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 <linux/percpu.h>
39 #include <asm/irq_remapping.h>
40 #include <asm/io_apic.h>
42 #include <asm/hw_irq.h>
43 #include <asm/msidef.h>
44 #include <asm/proto.h>
45 #include <asm/iommu.h>
49 #include "amd_iommu_proto.h"
50 #include "amd_iommu_types.h"
51 #include "irq_remapping.h"
53 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
55 #define LOOP_TIMEOUT 100000
58 * This bitmap is used to advertise the page sizes our hardware support
59 * to the IOMMU core, which will then use this information to split
60 * physically contiguous memory regions it is mapping into page sizes
63 * 512GB Pages are not supported due to a hardware bug
65 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
67 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
69 /* List of all available dev_data structures */
70 static LIST_HEAD(dev_data_list);
71 static DEFINE_SPINLOCK(dev_data_list_lock);
73 LIST_HEAD(ioapic_map);
77 * Domain for untranslated devices - only allocated
78 * if iommu=pt passed on kernel cmd line.
80 static const struct iommu_ops amd_iommu_ops;
82 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
83 int amd_iommu_max_glx_val = -1;
85 static struct dma_map_ops amd_iommu_dma_ops;
88 * This struct contains device specific data for the IOMMU
90 struct iommu_dev_data {
91 struct list_head list; /* For domain->dev_list */
92 struct list_head dev_data_list; /* For global dev_data_list */
93 struct protection_domain *domain; /* Domain the device is bound to */
94 u16 devid; /* PCI Device ID */
95 bool iommu_v2; /* Device can make use of IOMMUv2 */
96 bool passthrough; /* Device is identity mapped */
100 } ats; /* ATS state */
101 bool pri_tlp; /* PASID TLB required for
103 u32 errata; /* Bitmap for errata to apply */
107 * general struct to manage commands send to an IOMMU
113 struct kmem_cache *amd_iommu_irq_cache;
115 static void update_domain(struct protection_domain *domain);
116 static int protection_domain_init(struct protection_domain *domain);
119 * For dynamic growth the aperture size is split into ranges of 128MB of
120 * DMA address space each. This struct represents one such range.
122 struct aperture_range {
124 spinlock_t bitmap_lock;
126 /* address allocation bitmap */
127 unsigned long *bitmap;
128 unsigned long offset;
129 unsigned long next_bit;
132 * Array of PTE pages for the aperture. In this array we save all the
133 * leaf pages of the domain page table used for the aperture. This way
134 * we don't need to walk the page table to find a specific PTE. We can
135 * just calculate its address in constant time.
141 * Data container for a dma_ops specific protection domain
143 struct dma_ops_domain {
144 /* generic protection domain information */
145 struct protection_domain domain;
147 /* size of the aperture for the mappings */
148 unsigned long aperture_size;
150 /* aperture index we start searching for free addresses */
151 u32 __percpu *next_index;
153 /* address space relevant data */
154 struct aperture_range *aperture[APERTURE_MAX_RANGES];
157 /****************************************************************************
161 ****************************************************************************/
163 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
165 return container_of(dom, struct protection_domain, domain);
168 static struct iommu_dev_data *alloc_dev_data(u16 devid)
170 struct iommu_dev_data *dev_data;
173 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
177 dev_data->devid = devid;
179 spin_lock_irqsave(&dev_data_list_lock, flags);
180 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
181 spin_unlock_irqrestore(&dev_data_list_lock, flags);
186 static struct iommu_dev_data *search_dev_data(u16 devid)
188 struct iommu_dev_data *dev_data;
191 spin_lock_irqsave(&dev_data_list_lock, flags);
192 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
193 if (dev_data->devid == devid)
200 spin_unlock_irqrestore(&dev_data_list_lock, flags);
205 static struct iommu_dev_data *find_dev_data(u16 devid)
207 struct iommu_dev_data *dev_data;
209 dev_data = search_dev_data(devid);
211 if (dev_data == NULL)
212 dev_data = alloc_dev_data(devid);
217 static inline u16 get_device_id(struct device *dev)
219 struct pci_dev *pdev = to_pci_dev(dev);
221 return PCI_DEVID(pdev->bus->number, pdev->devfn);
224 static struct iommu_dev_data *get_dev_data(struct device *dev)
226 return dev->archdata.iommu;
229 static bool pci_iommuv2_capable(struct pci_dev *pdev)
231 static const int caps[] = {
234 PCI_EXT_CAP_ID_PASID,
238 for (i = 0; i < 3; ++i) {
239 pos = pci_find_ext_capability(pdev, caps[i]);
247 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
249 struct iommu_dev_data *dev_data;
251 dev_data = get_dev_data(&pdev->dev);
253 return dev_data->errata & (1 << erratum) ? true : false;
257 * This function actually applies the mapping to the page table of the
260 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
261 struct unity_map_entry *e)
265 for (addr = e->address_start; addr < e->address_end;
267 if (addr < dma_dom->aperture_size)
268 __set_bit(addr >> PAGE_SHIFT,
269 dma_dom->aperture[0]->bitmap);
274 * Inits the unity mappings required for a specific device
276 static void init_unity_mappings_for_device(struct device *dev,
277 struct dma_ops_domain *dma_dom)
279 struct unity_map_entry *e;
282 devid = get_device_id(dev);
284 list_for_each_entry(e, &amd_iommu_unity_map, list) {
285 if (!(devid >= e->devid_start && devid <= e->devid_end))
287 alloc_unity_mapping(dma_dom, e);
292 * This function checks if the driver got a valid device from the caller to
293 * avoid dereferencing invalid pointers.
295 static bool check_device(struct device *dev)
299 if (!dev || !dev->dma_mask)
303 if (!dev_is_pci(dev))
306 devid = get_device_id(dev);
308 /* Out of our scope? */
309 if (devid > amd_iommu_last_bdf)
312 if (amd_iommu_rlookup_table[devid] == NULL)
318 static void init_iommu_group(struct device *dev)
320 struct dma_ops_domain *dma_domain;
321 struct iommu_domain *domain;
322 struct iommu_group *group;
324 group = iommu_group_get_for_dev(dev);
328 domain = iommu_group_default_domain(group);
332 dma_domain = to_pdomain(domain)->priv;
334 init_unity_mappings_for_device(dev, dma_domain);
336 iommu_group_put(group);
339 static int iommu_init_device(struct device *dev)
341 struct pci_dev *pdev = to_pci_dev(dev);
342 struct iommu_dev_data *dev_data;
344 if (dev->archdata.iommu)
347 dev_data = find_dev_data(get_device_id(dev));
351 if (pci_iommuv2_capable(pdev)) {
352 struct amd_iommu *iommu;
354 iommu = amd_iommu_rlookup_table[dev_data->devid];
355 dev_data->iommu_v2 = iommu->is_iommu_v2;
358 dev->archdata.iommu = dev_data;
360 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
366 static void iommu_ignore_device(struct device *dev)
370 devid = get_device_id(dev);
371 alias = amd_iommu_alias_table[devid];
373 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
374 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
376 amd_iommu_rlookup_table[devid] = NULL;
377 amd_iommu_rlookup_table[alias] = NULL;
380 static void iommu_uninit_device(struct device *dev)
382 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
387 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
390 iommu_group_remove_device(dev);
393 dev->archdata.dma_ops = NULL;
396 * We keep dev_data around for unplugged devices and reuse it when the
397 * device is re-plugged - not doing so would introduce a ton of races.
401 #ifdef CONFIG_AMD_IOMMU_STATS
404 * Initialization code for statistics collection
407 DECLARE_STATS_COUNTER(compl_wait);
408 DECLARE_STATS_COUNTER(cnt_map_single);
409 DECLARE_STATS_COUNTER(cnt_unmap_single);
410 DECLARE_STATS_COUNTER(cnt_map_sg);
411 DECLARE_STATS_COUNTER(cnt_unmap_sg);
412 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
413 DECLARE_STATS_COUNTER(cnt_free_coherent);
414 DECLARE_STATS_COUNTER(cross_page);
415 DECLARE_STATS_COUNTER(domain_flush_single);
416 DECLARE_STATS_COUNTER(domain_flush_all);
417 DECLARE_STATS_COUNTER(alloced_io_mem);
418 DECLARE_STATS_COUNTER(total_map_requests);
419 DECLARE_STATS_COUNTER(complete_ppr);
420 DECLARE_STATS_COUNTER(invalidate_iotlb);
421 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
422 DECLARE_STATS_COUNTER(pri_requests);
424 static struct dentry *stats_dir;
425 static struct dentry *de_fflush;
427 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
429 if (stats_dir == NULL)
432 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
436 static void amd_iommu_stats_init(void)
438 stats_dir = debugfs_create_dir("amd-iommu", NULL);
439 if (stats_dir == NULL)
442 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
443 &amd_iommu_unmap_flush);
445 amd_iommu_stats_add(&compl_wait);
446 amd_iommu_stats_add(&cnt_map_single);
447 amd_iommu_stats_add(&cnt_unmap_single);
448 amd_iommu_stats_add(&cnt_map_sg);
449 amd_iommu_stats_add(&cnt_unmap_sg);
450 amd_iommu_stats_add(&cnt_alloc_coherent);
451 amd_iommu_stats_add(&cnt_free_coherent);
452 amd_iommu_stats_add(&cross_page);
453 amd_iommu_stats_add(&domain_flush_single);
454 amd_iommu_stats_add(&domain_flush_all);
455 amd_iommu_stats_add(&alloced_io_mem);
456 amd_iommu_stats_add(&total_map_requests);
457 amd_iommu_stats_add(&complete_ppr);
458 amd_iommu_stats_add(&invalidate_iotlb);
459 amd_iommu_stats_add(&invalidate_iotlb_all);
460 amd_iommu_stats_add(&pri_requests);
465 /****************************************************************************
467 * Interrupt handling functions
469 ****************************************************************************/
471 static void dump_dte_entry(u16 devid)
475 for (i = 0; i < 4; ++i)
476 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
477 amd_iommu_dev_table[devid].data[i]);
480 static void dump_command(unsigned long phys_addr)
482 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
485 for (i = 0; i < 4; ++i)
486 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
489 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
491 int type, devid, domid, flags;
492 volatile u32 *event = __evt;
497 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
498 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
499 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
500 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
501 address = (u64)(((u64)event[3]) << 32) | event[2];
504 /* Did we hit the erratum? */
505 if (++count == LOOP_TIMEOUT) {
506 pr_err("AMD-Vi: No event written to event log\n");
513 printk(KERN_ERR "AMD-Vi: Event logged [");
516 case EVENT_TYPE_ILL_DEV:
517 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
518 "address=0x%016llx flags=0x%04x]\n",
519 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
521 dump_dte_entry(devid);
523 case EVENT_TYPE_IO_FAULT:
524 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
525 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
526 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
527 domid, address, flags);
529 case EVENT_TYPE_DEV_TAB_ERR:
530 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
531 "address=0x%016llx flags=0x%04x]\n",
532 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
535 case EVENT_TYPE_PAGE_TAB_ERR:
536 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
537 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
538 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
539 domid, address, flags);
541 case EVENT_TYPE_ILL_CMD:
542 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
543 dump_command(address);
545 case EVENT_TYPE_CMD_HARD_ERR:
546 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
547 "flags=0x%04x]\n", address, flags);
549 case EVENT_TYPE_IOTLB_INV_TO:
550 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
551 "address=0x%016llx]\n",
552 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
555 case EVENT_TYPE_INV_DEV_REQ:
556 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
557 "address=0x%016llx flags=0x%04x]\n",
558 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
562 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
565 memset(__evt, 0, 4 * sizeof(u32));
568 static void iommu_poll_events(struct amd_iommu *iommu)
572 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
573 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
575 while (head != tail) {
576 iommu_print_event(iommu, iommu->evt_buf + head);
577 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
580 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
583 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
585 struct amd_iommu_fault fault;
587 INC_STATS_COUNTER(pri_requests);
589 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
590 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
594 fault.address = raw[1];
595 fault.pasid = PPR_PASID(raw[0]);
596 fault.device_id = PPR_DEVID(raw[0]);
597 fault.tag = PPR_TAG(raw[0]);
598 fault.flags = PPR_FLAGS(raw[0]);
600 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
603 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
607 if (iommu->ppr_log == NULL)
610 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
611 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
613 while (head != tail) {
618 raw = (u64 *)(iommu->ppr_log + head);
621 * Hardware bug: Interrupt may arrive before the entry is
622 * written to memory. If this happens we need to wait for the
625 for (i = 0; i < LOOP_TIMEOUT; ++i) {
626 if (PPR_REQ_TYPE(raw[0]) != 0)
631 /* Avoid memcpy function-call overhead */
636 * To detect the hardware bug we need to clear the entry
639 raw[0] = raw[1] = 0UL;
641 /* Update head pointer of hardware ring-buffer */
642 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
643 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
645 /* Handle PPR entry */
646 iommu_handle_ppr_entry(iommu, entry);
648 /* Refresh ring-buffer information */
649 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
650 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
654 irqreturn_t amd_iommu_int_thread(int irq, void *data)
656 struct amd_iommu *iommu = (struct amd_iommu *) data;
657 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
659 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
660 /* Enable EVT and PPR interrupts again */
661 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
662 iommu->mmio_base + MMIO_STATUS_OFFSET);
664 if (status & MMIO_STATUS_EVT_INT_MASK) {
665 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
666 iommu_poll_events(iommu);
669 if (status & MMIO_STATUS_PPR_INT_MASK) {
670 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
671 iommu_poll_ppr_log(iommu);
675 * Hardware bug: ERBT1312
676 * When re-enabling interrupt (by writing 1
677 * to clear the bit), the hardware might also try to set
678 * the interrupt bit in the event status register.
679 * In this scenario, the bit will be set, and disable
680 * subsequent interrupts.
682 * Workaround: The IOMMU driver should read back the
683 * status register and check if the interrupt bits are cleared.
684 * If not, driver will need to go through the interrupt handler
685 * again and re-clear the bits
687 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
692 irqreturn_t amd_iommu_int_handler(int irq, void *data)
694 return IRQ_WAKE_THREAD;
697 /****************************************************************************
699 * IOMMU command queuing functions
701 ****************************************************************************/
703 static int wait_on_sem(volatile u64 *sem)
707 while (*sem == 0 && i < LOOP_TIMEOUT) {
712 if (i == LOOP_TIMEOUT) {
713 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
720 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
721 struct iommu_cmd *cmd,
726 target = iommu->cmd_buf + tail;
727 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
729 /* Copy command to buffer */
730 memcpy(target, cmd, sizeof(*cmd));
732 /* Tell the IOMMU about it */
733 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
736 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
738 WARN_ON(address & 0x7ULL);
740 memset(cmd, 0, sizeof(*cmd));
741 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
742 cmd->data[1] = upper_32_bits(__pa(address));
744 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
747 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
749 memset(cmd, 0, sizeof(*cmd));
750 cmd->data[0] = devid;
751 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
754 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
755 size_t size, u16 domid, int pde)
760 pages = iommu_num_pages(address, size, PAGE_SIZE);
765 * If we have to flush more than one page, flush all
766 * TLB entries for this domain
768 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
772 address &= PAGE_MASK;
774 memset(cmd, 0, sizeof(*cmd));
775 cmd->data[1] |= domid;
776 cmd->data[2] = lower_32_bits(address);
777 cmd->data[3] = upper_32_bits(address);
778 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
779 if (s) /* size bit - we flush more than one 4kb page */
780 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
781 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
782 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
785 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
786 u64 address, size_t size)
791 pages = iommu_num_pages(address, size, PAGE_SIZE);
796 * If we have to flush more than one page, flush all
797 * TLB entries for this domain
799 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
803 address &= PAGE_MASK;
805 memset(cmd, 0, sizeof(*cmd));
806 cmd->data[0] = devid;
807 cmd->data[0] |= (qdep & 0xff) << 24;
808 cmd->data[1] = devid;
809 cmd->data[2] = lower_32_bits(address);
810 cmd->data[3] = upper_32_bits(address);
811 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
813 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
816 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
817 u64 address, bool size)
819 memset(cmd, 0, sizeof(*cmd));
821 address &= ~(0xfffULL);
823 cmd->data[0] = pasid;
824 cmd->data[1] = domid;
825 cmd->data[2] = lower_32_bits(address);
826 cmd->data[3] = upper_32_bits(address);
827 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
828 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
830 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
831 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
834 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
835 int qdep, u64 address, bool size)
837 memset(cmd, 0, sizeof(*cmd));
839 address &= ~(0xfffULL);
841 cmd->data[0] = devid;
842 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
843 cmd->data[0] |= (qdep & 0xff) << 24;
844 cmd->data[1] = devid;
845 cmd->data[1] |= (pasid & 0xff) << 16;
846 cmd->data[2] = lower_32_bits(address);
847 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
848 cmd->data[3] = upper_32_bits(address);
850 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
851 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
854 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
855 int status, int tag, bool gn)
857 memset(cmd, 0, sizeof(*cmd));
859 cmd->data[0] = devid;
861 cmd->data[1] = pasid;
862 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
864 cmd->data[3] = tag & 0x1ff;
865 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
867 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
870 static void build_inv_all(struct iommu_cmd *cmd)
872 memset(cmd, 0, sizeof(*cmd));
873 CMD_SET_TYPE(cmd, CMD_INV_ALL);
876 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
878 memset(cmd, 0, sizeof(*cmd));
879 cmd->data[0] = devid;
880 CMD_SET_TYPE(cmd, CMD_INV_IRT);
884 * Writes the command to the IOMMUs command buffer and informs the
885 * hardware about the new command.
887 static int iommu_queue_command_sync(struct amd_iommu *iommu,
888 struct iommu_cmd *cmd,
891 u32 left, tail, head, next_tail;
895 spin_lock_irqsave(&iommu->lock, flags);
897 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
898 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
899 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
900 left = (head - next_tail) % CMD_BUFFER_SIZE;
903 struct iommu_cmd sync_cmd;
904 volatile u64 sem = 0;
907 build_completion_wait(&sync_cmd, (u64)&sem);
908 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
910 spin_unlock_irqrestore(&iommu->lock, flags);
912 if ((ret = wait_on_sem(&sem)) != 0)
918 copy_cmd_to_buffer(iommu, cmd, tail);
920 /* We need to sync now to make sure all commands are processed */
921 iommu->need_sync = sync;
923 spin_unlock_irqrestore(&iommu->lock, flags);
928 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
930 return iommu_queue_command_sync(iommu, cmd, true);
934 * This function queues a completion wait command into the command
937 static int iommu_completion_wait(struct amd_iommu *iommu)
939 struct iommu_cmd cmd;
940 volatile u64 sem = 0;
943 if (!iommu->need_sync)
946 build_completion_wait(&cmd, (u64)&sem);
948 ret = iommu_queue_command_sync(iommu, &cmd, false);
952 return wait_on_sem(&sem);
955 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
957 struct iommu_cmd cmd;
959 build_inv_dte(&cmd, devid);
961 return iommu_queue_command(iommu, &cmd);
964 static void iommu_flush_dte_all(struct amd_iommu *iommu)
968 for (devid = 0; devid <= 0xffff; ++devid)
969 iommu_flush_dte(iommu, devid);
971 iommu_completion_wait(iommu);
975 * This function uses heavy locking and may disable irqs for some time. But
976 * this is no issue because it is only called during resume.
978 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
982 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
983 struct iommu_cmd cmd;
984 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
986 iommu_queue_command(iommu, &cmd);
989 iommu_completion_wait(iommu);
992 static void iommu_flush_all(struct amd_iommu *iommu)
994 struct iommu_cmd cmd;
998 iommu_queue_command(iommu, &cmd);
999 iommu_completion_wait(iommu);
1002 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1004 struct iommu_cmd cmd;
1006 build_inv_irt(&cmd, devid);
1008 iommu_queue_command(iommu, &cmd);
1011 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1015 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1016 iommu_flush_irt(iommu, devid);
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_irt_all(iommu);
1028 iommu_flush_tlb_all(iommu);
1033 * Command send function for flushing on-device TLB
1035 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1036 u64 address, size_t size)
1038 struct amd_iommu *iommu;
1039 struct iommu_cmd cmd;
1042 qdep = dev_data->ats.qdep;
1043 iommu = amd_iommu_rlookup_table[dev_data->devid];
1045 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1047 return iommu_queue_command(iommu, &cmd);
1051 * Command send function for invalidating a device table entry
1053 static int device_flush_dte(struct iommu_dev_data *dev_data)
1055 struct amd_iommu *iommu;
1059 iommu = amd_iommu_rlookup_table[dev_data->devid];
1060 alias = amd_iommu_alias_table[dev_data->devid];
1062 ret = iommu_flush_dte(iommu, dev_data->devid);
1063 if (!ret && alias != dev_data->devid)
1064 ret = iommu_flush_dte(iommu, alias);
1068 if (dev_data->ats.enabled)
1069 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1075 * TLB invalidation function which is called from the mapping functions.
1076 * It invalidates a single PTE if the range to flush is within a single
1077 * page. Otherwise it flushes the whole TLB of the IOMMU.
1079 static void __domain_flush_pages(struct protection_domain *domain,
1080 u64 address, size_t size, int pde)
1082 struct iommu_dev_data *dev_data;
1083 struct iommu_cmd cmd;
1086 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1088 for (i = 0; i < amd_iommus_present; ++i) {
1089 if (!domain->dev_iommu[i])
1093 * Devices of this domain are behind this IOMMU
1094 * We need a TLB flush
1096 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1099 list_for_each_entry(dev_data, &domain->dev_list, list) {
1101 if (!dev_data->ats.enabled)
1104 ret |= device_flush_iotlb(dev_data, address, size);
1110 static void domain_flush_pages(struct protection_domain *domain,
1111 u64 address, size_t size)
1113 __domain_flush_pages(domain, address, size, 0);
1116 /* Flush the whole IO/TLB for a given protection domain */
1117 static void domain_flush_tlb(struct protection_domain *domain)
1119 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1122 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1123 static void domain_flush_tlb_pde(struct protection_domain *domain)
1125 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1128 static void domain_flush_complete(struct protection_domain *domain)
1132 for (i = 0; i < amd_iommus_present; ++i) {
1133 if (!domain->dev_iommu[i])
1137 * Devices of this domain are behind this IOMMU
1138 * We need to wait for completion of all commands.
1140 iommu_completion_wait(amd_iommus[i]);
1146 * This function flushes the DTEs for all devices in domain
1148 static void domain_flush_devices(struct protection_domain *domain)
1150 struct iommu_dev_data *dev_data;
1152 list_for_each_entry(dev_data, &domain->dev_list, list)
1153 device_flush_dte(dev_data);
1156 /****************************************************************************
1158 * The functions below are used the create the page table mappings for
1159 * unity mapped regions.
1161 ****************************************************************************/
1164 * This function is used to add another level to an IO page table. Adding
1165 * another level increases the size of the address space by 9 bits to a size up
1168 static bool increase_address_space(struct protection_domain *domain,
1173 if (domain->mode == PAGE_MODE_6_LEVEL)
1174 /* address space already 64 bit large */
1177 pte = (void *)get_zeroed_page(gfp);
1181 *pte = PM_LEVEL_PDE(domain->mode,
1182 virt_to_phys(domain->pt_root));
1183 domain->pt_root = pte;
1185 domain->updated = true;
1190 static u64 *alloc_pte(struct protection_domain *domain,
1191 unsigned long address,
1192 unsigned long page_size,
1199 BUG_ON(!is_power_of_2(page_size));
1201 while (address > PM_LEVEL_SIZE(domain->mode))
1202 increase_address_space(domain, gfp);
1204 level = domain->mode - 1;
1205 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1206 address = PAGE_SIZE_ALIGN(address, page_size);
1207 end_lvl = PAGE_SIZE_LEVEL(page_size);
1209 while (level > end_lvl) {
1214 if (!IOMMU_PTE_PRESENT(__pte)) {
1215 page = (u64 *)get_zeroed_page(gfp);
1219 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1221 if (cmpxchg64(pte, __pte, __npte)) {
1222 free_page((unsigned long)page);
1227 /* No level skipping support yet */
1228 if (PM_PTE_LEVEL(*pte) != level)
1233 pte = IOMMU_PTE_PAGE(*pte);
1235 if (pte_page && level == end_lvl)
1238 pte = &pte[PM_LEVEL_INDEX(level, address)];
1245 * This function checks if there is a PTE for a given dma address. If
1246 * there is one, it returns the pointer to it.
1248 static u64 *fetch_pte(struct protection_domain *domain,
1249 unsigned long address,
1250 unsigned long *page_size)
1255 if (address > PM_LEVEL_SIZE(domain->mode))
1258 level = domain->mode - 1;
1259 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1260 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1265 if (!IOMMU_PTE_PRESENT(*pte))
1269 if (PM_PTE_LEVEL(*pte) == 7 ||
1270 PM_PTE_LEVEL(*pte) == 0)
1273 /* No level skipping support yet */
1274 if (PM_PTE_LEVEL(*pte) != level)
1279 /* Walk to the next level */
1280 pte = IOMMU_PTE_PAGE(*pte);
1281 pte = &pte[PM_LEVEL_INDEX(level, address)];
1282 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1285 if (PM_PTE_LEVEL(*pte) == 0x07) {
1286 unsigned long pte_mask;
1289 * If we have a series of large PTEs, make
1290 * sure to return a pointer to the first one.
1292 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1293 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1294 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1301 * Generic mapping functions. It maps a physical address into a DMA
1302 * address space. It allocates the page table pages if necessary.
1303 * In the future it can be extended to a generic mapping function
1304 * supporting all features of AMD IOMMU page tables like level skipping
1305 * and full 64 bit address spaces.
1307 static int iommu_map_page(struct protection_domain *dom,
1308 unsigned long bus_addr,
1309 unsigned long phys_addr,
1311 unsigned long page_size)
1316 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1317 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1319 if (!(prot & IOMMU_PROT_MASK))
1322 count = PAGE_SIZE_PTE_COUNT(page_size);
1323 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1328 for (i = 0; i < count; ++i)
1329 if (IOMMU_PTE_PRESENT(pte[i]))
1333 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1334 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1336 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1338 if (prot & IOMMU_PROT_IR)
1339 __pte |= IOMMU_PTE_IR;
1340 if (prot & IOMMU_PROT_IW)
1341 __pte |= IOMMU_PTE_IW;
1343 for (i = 0; i < count; ++i)
1351 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1352 unsigned long bus_addr,
1353 unsigned long page_size)
1355 unsigned long long unmapped;
1356 unsigned long unmap_size;
1359 BUG_ON(!is_power_of_2(page_size));
1363 while (unmapped < page_size) {
1365 pte = fetch_pte(dom, bus_addr, &unmap_size);
1370 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1371 for (i = 0; i < count; i++)
1375 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1376 unmapped += unmap_size;
1379 BUG_ON(unmapped && !is_power_of_2(unmapped));
1384 /****************************************************************************
1386 * The next functions belong to the address allocator for the dma_ops
1387 * interface functions. They work like the allocators in the other IOMMU
1388 * drivers. Its basically a bitmap which marks the allocated pages in
1389 * the aperture. Maybe it could be enhanced in the future to a more
1390 * efficient allocator.
1392 ****************************************************************************/
1395 * The address allocator core functions.
1397 * called with domain->lock held
1401 * Used to reserve address ranges in the aperture (e.g. for exclusion
1404 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1405 unsigned long start_page,
1408 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1410 if (start_page + pages > last_page)
1411 pages = last_page - start_page;
1413 for (i = start_page; i < start_page + pages; ++i) {
1414 int index = i / APERTURE_RANGE_PAGES;
1415 int page = i % APERTURE_RANGE_PAGES;
1416 __set_bit(page, dom->aperture[index]->bitmap);
1421 * This function is used to add a new aperture range to an existing
1422 * aperture in case of dma_ops domain allocation or address allocation
1425 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1426 bool populate, gfp_t gfp)
1428 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1429 unsigned long i, old_size, pte_pgsize;
1430 struct aperture_range *range;
1431 struct amd_iommu *iommu;
1432 unsigned long flags;
1434 #ifdef CONFIG_IOMMU_STRESS
1438 if (index >= APERTURE_MAX_RANGES)
1441 range = kzalloc(sizeof(struct aperture_range), gfp);
1445 range->bitmap = (void *)get_zeroed_page(gfp);
1449 range->offset = dma_dom->aperture_size;
1451 spin_lock_init(&range->bitmap_lock);
1454 unsigned long address = dma_dom->aperture_size;
1455 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1456 u64 *pte, *pte_page;
1458 for (i = 0; i < num_ptes; ++i) {
1459 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1464 range->pte_pages[i] = pte_page;
1466 address += APERTURE_RANGE_SIZE / 64;
1470 spin_lock_irqsave(&dma_dom->domain.lock, flags);
1472 /* First take the bitmap_lock and then publish the range */
1473 spin_lock(&range->bitmap_lock);
1475 old_size = dma_dom->aperture_size;
1476 dma_dom->aperture[index] = range;
1477 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1479 /* Reserve address range used for MSI messages */
1480 if (old_size < MSI_ADDR_BASE_LO &&
1481 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1482 unsigned long spage;
1485 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1486 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1488 dma_ops_reserve_addresses(dma_dom, spage, pages);
1491 /* Initialize the exclusion range if necessary */
1492 for_each_iommu(iommu) {
1493 if (iommu->exclusion_start &&
1494 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1495 && iommu->exclusion_start < dma_dom->aperture_size) {
1496 unsigned long startpage;
1497 int pages = iommu_num_pages(iommu->exclusion_start,
1498 iommu->exclusion_length,
1500 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1501 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1506 * Check for areas already mapped as present in the new aperture
1507 * range and mark those pages as reserved in the allocator. Such
1508 * mappings may already exist as a result of requested unity
1509 * mappings for devices.
1511 for (i = dma_dom->aperture[index]->offset;
1512 i < dma_dom->aperture_size;
1514 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1515 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1518 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1522 update_domain(&dma_dom->domain);
1524 spin_unlock(&range->bitmap_lock);
1526 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
1531 update_domain(&dma_dom->domain);
1533 free_page((unsigned long)range->bitmap);
1540 static dma_addr_t dma_ops_aperture_alloc(struct dma_ops_domain *dom,
1541 struct aperture_range *range,
1542 unsigned long pages,
1543 unsigned long dma_mask,
1544 unsigned long boundary_size,
1545 unsigned long align_mask,
1548 unsigned long offset, limit, flags;
1552 offset = range->offset >> PAGE_SHIFT;
1553 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1554 dma_mask >> PAGE_SHIFT);
1557 if (!spin_trylock_irqsave(&range->bitmap_lock, flags))
1560 spin_lock_irqsave(&range->bitmap_lock, flags);
1563 address = iommu_area_alloc(range->bitmap, limit, range->next_bit,
1564 pages, offset, boundary_size, align_mask);
1565 if (address == -1) {
1566 /* Nothing found, retry one time */
1567 address = iommu_area_alloc(range->bitmap, limit,
1568 0, pages, offset, boundary_size,
1574 range->next_bit = address + pages;
1576 spin_unlock_irqrestore(&range->bitmap_lock, flags);
1579 domain_flush_tlb(&dom->domain);
1580 domain_flush_complete(&dom->domain);
1586 static unsigned long dma_ops_area_alloc(struct device *dev,
1587 struct dma_ops_domain *dom,
1589 unsigned long align_mask,
1592 unsigned long boundary_size, mask;
1593 unsigned long address = -1;
1599 mask = dma_get_seg_boundary(dev);
1602 start = this_cpu_read(*dom->next_index);
1604 /* Sanity check - is it really necessary? */
1605 if (unlikely(start > APERTURE_MAX_RANGES)) {
1607 this_cpu_write(*dom->next_index, 0);
1610 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1611 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1613 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1614 struct aperture_range *range;
1617 index = (start + i) % APERTURE_MAX_RANGES;
1619 range = dom->aperture[index];
1621 if (!range || range->offset >= dma_mask)
1624 address = dma_ops_aperture_alloc(dom, range, pages,
1625 dma_mask, boundary_size,
1627 if (address != -1) {
1628 address = range->offset + (address << PAGE_SHIFT);
1629 this_cpu_write(*dom->next_index, index);
1634 if (address == -1 && first) {
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 = -1;
1652 while (address == -1) {
1653 address = dma_ops_area_alloc(dev, dom, pages,
1654 align_mask, dma_mask);
1656 if (address == -1 && alloc_new_range(dom, false, GFP_ATOMIC))
1660 if (unlikely(address == -1))
1661 address = DMA_ERROR_CODE;
1663 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1669 * The address free function.
1671 * called with domain->lock held
1673 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1674 unsigned long address,
1677 unsigned i = address >> APERTURE_RANGE_SHIFT;
1678 struct aperture_range *range = dom->aperture[i];
1679 unsigned long flags;
1681 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1683 #ifdef CONFIG_IOMMU_STRESS
1688 if (amd_iommu_unmap_flush) {
1689 domain_flush_tlb(&dom->domain);
1690 domain_flush_complete(&dom->domain);
1693 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1695 spin_lock_irqsave(&range->bitmap_lock, flags);
1696 if (address + pages > range->next_bit)
1697 range->next_bit = address + pages;
1698 bitmap_clear(range->bitmap, address, pages);
1699 spin_unlock_irqrestore(&range->bitmap_lock, flags);
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 #define DEFINE_FREE_PT_FN(LVL, FN) \
1766 static void free_pt_##LVL (unsigned long __pt) \
1774 for (i = 0; i < 512; ++i) { \
1775 /* PTE present? */ \
1776 if (!IOMMU_PTE_PRESENT(pt[i])) \
1780 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1781 PM_PTE_LEVEL(pt[i]) == 7) \
1784 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1787 free_page((unsigned long)pt); \
1790 DEFINE_FREE_PT_FN(l2, free_page)
1791 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1792 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1793 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1794 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1796 static void free_pagetable(struct protection_domain *domain)
1798 unsigned long root = (unsigned long)domain->pt_root;
1800 switch (domain->mode) {
1801 case PAGE_MODE_NONE:
1803 case PAGE_MODE_1_LEVEL:
1806 case PAGE_MODE_2_LEVEL:
1809 case PAGE_MODE_3_LEVEL:
1812 case PAGE_MODE_4_LEVEL:
1815 case PAGE_MODE_5_LEVEL:
1818 case PAGE_MODE_6_LEVEL:
1826 static void free_gcr3_tbl_level1(u64 *tbl)
1831 for (i = 0; i < 512; ++i) {
1832 if (!(tbl[i] & GCR3_VALID))
1835 ptr = __va(tbl[i] & PAGE_MASK);
1837 free_page((unsigned long)ptr);
1841 static void free_gcr3_tbl_level2(u64 *tbl)
1846 for (i = 0; i < 512; ++i) {
1847 if (!(tbl[i] & GCR3_VALID))
1850 ptr = __va(tbl[i] & PAGE_MASK);
1852 free_gcr3_tbl_level1(ptr);
1856 static void free_gcr3_table(struct protection_domain *domain)
1858 if (domain->glx == 2)
1859 free_gcr3_tbl_level2(domain->gcr3_tbl);
1860 else if (domain->glx == 1)
1861 free_gcr3_tbl_level1(domain->gcr3_tbl);
1863 BUG_ON(domain->glx != 0);
1865 free_page((unsigned long)domain->gcr3_tbl);
1869 * Free a domain, only used if something went wrong in the
1870 * allocation path and we need to free an already allocated page table
1872 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1879 free_percpu(dom->next_index);
1881 del_domain_from_list(&dom->domain);
1883 free_pagetable(&dom->domain);
1885 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1886 if (!dom->aperture[i])
1888 free_page((unsigned long)dom->aperture[i]->bitmap);
1889 kfree(dom->aperture[i]);
1896 * Allocates a new protection domain usable for the dma_ops functions.
1897 * It also initializes the page table and the address allocator data
1898 * structures required for the dma_ops interface
1900 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1902 struct dma_ops_domain *dma_dom;
1905 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1909 if (protection_domain_init(&dma_dom->domain))
1912 dma_dom->next_index = alloc_percpu(u32);
1913 if (!dma_dom->next_index)
1916 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1917 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1918 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1919 dma_dom->domain.priv = dma_dom;
1920 if (!dma_dom->domain.pt_root)
1923 add_domain_to_list(&dma_dom->domain);
1925 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1929 * mark the first page as allocated so we never return 0 as
1930 * a valid dma-address. So we can use 0 as error value
1932 dma_dom->aperture[0]->bitmap[0] = 1;
1934 for_each_possible_cpu(cpu)
1935 *per_cpu_ptr(dma_dom->next_index, cpu) = 0;
1940 dma_ops_domain_free(dma_dom);
1946 * little helper function to check whether a given protection domain is a
1949 static bool dma_ops_domain(struct protection_domain *domain)
1951 return domain->flags & PD_DMA_OPS_MASK;
1954 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1959 if (domain->mode != PAGE_MODE_NONE)
1960 pte_root = virt_to_phys(domain->pt_root);
1962 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1963 << DEV_ENTRY_MODE_SHIFT;
1964 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1966 flags = amd_iommu_dev_table[devid].data[1];
1969 flags |= DTE_FLAG_IOTLB;
1971 if (domain->flags & PD_IOMMUV2_MASK) {
1972 u64 gcr3 = __pa(domain->gcr3_tbl);
1973 u64 glx = domain->glx;
1976 pte_root |= DTE_FLAG_GV;
1977 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1979 /* First mask out possible old values for GCR3 table */
1980 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1983 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1986 /* Encode GCR3 table into DTE */
1987 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1990 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1993 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1997 flags &= ~(0xffffUL);
1998 flags |= domain->id;
2000 amd_iommu_dev_table[devid].data[1] = flags;
2001 amd_iommu_dev_table[devid].data[0] = pte_root;
2004 static void clear_dte_entry(u16 devid)
2006 /* remove entry from the device table seen by the hardware */
2007 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2008 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2010 amd_iommu_apply_erratum_63(devid);
2013 static void do_attach(struct iommu_dev_data *dev_data,
2014 struct protection_domain *domain)
2016 struct amd_iommu *iommu;
2020 iommu = amd_iommu_rlookup_table[dev_data->devid];
2021 alias = amd_iommu_alias_table[dev_data->devid];
2022 ats = dev_data->ats.enabled;
2024 /* Update data structures */
2025 dev_data->domain = domain;
2026 list_add(&dev_data->list, &domain->dev_list);
2028 /* Do reference counting */
2029 domain->dev_iommu[iommu->index] += 1;
2030 domain->dev_cnt += 1;
2032 /* Update device table */
2033 set_dte_entry(dev_data->devid, domain, ats);
2034 if (alias != dev_data->devid)
2035 set_dte_entry(dev_data->devid, domain, ats);
2037 device_flush_dte(dev_data);
2040 static void do_detach(struct iommu_dev_data *dev_data)
2042 struct amd_iommu *iommu;
2046 * First check if the device is still attached. It might already
2047 * be detached from its domain because the generic
2048 * iommu_detach_group code detached it and we try again here in
2049 * our alias handling.
2051 if (!dev_data->domain)
2054 iommu = amd_iommu_rlookup_table[dev_data->devid];
2055 alias = amd_iommu_alias_table[dev_data->devid];
2057 /* decrease reference counters */
2058 dev_data->domain->dev_iommu[iommu->index] -= 1;
2059 dev_data->domain->dev_cnt -= 1;
2061 /* Update data structures */
2062 dev_data->domain = NULL;
2063 list_del(&dev_data->list);
2064 clear_dte_entry(dev_data->devid);
2065 if (alias != dev_data->devid)
2066 clear_dte_entry(alias);
2068 /* Flush the DTE entry */
2069 device_flush_dte(dev_data);
2073 * If a device is not yet associated with a domain, this function does
2074 * assigns it visible for the hardware
2076 static int __attach_device(struct iommu_dev_data *dev_data,
2077 struct protection_domain *domain)
2082 * Must be called with IRQs disabled. Warn here to detect early
2085 WARN_ON(!irqs_disabled());
2088 spin_lock(&domain->lock);
2091 if (dev_data->domain != NULL)
2094 /* Attach alias group root */
2095 do_attach(dev_data, domain);
2102 spin_unlock(&domain->lock);
2108 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2110 pci_disable_ats(pdev);
2111 pci_disable_pri(pdev);
2112 pci_disable_pasid(pdev);
2115 /* FIXME: Change generic reset-function to do the same */
2116 static int pri_reset_while_enabled(struct pci_dev *pdev)
2121 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2125 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2126 control |= PCI_PRI_CTRL_RESET;
2127 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2132 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2137 /* FIXME: Hardcode number of outstanding requests for now */
2139 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2141 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2143 /* Only allow access to user-accessible pages */
2144 ret = pci_enable_pasid(pdev, 0);
2148 /* First reset the PRI state of the device */
2149 ret = pci_reset_pri(pdev);
2154 ret = pci_enable_pri(pdev, reqs);
2159 ret = pri_reset_while_enabled(pdev);
2164 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2171 pci_disable_pri(pdev);
2172 pci_disable_pasid(pdev);
2177 /* FIXME: Move this to PCI code */
2178 #define PCI_PRI_TLP_OFF (1 << 15)
2180 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2185 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2189 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2191 return (status & PCI_PRI_TLP_OFF) ? true : false;
2195 * If a device is not yet associated with a domain, this function
2196 * assigns it visible for the hardware
2198 static int attach_device(struct device *dev,
2199 struct protection_domain *domain)
2201 struct pci_dev *pdev = to_pci_dev(dev);
2202 struct iommu_dev_data *dev_data;
2203 unsigned long flags;
2206 dev_data = get_dev_data(dev);
2208 if (domain->flags & PD_IOMMUV2_MASK) {
2209 if (!dev_data->passthrough)
2212 if (dev_data->iommu_v2) {
2213 if (pdev_iommuv2_enable(pdev) != 0)
2216 dev_data->ats.enabled = true;
2217 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2218 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2220 } else if (amd_iommu_iotlb_sup &&
2221 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2222 dev_data->ats.enabled = true;
2223 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2226 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2227 ret = __attach_device(dev_data, domain);
2228 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2231 * We might boot into a crash-kernel here. The crashed kernel
2232 * left the caches in the IOMMU dirty. So we have to flush
2233 * here to evict all dirty stuff.
2235 domain_flush_tlb_pde(domain);
2241 * Removes a device from a protection domain (unlocked)
2243 static void __detach_device(struct iommu_dev_data *dev_data)
2245 struct protection_domain *domain;
2248 * Must be called with IRQs disabled. Warn here to detect early
2251 WARN_ON(!irqs_disabled());
2253 if (WARN_ON(!dev_data->domain))
2256 domain = dev_data->domain;
2258 spin_lock(&domain->lock);
2260 do_detach(dev_data);
2262 spin_unlock(&domain->lock);
2266 * Removes a device from a protection domain (with devtable_lock held)
2268 static void detach_device(struct device *dev)
2270 struct protection_domain *domain;
2271 struct iommu_dev_data *dev_data;
2272 unsigned long flags;
2274 dev_data = get_dev_data(dev);
2275 domain = dev_data->domain;
2277 /* lock device table */
2278 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2279 __detach_device(dev_data);
2280 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2282 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2283 pdev_iommuv2_disable(to_pci_dev(dev));
2284 else if (dev_data->ats.enabled)
2285 pci_disable_ats(to_pci_dev(dev));
2287 dev_data->ats.enabled = false;
2290 static int amd_iommu_add_device(struct device *dev)
2292 struct iommu_dev_data *dev_data;
2293 struct iommu_domain *domain;
2294 struct amd_iommu *iommu;
2298 if (!check_device(dev) || get_dev_data(dev))
2301 devid = get_device_id(dev);
2302 iommu = amd_iommu_rlookup_table[devid];
2304 ret = iommu_init_device(dev);
2306 if (ret != -ENOTSUPP)
2307 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2310 iommu_ignore_device(dev);
2311 dev->archdata.dma_ops = &nommu_dma_ops;
2314 init_iommu_group(dev);
2316 dev_data = get_dev_data(dev);
2320 if (iommu_pass_through || dev_data->iommu_v2)
2321 iommu_request_dm_for_dev(dev);
2323 /* Domains are initialized for this device - have a look what we ended up with */
2324 domain = iommu_get_domain_for_dev(dev);
2325 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2326 dev_data->passthrough = true;
2328 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2331 iommu_completion_wait(iommu);
2336 static void amd_iommu_remove_device(struct device *dev)
2338 struct amd_iommu *iommu;
2341 if (!check_device(dev))
2344 devid = get_device_id(dev);
2345 iommu = amd_iommu_rlookup_table[devid];
2347 iommu_uninit_device(dev);
2348 iommu_completion_wait(iommu);
2351 /*****************************************************************************
2353 * The next functions belong to the dma_ops mapping/unmapping code.
2355 *****************************************************************************/
2358 * In the dma_ops path we only have the struct device. This function
2359 * finds the corresponding IOMMU, the protection domain and the
2360 * requestor id for a given device.
2361 * If the device is not yet associated with a domain this is also done
2364 static struct protection_domain *get_domain(struct device *dev)
2366 struct protection_domain *domain;
2367 struct iommu_domain *io_domain;
2369 if (!check_device(dev))
2370 return ERR_PTR(-EINVAL);
2372 io_domain = iommu_get_domain_for_dev(dev);
2376 domain = to_pdomain(io_domain);
2377 if (!dma_ops_domain(domain))
2378 return ERR_PTR(-EBUSY);
2383 static void update_device_table(struct protection_domain *domain)
2385 struct iommu_dev_data *dev_data;
2387 list_for_each_entry(dev_data, &domain->dev_list, list)
2388 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2391 static void update_domain(struct protection_domain *domain)
2393 if (!domain->updated)
2396 update_device_table(domain);
2398 domain_flush_devices(domain);
2399 domain_flush_tlb_pde(domain);
2401 domain->updated = false;
2405 * This function fetches the PTE for a given address in the aperture
2407 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2408 unsigned long address)
2410 struct aperture_range *aperture;
2411 u64 *pte, *pte_page;
2413 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2417 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2419 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2421 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2423 pte += PM_LEVEL_INDEX(0, address);
2425 update_domain(&dom->domain);
2431 * This is the generic map function. It maps one 4kb page at paddr to
2432 * the given address in the DMA address space for the domain.
2434 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2435 unsigned long address,
2441 WARN_ON(address > dom->aperture_size);
2445 pte = dma_ops_get_pte(dom, address);
2447 return DMA_ERROR_CODE;
2449 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2451 if (direction == DMA_TO_DEVICE)
2452 __pte |= IOMMU_PTE_IR;
2453 else if (direction == DMA_FROM_DEVICE)
2454 __pte |= IOMMU_PTE_IW;
2455 else if (direction == DMA_BIDIRECTIONAL)
2456 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2462 return (dma_addr_t)address;
2466 * The generic unmapping function for on page in the DMA address space.
2468 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2469 unsigned long address)
2471 struct aperture_range *aperture;
2474 if (address >= dom->aperture_size)
2477 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2481 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2485 pte += PM_LEVEL_INDEX(0, address);
2487 WARN_ON_ONCE(!*pte);
2493 * This function contains common code for mapping of a physically
2494 * contiguous memory region into DMA address space. It is used by all
2495 * mapping functions provided with this IOMMU driver.
2496 * Must be called with the domain lock held.
2498 static dma_addr_t __map_single(struct device *dev,
2499 struct dma_ops_domain *dma_dom,
2506 dma_addr_t offset = paddr & ~PAGE_MASK;
2507 dma_addr_t address, start, ret;
2509 unsigned long align_mask = 0;
2512 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2515 INC_STATS_COUNTER(total_map_requests);
2518 INC_STATS_COUNTER(cross_page);
2521 align_mask = (1UL << get_order(size)) - 1;
2523 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2526 if (address == DMA_ERROR_CODE)
2530 for (i = 0; i < pages; ++i) {
2531 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2532 if (ret == DMA_ERROR_CODE)
2540 ADD_STATS_COUNTER(alloced_io_mem, size);
2542 if (unlikely(amd_iommu_np_cache)) {
2543 domain_flush_pages(&dma_dom->domain, address, size);
2544 domain_flush_complete(&dma_dom->domain);
2552 for (--i; i >= 0; --i) {
2554 dma_ops_domain_unmap(dma_dom, start);
2557 dma_ops_free_addresses(dma_dom, address, pages);
2559 return DMA_ERROR_CODE;
2563 * Does the reverse of the __map_single function. Must be called with
2564 * the domain lock held too
2566 static void __unmap_single(struct dma_ops_domain *dma_dom,
2567 dma_addr_t dma_addr,
2571 dma_addr_t flush_addr;
2572 dma_addr_t i, start;
2575 if ((dma_addr == DMA_ERROR_CODE) ||
2576 (dma_addr + size > dma_dom->aperture_size))
2579 flush_addr = dma_addr;
2580 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2581 dma_addr &= PAGE_MASK;
2584 for (i = 0; i < pages; ++i) {
2585 dma_ops_domain_unmap(dma_dom, start);
2589 SUB_STATS_COUNTER(alloced_io_mem, size);
2591 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2595 * The exported map_single function for dma_ops.
2597 static dma_addr_t map_page(struct device *dev, struct page *page,
2598 unsigned long offset, size_t size,
2599 enum dma_data_direction dir,
2600 struct dma_attrs *attrs)
2602 phys_addr_t paddr = page_to_phys(page) + offset;
2603 struct protection_domain *domain;
2606 INC_STATS_COUNTER(cnt_map_single);
2608 domain = get_domain(dev);
2609 if (PTR_ERR(domain) == -EINVAL)
2610 return (dma_addr_t)paddr;
2611 else if (IS_ERR(domain))
2612 return DMA_ERROR_CODE;
2614 dma_mask = *dev->dma_mask;
2616 return __map_single(dev, domain->priv, paddr, size, dir, false,
2621 * The exported unmap_single function for dma_ops.
2623 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2624 enum dma_data_direction dir, struct dma_attrs *attrs)
2626 struct protection_domain *domain;
2628 INC_STATS_COUNTER(cnt_unmap_single);
2630 domain = get_domain(dev);
2634 __unmap_single(domain->priv, dma_addr, size, dir);
2638 * The exported map_sg function for dma_ops (handles scatter-gather
2641 static int map_sg(struct device *dev, struct scatterlist *sglist,
2642 int nelems, enum dma_data_direction dir,
2643 struct dma_attrs *attrs)
2645 struct protection_domain *domain;
2647 struct scatterlist *s;
2649 int mapped_elems = 0;
2652 INC_STATS_COUNTER(cnt_map_sg);
2654 domain = get_domain(dev);
2658 dma_mask = *dev->dma_mask;
2660 for_each_sg(sglist, s, nelems, i) {
2663 s->dma_address = __map_single(dev, domain->priv,
2664 paddr, s->length, dir, false,
2667 if (s->dma_address) {
2668 s->dma_length = s->length;
2674 return mapped_elems;
2677 for_each_sg(sglist, s, mapped_elems, i) {
2679 __unmap_single(domain->priv, s->dma_address,
2680 s->dma_length, dir);
2681 s->dma_address = s->dma_length = 0;
2688 * The exported map_sg function for dma_ops (handles scatter-gather
2691 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2692 int nelems, enum dma_data_direction dir,
2693 struct dma_attrs *attrs)
2695 struct protection_domain *domain;
2696 struct scatterlist *s;
2699 INC_STATS_COUNTER(cnt_unmap_sg);
2701 domain = get_domain(dev);
2705 for_each_sg(sglist, s, nelems, i) {
2706 __unmap_single(domain->priv, s->dma_address,
2707 s->dma_length, dir);
2708 s->dma_address = s->dma_length = 0;
2713 * The exported alloc_coherent function for dma_ops.
2715 static void *alloc_coherent(struct device *dev, size_t size,
2716 dma_addr_t *dma_addr, gfp_t flag,
2717 struct dma_attrs *attrs)
2719 u64 dma_mask = dev->coherent_dma_mask;
2720 struct protection_domain *domain;
2723 INC_STATS_COUNTER(cnt_alloc_coherent);
2725 domain = get_domain(dev);
2726 if (PTR_ERR(domain) == -EINVAL) {
2727 page = alloc_pages(flag, get_order(size));
2728 *dma_addr = page_to_phys(page);
2729 return page_address(page);
2730 } else if (IS_ERR(domain))
2733 size = PAGE_ALIGN(size);
2734 dma_mask = dev->coherent_dma_mask;
2735 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2738 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2740 if (!gfpflags_allow_blocking(flag))
2743 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2750 dma_mask = *dev->dma_mask;
2752 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2753 size, DMA_BIDIRECTIONAL, true, dma_mask);
2755 if (*dma_addr == DMA_ERROR_CODE)
2758 return page_address(page);
2762 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2763 __free_pages(page, get_order(size));
2769 * The exported free_coherent function for dma_ops.
2771 static void free_coherent(struct device *dev, size_t size,
2772 void *virt_addr, dma_addr_t dma_addr,
2773 struct dma_attrs *attrs)
2775 struct protection_domain *domain;
2778 INC_STATS_COUNTER(cnt_free_coherent);
2780 page = virt_to_page(virt_addr);
2781 size = PAGE_ALIGN(size);
2783 domain = get_domain(dev);
2787 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2790 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2791 __free_pages(page, get_order(size));
2795 * This function is called by the DMA layer to find out if we can handle a
2796 * particular device. It is part of the dma_ops.
2798 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2800 return check_device(dev);
2803 static struct dma_map_ops amd_iommu_dma_ops = {
2804 .alloc = alloc_coherent,
2805 .free = free_coherent,
2806 .map_page = map_page,
2807 .unmap_page = unmap_page,
2809 .unmap_sg = unmap_sg,
2810 .dma_supported = amd_iommu_dma_supported,
2813 int __init amd_iommu_init_api(void)
2815 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2818 int __init amd_iommu_init_dma_ops(void)
2820 swiotlb = iommu_pass_through ? 1 : 0;
2824 * In case we don't initialize SWIOTLB (actually the common case
2825 * when AMD IOMMU is enabled), make sure there are global
2826 * dma_ops set as a fall-back for devices not handled by this
2827 * driver (for example non-PCI devices).
2830 dma_ops = &nommu_dma_ops;
2832 amd_iommu_stats_init();
2834 if (amd_iommu_unmap_flush)
2835 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2837 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2842 /*****************************************************************************
2844 * The following functions belong to the exported interface of AMD IOMMU
2846 * This interface allows access to lower level functions of the IOMMU
2847 * like protection domain handling and assignement of devices to domains
2848 * which is not possible with the dma_ops interface.
2850 *****************************************************************************/
2852 static void cleanup_domain(struct protection_domain *domain)
2854 struct iommu_dev_data *entry;
2855 unsigned long flags;
2857 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2859 while (!list_empty(&domain->dev_list)) {
2860 entry = list_first_entry(&domain->dev_list,
2861 struct iommu_dev_data, list);
2862 __detach_device(entry);
2865 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2868 static void protection_domain_free(struct protection_domain *domain)
2873 del_domain_from_list(domain);
2876 domain_id_free(domain->id);
2881 static int protection_domain_init(struct protection_domain *domain)
2883 spin_lock_init(&domain->lock);
2884 mutex_init(&domain->api_lock);
2885 domain->id = domain_id_alloc();
2888 INIT_LIST_HEAD(&domain->dev_list);
2893 static struct protection_domain *protection_domain_alloc(void)
2895 struct protection_domain *domain;
2897 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2901 if (protection_domain_init(domain))
2904 add_domain_to_list(domain);
2914 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2916 struct protection_domain *pdomain;
2917 struct dma_ops_domain *dma_domain;
2920 case IOMMU_DOMAIN_UNMANAGED:
2921 pdomain = protection_domain_alloc();
2925 pdomain->mode = PAGE_MODE_3_LEVEL;
2926 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2927 if (!pdomain->pt_root) {
2928 protection_domain_free(pdomain);
2932 pdomain->domain.geometry.aperture_start = 0;
2933 pdomain->domain.geometry.aperture_end = ~0ULL;
2934 pdomain->domain.geometry.force_aperture = true;
2937 case IOMMU_DOMAIN_DMA:
2938 dma_domain = dma_ops_domain_alloc();
2940 pr_err("AMD-Vi: Failed to allocate\n");
2943 pdomain = &dma_domain->domain;
2945 case IOMMU_DOMAIN_IDENTITY:
2946 pdomain = protection_domain_alloc();
2950 pdomain->mode = PAGE_MODE_NONE;
2956 return &pdomain->domain;
2959 static void amd_iommu_domain_free(struct iommu_domain *dom)
2961 struct protection_domain *domain;
2966 domain = to_pdomain(dom);
2968 if (domain->dev_cnt > 0)
2969 cleanup_domain(domain);
2971 BUG_ON(domain->dev_cnt != 0);
2973 if (domain->mode != PAGE_MODE_NONE)
2974 free_pagetable(domain);
2976 if (domain->flags & PD_IOMMUV2_MASK)
2977 free_gcr3_table(domain);
2979 protection_domain_free(domain);
2982 static void amd_iommu_detach_device(struct iommu_domain *dom,
2985 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2986 struct amd_iommu *iommu;
2989 if (!check_device(dev))
2992 devid = get_device_id(dev);
2994 if (dev_data->domain != NULL)
2997 iommu = amd_iommu_rlookup_table[devid];
3001 iommu_completion_wait(iommu);
3004 static int amd_iommu_attach_device(struct iommu_domain *dom,
3007 struct protection_domain *domain = to_pdomain(dom);
3008 struct iommu_dev_data *dev_data;
3009 struct amd_iommu *iommu;
3012 if (!check_device(dev))
3015 dev_data = dev->archdata.iommu;
3017 iommu = amd_iommu_rlookup_table[dev_data->devid];
3021 if (dev_data->domain)
3024 ret = attach_device(dev, domain);
3026 iommu_completion_wait(iommu);
3031 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3032 phys_addr_t paddr, size_t page_size, int iommu_prot)
3034 struct protection_domain *domain = to_pdomain(dom);
3038 if (domain->mode == PAGE_MODE_NONE)
3041 if (iommu_prot & IOMMU_READ)
3042 prot |= IOMMU_PROT_IR;
3043 if (iommu_prot & IOMMU_WRITE)
3044 prot |= IOMMU_PROT_IW;
3046 mutex_lock(&domain->api_lock);
3047 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3048 mutex_unlock(&domain->api_lock);
3053 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3056 struct protection_domain *domain = to_pdomain(dom);
3059 if (domain->mode == PAGE_MODE_NONE)
3062 mutex_lock(&domain->api_lock);
3063 unmap_size = iommu_unmap_page(domain, iova, page_size);
3064 mutex_unlock(&domain->api_lock);
3066 domain_flush_tlb_pde(domain);
3071 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3074 struct protection_domain *domain = to_pdomain(dom);
3075 unsigned long offset_mask, pte_pgsize;
3078 if (domain->mode == PAGE_MODE_NONE)
3081 pte = fetch_pte(domain, iova, &pte_pgsize);
3083 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3086 offset_mask = pte_pgsize - 1;
3087 __pte = *pte & PM_ADDR_MASK;
3089 return (__pte & ~offset_mask) | (iova & offset_mask);
3092 static bool amd_iommu_capable(enum iommu_cap cap)
3095 case IOMMU_CAP_CACHE_COHERENCY:
3097 case IOMMU_CAP_INTR_REMAP:
3098 return (irq_remapping_enabled == 1);
3099 case IOMMU_CAP_NOEXEC:
3106 static void amd_iommu_get_dm_regions(struct device *dev,
3107 struct list_head *head)
3109 struct unity_map_entry *entry;
3112 devid = get_device_id(dev);
3114 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3115 struct iommu_dm_region *region;
3117 if (devid < entry->devid_start || devid > entry->devid_end)
3120 region = kzalloc(sizeof(*region), GFP_KERNEL);
3122 pr_err("Out of memory allocating dm-regions for %s\n",
3127 region->start = entry->address_start;
3128 region->length = entry->address_end - entry->address_start;
3129 if (entry->prot & IOMMU_PROT_IR)
3130 region->prot |= IOMMU_READ;
3131 if (entry->prot & IOMMU_PROT_IW)
3132 region->prot |= IOMMU_WRITE;
3134 list_add_tail(®ion->list, head);
3138 static void amd_iommu_put_dm_regions(struct device *dev,
3139 struct list_head *head)
3141 struct iommu_dm_region *entry, *next;
3143 list_for_each_entry_safe(entry, next, head, list)
3147 static const struct iommu_ops amd_iommu_ops = {
3148 .capable = amd_iommu_capable,
3149 .domain_alloc = amd_iommu_domain_alloc,
3150 .domain_free = amd_iommu_domain_free,
3151 .attach_dev = amd_iommu_attach_device,
3152 .detach_dev = amd_iommu_detach_device,
3153 .map = amd_iommu_map,
3154 .unmap = amd_iommu_unmap,
3155 .map_sg = default_iommu_map_sg,
3156 .iova_to_phys = amd_iommu_iova_to_phys,
3157 .add_device = amd_iommu_add_device,
3158 .remove_device = amd_iommu_remove_device,
3159 .device_group = pci_device_group,
3160 .get_dm_regions = amd_iommu_get_dm_regions,
3161 .put_dm_regions = amd_iommu_put_dm_regions,
3162 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3165 /*****************************************************************************
3167 * The next functions do a basic initialization of IOMMU for pass through
3170 * In passthrough mode the IOMMU is initialized and enabled but not used for
3171 * DMA-API translation.
3173 *****************************************************************************/
3175 /* IOMMUv2 specific functions */
3176 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3178 return atomic_notifier_chain_register(&ppr_notifier, nb);
3180 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3182 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3184 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3186 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3188 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3190 struct protection_domain *domain = to_pdomain(dom);
3191 unsigned long flags;
3193 spin_lock_irqsave(&domain->lock, flags);
3195 /* Update data structure */
3196 domain->mode = PAGE_MODE_NONE;
3197 domain->updated = true;
3199 /* Make changes visible to IOMMUs */
3200 update_domain(domain);
3202 /* Page-table is not visible to IOMMU anymore, so free it */
3203 free_pagetable(domain);
3205 spin_unlock_irqrestore(&domain->lock, flags);
3207 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3209 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3211 struct protection_domain *domain = to_pdomain(dom);
3212 unsigned long flags;
3215 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3218 /* Number of GCR3 table levels required */
3219 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3222 if (levels > amd_iommu_max_glx_val)
3225 spin_lock_irqsave(&domain->lock, flags);
3228 * Save us all sanity checks whether devices already in the
3229 * domain support IOMMUv2. Just force that the domain has no
3230 * devices attached when it is switched into IOMMUv2 mode.
3233 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3237 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3238 if (domain->gcr3_tbl == NULL)
3241 domain->glx = levels;
3242 domain->flags |= PD_IOMMUV2_MASK;
3243 domain->updated = true;
3245 update_domain(domain);
3250 spin_unlock_irqrestore(&domain->lock, flags);
3254 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3256 static int __flush_pasid(struct protection_domain *domain, int pasid,
3257 u64 address, bool size)
3259 struct iommu_dev_data *dev_data;
3260 struct iommu_cmd cmd;
3263 if (!(domain->flags & PD_IOMMUV2_MASK))
3266 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3269 * IOMMU TLB needs to be flushed before Device TLB to
3270 * prevent device TLB refill from IOMMU TLB
3272 for (i = 0; i < amd_iommus_present; ++i) {
3273 if (domain->dev_iommu[i] == 0)
3276 ret = iommu_queue_command(amd_iommus[i], &cmd);
3281 /* Wait until IOMMU TLB flushes are complete */
3282 domain_flush_complete(domain);
3284 /* Now flush device TLBs */
3285 list_for_each_entry(dev_data, &domain->dev_list, list) {
3286 struct amd_iommu *iommu;
3290 There might be non-IOMMUv2 capable devices in an IOMMUv2
3293 if (!dev_data->ats.enabled)
3296 qdep = dev_data->ats.qdep;
3297 iommu = amd_iommu_rlookup_table[dev_data->devid];
3299 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3300 qdep, address, size);
3302 ret = iommu_queue_command(iommu, &cmd);
3307 /* Wait until all device TLBs are flushed */
3308 domain_flush_complete(domain);
3317 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3320 INC_STATS_COUNTER(invalidate_iotlb);
3322 return __flush_pasid(domain, pasid, address, false);
3325 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3328 struct protection_domain *domain = to_pdomain(dom);
3329 unsigned long flags;
3332 spin_lock_irqsave(&domain->lock, flags);
3333 ret = __amd_iommu_flush_page(domain, pasid, address);
3334 spin_unlock_irqrestore(&domain->lock, flags);
3338 EXPORT_SYMBOL(amd_iommu_flush_page);
3340 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3342 INC_STATS_COUNTER(invalidate_iotlb_all);
3344 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3348 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3350 struct protection_domain *domain = to_pdomain(dom);
3351 unsigned long flags;
3354 spin_lock_irqsave(&domain->lock, flags);
3355 ret = __amd_iommu_flush_tlb(domain, pasid);
3356 spin_unlock_irqrestore(&domain->lock, flags);
3360 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3362 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3369 index = (pasid >> (9 * level)) & 0x1ff;
3375 if (!(*pte & GCR3_VALID)) {
3379 root = (void *)get_zeroed_page(GFP_ATOMIC);
3383 *pte = __pa(root) | GCR3_VALID;
3386 root = __va(*pte & PAGE_MASK);
3394 static int __set_gcr3(struct protection_domain *domain, int pasid,
3399 if (domain->mode != PAGE_MODE_NONE)
3402 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3406 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3408 return __amd_iommu_flush_tlb(domain, pasid);
3411 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3415 if (domain->mode != PAGE_MODE_NONE)
3418 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3424 return __amd_iommu_flush_tlb(domain, pasid);
3427 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3430 struct protection_domain *domain = to_pdomain(dom);
3431 unsigned long flags;
3434 spin_lock_irqsave(&domain->lock, flags);
3435 ret = __set_gcr3(domain, pasid, cr3);
3436 spin_unlock_irqrestore(&domain->lock, flags);
3440 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3442 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3444 struct protection_domain *domain = to_pdomain(dom);
3445 unsigned long flags;
3448 spin_lock_irqsave(&domain->lock, flags);
3449 ret = __clear_gcr3(domain, pasid);
3450 spin_unlock_irqrestore(&domain->lock, flags);
3454 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3456 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3457 int status, int tag)
3459 struct iommu_dev_data *dev_data;
3460 struct amd_iommu *iommu;
3461 struct iommu_cmd cmd;
3463 INC_STATS_COUNTER(complete_ppr);
3465 dev_data = get_dev_data(&pdev->dev);
3466 iommu = amd_iommu_rlookup_table[dev_data->devid];
3468 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3469 tag, dev_data->pri_tlp);
3471 return iommu_queue_command(iommu, &cmd);
3473 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3475 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3477 struct protection_domain *pdomain;
3479 pdomain = get_domain(&pdev->dev);
3480 if (IS_ERR(pdomain))
3483 /* Only return IOMMUv2 domains */
3484 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3487 return &pdomain->domain;
3489 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3491 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3493 struct iommu_dev_data *dev_data;
3495 if (!amd_iommu_v2_supported())
3498 dev_data = get_dev_data(&pdev->dev);
3499 dev_data->errata |= (1 << erratum);
3501 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3503 int amd_iommu_device_info(struct pci_dev *pdev,
3504 struct amd_iommu_device_info *info)
3509 if (pdev == NULL || info == NULL)
3512 if (!amd_iommu_v2_supported())
3515 memset(info, 0, sizeof(*info));
3517 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3519 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3521 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3523 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3525 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3529 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3530 max_pasids = min(max_pasids, (1 << 20));
3532 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3533 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3535 features = pci_pasid_features(pdev);
3536 if (features & PCI_PASID_CAP_EXEC)
3537 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3538 if (features & PCI_PASID_CAP_PRIV)
3539 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3544 EXPORT_SYMBOL(amd_iommu_device_info);
3546 #ifdef CONFIG_IRQ_REMAP
3548 /*****************************************************************************
3550 * Interrupt Remapping Implementation
3552 *****************************************************************************/
3570 u16 devid; /* Device ID for IRTE table */
3571 u16 index; /* Index into IRTE table*/
3574 struct amd_ir_data {
3575 struct irq_2_irte irq_2_irte;
3576 union irte irte_entry;
3578 struct msi_msg msi_entry;
3582 static struct irq_chip amd_ir_chip;
3584 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3585 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3586 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3587 #define DTE_IRQ_REMAP_ENABLE 1ULL
3589 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3593 dte = amd_iommu_dev_table[devid].data[2];
3594 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3595 dte |= virt_to_phys(table->table);
3596 dte |= DTE_IRQ_REMAP_INTCTL;
3597 dte |= DTE_IRQ_TABLE_LEN;
3598 dte |= DTE_IRQ_REMAP_ENABLE;
3600 amd_iommu_dev_table[devid].data[2] = dte;
3603 #define IRTE_ALLOCATED (~1U)
3605 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3607 struct irq_remap_table *table = NULL;
3608 struct amd_iommu *iommu;
3609 unsigned long flags;
3612 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3614 iommu = amd_iommu_rlookup_table[devid];
3618 table = irq_lookup_table[devid];
3622 alias = amd_iommu_alias_table[devid];
3623 table = irq_lookup_table[alias];
3625 irq_lookup_table[devid] = table;
3626 set_dte_irq_entry(devid, table);
3627 iommu_flush_dte(iommu, devid);
3631 /* Nothing there yet, allocate new irq remapping table */
3632 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3636 /* Initialize table spin-lock */
3637 spin_lock_init(&table->lock);
3640 /* Keep the first 32 indexes free for IOAPIC interrupts */
3641 table->min_index = 32;
3643 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3644 if (!table->table) {
3650 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3655 for (i = 0; i < 32; ++i)
3656 table->table[i] = IRTE_ALLOCATED;
3659 irq_lookup_table[devid] = table;
3660 set_dte_irq_entry(devid, table);
3661 iommu_flush_dte(iommu, devid);
3662 if (devid != alias) {
3663 irq_lookup_table[alias] = table;
3664 set_dte_irq_entry(alias, table);
3665 iommu_flush_dte(iommu, alias);
3669 iommu_completion_wait(iommu);
3672 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3677 static int alloc_irq_index(u16 devid, int count)
3679 struct irq_remap_table *table;
3680 unsigned long flags;
3683 table = get_irq_table(devid, false);
3687 spin_lock_irqsave(&table->lock, flags);
3689 /* Scan table for free entries */
3690 for (c = 0, index = table->min_index;
3691 index < MAX_IRQS_PER_TABLE;
3693 if (table->table[index] == 0)
3700 table->table[index - c + 1] = IRTE_ALLOCATED;
3710 spin_unlock_irqrestore(&table->lock, flags);
3715 static int modify_irte(u16 devid, int index, union irte irte)
3717 struct irq_remap_table *table;
3718 struct amd_iommu *iommu;
3719 unsigned long flags;
3721 iommu = amd_iommu_rlookup_table[devid];
3725 table = get_irq_table(devid, false);
3729 spin_lock_irqsave(&table->lock, flags);
3730 table->table[index] = irte.val;
3731 spin_unlock_irqrestore(&table->lock, flags);
3733 iommu_flush_irt(iommu, devid);
3734 iommu_completion_wait(iommu);
3739 static void free_irte(u16 devid, int index)
3741 struct irq_remap_table *table;
3742 struct amd_iommu *iommu;
3743 unsigned long flags;
3745 iommu = amd_iommu_rlookup_table[devid];
3749 table = get_irq_table(devid, false);
3753 spin_lock_irqsave(&table->lock, flags);
3754 table->table[index] = 0;
3755 spin_unlock_irqrestore(&table->lock, flags);
3757 iommu_flush_irt(iommu, devid);
3758 iommu_completion_wait(iommu);
3761 static int get_devid(struct irq_alloc_info *info)
3765 switch (info->type) {
3766 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3767 devid = get_ioapic_devid(info->ioapic_id);
3769 case X86_IRQ_ALLOC_TYPE_HPET:
3770 devid = get_hpet_devid(info->hpet_id);
3772 case X86_IRQ_ALLOC_TYPE_MSI:
3773 case X86_IRQ_ALLOC_TYPE_MSIX:
3774 devid = get_device_id(&info->msi_dev->dev);
3784 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3786 struct amd_iommu *iommu;
3792 devid = get_devid(info);
3794 iommu = amd_iommu_rlookup_table[devid];
3796 return iommu->ir_domain;
3802 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3804 struct amd_iommu *iommu;
3810 switch (info->type) {
3811 case X86_IRQ_ALLOC_TYPE_MSI:
3812 case X86_IRQ_ALLOC_TYPE_MSIX:
3813 devid = get_device_id(&info->msi_dev->dev);
3815 iommu = amd_iommu_rlookup_table[devid];
3817 return iommu->msi_domain;
3827 struct irq_remap_ops amd_iommu_irq_ops = {
3828 .prepare = amd_iommu_prepare,
3829 .enable = amd_iommu_enable,
3830 .disable = amd_iommu_disable,
3831 .reenable = amd_iommu_reenable,
3832 .enable_faulting = amd_iommu_enable_faulting,
3833 .get_ir_irq_domain = get_ir_irq_domain,
3834 .get_irq_domain = get_irq_domain,
3837 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3838 struct irq_cfg *irq_cfg,
3839 struct irq_alloc_info *info,
3840 int devid, int index, int sub_handle)
3842 struct irq_2_irte *irte_info = &data->irq_2_irte;
3843 struct msi_msg *msg = &data->msi_entry;
3844 union irte *irte = &data->irte_entry;
3845 struct IO_APIC_route_entry *entry;
3847 data->irq_2_irte.devid = devid;
3848 data->irq_2_irte.index = index + sub_handle;
3850 /* Setup IRTE for IOMMU */
3852 irte->fields.vector = irq_cfg->vector;
3853 irte->fields.int_type = apic->irq_delivery_mode;
3854 irte->fields.destination = irq_cfg->dest_apicid;
3855 irte->fields.dm = apic->irq_dest_mode;
3856 irte->fields.valid = 1;
3858 switch (info->type) {
3859 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3860 /* Setup IOAPIC entry */
3861 entry = info->ioapic_entry;
3862 info->ioapic_entry = NULL;
3863 memset(entry, 0, sizeof(*entry));
3864 entry->vector = index;
3866 entry->trigger = info->ioapic_trigger;
3867 entry->polarity = info->ioapic_polarity;
3868 /* Mask level triggered irqs. */
3869 if (info->ioapic_trigger)
3873 case X86_IRQ_ALLOC_TYPE_HPET:
3874 case X86_IRQ_ALLOC_TYPE_MSI:
3875 case X86_IRQ_ALLOC_TYPE_MSIX:
3876 msg->address_hi = MSI_ADDR_BASE_HI;
3877 msg->address_lo = MSI_ADDR_BASE_LO;
3878 msg->data = irte_info->index;
3887 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3888 unsigned int nr_irqs, void *arg)
3890 struct irq_alloc_info *info = arg;
3891 struct irq_data *irq_data;
3892 struct amd_ir_data *data;
3893 struct irq_cfg *cfg;
3899 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3900 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3904 * With IRQ remapping enabled, don't need contiguous CPU vectors
3905 * to support multiple MSI interrupts.
3907 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3908 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3910 devid = get_devid(info);
3914 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3918 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3919 if (get_irq_table(devid, true))
3920 index = info->ioapic_pin;
3924 index = alloc_irq_index(devid, nr_irqs);
3927 pr_warn("Failed to allocate IRTE\n");
3928 goto out_free_parent;
3931 for (i = 0; i < nr_irqs; i++) {
3932 irq_data = irq_domain_get_irq_data(domain, virq + i);
3933 cfg = irqd_cfg(irq_data);
3934 if (!irq_data || !cfg) {
3940 data = kzalloc(sizeof(*data), GFP_KERNEL);
3944 irq_data->hwirq = (devid << 16) + i;
3945 irq_data->chip_data = data;
3946 irq_data->chip = &amd_ir_chip;
3947 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3948 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3954 for (i--; i >= 0; i--) {
3955 irq_data = irq_domain_get_irq_data(domain, virq + i);
3957 kfree(irq_data->chip_data);
3959 for (i = 0; i < nr_irqs; i++)
3960 free_irte(devid, index + i);
3962 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3966 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3967 unsigned int nr_irqs)
3969 struct irq_2_irte *irte_info;
3970 struct irq_data *irq_data;
3971 struct amd_ir_data *data;
3974 for (i = 0; i < nr_irqs; i++) {
3975 irq_data = irq_domain_get_irq_data(domain, virq + i);
3976 if (irq_data && irq_data->chip_data) {
3977 data = irq_data->chip_data;
3978 irte_info = &data->irq_2_irte;
3979 free_irte(irte_info->devid, irte_info->index);
3983 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3986 static void irq_remapping_activate(struct irq_domain *domain,
3987 struct irq_data *irq_data)
3989 struct amd_ir_data *data = irq_data->chip_data;
3990 struct irq_2_irte *irte_info = &data->irq_2_irte;
3992 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3995 static void irq_remapping_deactivate(struct irq_domain *domain,
3996 struct irq_data *irq_data)
3998 struct amd_ir_data *data = irq_data->chip_data;
3999 struct irq_2_irte *irte_info = &data->irq_2_irte;
4003 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4006 static struct irq_domain_ops amd_ir_domain_ops = {
4007 .alloc = irq_remapping_alloc,
4008 .free = irq_remapping_free,
4009 .activate = irq_remapping_activate,
4010 .deactivate = irq_remapping_deactivate,
4013 static int amd_ir_set_affinity(struct irq_data *data,
4014 const struct cpumask *mask, bool force)
4016 struct amd_ir_data *ir_data = data->chip_data;
4017 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4018 struct irq_cfg *cfg = irqd_cfg(data);
4019 struct irq_data *parent = data->parent_data;
4022 ret = parent->chip->irq_set_affinity(parent, mask, force);
4023 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4027 * Atomically updates the IRTE with the new destination, vector
4028 * and flushes the interrupt entry cache.
4030 ir_data->irte_entry.fields.vector = cfg->vector;
4031 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4032 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4035 * After this point, all the interrupts will start arriving
4036 * at the new destination. So, time to cleanup the previous
4037 * vector allocation.
4039 send_cleanup_vector(cfg);
4041 return IRQ_SET_MASK_OK_DONE;
4044 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4046 struct amd_ir_data *ir_data = irq_data->chip_data;
4048 *msg = ir_data->msi_entry;
4051 static struct irq_chip amd_ir_chip = {
4052 .irq_ack = ir_ack_apic_edge,
4053 .irq_set_affinity = amd_ir_set_affinity,
4054 .irq_compose_msi_msg = ir_compose_msi_msg,
4057 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4059 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4060 if (!iommu->ir_domain)
4063 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4064 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);