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);
117 /****************************************************************************
121 ****************************************************************************/
123 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
125 return container_of(dom, struct protection_domain, domain);
128 static struct iommu_dev_data *alloc_dev_data(u16 devid)
130 struct iommu_dev_data *dev_data;
133 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
137 dev_data->devid = devid;
139 spin_lock_irqsave(&dev_data_list_lock, flags);
140 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
141 spin_unlock_irqrestore(&dev_data_list_lock, flags);
146 static struct iommu_dev_data *search_dev_data(u16 devid)
148 struct iommu_dev_data *dev_data;
151 spin_lock_irqsave(&dev_data_list_lock, flags);
152 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
153 if (dev_data->devid == devid)
160 spin_unlock_irqrestore(&dev_data_list_lock, flags);
165 static struct iommu_dev_data *find_dev_data(u16 devid)
167 struct iommu_dev_data *dev_data;
169 dev_data = search_dev_data(devid);
171 if (dev_data == NULL)
172 dev_data = alloc_dev_data(devid);
177 static inline u16 get_device_id(struct device *dev)
179 struct pci_dev *pdev = to_pci_dev(dev);
181 return PCI_DEVID(pdev->bus->number, pdev->devfn);
184 static struct iommu_dev_data *get_dev_data(struct device *dev)
186 return dev->archdata.iommu;
189 static bool pci_iommuv2_capable(struct pci_dev *pdev)
191 static const int caps[] = {
194 PCI_EXT_CAP_ID_PASID,
198 for (i = 0; i < 3; ++i) {
199 pos = pci_find_ext_capability(pdev, caps[i]);
207 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
209 struct iommu_dev_data *dev_data;
211 dev_data = get_dev_data(&pdev->dev);
213 return dev_data->errata & (1 << erratum) ? true : false;
217 * This function actually applies the mapping to the page table of the
220 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
221 struct unity_map_entry *e)
225 for (addr = e->address_start; addr < e->address_end;
227 if (addr < dma_dom->aperture_size)
228 __set_bit(addr >> PAGE_SHIFT,
229 dma_dom->aperture[0]->bitmap);
234 * Inits the unity mappings required for a specific device
236 static void init_unity_mappings_for_device(struct device *dev,
237 struct dma_ops_domain *dma_dom)
239 struct unity_map_entry *e;
242 devid = get_device_id(dev);
244 list_for_each_entry(e, &amd_iommu_unity_map, list) {
245 if (!(devid >= e->devid_start && devid <= e->devid_end))
247 alloc_unity_mapping(dma_dom, e);
252 * This function checks if the driver got a valid device from the caller to
253 * avoid dereferencing invalid pointers.
255 static bool check_device(struct device *dev)
259 if (!dev || !dev->dma_mask)
263 if (!dev_is_pci(dev))
266 devid = get_device_id(dev);
268 /* Out of our scope? */
269 if (devid > amd_iommu_last_bdf)
272 if (amd_iommu_rlookup_table[devid] == NULL)
278 static void init_iommu_group(struct device *dev)
280 struct dma_ops_domain *dma_domain;
281 struct iommu_domain *domain;
282 struct iommu_group *group;
284 group = iommu_group_get_for_dev(dev);
288 domain = iommu_group_default_domain(group);
292 dma_domain = to_pdomain(domain)->priv;
294 init_unity_mappings_for_device(dev, dma_domain);
296 iommu_group_put(group);
299 static int iommu_init_device(struct device *dev)
301 struct pci_dev *pdev = to_pci_dev(dev);
302 struct iommu_dev_data *dev_data;
304 if (dev->archdata.iommu)
307 dev_data = find_dev_data(get_device_id(dev));
311 if (pci_iommuv2_capable(pdev)) {
312 struct amd_iommu *iommu;
314 iommu = amd_iommu_rlookup_table[dev_data->devid];
315 dev_data->iommu_v2 = iommu->is_iommu_v2;
318 dev->archdata.iommu = dev_data;
320 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
326 static void iommu_ignore_device(struct device *dev)
330 devid = get_device_id(dev);
331 alias = amd_iommu_alias_table[devid];
333 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
334 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
336 amd_iommu_rlookup_table[devid] = NULL;
337 amd_iommu_rlookup_table[alias] = NULL;
340 static void iommu_uninit_device(struct device *dev)
342 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
347 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
350 iommu_group_remove_device(dev);
353 dev->archdata.dma_ops = NULL;
356 * We keep dev_data around for unplugged devices and reuse it when the
357 * device is re-plugged - not doing so would introduce a ton of races.
361 #ifdef CONFIG_AMD_IOMMU_STATS
364 * Initialization code for statistics collection
367 DECLARE_STATS_COUNTER(compl_wait);
368 DECLARE_STATS_COUNTER(cnt_map_single);
369 DECLARE_STATS_COUNTER(cnt_unmap_single);
370 DECLARE_STATS_COUNTER(cnt_map_sg);
371 DECLARE_STATS_COUNTER(cnt_unmap_sg);
372 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
373 DECLARE_STATS_COUNTER(cnt_free_coherent);
374 DECLARE_STATS_COUNTER(cross_page);
375 DECLARE_STATS_COUNTER(domain_flush_single);
376 DECLARE_STATS_COUNTER(domain_flush_all);
377 DECLARE_STATS_COUNTER(alloced_io_mem);
378 DECLARE_STATS_COUNTER(total_map_requests);
379 DECLARE_STATS_COUNTER(complete_ppr);
380 DECLARE_STATS_COUNTER(invalidate_iotlb);
381 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
382 DECLARE_STATS_COUNTER(pri_requests);
384 static struct dentry *stats_dir;
385 static struct dentry *de_fflush;
387 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
389 if (stats_dir == NULL)
392 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
396 static void amd_iommu_stats_init(void)
398 stats_dir = debugfs_create_dir("amd-iommu", NULL);
399 if (stats_dir == NULL)
402 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
403 &amd_iommu_unmap_flush);
405 amd_iommu_stats_add(&compl_wait);
406 amd_iommu_stats_add(&cnt_map_single);
407 amd_iommu_stats_add(&cnt_unmap_single);
408 amd_iommu_stats_add(&cnt_map_sg);
409 amd_iommu_stats_add(&cnt_unmap_sg);
410 amd_iommu_stats_add(&cnt_alloc_coherent);
411 amd_iommu_stats_add(&cnt_free_coherent);
412 amd_iommu_stats_add(&cross_page);
413 amd_iommu_stats_add(&domain_flush_single);
414 amd_iommu_stats_add(&domain_flush_all);
415 amd_iommu_stats_add(&alloced_io_mem);
416 amd_iommu_stats_add(&total_map_requests);
417 amd_iommu_stats_add(&complete_ppr);
418 amd_iommu_stats_add(&invalidate_iotlb);
419 amd_iommu_stats_add(&invalidate_iotlb_all);
420 amd_iommu_stats_add(&pri_requests);
425 /****************************************************************************
427 * Interrupt handling functions
429 ****************************************************************************/
431 static void dump_dte_entry(u16 devid)
435 for (i = 0; i < 4; ++i)
436 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
437 amd_iommu_dev_table[devid].data[i]);
440 static void dump_command(unsigned long phys_addr)
442 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
445 for (i = 0; i < 4; ++i)
446 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
449 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
451 int type, devid, domid, flags;
452 volatile u32 *event = __evt;
457 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
458 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
459 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
460 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
461 address = (u64)(((u64)event[3]) << 32) | event[2];
464 /* Did we hit the erratum? */
465 if (++count == LOOP_TIMEOUT) {
466 pr_err("AMD-Vi: No event written to event log\n");
473 printk(KERN_ERR "AMD-Vi: Event logged [");
476 case EVENT_TYPE_ILL_DEV:
477 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
478 "address=0x%016llx flags=0x%04x]\n",
479 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
481 dump_dte_entry(devid);
483 case EVENT_TYPE_IO_FAULT:
484 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
485 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
486 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
487 domid, address, flags);
489 case EVENT_TYPE_DEV_TAB_ERR:
490 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
491 "address=0x%016llx flags=0x%04x]\n",
492 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
495 case EVENT_TYPE_PAGE_TAB_ERR:
496 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
497 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
498 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
499 domid, address, flags);
501 case EVENT_TYPE_ILL_CMD:
502 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
503 dump_command(address);
505 case EVENT_TYPE_CMD_HARD_ERR:
506 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
507 "flags=0x%04x]\n", address, flags);
509 case EVENT_TYPE_IOTLB_INV_TO:
510 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
511 "address=0x%016llx]\n",
512 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
515 case EVENT_TYPE_INV_DEV_REQ:
516 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
517 "address=0x%016llx flags=0x%04x]\n",
518 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
522 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
525 memset(__evt, 0, 4 * sizeof(u32));
528 static void iommu_poll_events(struct amd_iommu *iommu)
532 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
533 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
535 while (head != tail) {
536 iommu_print_event(iommu, iommu->evt_buf + head);
537 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
540 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
543 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
545 struct amd_iommu_fault fault;
547 INC_STATS_COUNTER(pri_requests);
549 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
550 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
554 fault.address = raw[1];
555 fault.pasid = PPR_PASID(raw[0]);
556 fault.device_id = PPR_DEVID(raw[0]);
557 fault.tag = PPR_TAG(raw[0]);
558 fault.flags = PPR_FLAGS(raw[0]);
560 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
563 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
567 if (iommu->ppr_log == NULL)
570 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
571 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
573 while (head != tail) {
578 raw = (u64 *)(iommu->ppr_log + head);
581 * Hardware bug: Interrupt may arrive before the entry is
582 * written to memory. If this happens we need to wait for the
585 for (i = 0; i < LOOP_TIMEOUT; ++i) {
586 if (PPR_REQ_TYPE(raw[0]) != 0)
591 /* Avoid memcpy function-call overhead */
596 * To detect the hardware bug we need to clear the entry
599 raw[0] = raw[1] = 0UL;
601 /* Update head pointer of hardware ring-buffer */
602 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
603 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
605 /* Handle PPR entry */
606 iommu_handle_ppr_entry(iommu, entry);
608 /* Refresh ring-buffer information */
609 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
610 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
614 irqreturn_t amd_iommu_int_thread(int irq, void *data)
616 struct amd_iommu *iommu = (struct amd_iommu *) data;
617 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
619 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
620 /* Enable EVT and PPR interrupts again */
621 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
622 iommu->mmio_base + MMIO_STATUS_OFFSET);
624 if (status & MMIO_STATUS_EVT_INT_MASK) {
625 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
626 iommu_poll_events(iommu);
629 if (status & MMIO_STATUS_PPR_INT_MASK) {
630 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
631 iommu_poll_ppr_log(iommu);
635 * Hardware bug: ERBT1312
636 * When re-enabling interrupt (by writing 1
637 * to clear the bit), the hardware might also try to set
638 * the interrupt bit in the event status register.
639 * In this scenario, the bit will be set, and disable
640 * subsequent interrupts.
642 * Workaround: The IOMMU driver should read back the
643 * status register and check if the interrupt bits are cleared.
644 * If not, driver will need to go through the interrupt handler
645 * again and re-clear the bits
647 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
652 irqreturn_t amd_iommu_int_handler(int irq, void *data)
654 return IRQ_WAKE_THREAD;
657 /****************************************************************************
659 * IOMMU command queuing functions
661 ****************************************************************************/
663 static int wait_on_sem(volatile u64 *sem)
667 while (*sem == 0 && i < LOOP_TIMEOUT) {
672 if (i == LOOP_TIMEOUT) {
673 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
680 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
681 struct iommu_cmd *cmd,
686 target = iommu->cmd_buf + tail;
687 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
689 /* Copy command to buffer */
690 memcpy(target, cmd, sizeof(*cmd));
692 /* Tell the IOMMU about it */
693 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
696 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
698 WARN_ON(address & 0x7ULL);
700 memset(cmd, 0, sizeof(*cmd));
701 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
702 cmd->data[1] = upper_32_bits(__pa(address));
704 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
707 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
709 memset(cmd, 0, sizeof(*cmd));
710 cmd->data[0] = devid;
711 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
714 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
715 size_t size, u16 domid, int pde)
720 pages = iommu_num_pages(address, size, PAGE_SIZE);
725 * If we have to flush more than one page, flush all
726 * TLB entries for this domain
728 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
732 address &= PAGE_MASK;
734 memset(cmd, 0, sizeof(*cmd));
735 cmd->data[1] |= domid;
736 cmd->data[2] = lower_32_bits(address);
737 cmd->data[3] = upper_32_bits(address);
738 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
739 if (s) /* size bit - we flush more than one 4kb page */
740 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
741 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
742 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
745 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
746 u64 address, size_t size)
751 pages = iommu_num_pages(address, size, PAGE_SIZE);
756 * If we have to flush more than one page, flush all
757 * TLB entries for this domain
759 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
763 address &= PAGE_MASK;
765 memset(cmd, 0, sizeof(*cmd));
766 cmd->data[0] = devid;
767 cmd->data[0] |= (qdep & 0xff) << 24;
768 cmd->data[1] = devid;
769 cmd->data[2] = lower_32_bits(address);
770 cmd->data[3] = upper_32_bits(address);
771 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
773 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
776 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
777 u64 address, bool size)
779 memset(cmd, 0, sizeof(*cmd));
781 address &= ~(0xfffULL);
783 cmd->data[0] = pasid;
784 cmd->data[1] = domid;
785 cmd->data[2] = lower_32_bits(address);
786 cmd->data[3] = upper_32_bits(address);
787 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
788 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
790 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
791 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
794 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
795 int qdep, u64 address, bool size)
797 memset(cmd, 0, sizeof(*cmd));
799 address &= ~(0xfffULL);
801 cmd->data[0] = devid;
802 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
803 cmd->data[0] |= (qdep & 0xff) << 24;
804 cmd->data[1] = devid;
805 cmd->data[1] |= (pasid & 0xff) << 16;
806 cmd->data[2] = lower_32_bits(address);
807 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
808 cmd->data[3] = upper_32_bits(address);
810 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
811 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
814 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
815 int status, int tag, bool gn)
817 memset(cmd, 0, sizeof(*cmd));
819 cmd->data[0] = devid;
821 cmd->data[1] = pasid;
822 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
824 cmd->data[3] = tag & 0x1ff;
825 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
827 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
830 static void build_inv_all(struct iommu_cmd *cmd)
832 memset(cmd, 0, sizeof(*cmd));
833 CMD_SET_TYPE(cmd, CMD_INV_ALL);
836 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
838 memset(cmd, 0, sizeof(*cmd));
839 cmd->data[0] = devid;
840 CMD_SET_TYPE(cmd, CMD_INV_IRT);
844 * Writes the command to the IOMMUs command buffer and informs the
845 * hardware about the new command.
847 static int iommu_queue_command_sync(struct amd_iommu *iommu,
848 struct iommu_cmd *cmd,
851 u32 left, tail, head, next_tail;
854 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
857 spin_lock_irqsave(&iommu->lock, flags);
859 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
860 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
861 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
862 left = (head - next_tail) % iommu->cmd_buf_size;
865 struct iommu_cmd sync_cmd;
866 volatile u64 sem = 0;
869 build_completion_wait(&sync_cmd, (u64)&sem);
870 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
872 spin_unlock_irqrestore(&iommu->lock, flags);
874 if ((ret = wait_on_sem(&sem)) != 0)
880 copy_cmd_to_buffer(iommu, cmd, tail);
882 /* We need to sync now to make sure all commands are processed */
883 iommu->need_sync = sync;
885 spin_unlock_irqrestore(&iommu->lock, flags);
890 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
892 return iommu_queue_command_sync(iommu, cmd, true);
896 * This function queues a completion wait command into the command
899 static int iommu_completion_wait(struct amd_iommu *iommu)
901 struct iommu_cmd cmd;
902 volatile u64 sem = 0;
905 if (!iommu->need_sync)
908 build_completion_wait(&cmd, (u64)&sem);
910 ret = iommu_queue_command_sync(iommu, &cmd, false);
914 return wait_on_sem(&sem);
917 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
919 struct iommu_cmd cmd;
921 build_inv_dte(&cmd, devid);
923 return iommu_queue_command(iommu, &cmd);
926 static void iommu_flush_dte_all(struct amd_iommu *iommu)
930 for (devid = 0; devid <= 0xffff; ++devid)
931 iommu_flush_dte(iommu, devid);
933 iommu_completion_wait(iommu);
937 * This function uses heavy locking and may disable irqs for some time. But
938 * this is no issue because it is only called during resume.
940 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
944 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
945 struct iommu_cmd cmd;
946 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
948 iommu_queue_command(iommu, &cmd);
951 iommu_completion_wait(iommu);
954 static void iommu_flush_all(struct amd_iommu *iommu)
956 struct iommu_cmd cmd;
960 iommu_queue_command(iommu, &cmd);
961 iommu_completion_wait(iommu);
964 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
966 struct iommu_cmd cmd;
968 build_inv_irt(&cmd, devid);
970 iommu_queue_command(iommu, &cmd);
973 static void iommu_flush_irt_all(struct amd_iommu *iommu)
977 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
978 iommu_flush_irt(iommu, devid);
980 iommu_completion_wait(iommu);
983 void iommu_flush_all_caches(struct amd_iommu *iommu)
985 if (iommu_feature(iommu, FEATURE_IA)) {
986 iommu_flush_all(iommu);
988 iommu_flush_dte_all(iommu);
989 iommu_flush_irt_all(iommu);
990 iommu_flush_tlb_all(iommu);
995 * Command send function for flushing on-device TLB
997 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
998 u64 address, size_t size)
1000 struct amd_iommu *iommu;
1001 struct iommu_cmd cmd;
1004 qdep = dev_data->ats.qdep;
1005 iommu = amd_iommu_rlookup_table[dev_data->devid];
1007 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1009 return iommu_queue_command(iommu, &cmd);
1013 * Command send function for invalidating a device table entry
1015 static int device_flush_dte(struct iommu_dev_data *dev_data)
1017 struct amd_iommu *iommu;
1021 iommu = amd_iommu_rlookup_table[dev_data->devid];
1022 alias = amd_iommu_alias_table[dev_data->devid];
1024 ret = iommu_flush_dte(iommu, dev_data->devid);
1025 if (!ret && alias != dev_data->devid)
1026 ret = iommu_flush_dte(iommu, alias);
1030 if (dev_data->ats.enabled)
1031 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1037 * TLB invalidation function which is called from the mapping functions.
1038 * It invalidates a single PTE if the range to flush is within a single
1039 * page. Otherwise it flushes the whole TLB of the IOMMU.
1041 static void __domain_flush_pages(struct protection_domain *domain,
1042 u64 address, size_t size, int pde)
1044 struct iommu_dev_data *dev_data;
1045 struct iommu_cmd cmd;
1048 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1050 for (i = 0; i < amd_iommus_present; ++i) {
1051 if (!domain->dev_iommu[i])
1055 * Devices of this domain are behind this IOMMU
1056 * We need a TLB flush
1058 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1061 list_for_each_entry(dev_data, &domain->dev_list, list) {
1063 if (!dev_data->ats.enabled)
1066 ret |= device_flush_iotlb(dev_data, address, size);
1072 static void domain_flush_pages(struct protection_domain *domain,
1073 u64 address, size_t size)
1075 __domain_flush_pages(domain, address, size, 0);
1078 /* Flush the whole IO/TLB for a given protection domain */
1079 static void domain_flush_tlb(struct protection_domain *domain)
1081 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1084 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1085 static void domain_flush_tlb_pde(struct protection_domain *domain)
1087 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1090 static void domain_flush_complete(struct protection_domain *domain)
1094 for (i = 0; i < amd_iommus_present; ++i) {
1095 if (!domain->dev_iommu[i])
1099 * Devices of this domain are behind this IOMMU
1100 * We need to wait for completion of all commands.
1102 iommu_completion_wait(amd_iommus[i]);
1108 * This function flushes the DTEs for all devices in domain
1110 static void domain_flush_devices(struct protection_domain *domain)
1112 struct iommu_dev_data *dev_data;
1114 list_for_each_entry(dev_data, &domain->dev_list, list)
1115 device_flush_dte(dev_data);
1118 /****************************************************************************
1120 * The functions below are used the create the page table mappings for
1121 * unity mapped regions.
1123 ****************************************************************************/
1126 * This function is used to add another level to an IO page table. Adding
1127 * another level increases the size of the address space by 9 bits to a size up
1130 static bool increase_address_space(struct protection_domain *domain,
1135 if (domain->mode == PAGE_MODE_6_LEVEL)
1136 /* address space already 64 bit large */
1139 pte = (void *)get_zeroed_page(gfp);
1143 *pte = PM_LEVEL_PDE(domain->mode,
1144 virt_to_phys(domain->pt_root));
1145 domain->pt_root = pte;
1147 domain->updated = true;
1152 static u64 *alloc_pte(struct protection_domain *domain,
1153 unsigned long address,
1154 unsigned long page_size,
1161 BUG_ON(!is_power_of_2(page_size));
1163 while (address > PM_LEVEL_SIZE(domain->mode))
1164 increase_address_space(domain, gfp);
1166 level = domain->mode - 1;
1167 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1168 address = PAGE_SIZE_ALIGN(address, page_size);
1169 end_lvl = PAGE_SIZE_LEVEL(page_size);
1171 while (level > end_lvl) {
1172 if (!IOMMU_PTE_PRESENT(*pte)) {
1173 page = (u64 *)get_zeroed_page(gfp);
1176 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1179 /* No level skipping support yet */
1180 if (PM_PTE_LEVEL(*pte) != level)
1185 pte = IOMMU_PTE_PAGE(*pte);
1187 if (pte_page && level == end_lvl)
1190 pte = &pte[PM_LEVEL_INDEX(level, address)];
1197 * This function checks if there is a PTE for a given dma address. If
1198 * there is one, it returns the pointer to it.
1200 static u64 *fetch_pte(struct protection_domain *domain,
1201 unsigned long address,
1202 unsigned long *page_size)
1207 if (address > PM_LEVEL_SIZE(domain->mode))
1210 level = domain->mode - 1;
1211 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1212 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1217 if (!IOMMU_PTE_PRESENT(*pte))
1221 if (PM_PTE_LEVEL(*pte) == 7 ||
1222 PM_PTE_LEVEL(*pte) == 0)
1225 /* No level skipping support yet */
1226 if (PM_PTE_LEVEL(*pte) != level)
1231 /* Walk to the next level */
1232 pte = IOMMU_PTE_PAGE(*pte);
1233 pte = &pte[PM_LEVEL_INDEX(level, address)];
1234 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1237 if (PM_PTE_LEVEL(*pte) == 0x07) {
1238 unsigned long pte_mask;
1241 * If we have a series of large PTEs, make
1242 * sure to return a pointer to the first one.
1244 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1245 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1246 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1253 * Generic mapping functions. It maps a physical address into a DMA
1254 * address space. It allocates the page table pages if necessary.
1255 * In the future it can be extended to a generic mapping function
1256 * supporting all features of AMD IOMMU page tables like level skipping
1257 * and full 64 bit address spaces.
1259 static int iommu_map_page(struct protection_domain *dom,
1260 unsigned long bus_addr,
1261 unsigned long phys_addr,
1263 unsigned long page_size)
1268 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1269 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1271 if (!(prot & IOMMU_PROT_MASK))
1274 count = PAGE_SIZE_PTE_COUNT(page_size);
1275 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1280 for (i = 0; i < count; ++i)
1281 if (IOMMU_PTE_PRESENT(pte[i]))
1285 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1286 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1288 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1290 if (prot & IOMMU_PROT_IR)
1291 __pte |= IOMMU_PTE_IR;
1292 if (prot & IOMMU_PROT_IW)
1293 __pte |= IOMMU_PTE_IW;
1295 for (i = 0; i < count; ++i)
1303 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1304 unsigned long bus_addr,
1305 unsigned long page_size)
1307 unsigned long long unmapped;
1308 unsigned long unmap_size;
1311 BUG_ON(!is_power_of_2(page_size));
1315 while (unmapped < page_size) {
1317 pte = fetch_pte(dom, bus_addr, &unmap_size);
1322 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1323 for (i = 0; i < count; i++)
1327 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1328 unmapped += unmap_size;
1331 BUG_ON(unmapped && !is_power_of_2(unmapped));
1336 /****************************************************************************
1338 * The next functions belong to the address allocator for the dma_ops
1339 * interface functions. They work like the allocators in the other IOMMU
1340 * drivers. Its basically a bitmap which marks the allocated pages in
1341 * the aperture. Maybe it could be enhanced in the future to a more
1342 * efficient allocator.
1344 ****************************************************************************/
1347 * The address allocator core functions.
1349 * called with domain->lock held
1353 * Used to reserve address ranges in the aperture (e.g. for exclusion
1356 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1357 unsigned long start_page,
1360 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1362 if (start_page + pages > last_page)
1363 pages = last_page - start_page;
1365 for (i = start_page; i < start_page + pages; ++i) {
1366 int index = i / APERTURE_RANGE_PAGES;
1367 int page = i % APERTURE_RANGE_PAGES;
1368 __set_bit(page, dom->aperture[index]->bitmap);
1373 * This function is used to add a new aperture range to an existing
1374 * aperture in case of dma_ops domain allocation or address allocation
1377 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1378 bool populate, gfp_t gfp)
1380 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1381 struct amd_iommu *iommu;
1382 unsigned long i, old_size, pte_pgsize;
1384 #ifdef CONFIG_IOMMU_STRESS
1388 if (index >= APERTURE_MAX_RANGES)
1391 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1392 if (!dma_dom->aperture[index])
1395 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1396 if (!dma_dom->aperture[index]->bitmap)
1399 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1402 unsigned long address = dma_dom->aperture_size;
1403 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1404 u64 *pte, *pte_page;
1406 for (i = 0; i < num_ptes; ++i) {
1407 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1412 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1414 address += APERTURE_RANGE_SIZE / 64;
1418 old_size = dma_dom->aperture_size;
1419 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1421 /* Reserve address range used for MSI messages */
1422 if (old_size < MSI_ADDR_BASE_LO &&
1423 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1424 unsigned long spage;
1427 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1428 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1430 dma_ops_reserve_addresses(dma_dom, spage, pages);
1433 /* Initialize the exclusion range if necessary */
1434 for_each_iommu(iommu) {
1435 if (iommu->exclusion_start &&
1436 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1437 && iommu->exclusion_start < dma_dom->aperture_size) {
1438 unsigned long startpage;
1439 int pages = iommu_num_pages(iommu->exclusion_start,
1440 iommu->exclusion_length,
1442 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1443 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1448 * Check for areas already mapped as present in the new aperture
1449 * range and mark those pages as reserved in the allocator. Such
1450 * mappings may already exist as a result of requested unity
1451 * mappings for devices.
1453 for (i = dma_dom->aperture[index]->offset;
1454 i < dma_dom->aperture_size;
1456 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1457 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1460 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1464 update_domain(&dma_dom->domain);
1469 update_domain(&dma_dom->domain);
1471 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1473 kfree(dma_dom->aperture[index]);
1474 dma_dom->aperture[index] = NULL;
1479 static unsigned long dma_ops_area_alloc(struct device *dev,
1480 struct dma_ops_domain *dom,
1482 unsigned long align_mask,
1484 unsigned long start)
1486 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1487 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1488 int i = start >> APERTURE_RANGE_SHIFT;
1489 unsigned long boundary_size, mask;
1490 unsigned long address = -1;
1491 unsigned long limit;
1493 next_bit >>= PAGE_SHIFT;
1495 mask = dma_get_seg_boundary(dev);
1497 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1498 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1500 for (;i < max_index; ++i) {
1501 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1503 if (dom->aperture[i]->offset >= dma_mask)
1506 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1507 dma_mask >> PAGE_SHIFT);
1509 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1510 limit, next_bit, pages, 0,
1511 boundary_size, align_mask);
1512 if (address != -1) {
1513 address = dom->aperture[i]->offset +
1514 (address << PAGE_SHIFT);
1515 dom->next_address = address + (pages << PAGE_SHIFT);
1525 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1526 struct dma_ops_domain *dom,
1528 unsigned long align_mask,
1531 unsigned long address;
1533 #ifdef CONFIG_IOMMU_STRESS
1534 dom->next_address = 0;
1535 dom->need_flush = true;
1538 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1539 dma_mask, dom->next_address);
1541 if (address == -1) {
1542 dom->next_address = 0;
1543 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1545 dom->need_flush = true;
1548 if (unlikely(address == -1))
1549 address = DMA_ERROR_CODE;
1551 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1557 * The address free function.
1559 * called with domain->lock held
1561 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1562 unsigned long address,
1565 unsigned i = address >> APERTURE_RANGE_SHIFT;
1566 struct aperture_range *range = dom->aperture[i];
1568 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1570 #ifdef CONFIG_IOMMU_STRESS
1575 if (address >= dom->next_address)
1576 dom->need_flush = true;
1578 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1580 bitmap_clear(range->bitmap, address, pages);
1584 /****************************************************************************
1586 * The next functions belong to the domain allocation. A domain is
1587 * allocated for every IOMMU as the default domain. If device isolation
1588 * is enabled, every device get its own domain. The most important thing
1589 * about domains is the page table mapping the DMA address space they
1592 ****************************************************************************/
1595 * This function adds a protection domain to the global protection domain list
1597 static void add_domain_to_list(struct protection_domain *domain)
1599 unsigned long flags;
1601 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1602 list_add(&domain->list, &amd_iommu_pd_list);
1603 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1607 * This function removes a protection domain to the global
1608 * protection domain list
1610 static void del_domain_from_list(struct protection_domain *domain)
1612 unsigned long flags;
1614 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1615 list_del(&domain->list);
1616 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1619 static u16 domain_id_alloc(void)
1621 unsigned long flags;
1624 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1625 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1627 if (id > 0 && id < MAX_DOMAIN_ID)
1628 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1631 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1636 static void domain_id_free(int id)
1638 unsigned long flags;
1640 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1641 if (id > 0 && id < MAX_DOMAIN_ID)
1642 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1643 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1646 #define DEFINE_FREE_PT_FN(LVL, FN) \
1647 static void free_pt_##LVL (unsigned long __pt) \
1655 for (i = 0; i < 512; ++i) { \
1656 /* PTE present? */ \
1657 if (!IOMMU_PTE_PRESENT(pt[i])) \
1661 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1662 PM_PTE_LEVEL(pt[i]) == 7) \
1665 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1668 free_page((unsigned long)pt); \
1671 DEFINE_FREE_PT_FN(l2, free_page)
1672 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1673 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1674 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1675 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1677 static void free_pagetable(struct protection_domain *domain)
1679 unsigned long root = (unsigned long)domain->pt_root;
1681 switch (domain->mode) {
1682 case PAGE_MODE_NONE:
1684 case PAGE_MODE_1_LEVEL:
1687 case PAGE_MODE_2_LEVEL:
1690 case PAGE_MODE_3_LEVEL:
1693 case PAGE_MODE_4_LEVEL:
1696 case PAGE_MODE_5_LEVEL:
1699 case PAGE_MODE_6_LEVEL:
1707 static void free_gcr3_tbl_level1(u64 *tbl)
1712 for (i = 0; i < 512; ++i) {
1713 if (!(tbl[i] & GCR3_VALID))
1716 ptr = __va(tbl[i] & PAGE_MASK);
1718 free_page((unsigned long)ptr);
1722 static void free_gcr3_tbl_level2(u64 *tbl)
1727 for (i = 0; i < 512; ++i) {
1728 if (!(tbl[i] & GCR3_VALID))
1731 ptr = __va(tbl[i] & PAGE_MASK);
1733 free_gcr3_tbl_level1(ptr);
1737 static void free_gcr3_table(struct protection_domain *domain)
1739 if (domain->glx == 2)
1740 free_gcr3_tbl_level2(domain->gcr3_tbl);
1741 else if (domain->glx == 1)
1742 free_gcr3_tbl_level1(domain->gcr3_tbl);
1744 BUG_ON(domain->glx != 0);
1746 free_page((unsigned long)domain->gcr3_tbl);
1750 * Free a domain, only used if something went wrong in the
1751 * allocation path and we need to free an already allocated page table
1753 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1760 del_domain_from_list(&dom->domain);
1762 free_pagetable(&dom->domain);
1764 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1765 if (!dom->aperture[i])
1767 free_page((unsigned long)dom->aperture[i]->bitmap);
1768 kfree(dom->aperture[i]);
1775 * Allocates a new protection domain usable for the dma_ops functions.
1776 * It also initializes the page table and the address allocator data
1777 * structures required for the dma_ops interface
1779 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1781 struct dma_ops_domain *dma_dom;
1783 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1787 if (protection_domain_init(&dma_dom->domain))
1790 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1791 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1792 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1793 dma_dom->domain.priv = dma_dom;
1794 if (!dma_dom->domain.pt_root)
1797 dma_dom->need_flush = false;
1799 add_domain_to_list(&dma_dom->domain);
1801 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1805 * mark the first page as allocated so we never return 0 as
1806 * a valid dma-address. So we can use 0 as error value
1808 dma_dom->aperture[0]->bitmap[0] = 1;
1809 dma_dom->next_address = 0;
1815 dma_ops_domain_free(dma_dom);
1821 * little helper function to check whether a given protection domain is a
1824 static bool dma_ops_domain(struct protection_domain *domain)
1826 return domain->flags & PD_DMA_OPS_MASK;
1829 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1834 if (domain->mode != PAGE_MODE_NONE)
1835 pte_root = virt_to_phys(domain->pt_root);
1837 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1838 << DEV_ENTRY_MODE_SHIFT;
1839 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1841 flags = amd_iommu_dev_table[devid].data[1];
1844 flags |= DTE_FLAG_IOTLB;
1846 if (domain->flags & PD_IOMMUV2_MASK) {
1847 u64 gcr3 = __pa(domain->gcr3_tbl);
1848 u64 glx = domain->glx;
1851 pte_root |= DTE_FLAG_GV;
1852 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1854 /* First mask out possible old values for GCR3 table */
1855 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1858 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1861 /* Encode GCR3 table into DTE */
1862 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1865 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1868 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1872 flags &= ~(0xffffUL);
1873 flags |= domain->id;
1875 amd_iommu_dev_table[devid].data[1] = flags;
1876 amd_iommu_dev_table[devid].data[0] = pte_root;
1879 static void clear_dte_entry(u16 devid)
1881 /* remove entry from the device table seen by the hardware */
1882 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1883 amd_iommu_dev_table[devid].data[1] = 0;
1885 amd_iommu_apply_erratum_63(devid);
1888 static void do_attach(struct iommu_dev_data *dev_data,
1889 struct protection_domain *domain)
1891 struct amd_iommu *iommu;
1895 iommu = amd_iommu_rlookup_table[dev_data->devid];
1896 alias = amd_iommu_alias_table[dev_data->devid];
1897 ats = dev_data->ats.enabled;
1899 /* Update data structures */
1900 dev_data->domain = domain;
1901 list_add(&dev_data->list, &domain->dev_list);
1903 /* Do reference counting */
1904 domain->dev_iommu[iommu->index] += 1;
1905 domain->dev_cnt += 1;
1907 /* Update device table */
1908 set_dte_entry(dev_data->devid, domain, ats);
1909 if (alias != dev_data->devid)
1910 set_dte_entry(dev_data->devid, domain, ats);
1912 device_flush_dte(dev_data);
1915 static void do_detach(struct iommu_dev_data *dev_data)
1917 struct amd_iommu *iommu;
1920 iommu = amd_iommu_rlookup_table[dev_data->devid];
1921 alias = amd_iommu_alias_table[dev_data->devid];
1923 /* decrease reference counters */
1924 dev_data->domain->dev_iommu[iommu->index] -= 1;
1925 dev_data->domain->dev_cnt -= 1;
1927 /* Update data structures */
1928 dev_data->domain = NULL;
1929 list_del(&dev_data->list);
1930 clear_dte_entry(dev_data->devid);
1931 if (alias != dev_data->devid)
1932 clear_dte_entry(alias);
1934 /* Flush the DTE entry */
1935 device_flush_dte(dev_data);
1939 * If a device is not yet associated with a domain, this function does
1940 * assigns it visible for the hardware
1942 static int __attach_device(struct iommu_dev_data *dev_data,
1943 struct protection_domain *domain)
1948 * Must be called with IRQs disabled. Warn here to detect early
1951 WARN_ON(!irqs_disabled());
1954 spin_lock(&domain->lock);
1957 if (dev_data->domain != NULL)
1960 /* Attach alias group root */
1961 do_attach(dev_data, domain);
1968 spin_unlock(&domain->lock);
1974 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1976 pci_disable_ats(pdev);
1977 pci_disable_pri(pdev);
1978 pci_disable_pasid(pdev);
1981 /* FIXME: Change generic reset-function to do the same */
1982 static int pri_reset_while_enabled(struct pci_dev *pdev)
1987 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1991 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1992 control |= PCI_PRI_CTRL_RESET;
1993 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1998 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2003 /* FIXME: Hardcode number of outstanding requests for now */
2005 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2007 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2009 /* Only allow access to user-accessible pages */
2010 ret = pci_enable_pasid(pdev, 0);
2014 /* First reset the PRI state of the device */
2015 ret = pci_reset_pri(pdev);
2020 ret = pci_enable_pri(pdev, reqs);
2025 ret = pri_reset_while_enabled(pdev);
2030 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2037 pci_disable_pri(pdev);
2038 pci_disable_pasid(pdev);
2043 /* FIXME: Move this to PCI code */
2044 #define PCI_PRI_TLP_OFF (1 << 15)
2046 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2051 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2055 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2057 return (status & PCI_PRI_TLP_OFF) ? true : false;
2061 * If a device is not yet associated with a domain, this function
2062 * assigns it visible for the hardware
2064 static int attach_device(struct device *dev,
2065 struct protection_domain *domain)
2067 struct pci_dev *pdev = to_pci_dev(dev);
2068 struct iommu_dev_data *dev_data;
2069 unsigned long flags;
2072 dev_data = get_dev_data(dev);
2074 if (domain->flags & PD_IOMMUV2_MASK) {
2075 if (!dev_data->passthrough)
2078 if (dev_data->iommu_v2) {
2079 if (pdev_iommuv2_enable(pdev) != 0)
2082 dev_data->ats.enabled = true;
2083 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2084 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2086 } else if (amd_iommu_iotlb_sup &&
2087 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2088 dev_data->ats.enabled = true;
2089 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2092 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2093 ret = __attach_device(dev_data, domain);
2094 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2097 * We might boot into a crash-kernel here. The crashed kernel
2098 * left the caches in the IOMMU dirty. So we have to flush
2099 * here to evict all dirty stuff.
2101 domain_flush_tlb_pde(domain);
2107 * Removes a device from a protection domain (unlocked)
2109 static void __detach_device(struct iommu_dev_data *dev_data)
2111 struct protection_domain *domain;
2114 * Must be called with IRQs disabled. Warn here to detect early
2117 WARN_ON(!irqs_disabled());
2119 if (WARN_ON(!dev_data->domain))
2122 domain = dev_data->domain;
2124 spin_lock(&domain->lock);
2126 do_detach(dev_data);
2128 spin_unlock(&domain->lock);
2132 * Removes a device from a protection domain (with devtable_lock held)
2134 static void detach_device(struct device *dev)
2136 struct protection_domain *domain;
2137 struct iommu_dev_data *dev_data;
2138 unsigned long flags;
2140 dev_data = get_dev_data(dev);
2141 domain = dev_data->domain;
2143 /* lock device table */
2144 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2145 __detach_device(dev_data);
2146 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2148 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2149 pdev_iommuv2_disable(to_pci_dev(dev));
2150 else if (dev_data->ats.enabled)
2151 pci_disable_ats(to_pci_dev(dev));
2153 dev_data->ats.enabled = false;
2156 static int amd_iommu_add_device(struct device *dev)
2158 struct iommu_dev_data *dev_data;
2159 struct iommu_domain *domain;
2160 struct amd_iommu *iommu;
2164 if (!check_device(dev) || get_dev_data(dev))
2167 devid = get_device_id(dev);
2168 iommu = amd_iommu_rlookup_table[devid];
2170 ret = iommu_init_device(dev);
2172 if (ret != -ENOTSUPP)
2173 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2176 iommu_ignore_device(dev);
2177 dev->archdata.dma_ops = &nommu_dma_ops;
2180 init_iommu_group(dev);
2182 dev_data = get_dev_data(dev);
2186 if (iommu_pass_through || dev_data->iommu_v2)
2187 iommu_request_dm_for_dev(dev);
2189 /* Domains are initialized for this device - have a look what we ended up with */
2190 domain = iommu_get_domain_for_dev(dev);
2191 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2192 dev_data->passthrough = true;
2194 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2197 iommu_completion_wait(iommu);
2202 static void amd_iommu_remove_device(struct device *dev)
2204 struct amd_iommu *iommu;
2207 if (!check_device(dev))
2210 devid = get_device_id(dev);
2211 iommu = amd_iommu_rlookup_table[devid];
2213 iommu_uninit_device(dev);
2214 iommu_completion_wait(iommu);
2217 /*****************************************************************************
2219 * The next functions belong to the dma_ops mapping/unmapping code.
2221 *****************************************************************************/
2224 * In the dma_ops path we only have the struct device. This function
2225 * finds the corresponding IOMMU, the protection domain and the
2226 * requestor id for a given device.
2227 * If the device is not yet associated with a domain this is also done
2230 static struct protection_domain *get_domain(struct device *dev)
2232 struct protection_domain *domain;
2233 struct iommu_domain *io_domain;
2235 if (!check_device(dev))
2236 return ERR_PTR(-EINVAL);
2238 io_domain = iommu_get_domain_for_dev(dev);
2242 domain = to_pdomain(io_domain);
2243 if (!dma_ops_domain(domain))
2244 return ERR_PTR(-EBUSY);
2249 static void update_device_table(struct protection_domain *domain)
2251 struct iommu_dev_data *dev_data;
2253 list_for_each_entry(dev_data, &domain->dev_list, list)
2254 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2257 static void update_domain(struct protection_domain *domain)
2259 if (!domain->updated)
2262 update_device_table(domain);
2264 domain_flush_devices(domain);
2265 domain_flush_tlb_pde(domain);
2267 domain->updated = false;
2271 * This function fetches the PTE for a given address in the aperture
2273 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2274 unsigned long address)
2276 struct aperture_range *aperture;
2277 u64 *pte, *pte_page;
2279 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2283 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2285 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2287 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2289 pte += PM_LEVEL_INDEX(0, address);
2291 update_domain(&dom->domain);
2297 * This is the generic map function. It maps one 4kb page at paddr to
2298 * the given address in the DMA address space for the domain.
2300 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2301 unsigned long address,
2307 WARN_ON(address > dom->aperture_size);
2311 pte = dma_ops_get_pte(dom, address);
2313 return DMA_ERROR_CODE;
2315 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2317 if (direction == DMA_TO_DEVICE)
2318 __pte |= IOMMU_PTE_IR;
2319 else if (direction == DMA_FROM_DEVICE)
2320 __pte |= IOMMU_PTE_IW;
2321 else if (direction == DMA_BIDIRECTIONAL)
2322 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2328 return (dma_addr_t)address;
2332 * The generic unmapping function for on page in the DMA address space.
2334 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2335 unsigned long address)
2337 struct aperture_range *aperture;
2340 if (address >= dom->aperture_size)
2343 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2347 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2351 pte += PM_LEVEL_INDEX(0, address);
2359 * This function contains common code for mapping of a physically
2360 * contiguous memory region into DMA address space. It is used by all
2361 * mapping functions provided with this IOMMU driver.
2362 * Must be called with the domain lock held.
2364 static dma_addr_t __map_single(struct device *dev,
2365 struct dma_ops_domain *dma_dom,
2372 dma_addr_t offset = paddr & ~PAGE_MASK;
2373 dma_addr_t address, start, ret;
2375 unsigned long align_mask = 0;
2378 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2381 INC_STATS_COUNTER(total_map_requests);
2384 INC_STATS_COUNTER(cross_page);
2387 align_mask = (1UL << get_order(size)) - 1;
2390 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2392 if (unlikely(address == DMA_ERROR_CODE)) {
2394 * setting next_address here will let the address
2395 * allocator only scan the new allocated range in the
2396 * first run. This is a small optimization.
2398 dma_dom->next_address = dma_dom->aperture_size;
2400 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2404 * aperture was successfully enlarged by 128 MB, try
2411 for (i = 0; i < pages; ++i) {
2412 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2413 if (ret == DMA_ERROR_CODE)
2421 ADD_STATS_COUNTER(alloced_io_mem, size);
2423 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2424 domain_flush_tlb(&dma_dom->domain);
2425 dma_dom->need_flush = false;
2426 } else if (unlikely(amd_iommu_np_cache))
2427 domain_flush_pages(&dma_dom->domain, address, size);
2434 for (--i; i >= 0; --i) {
2436 dma_ops_domain_unmap(dma_dom, start);
2439 dma_ops_free_addresses(dma_dom, address, pages);
2441 return DMA_ERROR_CODE;
2445 * Does the reverse of the __map_single function. Must be called with
2446 * the domain lock held too
2448 static void __unmap_single(struct dma_ops_domain *dma_dom,
2449 dma_addr_t dma_addr,
2453 dma_addr_t flush_addr;
2454 dma_addr_t i, start;
2457 if ((dma_addr == DMA_ERROR_CODE) ||
2458 (dma_addr + size > dma_dom->aperture_size))
2461 flush_addr = dma_addr;
2462 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2463 dma_addr &= PAGE_MASK;
2466 for (i = 0; i < pages; ++i) {
2467 dma_ops_domain_unmap(dma_dom, start);
2471 SUB_STATS_COUNTER(alloced_io_mem, size);
2473 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2475 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2476 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2477 dma_dom->need_flush = false;
2482 * The exported map_single function for dma_ops.
2484 static dma_addr_t map_page(struct device *dev, struct page *page,
2485 unsigned long offset, size_t size,
2486 enum dma_data_direction dir,
2487 struct dma_attrs *attrs)
2489 unsigned long flags;
2490 struct protection_domain *domain;
2493 phys_addr_t paddr = page_to_phys(page) + offset;
2495 INC_STATS_COUNTER(cnt_map_single);
2497 domain = get_domain(dev);
2498 if (PTR_ERR(domain) == -EINVAL)
2499 return (dma_addr_t)paddr;
2500 else if (IS_ERR(domain))
2501 return DMA_ERROR_CODE;
2503 dma_mask = *dev->dma_mask;
2505 spin_lock_irqsave(&domain->lock, flags);
2507 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2509 if (addr == DMA_ERROR_CODE)
2512 domain_flush_complete(domain);
2515 spin_unlock_irqrestore(&domain->lock, flags);
2521 * The exported unmap_single function for dma_ops.
2523 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2524 enum dma_data_direction dir, struct dma_attrs *attrs)
2526 unsigned long flags;
2527 struct protection_domain *domain;
2529 INC_STATS_COUNTER(cnt_unmap_single);
2531 domain = get_domain(dev);
2535 spin_lock_irqsave(&domain->lock, flags);
2537 __unmap_single(domain->priv, dma_addr, size, dir);
2539 domain_flush_complete(domain);
2541 spin_unlock_irqrestore(&domain->lock, flags);
2545 * The exported map_sg function for dma_ops (handles scatter-gather
2548 static int map_sg(struct device *dev, struct scatterlist *sglist,
2549 int nelems, enum dma_data_direction dir,
2550 struct dma_attrs *attrs)
2552 unsigned long flags;
2553 struct protection_domain *domain;
2555 struct scatterlist *s;
2557 int mapped_elems = 0;
2560 INC_STATS_COUNTER(cnt_map_sg);
2562 domain = get_domain(dev);
2566 dma_mask = *dev->dma_mask;
2568 spin_lock_irqsave(&domain->lock, flags);
2570 for_each_sg(sglist, s, nelems, i) {
2573 s->dma_address = __map_single(dev, domain->priv,
2574 paddr, s->length, dir, false,
2577 if (s->dma_address) {
2578 s->dma_length = s->length;
2584 domain_flush_complete(domain);
2587 spin_unlock_irqrestore(&domain->lock, flags);
2589 return mapped_elems;
2591 for_each_sg(sglist, s, mapped_elems, i) {
2593 __unmap_single(domain->priv, s->dma_address,
2594 s->dma_length, dir);
2595 s->dma_address = s->dma_length = 0;
2604 * The exported map_sg function for dma_ops (handles scatter-gather
2607 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2608 int nelems, enum dma_data_direction dir,
2609 struct dma_attrs *attrs)
2611 unsigned long flags;
2612 struct protection_domain *domain;
2613 struct scatterlist *s;
2616 INC_STATS_COUNTER(cnt_unmap_sg);
2618 domain = get_domain(dev);
2622 spin_lock_irqsave(&domain->lock, flags);
2624 for_each_sg(sglist, s, nelems, i) {
2625 __unmap_single(domain->priv, s->dma_address,
2626 s->dma_length, dir);
2627 s->dma_address = s->dma_length = 0;
2630 domain_flush_complete(domain);
2632 spin_unlock_irqrestore(&domain->lock, flags);
2636 * The exported alloc_coherent function for dma_ops.
2638 static void *alloc_coherent(struct device *dev, size_t size,
2639 dma_addr_t *dma_addr, gfp_t flag,
2640 struct dma_attrs *attrs)
2642 u64 dma_mask = dev->coherent_dma_mask;
2643 struct protection_domain *domain;
2644 unsigned long flags;
2647 INC_STATS_COUNTER(cnt_alloc_coherent);
2649 domain = get_domain(dev);
2650 if (PTR_ERR(domain) == -EINVAL) {
2651 page = alloc_pages(flag, get_order(size));
2652 *dma_addr = page_to_phys(page);
2653 return page_address(page);
2654 } else if (IS_ERR(domain))
2657 size = PAGE_ALIGN(size);
2658 dma_mask = dev->coherent_dma_mask;
2659 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2662 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2664 if (!(flag & __GFP_WAIT))
2667 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2674 dma_mask = *dev->dma_mask;
2676 spin_lock_irqsave(&domain->lock, flags);
2678 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2679 size, DMA_BIDIRECTIONAL, true, dma_mask);
2681 if (*dma_addr == DMA_ERROR_CODE) {
2682 spin_unlock_irqrestore(&domain->lock, flags);
2686 domain_flush_complete(domain);
2688 spin_unlock_irqrestore(&domain->lock, flags);
2690 return page_address(page);
2694 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2695 __free_pages(page, get_order(size));
2701 * The exported free_coherent function for dma_ops.
2703 static void free_coherent(struct device *dev, size_t size,
2704 void *virt_addr, dma_addr_t dma_addr,
2705 struct dma_attrs *attrs)
2707 struct protection_domain *domain;
2708 unsigned long flags;
2711 INC_STATS_COUNTER(cnt_free_coherent);
2713 page = virt_to_page(virt_addr);
2714 size = PAGE_ALIGN(size);
2716 domain = get_domain(dev);
2720 spin_lock_irqsave(&domain->lock, flags);
2722 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2724 domain_flush_complete(domain);
2726 spin_unlock_irqrestore(&domain->lock, flags);
2729 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2730 __free_pages(page, get_order(size));
2734 * This function is called by the DMA layer to find out if we can handle a
2735 * particular device. It is part of the dma_ops.
2737 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2739 return check_device(dev);
2742 static struct dma_map_ops amd_iommu_dma_ops = {
2743 .alloc = alloc_coherent,
2744 .free = free_coherent,
2745 .map_page = map_page,
2746 .unmap_page = unmap_page,
2748 .unmap_sg = unmap_sg,
2749 .dma_supported = amd_iommu_dma_supported,
2752 int __init amd_iommu_init_api(void)
2754 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2757 int __init amd_iommu_init_dma_ops(void)
2759 swiotlb = iommu_pass_through ? 1 : 0;
2763 * In case we don't initialize SWIOTLB (actually the common case
2764 * when AMD IOMMU is enabled), make sure there are global
2765 * dma_ops set as a fall-back for devices not handled by this
2766 * driver (for example non-PCI devices).
2769 dma_ops = &nommu_dma_ops;
2771 amd_iommu_stats_init();
2773 if (amd_iommu_unmap_flush)
2774 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2776 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2781 /*****************************************************************************
2783 * The following functions belong to the exported interface of AMD IOMMU
2785 * This interface allows access to lower level functions of the IOMMU
2786 * like protection domain handling and assignement of devices to domains
2787 * which is not possible with the dma_ops interface.
2789 *****************************************************************************/
2791 static void cleanup_domain(struct protection_domain *domain)
2793 struct iommu_dev_data *entry;
2794 unsigned long flags;
2796 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2798 while (!list_empty(&domain->dev_list)) {
2799 entry = list_first_entry(&domain->dev_list,
2800 struct iommu_dev_data, list);
2801 __detach_device(entry);
2804 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2807 static void protection_domain_free(struct protection_domain *domain)
2812 del_domain_from_list(domain);
2815 domain_id_free(domain->id);
2820 static int protection_domain_init(struct protection_domain *domain)
2822 spin_lock_init(&domain->lock);
2823 mutex_init(&domain->api_lock);
2824 domain->id = domain_id_alloc();
2827 INIT_LIST_HEAD(&domain->dev_list);
2832 static struct protection_domain *protection_domain_alloc(void)
2834 struct protection_domain *domain;
2836 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2840 if (protection_domain_init(domain))
2843 add_domain_to_list(domain);
2853 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2855 struct protection_domain *pdomain;
2856 struct dma_ops_domain *dma_domain;
2859 case IOMMU_DOMAIN_UNMANAGED:
2860 pdomain = protection_domain_alloc();
2864 pdomain->mode = PAGE_MODE_3_LEVEL;
2865 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2866 if (!pdomain->pt_root) {
2867 protection_domain_free(pdomain);
2871 pdomain->domain.geometry.aperture_start = 0;
2872 pdomain->domain.geometry.aperture_end = ~0ULL;
2873 pdomain->domain.geometry.force_aperture = true;
2876 case IOMMU_DOMAIN_DMA:
2877 dma_domain = dma_ops_domain_alloc();
2879 pr_err("AMD-Vi: Failed to allocate\n");
2882 pdomain = &dma_domain->domain;
2884 case IOMMU_DOMAIN_IDENTITY:
2885 pdomain = protection_domain_alloc();
2889 pdomain->mode = PAGE_MODE_NONE;
2895 return &pdomain->domain;
2898 static void amd_iommu_domain_free(struct iommu_domain *dom)
2900 struct protection_domain *domain;
2905 domain = to_pdomain(dom);
2907 if (domain->dev_cnt > 0)
2908 cleanup_domain(domain);
2910 BUG_ON(domain->dev_cnt != 0);
2912 if (domain->mode != PAGE_MODE_NONE)
2913 free_pagetable(domain);
2915 if (domain->flags & PD_IOMMUV2_MASK)
2916 free_gcr3_table(domain);
2918 protection_domain_free(domain);
2921 static void amd_iommu_detach_device(struct iommu_domain *dom,
2924 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2925 struct amd_iommu *iommu;
2928 if (!check_device(dev))
2931 devid = get_device_id(dev);
2933 if (dev_data->domain != NULL)
2936 iommu = amd_iommu_rlookup_table[devid];
2940 iommu_completion_wait(iommu);
2943 static int amd_iommu_attach_device(struct iommu_domain *dom,
2946 struct protection_domain *domain = to_pdomain(dom);
2947 struct iommu_dev_data *dev_data;
2948 struct amd_iommu *iommu;
2951 if (!check_device(dev))
2954 dev_data = dev->archdata.iommu;
2956 iommu = amd_iommu_rlookup_table[dev_data->devid];
2960 if (dev_data->domain)
2963 ret = attach_device(dev, domain);
2965 iommu_completion_wait(iommu);
2970 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2971 phys_addr_t paddr, size_t page_size, int iommu_prot)
2973 struct protection_domain *domain = to_pdomain(dom);
2977 if (domain->mode == PAGE_MODE_NONE)
2980 if (iommu_prot & IOMMU_READ)
2981 prot |= IOMMU_PROT_IR;
2982 if (iommu_prot & IOMMU_WRITE)
2983 prot |= IOMMU_PROT_IW;
2985 mutex_lock(&domain->api_lock);
2986 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
2987 mutex_unlock(&domain->api_lock);
2992 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2995 struct protection_domain *domain = to_pdomain(dom);
2998 if (domain->mode == PAGE_MODE_NONE)
3001 mutex_lock(&domain->api_lock);
3002 unmap_size = iommu_unmap_page(domain, iova, page_size);
3003 mutex_unlock(&domain->api_lock);
3005 domain_flush_tlb_pde(domain);
3010 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3013 struct protection_domain *domain = to_pdomain(dom);
3014 unsigned long offset_mask, pte_pgsize;
3017 if (domain->mode == PAGE_MODE_NONE)
3020 pte = fetch_pte(domain, iova, &pte_pgsize);
3022 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3025 offset_mask = pte_pgsize - 1;
3026 __pte = *pte & PM_ADDR_MASK;
3028 return (__pte & ~offset_mask) | (iova & offset_mask);
3031 static bool amd_iommu_capable(enum iommu_cap cap)
3034 case IOMMU_CAP_CACHE_COHERENCY:
3036 case IOMMU_CAP_INTR_REMAP:
3037 return (irq_remapping_enabled == 1);
3038 case IOMMU_CAP_NOEXEC:
3045 static void amd_iommu_get_dm_regions(struct device *dev,
3046 struct list_head *head)
3048 struct unity_map_entry *entry;
3051 devid = get_device_id(dev);
3053 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3054 struct iommu_dm_region *region;
3056 if (devid < entry->devid_start || devid > entry->devid_end)
3059 region = kzalloc(sizeof(*region), GFP_KERNEL);
3061 pr_err("Out of memory allocating dm-regions for %s\n",
3066 region->start = entry->address_start;
3067 region->length = entry->address_end - entry->address_start;
3068 if (entry->prot & IOMMU_PROT_IR)
3069 region->prot |= IOMMU_READ;
3070 if (entry->prot & IOMMU_PROT_IW)
3071 region->prot |= IOMMU_WRITE;
3073 list_add_tail(®ion->list, head);
3077 static void amd_iommu_put_dm_regions(struct device *dev,
3078 struct list_head *head)
3080 struct iommu_dm_region *entry, *next;
3082 list_for_each_entry_safe(entry, next, head, list)
3086 static const struct iommu_ops amd_iommu_ops = {
3087 .capable = amd_iommu_capable,
3088 .domain_alloc = amd_iommu_domain_alloc,
3089 .domain_free = amd_iommu_domain_free,
3090 .attach_dev = amd_iommu_attach_device,
3091 .detach_dev = amd_iommu_detach_device,
3092 .map = amd_iommu_map,
3093 .unmap = amd_iommu_unmap,
3094 .map_sg = default_iommu_map_sg,
3095 .iova_to_phys = amd_iommu_iova_to_phys,
3096 .add_device = amd_iommu_add_device,
3097 .remove_device = amd_iommu_remove_device,
3098 .get_dm_regions = amd_iommu_get_dm_regions,
3099 .put_dm_regions = amd_iommu_put_dm_regions,
3100 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3103 /*****************************************************************************
3105 * The next functions do a basic initialization of IOMMU for pass through
3108 * In passthrough mode the IOMMU is initialized and enabled but not used for
3109 * DMA-API translation.
3111 *****************************************************************************/
3113 /* IOMMUv2 specific functions */
3114 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3116 return atomic_notifier_chain_register(&ppr_notifier, nb);
3118 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3120 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3122 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3124 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3126 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3128 struct protection_domain *domain = to_pdomain(dom);
3129 unsigned long flags;
3131 spin_lock_irqsave(&domain->lock, flags);
3133 /* Update data structure */
3134 domain->mode = PAGE_MODE_NONE;
3135 domain->updated = true;
3137 /* Make changes visible to IOMMUs */
3138 update_domain(domain);
3140 /* Page-table is not visible to IOMMU anymore, so free it */
3141 free_pagetable(domain);
3143 spin_unlock_irqrestore(&domain->lock, flags);
3145 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3147 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3149 struct protection_domain *domain = to_pdomain(dom);
3150 unsigned long flags;
3153 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3156 /* Number of GCR3 table levels required */
3157 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3160 if (levels > amd_iommu_max_glx_val)
3163 spin_lock_irqsave(&domain->lock, flags);
3166 * Save us all sanity checks whether devices already in the
3167 * domain support IOMMUv2. Just force that the domain has no
3168 * devices attached when it is switched into IOMMUv2 mode.
3171 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3175 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3176 if (domain->gcr3_tbl == NULL)
3179 domain->glx = levels;
3180 domain->flags |= PD_IOMMUV2_MASK;
3181 domain->updated = true;
3183 update_domain(domain);
3188 spin_unlock_irqrestore(&domain->lock, flags);
3192 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3194 static int __flush_pasid(struct protection_domain *domain, int pasid,
3195 u64 address, bool size)
3197 struct iommu_dev_data *dev_data;
3198 struct iommu_cmd cmd;
3201 if (!(domain->flags & PD_IOMMUV2_MASK))
3204 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3207 * IOMMU TLB needs to be flushed before Device TLB to
3208 * prevent device TLB refill from IOMMU TLB
3210 for (i = 0; i < amd_iommus_present; ++i) {
3211 if (domain->dev_iommu[i] == 0)
3214 ret = iommu_queue_command(amd_iommus[i], &cmd);
3219 /* Wait until IOMMU TLB flushes are complete */
3220 domain_flush_complete(domain);
3222 /* Now flush device TLBs */
3223 list_for_each_entry(dev_data, &domain->dev_list, list) {
3224 struct amd_iommu *iommu;
3228 There might be non-IOMMUv2 capable devices in an IOMMUv2
3231 if (!dev_data->ats.enabled)
3234 qdep = dev_data->ats.qdep;
3235 iommu = amd_iommu_rlookup_table[dev_data->devid];
3237 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3238 qdep, address, size);
3240 ret = iommu_queue_command(iommu, &cmd);
3245 /* Wait until all device TLBs are flushed */
3246 domain_flush_complete(domain);
3255 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3258 INC_STATS_COUNTER(invalidate_iotlb);
3260 return __flush_pasid(domain, pasid, address, false);
3263 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3266 struct protection_domain *domain = to_pdomain(dom);
3267 unsigned long flags;
3270 spin_lock_irqsave(&domain->lock, flags);
3271 ret = __amd_iommu_flush_page(domain, pasid, address);
3272 spin_unlock_irqrestore(&domain->lock, flags);
3276 EXPORT_SYMBOL(amd_iommu_flush_page);
3278 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3280 INC_STATS_COUNTER(invalidate_iotlb_all);
3282 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3286 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3288 struct protection_domain *domain = to_pdomain(dom);
3289 unsigned long flags;
3292 spin_lock_irqsave(&domain->lock, flags);
3293 ret = __amd_iommu_flush_tlb(domain, pasid);
3294 spin_unlock_irqrestore(&domain->lock, flags);
3298 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3300 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3307 index = (pasid >> (9 * level)) & 0x1ff;
3313 if (!(*pte & GCR3_VALID)) {
3317 root = (void *)get_zeroed_page(GFP_ATOMIC);
3321 *pte = __pa(root) | GCR3_VALID;
3324 root = __va(*pte & PAGE_MASK);
3332 static int __set_gcr3(struct protection_domain *domain, int pasid,
3337 if (domain->mode != PAGE_MODE_NONE)
3340 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3344 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3346 return __amd_iommu_flush_tlb(domain, pasid);
3349 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3353 if (domain->mode != PAGE_MODE_NONE)
3356 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3362 return __amd_iommu_flush_tlb(domain, pasid);
3365 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3368 struct protection_domain *domain = to_pdomain(dom);
3369 unsigned long flags;
3372 spin_lock_irqsave(&domain->lock, flags);
3373 ret = __set_gcr3(domain, pasid, cr3);
3374 spin_unlock_irqrestore(&domain->lock, flags);
3378 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3380 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3382 struct protection_domain *domain = to_pdomain(dom);
3383 unsigned long flags;
3386 spin_lock_irqsave(&domain->lock, flags);
3387 ret = __clear_gcr3(domain, pasid);
3388 spin_unlock_irqrestore(&domain->lock, flags);
3392 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3394 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3395 int status, int tag)
3397 struct iommu_dev_data *dev_data;
3398 struct amd_iommu *iommu;
3399 struct iommu_cmd cmd;
3401 INC_STATS_COUNTER(complete_ppr);
3403 dev_data = get_dev_data(&pdev->dev);
3404 iommu = amd_iommu_rlookup_table[dev_data->devid];
3406 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3407 tag, dev_data->pri_tlp);
3409 return iommu_queue_command(iommu, &cmd);
3411 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3413 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3415 struct protection_domain *pdomain;
3417 pdomain = get_domain(&pdev->dev);
3418 if (IS_ERR(pdomain))
3421 /* Only return IOMMUv2 domains */
3422 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3425 return &pdomain->domain;
3427 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3429 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3431 struct iommu_dev_data *dev_data;
3433 if (!amd_iommu_v2_supported())
3436 dev_data = get_dev_data(&pdev->dev);
3437 dev_data->errata |= (1 << erratum);
3439 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3441 int amd_iommu_device_info(struct pci_dev *pdev,
3442 struct amd_iommu_device_info *info)
3447 if (pdev == NULL || info == NULL)
3450 if (!amd_iommu_v2_supported())
3453 memset(info, 0, sizeof(*info));
3455 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3457 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3459 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3461 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3463 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3467 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3468 max_pasids = min(max_pasids, (1 << 20));
3470 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3471 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3473 features = pci_pasid_features(pdev);
3474 if (features & PCI_PASID_CAP_EXEC)
3475 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3476 if (features & PCI_PASID_CAP_PRIV)
3477 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3482 EXPORT_SYMBOL(amd_iommu_device_info);
3484 #ifdef CONFIG_IRQ_REMAP
3486 /*****************************************************************************
3488 * Interrupt Remapping Implementation
3490 *****************************************************************************/
3508 u16 devid; /* Device ID for IRTE table */
3509 u16 index; /* Index into IRTE table*/
3512 struct amd_ir_data {
3513 struct irq_2_irte irq_2_irte;
3514 union irte irte_entry;
3516 struct msi_msg msi_entry;
3520 static struct irq_chip amd_ir_chip;
3522 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3523 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3524 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3525 #define DTE_IRQ_REMAP_ENABLE 1ULL
3527 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3531 dte = amd_iommu_dev_table[devid].data[2];
3532 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3533 dte |= virt_to_phys(table->table);
3534 dte |= DTE_IRQ_REMAP_INTCTL;
3535 dte |= DTE_IRQ_TABLE_LEN;
3536 dte |= DTE_IRQ_REMAP_ENABLE;
3538 amd_iommu_dev_table[devid].data[2] = dte;
3541 #define IRTE_ALLOCATED (~1U)
3543 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3545 struct irq_remap_table *table = NULL;
3546 struct amd_iommu *iommu;
3547 unsigned long flags;
3550 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3552 iommu = amd_iommu_rlookup_table[devid];
3556 table = irq_lookup_table[devid];
3560 alias = amd_iommu_alias_table[devid];
3561 table = irq_lookup_table[alias];
3563 irq_lookup_table[devid] = table;
3564 set_dte_irq_entry(devid, table);
3565 iommu_flush_dte(iommu, devid);
3569 /* Nothing there yet, allocate new irq remapping table */
3570 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3574 /* Initialize table spin-lock */
3575 spin_lock_init(&table->lock);
3578 /* Keep the first 32 indexes free for IOAPIC interrupts */
3579 table->min_index = 32;
3581 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3582 if (!table->table) {
3588 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3593 for (i = 0; i < 32; ++i)
3594 table->table[i] = IRTE_ALLOCATED;
3597 irq_lookup_table[devid] = table;
3598 set_dte_irq_entry(devid, table);
3599 iommu_flush_dte(iommu, devid);
3600 if (devid != alias) {
3601 irq_lookup_table[alias] = table;
3602 set_dte_irq_entry(alias, table);
3603 iommu_flush_dte(iommu, alias);
3607 iommu_completion_wait(iommu);
3610 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3615 static int alloc_irq_index(u16 devid, int count)
3617 struct irq_remap_table *table;
3618 unsigned long flags;
3621 table = get_irq_table(devid, false);
3625 spin_lock_irqsave(&table->lock, flags);
3627 /* Scan table for free entries */
3628 for (c = 0, index = table->min_index;
3629 index < MAX_IRQS_PER_TABLE;
3631 if (table->table[index] == 0)
3638 table->table[index - c + 1] = IRTE_ALLOCATED;
3648 spin_unlock_irqrestore(&table->lock, flags);
3653 static int modify_irte(u16 devid, int index, union irte irte)
3655 struct irq_remap_table *table;
3656 struct amd_iommu *iommu;
3657 unsigned long flags;
3659 iommu = amd_iommu_rlookup_table[devid];
3663 table = get_irq_table(devid, false);
3667 spin_lock_irqsave(&table->lock, flags);
3668 table->table[index] = irte.val;
3669 spin_unlock_irqrestore(&table->lock, flags);
3671 iommu_flush_irt(iommu, devid);
3672 iommu_completion_wait(iommu);
3677 static void free_irte(u16 devid, int index)
3679 struct irq_remap_table *table;
3680 struct amd_iommu *iommu;
3681 unsigned long flags;
3683 iommu = amd_iommu_rlookup_table[devid];
3687 table = get_irq_table(devid, false);
3691 spin_lock_irqsave(&table->lock, flags);
3692 table->table[index] = 0;
3693 spin_unlock_irqrestore(&table->lock, flags);
3695 iommu_flush_irt(iommu, devid);
3696 iommu_completion_wait(iommu);
3699 static int get_devid(struct irq_alloc_info *info)
3703 switch (info->type) {
3704 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3705 devid = get_ioapic_devid(info->ioapic_id);
3707 case X86_IRQ_ALLOC_TYPE_HPET:
3708 devid = get_hpet_devid(info->hpet_id);
3710 case X86_IRQ_ALLOC_TYPE_MSI:
3711 case X86_IRQ_ALLOC_TYPE_MSIX:
3712 devid = get_device_id(&info->msi_dev->dev);
3722 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3724 struct amd_iommu *iommu;
3730 devid = get_devid(info);
3732 iommu = amd_iommu_rlookup_table[devid];
3734 return iommu->ir_domain;
3740 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3742 struct amd_iommu *iommu;
3748 switch (info->type) {
3749 case X86_IRQ_ALLOC_TYPE_MSI:
3750 case X86_IRQ_ALLOC_TYPE_MSIX:
3751 devid = get_device_id(&info->msi_dev->dev);
3753 iommu = amd_iommu_rlookup_table[devid];
3755 return iommu->msi_domain;
3765 struct irq_remap_ops amd_iommu_irq_ops = {
3766 .prepare = amd_iommu_prepare,
3767 .enable = amd_iommu_enable,
3768 .disable = amd_iommu_disable,
3769 .reenable = amd_iommu_reenable,
3770 .enable_faulting = amd_iommu_enable_faulting,
3771 .get_ir_irq_domain = get_ir_irq_domain,
3772 .get_irq_domain = get_irq_domain,
3775 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3776 struct irq_cfg *irq_cfg,
3777 struct irq_alloc_info *info,
3778 int devid, int index, int sub_handle)
3780 struct irq_2_irte *irte_info = &data->irq_2_irte;
3781 struct msi_msg *msg = &data->msi_entry;
3782 union irte *irte = &data->irte_entry;
3783 struct IO_APIC_route_entry *entry;
3785 data->irq_2_irte.devid = devid;
3786 data->irq_2_irte.index = index + sub_handle;
3788 /* Setup IRTE for IOMMU */
3790 irte->fields.vector = irq_cfg->vector;
3791 irte->fields.int_type = apic->irq_delivery_mode;
3792 irte->fields.destination = irq_cfg->dest_apicid;
3793 irte->fields.dm = apic->irq_dest_mode;
3794 irte->fields.valid = 1;
3796 switch (info->type) {
3797 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3798 /* Setup IOAPIC entry */
3799 entry = info->ioapic_entry;
3800 info->ioapic_entry = NULL;
3801 memset(entry, 0, sizeof(*entry));
3802 entry->vector = index;
3804 entry->trigger = info->ioapic_trigger;
3805 entry->polarity = info->ioapic_polarity;
3806 /* Mask level triggered irqs. */
3807 if (info->ioapic_trigger)
3811 case X86_IRQ_ALLOC_TYPE_HPET:
3812 case X86_IRQ_ALLOC_TYPE_MSI:
3813 case X86_IRQ_ALLOC_TYPE_MSIX:
3814 msg->address_hi = MSI_ADDR_BASE_HI;
3815 msg->address_lo = MSI_ADDR_BASE_LO;
3816 msg->data = irte_info->index;
3825 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3826 unsigned int nr_irqs, void *arg)
3828 struct irq_alloc_info *info = arg;
3829 struct irq_data *irq_data;
3830 struct amd_ir_data *data;
3831 struct irq_cfg *cfg;
3837 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3838 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3842 * With IRQ remapping enabled, don't need contiguous CPU vectors
3843 * to support multiple MSI interrupts.
3845 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3846 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3848 devid = get_devid(info);
3852 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3856 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3857 if (get_irq_table(devid, true))
3858 index = info->ioapic_pin;
3862 index = alloc_irq_index(devid, nr_irqs);
3865 pr_warn("Failed to allocate IRTE\n");
3866 goto out_free_parent;
3869 for (i = 0; i < nr_irqs; i++) {
3870 irq_data = irq_domain_get_irq_data(domain, virq + i);
3871 cfg = irqd_cfg(irq_data);
3872 if (!irq_data || !cfg) {
3878 data = kzalloc(sizeof(*data), GFP_KERNEL);
3882 irq_data->hwirq = (devid << 16) + i;
3883 irq_data->chip_data = data;
3884 irq_data->chip = &amd_ir_chip;
3885 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3886 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3892 for (i--; i >= 0; i--) {
3893 irq_data = irq_domain_get_irq_data(domain, virq + i);
3895 kfree(irq_data->chip_data);
3897 for (i = 0; i < nr_irqs; i++)
3898 free_irte(devid, index + i);
3900 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3904 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3905 unsigned int nr_irqs)
3907 struct irq_2_irte *irte_info;
3908 struct irq_data *irq_data;
3909 struct amd_ir_data *data;
3912 for (i = 0; i < nr_irqs; i++) {
3913 irq_data = irq_domain_get_irq_data(domain, virq + i);
3914 if (irq_data && irq_data->chip_data) {
3915 data = irq_data->chip_data;
3916 irte_info = &data->irq_2_irte;
3917 free_irte(irte_info->devid, irte_info->index);
3921 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3924 static void irq_remapping_activate(struct irq_domain *domain,
3925 struct irq_data *irq_data)
3927 struct amd_ir_data *data = irq_data->chip_data;
3928 struct irq_2_irte *irte_info = &data->irq_2_irte;
3930 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3933 static void irq_remapping_deactivate(struct irq_domain *domain,
3934 struct irq_data *irq_data)
3936 struct amd_ir_data *data = irq_data->chip_data;
3937 struct irq_2_irte *irte_info = &data->irq_2_irte;
3941 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3944 static struct irq_domain_ops amd_ir_domain_ops = {
3945 .alloc = irq_remapping_alloc,
3946 .free = irq_remapping_free,
3947 .activate = irq_remapping_activate,
3948 .deactivate = irq_remapping_deactivate,
3951 static int amd_ir_set_affinity(struct irq_data *data,
3952 const struct cpumask *mask, bool force)
3954 struct amd_ir_data *ir_data = data->chip_data;
3955 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3956 struct irq_cfg *cfg = irqd_cfg(data);
3957 struct irq_data *parent = data->parent_data;
3960 ret = parent->chip->irq_set_affinity(parent, mask, force);
3961 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3965 * Atomically updates the IRTE with the new destination, vector
3966 * and flushes the interrupt entry cache.
3968 ir_data->irte_entry.fields.vector = cfg->vector;
3969 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
3970 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
3973 * After this point, all the interrupts will start arriving
3974 * at the new destination. So, time to cleanup the previous
3975 * vector allocation.
3977 send_cleanup_vector(cfg);
3979 return IRQ_SET_MASK_OK_DONE;
3982 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3984 struct amd_ir_data *ir_data = irq_data->chip_data;
3986 *msg = ir_data->msi_entry;
3989 static struct irq_chip amd_ir_chip = {
3990 .irq_ack = ir_ack_apic_edge,
3991 .irq_set_affinity = amd_ir_set_affinity,
3992 .irq_compose_msi_msg = ir_compose_msi_msg,
3995 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3997 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
3998 if (!iommu->ir_domain)
4001 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4002 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);