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 void swap_pci_ref(struct pci_dev **from, struct pci_dev *to)
265 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
267 static int iommu_init_device(struct device *dev)
269 struct pci_dev *dma_pdev, *pdev = to_pci_dev(dev);
270 struct iommu_dev_data *dev_data;
271 struct iommu_group *group;
275 if (dev->archdata.iommu)
278 dev_data = find_dev_data(get_device_id(dev));
282 alias = amd_iommu_alias_table[dev_data->devid];
283 if (alias != dev_data->devid) {
284 struct iommu_dev_data *alias_data;
286 alias_data = find_dev_data(alias);
287 if (alias_data == NULL) {
288 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
290 free_dev_data(dev_data);
293 dev_data->alias_data = alias_data;
295 dma_pdev = pci_get_bus_and_slot(alias >> 8, alias & 0xff);
297 dma_pdev = pci_dev_get(pdev);
299 /* Account for quirked devices */
300 swap_pci_ref(&dma_pdev, pci_get_dma_source(dma_pdev));
303 * If it's a multifunction device that does not support our
304 * required ACS flags, add to the same group as function 0.
306 if (dma_pdev->multifunction &&
307 !pci_acs_enabled(dma_pdev, REQ_ACS_FLAGS))
308 swap_pci_ref(&dma_pdev,
309 pci_get_slot(dma_pdev->bus,
310 PCI_DEVFN(PCI_SLOT(dma_pdev->devfn),
314 * Devices on the root bus go through the iommu. If that's not us,
315 * find the next upstream device and test ACS up to the root bus.
316 * Finding the next device may require skipping virtual buses.
318 while (!pci_is_root_bus(dma_pdev->bus)) {
319 struct pci_bus *bus = dma_pdev->bus;
322 if (!pci_is_root_bus(bus))
328 if (pci_acs_path_enabled(bus->self, NULL, REQ_ACS_FLAGS))
331 swap_pci_ref(&dma_pdev, pci_dev_get(bus->self));
335 group = iommu_group_get(&dma_pdev->dev);
336 pci_dev_put(dma_pdev);
338 group = iommu_group_alloc();
340 return PTR_ERR(group);
343 ret = iommu_group_add_device(group, dev);
345 iommu_group_put(group);
350 if (pci_iommuv2_capable(pdev)) {
351 struct amd_iommu *iommu;
353 iommu = amd_iommu_rlookup_table[dev_data->devid];
354 dev_data->iommu_v2 = iommu->is_iommu_v2;
357 dev->archdata.iommu = dev_data;
362 static void iommu_ignore_device(struct device *dev)
366 devid = get_device_id(dev);
367 alias = amd_iommu_alias_table[devid];
369 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
370 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
372 amd_iommu_rlookup_table[devid] = NULL;
373 amd_iommu_rlookup_table[alias] = NULL;
376 static void iommu_uninit_device(struct device *dev)
378 iommu_group_remove_device(dev);
381 * Nothing to do here - we keep dev_data around for unplugged devices
382 * and reuse it when the device is re-plugged - not doing so would
383 * introduce a ton of races.
387 void __init amd_iommu_uninit_devices(void)
389 struct iommu_dev_data *dev_data, *n;
390 struct pci_dev *pdev = NULL;
392 for_each_pci_dev(pdev) {
394 if (!check_device(&pdev->dev))
397 iommu_uninit_device(&pdev->dev);
400 /* Free all of our dev_data structures */
401 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
402 free_dev_data(dev_data);
405 int __init amd_iommu_init_devices(void)
407 struct pci_dev *pdev = NULL;
410 for_each_pci_dev(pdev) {
412 if (!check_device(&pdev->dev))
415 ret = iommu_init_device(&pdev->dev);
416 if (ret == -ENOTSUPP)
417 iommu_ignore_device(&pdev->dev);
426 amd_iommu_uninit_devices();
430 #ifdef CONFIG_AMD_IOMMU_STATS
433 * Initialization code for statistics collection
436 DECLARE_STATS_COUNTER(compl_wait);
437 DECLARE_STATS_COUNTER(cnt_map_single);
438 DECLARE_STATS_COUNTER(cnt_unmap_single);
439 DECLARE_STATS_COUNTER(cnt_map_sg);
440 DECLARE_STATS_COUNTER(cnt_unmap_sg);
441 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
442 DECLARE_STATS_COUNTER(cnt_free_coherent);
443 DECLARE_STATS_COUNTER(cross_page);
444 DECLARE_STATS_COUNTER(domain_flush_single);
445 DECLARE_STATS_COUNTER(domain_flush_all);
446 DECLARE_STATS_COUNTER(alloced_io_mem);
447 DECLARE_STATS_COUNTER(total_map_requests);
448 DECLARE_STATS_COUNTER(complete_ppr);
449 DECLARE_STATS_COUNTER(invalidate_iotlb);
450 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
451 DECLARE_STATS_COUNTER(pri_requests);
453 static struct dentry *stats_dir;
454 static struct dentry *de_fflush;
456 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
458 if (stats_dir == NULL)
461 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
465 static void amd_iommu_stats_init(void)
467 stats_dir = debugfs_create_dir("amd-iommu", NULL);
468 if (stats_dir == NULL)
471 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
472 &amd_iommu_unmap_flush);
474 amd_iommu_stats_add(&compl_wait);
475 amd_iommu_stats_add(&cnt_map_single);
476 amd_iommu_stats_add(&cnt_unmap_single);
477 amd_iommu_stats_add(&cnt_map_sg);
478 amd_iommu_stats_add(&cnt_unmap_sg);
479 amd_iommu_stats_add(&cnt_alloc_coherent);
480 amd_iommu_stats_add(&cnt_free_coherent);
481 amd_iommu_stats_add(&cross_page);
482 amd_iommu_stats_add(&domain_flush_single);
483 amd_iommu_stats_add(&domain_flush_all);
484 amd_iommu_stats_add(&alloced_io_mem);
485 amd_iommu_stats_add(&total_map_requests);
486 amd_iommu_stats_add(&complete_ppr);
487 amd_iommu_stats_add(&invalidate_iotlb);
488 amd_iommu_stats_add(&invalidate_iotlb_all);
489 amd_iommu_stats_add(&pri_requests);
494 /****************************************************************************
496 * Interrupt handling functions
498 ****************************************************************************/
500 static void dump_dte_entry(u16 devid)
504 for (i = 0; i < 4; ++i)
505 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
506 amd_iommu_dev_table[devid].data[i]);
509 static void dump_command(unsigned long phys_addr)
511 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
514 for (i = 0; i < 4; ++i)
515 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
518 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
520 int type, devid, domid, flags;
521 volatile u32 *event = __evt;
526 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
527 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
528 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
529 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
530 address = (u64)(((u64)event[3]) << 32) | event[2];
533 /* Did we hit the erratum? */
534 if (++count == LOOP_TIMEOUT) {
535 pr_err("AMD-Vi: No event written to event log\n");
542 printk(KERN_ERR "AMD-Vi: Event logged [");
545 case EVENT_TYPE_ILL_DEV:
546 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
547 "address=0x%016llx flags=0x%04x]\n",
548 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
550 dump_dte_entry(devid);
552 case EVENT_TYPE_IO_FAULT:
553 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
554 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
555 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
556 domid, address, flags);
558 case EVENT_TYPE_DEV_TAB_ERR:
559 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
560 "address=0x%016llx flags=0x%04x]\n",
561 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
564 case EVENT_TYPE_PAGE_TAB_ERR:
565 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
566 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
567 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
568 domid, address, flags);
570 case EVENT_TYPE_ILL_CMD:
571 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
572 dump_command(address);
574 case EVENT_TYPE_CMD_HARD_ERR:
575 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
576 "flags=0x%04x]\n", address, flags);
578 case EVENT_TYPE_IOTLB_INV_TO:
579 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
580 "address=0x%016llx]\n",
581 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
584 case EVENT_TYPE_INV_DEV_REQ:
585 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
586 "address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
591 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
594 memset(__evt, 0, 4 * sizeof(u32));
597 static void iommu_poll_events(struct amd_iommu *iommu)
602 spin_lock_irqsave(&iommu->lock, flags);
604 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
605 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
607 while (head != tail) {
608 iommu_print_event(iommu, iommu->evt_buf + head);
609 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
612 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
614 spin_unlock_irqrestore(&iommu->lock, flags);
617 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
619 struct amd_iommu_fault fault;
621 INC_STATS_COUNTER(pri_requests);
623 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
624 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
628 fault.address = raw[1];
629 fault.pasid = PPR_PASID(raw[0]);
630 fault.device_id = PPR_DEVID(raw[0]);
631 fault.tag = PPR_TAG(raw[0]);
632 fault.flags = PPR_FLAGS(raw[0]);
634 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
637 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
642 if (iommu->ppr_log == NULL)
645 /* enable ppr interrupts again */
646 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
648 spin_lock_irqsave(&iommu->lock, flags);
650 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
651 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
653 while (head != tail) {
658 raw = (u64 *)(iommu->ppr_log + head);
661 * Hardware bug: Interrupt may arrive before the entry is
662 * written to memory. If this happens we need to wait for the
665 for (i = 0; i < LOOP_TIMEOUT; ++i) {
666 if (PPR_REQ_TYPE(raw[0]) != 0)
671 /* Avoid memcpy function-call overhead */
676 * To detect the hardware bug we need to clear the entry
679 raw[0] = raw[1] = 0UL;
681 /* Update head pointer of hardware ring-buffer */
682 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
683 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
686 * Release iommu->lock because ppr-handling might need to
689 spin_unlock_irqrestore(&iommu->lock, flags);
691 /* Handle PPR entry */
692 iommu_handle_ppr_entry(iommu, entry);
694 spin_lock_irqsave(&iommu->lock, flags);
696 /* Refresh ring-buffer information */
697 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
698 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
701 spin_unlock_irqrestore(&iommu->lock, flags);
704 irqreturn_t amd_iommu_int_thread(int irq, void *data)
706 struct amd_iommu *iommu;
708 for_each_iommu(iommu) {
709 iommu_poll_events(iommu);
710 iommu_poll_ppr_log(iommu);
716 irqreturn_t amd_iommu_int_handler(int irq, void *data)
718 return IRQ_WAKE_THREAD;
721 /****************************************************************************
723 * IOMMU command queuing functions
725 ****************************************************************************/
727 static int wait_on_sem(volatile u64 *sem)
731 while (*sem == 0 && i < LOOP_TIMEOUT) {
736 if (i == LOOP_TIMEOUT) {
737 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
744 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
745 struct iommu_cmd *cmd,
750 target = iommu->cmd_buf + tail;
751 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
753 /* Copy command to buffer */
754 memcpy(target, cmd, sizeof(*cmd));
756 /* Tell the IOMMU about it */
757 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
760 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
762 WARN_ON(address & 0x7ULL);
764 memset(cmd, 0, sizeof(*cmd));
765 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
766 cmd->data[1] = upper_32_bits(__pa(address));
768 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
771 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
773 memset(cmd, 0, sizeof(*cmd));
774 cmd->data[0] = devid;
775 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
778 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
779 size_t size, u16 domid, int pde)
784 pages = iommu_num_pages(address, size, PAGE_SIZE);
789 * If we have to flush more than one page, flush all
790 * TLB entries for this domain
792 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
796 address &= PAGE_MASK;
798 memset(cmd, 0, sizeof(*cmd));
799 cmd->data[1] |= domid;
800 cmd->data[2] = lower_32_bits(address);
801 cmd->data[3] = upper_32_bits(address);
802 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
803 if (s) /* size bit - we flush more than one 4kb page */
804 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
805 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
806 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
809 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
810 u64 address, size_t size)
815 pages = iommu_num_pages(address, size, PAGE_SIZE);
820 * If we have to flush more than one page, flush all
821 * TLB entries for this domain
823 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
827 address &= PAGE_MASK;
829 memset(cmd, 0, sizeof(*cmd));
830 cmd->data[0] = devid;
831 cmd->data[0] |= (qdep & 0xff) << 24;
832 cmd->data[1] = devid;
833 cmd->data[2] = lower_32_bits(address);
834 cmd->data[3] = upper_32_bits(address);
835 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
837 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
840 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
841 u64 address, bool size)
843 memset(cmd, 0, sizeof(*cmd));
845 address &= ~(0xfffULL);
847 cmd->data[0] = pasid & PASID_MASK;
848 cmd->data[1] = domid;
849 cmd->data[2] = lower_32_bits(address);
850 cmd->data[3] = upper_32_bits(address);
851 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
852 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
854 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
855 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
858 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
859 int qdep, u64 address, bool size)
861 memset(cmd, 0, sizeof(*cmd));
863 address &= ~(0xfffULL);
865 cmd->data[0] = devid;
866 cmd->data[0] |= (pasid & 0xff) << 16;
867 cmd->data[0] |= (qdep & 0xff) << 24;
868 cmd->data[1] = devid;
869 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
870 cmd->data[2] = lower_32_bits(address);
871 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
872 cmd->data[3] = upper_32_bits(address);
874 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
875 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
878 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
879 int status, int tag, bool gn)
881 memset(cmd, 0, sizeof(*cmd));
883 cmd->data[0] = devid;
885 cmd->data[1] = pasid & PASID_MASK;
886 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
888 cmd->data[3] = tag & 0x1ff;
889 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
891 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
894 static void build_inv_all(struct iommu_cmd *cmd)
896 memset(cmd, 0, sizeof(*cmd));
897 CMD_SET_TYPE(cmd, CMD_INV_ALL);
901 * Writes the command to the IOMMUs command buffer and informs the
902 * hardware about the new command.
904 static int iommu_queue_command_sync(struct amd_iommu *iommu,
905 struct iommu_cmd *cmd,
908 u32 left, tail, head, next_tail;
911 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
914 spin_lock_irqsave(&iommu->lock, flags);
916 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
917 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
918 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
919 left = (head - next_tail) % iommu->cmd_buf_size;
922 struct iommu_cmd sync_cmd;
923 volatile u64 sem = 0;
926 build_completion_wait(&sync_cmd, (u64)&sem);
927 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
929 spin_unlock_irqrestore(&iommu->lock, flags);
931 if ((ret = wait_on_sem(&sem)) != 0)
937 copy_cmd_to_buffer(iommu, cmd, tail);
939 /* We need to sync now to make sure all commands are processed */
940 iommu->need_sync = sync;
942 spin_unlock_irqrestore(&iommu->lock, flags);
947 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
949 return iommu_queue_command_sync(iommu, cmd, true);
953 * This function queues a completion wait command into the command
956 static int iommu_completion_wait(struct amd_iommu *iommu)
958 struct iommu_cmd cmd;
959 volatile u64 sem = 0;
962 if (!iommu->need_sync)
965 build_completion_wait(&cmd, (u64)&sem);
967 ret = iommu_queue_command_sync(iommu, &cmd, false);
971 return wait_on_sem(&sem);
974 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
976 struct iommu_cmd cmd;
978 build_inv_dte(&cmd, devid);
980 return iommu_queue_command(iommu, &cmd);
983 static void iommu_flush_dte_all(struct amd_iommu *iommu)
987 for (devid = 0; devid <= 0xffff; ++devid)
988 iommu_flush_dte(iommu, devid);
990 iommu_completion_wait(iommu);
994 * This function uses heavy locking and may disable irqs for some time. But
995 * this is no issue because it is only called during resume.
997 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1001 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1002 struct iommu_cmd cmd;
1003 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1005 iommu_queue_command(iommu, &cmd);
1008 iommu_completion_wait(iommu);
1011 static void iommu_flush_all(struct amd_iommu *iommu)
1013 struct iommu_cmd cmd;
1015 build_inv_all(&cmd);
1017 iommu_queue_command(iommu, &cmd);
1018 iommu_completion_wait(iommu);
1021 void iommu_flush_all_caches(struct amd_iommu *iommu)
1023 if (iommu_feature(iommu, FEATURE_IA)) {
1024 iommu_flush_all(iommu);
1026 iommu_flush_dte_all(iommu);
1027 iommu_flush_tlb_all(iommu);
1032 * Command send function for flushing on-device TLB
1034 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1035 u64 address, size_t size)
1037 struct amd_iommu *iommu;
1038 struct iommu_cmd cmd;
1041 qdep = dev_data->ats.qdep;
1042 iommu = amd_iommu_rlookup_table[dev_data->devid];
1044 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1046 return iommu_queue_command(iommu, &cmd);
1050 * Command send function for invalidating a device table entry
1052 static int device_flush_dte(struct iommu_dev_data *dev_data)
1054 struct amd_iommu *iommu;
1057 iommu = amd_iommu_rlookup_table[dev_data->devid];
1059 ret = iommu_flush_dte(iommu, dev_data->devid);
1063 if (dev_data->ats.enabled)
1064 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1070 * TLB invalidation function which is called from the mapping functions.
1071 * It invalidates a single PTE if the range to flush is within a single
1072 * page. Otherwise it flushes the whole TLB of the IOMMU.
1074 static void __domain_flush_pages(struct protection_domain *domain,
1075 u64 address, size_t size, int pde)
1077 struct iommu_dev_data *dev_data;
1078 struct iommu_cmd cmd;
1081 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1083 for (i = 0; i < amd_iommus_present; ++i) {
1084 if (!domain->dev_iommu[i])
1088 * Devices of this domain are behind this IOMMU
1089 * We need a TLB flush
1091 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1094 list_for_each_entry(dev_data, &domain->dev_list, list) {
1096 if (!dev_data->ats.enabled)
1099 ret |= device_flush_iotlb(dev_data, address, size);
1105 static void domain_flush_pages(struct protection_domain *domain,
1106 u64 address, size_t size)
1108 __domain_flush_pages(domain, address, size, 0);
1111 /* Flush the whole IO/TLB for a given protection domain */
1112 static void domain_flush_tlb(struct protection_domain *domain)
1114 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1117 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1118 static void domain_flush_tlb_pde(struct protection_domain *domain)
1120 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1123 static void domain_flush_complete(struct protection_domain *domain)
1127 for (i = 0; i < amd_iommus_present; ++i) {
1128 if (!domain->dev_iommu[i])
1132 * Devices of this domain are behind this IOMMU
1133 * We need to wait for completion of all commands.
1135 iommu_completion_wait(amd_iommus[i]);
1141 * This function flushes the DTEs for all devices in domain
1143 static void domain_flush_devices(struct protection_domain *domain)
1145 struct iommu_dev_data *dev_data;
1147 list_for_each_entry(dev_data, &domain->dev_list, list)
1148 device_flush_dte(dev_data);
1151 /****************************************************************************
1153 * The functions below are used the create the page table mappings for
1154 * unity mapped regions.
1156 ****************************************************************************/
1159 * This function is used to add another level to an IO page table. Adding
1160 * another level increases the size of the address space by 9 bits to a size up
1163 static bool increase_address_space(struct protection_domain *domain,
1168 if (domain->mode == PAGE_MODE_6_LEVEL)
1169 /* address space already 64 bit large */
1172 pte = (void *)get_zeroed_page(gfp);
1176 *pte = PM_LEVEL_PDE(domain->mode,
1177 virt_to_phys(domain->pt_root));
1178 domain->pt_root = pte;
1180 domain->updated = true;
1185 static u64 *alloc_pte(struct protection_domain *domain,
1186 unsigned long address,
1187 unsigned long page_size,
1194 BUG_ON(!is_power_of_2(page_size));
1196 while (address > PM_LEVEL_SIZE(domain->mode))
1197 increase_address_space(domain, gfp);
1199 level = domain->mode - 1;
1200 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1201 address = PAGE_SIZE_ALIGN(address, page_size);
1202 end_lvl = PAGE_SIZE_LEVEL(page_size);
1204 while (level > end_lvl) {
1205 if (!IOMMU_PTE_PRESENT(*pte)) {
1206 page = (u64 *)get_zeroed_page(gfp);
1209 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1212 /* No level skipping support yet */
1213 if (PM_PTE_LEVEL(*pte) != level)
1218 pte = IOMMU_PTE_PAGE(*pte);
1220 if (pte_page && level == end_lvl)
1223 pte = &pte[PM_LEVEL_INDEX(level, address)];
1230 * This function checks if there is a PTE for a given dma address. If
1231 * there is one, it returns the pointer to it.
1233 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1238 if (address > PM_LEVEL_SIZE(domain->mode))
1241 level = domain->mode - 1;
1242 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1247 if (!IOMMU_PTE_PRESENT(*pte))
1251 if (PM_PTE_LEVEL(*pte) == 0x07) {
1252 unsigned long pte_mask, __pte;
1255 * If we have a series of large PTEs, make
1256 * sure to return a pointer to the first one.
1258 pte_mask = PTE_PAGE_SIZE(*pte);
1259 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1260 __pte = ((unsigned long)pte) & pte_mask;
1262 return (u64 *)__pte;
1265 /* No level skipping support yet */
1266 if (PM_PTE_LEVEL(*pte) != level)
1271 /* Walk to the next level */
1272 pte = IOMMU_PTE_PAGE(*pte);
1273 pte = &pte[PM_LEVEL_INDEX(level, address)];
1280 * Generic mapping functions. It maps a physical address into a DMA
1281 * address space. It allocates the page table pages if necessary.
1282 * In the future it can be extended to a generic mapping function
1283 * supporting all features of AMD IOMMU page tables like level skipping
1284 * and full 64 bit address spaces.
1286 static int iommu_map_page(struct protection_domain *dom,
1287 unsigned long bus_addr,
1288 unsigned long phys_addr,
1290 unsigned long page_size)
1295 if (!(prot & IOMMU_PROT_MASK))
1298 bus_addr = PAGE_ALIGN(bus_addr);
1299 phys_addr = PAGE_ALIGN(phys_addr);
1300 count = PAGE_SIZE_PTE_COUNT(page_size);
1301 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1303 for (i = 0; i < count; ++i)
1304 if (IOMMU_PTE_PRESENT(pte[i]))
1307 if (page_size > PAGE_SIZE) {
1308 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1309 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1311 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1313 if (prot & IOMMU_PROT_IR)
1314 __pte |= IOMMU_PTE_IR;
1315 if (prot & IOMMU_PROT_IW)
1316 __pte |= IOMMU_PTE_IW;
1318 for (i = 0; i < count; ++i)
1326 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1327 unsigned long bus_addr,
1328 unsigned long page_size)
1330 unsigned long long unmap_size, unmapped;
1333 BUG_ON(!is_power_of_2(page_size));
1337 while (unmapped < page_size) {
1339 pte = fetch_pte(dom, bus_addr);
1343 * No PTE for this address
1344 * move forward in 4kb steps
1346 unmap_size = PAGE_SIZE;
1347 } else if (PM_PTE_LEVEL(*pte) == 0) {
1348 /* 4kb PTE found for this address */
1349 unmap_size = PAGE_SIZE;
1354 /* Large PTE found which maps this address */
1355 unmap_size = PTE_PAGE_SIZE(*pte);
1356 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1357 for (i = 0; i < count; i++)
1361 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1362 unmapped += unmap_size;
1365 BUG_ON(!is_power_of_2(unmapped));
1371 * This function checks if a specific unity mapping entry is needed for
1372 * this specific IOMMU.
1374 static int iommu_for_unity_map(struct amd_iommu *iommu,
1375 struct unity_map_entry *entry)
1379 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1380 bdf = amd_iommu_alias_table[i];
1381 if (amd_iommu_rlookup_table[bdf] == iommu)
1389 * This function actually applies the mapping to the page table of the
1392 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1393 struct unity_map_entry *e)
1398 for (addr = e->address_start; addr < e->address_end;
1399 addr += PAGE_SIZE) {
1400 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1405 * if unity mapping is in aperture range mark the page
1406 * as allocated in the aperture
1408 if (addr < dma_dom->aperture_size)
1409 __set_bit(addr >> PAGE_SHIFT,
1410 dma_dom->aperture[0]->bitmap);
1417 * Init the unity mappings for a specific IOMMU in the system
1419 * Basically iterates over all unity mapping entries and applies them to
1420 * the default domain DMA of that IOMMU if necessary.
1422 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1424 struct unity_map_entry *entry;
1427 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1428 if (!iommu_for_unity_map(iommu, entry))
1430 ret = dma_ops_unity_map(iommu->default_dom, entry);
1439 * Inits the unity mappings required for a specific device
1441 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1444 struct unity_map_entry *e;
1447 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1448 if (!(devid >= e->devid_start && devid <= e->devid_end))
1450 ret = dma_ops_unity_map(dma_dom, e);
1458 /****************************************************************************
1460 * The next functions belong to the address allocator for the dma_ops
1461 * interface functions. They work like the allocators in the other IOMMU
1462 * drivers. Its basically a bitmap which marks the allocated pages in
1463 * the aperture. Maybe it could be enhanced in the future to a more
1464 * efficient allocator.
1466 ****************************************************************************/
1469 * The address allocator core functions.
1471 * called with domain->lock held
1475 * Used to reserve address ranges in the aperture (e.g. for exclusion
1478 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1479 unsigned long start_page,
1482 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1484 if (start_page + pages > last_page)
1485 pages = last_page - start_page;
1487 for (i = start_page; i < start_page + pages; ++i) {
1488 int index = i / APERTURE_RANGE_PAGES;
1489 int page = i % APERTURE_RANGE_PAGES;
1490 __set_bit(page, dom->aperture[index]->bitmap);
1495 * This function is used to add a new aperture range to an existing
1496 * aperture in case of dma_ops domain allocation or address allocation
1499 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1500 bool populate, gfp_t gfp)
1502 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1503 struct amd_iommu *iommu;
1504 unsigned long i, old_size;
1506 #ifdef CONFIG_IOMMU_STRESS
1510 if (index >= APERTURE_MAX_RANGES)
1513 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1514 if (!dma_dom->aperture[index])
1517 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1518 if (!dma_dom->aperture[index]->bitmap)
1521 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1524 unsigned long address = dma_dom->aperture_size;
1525 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1526 u64 *pte, *pte_page;
1528 for (i = 0; i < num_ptes; ++i) {
1529 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1534 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1536 address += APERTURE_RANGE_SIZE / 64;
1540 old_size = dma_dom->aperture_size;
1541 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1543 /* Reserve address range used for MSI messages */
1544 if (old_size < MSI_ADDR_BASE_LO &&
1545 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1546 unsigned long spage;
1549 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1550 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1552 dma_ops_reserve_addresses(dma_dom, spage, pages);
1555 /* Initialize the exclusion range if necessary */
1556 for_each_iommu(iommu) {
1557 if (iommu->exclusion_start &&
1558 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1559 && iommu->exclusion_start < dma_dom->aperture_size) {
1560 unsigned long startpage;
1561 int pages = iommu_num_pages(iommu->exclusion_start,
1562 iommu->exclusion_length,
1564 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1565 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1570 * Check for areas already mapped as present in the new aperture
1571 * range and mark those pages as reserved in the allocator. Such
1572 * mappings may already exist as a result of requested unity
1573 * mappings for devices.
1575 for (i = dma_dom->aperture[index]->offset;
1576 i < dma_dom->aperture_size;
1578 u64 *pte = fetch_pte(&dma_dom->domain, i);
1579 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1582 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1585 update_domain(&dma_dom->domain);
1590 update_domain(&dma_dom->domain);
1592 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1594 kfree(dma_dom->aperture[index]);
1595 dma_dom->aperture[index] = NULL;
1600 static unsigned long dma_ops_area_alloc(struct device *dev,
1601 struct dma_ops_domain *dom,
1603 unsigned long align_mask,
1605 unsigned long start)
1607 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1608 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1609 int i = start >> APERTURE_RANGE_SHIFT;
1610 unsigned long boundary_size;
1611 unsigned long address = -1;
1612 unsigned long limit;
1614 next_bit >>= PAGE_SHIFT;
1616 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1617 PAGE_SIZE) >> PAGE_SHIFT;
1619 for (;i < max_index; ++i) {
1620 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1622 if (dom->aperture[i]->offset >= dma_mask)
1625 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1626 dma_mask >> PAGE_SHIFT);
1628 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1629 limit, next_bit, pages, 0,
1630 boundary_size, align_mask);
1631 if (address != -1) {
1632 address = dom->aperture[i]->offset +
1633 (address << PAGE_SHIFT);
1634 dom->next_address = address + (pages << PAGE_SHIFT);
1644 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1645 struct dma_ops_domain *dom,
1647 unsigned long align_mask,
1650 unsigned long address;
1652 #ifdef CONFIG_IOMMU_STRESS
1653 dom->next_address = 0;
1654 dom->need_flush = true;
1657 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1658 dma_mask, dom->next_address);
1660 if (address == -1) {
1661 dom->next_address = 0;
1662 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1664 dom->need_flush = true;
1667 if (unlikely(address == -1))
1668 address = DMA_ERROR_CODE;
1670 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1676 * The address free function.
1678 * called with domain->lock held
1680 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1681 unsigned long address,
1684 unsigned i = address >> APERTURE_RANGE_SHIFT;
1685 struct aperture_range *range = dom->aperture[i];
1687 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1689 #ifdef CONFIG_IOMMU_STRESS
1694 if (address >= dom->next_address)
1695 dom->need_flush = true;
1697 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1699 bitmap_clear(range->bitmap, address, pages);
1703 /****************************************************************************
1705 * The next functions belong to the domain allocation. A domain is
1706 * allocated for every IOMMU as the default domain. If device isolation
1707 * is enabled, every device get its own domain. The most important thing
1708 * about domains is the page table mapping the DMA address space they
1711 ****************************************************************************/
1714 * This function adds a protection domain to the global protection domain list
1716 static void add_domain_to_list(struct protection_domain *domain)
1718 unsigned long flags;
1720 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1721 list_add(&domain->list, &amd_iommu_pd_list);
1722 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1726 * This function removes a protection domain to the global
1727 * protection domain list
1729 static void del_domain_from_list(struct protection_domain *domain)
1731 unsigned long flags;
1733 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1734 list_del(&domain->list);
1735 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1738 static u16 domain_id_alloc(void)
1740 unsigned long flags;
1743 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1744 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1746 if (id > 0 && id < MAX_DOMAIN_ID)
1747 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1750 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1755 static void domain_id_free(int id)
1757 unsigned long flags;
1759 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1760 if (id > 0 && id < MAX_DOMAIN_ID)
1761 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1762 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1765 static void free_pagetable(struct protection_domain *domain)
1770 p1 = domain->pt_root;
1775 for (i = 0; i < 512; ++i) {
1776 if (!IOMMU_PTE_PRESENT(p1[i]))
1779 p2 = IOMMU_PTE_PAGE(p1[i]);
1780 for (j = 0; j < 512; ++j) {
1781 if (!IOMMU_PTE_PRESENT(p2[j]))
1783 p3 = IOMMU_PTE_PAGE(p2[j]);
1784 free_page((unsigned long)p3);
1787 free_page((unsigned long)p2);
1790 free_page((unsigned long)p1);
1792 domain->pt_root = NULL;
1795 static void free_gcr3_tbl_level1(u64 *tbl)
1800 for (i = 0; i < 512; ++i) {
1801 if (!(tbl[i] & GCR3_VALID))
1804 ptr = __va(tbl[i] & PAGE_MASK);
1806 free_page((unsigned long)ptr);
1810 static void free_gcr3_tbl_level2(u64 *tbl)
1815 for (i = 0; i < 512; ++i) {
1816 if (!(tbl[i] & GCR3_VALID))
1819 ptr = __va(tbl[i] & PAGE_MASK);
1821 free_gcr3_tbl_level1(ptr);
1825 static void free_gcr3_table(struct protection_domain *domain)
1827 if (domain->glx == 2)
1828 free_gcr3_tbl_level2(domain->gcr3_tbl);
1829 else if (domain->glx == 1)
1830 free_gcr3_tbl_level1(domain->gcr3_tbl);
1831 else if (domain->glx != 0)
1834 free_page((unsigned long)domain->gcr3_tbl);
1838 * Free a domain, only used if something went wrong in the
1839 * allocation path and we need to free an already allocated page table
1841 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1848 del_domain_from_list(&dom->domain);
1850 free_pagetable(&dom->domain);
1852 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1853 if (!dom->aperture[i])
1855 free_page((unsigned long)dom->aperture[i]->bitmap);
1856 kfree(dom->aperture[i]);
1863 * Allocates a new protection domain usable for the dma_ops functions.
1864 * It also initializes the page table and the address allocator data
1865 * structures required for the dma_ops interface
1867 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1869 struct dma_ops_domain *dma_dom;
1871 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1875 spin_lock_init(&dma_dom->domain.lock);
1877 dma_dom->domain.id = domain_id_alloc();
1878 if (dma_dom->domain.id == 0)
1880 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1881 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1882 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1883 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1884 dma_dom->domain.priv = dma_dom;
1885 if (!dma_dom->domain.pt_root)
1888 dma_dom->need_flush = false;
1889 dma_dom->target_dev = 0xffff;
1891 add_domain_to_list(&dma_dom->domain);
1893 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1897 * mark the first page as allocated so we never return 0 as
1898 * a valid dma-address. So we can use 0 as error value
1900 dma_dom->aperture[0]->bitmap[0] = 1;
1901 dma_dom->next_address = 0;
1907 dma_ops_domain_free(dma_dom);
1913 * little helper function to check whether a given protection domain is a
1916 static bool dma_ops_domain(struct protection_domain *domain)
1918 return domain->flags & PD_DMA_OPS_MASK;
1921 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1926 if (domain->mode != PAGE_MODE_NONE)
1927 pte_root = virt_to_phys(domain->pt_root);
1929 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1930 << DEV_ENTRY_MODE_SHIFT;
1931 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1933 flags = amd_iommu_dev_table[devid].data[1];
1936 flags |= DTE_FLAG_IOTLB;
1938 if (domain->flags & PD_IOMMUV2_MASK) {
1939 u64 gcr3 = __pa(domain->gcr3_tbl);
1940 u64 glx = domain->glx;
1943 pte_root |= DTE_FLAG_GV;
1944 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1946 /* First mask out possible old values for GCR3 table */
1947 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1950 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1953 /* Encode GCR3 table into DTE */
1954 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1957 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1960 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1964 flags &= ~(0xffffUL);
1965 flags |= domain->id;
1967 amd_iommu_dev_table[devid].data[1] = flags;
1968 amd_iommu_dev_table[devid].data[0] = pte_root;
1971 static void clear_dte_entry(u16 devid)
1973 /* remove entry from the device table seen by the hardware */
1974 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1975 amd_iommu_dev_table[devid].data[1] = 0;
1977 amd_iommu_apply_erratum_63(devid);
1980 static void do_attach(struct iommu_dev_data *dev_data,
1981 struct protection_domain *domain)
1983 struct amd_iommu *iommu;
1986 iommu = amd_iommu_rlookup_table[dev_data->devid];
1987 ats = dev_data->ats.enabled;
1989 /* Update data structures */
1990 dev_data->domain = domain;
1991 list_add(&dev_data->list, &domain->dev_list);
1992 set_dte_entry(dev_data->devid, domain, ats);
1994 /* Do reference counting */
1995 domain->dev_iommu[iommu->index] += 1;
1996 domain->dev_cnt += 1;
1998 /* Flush the DTE entry */
1999 device_flush_dte(dev_data);
2002 static void do_detach(struct iommu_dev_data *dev_data)
2004 struct amd_iommu *iommu;
2006 iommu = amd_iommu_rlookup_table[dev_data->devid];
2008 /* decrease reference counters */
2009 dev_data->domain->dev_iommu[iommu->index] -= 1;
2010 dev_data->domain->dev_cnt -= 1;
2012 /* Update data structures */
2013 dev_data->domain = NULL;
2014 list_del(&dev_data->list);
2015 clear_dte_entry(dev_data->devid);
2017 /* Flush the DTE entry */
2018 device_flush_dte(dev_data);
2022 * If a device is not yet associated with a domain, this function does
2023 * assigns it visible for the hardware
2025 static int __attach_device(struct iommu_dev_data *dev_data,
2026 struct protection_domain *domain)
2031 spin_lock(&domain->lock);
2033 if (dev_data->alias_data != NULL) {
2034 struct iommu_dev_data *alias_data = dev_data->alias_data;
2036 /* Some sanity checks */
2038 if (alias_data->domain != NULL &&
2039 alias_data->domain != domain)
2042 if (dev_data->domain != NULL &&
2043 dev_data->domain != domain)
2046 /* Do real assignment */
2047 if (alias_data->domain == NULL)
2048 do_attach(alias_data, domain);
2050 atomic_inc(&alias_data->bind);
2053 if (dev_data->domain == NULL)
2054 do_attach(dev_data, domain);
2056 atomic_inc(&dev_data->bind);
2063 spin_unlock(&domain->lock);
2069 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2071 pci_disable_ats(pdev);
2072 pci_disable_pri(pdev);
2073 pci_disable_pasid(pdev);
2076 /* FIXME: Change generic reset-function to do the same */
2077 static int pri_reset_while_enabled(struct pci_dev *pdev)
2082 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2086 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2087 control |= PCI_PRI_CTRL_RESET;
2088 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2093 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2098 /* FIXME: Hardcode number of outstanding requests for now */
2100 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2102 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2104 /* Only allow access to user-accessible pages */
2105 ret = pci_enable_pasid(pdev, 0);
2109 /* First reset the PRI state of the device */
2110 ret = pci_reset_pri(pdev);
2115 ret = pci_enable_pri(pdev, reqs);
2120 ret = pri_reset_while_enabled(pdev);
2125 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2132 pci_disable_pri(pdev);
2133 pci_disable_pasid(pdev);
2138 /* FIXME: Move this to PCI code */
2139 #define PCI_PRI_TLP_OFF (1 << 15)
2141 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2146 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2150 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2152 return (status & PCI_PRI_TLP_OFF) ? true : false;
2156 * If a device is not yet associated with a domain, this function does
2157 * assigns it visible for the hardware
2159 static int attach_device(struct device *dev,
2160 struct protection_domain *domain)
2162 struct pci_dev *pdev = to_pci_dev(dev);
2163 struct iommu_dev_data *dev_data;
2164 unsigned long flags;
2167 dev_data = get_dev_data(dev);
2169 if (domain->flags & PD_IOMMUV2_MASK) {
2170 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2173 if (pdev_iommuv2_enable(pdev) != 0)
2176 dev_data->ats.enabled = true;
2177 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2178 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2179 } else if (amd_iommu_iotlb_sup &&
2180 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2181 dev_data->ats.enabled = true;
2182 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2185 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2186 ret = __attach_device(dev_data, domain);
2187 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2190 * We might boot into a crash-kernel here. The crashed kernel
2191 * left the caches in the IOMMU dirty. So we have to flush
2192 * here to evict all dirty stuff.
2194 domain_flush_tlb_pde(domain);
2200 * Removes a device from a protection domain (unlocked)
2202 static void __detach_device(struct iommu_dev_data *dev_data)
2204 struct protection_domain *domain;
2205 unsigned long flags;
2207 BUG_ON(!dev_data->domain);
2209 domain = dev_data->domain;
2211 spin_lock_irqsave(&domain->lock, flags);
2213 if (dev_data->alias_data != NULL) {
2214 struct iommu_dev_data *alias_data = dev_data->alias_data;
2216 if (atomic_dec_and_test(&alias_data->bind))
2217 do_detach(alias_data);
2220 if (atomic_dec_and_test(&dev_data->bind))
2221 do_detach(dev_data);
2223 spin_unlock_irqrestore(&domain->lock, flags);
2226 * If we run in passthrough mode the device must be assigned to the
2227 * passthrough domain if it is detached from any other domain.
2228 * Make sure we can deassign from the pt_domain itself.
2230 if (dev_data->passthrough &&
2231 (dev_data->domain == NULL && domain != pt_domain))
2232 __attach_device(dev_data, pt_domain);
2236 * Removes a device from a protection domain (with devtable_lock held)
2238 static void detach_device(struct device *dev)
2240 struct protection_domain *domain;
2241 struct iommu_dev_data *dev_data;
2242 unsigned long flags;
2244 dev_data = get_dev_data(dev);
2245 domain = dev_data->domain;
2247 /* lock device table */
2248 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2249 __detach_device(dev_data);
2250 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2252 if (domain->flags & PD_IOMMUV2_MASK)
2253 pdev_iommuv2_disable(to_pci_dev(dev));
2254 else if (dev_data->ats.enabled)
2255 pci_disable_ats(to_pci_dev(dev));
2257 dev_data->ats.enabled = false;
2261 * Find out the protection domain structure for a given PCI device. This
2262 * will give us the pointer to the page table root for example.
2264 static struct protection_domain *domain_for_device(struct device *dev)
2266 struct iommu_dev_data *dev_data;
2267 struct protection_domain *dom = NULL;
2268 unsigned long flags;
2270 dev_data = get_dev_data(dev);
2272 if (dev_data->domain)
2273 return dev_data->domain;
2275 if (dev_data->alias_data != NULL) {
2276 struct iommu_dev_data *alias_data = dev_data->alias_data;
2278 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2279 if (alias_data->domain != NULL) {
2280 __attach_device(dev_data, alias_data->domain);
2281 dom = alias_data->domain;
2283 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2289 static int device_change_notifier(struct notifier_block *nb,
2290 unsigned long action, void *data)
2292 struct dma_ops_domain *dma_domain;
2293 struct protection_domain *domain;
2294 struct iommu_dev_data *dev_data;
2295 struct device *dev = data;
2296 struct amd_iommu *iommu;
2297 unsigned long flags;
2300 if (!check_device(dev))
2303 devid = get_device_id(dev);
2304 iommu = amd_iommu_rlookup_table[devid];
2305 dev_data = get_dev_data(dev);
2308 case BUS_NOTIFY_UNBOUND_DRIVER:
2310 domain = domain_for_device(dev);
2314 if (dev_data->passthrough)
2318 case BUS_NOTIFY_ADD_DEVICE:
2320 iommu_init_device(dev);
2323 * dev_data is still NULL and
2324 * got initialized in iommu_init_device
2326 dev_data = get_dev_data(dev);
2328 if (iommu_pass_through || dev_data->iommu_v2) {
2329 dev_data->passthrough = true;
2330 attach_device(dev, pt_domain);
2334 domain = domain_for_device(dev);
2336 /* allocate a protection domain if a device is added */
2337 dma_domain = find_protection_domain(devid);
2340 dma_domain = dma_ops_domain_alloc();
2343 dma_domain->target_dev = devid;
2345 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2346 list_add_tail(&dma_domain->list, &iommu_pd_list);
2347 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2349 dev_data = get_dev_data(dev);
2351 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2354 case BUS_NOTIFY_DEL_DEVICE:
2356 iommu_uninit_device(dev);
2362 iommu_completion_wait(iommu);
2368 static struct notifier_block device_nb = {
2369 .notifier_call = device_change_notifier,
2372 void amd_iommu_init_notifier(void)
2374 bus_register_notifier(&pci_bus_type, &device_nb);
2377 /*****************************************************************************
2379 * The next functions belong to the dma_ops mapping/unmapping code.
2381 *****************************************************************************/
2384 * In the dma_ops path we only have the struct device. This function
2385 * finds the corresponding IOMMU, the protection domain and the
2386 * requestor id for a given device.
2387 * If the device is not yet associated with a domain this is also done
2390 static struct protection_domain *get_domain(struct device *dev)
2392 struct protection_domain *domain;
2393 struct dma_ops_domain *dma_dom;
2394 u16 devid = get_device_id(dev);
2396 if (!check_device(dev))
2397 return ERR_PTR(-EINVAL);
2399 domain = domain_for_device(dev);
2400 if (domain != NULL && !dma_ops_domain(domain))
2401 return ERR_PTR(-EBUSY);
2406 /* Device not bount yet - bind it */
2407 dma_dom = find_protection_domain(devid);
2409 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2410 attach_device(dev, &dma_dom->domain);
2411 DUMP_printk("Using protection domain %d for device %s\n",
2412 dma_dom->domain.id, dev_name(dev));
2414 return &dma_dom->domain;
2417 static void update_device_table(struct protection_domain *domain)
2419 struct iommu_dev_data *dev_data;
2421 list_for_each_entry(dev_data, &domain->dev_list, list)
2422 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2425 static void update_domain(struct protection_domain *domain)
2427 if (!domain->updated)
2430 update_device_table(domain);
2432 domain_flush_devices(domain);
2433 domain_flush_tlb_pde(domain);
2435 domain->updated = false;
2439 * This function fetches the PTE for a given address in the aperture
2441 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2442 unsigned long address)
2444 struct aperture_range *aperture;
2445 u64 *pte, *pte_page;
2447 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2451 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2453 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2455 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2457 pte += PM_LEVEL_INDEX(0, address);
2459 update_domain(&dom->domain);
2465 * This is the generic map function. It maps one 4kb page at paddr to
2466 * the given address in the DMA address space for the domain.
2468 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2469 unsigned long address,
2475 WARN_ON(address > dom->aperture_size);
2479 pte = dma_ops_get_pte(dom, address);
2481 return DMA_ERROR_CODE;
2483 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2485 if (direction == DMA_TO_DEVICE)
2486 __pte |= IOMMU_PTE_IR;
2487 else if (direction == DMA_FROM_DEVICE)
2488 __pte |= IOMMU_PTE_IW;
2489 else if (direction == DMA_BIDIRECTIONAL)
2490 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2496 return (dma_addr_t)address;
2500 * The generic unmapping function for on page in the DMA address space.
2502 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2503 unsigned long address)
2505 struct aperture_range *aperture;
2508 if (address >= dom->aperture_size)
2511 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2515 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2519 pte += PM_LEVEL_INDEX(0, address);
2527 * This function contains common code for mapping of a physically
2528 * contiguous memory region into DMA address space. It is used by all
2529 * mapping functions provided with this IOMMU driver.
2530 * Must be called with the domain lock held.
2532 static dma_addr_t __map_single(struct device *dev,
2533 struct dma_ops_domain *dma_dom,
2540 dma_addr_t offset = paddr & ~PAGE_MASK;
2541 dma_addr_t address, start, ret;
2543 unsigned long align_mask = 0;
2546 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2549 INC_STATS_COUNTER(total_map_requests);
2552 INC_STATS_COUNTER(cross_page);
2555 align_mask = (1UL << get_order(size)) - 1;
2558 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2560 if (unlikely(address == DMA_ERROR_CODE)) {
2562 * setting next_address here will let the address
2563 * allocator only scan the new allocated range in the
2564 * first run. This is a small optimization.
2566 dma_dom->next_address = dma_dom->aperture_size;
2568 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2572 * aperture was successfully enlarged by 128 MB, try
2579 for (i = 0; i < pages; ++i) {
2580 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2581 if (ret == DMA_ERROR_CODE)
2589 ADD_STATS_COUNTER(alloced_io_mem, size);
2591 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2592 domain_flush_tlb(&dma_dom->domain);
2593 dma_dom->need_flush = false;
2594 } else if (unlikely(amd_iommu_np_cache))
2595 domain_flush_pages(&dma_dom->domain, address, size);
2602 for (--i; i >= 0; --i) {
2604 dma_ops_domain_unmap(dma_dom, start);
2607 dma_ops_free_addresses(dma_dom, address, pages);
2609 return DMA_ERROR_CODE;
2613 * Does the reverse of the __map_single function. Must be called with
2614 * the domain lock held too
2616 static void __unmap_single(struct dma_ops_domain *dma_dom,
2617 dma_addr_t dma_addr,
2621 dma_addr_t flush_addr;
2622 dma_addr_t i, start;
2625 if ((dma_addr == DMA_ERROR_CODE) ||
2626 (dma_addr + size > dma_dom->aperture_size))
2629 flush_addr = dma_addr;
2630 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2631 dma_addr &= PAGE_MASK;
2634 for (i = 0; i < pages; ++i) {
2635 dma_ops_domain_unmap(dma_dom, start);
2639 SUB_STATS_COUNTER(alloced_io_mem, size);
2641 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2643 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2644 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2645 dma_dom->need_flush = false;
2650 * The exported map_single function for dma_ops.
2652 static dma_addr_t map_page(struct device *dev, struct page *page,
2653 unsigned long offset, size_t size,
2654 enum dma_data_direction dir,
2655 struct dma_attrs *attrs)
2657 unsigned long flags;
2658 struct protection_domain *domain;
2661 phys_addr_t paddr = page_to_phys(page) + offset;
2663 INC_STATS_COUNTER(cnt_map_single);
2665 domain = get_domain(dev);
2666 if (PTR_ERR(domain) == -EINVAL)
2667 return (dma_addr_t)paddr;
2668 else if (IS_ERR(domain))
2669 return DMA_ERROR_CODE;
2671 dma_mask = *dev->dma_mask;
2673 spin_lock_irqsave(&domain->lock, flags);
2675 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2677 if (addr == DMA_ERROR_CODE)
2680 domain_flush_complete(domain);
2683 spin_unlock_irqrestore(&domain->lock, flags);
2689 * The exported unmap_single function for dma_ops.
2691 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2692 enum dma_data_direction dir, struct dma_attrs *attrs)
2694 unsigned long flags;
2695 struct protection_domain *domain;
2697 INC_STATS_COUNTER(cnt_unmap_single);
2699 domain = get_domain(dev);
2703 spin_lock_irqsave(&domain->lock, flags);
2705 __unmap_single(domain->priv, dma_addr, size, dir);
2707 domain_flush_complete(domain);
2709 spin_unlock_irqrestore(&domain->lock, flags);
2713 * This is a special map_sg function which is used if we should map a
2714 * device which is not handled by an AMD IOMMU in the system.
2716 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2717 int nelems, int dir)
2719 struct scatterlist *s;
2722 for_each_sg(sglist, s, nelems, i) {
2723 s->dma_address = (dma_addr_t)sg_phys(s);
2724 s->dma_length = s->length;
2731 * The exported map_sg function for dma_ops (handles scatter-gather
2734 static int map_sg(struct device *dev, struct scatterlist *sglist,
2735 int nelems, enum dma_data_direction dir,
2736 struct dma_attrs *attrs)
2738 unsigned long flags;
2739 struct protection_domain *domain;
2741 struct scatterlist *s;
2743 int mapped_elems = 0;
2746 INC_STATS_COUNTER(cnt_map_sg);
2748 domain = get_domain(dev);
2749 if (PTR_ERR(domain) == -EINVAL)
2750 return map_sg_no_iommu(dev, sglist, nelems, dir);
2751 else if (IS_ERR(domain))
2754 dma_mask = *dev->dma_mask;
2756 spin_lock_irqsave(&domain->lock, flags);
2758 for_each_sg(sglist, s, nelems, i) {
2761 s->dma_address = __map_single(dev, domain->priv,
2762 paddr, s->length, dir, false,
2765 if (s->dma_address) {
2766 s->dma_length = s->length;
2772 domain_flush_complete(domain);
2775 spin_unlock_irqrestore(&domain->lock, flags);
2777 return mapped_elems;
2779 for_each_sg(sglist, s, mapped_elems, i) {
2781 __unmap_single(domain->priv, s->dma_address,
2782 s->dma_length, dir);
2783 s->dma_address = s->dma_length = 0;
2792 * The exported map_sg function for dma_ops (handles scatter-gather
2795 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2796 int nelems, enum dma_data_direction dir,
2797 struct dma_attrs *attrs)
2799 unsigned long flags;
2800 struct protection_domain *domain;
2801 struct scatterlist *s;
2804 INC_STATS_COUNTER(cnt_unmap_sg);
2806 domain = get_domain(dev);
2810 spin_lock_irqsave(&domain->lock, flags);
2812 for_each_sg(sglist, s, nelems, i) {
2813 __unmap_single(domain->priv, s->dma_address,
2814 s->dma_length, dir);
2815 s->dma_address = s->dma_length = 0;
2818 domain_flush_complete(domain);
2820 spin_unlock_irqrestore(&domain->lock, flags);
2824 * The exported alloc_coherent function for dma_ops.
2826 static void *alloc_coherent(struct device *dev, size_t size,
2827 dma_addr_t *dma_addr, gfp_t flag,
2828 struct dma_attrs *attrs)
2830 unsigned long flags;
2832 struct protection_domain *domain;
2834 u64 dma_mask = dev->coherent_dma_mask;
2836 INC_STATS_COUNTER(cnt_alloc_coherent);
2838 domain = get_domain(dev);
2839 if (PTR_ERR(domain) == -EINVAL) {
2840 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2841 *dma_addr = __pa(virt_addr);
2843 } else if (IS_ERR(domain))
2846 dma_mask = dev->coherent_dma_mask;
2847 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2850 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2854 paddr = virt_to_phys(virt_addr);
2857 dma_mask = *dev->dma_mask;
2859 spin_lock_irqsave(&domain->lock, flags);
2861 *dma_addr = __map_single(dev, domain->priv, paddr,
2862 size, DMA_BIDIRECTIONAL, true, dma_mask);
2864 if (*dma_addr == DMA_ERROR_CODE) {
2865 spin_unlock_irqrestore(&domain->lock, flags);
2869 domain_flush_complete(domain);
2871 spin_unlock_irqrestore(&domain->lock, flags);
2877 free_pages((unsigned long)virt_addr, get_order(size));
2883 * The exported free_coherent function for dma_ops.
2885 static void free_coherent(struct device *dev, size_t size,
2886 void *virt_addr, dma_addr_t dma_addr,
2887 struct dma_attrs *attrs)
2889 unsigned long flags;
2890 struct protection_domain *domain;
2892 INC_STATS_COUNTER(cnt_free_coherent);
2894 domain = get_domain(dev);
2898 spin_lock_irqsave(&domain->lock, flags);
2900 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2902 domain_flush_complete(domain);
2904 spin_unlock_irqrestore(&domain->lock, flags);
2907 free_pages((unsigned long)virt_addr, get_order(size));
2911 * This function is called by the DMA layer to find out if we can handle a
2912 * particular device. It is part of the dma_ops.
2914 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2916 return check_device(dev);
2920 * The function for pre-allocating protection domains.
2922 * If the driver core informs the DMA layer if a driver grabs a device
2923 * we don't need to preallocate the protection domains anymore.
2924 * For now we have to.
2926 static void __init prealloc_protection_domains(void)
2928 struct iommu_dev_data *dev_data;
2929 struct dma_ops_domain *dma_dom;
2930 struct pci_dev *dev = NULL;
2933 for_each_pci_dev(dev) {
2935 /* Do we handle this device? */
2936 if (!check_device(&dev->dev))
2939 dev_data = get_dev_data(&dev->dev);
2940 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2941 /* Make sure passthrough domain is allocated */
2942 alloc_passthrough_domain();
2943 dev_data->passthrough = true;
2944 attach_device(&dev->dev, pt_domain);
2945 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2946 dev_name(&dev->dev));
2949 /* Is there already any domain for it? */
2950 if (domain_for_device(&dev->dev))
2953 devid = get_device_id(&dev->dev);
2955 dma_dom = dma_ops_domain_alloc();
2958 init_unity_mappings_for_device(dma_dom, devid);
2959 dma_dom->target_dev = devid;
2961 attach_device(&dev->dev, &dma_dom->domain);
2963 list_add_tail(&dma_dom->list, &iommu_pd_list);
2967 static struct dma_map_ops amd_iommu_dma_ops = {
2968 .alloc = alloc_coherent,
2969 .free = free_coherent,
2970 .map_page = map_page,
2971 .unmap_page = unmap_page,
2973 .unmap_sg = unmap_sg,
2974 .dma_supported = amd_iommu_dma_supported,
2977 static unsigned device_dma_ops_init(void)
2979 struct iommu_dev_data *dev_data;
2980 struct pci_dev *pdev = NULL;
2981 unsigned unhandled = 0;
2983 for_each_pci_dev(pdev) {
2984 if (!check_device(&pdev->dev)) {
2986 iommu_ignore_device(&pdev->dev);
2992 dev_data = get_dev_data(&pdev->dev);
2994 if (!dev_data->passthrough)
2995 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2997 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
3004 * The function which clues the AMD IOMMU driver into dma_ops.
3007 void __init amd_iommu_init_api(void)
3009 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
3012 int __init amd_iommu_init_dma_ops(void)
3014 struct amd_iommu *iommu;
3018 * first allocate a default protection domain for every IOMMU we
3019 * found in the system. Devices not assigned to any other
3020 * protection domain will be assigned to the default one.
3022 for_each_iommu(iommu) {
3023 iommu->default_dom = dma_ops_domain_alloc();
3024 if (iommu->default_dom == NULL)
3026 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
3027 ret = iommu_init_unity_mappings(iommu);
3033 * Pre-allocate the protection domains for each device.
3035 prealloc_protection_domains();
3040 /* Make the driver finally visible to the drivers */
3041 unhandled = device_dma_ops_init();
3042 if (unhandled && max_pfn > MAX_DMA32_PFN) {
3043 /* There are unhandled devices - initialize swiotlb for them */
3047 amd_iommu_stats_init();
3049 if (amd_iommu_unmap_flush)
3050 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3052 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3058 for_each_iommu(iommu) {
3059 if (iommu->default_dom)
3060 dma_ops_domain_free(iommu->default_dom);
3066 /*****************************************************************************
3068 * The following functions belong to the exported interface of AMD IOMMU
3070 * This interface allows access to lower level functions of the IOMMU
3071 * like protection domain handling and assignement of devices to domains
3072 * which is not possible with the dma_ops interface.
3074 *****************************************************************************/
3076 static void cleanup_domain(struct protection_domain *domain)
3078 struct iommu_dev_data *dev_data, *next;
3079 unsigned long flags;
3081 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3083 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
3084 __detach_device(dev_data);
3085 atomic_set(&dev_data->bind, 0);
3088 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3091 static void protection_domain_free(struct protection_domain *domain)
3096 del_domain_from_list(domain);
3099 domain_id_free(domain->id);
3104 static struct protection_domain *protection_domain_alloc(void)
3106 struct protection_domain *domain;
3108 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3112 spin_lock_init(&domain->lock);
3113 mutex_init(&domain->api_lock);
3114 domain->id = domain_id_alloc();
3117 INIT_LIST_HEAD(&domain->dev_list);
3119 add_domain_to_list(domain);
3129 static int __init alloc_passthrough_domain(void)
3131 if (pt_domain != NULL)
3134 /* allocate passthrough domain */
3135 pt_domain = protection_domain_alloc();
3139 pt_domain->mode = PAGE_MODE_NONE;
3143 static int amd_iommu_domain_init(struct iommu_domain *dom)
3145 struct protection_domain *domain;
3147 domain = protection_domain_alloc();
3151 domain->mode = PAGE_MODE_3_LEVEL;
3152 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3153 if (!domain->pt_root)
3156 domain->iommu_domain = dom;
3160 dom->geometry.aperture_start = 0;
3161 dom->geometry.aperture_end = ~0ULL;
3162 dom->geometry.force_aperture = true;
3167 protection_domain_free(domain);
3172 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3174 struct protection_domain *domain = dom->priv;
3179 if (domain->dev_cnt > 0)
3180 cleanup_domain(domain);
3182 BUG_ON(domain->dev_cnt != 0);
3184 if (domain->mode != PAGE_MODE_NONE)
3185 free_pagetable(domain);
3187 if (domain->flags & PD_IOMMUV2_MASK)
3188 free_gcr3_table(domain);
3190 protection_domain_free(domain);
3195 static void amd_iommu_detach_device(struct iommu_domain *dom,
3198 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3199 struct amd_iommu *iommu;
3202 if (!check_device(dev))
3205 devid = get_device_id(dev);
3207 if (dev_data->domain != NULL)
3210 iommu = amd_iommu_rlookup_table[devid];
3214 iommu_completion_wait(iommu);
3217 static int amd_iommu_attach_device(struct iommu_domain *dom,
3220 struct protection_domain *domain = dom->priv;
3221 struct iommu_dev_data *dev_data;
3222 struct amd_iommu *iommu;
3225 if (!check_device(dev))
3228 dev_data = dev->archdata.iommu;
3230 iommu = amd_iommu_rlookup_table[dev_data->devid];
3234 if (dev_data->domain)
3237 ret = attach_device(dev, domain);
3239 iommu_completion_wait(iommu);
3244 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3245 phys_addr_t paddr, size_t page_size, int iommu_prot)
3247 struct protection_domain *domain = dom->priv;
3251 if (domain->mode == PAGE_MODE_NONE)
3254 if (iommu_prot & IOMMU_READ)
3255 prot |= IOMMU_PROT_IR;
3256 if (iommu_prot & IOMMU_WRITE)
3257 prot |= IOMMU_PROT_IW;
3259 mutex_lock(&domain->api_lock);
3260 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3261 mutex_unlock(&domain->api_lock);
3266 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3269 struct protection_domain *domain = dom->priv;
3272 if (domain->mode == PAGE_MODE_NONE)
3275 mutex_lock(&domain->api_lock);
3276 unmap_size = iommu_unmap_page(domain, iova, page_size);
3277 mutex_unlock(&domain->api_lock);
3279 domain_flush_tlb_pde(domain);
3284 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3287 struct protection_domain *domain = dom->priv;
3288 unsigned long offset_mask;
3292 if (domain->mode == PAGE_MODE_NONE)
3295 pte = fetch_pte(domain, iova);
3297 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3300 if (PM_PTE_LEVEL(*pte) == 0)
3301 offset_mask = PAGE_SIZE - 1;
3303 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3305 __pte = *pte & PM_ADDR_MASK;
3306 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3311 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3315 case IOMMU_CAP_CACHE_COHERENCY:
3322 static struct iommu_ops amd_iommu_ops = {
3323 .domain_init = amd_iommu_domain_init,
3324 .domain_destroy = amd_iommu_domain_destroy,
3325 .attach_dev = amd_iommu_attach_device,
3326 .detach_dev = amd_iommu_detach_device,
3327 .map = amd_iommu_map,
3328 .unmap = amd_iommu_unmap,
3329 .iova_to_phys = amd_iommu_iova_to_phys,
3330 .domain_has_cap = amd_iommu_domain_has_cap,
3331 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3334 /*****************************************************************************
3336 * The next functions do a basic initialization of IOMMU for pass through
3339 * In passthrough mode the IOMMU is initialized and enabled but not used for
3340 * DMA-API translation.
3342 *****************************************************************************/
3344 int __init amd_iommu_init_passthrough(void)
3346 struct iommu_dev_data *dev_data;
3347 struct pci_dev *dev = NULL;
3348 struct amd_iommu *iommu;
3352 ret = alloc_passthrough_domain();
3356 for_each_pci_dev(dev) {
3357 if (!check_device(&dev->dev))
3360 dev_data = get_dev_data(&dev->dev);
3361 dev_data->passthrough = true;
3363 devid = get_device_id(&dev->dev);
3365 iommu = amd_iommu_rlookup_table[devid];
3369 attach_device(&dev->dev, pt_domain);
3372 amd_iommu_stats_init();
3374 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3379 /* IOMMUv2 specific functions */
3380 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3382 return atomic_notifier_chain_register(&ppr_notifier, nb);
3384 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3386 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3388 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3390 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3392 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3394 struct protection_domain *domain = dom->priv;
3395 unsigned long flags;
3397 spin_lock_irqsave(&domain->lock, flags);
3399 /* Update data structure */
3400 domain->mode = PAGE_MODE_NONE;
3401 domain->updated = true;
3403 /* Make changes visible to IOMMUs */
3404 update_domain(domain);
3406 /* Page-table is not visible to IOMMU anymore, so free it */
3407 free_pagetable(domain);
3409 spin_unlock_irqrestore(&domain->lock, flags);
3411 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3413 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3415 struct protection_domain *domain = dom->priv;
3416 unsigned long flags;
3419 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3422 /* Number of GCR3 table levels required */
3423 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3426 if (levels > amd_iommu_max_glx_val)
3429 spin_lock_irqsave(&domain->lock, flags);
3432 * Save us all sanity checks whether devices already in the
3433 * domain support IOMMUv2. Just force that the domain has no
3434 * devices attached when it is switched into IOMMUv2 mode.
3437 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3441 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3442 if (domain->gcr3_tbl == NULL)
3445 domain->glx = levels;
3446 domain->flags |= PD_IOMMUV2_MASK;
3447 domain->updated = true;
3449 update_domain(domain);
3454 spin_unlock_irqrestore(&domain->lock, flags);
3458 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3460 static int __flush_pasid(struct protection_domain *domain, int pasid,
3461 u64 address, bool size)
3463 struct iommu_dev_data *dev_data;
3464 struct iommu_cmd cmd;
3467 if (!(domain->flags & PD_IOMMUV2_MASK))
3470 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3473 * IOMMU TLB needs to be flushed before Device TLB to
3474 * prevent device TLB refill from IOMMU TLB
3476 for (i = 0; i < amd_iommus_present; ++i) {
3477 if (domain->dev_iommu[i] == 0)
3480 ret = iommu_queue_command(amd_iommus[i], &cmd);
3485 /* Wait until IOMMU TLB flushes are complete */
3486 domain_flush_complete(domain);
3488 /* Now flush device TLBs */
3489 list_for_each_entry(dev_data, &domain->dev_list, list) {
3490 struct amd_iommu *iommu;
3493 BUG_ON(!dev_data->ats.enabled);
3495 qdep = dev_data->ats.qdep;
3496 iommu = amd_iommu_rlookup_table[dev_data->devid];
3498 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3499 qdep, address, size);
3501 ret = iommu_queue_command(iommu, &cmd);
3506 /* Wait until all device TLBs are flushed */
3507 domain_flush_complete(domain);
3516 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3519 INC_STATS_COUNTER(invalidate_iotlb);
3521 return __flush_pasid(domain, pasid, address, false);
3524 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3527 struct protection_domain *domain = dom->priv;
3528 unsigned long flags;
3531 spin_lock_irqsave(&domain->lock, flags);
3532 ret = __amd_iommu_flush_page(domain, pasid, address);
3533 spin_unlock_irqrestore(&domain->lock, flags);
3537 EXPORT_SYMBOL(amd_iommu_flush_page);
3539 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3541 INC_STATS_COUNTER(invalidate_iotlb_all);
3543 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3547 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3549 struct protection_domain *domain = dom->priv;
3550 unsigned long flags;
3553 spin_lock_irqsave(&domain->lock, flags);
3554 ret = __amd_iommu_flush_tlb(domain, pasid);
3555 spin_unlock_irqrestore(&domain->lock, flags);
3559 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3561 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3568 index = (pasid >> (9 * level)) & 0x1ff;
3574 if (!(*pte & GCR3_VALID)) {
3578 root = (void *)get_zeroed_page(GFP_ATOMIC);
3582 *pte = __pa(root) | GCR3_VALID;
3585 root = __va(*pte & PAGE_MASK);
3593 static int __set_gcr3(struct protection_domain *domain, int pasid,
3598 if (domain->mode != PAGE_MODE_NONE)
3601 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3605 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3607 return __amd_iommu_flush_tlb(domain, pasid);
3610 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3614 if (domain->mode != PAGE_MODE_NONE)
3617 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3623 return __amd_iommu_flush_tlb(domain, pasid);
3626 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3629 struct protection_domain *domain = dom->priv;
3630 unsigned long flags;
3633 spin_lock_irqsave(&domain->lock, flags);
3634 ret = __set_gcr3(domain, pasid, cr3);
3635 spin_unlock_irqrestore(&domain->lock, flags);
3639 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3641 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3643 struct protection_domain *domain = dom->priv;
3644 unsigned long flags;
3647 spin_lock_irqsave(&domain->lock, flags);
3648 ret = __clear_gcr3(domain, pasid);
3649 spin_unlock_irqrestore(&domain->lock, flags);
3653 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3655 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3656 int status, int tag)
3658 struct iommu_dev_data *dev_data;
3659 struct amd_iommu *iommu;
3660 struct iommu_cmd cmd;
3662 INC_STATS_COUNTER(complete_ppr);
3664 dev_data = get_dev_data(&pdev->dev);
3665 iommu = amd_iommu_rlookup_table[dev_data->devid];
3667 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3668 tag, dev_data->pri_tlp);
3670 return iommu_queue_command(iommu, &cmd);
3672 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3674 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3676 struct protection_domain *domain;
3678 domain = get_domain(&pdev->dev);
3682 /* Only return IOMMUv2 domains */
3683 if (!(domain->flags & PD_IOMMUV2_MASK))
3686 return domain->iommu_domain;
3688 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3690 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3692 struct iommu_dev_data *dev_data;
3694 if (!amd_iommu_v2_supported())
3697 dev_data = get_dev_data(&pdev->dev);
3698 dev_data->errata |= (1 << erratum);
3700 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3702 int amd_iommu_device_info(struct pci_dev *pdev,
3703 struct amd_iommu_device_info *info)
3708 if (pdev == NULL || info == NULL)
3711 if (!amd_iommu_v2_supported())
3714 memset(info, 0, sizeof(*info));
3716 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3718 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3720 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3722 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3724 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3728 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3729 max_pasids = min(max_pasids, (1 << 20));
3731 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3732 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3734 features = pci_pasid_features(pdev);
3735 if (features & PCI_PASID_CAP_EXEC)
3736 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3737 if (features & PCI_PASID_CAP_PRIV)
3738 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3743 EXPORT_SYMBOL(amd_iommu_device_info);