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);
388 static struct dentry *stats_dir;
389 static struct dentry *de_fflush;
391 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
393 if (stats_dir == NULL)
396 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
400 static void amd_iommu_stats_init(void)
402 stats_dir = debugfs_create_dir("amd-iommu", NULL);
403 if (stats_dir == NULL)
406 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
407 &amd_iommu_unmap_flush);
409 amd_iommu_stats_add(&compl_wait);
410 amd_iommu_stats_add(&cnt_map_single);
411 amd_iommu_stats_add(&cnt_unmap_single);
412 amd_iommu_stats_add(&cnt_map_sg);
413 amd_iommu_stats_add(&cnt_unmap_sg);
414 amd_iommu_stats_add(&cnt_alloc_coherent);
415 amd_iommu_stats_add(&cnt_free_coherent);
416 amd_iommu_stats_add(&cross_page);
417 amd_iommu_stats_add(&domain_flush_single);
418 amd_iommu_stats_add(&domain_flush_all);
419 amd_iommu_stats_add(&alloced_io_mem);
420 amd_iommu_stats_add(&total_map_requests);
421 amd_iommu_stats_add(&complete_ppr);
422 amd_iommu_stats_add(&invalidate_iotlb);
423 amd_iommu_stats_add(&invalidate_iotlb_all);
424 amd_iommu_stats_add(&pri_requests);
429 /****************************************************************************
431 * Interrupt handling functions
433 ****************************************************************************/
435 static void dump_dte_entry(u16 devid)
439 for (i = 0; i < 4; ++i)
440 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
441 amd_iommu_dev_table[devid].data[i]);
444 static void dump_command(unsigned long phys_addr)
446 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
449 for (i = 0; i < 4; ++i)
450 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
453 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
455 int type, devid, domid, flags;
456 volatile u32 *event = __evt;
461 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
462 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
463 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
464 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
465 address = (u64)(((u64)event[3]) << 32) | event[2];
468 /* Did we hit the erratum? */
469 if (++count == LOOP_TIMEOUT) {
470 pr_err("AMD-Vi: No event written to event log\n");
477 printk(KERN_ERR "AMD-Vi: Event logged [");
480 case EVENT_TYPE_ILL_DEV:
481 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
482 "address=0x%016llx flags=0x%04x]\n",
483 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
485 dump_dte_entry(devid);
487 case EVENT_TYPE_IO_FAULT:
488 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
489 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
490 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
491 domid, address, flags);
493 case EVENT_TYPE_DEV_TAB_ERR:
494 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
495 "address=0x%016llx flags=0x%04x]\n",
496 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
499 case EVENT_TYPE_PAGE_TAB_ERR:
500 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
501 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
502 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
503 domid, address, flags);
505 case EVENT_TYPE_ILL_CMD:
506 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
507 dump_command(address);
509 case EVENT_TYPE_CMD_HARD_ERR:
510 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
511 "flags=0x%04x]\n", address, flags);
513 case EVENT_TYPE_IOTLB_INV_TO:
514 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
515 "address=0x%016llx]\n",
516 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
519 case EVENT_TYPE_INV_DEV_REQ:
520 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
521 "address=0x%016llx flags=0x%04x]\n",
522 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
526 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
529 memset(__evt, 0, 4 * sizeof(u32));
532 static void iommu_poll_events(struct amd_iommu *iommu)
537 spin_lock_irqsave(&iommu->lock, flags);
539 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
540 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
542 while (head != tail) {
543 iommu_print_event(iommu, iommu->evt_buf + head);
544 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
547 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
549 spin_unlock_irqrestore(&iommu->lock, flags);
552 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
554 struct amd_iommu_fault fault;
556 INC_STATS_COUNTER(pri_requests);
558 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
559 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
563 fault.address = raw[1];
564 fault.pasid = PPR_PASID(raw[0]);
565 fault.device_id = PPR_DEVID(raw[0]);
566 fault.tag = PPR_TAG(raw[0]);
567 fault.flags = PPR_FLAGS(raw[0]);
569 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
572 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
577 if (iommu->ppr_log == NULL)
580 /* enable ppr interrupts again */
581 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
583 spin_lock_irqsave(&iommu->lock, flags);
585 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
586 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
588 while (head != tail) {
593 raw = (u64 *)(iommu->ppr_log + head);
596 * Hardware bug: Interrupt may arrive before the entry is
597 * written to memory. If this happens we need to wait for the
600 for (i = 0; i < LOOP_TIMEOUT; ++i) {
601 if (PPR_REQ_TYPE(raw[0]) != 0)
606 /* Avoid memcpy function-call overhead */
611 * To detect the hardware bug we need to clear the entry
614 raw[0] = raw[1] = 0UL;
616 /* Update head pointer of hardware ring-buffer */
617 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
618 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
621 * Release iommu->lock because ppr-handling might need to
624 spin_unlock_irqrestore(&iommu->lock, flags);
626 /* Handle PPR entry */
627 iommu_handle_ppr_entry(iommu, entry);
629 spin_lock_irqsave(&iommu->lock, flags);
631 /* Refresh ring-buffer information */
632 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
633 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
636 spin_unlock_irqrestore(&iommu->lock, flags);
639 irqreturn_t amd_iommu_int_thread(int irq, void *data)
641 struct amd_iommu *iommu;
643 for_each_iommu(iommu) {
644 iommu_poll_events(iommu);
645 iommu_poll_ppr_log(iommu);
651 irqreturn_t amd_iommu_int_handler(int irq, void *data)
653 return IRQ_WAKE_THREAD;
656 /****************************************************************************
658 * IOMMU command queuing functions
660 ****************************************************************************/
662 static int wait_on_sem(volatile u64 *sem)
666 while (*sem == 0 && i < LOOP_TIMEOUT) {
671 if (i == LOOP_TIMEOUT) {
672 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
679 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
680 struct iommu_cmd *cmd,
685 target = iommu->cmd_buf + tail;
686 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
688 /* Copy command to buffer */
689 memcpy(target, cmd, sizeof(*cmd));
691 /* Tell the IOMMU about it */
692 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
695 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
697 WARN_ON(address & 0x7ULL);
699 memset(cmd, 0, sizeof(*cmd));
700 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
701 cmd->data[1] = upper_32_bits(__pa(address));
703 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
706 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
708 memset(cmd, 0, sizeof(*cmd));
709 cmd->data[0] = devid;
710 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
713 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
714 size_t size, u16 domid, int pde)
719 pages = iommu_num_pages(address, size, PAGE_SIZE);
724 * If we have to flush more than one page, flush all
725 * TLB entries for this domain
727 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
731 address &= PAGE_MASK;
733 memset(cmd, 0, sizeof(*cmd));
734 cmd->data[1] |= domid;
735 cmd->data[2] = lower_32_bits(address);
736 cmd->data[3] = upper_32_bits(address);
737 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
738 if (s) /* size bit - we flush more than one 4kb page */
739 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
740 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
741 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
744 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
745 u64 address, size_t size)
750 pages = iommu_num_pages(address, size, PAGE_SIZE);
755 * If we have to flush more than one page, flush all
756 * TLB entries for this domain
758 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
762 address &= PAGE_MASK;
764 memset(cmd, 0, sizeof(*cmd));
765 cmd->data[0] = devid;
766 cmd->data[0] |= (qdep & 0xff) << 24;
767 cmd->data[1] = devid;
768 cmd->data[2] = lower_32_bits(address);
769 cmd->data[3] = upper_32_bits(address);
770 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
772 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
775 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
776 u64 address, bool size)
778 memset(cmd, 0, sizeof(*cmd));
780 address &= ~(0xfffULL);
782 cmd->data[0] = pasid & PASID_MASK;
783 cmd->data[1] = domid;
784 cmd->data[2] = lower_32_bits(address);
785 cmd->data[3] = upper_32_bits(address);
786 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
787 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
789 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
790 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
793 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
794 int qdep, u64 address, bool size)
796 memset(cmd, 0, sizeof(*cmd));
798 address &= ~(0xfffULL);
800 cmd->data[0] = devid;
801 cmd->data[0] |= (pasid & 0xff) << 16;
802 cmd->data[0] |= (qdep & 0xff) << 24;
803 cmd->data[1] = devid;
804 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
805 cmd->data[2] = lower_32_bits(address);
806 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
807 cmd->data[3] = upper_32_bits(address);
809 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
810 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
813 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
814 int status, int tag, bool gn)
816 memset(cmd, 0, sizeof(*cmd));
818 cmd->data[0] = devid;
820 cmd->data[1] = pasid & PASID_MASK;
821 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
823 cmd->data[3] = tag & 0x1ff;
824 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
826 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
829 static void build_inv_all(struct iommu_cmd *cmd)
831 memset(cmd, 0, sizeof(*cmd));
832 CMD_SET_TYPE(cmd, CMD_INV_ALL);
836 * Writes the command to the IOMMUs command buffer and informs the
837 * hardware about the new command.
839 static int iommu_queue_command_sync(struct amd_iommu *iommu,
840 struct iommu_cmd *cmd,
843 u32 left, tail, head, next_tail;
846 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
849 spin_lock_irqsave(&iommu->lock, flags);
851 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
852 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
853 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
854 left = (head - next_tail) % iommu->cmd_buf_size;
857 struct iommu_cmd sync_cmd;
858 volatile u64 sem = 0;
861 build_completion_wait(&sync_cmd, (u64)&sem);
862 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
864 spin_unlock_irqrestore(&iommu->lock, flags);
866 if ((ret = wait_on_sem(&sem)) != 0)
872 copy_cmd_to_buffer(iommu, cmd, tail);
874 /* We need to sync now to make sure all commands are processed */
875 iommu->need_sync = sync;
877 spin_unlock_irqrestore(&iommu->lock, flags);
882 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
884 return iommu_queue_command_sync(iommu, cmd, true);
888 * This function queues a completion wait command into the command
891 static int iommu_completion_wait(struct amd_iommu *iommu)
893 struct iommu_cmd cmd;
894 volatile u64 sem = 0;
897 if (!iommu->need_sync)
900 build_completion_wait(&cmd, (u64)&sem);
902 ret = iommu_queue_command_sync(iommu, &cmd, false);
906 return wait_on_sem(&sem);
909 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
911 struct iommu_cmd cmd;
913 build_inv_dte(&cmd, devid);
915 return iommu_queue_command(iommu, &cmd);
918 static void iommu_flush_dte_all(struct amd_iommu *iommu)
922 for (devid = 0; devid <= 0xffff; ++devid)
923 iommu_flush_dte(iommu, devid);
925 iommu_completion_wait(iommu);
929 * This function uses heavy locking and may disable irqs for some time. But
930 * this is no issue because it is only called during resume.
932 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
936 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
937 struct iommu_cmd cmd;
938 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
940 iommu_queue_command(iommu, &cmd);
943 iommu_completion_wait(iommu);
946 static void iommu_flush_all(struct amd_iommu *iommu)
948 struct iommu_cmd cmd;
952 iommu_queue_command(iommu, &cmd);
953 iommu_completion_wait(iommu);
956 void iommu_flush_all_caches(struct amd_iommu *iommu)
958 if (iommu_feature(iommu, FEATURE_IA)) {
959 iommu_flush_all(iommu);
961 iommu_flush_dte_all(iommu);
962 iommu_flush_tlb_all(iommu);
967 * Command send function for flushing on-device TLB
969 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
970 u64 address, size_t size)
972 struct amd_iommu *iommu;
973 struct iommu_cmd cmd;
976 qdep = dev_data->ats.qdep;
977 iommu = amd_iommu_rlookup_table[dev_data->devid];
979 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
981 return iommu_queue_command(iommu, &cmd);
985 * Command send function for invalidating a device table entry
987 static int device_flush_dte(struct iommu_dev_data *dev_data)
989 struct amd_iommu *iommu;
992 iommu = amd_iommu_rlookup_table[dev_data->devid];
994 ret = iommu_flush_dte(iommu, dev_data->devid);
998 if (dev_data->ats.enabled)
999 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1005 * TLB invalidation function which is called from the mapping functions.
1006 * It invalidates a single PTE if the range to flush is within a single
1007 * page. Otherwise it flushes the whole TLB of the IOMMU.
1009 static void __domain_flush_pages(struct protection_domain *domain,
1010 u64 address, size_t size, int pde)
1012 struct iommu_dev_data *dev_data;
1013 struct iommu_cmd cmd;
1016 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1018 for (i = 0; i < amd_iommus_present; ++i) {
1019 if (!domain->dev_iommu[i])
1023 * Devices of this domain are behind this IOMMU
1024 * We need a TLB flush
1026 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1029 list_for_each_entry(dev_data, &domain->dev_list, list) {
1031 if (!dev_data->ats.enabled)
1034 ret |= device_flush_iotlb(dev_data, address, size);
1040 static void domain_flush_pages(struct protection_domain *domain,
1041 u64 address, size_t size)
1043 __domain_flush_pages(domain, address, size, 0);
1046 /* Flush the whole IO/TLB for a given protection domain */
1047 static void domain_flush_tlb(struct protection_domain *domain)
1049 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1052 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1053 static void domain_flush_tlb_pde(struct protection_domain *domain)
1055 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1058 static void domain_flush_complete(struct protection_domain *domain)
1062 for (i = 0; i < amd_iommus_present; ++i) {
1063 if (!domain->dev_iommu[i])
1067 * Devices of this domain are behind this IOMMU
1068 * We need to wait for completion of all commands.
1070 iommu_completion_wait(amd_iommus[i]);
1076 * This function flushes the DTEs for all devices in domain
1078 static void domain_flush_devices(struct protection_domain *domain)
1080 struct iommu_dev_data *dev_data;
1082 list_for_each_entry(dev_data, &domain->dev_list, list)
1083 device_flush_dte(dev_data);
1086 /****************************************************************************
1088 * The functions below are used the create the page table mappings for
1089 * unity mapped regions.
1091 ****************************************************************************/
1094 * This function is used to add another level to an IO page table. Adding
1095 * another level increases the size of the address space by 9 bits to a size up
1098 static bool increase_address_space(struct protection_domain *domain,
1103 if (domain->mode == PAGE_MODE_6_LEVEL)
1104 /* address space already 64 bit large */
1107 pte = (void *)get_zeroed_page(gfp);
1111 *pte = PM_LEVEL_PDE(domain->mode,
1112 virt_to_phys(domain->pt_root));
1113 domain->pt_root = pte;
1115 domain->updated = true;
1120 static u64 *alloc_pte(struct protection_domain *domain,
1121 unsigned long address,
1122 unsigned long page_size,
1129 BUG_ON(!is_power_of_2(page_size));
1131 while (address > PM_LEVEL_SIZE(domain->mode))
1132 increase_address_space(domain, gfp);
1134 level = domain->mode - 1;
1135 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1136 address = PAGE_SIZE_ALIGN(address, page_size);
1137 end_lvl = PAGE_SIZE_LEVEL(page_size);
1139 while (level > end_lvl) {
1140 if (!IOMMU_PTE_PRESENT(*pte)) {
1141 page = (u64 *)get_zeroed_page(gfp);
1144 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1147 /* No level skipping support yet */
1148 if (PM_PTE_LEVEL(*pte) != level)
1153 pte = IOMMU_PTE_PAGE(*pte);
1155 if (pte_page && level == end_lvl)
1158 pte = &pte[PM_LEVEL_INDEX(level, address)];
1165 * This function checks if there is a PTE for a given dma address. If
1166 * there is one, it returns the pointer to it.
1168 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1173 if (address > PM_LEVEL_SIZE(domain->mode))
1176 level = domain->mode - 1;
1177 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1182 if (!IOMMU_PTE_PRESENT(*pte))
1186 if (PM_PTE_LEVEL(*pte) == 0x07) {
1187 unsigned long pte_mask, __pte;
1190 * If we have a series of large PTEs, make
1191 * sure to return a pointer to the first one.
1193 pte_mask = PTE_PAGE_SIZE(*pte);
1194 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1195 __pte = ((unsigned long)pte) & pte_mask;
1197 return (u64 *)__pte;
1200 /* No level skipping support yet */
1201 if (PM_PTE_LEVEL(*pte) != level)
1206 /* Walk to the next level */
1207 pte = IOMMU_PTE_PAGE(*pte);
1208 pte = &pte[PM_LEVEL_INDEX(level, address)];
1215 * Generic mapping functions. It maps a physical address into a DMA
1216 * address space. It allocates the page table pages if necessary.
1217 * In the future it can be extended to a generic mapping function
1218 * supporting all features of AMD IOMMU page tables like level skipping
1219 * and full 64 bit address spaces.
1221 static int iommu_map_page(struct protection_domain *dom,
1222 unsigned long bus_addr,
1223 unsigned long phys_addr,
1225 unsigned long page_size)
1230 if (!(prot & IOMMU_PROT_MASK))
1233 bus_addr = PAGE_ALIGN(bus_addr);
1234 phys_addr = PAGE_ALIGN(phys_addr);
1235 count = PAGE_SIZE_PTE_COUNT(page_size);
1236 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1238 for (i = 0; i < count; ++i)
1239 if (IOMMU_PTE_PRESENT(pte[i]))
1242 if (page_size > PAGE_SIZE) {
1243 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1244 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1246 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1248 if (prot & IOMMU_PROT_IR)
1249 __pte |= IOMMU_PTE_IR;
1250 if (prot & IOMMU_PROT_IW)
1251 __pte |= IOMMU_PTE_IW;
1253 for (i = 0; i < count; ++i)
1261 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1262 unsigned long bus_addr,
1263 unsigned long page_size)
1265 unsigned long long unmap_size, unmapped;
1268 BUG_ON(!is_power_of_2(page_size));
1272 while (unmapped < page_size) {
1274 pte = fetch_pte(dom, bus_addr);
1278 * No PTE for this address
1279 * move forward in 4kb steps
1281 unmap_size = PAGE_SIZE;
1282 } else if (PM_PTE_LEVEL(*pte) == 0) {
1283 /* 4kb PTE found for this address */
1284 unmap_size = PAGE_SIZE;
1289 /* Large PTE found which maps this address */
1290 unmap_size = PTE_PAGE_SIZE(*pte);
1291 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1292 for (i = 0; i < count; i++)
1296 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1297 unmapped += unmap_size;
1300 BUG_ON(!is_power_of_2(unmapped));
1306 * This function checks if a specific unity mapping entry is needed for
1307 * this specific IOMMU.
1309 static int iommu_for_unity_map(struct amd_iommu *iommu,
1310 struct unity_map_entry *entry)
1314 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1315 bdf = amd_iommu_alias_table[i];
1316 if (amd_iommu_rlookup_table[bdf] == iommu)
1324 * This function actually applies the mapping to the page table of the
1327 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1328 struct unity_map_entry *e)
1333 for (addr = e->address_start; addr < e->address_end;
1334 addr += PAGE_SIZE) {
1335 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1340 * if unity mapping is in aperture range mark the page
1341 * as allocated in the aperture
1343 if (addr < dma_dom->aperture_size)
1344 __set_bit(addr >> PAGE_SHIFT,
1345 dma_dom->aperture[0]->bitmap);
1352 * Init the unity mappings for a specific IOMMU in the system
1354 * Basically iterates over all unity mapping entries and applies them to
1355 * the default domain DMA of that IOMMU if necessary.
1357 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1359 struct unity_map_entry *entry;
1362 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1363 if (!iommu_for_unity_map(iommu, entry))
1365 ret = dma_ops_unity_map(iommu->default_dom, entry);
1374 * Inits the unity mappings required for a specific device
1376 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1379 struct unity_map_entry *e;
1382 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1383 if (!(devid >= e->devid_start && devid <= e->devid_end))
1385 ret = dma_ops_unity_map(dma_dom, e);
1393 /****************************************************************************
1395 * The next functions belong to the address allocator for the dma_ops
1396 * interface functions. They work like the allocators in the other IOMMU
1397 * drivers. Its basically a bitmap which marks the allocated pages in
1398 * the aperture. Maybe it could be enhanced in the future to a more
1399 * efficient allocator.
1401 ****************************************************************************/
1404 * The address allocator core functions.
1406 * called with domain->lock held
1410 * Used to reserve address ranges in the aperture (e.g. for exclusion
1413 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1414 unsigned long start_page,
1417 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1419 if (start_page + pages > last_page)
1420 pages = last_page - start_page;
1422 for (i = start_page; i < start_page + pages; ++i) {
1423 int index = i / APERTURE_RANGE_PAGES;
1424 int page = i % APERTURE_RANGE_PAGES;
1425 __set_bit(page, dom->aperture[index]->bitmap);
1430 * This function is used to add a new aperture range to an existing
1431 * aperture in case of dma_ops domain allocation or address allocation
1434 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1435 bool populate, gfp_t gfp)
1437 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1438 struct amd_iommu *iommu;
1439 unsigned long i, old_size;
1441 #ifdef CONFIG_IOMMU_STRESS
1445 if (index >= APERTURE_MAX_RANGES)
1448 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1449 if (!dma_dom->aperture[index])
1452 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1453 if (!dma_dom->aperture[index]->bitmap)
1456 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1459 unsigned long address = dma_dom->aperture_size;
1460 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1461 u64 *pte, *pte_page;
1463 for (i = 0; i < num_ptes; ++i) {
1464 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1469 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1471 address += APERTURE_RANGE_SIZE / 64;
1475 old_size = dma_dom->aperture_size;
1476 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1478 /* Reserve address range used for MSI messages */
1479 if (old_size < MSI_ADDR_BASE_LO &&
1480 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1481 unsigned long spage;
1484 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1485 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1487 dma_ops_reserve_addresses(dma_dom, spage, pages);
1490 /* Initialize the exclusion range if necessary */
1491 for_each_iommu(iommu) {
1492 if (iommu->exclusion_start &&
1493 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1494 && iommu->exclusion_start < dma_dom->aperture_size) {
1495 unsigned long startpage;
1496 int pages = iommu_num_pages(iommu->exclusion_start,
1497 iommu->exclusion_length,
1499 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1500 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1505 * Check for areas already mapped as present in the new aperture
1506 * range and mark those pages as reserved in the allocator. Such
1507 * mappings may already exist as a result of requested unity
1508 * mappings for devices.
1510 for (i = dma_dom->aperture[index]->offset;
1511 i < dma_dom->aperture_size;
1513 u64 *pte = fetch_pte(&dma_dom->domain, i);
1514 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1517 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1520 update_domain(&dma_dom->domain);
1525 update_domain(&dma_dom->domain);
1527 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1529 kfree(dma_dom->aperture[index]);
1530 dma_dom->aperture[index] = NULL;
1535 static unsigned long dma_ops_area_alloc(struct device *dev,
1536 struct dma_ops_domain *dom,
1538 unsigned long align_mask,
1540 unsigned long start)
1542 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1543 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1544 int i = start >> APERTURE_RANGE_SHIFT;
1545 unsigned long boundary_size;
1546 unsigned long address = -1;
1547 unsigned long limit;
1549 next_bit >>= PAGE_SHIFT;
1551 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1552 PAGE_SIZE) >> PAGE_SHIFT;
1554 for (;i < max_index; ++i) {
1555 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1557 if (dom->aperture[i]->offset >= dma_mask)
1560 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1561 dma_mask >> PAGE_SHIFT);
1563 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1564 limit, next_bit, pages, 0,
1565 boundary_size, align_mask);
1566 if (address != -1) {
1567 address = dom->aperture[i]->offset +
1568 (address << PAGE_SHIFT);
1569 dom->next_address = address + (pages << PAGE_SHIFT);
1579 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1580 struct dma_ops_domain *dom,
1582 unsigned long align_mask,
1585 unsigned long address;
1587 #ifdef CONFIG_IOMMU_STRESS
1588 dom->next_address = 0;
1589 dom->need_flush = true;
1592 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1593 dma_mask, dom->next_address);
1595 if (address == -1) {
1596 dom->next_address = 0;
1597 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1599 dom->need_flush = true;
1602 if (unlikely(address == -1))
1603 address = DMA_ERROR_CODE;
1605 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1611 * The address free function.
1613 * called with domain->lock held
1615 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1616 unsigned long address,
1619 unsigned i = address >> APERTURE_RANGE_SHIFT;
1620 struct aperture_range *range = dom->aperture[i];
1622 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1624 #ifdef CONFIG_IOMMU_STRESS
1629 if (address >= dom->next_address)
1630 dom->need_flush = true;
1632 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1634 bitmap_clear(range->bitmap, address, pages);
1638 /****************************************************************************
1640 * The next functions belong to the domain allocation. A domain is
1641 * allocated for every IOMMU as the default domain. If device isolation
1642 * is enabled, every device get its own domain. The most important thing
1643 * about domains is the page table mapping the DMA address space they
1646 ****************************************************************************/
1649 * This function adds a protection domain to the global protection domain list
1651 static void add_domain_to_list(struct protection_domain *domain)
1653 unsigned long flags;
1655 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1656 list_add(&domain->list, &amd_iommu_pd_list);
1657 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1661 * This function removes a protection domain to the global
1662 * protection domain list
1664 static void del_domain_from_list(struct protection_domain *domain)
1666 unsigned long flags;
1668 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1669 list_del(&domain->list);
1670 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1673 static u16 domain_id_alloc(void)
1675 unsigned long flags;
1678 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1679 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1681 if (id > 0 && id < MAX_DOMAIN_ID)
1682 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1685 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1690 static void domain_id_free(int id)
1692 unsigned long flags;
1694 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1695 if (id > 0 && id < MAX_DOMAIN_ID)
1696 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1697 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1700 static void free_pagetable(struct protection_domain *domain)
1705 p1 = domain->pt_root;
1710 for (i = 0; i < 512; ++i) {
1711 if (!IOMMU_PTE_PRESENT(p1[i]))
1714 p2 = IOMMU_PTE_PAGE(p1[i]);
1715 for (j = 0; j < 512; ++j) {
1716 if (!IOMMU_PTE_PRESENT(p2[j]))
1718 p3 = IOMMU_PTE_PAGE(p2[j]);
1719 free_page((unsigned long)p3);
1722 free_page((unsigned long)p2);
1725 free_page((unsigned long)p1);
1727 domain->pt_root = NULL;
1730 static void free_gcr3_tbl_level1(u64 *tbl)
1735 for (i = 0; i < 512; ++i) {
1736 if (!(tbl[i] & GCR3_VALID))
1739 ptr = __va(tbl[i] & PAGE_MASK);
1741 free_page((unsigned long)ptr);
1745 static void free_gcr3_tbl_level2(u64 *tbl)
1750 for (i = 0; i < 512; ++i) {
1751 if (!(tbl[i] & GCR3_VALID))
1754 ptr = __va(tbl[i] & PAGE_MASK);
1756 free_gcr3_tbl_level1(ptr);
1760 static void free_gcr3_table(struct protection_domain *domain)
1762 if (domain->glx == 2)
1763 free_gcr3_tbl_level2(domain->gcr3_tbl);
1764 else if (domain->glx == 1)
1765 free_gcr3_tbl_level1(domain->gcr3_tbl);
1766 else if (domain->glx != 0)
1769 free_page((unsigned long)domain->gcr3_tbl);
1773 * Free a domain, only used if something went wrong in the
1774 * allocation path and we need to free an already allocated page table
1776 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1783 del_domain_from_list(&dom->domain);
1785 free_pagetable(&dom->domain);
1787 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1788 if (!dom->aperture[i])
1790 free_page((unsigned long)dom->aperture[i]->bitmap);
1791 kfree(dom->aperture[i]);
1798 * Allocates a new protection domain usable for the dma_ops functions.
1799 * It also initializes the page table and the address allocator data
1800 * structures required for the dma_ops interface
1802 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1804 struct dma_ops_domain *dma_dom;
1806 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1810 spin_lock_init(&dma_dom->domain.lock);
1812 dma_dom->domain.id = domain_id_alloc();
1813 if (dma_dom->domain.id == 0)
1815 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1816 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1817 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1818 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1819 dma_dom->domain.priv = dma_dom;
1820 if (!dma_dom->domain.pt_root)
1823 dma_dom->need_flush = false;
1824 dma_dom->target_dev = 0xffff;
1826 add_domain_to_list(&dma_dom->domain);
1828 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1832 * mark the first page as allocated so we never return 0 as
1833 * a valid dma-address. So we can use 0 as error value
1835 dma_dom->aperture[0]->bitmap[0] = 1;
1836 dma_dom->next_address = 0;
1842 dma_ops_domain_free(dma_dom);
1848 * little helper function to check whether a given protection domain is a
1851 static bool dma_ops_domain(struct protection_domain *domain)
1853 return domain->flags & PD_DMA_OPS_MASK;
1856 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1861 if (domain->mode != PAGE_MODE_NONE)
1862 pte_root = virt_to_phys(domain->pt_root);
1864 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1865 << DEV_ENTRY_MODE_SHIFT;
1866 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1868 flags = amd_iommu_dev_table[devid].data[1];
1871 flags |= DTE_FLAG_IOTLB;
1873 if (domain->flags & PD_IOMMUV2_MASK) {
1874 u64 gcr3 = __pa(domain->gcr3_tbl);
1875 u64 glx = domain->glx;
1878 pte_root |= DTE_FLAG_GV;
1879 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1881 /* First mask out possible old values for GCR3 table */
1882 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1885 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1888 /* Encode GCR3 table into DTE */
1889 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1892 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1895 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1899 flags &= ~(0xffffUL);
1900 flags |= domain->id;
1902 amd_iommu_dev_table[devid].data[1] = flags;
1903 amd_iommu_dev_table[devid].data[0] = pte_root;
1906 static void clear_dte_entry(u16 devid)
1908 /* remove entry from the device table seen by the hardware */
1909 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1910 amd_iommu_dev_table[devid].data[1] = 0;
1912 amd_iommu_apply_erratum_63(devid);
1915 static void do_attach(struct iommu_dev_data *dev_data,
1916 struct protection_domain *domain)
1918 struct amd_iommu *iommu;
1921 iommu = amd_iommu_rlookup_table[dev_data->devid];
1922 ats = dev_data->ats.enabled;
1924 /* Update data structures */
1925 dev_data->domain = domain;
1926 list_add(&dev_data->list, &domain->dev_list);
1927 set_dte_entry(dev_data->devid, domain, ats);
1929 /* Do reference counting */
1930 domain->dev_iommu[iommu->index] += 1;
1931 domain->dev_cnt += 1;
1933 /* Flush the DTE entry */
1934 device_flush_dte(dev_data);
1937 static void do_detach(struct iommu_dev_data *dev_data)
1939 struct amd_iommu *iommu;
1941 iommu = amd_iommu_rlookup_table[dev_data->devid];
1943 /* decrease reference counters */
1944 dev_data->domain->dev_iommu[iommu->index] -= 1;
1945 dev_data->domain->dev_cnt -= 1;
1947 /* Update data structures */
1948 dev_data->domain = NULL;
1949 list_del(&dev_data->list);
1950 clear_dte_entry(dev_data->devid);
1952 /* Flush the DTE entry */
1953 device_flush_dte(dev_data);
1957 * If a device is not yet associated with a domain, this function does
1958 * assigns it visible for the hardware
1960 static int __attach_device(struct iommu_dev_data *dev_data,
1961 struct protection_domain *domain)
1966 spin_lock(&domain->lock);
1968 if (dev_data->alias_data != NULL) {
1969 struct iommu_dev_data *alias_data = dev_data->alias_data;
1971 /* Some sanity checks */
1973 if (alias_data->domain != NULL &&
1974 alias_data->domain != domain)
1977 if (dev_data->domain != NULL &&
1978 dev_data->domain != domain)
1981 /* Do real assignment */
1982 if (alias_data->domain == NULL)
1983 do_attach(alias_data, domain);
1985 atomic_inc(&alias_data->bind);
1988 if (dev_data->domain == NULL)
1989 do_attach(dev_data, domain);
1991 atomic_inc(&dev_data->bind);
1998 spin_unlock(&domain->lock);
2004 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2006 pci_disable_ats(pdev);
2007 pci_disable_pri(pdev);
2008 pci_disable_pasid(pdev);
2011 /* FIXME: Change generic reset-function to do the same */
2012 static int pri_reset_while_enabled(struct pci_dev *pdev)
2017 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2021 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2022 control |= PCI_PRI_CTRL_RESET;
2023 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2028 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2033 /* FIXME: Hardcode number of outstanding requests for now */
2035 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2037 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2039 /* Only allow access to user-accessible pages */
2040 ret = pci_enable_pasid(pdev, 0);
2044 /* First reset the PRI state of the device */
2045 ret = pci_reset_pri(pdev);
2050 ret = pci_enable_pri(pdev, reqs);
2055 ret = pri_reset_while_enabled(pdev);
2060 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2067 pci_disable_pri(pdev);
2068 pci_disable_pasid(pdev);
2073 /* FIXME: Move this to PCI code */
2074 #define PCI_PRI_TLP_OFF (1 << 15)
2076 bool pci_pri_tlp_required(struct pci_dev *pdev)
2081 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2085 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2087 return (status & PCI_PRI_TLP_OFF) ? true : false;
2091 * If a device is not yet associated with a domain, this function does
2092 * assigns it visible for the hardware
2094 static int attach_device(struct device *dev,
2095 struct protection_domain *domain)
2097 struct pci_dev *pdev = to_pci_dev(dev);
2098 struct iommu_dev_data *dev_data;
2099 unsigned long flags;
2102 dev_data = get_dev_data(dev);
2104 if (domain->flags & PD_IOMMUV2_MASK) {
2105 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2108 if (pdev_iommuv2_enable(pdev) != 0)
2111 dev_data->ats.enabled = true;
2112 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2113 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2114 } else if (amd_iommu_iotlb_sup &&
2115 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2116 dev_data->ats.enabled = true;
2117 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2120 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2121 ret = __attach_device(dev_data, domain);
2122 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2125 * We might boot into a crash-kernel here. The crashed kernel
2126 * left the caches in the IOMMU dirty. So we have to flush
2127 * here to evict all dirty stuff.
2129 domain_flush_tlb_pde(domain);
2135 * Removes a device from a protection domain (unlocked)
2137 static void __detach_device(struct iommu_dev_data *dev_data)
2139 struct protection_domain *domain;
2140 unsigned long flags;
2142 BUG_ON(!dev_data->domain);
2144 domain = dev_data->domain;
2146 spin_lock_irqsave(&domain->lock, flags);
2148 if (dev_data->alias_data != NULL) {
2149 struct iommu_dev_data *alias_data = dev_data->alias_data;
2151 if (atomic_dec_and_test(&alias_data->bind))
2152 do_detach(alias_data);
2155 if (atomic_dec_and_test(&dev_data->bind))
2156 do_detach(dev_data);
2158 spin_unlock_irqrestore(&domain->lock, flags);
2161 * If we run in passthrough mode the device must be assigned to the
2162 * passthrough domain if it is detached from any other domain.
2163 * Make sure we can deassign from the pt_domain itself.
2165 if (dev_data->passthrough &&
2166 (dev_data->domain == NULL && domain != pt_domain))
2167 __attach_device(dev_data, pt_domain);
2171 * Removes a device from a protection domain (with devtable_lock held)
2173 static void detach_device(struct device *dev)
2175 struct protection_domain *domain;
2176 struct iommu_dev_data *dev_data;
2177 unsigned long flags;
2179 dev_data = get_dev_data(dev);
2180 domain = dev_data->domain;
2182 /* lock device table */
2183 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2184 __detach_device(dev_data);
2185 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2187 if (domain->flags & PD_IOMMUV2_MASK)
2188 pdev_iommuv2_disable(to_pci_dev(dev));
2189 else if (dev_data->ats.enabled)
2190 pci_disable_ats(to_pci_dev(dev));
2192 dev_data->ats.enabled = false;
2196 * Find out the protection domain structure for a given PCI device. This
2197 * will give us the pointer to the page table root for example.
2199 static struct protection_domain *domain_for_device(struct device *dev)
2201 struct iommu_dev_data *dev_data;
2202 struct protection_domain *dom = NULL;
2203 unsigned long flags;
2205 dev_data = get_dev_data(dev);
2207 if (dev_data->domain)
2208 return dev_data->domain;
2210 if (dev_data->alias_data != NULL) {
2211 struct iommu_dev_data *alias_data = dev_data->alias_data;
2213 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2214 if (alias_data->domain != NULL) {
2215 __attach_device(dev_data, alias_data->domain);
2216 dom = alias_data->domain;
2218 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2224 static int device_change_notifier(struct notifier_block *nb,
2225 unsigned long action, void *data)
2227 struct dma_ops_domain *dma_domain;
2228 struct protection_domain *domain;
2229 struct iommu_dev_data *dev_data;
2230 struct device *dev = data;
2231 struct amd_iommu *iommu;
2232 unsigned long flags;
2235 if (!check_device(dev))
2238 devid = get_device_id(dev);
2239 iommu = amd_iommu_rlookup_table[devid];
2240 dev_data = get_dev_data(dev);
2243 case BUS_NOTIFY_UNBOUND_DRIVER:
2245 domain = domain_for_device(dev);
2249 if (dev_data->passthrough)
2253 case BUS_NOTIFY_ADD_DEVICE:
2255 iommu_init_device(dev);
2257 domain = domain_for_device(dev);
2259 /* allocate a protection domain if a device is added */
2260 dma_domain = find_protection_domain(devid);
2263 dma_domain = dma_ops_domain_alloc();
2266 dma_domain->target_dev = devid;
2268 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2269 list_add_tail(&dma_domain->list, &iommu_pd_list);
2270 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2272 dev_data = get_dev_data(dev);
2274 if (!dev_data->passthrough)
2275 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2277 dev->archdata.dma_ops = &nommu_dma_ops;
2280 case BUS_NOTIFY_DEL_DEVICE:
2282 iommu_uninit_device(dev);
2288 iommu_completion_wait(iommu);
2294 static struct notifier_block device_nb = {
2295 .notifier_call = device_change_notifier,
2298 void amd_iommu_init_notifier(void)
2300 bus_register_notifier(&pci_bus_type, &device_nb);
2303 /*****************************************************************************
2305 * The next functions belong to the dma_ops mapping/unmapping code.
2307 *****************************************************************************/
2310 * In the dma_ops path we only have the struct device. This function
2311 * finds the corresponding IOMMU, the protection domain and the
2312 * requestor id for a given device.
2313 * If the device is not yet associated with a domain this is also done
2316 static struct protection_domain *get_domain(struct device *dev)
2318 struct protection_domain *domain;
2319 struct dma_ops_domain *dma_dom;
2320 u16 devid = get_device_id(dev);
2322 if (!check_device(dev))
2323 return ERR_PTR(-EINVAL);
2325 domain = domain_for_device(dev);
2326 if (domain != NULL && !dma_ops_domain(domain))
2327 return ERR_PTR(-EBUSY);
2332 /* Device not bount yet - bind it */
2333 dma_dom = find_protection_domain(devid);
2335 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2336 attach_device(dev, &dma_dom->domain);
2337 DUMP_printk("Using protection domain %d for device %s\n",
2338 dma_dom->domain.id, dev_name(dev));
2340 return &dma_dom->domain;
2343 static void update_device_table(struct protection_domain *domain)
2345 struct iommu_dev_data *dev_data;
2347 list_for_each_entry(dev_data, &domain->dev_list, list)
2348 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2351 static void update_domain(struct protection_domain *domain)
2353 if (!domain->updated)
2356 update_device_table(domain);
2358 domain_flush_devices(domain);
2359 domain_flush_tlb_pde(domain);
2361 domain->updated = false;
2365 * This function fetches the PTE for a given address in the aperture
2367 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2368 unsigned long address)
2370 struct aperture_range *aperture;
2371 u64 *pte, *pte_page;
2373 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2377 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2379 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2381 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2383 pte += PM_LEVEL_INDEX(0, address);
2385 update_domain(&dom->domain);
2391 * This is the generic map function. It maps one 4kb page at paddr to
2392 * the given address in the DMA address space for the domain.
2394 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2395 unsigned long address,
2401 WARN_ON(address > dom->aperture_size);
2405 pte = dma_ops_get_pte(dom, address);
2407 return DMA_ERROR_CODE;
2409 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2411 if (direction == DMA_TO_DEVICE)
2412 __pte |= IOMMU_PTE_IR;
2413 else if (direction == DMA_FROM_DEVICE)
2414 __pte |= IOMMU_PTE_IW;
2415 else if (direction == DMA_BIDIRECTIONAL)
2416 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2422 return (dma_addr_t)address;
2426 * The generic unmapping function for on page in the DMA address space.
2428 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2429 unsigned long address)
2431 struct aperture_range *aperture;
2434 if (address >= dom->aperture_size)
2437 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2441 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2445 pte += PM_LEVEL_INDEX(0, address);
2453 * This function contains common code for mapping of a physically
2454 * contiguous memory region into DMA address space. It is used by all
2455 * mapping functions provided with this IOMMU driver.
2456 * Must be called with the domain lock held.
2458 static dma_addr_t __map_single(struct device *dev,
2459 struct dma_ops_domain *dma_dom,
2466 dma_addr_t offset = paddr & ~PAGE_MASK;
2467 dma_addr_t address, start, ret;
2469 unsigned long align_mask = 0;
2472 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2475 INC_STATS_COUNTER(total_map_requests);
2478 INC_STATS_COUNTER(cross_page);
2481 align_mask = (1UL << get_order(size)) - 1;
2484 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2486 if (unlikely(address == DMA_ERROR_CODE)) {
2488 * setting next_address here will let the address
2489 * allocator only scan the new allocated range in the
2490 * first run. This is a small optimization.
2492 dma_dom->next_address = dma_dom->aperture_size;
2494 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2498 * aperture was successfully enlarged by 128 MB, try
2505 for (i = 0; i < pages; ++i) {
2506 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2507 if (ret == DMA_ERROR_CODE)
2515 ADD_STATS_COUNTER(alloced_io_mem, size);
2517 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2518 domain_flush_tlb(&dma_dom->domain);
2519 dma_dom->need_flush = false;
2520 } else if (unlikely(amd_iommu_np_cache))
2521 domain_flush_pages(&dma_dom->domain, address, size);
2528 for (--i; i >= 0; --i) {
2530 dma_ops_domain_unmap(dma_dom, start);
2533 dma_ops_free_addresses(dma_dom, address, pages);
2535 return DMA_ERROR_CODE;
2539 * Does the reverse of the __map_single function. Must be called with
2540 * the domain lock held too
2542 static void __unmap_single(struct dma_ops_domain *dma_dom,
2543 dma_addr_t dma_addr,
2547 dma_addr_t flush_addr;
2548 dma_addr_t i, start;
2551 if ((dma_addr == DMA_ERROR_CODE) ||
2552 (dma_addr + size > dma_dom->aperture_size))
2555 flush_addr = dma_addr;
2556 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2557 dma_addr &= PAGE_MASK;
2560 for (i = 0; i < pages; ++i) {
2561 dma_ops_domain_unmap(dma_dom, start);
2565 SUB_STATS_COUNTER(alloced_io_mem, size);
2567 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2569 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2570 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2571 dma_dom->need_flush = false;
2576 * The exported map_single function for dma_ops.
2578 static dma_addr_t map_page(struct device *dev, struct page *page,
2579 unsigned long offset, size_t size,
2580 enum dma_data_direction dir,
2581 struct dma_attrs *attrs)
2583 unsigned long flags;
2584 struct protection_domain *domain;
2587 phys_addr_t paddr = page_to_phys(page) + offset;
2589 INC_STATS_COUNTER(cnt_map_single);
2591 domain = get_domain(dev);
2592 if (PTR_ERR(domain) == -EINVAL)
2593 return (dma_addr_t)paddr;
2594 else if (IS_ERR(domain))
2595 return DMA_ERROR_CODE;
2597 dma_mask = *dev->dma_mask;
2599 spin_lock_irqsave(&domain->lock, flags);
2601 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2603 if (addr == DMA_ERROR_CODE)
2606 domain_flush_complete(domain);
2609 spin_unlock_irqrestore(&domain->lock, flags);
2615 * The exported unmap_single function for dma_ops.
2617 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2618 enum dma_data_direction dir, struct dma_attrs *attrs)
2620 unsigned long flags;
2621 struct protection_domain *domain;
2623 INC_STATS_COUNTER(cnt_unmap_single);
2625 domain = get_domain(dev);
2629 spin_lock_irqsave(&domain->lock, flags);
2631 __unmap_single(domain->priv, dma_addr, size, dir);
2633 domain_flush_complete(domain);
2635 spin_unlock_irqrestore(&domain->lock, flags);
2639 * This is a special map_sg function which is used if we should map a
2640 * device which is not handled by an AMD IOMMU in the system.
2642 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2643 int nelems, int dir)
2645 struct scatterlist *s;
2648 for_each_sg(sglist, s, nelems, i) {
2649 s->dma_address = (dma_addr_t)sg_phys(s);
2650 s->dma_length = s->length;
2657 * The exported map_sg function for dma_ops (handles scatter-gather
2660 static int map_sg(struct device *dev, struct scatterlist *sglist,
2661 int nelems, enum dma_data_direction dir,
2662 struct dma_attrs *attrs)
2664 unsigned long flags;
2665 struct protection_domain *domain;
2667 struct scatterlist *s;
2669 int mapped_elems = 0;
2672 INC_STATS_COUNTER(cnt_map_sg);
2674 domain = get_domain(dev);
2675 if (PTR_ERR(domain) == -EINVAL)
2676 return map_sg_no_iommu(dev, sglist, nelems, dir);
2677 else if (IS_ERR(domain))
2680 dma_mask = *dev->dma_mask;
2682 spin_lock_irqsave(&domain->lock, flags);
2684 for_each_sg(sglist, s, nelems, i) {
2687 s->dma_address = __map_single(dev, domain->priv,
2688 paddr, s->length, dir, false,
2691 if (s->dma_address) {
2692 s->dma_length = s->length;
2698 domain_flush_complete(domain);
2701 spin_unlock_irqrestore(&domain->lock, flags);
2703 return mapped_elems;
2705 for_each_sg(sglist, s, mapped_elems, i) {
2707 __unmap_single(domain->priv, s->dma_address,
2708 s->dma_length, dir);
2709 s->dma_address = s->dma_length = 0;
2718 * The exported map_sg function for dma_ops (handles scatter-gather
2721 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2722 int nelems, enum dma_data_direction dir,
2723 struct dma_attrs *attrs)
2725 unsigned long flags;
2726 struct protection_domain *domain;
2727 struct scatterlist *s;
2730 INC_STATS_COUNTER(cnt_unmap_sg);
2732 domain = get_domain(dev);
2736 spin_lock_irqsave(&domain->lock, flags);
2738 for_each_sg(sglist, s, nelems, i) {
2739 __unmap_single(domain->priv, s->dma_address,
2740 s->dma_length, dir);
2741 s->dma_address = s->dma_length = 0;
2744 domain_flush_complete(domain);
2746 spin_unlock_irqrestore(&domain->lock, flags);
2750 * The exported alloc_coherent function for dma_ops.
2752 static void *alloc_coherent(struct device *dev, size_t size,
2753 dma_addr_t *dma_addr, gfp_t flag,
2754 struct dma_attrs *attrs)
2756 unsigned long flags;
2758 struct protection_domain *domain;
2760 u64 dma_mask = dev->coherent_dma_mask;
2762 INC_STATS_COUNTER(cnt_alloc_coherent);
2764 domain = get_domain(dev);
2765 if (PTR_ERR(domain) == -EINVAL) {
2766 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2767 *dma_addr = __pa(virt_addr);
2769 } else if (IS_ERR(domain))
2772 dma_mask = dev->coherent_dma_mask;
2773 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2776 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2780 paddr = virt_to_phys(virt_addr);
2783 dma_mask = *dev->dma_mask;
2785 spin_lock_irqsave(&domain->lock, flags);
2787 *dma_addr = __map_single(dev, domain->priv, paddr,
2788 size, DMA_BIDIRECTIONAL, true, dma_mask);
2790 if (*dma_addr == DMA_ERROR_CODE) {
2791 spin_unlock_irqrestore(&domain->lock, flags);
2795 domain_flush_complete(domain);
2797 spin_unlock_irqrestore(&domain->lock, flags);
2803 free_pages((unsigned long)virt_addr, get_order(size));
2809 * The exported free_coherent function for dma_ops.
2811 static void free_coherent(struct device *dev, size_t size,
2812 void *virt_addr, dma_addr_t dma_addr,
2813 struct dma_attrs *attrs)
2815 unsigned long flags;
2816 struct protection_domain *domain;
2818 INC_STATS_COUNTER(cnt_free_coherent);
2820 domain = get_domain(dev);
2824 spin_lock_irqsave(&domain->lock, flags);
2826 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2828 domain_flush_complete(domain);
2830 spin_unlock_irqrestore(&domain->lock, flags);
2833 free_pages((unsigned long)virt_addr, get_order(size));
2837 * This function is called by the DMA layer to find out if we can handle a
2838 * particular device. It is part of the dma_ops.
2840 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2842 return check_device(dev);
2846 * The function for pre-allocating protection domains.
2848 * If the driver core informs the DMA layer if a driver grabs a device
2849 * we don't need to preallocate the protection domains anymore.
2850 * For now we have to.
2852 static void __init prealloc_protection_domains(void)
2854 struct iommu_dev_data *dev_data;
2855 struct dma_ops_domain *dma_dom;
2856 struct pci_dev *dev = NULL;
2859 for_each_pci_dev(dev) {
2861 /* Do we handle this device? */
2862 if (!check_device(&dev->dev))
2865 dev_data = get_dev_data(&dev->dev);
2866 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2867 /* Make sure passthrough domain is allocated */
2868 alloc_passthrough_domain();
2869 dev_data->passthrough = true;
2870 attach_device(&dev->dev, pt_domain);
2871 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2872 dev_name(&dev->dev));
2875 /* Is there already any domain for it? */
2876 if (domain_for_device(&dev->dev))
2879 devid = get_device_id(&dev->dev);
2881 dma_dom = dma_ops_domain_alloc();
2884 init_unity_mappings_for_device(dma_dom, devid);
2885 dma_dom->target_dev = devid;
2887 attach_device(&dev->dev, &dma_dom->domain);
2889 list_add_tail(&dma_dom->list, &iommu_pd_list);
2893 static struct dma_map_ops amd_iommu_dma_ops = {
2894 .alloc = alloc_coherent,
2895 .free = free_coherent,
2896 .map_page = map_page,
2897 .unmap_page = unmap_page,
2899 .unmap_sg = unmap_sg,
2900 .dma_supported = amd_iommu_dma_supported,
2903 static unsigned device_dma_ops_init(void)
2905 struct iommu_dev_data *dev_data;
2906 struct pci_dev *pdev = NULL;
2907 unsigned unhandled = 0;
2909 for_each_pci_dev(pdev) {
2910 if (!check_device(&pdev->dev)) {
2912 iommu_ignore_device(&pdev->dev);
2918 dev_data = get_dev_data(&pdev->dev);
2920 if (!dev_data->passthrough)
2921 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2923 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
2930 * The function which clues the AMD IOMMU driver into dma_ops.
2933 void __init amd_iommu_init_api(void)
2935 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2938 int __init amd_iommu_init_dma_ops(void)
2940 struct amd_iommu *iommu;
2944 * first allocate a default protection domain for every IOMMU we
2945 * found in the system. Devices not assigned to any other
2946 * protection domain will be assigned to the default one.
2948 for_each_iommu(iommu) {
2949 iommu->default_dom = dma_ops_domain_alloc();
2950 if (iommu->default_dom == NULL)
2952 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2953 ret = iommu_init_unity_mappings(iommu);
2959 * Pre-allocate the protection domains for each device.
2961 prealloc_protection_domains();
2966 /* Make the driver finally visible to the drivers */
2967 unhandled = device_dma_ops_init();
2968 if (unhandled && max_pfn > MAX_DMA32_PFN) {
2969 /* There are unhandled devices - initialize swiotlb for them */
2973 amd_iommu_stats_init();
2979 for_each_iommu(iommu) {
2980 if (iommu->default_dom)
2981 dma_ops_domain_free(iommu->default_dom);
2987 /*****************************************************************************
2989 * The following functions belong to the exported interface of AMD IOMMU
2991 * This interface allows access to lower level functions of the IOMMU
2992 * like protection domain handling and assignement of devices to domains
2993 * which is not possible with the dma_ops interface.
2995 *****************************************************************************/
2997 static void cleanup_domain(struct protection_domain *domain)
2999 struct iommu_dev_data *dev_data, *next;
3000 unsigned long flags;
3002 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3004 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3005 __detach_device(dev_data);
3006 atomic_set(&dev_data->bind, 0);
3009 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3012 static void protection_domain_free(struct protection_domain *domain)
3017 del_domain_from_list(domain);
3020 domain_id_free(domain->id);
3025 static struct protection_domain *protection_domain_alloc(void)
3027 struct protection_domain *domain;
3029 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3033 spin_lock_init(&domain->lock);
3034 mutex_init(&domain->api_lock);
3035 domain->id = domain_id_alloc();
3038 INIT_LIST_HEAD(&domain->dev_list);
3040 add_domain_to_list(domain);
3050 static int __init alloc_passthrough_domain(void)
3052 if (pt_domain != NULL)
3055 /* allocate passthrough domain */
3056 pt_domain = protection_domain_alloc();
3060 pt_domain->mode = PAGE_MODE_NONE;
3064 static int amd_iommu_domain_init(struct iommu_domain *dom)
3066 struct protection_domain *domain;
3068 domain = protection_domain_alloc();
3072 domain->mode = PAGE_MODE_3_LEVEL;
3073 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3074 if (!domain->pt_root)
3077 domain->iommu_domain = dom;
3084 protection_domain_free(domain);
3089 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3091 struct protection_domain *domain = dom->priv;
3096 if (domain->dev_cnt > 0)
3097 cleanup_domain(domain);
3099 BUG_ON(domain->dev_cnt != 0);
3101 if (domain->mode != PAGE_MODE_NONE)
3102 free_pagetable(domain);
3104 if (domain->flags & PD_IOMMUV2_MASK)
3105 free_gcr3_table(domain);
3107 protection_domain_free(domain);
3112 static void amd_iommu_detach_device(struct iommu_domain *dom,
3115 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3116 struct amd_iommu *iommu;
3119 if (!check_device(dev))
3122 devid = get_device_id(dev);
3124 if (dev_data->domain != NULL)
3127 iommu = amd_iommu_rlookup_table[devid];
3131 iommu_completion_wait(iommu);
3134 static int amd_iommu_attach_device(struct iommu_domain *dom,
3137 struct protection_domain *domain = dom->priv;
3138 struct iommu_dev_data *dev_data;
3139 struct amd_iommu *iommu;
3142 if (!check_device(dev))
3145 dev_data = dev->archdata.iommu;
3147 iommu = amd_iommu_rlookup_table[dev_data->devid];
3151 if (dev_data->domain)
3154 ret = attach_device(dev, domain);
3156 iommu_completion_wait(iommu);
3161 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3162 phys_addr_t paddr, size_t page_size, int iommu_prot)
3164 struct protection_domain *domain = dom->priv;
3168 if (domain->mode == PAGE_MODE_NONE)
3171 if (iommu_prot & IOMMU_READ)
3172 prot |= IOMMU_PROT_IR;
3173 if (iommu_prot & IOMMU_WRITE)
3174 prot |= IOMMU_PROT_IW;
3176 mutex_lock(&domain->api_lock);
3177 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3178 mutex_unlock(&domain->api_lock);
3183 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3186 struct protection_domain *domain = dom->priv;
3189 if (domain->mode == PAGE_MODE_NONE)
3192 mutex_lock(&domain->api_lock);
3193 unmap_size = iommu_unmap_page(domain, iova, page_size);
3194 mutex_unlock(&domain->api_lock);
3196 domain_flush_tlb_pde(domain);
3201 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3204 struct protection_domain *domain = dom->priv;
3205 unsigned long offset_mask;
3209 if (domain->mode == PAGE_MODE_NONE)
3212 pte = fetch_pte(domain, iova);
3214 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3217 if (PM_PTE_LEVEL(*pte) == 0)
3218 offset_mask = PAGE_SIZE - 1;
3220 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3222 __pte = *pte & PM_ADDR_MASK;
3223 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3228 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3232 case IOMMU_CAP_CACHE_COHERENCY:
3239 static int amd_iommu_device_group(struct device *dev, unsigned int *groupid)
3241 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3242 struct pci_dev *pdev = to_pci_dev(dev);
3248 if (pdev->is_virtfn || !iommu_group_mf)
3249 devid = dev_data->devid;
3251 devid = calc_devid(pdev->bus->number,
3252 PCI_DEVFN(PCI_SLOT(pdev->devfn), 0));
3254 *groupid = amd_iommu_alias_table[devid];
3259 static struct iommu_ops amd_iommu_ops = {
3260 .domain_init = amd_iommu_domain_init,
3261 .domain_destroy = amd_iommu_domain_destroy,
3262 .attach_dev = amd_iommu_attach_device,
3263 .detach_dev = amd_iommu_detach_device,
3264 .map = amd_iommu_map,
3265 .unmap = amd_iommu_unmap,
3266 .iova_to_phys = amd_iommu_iova_to_phys,
3267 .domain_has_cap = amd_iommu_domain_has_cap,
3268 .device_group = amd_iommu_device_group,
3269 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3272 /*****************************************************************************
3274 * The next functions do a basic initialization of IOMMU for pass through
3277 * In passthrough mode the IOMMU is initialized and enabled but not used for
3278 * DMA-API translation.
3280 *****************************************************************************/
3282 int __init amd_iommu_init_passthrough(void)
3284 struct iommu_dev_data *dev_data;
3285 struct pci_dev *dev = NULL;
3286 struct amd_iommu *iommu;
3290 ret = alloc_passthrough_domain();
3294 for_each_pci_dev(dev) {
3295 if (!check_device(&dev->dev))
3298 dev_data = get_dev_data(&dev->dev);
3299 dev_data->passthrough = true;
3301 devid = get_device_id(&dev->dev);
3303 iommu = amd_iommu_rlookup_table[devid];
3307 attach_device(&dev->dev, pt_domain);
3310 amd_iommu_stats_init();
3312 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3317 /* IOMMUv2 specific functions */
3318 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3320 return atomic_notifier_chain_register(&ppr_notifier, nb);
3322 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3324 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3326 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3328 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3330 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3332 struct protection_domain *domain = dom->priv;
3333 unsigned long flags;
3335 spin_lock_irqsave(&domain->lock, flags);
3337 /* Update data structure */
3338 domain->mode = PAGE_MODE_NONE;
3339 domain->updated = true;
3341 /* Make changes visible to IOMMUs */
3342 update_domain(domain);
3344 /* Page-table is not visible to IOMMU anymore, so free it */
3345 free_pagetable(domain);
3347 spin_unlock_irqrestore(&domain->lock, flags);
3349 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3351 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3353 struct protection_domain *domain = dom->priv;
3354 unsigned long flags;
3357 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3360 /* Number of GCR3 table levels required */
3361 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3364 if (levels > amd_iommu_max_glx_val)
3367 spin_lock_irqsave(&domain->lock, flags);
3370 * Save us all sanity checks whether devices already in the
3371 * domain support IOMMUv2. Just force that the domain has no
3372 * devices attached when it is switched into IOMMUv2 mode.
3375 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3379 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3380 if (domain->gcr3_tbl == NULL)
3383 domain->glx = levels;
3384 domain->flags |= PD_IOMMUV2_MASK;
3385 domain->updated = true;
3387 update_domain(domain);
3392 spin_unlock_irqrestore(&domain->lock, flags);
3396 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3398 static int __flush_pasid(struct protection_domain *domain, int pasid,
3399 u64 address, bool size)
3401 struct iommu_dev_data *dev_data;
3402 struct iommu_cmd cmd;
3405 if (!(domain->flags & PD_IOMMUV2_MASK))
3408 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3411 * IOMMU TLB needs to be flushed before Device TLB to
3412 * prevent device TLB refill from IOMMU TLB
3414 for (i = 0; i < amd_iommus_present; ++i) {
3415 if (domain->dev_iommu[i] == 0)
3418 ret = iommu_queue_command(amd_iommus[i], &cmd);
3423 /* Wait until IOMMU TLB flushes are complete */
3424 domain_flush_complete(domain);
3426 /* Now flush device TLBs */
3427 list_for_each_entry(dev_data, &domain->dev_list, list) {
3428 struct amd_iommu *iommu;
3431 BUG_ON(!dev_data->ats.enabled);
3433 qdep = dev_data->ats.qdep;
3434 iommu = amd_iommu_rlookup_table[dev_data->devid];
3436 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3437 qdep, address, size);
3439 ret = iommu_queue_command(iommu, &cmd);
3444 /* Wait until all device TLBs are flushed */
3445 domain_flush_complete(domain);
3454 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3457 INC_STATS_COUNTER(invalidate_iotlb);
3459 return __flush_pasid(domain, pasid, address, false);
3462 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3465 struct protection_domain *domain = dom->priv;
3466 unsigned long flags;
3469 spin_lock_irqsave(&domain->lock, flags);
3470 ret = __amd_iommu_flush_page(domain, pasid, address);
3471 spin_unlock_irqrestore(&domain->lock, flags);
3475 EXPORT_SYMBOL(amd_iommu_flush_page);
3477 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3479 INC_STATS_COUNTER(invalidate_iotlb_all);
3481 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3485 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3487 struct protection_domain *domain = dom->priv;
3488 unsigned long flags;
3491 spin_lock_irqsave(&domain->lock, flags);
3492 ret = __amd_iommu_flush_tlb(domain, pasid);
3493 spin_unlock_irqrestore(&domain->lock, flags);
3497 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3499 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3506 index = (pasid >> (9 * level)) & 0x1ff;
3512 if (!(*pte & GCR3_VALID)) {
3516 root = (void *)get_zeroed_page(GFP_ATOMIC);
3520 *pte = __pa(root) | GCR3_VALID;
3523 root = __va(*pte & PAGE_MASK);
3531 static int __set_gcr3(struct protection_domain *domain, int pasid,
3536 if (domain->mode != PAGE_MODE_NONE)
3539 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3543 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3545 return __amd_iommu_flush_tlb(domain, pasid);
3548 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3552 if (domain->mode != PAGE_MODE_NONE)
3555 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3561 return __amd_iommu_flush_tlb(domain, pasid);
3564 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3567 struct protection_domain *domain = dom->priv;
3568 unsigned long flags;
3571 spin_lock_irqsave(&domain->lock, flags);
3572 ret = __set_gcr3(domain, pasid, cr3);
3573 spin_unlock_irqrestore(&domain->lock, flags);
3577 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3579 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3581 struct protection_domain *domain = dom->priv;
3582 unsigned long flags;
3585 spin_lock_irqsave(&domain->lock, flags);
3586 ret = __clear_gcr3(domain, pasid);
3587 spin_unlock_irqrestore(&domain->lock, flags);
3591 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3593 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3594 int status, int tag)
3596 struct iommu_dev_data *dev_data;
3597 struct amd_iommu *iommu;
3598 struct iommu_cmd cmd;
3600 INC_STATS_COUNTER(complete_ppr);
3602 dev_data = get_dev_data(&pdev->dev);
3603 iommu = amd_iommu_rlookup_table[dev_data->devid];
3605 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3606 tag, dev_data->pri_tlp);
3608 return iommu_queue_command(iommu, &cmd);
3610 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3612 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3614 struct protection_domain *domain;
3616 domain = get_domain(&pdev->dev);
3620 /* Only return IOMMUv2 domains */
3621 if (!(domain->flags & PD_IOMMUV2_MASK))
3624 return domain->iommu_domain;
3626 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3628 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3630 struct iommu_dev_data *dev_data;
3632 if (!amd_iommu_v2_supported())
3635 dev_data = get_dev_data(&pdev->dev);
3636 dev_data->errata |= (1 << erratum);
3638 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3640 int amd_iommu_device_info(struct pci_dev *pdev,
3641 struct amd_iommu_device_info *info)
3646 if (pdev == NULL || info == NULL)
3649 if (!amd_iommu_v2_supported())
3652 memset(info, 0, sizeof(*info));
3654 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3656 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3658 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3660 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3662 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3666 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3667 max_pasids = min(max_pasids, (1 << 20));
3669 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3670 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3672 features = pci_pasid_features(pdev);
3673 if (features & PCI_PASID_CAP_EXEC)
3674 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3675 if (features & PCI_PASID_CAP_PRIV)
3676 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3681 EXPORT_SYMBOL(amd_iommu_device_info);