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;
86 static struct dma_map_ops amd_iommu_dma_ops;
89 * general struct to manage commands send to an IOMMU
95 static void update_domain(struct protection_domain *domain);
96 static int __init alloc_passthrough_domain(void);
98 /****************************************************************************
102 ****************************************************************************/
104 static struct iommu_dev_data *alloc_dev_data(u16 devid)
106 struct iommu_dev_data *dev_data;
109 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
113 dev_data->devid = devid;
114 atomic_set(&dev_data->bind, 0);
116 spin_lock_irqsave(&dev_data_list_lock, flags);
117 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
118 spin_unlock_irqrestore(&dev_data_list_lock, flags);
123 static void free_dev_data(struct iommu_dev_data *dev_data)
127 spin_lock_irqsave(&dev_data_list_lock, flags);
128 list_del(&dev_data->dev_data_list);
129 spin_unlock_irqrestore(&dev_data_list_lock, flags);
134 static struct iommu_dev_data *search_dev_data(u16 devid)
136 struct iommu_dev_data *dev_data;
139 spin_lock_irqsave(&dev_data_list_lock, flags);
140 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
141 if (dev_data->devid == devid)
148 spin_unlock_irqrestore(&dev_data_list_lock, flags);
153 static struct iommu_dev_data *find_dev_data(u16 devid)
155 struct iommu_dev_data *dev_data;
157 dev_data = search_dev_data(devid);
159 if (dev_data == NULL)
160 dev_data = alloc_dev_data(devid);
165 static inline u16 get_device_id(struct device *dev)
167 struct pci_dev *pdev = to_pci_dev(dev);
169 return calc_devid(pdev->bus->number, pdev->devfn);
172 static struct iommu_dev_data *get_dev_data(struct device *dev)
174 return dev->archdata.iommu;
177 static bool pci_iommuv2_capable(struct pci_dev *pdev)
179 static const int caps[] = {
182 PCI_EXT_CAP_ID_PASID,
186 for (i = 0; i < 3; ++i) {
187 pos = pci_find_ext_capability(pdev, caps[i]);
195 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
197 struct iommu_dev_data *dev_data;
199 dev_data = get_dev_data(&pdev->dev);
201 return dev_data->errata & (1 << erratum) ? true : false;
205 * In this function the list of preallocated protection domains is traversed to
206 * find the domain for a specific device
208 static struct dma_ops_domain *find_protection_domain(u16 devid)
210 struct dma_ops_domain *entry, *ret = NULL;
212 u16 alias = amd_iommu_alias_table[devid];
214 if (list_empty(&iommu_pd_list))
217 spin_lock_irqsave(&iommu_pd_list_lock, flags);
219 list_for_each_entry(entry, &iommu_pd_list, list) {
220 if (entry->target_dev == devid ||
221 entry->target_dev == alias) {
227 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
233 * This function checks if the driver got a valid device from the caller to
234 * avoid dereferencing invalid pointers.
236 static bool check_device(struct device *dev)
240 if (!dev || !dev->dma_mask)
243 /* No device or no PCI device */
244 if (dev->bus != &pci_bus_type)
247 devid = get_device_id(dev);
249 /* Out of our scope? */
250 if (devid > amd_iommu_last_bdf)
253 if (amd_iommu_rlookup_table[devid] == NULL)
259 static int iommu_init_device(struct device *dev)
261 struct pci_dev *pdev = to_pci_dev(dev);
262 struct iommu_dev_data *dev_data;
265 if (dev->archdata.iommu)
268 dev_data = find_dev_data(get_device_id(dev));
272 alias = amd_iommu_alias_table[dev_data->devid];
273 if (alias != dev_data->devid) {
274 struct iommu_dev_data *alias_data;
276 alias_data = find_dev_data(alias);
277 if (alias_data == NULL) {
278 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
280 free_dev_data(dev_data);
283 dev_data->alias_data = alias_data;
286 if (pci_iommuv2_capable(pdev)) {
287 struct amd_iommu *iommu;
289 iommu = amd_iommu_rlookup_table[dev_data->devid];
290 dev_data->iommu_v2 = iommu->is_iommu_v2;
293 dev->archdata.iommu = dev_data;
298 static void iommu_ignore_device(struct device *dev)
302 devid = get_device_id(dev);
303 alias = amd_iommu_alias_table[devid];
305 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
306 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
308 amd_iommu_rlookup_table[devid] = NULL;
309 amd_iommu_rlookup_table[alias] = NULL;
312 static void iommu_uninit_device(struct device *dev)
315 * Nothing to do here - we keep dev_data around for unplugged devices
316 * and reuse it when the device is re-plugged - not doing so would
317 * introduce a ton of races.
321 void __init amd_iommu_uninit_devices(void)
323 struct iommu_dev_data *dev_data, *n;
324 struct pci_dev *pdev = NULL;
326 for_each_pci_dev(pdev) {
328 if (!check_device(&pdev->dev))
331 iommu_uninit_device(&pdev->dev);
334 /* Free all of our dev_data structures */
335 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
336 free_dev_data(dev_data);
339 int __init amd_iommu_init_devices(void)
341 struct pci_dev *pdev = NULL;
344 for_each_pci_dev(pdev) {
346 if (!check_device(&pdev->dev))
349 ret = iommu_init_device(&pdev->dev);
350 if (ret == -ENOTSUPP)
351 iommu_ignore_device(&pdev->dev);
360 amd_iommu_uninit_devices();
364 #ifdef CONFIG_AMD_IOMMU_STATS
367 * Initialization code for statistics collection
370 DECLARE_STATS_COUNTER(compl_wait);
371 DECLARE_STATS_COUNTER(cnt_map_single);
372 DECLARE_STATS_COUNTER(cnt_unmap_single);
373 DECLARE_STATS_COUNTER(cnt_map_sg);
374 DECLARE_STATS_COUNTER(cnt_unmap_sg);
375 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
376 DECLARE_STATS_COUNTER(cnt_free_coherent);
377 DECLARE_STATS_COUNTER(cross_page);
378 DECLARE_STATS_COUNTER(domain_flush_single);
379 DECLARE_STATS_COUNTER(domain_flush_all);
380 DECLARE_STATS_COUNTER(alloced_io_mem);
381 DECLARE_STATS_COUNTER(total_map_requests);
382 DECLARE_STATS_COUNTER(complete_ppr);
383 DECLARE_STATS_COUNTER(invalidate_iotlb);
384 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
385 DECLARE_STATS_COUNTER(pri_requests);
387 static struct dentry *stats_dir;
388 static struct dentry *de_fflush;
390 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
392 if (stats_dir == NULL)
395 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
399 static void amd_iommu_stats_init(void)
401 stats_dir = debugfs_create_dir("amd-iommu", NULL);
402 if (stats_dir == NULL)
405 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
406 &amd_iommu_unmap_flush);
408 amd_iommu_stats_add(&compl_wait);
409 amd_iommu_stats_add(&cnt_map_single);
410 amd_iommu_stats_add(&cnt_unmap_single);
411 amd_iommu_stats_add(&cnt_map_sg);
412 amd_iommu_stats_add(&cnt_unmap_sg);
413 amd_iommu_stats_add(&cnt_alloc_coherent);
414 amd_iommu_stats_add(&cnt_free_coherent);
415 amd_iommu_stats_add(&cross_page);
416 amd_iommu_stats_add(&domain_flush_single);
417 amd_iommu_stats_add(&domain_flush_all);
418 amd_iommu_stats_add(&alloced_io_mem);
419 amd_iommu_stats_add(&total_map_requests);
420 amd_iommu_stats_add(&complete_ppr);
421 amd_iommu_stats_add(&invalidate_iotlb);
422 amd_iommu_stats_add(&invalidate_iotlb_all);
423 amd_iommu_stats_add(&pri_requests);
428 /****************************************************************************
430 * Interrupt handling functions
432 ****************************************************************************/
434 static void dump_dte_entry(u16 devid)
438 for (i = 0; i < 4; ++i)
439 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
440 amd_iommu_dev_table[devid].data[i]);
443 static void dump_command(unsigned long phys_addr)
445 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
448 for (i = 0; i < 4; ++i)
449 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
452 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
454 int type, devid, domid, flags;
455 volatile u32 *event = __evt;
460 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
461 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
462 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
463 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
464 address = (u64)(((u64)event[3]) << 32) | event[2];
467 /* Did we hit the erratum? */
468 if (++count == LOOP_TIMEOUT) {
469 pr_err("AMD-Vi: No event written to event log\n");
476 printk(KERN_ERR "AMD-Vi: Event logged [");
479 case EVENT_TYPE_ILL_DEV:
480 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
481 "address=0x%016llx flags=0x%04x]\n",
482 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
484 dump_dte_entry(devid);
486 case EVENT_TYPE_IO_FAULT:
487 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
488 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
489 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
490 domid, address, flags);
492 case EVENT_TYPE_DEV_TAB_ERR:
493 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
494 "address=0x%016llx flags=0x%04x]\n",
495 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
498 case EVENT_TYPE_PAGE_TAB_ERR:
499 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
500 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
501 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
502 domid, address, flags);
504 case EVENT_TYPE_ILL_CMD:
505 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
506 dump_command(address);
508 case EVENT_TYPE_CMD_HARD_ERR:
509 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
510 "flags=0x%04x]\n", address, flags);
512 case EVENT_TYPE_IOTLB_INV_TO:
513 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
514 "address=0x%016llx]\n",
515 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
518 case EVENT_TYPE_INV_DEV_REQ:
519 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
520 "address=0x%016llx flags=0x%04x]\n",
521 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
525 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
528 memset(__evt, 0, 4 * sizeof(u32));
531 static void iommu_poll_events(struct amd_iommu *iommu)
536 spin_lock_irqsave(&iommu->lock, flags);
538 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
539 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
541 while (head != tail) {
542 iommu_print_event(iommu, iommu->evt_buf + head);
543 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
546 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
548 spin_unlock_irqrestore(&iommu->lock, flags);
551 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
553 struct amd_iommu_fault fault;
555 INC_STATS_COUNTER(pri_requests);
557 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
558 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
562 fault.address = raw[1];
563 fault.pasid = PPR_PASID(raw[0]);
564 fault.device_id = PPR_DEVID(raw[0]);
565 fault.tag = PPR_TAG(raw[0]);
566 fault.flags = PPR_FLAGS(raw[0]);
568 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
571 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
576 if (iommu->ppr_log == NULL)
579 /* enable ppr interrupts again */
580 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
582 spin_lock_irqsave(&iommu->lock, flags);
584 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
585 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
587 while (head != tail) {
592 raw = (u64 *)(iommu->ppr_log + head);
595 * Hardware bug: Interrupt may arrive before the entry is
596 * written to memory. If this happens we need to wait for the
599 for (i = 0; i < LOOP_TIMEOUT; ++i) {
600 if (PPR_REQ_TYPE(raw[0]) != 0)
605 /* Avoid memcpy function-call overhead */
610 * To detect the hardware bug we need to clear the entry
613 raw[0] = raw[1] = 0UL;
615 /* Update head pointer of hardware ring-buffer */
616 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
617 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
620 * Release iommu->lock because ppr-handling might need to
623 spin_unlock_irqrestore(&iommu->lock, flags);
625 /* Handle PPR entry */
626 iommu_handle_ppr_entry(iommu, entry);
628 spin_lock_irqsave(&iommu->lock, flags);
630 /* Refresh ring-buffer information */
631 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
632 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
635 spin_unlock_irqrestore(&iommu->lock, flags);
638 irqreturn_t amd_iommu_int_thread(int irq, void *data)
640 struct amd_iommu *iommu;
642 for_each_iommu(iommu) {
643 iommu_poll_events(iommu);
644 iommu_poll_ppr_log(iommu);
650 irqreturn_t amd_iommu_int_handler(int irq, void *data)
652 return IRQ_WAKE_THREAD;
655 /****************************************************************************
657 * IOMMU command queuing functions
659 ****************************************************************************/
661 static int wait_on_sem(volatile u64 *sem)
665 while (*sem == 0 && i < LOOP_TIMEOUT) {
670 if (i == LOOP_TIMEOUT) {
671 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
678 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
679 struct iommu_cmd *cmd,
684 target = iommu->cmd_buf + tail;
685 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
687 /* Copy command to buffer */
688 memcpy(target, cmd, sizeof(*cmd));
690 /* Tell the IOMMU about it */
691 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
694 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
696 WARN_ON(address & 0x7ULL);
698 memset(cmd, 0, sizeof(*cmd));
699 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
700 cmd->data[1] = upper_32_bits(__pa(address));
702 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
705 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
707 memset(cmd, 0, sizeof(*cmd));
708 cmd->data[0] = devid;
709 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
712 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
713 size_t size, u16 domid, int pde)
718 pages = iommu_num_pages(address, size, PAGE_SIZE);
723 * If we have to flush more than one page, flush all
724 * TLB entries for this domain
726 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
730 address &= PAGE_MASK;
732 memset(cmd, 0, sizeof(*cmd));
733 cmd->data[1] |= domid;
734 cmd->data[2] = lower_32_bits(address);
735 cmd->data[3] = upper_32_bits(address);
736 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
737 if (s) /* size bit - we flush more than one 4kb page */
738 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
739 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
740 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
743 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
744 u64 address, size_t size)
749 pages = iommu_num_pages(address, size, PAGE_SIZE);
754 * If we have to flush more than one page, flush all
755 * TLB entries for this domain
757 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
761 address &= PAGE_MASK;
763 memset(cmd, 0, sizeof(*cmd));
764 cmd->data[0] = devid;
765 cmd->data[0] |= (qdep & 0xff) << 24;
766 cmd->data[1] = devid;
767 cmd->data[2] = lower_32_bits(address);
768 cmd->data[3] = upper_32_bits(address);
769 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
771 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
774 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
775 u64 address, bool size)
777 memset(cmd, 0, sizeof(*cmd));
779 address &= ~(0xfffULL);
781 cmd->data[0] = pasid & PASID_MASK;
782 cmd->data[1] = domid;
783 cmd->data[2] = lower_32_bits(address);
784 cmd->data[3] = upper_32_bits(address);
785 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
786 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
788 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
789 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
792 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
793 int qdep, u64 address, bool size)
795 memset(cmd, 0, sizeof(*cmd));
797 address &= ~(0xfffULL);
799 cmd->data[0] = devid;
800 cmd->data[0] |= (pasid & 0xff) << 16;
801 cmd->data[0] |= (qdep & 0xff) << 24;
802 cmd->data[1] = devid;
803 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
804 cmd->data[2] = lower_32_bits(address);
805 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
806 cmd->data[3] = upper_32_bits(address);
808 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
809 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
812 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
813 int status, int tag, bool gn)
815 memset(cmd, 0, sizeof(*cmd));
817 cmd->data[0] = devid;
819 cmd->data[1] = pasid & PASID_MASK;
820 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
822 cmd->data[3] = tag & 0x1ff;
823 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
825 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
828 static void build_inv_all(struct iommu_cmd *cmd)
830 memset(cmd, 0, sizeof(*cmd));
831 CMD_SET_TYPE(cmd, CMD_INV_ALL);
835 * Writes the command to the IOMMUs command buffer and informs the
836 * hardware about the new command.
838 static int iommu_queue_command_sync(struct amd_iommu *iommu,
839 struct iommu_cmd *cmd,
842 u32 left, tail, head, next_tail;
845 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
848 spin_lock_irqsave(&iommu->lock, flags);
850 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
851 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
852 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
853 left = (head - next_tail) % iommu->cmd_buf_size;
856 struct iommu_cmd sync_cmd;
857 volatile u64 sem = 0;
860 build_completion_wait(&sync_cmd, (u64)&sem);
861 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
863 spin_unlock_irqrestore(&iommu->lock, flags);
865 if ((ret = wait_on_sem(&sem)) != 0)
871 copy_cmd_to_buffer(iommu, cmd, tail);
873 /* We need to sync now to make sure all commands are processed */
874 iommu->need_sync = sync;
876 spin_unlock_irqrestore(&iommu->lock, flags);
881 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
883 return iommu_queue_command_sync(iommu, cmd, true);
887 * This function queues a completion wait command into the command
890 static int iommu_completion_wait(struct amd_iommu *iommu)
892 struct iommu_cmd cmd;
893 volatile u64 sem = 0;
896 if (!iommu->need_sync)
899 build_completion_wait(&cmd, (u64)&sem);
901 ret = iommu_queue_command_sync(iommu, &cmd, false);
905 return wait_on_sem(&sem);
908 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
910 struct iommu_cmd cmd;
912 build_inv_dte(&cmd, devid);
914 return iommu_queue_command(iommu, &cmd);
917 static void iommu_flush_dte_all(struct amd_iommu *iommu)
921 for (devid = 0; devid <= 0xffff; ++devid)
922 iommu_flush_dte(iommu, devid);
924 iommu_completion_wait(iommu);
928 * This function uses heavy locking and may disable irqs for some time. But
929 * this is no issue because it is only called during resume.
931 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
935 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
936 struct iommu_cmd cmd;
937 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
939 iommu_queue_command(iommu, &cmd);
942 iommu_completion_wait(iommu);
945 static void iommu_flush_all(struct amd_iommu *iommu)
947 struct iommu_cmd cmd;
951 iommu_queue_command(iommu, &cmd);
952 iommu_completion_wait(iommu);
955 void iommu_flush_all_caches(struct amd_iommu *iommu)
957 if (iommu_feature(iommu, FEATURE_IA)) {
958 iommu_flush_all(iommu);
960 iommu_flush_dte_all(iommu);
961 iommu_flush_tlb_all(iommu);
966 * Command send function for flushing on-device TLB
968 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
969 u64 address, size_t size)
971 struct amd_iommu *iommu;
972 struct iommu_cmd cmd;
975 qdep = dev_data->ats.qdep;
976 iommu = amd_iommu_rlookup_table[dev_data->devid];
978 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
980 return iommu_queue_command(iommu, &cmd);
984 * Command send function for invalidating a device table entry
986 static int device_flush_dte(struct iommu_dev_data *dev_data)
988 struct amd_iommu *iommu;
991 iommu = amd_iommu_rlookup_table[dev_data->devid];
993 ret = iommu_flush_dte(iommu, dev_data->devid);
997 if (dev_data->ats.enabled)
998 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1004 * TLB invalidation function which is called from the mapping functions.
1005 * It invalidates a single PTE if the range to flush is within a single
1006 * page. Otherwise it flushes the whole TLB of the IOMMU.
1008 static void __domain_flush_pages(struct protection_domain *domain,
1009 u64 address, size_t size, int pde)
1011 struct iommu_dev_data *dev_data;
1012 struct iommu_cmd cmd;
1015 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1017 for (i = 0; i < amd_iommus_present; ++i) {
1018 if (!domain->dev_iommu[i])
1022 * Devices of this domain are behind this IOMMU
1023 * We need a TLB flush
1025 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1028 list_for_each_entry(dev_data, &domain->dev_list, list) {
1030 if (!dev_data->ats.enabled)
1033 ret |= device_flush_iotlb(dev_data, address, size);
1039 static void domain_flush_pages(struct protection_domain *domain,
1040 u64 address, size_t size)
1042 __domain_flush_pages(domain, address, size, 0);
1045 /* Flush the whole IO/TLB for a given protection domain */
1046 static void domain_flush_tlb(struct protection_domain *domain)
1048 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1051 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1052 static void domain_flush_tlb_pde(struct protection_domain *domain)
1054 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1057 static void domain_flush_complete(struct protection_domain *domain)
1061 for (i = 0; i < amd_iommus_present; ++i) {
1062 if (!domain->dev_iommu[i])
1066 * Devices of this domain are behind this IOMMU
1067 * We need to wait for completion of all commands.
1069 iommu_completion_wait(amd_iommus[i]);
1075 * This function flushes the DTEs for all devices in domain
1077 static void domain_flush_devices(struct protection_domain *domain)
1079 struct iommu_dev_data *dev_data;
1081 list_for_each_entry(dev_data, &domain->dev_list, list)
1082 device_flush_dte(dev_data);
1085 /****************************************************************************
1087 * The functions below are used the create the page table mappings for
1088 * unity mapped regions.
1090 ****************************************************************************/
1093 * This function is used to add another level to an IO page table. Adding
1094 * another level increases the size of the address space by 9 bits to a size up
1097 static bool increase_address_space(struct protection_domain *domain,
1102 if (domain->mode == PAGE_MODE_6_LEVEL)
1103 /* address space already 64 bit large */
1106 pte = (void *)get_zeroed_page(gfp);
1110 *pte = PM_LEVEL_PDE(domain->mode,
1111 virt_to_phys(domain->pt_root));
1112 domain->pt_root = pte;
1114 domain->updated = true;
1119 static u64 *alloc_pte(struct protection_domain *domain,
1120 unsigned long address,
1121 unsigned long page_size,
1128 BUG_ON(!is_power_of_2(page_size));
1130 while (address > PM_LEVEL_SIZE(domain->mode))
1131 increase_address_space(domain, gfp);
1133 level = domain->mode - 1;
1134 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1135 address = PAGE_SIZE_ALIGN(address, page_size);
1136 end_lvl = PAGE_SIZE_LEVEL(page_size);
1138 while (level > end_lvl) {
1139 if (!IOMMU_PTE_PRESENT(*pte)) {
1140 page = (u64 *)get_zeroed_page(gfp);
1143 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1146 /* No level skipping support yet */
1147 if (PM_PTE_LEVEL(*pte) != level)
1152 pte = IOMMU_PTE_PAGE(*pte);
1154 if (pte_page && level == end_lvl)
1157 pte = &pte[PM_LEVEL_INDEX(level, address)];
1164 * This function checks if there is a PTE for a given dma address. If
1165 * there is one, it returns the pointer to it.
1167 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1172 if (address > PM_LEVEL_SIZE(domain->mode))
1175 level = domain->mode - 1;
1176 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1181 if (!IOMMU_PTE_PRESENT(*pte))
1185 if (PM_PTE_LEVEL(*pte) == 0x07) {
1186 unsigned long pte_mask, __pte;
1189 * If we have a series of large PTEs, make
1190 * sure to return a pointer to the first one.
1192 pte_mask = PTE_PAGE_SIZE(*pte);
1193 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1194 __pte = ((unsigned long)pte) & pte_mask;
1196 return (u64 *)__pte;
1199 /* No level skipping support yet */
1200 if (PM_PTE_LEVEL(*pte) != level)
1205 /* Walk to the next level */
1206 pte = IOMMU_PTE_PAGE(*pte);
1207 pte = &pte[PM_LEVEL_INDEX(level, address)];
1214 * Generic mapping functions. It maps a physical address into a DMA
1215 * address space. It allocates the page table pages if necessary.
1216 * In the future it can be extended to a generic mapping function
1217 * supporting all features of AMD IOMMU page tables like level skipping
1218 * and full 64 bit address spaces.
1220 static int iommu_map_page(struct protection_domain *dom,
1221 unsigned long bus_addr,
1222 unsigned long phys_addr,
1224 unsigned long page_size)
1229 if (!(prot & IOMMU_PROT_MASK))
1232 bus_addr = PAGE_ALIGN(bus_addr);
1233 phys_addr = PAGE_ALIGN(phys_addr);
1234 count = PAGE_SIZE_PTE_COUNT(page_size);
1235 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1237 for (i = 0; i < count; ++i)
1238 if (IOMMU_PTE_PRESENT(pte[i]))
1241 if (page_size > PAGE_SIZE) {
1242 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1243 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1245 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1247 if (prot & IOMMU_PROT_IR)
1248 __pte |= IOMMU_PTE_IR;
1249 if (prot & IOMMU_PROT_IW)
1250 __pte |= IOMMU_PTE_IW;
1252 for (i = 0; i < count; ++i)
1260 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1261 unsigned long bus_addr,
1262 unsigned long page_size)
1264 unsigned long long unmap_size, unmapped;
1267 BUG_ON(!is_power_of_2(page_size));
1271 while (unmapped < page_size) {
1273 pte = fetch_pte(dom, bus_addr);
1277 * No PTE for this address
1278 * move forward in 4kb steps
1280 unmap_size = PAGE_SIZE;
1281 } else if (PM_PTE_LEVEL(*pte) == 0) {
1282 /* 4kb PTE found for this address */
1283 unmap_size = PAGE_SIZE;
1288 /* Large PTE found which maps this address */
1289 unmap_size = PTE_PAGE_SIZE(*pte);
1290 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1291 for (i = 0; i < count; i++)
1295 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1296 unmapped += unmap_size;
1299 BUG_ON(!is_power_of_2(unmapped));
1305 * This function checks if a specific unity mapping entry is needed for
1306 * this specific IOMMU.
1308 static int iommu_for_unity_map(struct amd_iommu *iommu,
1309 struct unity_map_entry *entry)
1313 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1314 bdf = amd_iommu_alias_table[i];
1315 if (amd_iommu_rlookup_table[bdf] == iommu)
1323 * This function actually applies the mapping to the page table of the
1326 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1327 struct unity_map_entry *e)
1332 for (addr = e->address_start; addr < e->address_end;
1333 addr += PAGE_SIZE) {
1334 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1339 * if unity mapping is in aperture range mark the page
1340 * as allocated in the aperture
1342 if (addr < dma_dom->aperture_size)
1343 __set_bit(addr >> PAGE_SHIFT,
1344 dma_dom->aperture[0]->bitmap);
1351 * Init the unity mappings for a specific IOMMU in the system
1353 * Basically iterates over all unity mapping entries and applies them to
1354 * the default domain DMA of that IOMMU if necessary.
1356 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1358 struct unity_map_entry *entry;
1361 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1362 if (!iommu_for_unity_map(iommu, entry))
1364 ret = dma_ops_unity_map(iommu->default_dom, entry);
1373 * Inits the unity mappings required for a specific device
1375 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1378 struct unity_map_entry *e;
1381 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1382 if (!(devid >= e->devid_start && devid <= e->devid_end))
1384 ret = dma_ops_unity_map(dma_dom, e);
1392 /****************************************************************************
1394 * The next functions belong to the address allocator for the dma_ops
1395 * interface functions. They work like the allocators in the other IOMMU
1396 * drivers. Its basically a bitmap which marks the allocated pages in
1397 * the aperture. Maybe it could be enhanced in the future to a more
1398 * efficient allocator.
1400 ****************************************************************************/
1403 * The address allocator core functions.
1405 * called with domain->lock held
1409 * Used to reserve address ranges in the aperture (e.g. for exclusion
1412 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1413 unsigned long start_page,
1416 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1418 if (start_page + pages > last_page)
1419 pages = last_page - start_page;
1421 for (i = start_page; i < start_page + pages; ++i) {
1422 int index = i / APERTURE_RANGE_PAGES;
1423 int page = i % APERTURE_RANGE_PAGES;
1424 __set_bit(page, dom->aperture[index]->bitmap);
1429 * This function is used to add a new aperture range to an existing
1430 * aperture in case of dma_ops domain allocation or address allocation
1433 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1434 bool populate, gfp_t gfp)
1436 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1437 struct amd_iommu *iommu;
1438 unsigned long i, old_size;
1440 #ifdef CONFIG_IOMMU_STRESS
1444 if (index >= APERTURE_MAX_RANGES)
1447 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1448 if (!dma_dom->aperture[index])
1451 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1452 if (!dma_dom->aperture[index]->bitmap)
1455 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1458 unsigned long address = dma_dom->aperture_size;
1459 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1460 u64 *pte, *pte_page;
1462 for (i = 0; i < num_ptes; ++i) {
1463 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1468 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1470 address += APERTURE_RANGE_SIZE / 64;
1474 old_size = dma_dom->aperture_size;
1475 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1477 /* Reserve address range used for MSI messages */
1478 if (old_size < MSI_ADDR_BASE_LO &&
1479 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1480 unsigned long spage;
1483 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1484 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1486 dma_ops_reserve_addresses(dma_dom, spage, pages);
1489 /* Initialize the exclusion range if necessary */
1490 for_each_iommu(iommu) {
1491 if (iommu->exclusion_start &&
1492 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1493 && iommu->exclusion_start < dma_dom->aperture_size) {
1494 unsigned long startpage;
1495 int pages = iommu_num_pages(iommu->exclusion_start,
1496 iommu->exclusion_length,
1498 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1499 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1504 * Check for areas already mapped as present in the new aperture
1505 * range and mark those pages as reserved in the allocator. Such
1506 * mappings may already exist as a result of requested unity
1507 * mappings for devices.
1509 for (i = dma_dom->aperture[index]->offset;
1510 i < dma_dom->aperture_size;
1512 u64 *pte = fetch_pte(&dma_dom->domain, i);
1513 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1516 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1519 update_domain(&dma_dom->domain);
1524 update_domain(&dma_dom->domain);
1526 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1528 kfree(dma_dom->aperture[index]);
1529 dma_dom->aperture[index] = NULL;
1534 static unsigned long dma_ops_area_alloc(struct device *dev,
1535 struct dma_ops_domain *dom,
1537 unsigned long align_mask,
1539 unsigned long start)
1541 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1542 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1543 int i = start >> APERTURE_RANGE_SHIFT;
1544 unsigned long boundary_size;
1545 unsigned long address = -1;
1546 unsigned long limit;
1548 next_bit >>= PAGE_SHIFT;
1550 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1551 PAGE_SIZE) >> PAGE_SHIFT;
1553 for (;i < max_index; ++i) {
1554 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1556 if (dom->aperture[i]->offset >= dma_mask)
1559 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1560 dma_mask >> PAGE_SHIFT);
1562 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1563 limit, next_bit, pages, 0,
1564 boundary_size, align_mask);
1565 if (address != -1) {
1566 address = dom->aperture[i]->offset +
1567 (address << PAGE_SHIFT);
1568 dom->next_address = address + (pages << PAGE_SHIFT);
1578 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1579 struct dma_ops_domain *dom,
1581 unsigned long align_mask,
1584 unsigned long address;
1586 #ifdef CONFIG_IOMMU_STRESS
1587 dom->next_address = 0;
1588 dom->need_flush = true;
1591 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1592 dma_mask, dom->next_address);
1594 if (address == -1) {
1595 dom->next_address = 0;
1596 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1598 dom->need_flush = true;
1601 if (unlikely(address == -1))
1602 address = DMA_ERROR_CODE;
1604 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1610 * The address free function.
1612 * called with domain->lock held
1614 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1615 unsigned long address,
1618 unsigned i = address >> APERTURE_RANGE_SHIFT;
1619 struct aperture_range *range = dom->aperture[i];
1621 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1623 #ifdef CONFIG_IOMMU_STRESS
1628 if (address >= dom->next_address)
1629 dom->need_flush = true;
1631 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1633 bitmap_clear(range->bitmap, address, pages);
1637 /****************************************************************************
1639 * The next functions belong to the domain allocation. A domain is
1640 * allocated for every IOMMU as the default domain. If device isolation
1641 * is enabled, every device get its own domain. The most important thing
1642 * about domains is the page table mapping the DMA address space they
1645 ****************************************************************************/
1648 * This function adds a protection domain to the global protection domain list
1650 static void add_domain_to_list(struct protection_domain *domain)
1652 unsigned long flags;
1654 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1655 list_add(&domain->list, &amd_iommu_pd_list);
1656 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1660 * This function removes a protection domain to the global
1661 * protection domain list
1663 static void del_domain_from_list(struct protection_domain *domain)
1665 unsigned long flags;
1667 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1668 list_del(&domain->list);
1669 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1672 static u16 domain_id_alloc(void)
1674 unsigned long flags;
1677 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1678 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1680 if (id > 0 && id < MAX_DOMAIN_ID)
1681 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1684 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1689 static void domain_id_free(int id)
1691 unsigned long flags;
1693 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1694 if (id > 0 && id < MAX_DOMAIN_ID)
1695 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1696 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1699 static void free_pagetable(struct protection_domain *domain)
1704 p1 = domain->pt_root;
1709 for (i = 0; i < 512; ++i) {
1710 if (!IOMMU_PTE_PRESENT(p1[i]))
1713 p2 = IOMMU_PTE_PAGE(p1[i]);
1714 for (j = 0; j < 512; ++j) {
1715 if (!IOMMU_PTE_PRESENT(p2[j]))
1717 p3 = IOMMU_PTE_PAGE(p2[j]);
1718 free_page((unsigned long)p3);
1721 free_page((unsigned long)p2);
1724 free_page((unsigned long)p1);
1726 domain->pt_root = NULL;
1729 static void free_gcr3_tbl_level1(u64 *tbl)
1734 for (i = 0; i < 512; ++i) {
1735 if (!(tbl[i] & GCR3_VALID))
1738 ptr = __va(tbl[i] & PAGE_MASK);
1740 free_page((unsigned long)ptr);
1744 static void free_gcr3_tbl_level2(u64 *tbl)
1749 for (i = 0; i < 512; ++i) {
1750 if (!(tbl[i] & GCR3_VALID))
1753 ptr = __va(tbl[i] & PAGE_MASK);
1755 free_gcr3_tbl_level1(ptr);
1759 static void free_gcr3_table(struct protection_domain *domain)
1761 if (domain->glx == 2)
1762 free_gcr3_tbl_level2(domain->gcr3_tbl);
1763 else if (domain->glx == 1)
1764 free_gcr3_tbl_level1(domain->gcr3_tbl);
1765 else if (domain->glx != 0)
1768 free_page((unsigned long)domain->gcr3_tbl);
1772 * Free a domain, only used if something went wrong in the
1773 * allocation path and we need to free an already allocated page table
1775 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1782 del_domain_from_list(&dom->domain);
1784 free_pagetable(&dom->domain);
1786 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1787 if (!dom->aperture[i])
1789 free_page((unsigned long)dom->aperture[i]->bitmap);
1790 kfree(dom->aperture[i]);
1797 * Allocates a new protection domain usable for the dma_ops functions.
1798 * It also initializes the page table and the address allocator data
1799 * structures required for the dma_ops interface
1801 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1803 struct dma_ops_domain *dma_dom;
1805 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1809 spin_lock_init(&dma_dom->domain.lock);
1811 dma_dom->domain.id = domain_id_alloc();
1812 if (dma_dom->domain.id == 0)
1814 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1815 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1816 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1817 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1818 dma_dom->domain.priv = dma_dom;
1819 if (!dma_dom->domain.pt_root)
1822 dma_dom->need_flush = false;
1823 dma_dom->target_dev = 0xffff;
1825 add_domain_to_list(&dma_dom->domain);
1827 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1831 * mark the first page as allocated so we never return 0 as
1832 * a valid dma-address. So we can use 0 as error value
1834 dma_dom->aperture[0]->bitmap[0] = 1;
1835 dma_dom->next_address = 0;
1841 dma_ops_domain_free(dma_dom);
1847 * little helper function to check whether a given protection domain is a
1850 static bool dma_ops_domain(struct protection_domain *domain)
1852 return domain->flags & PD_DMA_OPS_MASK;
1855 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1860 if (domain->mode != PAGE_MODE_NONE)
1861 pte_root = virt_to_phys(domain->pt_root);
1863 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1864 << DEV_ENTRY_MODE_SHIFT;
1865 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1867 flags = amd_iommu_dev_table[devid].data[1];
1870 flags |= DTE_FLAG_IOTLB;
1872 if (domain->flags & PD_IOMMUV2_MASK) {
1873 u64 gcr3 = __pa(domain->gcr3_tbl);
1874 u64 glx = domain->glx;
1877 pte_root |= DTE_FLAG_GV;
1878 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1880 /* First mask out possible old values for GCR3 table */
1881 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1884 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1887 /* Encode GCR3 table into DTE */
1888 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1891 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1894 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1898 flags &= ~(0xffffUL);
1899 flags |= domain->id;
1901 amd_iommu_dev_table[devid].data[1] = flags;
1902 amd_iommu_dev_table[devid].data[0] = pte_root;
1905 static void clear_dte_entry(u16 devid)
1907 /* remove entry from the device table seen by the hardware */
1908 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1909 amd_iommu_dev_table[devid].data[1] = 0;
1911 amd_iommu_apply_erratum_63(devid);
1914 static void do_attach(struct iommu_dev_data *dev_data,
1915 struct protection_domain *domain)
1917 struct amd_iommu *iommu;
1920 iommu = amd_iommu_rlookup_table[dev_data->devid];
1921 ats = dev_data->ats.enabled;
1923 /* Update data structures */
1924 dev_data->domain = domain;
1925 list_add(&dev_data->list, &domain->dev_list);
1926 set_dte_entry(dev_data->devid, domain, ats);
1928 /* Do reference counting */
1929 domain->dev_iommu[iommu->index] += 1;
1930 domain->dev_cnt += 1;
1932 /* Flush the DTE entry */
1933 device_flush_dte(dev_data);
1936 static void do_detach(struct iommu_dev_data *dev_data)
1938 struct amd_iommu *iommu;
1940 iommu = amd_iommu_rlookup_table[dev_data->devid];
1942 /* decrease reference counters */
1943 dev_data->domain->dev_iommu[iommu->index] -= 1;
1944 dev_data->domain->dev_cnt -= 1;
1946 /* Update data structures */
1947 dev_data->domain = NULL;
1948 list_del(&dev_data->list);
1949 clear_dte_entry(dev_data->devid);
1951 /* Flush the DTE entry */
1952 device_flush_dte(dev_data);
1956 * If a device is not yet associated with a domain, this function does
1957 * assigns it visible for the hardware
1959 static int __attach_device(struct iommu_dev_data *dev_data,
1960 struct protection_domain *domain)
1965 spin_lock(&domain->lock);
1967 if (dev_data->alias_data != NULL) {
1968 struct iommu_dev_data *alias_data = dev_data->alias_data;
1970 /* Some sanity checks */
1972 if (alias_data->domain != NULL &&
1973 alias_data->domain != domain)
1976 if (dev_data->domain != NULL &&
1977 dev_data->domain != domain)
1980 /* Do real assignment */
1981 if (alias_data->domain == NULL)
1982 do_attach(alias_data, domain);
1984 atomic_inc(&alias_data->bind);
1987 if (dev_data->domain == NULL)
1988 do_attach(dev_data, domain);
1990 atomic_inc(&dev_data->bind);
1997 spin_unlock(&domain->lock);
2003 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2005 pci_disable_ats(pdev);
2006 pci_disable_pri(pdev);
2007 pci_disable_pasid(pdev);
2010 /* FIXME: Change generic reset-function to do the same */
2011 static int pri_reset_while_enabled(struct pci_dev *pdev)
2016 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2020 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2021 control |= PCI_PRI_CTRL_RESET;
2022 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2027 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2032 /* FIXME: Hardcode number of outstanding requests for now */
2034 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2036 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2038 /* Only allow access to user-accessible pages */
2039 ret = pci_enable_pasid(pdev, 0);
2043 /* First reset the PRI state of the device */
2044 ret = pci_reset_pri(pdev);
2049 ret = pci_enable_pri(pdev, reqs);
2054 ret = pri_reset_while_enabled(pdev);
2059 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2066 pci_disable_pri(pdev);
2067 pci_disable_pasid(pdev);
2072 /* FIXME: Move this to PCI code */
2073 #define PCI_PRI_TLP_OFF (1 << 15)
2075 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2080 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2084 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2086 return (status & PCI_PRI_TLP_OFF) ? true : false;
2090 * If a device is not yet associated with a domain, this function does
2091 * assigns it visible for the hardware
2093 static int attach_device(struct device *dev,
2094 struct protection_domain *domain)
2096 struct pci_dev *pdev = to_pci_dev(dev);
2097 struct iommu_dev_data *dev_data;
2098 unsigned long flags;
2101 dev_data = get_dev_data(dev);
2103 if (domain->flags & PD_IOMMUV2_MASK) {
2104 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2107 if (pdev_iommuv2_enable(pdev) != 0)
2110 dev_data->ats.enabled = true;
2111 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2112 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2113 } else if (amd_iommu_iotlb_sup &&
2114 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2115 dev_data->ats.enabled = true;
2116 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2119 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2120 ret = __attach_device(dev_data, domain);
2121 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2124 * We might boot into a crash-kernel here. The crashed kernel
2125 * left the caches in the IOMMU dirty. So we have to flush
2126 * here to evict all dirty stuff.
2128 domain_flush_tlb_pde(domain);
2134 * Removes a device from a protection domain (unlocked)
2136 static void __detach_device(struct iommu_dev_data *dev_data)
2138 struct protection_domain *domain;
2139 unsigned long flags;
2141 BUG_ON(!dev_data->domain);
2143 domain = dev_data->domain;
2145 spin_lock_irqsave(&domain->lock, flags);
2147 if (dev_data->alias_data != NULL) {
2148 struct iommu_dev_data *alias_data = dev_data->alias_data;
2150 if (atomic_dec_and_test(&alias_data->bind))
2151 do_detach(alias_data);
2154 if (atomic_dec_and_test(&dev_data->bind))
2155 do_detach(dev_data);
2157 spin_unlock_irqrestore(&domain->lock, flags);
2160 * If we run in passthrough mode the device must be assigned to the
2161 * passthrough domain if it is detached from any other domain.
2162 * Make sure we can deassign from the pt_domain itself.
2164 if (dev_data->passthrough &&
2165 (dev_data->domain == NULL && domain != pt_domain))
2166 __attach_device(dev_data, pt_domain);
2170 * Removes a device from a protection domain (with devtable_lock held)
2172 static void detach_device(struct device *dev)
2174 struct protection_domain *domain;
2175 struct iommu_dev_data *dev_data;
2176 unsigned long flags;
2178 dev_data = get_dev_data(dev);
2179 domain = dev_data->domain;
2181 /* lock device table */
2182 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2183 __detach_device(dev_data);
2184 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2186 if (domain->flags & PD_IOMMUV2_MASK)
2187 pdev_iommuv2_disable(to_pci_dev(dev));
2188 else if (dev_data->ats.enabled)
2189 pci_disable_ats(to_pci_dev(dev));
2191 dev_data->ats.enabled = false;
2195 * Find out the protection domain structure for a given PCI device. This
2196 * will give us the pointer to the page table root for example.
2198 static struct protection_domain *domain_for_device(struct device *dev)
2200 struct iommu_dev_data *dev_data;
2201 struct protection_domain *dom = NULL;
2202 unsigned long flags;
2204 dev_data = get_dev_data(dev);
2206 if (dev_data->domain)
2207 return dev_data->domain;
2209 if (dev_data->alias_data != NULL) {
2210 struct iommu_dev_data *alias_data = dev_data->alias_data;
2212 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2213 if (alias_data->domain != NULL) {
2214 __attach_device(dev_data, alias_data->domain);
2215 dom = alias_data->domain;
2217 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2223 static int device_change_notifier(struct notifier_block *nb,
2224 unsigned long action, void *data)
2226 struct dma_ops_domain *dma_domain;
2227 struct protection_domain *domain;
2228 struct iommu_dev_data *dev_data;
2229 struct device *dev = data;
2230 struct amd_iommu *iommu;
2231 unsigned long flags;
2234 if (!check_device(dev))
2237 devid = get_device_id(dev);
2238 iommu = amd_iommu_rlookup_table[devid];
2239 dev_data = get_dev_data(dev);
2242 case BUS_NOTIFY_UNBOUND_DRIVER:
2244 domain = domain_for_device(dev);
2248 if (dev_data->passthrough)
2252 case BUS_NOTIFY_ADD_DEVICE:
2254 iommu_init_device(dev);
2256 domain = domain_for_device(dev);
2258 /* allocate a protection domain if a device is added */
2259 dma_domain = find_protection_domain(devid);
2262 dma_domain = dma_ops_domain_alloc();
2265 dma_domain->target_dev = devid;
2267 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2268 list_add_tail(&dma_domain->list, &iommu_pd_list);
2269 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2271 dev_data = get_dev_data(dev);
2273 if (!dev_data->passthrough)
2274 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2276 dev->archdata.dma_ops = &nommu_dma_ops;
2279 case BUS_NOTIFY_DEL_DEVICE:
2281 iommu_uninit_device(dev);
2287 iommu_completion_wait(iommu);
2293 static struct notifier_block device_nb = {
2294 .notifier_call = device_change_notifier,
2297 void amd_iommu_init_notifier(void)
2299 bus_register_notifier(&pci_bus_type, &device_nb);
2302 /*****************************************************************************
2304 * The next functions belong to the dma_ops mapping/unmapping code.
2306 *****************************************************************************/
2309 * In the dma_ops path we only have the struct device. This function
2310 * finds the corresponding IOMMU, the protection domain and the
2311 * requestor id for a given device.
2312 * If the device is not yet associated with a domain this is also done
2315 static struct protection_domain *get_domain(struct device *dev)
2317 struct protection_domain *domain;
2318 struct dma_ops_domain *dma_dom;
2319 u16 devid = get_device_id(dev);
2321 if (!check_device(dev))
2322 return ERR_PTR(-EINVAL);
2324 domain = domain_for_device(dev);
2325 if (domain != NULL && !dma_ops_domain(domain))
2326 return ERR_PTR(-EBUSY);
2331 /* Device not bount yet - bind it */
2332 dma_dom = find_protection_domain(devid);
2334 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2335 attach_device(dev, &dma_dom->domain);
2336 DUMP_printk("Using protection domain %d for device %s\n",
2337 dma_dom->domain.id, dev_name(dev));
2339 return &dma_dom->domain;
2342 static void update_device_table(struct protection_domain *domain)
2344 struct iommu_dev_data *dev_data;
2346 list_for_each_entry(dev_data, &domain->dev_list, list)
2347 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2350 static void update_domain(struct protection_domain *domain)
2352 if (!domain->updated)
2355 update_device_table(domain);
2357 domain_flush_devices(domain);
2358 domain_flush_tlb_pde(domain);
2360 domain->updated = false;
2364 * This function fetches the PTE for a given address in the aperture
2366 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2367 unsigned long address)
2369 struct aperture_range *aperture;
2370 u64 *pte, *pte_page;
2372 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2376 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2378 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2380 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2382 pte += PM_LEVEL_INDEX(0, address);
2384 update_domain(&dom->domain);
2390 * This is the generic map function. It maps one 4kb page at paddr to
2391 * the given address in the DMA address space for the domain.
2393 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2394 unsigned long address,
2400 WARN_ON(address > dom->aperture_size);
2404 pte = dma_ops_get_pte(dom, address);
2406 return DMA_ERROR_CODE;
2408 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2410 if (direction == DMA_TO_DEVICE)
2411 __pte |= IOMMU_PTE_IR;
2412 else if (direction == DMA_FROM_DEVICE)
2413 __pte |= IOMMU_PTE_IW;
2414 else if (direction == DMA_BIDIRECTIONAL)
2415 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2421 return (dma_addr_t)address;
2425 * The generic unmapping function for on page in the DMA address space.
2427 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2428 unsigned long address)
2430 struct aperture_range *aperture;
2433 if (address >= dom->aperture_size)
2436 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2440 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2444 pte += PM_LEVEL_INDEX(0, address);
2452 * This function contains common code for mapping of a physically
2453 * contiguous memory region into DMA address space. It is used by all
2454 * mapping functions provided with this IOMMU driver.
2455 * Must be called with the domain lock held.
2457 static dma_addr_t __map_single(struct device *dev,
2458 struct dma_ops_domain *dma_dom,
2465 dma_addr_t offset = paddr & ~PAGE_MASK;
2466 dma_addr_t address, start, ret;
2468 unsigned long align_mask = 0;
2471 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2474 INC_STATS_COUNTER(total_map_requests);
2477 INC_STATS_COUNTER(cross_page);
2480 align_mask = (1UL << get_order(size)) - 1;
2483 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2485 if (unlikely(address == DMA_ERROR_CODE)) {
2487 * setting next_address here will let the address
2488 * allocator only scan the new allocated range in the
2489 * first run. This is a small optimization.
2491 dma_dom->next_address = dma_dom->aperture_size;
2493 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2497 * aperture was successfully enlarged by 128 MB, try
2504 for (i = 0; i < pages; ++i) {
2505 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2506 if (ret == DMA_ERROR_CODE)
2514 ADD_STATS_COUNTER(alloced_io_mem, size);
2516 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2517 domain_flush_tlb(&dma_dom->domain);
2518 dma_dom->need_flush = false;
2519 } else if (unlikely(amd_iommu_np_cache))
2520 domain_flush_pages(&dma_dom->domain, address, size);
2527 for (--i; i >= 0; --i) {
2529 dma_ops_domain_unmap(dma_dom, start);
2532 dma_ops_free_addresses(dma_dom, address, pages);
2534 return DMA_ERROR_CODE;
2538 * Does the reverse of the __map_single function. Must be called with
2539 * the domain lock held too
2541 static void __unmap_single(struct dma_ops_domain *dma_dom,
2542 dma_addr_t dma_addr,
2546 dma_addr_t flush_addr;
2547 dma_addr_t i, start;
2550 if ((dma_addr == DMA_ERROR_CODE) ||
2551 (dma_addr + size > dma_dom->aperture_size))
2554 flush_addr = dma_addr;
2555 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2556 dma_addr &= PAGE_MASK;
2559 for (i = 0; i < pages; ++i) {
2560 dma_ops_domain_unmap(dma_dom, start);
2564 SUB_STATS_COUNTER(alloced_io_mem, size);
2566 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2568 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2569 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2570 dma_dom->need_flush = false;
2575 * The exported map_single function for dma_ops.
2577 static dma_addr_t map_page(struct device *dev, struct page *page,
2578 unsigned long offset, size_t size,
2579 enum dma_data_direction dir,
2580 struct dma_attrs *attrs)
2582 unsigned long flags;
2583 struct protection_domain *domain;
2586 phys_addr_t paddr = page_to_phys(page) + offset;
2588 INC_STATS_COUNTER(cnt_map_single);
2590 domain = get_domain(dev);
2591 if (PTR_ERR(domain) == -EINVAL)
2592 return (dma_addr_t)paddr;
2593 else if (IS_ERR(domain))
2594 return DMA_ERROR_CODE;
2596 dma_mask = *dev->dma_mask;
2598 spin_lock_irqsave(&domain->lock, flags);
2600 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2602 if (addr == DMA_ERROR_CODE)
2605 domain_flush_complete(domain);
2608 spin_unlock_irqrestore(&domain->lock, flags);
2614 * The exported unmap_single function for dma_ops.
2616 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2617 enum dma_data_direction dir, struct dma_attrs *attrs)
2619 unsigned long flags;
2620 struct protection_domain *domain;
2622 INC_STATS_COUNTER(cnt_unmap_single);
2624 domain = get_domain(dev);
2628 spin_lock_irqsave(&domain->lock, flags);
2630 __unmap_single(domain->priv, dma_addr, size, dir);
2632 domain_flush_complete(domain);
2634 spin_unlock_irqrestore(&domain->lock, flags);
2638 * This is a special map_sg function which is used if we should map a
2639 * device which is not handled by an AMD IOMMU in the system.
2641 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2642 int nelems, int dir)
2644 struct scatterlist *s;
2647 for_each_sg(sglist, s, nelems, i) {
2648 s->dma_address = (dma_addr_t)sg_phys(s);
2649 s->dma_length = s->length;
2656 * The exported map_sg function for dma_ops (handles scatter-gather
2659 static int map_sg(struct device *dev, struct scatterlist *sglist,
2660 int nelems, enum dma_data_direction dir,
2661 struct dma_attrs *attrs)
2663 unsigned long flags;
2664 struct protection_domain *domain;
2666 struct scatterlist *s;
2668 int mapped_elems = 0;
2671 INC_STATS_COUNTER(cnt_map_sg);
2673 domain = get_domain(dev);
2674 if (PTR_ERR(domain) == -EINVAL)
2675 return map_sg_no_iommu(dev, sglist, nelems, dir);
2676 else if (IS_ERR(domain))
2679 dma_mask = *dev->dma_mask;
2681 spin_lock_irqsave(&domain->lock, flags);
2683 for_each_sg(sglist, s, nelems, i) {
2686 s->dma_address = __map_single(dev, domain->priv,
2687 paddr, s->length, dir, false,
2690 if (s->dma_address) {
2691 s->dma_length = s->length;
2697 domain_flush_complete(domain);
2700 spin_unlock_irqrestore(&domain->lock, flags);
2702 return mapped_elems;
2704 for_each_sg(sglist, s, mapped_elems, i) {
2706 __unmap_single(domain->priv, s->dma_address,
2707 s->dma_length, dir);
2708 s->dma_address = s->dma_length = 0;
2717 * The exported map_sg function for dma_ops (handles scatter-gather
2720 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2721 int nelems, enum dma_data_direction dir,
2722 struct dma_attrs *attrs)
2724 unsigned long flags;
2725 struct protection_domain *domain;
2726 struct scatterlist *s;
2729 INC_STATS_COUNTER(cnt_unmap_sg);
2731 domain = get_domain(dev);
2735 spin_lock_irqsave(&domain->lock, flags);
2737 for_each_sg(sglist, s, nelems, i) {
2738 __unmap_single(domain->priv, s->dma_address,
2739 s->dma_length, dir);
2740 s->dma_address = s->dma_length = 0;
2743 domain_flush_complete(domain);
2745 spin_unlock_irqrestore(&domain->lock, flags);
2749 * The exported alloc_coherent function for dma_ops.
2751 static void *alloc_coherent(struct device *dev, size_t size,
2752 dma_addr_t *dma_addr, gfp_t flag,
2753 struct dma_attrs *attrs)
2755 unsigned long flags;
2757 struct protection_domain *domain;
2759 u64 dma_mask = dev->coherent_dma_mask;
2761 INC_STATS_COUNTER(cnt_alloc_coherent);
2763 domain = get_domain(dev);
2764 if (PTR_ERR(domain) == -EINVAL) {
2765 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2766 *dma_addr = __pa(virt_addr);
2768 } else if (IS_ERR(domain))
2771 dma_mask = dev->coherent_dma_mask;
2772 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2775 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2779 paddr = virt_to_phys(virt_addr);
2782 dma_mask = *dev->dma_mask;
2784 spin_lock_irqsave(&domain->lock, flags);
2786 *dma_addr = __map_single(dev, domain->priv, paddr,
2787 size, DMA_BIDIRECTIONAL, true, dma_mask);
2789 if (*dma_addr == DMA_ERROR_CODE) {
2790 spin_unlock_irqrestore(&domain->lock, flags);
2794 domain_flush_complete(domain);
2796 spin_unlock_irqrestore(&domain->lock, flags);
2802 free_pages((unsigned long)virt_addr, get_order(size));
2808 * The exported free_coherent function for dma_ops.
2810 static void free_coherent(struct device *dev, size_t size,
2811 void *virt_addr, dma_addr_t dma_addr,
2812 struct dma_attrs *attrs)
2814 unsigned long flags;
2815 struct protection_domain *domain;
2817 INC_STATS_COUNTER(cnt_free_coherent);
2819 domain = get_domain(dev);
2823 spin_lock_irqsave(&domain->lock, flags);
2825 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2827 domain_flush_complete(domain);
2829 spin_unlock_irqrestore(&domain->lock, flags);
2832 free_pages((unsigned long)virt_addr, get_order(size));
2836 * This function is called by the DMA layer to find out if we can handle a
2837 * particular device. It is part of the dma_ops.
2839 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2841 return check_device(dev);
2845 * The function for pre-allocating protection domains.
2847 * If the driver core informs the DMA layer if a driver grabs a device
2848 * we don't need to preallocate the protection domains anymore.
2849 * For now we have to.
2851 static void __init prealloc_protection_domains(void)
2853 struct iommu_dev_data *dev_data;
2854 struct dma_ops_domain *dma_dom;
2855 struct pci_dev *dev = NULL;
2858 for_each_pci_dev(dev) {
2860 /* Do we handle this device? */
2861 if (!check_device(&dev->dev))
2864 dev_data = get_dev_data(&dev->dev);
2865 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2866 /* Make sure passthrough domain is allocated */
2867 alloc_passthrough_domain();
2868 dev_data->passthrough = true;
2869 attach_device(&dev->dev, pt_domain);
2870 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2871 dev_name(&dev->dev));
2874 /* Is there already any domain for it? */
2875 if (domain_for_device(&dev->dev))
2878 devid = get_device_id(&dev->dev);
2880 dma_dom = dma_ops_domain_alloc();
2883 init_unity_mappings_for_device(dma_dom, devid);
2884 dma_dom->target_dev = devid;
2886 attach_device(&dev->dev, &dma_dom->domain);
2888 list_add_tail(&dma_dom->list, &iommu_pd_list);
2892 static struct dma_map_ops amd_iommu_dma_ops = {
2893 .alloc = alloc_coherent,
2894 .free = free_coherent,
2895 .map_page = map_page,
2896 .unmap_page = unmap_page,
2898 .unmap_sg = unmap_sg,
2899 .dma_supported = amd_iommu_dma_supported,
2902 static unsigned device_dma_ops_init(void)
2904 struct iommu_dev_data *dev_data;
2905 struct pci_dev *pdev = NULL;
2906 unsigned unhandled = 0;
2908 for_each_pci_dev(pdev) {
2909 if (!check_device(&pdev->dev)) {
2911 iommu_ignore_device(&pdev->dev);
2917 dev_data = get_dev_data(&pdev->dev);
2919 if (!dev_data->passthrough)
2920 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2922 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
2929 * The function which clues the AMD IOMMU driver into dma_ops.
2932 void __init amd_iommu_init_api(void)
2934 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2937 int __init amd_iommu_init_dma_ops(void)
2939 struct amd_iommu *iommu;
2943 * first allocate a default protection domain for every IOMMU we
2944 * found in the system. Devices not assigned to any other
2945 * protection domain will be assigned to the default one.
2947 for_each_iommu(iommu) {
2948 iommu->default_dom = dma_ops_domain_alloc();
2949 if (iommu->default_dom == NULL)
2951 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2952 ret = iommu_init_unity_mappings(iommu);
2958 * Pre-allocate the protection domains for each device.
2960 prealloc_protection_domains();
2965 /* Make the driver finally visible to the drivers */
2966 unhandled = device_dma_ops_init();
2967 if (unhandled && max_pfn > MAX_DMA32_PFN) {
2968 /* There are unhandled devices - initialize swiotlb for them */
2972 amd_iommu_stats_init();
2974 if (amd_iommu_unmap_flush)
2975 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2977 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2983 for_each_iommu(iommu) {
2984 if (iommu->default_dom)
2985 dma_ops_domain_free(iommu->default_dom);
2991 /*****************************************************************************
2993 * The following functions belong to the exported interface of AMD IOMMU
2995 * This interface allows access to lower level functions of the IOMMU
2996 * like protection domain handling and assignement of devices to domains
2997 * which is not possible with the dma_ops interface.
2999 *****************************************************************************/
3001 static void cleanup_domain(struct protection_domain *domain)
3003 struct iommu_dev_data *dev_data, *next;
3004 unsigned long flags;
3006 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3008 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3009 __detach_device(dev_data);
3010 atomic_set(&dev_data->bind, 0);
3013 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3016 static void protection_domain_free(struct protection_domain *domain)
3021 del_domain_from_list(domain);
3024 domain_id_free(domain->id);
3029 static struct protection_domain *protection_domain_alloc(void)
3031 struct protection_domain *domain;
3033 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3037 spin_lock_init(&domain->lock);
3038 mutex_init(&domain->api_lock);
3039 domain->id = domain_id_alloc();
3042 INIT_LIST_HEAD(&domain->dev_list);
3044 add_domain_to_list(domain);
3054 static int __init alloc_passthrough_domain(void)
3056 if (pt_domain != NULL)
3059 /* allocate passthrough domain */
3060 pt_domain = protection_domain_alloc();
3064 pt_domain->mode = PAGE_MODE_NONE;
3068 static int amd_iommu_domain_init(struct iommu_domain *dom)
3070 struct protection_domain *domain;
3072 domain = protection_domain_alloc();
3076 domain->mode = PAGE_MODE_3_LEVEL;
3077 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3078 if (!domain->pt_root)
3081 domain->iommu_domain = dom;
3088 protection_domain_free(domain);
3093 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3095 struct protection_domain *domain = dom->priv;
3100 if (domain->dev_cnt > 0)
3101 cleanup_domain(domain);
3103 BUG_ON(domain->dev_cnt != 0);
3105 if (domain->mode != PAGE_MODE_NONE)
3106 free_pagetable(domain);
3108 if (domain->flags & PD_IOMMUV2_MASK)
3109 free_gcr3_table(domain);
3111 protection_domain_free(domain);
3116 static void amd_iommu_detach_device(struct iommu_domain *dom,
3119 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3120 struct amd_iommu *iommu;
3123 if (!check_device(dev))
3126 devid = get_device_id(dev);
3128 if (dev_data->domain != NULL)
3131 iommu = amd_iommu_rlookup_table[devid];
3135 iommu_completion_wait(iommu);
3138 static int amd_iommu_attach_device(struct iommu_domain *dom,
3141 struct protection_domain *domain = dom->priv;
3142 struct iommu_dev_data *dev_data;
3143 struct amd_iommu *iommu;
3146 if (!check_device(dev))
3149 dev_data = dev->archdata.iommu;
3151 iommu = amd_iommu_rlookup_table[dev_data->devid];
3155 if (dev_data->domain)
3158 ret = attach_device(dev, domain);
3160 iommu_completion_wait(iommu);
3165 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3166 phys_addr_t paddr, size_t page_size, int iommu_prot)
3168 struct protection_domain *domain = dom->priv;
3172 if (domain->mode == PAGE_MODE_NONE)
3175 if (iommu_prot & IOMMU_READ)
3176 prot |= IOMMU_PROT_IR;
3177 if (iommu_prot & IOMMU_WRITE)
3178 prot |= IOMMU_PROT_IW;
3180 mutex_lock(&domain->api_lock);
3181 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3182 mutex_unlock(&domain->api_lock);
3187 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3190 struct protection_domain *domain = dom->priv;
3193 if (domain->mode == PAGE_MODE_NONE)
3196 mutex_lock(&domain->api_lock);
3197 unmap_size = iommu_unmap_page(domain, iova, page_size);
3198 mutex_unlock(&domain->api_lock);
3200 domain_flush_tlb_pde(domain);
3205 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3208 struct protection_domain *domain = dom->priv;
3209 unsigned long offset_mask;
3213 if (domain->mode == PAGE_MODE_NONE)
3216 pte = fetch_pte(domain, iova);
3218 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3221 if (PM_PTE_LEVEL(*pte) == 0)
3222 offset_mask = PAGE_SIZE - 1;
3224 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3226 __pte = *pte & PM_ADDR_MASK;
3227 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3232 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3236 case IOMMU_CAP_CACHE_COHERENCY:
3243 static int amd_iommu_device_group(struct device *dev, unsigned int *groupid)
3245 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3246 struct pci_dev *pdev = to_pci_dev(dev);
3252 if (pdev->is_virtfn || !iommu_group_mf)
3253 devid = dev_data->devid;
3255 devid = calc_devid(pdev->bus->number,
3256 PCI_DEVFN(PCI_SLOT(pdev->devfn), 0));
3258 *groupid = amd_iommu_alias_table[devid];
3263 static struct iommu_ops amd_iommu_ops = {
3264 .domain_init = amd_iommu_domain_init,
3265 .domain_destroy = amd_iommu_domain_destroy,
3266 .attach_dev = amd_iommu_attach_device,
3267 .detach_dev = amd_iommu_detach_device,
3268 .map = amd_iommu_map,
3269 .unmap = amd_iommu_unmap,
3270 .iova_to_phys = amd_iommu_iova_to_phys,
3271 .domain_has_cap = amd_iommu_domain_has_cap,
3272 .device_group = amd_iommu_device_group,
3273 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3276 /*****************************************************************************
3278 * The next functions do a basic initialization of IOMMU for pass through
3281 * In passthrough mode the IOMMU is initialized and enabled but not used for
3282 * DMA-API translation.
3284 *****************************************************************************/
3286 int __init amd_iommu_init_passthrough(void)
3288 struct iommu_dev_data *dev_data;
3289 struct pci_dev *dev = NULL;
3290 struct amd_iommu *iommu;
3294 ret = alloc_passthrough_domain();
3298 for_each_pci_dev(dev) {
3299 if (!check_device(&dev->dev))
3302 dev_data = get_dev_data(&dev->dev);
3303 dev_data->passthrough = true;
3305 devid = get_device_id(&dev->dev);
3307 iommu = amd_iommu_rlookup_table[devid];
3311 attach_device(&dev->dev, pt_domain);
3314 amd_iommu_stats_init();
3316 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3321 /* IOMMUv2 specific functions */
3322 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3324 return atomic_notifier_chain_register(&ppr_notifier, nb);
3326 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3328 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3330 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3332 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3334 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3336 struct protection_domain *domain = dom->priv;
3337 unsigned long flags;
3339 spin_lock_irqsave(&domain->lock, flags);
3341 /* Update data structure */
3342 domain->mode = PAGE_MODE_NONE;
3343 domain->updated = true;
3345 /* Make changes visible to IOMMUs */
3346 update_domain(domain);
3348 /* Page-table is not visible to IOMMU anymore, so free it */
3349 free_pagetable(domain);
3351 spin_unlock_irqrestore(&domain->lock, flags);
3353 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3355 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3357 struct protection_domain *domain = dom->priv;
3358 unsigned long flags;
3361 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3364 /* Number of GCR3 table levels required */
3365 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3368 if (levels > amd_iommu_max_glx_val)
3371 spin_lock_irqsave(&domain->lock, flags);
3374 * Save us all sanity checks whether devices already in the
3375 * domain support IOMMUv2. Just force that the domain has no
3376 * devices attached when it is switched into IOMMUv2 mode.
3379 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3383 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3384 if (domain->gcr3_tbl == NULL)
3387 domain->glx = levels;
3388 domain->flags |= PD_IOMMUV2_MASK;
3389 domain->updated = true;
3391 update_domain(domain);
3396 spin_unlock_irqrestore(&domain->lock, flags);
3400 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3402 static int __flush_pasid(struct protection_domain *domain, int pasid,
3403 u64 address, bool size)
3405 struct iommu_dev_data *dev_data;
3406 struct iommu_cmd cmd;
3409 if (!(domain->flags & PD_IOMMUV2_MASK))
3412 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3415 * IOMMU TLB needs to be flushed before Device TLB to
3416 * prevent device TLB refill from IOMMU TLB
3418 for (i = 0; i < amd_iommus_present; ++i) {
3419 if (domain->dev_iommu[i] == 0)
3422 ret = iommu_queue_command(amd_iommus[i], &cmd);
3427 /* Wait until IOMMU TLB flushes are complete */
3428 domain_flush_complete(domain);
3430 /* Now flush device TLBs */
3431 list_for_each_entry(dev_data, &domain->dev_list, list) {
3432 struct amd_iommu *iommu;
3435 BUG_ON(!dev_data->ats.enabled);
3437 qdep = dev_data->ats.qdep;
3438 iommu = amd_iommu_rlookup_table[dev_data->devid];
3440 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3441 qdep, address, size);
3443 ret = iommu_queue_command(iommu, &cmd);
3448 /* Wait until all device TLBs are flushed */
3449 domain_flush_complete(domain);
3458 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3461 INC_STATS_COUNTER(invalidate_iotlb);
3463 return __flush_pasid(domain, pasid, address, false);
3466 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3469 struct protection_domain *domain = dom->priv;
3470 unsigned long flags;
3473 spin_lock_irqsave(&domain->lock, flags);
3474 ret = __amd_iommu_flush_page(domain, pasid, address);
3475 spin_unlock_irqrestore(&domain->lock, flags);
3479 EXPORT_SYMBOL(amd_iommu_flush_page);
3481 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3483 INC_STATS_COUNTER(invalidate_iotlb_all);
3485 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3489 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3491 struct protection_domain *domain = dom->priv;
3492 unsigned long flags;
3495 spin_lock_irqsave(&domain->lock, flags);
3496 ret = __amd_iommu_flush_tlb(domain, pasid);
3497 spin_unlock_irqrestore(&domain->lock, flags);
3501 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3503 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3510 index = (pasid >> (9 * level)) & 0x1ff;
3516 if (!(*pte & GCR3_VALID)) {
3520 root = (void *)get_zeroed_page(GFP_ATOMIC);
3524 *pte = __pa(root) | GCR3_VALID;
3527 root = __va(*pte & PAGE_MASK);
3535 static int __set_gcr3(struct protection_domain *domain, int pasid,
3540 if (domain->mode != PAGE_MODE_NONE)
3543 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3547 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3549 return __amd_iommu_flush_tlb(domain, pasid);
3552 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3556 if (domain->mode != PAGE_MODE_NONE)
3559 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3565 return __amd_iommu_flush_tlb(domain, pasid);
3568 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3571 struct protection_domain *domain = dom->priv;
3572 unsigned long flags;
3575 spin_lock_irqsave(&domain->lock, flags);
3576 ret = __set_gcr3(domain, pasid, cr3);
3577 spin_unlock_irqrestore(&domain->lock, flags);
3581 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3583 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3585 struct protection_domain *domain = dom->priv;
3586 unsigned long flags;
3589 spin_lock_irqsave(&domain->lock, flags);
3590 ret = __clear_gcr3(domain, pasid);
3591 spin_unlock_irqrestore(&domain->lock, flags);
3595 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3597 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3598 int status, int tag)
3600 struct iommu_dev_data *dev_data;
3601 struct amd_iommu *iommu;
3602 struct iommu_cmd cmd;
3604 INC_STATS_COUNTER(complete_ppr);
3606 dev_data = get_dev_data(&pdev->dev);
3607 iommu = amd_iommu_rlookup_table[dev_data->devid];
3609 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3610 tag, dev_data->pri_tlp);
3612 return iommu_queue_command(iommu, &cmd);
3614 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3616 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3618 struct protection_domain *domain;
3620 domain = get_domain(&pdev->dev);
3624 /* Only return IOMMUv2 domains */
3625 if (!(domain->flags & PD_IOMMUV2_MASK))
3628 return domain->iommu_domain;
3630 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3632 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3634 struct iommu_dev_data *dev_data;
3636 if (!amd_iommu_v2_supported())
3639 dev_data = get_dev_data(&pdev->dev);
3640 dev_data->errata |= (1 << erratum);
3642 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3644 int amd_iommu_device_info(struct pci_dev *pdev,
3645 struct amd_iommu_device_info *info)
3650 if (pdev == NULL || info == NULL)
3653 if (!amd_iommu_v2_supported())
3656 memset(info, 0, sizeof(*info));
3658 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3660 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3662 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3664 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3666 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3670 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3671 max_pasids = min(max_pasids, (1 << 20));
3673 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3674 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3676 features = pci_pasid_features(pdev);
3677 if (features & PCI_PASID_CAP_EXEC)
3678 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3679 if (features & PCI_PASID_CAP_PRIV)
3680 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3685 EXPORT_SYMBOL(amd_iommu_device_info);