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) % EVT_BUFFER_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)) % CMD_BUFFER_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;
855 spin_lock_irqsave(&iommu->lock, flags);
857 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
858 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
859 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
860 left = (head - next_tail) % CMD_BUFFER_SIZE;
863 struct iommu_cmd sync_cmd;
864 volatile u64 sem = 0;
867 build_completion_wait(&sync_cmd, (u64)&sem);
868 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
870 spin_unlock_irqrestore(&iommu->lock, flags);
872 if ((ret = wait_on_sem(&sem)) != 0)
878 copy_cmd_to_buffer(iommu, cmd, tail);
880 /* We need to sync now to make sure all commands are processed */
881 iommu->need_sync = sync;
883 spin_unlock_irqrestore(&iommu->lock, flags);
888 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
890 return iommu_queue_command_sync(iommu, cmd, true);
894 * This function queues a completion wait command into the command
897 static int iommu_completion_wait(struct amd_iommu *iommu)
899 struct iommu_cmd cmd;
900 volatile u64 sem = 0;
903 if (!iommu->need_sync)
906 build_completion_wait(&cmd, (u64)&sem);
908 ret = iommu_queue_command_sync(iommu, &cmd, false);
912 return wait_on_sem(&sem);
915 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
917 struct iommu_cmd cmd;
919 build_inv_dte(&cmd, devid);
921 return iommu_queue_command(iommu, &cmd);
924 static void iommu_flush_dte_all(struct amd_iommu *iommu)
928 for (devid = 0; devid <= 0xffff; ++devid)
929 iommu_flush_dte(iommu, devid);
931 iommu_completion_wait(iommu);
935 * This function uses heavy locking and may disable irqs for some time. But
936 * this is no issue because it is only called during resume.
938 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
942 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
943 struct iommu_cmd cmd;
944 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
946 iommu_queue_command(iommu, &cmd);
949 iommu_completion_wait(iommu);
952 static void iommu_flush_all(struct amd_iommu *iommu)
954 struct iommu_cmd cmd;
958 iommu_queue_command(iommu, &cmd);
959 iommu_completion_wait(iommu);
962 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
964 struct iommu_cmd cmd;
966 build_inv_irt(&cmd, devid);
968 iommu_queue_command(iommu, &cmd);
971 static void iommu_flush_irt_all(struct amd_iommu *iommu)
975 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
976 iommu_flush_irt(iommu, devid);
978 iommu_completion_wait(iommu);
981 void iommu_flush_all_caches(struct amd_iommu *iommu)
983 if (iommu_feature(iommu, FEATURE_IA)) {
984 iommu_flush_all(iommu);
986 iommu_flush_dte_all(iommu);
987 iommu_flush_irt_all(iommu);
988 iommu_flush_tlb_all(iommu);
993 * Command send function for flushing on-device TLB
995 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
996 u64 address, size_t size)
998 struct amd_iommu *iommu;
999 struct iommu_cmd cmd;
1002 qdep = dev_data->ats.qdep;
1003 iommu = amd_iommu_rlookup_table[dev_data->devid];
1005 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1007 return iommu_queue_command(iommu, &cmd);
1011 * Command send function for invalidating a device table entry
1013 static int device_flush_dte(struct iommu_dev_data *dev_data)
1015 struct amd_iommu *iommu;
1019 iommu = amd_iommu_rlookup_table[dev_data->devid];
1020 alias = amd_iommu_alias_table[dev_data->devid];
1022 ret = iommu_flush_dte(iommu, dev_data->devid);
1023 if (!ret && alias != dev_data->devid)
1024 ret = iommu_flush_dte(iommu, alias);
1028 if (dev_data->ats.enabled)
1029 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1035 * TLB invalidation function which is called from the mapping functions.
1036 * It invalidates a single PTE if the range to flush is within a single
1037 * page. Otherwise it flushes the whole TLB of the IOMMU.
1039 static void __domain_flush_pages(struct protection_domain *domain,
1040 u64 address, size_t size, int pde)
1042 struct iommu_dev_data *dev_data;
1043 struct iommu_cmd cmd;
1046 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1048 for (i = 0; i < amd_iommus_present; ++i) {
1049 if (!domain->dev_iommu[i])
1053 * Devices of this domain are behind this IOMMU
1054 * We need a TLB flush
1056 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1059 list_for_each_entry(dev_data, &domain->dev_list, list) {
1061 if (!dev_data->ats.enabled)
1064 ret |= device_flush_iotlb(dev_data, address, size);
1070 static void domain_flush_pages(struct protection_domain *domain,
1071 u64 address, size_t size)
1073 __domain_flush_pages(domain, address, size, 0);
1076 /* Flush the whole IO/TLB for a given protection domain */
1077 static void domain_flush_tlb(struct protection_domain *domain)
1079 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1082 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1083 static void domain_flush_tlb_pde(struct protection_domain *domain)
1085 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1088 static void domain_flush_complete(struct protection_domain *domain)
1092 for (i = 0; i < amd_iommus_present; ++i) {
1093 if (!domain->dev_iommu[i])
1097 * Devices of this domain are behind this IOMMU
1098 * We need to wait for completion of all commands.
1100 iommu_completion_wait(amd_iommus[i]);
1106 * This function flushes the DTEs for all devices in domain
1108 static void domain_flush_devices(struct protection_domain *domain)
1110 struct iommu_dev_data *dev_data;
1112 list_for_each_entry(dev_data, &domain->dev_list, list)
1113 device_flush_dte(dev_data);
1116 /****************************************************************************
1118 * The functions below are used the create the page table mappings for
1119 * unity mapped regions.
1121 ****************************************************************************/
1124 * This function is used to add another level to an IO page table. Adding
1125 * another level increases the size of the address space by 9 bits to a size up
1128 static bool increase_address_space(struct protection_domain *domain,
1133 if (domain->mode == PAGE_MODE_6_LEVEL)
1134 /* address space already 64 bit large */
1137 pte = (void *)get_zeroed_page(gfp);
1141 *pte = PM_LEVEL_PDE(domain->mode,
1142 virt_to_phys(domain->pt_root));
1143 domain->pt_root = pte;
1145 domain->updated = true;
1150 static u64 *alloc_pte(struct protection_domain *domain,
1151 unsigned long address,
1152 unsigned long page_size,
1159 BUG_ON(!is_power_of_2(page_size));
1161 while (address > PM_LEVEL_SIZE(domain->mode))
1162 increase_address_space(domain, gfp);
1164 level = domain->mode - 1;
1165 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1166 address = PAGE_SIZE_ALIGN(address, page_size);
1167 end_lvl = PAGE_SIZE_LEVEL(page_size);
1169 while (level > end_lvl) {
1170 if (!IOMMU_PTE_PRESENT(*pte)) {
1171 page = (u64 *)get_zeroed_page(gfp);
1174 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1177 /* No level skipping support yet */
1178 if (PM_PTE_LEVEL(*pte) != level)
1183 pte = IOMMU_PTE_PAGE(*pte);
1185 if (pte_page && level == end_lvl)
1188 pte = &pte[PM_LEVEL_INDEX(level, address)];
1195 * This function checks if there is a PTE for a given dma address. If
1196 * there is one, it returns the pointer to it.
1198 static u64 *fetch_pte(struct protection_domain *domain,
1199 unsigned long address,
1200 unsigned long *page_size)
1205 if (address > PM_LEVEL_SIZE(domain->mode))
1208 level = domain->mode - 1;
1209 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1210 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1215 if (!IOMMU_PTE_PRESENT(*pte))
1219 if (PM_PTE_LEVEL(*pte) == 7 ||
1220 PM_PTE_LEVEL(*pte) == 0)
1223 /* No level skipping support yet */
1224 if (PM_PTE_LEVEL(*pte) != level)
1229 /* Walk to the next level */
1230 pte = IOMMU_PTE_PAGE(*pte);
1231 pte = &pte[PM_LEVEL_INDEX(level, address)];
1232 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1235 if (PM_PTE_LEVEL(*pte) == 0x07) {
1236 unsigned long pte_mask;
1239 * If we have a series of large PTEs, make
1240 * sure to return a pointer to the first one.
1242 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1243 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1244 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1251 * Generic mapping functions. It maps a physical address into a DMA
1252 * address space. It allocates the page table pages if necessary.
1253 * In the future it can be extended to a generic mapping function
1254 * supporting all features of AMD IOMMU page tables like level skipping
1255 * and full 64 bit address spaces.
1257 static int iommu_map_page(struct protection_domain *dom,
1258 unsigned long bus_addr,
1259 unsigned long phys_addr,
1261 unsigned long page_size)
1266 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1267 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1269 if (!(prot & IOMMU_PROT_MASK))
1272 count = PAGE_SIZE_PTE_COUNT(page_size);
1273 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1278 for (i = 0; i < count; ++i)
1279 if (IOMMU_PTE_PRESENT(pte[i]))
1283 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1284 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1286 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1288 if (prot & IOMMU_PROT_IR)
1289 __pte |= IOMMU_PTE_IR;
1290 if (prot & IOMMU_PROT_IW)
1291 __pte |= IOMMU_PTE_IW;
1293 for (i = 0; i < count; ++i)
1301 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1302 unsigned long bus_addr,
1303 unsigned long page_size)
1305 unsigned long long unmapped;
1306 unsigned long unmap_size;
1309 BUG_ON(!is_power_of_2(page_size));
1313 while (unmapped < page_size) {
1315 pte = fetch_pte(dom, bus_addr, &unmap_size);
1320 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1321 for (i = 0; i < count; i++)
1325 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1326 unmapped += unmap_size;
1329 BUG_ON(unmapped && !is_power_of_2(unmapped));
1334 /****************************************************************************
1336 * The next functions belong to the address allocator for the dma_ops
1337 * interface functions. They work like the allocators in the other IOMMU
1338 * drivers. Its basically a bitmap which marks the allocated pages in
1339 * the aperture. Maybe it could be enhanced in the future to a more
1340 * efficient allocator.
1342 ****************************************************************************/
1345 * The address allocator core functions.
1347 * called with domain->lock held
1351 * Used to reserve address ranges in the aperture (e.g. for exclusion
1354 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1355 unsigned long start_page,
1358 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1360 if (start_page + pages > last_page)
1361 pages = last_page - start_page;
1363 for (i = start_page; i < start_page + pages; ++i) {
1364 int index = i / APERTURE_RANGE_PAGES;
1365 int page = i % APERTURE_RANGE_PAGES;
1366 __set_bit(page, dom->aperture[index]->bitmap);
1371 * This function is used to add a new aperture range to an existing
1372 * aperture in case of dma_ops domain allocation or address allocation
1375 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1376 bool populate, gfp_t gfp)
1378 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1379 struct amd_iommu *iommu;
1380 unsigned long i, old_size, pte_pgsize;
1382 #ifdef CONFIG_IOMMU_STRESS
1386 if (index >= APERTURE_MAX_RANGES)
1389 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1390 if (!dma_dom->aperture[index])
1393 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1394 if (!dma_dom->aperture[index]->bitmap)
1397 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1400 unsigned long address = dma_dom->aperture_size;
1401 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1402 u64 *pte, *pte_page;
1404 for (i = 0; i < num_ptes; ++i) {
1405 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1410 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1412 address += APERTURE_RANGE_SIZE / 64;
1416 old_size = dma_dom->aperture_size;
1417 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1419 /* Reserve address range used for MSI messages */
1420 if (old_size < MSI_ADDR_BASE_LO &&
1421 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1422 unsigned long spage;
1425 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1426 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1428 dma_ops_reserve_addresses(dma_dom, spage, pages);
1431 /* Initialize the exclusion range if necessary */
1432 for_each_iommu(iommu) {
1433 if (iommu->exclusion_start &&
1434 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1435 && iommu->exclusion_start < dma_dom->aperture_size) {
1436 unsigned long startpage;
1437 int pages = iommu_num_pages(iommu->exclusion_start,
1438 iommu->exclusion_length,
1440 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1441 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1446 * Check for areas already mapped as present in the new aperture
1447 * range and mark those pages as reserved in the allocator. Such
1448 * mappings may already exist as a result of requested unity
1449 * mappings for devices.
1451 for (i = dma_dom->aperture[index]->offset;
1452 i < dma_dom->aperture_size;
1454 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1455 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1458 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1462 update_domain(&dma_dom->domain);
1467 update_domain(&dma_dom->domain);
1469 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1471 kfree(dma_dom->aperture[index]);
1472 dma_dom->aperture[index] = NULL;
1477 static unsigned long dma_ops_area_alloc(struct device *dev,
1478 struct dma_ops_domain *dom,
1480 unsigned long align_mask,
1482 unsigned long start)
1484 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1485 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1486 int i = start >> APERTURE_RANGE_SHIFT;
1487 unsigned long boundary_size, mask;
1488 unsigned long address = -1;
1489 unsigned long limit;
1491 next_bit >>= PAGE_SHIFT;
1493 mask = dma_get_seg_boundary(dev);
1495 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1496 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1498 for (;i < max_index; ++i) {
1499 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1501 if (dom->aperture[i]->offset >= dma_mask)
1504 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1505 dma_mask >> PAGE_SHIFT);
1507 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1508 limit, next_bit, pages, 0,
1509 boundary_size, align_mask);
1510 if (address != -1) {
1511 address = dom->aperture[i]->offset +
1512 (address << PAGE_SHIFT);
1513 dom->next_address = address + (pages << PAGE_SHIFT);
1523 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1524 struct dma_ops_domain *dom,
1526 unsigned long align_mask,
1529 unsigned long address;
1531 #ifdef CONFIG_IOMMU_STRESS
1532 dom->next_address = 0;
1533 dom->need_flush = true;
1536 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1537 dma_mask, dom->next_address);
1539 if (address == -1) {
1540 dom->next_address = 0;
1541 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1543 dom->need_flush = true;
1546 if (unlikely(address == -1))
1547 address = DMA_ERROR_CODE;
1549 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1555 * The address free function.
1557 * called with domain->lock held
1559 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1560 unsigned long address,
1563 unsigned i = address >> APERTURE_RANGE_SHIFT;
1564 struct aperture_range *range = dom->aperture[i];
1566 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1568 #ifdef CONFIG_IOMMU_STRESS
1573 if (address >= dom->next_address)
1574 dom->need_flush = true;
1576 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1578 bitmap_clear(range->bitmap, address, pages);
1582 /****************************************************************************
1584 * The next functions belong to the domain allocation. A domain is
1585 * allocated for every IOMMU as the default domain. If device isolation
1586 * is enabled, every device get its own domain. The most important thing
1587 * about domains is the page table mapping the DMA address space they
1590 ****************************************************************************/
1593 * This function adds a protection domain to the global protection domain list
1595 static void add_domain_to_list(struct protection_domain *domain)
1597 unsigned long flags;
1599 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1600 list_add(&domain->list, &amd_iommu_pd_list);
1601 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1605 * This function removes a protection domain to the global
1606 * protection domain list
1608 static void del_domain_from_list(struct protection_domain *domain)
1610 unsigned long flags;
1612 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1613 list_del(&domain->list);
1614 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1617 static u16 domain_id_alloc(void)
1619 unsigned long flags;
1622 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1623 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1625 if (id > 0 && id < MAX_DOMAIN_ID)
1626 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1629 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1634 static void domain_id_free(int id)
1636 unsigned long flags;
1638 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1639 if (id > 0 && id < MAX_DOMAIN_ID)
1640 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1641 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1644 #define DEFINE_FREE_PT_FN(LVL, FN) \
1645 static void free_pt_##LVL (unsigned long __pt) \
1653 for (i = 0; i < 512; ++i) { \
1654 /* PTE present? */ \
1655 if (!IOMMU_PTE_PRESENT(pt[i])) \
1659 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1660 PM_PTE_LEVEL(pt[i]) == 7) \
1663 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1666 free_page((unsigned long)pt); \
1669 DEFINE_FREE_PT_FN(l2, free_page)
1670 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1671 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1672 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1673 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1675 static void free_pagetable(struct protection_domain *domain)
1677 unsigned long root = (unsigned long)domain->pt_root;
1679 switch (domain->mode) {
1680 case PAGE_MODE_NONE:
1682 case PAGE_MODE_1_LEVEL:
1685 case PAGE_MODE_2_LEVEL:
1688 case PAGE_MODE_3_LEVEL:
1691 case PAGE_MODE_4_LEVEL:
1694 case PAGE_MODE_5_LEVEL:
1697 case PAGE_MODE_6_LEVEL:
1705 static void free_gcr3_tbl_level1(u64 *tbl)
1710 for (i = 0; i < 512; ++i) {
1711 if (!(tbl[i] & GCR3_VALID))
1714 ptr = __va(tbl[i] & PAGE_MASK);
1716 free_page((unsigned long)ptr);
1720 static void free_gcr3_tbl_level2(u64 *tbl)
1725 for (i = 0; i < 512; ++i) {
1726 if (!(tbl[i] & GCR3_VALID))
1729 ptr = __va(tbl[i] & PAGE_MASK);
1731 free_gcr3_tbl_level1(ptr);
1735 static void free_gcr3_table(struct protection_domain *domain)
1737 if (domain->glx == 2)
1738 free_gcr3_tbl_level2(domain->gcr3_tbl);
1739 else if (domain->glx == 1)
1740 free_gcr3_tbl_level1(domain->gcr3_tbl);
1742 BUG_ON(domain->glx != 0);
1744 free_page((unsigned long)domain->gcr3_tbl);
1748 * Free a domain, only used if something went wrong in the
1749 * allocation path and we need to free an already allocated page table
1751 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1758 del_domain_from_list(&dom->domain);
1760 free_pagetable(&dom->domain);
1762 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1763 if (!dom->aperture[i])
1765 free_page((unsigned long)dom->aperture[i]->bitmap);
1766 kfree(dom->aperture[i]);
1773 * Allocates a new protection domain usable for the dma_ops functions.
1774 * It also initializes the page table and the address allocator data
1775 * structures required for the dma_ops interface
1777 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1779 struct dma_ops_domain *dma_dom;
1781 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1785 if (protection_domain_init(&dma_dom->domain))
1788 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1789 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1790 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1791 dma_dom->domain.priv = dma_dom;
1792 if (!dma_dom->domain.pt_root)
1795 dma_dom->need_flush = false;
1797 add_domain_to_list(&dma_dom->domain);
1799 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1803 * mark the first page as allocated so we never return 0 as
1804 * a valid dma-address. So we can use 0 as error value
1806 dma_dom->aperture[0]->bitmap[0] = 1;
1807 dma_dom->next_address = 0;
1813 dma_ops_domain_free(dma_dom);
1819 * little helper function to check whether a given protection domain is a
1822 static bool dma_ops_domain(struct protection_domain *domain)
1824 return domain->flags & PD_DMA_OPS_MASK;
1827 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1832 if (domain->mode != PAGE_MODE_NONE)
1833 pte_root = virt_to_phys(domain->pt_root);
1835 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1836 << DEV_ENTRY_MODE_SHIFT;
1837 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1839 flags = amd_iommu_dev_table[devid].data[1];
1842 flags |= DTE_FLAG_IOTLB;
1844 if (domain->flags & PD_IOMMUV2_MASK) {
1845 u64 gcr3 = __pa(domain->gcr3_tbl);
1846 u64 glx = domain->glx;
1849 pte_root |= DTE_FLAG_GV;
1850 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1852 /* First mask out possible old values for GCR3 table */
1853 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1856 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1859 /* Encode GCR3 table into DTE */
1860 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1863 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1866 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1870 flags &= ~(0xffffUL);
1871 flags |= domain->id;
1873 amd_iommu_dev_table[devid].data[1] = flags;
1874 amd_iommu_dev_table[devid].data[0] = pte_root;
1877 static void clear_dte_entry(u16 devid)
1879 /* remove entry from the device table seen by the hardware */
1880 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1881 amd_iommu_dev_table[devid].data[1] = 0;
1883 amd_iommu_apply_erratum_63(devid);
1886 static void do_attach(struct iommu_dev_data *dev_data,
1887 struct protection_domain *domain)
1889 struct amd_iommu *iommu;
1893 iommu = amd_iommu_rlookup_table[dev_data->devid];
1894 alias = amd_iommu_alias_table[dev_data->devid];
1895 ats = dev_data->ats.enabled;
1897 /* Update data structures */
1898 dev_data->domain = domain;
1899 list_add(&dev_data->list, &domain->dev_list);
1901 /* Do reference counting */
1902 domain->dev_iommu[iommu->index] += 1;
1903 domain->dev_cnt += 1;
1905 /* Update device table */
1906 set_dte_entry(dev_data->devid, domain, ats);
1907 if (alias != dev_data->devid)
1908 set_dte_entry(dev_data->devid, domain, ats);
1910 device_flush_dte(dev_data);
1913 static void do_detach(struct iommu_dev_data *dev_data)
1915 struct amd_iommu *iommu;
1918 iommu = amd_iommu_rlookup_table[dev_data->devid];
1919 alias = amd_iommu_alias_table[dev_data->devid];
1921 /* decrease reference counters */
1922 dev_data->domain->dev_iommu[iommu->index] -= 1;
1923 dev_data->domain->dev_cnt -= 1;
1925 /* Update data structures */
1926 dev_data->domain = NULL;
1927 list_del(&dev_data->list);
1928 clear_dte_entry(dev_data->devid);
1929 if (alias != dev_data->devid)
1930 clear_dte_entry(alias);
1932 /* Flush the DTE entry */
1933 device_flush_dte(dev_data);
1937 * If a device is not yet associated with a domain, this function does
1938 * assigns it visible for the hardware
1940 static int __attach_device(struct iommu_dev_data *dev_data,
1941 struct protection_domain *domain)
1946 * Must be called with IRQs disabled. Warn here to detect early
1949 WARN_ON(!irqs_disabled());
1952 spin_lock(&domain->lock);
1955 if (dev_data->domain != NULL)
1958 /* Attach alias group root */
1959 do_attach(dev_data, domain);
1966 spin_unlock(&domain->lock);
1972 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1974 pci_disable_ats(pdev);
1975 pci_disable_pri(pdev);
1976 pci_disable_pasid(pdev);
1979 /* FIXME: Change generic reset-function to do the same */
1980 static int pri_reset_while_enabled(struct pci_dev *pdev)
1985 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1989 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1990 control |= PCI_PRI_CTRL_RESET;
1991 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1996 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2001 /* FIXME: Hardcode number of outstanding requests for now */
2003 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2005 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2007 /* Only allow access to user-accessible pages */
2008 ret = pci_enable_pasid(pdev, 0);
2012 /* First reset the PRI state of the device */
2013 ret = pci_reset_pri(pdev);
2018 ret = pci_enable_pri(pdev, reqs);
2023 ret = pri_reset_while_enabled(pdev);
2028 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2035 pci_disable_pri(pdev);
2036 pci_disable_pasid(pdev);
2041 /* FIXME: Move this to PCI code */
2042 #define PCI_PRI_TLP_OFF (1 << 15)
2044 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2049 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2053 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2055 return (status & PCI_PRI_TLP_OFF) ? true : false;
2059 * If a device is not yet associated with a domain, this function
2060 * assigns it visible for the hardware
2062 static int attach_device(struct device *dev,
2063 struct protection_domain *domain)
2065 struct pci_dev *pdev = to_pci_dev(dev);
2066 struct iommu_dev_data *dev_data;
2067 unsigned long flags;
2070 dev_data = get_dev_data(dev);
2072 if (domain->flags & PD_IOMMUV2_MASK) {
2073 if (!dev_data->passthrough)
2076 if (dev_data->iommu_v2) {
2077 if (pdev_iommuv2_enable(pdev) != 0)
2080 dev_data->ats.enabled = true;
2081 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2082 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2084 } else if (amd_iommu_iotlb_sup &&
2085 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2086 dev_data->ats.enabled = true;
2087 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2090 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2091 ret = __attach_device(dev_data, domain);
2092 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2095 * We might boot into a crash-kernel here. The crashed kernel
2096 * left the caches in the IOMMU dirty. So we have to flush
2097 * here to evict all dirty stuff.
2099 domain_flush_tlb_pde(domain);
2105 * Removes a device from a protection domain (unlocked)
2107 static void __detach_device(struct iommu_dev_data *dev_data)
2109 struct protection_domain *domain;
2112 * Must be called with IRQs disabled. Warn here to detect early
2115 WARN_ON(!irqs_disabled());
2117 if (WARN_ON(!dev_data->domain))
2120 domain = dev_data->domain;
2122 spin_lock(&domain->lock);
2124 do_detach(dev_data);
2126 spin_unlock(&domain->lock);
2130 * Removes a device from a protection domain (with devtable_lock held)
2132 static void detach_device(struct device *dev)
2134 struct protection_domain *domain;
2135 struct iommu_dev_data *dev_data;
2136 unsigned long flags;
2138 dev_data = get_dev_data(dev);
2139 domain = dev_data->domain;
2141 /* lock device table */
2142 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2143 __detach_device(dev_data);
2144 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2146 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2147 pdev_iommuv2_disable(to_pci_dev(dev));
2148 else if (dev_data->ats.enabled)
2149 pci_disable_ats(to_pci_dev(dev));
2151 dev_data->ats.enabled = false;
2154 static int amd_iommu_add_device(struct device *dev)
2156 struct iommu_dev_data *dev_data;
2157 struct iommu_domain *domain;
2158 struct amd_iommu *iommu;
2162 if (!check_device(dev) || get_dev_data(dev))
2165 devid = get_device_id(dev);
2166 iommu = amd_iommu_rlookup_table[devid];
2168 ret = iommu_init_device(dev);
2170 if (ret != -ENOTSUPP)
2171 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2174 iommu_ignore_device(dev);
2175 dev->archdata.dma_ops = &nommu_dma_ops;
2178 init_iommu_group(dev);
2180 dev_data = get_dev_data(dev);
2184 if (iommu_pass_through || dev_data->iommu_v2)
2185 iommu_request_dm_for_dev(dev);
2187 /* Domains are initialized for this device - have a look what we ended up with */
2188 domain = iommu_get_domain_for_dev(dev);
2189 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2190 dev_data->passthrough = true;
2192 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2195 iommu_completion_wait(iommu);
2200 static void amd_iommu_remove_device(struct device *dev)
2202 struct amd_iommu *iommu;
2205 if (!check_device(dev))
2208 devid = get_device_id(dev);
2209 iommu = amd_iommu_rlookup_table[devid];
2211 iommu_uninit_device(dev);
2212 iommu_completion_wait(iommu);
2215 /*****************************************************************************
2217 * The next functions belong to the dma_ops mapping/unmapping code.
2219 *****************************************************************************/
2222 * In the dma_ops path we only have the struct device. This function
2223 * finds the corresponding IOMMU, the protection domain and the
2224 * requestor id for a given device.
2225 * If the device is not yet associated with a domain this is also done
2228 static struct protection_domain *get_domain(struct device *dev)
2230 struct protection_domain *domain;
2231 struct iommu_domain *io_domain;
2233 if (!check_device(dev))
2234 return ERR_PTR(-EINVAL);
2236 io_domain = iommu_get_domain_for_dev(dev);
2240 domain = to_pdomain(io_domain);
2241 if (!dma_ops_domain(domain))
2242 return ERR_PTR(-EBUSY);
2247 static void update_device_table(struct protection_domain *domain)
2249 struct iommu_dev_data *dev_data;
2251 list_for_each_entry(dev_data, &domain->dev_list, list)
2252 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2255 static void update_domain(struct protection_domain *domain)
2257 if (!domain->updated)
2260 update_device_table(domain);
2262 domain_flush_devices(domain);
2263 domain_flush_tlb_pde(domain);
2265 domain->updated = false;
2269 * This function fetches the PTE for a given address in the aperture
2271 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2272 unsigned long address)
2274 struct aperture_range *aperture;
2275 u64 *pte, *pte_page;
2277 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2281 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2283 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2285 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2287 pte += PM_LEVEL_INDEX(0, address);
2289 update_domain(&dom->domain);
2295 * This is the generic map function. It maps one 4kb page at paddr to
2296 * the given address in the DMA address space for the domain.
2298 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2299 unsigned long address,
2305 WARN_ON(address > dom->aperture_size);
2309 pte = dma_ops_get_pte(dom, address);
2311 return DMA_ERROR_CODE;
2313 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2315 if (direction == DMA_TO_DEVICE)
2316 __pte |= IOMMU_PTE_IR;
2317 else if (direction == DMA_FROM_DEVICE)
2318 __pte |= IOMMU_PTE_IW;
2319 else if (direction == DMA_BIDIRECTIONAL)
2320 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2326 return (dma_addr_t)address;
2330 * The generic unmapping function for on page in the DMA address space.
2332 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2333 unsigned long address)
2335 struct aperture_range *aperture;
2338 if (address >= dom->aperture_size)
2341 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2345 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2349 pte += PM_LEVEL_INDEX(0, address);
2357 * This function contains common code for mapping of a physically
2358 * contiguous memory region into DMA address space. It is used by all
2359 * mapping functions provided with this IOMMU driver.
2360 * Must be called with the domain lock held.
2362 static dma_addr_t __map_single(struct device *dev,
2363 struct dma_ops_domain *dma_dom,
2370 dma_addr_t offset = paddr & ~PAGE_MASK;
2371 dma_addr_t address, start, ret;
2373 unsigned long align_mask = 0;
2376 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2379 INC_STATS_COUNTER(total_map_requests);
2382 INC_STATS_COUNTER(cross_page);
2385 align_mask = (1UL << get_order(size)) - 1;
2388 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2390 if (unlikely(address == DMA_ERROR_CODE)) {
2392 * setting next_address here will let the address
2393 * allocator only scan the new allocated range in the
2394 * first run. This is a small optimization.
2396 dma_dom->next_address = dma_dom->aperture_size;
2398 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2402 * aperture was successfully enlarged by 128 MB, try
2409 for (i = 0; i < pages; ++i) {
2410 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2411 if (ret == DMA_ERROR_CODE)
2419 ADD_STATS_COUNTER(alloced_io_mem, size);
2421 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2422 domain_flush_tlb(&dma_dom->domain);
2423 dma_dom->need_flush = false;
2424 } else if (unlikely(amd_iommu_np_cache))
2425 domain_flush_pages(&dma_dom->domain, address, size);
2432 for (--i; i >= 0; --i) {
2434 dma_ops_domain_unmap(dma_dom, start);
2437 dma_ops_free_addresses(dma_dom, address, pages);
2439 return DMA_ERROR_CODE;
2443 * Does the reverse of the __map_single function. Must be called with
2444 * the domain lock held too
2446 static void __unmap_single(struct dma_ops_domain *dma_dom,
2447 dma_addr_t dma_addr,
2451 dma_addr_t flush_addr;
2452 dma_addr_t i, start;
2455 if ((dma_addr == DMA_ERROR_CODE) ||
2456 (dma_addr + size > dma_dom->aperture_size))
2459 flush_addr = dma_addr;
2460 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2461 dma_addr &= PAGE_MASK;
2464 for (i = 0; i < pages; ++i) {
2465 dma_ops_domain_unmap(dma_dom, start);
2469 SUB_STATS_COUNTER(alloced_io_mem, size);
2471 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2473 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2474 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2475 dma_dom->need_flush = false;
2480 * The exported map_single function for dma_ops.
2482 static dma_addr_t map_page(struct device *dev, struct page *page,
2483 unsigned long offset, size_t size,
2484 enum dma_data_direction dir,
2485 struct dma_attrs *attrs)
2487 unsigned long flags;
2488 struct protection_domain *domain;
2491 phys_addr_t paddr = page_to_phys(page) + offset;
2493 INC_STATS_COUNTER(cnt_map_single);
2495 domain = get_domain(dev);
2496 if (PTR_ERR(domain) == -EINVAL)
2497 return (dma_addr_t)paddr;
2498 else if (IS_ERR(domain))
2499 return DMA_ERROR_CODE;
2501 dma_mask = *dev->dma_mask;
2503 spin_lock_irqsave(&domain->lock, flags);
2505 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2507 if (addr == DMA_ERROR_CODE)
2510 domain_flush_complete(domain);
2513 spin_unlock_irqrestore(&domain->lock, flags);
2519 * The exported unmap_single function for dma_ops.
2521 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2522 enum dma_data_direction dir, struct dma_attrs *attrs)
2524 unsigned long flags;
2525 struct protection_domain *domain;
2527 INC_STATS_COUNTER(cnt_unmap_single);
2529 domain = get_domain(dev);
2533 spin_lock_irqsave(&domain->lock, flags);
2535 __unmap_single(domain->priv, dma_addr, size, dir);
2537 domain_flush_complete(domain);
2539 spin_unlock_irqrestore(&domain->lock, flags);
2543 * The exported map_sg function for dma_ops (handles scatter-gather
2546 static int map_sg(struct device *dev, struct scatterlist *sglist,
2547 int nelems, enum dma_data_direction dir,
2548 struct dma_attrs *attrs)
2550 unsigned long flags;
2551 struct protection_domain *domain;
2553 struct scatterlist *s;
2555 int mapped_elems = 0;
2558 INC_STATS_COUNTER(cnt_map_sg);
2560 domain = get_domain(dev);
2564 dma_mask = *dev->dma_mask;
2566 spin_lock_irqsave(&domain->lock, flags);
2568 for_each_sg(sglist, s, nelems, i) {
2571 s->dma_address = __map_single(dev, domain->priv,
2572 paddr, s->length, dir, false,
2575 if (s->dma_address) {
2576 s->dma_length = s->length;
2582 domain_flush_complete(domain);
2585 spin_unlock_irqrestore(&domain->lock, flags);
2587 return mapped_elems;
2589 for_each_sg(sglist, s, mapped_elems, i) {
2591 __unmap_single(domain->priv, s->dma_address,
2592 s->dma_length, dir);
2593 s->dma_address = s->dma_length = 0;
2602 * The exported map_sg function for dma_ops (handles scatter-gather
2605 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2606 int nelems, enum dma_data_direction dir,
2607 struct dma_attrs *attrs)
2609 unsigned long flags;
2610 struct protection_domain *domain;
2611 struct scatterlist *s;
2614 INC_STATS_COUNTER(cnt_unmap_sg);
2616 domain = get_domain(dev);
2620 spin_lock_irqsave(&domain->lock, flags);
2622 for_each_sg(sglist, s, nelems, i) {
2623 __unmap_single(domain->priv, s->dma_address,
2624 s->dma_length, dir);
2625 s->dma_address = s->dma_length = 0;
2628 domain_flush_complete(domain);
2630 spin_unlock_irqrestore(&domain->lock, flags);
2634 * The exported alloc_coherent function for dma_ops.
2636 static void *alloc_coherent(struct device *dev, size_t size,
2637 dma_addr_t *dma_addr, gfp_t flag,
2638 struct dma_attrs *attrs)
2640 u64 dma_mask = dev->coherent_dma_mask;
2641 struct protection_domain *domain;
2642 unsigned long flags;
2645 INC_STATS_COUNTER(cnt_alloc_coherent);
2647 domain = get_domain(dev);
2648 if (PTR_ERR(domain) == -EINVAL) {
2649 page = alloc_pages(flag, get_order(size));
2650 *dma_addr = page_to_phys(page);
2651 return page_address(page);
2652 } else if (IS_ERR(domain))
2655 size = PAGE_ALIGN(size);
2656 dma_mask = dev->coherent_dma_mask;
2657 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2660 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2662 if (!(flag & __GFP_WAIT))
2665 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2672 dma_mask = *dev->dma_mask;
2674 spin_lock_irqsave(&domain->lock, flags);
2676 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2677 size, DMA_BIDIRECTIONAL, true, dma_mask);
2679 if (*dma_addr == DMA_ERROR_CODE) {
2680 spin_unlock_irqrestore(&domain->lock, flags);
2684 domain_flush_complete(domain);
2686 spin_unlock_irqrestore(&domain->lock, flags);
2688 return page_address(page);
2692 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2693 __free_pages(page, get_order(size));
2699 * The exported free_coherent function for dma_ops.
2701 static void free_coherent(struct device *dev, size_t size,
2702 void *virt_addr, dma_addr_t dma_addr,
2703 struct dma_attrs *attrs)
2705 struct protection_domain *domain;
2706 unsigned long flags;
2709 INC_STATS_COUNTER(cnt_free_coherent);
2711 page = virt_to_page(virt_addr);
2712 size = PAGE_ALIGN(size);
2714 domain = get_domain(dev);
2718 spin_lock_irqsave(&domain->lock, flags);
2720 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2722 domain_flush_complete(domain);
2724 spin_unlock_irqrestore(&domain->lock, flags);
2727 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2728 __free_pages(page, get_order(size));
2732 * This function is called by the DMA layer to find out if we can handle a
2733 * particular device. It is part of the dma_ops.
2735 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2737 return check_device(dev);
2740 static struct dma_map_ops amd_iommu_dma_ops = {
2741 .alloc = alloc_coherent,
2742 .free = free_coherent,
2743 .map_page = map_page,
2744 .unmap_page = unmap_page,
2746 .unmap_sg = unmap_sg,
2747 .dma_supported = amd_iommu_dma_supported,
2750 int __init amd_iommu_init_api(void)
2752 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2755 int __init amd_iommu_init_dma_ops(void)
2757 swiotlb = iommu_pass_through ? 1 : 0;
2761 * In case we don't initialize SWIOTLB (actually the common case
2762 * when AMD IOMMU is enabled), make sure there are global
2763 * dma_ops set as a fall-back for devices not handled by this
2764 * driver (for example non-PCI devices).
2767 dma_ops = &nommu_dma_ops;
2769 amd_iommu_stats_init();
2771 if (amd_iommu_unmap_flush)
2772 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2774 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2779 /*****************************************************************************
2781 * The following functions belong to the exported interface of AMD IOMMU
2783 * This interface allows access to lower level functions of the IOMMU
2784 * like protection domain handling and assignement of devices to domains
2785 * which is not possible with the dma_ops interface.
2787 *****************************************************************************/
2789 static void cleanup_domain(struct protection_domain *domain)
2791 struct iommu_dev_data *entry;
2792 unsigned long flags;
2794 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2796 while (!list_empty(&domain->dev_list)) {
2797 entry = list_first_entry(&domain->dev_list,
2798 struct iommu_dev_data, list);
2799 __detach_device(entry);
2802 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2805 static void protection_domain_free(struct protection_domain *domain)
2810 del_domain_from_list(domain);
2813 domain_id_free(domain->id);
2818 static int protection_domain_init(struct protection_domain *domain)
2820 spin_lock_init(&domain->lock);
2821 mutex_init(&domain->api_lock);
2822 domain->id = domain_id_alloc();
2825 INIT_LIST_HEAD(&domain->dev_list);
2830 static struct protection_domain *protection_domain_alloc(void)
2832 struct protection_domain *domain;
2834 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2838 if (protection_domain_init(domain))
2841 add_domain_to_list(domain);
2851 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2853 struct protection_domain *pdomain;
2854 struct dma_ops_domain *dma_domain;
2857 case IOMMU_DOMAIN_UNMANAGED:
2858 pdomain = protection_domain_alloc();
2862 pdomain->mode = PAGE_MODE_3_LEVEL;
2863 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2864 if (!pdomain->pt_root) {
2865 protection_domain_free(pdomain);
2869 pdomain->domain.geometry.aperture_start = 0;
2870 pdomain->domain.geometry.aperture_end = ~0ULL;
2871 pdomain->domain.geometry.force_aperture = true;
2874 case IOMMU_DOMAIN_DMA:
2875 dma_domain = dma_ops_domain_alloc();
2877 pr_err("AMD-Vi: Failed to allocate\n");
2880 pdomain = &dma_domain->domain;
2882 case IOMMU_DOMAIN_IDENTITY:
2883 pdomain = protection_domain_alloc();
2887 pdomain->mode = PAGE_MODE_NONE;
2893 return &pdomain->domain;
2896 static void amd_iommu_domain_free(struct iommu_domain *dom)
2898 struct protection_domain *domain;
2903 domain = to_pdomain(dom);
2905 if (domain->dev_cnt > 0)
2906 cleanup_domain(domain);
2908 BUG_ON(domain->dev_cnt != 0);
2910 if (domain->mode != PAGE_MODE_NONE)
2911 free_pagetable(domain);
2913 if (domain->flags & PD_IOMMUV2_MASK)
2914 free_gcr3_table(domain);
2916 protection_domain_free(domain);
2919 static void amd_iommu_detach_device(struct iommu_domain *dom,
2922 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2923 struct amd_iommu *iommu;
2926 if (!check_device(dev))
2929 devid = get_device_id(dev);
2931 if (dev_data->domain != NULL)
2934 iommu = amd_iommu_rlookup_table[devid];
2938 iommu_completion_wait(iommu);
2941 static int amd_iommu_attach_device(struct iommu_domain *dom,
2944 struct protection_domain *domain = to_pdomain(dom);
2945 struct iommu_dev_data *dev_data;
2946 struct amd_iommu *iommu;
2949 if (!check_device(dev))
2952 dev_data = dev->archdata.iommu;
2954 iommu = amd_iommu_rlookup_table[dev_data->devid];
2958 if (dev_data->domain)
2961 ret = attach_device(dev, domain);
2963 iommu_completion_wait(iommu);
2968 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2969 phys_addr_t paddr, size_t page_size, int iommu_prot)
2971 struct protection_domain *domain = to_pdomain(dom);
2975 if (domain->mode == PAGE_MODE_NONE)
2978 if (iommu_prot & IOMMU_READ)
2979 prot |= IOMMU_PROT_IR;
2980 if (iommu_prot & IOMMU_WRITE)
2981 prot |= IOMMU_PROT_IW;
2983 mutex_lock(&domain->api_lock);
2984 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
2985 mutex_unlock(&domain->api_lock);
2990 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2993 struct protection_domain *domain = to_pdomain(dom);
2996 if (domain->mode == PAGE_MODE_NONE)
2999 mutex_lock(&domain->api_lock);
3000 unmap_size = iommu_unmap_page(domain, iova, page_size);
3001 mutex_unlock(&domain->api_lock);
3003 domain_flush_tlb_pde(domain);
3008 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3011 struct protection_domain *domain = to_pdomain(dom);
3012 unsigned long offset_mask, pte_pgsize;
3015 if (domain->mode == PAGE_MODE_NONE)
3018 pte = fetch_pte(domain, iova, &pte_pgsize);
3020 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3023 offset_mask = pte_pgsize - 1;
3024 __pte = *pte & PM_ADDR_MASK;
3026 return (__pte & ~offset_mask) | (iova & offset_mask);
3029 static bool amd_iommu_capable(enum iommu_cap cap)
3032 case IOMMU_CAP_CACHE_COHERENCY:
3034 case IOMMU_CAP_INTR_REMAP:
3035 return (irq_remapping_enabled == 1);
3036 case IOMMU_CAP_NOEXEC:
3043 static void amd_iommu_get_dm_regions(struct device *dev,
3044 struct list_head *head)
3046 struct unity_map_entry *entry;
3049 devid = get_device_id(dev);
3051 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3052 struct iommu_dm_region *region;
3054 if (devid < entry->devid_start || devid > entry->devid_end)
3057 region = kzalloc(sizeof(*region), GFP_KERNEL);
3059 pr_err("Out of memory allocating dm-regions for %s\n",
3064 region->start = entry->address_start;
3065 region->length = entry->address_end - entry->address_start;
3066 if (entry->prot & IOMMU_PROT_IR)
3067 region->prot |= IOMMU_READ;
3068 if (entry->prot & IOMMU_PROT_IW)
3069 region->prot |= IOMMU_WRITE;
3071 list_add_tail(®ion->list, head);
3075 static void amd_iommu_put_dm_regions(struct device *dev,
3076 struct list_head *head)
3078 struct iommu_dm_region *entry, *next;
3080 list_for_each_entry_safe(entry, next, head, list)
3084 static const struct iommu_ops amd_iommu_ops = {
3085 .capable = amd_iommu_capable,
3086 .domain_alloc = amd_iommu_domain_alloc,
3087 .domain_free = amd_iommu_domain_free,
3088 .attach_dev = amd_iommu_attach_device,
3089 .detach_dev = amd_iommu_detach_device,
3090 .map = amd_iommu_map,
3091 .unmap = amd_iommu_unmap,
3092 .map_sg = default_iommu_map_sg,
3093 .iova_to_phys = amd_iommu_iova_to_phys,
3094 .add_device = amd_iommu_add_device,
3095 .remove_device = amd_iommu_remove_device,
3096 .get_dm_regions = amd_iommu_get_dm_regions,
3097 .put_dm_regions = amd_iommu_put_dm_regions,
3098 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3101 /*****************************************************************************
3103 * The next functions do a basic initialization of IOMMU for pass through
3106 * In passthrough mode the IOMMU is initialized and enabled but not used for
3107 * DMA-API translation.
3109 *****************************************************************************/
3111 /* IOMMUv2 specific functions */
3112 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3114 return atomic_notifier_chain_register(&ppr_notifier, nb);
3116 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3118 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3120 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3122 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3124 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3126 struct protection_domain *domain = to_pdomain(dom);
3127 unsigned long flags;
3129 spin_lock_irqsave(&domain->lock, flags);
3131 /* Update data structure */
3132 domain->mode = PAGE_MODE_NONE;
3133 domain->updated = true;
3135 /* Make changes visible to IOMMUs */
3136 update_domain(domain);
3138 /* Page-table is not visible to IOMMU anymore, so free it */
3139 free_pagetable(domain);
3141 spin_unlock_irqrestore(&domain->lock, flags);
3143 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3145 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3147 struct protection_domain *domain = to_pdomain(dom);
3148 unsigned long flags;
3151 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3154 /* Number of GCR3 table levels required */
3155 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3158 if (levels > amd_iommu_max_glx_val)
3161 spin_lock_irqsave(&domain->lock, flags);
3164 * Save us all sanity checks whether devices already in the
3165 * domain support IOMMUv2. Just force that the domain has no
3166 * devices attached when it is switched into IOMMUv2 mode.
3169 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3173 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3174 if (domain->gcr3_tbl == NULL)
3177 domain->glx = levels;
3178 domain->flags |= PD_IOMMUV2_MASK;
3179 domain->updated = true;
3181 update_domain(domain);
3186 spin_unlock_irqrestore(&domain->lock, flags);
3190 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3192 static int __flush_pasid(struct protection_domain *domain, int pasid,
3193 u64 address, bool size)
3195 struct iommu_dev_data *dev_data;
3196 struct iommu_cmd cmd;
3199 if (!(domain->flags & PD_IOMMUV2_MASK))
3202 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3205 * IOMMU TLB needs to be flushed before Device TLB to
3206 * prevent device TLB refill from IOMMU TLB
3208 for (i = 0; i < amd_iommus_present; ++i) {
3209 if (domain->dev_iommu[i] == 0)
3212 ret = iommu_queue_command(amd_iommus[i], &cmd);
3217 /* Wait until IOMMU TLB flushes are complete */
3218 domain_flush_complete(domain);
3220 /* Now flush device TLBs */
3221 list_for_each_entry(dev_data, &domain->dev_list, list) {
3222 struct amd_iommu *iommu;
3226 There might be non-IOMMUv2 capable devices in an IOMMUv2
3229 if (!dev_data->ats.enabled)
3232 qdep = dev_data->ats.qdep;
3233 iommu = amd_iommu_rlookup_table[dev_data->devid];
3235 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3236 qdep, address, size);
3238 ret = iommu_queue_command(iommu, &cmd);
3243 /* Wait until all device TLBs are flushed */
3244 domain_flush_complete(domain);
3253 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3256 INC_STATS_COUNTER(invalidate_iotlb);
3258 return __flush_pasid(domain, pasid, address, false);
3261 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3264 struct protection_domain *domain = to_pdomain(dom);
3265 unsigned long flags;
3268 spin_lock_irqsave(&domain->lock, flags);
3269 ret = __amd_iommu_flush_page(domain, pasid, address);
3270 spin_unlock_irqrestore(&domain->lock, flags);
3274 EXPORT_SYMBOL(amd_iommu_flush_page);
3276 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3278 INC_STATS_COUNTER(invalidate_iotlb_all);
3280 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3284 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3286 struct protection_domain *domain = to_pdomain(dom);
3287 unsigned long flags;
3290 spin_lock_irqsave(&domain->lock, flags);
3291 ret = __amd_iommu_flush_tlb(domain, pasid);
3292 spin_unlock_irqrestore(&domain->lock, flags);
3296 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3298 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3305 index = (pasid >> (9 * level)) & 0x1ff;
3311 if (!(*pte & GCR3_VALID)) {
3315 root = (void *)get_zeroed_page(GFP_ATOMIC);
3319 *pte = __pa(root) | GCR3_VALID;
3322 root = __va(*pte & PAGE_MASK);
3330 static int __set_gcr3(struct protection_domain *domain, int pasid,
3335 if (domain->mode != PAGE_MODE_NONE)
3338 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3342 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3344 return __amd_iommu_flush_tlb(domain, pasid);
3347 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3351 if (domain->mode != PAGE_MODE_NONE)
3354 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3360 return __amd_iommu_flush_tlb(domain, pasid);
3363 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3366 struct protection_domain *domain = to_pdomain(dom);
3367 unsigned long flags;
3370 spin_lock_irqsave(&domain->lock, flags);
3371 ret = __set_gcr3(domain, pasid, cr3);
3372 spin_unlock_irqrestore(&domain->lock, flags);
3376 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3378 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3380 struct protection_domain *domain = to_pdomain(dom);
3381 unsigned long flags;
3384 spin_lock_irqsave(&domain->lock, flags);
3385 ret = __clear_gcr3(domain, pasid);
3386 spin_unlock_irqrestore(&domain->lock, flags);
3390 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3392 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3393 int status, int tag)
3395 struct iommu_dev_data *dev_data;
3396 struct amd_iommu *iommu;
3397 struct iommu_cmd cmd;
3399 INC_STATS_COUNTER(complete_ppr);
3401 dev_data = get_dev_data(&pdev->dev);
3402 iommu = amd_iommu_rlookup_table[dev_data->devid];
3404 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3405 tag, dev_data->pri_tlp);
3407 return iommu_queue_command(iommu, &cmd);
3409 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3411 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3413 struct protection_domain *pdomain;
3415 pdomain = get_domain(&pdev->dev);
3416 if (IS_ERR(pdomain))
3419 /* Only return IOMMUv2 domains */
3420 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3423 return &pdomain->domain;
3425 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3427 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3429 struct iommu_dev_data *dev_data;
3431 if (!amd_iommu_v2_supported())
3434 dev_data = get_dev_data(&pdev->dev);
3435 dev_data->errata |= (1 << erratum);
3437 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3439 int amd_iommu_device_info(struct pci_dev *pdev,
3440 struct amd_iommu_device_info *info)
3445 if (pdev == NULL || info == NULL)
3448 if (!amd_iommu_v2_supported())
3451 memset(info, 0, sizeof(*info));
3453 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3455 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3457 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3459 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3461 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3465 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3466 max_pasids = min(max_pasids, (1 << 20));
3468 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3469 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3471 features = pci_pasid_features(pdev);
3472 if (features & PCI_PASID_CAP_EXEC)
3473 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3474 if (features & PCI_PASID_CAP_PRIV)
3475 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3480 EXPORT_SYMBOL(amd_iommu_device_info);
3482 #ifdef CONFIG_IRQ_REMAP
3484 /*****************************************************************************
3486 * Interrupt Remapping Implementation
3488 *****************************************************************************/
3506 u16 devid; /* Device ID for IRTE table */
3507 u16 index; /* Index into IRTE table*/
3510 struct amd_ir_data {
3511 struct irq_2_irte irq_2_irte;
3512 union irte irte_entry;
3514 struct msi_msg msi_entry;
3518 static struct irq_chip amd_ir_chip;
3520 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3521 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3522 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3523 #define DTE_IRQ_REMAP_ENABLE 1ULL
3525 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3529 dte = amd_iommu_dev_table[devid].data[2];
3530 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3531 dte |= virt_to_phys(table->table);
3532 dte |= DTE_IRQ_REMAP_INTCTL;
3533 dte |= DTE_IRQ_TABLE_LEN;
3534 dte |= DTE_IRQ_REMAP_ENABLE;
3536 amd_iommu_dev_table[devid].data[2] = dte;
3539 #define IRTE_ALLOCATED (~1U)
3541 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3543 struct irq_remap_table *table = NULL;
3544 struct amd_iommu *iommu;
3545 unsigned long flags;
3548 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3550 iommu = amd_iommu_rlookup_table[devid];
3554 table = irq_lookup_table[devid];
3558 alias = amd_iommu_alias_table[devid];
3559 table = irq_lookup_table[alias];
3561 irq_lookup_table[devid] = table;
3562 set_dte_irq_entry(devid, table);
3563 iommu_flush_dte(iommu, devid);
3567 /* Nothing there yet, allocate new irq remapping table */
3568 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3572 /* Initialize table spin-lock */
3573 spin_lock_init(&table->lock);
3576 /* Keep the first 32 indexes free for IOAPIC interrupts */
3577 table->min_index = 32;
3579 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3580 if (!table->table) {
3586 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3591 for (i = 0; i < 32; ++i)
3592 table->table[i] = IRTE_ALLOCATED;
3595 irq_lookup_table[devid] = table;
3596 set_dte_irq_entry(devid, table);
3597 iommu_flush_dte(iommu, devid);
3598 if (devid != alias) {
3599 irq_lookup_table[alias] = table;
3600 set_dte_irq_entry(alias, table);
3601 iommu_flush_dte(iommu, alias);
3605 iommu_completion_wait(iommu);
3608 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3613 static int alloc_irq_index(u16 devid, int count)
3615 struct irq_remap_table *table;
3616 unsigned long flags;
3619 table = get_irq_table(devid, false);
3623 spin_lock_irqsave(&table->lock, flags);
3625 /* Scan table for free entries */
3626 for (c = 0, index = table->min_index;
3627 index < MAX_IRQS_PER_TABLE;
3629 if (table->table[index] == 0)
3636 table->table[index - c + 1] = IRTE_ALLOCATED;
3646 spin_unlock_irqrestore(&table->lock, flags);
3651 static int modify_irte(u16 devid, int index, union irte irte)
3653 struct irq_remap_table *table;
3654 struct amd_iommu *iommu;
3655 unsigned long flags;
3657 iommu = amd_iommu_rlookup_table[devid];
3661 table = get_irq_table(devid, false);
3665 spin_lock_irqsave(&table->lock, flags);
3666 table->table[index] = irte.val;
3667 spin_unlock_irqrestore(&table->lock, flags);
3669 iommu_flush_irt(iommu, devid);
3670 iommu_completion_wait(iommu);
3675 static void free_irte(u16 devid, int index)
3677 struct irq_remap_table *table;
3678 struct amd_iommu *iommu;
3679 unsigned long flags;
3681 iommu = amd_iommu_rlookup_table[devid];
3685 table = get_irq_table(devid, false);
3689 spin_lock_irqsave(&table->lock, flags);
3690 table->table[index] = 0;
3691 spin_unlock_irqrestore(&table->lock, flags);
3693 iommu_flush_irt(iommu, devid);
3694 iommu_completion_wait(iommu);
3697 static int get_devid(struct irq_alloc_info *info)
3701 switch (info->type) {
3702 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3703 devid = get_ioapic_devid(info->ioapic_id);
3705 case X86_IRQ_ALLOC_TYPE_HPET:
3706 devid = get_hpet_devid(info->hpet_id);
3708 case X86_IRQ_ALLOC_TYPE_MSI:
3709 case X86_IRQ_ALLOC_TYPE_MSIX:
3710 devid = get_device_id(&info->msi_dev->dev);
3720 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3722 struct amd_iommu *iommu;
3728 devid = get_devid(info);
3730 iommu = amd_iommu_rlookup_table[devid];
3732 return iommu->ir_domain;
3738 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3740 struct amd_iommu *iommu;
3746 switch (info->type) {
3747 case X86_IRQ_ALLOC_TYPE_MSI:
3748 case X86_IRQ_ALLOC_TYPE_MSIX:
3749 devid = get_device_id(&info->msi_dev->dev);
3751 iommu = amd_iommu_rlookup_table[devid];
3753 return iommu->msi_domain;
3763 struct irq_remap_ops amd_iommu_irq_ops = {
3764 .prepare = amd_iommu_prepare,
3765 .enable = amd_iommu_enable,
3766 .disable = amd_iommu_disable,
3767 .reenable = amd_iommu_reenable,
3768 .enable_faulting = amd_iommu_enable_faulting,
3769 .get_ir_irq_domain = get_ir_irq_domain,
3770 .get_irq_domain = get_irq_domain,
3773 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3774 struct irq_cfg *irq_cfg,
3775 struct irq_alloc_info *info,
3776 int devid, int index, int sub_handle)
3778 struct irq_2_irte *irte_info = &data->irq_2_irte;
3779 struct msi_msg *msg = &data->msi_entry;
3780 union irte *irte = &data->irte_entry;
3781 struct IO_APIC_route_entry *entry;
3783 data->irq_2_irte.devid = devid;
3784 data->irq_2_irte.index = index + sub_handle;
3786 /* Setup IRTE for IOMMU */
3788 irte->fields.vector = irq_cfg->vector;
3789 irte->fields.int_type = apic->irq_delivery_mode;
3790 irte->fields.destination = irq_cfg->dest_apicid;
3791 irte->fields.dm = apic->irq_dest_mode;
3792 irte->fields.valid = 1;
3794 switch (info->type) {
3795 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3796 /* Setup IOAPIC entry */
3797 entry = info->ioapic_entry;
3798 info->ioapic_entry = NULL;
3799 memset(entry, 0, sizeof(*entry));
3800 entry->vector = index;
3802 entry->trigger = info->ioapic_trigger;
3803 entry->polarity = info->ioapic_polarity;
3804 /* Mask level triggered irqs. */
3805 if (info->ioapic_trigger)
3809 case X86_IRQ_ALLOC_TYPE_HPET:
3810 case X86_IRQ_ALLOC_TYPE_MSI:
3811 case X86_IRQ_ALLOC_TYPE_MSIX:
3812 msg->address_hi = MSI_ADDR_BASE_HI;
3813 msg->address_lo = MSI_ADDR_BASE_LO;
3814 msg->data = irte_info->index;
3823 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3824 unsigned int nr_irqs, void *arg)
3826 struct irq_alloc_info *info = arg;
3827 struct irq_data *irq_data;
3828 struct amd_ir_data *data;
3829 struct irq_cfg *cfg;
3835 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3836 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3840 * With IRQ remapping enabled, don't need contiguous CPU vectors
3841 * to support multiple MSI interrupts.
3843 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3844 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3846 devid = get_devid(info);
3850 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3854 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3855 if (get_irq_table(devid, true))
3856 index = info->ioapic_pin;
3860 index = alloc_irq_index(devid, nr_irqs);
3863 pr_warn("Failed to allocate IRTE\n");
3864 goto out_free_parent;
3867 for (i = 0; i < nr_irqs; i++) {
3868 irq_data = irq_domain_get_irq_data(domain, virq + i);
3869 cfg = irqd_cfg(irq_data);
3870 if (!irq_data || !cfg) {
3876 data = kzalloc(sizeof(*data), GFP_KERNEL);
3880 irq_data->hwirq = (devid << 16) + i;
3881 irq_data->chip_data = data;
3882 irq_data->chip = &amd_ir_chip;
3883 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3884 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3890 for (i--; i >= 0; i--) {
3891 irq_data = irq_domain_get_irq_data(domain, virq + i);
3893 kfree(irq_data->chip_data);
3895 for (i = 0; i < nr_irqs; i++)
3896 free_irte(devid, index + i);
3898 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3902 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3903 unsigned int nr_irqs)
3905 struct irq_2_irte *irte_info;
3906 struct irq_data *irq_data;
3907 struct amd_ir_data *data;
3910 for (i = 0; i < nr_irqs; i++) {
3911 irq_data = irq_domain_get_irq_data(domain, virq + i);
3912 if (irq_data && irq_data->chip_data) {
3913 data = irq_data->chip_data;
3914 irte_info = &data->irq_2_irte;
3915 free_irte(irte_info->devid, irte_info->index);
3919 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3922 static void irq_remapping_activate(struct irq_domain *domain,
3923 struct irq_data *irq_data)
3925 struct amd_ir_data *data = irq_data->chip_data;
3926 struct irq_2_irte *irte_info = &data->irq_2_irte;
3928 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3931 static void irq_remapping_deactivate(struct irq_domain *domain,
3932 struct irq_data *irq_data)
3934 struct amd_ir_data *data = irq_data->chip_data;
3935 struct irq_2_irte *irte_info = &data->irq_2_irte;
3939 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
3942 static struct irq_domain_ops amd_ir_domain_ops = {
3943 .alloc = irq_remapping_alloc,
3944 .free = irq_remapping_free,
3945 .activate = irq_remapping_activate,
3946 .deactivate = irq_remapping_deactivate,
3949 static int amd_ir_set_affinity(struct irq_data *data,
3950 const struct cpumask *mask, bool force)
3952 struct amd_ir_data *ir_data = data->chip_data;
3953 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3954 struct irq_cfg *cfg = irqd_cfg(data);
3955 struct irq_data *parent = data->parent_data;
3958 ret = parent->chip->irq_set_affinity(parent, mask, force);
3959 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3963 * Atomically updates the IRTE with the new destination, vector
3964 * and flushes the interrupt entry cache.
3966 ir_data->irte_entry.fields.vector = cfg->vector;
3967 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
3968 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
3971 * After this point, all the interrupts will start arriving
3972 * at the new destination. So, time to cleanup the previous
3973 * vector allocation.
3975 send_cleanup_vector(cfg);
3977 return IRQ_SET_MASK_OK_DONE;
3980 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3982 struct amd_ir_data *ir_data = irq_data->chip_data;
3984 *msg = ir_data->msi_entry;
3987 static struct irq_chip amd_ir_chip = {
3988 .irq_ack = ir_ack_apic_edge,
3989 .irq_set_affinity = amd_ir_set_affinity,
3990 .irq_compose_msi_msg = ir_compose_msi_msg,
3993 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3995 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
3996 if (!iommu->ir_domain)
3999 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4000 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);