2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <linux/dma-contiguous.h>
37 #include <linux/irqdomain.h>
38 #include <asm/irq_remapping.h>
39 #include <asm/io_apic.h>
41 #include <asm/hw_irq.h>
42 #include <asm/msidef.h>
43 #include <asm/proto.h>
44 #include <asm/iommu.h>
48 #include "amd_iommu_proto.h"
49 #include "amd_iommu_types.h"
50 #include "irq_remapping.h"
52 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
54 #define LOOP_TIMEOUT 100000
57 * This bitmap is used to advertise the page sizes our hardware support
58 * to the IOMMU core, which will then use this information to split
59 * physically contiguous memory regions it is mapping into page sizes
62 * 512GB Pages are not supported due to a hardware bug
64 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
66 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
68 /* List of all available dev_data structures */
69 static LIST_HEAD(dev_data_list);
70 static DEFINE_SPINLOCK(dev_data_list_lock);
72 LIST_HEAD(ioapic_map);
76 * Domain for untranslated devices - only allocated
77 * if iommu=pt passed on kernel cmd line.
79 static const struct iommu_ops amd_iommu_ops;
81 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
82 int amd_iommu_max_glx_val = -1;
84 static struct dma_map_ops amd_iommu_dma_ops;
87 * This struct contains device specific data for the IOMMU
89 struct iommu_dev_data {
90 struct list_head list; /* For domain->dev_list */
91 struct list_head dev_data_list; /* For global dev_data_list */
92 struct protection_domain *domain; /* Domain the device is bound to */
93 u16 devid; /* PCI Device ID */
94 bool iommu_v2; /* Device can make use of IOMMUv2 */
95 bool passthrough; /* Device is identity mapped */
99 } ats; /* ATS state */
100 bool pri_tlp; /* PASID TLB required for
102 u32 errata; /* Bitmap for errata to apply */
106 * general struct to manage commands send to an IOMMU
112 struct kmem_cache *amd_iommu_irq_cache;
114 static void update_domain(struct protection_domain *domain);
115 static int protection_domain_init(struct protection_domain *domain);
118 * For dynamic growth the aperture size is split into ranges of 128MB of
119 * DMA address space each. This struct represents one such range.
121 struct aperture_range {
123 spinlock_t bitmap_lock;
125 /* address allocation bitmap */
126 unsigned long *bitmap;
127 unsigned long offset;
128 unsigned long next_bit;
131 * Array of PTE pages for the aperture. In this array we save all the
132 * leaf pages of the domain page table used for the aperture. This way
133 * we don't need to walk the page table to find a specific PTE. We can
134 * just calculate its address in constant time.
140 * Data container for a dma_ops specific protection domain
142 struct dma_ops_domain {
143 /* generic protection domain information */
144 struct protection_domain domain;
146 /* size of the aperture for the mappings */
147 unsigned long aperture_size;
149 /* address we start to search for free addresses */
150 unsigned long next_address;
152 /* address space relevant data */
153 struct aperture_range *aperture[APERTURE_MAX_RANGES];
155 /* This will be set to true when TLB needs to be flushed */
159 /****************************************************************************
163 ****************************************************************************/
165 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
167 return container_of(dom, struct protection_domain, domain);
170 static struct iommu_dev_data *alloc_dev_data(u16 devid)
172 struct iommu_dev_data *dev_data;
175 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
179 dev_data->devid = devid;
181 spin_lock_irqsave(&dev_data_list_lock, flags);
182 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
183 spin_unlock_irqrestore(&dev_data_list_lock, flags);
188 static struct iommu_dev_data *search_dev_data(u16 devid)
190 struct iommu_dev_data *dev_data;
193 spin_lock_irqsave(&dev_data_list_lock, flags);
194 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
195 if (dev_data->devid == devid)
202 spin_unlock_irqrestore(&dev_data_list_lock, flags);
207 static struct iommu_dev_data *find_dev_data(u16 devid)
209 struct iommu_dev_data *dev_data;
211 dev_data = search_dev_data(devid);
213 if (dev_data == NULL)
214 dev_data = alloc_dev_data(devid);
219 static inline u16 get_device_id(struct device *dev)
221 struct pci_dev *pdev = to_pci_dev(dev);
223 return PCI_DEVID(pdev->bus->number, pdev->devfn);
226 static struct iommu_dev_data *get_dev_data(struct device *dev)
228 return dev->archdata.iommu;
231 static bool pci_iommuv2_capable(struct pci_dev *pdev)
233 static const int caps[] = {
236 PCI_EXT_CAP_ID_PASID,
240 for (i = 0; i < 3; ++i) {
241 pos = pci_find_ext_capability(pdev, caps[i]);
249 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
251 struct iommu_dev_data *dev_data;
253 dev_data = get_dev_data(&pdev->dev);
255 return dev_data->errata & (1 << erratum) ? true : false;
259 * This function actually applies the mapping to the page table of the
262 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
263 struct unity_map_entry *e)
267 for (addr = e->address_start; addr < e->address_end;
269 if (addr < dma_dom->aperture_size)
270 __set_bit(addr >> PAGE_SHIFT,
271 dma_dom->aperture[0]->bitmap);
276 * Inits the unity mappings required for a specific device
278 static void init_unity_mappings_for_device(struct device *dev,
279 struct dma_ops_domain *dma_dom)
281 struct unity_map_entry *e;
284 devid = get_device_id(dev);
286 list_for_each_entry(e, &amd_iommu_unity_map, list) {
287 if (!(devid >= e->devid_start && devid <= e->devid_end))
289 alloc_unity_mapping(dma_dom, e);
294 * This function checks if the driver got a valid device from the caller to
295 * avoid dereferencing invalid pointers.
297 static bool check_device(struct device *dev)
301 if (!dev || !dev->dma_mask)
305 if (!dev_is_pci(dev))
308 devid = get_device_id(dev);
310 /* Out of our scope? */
311 if (devid > amd_iommu_last_bdf)
314 if (amd_iommu_rlookup_table[devid] == NULL)
320 static void init_iommu_group(struct device *dev)
322 struct dma_ops_domain *dma_domain;
323 struct iommu_domain *domain;
324 struct iommu_group *group;
326 group = iommu_group_get_for_dev(dev);
330 domain = iommu_group_default_domain(group);
334 dma_domain = to_pdomain(domain)->priv;
336 init_unity_mappings_for_device(dev, dma_domain);
338 iommu_group_put(group);
341 static int iommu_init_device(struct device *dev)
343 struct pci_dev *pdev = to_pci_dev(dev);
344 struct iommu_dev_data *dev_data;
346 if (dev->archdata.iommu)
349 dev_data = find_dev_data(get_device_id(dev));
353 if (pci_iommuv2_capable(pdev)) {
354 struct amd_iommu *iommu;
356 iommu = amd_iommu_rlookup_table[dev_data->devid];
357 dev_data->iommu_v2 = iommu->is_iommu_v2;
360 dev->archdata.iommu = dev_data;
362 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
368 static void iommu_ignore_device(struct device *dev)
372 devid = get_device_id(dev);
373 alias = amd_iommu_alias_table[devid];
375 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
376 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
378 amd_iommu_rlookup_table[devid] = NULL;
379 amd_iommu_rlookup_table[alias] = NULL;
382 static void iommu_uninit_device(struct device *dev)
384 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
389 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
392 iommu_group_remove_device(dev);
395 dev->archdata.dma_ops = NULL;
398 * We keep dev_data around for unplugged devices and reuse it when the
399 * device is re-plugged - not doing so would introduce a ton of races.
403 #ifdef CONFIG_AMD_IOMMU_STATS
406 * Initialization code for statistics collection
409 DECLARE_STATS_COUNTER(compl_wait);
410 DECLARE_STATS_COUNTER(cnt_map_single);
411 DECLARE_STATS_COUNTER(cnt_unmap_single);
412 DECLARE_STATS_COUNTER(cnt_map_sg);
413 DECLARE_STATS_COUNTER(cnt_unmap_sg);
414 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
415 DECLARE_STATS_COUNTER(cnt_free_coherent);
416 DECLARE_STATS_COUNTER(cross_page);
417 DECLARE_STATS_COUNTER(domain_flush_single);
418 DECLARE_STATS_COUNTER(domain_flush_all);
419 DECLARE_STATS_COUNTER(alloced_io_mem);
420 DECLARE_STATS_COUNTER(total_map_requests);
421 DECLARE_STATS_COUNTER(complete_ppr);
422 DECLARE_STATS_COUNTER(invalidate_iotlb);
423 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
424 DECLARE_STATS_COUNTER(pri_requests);
426 static struct dentry *stats_dir;
427 static struct dentry *de_fflush;
429 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
431 if (stats_dir == NULL)
434 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
438 static void amd_iommu_stats_init(void)
440 stats_dir = debugfs_create_dir("amd-iommu", NULL);
441 if (stats_dir == NULL)
444 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
445 &amd_iommu_unmap_flush);
447 amd_iommu_stats_add(&compl_wait);
448 amd_iommu_stats_add(&cnt_map_single);
449 amd_iommu_stats_add(&cnt_unmap_single);
450 amd_iommu_stats_add(&cnt_map_sg);
451 amd_iommu_stats_add(&cnt_unmap_sg);
452 amd_iommu_stats_add(&cnt_alloc_coherent);
453 amd_iommu_stats_add(&cnt_free_coherent);
454 amd_iommu_stats_add(&cross_page);
455 amd_iommu_stats_add(&domain_flush_single);
456 amd_iommu_stats_add(&domain_flush_all);
457 amd_iommu_stats_add(&alloced_io_mem);
458 amd_iommu_stats_add(&total_map_requests);
459 amd_iommu_stats_add(&complete_ppr);
460 amd_iommu_stats_add(&invalidate_iotlb);
461 amd_iommu_stats_add(&invalidate_iotlb_all);
462 amd_iommu_stats_add(&pri_requests);
467 /****************************************************************************
469 * Interrupt handling functions
471 ****************************************************************************/
473 static void dump_dte_entry(u16 devid)
477 for (i = 0; i < 4; ++i)
478 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
479 amd_iommu_dev_table[devid].data[i]);
482 static void dump_command(unsigned long phys_addr)
484 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
487 for (i = 0; i < 4; ++i)
488 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
491 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
493 int type, devid, domid, flags;
494 volatile u32 *event = __evt;
499 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
500 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
501 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
502 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
503 address = (u64)(((u64)event[3]) << 32) | event[2];
506 /* Did we hit the erratum? */
507 if (++count == LOOP_TIMEOUT) {
508 pr_err("AMD-Vi: No event written to event log\n");
515 printk(KERN_ERR "AMD-Vi: Event logged [");
518 case EVENT_TYPE_ILL_DEV:
519 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
520 "address=0x%016llx flags=0x%04x]\n",
521 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
523 dump_dte_entry(devid);
525 case EVENT_TYPE_IO_FAULT:
526 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
527 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
528 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
529 domid, address, flags);
531 case EVENT_TYPE_DEV_TAB_ERR:
532 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
533 "address=0x%016llx flags=0x%04x]\n",
534 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
537 case EVENT_TYPE_PAGE_TAB_ERR:
538 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
539 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
540 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
541 domid, address, flags);
543 case EVENT_TYPE_ILL_CMD:
544 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
545 dump_command(address);
547 case EVENT_TYPE_CMD_HARD_ERR:
548 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
549 "flags=0x%04x]\n", address, flags);
551 case EVENT_TYPE_IOTLB_INV_TO:
552 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
553 "address=0x%016llx]\n",
554 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
557 case EVENT_TYPE_INV_DEV_REQ:
558 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
559 "address=0x%016llx flags=0x%04x]\n",
560 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
564 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
567 memset(__evt, 0, 4 * sizeof(u32));
570 static void iommu_poll_events(struct amd_iommu *iommu)
574 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
575 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
577 while (head != tail) {
578 iommu_print_event(iommu, iommu->evt_buf + head);
579 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
582 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
585 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
587 struct amd_iommu_fault fault;
589 INC_STATS_COUNTER(pri_requests);
591 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
592 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
596 fault.address = raw[1];
597 fault.pasid = PPR_PASID(raw[0]);
598 fault.device_id = PPR_DEVID(raw[0]);
599 fault.tag = PPR_TAG(raw[0]);
600 fault.flags = PPR_FLAGS(raw[0]);
602 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
605 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
609 if (iommu->ppr_log == NULL)
612 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
613 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
615 while (head != tail) {
620 raw = (u64 *)(iommu->ppr_log + head);
623 * Hardware bug: Interrupt may arrive before the entry is
624 * written to memory. If this happens we need to wait for the
627 for (i = 0; i < LOOP_TIMEOUT; ++i) {
628 if (PPR_REQ_TYPE(raw[0]) != 0)
633 /* Avoid memcpy function-call overhead */
638 * To detect the hardware bug we need to clear the entry
641 raw[0] = raw[1] = 0UL;
643 /* Update head pointer of hardware ring-buffer */
644 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
645 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
647 /* Handle PPR entry */
648 iommu_handle_ppr_entry(iommu, entry);
650 /* Refresh ring-buffer information */
651 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
652 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
656 irqreturn_t amd_iommu_int_thread(int irq, void *data)
658 struct amd_iommu *iommu = (struct amd_iommu *) data;
659 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
661 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
662 /* Enable EVT and PPR interrupts again */
663 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
664 iommu->mmio_base + MMIO_STATUS_OFFSET);
666 if (status & MMIO_STATUS_EVT_INT_MASK) {
667 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
668 iommu_poll_events(iommu);
671 if (status & MMIO_STATUS_PPR_INT_MASK) {
672 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
673 iommu_poll_ppr_log(iommu);
677 * Hardware bug: ERBT1312
678 * When re-enabling interrupt (by writing 1
679 * to clear the bit), the hardware might also try to set
680 * the interrupt bit in the event status register.
681 * In this scenario, the bit will be set, and disable
682 * subsequent interrupts.
684 * Workaround: The IOMMU driver should read back the
685 * status register and check if the interrupt bits are cleared.
686 * If not, driver will need to go through the interrupt handler
687 * again and re-clear the bits
689 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
694 irqreturn_t amd_iommu_int_handler(int irq, void *data)
696 return IRQ_WAKE_THREAD;
699 /****************************************************************************
701 * IOMMU command queuing functions
703 ****************************************************************************/
705 static int wait_on_sem(volatile u64 *sem)
709 while (*sem == 0 && i < LOOP_TIMEOUT) {
714 if (i == LOOP_TIMEOUT) {
715 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
722 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
723 struct iommu_cmd *cmd,
728 target = iommu->cmd_buf + tail;
729 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
731 /* Copy command to buffer */
732 memcpy(target, cmd, sizeof(*cmd));
734 /* Tell the IOMMU about it */
735 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
738 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
740 WARN_ON(address & 0x7ULL);
742 memset(cmd, 0, sizeof(*cmd));
743 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
744 cmd->data[1] = upper_32_bits(__pa(address));
746 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
749 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
751 memset(cmd, 0, sizeof(*cmd));
752 cmd->data[0] = devid;
753 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
756 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
757 size_t size, u16 domid, int pde)
762 pages = iommu_num_pages(address, size, PAGE_SIZE);
767 * If we have to flush more than one page, flush all
768 * TLB entries for this domain
770 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
774 address &= PAGE_MASK;
776 memset(cmd, 0, sizeof(*cmd));
777 cmd->data[1] |= domid;
778 cmd->data[2] = lower_32_bits(address);
779 cmd->data[3] = upper_32_bits(address);
780 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
781 if (s) /* size bit - we flush more than one 4kb page */
782 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
783 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
784 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
787 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
788 u64 address, size_t size)
793 pages = iommu_num_pages(address, size, PAGE_SIZE);
798 * If we have to flush more than one page, flush all
799 * TLB entries for this domain
801 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
805 address &= PAGE_MASK;
807 memset(cmd, 0, sizeof(*cmd));
808 cmd->data[0] = devid;
809 cmd->data[0] |= (qdep & 0xff) << 24;
810 cmd->data[1] = devid;
811 cmd->data[2] = lower_32_bits(address);
812 cmd->data[3] = upper_32_bits(address);
813 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
815 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
818 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
819 u64 address, bool size)
821 memset(cmd, 0, sizeof(*cmd));
823 address &= ~(0xfffULL);
825 cmd->data[0] = pasid;
826 cmd->data[1] = domid;
827 cmd->data[2] = lower_32_bits(address);
828 cmd->data[3] = upper_32_bits(address);
829 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
830 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
832 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
833 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
836 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
837 int qdep, u64 address, bool size)
839 memset(cmd, 0, sizeof(*cmd));
841 address &= ~(0xfffULL);
843 cmd->data[0] = devid;
844 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
845 cmd->data[0] |= (qdep & 0xff) << 24;
846 cmd->data[1] = devid;
847 cmd->data[1] |= (pasid & 0xff) << 16;
848 cmd->data[2] = lower_32_bits(address);
849 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
850 cmd->data[3] = upper_32_bits(address);
852 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
853 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
856 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
857 int status, int tag, bool gn)
859 memset(cmd, 0, sizeof(*cmd));
861 cmd->data[0] = devid;
863 cmd->data[1] = pasid;
864 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
866 cmd->data[3] = tag & 0x1ff;
867 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
869 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
872 static void build_inv_all(struct iommu_cmd *cmd)
874 memset(cmd, 0, sizeof(*cmd));
875 CMD_SET_TYPE(cmd, CMD_INV_ALL);
878 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
880 memset(cmd, 0, sizeof(*cmd));
881 cmd->data[0] = devid;
882 CMD_SET_TYPE(cmd, CMD_INV_IRT);
886 * Writes the command to the IOMMUs command buffer and informs the
887 * hardware about the new command.
889 static int iommu_queue_command_sync(struct amd_iommu *iommu,
890 struct iommu_cmd *cmd,
893 u32 left, tail, head, next_tail;
897 spin_lock_irqsave(&iommu->lock, flags);
899 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
900 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
901 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
902 left = (head - next_tail) % CMD_BUFFER_SIZE;
905 struct iommu_cmd sync_cmd;
906 volatile u64 sem = 0;
909 build_completion_wait(&sync_cmd, (u64)&sem);
910 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
912 spin_unlock_irqrestore(&iommu->lock, flags);
914 if ((ret = wait_on_sem(&sem)) != 0)
920 copy_cmd_to_buffer(iommu, cmd, tail);
922 /* We need to sync now to make sure all commands are processed */
923 iommu->need_sync = sync;
925 spin_unlock_irqrestore(&iommu->lock, flags);
930 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
932 return iommu_queue_command_sync(iommu, cmd, true);
936 * This function queues a completion wait command into the command
939 static int iommu_completion_wait(struct amd_iommu *iommu)
941 struct iommu_cmd cmd;
942 volatile u64 sem = 0;
945 if (!iommu->need_sync)
948 build_completion_wait(&cmd, (u64)&sem);
950 ret = iommu_queue_command_sync(iommu, &cmd, false);
954 return wait_on_sem(&sem);
957 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
959 struct iommu_cmd cmd;
961 build_inv_dte(&cmd, devid);
963 return iommu_queue_command(iommu, &cmd);
966 static void iommu_flush_dte_all(struct amd_iommu *iommu)
970 for (devid = 0; devid <= 0xffff; ++devid)
971 iommu_flush_dte(iommu, devid);
973 iommu_completion_wait(iommu);
977 * This function uses heavy locking and may disable irqs for some time. But
978 * this is no issue because it is only called during resume.
980 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
984 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
985 struct iommu_cmd cmd;
986 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
988 iommu_queue_command(iommu, &cmd);
991 iommu_completion_wait(iommu);
994 static void iommu_flush_all(struct amd_iommu *iommu)
996 struct iommu_cmd cmd;
1000 iommu_queue_command(iommu, &cmd);
1001 iommu_completion_wait(iommu);
1004 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1006 struct iommu_cmd cmd;
1008 build_inv_irt(&cmd, devid);
1010 iommu_queue_command(iommu, &cmd);
1013 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1017 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1018 iommu_flush_irt(iommu, devid);
1020 iommu_completion_wait(iommu);
1023 void iommu_flush_all_caches(struct amd_iommu *iommu)
1025 if (iommu_feature(iommu, FEATURE_IA)) {
1026 iommu_flush_all(iommu);
1028 iommu_flush_dte_all(iommu);
1029 iommu_flush_irt_all(iommu);
1030 iommu_flush_tlb_all(iommu);
1035 * Command send function for flushing on-device TLB
1037 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1038 u64 address, size_t size)
1040 struct amd_iommu *iommu;
1041 struct iommu_cmd cmd;
1044 qdep = dev_data->ats.qdep;
1045 iommu = amd_iommu_rlookup_table[dev_data->devid];
1047 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1049 return iommu_queue_command(iommu, &cmd);
1053 * Command send function for invalidating a device table entry
1055 static int device_flush_dte(struct iommu_dev_data *dev_data)
1057 struct amd_iommu *iommu;
1061 iommu = amd_iommu_rlookup_table[dev_data->devid];
1062 alias = amd_iommu_alias_table[dev_data->devid];
1064 ret = iommu_flush_dte(iommu, dev_data->devid);
1065 if (!ret && alias != dev_data->devid)
1066 ret = iommu_flush_dte(iommu, alias);
1070 if (dev_data->ats.enabled)
1071 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1077 * TLB invalidation function which is called from the mapping functions.
1078 * It invalidates a single PTE if the range to flush is within a single
1079 * page. Otherwise it flushes the whole TLB of the IOMMU.
1081 static void __domain_flush_pages(struct protection_domain *domain,
1082 u64 address, size_t size, int pde)
1084 struct iommu_dev_data *dev_data;
1085 struct iommu_cmd cmd;
1088 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1090 for (i = 0; i < amd_iommus_present; ++i) {
1091 if (!domain->dev_iommu[i])
1095 * Devices of this domain are behind this IOMMU
1096 * We need a TLB flush
1098 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1101 list_for_each_entry(dev_data, &domain->dev_list, list) {
1103 if (!dev_data->ats.enabled)
1106 ret |= device_flush_iotlb(dev_data, address, size);
1112 static void domain_flush_pages(struct protection_domain *domain,
1113 u64 address, size_t size)
1115 __domain_flush_pages(domain, address, size, 0);
1118 /* Flush the whole IO/TLB for a given protection domain */
1119 static void domain_flush_tlb(struct protection_domain *domain)
1121 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1124 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1125 static void domain_flush_tlb_pde(struct protection_domain *domain)
1127 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1130 static void domain_flush_complete(struct protection_domain *domain)
1134 for (i = 0; i < amd_iommus_present; ++i) {
1135 if (!domain->dev_iommu[i])
1139 * Devices of this domain are behind this IOMMU
1140 * We need to wait for completion of all commands.
1142 iommu_completion_wait(amd_iommus[i]);
1148 * This function flushes the DTEs for all devices in domain
1150 static void domain_flush_devices(struct protection_domain *domain)
1152 struct iommu_dev_data *dev_data;
1154 list_for_each_entry(dev_data, &domain->dev_list, list)
1155 device_flush_dte(dev_data);
1158 /****************************************************************************
1160 * The functions below are used the create the page table mappings for
1161 * unity mapped regions.
1163 ****************************************************************************/
1166 * This function is used to add another level to an IO page table. Adding
1167 * another level increases the size of the address space by 9 bits to a size up
1170 static bool increase_address_space(struct protection_domain *domain,
1175 if (domain->mode == PAGE_MODE_6_LEVEL)
1176 /* address space already 64 bit large */
1179 pte = (void *)get_zeroed_page(gfp);
1183 *pte = PM_LEVEL_PDE(domain->mode,
1184 virt_to_phys(domain->pt_root));
1185 domain->pt_root = pte;
1187 domain->updated = true;
1192 static u64 *alloc_pte(struct protection_domain *domain,
1193 unsigned long address,
1194 unsigned long page_size,
1201 BUG_ON(!is_power_of_2(page_size));
1203 while (address > PM_LEVEL_SIZE(domain->mode))
1204 increase_address_space(domain, gfp);
1206 level = domain->mode - 1;
1207 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1208 address = PAGE_SIZE_ALIGN(address, page_size);
1209 end_lvl = PAGE_SIZE_LEVEL(page_size);
1211 while (level > end_lvl) {
1212 if (!IOMMU_PTE_PRESENT(*pte)) {
1213 page = (u64 *)get_zeroed_page(gfp);
1216 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1219 /* No level skipping support yet */
1220 if (PM_PTE_LEVEL(*pte) != level)
1225 pte = IOMMU_PTE_PAGE(*pte);
1227 if (pte_page && level == end_lvl)
1230 pte = &pte[PM_LEVEL_INDEX(level, address)];
1237 * This function checks if there is a PTE for a given dma address. If
1238 * there is one, it returns the pointer to it.
1240 static u64 *fetch_pte(struct protection_domain *domain,
1241 unsigned long address,
1242 unsigned long *page_size)
1247 if (address > PM_LEVEL_SIZE(domain->mode))
1250 level = domain->mode - 1;
1251 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1252 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1257 if (!IOMMU_PTE_PRESENT(*pte))
1261 if (PM_PTE_LEVEL(*pte) == 7 ||
1262 PM_PTE_LEVEL(*pte) == 0)
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)];
1274 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1277 if (PM_PTE_LEVEL(*pte) == 0x07) {
1278 unsigned long pte_mask;
1281 * If we have a series of large PTEs, make
1282 * sure to return a pointer to the first one.
1284 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1285 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1286 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1293 * Generic mapping functions. It maps a physical address into a DMA
1294 * address space. It allocates the page table pages if necessary.
1295 * In the future it can be extended to a generic mapping function
1296 * supporting all features of AMD IOMMU page tables like level skipping
1297 * and full 64 bit address spaces.
1299 static int iommu_map_page(struct protection_domain *dom,
1300 unsigned long bus_addr,
1301 unsigned long phys_addr,
1303 unsigned long page_size)
1308 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1309 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1311 if (!(prot & IOMMU_PROT_MASK))
1314 count = PAGE_SIZE_PTE_COUNT(page_size);
1315 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1320 for (i = 0; i < count; ++i)
1321 if (IOMMU_PTE_PRESENT(pte[i]))
1325 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1326 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1328 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1330 if (prot & IOMMU_PROT_IR)
1331 __pte |= IOMMU_PTE_IR;
1332 if (prot & IOMMU_PROT_IW)
1333 __pte |= IOMMU_PTE_IW;
1335 for (i = 0; i < count; ++i)
1343 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1344 unsigned long bus_addr,
1345 unsigned long page_size)
1347 unsigned long long unmapped;
1348 unsigned long unmap_size;
1351 BUG_ON(!is_power_of_2(page_size));
1355 while (unmapped < page_size) {
1357 pte = fetch_pte(dom, bus_addr, &unmap_size);
1362 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1363 for (i = 0; i < count; i++)
1367 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1368 unmapped += unmap_size;
1371 BUG_ON(unmapped && !is_power_of_2(unmapped));
1376 /****************************************************************************
1378 * The next functions belong to the address allocator for the dma_ops
1379 * interface functions. They work like the allocators in the other IOMMU
1380 * drivers. Its basically a bitmap which marks the allocated pages in
1381 * the aperture. Maybe it could be enhanced in the future to a more
1382 * efficient allocator.
1384 ****************************************************************************/
1387 * The address allocator core functions.
1389 * called with domain->lock held
1393 * Used to reserve address ranges in the aperture (e.g. for exclusion
1396 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1397 unsigned long start_page,
1400 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1402 if (start_page + pages > last_page)
1403 pages = last_page - start_page;
1405 for (i = start_page; i < start_page + pages; ++i) {
1406 int index = i / APERTURE_RANGE_PAGES;
1407 int page = i % APERTURE_RANGE_PAGES;
1408 __set_bit(page, dom->aperture[index]->bitmap);
1413 * This function is used to add a new aperture range to an existing
1414 * aperture in case of dma_ops domain allocation or address allocation
1417 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1418 bool populate, gfp_t gfp)
1420 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1421 struct amd_iommu *iommu;
1422 unsigned long i, old_size, pte_pgsize;
1424 #ifdef CONFIG_IOMMU_STRESS
1428 if (index >= APERTURE_MAX_RANGES)
1431 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1432 if (!dma_dom->aperture[index])
1435 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1436 if (!dma_dom->aperture[index]->bitmap)
1439 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1441 spin_lock_init(&dma_dom->aperture[index]->bitmap_lock);
1444 unsigned long address = dma_dom->aperture_size;
1445 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1446 u64 *pte, *pte_page;
1448 for (i = 0; i < num_ptes; ++i) {
1449 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1454 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1456 address += APERTURE_RANGE_SIZE / 64;
1460 old_size = dma_dom->aperture_size;
1461 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1463 /* Reserve address range used for MSI messages */
1464 if (old_size < MSI_ADDR_BASE_LO &&
1465 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1466 unsigned long spage;
1469 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1470 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1472 dma_ops_reserve_addresses(dma_dom, spage, pages);
1475 /* Initialize the exclusion range if necessary */
1476 for_each_iommu(iommu) {
1477 if (iommu->exclusion_start &&
1478 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1479 && iommu->exclusion_start < dma_dom->aperture_size) {
1480 unsigned long startpage;
1481 int pages = iommu_num_pages(iommu->exclusion_start,
1482 iommu->exclusion_length,
1484 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1485 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1490 * Check for areas already mapped as present in the new aperture
1491 * range and mark those pages as reserved in the allocator. Such
1492 * mappings may already exist as a result of requested unity
1493 * mappings for devices.
1495 for (i = dma_dom->aperture[index]->offset;
1496 i < dma_dom->aperture_size;
1498 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1499 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1502 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1506 update_domain(&dma_dom->domain);
1511 update_domain(&dma_dom->domain);
1513 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1515 kfree(dma_dom->aperture[index]);
1516 dma_dom->aperture[index] = NULL;
1521 static dma_addr_t dma_ops_aperture_alloc(struct dma_ops_domain *dom,
1522 struct aperture_range *range,
1523 unsigned long pages,
1524 unsigned long dma_mask,
1525 unsigned long boundary_size,
1526 unsigned long align_mask)
1528 unsigned long offset, limit, flags;
1532 offset = range->offset >> PAGE_SHIFT;
1533 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1534 dma_mask >> PAGE_SHIFT);
1536 spin_lock_irqsave(&range->bitmap_lock, flags);
1537 address = iommu_area_alloc(range->bitmap, limit, range->next_bit,
1538 pages, offset, boundary_size, align_mask);
1539 if (address == -1) {
1540 /* Nothing found, retry one time */
1541 address = iommu_area_alloc(range->bitmap, limit,
1542 0, pages, offset, boundary_size,
1548 range->next_bit = address + pages;
1550 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1553 domain_flush_tlb(&dom->domain);
1554 domain_flush_complete(&dom->domain);
1560 static unsigned long dma_ops_area_alloc(struct device *dev,
1561 struct dma_ops_domain *dom,
1563 unsigned long align_mask,
1566 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1567 int i = dom->next_address >> APERTURE_RANGE_SHIFT;
1568 unsigned long next_bit, boundary_size, mask;
1569 unsigned long address = -1;
1571 mask = dma_get_seg_boundary(dev);
1573 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1574 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1576 for (;i < max_index; ++i) {
1577 struct aperture_range *range = dom->aperture[i];
1579 if (range->offset >= dma_mask)
1582 next_bit = range->next_bit;
1584 address = dma_ops_aperture_alloc(dom, dom->aperture[i], pages,
1585 dma_mask, boundary_size,
1587 if (address != -1) {
1588 address = dom->aperture[i]->offset +
1589 (address << PAGE_SHIFT);
1590 dom->next_address = address + (pages << PAGE_SHIFT);
1594 if (next_bit > dom->aperture[i]->next_bit)
1595 dom->need_flush = true;
1601 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1602 struct dma_ops_domain *dom,
1604 unsigned long align_mask,
1607 unsigned long address;
1609 #ifdef CONFIG_IOMMU_STRESS
1610 dom->next_address = 0;
1611 dom->need_flush = true;
1614 address = dma_ops_area_alloc(dev, dom, pages, align_mask, dma_mask);
1616 if (address == -1) {
1617 dom->next_address = 0;
1618 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1620 dom->need_flush = true;
1623 if (unlikely(address == -1))
1624 address = DMA_ERROR_CODE;
1626 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1632 * The address free function.
1634 * called with domain->lock held
1636 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1637 unsigned long address,
1640 unsigned i = address >> APERTURE_RANGE_SHIFT;
1641 struct aperture_range *range = dom->aperture[i];
1642 unsigned long flags;
1644 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1646 #ifdef CONFIG_IOMMU_STRESS
1651 if (address >= dom->next_address)
1652 dom->need_flush = true;
1654 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1656 spin_lock_irqsave(&range->bitmap_lock, flags);
1657 bitmap_clear(range->bitmap, address, pages);
1658 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1662 /****************************************************************************
1664 * The next functions belong to the domain allocation. A domain is
1665 * allocated for every IOMMU as the default domain. If device isolation
1666 * is enabled, every device get its own domain. The most important thing
1667 * about domains is the page table mapping the DMA address space they
1670 ****************************************************************************/
1673 * This function adds a protection domain to the global protection domain list
1675 static void add_domain_to_list(struct protection_domain *domain)
1677 unsigned long flags;
1679 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1680 list_add(&domain->list, &amd_iommu_pd_list);
1681 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1685 * This function removes a protection domain to the global
1686 * protection domain list
1688 static void del_domain_from_list(struct protection_domain *domain)
1690 unsigned long flags;
1692 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1693 list_del(&domain->list);
1694 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1697 static u16 domain_id_alloc(void)
1699 unsigned long flags;
1702 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1703 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1705 if (id > 0 && id < MAX_DOMAIN_ID)
1706 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1709 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1714 static void domain_id_free(int id)
1716 unsigned long flags;
1718 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1719 if (id > 0 && id < MAX_DOMAIN_ID)
1720 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1721 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1724 #define DEFINE_FREE_PT_FN(LVL, FN) \
1725 static void free_pt_##LVL (unsigned long __pt) \
1733 for (i = 0; i < 512; ++i) { \
1734 /* PTE present? */ \
1735 if (!IOMMU_PTE_PRESENT(pt[i])) \
1739 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1740 PM_PTE_LEVEL(pt[i]) == 7) \
1743 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1746 free_page((unsigned long)pt); \
1749 DEFINE_FREE_PT_FN(l2, free_page)
1750 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1751 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1752 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1753 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1755 static void free_pagetable(struct protection_domain *domain)
1757 unsigned long root = (unsigned long)domain->pt_root;
1759 switch (domain->mode) {
1760 case PAGE_MODE_NONE:
1762 case PAGE_MODE_1_LEVEL:
1765 case PAGE_MODE_2_LEVEL:
1768 case PAGE_MODE_3_LEVEL:
1771 case PAGE_MODE_4_LEVEL:
1774 case PAGE_MODE_5_LEVEL:
1777 case PAGE_MODE_6_LEVEL:
1785 static void free_gcr3_tbl_level1(u64 *tbl)
1790 for (i = 0; i < 512; ++i) {
1791 if (!(tbl[i] & GCR3_VALID))
1794 ptr = __va(tbl[i] & PAGE_MASK);
1796 free_page((unsigned long)ptr);
1800 static void free_gcr3_tbl_level2(u64 *tbl)
1805 for (i = 0; i < 512; ++i) {
1806 if (!(tbl[i] & GCR3_VALID))
1809 ptr = __va(tbl[i] & PAGE_MASK);
1811 free_gcr3_tbl_level1(ptr);
1815 static void free_gcr3_table(struct protection_domain *domain)
1817 if (domain->glx == 2)
1818 free_gcr3_tbl_level2(domain->gcr3_tbl);
1819 else if (domain->glx == 1)
1820 free_gcr3_tbl_level1(domain->gcr3_tbl);
1822 BUG_ON(domain->glx != 0);
1824 free_page((unsigned long)domain->gcr3_tbl);
1828 * Free a domain, only used if something went wrong in the
1829 * allocation path and we need to free an already allocated page table
1831 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1838 del_domain_from_list(&dom->domain);
1840 free_pagetable(&dom->domain);
1842 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1843 if (!dom->aperture[i])
1845 free_page((unsigned long)dom->aperture[i]->bitmap);
1846 kfree(dom->aperture[i]);
1853 * Allocates a new protection domain usable for the dma_ops functions.
1854 * It also initializes the page table and the address allocator data
1855 * structures required for the dma_ops interface
1857 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1859 struct dma_ops_domain *dma_dom;
1861 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1865 if (protection_domain_init(&dma_dom->domain))
1868 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1869 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1870 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1871 dma_dom->domain.priv = dma_dom;
1872 if (!dma_dom->domain.pt_root)
1875 dma_dom->need_flush = false;
1877 add_domain_to_list(&dma_dom->domain);
1879 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1883 * mark the first page as allocated so we never return 0 as
1884 * a valid dma-address. So we can use 0 as error value
1886 dma_dom->aperture[0]->bitmap[0] = 1;
1887 dma_dom->next_address = 0;
1893 dma_ops_domain_free(dma_dom);
1899 * little helper function to check whether a given protection domain is a
1902 static bool dma_ops_domain(struct protection_domain *domain)
1904 return domain->flags & PD_DMA_OPS_MASK;
1907 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1912 if (domain->mode != PAGE_MODE_NONE)
1913 pte_root = virt_to_phys(domain->pt_root);
1915 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1916 << DEV_ENTRY_MODE_SHIFT;
1917 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1919 flags = amd_iommu_dev_table[devid].data[1];
1922 flags |= DTE_FLAG_IOTLB;
1924 if (domain->flags & PD_IOMMUV2_MASK) {
1925 u64 gcr3 = __pa(domain->gcr3_tbl);
1926 u64 glx = domain->glx;
1929 pte_root |= DTE_FLAG_GV;
1930 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1932 /* First mask out possible old values for GCR3 table */
1933 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1936 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1939 /* Encode GCR3 table into DTE */
1940 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1943 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1946 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1950 flags &= ~(0xffffUL);
1951 flags |= domain->id;
1953 amd_iommu_dev_table[devid].data[1] = flags;
1954 amd_iommu_dev_table[devid].data[0] = pte_root;
1957 static void clear_dte_entry(u16 devid)
1959 /* remove entry from the device table seen by the hardware */
1960 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1961 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1963 amd_iommu_apply_erratum_63(devid);
1966 static void do_attach(struct iommu_dev_data *dev_data,
1967 struct protection_domain *domain)
1969 struct amd_iommu *iommu;
1973 iommu = amd_iommu_rlookup_table[dev_data->devid];
1974 alias = amd_iommu_alias_table[dev_data->devid];
1975 ats = dev_data->ats.enabled;
1977 /* Update data structures */
1978 dev_data->domain = domain;
1979 list_add(&dev_data->list, &domain->dev_list);
1981 /* Do reference counting */
1982 domain->dev_iommu[iommu->index] += 1;
1983 domain->dev_cnt += 1;
1985 /* Update device table */
1986 set_dte_entry(dev_data->devid, domain, ats);
1987 if (alias != dev_data->devid)
1988 set_dte_entry(dev_data->devid, domain, ats);
1990 device_flush_dte(dev_data);
1993 static void do_detach(struct iommu_dev_data *dev_data)
1995 struct amd_iommu *iommu;
1999 * First check if the device is still attached. It might already
2000 * be detached from its domain because the generic
2001 * iommu_detach_group code detached it and we try again here in
2002 * our alias handling.
2004 if (!dev_data->domain)
2007 iommu = amd_iommu_rlookup_table[dev_data->devid];
2008 alias = amd_iommu_alias_table[dev_data->devid];
2010 /* decrease reference counters */
2011 dev_data->domain->dev_iommu[iommu->index] -= 1;
2012 dev_data->domain->dev_cnt -= 1;
2014 /* Update data structures */
2015 dev_data->domain = NULL;
2016 list_del(&dev_data->list);
2017 clear_dte_entry(dev_data->devid);
2018 if (alias != dev_data->devid)
2019 clear_dte_entry(alias);
2021 /* Flush the DTE entry */
2022 device_flush_dte(dev_data);
2026 * If a device is not yet associated with a domain, this function does
2027 * assigns it visible for the hardware
2029 static int __attach_device(struct iommu_dev_data *dev_data,
2030 struct protection_domain *domain)
2035 * Must be called with IRQs disabled. Warn here to detect early
2038 WARN_ON(!irqs_disabled());
2041 spin_lock(&domain->lock);
2044 if (dev_data->domain != NULL)
2047 /* Attach alias group root */
2048 do_attach(dev_data, domain);
2055 spin_unlock(&domain->lock);
2061 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2063 pci_disable_ats(pdev);
2064 pci_disable_pri(pdev);
2065 pci_disable_pasid(pdev);
2068 /* FIXME: Change generic reset-function to do the same */
2069 static int pri_reset_while_enabled(struct pci_dev *pdev)
2074 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2078 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2079 control |= PCI_PRI_CTRL_RESET;
2080 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2085 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2090 /* FIXME: Hardcode number of outstanding requests for now */
2092 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2094 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2096 /* Only allow access to user-accessible pages */
2097 ret = pci_enable_pasid(pdev, 0);
2101 /* First reset the PRI state of the device */
2102 ret = pci_reset_pri(pdev);
2107 ret = pci_enable_pri(pdev, reqs);
2112 ret = pri_reset_while_enabled(pdev);
2117 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2124 pci_disable_pri(pdev);
2125 pci_disable_pasid(pdev);
2130 /* FIXME: Move this to PCI code */
2131 #define PCI_PRI_TLP_OFF (1 << 15)
2133 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2138 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2142 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2144 return (status & PCI_PRI_TLP_OFF) ? true : false;
2148 * If a device is not yet associated with a domain, this function
2149 * assigns it visible for the hardware
2151 static int attach_device(struct device *dev,
2152 struct protection_domain *domain)
2154 struct pci_dev *pdev = to_pci_dev(dev);
2155 struct iommu_dev_data *dev_data;
2156 unsigned long flags;
2159 dev_data = get_dev_data(dev);
2161 if (domain->flags & PD_IOMMUV2_MASK) {
2162 if (!dev_data->passthrough)
2165 if (dev_data->iommu_v2) {
2166 if (pdev_iommuv2_enable(pdev) != 0)
2169 dev_data->ats.enabled = true;
2170 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2171 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2173 } else if (amd_iommu_iotlb_sup &&
2174 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2175 dev_data->ats.enabled = true;
2176 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2179 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2180 ret = __attach_device(dev_data, domain);
2181 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2184 * We might boot into a crash-kernel here. The crashed kernel
2185 * left the caches in the IOMMU dirty. So we have to flush
2186 * here to evict all dirty stuff.
2188 domain_flush_tlb_pde(domain);
2194 * Removes a device from a protection domain (unlocked)
2196 static void __detach_device(struct iommu_dev_data *dev_data)
2198 struct protection_domain *domain;
2201 * Must be called with IRQs disabled. Warn here to detect early
2204 WARN_ON(!irqs_disabled());
2206 if (WARN_ON(!dev_data->domain))
2209 domain = dev_data->domain;
2211 spin_lock(&domain->lock);
2213 do_detach(dev_data);
2215 spin_unlock(&domain->lock);
2219 * Removes a device from a protection domain (with devtable_lock held)
2221 static void detach_device(struct device *dev)
2223 struct protection_domain *domain;
2224 struct iommu_dev_data *dev_data;
2225 unsigned long flags;
2227 dev_data = get_dev_data(dev);
2228 domain = dev_data->domain;
2230 /* lock device table */
2231 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2232 __detach_device(dev_data);
2233 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2235 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2236 pdev_iommuv2_disable(to_pci_dev(dev));
2237 else if (dev_data->ats.enabled)
2238 pci_disable_ats(to_pci_dev(dev));
2240 dev_data->ats.enabled = false;
2243 static int amd_iommu_add_device(struct device *dev)
2245 struct iommu_dev_data *dev_data;
2246 struct iommu_domain *domain;
2247 struct amd_iommu *iommu;
2251 if (!check_device(dev) || get_dev_data(dev))
2254 devid = get_device_id(dev);
2255 iommu = amd_iommu_rlookup_table[devid];
2257 ret = iommu_init_device(dev);
2259 if (ret != -ENOTSUPP)
2260 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2263 iommu_ignore_device(dev);
2264 dev->archdata.dma_ops = &nommu_dma_ops;
2267 init_iommu_group(dev);
2269 dev_data = get_dev_data(dev);
2273 if (iommu_pass_through || dev_data->iommu_v2)
2274 iommu_request_dm_for_dev(dev);
2276 /* Domains are initialized for this device - have a look what we ended up with */
2277 domain = iommu_get_domain_for_dev(dev);
2278 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2279 dev_data->passthrough = true;
2281 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2284 iommu_completion_wait(iommu);
2289 static void amd_iommu_remove_device(struct device *dev)
2291 struct amd_iommu *iommu;
2294 if (!check_device(dev))
2297 devid = get_device_id(dev);
2298 iommu = amd_iommu_rlookup_table[devid];
2300 iommu_uninit_device(dev);
2301 iommu_completion_wait(iommu);
2304 /*****************************************************************************
2306 * The next functions belong to the dma_ops mapping/unmapping code.
2308 *****************************************************************************/
2311 * In the dma_ops path we only have the struct device. This function
2312 * finds the corresponding IOMMU, the protection domain and the
2313 * requestor id for a given device.
2314 * If the device is not yet associated with a domain this is also done
2317 static struct protection_domain *get_domain(struct device *dev)
2319 struct protection_domain *domain;
2320 struct iommu_domain *io_domain;
2322 if (!check_device(dev))
2323 return ERR_PTR(-EINVAL);
2325 io_domain = iommu_get_domain_for_dev(dev);
2329 domain = to_pdomain(io_domain);
2330 if (!dma_ops_domain(domain))
2331 return ERR_PTR(-EBUSY);
2336 static void update_device_table(struct protection_domain *domain)
2338 struct iommu_dev_data *dev_data;
2340 list_for_each_entry(dev_data, &domain->dev_list, list)
2341 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2344 static void update_domain(struct protection_domain *domain)
2346 if (!domain->updated)
2349 update_device_table(domain);
2351 domain_flush_devices(domain);
2352 domain_flush_tlb_pde(domain);
2354 domain->updated = false;
2358 * This function fetches the PTE for a given address in the aperture
2360 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2361 unsigned long address)
2363 struct aperture_range *aperture;
2364 u64 *pte, *pte_page;
2366 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2370 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2372 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2374 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2376 pte += PM_LEVEL_INDEX(0, address);
2378 update_domain(&dom->domain);
2384 * This is the generic map function. It maps one 4kb page at paddr to
2385 * the given address in the DMA address space for the domain.
2387 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2388 unsigned long address,
2394 WARN_ON(address > dom->aperture_size);
2398 pte = dma_ops_get_pte(dom, address);
2400 return DMA_ERROR_CODE;
2402 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2404 if (direction == DMA_TO_DEVICE)
2405 __pte |= IOMMU_PTE_IR;
2406 else if (direction == DMA_FROM_DEVICE)
2407 __pte |= IOMMU_PTE_IW;
2408 else if (direction == DMA_BIDIRECTIONAL)
2409 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2415 return (dma_addr_t)address;
2419 * The generic unmapping function for on page in the DMA address space.
2421 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2422 unsigned long address)
2424 struct aperture_range *aperture;
2427 if (address >= dom->aperture_size)
2430 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2434 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2438 pte += PM_LEVEL_INDEX(0, address);
2440 WARN_ON_ONCE(!*pte);
2446 * This function contains common code for mapping of a physically
2447 * contiguous memory region into DMA address space. It is used by all
2448 * mapping functions provided with this IOMMU driver.
2449 * Must be called with the domain lock held.
2451 static dma_addr_t __map_single(struct device *dev,
2452 struct dma_ops_domain *dma_dom,
2459 dma_addr_t offset = paddr & ~PAGE_MASK;
2460 dma_addr_t address, start, ret;
2462 unsigned long align_mask = 0;
2465 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2468 INC_STATS_COUNTER(total_map_requests);
2471 INC_STATS_COUNTER(cross_page);
2474 align_mask = (1UL << get_order(size)) - 1;
2477 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2479 if (unlikely(address == DMA_ERROR_CODE)) {
2481 * setting next_address here will let the address
2482 * allocator only scan the new allocated range in the
2483 * first run. This is a small optimization.
2485 dma_dom->next_address = dma_dom->aperture_size;
2487 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2491 * aperture was successfully enlarged by 128 MB, try
2498 for (i = 0; i < pages; ++i) {
2499 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2500 if (ret == DMA_ERROR_CODE)
2508 ADD_STATS_COUNTER(alloced_io_mem, size);
2510 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2511 domain_flush_tlb(&dma_dom->domain);
2512 dma_dom->need_flush = false;
2513 } else if (unlikely(amd_iommu_np_cache))
2514 domain_flush_pages(&dma_dom->domain, address, size);
2521 for (--i; i >= 0; --i) {
2523 dma_ops_domain_unmap(dma_dom, start);
2526 domain_flush_pages(&dma_dom->domain, address, size);
2528 dma_ops_free_addresses(dma_dom, address, pages);
2530 return DMA_ERROR_CODE;
2534 * Does the reverse of the __map_single function. Must be called with
2535 * the domain lock held too
2537 static void __unmap_single(struct dma_ops_domain *dma_dom,
2538 dma_addr_t dma_addr,
2542 dma_addr_t flush_addr;
2543 dma_addr_t i, start;
2546 if ((dma_addr == DMA_ERROR_CODE) ||
2547 (dma_addr + size > dma_dom->aperture_size))
2550 flush_addr = dma_addr;
2551 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2552 dma_addr &= PAGE_MASK;
2555 for (i = 0; i < pages; ++i) {
2556 dma_ops_domain_unmap(dma_dom, start);
2560 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2561 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2562 dma_dom->need_flush = false;
2565 SUB_STATS_COUNTER(alloced_io_mem, size);
2567 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2571 * The exported map_single function for dma_ops.
2573 static dma_addr_t map_page(struct device *dev, struct page *page,
2574 unsigned long offset, size_t size,
2575 enum dma_data_direction dir,
2576 struct dma_attrs *attrs)
2578 unsigned long flags;
2579 struct protection_domain *domain;
2582 phys_addr_t paddr = page_to_phys(page) + offset;
2584 INC_STATS_COUNTER(cnt_map_single);
2586 domain = get_domain(dev);
2587 if (PTR_ERR(domain) == -EINVAL)
2588 return (dma_addr_t)paddr;
2589 else if (IS_ERR(domain))
2590 return DMA_ERROR_CODE;
2592 dma_mask = *dev->dma_mask;
2594 spin_lock_irqsave(&domain->lock, flags);
2596 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2598 if (addr == DMA_ERROR_CODE)
2601 domain_flush_complete(domain);
2604 spin_unlock_irqrestore(&domain->lock, flags);
2610 * The exported unmap_single function for dma_ops.
2612 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2613 enum dma_data_direction dir, struct dma_attrs *attrs)
2615 unsigned long flags;
2616 struct protection_domain *domain;
2618 INC_STATS_COUNTER(cnt_unmap_single);
2620 domain = get_domain(dev);
2624 spin_lock_irqsave(&domain->lock, flags);
2626 __unmap_single(domain->priv, dma_addr, size, dir);
2628 domain_flush_complete(domain);
2630 spin_unlock_irqrestore(&domain->lock, flags);
2634 * The exported map_sg function for dma_ops (handles scatter-gather
2637 static int map_sg(struct device *dev, struct scatterlist *sglist,
2638 int nelems, enum dma_data_direction dir,
2639 struct dma_attrs *attrs)
2641 unsigned long flags;
2642 struct protection_domain *domain;
2644 struct scatterlist *s;
2646 int mapped_elems = 0;
2649 INC_STATS_COUNTER(cnt_map_sg);
2651 domain = get_domain(dev);
2655 dma_mask = *dev->dma_mask;
2657 spin_lock_irqsave(&domain->lock, flags);
2659 for_each_sg(sglist, s, nelems, i) {
2662 s->dma_address = __map_single(dev, domain->priv,
2663 paddr, s->length, dir, false,
2666 if (s->dma_address) {
2667 s->dma_length = s->length;
2673 domain_flush_complete(domain);
2676 spin_unlock_irqrestore(&domain->lock, flags);
2678 return mapped_elems;
2680 for_each_sg(sglist, s, mapped_elems, i) {
2682 __unmap_single(domain->priv, s->dma_address,
2683 s->dma_length, dir);
2684 s->dma_address = s->dma_length = 0;
2693 * The exported map_sg function for dma_ops (handles scatter-gather
2696 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2697 int nelems, enum dma_data_direction dir,
2698 struct dma_attrs *attrs)
2700 unsigned long flags;
2701 struct protection_domain *domain;
2702 struct scatterlist *s;
2705 INC_STATS_COUNTER(cnt_unmap_sg);
2707 domain = get_domain(dev);
2711 spin_lock_irqsave(&domain->lock, flags);
2713 for_each_sg(sglist, s, nelems, i) {
2714 __unmap_single(domain->priv, s->dma_address,
2715 s->dma_length, dir);
2716 s->dma_address = s->dma_length = 0;
2719 domain_flush_complete(domain);
2721 spin_unlock_irqrestore(&domain->lock, flags);
2725 * The exported alloc_coherent function for dma_ops.
2727 static void *alloc_coherent(struct device *dev, size_t size,
2728 dma_addr_t *dma_addr, gfp_t flag,
2729 struct dma_attrs *attrs)
2731 u64 dma_mask = dev->coherent_dma_mask;
2732 struct protection_domain *domain;
2733 unsigned long flags;
2736 INC_STATS_COUNTER(cnt_alloc_coherent);
2738 domain = get_domain(dev);
2739 if (PTR_ERR(domain) == -EINVAL) {
2740 page = alloc_pages(flag, get_order(size));
2741 *dma_addr = page_to_phys(page);
2742 return page_address(page);
2743 } else if (IS_ERR(domain))
2746 size = PAGE_ALIGN(size);
2747 dma_mask = dev->coherent_dma_mask;
2748 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2751 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2753 if (!gfpflags_allow_blocking(flag))
2756 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2763 dma_mask = *dev->dma_mask;
2765 spin_lock_irqsave(&domain->lock, flags);
2767 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2768 size, DMA_BIDIRECTIONAL, true, dma_mask);
2770 if (*dma_addr == DMA_ERROR_CODE) {
2771 spin_unlock_irqrestore(&domain->lock, flags);
2775 domain_flush_complete(domain);
2777 spin_unlock_irqrestore(&domain->lock, flags);
2779 return page_address(page);
2783 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2784 __free_pages(page, get_order(size));
2790 * The exported free_coherent function for dma_ops.
2792 static void free_coherent(struct device *dev, size_t size,
2793 void *virt_addr, dma_addr_t dma_addr,
2794 struct dma_attrs *attrs)
2796 struct protection_domain *domain;
2797 unsigned long flags;
2800 INC_STATS_COUNTER(cnt_free_coherent);
2802 page = virt_to_page(virt_addr);
2803 size = PAGE_ALIGN(size);
2805 domain = get_domain(dev);
2809 spin_lock_irqsave(&domain->lock, flags);
2811 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2813 domain_flush_complete(domain);
2815 spin_unlock_irqrestore(&domain->lock, flags);
2818 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2819 __free_pages(page, get_order(size));
2823 * This function is called by the DMA layer to find out if we can handle a
2824 * particular device. It is part of the dma_ops.
2826 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2828 return check_device(dev);
2831 static struct dma_map_ops amd_iommu_dma_ops = {
2832 .alloc = alloc_coherent,
2833 .free = free_coherent,
2834 .map_page = map_page,
2835 .unmap_page = unmap_page,
2837 .unmap_sg = unmap_sg,
2838 .dma_supported = amd_iommu_dma_supported,
2841 int __init amd_iommu_init_api(void)
2843 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2846 int __init amd_iommu_init_dma_ops(void)
2848 swiotlb = iommu_pass_through ? 1 : 0;
2852 * In case we don't initialize SWIOTLB (actually the common case
2853 * when AMD IOMMU is enabled), make sure there are global
2854 * dma_ops set as a fall-back for devices not handled by this
2855 * driver (for example non-PCI devices).
2858 dma_ops = &nommu_dma_ops;
2860 amd_iommu_stats_init();
2862 if (amd_iommu_unmap_flush)
2863 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2865 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2870 /*****************************************************************************
2872 * The following functions belong to the exported interface of AMD IOMMU
2874 * This interface allows access to lower level functions of the IOMMU
2875 * like protection domain handling and assignement of devices to domains
2876 * which is not possible with the dma_ops interface.
2878 *****************************************************************************/
2880 static void cleanup_domain(struct protection_domain *domain)
2882 struct iommu_dev_data *entry;
2883 unsigned long flags;
2885 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2887 while (!list_empty(&domain->dev_list)) {
2888 entry = list_first_entry(&domain->dev_list,
2889 struct iommu_dev_data, list);
2890 __detach_device(entry);
2893 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2896 static void protection_domain_free(struct protection_domain *domain)
2901 del_domain_from_list(domain);
2904 domain_id_free(domain->id);
2909 static int protection_domain_init(struct protection_domain *domain)
2911 spin_lock_init(&domain->lock);
2912 mutex_init(&domain->api_lock);
2913 domain->id = domain_id_alloc();
2916 INIT_LIST_HEAD(&domain->dev_list);
2921 static struct protection_domain *protection_domain_alloc(void)
2923 struct protection_domain *domain;
2925 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2929 if (protection_domain_init(domain))
2932 add_domain_to_list(domain);
2942 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2944 struct protection_domain *pdomain;
2945 struct dma_ops_domain *dma_domain;
2948 case IOMMU_DOMAIN_UNMANAGED:
2949 pdomain = protection_domain_alloc();
2953 pdomain->mode = PAGE_MODE_3_LEVEL;
2954 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2955 if (!pdomain->pt_root) {
2956 protection_domain_free(pdomain);
2960 pdomain->domain.geometry.aperture_start = 0;
2961 pdomain->domain.geometry.aperture_end = ~0ULL;
2962 pdomain->domain.geometry.force_aperture = true;
2965 case IOMMU_DOMAIN_DMA:
2966 dma_domain = dma_ops_domain_alloc();
2968 pr_err("AMD-Vi: Failed to allocate\n");
2971 pdomain = &dma_domain->domain;
2973 case IOMMU_DOMAIN_IDENTITY:
2974 pdomain = protection_domain_alloc();
2978 pdomain->mode = PAGE_MODE_NONE;
2984 return &pdomain->domain;
2987 static void amd_iommu_domain_free(struct iommu_domain *dom)
2989 struct protection_domain *domain;
2994 domain = to_pdomain(dom);
2996 if (domain->dev_cnt > 0)
2997 cleanup_domain(domain);
2999 BUG_ON(domain->dev_cnt != 0);
3001 if (domain->mode != PAGE_MODE_NONE)
3002 free_pagetable(domain);
3004 if (domain->flags & PD_IOMMUV2_MASK)
3005 free_gcr3_table(domain);
3007 protection_domain_free(domain);
3010 static void amd_iommu_detach_device(struct iommu_domain *dom,
3013 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3014 struct amd_iommu *iommu;
3017 if (!check_device(dev))
3020 devid = get_device_id(dev);
3022 if (dev_data->domain != NULL)
3025 iommu = amd_iommu_rlookup_table[devid];
3029 iommu_completion_wait(iommu);
3032 static int amd_iommu_attach_device(struct iommu_domain *dom,
3035 struct protection_domain *domain = to_pdomain(dom);
3036 struct iommu_dev_data *dev_data;
3037 struct amd_iommu *iommu;
3040 if (!check_device(dev))
3043 dev_data = dev->archdata.iommu;
3045 iommu = amd_iommu_rlookup_table[dev_data->devid];
3049 if (dev_data->domain)
3052 ret = attach_device(dev, domain);
3054 iommu_completion_wait(iommu);
3059 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3060 phys_addr_t paddr, size_t page_size, int iommu_prot)
3062 struct protection_domain *domain = to_pdomain(dom);
3066 if (domain->mode == PAGE_MODE_NONE)
3069 if (iommu_prot & IOMMU_READ)
3070 prot |= IOMMU_PROT_IR;
3071 if (iommu_prot & IOMMU_WRITE)
3072 prot |= IOMMU_PROT_IW;
3074 mutex_lock(&domain->api_lock);
3075 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3076 mutex_unlock(&domain->api_lock);
3081 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3084 struct protection_domain *domain = to_pdomain(dom);
3087 if (domain->mode == PAGE_MODE_NONE)
3090 mutex_lock(&domain->api_lock);
3091 unmap_size = iommu_unmap_page(domain, iova, page_size);
3092 mutex_unlock(&domain->api_lock);
3094 domain_flush_tlb_pde(domain);
3099 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3102 struct protection_domain *domain = to_pdomain(dom);
3103 unsigned long offset_mask, pte_pgsize;
3106 if (domain->mode == PAGE_MODE_NONE)
3109 pte = fetch_pte(domain, iova, &pte_pgsize);
3111 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3114 offset_mask = pte_pgsize - 1;
3115 __pte = *pte & PM_ADDR_MASK;
3117 return (__pte & ~offset_mask) | (iova & offset_mask);
3120 static bool amd_iommu_capable(enum iommu_cap cap)
3123 case IOMMU_CAP_CACHE_COHERENCY:
3125 case IOMMU_CAP_INTR_REMAP:
3126 return (irq_remapping_enabled == 1);
3127 case IOMMU_CAP_NOEXEC:
3134 static void amd_iommu_get_dm_regions(struct device *dev,
3135 struct list_head *head)
3137 struct unity_map_entry *entry;
3140 devid = get_device_id(dev);
3142 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3143 struct iommu_dm_region *region;
3145 if (devid < entry->devid_start || devid > entry->devid_end)
3148 region = kzalloc(sizeof(*region), GFP_KERNEL);
3150 pr_err("Out of memory allocating dm-regions for %s\n",
3155 region->start = entry->address_start;
3156 region->length = entry->address_end - entry->address_start;
3157 if (entry->prot & IOMMU_PROT_IR)
3158 region->prot |= IOMMU_READ;
3159 if (entry->prot & IOMMU_PROT_IW)
3160 region->prot |= IOMMU_WRITE;
3162 list_add_tail(®ion->list, head);
3166 static void amd_iommu_put_dm_regions(struct device *dev,
3167 struct list_head *head)
3169 struct iommu_dm_region *entry, *next;
3171 list_for_each_entry_safe(entry, next, head, list)
3175 static const struct iommu_ops amd_iommu_ops = {
3176 .capable = amd_iommu_capable,
3177 .domain_alloc = amd_iommu_domain_alloc,
3178 .domain_free = amd_iommu_domain_free,
3179 .attach_dev = amd_iommu_attach_device,
3180 .detach_dev = amd_iommu_detach_device,
3181 .map = amd_iommu_map,
3182 .unmap = amd_iommu_unmap,
3183 .map_sg = default_iommu_map_sg,
3184 .iova_to_phys = amd_iommu_iova_to_phys,
3185 .add_device = amd_iommu_add_device,
3186 .remove_device = amd_iommu_remove_device,
3187 .device_group = pci_device_group,
3188 .get_dm_regions = amd_iommu_get_dm_regions,
3189 .put_dm_regions = amd_iommu_put_dm_regions,
3190 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3193 /*****************************************************************************
3195 * The next functions do a basic initialization of IOMMU for pass through
3198 * In passthrough mode the IOMMU is initialized and enabled but not used for
3199 * DMA-API translation.
3201 *****************************************************************************/
3203 /* IOMMUv2 specific functions */
3204 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3206 return atomic_notifier_chain_register(&ppr_notifier, nb);
3208 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3210 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3212 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3214 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3216 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3218 struct protection_domain *domain = to_pdomain(dom);
3219 unsigned long flags;
3221 spin_lock_irqsave(&domain->lock, flags);
3223 /* Update data structure */
3224 domain->mode = PAGE_MODE_NONE;
3225 domain->updated = true;
3227 /* Make changes visible to IOMMUs */
3228 update_domain(domain);
3230 /* Page-table is not visible to IOMMU anymore, so free it */
3231 free_pagetable(domain);
3233 spin_unlock_irqrestore(&domain->lock, flags);
3235 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3237 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3239 struct protection_domain *domain = to_pdomain(dom);
3240 unsigned long flags;
3243 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3246 /* Number of GCR3 table levels required */
3247 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3250 if (levels > amd_iommu_max_glx_val)
3253 spin_lock_irqsave(&domain->lock, flags);
3256 * Save us all sanity checks whether devices already in the
3257 * domain support IOMMUv2. Just force that the domain has no
3258 * devices attached when it is switched into IOMMUv2 mode.
3261 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3265 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3266 if (domain->gcr3_tbl == NULL)
3269 domain->glx = levels;
3270 domain->flags |= PD_IOMMUV2_MASK;
3271 domain->updated = true;
3273 update_domain(domain);
3278 spin_unlock_irqrestore(&domain->lock, flags);
3282 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3284 static int __flush_pasid(struct protection_domain *domain, int pasid,
3285 u64 address, bool size)
3287 struct iommu_dev_data *dev_data;
3288 struct iommu_cmd cmd;
3291 if (!(domain->flags & PD_IOMMUV2_MASK))
3294 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3297 * IOMMU TLB needs to be flushed before Device TLB to
3298 * prevent device TLB refill from IOMMU TLB
3300 for (i = 0; i < amd_iommus_present; ++i) {
3301 if (domain->dev_iommu[i] == 0)
3304 ret = iommu_queue_command(amd_iommus[i], &cmd);
3309 /* Wait until IOMMU TLB flushes are complete */
3310 domain_flush_complete(domain);
3312 /* Now flush device TLBs */
3313 list_for_each_entry(dev_data, &domain->dev_list, list) {
3314 struct amd_iommu *iommu;
3318 There might be non-IOMMUv2 capable devices in an IOMMUv2
3321 if (!dev_data->ats.enabled)
3324 qdep = dev_data->ats.qdep;
3325 iommu = amd_iommu_rlookup_table[dev_data->devid];
3327 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3328 qdep, address, size);
3330 ret = iommu_queue_command(iommu, &cmd);
3335 /* Wait until all device TLBs are flushed */
3336 domain_flush_complete(domain);
3345 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3348 INC_STATS_COUNTER(invalidate_iotlb);
3350 return __flush_pasid(domain, pasid, address, false);
3353 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3356 struct protection_domain *domain = to_pdomain(dom);
3357 unsigned long flags;
3360 spin_lock_irqsave(&domain->lock, flags);
3361 ret = __amd_iommu_flush_page(domain, pasid, address);
3362 spin_unlock_irqrestore(&domain->lock, flags);
3366 EXPORT_SYMBOL(amd_iommu_flush_page);
3368 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3370 INC_STATS_COUNTER(invalidate_iotlb_all);
3372 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3376 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3378 struct protection_domain *domain = to_pdomain(dom);
3379 unsigned long flags;
3382 spin_lock_irqsave(&domain->lock, flags);
3383 ret = __amd_iommu_flush_tlb(domain, pasid);
3384 spin_unlock_irqrestore(&domain->lock, flags);
3388 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3390 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3397 index = (pasid >> (9 * level)) & 0x1ff;
3403 if (!(*pte & GCR3_VALID)) {
3407 root = (void *)get_zeroed_page(GFP_ATOMIC);
3411 *pte = __pa(root) | GCR3_VALID;
3414 root = __va(*pte & PAGE_MASK);
3422 static int __set_gcr3(struct protection_domain *domain, int pasid,
3427 if (domain->mode != PAGE_MODE_NONE)
3430 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3434 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3436 return __amd_iommu_flush_tlb(domain, pasid);
3439 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3443 if (domain->mode != PAGE_MODE_NONE)
3446 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3452 return __amd_iommu_flush_tlb(domain, pasid);
3455 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3458 struct protection_domain *domain = to_pdomain(dom);
3459 unsigned long flags;
3462 spin_lock_irqsave(&domain->lock, flags);
3463 ret = __set_gcr3(domain, pasid, cr3);
3464 spin_unlock_irqrestore(&domain->lock, flags);
3468 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3470 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3472 struct protection_domain *domain = to_pdomain(dom);
3473 unsigned long flags;
3476 spin_lock_irqsave(&domain->lock, flags);
3477 ret = __clear_gcr3(domain, pasid);
3478 spin_unlock_irqrestore(&domain->lock, flags);
3482 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3484 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3485 int status, int tag)
3487 struct iommu_dev_data *dev_data;
3488 struct amd_iommu *iommu;
3489 struct iommu_cmd cmd;
3491 INC_STATS_COUNTER(complete_ppr);
3493 dev_data = get_dev_data(&pdev->dev);
3494 iommu = amd_iommu_rlookup_table[dev_data->devid];
3496 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3497 tag, dev_data->pri_tlp);
3499 return iommu_queue_command(iommu, &cmd);
3501 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3503 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3505 struct protection_domain *pdomain;
3507 pdomain = get_domain(&pdev->dev);
3508 if (IS_ERR(pdomain))
3511 /* Only return IOMMUv2 domains */
3512 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3515 return &pdomain->domain;
3517 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3519 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3521 struct iommu_dev_data *dev_data;
3523 if (!amd_iommu_v2_supported())
3526 dev_data = get_dev_data(&pdev->dev);
3527 dev_data->errata |= (1 << erratum);
3529 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3531 int amd_iommu_device_info(struct pci_dev *pdev,
3532 struct amd_iommu_device_info *info)
3537 if (pdev == NULL || info == NULL)
3540 if (!amd_iommu_v2_supported())
3543 memset(info, 0, sizeof(*info));
3545 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3547 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3549 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3551 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3553 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3557 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3558 max_pasids = min(max_pasids, (1 << 20));
3560 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3561 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3563 features = pci_pasid_features(pdev);
3564 if (features & PCI_PASID_CAP_EXEC)
3565 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3566 if (features & PCI_PASID_CAP_PRIV)
3567 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3572 EXPORT_SYMBOL(amd_iommu_device_info);
3574 #ifdef CONFIG_IRQ_REMAP
3576 /*****************************************************************************
3578 * Interrupt Remapping Implementation
3580 *****************************************************************************/
3598 u16 devid; /* Device ID for IRTE table */
3599 u16 index; /* Index into IRTE table*/
3602 struct amd_ir_data {
3603 struct irq_2_irte irq_2_irte;
3604 union irte irte_entry;
3606 struct msi_msg msi_entry;
3610 static struct irq_chip amd_ir_chip;
3612 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3613 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3614 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3615 #define DTE_IRQ_REMAP_ENABLE 1ULL
3617 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3621 dte = amd_iommu_dev_table[devid].data[2];
3622 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3623 dte |= virt_to_phys(table->table);
3624 dte |= DTE_IRQ_REMAP_INTCTL;
3625 dte |= DTE_IRQ_TABLE_LEN;
3626 dte |= DTE_IRQ_REMAP_ENABLE;
3628 amd_iommu_dev_table[devid].data[2] = dte;
3631 #define IRTE_ALLOCATED (~1U)
3633 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3635 struct irq_remap_table *table = NULL;
3636 struct amd_iommu *iommu;
3637 unsigned long flags;
3640 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3642 iommu = amd_iommu_rlookup_table[devid];
3646 table = irq_lookup_table[devid];
3650 alias = amd_iommu_alias_table[devid];
3651 table = irq_lookup_table[alias];
3653 irq_lookup_table[devid] = table;
3654 set_dte_irq_entry(devid, table);
3655 iommu_flush_dte(iommu, devid);
3659 /* Nothing there yet, allocate new irq remapping table */
3660 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3664 /* Initialize table spin-lock */
3665 spin_lock_init(&table->lock);
3668 /* Keep the first 32 indexes free for IOAPIC interrupts */
3669 table->min_index = 32;
3671 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3672 if (!table->table) {
3678 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3683 for (i = 0; i < 32; ++i)
3684 table->table[i] = IRTE_ALLOCATED;
3687 irq_lookup_table[devid] = table;
3688 set_dte_irq_entry(devid, table);
3689 iommu_flush_dte(iommu, devid);
3690 if (devid != alias) {
3691 irq_lookup_table[alias] = table;
3692 set_dte_irq_entry(alias, table);
3693 iommu_flush_dte(iommu, alias);
3697 iommu_completion_wait(iommu);
3700 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3705 static int alloc_irq_index(u16 devid, int count)
3707 struct irq_remap_table *table;
3708 unsigned long flags;
3711 table = get_irq_table(devid, false);
3715 spin_lock_irqsave(&table->lock, flags);
3717 /* Scan table for free entries */
3718 for (c = 0, index = table->min_index;
3719 index < MAX_IRQS_PER_TABLE;
3721 if (table->table[index] == 0)
3728 table->table[index - c + 1] = IRTE_ALLOCATED;
3738 spin_unlock_irqrestore(&table->lock, flags);
3743 static int modify_irte(u16 devid, int index, union irte irte)
3745 struct irq_remap_table *table;
3746 struct amd_iommu *iommu;
3747 unsigned long flags;
3749 iommu = amd_iommu_rlookup_table[devid];
3753 table = get_irq_table(devid, false);
3757 spin_lock_irqsave(&table->lock, flags);
3758 table->table[index] = irte.val;
3759 spin_unlock_irqrestore(&table->lock, flags);
3761 iommu_flush_irt(iommu, devid);
3762 iommu_completion_wait(iommu);
3767 static void free_irte(u16 devid, int index)
3769 struct irq_remap_table *table;
3770 struct amd_iommu *iommu;
3771 unsigned long flags;
3773 iommu = amd_iommu_rlookup_table[devid];
3777 table = get_irq_table(devid, false);
3781 spin_lock_irqsave(&table->lock, flags);
3782 table->table[index] = 0;
3783 spin_unlock_irqrestore(&table->lock, flags);
3785 iommu_flush_irt(iommu, devid);
3786 iommu_completion_wait(iommu);
3789 static int get_devid(struct irq_alloc_info *info)
3793 switch (info->type) {
3794 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3795 devid = get_ioapic_devid(info->ioapic_id);
3797 case X86_IRQ_ALLOC_TYPE_HPET:
3798 devid = get_hpet_devid(info->hpet_id);
3800 case X86_IRQ_ALLOC_TYPE_MSI:
3801 case X86_IRQ_ALLOC_TYPE_MSIX:
3802 devid = get_device_id(&info->msi_dev->dev);
3812 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3814 struct amd_iommu *iommu;
3820 devid = get_devid(info);
3822 iommu = amd_iommu_rlookup_table[devid];
3824 return iommu->ir_domain;
3830 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3832 struct amd_iommu *iommu;
3838 switch (info->type) {
3839 case X86_IRQ_ALLOC_TYPE_MSI:
3840 case X86_IRQ_ALLOC_TYPE_MSIX:
3841 devid = get_device_id(&info->msi_dev->dev);
3843 iommu = amd_iommu_rlookup_table[devid];
3845 return iommu->msi_domain;
3855 struct irq_remap_ops amd_iommu_irq_ops = {
3856 .prepare = amd_iommu_prepare,
3857 .enable = amd_iommu_enable,
3858 .disable = amd_iommu_disable,
3859 .reenable = amd_iommu_reenable,
3860 .enable_faulting = amd_iommu_enable_faulting,
3861 .get_ir_irq_domain = get_ir_irq_domain,
3862 .get_irq_domain = get_irq_domain,
3865 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3866 struct irq_cfg *irq_cfg,
3867 struct irq_alloc_info *info,
3868 int devid, int index, int sub_handle)
3870 struct irq_2_irte *irte_info = &data->irq_2_irte;
3871 struct msi_msg *msg = &data->msi_entry;
3872 union irte *irte = &data->irte_entry;
3873 struct IO_APIC_route_entry *entry;
3875 data->irq_2_irte.devid = devid;
3876 data->irq_2_irte.index = index + sub_handle;
3878 /* Setup IRTE for IOMMU */
3880 irte->fields.vector = irq_cfg->vector;
3881 irte->fields.int_type = apic->irq_delivery_mode;
3882 irte->fields.destination = irq_cfg->dest_apicid;
3883 irte->fields.dm = apic->irq_dest_mode;
3884 irte->fields.valid = 1;
3886 switch (info->type) {
3887 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3888 /* Setup IOAPIC entry */
3889 entry = info->ioapic_entry;
3890 info->ioapic_entry = NULL;
3891 memset(entry, 0, sizeof(*entry));
3892 entry->vector = index;
3894 entry->trigger = info->ioapic_trigger;
3895 entry->polarity = info->ioapic_polarity;
3896 /* Mask level triggered irqs. */
3897 if (info->ioapic_trigger)
3901 case X86_IRQ_ALLOC_TYPE_HPET:
3902 case X86_IRQ_ALLOC_TYPE_MSI:
3903 case X86_IRQ_ALLOC_TYPE_MSIX:
3904 msg->address_hi = MSI_ADDR_BASE_HI;
3905 msg->address_lo = MSI_ADDR_BASE_LO;
3906 msg->data = irte_info->index;
3915 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3916 unsigned int nr_irqs, void *arg)
3918 struct irq_alloc_info *info = arg;
3919 struct irq_data *irq_data;
3920 struct amd_ir_data *data;
3921 struct irq_cfg *cfg;
3927 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3928 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3932 * With IRQ remapping enabled, don't need contiguous CPU vectors
3933 * to support multiple MSI interrupts.
3935 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3936 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3938 devid = get_devid(info);
3942 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3946 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3947 if (get_irq_table(devid, true))
3948 index = info->ioapic_pin;
3952 index = alloc_irq_index(devid, nr_irqs);
3955 pr_warn("Failed to allocate IRTE\n");
3956 goto out_free_parent;
3959 for (i = 0; i < nr_irqs; i++) {
3960 irq_data = irq_domain_get_irq_data(domain, virq + i);
3961 cfg = irqd_cfg(irq_data);
3962 if (!irq_data || !cfg) {
3968 data = kzalloc(sizeof(*data), GFP_KERNEL);
3972 irq_data->hwirq = (devid << 16) + i;
3973 irq_data->chip_data = data;
3974 irq_data->chip = &amd_ir_chip;
3975 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3976 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3982 for (i--; i >= 0; i--) {
3983 irq_data = irq_domain_get_irq_data(domain, virq + i);
3985 kfree(irq_data->chip_data);
3987 for (i = 0; i < nr_irqs; i++)
3988 free_irte(devid, index + i);
3990 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3994 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3995 unsigned int nr_irqs)
3997 struct irq_2_irte *irte_info;
3998 struct irq_data *irq_data;
3999 struct amd_ir_data *data;
4002 for (i = 0; i < nr_irqs; i++) {
4003 irq_data = irq_domain_get_irq_data(domain, virq + i);
4004 if (irq_data && irq_data->chip_data) {
4005 data = irq_data->chip_data;
4006 irte_info = &data->irq_2_irte;
4007 free_irte(irte_info->devid, irte_info->index);
4011 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4014 static void irq_remapping_activate(struct irq_domain *domain,
4015 struct irq_data *irq_data)
4017 struct amd_ir_data *data = irq_data->chip_data;
4018 struct irq_2_irte *irte_info = &data->irq_2_irte;
4020 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4023 static void irq_remapping_deactivate(struct irq_domain *domain,
4024 struct irq_data *irq_data)
4026 struct amd_ir_data *data = irq_data->chip_data;
4027 struct irq_2_irte *irte_info = &data->irq_2_irte;
4031 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4034 static struct irq_domain_ops amd_ir_domain_ops = {
4035 .alloc = irq_remapping_alloc,
4036 .free = irq_remapping_free,
4037 .activate = irq_remapping_activate,
4038 .deactivate = irq_remapping_deactivate,
4041 static int amd_ir_set_affinity(struct irq_data *data,
4042 const struct cpumask *mask, bool force)
4044 struct amd_ir_data *ir_data = data->chip_data;
4045 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4046 struct irq_cfg *cfg = irqd_cfg(data);
4047 struct irq_data *parent = data->parent_data;
4050 ret = parent->chip->irq_set_affinity(parent, mask, force);
4051 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4055 * Atomically updates the IRTE with the new destination, vector
4056 * and flushes the interrupt entry cache.
4058 ir_data->irte_entry.fields.vector = cfg->vector;
4059 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4060 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4063 * After this point, all the interrupts will start arriving
4064 * at the new destination. So, time to cleanup the previous
4065 * vector allocation.
4067 send_cleanup_vector(cfg);
4069 return IRQ_SET_MASK_OK_DONE;
4072 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4074 struct amd_ir_data *ir_data = irq_data->chip_data;
4076 *msg = ir_data->msi_entry;
4079 static struct irq_chip amd_ir_chip = {
4080 .irq_ack = ir_ack_apic_edge,
4081 .irq_set_affinity = amd_ir_set_affinity,
4082 .irq_compose_msi_msg = ir_compose_msi_msg,
4085 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4087 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4088 if (!iommu->ir_domain)
4091 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4092 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);