478 ****************************************************************************/
480 static void dump_dte_entry(u16 devid)
484 for (i = 0; i < 4; ++i)
485 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
486 amd_iommu_dev_table[devid].data[i]);
489 static void dump_command(unsigned long phys_addr)
491 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
494 for (i = 0; i < 4; ++i)
495 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
498 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
500 int type, devid, domid, flags;
501 volatile u32 *event = __evt;
506 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
507 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
508 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
509 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
510 address = (u64)(((u64)event[3]) << 32) | event[2];
513 /* Did we hit the erratum? */
514 if (++count == LOOP_TIMEOUT) {
515 pr_err("AMD-Vi: No event written to event log\n");
522 printk(KERN_ERR "AMD-Vi: Event logged [");
525 case EVENT_TYPE_ILL_DEV:
526 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
527 "address=0x%016llx flags=0x%04x]\n",
528 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
530 dump_dte_entry(devid);
532 case EVENT_TYPE_IO_FAULT:
533 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
534 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
535 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
536 domid, address, flags);
538 case EVENT_TYPE_DEV_TAB_ERR:
539 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
540 "address=0x%016llx flags=0x%04x]\n",
541 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
544 case EVENT_TYPE_PAGE_TAB_ERR:
545 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
546 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
547 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
548 domid, address, flags);
550 case EVENT_TYPE_ILL_CMD:
551 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
552 dump_command(address);
554 case EVENT_TYPE_CMD_HARD_ERR:
555 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
556 "flags=0x%04x]\n", address, flags);
558 case EVENT_TYPE_IOTLB_INV_TO:
559 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
560 "address=0x%016llx]\n",
561 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
564 case EVENT_TYPE_INV_DEV_REQ:
565 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
566 "address=0x%016llx flags=0x%04x]\n",
567 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
571 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
574 memset(__evt, 0, 4 * sizeof(u32));
577 static void iommu_poll_events(struct amd_iommu *iommu)
582 spin_lock_irqsave(&iommu->lock, flags);
584 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
585 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
587 while (head != tail) {
588 iommu_print_event(iommu, iommu->evt_buf + head);
589 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
592 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
594 spin_unlock_irqrestore(&iommu->lock, flags);
597 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
599 struct amd_iommu_fault fault;
601 INC_STATS_COUNTER(pri_requests);
603 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
604 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
608 fault.address = raw[1];
609 fault.pasid = PPR_PASID(raw[0]);
610 fault.device_id = PPR_DEVID(raw[0]);
611 fault.tag = PPR_TAG(raw[0]);
612 fault.flags = PPR_FLAGS(raw[0]);
614 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
617 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
622 if (iommu->ppr_log == NULL)
625 /* enable ppr interrupts again */
626 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
628 spin_lock_irqsave(&iommu->lock, flags);
630 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
631 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
633 while (head != tail) {
638 raw = (u64 *)(iommu->ppr_log + head);
641 * Hardware bug: Interrupt may arrive before the entry is
642 * written to memory. If this happens we need to wait for the
645 for (i = 0; i < LOOP_TIMEOUT; ++i) {
646 if (PPR_REQ_TYPE(raw[0]) != 0)
651 /* Avoid memcpy function-call overhead */
656 * To detect the hardware bug we need to clear the entry
659 raw[0] = raw[1] = 0UL;
661 /* Update head pointer of hardware ring-buffer */
662 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
663 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
666 * Release iommu->lock because ppr-handling might need to
669 spin_unlock_irqrestore(&iommu->lock, flags);
671 /* Handle PPR entry */
672 iommu_handle_ppr_entry(iommu, entry);
674 spin_lock_irqsave(&iommu->lock, flags);
676 /* Refresh ring-buffer information */
677 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
678 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
681 spin_unlock_irqrestore(&iommu->lock, flags);
684 irqreturn_t amd_iommu_int_thread(int irq, void *data)
686 struct amd_iommu *iommu;
688 for_each_iommu(iommu) {
689 iommu_poll_events(iommu);
690 iommu_poll_ppr_log(iommu);
696 irqreturn_t amd_iommu_int_handler(int irq, void *data)
698 return IRQ_WAKE_THREAD;
701 /****************************************************************************
703 * IOMMU command queuing functions
705 ****************************************************************************/
707 static int wait_on_sem(volatile u64 *sem)
711 while (*sem == 0 && i < LOOP_TIMEOUT) {
716 if (i == LOOP_TIMEOUT) {
717 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
724 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
725 struct iommu_cmd *cmd,
730 target = iommu->cmd_buf + tail;
731 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
733 /* Copy command to buffer */
734 memcpy(target, cmd, sizeof(*cmd));
736 /* Tell the IOMMU about it */
737 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
740 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
742 WARN_ON(address & 0x7ULL);
744 memset(cmd, 0, sizeof(*cmd));
745 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
746 cmd->data[1] = upper_32_bits(__pa(address));
748 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
751 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
753 memset(cmd, 0, sizeof(*cmd));
754 cmd->data[0] = devid;
755 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
758 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
759 size_t size, u16 domid, int pde)
764 pages = iommu_num_pages(address, size, PAGE_SIZE);
769 * If we have to flush more than one page, flush all
770 * TLB entries for this domain
772 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
776 address &= PAGE_MASK;
778 memset(cmd, 0, sizeof(*cmd));
779 cmd->data[1] |= domid;
780 cmd->data[2] = lower_32_bits(address);
781 cmd->data[3] = upper_32_bits(address);
782 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
783 if (s) /* size bit - we flush more than one 4kb page */
784 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
785 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
786 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
789 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
790 u64 address, size_t size)
795 pages = iommu_num_pages(address, size, PAGE_SIZE);
800 * If we have to flush more than one page, flush all
801 * TLB entries for this domain
803 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
807 address &= PAGE_MASK;
809 memset(cmd, 0, sizeof(*cmd));
810 cmd->data[0] = devid;
811 cmd->data[0] |= (qdep & 0xff) << 24;
812 cmd->data[1] = devid;
813 cmd->data[2] = lower_32_bits(address);
814 cmd->data[3] = upper_32_bits(address);
815 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
817 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
820 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
821 u64 address, bool size)
823 memset(cmd, 0, sizeof(*cmd));
825 address &= ~(0xfffULL);
827 cmd->data[0] = pasid & PASID_MASK;
828 cmd->data[1] = domid;
829 cmd->data[2] = lower_32_bits(address);
830 cmd->data[3] = upper_32_bits(address);
831 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
832 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
834 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
835 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
838 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
839 int qdep, u64 address, bool size)
841 memset(cmd, 0, sizeof(*cmd));
843 address &= ~(0xfffULL);
845 cmd->data[0] = devid;
846 cmd->data[0] |= (pasid & 0xff) << 16;
847 cmd->data[0] |= (qdep & 0xff) << 24;
848 cmd->data[1] = devid;
849 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
850 cmd->data[2] = lower_32_bits(address);
851 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
852 cmd->data[3] = upper_32_bits(address);
854 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
855 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
858 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
859 int status, int tag, bool gn)
861 memset(cmd, 0, sizeof(*cmd));
863 cmd->data[0] = devid;
865 cmd->data[1] = pasid & PASID_MASK;
866 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
868 cmd->data[3] = tag & 0x1ff;
869 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
871 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
874 static void build_inv_all(struct iommu_cmd *cmd)
876 memset(cmd, 0, sizeof(*cmd));
877 CMD_SET_TYPE(cmd, CMD_INV_ALL);
881 * Writes the command to the IOMMUs command buffer and informs the
882 * hardware about the new command.
884 static int iommu_queue_command_sync(struct amd_iommu *iommu,
885 struct iommu_cmd *cmd,
888 u32 left, tail, head, next_tail;
891 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
894 spin_lock_irqsave(&iommu->lock, flags);
896 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
897 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
898 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
899 left = (head - next_tail) % iommu->cmd_buf_size;
902 struct iommu_cmd sync_cmd;
903 volatile u64 sem = 0;
906 build_completion_wait(&sync_cmd, (u64)&sem);
907 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
909 spin_unlock_irqrestore(&iommu->lock, flags);
911 if ((ret = wait_on_sem(&sem)) != 0)
917 copy_cmd_to_buffer(iommu, cmd, tail);
919 /* We need to sync now to make sure all commands are processed */
920 iommu->need_sync = sync;
922 spin_unlock_irqrestore(&iommu->lock, flags);
927 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
929 return iommu_queue_command_sync(iommu, cmd, true);
933 * This function queues a completion wait command into the command
936 static int iommu_completion_wait(struct amd_iommu *iommu)
938 struct iommu_cmd cmd;
939 volatile u64 sem = 0;
942 if (!iommu->need_sync)
945 build_completion_wait(&cmd, (u64)&sem);
947 ret = iommu_queue_command_sync(iommu, &cmd, false);
951 return wait_on_sem(&sem);
954 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
956 struct iommu_cmd cmd;
958 build_inv_dte(&cmd, devid);
960 return iommu_queue_command(iommu, &cmd);
963 static void iommu_flush_dte_all(struct amd_iommu *iommu)
967 for (devid = 0; devid <= 0xffff; ++devid)
968 iommu_flush_dte(iommu, devid);
970 iommu_completion_wait(iommu);
974 * This function uses heavy locking and may disable irqs for some time. But
975 * this is no issue because it is only called during resume.
977 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
981 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
982 struct iommu_cmd cmd;
983 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
985 iommu_queue_command(iommu, &cmd);
988 iommu_completion_wait(iommu);
991 static void iommu_flush_all(struct amd_iommu *iommu)
993 struct iommu_cmd cmd;
997 iommu_queue_command(iommu, &cmd);
998 iommu_completion_wait(iommu);
1001 void iommu_flush_all_caches(struct amd_iommu *iommu)
1003 if (iommu_feature(iommu, FEATURE_IA)) {
1004 iommu_flush_all(iommu);
1006 iommu_flush_dte_all(iommu);
1007 iommu_flush_tlb_all(iommu);
1012 * Command send function for flushing on-device TLB
1014 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1015 u64 address, size_t size)
1017 struct amd_iommu *iommu;
1018 struct iommu_cmd cmd;
1021 qdep = dev_data->ats.qdep;
1022 iommu = amd_iommu_rlookup_table[dev_data->devid];
1024 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1026 return iommu_queue_command(iommu, &cmd);
1030 * Command send function for invalidating a device table entry
1032 static int device_flush_dte(struct iommu_dev_data *dev_data)
1034 struct amd_iommu *iommu;
1037 iommu = amd_iommu_rlookup_table[dev_data->devid];
1039 ret = iommu_flush_dte(iommu, dev_data->devid);
1043 if (dev_data->ats.enabled)
1044 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1050 * TLB invalidation function which is called from the mapping functions.
1051 * It invalidates a single PTE if the range to flush is within a single
1052 * page. Otherwise it flushes the whole TLB of the IOMMU.
1054 static void __domain_flush_pages(struct protection_domain *domain,
1055 u64 address, size_t size, int pde)
1057 struct iommu_dev_data *dev_data;
1058 struct iommu_cmd cmd;
1061 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1063 for (i = 0; i < amd_iommus_present; ++i) {
1064 if (!domain->dev_iommu[i])
1068 * Devices of this domain are behind this IOMMU
1069 * We need a TLB flush
1071 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1074 list_for_each_entry(dev_data, &domain->dev_list, list) {
1076 if (!dev_data->ats.enabled)
1079 ret |= device_flush_iotlb(dev_data, address, size);
1085 static void domain_flush_pages(struct protection_domain *domain,
1086 u64 address, size_t size)
1088 __domain_flush_pages(domain, address, size, 0);
1091 /* Flush the whole IO/TLB for a given protection domain */
1092 static void domain_flush_tlb(struct protection_domain *domain)
1094 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1097 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1098 static void domain_flush_tlb_pde(struct protection_domain *domain)
1100 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1103 static void domain_flush_complete(struct protection_domain *domain)
1107 for (i = 0; i < amd_iommus_present; ++i) {
1108 if (!domain->dev_iommu[i])
1112 * Devices of this domain are behind this IOMMU
1113 * We need to wait for completion of all commands.
1115 iommu_completion_wait(amd_iommus[i]);
1121 * This function flushes the DTEs for all devices in domain
1123 static void domain_flush_devices(struct protection_domain *domain)
1125 struct iommu_dev_data *dev_data;
1127 list_for_each_entry(dev_data, &domain->dev_list, list)
1128 device_flush_dte(dev_data);
1131 /****************************************************************************
1133 * The functions below are used the create the page table mappings for
1134 * unity mapped regions.
1136 ****************************************************************************/
1139 * This function is used to add another level to an IO page table. Adding
1140 * another level increases the size of the address space by 9 bits to a size up
1143 static bool increase_address_space(struct protection_domain *domain,
1148 if (domain->mode == PAGE_MODE_6_LEVEL)
1149 /* address space already 64 bit large */
1152 pte = (void *)get_zeroed_page(gfp);
1156 *pte = PM_LEVEL_PDE(domain->mode,
1157 virt_to_phys(domain->pt_root));
1158 domain->pt_root = pte;
1160 domain->updated = true;
1165 static u64 *alloc_pte(struct protection_domain *domain,
1166 unsigned long address,
1167 unsigned long page_size,
1174 BUG_ON(!is_power_of_2(page_size));
1176 while (address > PM_LEVEL_SIZE(domain->mode))
1177 increase_address_space(domain, gfp);
1179 level = domain->mode - 1;
1180 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1181 address = PAGE_SIZE_ALIGN(address, page_size);
1182 end_lvl = PAGE_SIZE_LEVEL(page_size);
1184 while (level > end_lvl) {
1185 if (!IOMMU_PTE_PRESENT(*pte)) {
1186 page = (u64 *)get_zeroed_page(gfp);
1189 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1192 /* No level skipping support yet */
1193 if (PM_PTE_LEVEL(*pte) != level)
1198 pte = IOMMU_PTE_PAGE(*pte);
1200 if (pte_page && level == end_lvl)
1203 pte = &pte[PM_LEVEL_INDEX(level, address)];
1210 * This function checks if there is a PTE for a given dma address. If
1211 * there is one, it returns the pointer to it.
1213 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1218 if (address > PM_LEVEL_SIZE(domain->mode))
1221 level = domain->mode - 1;
1222 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1227 if (!IOMMU_PTE_PRESENT(*pte))
1231 if (PM_PTE_LEVEL(*pte) == 0x07) {
1232 unsigned long pte_mask, __pte;
1235 * If we have a series of large PTEs, make
1236 * sure to return a pointer to the first one.
1238 pte_mask = PTE_PAGE_SIZE(*pte);
1239 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1240 __pte = ((unsigned long)pte) & pte_mask;
1242 return (u64 *)__pte;
1245 /* No level skipping support yet */
1246 if (PM_PTE_LEVEL(*pte) != level)
1251 /* Walk to the next level */
1252 pte = IOMMU_PTE_PAGE(*pte);
1253 pte = &pte[PM_LEVEL_INDEX(level, address)];
1260 * Generic mapping functions. It maps a physical address into a DMA
1261 * address space. It allocates the page table pages if necessary.
1262 * In the future it can be extended to a generic mapping function
1263 * supporting all features of AMD IOMMU page tables like level skipping
1264 * and full 64 bit address spaces.
1266 static int iommu_map_page(struct protection_domain *dom,
1267 unsigned long bus_addr,
1268 unsigned long phys_addr,
1270 unsigned long page_size)
1275 if (!(prot & IOMMU_PROT_MASK))
1278 bus_addr = PAGE_ALIGN(bus_addr);
1279 phys_addr = PAGE_ALIGN(phys_addr);
1280 count = PAGE_SIZE_PTE_COUNT(page_size);
1281 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1283 for (i = 0; i < count; ++i)
1284 if (IOMMU_PTE_PRESENT(pte[i]))
1287 if (page_size > PAGE_SIZE) {
1288 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1289 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1291 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1293 if (prot & IOMMU_PROT_IR)
1294 __pte |= IOMMU_PTE_IR;
1295 if (prot & IOMMU_PROT_IW)
1296 __pte |= IOMMU_PTE_IW;
1298 for (i = 0; i < count; ++i)
1306 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1307 unsigned long bus_addr,
1308 unsigned long page_size)
1310 unsigned long long unmap_size, unmapped;
1313 BUG_ON(!is_power_of_2(page_size));
1317 while (unmapped < page_size) {
1319 pte = fetch_pte(dom, bus_addr);
1323 * No PTE for this address
1324 * move forward in 4kb steps
1326 unmap_size = PAGE_SIZE;
1327 } else if (PM_PTE_LEVEL(*pte) == 0) {
1328 /* 4kb PTE found for this address */
1329 unmap_size = PAGE_SIZE;
1334 /* Large PTE found which maps this address */
1335 unmap_size = PTE_PAGE_SIZE(*pte);
1336 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1337 for (i = 0; i < count; i++)
1341 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1342 unmapped += unmap_size;
1345 BUG_ON(!is_power_of_2(unmapped));
1351 * This function checks if a specific unity mapping entry is needed for
1352 * this specific IOMMU.
1354 static int iommu_for_unity_map(struct amd_iommu *iommu,
1355 struct unity_map_entry *entry)
1359 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1360 bdf = amd_iommu_alias_table[i];
1361 if (amd_iommu_rlookup_table[bdf] == iommu)
1369 * This function actually applies the mapping to the page table of the
1372 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1373 struct unity_map_entry *e)
1378 for (addr = e->address_start; addr < e->address_end;
1379 addr += PAGE_SIZE) {
1380 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1385 * if unity mapping is in aperture range mark the page
1386 * as allocated in the aperture
1388 if (addr < dma_dom->aperture_size)
1389 __set_bit(addr >> PAGE_SHIFT,
1390 dma_dom->aperture[0]->bitmap);
1397 * Init the unity mappings for a specific IOMMU in the system
1399 * Basically iterates over all unity mapping entries and applies them to
1400 * the default domain DMA of that IOMMU if necessary.
1402 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1404 struct unity_map_entry *entry;
1407 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1408 if (!iommu_for_unity_map(iommu, entry))
1410 ret = dma_ops_unity_map(iommu->default_dom, entry);
1419 * Inits the unity mappings required for a specific device
1421 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1424 struct unity_map_entry *e;
1427 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1428 if (!(devid >= e->devid_start && devid <= e->devid_end))
1430 ret = dma_ops_unity_map(dma_dom, e);
1438 /****************************************************************************
1440 * The next functions belong to the address allocator for the dma_ops
1441 * interface functions. They work like the allocators in the other IOMMU
1442 * drivers. Its basically a bitmap which marks the allocated pages in
1443 * the aperture. Maybe it could be enhanced in the future to a more
1444 * efficient allocator.
1446 ****************************************************************************/
1449 * The address allocator core functions.
1451 * called with domain->lock held
1455 * Used to reserve address ranges in the aperture (e.g. for exclusion
1458 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1459 unsigned long start_page,
1462 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1464 if (start_page + pages > last_page)
1465 pages = last_page - start_page;
1467 for (i = start_page; i < start_page + pages; ++i) {
1468 int index = i / APERTURE_RANGE_PAGES;
1469 int page = i % APERTURE_RANGE_PAGES;
1470 __set_bit(page, dom->aperture[index]->bitmap);
1475 * This function is used to add a new aperture range to an existing
1476 * aperture in case of dma_ops domain allocation or address allocation
1479 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1480 bool populate, gfp_t gfp)
1482 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1483 struct amd_iommu *iommu;
1484 unsigned long i, old_size;
1486 #ifdef CONFIG_IOMMU_STRESS
1490 if (index >= APERTURE_MAX_RANGES)
1493 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1494 if (!dma_dom->aperture[index])
1497 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1498 if (!dma_dom->aperture[index]->bitmap)
1501 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1504 unsigned long address = dma_dom->aperture_size;
1505 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1506 u64 *pte, *pte_page;
1508 for (i = 0; i < num_ptes; ++i) {
1509 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1514 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1516 address += APERTURE_RANGE_SIZE / 64;
1520 old_size = dma_dom->aperture_size;
1521 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1523 /* Reserve address range used for MSI messages */
1524 if (old_size < MSI_ADDR_BASE_LO &&
1525 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1526 unsigned long spage;
1529 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1530 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1532 dma_ops_reserve_addresses(dma_dom, spage, pages);
1535 /* Initialize the exclusion range if necessary */
1536 for_each_iommu(iommu) {
1537 if (iommu->exclusion_start &&
1538 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1539 && iommu->exclusion_start < dma_dom->aperture_size) {
1540 unsigned long startpage;
1541 int pages = iommu_num_pages(iommu->exclusion_start,
1542 iommu->exclusion_length,
1544 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1545 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1550 * Check for areas already mapped as present in the new aperture
1551 * range and mark those pages as reserved in the allocator. Such
1552 * mappings may already exist as a result of requested unity
1553 * mappings for devices.
1555 for (i = dma_dom->aperture[index]->offset;
1556 i < dma_dom->aperture_size;
1558 u64 *pte = fetch_pte(&dma_dom->domain, i);
1559 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1562 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1565 update_domain(&dma_dom->domain);
1570 update_domain(&dma_dom->domain);
1572 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1574 kfree(dma_dom->aperture[index]);
1575 dma_dom->aperture[index] = NULL;
1580 static unsigned long dma_ops_area_alloc(struct device *dev,
1581 struct dma_ops_domain *dom,
1583 unsigned long align_mask,
1585 unsigned long start)
1587 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1588 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1589 int i = start >> APERTURE_RANGE_SHIFT;
1590 unsigned long boundary_size;
1591 unsigned long address = -1;
1592 unsigned long limit;
1594 next_bit >>= PAGE_SHIFT;
1596 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1597 PAGE_SIZE) >> PAGE_SHIFT;
1599 for (;i < max_index; ++i) {
1600 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1602 if (dom->aperture[i]->offset >= dma_mask)
1605 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1606 dma_mask >> PAGE_SHIFT);
1608 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1609 limit, next_bit, pages, 0,
1610 boundary_size, align_mask);
1611 if (address != -1) {
1612 address = dom->aperture[i]->offset +
1613 (address << PAGE_SHIFT);
1614 dom->next_address = address + (pages << PAGE_SHIFT);
1624 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1625 struct dma_ops_domain *dom,
1627 unsigned long align_mask,
1630 unsigned long address;
1632 #ifdef CONFIG_IOMMU_STRESS
1633 dom->next_address = 0;
1634 dom->need_flush = true;
1637 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1638 dma_mask, dom->next_address);
1640 if (address == -1) {
1641 dom->next_address = 0;
1642 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1644 dom->need_flush = true;
1647 if (unlikely(address == -1))
1648 address = DMA_ERROR_CODE;
1650 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1656 * The address free function.
1658 * called with domain->lock held
1660 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1661 unsigned long address,
1664 unsigned i = address >> APERTURE_RANGE_SHIFT;
1665 struct aperture_range *range = dom->aperture[i];
1667 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1669 #ifdef CONFIG_IOMMU_STRESS
1674 if (address >= dom->next_address)
1675 dom->need_flush = true;
1677 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1679 bitmap_clear(range->bitmap, address, pages);
1683 /****************************************************************************
1685 * The next functions belong to the domain allocation. A domain is
1686 * allocated for every IOMMU as the default domain. If device isolation
1687 * is enabled, every device get its own domain. The most important thing
1688 * about domains is the page table mapping the DMA address space they
1691 ****************************************************************************/
1694 * This function adds a protection domain to the global protection domain list
1696 static void add_domain_to_list(struct protection_domain *domain)
1698 unsigned long flags;
1700 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1701 list_add(&domain->list, &amd_iommu_pd_list);
1702 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1706 * This function removes a protection domain to the global
1707 * protection domain list
1709 static void del_domain_from_list(struct protection_domain *domain)
1711 unsigned long flags;
1713 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1714 list_del(&domain->list);
1715 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1718 static u16 domain_id_alloc(void)
1720 unsigned long flags;
1723 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1724 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1726 if (id > 0 && id < MAX_DOMAIN_ID)
1727 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1730 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1735 static void domain_id_free(int id)
1737 unsigned long flags;
1739 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1740 if (id > 0 && id < MAX_DOMAIN_ID)
1741 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1742 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1745 static void free_pagetable(struct protection_domain *domain)
1750 p1 = domain->pt_root;
1755 for (i = 0; i < 512; ++i) {
1756 if (!IOMMU_PTE_PRESENT(p1[i]))
1759 p2 = IOMMU_PTE_PAGE(p1[i]);
1760 for (j = 0; j < 512; ++j) {
1761 if (!IOMMU_PTE_PRESENT(p2[j]))
1763 p3 = IOMMU_PTE_PAGE(p2[j]);
1764 free_page((unsigned long)p3);
1767 free_page((unsigned long)p2);
1770 free_page((unsigned long)p1);
1772 domain->pt_root = NULL;
1775 static void free_gcr3_tbl_level1(u64 *tbl)
1780 for (i = 0; i < 512; ++i) {
1781 if (!(tbl[i] & GCR3_VALID))
1784 ptr = __va(tbl[i] & PAGE_MASK);
1786 free_page((unsigned long)ptr);
1790 static void free_gcr3_tbl_level2(u64 *tbl)
1795 for (i = 0; i < 512; ++i) {
1796 if (!(tbl[i] & GCR3_VALID))
1799 ptr = __va(tbl[i] & PAGE_MASK);
1801 free_gcr3_tbl_level1(ptr);
1805 static void free_gcr3_table(struct protection_domain *domain)
1807 if (domain->glx == 2)
1808 free_gcr3_tbl_level2(domain->gcr3_tbl);
1809 else if (domain->glx == 1)
1810 free_gcr3_tbl_level1(domain->gcr3_tbl);
1811 else if (domain->glx != 0)
1814 free_page((unsigned long)domain->gcr3_tbl);
1818 * Free a domain, only used if something went wrong in the
1819 * allocation path and we need to free an already allocated page table
1821 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1828 del_domain_from_list(&dom->domain);
1830 free_pagetable(&dom->domain);
1832 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1833 if (!dom->aperture[i])
1835 free_page((unsigned long)dom->aperture[i]->bitmap);
1836 kfree(dom->aperture[i]);
1843 * Allocates a new protection domain usable for the dma_ops functions.
1844 * It also initializes the page table and the address allocator data
1845 * structures required for the dma_ops interface
1847 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1849 struct dma_ops_domain *dma_dom;
1851 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1855 spin_lock_init(&dma_dom->domain.lock);
1857 dma_dom->domain.id = domain_id_alloc();
1858 if (dma_dom->domain.id == 0)
1860 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1861 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1862 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1863 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1864 dma_dom->domain.priv = dma_dom;
1865 if (!dma_dom->domain.pt_root)
1868 dma_dom->need_flush = false;
1869 dma_dom->target_dev = 0xffff;
1871 add_domain_to_list(&dma_dom->domain);
1873 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1877 * mark the first page as allocated so we never return 0 as
1878 * a valid dma-address. So we can use 0 as error value
1880 dma_dom->aperture[0]->bitmap[0] = 1;
1881 dma_dom->next_address = 0;
1887 dma_ops_domain_free(dma_dom);
1893 * little helper function to check whether a given protection domain is a
1896 static bool dma_ops_domain(struct protection_domain *domain)
1898 return domain->flags & PD_DMA_OPS_MASK;
1901 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1906 if (domain->mode != PAGE_MODE_NONE)
1907 pte_root = virt_to_phys(domain->pt_root);
1909 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1910 << DEV_ENTRY_MODE_SHIFT;
1911 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1913 flags = amd_iommu_dev_table[devid].data[1];
1916 flags |= DTE_FLAG_IOTLB;
1918 if (domain->flags & PD_IOMMUV2_MASK) {
1919 u64 gcr3 = __pa(domain->gcr3_tbl);
1920 u64 glx = domain->glx;
1923 pte_root |= DTE_FLAG_GV;
1924 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1926 /* First mask out possible old values for GCR3 table */
1927 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1930 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1933 /* Encode GCR3 table into DTE */
1934 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1937 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1940 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1944 flags &= ~(0xffffUL);
1945 flags |= domain->id;
1947 amd_iommu_dev_table[devid].data[1] = flags;
1948 amd_iommu_dev_table[devid].data[0] = pte_root;
1951 static void clear_dte_entry(u16 devid)
1953 /* remove entry from the device table seen by the hardware */
1954 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1955 amd_iommu_dev_table[devid].data[1] = 0;
1957 amd_iommu_apply_erratum_63(devid);
1960 static void do_attach(struct iommu_dev_data *dev_data,
1961 struct protection_domain *domain)
1963 struct amd_iommu *iommu;
1966 iommu = amd_iommu_rlookup_table[dev_data->devid];
1967 ats = dev_data->ats.enabled;
1969 /* Update data structures */
1970 dev_data->domain = domain;
1971 list_add(&dev_data->list, &domain->dev_list);
1972 set_dte_entry(dev_data->devid, domain, ats);
1974 /* Do reference counting */
1975 domain->dev_iommu[iommu->index] += 1;
1976 domain->dev_cnt += 1;
1978 /* Flush the DTE entry */
1979 device_flush_dte(dev_data);
1982 static void do_detach(struct iommu_dev_data *dev_data)
1984 struct amd_iommu *iommu;
1986 iommu = amd_iommu_rlookup_table[dev_data->devid];
1988 /* decrease reference counters */
1989 dev_data->domain->dev_iommu[iommu->index] -= 1;
1990 dev_data->domain->dev_cnt -= 1;
1992 /* Update data structures */
1993 dev_data->domain = NULL;
1994 list_del(&dev_data->list);
1995 clear_dte_entry(dev_data->devid);
1997 /* Flush the DTE entry */
1998 device_flush_dte(dev_data);
2002 * If a device is not yet associated with a domain, this function does
2003 * assigns it visible for the hardware
2005 static int __attach_device(struct iommu_dev_data *dev_data,
2006 struct protection_domain *domain)
2011 spin_lock(&domain->lock);
2013 if (dev_data->alias_data != NULL) {
2014 struct iommu_dev_data *alias_data = dev_data->alias_data;
2016 /* Some sanity checks */
2018 if (alias_data->domain != NULL &&
2019 alias_data->domain != domain)
2022 if (dev_data->domain != NULL &&
2023 dev_data->domain != domain)
2026 /* Do real assignment */
2027 if (alias_data->domain == NULL)
2028 do_attach(alias_data, domain);
2030 atomic_inc(&alias_data->bind);
2033 if (dev_data->domain == NULL)
2034 do_attach(dev_data, domain);
2036 atomic_inc(&dev_data->bind);
2043 spin_unlock(&domain->lock);
2049 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2051 pci_disable_ats(pdev);
2052 pci_disable_pri(pdev);
2053 pci_disable_pasid(pdev);
2056 /* FIXME: Change generic reset-function to do the same */
2057 static int pri_reset_while_enabled(struct pci_dev *pdev)
2062 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2066 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2067 control |= PCI_PRI_CTRL_RESET;
2068 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2073 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2078 /* FIXME: Hardcode number of outstanding requests for now */
2080 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2082 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2084 /* Only allow access to user-accessible pages */
2085 ret = pci_enable_pasid(pdev, 0);
2089 /* First reset the PRI state of the device */
2090 ret = pci_reset_pri(pdev);
2095 ret = pci_enable_pri(pdev, reqs);
2100 ret = pri_reset_while_enabled(pdev);
2105 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2112 pci_disable_pri(pdev);
2113 pci_disable_pasid(pdev);
2118 /* FIXME: Move this to PCI code */
2119 #define PCI_PRI_TLP_OFF (1 << 15)
2121 bool pci_pri_tlp_required(struct pci_dev *pdev)
2126 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2130 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2132 return (status & PCI_PRI_TLP_OFF) ? true : false;
2136 * If a device is not yet associated with a domain, this function does
2137 * assigns it visible for the hardware
2139 static int attach_device(struct device *dev,
2140 struct protection_domain *domain)
2142 struct pci_dev *pdev = to_pci_dev(dev);
2143 struct iommu_dev_data *dev_data;
2144 unsigned long flags;
2147 dev_data = get_dev_data(dev);
2149 if (domain->flags & PD_IOMMUV2_MASK) {
2150 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2153 if (pdev_iommuv2_enable(pdev) != 0)
2156 dev_data->ats.enabled = true;
2157 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2158 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2159 } else if (amd_iommu_iotlb_sup &&
2160 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2161 dev_data->ats.enabled = true;
2162 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2165 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2166 ret = __attach_device(dev_data, domain);
2167 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2170 * We might boot into a crash-kernel here. The crashed kernel
2171 * left the caches in the IOMMU dirty. So we have to flush
2172 * here to evict all dirty stuff.
2174 domain_flush_tlb_pde(domain);
2180 * Removes a device from a protection domain (unlocked)
2182 static void __detach_device(struct iommu_dev_data *dev_data)
2184 struct protection_domain *domain;
2185 unsigned long flags;
2187 BUG_ON(!dev_data->domain);
2189 domain = dev_data->domain;
2191 spin_lock_irqsave(&domain->lock, flags);
2193 if (dev_data->alias_data != NULL) {
2194 struct iommu_dev_data *alias_data = dev_data->alias_data;
2196 if (atomic_dec_and_test(&alias_data->bind))
2197 do_detach(alias_data);
2200 if (atomic_dec_and_test(&dev_data->bind))
2201 do_detach(dev_data);
2203 spin_unlock_irqrestore(&domain->lock, flags);
2206 * If we run in passthrough mode the device must be assigned to the
2207 * passthrough domain if it is detached from any other domain.
2208 * Make sure we can deassign from the pt_domain itself.
2210 if (dev_data->passthrough &&
2211 (dev_data->domain == NULL && domain != pt_domain))
2212 __attach_device(dev_data, pt_domain);
2216 * Removes a device from a protection domain (with devtable_lock held)
2218 static void detach_device(struct device *dev)
2220 struct protection_domain *domain;
2221 struct iommu_dev_data *dev_data;
2222 unsigned long flags;
2224 dev_data = get_dev_data(dev);
2225 domain = dev_data->domain;
2227 /* lock device table */
2228 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2229 __detach_device(dev_data);
2230 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2232 if (domain->flags & PD_IOMMUV2_MASK)
2233 pdev_iommuv2_disable(to_pci_dev(dev));
2234 else if (dev_data->ats.enabled)
2235 pci_disable_ats(to_pci_dev(dev));
2237 dev_data->ats.enabled = false;
2241 * Find out the protection domain structure for a given PCI device. This
2242 * will give us the pointer to the page table root for example.
2244 static struct protection_domain *domain_for_device(struct device *dev)
2246 struct iommu_dev_data *dev_data;
2247 struct protection_domain *dom = NULL;
2248 unsigned long flags;
2250 dev_data = get_dev_data(dev);
2252 if (dev_data->domain)
2253 return dev_data->domain;
2255 if (dev_data->alias_data != NULL) {
2256 struct iommu_dev_data *alias_data = dev_data->alias_data;
2258 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2259 if (alias_data->domain != NULL) {
2260 __attach_device(dev_data, alias_data->domain);
2261 dom = alias_data->domain;
2263 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2269 static int device_change_notifier(struct notifier_block *nb,
2270 unsigned long action, void *data)
2272 struct dma_ops_domain *dma_domain;
2273 struct protection_domain *domain;
2274 struct iommu_dev_data *dev_data;
2275 struct device *dev = data;
2276 struct amd_iommu *iommu;
2277 unsigned long flags;
2280 if (!check_device(dev))
2283 devid = get_device_id(dev);
2284 iommu = amd_iommu_rlookup_table[devid];
2285 dev_data = get_dev_data(dev);
2288 case BUS_NOTIFY_UNBOUND_DRIVER:
2290 domain = domain_for_device(dev);
2294 if (dev_data->passthrough)
2298 case BUS_NOTIFY_ADD_DEVICE:
2300 iommu_init_device(dev);
2302 domain = domain_for_device(dev);
2304 /* allocate a protection domain if a device is added */
2305 dma_domain = find_protection_domain(devid);
2308 dma_domain = dma_ops_domain_alloc();
2311 dma_domain->target_dev = devid;
2313 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2314 list_add_tail(&dma_domain->list, &iommu_pd_list);
2315 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2318 case BUS_NOTIFY_DEL_DEVICE:
2320 iommu_uninit_device(dev);
2326 iommu_completion_wait(iommu);
2332 static struct notifier_block device_nb = {
2333 .notifier_call = device_change_notifier,
2336 void amd_iommu_init_notifier(void)
2338 bus_register_notifier(&pci_bus_type, &device_nb);
2341 /*****************************************************************************
2343 * The next functions belong to the dma_ops mapping/unmapping code.
2345 *****************************************************************************/
2348 * In the dma_ops path we only have the struct device. This function
2349 * finds the corresponding IOMMU, the protection domain and the
2350 * requestor id for a given device.
2351 * If the device is not yet associated with a domain this is also done
2354 static struct protection_domain *get_domain(struct device *dev)
2356 struct protection_domain *domain;
2357 struct dma_ops_domain *dma_dom;
2358 u16 devid = get_device_id(dev);
2360 if (!check_device(dev))
2361 return ERR_PTR(-EINVAL);
2363 domain = domain_for_device(dev);
2364 if (domain != NULL && !dma_ops_domain(domain))
2365 return ERR_PTR(-EBUSY);
2370 /* Device not bount yet - bind it */
2371 dma_dom = find_protection_domain(devid);
2373 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2374 attach_device(dev, &dma_dom->domain);
2375 DUMP_printk("Using protection domain %d for device %s\n",
2376 dma_dom->domain.id, dev_name(dev));
2378 return &dma_dom->domain;
2381 static void update_device_table(struct protection_domain *domain)
2383 struct iommu_dev_data *dev_data;
2385 list_for_each_entry(dev_data, &domain->dev_list, list)
2386 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2389 static void update_domain(struct protection_domain *domain)
2391 if (!domain->updated)
2394 update_device_table(domain);
2396 domain_flush_devices(domain);
2397 domain_flush_tlb_pde(domain);
2399 domain->updated = false;
2403 * This function fetches the PTE for a given address in the aperture
2405 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2406 unsigned long address)
2408 struct aperture_range *aperture;
2409 u64 *pte, *pte_page;
2411 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2415 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2417 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2419 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2421 pte += PM_LEVEL_INDEX(0, address);
2423 update_domain(&dom->domain);
2429 * This is the generic map function. It maps one 4kb page at paddr to
2430 * the given address in the DMA address space for the domain.
2432 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2433 unsigned long address,
2439 WARN_ON(address > dom->aperture_size);
2443 pte = dma_ops_get_pte(dom, address);
2445 return DMA_ERROR_CODE;
2447 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2449 if (direction == DMA_TO_DEVICE)
2450 __pte |= IOMMU_PTE_IR;
2451 else if (direction == DMA_FROM_DEVICE)
2452 __pte |= IOMMU_PTE_IW;
2453 else if (direction == DMA_BIDIRECTIONAL)
2454 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2460 return (dma_addr_t)address;
2464 * The generic unmapping function for on page in the DMA address space.
2466 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2467 unsigned long address)
2469 struct aperture_range *aperture;
2472 if (address >= dom->aperture_size)
2475 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2479 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2483 pte += PM_LEVEL_INDEX(0, address);
2491 * This function contains common code for mapping of a physically
2492 * contiguous memory region into DMA address space. It is used by all
2493 * mapping functions provided with this IOMMU driver.
2494 * Must be called with the domain lock held.
2496 static dma_addr_t __map_single(struct device *dev,
2497 struct dma_ops_domain *dma_dom,
2504 dma_addr_t offset = paddr & ~PAGE_MASK;
2505 dma_addr_t address, start, ret;
2507 unsigned long align_mask = 0;
2510 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2513 INC_STATS_COUNTER(total_map_requests);
2516 INC_STATS_COUNTER(cross_page);
2519 align_mask = (1UL << get_order(size)) - 1;
2522 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2524 if (unlikely(address == DMA_ERROR_CODE)) {
2526 * setting next_address here will let the address
2527 * allocator only scan the new allocated range in the
2528 * first run. This is a small optimization.
2530 dma_dom->next_address = dma_dom->aperture_size;
2532 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2536 * aperture was successfully enlarged by 128 MB, try
2543 for (i = 0; i < pages; ++i) {
2544 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2545 if (ret == DMA_ERROR_CODE)
2553 ADD_STATS_COUNTER(alloced_io_mem, size);
2555 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2556 domain_flush_tlb(&dma_dom->domain);
2557 dma_dom->need_flush = false;
2558 } else if (unlikely(amd_iommu_np_cache))
2559 domain_flush_pages(&dma_dom->domain, address, size);
2566 for (--i; i >= 0; --i) {
2568 dma_ops_domain_unmap(dma_dom, start);
2571 dma_ops_free_addresses(dma_dom, address, pages);
2573 return DMA_ERROR_CODE;
2577 * Does the reverse of the __map_single function. Must be called with
2578 * the domain lock held too
2580 static void __unmap_single(struct dma_ops_domain *dma_dom,
2581 dma_addr_t dma_addr,
2585 dma_addr_t flush_addr;
2586 dma_addr_t i, start;
2589 if ((dma_addr == DMA_ERROR_CODE) ||
2590 (dma_addr + size > dma_dom->aperture_size))
2593 flush_addr = dma_addr;
2594 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2595 dma_addr &= PAGE_MASK;
2598 for (i = 0; i < pages; ++i) {
2599 dma_ops_domain_unmap(dma_dom, start);
2603 SUB_STATS_COUNTER(alloced_io_mem, size);
2605 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2607 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2608 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2609 dma_dom->need_flush = false;
2614 * The exported map_single function for dma_ops.
2616 static dma_addr_t map_page(struct device *dev, struct page *page,
2617 unsigned long offset, size_t size,
2618 enum dma_data_direction dir,
2619 struct dma_attrs *attrs)
2621 unsigned long flags;
2622 struct protection_domain *domain;
2625 phys_addr_t paddr = page_to_phys(page) + offset;
2627 INC_STATS_COUNTER(cnt_map_single);
2629 domain = get_domain(dev);
2630 if (PTR_ERR(domain) == -EINVAL)
2631 return (dma_addr_t)paddr;
2632 else if (IS_ERR(domain))
2633 return DMA_ERROR_CODE;
2635 dma_mask = *dev->dma_mask;
2637 spin_lock_irqsave(&domain->lock, flags);
2639 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2641 if (addr == DMA_ERROR_CODE)
2644 domain_flush_complete(domain);
2647 spin_unlock_irqrestore(&domain->lock, flags);
2653 * The exported unmap_single function for dma_ops.
2655 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2656 enum dma_data_direction dir, struct dma_attrs *attrs)
2658 unsigned long flags;
2659 struct protection_domain *domain;
2661 INC_STATS_COUNTER(cnt_unmap_single);
2663 domain = get_domain(dev);
2667 spin_lock_irqsave(&domain->lock, flags);
2669 __unmap_single(domain->priv, dma_addr, size, dir);
2671 domain_flush_complete(domain);
2673 spin_unlock_irqrestore(&domain->lock, flags);
2677 * This is a special map_sg function which is used if we should map a
2678 * device which is not handled by an AMD IOMMU in the system.
2680 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2681 int nelems, int dir)
2683 struct scatterlist *s;
2686 for_each_sg(sglist, s, nelems, i) {
2687 s->dma_address = (dma_addr_t)sg_phys(s);
2688 s->dma_length = s->length;
2695 * The exported map_sg function for dma_ops (handles scatter-gather
2698 static int map_sg(struct device *dev, struct scatterlist *sglist,
2699 int nelems, enum dma_data_direction dir,
2700 struct dma_attrs *attrs)
2702 unsigned long flags;
2703 struct protection_domain *domain;
2705 struct scatterlist *s;
2707 int mapped_elems = 0;
2710 INC_STATS_COUNTER(cnt_map_sg);
2712 domain = get_domain(dev);
2713 if (PTR_ERR(domain) == -EINVAL)
2714 return map_sg_no_iommu(dev, sglist, nelems, dir);
2715 else if (IS_ERR(domain))
2718 dma_mask = *dev->dma_mask;
2720 spin_lock_irqsave(&domain->lock, flags);
2722 for_each_sg(sglist, s, nelems, i) {
2725 s->dma_address = __map_single(dev, domain->priv,
2726 paddr, s->length, dir, false,
2729 if (s->dma_address) {
2730 s->dma_length = s->length;
2736 domain_flush_complete(domain);
2739 spin_unlock_irqrestore(&domain->lock, flags);
2741 return mapped_elems;
2743 for_each_sg(sglist, s, mapped_elems, i) {
2745 __unmap_single(domain->priv, s->dma_address,
2746 s->dma_length, dir);
2747 s->dma_address = s->dma_length = 0;
2756 * The exported map_sg function for dma_ops (handles scatter-gather
2759 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2760 int nelems, enum dma_data_direction dir,
2761 struct dma_attrs *attrs)
2763 unsigned long flags;
2764 struct protection_domain *domain;
2765 struct scatterlist *s;
2768 INC_STATS_COUNTER(cnt_unmap_sg);
2770 domain = get_domain(dev);
2774 spin_lock_irqsave(&domain->lock, flags);
2776 for_each_sg(sglist, s, nelems, i) {
2777 __unmap_single(domain->priv, s->dma_address,
2778 s->dma_length, dir);
2779 s->dma_address = s->dma_length = 0;
2782 domain_flush_complete(domain);
2784 spin_unlock_irqrestore(&domain->lock, flags);
2788 * The exported alloc_coherent function for dma_ops.
2790 static void *alloc_coherent(struct device *dev, size_t size,
2791 dma_addr_t *dma_addr, gfp_t flag,
2792 struct dma_attrs *attrs)
2794 unsigned long flags;
2796 struct protection_domain *domain;
2798 u64 dma_mask = dev->coherent_dma_mask;
2800 INC_STATS_COUNTER(cnt_alloc_coherent);
2802 domain = get_domain(dev);
2803 if (PTR_ERR(domain) == -EINVAL) {
2804 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2805 *dma_addr = __pa(virt_addr);
2807 } else if (IS_ERR(domain))
2810 dma_mask = dev->coherent_dma_mask;
2811 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2814 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2818 paddr = virt_to_phys(virt_addr);
2821 dma_mask = *dev->dma_mask;
2823 spin_lock_irqsave(&domain->lock, flags);
2825 *dma_addr = __map_single(dev, domain->priv, paddr,
2826 size, DMA_BIDIRECTIONAL, true, dma_mask);
2828 if (*dma_addr == DMA_ERROR_CODE) {
2829 spin_unlock_irqrestore(&domain->lock, flags);
2833 domain_flush_complete(domain);
2835 spin_unlock_irqrestore(&domain->lock, flags);
2841 free_pages((unsigned long)virt_addr, get_order(size));
2847 * The exported free_coherent function for dma_ops.
2849 static void free_coherent(struct device *dev, size_t size,
2850 void *virt_addr, dma_addr_t dma_addr,
2851 struct dma_attrs *attrs)
2853 unsigned long flags;
2854 struct protection_domain *domain;
2856 INC_STATS_COUNTER(cnt_free_coherent);
2858 domain = get_domain(dev);
2862 spin_lock_irqsave(&domain->lock, flags);
2864 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2866 domain_flush_complete(domain);
2868 spin_unlock_irqrestore(&domain->lock, flags);
2871 free_pages((unsigned long)virt_addr, get_order(size));
2875 * This function is called by the DMA layer to find out if we can handle a
2876 * particular device. It is part of the dma_ops.
2878 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2880 return check_device(dev);
2884 * The function for pre-allocating protection domains.
2886 * If the driver core informs the DMA layer if a driver grabs a device
2887 * we don't need to preallocate the protection domains anymore.
2888 * For now we have to.
2890 static void __init prealloc_protection_domains(void)
2892 struct iommu_dev_data *dev_data;
2893 struct dma_ops_domain *dma_dom;
2894 struct pci_dev *dev = NULL;
2897 for_each_pci_dev(dev) {
2899 /* Do we handle this device? */
2900 if (!check_device(&dev->dev))
2903 dev_data = get_dev_data(&dev->dev);
2904 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2905 /* Make sure passthrough domain is allocated */
2906 alloc_passthrough_domain();
2907 dev_data->passthrough = true;
2908 attach_device(&dev->dev, pt_domain);
2909 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2910 dev_name(&dev->dev));
2913 /* Is there already any domain for it? */
2914 if (domain_for_device(&dev->dev))
2917 devid = get_device_id(&dev->dev);
2919 dma_dom = dma_ops_domain_alloc();
2922 init_unity_mappings_for_device(dma_dom, devid);
2923 dma_dom->target_dev = devid;
2925 attach_device(&dev->dev, &dma_dom->domain);
2927 list_add_tail(&dma_dom->list, &iommu_pd_list);
2931 static struct dma_map_ops amd_iommu_dma_ops = {
2932 .alloc = alloc_coherent,
2933 .free = free_coherent,
2934 .map_page = map_page,
2935 .unmap_page = unmap_page,
2937 .unmap_sg = unmap_sg,
2938 .dma_supported = amd_iommu_dma_supported,
2941 static unsigned device_dma_ops_init(void)
2943 struct iommu_dev_data *dev_data;
2944 struct pci_dev *pdev = NULL;
2945 unsigned unhandled = 0;
2947 for_each_pci_dev(pdev) {
2948 if (!check_device(&pdev->dev)) {
2950 iommu_ignore_device(&pdev->dev);
2956 dev_data = get_dev_data(&pdev->dev);
2958 if (!dev_data->passthrough)
2959 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2961 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
2968 * The function which clues the AMD IOMMU driver into dma_ops.
2971 void __init amd_iommu_init_api(void)
2973 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2976 int __init amd_iommu_init_dma_ops(void)
2978 struct amd_iommu *iommu;
2982 * first allocate a default protection domain for every IOMMU we
2983 * found in the system. Devices not assigned to any other
2984 * protection domain will be assigned to the default one.
2986 for_each_iommu(iommu) {
2987 iommu->default_dom = dma_ops_domain_alloc();
2988 if (iommu->default_dom == NULL)
2990 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2991 ret = iommu_init_unity_mappings(iommu);
2997 * Pre-allocate the protection domains for each device.
2999 prealloc_protection_domains();
3004 /* Make the driver finally visible to the drivers */
3005 unhandled = device_dma_ops_init();
3006 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3007 /* There are unhandled devices - initialize swiotlb for them */
3011 amd_iommu_stats_init();
3017 for_each_iommu(iommu) {
3018 if (iommu->default_dom)
3019 dma_ops_domain_free(iommu->default_dom);
3025 /*****************************************************************************
3027 * The following functions belong to the exported interface of AMD IOMMU
3029 * This interface allows access to lower level functions of the IOMMU
3030 * like protection domain handling and assignement of devices to domains
3031 * which is not possible with the dma_ops interface.
3033 *****************************************************************************/
3035 static void cleanup_domain(struct protection_domain *domain)
3037 struct iommu_dev_data *dev_data, *next;
3038 unsigned long flags;
3040 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3042 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3043 __detach_device(dev_data);
3044 atomic_set(&dev_data->bind, 0);
3047 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3050 static void protection_domain_free(struct protection_domain *domain)
3055 del_domain_from_list(domain);
3058 domain_id_free(domain->id);
3063 static struct protection_domain *protection_domain_alloc(void)
3065 struct protection_domain *domain;
3067 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3071 spin_lock_init(&domain->lock);
3072 mutex_init(&domain->api_lock);
3073 domain->id = domain_id_alloc();
3076 INIT_LIST_HEAD(&domain->dev_list);
3078 add_domain_to_list(domain);
3088 static int __init alloc_passthrough_domain(void)
3090 if (pt_domain != NULL)
3093 /* allocate passthrough domain */
3094 pt_domain = protection_domain_alloc();
3098 pt_domain->mode = PAGE_MODE_NONE;
3102 static int amd_iommu_domain_init(struct iommu_domain *dom)
3104 struct protection_domain *domain;
3106 domain = protection_domain_alloc();
3110 domain->mode = PAGE_MODE_3_LEVEL;
3111 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3112 if (!domain->pt_root)
3115 domain->iommu_domain = dom;
3122 protection_domain_free(domain);
3127 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3129 struct protection_domain *domain = dom->priv;
3134 if (domain->dev_cnt > 0)
3135 cleanup_domain(domain);
3137 BUG_ON(domain->dev_cnt != 0);
3139 if (domain->mode != PAGE_MODE_NONE)
3140 free_pagetable(domain);
3142 if (domain->flags & PD_IOMMUV2_MASK)
3143 free_gcr3_table(domain);
3145 protection_domain_free(domain);
3150 static void amd_iommu_detach_device(struct iommu_domain *dom,
3153 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3154 struct amd_iommu *iommu;
3157 if (!check_device(dev))
3160 devid = get_device_id(dev);
3162 if (dev_data->domain != NULL)
3165 iommu = amd_iommu_rlookup_table[devid];
3169 iommu_completion_wait(iommu);
3172 static int amd_iommu_attach_device(struct iommu_domain *dom,
3175 struct protection_domain *domain = dom->priv;
3176 struct iommu_dev_data *dev_data;
3177 struct amd_iommu *iommu;
3180 if (!check_device(dev))
3183 dev_data = dev->archdata.iommu;
3185 iommu = amd_iommu_rlookup_table[dev_data->devid];
3189 if (dev_data->domain)
3192 ret = attach_device(dev, domain);
3194 iommu_completion_wait(iommu);
3199 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3200 phys_addr_t paddr, size_t page_size, int iommu_prot)
3202 struct protection_domain *domain = dom->priv;
3206 if (domain->mode == PAGE_MODE_NONE)
3209 if (iommu_prot & IOMMU_READ)
3210 prot |= IOMMU_PROT_IR;
3211 if (iommu_prot & IOMMU_WRITE)
3212 prot |= IOMMU_PROT_IW;
3214 mutex_lock(&domain->api_lock);
3215 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3216 mutex_unlock(&domain->api_lock);
3221 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3224 struct protection_domain *domain = dom->priv;
3227 if (domain->mode == PAGE_MODE_NONE)
3230 mutex_lock(&domain->api_lock);
3231 unmap_size = iommu_unmap_page(domain, iova, page_size);
3232 mutex_unlock(&domain->api_lock);
3234 domain_flush_tlb_pde(domain);
3239 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3242 struct protection_domain *domain = dom->priv;
3243 unsigned long offset_mask;
3247 if (domain->mode == PAGE_MODE_NONE)
3250 pte = fetch_pte(domain, iova);
3252 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3255 if (PM_PTE_LEVEL(*pte) == 0)
3256 offset_mask = PAGE_SIZE - 1;
3258 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3260 __pte = *pte & PM_ADDR_MASK;
3261 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3266 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3270 case IOMMU_CAP_CACHE_COHERENCY:
3277 static struct iommu_ops amd_iommu_ops = {
3278 .domain_init = amd_iommu_domain_init,
3279 .domain_destroy = amd_iommu_domain_destroy,
3280 .attach_dev = amd_iommu_attach_device,
3281 .detach_dev = amd_iommu_detach_device,
3282 .map = amd_iommu_map,
3283 .unmap = amd_iommu_unmap,
3284 .iova_to_phys = amd_iommu_iova_to_phys,
3285 .domain_has_cap = amd_iommu_domain_has_cap,
3286 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3289 /*****************************************************************************
3291 * The next functions do a basic initialization of IOMMU for pass through
3294 * In passthrough mode the IOMMU is initialized and enabled but not used for
3295 * DMA-API translation.
3297 *****************************************************************************/
3299 int __init amd_iommu_init_passthrough(void)
3301 struct iommu_dev_data *dev_data;
3302 struct pci_dev *dev = NULL;
3303 struct amd_iommu *iommu;
3307 ret = alloc_passthrough_domain();
3311 for_each_pci_dev(dev) {
3312 if (!check_device(&dev->dev))
3315 dev_data = get_dev_data(&dev->dev);
3316 dev_data->passthrough = true;
3318 devid = get_device_id(&dev->dev);
3320 iommu = amd_iommu_rlookup_table[devid];
3324 attach_device(&dev->dev, pt_domain);
3327 amd_iommu_stats_init();
3329 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3334 /* IOMMUv2 specific functions */
3335 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3337 return atomic_notifier_chain_register(&ppr_notifier, nb);
3339 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3341 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3343 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3345 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3347 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3349 struct protection_domain *domain = dom->priv;
3350 unsigned long flags;
3352 spin_lock_irqsave(&domain->lock, flags);
3354 /* Update data structure */
3355 domain->mode = PAGE_MODE_NONE;
3356 domain->updated = true;
3358 /* Make changes visible to IOMMUs */
3359 update_domain(domain);
3361 /* Page-table is not visible to IOMMU anymore, so free it */
3362 free_pagetable(domain);
3364 spin_unlock_irqrestore(&domain->lock, flags);
3366 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3368 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3370 struct protection_domain *domain = dom->priv;
3371 unsigned long flags;
3374 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3377 /* Number of GCR3 table levels required */
3378 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3381 if (levels > amd_iommu_max_glx_val)
3384 spin_lock_irqsave(&domain->lock, flags);
3387 * Save us all sanity checks whether devices already in the
3388 * domain support IOMMUv2. Just force that the domain has no
3389 * devices attached when it is switched into IOMMUv2 mode.
3392 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3396 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3397 if (domain->gcr3_tbl == NULL)
3400 domain->glx = levels;
3401 domain->flags |= PD_IOMMUV2_MASK;
3402 domain->updated = true;
3404 update_domain(domain);
3409 spin_unlock_irqrestore(&domain->lock, flags);
3413 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3415 static int __flush_pasid(struct protection_domain *domain, int pasid,
3416 u64 address, bool size)
3418 struct iommu_dev_data *dev_data;
3419 struct iommu_cmd cmd;
3422 if (!(domain->flags & PD_IOMMUV2_MASK))
3425 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3428 * IOMMU TLB needs to be flushed before Device TLB to
3429 * prevent device TLB refill from IOMMU TLB
3431 for (i = 0; i < amd_iommus_present; ++i) {
3432 if (domain->dev_iommu[i] == 0)
3435 ret = iommu_queue_command(amd_iommus[i], &cmd);
3440 /* Wait until IOMMU TLB flushes are complete */
3441 domain_flush_complete(domain);
3443 /* Now flush device TLBs */
3444 list_for_each_entry(dev_data, &domain->dev_list, list) {
3445 struct amd_iommu *iommu;
3448 BUG_ON(!dev_data->ats.enabled);
3450 qdep = dev_data->ats.qdep;
3451 iommu = amd_iommu_rlookup_table[dev_data->devid];
3453 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3454 qdep, address, size);
3456 ret = iommu_queue_command(iommu, &cmd);
3461 /* Wait until all device TLBs are flushed */
3462 domain_flush_complete(domain);
3471 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3474 INC_STATS_COUNTER(invalidate_iotlb);
3476 return __flush_pasid(domain, pasid, address, false);
3479 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3482 struct protection_domain *domain = dom->priv;
3483 unsigned long flags;
3486 spin_lock_irqsave(&domain->lock, flags);
3487 ret = __amd_iommu_flush_page(domain, pasid, address);
3488 spin_unlock_irqrestore(&domain->lock, flags);
3492 EXPORT_SYMBOL(amd_iommu_flush_page);
3494 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3496 INC_STATS_COUNTER(invalidate_iotlb_all);
3498 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3502 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3504 struct protection_domain *domain = dom->priv;
3505 unsigned long flags;
3508 spin_lock_irqsave(&domain->lock, flags);
3509 ret = __amd_iommu_flush_tlb(domain, pasid);
3510 spin_unlock_irqrestore(&domain->lock, flags);
3514 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3516 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3523 index = (pasid >> (9 * level)) & 0x1ff;
3529 if (!(*pte & GCR3_VALID)) {
3533 root = (void *)get_zeroed_page(GFP_ATOMIC);
3537 *pte = __pa(root) | GCR3_VALID;
3540 root = __va(*pte & PAGE_MASK);
3548 static int __set_gcr3(struct protection_domain *domain, int pasid,
3553 if (domain->mode != PAGE_MODE_NONE)
3556 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3560 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3562 return __amd_iommu_flush_tlb(domain, pasid);
3565 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3569 if (domain->mode != PAGE_MODE_NONE)
3572 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3578 return __amd_iommu_flush_tlb(domain, pasid);
3581 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3584 struct protection_domain *domain = dom->priv;
3585 unsigned long flags;
3588 spin_lock_irqsave(&domain->lock, flags);
3589 ret = __set_gcr3(domain, pasid, cr3);
3590 spin_unlock_irqrestore(&domain->lock, flags);
3594 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3596 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3598 struct protection_domain *domain = dom->priv;
3599 unsigned long flags;
3602 spin_lock_irqsave(&domain->lock, flags);
3603 ret = __clear_gcr3(domain, pasid);
3604 spin_unlock_irqrestore(&domain->lock, flags);
3608 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3610 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3611 int status, int tag)
3613 struct iommu_dev_data *dev_data;
3614 struct amd_iommu *iommu;
3615 struct iommu_cmd cmd;
3617 INC_STATS_COUNTER(complete_ppr);
3619 dev_data = get_dev_data(&pdev->dev);
3620 iommu = amd_iommu_rlookup_table[dev_data->devid];
3622 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3623 tag, dev_data->pri_tlp);
3625 return iommu_queue_command(iommu, &cmd);
3627 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3629 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3631 struct protection_domain *domain;
3633 domain = get_domain(&pdev->dev);
3637 /* Only return IOMMUv2 domains */
3638 if (!(domain->flags & PD_IOMMUV2_MASK))
3641 return domain->iommu_domain;
3643 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3645 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3647 struct iommu_dev_data *dev_data;
3649 if (!amd_iommu_v2_supported())
3652 dev_data = get_dev_data(&pdev->dev);
3653 dev_data->errata |= (1 << erratum);
3655 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3657 int amd_iommu_device_info(struct pci_dev *pdev,
3658 struct amd_iommu_device_info *info)
3663 if (pdev == NULL || info == NULL)
3666 if (!amd_iommu_v2_supported())
3669 memset(info, 0, sizeof(*info));
3671 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3673 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3675 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3677 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3679 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3683 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3684 max_pasids = min(max_pasids, (1 << 20));
3686 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3687 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3689 features = pci_pasid_features(pdev);
3690 if (features & PCI_PASID_CAP_EXEC)
3691 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3692 if (features & PCI_PASID_CAP_PRIV)
3693 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3698 EXPORT_SYMBOL(amd_iommu_device_info);